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)
362 * Only ll_inode_size_lock is taken at this level.
364 ll_inode_size_lock(inode);
365 result = vmtruncate(inode, size);
366 ll_inode_size_unlock(inode);
371 static int vvp_io_setattr_trunc(const struct lu_env *env,
372 const struct cl_io_slice *ios,
373 struct inode *inode, loff_t size)
375 inode_dio_wait(inode);
379 static int vvp_io_setattr_time(const struct lu_env *env,
380 const struct cl_io_slice *ios)
382 struct cl_io *io = ios->cis_io;
383 struct cl_object *obj = io->ci_obj;
384 struct cl_attr *attr = ccc_env_thread_attr(env);
386 unsigned valid = CAT_CTIME;
388 cl_object_attr_lock(obj);
389 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
390 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
391 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
394 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
395 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
398 result = cl_object_attr_set(env, obj, attr, valid);
399 cl_object_attr_unlock(obj);
404 static int vvp_io_setattr_start(const struct lu_env *env,
405 const struct cl_io_slice *ios)
407 struct cl_io *io = ios->cis_io;
408 struct inode *inode = ccc_object_inode(io->ci_obj);
411 mutex_lock(&inode->i_mutex);
412 if (cl_io_is_trunc(io))
413 result = vvp_io_setattr_trunc(env, ios, inode,
414 io->u.ci_setattr.sa_attr.lvb_size);
416 result = vvp_io_setattr_time(env, ios);
420 static void vvp_io_setattr_end(const struct lu_env *env,
421 const struct cl_io_slice *ios)
423 struct cl_io *io = ios->cis_io;
424 struct inode *inode = ccc_object_inode(io->ci_obj);
426 if (cl_io_is_trunc(io)) {
427 /* Truncate in memory pages - they must be clean pages
428 * because osc has already notified to destroy osc_extents. */
429 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
430 inode_dio_write_done(inode);
432 mutex_unlock(&inode->i_mutex);
435 static void vvp_io_setattr_fini(const struct lu_env *env,
436 const struct cl_io_slice *ios)
438 vvp_io_fini(env, ios);
441 #ifdef HAVE_FILE_READV
442 static ssize_t lustre_generic_file_read(struct file *file,
443 struct ccc_io *vio, loff_t *ppos)
445 return generic_file_readv(file, vio->cui_iov, vio->cui_nrsegs, ppos);
448 static ssize_t lustre_generic_file_write(struct file *file,
449 struct ccc_io *vio, loff_t *ppos)
451 return generic_file_writev(file, vio->cui_iov, vio->cui_nrsegs, ppos);
454 static ssize_t lustre_generic_file_read(struct file *file,
455 struct ccc_io *vio, loff_t *ppos)
457 return generic_file_aio_read(vio->cui_iocb, vio->cui_iov,
458 vio->cui_nrsegs, *ppos);
461 static ssize_t lustre_generic_file_write(struct file *file,
462 struct ccc_io *vio, loff_t *ppos)
464 return generic_file_aio_write(vio->cui_iocb, vio->cui_iov,
465 vio->cui_nrsegs, *ppos);
469 static int vvp_io_read_start(const struct lu_env *env,
470 const struct cl_io_slice *ios)
472 struct vvp_io *vio = cl2vvp_io(env, ios);
473 struct ccc_io *cio = cl2ccc_io(env, ios);
474 struct cl_io *io = ios->cis_io;
475 struct cl_object *obj = io->ci_obj;
476 struct inode *inode = ccc_object_inode(obj);
477 struct ll_ra_read *bead = &vio->cui_bead;
478 struct file *file = cio->cui_fd->fd_file;
481 loff_t pos = io->u.ci_rd.rd.crw_pos;
482 long cnt = io->u.ci_rd.rd.crw_count;
483 long tot = cio->cui_tot_count;
486 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
488 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
490 if (!can_populate_pages(env, io, inode))
493 result = ccc_prep_size(env, obj, io, pos, tot, &exceed);
496 else if (exceed != 0)
499 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
500 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
501 inode->i_ino, cnt, pos, i_size_read(inode));
