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);
93 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
95 /* check layout version */
96 ll_layout_refresh(ccc_object_inode(obj), &gen);
97 if (cio->cui_layout_gen > 0)
98 io->ci_need_restart = cio->cui_layout_gen == gen;
101 static void vvp_io_fault_fini(const struct lu_env *env,
102 const struct cl_io_slice *ios)
104 struct cl_io *io = ios->cis_io;
105 struct cl_page *page = io->u.ci_fault.ft_page;
107 CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj));
110 lu_ref_del(&page->cp_reference, "fault", io);
111 cl_page_put(env, page);
112 io->u.ci_fault.ft_page = NULL;
114 vvp_io_fini(env, ios);
117 enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
120 * we only want to hold PW locks if the mmap() can generate
121 * writes back to the file and that only happens in shared
124 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
129 static int vvp_mmap_locks(const struct lu_env *env,
130 struct ccc_io *vio, struct cl_io *io)
132 struct ccc_thread_info *cti = ccc_env_info(env);
133 struct mm_struct *mm = current->mm;
134 struct vm_area_struct *vma;
135 struct cl_lock_descr *descr = &cti->cti_descr;
136 ldlm_policy_data_t policy;
143 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
145 if (!cl_is_normalio(env, io))
148 if (vio->cui_iov == NULL) /* nfs or loop back device write */
151 /* No MM (e.g. NFS)? No vmas too. */
155 for (seg = 0; seg < vio->cui_nrsegs; seg++) {
156 const struct iovec *iv = &vio->cui_iov[seg];
158 addr = (unsigned long)iv->iov_base;
163 count += addr & (~CFS_PAGE_MASK);
164 addr &= CFS_PAGE_MASK;
166 down_read(&mm->mmap_sem);
167 while((vma = our_vma(mm, addr, count)) != NULL) {
168 struct inode *inode = vma->vm_file->f_dentry->d_inode;
169 int flags = CEF_MUST;
171 if (ll_file_nolock(vma->vm_file)) {
173 * For no lock case, a lockless lock will be
180 * XXX: Required lock mode can be weakened: CIT_WRITE
181 * io only ever reads user level buffer, and CIT_READ
184 policy_from_vma(&policy, vma, addr, count);
185 descr->cld_mode = vvp_mode_from_vma(vma);
186 descr->cld_obj = ll_i2info(inode)->lli_clob;
187 descr->cld_start = cl_index(descr->cld_obj,
188 policy.l_extent.start);
189 descr->cld_end = cl_index(descr->cld_obj,
190 policy.l_extent.end);
191 descr->cld_enq_flags = flags;
192 result = cl_io_lock_alloc_add(env, io, descr);
194 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
195 descr->cld_mode, descr->cld_start,
201 if (vma->vm_end - addr >= count)
204 count -= vma->vm_end - addr;
207 up_read(&mm->mmap_sem);
212 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
213 enum cl_lock_mode mode, loff_t start, loff_t end)
215 struct ccc_io *cio = ccc_env_io(env);
219 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
222 ccc_io_update_iov(env, cio, io);
224 if (io->u.ci_rw.crw_nonblock)
225 ast_flags |= CEF_NONBLOCK;
226 result = vvp_mmap_locks(env, cio, io);
228 result = ccc_io_one_lock(env, io, ast_flags, mode, start, end);
232 static int vvp_io_read_lock(const struct lu_env *env,
233 const struct cl_io_slice *ios)
235 struct cl_io *io = ios->cis_io;
236 struct ll_inode_info *lli = ll_i2info(ccc_object_inode(io->ci_obj));
240 /* XXX: Layer violation, we shouldn't see lsm at llite level. */
241 if (lli->lli_has_smd) /* lsm-less file doesn't need to lock */
242 result = vvp_io_rw_lock(env, io, CLM_READ,
243 io->u.ci_rd.rd.crw_pos,
244 io->u.ci_rd.rd.crw_pos +
245 io->u.ci_rd.rd.crw_count - 1);
251 static int vvp_io_fault_lock(const struct lu_env *env,
252 const struct cl_io_slice *ios)
254 struct cl_io *io = ios->cis_io;
255 struct vvp_io *vio = cl2vvp_io(env, ios);
257 * XXX LDLM_FL_CBPENDING
259 return ccc_io_one_lock_index
260 (env, io, 0, vvp_mode_from_vma(vio->u.fault.ft_vma),
261 io->u.ci_fault.ft_index, io->u.ci_fault.ft_index);
264 static int vvp_io_write_lock(const struct lu_env *env,
265 const struct cl_io_slice *ios)
267 struct cl_io *io = ios->cis_io;
271 if (io->u.ci_wr.wr_append) {
273 end = OBD_OBJECT_EOF;
275 start = io->u.ci_wr.wr.crw_pos;
276 end = start + io->u.ci_wr.wr.crw_count - 1;
278 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
281 static int vvp_io_setattr_iter_init(const struct lu_env *env,
282 const struct cl_io_slice *ios)
284 struct ccc_io *cio = ccc_env_io(env);
285 struct inode *inode = ccc_object_inode(ios->cis_obj);
288 * We really need to get our PW lock before we change inode->i_size.
