1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
32 * Copyright (c) 2011 Whamcloud, Inc.
36 * This file is part of Lustre, http://www.lustre.org/
37 * Lustre is a trademark of Sun Microsystems, Inc.
39 * Implementation of cl_io for VVP layer.
41 * Author: Nikita Danilov <nikita.danilov@sun.com>
44 #define DEBUG_SUBSYSTEM S_LLITE
47 # error This file is kernel only.
51 #include <lustre_lite.h>
53 #include "vvp_internal.h"
55 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
56 const struct cl_io_slice *slice);
59 * True, if \a io is a normal io, False for sendfile() / splice_{read|write}
61 int cl_is_normalio(const struct lu_env *env, const struct cl_io *io)
63 struct vvp_io *vio = vvp_env_io(env);
65 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
67 return vio->cui_io_subtype == IO_NORMAL;
70 /*****************************************************************************
76 static int vvp_io_fault_iter_init(const struct lu_env *env,
77 const struct cl_io_slice *ios)
79 struct vvp_io *vio = cl2vvp_io(env, ios);
80 struct inode *inode = ccc_object_inode(ios->cis_obj);
83 cl2ccc_io(env, ios)->cui_fd->fd_file->f_dentry->d_inode);
84 vio->u.fault.ft_mtime = LTIME_S(inode->i_mtime);
88 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
90 struct cl_io *io = ios->cis_io;
91 struct cl_object *obj = io->ci_obj;
93 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
94 if (io->ci_type == CIT_READ) {
95 struct vvp_io *vio = cl2vvp_io(env, ios);
96 struct ccc_io *cio = cl2ccc_io(env, ios);
98 if (vio->cui_ra_window_set)
99 ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead);
104 static void vvp_io_fault_fini(const struct lu_env *env,
105 const struct cl_io_slice *ios)
107 struct cl_io *io = ios->cis_io;
108 struct cl_page *page = io->u.ci_fault.ft_page;
110 CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj));
113 lu_ref_del(&page->cp_reference, "fault", io);
114 cl_page_put(env, page);
115 io->u.ci_fault.ft_page = NULL;
117 vvp_io_fini(env, ios);
120 enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
123 * we only want to hold PW locks if the mmap() can generate
124 * writes back to the file and that only happens in shared
127 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
132 static int vvp_mmap_locks(const struct lu_env *env,
133 struct ccc_io *vio, struct cl_io *io)
135 struct ccc_thread_info *cti = ccc_env_info(env);
136 struct vm_area_struct *vma;
137 struct cl_lock_descr *descr = &cti->cti_descr;
138 ldlm_policy_data_t policy;
145 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
147 if (!cl_is_normalio(env, io))
150 for (seg = 0; seg < vio->cui_nrsegs; seg++) {
151 const struct iovec *iv = &vio->cui_iov[seg];
153 addr = (unsigned long)iv->iov_base;
158 count += addr & (~CFS_PAGE_MASK);
159 addr &= CFS_PAGE_MASK;
160 while((vma = our_vma(addr, count)) != NULL) {
161 struct inode *inode = vma->vm_file->f_dentry->d_inode;
162 int flags = CEF_MUST;
164 if (ll_file_nolock(vma->vm_file)) {
166 * For no lock case, a lockless lock will be
173 * XXX: Required lock mode can be weakened: CIT_WRITE
174 * io only ever reads user level buffer, and CIT_READ
177 policy_from_vma(&policy, vma, addr, count);
178 descr->cld_mode = vvp_mode_from_vma(vma);
179 descr->cld_obj = ll_i2info(inode)->lli_clob;
180 descr->cld_start = cl_index(descr->cld_obj,
181 policy.l_extent.start);
182 descr->cld_end = cl_index(descr->cld_obj,
183 policy.l_extent.end);
184 descr->cld_enq_flags = flags;