503 /* turn off the kernel's read-ahead */
504 cio->cui_fd->fd_file->f_ra.ra_pages = 0;
506 /* initialize read-ahead window once per syscall */
507 if (!vio->cui_ra_window_set) {
508 vio->cui_ra_window_set = 1;
509 bead->lrr_start = cl_index(obj, pos);
511 * XXX: explicit CFS_PAGE_SIZE
513 bead->lrr_count = cl_index(obj, tot + CFS_PAGE_SIZE - 1);
514 ll_ra_read_in(file, bead);
519 switch (vio->cui_io_subtype) {
521 result = lustre_generic_file_read(file, cio, &pos);
523 #ifdef HAVE_KERNEL_SENDFILE
525 result = generic_file_sendfile(file, &pos, cnt,
526 vio->u.sendfile.cui_actor,
527 vio->u.sendfile.cui_target);
530 #ifdef HAVE_KERNEL_SPLICE_READ
532 result = generic_file_splice_read(file, &pos,
533 vio->u.splice.cui_pipe, cnt,
534 vio->u.splice.cui_flags);
535 /* LU-1109: do splice read stripe by stripe otherwise if it
536 * may make nfsd stuck if this read occupied all internal pipe
542 CERROR("Wrong IO type %u\n", vio->cui_io_subtype);
550 io->ci_nob += result;
551 ll_rw_stats_tally(ll_i2sbi(inode), current->pid, cio->cui_fd,
559 static void vvp_io_read_fini(const struct lu_env *env, const struct cl_io_slice *ios)
561 struct vvp_io *vio = cl2vvp_io(env, ios);
562 struct ccc_io *cio = cl2ccc_io(env, ios);
564 if (vio->cui_ra_window_set)
565 ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead);
567 vvp_io_fini(env, ios);
570 static int vvp_io_write_start(const struct lu_env *env,
571 const struct cl_io_slice *ios)
573 struct ccc_io *cio = cl2ccc_io(env, ios);
574 struct cl_io *io = ios->cis_io;
575 struct cl_object *obj = io->ci_obj;
576 struct inode *inode = ccc_object_inode(obj);
577 struct file *file = cio->cui_fd->fd_file;
579 loff_t pos = io->u.ci_wr.wr.crw_pos;
580 size_t cnt = io->u.ci_wr.wr.crw_count;
584 if (!can_populate_pages(env, io, inode))
587 if (cl_io_is_append(io)) {
589 * PARALLEL IO This has to be changed for parallel IO doing
590 * out-of-order writes.
592 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
593 #ifndef HAVE_FILE_WRITEV
594 cio->cui_iocb->ki_pos = pos;
598 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
600 if (cio->cui_iov == NULL) /* from a temp io in ll_cl_init(). */
603 result = lustre_generic_file_write(file, cio, &pos);
608 io->ci_nob += result;
609 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
610 cio->cui_fd, pos, result, WRITE);
617 #ifndef HAVE_VM_OP_FAULT
618 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
622 vmpage = filemap_nopage(cfio->ft_vma, cfio->nopage.ft_address,
623 cfio->nopage.ft_type);
625 if (vmpage == NOPAGE_SIGBUS) {
626 CDEBUG(D_PAGE, "got addr %lu type %lx - SIGBUS\n",
627 cfio->nopage.ft_address,(long)cfio->nopage.ft_type);
629 } else if (vmpage == NOPAGE_OOM) {
630 CDEBUG(D_PAGE, "got addr %lu type %lx - OOM\n",
631 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
635 LL_CDEBUG_PAGE(D_PAGE, vmpage, "got addr %lu type %lx\n",
636 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
638 cfio->ft_vmpage = vmpage;
644 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
646 struct vm_fault *vmf = cfio->fault.ft_vmf;
648 cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, vmf);
651 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
652 vmf->virtual_address);
653 if (unlikely(!(cfio->fault.ft_flags & VM_FAULT_LOCKED))) {
654 lock_page(vmf->page);
655 cfio->fault.ft_flags &= VM_FAULT_LOCKED;
658 cfio->ft_vmpage = vmf->page;
662 if (cfio->fault.ft_flags & VM_FAULT_SIGBUS) {
663 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
667 if (cfio->fault.ft_flags & VM_FAULT_OOM) {
668 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
672 if (cfio->fault.ft_flags & VM_FAULT_RETRY)
675 CERROR("unknow error in page fault %d!\n", cfio->fault.ft_flags);
681 static int vvp_io_fault_start(const struct lu_env *env,
