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, Whamcloud, Inc.
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;
68 /*****************************************************************************
74 static int vvp_io_fault_iter_init(const struct lu_env *env,
75 const struct cl_io_slice *ios)
77 struct vvp_io *vio = cl2vvp_io(env, ios);
78 struct inode *inode = ccc_object_inode(ios->cis_obj);
81 cl2ccc_io(env, ios)->cui_fd->fd_file->f_dentry->d_inode);
82 vio->u.fault.ft_mtime = LTIME_S(inode->i_mtime);
86 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
88 struct cl_io *io = ios->cis_io;
89 struct cl_object *obj = io->ci_obj;
90 struct ccc_io *cio = cl2ccc_io(env, ios);
92 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
94 CDEBUG(D_VFSTRACE, "ignore/verify layout %d/%d, layout version %d.\n",
95 io->ci_ignore_layout, io->ci_verify_layout, cio->cui_layout_gen);
97 if (!io->ci_ignore_layout && io->ci_verify_layout) {
100 /* check layout version */
101 ll_layout_refresh(ccc_object_inode(obj), &gen);
102 io->ci_need_restart = cio->cui_layout_gen != gen;
103 if (io->ci_need_restart)
104 CDEBUG(D_VFSTRACE, "layout changed from %d to %d.\n",
105 cio->cui_layout_gen, gen);
109 static void vvp_io_fault_fini(const struct lu_env *env,
110 const struct cl_io_slice *ios)
112 struct cl_io *io = ios->cis_io;
113 struct cl_page *page = io->u.ci_fault.ft_page;
115 CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj));
118 lu_ref_del(&page->cp_reference, "fault", io);
119 cl_page_put(env, page);
120 io->u.ci_fault.ft_page = NULL;
122 vvp_io_fini(env, ios);
125 enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
128 * we only want to hold PW locks if the mmap() can generate
129 * writes back to the file and that only happens in shared
132 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
137 static int vvp_mmap_locks(const struct lu_env *env,
138 struct ccc_io *vio, struct cl_io *io)
140 struct ccc_thread_info *cti = ccc_env_info(env);
141 struct mm_struct *mm = current->mm;
142 struct vm_area_struct *vma;
143 struct cl_lock_descr *descr = &cti->cti_descr;
144 ldlm_policy_data_t policy;
151 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
153 if (!cl_is_normalio(env, io))
156 if (vio->cui_iov == NULL) /* nfs or loop back device write */
159 /* No MM (e.g. NFS)? No vmas too. */
163 for (seg = 0; seg < vio->cui_nrsegs; seg++) {
164 const struct iovec *iv = &vio->cui_iov[seg];
166 addr = (unsigned long)iv->iov_base;
171 count += addr & (~CFS_PAGE_MASK);
172 addr &= CFS_PAGE_MASK;
174 down_read(&mm->mmap_sem);
175 while((vma = our_vma(mm, addr, count)) != NULL) {
176 struct inode *inode = vma->vm_file->f_dentry->d_inode;
177 int flags = CEF_MUST;
179 if (ll_file_nolock(vma->vm_file)) {
181 * For no lock case, a lockless lock will be
188 * XXX: Required lock mode can be weakened: CIT_WRITE
189 * io only ever reads user level buffer, and CIT_READ
192 policy_from_vma(&policy, vma, addr, count);
193 descr->cld_mode = vvp_mode_from_vma(vma);
194 descr->cld_obj = ll_i2info(inode)->lli_clob;
195 descr->cld_start = cl_index(descr->cld_obj,
196 policy.l_extent.start);
197 descr->cld_end = cl_index(descr->cld_obj,
198 policy.l_extent.end);
199 descr->cld_enq_flags = flags;
200 result = cl_io_lock_alloc_add(env, io, descr);