185 result = cl_io_lock_alloc_add(env, io, descr);
187 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
188 descr->cld_mode, descr->cld_start,
194 if (vma->vm_end - addr >= count)
197 count -= vma->vm_end - addr;
204 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
205 enum cl_lock_mode mode, loff_t start, loff_t end)
207 struct ccc_io *cio = ccc_env_io(env);
211 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
214 ccc_io_update_iov(env, cio, io);
216 if (io->u.ci_rw.crw_nonblock)
217 ast_flags |= CEF_NONBLOCK;
218 result = vvp_mmap_locks(env, cio, io);
220 result = ccc_io_one_lock(env, io, ast_flags, mode, start, end);
224 static int vvp_io_read_lock(const struct lu_env *env,
225 const struct cl_io_slice *ios)
227 struct cl_io *io = ios->cis_io;
228 struct ll_inode_info *lli = ll_i2info(ccc_object_inode(io->ci_obj));
232 /* XXX: Layer violation, we shouldn't see lsm at llite level. */
233 if (lli->lli_smd != NULL) /* lsm-less file, don't need to lock */
234 result = vvp_io_rw_lock(env, io, CLM_READ,
235 io->u.ci_rd.rd.crw_pos,
236 io->u.ci_rd.rd.crw_pos +
237 io->u.ci_rd.rd.crw_count - 1);
243 static int vvp_io_fault_lock(const struct lu_env *env,
244 const struct cl_io_slice *ios)
246 struct cl_io *io = ios->cis_io;
247 struct vvp_io *vio = cl2vvp_io(env, ios);
249 * XXX LDLM_FL_CBPENDING
251 return ccc_io_one_lock_index
252 (env, io, 0, vvp_mode_from_vma(vio->u.fault.ft_vma),
253 io->u.ci_fault.ft_index, io->u.ci_fault.ft_index);
256 static int vvp_io_write_lock(const struct lu_env *env,
257 const struct cl_io_slice *ios)
259 struct cl_io *io = ios->cis_io;
263 if (io->u.ci_wr.wr_append) {
265 end = OBD_OBJECT_EOF;
267 start = io->u.ci_wr.wr.crw_pos;
268 end = start + io->u.ci_wr.wr.crw_count - 1;
270 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
273 static int vvp_io_setattr_iter_init(const struct lu_env *env,
274 const struct cl_io_slice *ios)
276 struct ccc_io *cio = ccc_env_io(env);
277 struct inode *inode = ccc_object_inode(ios->cis_obj);
280 * We really need to get our PW lock before we change inode->i_size.
281 * If we don't we can race with other i_size updaters on our node,
282 * like ll_file_read. We can also race with i_size propogation to
283 * other nodes through dirtying and writeback of final cached pages.
284 * This last one is especially bad for racing o_append users on other
287 UNLOCK_INODE_MUTEX(inode);
288 if (cl_io_is_trunc(ios->cis_io))
289 UP_WRITE_I_ALLOC_SEM(inode);
290 cio->u.setattr.cui_locks_released = 1;
295 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
297 * Handles "lockless io" mode when extent locking is done by server.
299 static int vvp_io_setattr_lock(const struct lu_env *env,
300 const struct cl_io_slice *ios)
302 struct ccc_io *cio = ccc_env_io(env);
303 struct cl_io *io = ios->cis_io;
307 if (cl_io_is_trunc(io)) {
308 new_size = io->u.ci_setattr.sa_attr.lvb_size;
310 enqflags = CEF_DISCARD_DATA;
312 if ((io->u.ci_setattr.sa_attr.lvb_mtime >=
313 io->u.ci_setattr.sa_attr.lvb_ctime) ||
314 (io->u.ci_setattr.sa_attr.lvb_atime >=
315 io->u.ci_setattr.sa_attr.lvb_ctime))
319 cio->u.setattr.cui_local_lock = SETATTR_EXTENT_LOCK;
320 return ccc_io_one_lock(env, io, enqflags, CLM_WRITE,
321 new_size, OBD_OBJECT_EOF);
324 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
328 * Only ll_inode_size_lock is taken at this level. lov_stripe_lock()
329 * is grabbed by ll_truncate() only over call to obd_adjust_kms().
