4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Intel Corporation.
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>
39 * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
42 #define DEBUG_SUBSYSTEM S_LLITE
46 #include <lustre_lite.h>
48 #include "vvp_internal.h"
50 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
51 const struct cl_io_slice *slice);
54 * True, if \a io is a normal io, False for sendfile() / splice_{read|write}
56 int cl_is_normalio(const struct lu_env *env, const struct cl_io *io)
58 struct vvp_io *vio = vvp_env_io(env);
60 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
62 return vio->cui_io_subtype == IO_NORMAL;
66 * For swapping layout. The file's layout may have changed.
67 * To avoid populating pages to a wrong stripe, we have to verify the
68 * correctness of layout. It works because swapping layout processes
69 * have to acquire group lock.
71 static bool can_populate_pages(const struct lu_env *env, struct cl_io *io,
74 struct ll_inode_info *lli = ll_i2info(inode);
75 struct ccc_io *cio = ccc_env_io(env);
78 switch (io->ci_type) {
81 /* don't need lock here to check lli_layout_gen as we have held
82 * extent lock and GROUP lock has to hold to swap layout */
83 if (lli->lli_layout_gen != cio->cui_layout_gen) {
84 io->ci_need_restart = 1;
85 /* this will return application a short read/write */
90 /* fault is okay because we've already had a page. */
98 /*****************************************************************************
104 static int vvp_io_fault_iter_init(const struct lu_env *env,
105 const struct cl_io_slice *ios)
107 struct vvp_io *vio = cl2vvp_io(env, ios);
108 struct inode *inode = ccc_object_inode(ios->cis_obj);
111 cl2ccc_io(env, ios)->cui_fd->fd_file->f_dentry->d_inode);
112 vio->u.fault.ft_mtime = LTIME_S(inode->i_mtime);
116 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
118 struct cl_io *io = ios->cis_io;
119 struct cl_object *obj = io->ci_obj;
120 struct ccc_io *cio = cl2ccc_io(env, ios);
122 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
124 CDEBUG(D_VFSTRACE, "ignore/verify layout %d/%d, layout version %d.\n",
125 io->ci_ignore_layout, io->ci_verify_layout, cio->cui_layout_gen);
127 if (!io->ci_ignore_layout && io->ci_verify_layout) {
130 /* check layout version */
131 ll_layout_refresh(ccc_object_inode(obj), &gen);
132 io->ci_need_restart = cio->cui_layout_gen != gen;
133 if (io->ci_need_restart)
134 CDEBUG(D_VFSTRACE, "layout changed from %d to %d.\n",
135 cio->cui_layout_gen, gen);
139 static void vvp_io_fault_fini(const struct lu_env *env,
140 const struct cl_io_slice *ios)
142 struct cl_io *io = ios->cis_io;
143 struct cl_page *page = io->u.ci_fault.ft_page;
145 CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj));
148 lu_ref_del(&page->cp_reference, "fault", io);
149 cl_page_put(env, page);
150 io->u.ci_fault.ft_page = NULL;
152 vvp_io_fini(env, ios);
155 enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
158 * we only want to hold PW locks if the mmap() can generate
159 * writes back to the file and that only happens in shared
162 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
167 static int vvp_mmap_locks(const struct lu_env *env,
168 struct ccc_io *vio, struct cl_io *io)
170 struct ccc_thread_info *cti = ccc_env_info(env);
171 struct mm_struct *mm = current->mm;
172 struct vm_area_struct *vma;
173 struct cl_lock_descr *descr = &cti->cti_descr;
174 ldlm_policy_data_t policy;
180 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
182 if (!cl_is_normalio(env, io))
185 if (vio->cui_iov == NULL) /* nfs or loop back device write */
188 /* No MM (e.g. NFS)? No vmas too. */
