4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2012, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * Implementation of cl_io for VVP layer.
38 * Author: Nikita Danilov <nikita.danilov@sun.com>
39 * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
42 #define DEBUG_SUBSYSTEM S_LLITE
46 #include <lustre_lite.h>
48 #include "vvp_internal.h"
50 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
51 const struct cl_io_slice *slice);
54 * True, if \a io is a normal io, False for sendfile() / splice_{read|write}
56 int cl_is_normalio(const struct lu_env *env, const struct cl_io *io)
58 struct vvp_io *vio = vvp_env_io(env);
60 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
62 return vio->cui_io_subtype == IO_NORMAL;
66 * For swapping layout. The file's layout may have changed.
67 * To avoid populating pages to a wrong stripe, we have to verify the
68 * correctness of layout. It works because swapping layout processes
69 * have to acquire group lock.
71 static bool can_populate_pages(const struct lu_env *env, struct cl_io *io,
74 struct ll_inode_info *lli = ll_i2info(inode);
75 struct ccc_io *cio = ccc_env_io(env);
78 switch (io->ci_type) {
81 /* don't need lock here to check lli_layout_gen as we have held
82 * extent lock and GROUP lock has to hold to swap layout */
83 if (lli->lli_layout_gen != cio->cui_layout_gen) {
84 io->ci_need_restart = 1;
85 /* this will return application a short read/write */
90 /* fault is okay because we've already had a page. */
98 /*****************************************************************************
104 static int vvp_io_fault_iter_init(const struct lu_env *env,
105 const struct cl_io_slice *ios)
107 struct vvp_io *vio = cl2vvp_io(env, ios);
108 struct inode *inode = ccc_object_inode(ios->cis_obj);
111 cl2ccc_io(env, ios)->cui_fd->fd_file->f_dentry->d_inode);
112 vio->u.fault.ft_mtime = LTIME_S(inode->i_mtime);
116 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
118 struct cl_io *io = ios->cis_io;
119 struct cl_object *obj = io->ci_obj;
120 struct ccc_io *cio = cl2ccc_io(env, ios);
122 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
124 CDEBUG(D_VFSTRACE, DFID
125 " ignore/verify layout %d/%d, layout version %d restore needed %d\n",
126 PFID(lu_object_fid(&obj->co_lu)),
127 io->ci_ignore_layout, io->ci_verify_layout,
128 cio->cui_layout_gen, io->ci_restore_needed);
130 if (io->ci_restore_needed == 1) {
133 /* file was detected release, we need to restore it
134 * before finishing the io
136 rc = ll_layout_restore(ccc_object_inode(obj));
137 /* if restore registration failed, no restart,
138 * we will return -ENODATA */
139 /* The layout will change after restore, so we need to
140 * block on layout lock hold by the MDT
141 * as MDT will not send new layout in lvb (see LU-3124)
142 * we have to explicitly fetch it, all this will be done
143 * by ll_layout_refresh()
146 io->ci_restore_needed = 0;
147 io->ci_need_restart = 1;
148 io->ci_verify_layout = 1;
150 io->ci_restore_needed = 1;
151 io->ci_need_restart = 0;
152 io->ci_verify_layout = 0;
157 if (!io->ci_ignore_layout && io->ci_verify_layout) {
160 /* check layout version */
161 ll_layout_refresh(ccc_object_inode(obj), &gen);
162 io->ci_need_restart = cio->cui_layout_gen != gen;
163 if (io->ci_need_restart) {
165 DFID" layout changed from %d to %d.\n",
166 PFID(lu_object_fid(&obj->co_lu)),
167 cio->cui_layout_gen, gen);
168 /* today successful restore is the only possible
170 /* restore was done, clear restoring state */
171 ll_i2info(ccc_object_inode(obj))->lli_flags &=
172 ~LLIF_FILE_RESTORING;
177 static void vvp_io_fault_fini(const struct lu_env *env,
178 const struct cl_io_slice *ios)
180 struct cl_io *io = ios->cis_io;
181 struct cl_page *page = io->u.ci_fault.ft_page;
183 CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj));