331 ll_inode_size_lock(inode, 0);
332 result = vmtruncate(inode, size);
333 ll_inode_size_unlock(inode, 0);
338 static int vvp_io_setattr_trunc(const struct lu_env *env,
339 const struct cl_io_slice *ios,
340 struct inode *inode, loff_t size)
342 struct vvp_io *vio = cl2vvp_io(env, ios);
343 struct cl_io *io = ios->cis_io;
344 struct cl_object *obj = ios->cis_obj;
345 pgoff_t start = cl_index(obj, size);
348 DOWN_WRITE_I_ALLOC_SEM(inode);
350 result = vvp_do_vmtruncate(inode, size);
353 * If a page is partially truncated, keep it owned across truncate to
356 * XXX this properly belongs to osc, because races in question are OST
359 if (cl_offset(obj, start) != size) {
360 struct cl_object_header *hdr;
362 hdr = cl_object_header(obj);
363 cfs_spin_lock(&hdr->coh_page_guard);
364 vio->cui_partpage = cl_page_lookup(hdr, start);
365 cfs_spin_unlock(&hdr->coh_page_guard);
367 if (vio->cui_partpage != NULL)
369 * Wait for the transfer completion for a partially
370 * truncated page to avoid dead-locking an OST with
371 * the concurrent page-wise overlapping WRITE and
372 * PUNCH requests. BUG:17397.
374 * Partial page is disowned in vvp_io_trunc_end().
376 cl_page_own(env, io, vio->cui_partpage);
378 vio->cui_partpage = NULL;
382 static int vvp_io_setattr_time(const struct lu_env *env,
383 const struct cl_io_slice *ios)
385 struct cl_io *io = ios->cis_io;
386 struct cl_object *obj = io->ci_obj;
387 struct cl_attr *attr = ccc_env_thread_attr(env);
389 unsigned valid = CAT_CTIME;
391 cl_object_attr_lock(obj);
392 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
393 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
394 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
397 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
398 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
401 result = cl_object_attr_set(env, obj, attr, valid);
402 cl_object_attr_unlock(obj);
407 static int vvp_io_setattr_start(const struct lu_env *env,
408 const struct cl_io_slice *ios)
410 struct ccc_io *cio = cl2ccc_io(env, ios);
411 struct cl_io *io = ios->cis_io;
412 struct inode *inode = ccc_object_inode(io->ci_obj);
414 LASSERT(cio->u.setattr.cui_locks_released);
416 LOCK_INODE_MUTEX(inode);
417 cio->u.setattr.cui_locks_released = 0;
419 if (cl_io_is_trunc(io))
420 return vvp_io_setattr_trunc(env, ios, inode,
421 io->u.ci_setattr.sa_attr.lvb_size);
423 return vvp_io_setattr_time(env, ios);
426 static void vvp_io_setattr_end(const struct lu_env *env,
427 const struct cl_io_slice *ios)
429 struct vvp_io *vio = cl2vvp_io(env, ios);
430 struct cl_io *io = ios->cis_io;
431 struct inode *inode = ccc_object_inode(io->ci_obj);
433 if (!cl_io_is_trunc(io))
435 if (vio->cui_partpage != NULL) {
436 cl_page_disown(env, ios->cis_io, vio->cui_partpage);
437 cl_page_put(env, vio->cui_partpage);
438 vio->cui_partpage = NULL;
442 * Do vmtruncate again, to remove possible stale pages populated by
443 * competing read threads. bz20645.
