1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright 2008 Sun Microsystems, Inc. All rights reserved.
30 * Use is subject to license terms.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * Implementation of cl_io for VVP layer.
38 * Author: Nikita Danilov <nikita.danilov@sun.com>
41 #define DEBUG_SUBSYSTEM S_LLITE
44 # error This file is kernel only.
48 #include <lustre_lite.h>
50 #include "vvp_internal.h"
52 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
53 const struct cl_io_slice *slice);
56 * True, if \a io is a normal io, False for sendfile() / splice_{read|write}
58 int cl_is_normalio(const struct lu_env *env, const struct cl_io *io)
60 struct vvp_io *vio = vvp_env_io(env);
62 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
64 return vio->cui_io_subtype == IO_NORMAL;
67 /*****************************************************************************
73 static int vvp_io_fault_iter_init(const struct lu_env *env,
74 const struct cl_io_slice *ios)
76 struct vvp_io *vio = cl2vvp_io(env, ios);
77 struct inode *inode = ccc_object_inode(ios->cis_obj);
80 cl2ccc_io(env, ios)->cui_fd->fd_file->f_dentry->d_inode);
81 vio->u.fault.ft_mtime = LTIME_S(inode->i_mtime);
85 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
87 struct cl_io *io = ios->cis_io;
88 struct cl_object *obj = io->ci_obj;
90 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
91 if (io->ci_type == CIT_READ) {
92 struct vvp_io *vio = cl2vvp_io(env, ios);
93 struct ccc_io *cio = cl2ccc_io(env, ios);
95 if (vio->cui_ra_window_set)
96 ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead);
101 static void vvp_io_fault_fini(const struct lu_env *env,
102 const struct cl_io_slice *ios)
104 struct cl_io *io = ios->cis_io;
105 struct cl_page *page = io->u.ci_fault.ft_page;
107 CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj));
110 lu_ref_del(&page->cp_reference, "fault", io);
111 cl_page_put(env, page);
112 io->u.ci_fault.ft_page = NULL;
114 vvp_io_fini(env, ios);
117 enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
120 * we only want to hold PW locks if the mmap() can generate
121 * writes back to the file and that only happens in shared
124 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
129 static int vvp_mmap_locks(const struct lu_env *env,
130 struct ccc_io *vio, struct cl_io *io)
132 struct ccc_thread_info *cti = ccc_env_info(env);
133 struct vm_area_struct *vma;
134 struct cl_lock_descr *descr = &cti->cti_descr;
135 ldlm_policy_data_t policy;
142 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
144 if (!cl_is_normalio(env, io))
147 for (seg = 0; seg < vio->cui_nrsegs; seg++) {
148 const struct iovec *iv = &vio->cui_iov[seg];
150 addr = (unsigned long)iv->iov_base;
155 count += addr & (~CFS_PAGE_MASK);
156 addr &= CFS_PAGE_MASK;
157 while((vma = our_vma(addr, count)) != NULL) {
158 struct inode *inode = vma->vm_file->f_dentry->d_inode;
159 int flags = CEF_MUST;
161 if (ll_file_nolock(vma->vm_file)) {
163 * For no lock case, a lockless lock will be
170 * XXX: Required lock mode can be weakened: CIT_WRITE
171 * io only ever reads user level buffer, and CIT_READ
174 policy_from_vma(&policy, vma, addr, count);
175 descr->cld_mode = vvp_mode_from_vma(vma);
176 descr->cld_obj = ll_i2info(inode)->lli_clob;
177 descr->cld_start = cl_index(descr->cld_obj,
178 policy.l_extent.start);
179 descr->cld_end = cl_index(descr->cld_obj,
180 policy.l_extent.end);
181 descr->cld_enq_flags = flags;
182 result = cl_io_lock_alloc_add(env, io, descr);
186 if (vma->vm_end - addr >= count)
189 count -= vma->vm_end - addr;
196 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
197 enum cl_lock_mode mode, loff_t start, loff_t end)
199 struct ccc_io *cio = ccc_env_io(env);
203 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
206 ccc_io_update_iov(env, cio, io);
208 if (io->u.ci_rw.crw_nonblock)
209 ast_flags |= CEF_NONBLOCK;
210 result = vvp_mmap_locks(env, cio, io);
212 result = ccc_io_one_lock(env, io, ast_flags, mode, start, end);