186 lu_ref_del(&page->cp_reference, "fault", io);
187 cl_page_put(env, page);
188 io->u.ci_fault.ft_page = NULL;
190 vvp_io_fini(env, ios);
193 enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
196 * we only want to hold PW locks if the mmap() can generate
197 * writes back to the file and that only happens in shared
200 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
205 static int vvp_mmap_locks(const struct lu_env *env,
206 struct ccc_io *vio, struct cl_io *io)
208 struct ccc_thread_info *cti = ccc_env_info(env);
209 struct mm_struct *mm = current->mm;
210 struct vm_area_struct *vma;
211 struct cl_lock_descr *descr = &cti->cti_descr;
212 ldlm_policy_data_t policy;
218 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
220 if (!cl_is_normalio(env, io))
223 if (vio->cui_iov == NULL) /* nfs or loop back device write */
226 /* No MM (e.g. NFS)? No vmas too. */
230 for (seg = 0; seg < vio->cui_nrsegs; seg++) {
231 const struct iovec *iv = &vio->cui_iov[seg];
233 addr = (unsigned long)iv->iov_base;
238 count += addr & (~CFS_PAGE_MASK);
239 addr &= CFS_PAGE_MASK;
241 down_read(&mm->mmap_sem);
242 while((vma = our_vma(mm, addr, count)) != NULL) {
243 struct inode *inode = vma->vm_file->f_dentry->d_inode;
244 int flags = CEF_MUST;
246 if (ll_file_nolock(vma->vm_file)) {
248 * For no lock case, a lockless lock will be
255 * XXX: Required lock mode can be weakened: CIT_WRITE
256 * io only ever reads user level buffer, and CIT_READ
259 policy_from_vma(&policy, vma, addr, count);
260 descr->cld_mode = vvp_mode_from_vma(vma);
261 descr->cld_obj = ll_i2info(inode)->lli_clob;
262 descr->cld_start = cl_index(descr->cld_obj,
263 policy.l_extent.start);
264 descr->cld_end = cl_index(descr->cld_obj,
265 policy.l_extent.end);
266 descr->cld_enq_flags = flags;
267 result = cl_io_lock_alloc_add(env, io, descr);
269 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
270 descr->cld_mode, descr->cld_start,
276 if (vma->vm_end - addr >= count)
279 count -= vma->vm_end - addr;
282 up_read(&mm->mmap_sem);
287 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
288 enum cl_lock_mode mode, loff_t start, loff_t end)
290 struct ccc_io *cio = ccc_env_io(env);
294 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
296 ccc_io_update_iov(env, cio, io);
298 if (io->u.ci_rw.crw_nonblock)
299 ast_flags |= CEF_NONBLOCK;
300 result = vvp_mmap_locks(env, cio, io);
302 result = ccc_io_one_lock(env, io, ast_flags, mode, start, end);
306 static int vvp_io_read_lock(const struct lu_env *env,
307 const struct cl_io_slice *ios)
309 struct cl_io *io = ios->cis_io;
310 struct ll_inode_info *lli = ll_i2info(ccc_object_inode(io->ci_obj));
313 /* XXX: Layer violation, we shouldn't see lsm at llite level. */
314 if (lli->lli_has_smd) /* lsm-less file doesn't need to lock */
315 result = vvp_io_rw_lock(env, io, CLM_READ,
316 io->u.ci_rd.rd.crw_pos,
317 io->u.ci_rd.rd.crw_pos +
318 io->u.ci_rd.rd.crw_count - 1);
324 static int vvp_io_fault_lock(const struct lu_env *env,
325 const struct cl_io_slice *ios)
327 struct cl_io *io = ios->cis_io;
328 struct vvp_io *vio = cl2vvp_io(env, ios);
330 * XXX LDLM_FL_CBPENDING
332 return ccc_io_one_lock_index
333 (env, io, 0, vvp_mode_from_vma(vio->u.fault.ft_vma),
334 io->u.ci_fault.ft_index, io->u.ci_fault.ft_index);
337 static int vvp_io_write_lock(const struct lu_env *env,
338 const struct cl_io_slice *ios)
340 struct cl_io *io = ios->cis_io;
344 if (io->u.ci_wr.wr_append) {
346 end = OBD_OBJECT_EOF;
348 start = io->u.ci_wr.wr.crw_pos;
349 end = start + io->u.ci_wr.wr.crw_count - 1;
351 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
354 static int vvp_io_setattr_iter_init(const struct lu_env *env,
355 const struct cl_io_slice *ios)
361 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
363 * Handles "lockless io" mode when extent locking is done by server.
