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>
41 #define DEBUG_SUBSYSTEM S_LLITE
44 # error This file is kernel only.
48 #include <lustre_lite.h>
50 #include "vvp_internal.h"
52 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
53 const struct cl_io_slice *slice);
56 * True, if \a io is a normal io, False for sendfile() / splice_{read|write}
58 int cl_is_normalio(const struct lu_env *env, const struct cl_io *io)
60 struct vvp_io *vio = vvp_env_io(env);
62 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
64 return vio->cui_io_subtype == IO_NORMAL;
67 /*****************************************************************************
73 static int vvp_io_fault_iter_init(const struct lu_env *env,
74 const struct cl_io_slice *ios)
76 struct vvp_io *vio = cl2vvp_io(env, ios);
77 struct inode *inode = ccc_object_inode(ios->cis_obj);
80 cl2ccc_io(env, ios)->cui_fd->fd_file->f_dentry->d_inode);
81 vio->u.fault.ft_mtime = LTIME_S(inode->i_mtime);
85 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
87 struct cl_io *io = ios->cis_io;
88 struct cl_object *obj = io->ci_obj;
90 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
91 if (io->ci_type == CIT_READ) {
92 struct vvp_io *vio = cl2vvp_io(env, ios);
93 struct ccc_io *cio = cl2ccc_io(env, ios);
95 if (vio->cui_ra_window_set)
96 ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead);
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_smd != NULL) /* lsm-less file, don'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 UNLOCK_INODE_MUTEX(inode);
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, 0);
340 result = vmtruncate(inode, size);
341 ll_inode_size_unlock(inode, 0);
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 struct vvp_io *vio = cl2vvp_io(env, ios);
351 struct cl_io *io = ios->cis_io;
352 struct cl_object *obj = ios->cis_obj;
353 pgoff_t start = cl_index(obj, size);
356 DOWN_WRITE_I_ALLOC_SEM(inode);
358 result = vvp_do_vmtruncate(inode, size);
361 * If a page is partially truncated, keep it owned across truncate to
364 * XXX this properly belongs to osc, because races in question are OST
367 if (cl_offset(obj, start) != size) {
368 struct cl_object_header *hdr;
370 hdr = cl_object_header(obj);
371 cfs_spin_lock(&hdr->coh_page_guard);
372 vio->cui_partpage = cl_page_lookup(hdr, start);
373 cfs_spin_unlock(&hdr->coh_page_guard);
375 if (vio->cui_partpage != NULL)
377 * Wait for the transfer completion for a partially
378 * truncated page to avoid dead-locking an OST with
379 * the concurrent page-wise overlapping WRITE and
380 * PUNCH requests. BUG:17397.
382 * Partial page is disowned in vvp_io_trunc_end().
384 cl_page_own(env, io, vio->cui_partpage);
386 vio->cui_partpage = NULL;
390 static int vvp_io_setattr_time(const struct lu_env *env,
391 const struct cl_io_slice *ios)
393 struct cl_io *io = ios->cis_io;
394 struct cl_object *obj = io->ci_obj;
395 struct cl_attr *attr = ccc_env_thread_attr(env);
397 unsigned valid = CAT_CTIME;
399 cl_object_attr_lock(obj);
400 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
401 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
402 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
405 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
406 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
409 result = cl_object_attr_set(env, obj, attr, valid);
410 cl_object_attr_unlock(obj);
415 static int vvp_io_setattr_start(const struct lu_env *env,
416 const struct cl_io_slice *ios)
418 struct ccc_io *cio = cl2ccc_io(env, ios);
419 struct cl_io *io = ios->cis_io;
420 struct inode *inode = ccc_object_inode(io->ci_obj);
422 LASSERT(cio->u.setattr.cui_locks_released);
424 LOCK_INODE_MUTEX(inode);
425 cio->u.setattr.cui_locks_released = 0;
427 if (cl_io_is_trunc(io))
428 return vvp_io_setattr_trunc(env, ios, inode,
429 io->u.ci_setattr.sa_attr.lvb_size);
431 return vvp_io_setattr_time(env, ios);
434 static void vvp_io_setattr_end(const struct lu_env *env,
435 const struct cl_io_slice *ios)
437 struct vvp_io *vio = cl2vvp_io(env, ios);
438 struct cl_io *io = ios->cis_io;
439 struct inode *inode = ccc_object_inode(io->ci_obj);
441 if (!cl_io_is_trunc(io))
443 if (vio->cui_partpage != NULL) {
444 cl_page_disown(env, ios->cis_io, vio->cui_partpage);
445 cl_page_put(env, vio->cui_partpage);
446 vio->cui_partpage = NULL;
450 * Do vmtruncate again, to remove possible stale pages populated by
451 * competing read threads. bz20645.