445 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
448 static void vvp_io_setattr_fini(const struct lu_env *env,
449 const struct cl_io_slice *ios)
451 struct ccc_io *cio = ccc_env_io(env);
452 struct cl_io *io = ios->cis_io;
453 struct inode *inode = ccc_object_inode(ios->cis_io->ci_obj);
455 if (cio->u.setattr.cui_locks_released) {
456 LOCK_INODE_MUTEX(inode);
457 if (cl_io_is_trunc(io))
458 DOWN_WRITE_I_ALLOC_SEM(inode);
459 cio->u.setattr.cui_locks_released = 0;
461 vvp_io_fini(env, ios);
464 #ifdef HAVE_FILE_READV
465 static ssize_t lustre_generic_file_read(struct file *file,
466 struct ccc_io *vio, loff_t *ppos)
468 return generic_file_readv(file, vio->cui_iov, vio->cui_nrsegs, ppos);
471 static ssize_t lustre_generic_file_write(struct file *file,
472 struct ccc_io *vio, loff_t *ppos)
474 return generic_file_writev(file, vio->cui_iov, vio->cui_nrsegs, ppos);
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);
492 static int vvp_io_read_start(const struct lu_env *env,
493 const struct cl_io_slice *ios)
495 struct vvp_io *vio = cl2vvp_io(env, ios);
496 struct ccc_io *cio = cl2ccc_io(env, ios);
497 struct cl_io *io = ios->cis_io;
498 struct cl_object *obj = io->ci_obj;
499 struct inode *inode = ccc_object_inode(obj);
500 struct ll_ra_read *bead = &vio->cui_bead;
501 struct file *file = cio->cui_fd->fd_file;
504 loff_t pos = io->u.ci_rd.rd.crw_pos;
505 long cnt = io->u.ci_rd.rd.crw_count;
506 long tot = cio->cui_tot_count;
509 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
511 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
513 result = ccc_prep_size(env, obj, io, pos, tot, &exceed);
516 else if (exceed != 0)
519 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
520 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
521 inode->i_ino, cnt, pos, i_size_read(inode));
523 /* turn off the kernel's read-ahead */
524 cio->cui_fd->fd_file->f_ra.ra_pages = 0;
526 /* initialize read-ahead window once per syscall */
527 if (!vio->cui_ra_window_set) {
528 vio->cui_ra_window_set = 1;
529 bead->lrr_start = cl_index(obj, pos);
531 * XXX: explicit CFS_PAGE_SIZE
533 bead->lrr_count = cl_index(obj, tot + CFS_PAGE_SIZE - 1);
534 ll_ra_read_in(file, bead);
539 switch (vio->cui_io_subtype) {
541 result = lustre_generic_file_read(file, cio, &pos);
543 #ifdef HAVE_KERNEL_SENDFILE
545 result = generic_file_sendfile(file, &pos, cnt,
546 vio->u.sendfile.cui_actor,
547 vio->u.sendfile.cui_target);
550 #ifdef HAVE_KERNEL_SPLICE_READ
552 result = generic_file_splice_read(file, &pos,
553 vio->u.splice.cui_pipe, cnt,
554 vio->u.splice.cui_flags);
558 CERROR("Wrong IO type %u\n", vio->cui_io_subtype);
566 io->ci_nob += result;
567 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
568 cio->cui_fd, pos, result, 0);
574 static int vvp_io_write_start(const struct lu_env *env,
575 const struct cl_io_slice *ios)
577 struct ccc_io *cio = cl2ccc_io(env, ios);
578 struct cl_io *io = ios->cis_io;
579 struct cl_object *obj = io->ci_obj;
580 struct inode *inode = ccc_object_inode(obj);
581 struct file *file = cio->cui_fd->fd_file;
583 loff_t pos = io->u.ci_wr.wr.crw_pos;
584 size_t cnt = io->u.ci_wr.wr.crw_count;
588 if (cl_io_is_append(io)) {
590 * PARALLEL IO This has to be changed for parallel IO doing
591 * out-of-order writes.
