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);
410 mutex_lock(&inode->i_mutex);
411 if (cl_io_is_trunc(io))
412 return vvp_io_setattr_trunc(env, ios, inode,
413 io->u.ci_setattr.sa_attr.lvb_size);
415 return 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 #ifdef HAVE_FILE_READV
440 static ssize_t lustre_generic_file_read(struct file *file,
441 struct ccc_io *vio, loff_t *ppos)
443 return generic_file_readv(file, vio->cui_iov, 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_writev(file, vio->cui_iov, vio->cui_nrsegs, ppos);
452 static ssize_t lustre_generic_file_read(struct file *file,
453 struct ccc_io *vio, loff_t *ppos)
455 return generic_file_aio_read(vio->cui_iocb, vio->cui_iov,
456 vio->cui_nrsegs, *ppos);
459 static ssize_t lustre_generic_file_write(struct file *file,
460 struct ccc_io *vio, loff_t *ppos)
462 return generic_file_aio_write(vio->cui_iocb, vio->cui_iov,
463 vio->cui_nrsegs, *ppos);
467 static int vvp_io_read_start(const struct lu_env *env,
468 const struct cl_io_slice *ios)
470 struct vvp_io *vio = cl2vvp_io(env, ios);
471 struct ccc_io *cio = cl2ccc_io(env, ios);
472 struct cl_io *io = ios->cis_io;
473 struct cl_object *obj = io->ci_obj;
474 struct inode *inode = ccc_object_inode(obj);
475 struct ll_ra_read *bead = &vio->cui_bead;
476 struct file *file = cio->cui_fd->fd_file;
479 loff_t pos = io->u.ci_rd.rd.crw_pos;
480 long cnt = io->u.ci_rd.rd.crw_count;
481 long tot = cio->cui_tot_count;
484 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
486 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
488 if (!can_populate_pages(env, io, inode))
491 result = ccc_prep_size(env, obj, io, pos, tot, &exceed);
494 else if (exceed != 0)
497 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
498 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
499 inode->i_ino, cnt, pos, i_size_read(inode));
501 /* turn off the kernel's read-ahead */
502 cio->cui_fd->fd_file->f_ra.ra_pages = 0;
504 /* initialize read-ahead window once per syscall */
505 if (!vio->cui_ra_window_set) {
506 vio->cui_ra_window_set = 1;
507 bead->lrr_start = cl_index(obj, pos);
509 * XXX: explicit CFS_PAGE_SIZE
511 bead->lrr_count = cl_index(obj, tot + CFS_PAGE_SIZE - 1);
512 ll_ra_read_in(file, bead);
517 switch (vio->cui_io_subtype) {
519 result = lustre_generic_file_read(file, cio, &pos);
521 #ifdef HAVE_KERNEL_SENDFILE
523 result = generic_file_sendfile(file, &pos, cnt,
524 vio->u.sendfile.cui_actor,
525 vio->u.sendfile.cui_target);
528 #ifdef HAVE_KERNEL_SPLICE_READ
530 result = generic_file_splice_read(file, &pos,
531 vio->u.splice.cui_pipe, cnt,
532 vio->u.splice.cui_flags);
533 /* LU-1109: do splice read stripe by stripe otherwise if it
534 * may make nfsd stuck if this read occupied all internal pipe
540 CERROR("Wrong IO type %u\n", vio->cui_io_subtype);
548 io->ci_nob += result;
549 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
550 cio->cui_fd, pos, result, 0);
556 static void vvp_io_read_fini(const struct lu_env *env, const struct cl_io_slice *ios)
558 struct vvp_io *vio = cl2vvp_io(env, ios);
559 struct ccc_io *cio = cl2ccc_io(env, ios);
561 if (vio->cui_ra_window_set)
562 ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead);
564 vvp_io_fini(env, ios);
567 static int vvp_io_write_start(const struct lu_env *env,
568 const struct cl_io_slice *ios)
570 struct ccc_io *cio = cl2ccc_io(env, ios);
571 struct cl_io *io = ios->cis_io;
572 struct cl_object *obj = io->ci_obj;
573 struct inode *inode = ccc_object_inode(obj);
574 struct file *file = cio->cui_fd->fd_file;
576 loff_t pos = io->u.ci_wr.wr.crw_pos;
577 size_t cnt = io->u.ci_wr.wr.crw_count;
581 if (!can_populate_pages(env, io, inode))
584 if (cl_io_is_append(io)) {
586 * PARALLEL IO This has to be changed for parallel IO doing
587 * out-of-order writes.