453 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
456 static void vvp_io_setattr_fini(const struct lu_env *env,
457 const struct cl_io_slice *ios)
459 struct ccc_io *cio = ccc_env_io(env);
460 struct cl_io *io = ios->cis_io;
461 struct inode *inode = ccc_object_inode(ios->cis_io->ci_obj);
463 if (cio->u.setattr.cui_locks_released) {
464 LOCK_INODE_MUTEX(inode);
465 if (cl_io_is_trunc(io))
466 DOWN_WRITE_I_ALLOC_SEM(inode);
467 cio->u.setattr.cui_locks_released = 0;
469 vvp_io_fini(env, ios);
472 #ifdef HAVE_FILE_READV
473 static ssize_t lustre_generic_file_read(struct file *file,
474 struct ccc_io *vio, loff_t *ppos)
476 return generic_file_readv(file, vio->cui_iov, vio->cui_nrsegs, ppos);
479 static ssize_t lustre_generic_file_write(struct file *file,
480 struct ccc_io *vio, loff_t *ppos)
482 return generic_file_writev(file, vio->cui_iov, vio->cui_nrsegs, ppos);
485 static ssize_t lustre_generic_file_read(struct file *file,
486 struct ccc_io *vio, loff_t *ppos)
488 return generic_file_aio_read(vio->cui_iocb, vio->cui_iov,
489 vio->cui_nrsegs, *ppos);
492 static ssize_t lustre_generic_file_write(struct file *file,
493 struct ccc_io *vio, loff_t *ppos)
495 return generic_file_aio_write(vio->cui_iocb, vio->cui_iov,
496 vio->cui_nrsegs, *ppos);
500 static int vvp_io_read_start(const struct lu_env *env,
501 const struct cl_io_slice *ios)
503 struct vvp_io *vio = cl2vvp_io(env, ios);
504 struct ccc_io *cio = cl2ccc_io(env, ios);
505 struct cl_io *io = ios->cis_io;
506 struct cl_object *obj = io->ci_obj;
507 struct inode *inode = ccc_object_inode(obj);
508 struct ll_ra_read *bead = &vio->cui_bead;
509 struct file *file = cio->cui_fd->fd_file;
512 loff_t pos = io->u.ci_rd.rd.crw_pos;
513 long cnt = io->u.ci_rd.rd.crw_count;
514 long tot = cio->cui_tot_count;
517 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
519 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
521 result = ccc_prep_size(env, obj, io, pos, tot, &exceed);
524 else if (exceed != 0)
527 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
528 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
529 inode->i_ino, cnt, pos, i_size_read(inode));
531 /* turn off the kernel's read-ahead */
532 cio->cui_fd->fd_file->f_ra.ra_pages = 0;
534 /* initialize read-ahead window once per syscall */
535 if (!vio->cui_ra_window_set) {
536 vio->cui_ra_window_set = 1;
537 bead->lrr_start = cl_index(obj, pos);
539 * XXX: explicit CFS_PAGE_SIZE
541 bead->lrr_count = cl_index(obj, tot + CFS_PAGE_SIZE - 1);
542 ll_ra_read_in(file, bead);
547 switch (vio->cui_io_subtype) {
549 result = lustre_generic_file_read(file, cio, &pos);
551 #ifdef HAVE_KERNEL_SENDFILE
553 result = generic_file_sendfile(file, &pos, cnt,
554 vio->u.sendfile.cui_actor,
555 vio->u.sendfile.cui_target);
558 #ifdef HAVE_KERNEL_SPLICE_READ
560 result = generic_file_splice_read(file, &pos,
561 vio->u.splice.cui_pipe, cnt,
562 vio->u.splice.cui_flags);
563 /* LU-1109: do splice read stripe by stripe otherwise if it
564 * may make nfsd stuck if this read occupied all internal pipe
570 CERROR("Wrong IO type %u\n", vio->cui_io_subtype);
578 io->ci_nob += result;
579 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
580 cio->cui_fd, pos, result, 0);
586 static int vvp_io_write_start(const struct lu_env *env,
587 const struct cl_io_slice *ios)
589 struct ccc_io *cio = cl2ccc_io(env, ios);
590 struct cl_io *io = ios->cis_io;
591 struct cl_object *obj = io->ci_obj;
592 struct inode *inode = ccc_object_inode(obj);
593 struct file *file = cio->cui_fd->fd_file;
595 loff_t pos = io->u.ci_wr.wr.crw_pos;
596 size_t cnt = io->u.ci_wr.wr.crw_count;
600 if (cl_io_is_append(io)) {
602 * PARALLEL IO This has to be changed for parallel IO doing
603 * out-of-order writes.