216 static int vvp_io_read_lock(const struct lu_env *env,
217 const struct cl_io_slice *ios)
219 struct cl_io *io = ios->cis_io;
220 struct ll_inode_info *lli = ll_i2info(ccc_object_inode(io->ci_obj));
224 /* XXX: Layer violation, we shouldn't see lsm at llite level. */
225 if (lli->lli_smd != NULL) /* lsm-less file, don't need to lock */
226 result = vvp_io_rw_lock(env, io, CLM_READ,
227 io->u.ci_rd.rd.crw_pos,
228 io->u.ci_rd.rd.crw_pos +
229 io->u.ci_rd.rd.crw_count - 1);
235 static int vvp_io_fault_lock(const struct lu_env *env,
236 const struct cl_io_slice *ios)
238 struct cl_io *io = ios->cis_io;
239 struct vvp_io *vio = cl2vvp_io(env, ios);
241 * XXX LDLM_FL_CBPENDING
243 return ccc_io_one_lock_index
244 (env, io, 0, vvp_mode_from_vma(vio->u.fault.ft_vma),
245 io->u.ci_fault.ft_index, io->u.ci_fault.ft_index);
248 static int vvp_io_write_lock(const struct lu_env *env,
249 const struct cl_io_slice *ios)
251 struct cl_io *io = ios->cis_io;
255 if (io->u.ci_wr.wr_append) {
257 end = OBD_OBJECT_EOF;
259 start = io->u.ci_wr.wr.crw_pos;
260 end = start + io->u.ci_wr.wr.crw_count - 1;
262 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
265 static int vvp_io_trunc_iter_init(const struct lu_env *env,
266 const struct cl_io_slice *ios)
268 struct ccc_io *vio = cl2ccc_io(env, ios);
269 struct inode *inode = ccc_object_inode(ios->cis_obj);
272 * We really need to get our PW lock before we change inode->i_size.
273 * If we don't we can race with other i_size updaters on our node,
274 * like ll_file_read. We can also race with i_size propogation to
275 * other nodes through dirtying and writeback of final cached pages.
276 * This last one is especially bad for racing o_append users on other
280 UNLOCK_INODE_MUTEX(inode);
281 UP_WRITE_I_ALLOC_SEM(inode);
282 vio->u.trunc.cui_locks_released = 1;
287 * Implementation of cl_io_operations::cio_lock() method for CIT_TRUNC io.
289 * Handles "lockless io" mode when extent locking is done by server.
291 static int vvp_io_trunc_lock(const struct lu_env *env,
292 const struct cl_io_slice *ios)
294 struct ccc_io *vio = cl2ccc_io(env, ios);
295 struct cl_io *io = ios->cis_io;
296 size_t new_size = io->u.ci_truncate.tr_size;
297 __u32 enqflags = new_size == 0 ? CEF_DISCARD_DATA : 0;
300 vio->u.trunc.cui_local_lock = TRUNC_EXTENT;
301 result = ccc_io_one_lock(env, io, enqflags, CLM_WRITE,
302 new_size, OBD_OBJECT_EOF);
306 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
310 * Only ll_inode_size_lock is taken at this level. lov_stripe_lock()
311 * is grabbed by ll_truncate() only over call to obd_adjust_kms(). If
312 * vmtruncate returns 0, then ll_truncate dropped ll_inode_size_lock()
314 ll_inode_size_lock(inode, 0);
315 result = vmtruncate(inode, size);
317 ll_inode_size_unlock(inode, 0);
322 static int vvp_io_trunc_start(const struct lu_env *env,
323 const struct cl_io_slice *ios)
325 struct ccc_io *cio = cl2ccc_io(env, ios);
326 struct vvp_io *vio = cl2vvp_io(env, ios);
327 struct cl_io *io = ios->cis_io;
328 struct inode *inode = ccc_object_inode(io->ci_obj);
329 struct cl_object *obj = ios->cis_obj;
330 size_t size = io->u.ci_truncate.tr_size;
331 pgoff_t start = cl_index(obj, size);
334 LASSERT(cio->u.trunc.cui_locks_released);
336 LOCK_INODE_MUTEX(inode);
337 DOWN_WRITE_I_ALLOC_SEM(inode);
338 cio->u.trunc.cui_locks_released = 0;
340 result = vvp_do_vmtruncate(inode, size);
343 * If a page is partially truncated, keep it owned across truncate to
346 * XXX this properly belongs to osc, because races in question are OST
349 if (cl_offset(obj, start) != size) {
350 struct cl_object_header *hdr;
352 hdr = cl_object_header(obj);
353 spin_lock(&hdr->coh_page_guard);
354 vio->cui_partpage = cl_page_lookup(hdr, start);
355 spin_unlock(&hdr->coh_page_guard);
357 if (vio->cui_partpage != NULL)
359 * Wait for the transfer completion for a partially
360 * truncated page to avoid dead-locking an OST with
361 * the concurrent page-wise overlapping WRITE and
362 * PUNCH requests. BUG:17397.
