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);
124 struct inode *inode = ccc_object_inode(obj);
126 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
128 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
129 "restore needed %d\n",
130 PFID(lu_object_fid(&obj->co_lu)),
131 io->ci_ignore_layout, io->ci_verify_layout,
132 cio->cui_layout_gen, io->ci_restore_needed);
134 if (io->ci_restore_needed == 1) {
137 /* file was detected release, we need to restore it
138 * before finishing the io
140 rc = ll_layout_restore(inode, 0, OBD_OBJECT_EOF);
141 /* if restore registration failed, no restart,
142 * we will return -ENODATA */
143 /* The layout will change after restore, so we need to
144 * block on layout lock hold by the MDT
145 * as MDT will not send new layout in lvb (see LU-3124)
146 * we have to explicitly fetch it, all this will be done
147 * by ll_layout_refresh()
150 io->ci_restore_needed = 0;
151 io->ci_need_restart = 1;
152 io->ci_verify_layout = 1;
154 io->ci_restore_needed = 1;
155 io->ci_need_restart = 0;
156 io->ci_verify_layout = 0;
161 if (!io->ci_ignore_layout && io->ci_verify_layout) {
164 /* check layout version */
165 ll_layout_refresh(inode, &gen);
166 io->ci_need_restart = cio->cui_layout_gen != gen;
167 if (io->ci_need_restart) {
169 DFID" layout changed from %d to %d.\n",
170 PFID(lu_object_fid(&obj->co_lu)),
171 cio->cui_layout_gen, gen);
172 /* today successful restore is the only possible
174 /* restore was done, clear restoring state */
175 ll_i2info(ccc_object_inode(obj))->lli_flags &=
176 ~LLIF_FILE_RESTORING;
181 static void vvp_io_fault_fini(const struct lu_env *env,
182 const struct cl_io_slice *ios)
184 struct cl_io *io = ios->cis_io;
185 struct cl_page *page = io->u.ci_fault.ft_page;
187 CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj));
190 lu_ref_del(&page->cp_reference, "fault", io);
191 cl_page_put(env, page);
192 io->u.ci_fault.ft_page = NULL;
194 vvp_io_fini(env, ios);
197 enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
200 * we only want to hold PW locks if the mmap() can generate
201 * writes back to the file and that only happens in shared
204 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
209 static int vvp_mmap_locks(const struct lu_env *env,
210 struct ccc_io *vio, struct cl_io *io)
212 struct ccc_thread_info *cti = ccc_env_info(env);
213 struct mm_struct *mm = current->mm;
214 struct vm_area_struct *vma;
215 struct cl_lock_descr *descr = &cti->cti_descr;
216 ldlm_policy_data_t policy;
223 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
225 if (!cl_is_normalio(env, io))
228 if (vio->cui_iov == NULL) /* nfs or loop back device write */
231 /* No MM (e.g. NFS)? No vmas too. */
235 for (seg = 0; seg < vio->cui_nrsegs; seg++) {
236 const struct iovec *iv = &vio->cui_iov[seg];
238 addr = (unsigned long)iv->iov_base;
243 count += addr & (~CFS_PAGE_MASK);
244 addr &= CFS_PAGE_MASK;
246 down_read(&mm->mmap_sem);
247 while((vma = our_vma(mm, addr, count)) != NULL) {
248 struct inode *inode = vma->vm_file->f_dentry->d_inode;
249 int flags = CEF_MUST;
251 if (ll_file_nolock(vma->vm_file)) {
253 * For no lock case, a lockless lock will be