365 static int vvp_io_setattr_lock(const struct lu_env *env,
366 const struct cl_io_slice *ios)
368 struct ccc_io *cio = ccc_env_io(env);
369 struct cl_io *io = ios->cis_io;
373 if (cl_io_is_trunc(io)) {
374 new_size = io->u.ci_setattr.sa_attr.lvb_size;
376 enqflags = CEF_DISCARD_DATA;
378 if ((io->u.ci_setattr.sa_attr.lvb_mtime >=
379 io->u.ci_setattr.sa_attr.lvb_ctime) ||
380 (io->u.ci_setattr.sa_attr.lvb_atime >=
381 io->u.ci_setattr.sa_attr.lvb_ctime))
385 cio->u.setattr.cui_local_lock = SETATTR_EXTENT_LOCK;
386 return ccc_io_one_lock(env, io, enqflags, CLM_WRITE,
387 new_size, OBD_OBJECT_EOF);
390 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
394 * Only ll_inode_size_lock is taken at this level.
396 ll_inode_size_lock(inode);
397 result = inode_newsize_ok(inode, size);
399 ll_inode_size_unlock(inode);
402 truncate_setsize(inode, size);
403 ll_inode_size_unlock(inode);
407 static int vvp_io_setattr_trunc(const struct lu_env *env,
408 const struct cl_io_slice *ios,
409 struct inode *inode, loff_t size)
411 inode_dio_wait(inode);
415 static int vvp_io_setattr_time(const struct lu_env *env,
416 const struct cl_io_slice *ios)
418 struct cl_io *io = ios->cis_io;
419 struct cl_object *obj = io->ci_obj;
420 struct cl_attr *attr = ccc_env_thread_attr(env);
422 unsigned valid = CAT_CTIME;
424 cl_object_attr_lock(obj);
425 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
426 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
427 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
430 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
431 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
434 result = cl_object_attr_set(env, obj, attr, valid);
435 cl_object_attr_unlock(obj);
440 static int vvp_io_setattr_start(const struct lu_env *env,
441 const struct cl_io_slice *ios)
443 struct cl_io *io = ios->cis_io;
444 struct inode *inode = ccc_object_inode(io->ci_obj);
447 mutex_lock(&inode->i_mutex);
448 if (cl_io_is_trunc(io))
449 result = vvp_io_setattr_trunc(env, ios, inode,
450 io->u.ci_setattr.sa_attr.lvb_size);
452 result = vvp_io_setattr_time(env, ios);
456 static void vvp_io_setattr_end(const struct lu_env *env,
457 const struct cl_io_slice *ios)
459 struct cl_io *io = ios->cis_io;
460 struct inode *inode = ccc_object_inode(io->ci_obj);
462 if (cl_io_is_trunc(io)) {
463 /* Truncate in memory pages - they must be clean pages
464 * because osc has already notified to destroy osc_extents. */
465 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
466 inode_dio_write_done(inode);
468 mutex_unlock(&inode->i_mutex);
471 static void vvp_io_setattr_fini(const struct lu_env *env,
472 const struct cl_io_slice *ios)
474 vvp_io_fini(env, ios);
477 static ssize_t lustre_generic_file_read(struct file *file,
478 struct ccc_io *vio, loff_t *ppos)
480 return generic_file_aio_read(vio->cui_iocb, vio->cui_iov,
481 vio->cui_nrsegs, *ppos);
484 static ssize_t lustre_generic_file_write(struct file *file,
485 struct ccc_io *vio, loff_t *ppos)
487 return generic_file_aio_write(vio->cui_iocb, vio->cui_iov,
488 vio->cui_nrsegs, *ppos);
491 static int vvp_io_read_start(const struct lu_env *env,
492 const struct cl_io_slice *ios)
494 struct vvp_io *vio = cl2vvp_io(env, ios);
495 struct ccc_io *cio = cl2ccc_io(env, ios);
496 struct cl_io *io = ios->cis_io;
497 struct cl_object *obj = io->ci_obj;
498 struct inode *inode = ccc_object_inode(obj);
499 struct ll_ra_read *bead = &vio->cui_bead;
500 struct file *file = cio->cui_fd->fd_file;
503 loff_t pos = io->u.ci_rd.rd.crw_pos;
504 long cnt = io->u.ci_rd.rd.crw_count;
505 long tot = cio->cui_tot_count;
