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, 2013, 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
45 # error This file is kernel only.
49 #include <lustre_lite.h>
51 #include "vvp_internal.h"
53 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
54 const struct cl_io_slice *slice);
57 * True, if \a io is a normal io, False for sendfile() / splice_{read|write}
59 int cl_is_normalio(const struct lu_env *env, const struct cl_io *io)
61 struct vvp_io *vio = vvp_env_io(env);
63 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
65 return vio->cui_io_subtype == IO_NORMAL;
69 * For swapping layout. The file's layout may have changed.
70 * To avoid populating pages to a wrong stripe, we have to verify the
71 * correctness of layout. It works because swapping layout processes
72 * have to acquire group lock.
74 static bool can_populate_pages(const struct lu_env *env, struct cl_io *io,
77 struct ll_inode_info *lli = ll_i2info(inode);
78 struct ccc_io *cio = ccc_env_io(env);
81 switch (io->ci_type) {
84 /* don't need lock here to check lli_layout_gen as we have held
85 * extent lock and GROUP lock has to hold to swap layout */
86 if (lli->lli_layout_gen != cio->cui_layout_gen) {
87 io->ci_need_restart = 1;
88 /* this will return application a short read/write */
93 /* fault is okay because we've already had a page. */
101 /*****************************************************************************
107 static int vvp_io_fault_iter_init(const struct lu_env *env,
108 const struct cl_io_slice *ios)
110 struct vvp_io *vio = cl2vvp_io(env, ios);
111 struct inode *inode = ccc_object_inode(ios->cis_obj);
114 cl2ccc_io(env, ios)->cui_fd->fd_file->f_dentry->d_inode);
115 vio->u.fault.ft_mtime = LTIME_S(inode->i_mtime);
119 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
121 struct cl_io *io = ios->cis_io;
122 struct cl_object *obj = io->ci_obj;
123 struct ccc_io *cio = cl2ccc_io(env, ios);
125 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
127 CDEBUG(D_VFSTRACE, "ignore/verify layout %d/%d, layout version %d.\n",
128 io->ci_ignore_layout, io->ci_verify_layout, cio->cui_layout_gen);
130 if (!io->ci_ignore_layout && io->ci_verify_layout) {
133 /* check layout version */
134 ll_layout_refresh(ccc_object_inode(obj), &gen);
135 io->ci_need_restart = cio->cui_layout_gen != gen;
136 if (io->ci_need_restart)
137 CDEBUG(D_VFSTRACE, "layout changed from %d to %d.\n",
138 cio->cui_layout_gen, gen);
142 static void vvp_io_fault_fini(const struct lu_env *env,
143 const struct cl_io_slice *ios)
145 struct cl_io *io = ios->cis_io;
146 struct cl_page *page = io->u.ci_fault.ft_page;
148 CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj));
151 lu_ref_del(&page->cp_reference, "fault", io);
152 cl_page_put(env, page);
153 io->u.ci_fault.ft_page = NULL;
155 vvp_io_fini(env, ios);
158 enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
161 * we only want to hold PW locks if the mmap() can generate
162 * writes back to the file and that only happens in shared
165 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
170 static int vvp_mmap_locks(const struct lu_env *env,
171 struct ccc_io *vio, struct cl_io *io)
173 struct ccc_thread_info *cti = ccc_env_info(env);
174 struct mm_struct *mm = current->mm;
175 struct vm_area_struct *vma;
176 struct cl_lock_descr *descr = &cti->cti_descr;
177 ldlm_policy_data_t policy;
184 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
186 if (!cl_is_normalio(env, io))
189 if (vio->cui_iov == NULL) /* nfs or loop back device write */
192 /* No MM (e.g. NFS)? No vmas too. */
196 for (seg = 0; seg < vio->cui_nrsegs; seg++) {