260 * XXX: Required lock mode can be weakened: CIT_WRITE
261 * io only ever reads user level buffer, and CIT_READ
264 policy_from_vma(&policy, vma, addr, count);
265 descr->cld_mode = vvp_mode_from_vma(vma);
266 descr->cld_obj = ll_i2info(inode)->lli_clob;
267 descr->cld_start = cl_index(descr->cld_obj,
268 policy.l_extent.start);
269 descr->cld_end = cl_index(descr->cld_obj,
270 policy.l_extent.end);
271 descr->cld_enq_flags = flags;
272 result = cl_io_lock_alloc_add(env, io, descr);
274 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
275 descr->cld_mode, descr->cld_start,
281 if (vma->vm_end - addr >= count)
284 count -= vma->vm_end - addr;
287 up_read(&mm->mmap_sem);
292 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
293 enum cl_lock_mode mode, loff_t start, loff_t end)
295 struct ccc_io *cio = ccc_env_io(env);
299 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
302 ccc_io_update_iov(env, cio, io);
304 if (io->u.ci_rw.crw_nonblock)
305 ast_flags |= CEF_NONBLOCK;
306 result = vvp_mmap_locks(env, cio, io);
308 result = ccc_io_one_lock(env, io, ast_flags, mode, start, end);
312 static int vvp_io_read_lock(const struct lu_env *env,
313 const struct cl_io_slice *ios)
315 struct cl_io *io = ios->cis_io;
316 struct ll_inode_info *lli = ll_i2info(ccc_object_inode(io->ci_obj));
320 /* XXX: Layer violation, we shouldn't see lsm at llite level. */
321 if (lli->lli_has_smd) /* lsm-less file doesn't need to lock */
322 result = vvp_io_rw_lock(env, io, CLM_READ,
323 io->u.ci_rd.rd.crw_pos,
324 io->u.ci_rd.rd.crw_pos +
325 io->u.ci_rd.rd.crw_count - 1);
331 static int vvp_io_fault_lock(const struct lu_env *env,
332 const struct cl_io_slice *ios)
334 struct cl_io *io = ios->cis_io;
335 struct vvp_io *vio = cl2vvp_io(env, ios);
337 * XXX LDLM_FL_CBPENDING
339 return ccc_io_one_lock_index
340 (env, io, 0, vvp_mode_from_vma(vio->u.fault.ft_vma),
341 io->u.ci_fault.ft_index, io->u.ci_fault.ft_index);
344 static int vvp_io_write_lock(const struct lu_env *env,
345 const struct cl_io_slice *ios)
347 struct cl_io *io = ios->cis_io;
351 if (io->u.ci_wr.wr_append) {
353 end = OBD_OBJECT_EOF;
355 start = io->u.ci_wr.wr.crw_pos;
356 end = start + io->u.ci_wr.wr.crw_count - 1;
358 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
361 static int vvp_io_setattr_iter_init(const struct lu_env *env,
362 const struct cl_io_slice *ios)
368 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
370 * Handles "lockless io" mode when extent locking is done by server.
372 static int vvp_io_setattr_lock(const struct lu_env *env,
373 const struct cl_io_slice *ios)
375 struct ccc_io *cio = ccc_env_io(env);
376 struct cl_io *io = ios->cis_io;
380 if (cl_io_is_trunc(io)) {
381 new_size = io->u.ci_setattr.sa_attr.lvb_size;
383 enqflags = CEF_DISCARD_DATA;
385 if ((io->u.ci_setattr.sa_attr.lvb_mtime >=
386 io->u.ci_setattr.sa_attr.lvb_ctime) ||
387 (io->u.ci_setattr.sa_attr.lvb_atime >=
388 io->u.ci_setattr.sa_attr.lvb_ctime))
392 cio->u.setattr.cui_local_lock = SETATTR_EXTENT_LOCK;
393 return ccc_io_one_lock(env, io, enqflags, CLM_WRITE,
394 new_size, OBD_OBJECT_EOF);
397 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
403 * Only ll_inode_size_lock is taken at this level.