289 * If we don't we can race with other i_size updaters on our node,
290 * like ll_file_read. We can also race with i_size propogation to
291 * other nodes through dirtying and writeback of final cached pages.
292 * This last one is especially bad for racing o_append users on other
295 mutex_unlock(&inode->i_mutex);
296 if (cl_io_is_trunc(ios->cis_io))
297 UP_WRITE_I_ALLOC_SEM(inode);
298 cio->u.setattr.cui_locks_released = 1;
303 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
305 * Handles "lockless io" mode when extent locking is done by server.
307 static int vvp_io_setattr_lock(const struct lu_env *env,
308 const struct cl_io_slice *ios)
310 struct ccc_io *cio = ccc_env_io(env);
311 struct cl_io *io = ios->cis_io;
315 if (cl_io_is_trunc(io)) {
316 new_size = io->u.ci_setattr.sa_attr.lvb_size;
318 enqflags = CEF_DISCARD_DATA;
320 if ((io->u.ci_setattr.sa_attr.lvb_mtime >=
321 io->u.ci_setattr.sa_attr.lvb_ctime) ||
322 (io->u.ci_setattr.sa_attr.lvb_atime >=
323 io->u.ci_setattr.sa_attr.lvb_ctime))
327 cio->u.setattr.cui_local_lock = SETATTR_EXTENT_LOCK;
328 return ccc_io_one_lock(env, io, enqflags, CLM_WRITE,
329 new_size, OBD_OBJECT_EOF);
332 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
336 * Only ll_inode_size_lock is taken at this level. lov_stripe_lock()
337 * is grabbed by ll_truncate() only over call to obd_adjust_kms().
339 ll_inode_size_lock(inode);
340 result = vmtruncate(inode, size);
341 ll_inode_size_unlock(inode);
346 static int vvp_io_setattr_trunc(const struct lu_env *env,
347 const struct cl_io_slice *ios,
348 struct inode *inode, loff_t size)
350 DOWN_WRITE_I_ALLOC_SEM(inode);
354 static int vvp_io_setattr_time(const struct lu_env *env,
355 const struct cl_io_slice *ios)
357 struct cl_io *io = ios->cis_io;
358 struct cl_object *obj = io->ci_obj;
359 struct cl_attr *attr = ccc_env_thread_attr(env);
361 unsigned valid = CAT_CTIME;
363 cl_object_attr_lock(obj);
364 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
365 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
366 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
369 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
370 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
373 result = cl_object_attr_set(env, obj, attr, valid);
374 cl_object_attr_unlock(obj);
379 static int vvp_io_setattr_start(const struct lu_env *env,
380 const struct cl_io_slice *ios)
382 struct ccc_io *cio = cl2ccc_io(env, ios);
383 struct cl_io *io = ios->cis_io;
384 struct inode *inode = ccc_object_inode(io->ci_obj);
386 LASSERT(cio->u.setattr.cui_locks_released);
388 mutex_lock(&inode->i_mutex);
389 cio->u.setattr.cui_locks_released = 0;
391 if (cl_io_is_trunc(io))
392 return vvp_io_setattr_trunc(env, ios, inode,
393 io->u.ci_setattr.sa_attr.lvb_size);
395 return vvp_io_setattr_time(env, ios);
398 static void vvp_io_setattr_end(const struct lu_env *env,
399 const struct cl_io_slice *ios)
401 struct cl_io *io = ios->cis_io;
402 struct inode *inode = ccc_object_inode(io->ci_obj);
404 if (!cl_io_is_trunc(io))
407 /* Truncate in memory pages - they must be clean pages because osc
408 * has already notified to destroy osc_extents. */
409 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
412 static void vvp_io_setattr_fini(const struct lu_env *env,
413 const struct cl_io_slice *ios)
415 struct ccc_io *cio = ccc_env_io(env);
416 struct cl_io *io = ios->cis_io;
417 struct inode *inode = ccc_object_inode(ios->cis_io->ci_obj);
419 if (cio->u.setattr.cui_locks_released) {
420 mutex_lock(&inode->i_mutex);
421 if (cl_io_is_trunc(io))
422 DOWN_WRITE_I_ALLOC_SEM(inode);