364 * Partial page is disowned in vvp_io_trunc_end().
366 cl_page_own(env, io, vio->cui_partpage);
368 vio->cui_partpage = NULL;
372 static void vvp_io_trunc_end(const struct lu_env *env,
373 const struct cl_io_slice *ios)
375 struct vvp_io *vio = cl2vvp_io(env, ios);
376 struct cl_io *io = ios->cis_io;
377 struct inode *inode = ccc_object_inode(io->ci_obj);
378 size_t size = io->u.ci_truncate.tr_size;
380 if (vio->cui_partpage != NULL) {
381 cl_page_disown(env, ios->cis_io, vio->cui_partpage);
382 cl_page_put(env, vio->cui_partpage);
383 vio->cui_partpage = NULL;
387 * Do vmtruncate again, to remove possible stale pages populated by
388 * competing read threads. bz20645.
390 LASSERT(size == i_size_read(inode));
391 vvp_do_vmtruncate(inode, size);
394 static void vvp_io_trunc_fini(const struct lu_env *env,
395 const struct cl_io_slice *ios)
397 struct ccc_io *cio = ccc_env_io(env);
398 struct inode *inode = ccc_object_inode(ios->cis_io->ci_obj);
400 if (cio->u.trunc.cui_locks_released) {
401 LOCK_INODE_MUTEX(inode);
402 DOWN_WRITE_I_ALLOC_SEM(inode);
403 cio->u.trunc.cui_locks_released = 0;
405 vvp_io_fini(env, ios);
408 #ifdef HAVE_FILE_READV
409 static ssize_t lustre_generic_file_read(struct file *file,
410 struct ccc_io *vio, loff_t *ppos)
412 return generic_file_readv(file, vio->cui_iov, vio->cui_nrsegs, ppos);
415 static ssize_t lustre_generic_file_write(struct file *file,
416 struct ccc_io *vio, loff_t *ppos)
418 return generic_file_writev(file, vio->cui_iov, vio->cui_nrsegs, ppos);
421 static ssize_t lustre_generic_file_read(struct file *file,
422 struct ccc_io *vio, loff_t *ppos)
424 return generic_file_aio_read(vio->cui_iocb, vio->cui_iov,
425 vio->cui_nrsegs, *ppos);
428 static ssize_t lustre_generic_file_write(struct file *file,
429 struct ccc_io *vio, loff_t *ppos)
431 return generic_file_aio_write(vio->cui_iocb, vio->cui_iov,
432 vio->cui_nrsegs, *ppos);
436 static int vvp_io_read_start(const struct lu_env *env,
437 const struct cl_io_slice *ios)
439 struct vvp_io *vio = cl2vvp_io(env, ios);
440 struct ccc_io *cio = cl2ccc_io(env, ios);
441 struct cl_io *io = ios->cis_io;
442 struct cl_object *obj = io->ci_obj;
443 struct inode *inode = ccc_object_inode(obj);
444 struct ll_ra_read *bead = &vio->cui_bead;
445 struct file *file = cio->cui_fd->fd_file;
448 loff_t pos = io->u.ci_rd.rd.crw_pos;
449 long cnt = io->u.ci_rd.rd.crw_count;
450 long tot = cio->cui_tot_count;
453 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
455 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
457 result = ccc_prep_size(env, obj, io, pos, tot, 1, &exceed);
460 else if (exceed != 0)
463 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
464 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
465 inode->i_ino, cnt, pos, i_size_read(inode));
467 /* turn off the kernel's read-ahead */
468 cio->cui_fd->fd_file->f_ra.ra_pages = 0;
470 /* initialize read-ahead window once per syscall */
471 if (!vio->cui_ra_window_set) {
472 vio->cui_ra_window_set = 1;
473 bead->lrr_start = cl_index(obj, pos);
475 * XXX: explicit CFS_PAGE_SIZE
477 bead->lrr_count = cl_index(obj, tot + CFS_PAGE_SIZE - 1);