192 for (seg = 0; seg < vio->cui_nrsegs; seg++) {
193 const struct iovec *iv = &vio->cui_iov[seg];
195 addr = (unsigned long)iv->iov_base;
200 count += addr & (~CFS_PAGE_MASK);
201 addr &= CFS_PAGE_MASK;
203 down_read(&mm->mmap_sem);
204 while((vma = our_vma(mm, addr, count)) != NULL) {
205 struct inode *inode = vma->vm_file->f_dentry->d_inode;
206 int flags = CEF_MUST;
208 if (ll_file_nolock(vma->vm_file)) {
210 * For no lock case, a lockless lock will be
217 * XXX: Required lock mode can be weakened: CIT_WRITE
218 * io only ever reads user level buffer, and CIT_READ
221 policy_from_vma(&policy, vma, addr, count);
222 descr->cld_mode = vvp_mode_from_vma(vma);
223 descr->cld_obj = ll_i2info(inode)->lli_clob;
224 descr->cld_start = cl_index(descr->cld_obj,
225 policy.l_extent.start);
226 descr->cld_end = cl_index(descr->cld_obj,
227 policy.l_extent.end);
228 descr->cld_enq_flags = flags;
229 result = cl_io_lock_alloc_add(env, io, descr);
231 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
232 descr->cld_mode, descr->cld_start,
238 if (vma->vm_end - addr >= count)
241 count -= vma->vm_end - addr;
244 up_read(&mm->mmap_sem);
249 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
250 enum cl_lock_mode mode, loff_t start, loff_t end)
252 struct ccc_io *cio = ccc_env_io(env);
256 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
258 ccc_io_update_iov(env, cio, io);
260 if (io->u.ci_rw.crw_nonblock)
261 ast_flags |= CEF_NONBLOCK;
262 result = vvp_mmap_locks(env, cio, io);
264 result = ccc_io_one_lock(env, io, ast_flags, mode, start, end);
268 static int vvp_io_read_lock(const struct lu_env *env,
269 const struct cl_io_slice *ios)
271 struct cl_io *io = ios->cis_io;
272 struct ll_inode_info *lli = ll_i2info(ccc_object_inode(io->ci_obj));
275 /* XXX: Layer violation, we shouldn't see lsm at llite level. */
276 if (lli->lli_has_smd) /* lsm-less file doesn't need to lock */
277 result = vvp_io_rw_lock(env, io, CLM_READ,
278 io->u.ci_rd.rd.crw_pos,
279 io->u.ci_rd.rd.crw_pos +
280 io->u.ci_rd.rd.crw_count - 1);
286 static int vvp_io_fault_lock(const struct lu_env *env,
287 const struct cl_io_slice *ios)
289 struct cl_io *io = ios->cis_io;
290 struct vvp_io *vio = cl2vvp_io(env, ios);
292 * XXX LDLM_FL_CBPENDING
294 return ccc_io_one_lock_index
295 (env, io, 0, vvp_mode_from_vma(vio->u.fault.ft_vma),
296 io->u.ci_fault.ft_index, io->u.ci_fault.ft_index);
299 static int vvp_io_write_lock(const struct lu_env *env,
300 const struct cl_io_slice *ios)
302 struct cl_io *io = ios->cis_io;
306 if (io->u.ci_wr.wr_append) {
308 end = OBD_OBJECT_EOF;
310 start = io->u.ci_wr.wr.crw_pos;
311 end = start + io->u.ci_wr.wr.crw_count - 1;
313 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
316 static int vvp_io_setattr_iter_init(const struct lu_env *env,
317 const struct cl_io_slice *ios)
323 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
325 * Handles "lockless io" mode when extent locking is done by server.
327 static int vvp_io_setattr_lock(const struct lu_env *env,
328 const struct cl_io_slice *ios)
330 struct ccc_io *cio = ccc_env_io(env);
331 struct cl_io *io = ios->cis_io;
335 if (cl_io_is_trunc(io)) {
336 new_size = io->u.ci_setattr.sa_attr.lvb_size;
338 enqflags = CEF_DISCARD_DATA;
340 if ((io->u.ci_setattr.sa_attr.lvb_mtime >=
341 io->u.ci_setattr.sa_attr.lvb_ctime) ||
342 (io->u.ci_setattr.sa_attr.lvb_atime >=
343 io->u.ci_setattr.sa_attr.lvb_ctime))
347 cio->u.setattr.cui_local_lock = SETATTR_EXTENT_LOCK;
348 return ccc_io_one_lock(env, io, enqflags, CLM_WRITE,
349 new_size, OBD_OBJECT_EOF);
352 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
356 * Only ll_inode_size_lock is taken at this level.