423 cio->u.setattr.cui_locks_released = 0;
425 vvp_io_fini(env, ios);
428 #ifdef HAVE_FILE_READV
429 static ssize_t lustre_generic_file_read(struct file *file,
430 struct ccc_io *vio, loff_t *ppos)
432 return generic_file_readv(file, vio->cui_iov, vio->cui_nrsegs, ppos);
435 static ssize_t lustre_generic_file_write(struct file *file,
436 struct ccc_io *vio, loff_t *ppos)
438 return generic_file_writev(file, vio->cui_iov, vio->cui_nrsegs, ppos);
441 static ssize_t lustre_generic_file_read(struct file *file,
442 struct ccc_io *vio, loff_t *ppos)
444 return generic_file_aio_read(vio->cui_iocb, vio->cui_iov,
445 vio->cui_nrsegs, *ppos);
448 static ssize_t lustre_generic_file_write(struct file *file,
449 struct ccc_io *vio, loff_t *ppos)
451 return generic_file_aio_write(vio->cui_iocb, vio->cui_iov,
452 vio->cui_nrsegs, *ppos);
456 static int vvp_io_read_start(const struct lu_env *env,
457 const struct cl_io_slice *ios)
459 struct vvp_io *vio = cl2vvp_io(env, ios);
460 struct ccc_io *cio = cl2ccc_io(env, ios);
461 struct cl_io *io = ios->cis_io;
462 struct cl_object *obj = io->ci_obj;
463 struct inode *inode = ccc_object_inode(obj);
464 struct ll_ra_read *bead = &vio->cui_bead;
465 struct file *file = cio->cui_fd->fd_file;
468 loff_t pos = io->u.ci_rd.rd.crw_pos;
469 long cnt = io->u.ci_rd.rd.crw_count;
470 long tot = cio->cui_tot_count;
473 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
475 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
477 result = ccc_prep_size(env, obj, io, pos, tot, &exceed);
480 else if (exceed != 0)
483 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
484 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
485 inode->i_ino, cnt, pos, i_size_read(inode));
487 /* turn off the kernel's read-ahead */
488 cio->cui_fd->fd_file->f_ra.ra_pages = 0;
490 /* initialize read-ahead window once per syscall */
491 if (!vio->cui_ra_window_set) {
492 vio->cui_ra_window_set = 1;
493 bead->lrr_start = cl_index(obj, pos);
495 * XXX: explicit CFS_PAGE_SIZE
497 bead->lrr_count = cl_index(obj, tot + CFS_PAGE_SIZE - 1);
498 ll_ra_read_in(file, bead);
503 switch (vio->cui_io_subtype) {
505 result = lustre_generic_file_read(file, cio, &pos);
507 #ifdef HAVE_KERNEL_SENDFILE
509 result = generic_file_sendfile(file, &pos, cnt,
510 vio->u.sendfile.cui_actor,
511 vio->u.sendfile.cui_target);
514 #ifdef HAVE_KERNEL_SPLICE_READ
516 result = generic_file_splice_read(file, &pos,
517 vio->u.splice.cui_pipe, cnt,
518 vio->u.splice.cui_flags);
519 /* LU-1109: do splice read stripe by stripe otherwise if it
520 * may make nfsd stuck if this read occupied all internal pipe
526 CERROR("Wrong IO type %u\n", vio->cui_io_subtype);
534 io->ci_nob += result;
535 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
536 cio->cui_fd, pos, result, 0);
542 static void vvp_io_read_fini(const struct lu_env *env, const struct cl_io_slice *ios)
544 struct vvp_io *vio = cl2vvp_io(env, ios);
545 struct ccc_io *cio = cl2ccc_io(env, ios);
547 if (vio->cui_ra_window_set)
548 ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead);
550 vvp_io_fini(env, ios);
553 static int vvp_io_write_start(const struct lu_env *env,
554 const struct cl_io_slice *ios)
556 struct ccc_io *cio = cl2ccc_io(env, ios);
557 struct cl_io *io = ios->cis_io;
558 struct cl_object *obj = io->ci_obj;
559 struct inode *inode = ccc_object_inode(obj);
560 struct file *file = cio->cui_fd->fd_file;
562 loff_t pos = io->u.ci_wr.wr.crw_pos;
563 size_t cnt = io->u.ci_wr.wr.crw_count;
567 if (cl_io_is_append(io)) {
569 * PARALLEL IO This has to be changed for parallel IO doing
570 * out-of-order writes.