508 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
510 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
512 if (!can_populate_pages(env, io, inode))
515 result = ccc_prep_size(env, obj, io, pos, tot, &exceed);
518 else if (exceed != 0)
521 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
522 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
523 inode->i_ino, cnt, pos, i_size_read(inode));
525 /* turn off the kernel's read-ahead */
526 cio->cui_fd->fd_file->f_ra.ra_pages = 0;
528 /* initialize read-ahead window once per syscall */
529 if (!vio->cui_ra_window_set) {
530 vio->cui_ra_window_set = 1;
531 bead->lrr_start = cl_index(obj, pos);
533 * XXX: explicit PAGE_CACHE_SIZE
535 bead->lrr_count = cl_index(obj, tot + PAGE_CACHE_SIZE - 1);
536 ll_ra_read_in(file, bead);
541 switch (vio->cui_io_subtype) {
543 result = lustre_generic_file_read(file, cio, &pos);
546 result = generic_file_splice_read(file, &pos,
547 vio->u.splice.cui_pipe, cnt,
548 vio->u.splice.cui_flags);
549 /* LU-1109: do splice read stripe by stripe otherwise if it
550 * may make nfsd stuck if this read occupied all internal pipe
555 CERROR("Wrong IO type %u\n", vio->cui_io_subtype);
563 io->ci_nob += result;
564 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
565 cio->cui_fd, pos, result, READ);
571 static void vvp_io_read_fini(const struct lu_env *env, const struct cl_io_slice *ios)
573 struct vvp_io *vio = cl2vvp_io(env, ios);
574 struct ccc_io *cio = cl2ccc_io(env, ios);
576 if (vio->cui_ra_window_set)
577 ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead);
579 vvp_io_fini(env, ios);
582 static int vvp_io_write_start(const struct lu_env *env,
583 const struct cl_io_slice *ios)
585 struct ccc_io *cio = cl2ccc_io(env, ios);
586 struct cl_io *io = ios->cis_io;
587 struct cl_object *obj = io->ci_obj;
588 struct inode *inode = ccc_object_inode(obj);
589 struct file *file = cio->cui_fd->fd_file;
591 loff_t pos = io->u.ci_wr.wr.crw_pos;
592 size_t cnt = io->u.ci_wr.wr.crw_count;
594 if (!can_populate_pages(env, io, inode))
597 if (cl_io_is_append(io)) {
599 * PARALLEL IO This has to be changed for parallel IO doing
600 * out-of-order writes.
602 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
603 cio->cui_iocb->ki_pos = pos;
606 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
608 if (cio->cui_iov == NULL) /* from a temp io in ll_cl_init(). */
611 result = lustre_generic_file_write(file, cio, &pos);
616 io->ci_nob += result;
617 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
618 cio->cui_fd, pos, result, WRITE);
624 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
626 struct vm_fault *vmf = cfio->fault.ft_vmf;
628 cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, vmf);
632 "page %p map %p index %lu flags %lx count %u priv %0lx: got addr %p type NOPAGE\n",
633 vmf->page, vmf->page->mapping, vmf->page->index,
634 (long)vmf->page->flags, page_count(vmf->page),
635 page_private(vmf->page), vmf->virtual_address);
636 if (unlikely(!(cfio->fault.ft_flags & VM_FAULT_LOCKED))) {
637 lock_page(vmf->page);
638 cfio->fault.ft_flags &= VM_FAULT_LOCKED;
641 cfio->ft_vmpage = vmf->page;
645 if (cfio->fault.ft_flags & VM_FAULT_SIGBUS) {
646 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
650 if (cfio->fault.ft_flags & VM_FAULT_OOM) {
651 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
655 if (cfio->fault.ft_flags & VM_FAULT_RETRY)
658 CERROR("unknow error in page fault %d!\n", cfio->fault.ft_flags);
663 static int vvp_io_fault_start(const struct lu_env *env,
664 const struct cl_io_slice *ios)
666 struct vvp_io *vio = cl2vvp_io(env, ios);