593 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
594 #ifndef HAVE_FILE_WRITEV
595 cio->cui_iocb->ki_pos = pos;
599 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
601 if (cio->cui_iov == NULL) /* from a temp io in ll_cl_init(). */
604 result = lustre_generic_file_write(file, cio, &pos);
609 io->ci_nob += result;
610 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
611 cio->cui_fd, pos, result, 0);
617 #ifndef HAVE_VM_OP_FAULT
618 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
622 vmpage = filemap_nopage(cfio->ft_vma, cfio->nopage.ft_address,
623 cfio->nopage.ft_type);
625 if (vmpage == NOPAGE_SIGBUS) {
626 CDEBUG(D_PAGE, "got addr %lu type %lx - SIGBUS\n",
627 cfio->nopage.ft_address,(long)cfio->nopage.ft_type);
629 } else if (vmpage == NOPAGE_OOM) {
630 CDEBUG(D_PAGE, "got addr %lu type %lx - OOM\n",
631 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
635 LL_CDEBUG_PAGE(D_PAGE, vmpage, "got addr %lu type %lx\n",
636 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
638 cfio->ft_vmpage = vmpage;
644 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
646 struct vm_fault *vmf = cfio->fault.ft_vmf;
648 cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, vmf);
651 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
652 vmf->virtual_address);
653 if (unlikely(!(cfio->fault.ft_flags & VM_FAULT_LOCKED))) {
654 lock_page(vmf->page);
655 cfio->fault.ft_flags &= VM_FAULT_LOCKED;
658 cfio->ft_vmpage = vmf->page;
662 if (cfio->fault.ft_flags & VM_FAULT_SIGBUS) {
663 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
667 if (cfio->fault.ft_flags & VM_FAULT_OOM) {
668 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
672 if (cfio->fault.ft_flags & VM_FAULT_RETRY)
675 CERROR("unknow error in page fault %d!\n", cfio->fault.ft_flags);
681 static int vvp_io_fault_start(const struct lu_env *env,
682 const struct cl_io_slice *ios)
684 struct vvp_io *vio = cl2vvp_io(env, ios);
685 struct cl_io *io = ios->cis_io;
686 struct cl_object *obj = io->ci_obj;
687 struct inode *inode = ccc_object_inode(obj);
688 struct cl_fault_io *fio = &io->u.ci_fault;
689 struct vvp_fault_io *cfio = &vio->u.fault;
692 cfs_page_t *vmpage = NULL;
693 struct cl_page *page;
695 pgoff_t last; /* last page in a file data region */
697 if (fio->ft_executable &&
698 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
700 " changed while waiting for the page fault lock\n",
701 PFID(lu_object_fid(&obj->co_lu)));
703 /* offset of the last byte on the page */
704 offset = cl_offset(obj, fio->ft_index + 1) - 1;
705 LASSERT(cl_index(obj, offset) == fio->ft_index);
706 result = ccc_prep_size(env, obj, io, 0, offset + 1, NULL);
710 /* must return locked page */
711 if (fio->ft_mkwrite) {
712 LASSERT(cfio->ft_vmpage != NULL);
713 lock_page(cfio->ft_vmpage);
715 result = vvp_io_kernel_fault(cfio);
720 vmpage = cfio->ft_vmpage;
721 LASSERT(PageLocked(vmpage));
723 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
724 ll_invalidate_page(vmpage);
726 /* Though we have already held a cl_lock upon this page, but
727 * it still can be truncated locally. */
728 if (unlikely(vmpage->mapping == NULL)) {
729 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
731 /* return +1 to stop cl_io_loop() and ll_fault() will catch
733 GOTO(out, result = +1);
736 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
738 GOTO(out, result = PTR_ERR(page));
740 /* if page is going to be written, we should add this page into cache
742 if (fio->ft_mkwrite) {
743 wait_on_page_writeback(vmpage);
744 if (set_page_dirty(vmpage)) {
747 /* vvp_page_assume() calls wait_on_page_writeback(). */
748 cl_page_assume(env, io, page);
750 cp = cl2ccc_page(cl_page_at(page, &vvp_device_type));
751 vvp_write_pending(cl2ccc(obj), cp);
753 /* Do not set Dirty bit here so that in case IO is
754 * started before the page is really made dirty, we
755 * still have chance to detect it. */
756 result = cl_page_cache_add(env, io, page, CRT_WRITE);
758 cl_page_unassume(env, io, page);
759 cl_page_put(env, page);
761 /* we're in big trouble, what can we do now? */
762 if (result == -EDQUOT)
769 size = i_size_read(inode);
770 last = cl_index(obj, size - 1);
771 LASSERT(fio->ft_index <= last);
772 if (fio->ft_index == last)
774 * Last page is mapped partially.
