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, 2012, Whamcloud, Inc.
35 * This file is part of Lustre, http://www.lustre.org/
36 * Lustre is a trademark of Sun Microsystems, Inc.
38 * Implementation of cl_io for VVP layer.
40 * Author: Nikita Danilov <nikita.danilov@sun.com>
43 #define DEBUG_SUBSYSTEM S_LLITE
46 # error This file is kernel only.
50 #include <lustre_lite.h>
52 #include "vvp_internal.h"
54 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
55 const struct cl_io_slice *slice);
58 * True, if \a io is a normal io, False for sendfile() / splice_{read|write}
60 int cl_is_normalio(const struct lu_env *env, const struct cl_io *io)
62 struct vvp_io *vio = vvp_env_io(env);
64 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
66 return vio->cui_io_subtype == IO_NORMAL;
69 /*****************************************************************************
75 static int vvp_io_fault_iter_init(const struct lu_env *env,
76 const struct cl_io_slice *ios)
78 struct vvp_io *vio = cl2vvp_io(env, ios);
79 struct inode *inode = ccc_object_inode(ios->cis_obj);
82 cl2ccc_io(env, ios)->cui_fd->fd_file->f_dentry->d_inode);
83 vio->u.fault.ft_mtime = LTIME_S(inode->i_mtime);
87 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
89 struct cl_io *io = ios->cis_io;
90 struct cl_object *obj = io->ci_obj;
92 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
93 if (io->ci_type == CIT_READ) {
94 struct vvp_io *vio = cl2vvp_io(env, ios);
95 struct ccc_io *cio = cl2ccc_io(env, ios);
97 if (vio->cui_ra_window_set)
98 ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead);
103 static void vvp_io_fault_fini(const struct lu_env *env,
104 const struct cl_io_slice *ios)
106 struct cl_io *io = ios->cis_io;
107 struct cl_page *page = io->u.ci_fault.ft_page;
109 CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj));
112 lu_ref_del(&page->cp_reference, "fault", io);
113 cl_page_put(env, page);
114 io->u.ci_fault.ft_page = NULL;
116 vvp_io_fini(env, ios);
119 enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
122 * we only want to hold PW locks if the mmap() can generate
123 * writes back to the file and that only happens in shared
126 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
131 static int vvp_mmap_locks(const struct lu_env *env,
132 struct ccc_io *vio, struct cl_io *io)
134 struct ccc_thread_info *cti = ccc_env_info(env);
135 struct vm_area_struct *vma;
136 struct cl_lock_descr *descr = &cti->cti_descr;
137 ldlm_policy_data_t policy;
144 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
146 if (!cl_is_normalio(env, io))
149 for (seg = 0; seg < vio->cui_nrsegs; seg++) {
150 const struct iovec *iv = &vio->cui_iov[seg];
152 addr = (unsigned long)iv->iov_base;
157 count += addr & (~CFS_PAGE_MASK);
158 addr &= CFS_PAGE_MASK;
159 while((vma = our_vma(addr, count)) != NULL) {
160 struct inode *inode = vma->vm_file->f_dentry->d_inode;
161 int flags = CEF_MUST;
163 if (ll_file_nolock(vma->vm_file)) {
165 * For no lock case, a lockless lock will be
172 * XXX: Required lock mode can be weakened: CIT_WRITE
173 * io only ever reads user level buffer, and CIT_READ
176 policy_from_vma(&policy, vma, addr, count);
177 descr->cld_mode = vvp_mode_from_vma(vma);
178 descr->cld_obj = ll_i2info(inode)->lli_clob;
179 descr->cld_start = cl_index(descr->cld_obj,
180 policy.l_extent.start);
181 descr->cld_end = cl_index(descr->cld_obj,
182 policy.l_extent.end);
183 descr->cld_enq_flags = flags;