405 ll_inode_size_lock(inode);
406 result = inode_newsize_ok(inode, size);
408 ll_inode_size_unlock(inode);
411 oldsize = inode->i_size;
412 i_size_write(inode, size);
414 truncate_pagecache(inode, oldsize, size);
415 ll_inode_size_unlock(inode);
419 static int vvp_io_setattr_trunc(const struct lu_env *env,
420 const struct cl_io_slice *ios,
421 struct inode *inode, loff_t size)
423 inode_dio_wait(inode);
427 static int vvp_io_setattr_time(const struct lu_env *env,
428 const struct cl_io_slice *ios)
430 struct cl_io *io = ios->cis_io;
431 struct cl_object *obj = io->ci_obj;
432 struct cl_attr *attr = ccc_env_thread_attr(env);
434 unsigned valid = CAT_CTIME;
436 cl_object_attr_lock(obj);
437 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
438 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
439 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
442 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
443 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
446 result = cl_object_attr_set(env, obj, attr, valid);
447 cl_object_attr_unlock(obj);
452 static int vvp_io_setattr_start(const struct lu_env *env,
453 const struct cl_io_slice *ios)
455 struct cl_io *io = ios->cis_io;
456 struct inode *inode = ccc_object_inode(io->ci_obj);
459 mutex_lock(&inode->i_mutex);
460 if (cl_io_is_trunc(io))
461 result = vvp_io_setattr_trunc(env, ios, inode,
462 io->u.ci_setattr.sa_attr.lvb_size);
464 result = vvp_io_setattr_time(env, ios);
468 static void vvp_io_setattr_end(const struct lu_env *env,
469 const struct cl_io_slice *ios)
471 struct cl_io *io = ios->cis_io;
472 struct inode *inode = ccc_object_inode(io->ci_obj);
474 if (cl_io_is_trunc(io)) {
475 /* Truncate in memory pages - they must be clean pages
476 * because osc has already notified to destroy osc_extents. */
477 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
478 inode_dio_write_done(inode);
480 mutex_unlock(&inode->i_mutex);
483 static void vvp_io_setattr_fini(const struct lu_env *env,
484 const struct cl_io_slice *ios)
486 vvp_io_fini(env, ios);
489 static ssize_t lustre_generic_file_read(struct file *file,
490 struct ccc_io *vio, loff_t *ppos)
492 return generic_file_aio_read(vio->cui_iocb, vio->cui_iov,
493 vio->cui_nrsegs, *ppos);
496 static ssize_t lustre_generic_file_write(struct file *file,
497 struct ccc_io *vio, loff_t *ppos)
499 return generic_file_aio_write(vio->cui_iocb, vio->cui_iov,
500 vio->cui_nrsegs, *ppos);
503 static int vvp_io_read_start(const struct lu_env *env,
504 const struct cl_io_slice *ios)
506 struct vvp_io *vio = cl2vvp_io(env, ios);
507 struct ccc_io *cio = cl2ccc_io(env, ios);
508 struct cl_io *io = ios->cis_io;
509 struct cl_object *obj = io->ci_obj;
510 struct inode *inode = ccc_object_inode(obj);
511 struct ll_ra_read *bead = &vio->cui_bead;
512 struct file *file = cio->cui_fd->fd_file;
515 loff_t pos = io->u.ci_rd.rd.crw_pos;
516 long cnt = io->u.ci_rd.rd.crw_count;
517 long tot = cio->cui_tot_count;
520 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
522 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
524 if (!can_populate_pages(env, io, inode))
527 result = ccc_prep_size(env, obj, io, pos, tot, &exceed);
530 else if (exceed != 0)
533 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
534 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
535 inode->i_ino, cnt, pos, i_size_read(inode));
537 /* turn off the kernel's read-ahead */
538 cio->cui_fd->fd_file->f_ra.ra_pages = 0;
540 /* initialize read-ahead window once per syscall */
541 if (!vio->cui_ra_window_set) {
542 vio->cui_ra_window_set = 1;
543 bead->lrr_start = cl_index(obj, pos);
544 bead->lrr_count = cl_index(obj, tot + PAGE_CACHE_SIZE - 1);
545 ll_ra_read_in(file, bead);
550 switch (vio->cui_io_subtype) {
552 result = lustre_generic_file_read(file, cio, &pos);
555 result = generic_file_splice_read(file, &pos,
556 vio->u.splice.cui_pipe, cnt,
557 vio->u.splice.cui_flags);
558 /* LU-1109: do splice read stripe by stripe otherwise if it
559 * may make nfsd stuck if this read occupied all internal pipe