202 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
203 descr->cld_mode, descr->cld_start,
209 if (vma->vm_end - addr >= count)
212 count -= vma->vm_end - addr;
215 up_read(&mm->mmap_sem);
220 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
221 enum cl_lock_mode mode, loff_t start, loff_t end)
223 struct ccc_io *cio = ccc_env_io(env);
227 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
230 ccc_io_update_iov(env, cio, io);
232 if (io->u.ci_rw.crw_nonblock)
233 ast_flags |= CEF_NONBLOCK;
234 result = vvp_mmap_locks(env, cio, io);
236 result = ccc_io_one_lock(env, io, ast_flags, mode, start, end);
240 static int vvp_io_read_lock(const struct lu_env *env,
241 const struct cl_io_slice *ios)
243 struct cl_io *io = ios->cis_io;
244 struct ll_inode_info *lli = ll_i2info(ccc_object_inode(io->ci_obj));
248 /* XXX: Layer violation, we shouldn't see lsm at llite level. */
249 if (lli->lli_has_smd) /* lsm-less file doesn't need to lock */
250 result = vvp_io_rw_lock(env, io, CLM_READ,
251 io->u.ci_rd.rd.crw_pos,
252 io->u.ci_rd.rd.crw_pos +
253 io->u.ci_rd.rd.crw_count - 1);
259 static int vvp_io_fault_lock(const struct lu_env *env,
260 const struct cl_io_slice *ios)
262 struct cl_io *io = ios->cis_io;
263 struct vvp_io *vio = cl2vvp_io(env, ios);
265 * XXX LDLM_FL_CBPENDING
267 return ccc_io_one_lock_index
268 (env, io, 0, vvp_mode_from_vma(vio->u.fault.ft_vma),
269 io->u.ci_fault.ft_index, io->u.ci_fault.ft_index);
272 static int vvp_io_write_lock(const struct lu_env *env,
273 const struct cl_io_slice *ios)
275 struct cl_io *io = ios->cis_io;
279 if (io->u.ci_wr.wr_append) {
281 end = OBD_OBJECT_EOF;
283 start = io->u.ci_wr.wr.crw_pos;
284 end = start + io->u.ci_wr.wr.crw_count - 1;
286 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
289 static int vvp_io_setattr_iter_init(const struct lu_env *env,
290 const struct cl_io_slice *ios)
292 struct ccc_io *cio = ccc_env_io(env);
293 struct inode *inode = ccc_object_inode(ios->cis_obj);
296 * We really need to get our PW lock before we change inode->i_size.
297 * If we don't we can race with other i_size updaters on our node,
298 * like ll_file_read. We can also race with i_size propogation to
299 * other nodes through dirtying and writeback of final cached pages.
300 * This last one is especially bad for racing o_append users on other
303 if (cl_io_is_trunc(ios->cis_io))
304 inode_dio_write_done(inode);
305 mutex_unlock(&inode->i_mutex);
306 cio->u.setattr.cui_locks_released = 1;
311 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
313 * Handles "lockless io" mode when extent locking is done by server.
315 static int vvp_io_setattr_lock(const struct lu_env *env,
316 const struct cl_io_slice *ios)
318 struct ccc_io *cio = ccc_env_io(env);
319 struct cl_io *io = ios->cis_io;
323 if (cl_io_is_trunc(io)) {
324 new_size = io->u.ci_setattr.sa_attr.lvb_size;
326 enqflags = CEF_DISCARD_DATA;
328 if ((io->u.ci_setattr.sa_attr.lvb_mtime >=
329 io->u.ci_setattr.sa_attr.lvb_ctime) ||
330 (io->u.ci_setattr.sa_attr.lvb_atime >=
331 io->u.ci_setattr.sa_attr.lvb_ctime))
335 cio->u.setattr.cui_local_lock = SETATTR_EXTENT_LOCK;
336 return ccc_io_one_lock(env, io, enqflags, CLM_WRITE,
337 new_size, OBD_OBJECT_EOF);
340 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
344 * Only ll_inode_size_lock is taken at this level.