478 ll_ra_read_in(file, bead);
483 switch (vio->cui_io_subtype) {
485 result = lustre_generic_file_read(file, cio, &pos);
487 #ifdef HAVE_KERNEL_SENDFILE
489 result = generic_file_sendfile(file, &pos, cnt,
490 vio->u.sendfile.cui_actor,
491 vio->u.sendfile.cui_target);
494 #ifdef HAVE_KERNEL_SPLICE_READ
496 result = generic_file_splice_read(file, &pos,
497 vio->u.splice.cui_pipe, cnt,
498 vio->u.splice.cui_flags);
502 CERROR("Wrong IO type %u\n", vio->cui_io_subtype);
510 io->ci_nob += result;
511 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
512 cio->cui_fd, pos, result, 0);
518 static int vvp_io_write_start(const struct lu_env *env,
519 const struct cl_io_slice *ios)
521 struct ccc_io *cio = cl2ccc_io(env, ios);
522 struct cl_io *io = ios->cis_io;
523 struct cl_object *obj = io->ci_obj;
524 struct inode *inode = ccc_object_inode(obj);
525 struct file *file = cio->cui_fd->fd_file;
527 loff_t pos = io->u.ci_wr.wr.crw_pos;
528 size_t cnt = io->u.ci_wr.wr.crw_count;
532 if (cl_io_is_append(io)) {
534 * PARALLEL IO This has to be changed for parallel IO doing
535 * out-of-order writes.
537 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
538 #ifndef HAVE_FILE_WRITEV
539 cio->cui_iocb->ki_pos = pos;
543 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
545 if (cio->cui_iov == NULL) /* from a temp io in ll_cl_init(). */
548 result = lustre_generic_file_write(file, cio, &pos);
553 io->ci_nob += result;
554 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
555 cio->cui_fd, pos, result, 0);
561 #ifndef HAVE_VM_OP_FAULT
562 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
566 vmpage = filemap_nopage(cfio->ft_vma, cfio->nopage.ft_address,
567 cfio->nopage.ft_type);
569 if (vmpage == NOPAGE_SIGBUS) {
570 CDEBUG(D_PAGE, "got addr %lu type %lx - SIGBUS\n",
571 cfio->nopage.ft_address,(long)cfio->nopage.ft_type);
573 } else if (vmpage == NOPAGE_OOM) {
574 CDEBUG(D_PAGE, "got addr %lu type %lx - OOM\n",
575 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
579 LL_CDEBUG_PAGE(D_PAGE, vmpage, "got addr %lu type %lx\n",
580 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
582 cfio->ft_vmpage = vmpage;
587 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
589 cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, cfio->fault.ft_vmf);
591 if (cfio->fault.ft_vmf->page) {
592 LL_CDEBUG_PAGE(D_PAGE, cfio->fault.ft_vmf->page,
593 "got addr %p type NOPAGE\n",
594 cfio->fault.ft_vmf->virtual_address);
595 /*XXX workaround to bug in CLIO - he deadlocked with
596 lock cancel if page locked */
597 if (likely(cfio->fault.ft_flags & VM_FAULT_LOCKED)) {
598 unlock_page(cfio->fault.ft_vmf->page);
599 cfio->fault.ft_flags &= ~VM_FAULT_LOCKED;
602 cfio->ft_vmpage = cfio->fault.ft_vmf->page;
606 if (unlikely (cfio->fault.ft_flags & VM_FAULT_ERROR)) {
607 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n",
608 cfio->fault.ft_vmf->virtual_address);
612 if (unlikely (cfio->fault.ft_flags & VM_FAULT_NOPAGE)) {
613 CDEBUG(D_PAGE, "got addr %p - OOM\n",
614 cfio->fault.ft_vmf->virtual_address);
618 CERROR("unknow error in page fault!\n");