564 CERROR("Wrong IO type %u\n", vio->cui_io_subtype);
572 io->ci_nob += result;
573 ll_rw_stats_tally(ll_i2sbi(inode), current->pid, cio->cui_fd,
581 static void vvp_io_read_fini(const struct lu_env *env, const struct cl_io_slice *ios)
583 struct vvp_io *vio = cl2vvp_io(env, ios);
584 struct ccc_io *cio = cl2ccc_io(env, ios);
586 if (vio->cui_ra_window_set)
587 ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead);
589 vvp_io_fini(env, ios);
592 static int vvp_io_write_start(const struct lu_env *env,
593 const struct cl_io_slice *ios)
595 struct ccc_io *cio = cl2ccc_io(env, ios);
596 struct cl_io *io = ios->cis_io;
597 struct cl_object *obj = io->ci_obj;
598 struct inode *inode = ccc_object_inode(obj);
599 struct file *file = cio->cui_fd->fd_file;
601 loff_t pos = io->u.ci_wr.wr.crw_pos;
602 size_t cnt = io->u.ci_wr.wr.crw_count;
606 if (!can_populate_pages(env, io, inode))
609 if (cl_io_is_append(io)) {
611 * PARALLEL IO This has to be changed for parallel IO doing
612 * out-of-order writes.
614 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
615 cio->cui_iocb->ki_pos = pos;
618 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
620 if (cio->cui_iov == NULL) /* from a temp io in ll_cl_init(). */
623 result = lustre_generic_file_write(file, cio, &pos);
628 io->ci_nob += result;
629 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
630 cio->cui_fd, pos, result, WRITE);
637 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
639 struct vm_fault *vmf = cfio->fault.ft_vmf;
641 cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, vmf);
644 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
645 vmf->virtual_address);
646 if (unlikely(!(cfio->fault.ft_flags & VM_FAULT_LOCKED))) {
647 lock_page(vmf->page);
648 cfio->fault.ft_flags &= VM_FAULT_LOCKED;
651 cfio->ft_vmpage = vmf->page;
655 if (cfio->fault.ft_flags & VM_FAULT_SIGBUS) {
656 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
660 if (cfio->fault.ft_flags & VM_FAULT_OOM) {
661 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
665 if (cfio->fault.ft_flags & VM_FAULT_RETRY)
668 CERROR("unknow error in page fault %d!\n", cfio->fault.ft_flags);
672 static int vvp_io_fault_start(const struct lu_env *env,
673 const struct cl_io_slice *ios)
675 struct vvp_io *vio = cl2vvp_io(env, ios);
676 struct cl_io *io = ios->cis_io;
677 struct cl_object *obj = io->ci_obj;
678 struct inode *inode = ccc_object_inode(obj);
679 struct cl_fault_io *fio = &io->u.ci_fault;
680 struct vvp_fault_io *cfio = &vio->u.fault;
683 struct page *vmpage = NULL;
684 struct cl_page *page;
686 pgoff_t last; /* last page in a file data region */
688 if (fio->ft_executable &&
689 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
691 " changed while waiting for the page fault lock\n",
692 PFID(lu_object_fid(&obj->co_lu)));
694 /* offset of the last byte on the page */
695 offset = cl_offset(obj, fio->ft_index + 1) - 1;
696 LASSERT(cl_index(obj, offset) == fio->ft_index);
697 result = ccc_prep_size(env, obj, io, 0, offset + 1, NULL);
701 /* must return locked page */
702 if (fio->ft_mkwrite) {
703 LASSERT(cfio->ft_vmpage != NULL);
704 lock_page(cfio->ft_vmpage);
706 result = vvp_io_kernel_fault(cfio);
711 vmpage = cfio->ft_vmpage;
712 LASSERT(PageLocked(vmpage));
714 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
715 ll_invalidate_page(vmpage);
717 size = i_size_read(inode);
718 /* Though we have already held a cl_lock upon this page, but
719 * it still can be truncated locally. */
720 if (unlikely((vmpage->mapping != inode->i_mapping) ||
721 (page_offset(vmpage) > size))) {
722 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
724 /* return +1 to stop cl_io_loop() and ll_fault() will catch
726 GOTO(out, result = +1);
730 if (fio->ft_mkwrite ) {
733 * Capture the size while holding the lli_trunc_sem from above
734 * we want to make sure that we complete the mkwrite action
735 * while holding this lock. We need to make sure that we are
736 * not past the end of the file.
