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, Intel Corporation.
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
46 #include <lustre_lite.h>
48 #include "vvp_internal.h"
50 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
51 const struct cl_io_slice *slice);
54 * True, if \a io is a normal io, False for sendfile() / splice_{read|write}
56 int cl_is_normalio(const struct lu_env *env, const struct cl_io *io)
58 struct vvp_io *vio = vvp_env_io(env);
60 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
62 return vio->cui_io_subtype == IO_NORMAL;
66 * For swapping layout. The file's layout may have changed.
67 * To avoid populating pages to a wrong stripe, we have to verify the
68 * correctness of layout. It works because swapping layout processes
69 * have to acquire group lock.
71 static bool can_populate_pages(const struct lu_env *env, struct cl_io *io,
74 struct ll_inode_info *lli = ll_i2info(inode);
75 struct ccc_io *cio = ccc_env_io(env);
78 switch (io->ci_type) {
81 /* don't need lock here to check lli_layout_gen as we have held
82 * extent lock and GROUP lock has to hold to swap layout */
83 if (lli->lli_layout_gen != cio->cui_layout_gen) {
84 io->ci_need_restart = 1;
85 /* this will return application a short read/write */
90 /* fault is okay because we've already had a page. */
98 /*****************************************************************************
104 static int vvp_io_fault_iter_init(const struct lu_env *env,
105 const struct cl_io_slice *ios)
107 struct vvp_io *vio = cl2vvp_io(env, ios);
108 struct inode *inode = ccc_object_inode(ios->cis_obj);
111 cl2ccc_io(env, ios)->cui_fd->fd_file->f_dentry->d_inode);
112 vio->u.fault.ft_mtime = LTIME_S(inode->i_mtime);
116 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
118 struct cl_io *io = ios->cis_io;
119 struct cl_object *obj = io->ci_obj;
120 struct ccc_io *cio = cl2ccc_io(env, ios);
122 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
124 CDEBUG(D_VFSTRACE, "ignore/verify layout %d/%d, layout version %d.\n",
125 io->ci_ignore_layout, io->ci_verify_layout, cio->cui_layout_gen);
127 if (!io->ci_ignore_layout && io->ci_verify_layout) {
130 /* check layout version */
131 ll_layout_refresh(ccc_object_inode(obj), &gen);
132 io->ci_need_restart = cio->cui_layout_gen != gen;
133 if (io->ci_need_restart)
134 CDEBUG(D_VFSTRACE, "layout changed from %d to %d.\n",
135 cio->cui_layout_gen, gen);
139 static void vvp_io_fault_fini(const struct lu_env *env,
140 const struct cl_io_slice *ios)
142 struct cl_io *io = ios->cis_io;
143 struct cl_page *page = io->u.ci_fault.ft_page;
145 CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj));
148 lu_ref_del(&page->cp_reference, "fault", io);
149 cl_page_put(env, page);
150 io->u.ci_fault.ft_page = NULL;
152 vvp_io_fini(env, ios);
155 enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
158 * we only want to hold PW locks if the mmap() can generate
159 * writes back to the file and that only happens in shared
162 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
167 static int vvp_mmap_locks(const struct lu_env *env,
168 struct ccc_io *vio, struct cl_io *io)
170 struct ccc_thread_info *cti = ccc_env_info(env);
171 struct mm_struct *mm = current->mm;
172 struct vm_area_struct *vma;
173 struct cl_lock_descr *descr = &cti->cti_descr;
174 ldlm_policy_data_t policy;
181 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
183 if (!cl_is_normalio(env, io))
186 if (vio->cui_iov == NULL) /* nfs or loop back device write */
189 /* No MM (e.g. NFS)? No vmas too. */
193 for (seg = 0; seg < vio->cui_nrsegs; seg++) {