589 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
590 #ifndef HAVE_FILE_WRITEV
591 cio->cui_iocb->ki_pos = pos;
595 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
597 if (cio->cui_iov == NULL) /* from a temp io in ll_cl_init(). */
600 result = lustre_generic_file_write(file, cio, &pos);
605 io->ci_nob += result;
606 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
607 cio->cui_fd, pos, result, 0);
613 #ifndef HAVE_VM_OP_FAULT
614 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
618 vmpage = filemap_nopage(cfio->ft_vma, cfio->nopage.ft_address,
619 cfio->nopage.ft_type);
621 if (vmpage == NOPAGE_SIGBUS) {
622 CDEBUG(D_PAGE, "got addr %lu type %lx - SIGBUS\n",
623 cfio->nopage.ft_address,(long)cfio->nopage.ft_type);
625 } else if (vmpage == NOPAGE_OOM) {
626 CDEBUG(D_PAGE, "got addr %lu type %lx - OOM\n",
627 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
631 LL_CDEBUG_PAGE(D_PAGE, vmpage, "got addr %lu type %lx\n",
632 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
634 cfio->ft_vmpage = vmpage;
640 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
642 struct vm_fault *vmf = cfio->fault.ft_vmf;
644 cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, vmf);
647 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
648 vmf->virtual_address);
649 if (unlikely(!(cfio->fault.ft_flags & VM_FAULT_LOCKED))) {
650 lock_page(vmf->page);
651 cfio->fault.ft_flags &= VM_FAULT_LOCKED;
654 cfio->ft_vmpage = vmf->page;
658 if (cfio->fault.ft_flags & VM_FAULT_SIGBUS) {
659 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
663 if (cfio->fault.ft_flags & VM_FAULT_OOM) {
664 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
668 if (cfio->fault.ft_flags & VM_FAULT_RETRY)
671 CERROR("unknow error in page fault %d!\n", cfio->fault.ft_flags);
677 static int vvp_io_fault_start(const struct lu_env *env,
678 const struct cl_io_slice *ios)
680 struct vvp_io *vio = cl2vvp_io(env, ios);
681 struct cl_io *io = ios->cis_io;
682 struct cl_object *obj = io->ci_obj;
683 struct inode *inode = ccc_object_inode(obj);
684 struct cl_fault_io *fio = &io->u.ci_fault;
685 struct vvp_fault_io *cfio = &vio->u.fault;
688 cfs_page_t *vmpage = NULL;
689 struct cl_page *page;
691 pgoff_t last; /* last page in a file data region */
693 if (fio->ft_executable &&
694 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
696 " changed while waiting for the page fault lock\n",
697 PFID(lu_object_fid(&obj->co_lu)));
699 /* offset of the last byte on the page */
700 offset = cl_offset(obj, fio->ft_index + 1) - 1;
701 LASSERT(cl_index(obj, offset) == fio->ft_index);
702 result = ccc_prep_size(env, obj, io, 0, offset + 1, NULL);
706 /* must return locked page */
707 if (fio->ft_mkwrite) {
708 LASSERT(cfio->ft_vmpage != NULL);
709 lock_page(cfio->ft_vmpage);
711 result = vvp_io_kernel_fault(cfio);
716 vmpage = cfio->ft_vmpage;
717 LASSERT(PageLocked(vmpage));
719 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
720 ll_invalidate_page(vmpage);
722 size = i_size_read(inode);
723 /* Though we have already held a cl_lock upon this page, but
724 * it still can be truncated locally. */
725 if (unlikely((vmpage->mapping != inode->i_mapping) ||
726 (page_offset(vmpage) > size))) {
727 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
729 /* return +1 to stop cl_io_loop() and ll_fault() will catch
731 GOTO(out, result = +1);
735 if (fio->ft_mkwrite ) {
738 * Capture the size while holding the lli_trunc_sem from above
739 * we want to make sure that we complete the mkwrite action
740 * while holding this lock. We need to make sure that we are
741 * not past the end of the file.