184 result = cl_io_lock_alloc_add(env, io, descr);
186 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
187 descr->cld_mode, descr->cld_start,
193 if (vma->vm_end - addr >= count)
196 count -= vma->vm_end - addr;
203 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
204 enum cl_lock_mode mode, loff_t start, loff_t end)
206 struct ccc_io *cio = ccc_env_io(env);
210 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
213 ccc_io_update_iov(env, cio, io);
215 if (io->u.ci_rw.crw_nonblock)
216 ast_flags |= CEF_NONBLOCK;
217 result = vvp_mmap_locks(env, cio, io);
219 result = ccc_io_one_lock(env, io, ast_flags, mode, start, end);
223 static int vvp_io_read_lock(const struct lu_env *env,
224 const struct cl_io_slice *ios)
226 struct cl_io *io = ios->cis_io;
227 struct ll_inode_info *lli = ll_i2info(ccc_object_inode(io->ci_obj));
231 /* XXX: Layer violation, we shouldn't see lsm at llite level. */
232 if (lli->lli_smd != NULL) /* lsm-less file, don't need to lock */
233 result = vvp_io_rw_lock(env, io, CLM_READ,
234 io->u.ci_rd.rd.crw_pos,
235 io->u.ci_rd.rd.crw_pos +
236 io->u.ci_rd.rd.crw_count - 1);
242 static int vvp_io_fault_lock(const struct lu_env *env,
243 const struct cl_io_slice *ios)
245 struct cl_io *io = ios->cis_io;
246 struct vvp_io *vio = cl2vvp_io(env, ios);
248 * XXX LDLM_FL_CBPENDING
250 return ccc_io_one_lock_index
251 (env, io, 0, vvp_mode_from_vma(vio->u.fault.ft_vma),
252 io->u.ci_fault.ft_index, io->u.ci_fault.ft_index);
255 static int vvp_io_write_lock(const struct lu_env *env,
256 const struct cl_io_slice *ios)
258 struct cl_io *io = ios->cis_io;
262 if (io->u.ci_wr.wr_append) {
264 end = OBD_OBJECT_EOF;
266 start = io->u.ci_wr.wr.crw_pos;
267 end = start + io->u.ci_wr.wr.crw_count - 1;
269 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
272 static int vvp_io_setattr_iter_init(const struct lu_env *env,
273 const struct cl_io_slice *ios)
275 struct ccc_io *cio = ccc_env_io(env);
276 struct inode *inode = ccc_object_inode(ios->cis_obj);
279 * We really need to get our PW lock before we change inode->i_size.
280 * If we don't we can race with other i_size updaters on our node,
281 * like ll_file_read. We can also race with i_size propogation to
282 * other nodes through dirtying and writeback of final cached pages.
283 * This last one is especially bad for racing o_append users on other
286 UNLOCK_INODE_MUTEX(inode);
287 if (cl_io_is_trunc(ios->cis_io))
288 UP_WRITE_I_ALLOC_SEM(inode);
289 cio->u.setattr.cui_locks_released = 1;
294 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
296 * Handles "lockless io" mode when extent locking is done by server.
298 static int vvp_io_setattr_lock(const struct lu_env *env,
299 const struct cl_io_slice *ios)
301 struct ccc_io *cio = ccc_env_io(env);
302 struct cl_io *io = ios->cis_io;
306 if (cl_io_is_trunc(io)) {
307 new_size = io->u.ci_setattr.sa_attr.lvb_size;
309 enqflags = CEF_DISCARD_DATA;
311 if ((io->u.ci_setattr.sa_attr.lvb_mtime >=
312 io->u.ci_setattr.sa_attr.lvb_ctime) ||
313 (io->u.ci_setattr.sa_attr.lvb_atime >=
314 io->u.ci_setattr.sa_attr.lvb_ctime))
318 cio->u.setattr.cui_local_lock = SETATTR_EXTENT_LOCK;
319 return ccc_io_one_lock(env, io, enqflags, CLM_WRITE,
320 new_size, OBD_OBJECT_EOF);
323 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
327 * Only ll_inode_size_lock is taken at this level. lov_stripe_lock()
328 * is grabbed by ll_truncate() only over call to obd_adjust_kms().