346 ll_inode_size_lock(inode);
347 result = vmtruncate(inode, size);
348 ll_inode_size_unlock(inode);
353 static int vvp_io_setattr_trunc(const struct lu_env *env,
354 const struct cl_io_slice *ios,
355 struct inode *inode, loff_t size)
357 inode_dio_wait(inode);
361 static int vvp_io_setattr_time(const struct lu_env *env,
362 const struct cl_io_slice *ios)
364 struct cl_io *io = ios->cis_io;
365 struct cl_object *obj = io->ci_obj;
366 struct cl_attr *attr = ccc_env_thread_attr(env);
368 unsigned valid = CAT_CTIME;
370 cl_object_attr_lock(obj);
371 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
372 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
373 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
376 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
377 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
380 result = cl_object_attr_set(env, obj, attr, valid);
381 cl_object_attr_unlock(obj);
386 static int vvp_io_setattr_start(const struct lu_env *env,
387 const struct cl_io_slice *ios)
389 struct ccc_io *cio = cl2ccc_io(env, ios);
390 struct cl_io *io = ios->cis_io;
391 struct inode *inode = ccc_object_inode(io->ci_obj);
393 LASSERT(cio->u.setattr.cui_locks_released);
395 mutex_lock(&inode->i_mutex);
396 cio->u.setattr.cui_locks_released = 0;
398 if (cl_io_is_trunc(io))
399 return vvp_io_setattr_trunc(env, ios, inode,
400 io->u.ci_setattr.sa_attr.lvb_size);
402 return vvp_io_setattr_time(env, ios);
405 static void vvp_io_setattr_end(const struct lu_env *env,
406 const struct cl_io_slice *ios)
408 struct cl_io *io = ios->cis_io;
409 struct inode *inode = ccc_object_inode(io->ci_obj);
411 if (!cl_io_is_trunc(io))
414 /* Truncate in memory pages - they must be clean pages because osc
415 * has already notified to destroy osc_extents. */
416 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
419 static void vvp_io_setattr_fini(const struct lu_env *env,
420 const struct cl_io_slice *ios)
422 struct ccc_io *cio = ccc_env_io(env);
423 struct cl_io *io = ios->cis_io;
424 struct inode *inode = ccc_object_inode(ios->cis_io->ci_obj);
426 if (cio->u.setattr.cui_locks_released) {
427 mutex_lock(&inode->i_mutex);
428 if (cl_io_is_trunc(io))
429 inode_dio_wait(inode);
430 cio->u.setattr.cui_locks_released = 0;
432 vvp_io_fini(env, ios);
435 #ifdef HAVE_FILE_READV
436 static ssize_t lustre_generic_file_read(struct file *file,
437 struct ccc_io *vio, loff_t *ppos)
439 return generic_file_readv(file, vio->cui_iov, vio->cui_nrsegs, ppos);
442 static ssize_t lustre_generic_file_write(struct file *file,
443 struct ccc_io *vio, loff_t *ppos)
445 return generic_file_writev(file, vio->cui_iov, vio->cui_nrsegs, ppos);
448 static ssize_t lustre_generic_file_read(struct file *file,
449 struct ccc_io *vio, loff_t *ppos)
451 return generic_file_aio_read(vio->cui_iocb, vio->cui_iov,
452 vio->cui_nrsegs, *ppos);
455 static ssize_t lustre_generic_file_write(struct file *file,
456 struct ccc_io *vio, loff_t *ppos)
458 return generic_file_aio_write(vio->cui_iocb, vio->cui_iov,
459 vio->cui_nrsegs, *ppos);
463 static int vvp_io_read_start(const struct lu_env *env,
464 const struct cl_io_slice *ios)
466 struct vvp_io *vio = cl2vvp_io(env, ios);
467 struct ccc_io *cio = cl2ccc_io(env, ios);
468 struct cl_io *io = ios->cis_io;
469 struct cl_object *obj = io->ci_obj;
470 struct inode *inode = ccc_object_inode(obj);
471 struct ll_ra_read *bead = &vio->cui_bead;
472 struct file *file = cio->cui_fd->fd_file;
475 loff_t pos = io->u.ci_rd.rd.crw_pos;
476 long cnt = io->u.ci_rd.rd.crw_count;
477 long tot = cio->cui_tot_count;
480 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
482 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
484 result = ccc_prep_size(env, obj, io, pos, tot, &exceed);
487 else if (exceed != 0)