358 ll_inode_size_lock(inode);
359 result = inode_newsize_ok(inode, size);
361 ll_inode_size_unlock(inode);
364 truncate_setsize(inode, size);
365 ll_inode_size_unlock(inode);
369 static int vvp_io_setattr_trunc(const struct lu_env *env,
370 const struct cl_io_slice *ios,
371 struct inode *inode, loff_t size)
373 inode_dio_wait(inode);
377 static int vvp_io_setattr_time(const struct lu_env *env,
378 const struct cl_io_slice *ios)
380 struct cl_io *io = ios->cis_io;
381 struct cl_object *obj = io->ci_obj;
382 struct cl_attr *attr = ccc_env_thread_attr(env);
384 unsigned valid = CAT_CTIME;
386 cl_object_attr_lock(obj);
387 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
388 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
389 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
392 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
393 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
396 result = cl_object_attr_set(env, obj, attr, valid);
397 cl_object_attr_unlock(obj);
402 static int vvp_io_setattr_start(const struct lu_env *env,
403 const struct cl_io_slice *ios)
405 struct cl_io *io = ios->cis_io;
406 struct inode *inode = ccc_object_inode(io->ci_obj);
409 mutex_lock(&inode->i_mutex);
410 if (cl_io_is_trunc(io))
411 result = vvp_io_setattr_trunc(env, ios, inode,
412 io->u.ci_setattr.sa_attr.lvb_size);
414 result = vvp_io_setattr_time(env, ios);
418 static void vvp_io_setattr_end(const struct lu_env *env,
419 const struct cl_io_slice *ios)
421 struct cl_io *io = ios->cis_io;
422 struct inode *inode = ccc_object_inode(io->ci_obj);
424 if (cl_io_is_trunc(io)) {
425 /* Truncate in memory pages - they must be clean pages
426 * because osc has already notified to destroy osc_extents. */
427 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
428 inode_dio_write_done(inode);
430 mutex_unlock(&inode->i_mutex);
433 static void vvp_io_setattr_fini(const struct lu_env *env,
434 const struct cl_io_slice *ios)
436 vvp_io_fini(env, ios);
439 static ssize_t lustre_generic_file_read(struct file *file,
440 struct ccc_io *vio, loff_t *ppos)
442 return generic_file_aio_read(vio->cui_iocb, vio->cui_iov,
443 vio->cui_nrsegs, *ppos);
446 static ssize_t lustre_generic_file_write(struct file *file,
447 struct ccc_io *vio, loff_t *ppos)
449 return generic_file_aio_write(vio->cui_iocb, vio->cui_iov,
450 vio->cui_nrsegs, *ppos);
453 static int vvp_io_read_start(const struct lu_env *env,
454 const struct cl_io_slice *ios)
456 struct vvp_io *vio = cl2vvp_io(env, ios);
457 struct ccc_io *cio = cl2ccc_io(env, ios);
458 struct cl_io *io = ios->cis_io;
459 struct cl_object *obj = io->ci_obj;
460 struct inode *inode = ccc_object_inode(obj);
461 struct ll_ra_read *bead = &vio->cui_bead;
462 struct file *file = cio->cui_fd->fd_file;
465 loff_t pos = io->u.ci_rd.rd.crw_pos;
466 long cnt = io->u.ci_rd.rd.crw_count;
467 long tot = cio->cui_tot_count;
470 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
472 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
474 if (!can_populate_pages(env, io, inode))
477 result = ccc_prep_size(env, obj, io, pos, tot, &exceed);
480 else if (exceed != 0)
483 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
484 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
485 inode->i_ino, cnt, pos, i_size_read(inode));
487 /* turn off the kernel's read-ahead */