197 const struct iovec *iv = &vio->cui_iov[seg];
199 addr = (unsigned long)iv->iov_base;
204 count += addr & (~CFS_PAGE_MASK);
205 addr &= CFS_PAGE_MASK;
207 down_read(&mm->mmap_sem);
208 while((vma = our_vma(mm, addr, count)) != NULL) {
209 struct inode *inode = vma->vm_file->f_dentry->d_inode;
210 int flags = CEF_MUST;
212 if (ll_file_nolock(vma->vm_file)) {
214 * For no lock case, a lockless lock will be
221 * XXX: Required lock mode can be weakened: CIT_WRITE
222 * io only ever reads user level buffer, and CIT_READ
225 policy_from_vma(&policy, vma, addr, count);
226 descr->cld_mode = vvp_mode_from_vma(vma);
227 descr->cld_obj = ll_i2info(inode)->lli_clob;
228 descr->cld_start = cl_index(descr->cld_obj,
229 policy.l_extent.start);
230 descr->cld_end = cl_index(descr->cld_obj,
231 policy.l_extent.end);
232 descr->cld_enq_flags = flags;
233 result = cl_io_lock_alloc_add(env, io, descr);
235 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
236 descr->cld_mode, descr->cld_start,
242 if (vma->vm_end - addr >= count)
245 count -= vma->vm_end - addr;
248 up_read(&mm->mmap_sem);
253 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
254 enum cl_lock_mode mode, loff_t start, loff_t end)
256 struct ccc_io *cio = ccc_env_io(env);
260 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
263 ccc_io_update_iov(env, cio, io);
265 if (io->u.ci_rw.crw_nonblock)
266 ast_flags |= CEF_NONBLOCK;
267 result = vvp_mmap_locks(env, cio, io);
269 result = ccc_io_one_lock(env, io, ast_flags, mode, start, end);
273 static int vvp_io_read_lock(const struct lu_env *env,
274 const struct cl_io_slice *ios)
276 struct cl_io *io = ios->cis_io;
277 struct ll_inode_info *lli = ll_i2info(ccc_object_inode(io->ci_obj));
281 /* XXX: Layer violation, we shouldn't see lsm at llite level. */
282 if (lli->lli_has_smd) /* lsm-less file doesn't need to lock */
283 result = vvp_io_rw_lock(env, io, CLM_READ,
284 io->u.ci_rd.rd.crw_pos,
285 io->u.ci_rd.rd.crw_pos +
286 io->u.ci_rd.rd.crw_count - 1);
292 static int vvp_io_fault_lock(const struct lu_env *env,
293 const struct cl_io_slice *ios)
295 struct cl_io *io = ios->cis_io;
296 struct vvp_io *vio = cl2vvp_io(env, ios);
298 * XXX LDLM_FL_CBPENDING
300 return ccc_io_one_lock_index
301 (env, io, 0, vvp_mode_from_vma(vio->u.fault.ft_vma),
302 io->u.ci_fault.ft_index, io->u.ci_fault.ft_index);
305 static int vvp_io_write_lock(const struct lu_env *env,
306 const struct cl_io_slice *ios)
308 struct cl_io *io = ios->cis_io;
312 if (io->u.ci_wr.wr_append) {
314 end = OBD_OBJECT_EOF;
316 start = io->u.ci_wr.wr.crw_pos;
317 end = start + io->u.ci_wr.wr.crw_count - 1;
319 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
322 static int vvp_io_setattr_iter_init(const struct lu_env *env,
323 const struct cl_io_slice *ios)
329 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
331 * Handles "lockless io" mode when extent locking is done by server.
333 static int vvp_io_setattr_lock(const struct lu_env *env,
334 const struct cl_io_slice *ios)
336 struct ccc_io *cio = ccc_env_io(env);
337 struct cl_io *io = ios->cis_io;
341 if (cl_io_is_trunc(io)) {
342 new_size = io->u.ci_setattr.sa_attr.lvb_size;
344 enqflags = CEF_DISCARD_DATA;
346 if ((io->u.ci_setattr.sa_attr.lvb_mtime >=
347 io->u.ci_setattr.sa_attr.lvb_ctime) ||
348 (io->u.ci_setattr.sa_attr.lvb_atime >=
349 io->u.ci_setattr.sa_attr.lvb_ctime))
353 cio->u.setattr.cui_local_lock = SETATTR_EXTENT_LOCK;
354 return ccc_io_one_lock(env, io, enqflags, CLM_WRITE,
355 new_size, OBD_OBJECT_EOF);
358 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
364 * Only ll_inode_size_lock is taken at this level.