738 last_index = cl_index(obj, size - 1);
739 if (last_index < fio->ft_index) {
741 "llite: mkwrite and truncate race happened: "
743 vmpage->mapping,fio->ft_index,last_index);
745 * We need to return if we are
746 * passed the end of the file. This will propagate
747 * up the call stack to ll_page_mkwrite where
748 * we will return VM_FAULT_NOPAGE. Any non-negative
749 * value returned here will be silently
750 * converted to 0. If the vmpage->mapping is null
751 * the error code would be converted back to ENODATA
752 * in ll_page_mkwrite0. Thus we return -ENODATA
753 * to handle both cases
755 GOTO(out, result = -ENODATA);
759 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
761 GOTO(out, result = PTR_ERR(page));
763 /* if page is going to be written, we should add this page into cache
765 if (fio->ft_mkwrite) {
766 wait_on_page_writeback(vmpage);
767 if (set_page_dirty(vmpage)) {
770 /* vvp_page_assume() calls wait_on_page_writeback(). */
771 cl_page_assume(env, io, page);
773 cp = cl2ccc_page(cl_page_at(page, &vvp_device_type));
774 vvp_write_pending(cl2ccc(obj), cp);
776 /* Do not set Dirty bit here so that in case IO is
777 * started before the page is really made dirty, we
778 * still have chance to detect it. */
779 result = cl_page_cache_add(env, io, page, CRT_WRITE);
780 LASSERT(cl_page_is_owned(page, io));
784 cl_page_unmap(env, io, page);
785 cl_page_discard(env, io, page);
786 cl_page_disown(env, io, page);
788 cl_page_put(env, page);
790 /* we're in big trouble, what can we do now? */
791 if (result == -EDQUOT)
795 cl_page_disown(env, io, page);
799 last = cl_index(obj, size - 1);
801 * The ft_index is only used in the case of
802 * a mkwrite action. We need to check
803 * our assertions are correct, since
804 * we should have caught this above
806 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last);
807 if (fio->ft_index == last)
809 * Last page is mapped partially.
811 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
813 fio->ft_nob = cl_page_size(obj);
815 lu_ref_add(&page->cp_reference, "fault", io);
820 /* return unlocked vmpage to avoid deadlocking */
823 cfio->fault.ft_flags &= ~VM_FAULT_LOCKED;
827 static int vvp_io_fsync_start(const struct lu_env *env,
828 const struct cl_io_slice *ios)
830 /* we should mark TOWRITE bit to each dirty page in radix tree to
831 * verify pages have been written, but this is difficult because of
836 static int vvp_io_read_page(const struct lu_env *env,
837 const struct cl_io_slice *ios,
838 const struct cl_page_slice *slice)
840 struct cl_io *io = ios->cis_io;
841 struct cl_object *obj = slice->cpl_obj;
842 struct ccc_page *cp = cl2ccc_page(slice);
843 struct cl_page *page = slice->cpl_page;
844 struct inode *inode = ccc_object_inode(obj);
845 struct ll_sb_info *sbi = ll_i2sbi(inode);
846 struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd;
847 struct ll_readahead_state *ras = &fd->fd_ras;
848 struct page *vmpage = cp->cpg_page;
849 struct cl_2queue *queue = &io->ci_queue;
852 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
853 LASSERT(slice->cpl_obj == obj);
857 if (sbi->ll_ra_info.ra_max_pages_per_file &&
858 sbi->ll_ra_info.ra_max_pages)
859 ras_update(sbi, inode, ras, page->cp_index,
860 cp->cpg_defer_uptodate);
862 /* Sanity check whether the page is protected by a lock. */
863 rc = cl_page_is_under_lock(env, io, page);
865 CL_PAGE_HEADER(D_WARNING, env, page, "%s: %d\n",
866 rc == -ENODATA ? "without a lock" :
872 if (cp->cpg_defer_uptodate) {
874 cl_page_export(env, page, 1);
877 * Add page into the queue even when it is marked uptodate above.