490 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
491 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
492 inode->i_ino, cnt, pos, i_size_read(inode));
494 /* turn off the kernel's read-ahead */
495 cio->cui_fd->fd_file->f_ra.ra_pages = 0;
497 /* initialize read-ahead window once per syscall */
498 if (!vio->cui_ra_window_set) {
499 vio->cui_ra_window_set = 1;
500 bead->lrr_start = cl_index(obj, pos);
502 * XXX: explicit CFS_PAGE_SIZE
504 bead->lrr_count = cl_index(obj, tot + CFS_PAGE_SIZE - 1);
505 ll_ra_read_in(file, bead);
510 switch (vio->cui_io_subtype) {
512 result = lustre_generic_file_read(file, cio, &pos);
514 #ifdef HAVE_KERNEL_SENDFILE
516 result = generic_file_sendfile(file, &pos, cnt,
517 vio->u.sendfile.cui_actor,
518 vio->u.sendfile.cui_target);
521 #ifdef HAVE_KERNEL_SPLICE_READ
523 result = generic_file_splice_read(file, &pos,
524 vio->u.splice.cui_pipe, cnt,
525 vio->u.splice.cui_flags);
526 /* LU-1109: do splice read stripe by stripe otherwise if it
527 * may make nfsd stuck if this read occupied all internal pipe
533 CERROR("Wrong IO type %u\n", vio->cui_io_subtype);
541 io->ci_nob += result;
542 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
543 cio->cui_fd, pos, result, 0);
549 static void vvp_io_read_fini(const struct lu_env *env, const struct cl_io_slice *ios)
551 struct vvp_io *vio = cl2vvp_io(env, ios);
552 struct ccc_io *cio = cl2ccc_io(env, ios);
554 if (vio->cui_ra_window_set)
555 ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead);
557 vvp_io_fini(env, ios);
560 static int vvp_io_write_start(const struct lu_env *env,
561 const struct cl_io_slice *ios)
563 struct ccc_io *cio = cl2ccc_io(env, ios);
564 struct cl_io *io = ios->cis_io;
565 struct cl_object *obj = io->ci_obj;
566 struct inode *inode = ccc_object_inode(obj);
567 struct file *file = cio->cui_fd->fd_file;
569 loff_t pos = io->u.ci_wr.wr.crw_pos;
570 size_t cnt = io->u.ci_wr.wr.crw_count;
574 if (cl_io_is_append(io)) {
576 * PARALLEL IO This has to be changed for parallel IO doing
577 * out-of-order writes.
579 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
580 #ifndef HAVE_FILE_WRITEV
581 cio->cui_iocb->ki_pos = pos;
585 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
587 if (cio->cui_iov == NULL) /* from a temp io in ll_cl_init(). */
590 result = lustre_generic_file_write(file, cio, &pos);
595 io->ci_nob += result;
596 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
597 cio->cui_fd, pos, result, 0);
603 #ifndef HAVE_VM_OP_FAULT
604 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
608 vmpage = filemap_nopage(cfio->ft_vma, cfio->nopage.ft_address,
609 cfio->nopage.ft_type);
611 if (vmpage == NOPAGE_SIGBUS) {
612 CDEBUG(D_PAGE, "got addr %lu type %lx - SIGBUS\n",
613 cfio->nopage.ft_address,(long)cfio->nopage.ft_type);
615 } else if (vmpage == NOPAGE_OOM) {
616 CDEBUG(D_PAGE, "got addr %lu type %lx - OOM\n",
617 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
621 LL_CDEBUG_PAGE(D_PAGE, vmpage, "got addr %lu type %lx\n",
622 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
624 cfio->ft_vmpage = vmpage;
630 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
632 struct vm_fault *vmf = cfio->fault.ft_vmf;
634 cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, vmf);
637 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
638 vmf->virtual_address);
639 if (unlikely(!(cfio->fault.ft_flags & VM_FAULT_LOCKED))) {
640 lock_page(vmf->page);
641 cfio->fault.ft_flags &= VM_FAULT_LOCKED;
644 cfio->ft_vmpage = vmf->page;
648 if (cfio->fault.ft_flags & VM_FAULT_SIGBUS) {
649 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
653 if (cfio->fault.ft_flags & VM_FAULT_OOM) {
654 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
658 if (cfio->fault.ft_flags & VM_FAULT_RETRY)