366 ll_inode_size_lock(inode);
367 result = inode_newsize_ok(inode, size);
369 ll_inode_size_unlock(inode);
372 oldsize = inode->i_size;
373 i_size_write(inode, size);
375 truncate_pagecache(inode, oldsize, size);
376 ll_inode_size_unlock(inode);
380 static int vvp_io_setattr_trunc(const struct lu_env *env,
381 const struct cl_io_slice *ios,
382 struct inode *inode, loff_t size)
384 inode_dio_wait(inode);
388 static int vvp_io_setattr_time(const struct lu_env *env,
389 const struct cl_io_slice *ios)
391 struct cl_io *io = ios->cis_io;
392 struct cl_object *obj = io->ci_obj;
393 struct cl_attr *attr = ccc_env_thread_attr(env);
395 unsigned valid = CAT_CTIME;
397 cl_object_attr_lock(obj);
398 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
399 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
400 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
403 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
404 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
407 result = cl_object_attr_set(env, obj, attr, valid);
408 cl_object_attr_unlock(obj);
413 static int vvp_io_setattr_start(const struct lu_env *env,
414 const struct cl_io_slice *ios)
416 struct cl_io *io = ios->cis_io;
417 struct inode *inode = ccc_object_inode(io->ci_obj);
420 mutex_lock(&inode->i_mutex);
421 if (cl_io_is_trunc(io))
422 result = vvp_io_setattr_trunc(env, ios, inode,
423 io->u.ci_setattr.sa_attr.lvb_size);
425 result = vvp_io_setattr_time(env, ios);
429 static void vvp_io_setattr_end(const struct lu_env *env,
430 const struct cl_io_slice *ios)
432 struct cl_io *io = ios->cis_io;
433 struct inode *inode = ccc_object_inode(io->ci_obj);
435 if (cl_io_is_trunc(io)) {
436 /* Truncate in memory pages - they must be clean pages
437 * because osc has already notified to destroy osc_extents. */
438 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
439 inode_dio_write_done(inode);
441 mutex_unlock(&inode->i_mutex);
444 static void vvp_io_setattr_fini(const struct lu_env *env,
445 const struct cl_io_slice *ios)
447 vvp_io_fini(env, ios);
450 #ifdef HAVE_FILE_READV
451 static ssize_t lustre_generic_file_read(struct file *file,
452 struct ccc_io *vio, loff_t *ppos)
454 return generic_file_readv(file, vio->cui_iov, vio->cui_nrsegs, ppos);
457 static ssize_t lustre_generic_file_write(struct file *file,
458 struct ccc_io *vio, loff_t *ppos)
460 return generic_file_writev(file, vio->cui_iov, vio->cui_nrsegs, ppos);
463 static ssize_t lustre_generic_file_read(struct file *file,
464 struct ccc_io *vio, loff_t *ppos)
466 return generic_file_aio_read(vio->cui_iocb, vio->cui_iov,
467 vio->cui_nrsegs, *ppos);
470 static ssize_t lustre_generic_file_write(struct file *file,
471 struct ccc_io *vio, loff_t *ppos)
473 return generic_file_aio_write(vio->cui_iocb, vio->cui_iov,
474 vio->cui_nrsegs, *ppos);
478 static int vvp_io_read_start(const struct lu_env *env,
479 const struct cl_io_slice *ios)
481 struct vvp_io *vio = cl2vvp_io(env, ios);
482 struct ccc_io *cio = cl2ccc_io(env, ios);
483 struct cl_io *io = ios->cis_io;
484 struct cl_object *obj = io->ci_obj;
485 struct inode *inode = ccc_object_inode(obj);
486 struct ll_ra_read *bead = &vio->cui_bead;
487 struct file *file = cio->cui_fd->fd_file;
490 loff_t pos = io->u.ci_rd.rd.crw_pos;
491 long cnt = io->u.ci_rd.rd.crw_count;
492 long tot = cio->cui_tot_count;
495 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
497 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
499 if (!can_populate_pages(env, io, inode))
502 result = ccc_prep_size(env, obj, io, pos, tot, &exceed);
505 else if (exceed != 0)
508 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
509 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
510 inode->i_ino, cnt, pos, i_size_read(inode));