624 static int vvp_io_fault_start(const struct lu_env *env,
625 const struct cl_io_slice *ios)
627 struct vvp_io *vio = cl2vvp_io(env, ios);
628 struct cl_io *io = ios->cis_io;
629 struct cl_object *obj = io->ci_obj;
630 struct inode *inode = ccc_object_inode(obj);
631 struct cl_fault_io *fio = &io->u.ci_fault;
632 struct vvp_fault_io *cfio = &vio->u.fault;
634 int kernel_result = 0;
636 struct cl_page *page;
638 pgoff_t last; /* last page in a file data region */
640 if (fio->ft_executable &&
641 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
643 " changed while waiting for the page fault lock\n",
644 PFID(lu_object_fid(&obj->co_lu)));
646 /* offset of the last byte on the page */
647 offset = cl_offset(obj, fio->ft_index + 1) - 1;
648 LASSERT(cl_index(obj, offset) == fio->ft_index);
649 result = ccc_prep_size(env, obj, io, 0, offset + 1, 0, NULL);
653 /* must return locked page */
654 kernel_result = vvp_io_kernel_fault(cfio);
655 if (kernel_result != 0)
656 return kernel_result;
657 /* Temporarily lock vmpage to keep cl_page_find() happy. */
658 lock_page(cfio->ft_vmpage);
659 page = cl_page_find(env, obj, fio->ft_index, cfio->ft_vmpage,
661 unlock_page(cfio->ft_vmpage);
663 page_cache_release(cfio->ft_vmpage);
664 cfio->ft_vmpage = NULL;
665 return PTR_ERR(page);
668 size = i_size_read(inode);
669 last = cl_index(obj, size - 1);
670 if (fio->ft_index == last)
672 * Last page is mapped partially.
674 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
676 fio->ft_nob = cl_page_size(obj);
678 lu_ref_add(&page->cp_reference, "fault", io);
681 * Certain 2.6 kernels return not-NULL from
682 * filemap_nopage() when page is beyond the file size,
683 * on the grounds that "An external ptracer can access
684 * pages that normally aren't accessible.." Don't
685 * propagate such page fault to the lower layers to
686 * avoid side-effects like KMS updates.
688 if (fio->ft_index > last)
694 static int vvp_io_read_page(const struct lu_env *env,
695 const struct cl_io_slice *ios,
696 const struct cl_page_slice *slice)
698 struct cl_io *io = ios->cis_io;
699 struct cl_object *obj = slice->cpl_obj;
700 struct ccc_page *cp = cl2ccc_page(slice);
701 struct cl_page *page = slice->cpl_page;
702 struct inode *inode = ccc_object_inode(obj);
703 struct ll_sb_info *sbi = ll_i2sbi(inode);
704 struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd;
705 struct ll_readahead_state *ras = &fd->fd_ras;
706 cfs_page_t *vmpage = cp->cpg_page;
707 struct cl_2queue *queue = &io->ci_queue;
710 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
711 LASSERT(slice->cpl_obj == obj);
715 if (sbi->ll_ra_info.ra_max_pages_per_file)
716 ras_update(sbi, inode, ras, page->cp_index,
717 cp->cpg_defer_uptodate);
719 /* Sanity check whether the page is protected by a lock. */
720 rc = cl_page_is_under_lock(env, io, page);
722 CL_PAGE_HEADER(D_WARNING, env, page, "%s: %i\n",
723 rc == -ENODATA ? "without a lock" :
729 if (cp->cpg_defer_uptodate) {
731 cl_page_export(env, page, 1);
734 * Add page into the queue even when it is marked uptodate above.
735 * this will unlock it automatically as part of cl_page_list_disown().