776 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
778 fio->ft_nob = cl_page_size(obj);
780 lu_ref_add(&page->cp_reference, "fault", io);
785 /* return unlocked vmpage to avoid deadlocking */
787 #ifdef HAVE_VM_OP_FAULT
788 cfio->fault.ft_flags &= ~VM_FAULT_LOCKED;
793 static int vvp_io_read_page(const struct lu_env *env,
794 const struct cl_io_slice *ios,
795 const struct cl_page_slice *slice)
797 struct cl_io *io = ios->cis_io;
798 struct cl_object *obj = slice->cpl_obj;
799 struct ccc_page *cp = cl2ccc_page(slice);
800 struct cl_page *page = slice->cpl_page;
801 struct inode *inode = ccc_object_inode(obj);
802 struct ll_sb_info *sbi = ll_i2sbi(inode);
803 struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd;
804 struct ll_readahead_state *ras = &fd->fd_ras;
805 cfs_page_t *vmpage = cp->cpg_page;
806 struct cl_2queue *queue = &io->ci_queue;
809 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
810 LASSERT(slice->cpl_obj == obj);
814 if (sbi->ll_ra_info.ra_max_pages_per_file &&
815 sbi->ll_ra_info.ra_max_pages)
816 ras_update(sbi, inode, ras, page->cp_index,
817 cp->cpg_defer_uptodate);
819 /* Sanity check whether the page is protected by a lock. */
820 rc = cl_page_is_under_lock(env, io, page);
822 CL_PAGE_HEADER(D_WARNING, env, page, "%s: %d\n",
823 rc == -ENODATA ? "without a lock" :
829 if (cp->cpg_defer_uptodate) {
831 cl_page_export(env, page, 1);
834 * Add page into the queue even when it is marked uptodate above.
835 * this will unlock it automatically as part of cl_page_list_disown().
837 cl_2queue_add(queue, page);
838 if (sbi->ll_ra_info.ra_max_pages_per_file &&
839 sbi->ll_ra_info.ra_max_pages)
840 ll_readahead(env, io, ras,
841 vmpage->mapping, &queue->c2_qin, fd->fd_flags);
846 static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io,
847 struct cl_page *page, struct ccc_page *cp,
848 enum cl_req_type crt)
850 struct cl_2queue *queue;
853 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
855 queue = &io->ci_queue;
856 cl_2queue_init_page(queue, page);
858 result = cl_io_submit_sync(env, io, crt, queue, CRP_NORMAL, 0);
859 LASSERT(cl_page_is_owned(page, io));
863 * in CRT_WRITE case page is left locked even in case of
866 cl_page_list_disown(env, io, &queue->c2_qin);
867 cl_2queue_fini(env, queue);
873 * Prepare partially written-to page for a write.
875 static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io,
876 struct cl_object *obj, struct cl_page *pg,
878 unsigned from, unsigned to)
880 struct cl_attr *attr = ccc_env_thread_attr(env);
881 loff_t offset = cl_offset(obj, pg->cp_index);
884 cl_object_attr_lock(obj);
885 result = cl_object_attr_get(env, obj, attr);
886 cl_object_attr_unlock(obj);
889 * If are writing to a new page, no need to read old data.