572 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
573 #ifndef HAVE_FILE_WRITEV
574 cio->cui_iocb->ki_pos = pos;
578 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
580 if (cio->cui_iov == NULL) /* from a temp io in ll_cl_init(). */
583 result = lustre_generic_file_write(file, cio, &pos);
588 io->ci_nob += result;
589 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
590 cio->cui_fd, pos, result, 0);
596 #ifndef HAVE_VM_OP_FAULT
597 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
601 vmpage = filemap_nopage(cfio->ft_vma, cfio->nopage.ft_address,
602 cfio->nopage.ft_type);
604 if (vmpage == NOPAGE_SIGBUS) {
605 CDEBUG(D_PAGE, "got addr %lu type %lx - SIGBUS\n",
606 cfio->nopage.ft_address,(long)cfio->nopage.ft_type);
608 } else if (vmpage == NOPAGE_OOM) {
609 CDEBUG(D_PAGE, "got addr %lu type %lx - OOM\n",
610 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
614 LL_CDEBUG_PAGE(D_PAGE, vmpage, "got addr %lu type %lx\n",
615 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
617 cfio->ft_vmpage = vmpage;
623 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
625 struct vm_fault *vmf = cfio->fault.ft_vmf;
627 cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, vmf);
630 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
631 vmf->virtual_address);
632 if (unlikely(!(cfio->fault.ft_flags & VM_FAULT_LOCKED))) {
633 lock_page(vmf->page);
634 cfio->fault.ft_flags &= VM_FAULT_LOCKED;
637 cfio->ft_vmpage = vmf->page;
641 if (cfio->fault.ft_flags & VM_FAULT_SIGBUS) {
642 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
646 if (cfio->fault.ft_flags & VM_FAULT_OOM) {
647 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
651 if (cfio->fault.ft_flags & VM_FAULT_RETRY)
654 CERROR("unknow error in page fault %d!\n", cfio->fault.ft_flags);
660 static int vvp_io_fault_start(const struct lu_env *env,
661 const struct cl_io_slice *ios)
663 struct vvp_io *vio = cl2vvp_io(env, ios);
664 struct cl_io *io = ios->cis_io;
665 struct cl_object *obj = io->ci_obj;
666 struct inode *inode = ccc_object_inode(obj);
667 struct cl_fault_io *fio = &io->u.ci_fault;
668 struct vvp_fault_io *cfio = &vio->u.fault;
671 cfs_page_t *vmpage = NULL;
672 struct cl_page *page;
674 pgoff_t last; /* last page in a file data region */
676 if (fio->ft_executable &&
677 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
679 " changed while waiting for the page fault lock\n",
680 PFID(lu_object_fid(&obj->co_lu)));
682 /* offset of the last byte on the page */
683 offset = cl_offset(obj, fio->ft_index + 1) - 1;
684 LASSERT(cl_index(obj, offset) == fio->ft_index);
685 result = ccc_prep_size(env, obj, io, 0, offset + 1, NULL);
689 /* must return locked page */
690 if (fio->ft_mkwrite) {
691 /* we grab alloc_sem to exclude truncate case.