605 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
606 #ifndef HAVE_FILE_WRITEV
607 cio->cui_iocb->ki_pos = pos;
611 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
613 if (cio->cui_iov == NULL) /* from a temp io in ll_cl_init(). */
616 result = lustre_generic_file_write(file, cio, &pos);
621 io->ci_nob += result;
622 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
623 cio->cui_fd, pos, result, 0);
629 #ifndef HAVE_VM_OP_FAULT
630 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
634 vmpage = filemap_nopage(cfio->ft_vma, cfio->nopage.ft_address,
635 cfio->nopage.ft_type);
637 if (vmpage == NOPAGE_SIGBUS) {
638 CDEBUG(D_PAGE, "got addr %lu type %lx - SIGBUS\n",
639 cfio->nopage.ft_address,(long)cfio->nopage.ft_type);
641 } else if (vmpage == NOPAGE_OOM) {
642 CDEBUG(D_PAGE, "got addr %lu type %lx - OOM\n",
643 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
647 LL_CDEBUG_PAGE(D_PAGE, vmpage, "got addr %lu type %lx\n",
648 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
650 cfio->ft_vmpage = vmpage;
656 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
658 struct vm_fault *vmf = cfio->fault.ft_vmf;
660 cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, vmf);
663 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
664 vmf->virtual_address);
665 if (unlikely(!(cfio->fault.ft_flags & VM_FAULT_LOCKED))) {
666 lock_page(vmf->page);
667 cfio->fault.ft_flags &= VM_FAULT_LOCKED;
670 cfio->ft_vmpage = vmf->page;
674 if (cfio->fault.ft_flags & VM_FAULT_SIGBUS) {
675 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
679 if (cfio->fault.ft_flags & VM_FAULT_OOM) {
680 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
684 if (cfio->fault.ft_flags & VM_FAULT_RETRY)
687 CERROR("unknow error in page fault %d!\n", cfio->fault.ft_flags);
693 static int vvp_io_fault_start(const struct lu_env *env,
694 const struct cl_io_slice *ios)
696 struct vvp_io *vio = cl2vvp_io(env, ios);
697 struct cl_io *io = ios->cis_io;
698 struct cl_object *obj = io->ci_obj;
699 struct inode *inode = ccc_object_inode(obj);
700 struct cl_fault_io *fio = &io->u.ci_fault;
701 struct vvp_fault_io *cfio = &vio->u.fault;
704 cfs_page_t *vmpage = NULL;
705 struct cl_page *page;
707 pgoff_t last; /* last page in a file data region */
709 if (fio->ft_executable &&
710 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
712 " changed while waiting for the page fault lock\n",
713 PFID(lu_object_fid(&obj->co_lu)));
715 /* offset of the last byte on the page */
716 offset = cl_offset(obj, fio->ft_index + 1) - 1;
717 LASSERT(cl_index(obj, offset) == fio->ft_index);
718 result = ccc_prep_size(env, obj, io, 0, offset + 1, NULL);
722 /* must return locked page */
723 if (fio->ft_mkwrite) {
724 LASSERT(cfio->ft_vmpage != NULL);
725 lock_page(cfio->ft_vmpage);
727 result = vvp_io_kernel_fault(cfio);
732 vmpage = cfio->ft_vmpage;
733 LASSERT(PageLocked(vmpage));
735 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
736 ll_invalidate_page(vmpage);
738 /* Though we have already held a cl_lock upon this page, but
739 * it still can be truncated locally. */
740 if (unlikely(vmpage->mapping == NULL)) {
741 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
743 /* return +1 to stop cl_io_loop() and ll_fault() will catch
745 GOTO(out, result = +1);
748 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
750 GOTO(out, result = PTR_ERR(page));