737 cl_2queue_add(queue, page);
738 if (sbi->ll_ra_info.ra_max_pages_per_file)
739 ll_readahead(env, io, ras,
740 vmpage->mapping, &queue->c2_qin, fd->fd_flags);
745 static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io,
746 struct cl_page *page, struct ccc_page *cp,
747 int to, enum cl_req_type crt)
749 struct cl_2queue *queue;
752 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
754 queue = &io->ci_queue;
756 cl_2queue_init_page(queue, page);
757 cl_page_clip(env, page, 0, to);
759 result = cl_io_submit_sync(env, io, crt, queue, CRP_NORMAL, 0);
760 LASSERT(cl_page_is_owned(page, io));
761 cl_page_clip(env, page, 0, CFS_PAGE_SIZE);
765 * in CRT_WRITE case page is left locked even in case of
768 cl_page_list_disown(env, io, &queue->c2_qin);
769 cl_2queue_fini(env, queue);
775 * Prepare partially written-to page for a write.
777 static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io,
778 struct cl_object *obj, struct cl_page *pg,
780 unsigned from, unsigned to)
782 struct cl_attr *attr = ccc_env_thread_attr(env);
783 loff_t offset = cl_offset(obj, pg->cp_index);
786 cl_object_attr_lock(obj);
787 result = cl_object_attr_get(env, obj, attr);
788 cl_object_attr_unlock(obj);
791 * If are writing to a new page, no need to read old data.
792 * The extent locking will have updated the KMS, and for our
793 * purposes here we can treat it like i_size.
795 if (attr->cat_kms <= offset) {
796 char *kaddr = kmap_atomic(cp->cpg_page, KM_USER0);
798 memset(kaddr, 0, cl_page_size(obj));
799 kunmap_atomic(kaddr, KM_USER0);
800 } else if (cp->cpg_defer_uptodate)
803 result = vvp_page_sync_io(env, io, pg, cp,
804 CFS_PAGE_SIZE, CRT_READ);
806 * In older implementations, obdo_refresh_inode is called here
807 * to update the inode because the write might modify the
808 * object info at OST. However, this has been proven useless,
809 * since LVB functions will be called when user space program
810 * tries to retrieve inode attribute. Also, see bug 15909 for
814 cl_page_export(env, pg, 1);
819 static int vvp_io_prepare_write(const struct lu_env *env,
820 const struct cl_io_slice *ios,
821 const struct cl_page_slice *slice,
822 unsigned from, unsigned to)
824 struct cl_object *obj = slice->cpl_obj;
825 struct ccc_page *cp = cl2ccc_page(slice);
826 struct cl_page *pg = slice->cpl_page;
827 cfs_page_t *vmpage = cp->cpg_page;
833 LINVRNT(cl_page_is_vmlocked(env, pg));
834 LASSERT(vmpage->mapping->host == ccc_object_inode(obj));
838 CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to);
839 if (!PageUptodate(vmpage)) {
841 * We're completely overwriting an existing page, so _don't_
842 * set it up to date until commit_write
844 if (from == 0 && to == CFS_PAGE_SIZE) {
845 CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n");
846 POISON_PAGE(page, 0x11);
848 result = vvp_io_prepare_partial(env, ios->cis_io, obj,
851 CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n");
855 static int vvp_io_commit_write(const struct lu_env *env,
856 const struct cl_io_slice *ios,
857 const struct cl_page_slice *slice,
858 unsigned from, unsigned to)
860 struct cl_object *obj = slice->cpl_obj;
861 struct cl_io *io = ios->cis_io;
862 struct ccc_page *cp = cl2ccc_page(slice);
863 struct cl_page *pg = slice->cpl_page;
864 struct inode *inode = ccc_object_inode(obj);
865 struct ll_sb_info *sbi = ll_i2sbi(inode);
866 cfs_page_t *vmpage = cp->cpg_page;
874 LINVRNT(cl_page_is_vmlocked(env, pg));
875 LASSERT(vmpage->mapping->host == inode);
877 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "commiting page write\n");
878 CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to);
881 * queue a write for some time in the future the first time we