667 struct cl_io *io = ios->cis_io;
668 struct cl_object *obj = io->ci_obj;
669 struct inode *inode = ccc_object_inode(obj);
670 struct cl_fault_io *fio = &io->u.ci_fault;
671 struct vvp_fault_io *cfio = &vio->u.fault;
674 struct page *vmpage = NULL;
675 struct cl_page *page;
677 pgoff_t last; /* last page in a file data region */
679 if (fio->ft_executable &&
680 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
682 " changed while waiting for the page fault lock\n",
683 PFID(lu_object_fid(&obj->co_lu)));
685 /* offset of the last byte on the page */
686 offset = cl_offset(obj, fio->ft_index + 1) - 1;
687 LASSERT(cl_index(obj, offset) == fio->ft_index);
688 result = ccc_prep_size(env, obj, io, 0, offset + 1, NULL);
692 /* must return locked page */
693 if (fio->ft_mkwrite) {
694 LASSERT(cfio->ft_vmpage != NULL);
695 lock_page(cfio->ft_vmpage);
697 result = vvp_io_kernel_fault(cfio);
702 vmpage = cfio->ft_vmpage;
703 LASSERT(PageLocked(vmpage));
705 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
706 ll_invalidate_page(vmpage);
708 size = i_size_read(inode);
709 /* Though we have already held a cl_lock upon this page, but
710 * it still can be truncated locally. */
711 if (unlikely((vmpage->mapping != inode->i_mapping) ||
712 (page_offset(vmpage) > size))) {
713 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
715 /* return +1 to stop cl_io_loop() and ll_fault() will catch
717 GOTO(out, result = +1);
721 if (fio->ft_mkwrite ) {
724 * Capture the size while holding the lli_trunc_sem from above
725 * we want to make sure that we complete the mkwrite action
726 * while holding this lock. We need to make sure that we are
727 * not past the end of the file.
729 last_index = cl_index(obj, size - 1);
730 if (last_index < fio->ft_index) {
732 "llite: mkwrite and truncate race happened: "
734 vmpage->mapping,fio->ft_index,last_index);
736 * We need to return if we are
737 * passed the end of the file. This will propagate
738 * up the call stack to ll_page_mkwrite where
739 * we will return VM_FAULT_NOPAGE. Any non-negative
740 * value returned here will be silently
741 * converted to 0. If the vmpage->mapping is null
742 * the error code would be converted back to ENODATA
743 * in ll_page_mkwrite0. Thus we return -ENODATA
744 * to handle both cases
746 GOTO(out, result = -ENODATA);
750 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
752 GOTO(out, result = PTR_ERR(page));
754 /* if page is going to be written, we should add this page into cache
756 if (fio->ft_mkwrite) {
757 wait_on_page_writeback(vmpage);
758 if (set_page_dirty(vmpage)) {
761 /* vvp_page_assume() calls wait_on_page_writeback(). */
762 cl_page_assume(env, io, page);
764 cp = cl2ccc_page(cl_page_at(page, &vvp_device_type));
765 vvp_write_pending(cl2ccc(obj), cp);
767 /* Do not set Dirty bit here so that in case IO is
768 * started before the page is really made dirty, we
769 * still have chance to detect it. */
770 result = cl_page_cache_add(env, io, page, CRT_WRITE);
771 LASSERT(cl_page_is_owned(page, io));
775 cl_page_unmap(env, io, page);
776 cl_page_discard(env, io, page);
777 cl_page_disown(env, io, page);
779 cl_page_put(env, page);
781 /* we're in big trouble, what can we do now? */
782 if (result == -EDQUOT)
786 cl_page_disown(env, io, page);
790 last = cl_index(obj, size - 1);
792 * The ft_index is only used in the case of
793 * a mkwrite action. We need to check
794 * our assertions are correct, since
795 * we should have caught this above
797 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last);
798 if (fio->ft_index == last)
800 * Last page is mapped partially.