752 /* if page is going to be written, we should add this page into cache
754 if (fio->ft_mkwrite) {
755 wait_on_page_writeback(vmpage);
756 if (set_page_dirty(vmpage)) {
759 /* vvp_page_assume() calls wait_on_page_writeback(). */
760 cl_page_assume(env, io, page);
762 cp = cl2ccc_page(cl_page_at(page, &vvp_device_type));
763 vvp_write_pending(cl2ccc(obj), cp);
765 /* Do not set Dirty bit here so that in case IO is
766 * started before the page is really made dirty, we
767 * still have chance to detect it. */
768 result = cl_page_cache_add(env, io, page, CRT_WRITE);
770 cl_page_unassume(env, io, page);
771 cl_page_put(env, page);
773 /* we're in big trouble, what can we do now? */
774 if (result == -EDQUOT)
781 size = i_size_read(inode);
782 last = cl_index(obj, size - 1);
783 LASSERT(fio->ft_index <= last);
784 if (fio->ft_index == last)
786 * Last page is mapped partially.
788 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
790 fio->ft_nob = cl_page_size(obj);
792 lu_ref_add(&page->cp_reference, "fault", io);
797 /* return unlocked vmpage to avoid deadlocking */
799 #ifdef HAVE_VM_OP_FAULT
800 cfio->fault.ft_flags &= ~VM_FAULT_LOCKED;
805 static void vvp_io_fsync_end(const struct lu_env *env,
806 const struct cl_io_slice *ios)
808 /* never try to verify there is no dirty pages in sync range
809 * because page_mkwrite() can generate new dirty pages any time. */
812 static int vvp_io_read_page(const struct lu_env *env,
813 const struct cl_io_slice *ios,
814 const struct cl_page_slice *slice)
816 struct cl_io *io = ios->cis_io;
817 struct cl_object *obj = slice->cpl_obj;
818 struct ccc_page *cp = cl2ccc_page(slice);
819 struct cl_page *page = slice->cpl_page;
820 struct inode *inode = ccc_object_inode(obj);
821 struct ll_sb_info *sbi = ll_i2sbi(inode);
822 struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd;
823 struct ll_readahead_state *ras = &fd->fd_ras;
824 cfs_page_t *vmpage = cp->cpg_page;
825 struct cl_2queue *queue = &io->ci_queue;
828 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
829 LASSERT(slice->cpl_obj == obj);
833 if (sbi->ll_ra_info.ra_max_pages_per_file &&
834 sbi->ll_ra_info.ra_max_pages)
835 ras_update(sbi, inode, ras, page->cp_index,
836 cp->cpg_defer_uptodate);
838 /* Sanity check whether the page is protected by a lock. */
839 rc = cl_page_is_under_lock(env, io, page);
841 CL_PAGE_HEADER(D_WARNING, env, page, "%s: %d\n",
842 rc == -ENODATA ? "without a lock" :
848 if (cp->cpg_defer_uptodate) {
850 cl_page_export(env, page, 1);
853 * Add page into the queue even when it is marked uptodate above.
854 * this will unlock it automatically as part of cl_page_list_disown().
856 cl_2queue_add(queue, page);
857 if (sbi->ll_ra_info.ra_max_pages_per_file &&
858 sbi->ll_ra_info.ra_max_pages)
859 ll_readahead(env, io, ras,
860 vmpage->mapping, &queue->c2_qin, fd->fd_flags);
865 static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io,
866 struct cl_page *page, struct ccc_page *cp,
867 enum cl_req_type crt)
869 struct cl_2queue *queue;
872 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
874 queue = &io->ci_queue;
875 cl_2queue_init_page(queue, page);
877 result = cl_io_submit_sync(env, io, crt, queue, CRP_NORMAL, 0);
878 LASSERT(cl_page_is_owned(page, io));
882 * in CRT_WRITE case page is left locked even in case of
885 cl_page_list_disown(env, io, &queue->c2_qin);
886 cl_2queue_fini(env, queue);
892 * Prepare partially written-to page for a write.