330 ll_inode_size_lock(inode, 0);
331 result = vmtruncate(inode, size);
332 ll_inode_size_unlock(inode, 0);
337 static int vvp_io_setattr_trunc(const struct lu_env *env,
338 const struct cl_io_slice *ios,
339 struct inode *inode, loff_t size)
341 struct vvp_io *vio = cl2vvp_io(env, ios);
342 struct cl_io *io = ios->cis_io;
343 struct cl_object *obj = ios->cis_obj;
344 pgoff_t start = cl_index(obj, size);
347 DOWN_WRITE_I_ALLOC_SEM(inode);
349 result = vvp_do_vmtruncate(inode, size);
352 * If a page is partially truncated, keep it owned across truncate to
355 * XXX this properly belongs to osc, because races in question are OST
358 if (cl_offset(obj, start) != size) {
359 struct cl_object_header *hdr;
361 hdr = cl_object_header(obj);
362 cfs_spin_lock(&hdr->coh_page_guard);
363 vio->cui_partpage = cl_page_lookup(hdr, start);
364 cfs_spin_unlock(&hdr->coh_page_guard);
366 if (vio->cui_partpage != NULL)
368 * Wait for the transfer completion for a partially
369 * truncated page to avoid dead-locking an OST with
370 * the concurrent page-wise overlapping WRITE and
371 * PUNCH requests. BUG:17397.
373 * Partial page is disowned in vvp_io_trunc_end().
375 cl_page_own(env, io, vio->cui_partpage);
377 vio->cui_partpage = NULL;
381 static int vvp_io_setattr_time(const struct lu_env *env,
382 const struct cl_io_slice *ios)
384 struct cl_io *io = ios->cis_io;
385 struct cl_object *obj = io->ci_obj;
386 struct cl_attr *attr = ccc_env_thread_attr(env);
388 unsigned valid = CAT_CTIME;
390 cl_object_attr_lock(obj);
391 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
392 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
393 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
396 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
397 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
400 result = cl_object_attr_set(env, obj, attr, valid);
401 cl_object_attr_unlock(obj);
406 static int vvp_io_setattr_start(const struct lu_env *env,
407 const struct cl_io_slice *ios)
409 struct ccc_io *cio = cl2ccc_io(env, ios);
410 struct cl_io *io = ios->cis_io;
411 struct inode *inode = ccc_object_inode(io->ci_obj);
413 LASSERT(cio->u.setattr.cui_locks_released);
415 LOCK_INODE_MUTEX(inode);
416 cio->u.setattr.cui_locks_released = 0;
418 if (cl_io_is_trunc(io))
419 return vvp_io_setattr_trunc(env, ios, inode,
420 io->u.ci_setattr.sa_attr.lvb_size);
422 return vvp_io_setattr_time(env, ios);
425 static void vvp_io_setattr_end(const struct lu_env *env,
426 const struct cl_io_slice *ios)
428 struct vvp_io *vio = cl2vvp_io(env, ios);
429 struct cl_io *io = ios->cis_io;
430 struct inode *inode = ccc_object_inode(io->ci_obj);
432 if (!cl_io_is_trunc(io))
434 if (vio->cui_partpage != NULL) {
435 cl_page_disown(env, ios->cis_io, vio->cui_partpage);
436 cl_page_put(env, vio->cui_partpage);
437 vio->cui_partpage = NULL;
441 * Do vmtruncate again, to remove possible stale pages populated by
442 * competing read threads. bz20645.
444 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
447 static void vvp_io_setattr_fini(const struct lu_env *env,
448 const struct cl_io_slice *ios)
450 struct ccc_io *cio = ccc_env_io(env);
451 struct cl_io *io = ios->cis_io;
452 struct inode *inode = ccc_object_inode(ios->cis_io->ci_obj);
454 if (cio->u.setattr.cui_locks_released) {
455 LOCK_INODE_MUTEX(inode);
456 if (cl_io_is_trunc(io))