512 /* turn off the kernel's read-ahead */
513 cio->cui_fd->fd_file->f_ra.ra_pages = 0;
515 /* initialize read-ahead window once per syscall */
516 if (!vio->cui_ra_window_set) {
517 vio->cui_ra_window_set = 1;
518 bead->lrr_start = cl_index(obj, pos);
520 * XXX: explicit CFS_PAGE_SIZE
522 bead->lrr_count = cl_index(obj, tot + CFS_PAGE_SIZE - 1);
523 ll_ra_read_in(file, bead);
528 switch (vio->cui_io_subtype) {
530 result = lustre_generic_file_read(file, cio, &pos);
532 #ifdef HAVE_KERNEL_SENDFILE
534 result = generic_file_sendfile(file, &pos, cnt,
535 vio->u.sendfile.cui_actor,
536 vio->u.sendfile.cui_target);
539 #ifdef HAVE_KERNEL_SPLICE_READ
541 result = generic_file_splice_read(file, &pos,
542 vio->u.splice.cui_pipe, cnt,
543 vio->u.splice.cui_flags);
544 /* LU-1109: do splice read stripe by stripe otherwise if it
545 * may make nfsd stuck if this read occupied all internal pipe
551 CERROR("Wrong IO type %u\n", vio->cui_io_subtype);
559 io->ci_nob += result;
560 ll_rw_stats_tally(ll_i2sbi(inode), current->pid, cio->cui_fd,
568 static void vvp_io_read_fini(const struct lu_env *env, const struct cl_io_slice *ios)
570 struct vvp_io *vio = cl2vvp_io(env, ios);
571 struct ccc_io *cio = cl2ccc_io(env, ios);
573 if (vio->cui_ra_window_set)
574 ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead);
576 vvp_io_fini(env, ios);
579 static int vvp_io_write_start(const struct lu_env *env,
580 const struct cl_io_slice *ios)
582 struct ccc_io *cio = cl2ccc_io(env, ios);
583 struct cl_io *io = ios->cis_io;
584 struct cl_object *obj = io->ci_obj;
585 struct inode *inode = ccc_object_inode(obj);
586 struct file *file = cio->cui_fd->fd_file;
588 loff_t pos = io->u.ci_wr.wr.crw_pos;
589 size_t cnt = io->u.ci_wr.wr.crw_count;
593 if (!can_populate_pages(env, io, inode))
596 if (cl_io_is_append(io)) {
598 * PARALLEL IO This has to be changed for parallel IO doing
599 * out-of-order writes.
601 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
602 #ifndef HAVE_FILE_WRITEV
603 cio->cui_iocb->ki_pos = pos;
607 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
609 if (cio->cui_iov == NULL) /* from a temp io in ll_cl_init(). */
612 result = lustre_generic_file_write(file, cio, &pos);
617 io->ci_nob += result;
618 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
619 cio->cui_fd, pos, result, WRITE);
626 #ifndef HAVE_VM_OP_FAULT
627 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
631 vmpage = filemap_nopage(cfio->ft_vma, cfio->nopage.ft_address,
632 cfio->nopage.ft_type);
634 if (vmpage == NOPAGE_SIGBUS) {
635 CDEBUG(D_PAGE, "got addr %lu type %lx - SIGBUS\n",
636 cfio->nopage.ft_address,(long)cfio->nopage.ft_type);
638 } else if (vmpage == NOPAGE_OOM) {
639 CDEBUG(D_PAGE, "got addr %lu type %lx - OOM\n",
640 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
644 LL_CDEBUG_PAGE(D_PAGE, vmpage, "got addr %lu type %lx\n",
645 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
647 cfio->ft_vmpage = vmpage;
653 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
655 struct vm_fault *vmf = cfio->fault.ft_vmf;
657 cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, vmf);
660 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
661 vmf->virtual_address);
662 if (unlikely(!(cfio->fault.ft_flags & VM_FAULT_LOCKED))) {
663 lock_page(vmf->page);
664 cfio->fault.ft_flags &= VM_FAULT_LOCKED;
667 cfio->ft_vmpage = vmf->page;
671 if (cfio->fault.ft_flags & VM_FAULT_SIGBUS) {
672 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
676 if (cfio->fault.ft_flags & VM_FAULT_OOM) {
677 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
681 if (cfio->fault.ft_flags & VM_FAULT_RETRY)
684 CERROR("unknow error in page fault %d!\n", cfio->fault.ft_flags);