894 static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io,
895 struct cl_object *obj, struct cl_page *pg,
897 unsigned from, unsigned to)
899 struct cl_attr *attr = ccc_env_thread_attr(env);
900 loff_t offset = cl_offset(obj, pg->cp_index);
903 cl_object_attr_lock(obj);
904 result = cl_object_attr_get(env, obj, attr);
905 cl_object_attr_unlock(obj);
908 * If are writing to a new page, no need to read old data.
909 * The extent locking will have updated the KMS, and for our
910 * purposes here we can treat it like i_size.
912 if (attr->cat_kms <= offset) {
913 char *kaddr = kmap_atomic(cp->cpg_page, KM_USER0);
915 memset(kaddr, 0, cl_page_size(obj));
916 kunmap_atomic(kaddr, KM_USER0);
917 } else if (cp->cpg_defer_uptodate)
920 result = vvp_page_sync_io(env, io, pg, cp, CRT_READ);
922 * In older implementations, obdo_refresh_inode is called here
923 * to update the inode because the write might modify the
924 * object info at OST. However, this has been proven useless,
925 * since LVB functions will be called when user space program
926 * tries to retrieve inode attribute. Also, see bug 15909 for
930 cl_page_export(env, pg, 1);
935 static int vvp_io_prepare_write(const struct lu_env *env,
936 const struct cl_io_slice *ios,
937 const struct cl_page_slice *slice,
938 unsigned from, unsigned to)
940 struct cl_object *obj = slice->cpl_obj;
941 struct ccc_page *cp = cl2ccc_page(slice);
942 struct cl_page *pg = slice->cpl_page;
943 cfs_page_t *vmpage = cp->cpg_page;
949 LINVRNT(cl_page_is_vmlocked(env, pg));
950 LASSERT(vmpage->mapping->host == ccc_object_inode(obj));
954 CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to);
955 if (!PageUptodate(vmpage)) {
957 * We're completely overwriting an existing page, so _don't_
958 * set it up to date until commit_write
960 if (from == 0 && to == CFS_PAGE_SIZE) {
961 CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n");
962 POISON_PAGE(page, 0x11);
964 result = vvp_io_prepare_partial(env, ios->cis_io, obj,
967 CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n");
971 static int vvp_io_commit_write(const struct lu_env *env,
972 const struct cl_io_slice *ios,
973 const struct cl_page_slice *slice,
974 unsigned from, unsigned to)
976 struct cl_object *obj = slice->cpl_obj;
977 struct cl_io *io = ios->cis_io;
978 struct ccc_page *cp = cl2ccc_page(slice);
979 struct cl_page *pg = slice->cpl_page;
980 struct inode *inode = ccc_object_inode(obj);
981 struct ll_sb_info *sbi = ll_i2sbi(inode);
982 cfs_page_t *vmpage = cp->cpg_page;
990 LINVRNT(cl_page_is_vmlocked(env, pg));
991 LASSERT(vmpage->mapping->host == inode);
993 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "commiting page write\n");
994 CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to);
997 * queue a write for some time in the future the first time we
1000 * This is different from what other file systems do: they usually
1001 * just mark page (and some of its buffers) dirty and rely on
1002 * balance_dirty_pages() to start a write-back. Lustre wants write-back
1003 * to be started earlier for the following reasons:
1005 * (1) with a large number of clients we need to limit the amount
1006 * of cached data on the clients a lot;
1008 * (2) large compute jobs generally want compute-only then io-only
1009 * and the IO should complete as quickly as possible;
1011 * (3) IO is batched up to the RPC size and is async until the
1012 * client max cache is hit
1013 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
1016 if (!PageDirty(vmpage)) {
1017 tallyop = LPROC_LL_DIRTY_MISSES;
1018 vvp_write_pending(cl2ccc(obj), cp);