457 DOWN_WRITE_I_ALLOC_SEM(inode);
458 cio->u.setattr.cui_locks_released = 0;
460 vvp_io_fini(env, ios);
463 #ifdef HAVE_FILE_READV
464 static ssize_t lustre_generic_file_read(struct file *file,
465 struct ccc_io *vio, loff_t *ppos)
467 return generic_file_readv(file, vio->cui_iov, vio->cui_nrsegs, ppos);
470 static ssize_t lustre_generic_file_write(struct file *file,
471 struct ccc_io *vio, loff_t *ppos)
473 return generic_file_writev(file, vio->cui_iov, vio->cui_nrsegs, ppos);
476 static ssize_t lustre_generic_file_read(struct file *file,
477 struct ccc_io *vio, loff_t *ppos)
479 return generic_file_aio_read(vio->cui_iocb, vio->cui_iov,
480 vio->cui_nrsegs, *ppos);
483 static ssize_t lustre_generic_file_write(struct file *file,
484 struct ccc_io *vio, loff_t *ppos)
486 return generic_file_aio_write(vio->cui_iocb, vio->cui_iov,
487 vio->cui_nrsegs, *ppos);
491 static int vvp_io_read_start(const struct lu_env *env,
492 const struct cl_io_slice *ios)
494 struct vvp_io *vio = cl2vvp_io(env, ios);
495 struct ccc_io *cio = cl2ccc_io(env, ios);
496 struct cl_io *io = ios->cis_io;
497 struct cl_object *obj = io->ci_obj;
498 struct inode *inode = ccc_object_inode(obj);
499 struct ll_ra_read *bead = &vio->cui_bead;
500 struct file *file = cio->cui_fd->fd_file;
503 loff_t pos = io->u.ci_rd.rd.crw_pos;
504 long cnt = io->u.ci_rd.rd.crw_count;
505 long tot = cio->cui_tot_count;
508 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
510 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
512 result = ccc_prep_size(env, obj, io, pos, tot, &exceed);
515 else if (exceed != 0)
518 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
519 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
520 inode->i_ino, cnt, pos, i_size_read(inode));
522 /* turn off the kernel's read-ahead */
523 cio->cui_fd->fd_file->f_ra.ra_pages = 0;
525 /* initialize read-ahead window once per syscall */
526 if (!vio->cui_ra_window_set) {
527 vio->cui_ra_window_set = 1;
528 bead->lrr_start = cl_index(obj, pos);
530 * XXX: explicit CFS_PAGE_SIZE
532 bead->lrr_count = cl_index(obj, tot + CFS_PAGE_SIZE - 1);
533 ll_ra_read_in(file, bead);
538 switch (vio->cui_io_subtype) {
540 result = lustre_generic_file_read(file, cio, &pos);
542 #ifdef HAVE_KERNEL_SENDFILE
544 result = generic_file_sendfile(file, &pos, cnt,
545 vio->u.sendfile.cui_actor,
546 vio->u.sendfile.cui_target);
549 #ifdef HAVE_KERNEL_SPLICE_READ
551 result = generic_file_splice_read(file, &pos,
552 vio->u.splice.cui_pipe, cnt,
553 vio->u.splice.cui_flags);
554 /* LU-1109: do splice read stripe by stripe otherwise if it
555 * may make nfsd stuck if this read occupied all internal pipe
561 CERROR("Wrong IO type %u\n", vio->cui_io_subtype);
569 io->ci_nob += result;
570 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
571 cio->cui_fd, pos, result, 0);
577 static int vvp_io_write_start(const struct lu_env *env,
578 const struct cl_io_slice *ios)
580 struct ccc_io *cio = cl2ccc_io(env, ios);
581 struct cl_io *io = ios->cis_io;
582 struct cl_object *obj = io->ci_obj;
583 struct inode *inode = ccc_object_inode(obj);
584 struct file *file = cio->cui_fd->fd_file;
586 loff_t pos = io->u.ci_wr.wr.crw_pos;
587 size_t cnt = io->u.ci_wr.wr.crw_count;
591 if (cl_io_is_append(io)) {
593 * PARALLEL IO This has to be changed for parallel IO doing
594 * out-of-order writes.