488 cio->cui_fd->fd_file->f_ra.ra_pages = 0;
490 /* initialize read-ahead window once per syscall */
491 if (!vio->cui_ra_window_set) {
492 vio->cui_ra_window_set = 1;
493 bead->lrr_start = cl_index(obj, pos);
495 * XXX: explicit PAGE_CACHE_SIZE
497 bead->lrr_count = cl_index(obj, tot + PAGE_CACHE_SIZE - 1);
498 ll_ra_read_in(file, bead);
503 switch (vio->cui_io_subtype) {
505 result = lustre_generic_file_read(file, cio, &pos);
508 result = generic_file_splice_read(file, &pos,
509 vio->u.splice.cui_pipe, cnt,
510 vio->u.splice.cui_flags);
511 /* LU-1109: do splice read stripe by stripe otherwise if it
512 * may make nfsd stuck if this read occupied all internal pipe
517 CERROR("Wrong IO type %u\n", vio->cui_io_subtype);
525 io->ci_nob += result;
526 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
527 cio->cui_fd, pos, result, READ);
533 static void vvp_io_read_fini(const struct lu_env *env, const struct cl_io_slice *ios)
535 struct vvp_io *vio = cl2vvp_io(env, ios);
536 struct ccc_io *cio = cl2ccc_io(env, ios);
538 if (vio->cui_ra_window_set)
539 ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead);
541 vvp_io_fini(env, ios);
544 static int vvp_io_write_start(const struct lu_env *env,
545 const struct cl_io_slice *ios)
547 struct ccc_io *cio = cl2ccc_io(env, ios);
548 struct cl_io *io = ios->cis_io;
549 struct cl_object *obj = io->ci_obj;
550 struct inode *inode = ccc_object_inode(obj);
551 struct file *file = cio->cui_fd->fd_file;
553 loff_t pos = io->u.ci_wr.wr.crw_pos;
554 size_t cnt = io->u.ci_wr.wr.crw_count;
556 if (!can_populate_pages(env, io, inode))
559 if (cl_io_is_append(io)) {
561 * PARALLEL IO This has to be changed for parallel IO doing
562 * out-of-order writes.
564 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
565 cio->cui_iocb->ki_pos = pos;
568 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
570 if (cio->cui_iov == NULL) /* from a temp io in ll_cl_init(). */
573 result = lustre_generic_file_write(file, cio, &pos);
578 io->ci_nob += result;
579 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
580 cio->cui_fd, pos, result, WRITE);
586 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
588 struct vm_fault *vmf = cfio->fault.ft_vmf;
590 cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, vmf);
593 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
594 vmf->virtual_address);
595 if (unlikely(!(cfio->fault.ft_flags & VM_FAULT_LOCKED))) {
596 lock_page(vmf->page);
597 cfio->fault.ft_flags &= VM_FAULT_LOCKED;
600 cfio->ft_vmpage = vmf->page;
604 if (cfio->fault.ft_flags & VM_FAULT_SIGBUS) {
605 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
609 if (cfio->fault.ft_flags & VM_FAULT_OOM) {
610 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
614 if (cfio->fault.ft_flags & VM_FAULT_RETRY)
617 CERROR("unknow error in page fault %d!\n", cfio->fault.ft_flags);
622 static int vvp_io_fault_start(const struct lu_env *env,
623 const struct cl_io_slice *ios)
625 struct vvp_io *vio = cl2vvp_io(env, ios);
626 struct cl_io *io = ios->cis_io;
627 struct cl_object *obj = io->ci_obj;
628 struct inode *inode = ccc_object_inode(obj);
629 struct cl_fault_io *fio = &io->u.ci_fault;
630 struct vvp_fault_io *cfio = &vio->u.fault;
633 struct page *vmpage = NULL;
634 struct cl_page *page;