692 * Otherwise, we could add dirty pages into osc cache
693 * while truncate is on-going. */
694 DOWN_READ_I_ALLOC_SEM(inode);
696 LASSERT(cfio->ft_vmpage != NULL);
697 lock_page(cfio->ft_vmpage);
699 result = vvp_io_kernel_fault(cfio);
704 vmpage = cfio->ft_vmpage;
705 LASSERT(PageLocked(vmpage));
707 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
708 ll_invalidate_page(vmpage);
710 /* Though we have already held a cl_lock upon this page, but
711 * it still can be truncated locally. */
712 if (unlikely(vmpage->mapping == NULL)) {
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);
720 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
722 GOTO(out, result = PTR_ERR(page));
724 /* if page is going to be written, we should add this page into cache
726 if (fio->ft_mkwrite) {
727 wait_on_page_writeback(vmpage);
728 if (set_page_dirty(vmpage)) {
731 /* vvp_page_assume() calls wait_on_page_writeback(). */
732 cl_page_assume(env, io, page);
734 cp = cl2ccc_page(cl_page_at(page, &vvp_device_type));
735 vvp_write_pending(cl2ccc(obj), cp);
737 /* Do not set Dirty bit here so that in case IO is
738 * started before the page is really made dirty, we
739 * still have chance to detect it. */
740 result = cl_page_cache_add(env, io, page, CRT_WRITE);
741 LASSERT(cl_page_is_owned(page, io));
745 cl_page_unmap(env, io, page);
746 cl_page_discard(env, io, page);
747 cl_page_disown(env, io, page);
749 cl_page_put(env, page);
751 /* we're in big trouble, what can we do now? */
752 if (result == -EDQUOT)
756 cl_page_disown(env, io, page);
760 size = i_size_read(inode);
761 last = cl_index(obj, size - 1);
762 LASSERT(fio->ft_index <= last);
763 if (fio->ft_index == last)
765 * Last page is mapped partially.
767 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
769 fio->ft_nob = cl_page_size(obj);
771 lu_ref_add(&page->cp_reference, "fault", io);
776 /* return unlocked vmpage to avoid deadlocking */
780 UP_READ_I_ALLOC_SEM(inode);
781 #ifdef HAVE_VM_OP_FAULT
782 cfio->fault.ft_flags &= ~VM_FAULT_LOCKED;
787 static int vvp_io_fsync_start(const struct lu_env *env,
788 const struct cl_io_slice *ios)
790 /* we should mark TOWRITE bit to each dirty page in radix tree to
791 * verify pages have been written, but this is difficult because of
796 static int vvp_io_read_page(const struct lu_env *env,
797 const struct cl_io_slice *ios,
798 const struct cl_page_slice *slice)
800 struct cl_io *io = ios->cis_io;
801 struct cl_object *obj = slice->cpl_obj;
802 struct ccc_page *cp = cl2ccc_page(slice);
803 struct cl_page *page = slice->cpl_page;
804 struct inode *inode = ccc_object_inode(obj);
805 struct ll_sb_info *sbi = ll_i2sbi(inode);
806 struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd;
807 struct ll_readahead_state *ras = &fd->fd_ras;
808 cfs_page_t *vmpage = cp->cpg_page;
809 struct cl_2queue *queue = &io->ci_queue;
812 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
813 LASSERT(slice->cpl_obj == obj);
817 if (sbi->ll_ra_info.ra_max_pages_per_file &&
818 sbi->ll_ra_info.ra_max_pages)
819 ras_update(sbi, inode, ras, page->cp_index,
820 cp->cpg_defer_uptodate);
822 /* Sanity check whether the page is protected by a lock. */
823 rc = cl_page_is_under_lock(env, io, page);
825 CL_PAGE_HEADER(D_WARNING, env, page, "%s: %d\n",
826 rc == -ENODATA ? "without a lock" :
832 if (cp->cpg_defer_uptodate) {
834 cl_page_export(env, page, 1);
837 * Add page into the queue even when it is marked uptodate above.
838 * this will unlock it automatically as part of cl_page_list_disown().
840 cl_2queue_add(queue, page);
841 if (sbi->ll_ra_info.ra_max_pages_per_file &&
842 sbi->ll_ra_info.ra_max_pages)
843 ll_readahead(env, io, ras,
844 vmpage->mapping, &queue->c2_qin, fd->fd_flags);
849 static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io,
850 struct cl_page *page, struct ccc_page *cp,
851 enum cl_req_type crt)
853 struct cl_2queue *queue;
856 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
858 queue = &io->ci_queue;
859 cl_2queue_init_page(queue, page);
861 result = cl_io_submit_sync(env, io, crt, queue, 0);
862 LASSERT(cl_page_is_owned(page, io));
866 * in CRT_WRITE case page is left locked even in case of
869 cl_page_list_disown(env, io, &queue->c2_qin);
870 cl_2queue_fini(env, queue);
876 * Prepare partially written-to page for a write.