690 static int vvp_io_fault_start(const struct lu_env *env,
691 const struct cl_io_slice *ios)
693 struct vvp_io *vio = cl2vvp_io(env, ios);
694 struct cl_io *io = ios->cis_io;
695 struct cl_object *obj = io->ci_obj;
696 struct inode *inode = ccc_object_inode(obj);
697 struct cl_fault_io *fio = &io->u.ci_fault;
698 struct vvp_fault_io *cfio = &vio->u.fault;
701 cfs_page_t *vmpage = NULL;
702 struct cl_page *page;
704 pgoff_t last; /* last page in a file data region */
706 if (fio->ft_executable &&
707 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
709 " changed while waiting for the page fault lock\n",
710 PFID(lu_object_fid(&obj->co_lu)));
712 /* offset of the last byte on the page */
713 offset = cl_offset(obj, fio->ft_index + 1) - 1;
714 LASSERT(cl_index(obj, offset) == fio->ft_index);
715 result = ccc_prep_size(env, obj, io, 0, offset + 1, NULL);
719 /* must return locked page */
720 if (fio->ft_mkwrite) {
721 LASSERT(cfio->ft_vmpage != NULL);
722 lock_page(cfio->ft_vmpage);
724 result = vvp_io_kernel_fault(cfio);
729 vmpage = cfio->ft_vmpage;
730 LASSERT(PageLocked(vmpage));
732 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
733 ll_invalidate_page(vmpage);
735 size = i_size_read(inode);
736 /* Though we have already held a cl_lock upon this page, but
737 * it still can be truncated locally. */
738 if (unlikely((vmpage->mapping != inode->i_mapping) ||
739 (page_offset(vmpage) > size))) {
740 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
742 /* return +1 to stop cl_io_loop() and ll_fault() will catch
744 GOTO(out, result = +1);
748 if (fio->ft_mkwrite ) {
751 * Capture the size while holding the lli_trunc_sem from above
752 * we want to make sure that we complete the mkwrite action
753 * while holding this lock. We need to make sure that we are
754 * not past the end of the file.
756 last_index = cl_index(obj, size - 1);
757 if (last_index < fio->ft_index) {
759 "llite: mkwrite and truncate race happened: "
761 vmpage->mapping,fio->ft_index,last_index);
763 * We need to return if we are
764 * passed the end of the file. This will propagate
765 * up the call stack to ll_page_mkwrite where
766 * we will return VM_FAULT_NOPAGE. Any non-negative
767 * value returned here will be silently
768 * converted to 0. If the vmpage->mapping is null
769 * the error code would be converted back to ENODATA
770 * in ll_page_mkwrite0. Thus we return -ENODATA
771 * to handle both cases
773 GOTO(out, result = -ENODATA);
777 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
779 GOTO(out, result = PTR_ERR(page));
781 /* if page is going to be written, we should add this page into cache
783 if (fio->ft_mkwrite) {
784 wait_on_page_writeback(vmpage);
785 if (set_page_dirty(vmpage)) {
788 /* vvp_page_assume() calls wait_on_page_writeback(). */
789 cl_page_assume(env, io, page);
791 cp = cl2ccc_page(cl_page_at(page, &vvp_device_type));
792 vvp_write_pending(cl2ccc(obj), cp);
794 /* Do not set Dirty bit here so that in case IO is
795 * started before the page is really made dirty, we
796 * still have chance to detect it. */
797 result = cl_page_cache_add(env, io, page, CRT_WRITE);
798 LASSERT(cl_page_is_owned(page, io));
802 cl_page_unmap(env, io, page);
803 cl_page_discard(env, io, page);
804 cl_page_disown(env, io, page);
806 cl_page_put(env, page);
808 /* we're in big trouble, what can we do now? */
809 if (result == -EDQUOT)
813 cl_page_disown(env, io, page);
817 last = cl_index(obj, size - 1);
819 * The ft_index is only used in the case of
820 * a mkwrite action. We need to check
821 * our assertions are correct, since
822 * we should have caught this above
824 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last);
825 if (fio->ft_index == last)
827 * Last page is mapped partially.