636 pgoff_t last; /* last page in a file data region */
638 if (fio->ft_executable &&
639 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
641 " changed while waiting for the page fault lock\n",
642 PFID(lu_object_fid(&obj->co_lu)));
644 /* offset of the last byte on the page */
645 offset = cl_offset(obj, fio->ft_index + 1) - 1;
646 LASSERT(cl_index(obj, offset) == fio->ft_index);
647 result = ccc_prep_size(env, obj, io, 0, offset + 1, NULL);
651 /* must return locked page */
652 if (fio->ft_mkwrite) {
653 LASSERT(cfio->ft_vmpage != NULL);
654 lock_page(cfio->ft_vmpage);
656 result = vvp_io_kernel_fault(cfio);
661 vmpage = cfio->ft_vmpage;
662 LASSERT(PageLocked(vmpage));
664 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
665 ll_invalidate_page(vmpage);
667 size = i_size_read(inode);
668 /* Though we have already held a cl_lock upon this page, but
669 * it still can be truncated locally. */
670 if (unlikely((vmpage->mapping != inode->i_mapping) ||
671 (page_offset(vmpage) > size))) {
672 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
674 /* return +1 to stop cl_io_loop() and ll_fault() will catch
676 GOTO(out, result = +1);
680 if (fio->ft_mkwrite ) {
683 * Capture the size while holding the lli_trunc_sem from above
684 * we want to make sure that we complete the mkwrite action
685 * while holding this lock. We need to make sure that we are
686 * not past the end of the file.
688 last_index = cl_index(obj, size - 1);
689 if (last_index < fio->ft_index) {
691 "llite: mkwrite and truncate race happened: "
693 vmpage->mapping,fio->ft_index,last_index);
695 * We need to return if we are
696 * passed the end of the file. This will propagate
697 * up the call stack to ll_page_mkwrite where
698 * we will return VM_FAULT_NOPAGE. Any non-negative
699 * value returned here will be silently
700 * converted to 0. If the vmpage->mapping is null
701 * the error code would be converted back to ENODATA
702 * in ll_page_mkwrite0. Thus we return -ENODATA
703 * to handle both cases
705 GOTO(out, result = -ENODATA);
709 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
711 GOTO(out, result = PTR_ERR(page));
713 /* if page is going to be written, we should add this page into cache
715 if (fio->ft_mkwrite) {
716 wait_on_page_writeback(vmpage);
717 if (set_page_dirty(vmpage)) {
720 /* vvp_page_assume() calls wait_on_page_writeback(). */
721 cl_page_assume(env, io, page);
723 cp = cl2ccc_page(cl_page_at(page, &vvp_device_type));
724 vvp_write_pending(cl2ccc(obj), cp);
726 /* Do not set Dirty bit here so that in case IO is
727 * started before the page is really made dirty, we
728 * still have chance to detect it. */
729 result = cl_page_cache_add(env, io, page, CRT_WRITE);
730 LASSERT(cl_page_is_owned(page, io));
734 cl_page_unmap(env, io, page);
735 cl_page_discard(env, io, page);
736 cl_page_disown(env, io, page);
738 cl_page_put(env, page);
740 /* we're in big trouble, what can we do now? */
741 if (result == -EDQUOT)
745 cl_page_disown(env, io, page);
749 last = cl_index(obj, size - 1);
751 * The ft_index is only used in the case of
752 * a mkwrite action. We need to check
753 * our assertions are correct, since
754 * we should have caught this above
756 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last);
757 if (fio->ft_index == last)
759 * Last page is mapped partially.