802 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
804 fio->ft_nob = cl_page_size(obj);
806 lu_ref_add(&page->cp_reference, "fault", io);
810 /* return unlocked vmpage to avoid deadlocking */
813 cfio->fault.ft_flags &= ~VM_FAULT_LOCKED;
817 static int vvp_io_fsync_start(const struct lu_env *env,
818 const struct cl_io_slice *ios)
820 /* we should mark TOWRITE bit to each dirty page in radix tree to
821 * verify pages have been written, but this is difficult because of
826 static int vvp_io_read_page(const struct lu_env *env,
827 const struct cl_io_slice *ios,
828 const struct cl_page_slice *slice)
830 struct cl_io *io = ios->cis_io;
831 struct cl_object *obj = slice->cpl_obj;
832 struct ccc_page *cp = cl2ccc_page(slice);
833 struct cl_page *page = slice->cpl_page;
834 struct inode *inode = ccc_object_inode(obj);
835 struct ll_sb_info *sbi = ll_i2sbi(inode);
836 struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd;
837 struct ll_readahead_state *ras = &fd->fd_ras;
838 struct page *vmpage = cp->cpg_page;
839 struct cl_2queue *queue = &io->ci_queue;
842 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
843 LASSERT(slice->cpl_obj == obj);
845 if (sbi->ll_ra_info.ra_max_pages_per_file &&
846 sbi->ll_ra_info.ra_max_pages)
847 ras_update(sbi, inode, ras, page->cp_index,
848 cp->cpg_defer_uptodate);
850 /* Sanity check whether the page is protected by a lock. */
851 rc = cl_page_is_under_lock(env, io, page);
853 CL_PAGE_HEADER(D_WARNING, env, page, "%s: %d\n",
854 rc == -ENODATA ? "without a lock" :
860 if (cp->cpg_defer_uptodate) {
862 cl_page_export(env, page, 1);
865 * Add page into the queue even when it is marked uptodate above.
866 * this will unlock it automatically as part of cl_page_list_disown().
868 cl_2queue_add(queue, page);
869 if (sbi->ll_ra_info.ra_max_pages_per_file &&
870 sbi->ll_ra_info.ra_max_pages)
871 ll_readahead(env, io, ras,
872 vmpage->mapping, &queue->c2_qin, fd->fd_flags);
877 static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io,
878 struct cl_page *page, struct ccc_page *cp,
879 enum cl_req_type crt)
881 struct cl_2queue *queue;
884 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
886 queue = &io->ci_queue;
887 cl_2queue_init_page(queue, page);
889 result = cl_io_submit_sync(env, io, crt, queue, 0);
890 LASSERT(cl_page_is_owned(page, io));
894 * in CRT_WRITE case page is left locked even in case of
897 cl_page_list_disown(env, io, &queue->c2_qin);
898 cl_2queue_fini(env, queue);
904 * Prepare partially written-to page for a write.
906 static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io,
907 struct cl_object *obj, struct cl_page *pg,
909 unsigned from, unsigned to)
911 struct cl_attr *attr = ccc_env_thread_attr(env);
912 loff_t offset = cl_offset(obj, pg->cp_index);
915 cl_object_attr_lock(obj);
916 result = cl_object_attr_get(env, obj, attr);
917 cl_object_attr_unlock(obj);
920 * If are writing to a new page, no need to read old data.
921 * The extent locking will have updated the KMS, and for our
922 * purposes here we can treat it like i_size.
924 if (attr->cat_kms <= offset) {
925 char *kaddr = kmap_atomic(cp->cpg_page);
927 memset(kaddr, 0, cl_page_size(obj));
928 kunmap_atomic(kaddr);
929 } else if (cp->cpg_defer_uptodate)
932 result = vvp_page_sync_io(env, io, pg, cp, CRT_READ);
934 * In older implementations, obdo_refresh_inode is called here
935 * to update the inode because the write might modify the
936 * object info at OST. However, this has been proven useless,
937 * since LVB functions will be called when user space program
938 * tries to retrieve inode attribute. Also, see bug 15909 for
942 cl_page_export(env, pg, 1);
947 static int vvp_io_prepare_write(const struct lu_env *env,
948 const struct cl_io_slice *ios,
949 const struct cl_page_slice *slice,
950 unsigned from, unsigned to)