661 CERROR("unknow error in page fault %d!\n", cfio->fault.ft_flags);
667 static int vvp_io_fault_start(const struct lu_env *env,
668 const struct cl_io_slice *ios)
670 struct vvp_io *vio = cl2vvp_io(env, ios);
671 struct cl_io *io = ios->cis_io;
672 struct cl_object *obj = io->ci_obj;
673 struct inode *inode = ccc_object_inode(obj);
674 struct cl_fault_io *fio = &io->u.ci_fault;
675 struct vvp_fault_io *cfio = &vio->u.fault;
678 cfs_page_t *vmpage = NULL;
679 struct cl_page *page;
681 pgoff_t last; /* last page in a file data region */
683 if (fio->ft_executable &&
684 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
686 " changed while waiting for the page fault lock\n",
687 PFID(lu_object_fid(&obj->co_lu)));
689 /* offset of the last byte on the page */
690 offset = cl_offset(obj, fio->ft_index + 1) - 1;
691 LASSERT(cl_index(obj, offset) == fio->ft_index);
692 result = ccc_prep_size(env, obj, io, 0, offset + 1, NULL);
696 /* must return locked page */
697 if (fio->ft_mkwrite) {
698 LASSERT(cfio->ft_vmpage != NULL);
699 lock_page(cfio->ft_vmpage);
701 result = vvp_io_kernel_fault(cfio);
706 vmpage = cfio->ft_vmpage;
707 LASSERT(PageLocked(vmpage));
709 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
710 ll_invalidate_page(vmpage);
712 size = i_size_read(inode);
713 /* Though we have already held a cl_lock upon this page, but
714 * it still can be truncated locally. */
715 if (unlikely((vmpage->mapping != inode->i_mapping) ||
716 (page_offset(vmpage) > size))) {
717 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
719 /* return +1 to stop cl_io_loop() and ll_fault() will catch
721 GOTO(out, result = +1);
725 if (fio->ft_mkwrite ) {
728 * Capture the size while holding the lli_trunc_sem from above
729 * we want to make sure that we complete the mkwrite action
730 * while holding this lock. We need to make sure that we are
731 * not past the end of the file.
733 last_index = cl_index(obj, size - 1);
734 if (last_index < fio->ft_index) {
736 "llite: mkwrite and truncate race happened: "
738 vmpage->mapping,fio->ft_index,last_index);
740 * We need to return if we are
741 * passed the end of the file. This will propagate
742 * up the call stack to ll_page_mkwrite where
743 * we will return VM_FAULT_NOPAGE. Any non-negative
744 * value returned here will be silently
745 * converted to 0. If the vmpage->mapping is null
746 * the error code would be converted back to ENODATA
747 * in ll_page_mkwrite0. Thus we return -ENODATA
748 * to handle both cases
750 GOTO(out, result = -ENODATA);
754 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
756 GOTO(out, result = PTR_ERR(page));
758 /* if page is going to be written, we should add this page into cache
760 if (fio->ft_mkwrite) {
761 wait_on_page_writeback(vmpage);
762 if (set_page_dirty(vmpage)) {
765 /* vvp_page_assume() calls wait_on_page_writeback(). */
766 cl_page_assume(env, io, page);
768 cp = cl2ccc_page(cl_page_at(page, &vvp_device_type));
769 vvp_write_pending(cl2ccc(obj), cp);
771 /* Do not set Dirty bit here so that in case IO is
772 * started before the page is really made dirty, we
773 * still have chance to detect it. */
774 result = cl_page_cache_add(env, io, page, CRT_WRITE);
775 LASSERT(cl_page_is_owned(page, io));
779 cl_page_unmap(env, io, page);
780 cl_page_discard(env, io, page);
781 cl_page_disown(env, io, page);
783 cl_page_put(env, page);
785 /* we're in big trouble, what can we do now? */
786 if (result == -EDQUOT)
790 cl_page_disown(env, io, page);
794 last = cl_index(obj, size - 1);
796 * The ft_index is only used in the case of
797 * a mkwrite action. We need to check
798 * our assertions are correct, since
799 * we should have caught this above
801 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last);
802 if (fio->ft_index == last)
804 * Last page is mapped partially.