596 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
597 #ifndef HAVE_FILE_WRITEV
598 cio->cui_iocb->ki_pos = pos;
602 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
604 if (cio->cui_iov == NULL) /* from a temp io in ll_cl_init(). */
607 result = lustre_generic_file_write(file, cio, &pos);
612 io->ci_nob += result;
613 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
614 cio->cui_fd, pos, result, 0);
620 #ifndef HAVE_VM_OP_FAULT
621 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
625 vmpage = filemap_nopage(cfio->ft_vma, cfio->nopage.ft_address,
626 cfio->nopage.ft_type);
628 if (vmpage == NOPAGE_SIGBUS) {
629 CDEBUG(D_PAGE, "got addr %lu type %lx - SIGBUS\n",
630 cfio->nopage.ft_address,(long)cfio->nopage.ft_type);
632 } else if (vmpage == NOPAGE_OOM) {
633 CDEBUG(D_PAGE, "got addr %lu type %lx - OOM\n",
634 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
638 LL_CDEBUG_PAGE(D_PAGE, vmpage, "got addr %lu type %lx\n",
639 cfio->nopage.ft_address, (long)cfio->nopage.ft_type);
641 cfio->ft_vmpage = vmpage;
647 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
649 struct vm_fault *vmf = cfio->fault.ft_vmf;
651 cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, vmf);
654 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
655 vmf->virtual_address);
656 if (unlikely(!(cfio->fault.ft_flags & VM_FAULT_LOCKED))) {
657 lock_page(vmf->page);
658 cfio->fault.ft_flags &= VM_FAULT_LOCKED;
661 cfio->ft_vmpage = vmf->page;
665 if (cfio->fault.ft_flags & VM_FAULT_SIGBUS) {
666 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
670 if (cfio->fault.ft_flags & VM_FAULT_OOM) {
671 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
675 if (cfio->fault.ft_flags & VM_FAULT_RETRY)
678 CERROR("unknow error in page fault %d!\n", cfio->fault.ft_flags);
684 static int vvp_io_fault_start(const struct lu_env *env,
685 const struct cl_io_slice *ios)
687 struct vvp_io *vio = cl2vvp_io(env, ios);
688 struct cl_io *io = ios->cis_io;
689 struct cl_object *obj = io->ci_obj;
690 struct inode *inode = ccc_object_inode(obj);
691 struct cl_fault_io *fio = &io->u.ci_fault;
692 struct vvp_fault_io *cfio = &vio->u.fault;
695 cfs_page_t *vmpage = NULL;
696 struct cl_page *page;
698 pgoff_t last; /* last page in a file data region */
700 if (fio->ft_executable &&
701 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
703 " changed while waiting for the page fault lock\n",
704 PFID(lu_object_fid(&obj->co_lu)));
706 /* offset of the last byte on the page */
707 offset = cl_offset(obj, fio->ft_index + 1) - 1;
708 LASSERT(cl_index(obj, offset) == fio->ft_index);
709 result = ccc_prep_size(env, obj, io, 0, offset + 1, NULL);
713 /* must return locked page */
714 if (fio->ft_mkwrite) {
715 LASSERT(cfio->ft_vmpage != NULL);
716 lock_page(cfio->ft_vmpage);
718 result = vvp_io_kernel_fault(cfio);
723 vmpage = cfio->ft_vmpage;
724 LASSERT(PageLocked(vmpage));
726 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
727 ll_invalidate_page(vmpage);
729 /* Though we have already held a cl_lock upon this page, but
730 * it still can be truncated locally. */
731 if (unlikely(vmpage->mapping == NULL)) {
732 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
734 /* return +1 to stop cl_io_loop() and ll_fault() will catch
736 GOTO(out, result = +1);
739 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
741 GOTO(out, result = PTR_ERR(page));
743 /* if page is going to be written, we should add this page into cache
745 if (fio->ft_mkwrite) {
746 wait_on_page_writeback(vmpage);
747 if (set_page_dirty(vmpage)) {
750 /* vvp_page_assume() calls wait_on_page_writeback(). */
751 cl_page_assume(env, io, page);
753 cp = cl2ccc_page(cl_page_at(page, &vvp_device_type));
754 vvp_write_pending(cl2ccc(obj), cp);
756 /* Do not set Dirty bit here so that in case IO is
757 * started before the page is really made dirty, we
758 * still have chance to detect it. */
759 result = cl_page_cache_add(env, io, page, CRT_WRITE);
761 cl_page_unassume(env, io, page);
762 cl_page_put(env, page);
764 /* we're in big trouble, what can we do now? */
765 if (result == -EDQUOT)
772 size = i_size_read(inode);
773 last = cl_index(obj, size - 1);
774 LASSERT(fio->ft_index <= last);
775 if (fio->ft_index == last)
777 * Last page is mapped partially.