761 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
763 fio->ft_nob = cl_page_size(obj);
765 lu_ref_add(&page->cp_reference, "fault", io);
769 /* return unlocked vmpage to avoid deadlocking */
772 cfio->fault.ft_flags &= ~VM_FAULT_LOCKED;
776 static int vvp_io_fsync_start(const struct lu_env *env,
777 const struct cl_io_slice *ios)
779 /* we should mark TOWRITE bit to each dirty page in radix tree to
780 * verify pages have been written, but this is difficult because of
785 static int vvp_io_read_page(const struct lu_env *env,
786 const struct cl_io_slice *ios,
787 const struct cl_page_slice *slice)
789 struct cl_io *io = ios->cis_io;
790 struct cl_object *obj = slice->cpl_obj;
791 struct ccc_page *cp = cl2ccc_page(slice);
792 struct cl_page *page = slice->cpl_page;
793 struct inode *inode = ccc_object_inode(obj);
794 struct ll_sb_info *sbi = ll_i2sbi(inode);
795 struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd;
796 struct ll_readahead_state *ras = &fd->fd_ras;
797 struct page *vmpage = cp->cpg_page;
798 struct cl_2queue *queue = &io->ci_queue;
801 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
802 LASSERT(slice->cpl_obj == obj);
804 if (sbi->ll_ra_info.ra_max_pages_per_file &&
805 sbi->ll_ra_info.ra_max_pages)
806 ras_update(sbi, inode, ras, page->cp_index,
807 cp->cpg_defer_uptodate);
809 /* Sanity check whether the page is protected by a lock. */
810 rc = cl_page_is_under_lock(env, io, page);
812 CL_PAGE_HEADER(D_WARNING, env, page, "%s: %d\n",
813 rc == -ENODATA ? "without a lock" :
819 if (cp->cpg_defer_uptodate) {
821 cl_page_export(env, page, 1);
824 * Add page into the queue even when it is marked uptodate above.
825 * this will unlock it automatically as part of cl_page_list_disown().
827 cl_2queue_add(queue, page);
828 if (sbi->ll_ra_info.ra_max_pages_per_file &&
829 sbi->ll_ra_info.ra_max_pages)
830 ll_readahead(env, io, ras,
831 vmpage->mapping, &queue->c2_qin, fd->fd_flags);
836 static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io,
837 struct cl_page *page, struct ccc_page *cp,
838 enum cl_req_type crt)
840 struct cl_2queue *queue;
843 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
845 queue = &io->ci_queue;
846 cl_2queue_init_page(queue, page);
848 result = cl_io_submit_sync(env, io, crt, queue, 0);
849 LASSERT(cl_page_is_owned(page, io));
853 * in CRT_WRITE case page is left locked even in case of
856 cl_page_list_disown(env, io, &queue->c2_qin);
857 cl_2queue_fini(env, queue);
863 * Prepare partially written-to page for a write.
865 static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io,
866 struct cl_object *obj, struct cl_page *pg,
868 unsigned from, unsigned to)
870 struct cl_attr *attr = ccc_env_thread_attr(env);
871 loff_t offset = cl_offset(obj, pg->cp_index);
874 cl_object_attr_lock(obj);
875 result = cl_object_attr_get(env, obj, attr);
876 cl_object_attr_unlock(obj);
879 * If are writing to a new page, no need to read old data.
880 * The extent locking will have updated the KMS, and for our
881 * purposes here we can treat it like i_size.
883 if (attr->cat_kms <= offset) {
884 char *kaddr = kmap_atomic(cp->cpg_page);
886 memset(kaddr, 0, cl_page_size(obj));
887 kunmap_atomic(kaddr);
888 } else if (cp->cpg_defer_uptodate)
891 result = vvp_page_sync_io(env, io, pg, cp, CRT_READ);
893 * In older implementations, obdo_refresh_inode is called here
894 * to update the inode because the write might modify the
895 * object info at OST. However, this has been proven useless,
896 * since LVB functions will be called when user space program
897 * tries to retrieve inode attribute. Also, see bug 15909 for
901 cl_page_export(env, pg, 1);
906 static int vvp_io_prepare_write(const struct lu_env *env,
907 const struct cl_io_slice *ios,
908 const struct cl_page_slice *slice,