682 const struct cl_io_slice *ios)
684 struct vvp_io *vio = cl2vvp_io(env, ios);
685 struct cl_io *io = ios->cis_io;
686 struct cl_object *obj = io->ci_obj;
687 struct inode *inode = ccc_object_inode(obj);
688 struct cl_fault_io *fio = &io->u.ci_fault;
689 struct vvp_fault_io *cfio = &vio->u.fault;
692 cfs_page_t *vmpage = NULL;
693 struct cl_page *page;
695 pgoff_t last; /* last page in a file data region */
697 if (fio->ft_executable &&
698 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
700 " changed while waiting for the page fault lock\n",
701 PFID(lu_object_fid(&obj->co_lu)));
703 /* offset of the last byte on the page */
704 offset = cl_offset(obj, fio->ft_index + 1) - 1;
705 LASSERT(cl_index(obj, offset) == fio->ft_index);
706 result = ccc_prep_size(env, obj, io, 0, offset + 1, NULL);
710 /* must return locked page */
711 if (fio->ft_mkwrite) {
712 LASSERT(cfio->ft_vmpage != NULL);
713 lock_page(cfio->ft_vmpage);
715 result = vvp_io_kernel_fault(cfio);
720 vmpage = cfio->ft_vmpage;
721 LASSERT(PageLocked(vmpage));
723 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
724 ll_invalidate_page(vmpage);
726 size = i_size_read(inode);
727 /* Though we have already held a cl_lock upon this page, but
728 * it still can be truncated locally. */
729 if (unlikely((vmpage->mapping != inode->i_mapping) ||
730 (page_offset(vmpage) > size))) {
731 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
733 /* return +1 to stop cl_io_loop() and ll_fault() will catch
735 GOTO(out, result = +1);
739 if (fio->ft_mkwrite ) {
742 * Capture the size while holding the lli_trunc_sem from above
743 * we want to make sure that we complete the mkwrite action
744 * while holding this lock. We need to make sure that we are
745 * not past the end of the file.
747 last_index = cl_index(obj, size - 1);
748 if (last_index < fio->ft_index) {
750 "llite: mkwrite and truncate race happened: "
752 vmpage->mapping,fio->ft_index,last_index);
754 * We need to return if we are
755 * passed the end of the file. This will propagate
756 * up the call stack to ll_page_mkwrite where
757 * we will return VM_FAULT_NOPAGE. Any non-negative
758 * value returned here will be silently
759 * converted to 0. If the vmpage->mapping is null
760 * the error code would be converted back to ENODATA
761 * in ll_page_mkwrite0. Thus we return -ENODATA
762 * to handle both cases
764 GOTO(out, result = -ENODATA);
768 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
770 GOTO(out, result = PTR_ERR(page));
772 /* if page is going to be written, we should add this page into cache
774 if (fio->ft_mkwrite) {
775 wait_on_page_writeback(vmpage);
776 if (set_page_dirty(vmpage)) {
779 /* vvp_page_assume() calls wait_on_page_writeback(). */
780 cl_page_assume(env, io, page);
782 cp = cl2ccc_page(cl_page_at(page, &vvp_device_type));
783 vvp_write_pending(cl2ccc(obj), cp);
785 /* Do not set Dirty bit here so that in case IO is
786 * started before the page is really made dirty, we
787 * still have chance to detect it. */
788 result = cl_page_cache_add(env, io, page, CRT_WRITE);
789 LASSERT(cl_page_is_owned(page, io));
793 cl_page_unmap(env, io, page);
794 cl_page_discard(env, io, page);
795 cl_page_disown(env, io, page);
797 cl_page_put(env, page);
799 /* we're in big trouble, what can we do now? */
800 if (result == -EDQUOT)
804 cl_page_disown(env, io, page);
808 last = cl_index(obj, size - 1);
810 * The ft_index is only used in the case of
811 * a mkwrite action. We need to check
812 * our assertions are correct, since
813 * we should have caught this above
815 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last);
816 if (fio->ft_index == last)
818 * Last page is mapped partially.