878 * this will unlock it automatically as part of cl_page_list_disown().
880 cl_2queue_add(queue, page);
881 if (sbi->ll_ra_info.ra_max_pages_per_file &&
882 sbi->ll_ra_info.ra_max_pages)
883 ll_readahead(env, io, ras,
884 vmpage->mapping, &queue->c2_qin, fd->fd_flags);
889 static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io,
890 struct cl_page *page, struct ccc_page *cp,
891 enum cl_req_type crt)
893 struct cl_2queue *queue;
896 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
898 queue = &io->ci_queue;
899 cl_2queue_init_page(queue, page);
901 result = cl_io_submit_sync(env, io, crt, queue, 0);
902 LASSERT(cl_page_is_owned(page, io));
906 * in CRT_WRITE case page is left locked even in case of
909 cl_page_list_disown(env, io, &queue->c2_qin);
910 cl_2queue_fini(env, queue);
916 * Prepare partially written-to page for a write.
918 static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io,
919 struct cl_object *obj, struct cl_page *pg,
921 unsigned from, unsigned to)
923 struct cl_attr *attr = ccc_env_thread_attr(env);
924 loff_t offset = cl_offset(obj, pg->cp_index);
927 cl_object_attr_lock(obj);
928 result = cl_object_attr_get(env, obj, attr);
929 cl_object_attr_unlock(obj);
932 * If are writing to a new page, no need to read old data.
933 * The extent locking will have updated the KMS, and for our
934 * purposes here we can treat it like i_size.
936 if (attr->cat_kms <= offset) {
937 char *kaddr = ll_kmap_atomic(cp->cpg_page, KM_USER0);
939 memset(kaddr, 0, cl_page_size(obj));
940 ll_kunmap_atomic(kaddr, KM_USER0);
941 } else if (cp->cpg_defer_uptodate)
944 result = vvp_page_sync_io(env, io, pg, cp, CRT_READ);
946 * In older implementations, obdo_refresh_inode is called here
947 * to update the inode because the write might modify the
948 * object info at OST. However, this has been proven useless,
949 * since LVB functions will be called when user space program
950 * tries to retrieve inode attribute. Also, see bug 15909 for
954 cl_page_export(env, pg, 1);
959 static int vvp_io_prepare_write(const struct lu_env *env,
960 const struct cl_io_slice *ios,
961 const struct cl_page_slice *slice,
962 unsigned from, unsigned to)
964 struct cl_object *obj = slice->cpl_obj;
965 struct ccc_page *cp = cl2ccc_page(slice);
966 struct cl_page *pg = slice->cpl_page;
967 struct page *vmpage = cp->cpg_page;
973 LINVRNT(cl_page_is_vmlocked(env, pg));
974 LASSERT(vmpage->mapping->host == ccc_object_inode(obj));
978 CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to);
979 if (!PageUptodate(vmpage)) {
981 * We're completely overwriting an existing page, so _don't_
982 * set it up to date until commit_write
984 if (from == 0 && to == PAGE_CACHE_SIZE) {
985 CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n");
986 POISON_PAGE(page, 0x11);
988 result = vvp_io_prepare_partial(env, ios->cis_io, obj,
991 CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n");
995 static int vvp_io_commit_write(const struct lu_env *env,
996 const struct cl_io_slice *ios,