878 static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io,
879 struct cl_object *obj, struct cl_page *pg,
881 unsigned from, unsigned to)
883 struct cl_attr *attr = ccc_env_thread_attr(env);
884 loff_t offset = cl_offset(obj, pg->cp_index);
887 cl_object_attr_lock(obj);
888 result = cl_object_attr_get(env, obj, attr);
889 cl_object_attr_unlock(obj);
892 * If are writing to a new page, no need to read old data.
893 * The extent locking will have updated the KMS, and for our
894 * purposes here we can treat it like i_size.
896 if (attr->cat_kms <= offset) {
897 char *kaddr = kmap_atomic(cp->cpg_page, KM_USER0);
899 memset(kaddr, 0, cl_page_size(obj));
900 kunmap_atomic(kaddr, KM_USER0);
901 } else if (cp->cpg_defer_uptodate)
904 result = vvp_page_sync_io(env, io, pg, cp, CRT_READ);
906 * In older implementations, obdo_refresh_inode is called here
907 * to update the inode because the write might modify the
908 * object info at OST. However, this has been proven useless,
909 * since LVB functions will be called when user space program
910 * tries to retrieve inode attribute. Also, see bug 15909 for
914 cl_page_export(env, pg, 1);
919 static int vvp_io_prepare_write(const struct lu_env *env,
920 const struct cl_io_slice *ios,
921 const struct cl_page_slice *slice,
922 unsigned from, unsigned to)
924 struct cl_object *obj = slice->cpl_obj;
925 struct ccc_page *cp = cl2ccc_page(slice);
926 struct cl_page *pg = slice->cpl_page;
927 cfs_page_t *vmpage = cp->cpg_page;
933 LINVRNT(cl_page_is_vmlocked(env, pg));
934 LASSERT(vmpage->mapping->host == ccc_object_inode(obj));
938 CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to);
939 if (!PageUptodate(vmpage)) {
941 * We're completely overwriting an existing page, so _don't_
942 * set it up to date until commit_write
944 if (from == 0 && to == CFS_PAGE_SIZE) {
945 CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n");
946 POISON_PAGE(page, 0x11);
948 result = vvp_io_prepare_partial(env, ios->cis_io, obj,
951 CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n");
955 static int vvp_io_commit_write(const struct lu_env *env,
956 const struct cl_io_slice *ios,
957 const struct cl_page_slice *slice,
958 unsigned from, unsigned to)
960 struct cl_object *obj = slice->cpl_obj;
961 struct cl_io *io = ios->cis_io;
962 struct ccc_page *cp = cl2ccc_page(slice);
963 struct cl_page *pg = slice->cpl_page;
964 struct inode *inode = ccc_object_inode(obj);
965 struct ll_sb_info *sbi = ll_i2sbi(inode);
966 cfs_page_t *vmpage = cp->cpg_page;
974 LINVRNT(cl_page_is_vmlocked(env, pg));
975 LASSERT(vmpage->mapping->host == inode);
977 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "commiting page write\n");
978 CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to);
981 * queue a write for some time in the future the first time we
984 * This is different from what other file systems do: they usually
985 * just mark page (and some of its buffers) dirty and rely on
986 * balance_dirty_pages() to start a write-back. Lustre wants write-back
987 * to be started earlier for the following reasons:
989 * (1) with a large number of clients we need to limit the amount
990 * of cached data on the clients a lot;
992 * (2) large compute jobs generally want compute-only then io-only
993 * and the IO should complete as quickly as possible;
995 * (3) IO is batched up to the RPC size and is async until the
996 * client max cache is hit
997 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
1000 if (!PageDirty(vmpage)) {
1001 tallyop = LPROC_LL_DIRTY_MISSES;
1002 result = cl_page_cache_add(env, io, pg, CRT_WRITE);
1004 /* page was added into cache successfully. */
1005 set_page_dirty(vmpage);
1006 vvp_write_pending(cl2ccc(obj), cp);