806 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
808 fio->ft_nob = cl_page_size(obj);
810 lu_ref_add(&page->cp_reference, "fault", io);
815 /* return unlocked vmpage to avoid deadlocking */
818 #ifdef HAVE_VM_OP_FAULT
819 cfio->fault.ft_flags &= ~VM_FAULT_LOCKED;
824 static int vvp_io_fsync_start(const struct lu_env *env,
825 const struct cl_io_slice *ios)
827 /* we should mark TOWRITE bit to each dirty page in radix tree to
828 * verify pages have been written, but this is difficult because of
833 static int vvp_io_read_page(const struct lu_env *env,
834 const struct cl_io_slice *ios,
835 const struct cl_page_slice *slice)
837 struct cl_io *io = ios->cis_io;
838 struct cl_object *obj = slice->cpl_obj;
839 struct ccc_page *cp = cl2ccc_page(slice);
840 struct cl_page *page = slice->cpl_page;
841 struct inode *inode = ccc_object_inode(obj);
842 struct ll_sb_info *sbi = ll_i2sbi(inode);
843 struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd;
844 struct ll_readahead_state *ras = &fd->fd_ras;
845 cfs_page_t *vmpage = cp->cpg_page;
846 struct cl_2queue *queue = &io->ci_queue;
849 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
850 LASSERT(slice->cpl_obj == obj);
854 if (sbi->ll_ra_info.ra_max_pages_per_file &&
855 sbi->ll_ra_info.ra_max_pages)
856 ras_update(sbi, inode, ras, page->cp_index,
857 cp->cpg_defer_uptodate);
859 /* Sanity check whether the page is protected by a lock. */
860 rc = cl_page_is_under_lock(env, io, page);
862 CL_PAGE_HEADER(D_WARNING, env, page, "%s: %d\n",
863 rc == -ENODATA ? "without a lock" :
869 if (cp->cpg_defer_uptodate) {
871 cl_page_export(env, page, 1);
874 * Add page into the queue even when it is marked uptodate above.
875 * this will unlock it automatically as part of cl_page_list_disown().
877 cl_2queue_add(queue, page);
878 if (sbi->ll_ra_info.ra_max_pages_per_file &&
879 sbi->ll_ra_info.ra_max_pages)
880 ll_readahead(env, io, ras,
881 vmpage->mapping, &queue->c2_qin, fd->fd_flags);
886 static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io,
887 struct cl_page *page, struct ccc_page *cp,
888 enum cl_req_type crt)
890 struct cl_2queue *queue;
893 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
895 queue = &io->ci_queue;
896 cl_2queue_init_page(queue, page);
898 result = cl_io_submit_sync(env, io, crt, queue, 0);
899 LASSERT(cl_page_is_owned(page, io));
903 * in CRT_WRITE case page is left locked even in case of
906 cl_page_list_disown(env, io, &queue->c2_qin);
907 cl_2queue_fini(env, queue);
913 * Prepare partially written-to page for a write.
915 static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io,
916 struct cl_object *obj, struct cl_page *pg,
918 unsigned from, unsigned to)
920 struct cl_attr *attr = ccc_env_thread_attr(env);
921 loff_t offset = cl_offset(obj, pg->cp_index);
924 cl_object_attr_lock(obj);
925 result = cl_object_attr_get(env, obj, attr);
926 cl_object_attr_unlock(obj);
929 * If are writing to a new page, no need to read old data.
930 * The extent locking will have updated the KMS, and for our
931 * purposes here we can treat it like i_size.