890 * The extent locking will have updated the KMS, and for our
891 * purposes here we can treat it like i_size.
893 if (attr->cat_kms <= offset) {
894 char *kaddr = kmap_atomic(cp->cpg_page, KM_USER0);
896 memset(kaddr, 0, cl_page_size(obj));
897 kunmap_atomic(kaddr, KM_USER0);
898 } else if (cp->cpg_defer_uptodate)
901 result = vvp_page_sync_io(env, io, pg, cp, CRT_READ);
903 * In older implementations, obdo_refresh_inode is called here
904 * to update the inode because the write might modify the
905 * object info at OST. However, this has been proven useless,
906 * since LVB functions will be called when user space program
907 * tries to retrieve inode attribute. Also, see bug 15909 for
911 cl_page_export(env, pg, 1);
916 static int vvp_io_prepare_write(const struct lu_env *env,
917 const struct cl_io_slice *ios,
918 const struct cl_page_slice *slice,
919 unsigned from, unsigned to)
921 struct cl_object *obj = slice->cpl_obj;
922 struct ccc_page *cp = cl2ccc_page(slice);
923 struct cl_page *pg = slice->cpl_page;
924 cfs_page_t *vmpage = cp->cpg_page;
930 LINVRNT(cl_page_is_vmlocked(env, pg));
931 LASSERT(vmpage->mapping->host == ccc_object_inode(obj));
935 CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to);
936 if (!PageUptodate(vmpage)) {
938 * We're completely overwriting an existing page, so _don't_
939 * set it up to date until commit_write
941 if (from == 0 && to == CFS_PAGE_SIZE) {
942 CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n");
943 POISON_PAGE(page, 0x11);
945 result = vvp_io_prepare_partial(env, ios->cis_io, obj,
948 CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n");
952 static int vvp_io_commit_write(const struct lu_env *env,
953 const struct cl_io_slice *ios,
954 const struct cl_page_slice *slice,
955 unsigned from, unsigned to)
957 struct cl_object *obj = slice->cpl_obj;
958 struct cl_io *io = ios->cis_io;
959 struct ccc_page *cp = cl2ccc_page(slice);
960 struct cl_page *pg = slice->cpl_page;
961 struct inode *inode = ccc_object_inode(obj);
962 struct ll_sb_info *sbi = ll_i2sbi(inode);
963 cfs_page_t *vmpage = cp->cpg_page;
971 LINVRNT(cl_page_is_vmlocked(env, pg));
972 LASSERT(vmpage->mapping->host == inode);
974 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "commiting page write\n");
975 CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to);
978 * queue a write for some time in the future the first time we
981 * This is different from what other file systems do: they usually
982 * just mark page (and some of its buffers) dirty and rely on
983 * balance_dirty_pages() to start a write-back. Lustre wants write-back
984 * to be started earlier for the following reasons:
986 * (1) with a large number of clients we need to limit the amount
987 * of cached data on the clients a lot;
989 * (2) large compute jobs generally want compute-only then io-only
990 * and the IO should complete as quickly as possible;
992 * (3) IO is batched up to the RPC size and is async until the
993 * client max cache is hit
994 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
997 if (!PageDirty(vmpage)) {
998 tallyop = LPROC_LL_DIRTY_MISSES;
999 vvp_write_pending(cl2ccc(obj), cp);
1000 set_page_dirty(vmpage);
1001 /* ll_set_page_dirty() does the same for now, but
1002 * it will not soon. */