952 struct cl_object *obj = slice->cpl_obj;
953 struct ccc_page *cp = cl2ccc_page(slice);
954 struct cl_page *pg = slice->cpl_page;
955 struct page *vmpage = cp->cpg_page;
959 LINVRNT(cl_page_is_vmlocked(env, pg));
960 LASSERT(vmpage->mapping->host == ccc_object_inode(obj));
964 CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to);
965 if (!PageUptodate(vmpage)) {
967 * We're completely overwriting an existing page, so _don't_
968 * set it up to date until commit_write
970 if (from == 0 && to == PAGE_CACHE_SIZE) {
971 CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n");
972 POISON_PAGE(page, 0x11);
974 result = vvp_io_prepare_partial(env, ios->cis_io, obj,
977 CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n");
981 static int vvp_io_commit_write(const struct lu_env *env,
982 const struct cl_io_slice *ios,
983 const struct cl_page_slice *slice,
984 unsigned from, unsigned to)
986 struct cl_object *obj = slice->cpl_obj;
987 struct cl_io *io = ios->cis_io;
988 struct ccc_page *cp = cl2ccc_page(slice);
989 struct cl_page *pg = slice->cpl_page;
990 struct inode *inode = ccc_object_inode(obj);
991 struct ll_sb_info *sbi = ll_i2sbi(inode);
992 struct ll_inode_info *lli = ll_i2info(inode);
993 struct page *vmpage = cp->cpg_page;
999 LINVRNT(cl_page_is_vmlocked(env, pg));
1000 LASSERT(vmpage->mapping->host == inode);
1002 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "committing page write\n");
1003 CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to);
1006 * queue a write for some time in the future the first time we
1009 * This is different from what other file systems do: they usually
1010 * just mark page (and some of its buffers) dirty and rely on
1011 * balance_dirty_pages() to start a write-back. Lustre wants write-back
1012 * to be started earlier for the following reasons:
1014 * (1) with a large number of clients we need to limit the amount
1015 * of cached data on the clients a lot;
1017 * (2) large compute jobs generally want compute-only then io-only
1018 * and the IO should complete as quickly as possible;
1020 * (3) IO is batched up to the RPC size and is async until the
1021 * client max cache is hit
1022 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
1025 if (!PageDirty(vmpage)) {
1026 tallyop = LPROC_LL_DIRTY_MISSES;
1027 result = cl_page_cache_add(env, io, pg, CRT_WRITE);
1029 /* page was added into cache successfully. */
1030 set_page_dirty(vmpage);
1031 vvp_write_pending(cl2ccc(obj), cp);
1032 } else if (result == -EDQUOT) {
1033 pgoff_t last_index = i_size_read(inode) >> PAGE_CACHE_SHIFT;
1034 bool need_clip = true;
1037 * Client ran out of disk space grant. Possible
1040 * (a) do a sync write, renewing grant;
1042 * (b) stop writing on this stripe, switch to the
1045 * (b) is a part of "parallel io" design that is the
1046 * ultimate goal. (a) is what "old" client did, and
1047 * what the new code continues to do for the time
1050 if (last_index > pg->cp_index) {
1051 to = PAGE_CACHE_SIZE;
1053 } else if (last_index == pg->cp_index) {
1054 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
1059 cl_page_clip(env, pg, 0, to);
1060 result = vvp_page_sync_io(env, io, pg, cp, CRT_WRITE);
1062 CERROR("Write page %lu of inode %p failed %d\n",
1063 pg->cp_index, inode, result);
1066 tallyop = LPROC_LL_DIRTY_HITS;
1069 ll_stats_ops_tally(sbi, tallyop, 1);
1071 /* Inode should be marked DIRTY even if no new page was marked DIRTY
1072 * because page could have been not flushed between 2 modifications.
1073 * It is important the file is marked DIRTY as soon as the I/O is done
1074 * Indeed, when cache is flushed, file could be already closed and it
1075 * is too late to warn the MDT.
1076 * It is acceptable that file is marked DIRTY even if I/O is dropped
1077 * for some reasons before being flushed to OST.