820 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
822 fio->ft_nob = cl_page_size(obj);
824 lu_ref_add(&page->cp_reference, "fault", io);
829 /* return unlocked vmpage to avoid deadlocking */
832 #ifdef HAVE_VM_OP_FAULT
833 cfio->fault.ft_flags &= ~VM_FAULT_LOCKED;
838 static int vvp_io_fsync_start(const struct lu_env *env,
839 const struct cl_io_slice *ios)
841 /* we should mark TOWRITE bit to each dirty page in radix tree to
842 * verify pages have been written, but this is difficult because of
847 static int vvp_io_read_page(const struct lu_env *env,
848 const struct cl_io_slice *ios,
849 const struct cl_page_slice *slice)
851 struct cl_io *io = ios->cis_io;
852 struct cl_object *obj = slice->cpl_obj;
853 struct ccc_page *cp = cl2ccc_page(slice);
854 struct cl_page *page = slice->cpl_page;
855 struct inode *inode = ccc_object_inode(obj);
856 struct ll_sb_info *sbi = ll_i2sbi(inode);
857 struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd;
858 struct ll_readahead_state *ras = &fd->fd_ras;
859 cfs_page_t *vmpage = cp->cpg_page;
860 struct cl_2queue *queue = &io->ci_queue;
863 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
864 LASSERT(slice->cpl_obj == obj);
868 if (sbi->ll_ra_info.ra_max_pages_per_file &&
869 sbi->ll_ra_info.ra_max_pages)
870 ras_update(sbi, inode, ras, page->cp_index,
871 cp->cpg_defer_uptodate);
873 /* Sanity check whether the page is protected by a lock. */
874 rc = cl_page_is_under_lock(env, io, page);
876 CL_PAGE_HEADER(D_WARNING, env, page, "%s: %d\n",
877 rc == -ENODATA ? "without a lock" :
883 if (cp->cpg_defer_uptodate) {
885 cl_page_export(env, page, 1);
888 * Add page into the queue even when it is marked uptodate above.
889 * this will unlock it automatically as part of cl_page_list_disown().
891 cl_2queue_add(queue, page);
892 if (sbi->ll_ra_info.ra_max_pages_per_file &&
893 sbi->ll_ra_info.ra_max_pages)
894 ll_readahead(env, io, ras,
895 vmpage->mapping, &queue->c2_qin, fd->fd_flags);
900 static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io,
901 struct cl_page *page, struct ccc_page *cp,
902 enum cl_req_type crt)
904 struct cl_2queue *queue;
907 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
909 queue = &io->ci_queue;
910 cl_2queue_init_page(queue, page);
912 result = cl_io_submit_sync(env, io, crt, queue, 0);
913 LASSERT(cl_page_is_owned(page, io));
917 * in CRT_WRITE case page is left locked even in case of
920 cl_page_list_disown(env, io, &queue->c2_qin);
921 cl_2queue_fini(env, queue);
927 * Prepare partially written-to page for a write.
929 static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io,
930 struct cl_object *obj, struct cl_page *pg,
932 unsigned from, unsigned to)
934 struct cl_attr *attr = ccc_env_thread_attr(env);
935 loff_t offset = cl_offset(obj, pg->cp_index);
938 cl_object_attr_lock(obj);
939 result = cl_object_attr_get(env, obj, attr);
940 cl_object_attr_unlock(obj);
943 * If are writing to a new page, no need to read old data.
944 * The extent locking will have updated the KMS, and for our
945 * purposes here we can treat it like i_size.