997 const struct cl_page_slice *slice,
998 unsigned from, unsigned to)
1000 struct cl_object *obj = slice->cpl_obj;
1001 struct cl_io *io = ios->cis_io;
1002 struct ccc_page *cp = cl2ccc_page(slice);
1003 struct cl_page *pg = slice->cpl_page;
1004 struct inode *inode = ccc_object_inode(obj);
1005 struct ll_sb_info *sbi = ll_i2sbi(inode);
1006 struct ll_inode_info *lli = ll_i2info(inode);
1007 struct page *vmpage = cp->cpg_page;
1015 LINVRNT(cl_page_is_vmlocked(env, pg));
1016 LASSERT(vmpage->mapping->host == inode);
1018 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "commiting page write\n");
1019 CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to);
1022 * queue a write for some time in the future the first time we
1025 * This is different from what other file systems do: they usually
1026 * just mark page (and some of its buffers) dirty and rely on
1027 * balance_dirty_pages() to start a write-back. Lustre wants write-back
1028 * to be started earlier for the following reasons:
1030 * (1) with a large number of clients we need to limit the amount
1031 * of cached data on the clients a lot;
1033 * (2) large compute jobs generally want compute-only then io-only
1034 * and the IO should complete as quickly as possible;
1036 * (3) IO is batched up to the RPC size and is async until the
1037 * client max cache is hit
1038 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
1041 if (!PageDirty(vmpage)) {
1042 tallyop = LPROC_LL_DIRTY_MISSES;
1043 result = cl_page_cache_add(env, io, pg, CRT_WRITE);
1045 /* page was added into cache successfully. */
1046 set_page_dirty(vmpage);
1047 vvp_write_pending(cl2ccc(obj), cp);
1048 } else if (result == -EDQUOT) {
1049 pgoff_t last_index = i_size_read(inode) >> PAGE_CACHE_SHIFT;
1050 bool need_clip = true;
1053 * Client ran out of disk space grant. Possible
1056 * (a) do a sync write, renewing grant;
1058 * (b) stop writing on this stripe, switch to the
1061 * (b) is a part of "parallel io" design that is the
1062 * ultimate goal. (a) is what "old" client did, and
1063 * what the new code continues to do for the time
1066 if (last_index > pg->cp_index) {
1067 to = PAGE_CACHE_SIZE;
1069 } else if (last_index == pg->cp_index) {
1070 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
1075 cl_page_clip(env, pg, 0, to);
1076 result = vvp_page_sync_io(env, io, pg, cp, CRT_WRITE);
1078 CERROR("Write page %lu of inode %p failed %d\n",
1079 pg->cp_index, inode, result);
1082 tallyop = LPROC_LL_DIRTY_HITS;
1085 ll_stats_ops_tally(sbi, tallyop, 1);
1087 /* Inode should be marked DIRTY even if no new page was marked DIRTY
1088 * because page could have been not flushed between 2 modifications.
1089 * It is important the file is marked DIRTY as soon as the I/O is done
1090 * Indeed, when cache is flushed, file could be already closed and it
1091 * is too late to warn the MDT.
1092 * It is acceptable that file is marked DIRTY even if I/O is dropped
1093 * for some reasons before being flushed to OST.