829 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
831 fio->ft_nob = cl_page_size(obj);
833 lu_ref_add(&page->cp_reference, "fault", io);
838 /* return unlocked vmpage to avoid deadlocking */
841 #ifdef HAVE_VM_OP_FAULT
842 cfio->fault.ft_flags &= ~VM_FAULT_LOCKED;
847 static int vvp_io_fsync_start(const struct lu_env *env,
848 const struct cl_io_slice *ios)
850 /* we should mark TOWRITE bit to each dirty page in radix tree to
851 * verify pages have been written, but this is difficult because of
856 static int vvp_io_read_page(const struct lu_env *env,
857 const struct cl_io_slice *ios,
858 const struct cl_page_slice *slice)
860 struct cl_io *io = ios->cis_io;
861 struct cl_object *obj = slice->cpl_obj;
862 struct ccc_page *cp = cl2ccc_page(slice);
863 struct cl_page *page = slice->cpl_page;
864 struct inode *inode = ccc_object_inode(obj);
865 struct ll_sb_info *sbi = ll_i2sbi(inode);
866 struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd;
867 struct ll_readahead_state *ras = &fd->fd_ras;
868 cfs_page_t *vmpage = cp->cpg_page;
869 struct cl_2queue *queue = &io->ci_queue;
872 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
873 LASSERT(slice->cpl_obj == obj);
877 if (sbi->ll_ra_info.ra_max_pages_per_file &&
878 sbi->ll_ra_info.ra_max_pages)
879 ras_update(sbi, inode, ras, page->cp_index,
880 cp->cpg_defer_uptodate);
882 /* Sanity check whether the page is protected by a lock. */
883 rc = cl_page_is_under_lock(env, io, page);
885 CL_PAGE_HEADER(D_WARNING, env, page, "%s: %d\n",
886 rc == -ENODATA ? "without a lock" :
892 if (cp->cpg_defer_uptodate) {
894 cl_page_export(env, page, 1);
897 * Add page into the queue even when it is marked uptodate above.
898 * this will unlock it automatically as part of cl_page_list_disown().
900 cl_2queue_add(queue, page);
901 if (sbi->ll_ra_info.ra_max_pages_per_file &&
902 sbi->ll_ra_info.ra_max_pages)
903 ll_readahead(env, io, ras,
904 vmpage->mapping, &queue->c2_qin, fd->fd_flags);
909 static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io,
910 struct cl_page *page, struct ccc_page *cp,
911 enum cl_req_type crt)
913 struct cl_2queue *queue;
916 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
918 queue = &io->ci_queue;
919 cl_2queue_init_page(queue, page);
921 result = cl_io_submit_sync(env, io, crt, queue, 0);
922 LASSERT(cl_page_is_owned(page, io));
926 * in CRT_WRITE case page is left locked even in case of
929 cl_page_list_disown(env, io, &queue->c2_qin);
930 cl_2queue_fini(env, queue);
936 * Prepare partially written-to page for a write.
938 static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io,
939 struct cl_object *obj, struct cl_page *pg,
941 unsigned from, unsigned to)
943 struct cl_attr *attr = ccc_env_thread_attr(env);
944 loff_t offset = cl_offset(obj, pg->cp_index);
947 cl_object_attr_lock(obj);
948 result = cl_object_attr_get(env, obj, attr);
949 cl_object_attr_unlock(obj);
952 * If are writing to a new page, no need to read old data.
953 * The extent locking will have updated the KMS, and for our
954 * purposes here we can treat it like i_size.