1019 set_page_dirty(vmpage);
1020 /* ll_set_page_dirty() does the same for now, but
1021 * it will not soon. */
1022 vvp_write_pending(cl2ccc(obj), cp);
1023 result = cl_page_cache_add(env, io, pg, CRT_WRITE);
1024 if (result == -EDQUOT) {
1025 pgoff_t last_index = i_size_read(inode) >> CFS_PAGE_SHIFT;
1026 bool need_clip = true;
1029 * Client ran out of disk space grant. Possible
1032 * (a) do a sync write, renewing grant;
1034 * (b) stop writing on this stripe, switch to the
1037 * (b) is a part of "parallel io" design that is the
1038 * ultimate goal. (a) is what "old" client did, and
1039 * what the new code continues to do for the time
1042 if (last_index > pg->cp_index) {
1045 } else if (last_index == pg->cp_index) {
1046 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
1051 cl_page_clip(env, pg, 0, to);
1052 result = vvp_page_sync_io(env, io, pg, cp, CRT_WRITE);
1054 CERROR("Write page %lu of inode %p failed %d\n",
1055 pg->cp_index, inode, result);
1058 tallyop = LPROC_LL_DIRTY_HITS;
1061 ll_stats_ops_tally(sbi, tallyop, 1);
1063 size = cl_offset(obj, pg->cp_index) + to;
1065 ll_inode_size_lock(inode, 0);
1067 if (size > i_size_read(inode)) {
1068 cl_isize_write_nolock(inode, size);
1069 CDEBUG(D_VFSTRACE, DFID" updating i_size %lu\n",
1070 PFID(lu_object_fid(&obj->co_lu)),
1071 (unsigned long)size);
1073 cl_page_export(env, pg, 1);
1075 if (size > i_size_read(inode))
1076 cl_page_discard(env, io, pg);
1078 ll_inode_size_unlock(inode, 0);
1082 static const struct cl_io_operations vvp_io_ops = {
1085 .cio_fini = vvp_io_fini,
1086 .cio_lock = vvp_io_read_lock,
1087 .cio_start = vvp_io_read_start,
1088 .cio_advance = ccc_io_advance
1091 .cio_fini = vvp_io_fini,
1092 .cio_lock = vvp_io_write_lock,
1093 .cio_start = vvp_io_write_start,
1094 .cio_advance = ccc_io_advance
1097 .cio_fini = vvp_io_setattr_fini,
1098 .cio_iter_init = vvp_io_setattr_iter_init,
1099 .cio_lock = vvp_io_setattr_lock,
1100 .cio_start = vvp_io_setattr_start,
1101 .cio_end = vvp_io_setattr_end
1104 .cio_fini = vvp_io_fault_fini,
1105 .cio_iter_init = vvp_io_fault_iter_init,
1106 .cio_lock = vvp_io_fault_lock,
1107 .cio_start = vvp_io_fault_start,
1108 .cio_end = ccc_io_end
1111 .cio_end = vvp_io_fsync_end,
1112 .cio_fini = vvp_io_fini
1115 .cio_fini = vvp_io_fini
1118 .cio_read_page = vvp_io_read_page,
1119 .cio_prepare_write = vvp_io_prepare_write,
1120 .cio_commit_write = vvp_io_commit_write
1123 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1126 struct vvp_io *vio = vvp_env_io(env);
1127 struct ccc_io *cio = ccc_env_io(env);
1130 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
1133 CL_IO_SLICE_CLEAN(cio, cui_cl);
1134 cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops);
1135 vio->cui_ra_window_set = 0;
1137 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1139 struct ll_inode_info *lli = ll_i2info(ccc_object_inode(obj));
1141 count = io->u.ci_rw.crw_count;
1142 /* "If nbyte is 0, read() will return 0 and have no other
1143 * results." -- Single Unix Spec */
1147 cio->cui_tot_count = count;
1148 cio->cui_tot_nrsegs = 0;
1150 /* for read/write, we store the jobid in the inode, and
1151 * it'll be fetched by osc when building RPC.
1153 * it's not accurate if the file is shared by different
1156 lustre_get_jobid(lli->lli_jobid);
1157 } else if (io->ci_type == CIT_SETATTR) {
1158 if (!cl_io_is_trunc(io))
1159 io->ci_lockreq = CILR_MANDATORY;
1164 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
1165 const struct cl_io_slice *slice)
1167 /* Caling just for assertion */
1168 cl2ccc_io(env, slice);
1169 return vvp_env_io(env);