884 * This is different from what other file systems do: they usually
885 * just mark page (and some of its buffers) dirty and rely on
886 * balance_dirty_pages() to start a write-back. Lustre wants write-back
887 * to be started earlier for the following reasons:
889 * (1) with a large number of clients we need to limit the amount
890 * of cached data on the clients a lot;
892 * (2) large compute jobs generally want compute-only then io-only
893 * and the IO should complete as quickly as possible;
895 * (3) IO is batched up to the RPC size and is async until the
896 * client max cache is hit
897 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
900 if (!PageDirty(vmpage)) {
901 tallyop = LPROC_LL_DIRTY_MISSES;
902 vvp_write_pending(cl2ccc(obj), cp);
903 set_page_dirty(vmpage);
904 /* ll_set_page_dirty() does the same for now, but
905 * it will not soon. */
906 vvp_write_pending(cl2ccc(obj), cp);
907 result = cl_page_cache_add(env, io, pg, CRT_WRITE);
908 if (result == -EDQUOT)
910 * Client ran out of disk space grant. Possible
913 * (a) do a sync write, renewing grant;
915 * (b) stop writing on this stripe, switch to the
918 * (b) is a part of "parallel io" design that is the
919 * ultimate goal. (a) is what "old" client did, and
920 * what the new code continues to do for the time
923 result = vvp_page_sync_io(env, io, pg, cp,
926 CERROR("Write page %lu of inode %p failed %d\n",
927 pg->cp_index, inode, result);
929 tallyop = LPROC_LL_DIRTY_HITS;
932 ll_stats_ops_tally(sbi, tallyop, 1);
934 size = cl_offset(obj, pg->cp_index) + to;
937 if (size > i_size_read(inode))
938 i_size_write(inode, size);
939 cl_page_export(env, pg, 1);
940 } else if (size > i_size_read(inode))
941 cl_page_discard(env, io, pg);
945 static const struct cl_io_operations vvp_io_ops = {
948 .cio_fini = vvp_io_fini,
949 .cio_lock = vvp_io_read_lock,
950 .cio_start = vvp_io_read_start,
951 .cio_advance = ccc_io_advance
954 .cio_fini = vvp_io_fini,
955 .cio_lock = vvp_io_write_lock,
956 .cio_start = vvp_io_write_start,
957 .cio_advance = ccc_io_advance
960 .cio_fini = vvp_io_trunc_fini,
961 .cio_iter_init = vvp_io_trunc_iter_init,
962 .cio_lock = vvp_io_trunc_lock,
963 .cio_start = vvp_io_trunc_start,
964 .cio_end = vvp_io_trunc_end
967 .cio_fini = vvp_io_fault_fini,
968 .cio_iter_init = vvp_io_fault_iter_init,
969 .cio_lock = vvp_io_fault_lock,
970 .cio_start = vvp_io_fault_start,
971 .cio_end = ccc_io_end
974 .cio_fini = vvp_io_fini
977 .cio_read_page = vvp_io_read_page,
978 .cio_prepare_write = vvp_io_prepare_write,
979 .cio_commit_write = vvp_io_commit_write
982 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
985 struct vvp_io *vio = vvp_env_io(env);
986 struct ccc_io *cio = ccc_env_io(env);
987 struct inode *inode = ccc_object_inode(obj);
988 struct ll_sb_info *sbi = ll_i2sbi(inode);
991 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
994 CL_IO_SLICE_CLEAN(cio, cui_cl);
995 cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops);
996 vio->cui_ra_window_set = 0;
998 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1002 count = io->u.ci_rw.crw_count;
1003 op = io->ci_type == CIT_READ ?
1004 LPROC_LL_READ_BYTES : LPROC_LL_WRITE_BYTES;
1005 /* "If nbyte is 0, read() will return 0 and have no other
1006 * results." -- Single Unix Spec */
1010 cio->cui_tot_count = count;
1011 cio->cui_tot_nrsegs = 0;
1012 ll_stats_ops_tally(sbi, op, count);
1014 } else if (io->ci_type == CIT_TRUNC) {
1015 /* lockless truncate? */
1016 ll_stats_ops_tally(sbi, LPROC_LL_TRUNC, 1);
1021 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
1022 const struct cl_io_slice *slice)
1024 /* Caling just for assertion */
1025 cl2ccc_io(env, slice);
1026 return vvp_env_io(env);