956 if (attr->cat_kms <= offset) {
957 char *kaddr = ll_kmap_atomic(cp->cpg_page, KM_USER0);
959 memset(kaddr, 0, cl_page_size(obj));
960 ll_kunmap_atomic(kaddr, KM_USER0);
961 } else if (cp->cpg_defer_uptodate)
964 result = vvp_page_sync_io(env, io, pg, cp, CRT_READ);
966 * In older implementations, obdo_refresh_inode is called here
967 * to update the inode because the write might modify the
968 * object info at OST. However, this has been proven useless,
969 * since LVB functions will be called when user space program
970 * tries to retrieve inode attribute. Also, see bug 15909 for
974 cl_page_export(env, pg, 1);
979 static int vvp_io_prepare_write(const struct lu_env *env,
980 const struct cl_io_slice *ios,
981 const struct cl_page_slice *slice,
982 unsigned from, unsigned to)
984 struct cl_object *obj = slice->cpl_obj;
985 struct ccc_page *cp = cl2ccc_page(slice);
986 struct cl_page *pg = slice->cpl_page;
987 cfs_page_t *vmpage = cp->cpg_page;
993 LINVRNT(cl_page_is_vmlocked(env, pg));
994 LASSERT(vmpage->mapping->host == ccc_object_inode(obj));
998 CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to);
999 if (!PageUptodate(vmpage)) {
1001 * We're completely overwriting an existing page, so _don't_
1002 * set it up to date until commit_write
1004 if (from == 0 && to == CFS_PAGE_SIZE) {
1005 CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n");
1006 POISON_PAGE(page, 0x11);
1008 result = vvp_io_prepare_partial(env, ios->cis_io, obj,
1011 CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n");
1015 static int vvp_io_commit_write(const struct lu_env *env,
1016 const struct cl_io_slice *ios,
1017 const struct cl_page_slice *slice,
1018 unsigned from, unsigned to)
1020 struct cl_object *obj = slice->cpl_obj;
1021 struct cl_io *io = ios->cis_io;
1022 struct ccc_page *cp = cl2ccc_page(slice);
1023 struct cl_page *pg = slice->cpl_page;
1024 struct inode *inode = ccc_object_inode(obj);
1025 struct ll_sb_info *sbi = ll_i2sbi(inode);
1026 struct ll_inode_info *lli = ll_i2info(inode);
1027 cfs_page_t *vmpage = cp->cpg_page;
1035 LINVRNT(cl_page_is_vmlocked(env, pg));
1036 LASSERT(vmpage->mapping->host == inode);
1038 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "commiting page write\n");
1039 CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to);
1042 * queue a write for some time in the future the first time we
1045 * This is different from what other file systems do: they usually
1046 * just mark page (and some of its buffers) dirty and rely on
1047 * balance_dirty_pages() to start a write-back. Lustre wants write-back
1048 * to be started earlier for the following reasons:
1050 * (1) with a large number of clients we need to limit the amount
1051 * of cached data on the clients a lot;
1053 * (2) large compute jobs generally want compute-only then io-only
1054 * and the IO should complete as quickly as possible;
1056 * (3) IO is batched up to the RPC size and is async until the
1057 * client max cache is hit
1058 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
1061 if (!PageDirty(vmpage)) {
1062 tallyop = LPROC_LL_DIRTY_MISSES;
1063 result = cl_page_cache_add(env, io, pg, CRT_WRITE);
1065 /* page was added into cache successfully. */
1066 set_page_dirty(vmpage);
1067 vvp_write_pending(cl2ccc(obj), cp);
1068 } else if (result == -EDQUOT) {
1069 pgoff_t last_index = i_size_read(inode) >> CFS_PAGE_SHIFT;
1070 bool need_clip = true;
1073 * Client ran out of disk space grant. Possible
1076 * (a) do a sync write, renewing grant;
1078 * (b) stop writing on this stripe, switch to the
1081 * (b) is a part of "parallel io" design that is the
1082 * ultimate goal. (a) is what "old" client did, and
1083 * what the new code continues to do for the time
1086 if (last_index > pg->cp_index) {
1089 } else if (last_index == pg->cp_index) {
1090 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
1095 cl_page_clip(env, pg, 0, to);
1096 result = vvp_page_sync_io(env, io, pg, cp, CRT_WRITE);
1098 CERROR("Write page %lu of inode %p failed %d\n",
1099 pg->cp_index, inode, result);
1102 tallyop = LPROC_LL_DIRTY_HITS;
1105 ll_stats_ops_tally(sbi, tallyop, 1);
1107 /* Inode should be marked DIRTY even if no new page was marked DIRTY
1108 * because page could have been not flushed between 2 modifications.
1109 * It is important the file is marked DIRTY as soon as the I/O is done
1110 * Indeed, when cache is flushed, file could be already closed and it
1111 * is too late to warn the MDT.
1112 * It is acceptable that file is marked DIRTY even if I/O is dropped
1113 * for some reasons before being flushed to OST.