779 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
781 fio->ft_nob = cl_page_size(obj);
783 lu_ref_add(&page->cp_reference, "fault", io);
788 /* return unlocked vmpage to avoid deadlocking */
790 #ifdef HAVE_VM_OP_FAULT
791 cfio->fault.ft_flags &= ~VM_FAULT_LOCKED;
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, CRP_NORMAL, 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 vvp_write_pending(cl2ccc(obj), cp);
1003 set_page_dirty(vmpage);
1004 /* ll_set_page_dirty() does the same for now, but
1005 * it will not soon. */
1006 vvp_write_pending(cl2ccc(obj), cp);
1007 result = cl_page_cache_add(env, io, pg, CRT_WRITE);
1008 if (result == -EDQUOT) {
1009 pgoff_t last_index = i_size_read(inode) >> CFS_PAGE_SHIFT;
1010 bool need_clip = true;
1013 * Client ran out of disk space grant. Possible
1016 * (a) do a sync write, renewing grant;
1018 * (b) stop writing on this stripe, switch to the
1021 * (b) is a part of "parallel io" design that is the
1022 * ultimate goal. (a) is what "old" client did, and
1023 * what the new code continues to do for the time
1026 if (last_index > pg->cp_index) {
1029 } else if (last_index == pg->cp_index) {
1030 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
1035 cl_page_clip(env, pg, 0, to);
1036 result = vvp_page_sync_io(env, io, pg, cp, CRT_WRITE);
1038 CERROR("Write page %lu of inode %p failed %d\n",
1039 pg->cp_index, inode, result);
1042 tallyop = LPROC_LL_DIRTY_HITS;
1045 ll_stats_ops_tally(sbi, tallyop, 1);
1047 size = cl_offset(obj, pg->cp_index) + to;
1049 ll_inode_size_lock(inode, 0);
1051 if (size > i_size_read(inode)) {
1052 cl_isize_write_nolock(inode, size);
1053 CDEBUG(D_VFSTRACE, DFID" updating i_size %lu\n",
1054 PFID(lu_object_fid(&obj->co_lu)),
1055 (unsigned long)size);
1057 cl_page_export(env, pg, 1);
1059 if (size > i_size_read(inode))
1060 cl_page_discard(env, io, pg);
1062 ll_inode_size_unlock(inode, 0);
1066 static const struct cl_io_operations vvp_io_ops = {
1069 .cio_fini = vvp_io_fini,
1070 .cio_lock = vvp_io_read_lock,
1071 .cio_start = vvp_io_read_start,
1072 .cio_advance = ccc_io_advance
1075 .cio_fini = vvp_io_fini,
1076 .cio_lock = vvp_io_write_lock,
1077 .cio_start = vvp_io_write_start,
1078 .cio_advance = ccc_io_advance
1081 .cio_fini = vvp_io_setattr_fini,
1082 .cio_iter_init = vvp_io_setattr_iter_init,
1083 .cio_lock = vvp_io_setattr_lock,
1084 .cio_start = vvp_io_setattr_start,
1085 .cio_end = vvp_io_setattr_end
1088 .cio_fini = vvp_io_fault_fini,
1089 .cio_iter_init = vvp_io_fault_iter_init,
1090 .cio_lock = vvp_io_fault_lock,
1091 .cio_start = vvp_io_fault_start,
1092 .cio_end = ccc_io_end
1095 .cio_fini = vvp_io_fini
1098 .cio_read_page = vvp_io_read_page,
1099 .cio_prepare_write = vvp_io_prepare_write,
1100 .cio_commit_write = vvp_io_commit_write
1103 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1106 struct vvp_io *vio = vvp_env_io(env);
1107 struct ccc_io *cio = ccc_env_io(env);
1108 struct inode *inode = ccc_object_inode(obj);
1109 struct ll_sb_info *sbi = ll_i2sbi(inode);
1112 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
1115 CL_IO_SLICE_CLEAN(cio, cui_cl);
1116 cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops);
1117 vio->cui_ra_window_set = 0;
1119 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1122 count = io->u.ci_rw.crw_count;
1123 /* "If nbyte is 0, read() will return 0 and have no other
1124 * results." -- Single Unix Spec */
1128 cio->cui_tot_count = count;
1129 cio->cui_tot_nrsegs = 0;
1131 } else if (io->ci_type == CIT_SETATTR) {
1132 if (cl_io_is_trunc(io))
1133 /* lockless truncate? */
1134 ll_stats_ops_tally(sbi, LPROC_LL_TRUNC, 1);
1136 io->ci_lockreq = CILR_MANDATORY;
1141 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
1142 const struct cl_io_slice *slice)
1144 /* Caling just for assertion */
1145 cl2ccc_io(env, slice);
1146 return vvp_env_io(env);