909 unsigned from, unsigned to)
911 struct cl_object *obj = slice->cpl_obj;
912 struct ccc_page *cp = cl2ccc_page(slice);
913 struct cl_page *pg = slice->cpl_page;
914 struct page *vmpage = cp->cpg_page;
918 LINVRNT(cl_page_is_vmlocked(env, pg));
919 LASSERT(vmpage->mapping->host == ccc_object_inode(obj));
923 CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to);
924 if (!PageUptodate(vmpage)) {
926 * We're completely overwriting an existing page, so _don't_
927 * set it up to date until commit_write
929 if (from == 0 && to == PAGE_CACHE_SIZE) {
930 CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n");
931 POISON_PAGE(page, 0x11);
933 result = vvp_io_prepare_partial(env, ios->cis_io, obj,
936 CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n");
940 static int vvp_io_commit_write(const struct lu_env *env,
941 const struct cl_io_slice *ios,
942 const struct cl_page_slice *slice,
943 unsigned from, unsigned to)
945 struct cl_object *obj = slice->cpl_obj;
946 struct cl_io *io = ios->cis_io;
947 struct ccc_page *cp = cl2ccc_page(slice);
948 struct cl_page *pg = slice->cpl_page;
949 struct inode *inode = ccc_object_inode(obj);
950 struct ll_sb_info *sbi = ll_i2sbi(inode);
951 struct ll_inode_info *lli = ll_i2info(inode);
952 struct page *vmpage = cp->cpg_page;
958 LINVRNT(cl_page_is_vmlocked(env, pg));
959 LASSERT(vmpage->mapping->host == inode);
961 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "committing page write\n");
962 CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to);
965 * queue a write for some time in the future the first time we
968 * This is different from what other file systems do: they usually
969 * just mark page (and some of its buffers) dirty and rely on
970 * balance_dirty_pages() to start a write-back. Lustre wants write-back
971 * to be started earlier for the following reasons:
973 * (1) with a large number of clients we need to limit the amount
974 * of cached data on the clients a lot;
976 * (2) large compute jobs generally want compute-only then io-only
977 * and the IO should complete as quickly as possible;
979 * (3) IO is batched up to the RPC size and is async until the
980 * client max cache is hit
981 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
984 if (!PageDirty(vmpage)) {
985 tallyop = LPROC_LL_DIRTY_MISSES;
986 result = cl_page_cache_add(env, io, pg, CRT_WRITE);
988 /* page was added into cache successfully. */
989 set_page_dirty(vmpage);
990 vvp_write_pending(cl2ccc(obj), cp);
991 } else if (result == -EDQUOT) {
992 pgoff_t last_index = i_size_read(inode) >> PAGE_CACHE_SHIFT;
993 bool need_clip = true;
996 * Client ran out of disk space grant. Possible
999 * (a) do a sync write, renewing grant;
1001 * (b) stop writing on this stripe, switch to the
1004 * (b) is a part of "parallel io" design that is the
1005 * ultimate goal. (a) is what "old" client did, and
1006 * what the new code continues to do for the time
1009 if (last_index > pg->cp_index) {
1010 to = PAGE_CACHE_SIZE;
1012 } else if (last_index == pg->cp_index) {
1013 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
1018 cl_page_clip(env, pg, 0, to);
1019 result = vvp_page_sync_io(env, io, pg, cp, CRT_WRITE);
1021 CERROR("Write page %lu of inode %p failed %d\n",
1022 pg->cp_index, inode, result);
1025 tallyop = LPROC_LL_DIRTY_HITS;
1028 ll_stats_ops_tally(sbi, tallyop, 1);
1030 /* Inode should be marked DIRTY even if no new page was marked DIRTY
1031 * because page could have been not flushed between 2 modifications.
1032 * It is important the file is marked DIRTY as soon as the I/O is done
1033 * Indeed, when cache is flushed, file could be already closed and it
1034 * is too late to warn the MDT.
1035 * It is acceptable that file is marked DIRTY even if I/O is dropped
1036 * for some reasons before being flushed to OST.