194 const struct iovec *iv = &vio->cui_iov[seg];
196 addr = (unsigned long)iv->iov_base;
201 count += addr & (~CFS_PAGE_MASK);
202 addr &= CFS_PAGE_MASK;
204 down_read(&mm->mmap_sem);
205 while((vma = our_vma(mm, addr, count)) != NULL) {
206 struct inode *inode = vma->vm_file->f_dentry->d_inode;
207 int flags = CEF_MUST;
209 if (ll_file_nolock(vma->vm_file)) {
211 * For no lock case, a lockless lock will be
218 * XXX: Required lock mode can be weakened: CIT_WRITE
219 * io only ever reads user level buffer, and CIT_READ
222 policy_from_vma(&policy, vma, addr, count);
223 descr->cld_mode = vvp_mode_from_vma(vma);
224 descr->cld_obj = ll_i2info(inode)->lli_clob;
225 descr->cld_start = cl_index(descr->cld_obj,
226 policy.l_extent.start);
227 descr->cld_end = cl_index(descr->cld_obj,
228 policy.l_extent.end);
229 descr->cld_enq_flags = flags;
230 result = cl_io_lock_alloc_add(env, io, descr);
232 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
233 descr->cld_mode, descr->cld_start,
239 if (vma->vm_end - addr >= count)
242 count -= vma->vm_end - addr;
245 up_read(&mm->mmap_sem);
250 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
251 enum cl_lock_mode mode, loff_t start, loff_t end)
253 struct ccc_io *cio = ccc_env_io(env);
257 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
260 ccc_io_update_iov(env, cio, io);
262 if (io->u.ci_rw.crw_nonblock)
263 ast_flags |= CEF_NONBLOCK;
264 result = vvp_mmap_locks(env, cio, io);
266 result = ccc_io_one_lock(env, io, ast_flags, mode, start, end);
270 static int vvp_io_read_lock(const struct lu_env *env,
271 const struct cl_io_slice *ios)
273 struct cl_io *io = ios->cis_io;
274 struct ll_inode_info *lli = ll_i2info(ccc_object_inode(io->ci_obj));
278 /* XXX: Layer violation, we shouldn't see lsm at llite level. */
279 if (lli->lli_has_smd) /* lsm-less file doesn't need to lock */
280 result = vvp_io_rw_lock(env, io, CLM_READ,
281 io->u.ci_rd.rd.crw_pos,
282 io->u.ci_rd.rd.crw_pos +
283 io->u.ci_rd.rd.crw_count - 1);
289 static int vvp_io_fault_lock(const struct lu_env *env,
290 const struct cl_io_slice *ios)
292 struct cl_io *io = ios->cis_io;
293 struct vvp_io *vio = cl2vvp_io(env, ios);
295 * XXX LDLM_FL_CBPENDING
297 return ccc_io_one_lock_index
298 (env, io, 0, vvp_mode_from_vma(vio->u.fault.ft_vma),
299 io->u.ci_fault.ft_index, io->u.ci_fault.ft_index);
302 static int vvp_io_write_lock(const struct lu_env *env,
303 const struct cl_io_slice *ios)
305 struct cl_io *io = ios->cis_io;
309 if (io->u.ci_wr.wr_append) {
311 end = OBD_OBJECT_EOF;
313 start = io->u.ci_wr.wr.crw_pos;
314 end = start + io->u.ci_wr.wr.crw_count - 1;
316 return vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
319 static int vvp_io_setattr_iter_init(const struct lu_env *env,
320 const struct cl_io_slice *ios)
326 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
328 * Handles "lockless io" mode when extent locking is done by server.
330 static int vvp_io_setattr_lock(const struct lu_env *env,
331 const struct cl_io_slice *ios)
333 struct ccc_io *cio = ccc_env_io(env);
334 struct cl_io *io = ios->cis_io;
338 if (cl_io_is_trunc(io)) {
339 new_size = io->u.ci_setattr.sa_attr.lvb_size;
341 enqflags = CEF_DISCARD_DATA;
343 if ((io->u.ci_setattr.sa_attr.lvb_mtime >=
344 io->u.ci_setattr.sa_attr.lvb_ctime) ||
345 (io->u.ci_setattr.sa_attr.lvb_atime >=
346 io->u.ci_setattr.sa_attr.lvb_ctime))
350 cio->u.setattr.cui_local_lock = SETATTR_EXTENT_LOCK;
351 return ccc_io_one_lock(env, io, enqflags, CLM_WRITE,
352 new_size, OBD_OBJECT_EOF);
355 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
359 * Only ll_inode_size_lock is taken at this level.