1007 } else if (result == -EDQUOT) {
1008 pgoff_t last_index = i_size_read(inode) >> CFS_PAGE_SHIFT;
1009 bool need_clip = true;
1012 * Client ran out of disk space grant. Possible
1015 * (a) do a sync write, renewing grant;
1017 * (b) stop writing on this stripe, switch to the
1020 * (b) is a part of "parallel io" design that is the
1021 * ultimate goal. (a) is what "old" client did, and
1022 * what the new code continues to do for the time
1025 if (last_index > pg->cp_index) {
1028 } else if (last_index == pg->cp_index) {
1029 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
1034 cl_page_clip(env, pg, 0, to);
1035 result = vvp_page_sync_io(env, io, pg, cp, CRT_WRITE);
1037 CERROR("Write page %lu of inode %p failed %d\n",
1038 pg->cp_index, inode, result);
1041 tallyop = LPROC_LL_DIRTY_HITS;
1044 ll_stats_ops_tally(sbi, tallyop, 1);
1046 size = cl_offset(obj, pg->cp_index) + to;
1048 ll_inode_size_lock(inode);
1050 if (size > i_size_read(inode)) {
1051 cl_isize_write_nolock(inode, size);
1052 CDEBUG(D_VFSTRACE, DFID" updating i_size %lu\n",
1053 PFID(lu_object_fid(&obj->co_lu)),
1054 (unsigned long)size);
1056 cl_page_export(env, pg, 1);
1058 if (size > i_size_read(inode))
1059 cl_page_discard(env, io, pg);
1061 ll_inode_size_unlock(inode);
1065 static const struct cl_io_operations vvp_io_ops = {
1068 .cio_fini = vvp_io_read_fini,
1069 .cio_lock = vvp_io_read_lock,
1070 .cio_start = vvp_io_read_start,
1071 .cio_advance = ccc_io_advance
1074 .cio_fini = vvp_io_fini,
1075 .cio_lock = vvp_io_write_lock,
1076 .cio_start = vvp_io_write_start,
1077 .cio_advance = ccc_io_advance
1080 .cio_fini = vvp_io_setattr_fini,
1081 .cio_iter_init = vvp_io_setattr_iter_init,
1082 .cio_lock = vvp_io_setattr_lock,
1083 .cio_start = vvp_io_setattr_start,
1084 .cio_end = vvp_io_setattr_end
1087 .cio_fini = vvp_io_fault_fini,
1088 .cio_iter_init = vvp_io_fault_iter_init,
1089 .cio_lock = vvp_io_fault_lock,
1090 .cio_start = vvp_io_fault_start,
1091 .cio_end = ccc_io_end
1094 .cio_start = vvp_io_fsync_start,
1095 .cio_fini = vvp_io_fini
1098 .cio_fini = vvp_io_fini
1101 .cio_read_page = vvp_io_read_page,
1102 .cio_prepare_write = vvp_io_prepare_write,
1103 .cio_commit_write = vvp_io_commit_write
1106 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1109 struct vvp_io *vio = vvp_env_io(env);
1110 struct ccc_io *cio = ccc_env_io(env);
1111 struct inode *inode = ccc_object_inode(obj);
1114 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
1117 CL_IO_SLICE_CLEAN(cio, cui_cl);
1118 cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops);
1119 vio->cui_ra_window_set = 0;
1121 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1123 struct ll_inode_info *lli = ll_i2info(inode);
1125 count = io->u.ci_rw.crw_count;
1126 /* "If nbyte is 0, read() will return 0 and have no other
1127 * results." -- Single Unix Spec */
1131 cio->cui_tot_count = count;
1132 cio->cui_tot_nrsegs = 0;
1134 /* for read/write, we store the jobid in the inode, and
1135 * it'll be fetched by osc when building RPC.
1137 * it's not accurate if the file is shared by different
1140 lustre_get_jobid(lli->lli_jobid);
1141 } else if (io->ci_type == CIT_SETATTR) {
1142 if (!cl_io_is_trunc(io))
1143 io->ci_lockreq = CILR_MANDATORY;
1146 /* Enqueue layout lock and get layout version. We need to do this
1147 * even for operations requiring to open file, such as read and write,
1148 * because it might not grant layout lock in IT_OPEN. */
1149 if (result == 0 && !io->ci_ignore_layout)
1150 result = ll_layout_refresh(inode, &cio->cui_layout_gen);
1155 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
1156 const struct cl_io_slice *slice)
1158 /* Caling just for assertion */
1159 cl2ccc_io(env, slice);
1160 return vvp_env_io(env);