947 if (attr->cat_kms <= offset) {
948 char *kaddr = ll_kmap_atomic(cp->cpg_page, KM_USER0);
950 memset(kaddr, 0, cl_page_size(obj));
951 ll_kunmap_atomic(kaddr, KM_USER0);
952 } else if (cp->cpg_defer_uptodate)
955 result = vvp_page_sync_io(env, io, pg, cp, CRT_READ);
957 * In older implementations, obdo_refresh_inode is called here
958 * to update the inode because the write might modify the
959 * object info at OST. However, this has been proven useless,
960 * since LVB functions will be called when user space program
961 * tries to retrieve inode attribute. Also, see bug 15909 for
965 cl_page_export(env, pg, 1);
970 static int vvp_io_prepare_write(const struct lu_env *env,
971 const struct cl_io_slice *ios,
972 const struct cl_page_slice *slice,
973 unsigned from, unsigned to)
975 struct cl_object *obj = slice->cpl_obj;
976 struct ccc_page *cp = cl2ccc_page(slice);
977 struct cl_page *pg = slice->cpl_page;
978 cfs_page_t *vmpage = cp->cpg_page;
984 LINVRNT(cl_page_is_vmlocked(env, pg));
985 LASSERT(vmpage->mapping->host == ccc_object_inode(obj));
989 CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to);
990 if (!PageUptodate(vmpage)) {
992 * We're completely overwriting an existing page, so _don't_
993 * set it up to date until commit_write
995 if (from == 0 && to == CFS_PAGE_SIZE) {
996 CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n");
997 POISON_PAGE(page, 0x11);
999 result = vvp_io_prepare_partial(env, ios->cis_io, obj,
1002 CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n");
1006 static int vvp_io_commit_write(const struct lu_env *env,
1007 const struct cl_io_slice *ios,
1008 const struct cl_page_slice *slice,
1009 unsigned from, unsigned to)
1011 struct cl_object *obj = slice->cpl_obj;
1012 struct cl_io *io = ios->cis_io;
1013 struct ccc_page *cp = cl2ccc_page(slice);
1014 struct cl_page *pg = slice->cpl_page;
1015 struct inode *inode = ccc_object_inode(obj);
1016 struct ll_sb_info *sbi = ll_i2sbi(inode);
1017 struct ll_inode_info *lli = ll_i2info(inode);
1018 cfs_page_t *vmpage = cp->cpg_page;
1026 LINVRNT(cl_page_is_vmlocked(env, pg));
1027 LASSERT(vmpage->mapping->host == inode);
1029 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "commiting page write\n");
1030 CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to);
1033 * queue a write for some time in the future the first time we
1036 * This is different from what other file systems do: they usually
1037 * just mark page (and some of its buffers) dirty and rely on
1038 * balance_dirty_pages() to start a write-back. Lustre wants write-back
1039 * to be started earlier for the following reasons:
1041 * (1) with a large number of clients we need to limit the amount
1042 * of cached data on the clients a lot;
1044 * (2) large compute jobs generally want compute-only then io-only
1045 * and the IO should complete as quickly as possible;
1047 * (3) IO is batched up to the RPC size and is async until the
1048 * client max cache is hit
1049 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
1052 if (!PageDirty(vmpage)) {
1053 tallyop = LPROC_LL_DIRTY_MISSES;
1054 result = cl_page_cache_add(env, io, pg, CRT_WRITE);
1056 /* page was added into cache successfully. */
1057 set_page_dirty(vmpage);
1058 vvp_write_pending(cl2ccc(obj), cp);
1059 } else if (result == -EDQUOT) {
1060 pgoff_t last_index = i_size_read(inode) >> CFS_PAGE_SHIFT;
1061 bool need_clip = true;
1064 * Client ran out of disk space grant. Possible
1067 * (a) do a sync write, renewing grant;
1069 * (b) stop writing on this stripe, switch to the
1072 * (b) is a part of "parallel io" design that is the
1073 * ultimate goal. (a) is what "old" client did, and
1074 * what the new code continues to do for the time
1077 if (last_index > pg->cp_index) {
1080 } else if (last_index == pg->cp_index) {
1081 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
1086 cl_page_clip(env, pg, 0, to);
1087 result = vvp_page_sync_io(env, io, pg, cp, CRT_WRITE);
1089 CERROR("Write page %lu of inode %p failed %d\n",
1090 pg->cp_index, inode, result);
1093 tallyop = LPROC_LL_DIRTY_HITS;
1096 ll_stats_ops_tally(sbi, tallyop, 1);
1098 /* Inode should be marked DIRTY even if no new page was marked DIRTY
1099 * because page could have been not flushed between 2 modifications.
1100 * It is important the file is marked DIRTY as soon as the I/O is done
1101 * Indeed, when cache is flushed, file could be already closed and it
1102 * is too late to warn the MDT.
1103 * It is acceptable that file is marked DIRTY even if I/O is dropped
1104 * for some reasons before being flushed to OST.