1080 spin_lock(&lli->lli_lock);
1081 lli->lli_flags |= LLIF_DATA_MODIFIED;
1082 spin_unlock(&lli->lli_lock);
1085 size = cl_offset(obj, pg->cp_index) + to;
1087 ll_inode_size_lock(inode);
1089 if (size > i_size_read(inode)) {
1090 cl_isize_write_nolock(inode, size);
1091 CDEBUG(D_VFSTRACE, DFID" updating i_size %lu\n",
1092 PFID(lu_object_fid(&obj->co_lu)),
1093 (unsigned long)size);
1095 cl_page_export(env, pg, 1);
1097 if (size > i_size_read(inode))
1098 cl_page_discard(env, io, pg);
1100 ll_inode_size_unlock(inode);
1104 static const struct cl_io_operations vvp_io_ops = {
1107 .cio_fini = vvp_io_read_fini,
1108 .cio_lock = vvp_io_read_lock,
1109 .cio_start = vvp_io_read_start,
1110 .cio_advance = ccc_io_advance
1113 .cio_fini = vvp_io_fini,
1114 .cio_lock = vvp_io_write_lock,
1115 .cio_start = vvp_io_write_start,
1116 .cio_advance = ccc_io_advance
1119 .cio_fini = vvp_io_setattr_fini,
1120 .cio_iter_init = vvp_io_setattr_iter_init,
1121 .cio_lock = vvp_io_setattr_lock,
1122 .cio_start = vvp_io_setattr_start,
1123 .cio_end = vvp_io_setattr_end
1126 .cio_fini = vvp_io_fault_fini,
1127 .cio_iter_init = vvp_io_fault_iter_init,
1128 .cio_lock = vvp_io_fault_lock,
1129 .cio_start = vvp_io_fault_start,
1130 .cio_end = ccc_io_end
1133 .cio_start = vvp_io_fsync_start,
1134 .cio_fini = vvp_io_fini
1137 .cio_fini = vvp_io_fini
1140 .cio_read_page = vvp_io_read_page,
1141 .cio_prepare_write = vvp_io_prepare_write,
1142 .cio_commit_write = vvp_io_commit_write
1145 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1148 struct vvp_io *vio = vvp_env_io(env);
1149 struct ccc_io *cio = ccc_env_io(env);
1150 struct inode *inode = ccc_object_inode(obj);
1153 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
1155 CDEBUG(D_VFSTRACE, DFID
1156 " ignore/verify layout %d/%d, layout version %d restore needed %d\n",
1157 PFID(lu_object_fid(&obj->co_lu)),
1158 io->ci_ignore_layout, io->ci_verify_layout,
1159 cio->cui_layout_gen, io->ci_restore_needed);
1161 CL_IO_SLICE_CLEAN(cio, cui_cl);
1162 cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops);
1163 vio->cui_ra_window_set = 0;
1165 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1167 struct ll_inode_info *lli = ll_i2info(inode);
1169 count = io->u.ci_rw.crw_count;
1170 /* "If nbyte is 0, read() will return 0 and have no other
1171 * results." -- Single Unix Spec */
1175 cio->cui_tot_count = count;
1176 cio->cui_tot_nrsegs = 0;
1178 /* for read/write, we store the jobid in the inode, and
1179 * it'll be fetched by osc when building RPC.
1181 * it's not accurate if the file is shared by different
1184 lustre_get_jobid(lli->lli_jobid);
1185 } else if (io->ci_type == CIT_SETATTR) {
1186 if (!cl_io_is_trunc(io))
1187 io->ci_lockreq = CILR_MANDATORY;
1190 /* ignore layout change for generic CIT_MISC but not for glimpse.
1191 * io context for glimpse must set ci_verify_layout to true,
1192 * see cl_glimpse_size0() for details. */
1193 if (io->ci_type == CIT_MISC && !io->ci_verify_layout)
1194 io->ci_ignore_layout = 1;
1196 /* Enqueue layout lock and get layout version. We need to do this
1197 * even for operations requiring to open file, such as read and write,
1198 * because it might not grant layout lock in IT_OPEN. */
1199 if (result == 0 && !io->ci_ignore_layout) {
1200 result = ll_layout_refresh(inode, &cio->cui_layout_gen);
1201 if (result == -ENOENT)
1202 /* If the inode on MDS has been removed, but the objects
1203 * on OSTs haven't been destroyed (async unlink), layout
1204 * fetch will return -ENOENT, we'd ingore this error
1205 * and continue with dirty flush. LU-3230. */
1208 CERROR("%s: refresh file layout " DFID " error %d.\n",
1209 ll_get_fsname(inode->i_sb, NULL, 0),
1210 PFID(lu_object_fid(&obj->co_lu)), result);
1216 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
1217 const struct cl_io_slice *slice)
1219 /* Caling just for assertion */
1220 cl2ccc_io(env, slice);
1221 return vvp_env_io(env);