1003 vvp_write_pending(cl2ccc(obj), cp);
1004 result = cl_page_cache_add(env, io, pg, CRT_WRITE);
1005 if (result == -EDQUOT) {
1006 pgoff_t last_index = i_size_read(inode) >> CFS_PAGE_SHIFT;
1007 bool need_clip = true;
1010 * Client ran out of disk space grant. Possible
1013 * (a) do a sync write, renewing grant;
1015 * (b) stop writing on this stripe, switch to the
1018 * (b) is a part of "parallel io" design that is the
1019 * ultimate goal. (a) is what "old" client did, and
1020 * what the new code continues to do for the time
1023 if (last_index > pg->cp_index) {
1026 } else if (last_index == pg->cp_index) {
1027 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
1032 cl_page_clip(env, pg, 0, to);
1033 result = vvp_page_sync_io(env, io, pg, cp, CRT_WRITE);
1035 CERROR("Write page %lu of inode %p failed %d\n",
1036 pg->cp_index, inode, result);
1039 tallyop = LPROC_LL_DIRTY_HITS;
1042 ll_stats_ops_tally(sbi, tallyop, 1);
1044 size = cl_offset(obj, pg->cp_index) + to;
1046 ll_inode_size_lock(inode, 0);
1048 if (size > i_size_read(inode)) {
1049 cl_isize_write_nolock(inode, size);
1050 CDEBUG(D_VFSTRACE, DFID" updating i_size %lu\n",
1051 PFID(lu_object_fid(&obj->co_lu)),
1052 (unsigned long)size);
1054 cl_page_export(env, pg, 1);
1056 if (size > i_size_read(inode))
1057 cl_page_discard(env, io, pg);
1059 ll_inode_size_unlock(inode, 0);
1063 static const struct cl_io_operations vvp_io_ops = {
1066 .cio_fini = vvp_io_fini,
1067 .cio_lock = vvp_io_read_lock,
1068 .cio_start = vvp_io_read_start,
1069 .cio_advance = ccc_io_advance
1072 .cio_fini = vvp_io_fini,
1073 .cio_lock = vvp_io_write_lock,
1074 .cio_start = vvp_io_write_start,
1075 .cio_advance = ccc_io_advance
1078 .cio_fini = vvp_io_setattr_fini,
1079 .cio_iter_init = vvp_io_setattr_iter_init,
1080 .cio_lock = vvp_io_setattr_lock,
1081 .cio_start = vvp_io_setattr_start,
1082 .cio_end = vvp_io_setattr_end
1085 .cio_fini = vvp_io_fault_fini,
1086 .cio_iter_init = vvp_io_fault_iter_init,
1087 .cio_lock = vvp_io_fault_lock,
1088 .cio_start = vvp_io_fault_start,
1089 .cio_end = ccc_io_end
1092 .cio_fini = vvp_io_fini
1095 .cio_read_page = vvp_io_read_page,
1096 .cio_prepare_write = vvp_io_prepare_write,
1097 .cio_commit_write = vvp_io_commit_write
1100 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1103 struct vvp_io *vio = vvp_env_io(env);
1104 struct ccc_io *cio = ccc_env_io(env);
1105 struct inode *inode = ccc_object_inode(obj);
1106 struct ll_sb_info *sbi = ll_i2sbi(inode);
1109 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
1112 CL_IO_SLICE_CLEAN(cio, cui_cl);
1113 cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops);
1114 vio->cui_ra_window_set = 0;
1116 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1119 count = io->u.ci_rw.crw_count;
1120 /* "If nbyte is 0, read() will return 0 and have no other
1121 * results." -- Single Unix Spec */
1125 cio->cui_tot_count = count;
1126 cio->cui_tot_nrsegs = 0;
1128 } else if (io->ci_type == CIT_SETATTR) {
1129 if (cl_io_is_trunc(io))
1130 /* lockless truncate? */
1131 ll_stats_ops_tally(sbi, LPROC_LL_TRUNC, 1);
1133 io->ci_lockreq = CILR_MANDATORY;
1138 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
1139 const struct cl_io_slice *slice)
1141 /* Caling just for assertion */
1142 cl2ccc_io(env, slice);
1143 return vvp_env_io(env);