1096 spin_lock(&lli->lli_lock);
1097 lli->lli_flags |= LLIF_DATA_MODIFIED;
1098 spin_unlock(&lli->lli_lock);
1101 size = cl_offset(obj, pg->cp_index) + to;
1103 ll_inode_size_lock(inode);
1105 if (size > i_size_read(inode)) {
1106 cl_isize_write_nolock(inode, size);
1107 CDEBUG(D_VFSTRACE, DFID" updating i_size %lu\n",
1108 PFID(lu_object_fid(&obj->co_lu)),
1109 (unsigned long)size);
1111 cl_page_export(env, pg, 1);
1113 if (size > i_size_read(inode))
1114 cl_page_discard(env, io, pg);
1116 ll_inode_size_unlock(inode);
1120 static const struct cl_io_operations vvp_io_ops = {
1123 .cio_fini = vvp_io_read_fini,
1124 .cio_lock = vvp_io_read_lock,
1125 .cio_start = vvp_io_read_start,
1126 .cio_advance = ccc_io_advance
1129 .cio_fini = vvp_io_fini,
1130 .cio_lock = vvp_io_write_lock,
1131 .cio_start = vvp_io_write_start,
1132 .cio_advance = ccc_io_advance
1135 .cio_fini = vvp_io_setattr_fini,
1136 .cio_iter_init = vvp_io_setattr_iter_init,
1137 .cio_lock = vvp_io_setattr_lock,
1138 .cio_start = vvp_io_setattr_start,
1139 .cio_end = vvp_io_setattr_end
1142 .cio_fini = vvp_io_fault_fini,
1143 .cio_iter_init = vvp_io_fault_iter_init,
1144 .cio_lock = vvp_io_fault_lock,
1145 .cio_start = vvp_io_fault_start,
1146 .cio_end = ccc_io_end
1149 .cio_start = vvp_io_fsync_start,
1150 .cio_fini = vvp_io_fini
1153 .cio_fini = vvp_io_fini
1156 .cio_read_page = vvp_io_read_page,
1157 .cio_prepare_write = vvp_io_prepare_write,
1158 .cio_commit_write = vvp_io_commit_write
1161 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1164 struct vvp_io *vio = vvp_env_io(env);
1165 struct ccc_io *cio = ccc_env_io(env);
1166 struct inode *inode = ccc_object_inode(obj);
1169 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
1172 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1173 "restore needed %d\n",
1174 PFID(lu_object_fid(&obj->co_lu)),
1175 io->ci_ignore_layout, io->ci_verify_layout,
1176 cio->cui_layout_gen, io->ci_restore_needed);
1178 CL_IO_SLICE_CLEAN(cio, cui_cl);
1179 cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops);
1180 vio->cui_ra_window_set = 0;
1182 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1184 struct ll_inode_info *lli = ll_i2info(inode);
1186 count = io->u.ci_rw.crw_count;
1187 /* "If nbyte is 0, read() will return 0 and have no other
1188 * results." -- Single Unix Spec */
1192 cio->cui_tot_count = count;
1193 cio->cui_tot_nrsegs = 0;
1195 /* for read/write, we store the jobid in the inode, and
1196 * it'll be fetched by osc when building RPC.
1198 * it's not accurate if the file is shared by different
1201 lustre_get_jobid(lli->lli_jobid);
1202 } else if (io->ci_type == CIT_SETATTR) {
1203 if (!cl_io_is_trunc(io))
1204 io->ci_lockreq = CILR_MANDATORY;
1207 /* ignore layout change for generic CIT_MISC but not for glimpse.
1208 * io context for glimpse must set ci_verify_layout to true,
1209 * see cl_glimpse_size0() for details. */
1210 if (io->ci_type == CIT_MISC && !io->ci_verify_layout)
1211 io->ci_ignore_layout = 1;
1213 /* Enqueue layout lock and get layout version. We need to do this
1214 * even for operations requiring to open file, such as read and write,
1215 * because it might not grant layout lock in IT_OPEN. */
1216 if (result == 0 && !io->ci_ignore_layout) {
1217 result = ll_layout_refresh(inode, &cio->cui_layout_gen);
1218 if (result == -ENOENT)
1219 /* If the inode on MDS has been removed, but the objects
1220 * on OSTs haven't been destroyed (async unlink), layout
1221 * fetch will return -ENOENT, we'd ingore this error
1222 * and continue with dirty flush. LU-3230. */
1225 CERROR("%s: refresh file layout " DFID " error %d.\n",
1226 ll_get_fsname(inode->i_sb, NULL, 0),
1227 PFID(lu_object_fid(&obj->co_lu)), result);
1233 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
1234 const struct cl_io_slice *slice)
1236 /* Caling just for assertion */
1237 cl2ccc_io(env, slice);
1238 return vvp_env_io(env);