743 last_index = cl_index(obj, size - 1);
744 if (last_index < fio->ft_index) {
746 "llite: mkwrite and truncate race happened: "
748 vmpage->mapping,fio->ft_index,last_index);
750 * We need to return if we are
751 * passed the end of the file. This will propagate
752 * up the call stack to ll_page_mkwrite where
753 * we will return VM_FAULT_NOPAGE. Any non-negative
754 * value returned here will be silently
755 * converted to 0. If the vmpage->mapping is null
756 * the error code would be converted back to ENODATA
757 * in ll_page_mkwrite0. Thus we return -ENODATA
758 * to handle both cases
760 GOTO(out, result = -ENODATA);
764 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
766 GOTO(out, result = PTR_ERR(page));
768 /* if page is going to be written, we should add this page into cache
770 if (fio->ft_mkwrite) {
771 wait_on_page_writeback(vmpage);
772 if (set_page_dirty(vmpage)) {
775 /* vvp_page_assume() calls wait_on_page_writeback(). */
776 cl_page_assume(env, io, page);
778 cp = cl2ccc_page(cl_page_at(page, &vvp_device_type));
779 vvp_write_pending(cl2ccc(obj), cp);
781 /* Do not set Dirty bit here so that in case IO is
782 * started before the page is really made dirty, we
783 * still have chance to detect it. */
784 result = cl_page_cache_add(env, io, page, CRT_WRITE);
785 LASSERT(cl_page_is_owned(page, io));
789 cl_page_unmap(env, io, page);
790 cl_page_discard(env, io, page);
791 cl_page_disown(env, io, page);
793 cl_page_put(env, page);
795 /* we're in big trouble, what can we do now? */
796 if (result == -EDQUOT)
800 cl_page_disown(env, io, page);
804 last = cl_index(obj, size - 1);
806 * The ft_index is only used in the case of
807 * a mkwrite action. We need to check
808 * our assertions are correct, since
809 * we should have caught this above
811 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last);
812 if (fio->ft_index == last)
814 * Last page is mapped partially.
816 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
818 fio->ft_nob = cl_page_size(obj);
820 lu_ref_add(&page->cp_reference, "fault", io);
825 /* return unlocked vmpage to avoid deadlocking */
828 #ifdef HAVE_VM_OP_FAULT
829 cfio->fault.ft_flags &= ~VM_FAULT_LOCKED;
834 static int vvp_io_fsync_start(const struct lu_env *env,
835 const struct cl_io_slice *ios)
837 /* we should mark TOWRITE bit to each dirty page in radix tree to
838 * verify pages have been written, but this is difficult because of
843 static int vvp_io_read_page(const struct lu_env *env,
844 const struct cl_io_slice *ios,
845 const struct cl_page_slice *slice)
847 struct cl_io *io = ios->cis_io;
848 struct cl_object *obj = slice->cpl_obj;
849 struct ccc_page *cp = cl2ccc_page(slice);
850 struct cl_page *page = slice->cpl_page;
851 struct inode *inode = ccc_object_inode(obj);
852 struct ll_sb_info *sbi = ll_i2sbi(inode);
853 struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd;
854 struct ll_readahead_state *ras = &fd->fd_ras;
855 cfs_page_t *vmpage = cp->cpg_page;
856 struct cl_2queue *queue = &io->ci_queue;
859 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
860 LASSERT(slice->cpl_obj == obj);
864 if (sbi->ll_ra_info.ra_max_pages_per_file &&
865 sbi->ll_ra_info.ra_max_pages)
866 ras_update(sbi, inode, ras, page->cp_index,
867 cp->cpg_defer_uptodate);
869 /* Sanity check whether the page is protected by a lock. */
870 rc = cl_page_is_under_lock(env, io, page);
872 CL_PAGE_HEADER(D_WARNING, env, page, "%s: %d\n",
873 rc == -ENODATA ? "without a lock" :
879 if (cp->cpg_defer_uptodate) {
881 cl_page_export(env, page, 1);
884 * Add page into the queue even when it is marked uptodate above.