933 if (attr->cat_kms <= offset) {
934 char *kaddr = ll_kmap_atomic(cp->cpg_page, KM_USER0);
936 memset(kaddr, 0, cl_page_size(obj));
937 ll_kunmap_atomic(kaddr, KM_USER0);
938 } else if (cp->cpg_defer_uptodate)
941 result = vvp_page_sync_io(env, io, pg, cp, CRT_READ);
943 * In older implementations, obdo_refresh_inode is called here
944 * to update the inode because the write might modify the
945 * object info at OST. However, this has been proven useless,
946 * since LVB functions will be called when user space program
947 * tries to retrieve inode attribute. Also, see bug 15909 for
951 cl_page_export(env, pg, 1);
956 static int vvp_io_prepare_write(const struct lu_env *env,
957 const struct cl_io_slice *ios,
958 const struct cl_page_slice *slice,
959 unsigned from, unsigned to)
961 struct cl_object *obj = slice->cpl_obj;
962 struct ccc_page *cp = cl2ccc_page(slice);
963 struct cl_page *pg = slice->cpl_page;
964 cfs_page_t *vmpage = cp->cpg_page;
970 LINVRNT(cl_page_is_vmlocked(env, pg));
971 LASSERT(vmpage->mapping->host == ccc_object_inode(obj));
975 CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to);
976 if (!PageUptodate(vmpage)) {
978 * We're completely overwriting an existing page, so _don't_
979 * set it up to date until commit_write
981 if (from == 0 && to == CFS_PAGE_SIZE) {
982 CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n");
983 POISON_PAGE(page, 0x11);
985 result = vvp_io_prepare_partial(env, ios->cis_io, obj,
988 CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n");
992 static int vvp_io_commit_write(const struct lu_env *env,
993 const struct cl_io_slice *ios,
994 const struct cl_page_slice *slice,
995 unsigned from, unsigned to)
997 struct cl_object *obj = slice->cpl_obj;
998 struct cl_io *io = ios->cis_io;
999 struct ccc_page *cp = cl2ccc_page(slice);
1000 struct cl_page *pg = slice->cpl_page;
1001 struct inode *inode = ccc_object_inode(obj);
1002 struct ll_sb_info *sbi = ll_i2sbi(inode);
1003 cfs_page_t *vmpage = cp->cpg_page;
1011 LINVRNT(cl_page_is_vmlocked(env, pg));
1012 LASSERT(vmpage->mapping->host == inode);
1014 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "commiting page write\n");
1015 CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to);
1018 * queue a write for some time in the future the first time we
1021 * This is different from what other file systems do: they usually
1022 * just mark page (and some of its buffers) dirty and rely on
1023 * balance_dirty_pages() to start a write-back. Lustre wants write-back
1024 * to be started earlier for the following reasons:
1026 * (1) with a large number of clients we need to limit the amount
1027 * of cached data on the clients a lot;
1029 * (2) large compute jobs generally want compute-only then io-only
1030 * and the IO should complete as quickly as possible;
1032 * (3) IO is batched up to the RPC size and is async until the
1033 * client max cache is hit
1034 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
1037 if (!PageDirty(vmpage)) {
1038 tallyop = LPROC_LL_DIRTY_MISSES;
1039 result = cl_page_cache_add(env, io, pg, CRT_WRITE);
1041 /* page was added into cache successfully. */
1042 set_page_dirty(vmpage);
1043 vvp_write_pending(cl2ccc(obj), cp);
1044 } else if (result == -EDQUOT) {
1045 pgoff_t last_index = i_size_read(inode) >> CFS_PAGE_SHIFT;
1046 bool need_clip = true;
1049 * Client ran out of disk space grant. Possible
1052 * (a) do a sync write, renewing grant;
1054 * (b) stop writing on this stripe, switch to the