1116 spin_lock(&lli->lli_lock);
1117 lli->lli_flags |= LLIF_DATA_MODIFIED;
1118 spin_unlock(&lli->lli_lock);
1121 size = cl_offset(obj, pg->cp_index) + to;
1123 ll_inode_size_lock(inode);
1125 if (size > i_size_read(inode)) {
1126 cl_isize_write_nolock(inode, size);
1127 CDEBUG(D_VFSTRACE, DFID" updating i_size %lu\n",
1128 PFID(lu_object_fid(&obj->co_lu)),
1129 (unsigned long)size);
1131 cl_page_export(env, pg, 1);
1133 if (size > i_size_read(inode))
1134 cl_page_discard(env, io, pg);
1136 ll_inode_size_unlock(inode);
1140 static const struct cl_io_operations vvp_io_ops = {
1143 .cio_fini = vvp_io_read_fini,
1144 .cio_lock = vvp_io_read_lock,
1145 .cio_start = vvp_io_read_start,
1146 .cio_advance = ccc_io_advance
1149 .cio_fini = vvp_io_fini,
1150 .cio_lock = vvp_io_write_lock,
1151 .cio_start = vvp_io_write_start,
1152 .cio_advance = ccc_io_advance
1155 .cio_fini = vvp_io_setattr_fini,
1156 .cio_iter_init = vvp_io_setattr_iter_init,
1157 .cio_lock = vvp_io_setattr_lock,
1158 .cio_start = vvp_io_setattr_start,
1159 .cio_end = vvp_io_setattr_end
1162 .cio_fini = vvp_io_fault_fini,
1163 .cio_iter_init = vvp_io_fault_iter_init,
1164 .cio_lock = vvp_io_fault_lock,
1165 .cio_start = vvp_io_fault_start,
1166 .cio_end = ccc_io_end
1169 .cio_start = vvp_io_fsync_start,
1170 .cio_fini = vvp_io_fini
1173 .cio_fini = vvp_io_fini
1176 .cio_read_page = vvp_io_read_page,
1177 .cio_prepare_write = vvp_io_prepare_write,
1178 .cio_commit_write = vvp_io_commit_write
1181 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1184 struct vvp_io *vio = vvp_env_io(env);
1185 struct ccc_io *cio = ccc_env_io(env);
1186 struct inode *inode = ccc_object_inode(obj);
1189 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
1192 CL_IO_SLICE_CLEAN(cio, cui_cl);
1193 cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops);
1194 vio->cui_ra_window_set = 0;
1196 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1198 struct ll_inode_info *lli = ll_i2info(inode);
1200 count = io->u.ci_rw.crw_count;
1201 /* "If nbyte is 0, read() will return 0 and have no other
1202 * results." -- Single Unix Spec */
1206 cio->cui_tot_count = count;
1207 cio->cui_tot_nrsegs = 0;
1209 /* for read/write, we store the jobid in the inode, and
1210 * it'll be fetched by osc when building RPC.
1212 * it's not accurate if the file is shared by different
1215 lustre_get_jobid(lli->lli_jobid);
1216 } else if (io->ci_type == CIT_SETATTR) {
1217 if (!cl_io_is_trunc(io))
1218 io->ci_lockreq = CILR_MANDATORY;
1221 /* ignore layout change for generic CIT_MISC but not for glimpse.
1222 * io context for glimpse must set ci_verify_layout to true,
1223 * see cl_glimpse_size0() for details. */
1224 if (io->ci_type == CIT_MISC && !io->ci_verify_layout)
1225 io->ci_ignore_layout = 1;
1227 /* Enqueue layout lock and get layout version. We need to do this
1228 * even for operations requiring to open file, such as read and write,
1229 * because it might not grant layout lock in IT_OPEN. */
1230 if (result == 0 && !io->ci_ignore_layout) {
1231 result = ll_layout_refresh(inode, &cio->cui_layout_gen);
1232 if (result == -ENOENT)
1233 /* If the inode on MDS has been removed, but the objects
1234 * on OSTs haven't been destroyed (async unlink), layout
1235 * fetch will return -ENOENT, we'd ingore this error
1236 * and continue with dirty flush. LU-3230. */
1239 CERROR("%s: refresh file layout " DFID " error %d.\n",
1240 ll_get_fsname(inode->i_sb, NULL, 0),
1241 PFID(lu_object_fid(&obj->co_lu)), result);
1247 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
1248 const struct cl_io_slice *slice)
1250 /* Caling just for assertion */
1251 cl2ccc_io(env, slice);
1252 return vvp_env_io(env);