1039 spin_lock(&lli->lli_lock);
1040 lli->lli_flags |= LLIF_DATA_MODIFIED;
1041 spin_unlock(&lli->lli_lock);
1044 size = cl_offset(obj, pg->cp_index) + to;
1046 ll_inode_size_lock(inode);
1048 if (size > i_size_read(inode)) {
1049 cl_isize_write_nolock(inode, size);
1050 CDEBUG(D_VFSTRACE, DFID" updating i_size %lu\n",
1051 PFID(lu_object_fid(&obj->co_lu)),
1052 (unsigned long)size);
1054 cl_page_export(env, pg, 1);
1056 if (size > i_size_read(inode))
1057 cl_page_discard(env, io, pg);
1059 ll_inode_size_unlock(inode);
1063 static const struct cl_io_operations vvp_io_ops = {
1066 .cio_fini = vvp_io_read_fini,
1067 .cio_lock = vvp_io_read_lock,
1068 .cio_start = vvp_io_read_start,
1069 .cio_advance = ccc_io_advance
1072 .cio_fini = vvp_io_fini,
1073 .cio_lock = vvp_io_write_lock,
1074 .cio_start = vvp_io_write_start,
1075 .cio_advance = ccc_io_advance
1078 .cio_fini = vvp_io_setattr_fini,
1079 .cio_iter_init = vvp_io_setattr_iter_init,
1080 .cio_lock = vvp_io_setattr_lock,
1081 .cio_start = vvp_io_setattr_start,
1082 .cio_end = vvp_io_setattr_end
1085 .cio_fini = vvp_io_fault_fini,
1086 .cio_iter_init = vvp_io_fault_iter_init,
1087 .cio_lock = vvp_io_fault_lock,
1088 .cio_start = vvp_io_fault_start,
1089 .cio_end = ccc_io_end
1092 .cio_start = vvp_io_fsync_start,
1093 .cio_fini = vvp_io_fini
1096 .cio_fini = vvp_io_fini
1099 .cio_read_page = vvp_io_read_page,
1100 .cio_prepare_write = vvp_io_prepare_write,
1101 .cio_commit_write = vvp_io_commit_write
1104 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1107 struct vvp_io *vio = vvp_env_io(env);
1108 struct ccc_io *cio = ccc_env_io(env);
1109 struct inode *inode = ccc_object_inode(obj);
1112 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
1114 CL_IO_SLICE_CLEAN(cio, cui_cl);
1115 cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops);
1116 vio->cui_ra_window_set = 0;
1118 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1120 struct ll_inode_info *lli = ll_i2info(inode);
1122 count = io->u.ci_rw.crw_count;
1123 /* "If nbyte is 0, read() will return 0 and have no other
1124 * results." -- Single Unix Spec */
1128 cio->cui_tot_count = count;
1129 cio->cui_tot_nrsegs = 0;
1131 /* for read/write, we store the jobid in the inode, and
1132 * it'll be fetched by osc when building RPC.
1134 * it's not accurate if the file is shared by different
1137 lustre_get_jobid(lli->lli_jobid);
1138 } else if (io->ci_type == CIT_SETATTR) {
1139 if (!cl_io_is_trunc(io))
1140 io->ci_lockreq = CILR_MANDATORY;
1143 /* ignore layout change for generic CIT_MISC but not for glimpse.
1144 * io context for glimpse must set ci_verify_layout to true,
1145 * see cl_glimpse_size0() for details. */
1146 if (io->ci_type == CIT_MISC && !io->ci_verify_layout)
1147 io->ci_ignore_layout = 1;
1149 /* Enqueue layout lock and get layout version. We need to do this
1150 * even for operations requiring to open file, such as read and write,
1151 * because it might not grant layout lock in IT_OPEN. */
1152 if (result == 0 && !io->ci_ignore_layout) {
1153 result = ll_layout_refresh(inode, &cio->cui_layout_gen);
1154 if (result == -ENOENT)
1155 /* If the inode on MDS has been removed, but the objects
1156 * on OSTs haven't been destroyed (async unlink), layout
1157 * fetch will return -ENOENT, we'd ingore this error
1158 * and continue with dirty flush. LU-3230. */
1161 CERROR("%s: refresh file layout " DFID " error %d.\n",
1162 ll_get_fsname(inode->i_sb, NULL, 0),
1163 PFID(lu_object_fid(&obj->co_lu)), result);
1169 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
1170 const struct cl_io_slice *slice)
1172 /* Caling just for assertion */
1173 cl2ccc_io(env, slice);
1174 return vvp_env_io(env);