1057 * (b) is a part of "parallel io" design that is the
1058 * ultimate goal. (a) is what "old" client did, and
1059 * what the new code continues to do for the time
1062 if (last_index > pg->cp_index) {
1065 } else if (last_index == pg->cp_index) {
1066 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
1071 cl_page_clip(env, pg, 0, to);
1072 result = vvp_page_sync_io(env, io, pg, cp, CRT_WRITE);
1074 CERROR("Write page %lu of inode %p failed %d\n",
1075 pg->cp_index, inode, result);
1078 tallyop = LPROC_LL_DIRTY_HITS;
1081 ll_stats_ops_tally(sbi, tallyop, 1);
1083 size = cl_offset(obj, pg->cp_index) + to;
1085 ll_inode_size_lock(inode);
1087 if (size > i_size_read(inode)) {
1088 cl_isize_write_nolock(inode, size);
1089 CDEBUG(D_VFSTRACE, DFID" updating i_size %lu\n",
1090 PFID(lu_object_fid(&obj->co_lu)),
1091 (unsigned long)size);
1093 cl_page_export(env, pg, 1);
1095 if (size > i_size_read(inode))
1096 cl_page_discard(env, io, pg);
1098 ll_inode_size_unlock(inode);
1102 static const struct cl_io_operations vvp_io_ops = {
1105 .cio_fini = vvp_io_read_fini,
1106 .cio_lock = vvp_io_read_lock,
1107 .cio_start = vvp_io_read_start,
1108 .cio_advance = ccc_io_advance
1111 .cio_fini = vvp_io_fini,
1112 .cio_lock = vvp_io_write_lock,
1113 .cio_start = vvp_io_write_start,
1114 .cio_advance = ccc_io_advance
1117 .cio_fini = vvp_io_setattr_fini,
1118 .cio_iter_init = vvp_io_setattr_iter_init,
1119 .cio_lock = vvp_io_setattr_lock,
1120 .cio_start = vvp_io_setattr_start,
1121 .cio_end = vvp_io_setattr_end
1124 .cio_fini = vvp_io_fault_fini,
1125 .cio_iter_init = vvp_io_fault_iter_init,
1126 .cio_lock = vvp_io_fault_lock,
1127 .cio_start = vvp_io_fault_start,
1128 .cio_end = ccc_io_end
1131 .cio_start = vvp_io_fsync_start,
1132 .cio_fini = vvp_io_fini
1135 .cio_fini = vvp_io_fini
1138 .cio_read_page = vvp_io_read_page,
1139 .cio_prepare_write = vvp_io_prepare_write,
1140 .cio_commit_write = vvp_io_commit_write
1143 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1146 struct vvp_io *vio = vvp_env_io(env);
1147 struct ccc_io *cio = ccc_env_io(env);
1148 struct inode *inode = ccc_object_inode(obj);
1151 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
1154 CL_IO_SLICE_CLEAN(cio, cui_cl);
1155 cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops);
1156 vio->cui_ra_window_set = 0;
1158 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1160 struct ll_inode_info *lli = ll_i2info(inode);
1162 count = io->u.ci_rw.crw_count;
1163 /* "If nbyte is 0, read() will return 0 and have no other
1164 * results." -- Single Unix Spec */
1168 cio->cui_tot_count = count;
1169 cio->cui_tot_nrsegs = 0;
1171 /* for read/write, we store the jobid in the inode, and
1172 * it'll be fetched by osc when building RPC.
1174 * it's not accurate if the file is shared by different
1177 lustre_get_jobid(lli->lli_jobid);
1178 } else if (io->ci_type == CIT_SETATTR) {
1179 if (!cl_io_is_trunc(io))
1180 io->ci_lockreq = CILR_MANDATORY;
1183 /* Enqueue layout lock and get layout version. We need to do this
1184 * even for operations requiring to open file, such as read and write,
1185 * because it might not grant layout lock in IT_OPEN. */
1186 if (result == 0 && !io->ci_ignore_layout)
1187 result = ll_layout_refresh(inode, &cio->cui_layout_gen);
1192 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
1193 const struct cl_io_slice *slice)
1195 /* Caling just for assertion */
1196 cl2ccc_io(env, slice);
1197 return vvp_env_io(env);