885 * this will unlock it automatically as part of cl_page_list_disown().
887 cl_2queue_add(queue, page);
888 if (sbi->ll_ra_info.ra_max_pages_per_file &&
889 sbi->ll_ra_info.ra_max_pages)
890 ll_readahead(env, io, ras,
891 vmpage->mapping, &queue->c2_qin, fd->fd_flags);
896 static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io,
897 struct cl_page *page, struct ccc_page *cp,
898 enum cl_req_type crt)
900 struct cl_2queue *queue;
903 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
905 queue = &io->ci_queue;
906 cl_2queue_init_page(queue, page);
908 result = cl_io_submit_sync(env, io, crt, queue, 0);
909 LASSERT(cl_page_is_owned(page, io));
913 * in CRT_WRITE case page is left locked even in case of
916 cl_page_list_disown(env, io, &queue->c2_qin);
917 cl_2queue_fini(env, queue);
923 * Prepare partially written-to page for a write.
925 static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io,
926 struct cl_object *obj, struct cl_page *pg,
928 unsigned from, unsigned to)
930 struct cl_attr *attr = ccc_env_thread_attr(env);
931 loff_t offset = cl_offset(obj, pg->cp_index);
934 cl_object_attr_lock(obj);
935 result = cl_object_attr_get(env, obj, attr);
936 cl_object_attr_unlock(obj);
939 * If are writing to a new page, no need to read old data.
940 * The extent locking will have updated the KMS, and for our
941 * purposes here we can treat it like i_size.
943 if (attr->cat_kms <= offset) {
944 char *kaddr = ll_kmap_atomic(cp->cpg_page, KM_USER0);
946 memset(kaddr, 0, cl_page_size(obj));
947 ll_kunmap_atomic(kaddr, KM_USER0);
948 } else if (cp->cpg_defer_uptodate)
951 result = vvp_page_sync_io(env, io, pg, cp, CRT_READ);
953 * In older implementations, obdo_refresh_inode is called here
954 * to update the inode because the write might modify the
955 * object info at OST. However, this has been proven useless,
956 * since LVB functions will be called when user space program
957 * tries to retrieve inode attribute. Also, see bug 15909 for
961 cl_page_export(env, pg, 1);
966 static int vvp_io_prepare_write(const struct lu_env *env,
967 const struct cl_io_slice *ios,
968 const struct cl_page_slice *slice,
969 unsigned from, unsigned to)
971 struct cl_object *obj = slice->cpl_obj;
972 struct ccc_page *cp = cl2ccc_page(slice);
973 struct cl_page *pg = slice->cpl_page;
974 cfs_page_t *vmpage = cp->cpg_page;
980 LINVRNT(cl_page_is_vmlocked(env, pg));
981 LASSERT(vmpage->mapping->host == ccc_object_inode(obj));
985 CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to);
986 if (!PageUptodate(vmpage)) {
988 * We're completely overwriting an existing page, so _don't_
989 * set it up to date until commit_write
991 if (from == 0 && to == CFS_PAGE_SIZE) {
992 CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n");
993 POISON_PAGE(page, 0x11);
995 result = vvp_io_prepare_partial(env, ios->cis_io, obj,
998 CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n");
1002 static int vvp_io_commit_write(const struct lu_env *env,