1107 spin_lock(&lli->lli_lock);
1108 lli->lli_flags |= LLIF_DATA_MODIFIED;
1109 spin_unlock(&lli->lli_lock);
1112 size = cl_offset(obj, pg->cp_index) + to;
1114 ll_inode_size_lock(inode);
1116 if (size > i_size_read(inode)) {
1117 cl_isize_write_nolock(inode, size);
1118 CDEBUG(D_VFSTRACE, DFID" updating i_size %lu\n",
1119 PFID(lu_object_fid(&obj->co_lu)),
1120 (unsigned long)size);
1122 cl_page_export(env, pg, 1);
1124 if (size > i_size_read(inode))
1125 cl_page_discard(env, io, pg);
1127 ll_inode_size_unlock(inode);
1131 static const struct cl_io_operations vvp_io_ops = {
1134 .cio_fini = vvp_io_read_fini,
1135 .cio_lock = vvp_io_read_lock,
1136 .cio_start = vvp_io_read_start,
1137 .cio_advance = ccc_io_advance
1140 .cio_fini = vvp_io_fini,
1141 .cio_lock = vvp_io_write_lock,
1142 .cio_start = vvp_io_write_start,
1143 .cio_advance = ccc_io_advance
1146 .cio_fini = vvp_io_setattr_fini,
1147 .cio_iter_init = vvp_io_setattr_iter_init,
1148 .cio_lock = vvp_io_setattr_lock,
1149 .cio_start = vvp_io_setattr_start,
1150 .cio_end = vvp_io_setattr_end
1153 .cio_fini = vvp_io_fault_fini,
1154 .cio_iter_init = vvp_io_fault_iter_init,
1155 .cio_lock = vvp_io_fault_lock,
1156 .cio_start = vvp_io_fault_start,
1157 .cio_end = ccc_io_end
1160 .cio_start = vvp_io_fsync_start,
1161 .cio_fini = vvp_io_fini
1164 .cio_fini = vvp_io_fini
1167 .cio_read_page = vvp_io_read_page,
1168 .cio_prepare_write = vvp_io_prepare_write,
1169 .cio_commit_write = vvp_io_commit_write
1172 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1175 struct vvp_io *vio = vvp_env_io(env);
1176 struct ccc_io *cio = ccc_env_io(env);
1177 struct inode *inode = ccc_object_inode(obj);
1180 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
1183 CL_IO_SLICE_CLEAN(cio, cui_cl);
1184 cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops);
1185 vio->cui_ra_window_set = 0;
1187 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1189 struct ll_inode_info *lli = ll_i2info(inode);
1191 count = io->u.ci_rw.crw_count;
1192 /* "If nbyte is 0, read() will return 0 and have no other
1193 * results." -- Single Unix Spec */
1197 cio->cui_tot_count = count;
1198 cio->cui_tot_nrsegs = 0;
1200 /* for read/write, we store the jobid in the inode, and
1201 * it'll be fetched by osc when building RPC.
1203 * it's not accurate if the file is shared by different
1206 lustre_get_jobid(lli->lli_jobid);
1207 } else if (io->ci_type == CIT_SETATTR) {
1208 if (!cl_io_is_trunc(io))
1209 io->ci_lockreq = CILR_MANDATORY;
1212 /* ignore layout change for generic CIT_MISC but not for glimpse.
1213 * io context for glimpse must set ci_verify_layout to true,
1214 * see cl_glimpse_size0() for details. */
1215 if (io->ci_type == CIT_MISC && !io->ci_verify_layout)
1216 io->ci_ignore_layout = 1;
1218 /* Enqueue layout lock and get layout version. We need to do this
1219 * even for operations requiring to open file, such as read and write,
1220 * because it might not grant layout lock in IT_OPEN. */
1221 if (result == 0 && !io->ci_ignore_layout) {
1222 result = ll_layout_refresh(inode, &cio->cui_layout_gen);
1223 if (result == -ENOENT)
1224 /* If the inode on MDS has been removed, but the objects
1225 * on OSTs haven't been destroyed (async unlink), layout
1226 * fetch will return -ENOENT, we'd ingore this error
1227 * and continue with dirty flush. LU-3230. */
1230 CERROR("%s: refresh file layout " DFID " error %d.\n",
1231 ll_get_fsname(inode->i_sb, NULL, 0),
1232 PFID(lu_object_fid(&obj->co_lu)), result);
1238 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
1239 const struct cl_io_slice *slice)
1241 /* Caling just for assertion */
1242 cl2ccc_io(env, slice);
1243 return vvp_env_io(env);