1003 const struct cl_io_slice *ios,
1004 const struct cl_page_slice *slice,
1005 unsigned from, unsigned to)
1007 struct cl_object *obj = slice->cpl_obj;
1008 struct cl_io *io = ios->cis_io;
1009 struct ccc_page *cp = cl2ccc_page(slice);
1010 struct cl_page *pg = slice->cpl_page;
1011 struct inode *inode = ccc_object_inode(obj);
1012 struct ll_sb_info *sbi = ll_i2sbi(inode);
1013 struct ll_inode_info *lli = ll_i2info(inode);
1014 cfs_page_t *vmpage = cp->cpg_page;
1022 LINVRNT(cl_page_is_vmlocked(env, pg));
1023 LASSERT(vmpage->mapping->host == inode);
1025 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "commiting page write\n");
1026 CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to);
1029 * queue a write for some time in the future the first time we
1032 * This is different from what other file systems do: they usually
1033 * just mark page (and some of its buffers) dirty and rely on
1034 * balance_dirty_pages() to start a write-back. Lustre wants write-back
1035 * to be started earlier for the following reasons:
1037 * (1) with a large number of clients we need to limit the amount
1038 * of cached data on the clients a lot;
1040 * (2) large compute jobs generally want compute-only then io-only
1041 * and the IO should complete as quickly as possible;
1043 * (3) IO is batched up to the RPC size and is async until the
1044 * client max cache is hit
1045 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
1048 if (!PageDirty(vmpage)) {
1049 tallyop = LPROC_LL_DIRTY_MISSES;
1050 result = cl_page_cache_add(env, io, pg, CRT_WRITE);
1052 /* page was added into cache successfully. */
1053 set_page_dirty(vmpage);
1054 vvp_write_pending(cl2ccc(obj), cp);
1055 } else if (result == -EDQUOT) {
1056 pgoff_t last_index = i_size_read(inode) >> CFS_PAGE_SHIFT;
1057 bool need_clip = true;
1060 * Client ran out of disk space grant. Possible
1063 * (a) do a sync write, renewing grant;
1065 * (b) stop writing on this stripe, switch to the
1068 * (b) is a part of "parallel io" design that is the
1069 * ultimate goal. (a) is what "old" client did, and
1070 * what the new code continues to do for the time
1073 if (last_index > pg->cp_index) {
1076 } else if (last_index == pg->cp_index) {
1077 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
1082 cl_page_clip(env, pg, 0, to);
1083 result = vvp_page_sync_io(env, io, pg, cp, CRT_WRITE);
1085 CERROR("Write page %lu of inode %p failed %d\n",
1086 pg->cp_index, inode, result);
1089 tallyop = LPROC_LL_DIRTY_HITS;
1092 ll_stats_ops_tally(sbi, tallyop, 1);
1094 /* Inode should be marked DIRTY even if no new page was marked DIRTY
1095 * because page could have been not flushed between 2 modifications.
1096 * It is important the file is marked DIRTY as soon as the I/O is done
1097 * Indeed, when cache is flushed, file could be already closed and it
1098 * is too late to warn the MDT.
1099 * It is acceptable that file is marked DIRTY even if I/O is dropped
1100 * for some reasons before being flushed to OST.
1103 spin_lock(&lli->lli_lock);
1104 lli->lli_flags |= LLIF_DATA_MODIFIED;
1105 spin_unlock(&lli->lli_lock);
1108 size = cl_offset(obj, pg->cp_index) + to;
1110 ll_inode_size_lock(inode);
1112 if (size > i_size_read(inode)) {
1113 cl_isize_write_nolock(inode, size);
1114 CDEBUG(D_VFSTRACE, DFID" updating i_size %lu\n",
1115 PFID(lu_object_fid(&obj->co_lu)),
1116 (unsigned long)size);
1118 cl_page_export(env, pg, 1);
1120 if (size > i_size_read(inode))
1121 cl_page_discard(env, io, pg);
1123 ll_inode_size_unlock(inode);
1127 static const struct cl_io_operations vvp_io_ops = {
1130 .cio_fini = vvp_io_read_fini,
1131 .cio_lock = vvp_io_read_lock,
1132 .cio_start = vvp_io_read_start,
1133 .cio_advance = ccc_io_advance
1136 .cio_fini = vvp_io_fini,
1137 .cio_lock = vvp_io_write_lock,
1138 .cio_start = vvp_io_write_start,
1139 .cio_advance = ccc_io_advance
1142 .cio_fini = vvp_io_setattr_fini,
1143 .cio_iter_init = vvp_io_setattr_iter_init,
1144 .cio_lock = vvp_io_setattr_lock,
1145 .cio_start = vvp_io_setattr_start,
1146 .cio_end = vvp_io_setattr_end
1149 .cio_fini = vvp_io_fault_fini,
1150 .cio_iter_init = vvp_io_fault_iter_init,
1151 .cio_lock = vvp_io_fault_lock,
1152 .cio_start = vvp_io_fault_start,
1153 .cio_end = ccc_io_end
1156 .cio_start = vvp_io_fsync_start,
1157 .cio_fini = vvp_io_fini
1160 .cio_fini = vvp_io_fini
1163 .cio_read_page = vvp_io_read_page,
1164 .cio_prepare_write = vvp_io_prepare_write,
1165 .cio_commit_write = vvp_io_commit_write
1168 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1171 struct vvp_io *vio = vvp_env_io(env);
1172 struct ccc_io *cio = ccc_env_io(env);
1173 struct inode *inode = ccc_object_inode(obj);
1176 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
1179 CL_IO_SLICE_CLEAN(cio, cui_cl);
1180 cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops);
1181 vio->cui_ra_window_set = 0;
1183 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1185 struct ll_inode_info *lli = ll_i2info(inode);
1187 count = io->u.ci_rw.crw_count;
1188 /* "If nbyte is 0, read() will return 0 and have no other
1189 * results." -- Single Unix Spec */
1193 cio->cui_tot_count = count;
1194 cio->cui_tot_nrsegs = 0;
1196 /* for read/write, we store the jobid in the inode, and
1197 * it'll be fetched by osc when building RPC.
1199 * it's not accurate if the file is shared by different
1202 lustre_get_jobid(lli->lli_jobid);
1203 } else if (io->ci_type == CIT_SETATTR) {
1204 if (!cl_io_is_trunc(io))
1205 io->ci_lockreq = CILR_MANDATORY;
1208 /* ignore layout change for generic CIT_MISC but not for glimpse.
1209 * io context for glimpse must set ci_verify_layout to true,
1210 * see cl_glimpse_size0() for details. */
1211 if (io->ci_type == CIT_MISC && !io->ci_verify_layout)
1212 io->ci_ignore_layout = 1;
1214 /* Enqueue layout lock and get layout version. We need to do this
1215 * even for operations requiring to open file, such as read and write,
1216 * because it might not grant layout lock in IT_OPEN. */
1217 if (result == 0 && !io->ci_ignore_layout) {
1218 result = ll_layout_refresh(inode, &cio->cui_layout_gen);
1219 if (result == -ENOENT)
1220 /* If the inode on MDS has been removed, but the objects
1221 * on OSTs haven't been destroyed (async unlink), layout
1222 * fetch will return -ENOENT, we'd ingore this error
1223 * and continue with dirty flush. LU-3230. */
1226 CERROR("%s: refresh file layout " DFID " error %d.\n",
1227 ll_get_fsname(inode->i_sb, NULL, 0),
1228 PFID(lu_object_fid(&obj->co_lu)), result);
1234 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
1235 const struct cl_io_slice *slice)
1237 /* Caling just for assertion */
1238 cl2ccc_io(env, slice);
1239 return vvp_env_io(env);