361 ll_inode_size_lock(inode);
362 result = inode_newsize_ok(inode, size);
364 ll_inode_size_unlock(inode);
367 truncate_setsize(inode, size);
368 ll_inode_size_unlock(inode);
372 static int vvp_io_setattr_trunc(const struct lu_env *env,
373 const struct cl_io_slice *ios,
374 struct inode *inode, loff_t size)
376 inode_dio_wait(inode);
380 static int vvp_io_setattr_time(const struct lu_env *env,
381 const struct cl_io_slice *ios)
383 struct cl_io *io = ios->cis_io;
384 struct cl_object *obj = io->ci_obj;
385 struct cl_attr *attr = ccc_env_thread_attr(env);
387 unsigned valid = CAT_CTIME;
389 cl_object_attr_lock(obj);
390 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
391 if (io->u.ci_setattr.sa_valid & ATTR_ATIME_SET) {
392 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
395 if (io->u.ci_setattr.sa_valid & ATTR_MTIME_SET) {
396 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
399 result = cl_object_attr_set(env, obj, attr, valid);
400 cl_object_attr_unlock(obj);
405 static int vvp_io_setattr_start(const struct lu_env *env,
406 const struct cl_io_slice *ios)
408 struct cl_io *io = ios->cis_io;
409 struct inode *inode = ccc_object_inode(io->ci_obj);
411 mutex_lock(&inode->i_mutex);
412 if (cl_io_is_trunc(io))
413 return vvp_io_setattr_trunc(env, ios, inode,
414 io->u.ci_setattr.sa_attr.lvb_size);
416 return vvp_io_setattr_time(env, ios);
419 static void vvp_io_setattr_end(const struct lu_env *env,
420 const struct cl_io_slice *ios)
422 struct cl_io *io = ios->cis_io;
423 struct inode *inode = ccc_object_inode(io->ci_obj);
425 if (cl_io_is_trunc(io)) {
426 /* Truncate in memory pages - they must be clean pages
427 * because osc has already notified to destroy osc_extents. */
428 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
429 inode_dio_write_done(inode);
431 mutex_unlock(&inode->i_mutex);
434 static void vvp_io_setattr_fini(const struct lu_env *env,
435 const struct cl_io_slice *ios)
437 vvp_io_fini(env, ios);
440 static ssize_t lustre_generic_file_read(struct file *file,
441 struct ccc_io *vio, loff_t *ppos)
443 return generic_file_aio_read(vio->cui_iocb, vio->cui_iov,
444 vio->cui_nrsegs, *ppos);
447 static ssize_t lustre_generic_file_write(struct file *file,
448 struct ccc_io *vio, loff_t *ppos)
450 return generic_file_aio_write(vio->cui_iocb, vio->cui_iov,
451 vio->cui_nrsegs, *ppos);
454 static int vvp_io_read_start(const struct lu_env *env,
455 const struct cl_io_slice *ios)
457 struct vvp_io *vio = cl2vvp_io(env, ios);
458 struct ccc_io *cio = cl2ccc_io(env, ios);
459 struct cl_io *io = ios->cis_io;
460 struct cl_object *obj = io->ci_obj;
461 struct inode *inode = ccc_object_inode(obj);
462 struct ll_ra_read *bead = &vio->cui_bead;
463 struct file *file = cio->cui_fd->fd_file;
466 loff_t pos = io->u.ci_rd.rd.crw_pos;
467 long cnt = io->u.ci_rd.rd.crw_count;
468 long tot = cio->cui_tot_count;
471 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
473 CDEBUG(D_VFSTRACE, "read: -> [%lli, %lli)\n", pos, pos + cnt);
475 if (!can_populate_pages(env, io, inode))
478 result = ccc_prep_size(env, obj, io, pos, tot, &exceed);
481 else if (exceed != 0)
484 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
485 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
486 inode->i_ino, cnt, pos, i_size_read(inode));
488 /* turn off the kernel's read-ahead */
489 cio->cui_fd->fd_file->f_ra.ra_pages = 0;
491 /* initialize read-ahead window once per syscall */
492 if (!vio->cui_ra_window_set) {
493 vio->cui_ra_window_set = 1;
494 bead->lrr_start = cl_index(obj, pos);
496 * XXX: explicit PAGE_CACHE_SIZE
498 bead->lrr_count = cl_index(obj, tot + PAGE_CACHE_SIZE - 1);
499 ll_ra_read_in(file, bead);
504 switch (vio->cui_io_subtype) {
506 result = lustre_generic_file_read(file, cio, &pos);
509 result = generic_file_splice_read(file, &pos,
510 vio->u.splice.cui_pipe, cnt,
511 vio->u.splice.cui_flags);
512 /* LU-1109: do splice read stripe by stripe otherwise if it
513 * may make nfsd stuck if this read occupied all internal pipe
518 CERROR("Wrong IO type %u\n", vio->cui_io_subtype);
526 io->ci_nob += result;
527 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
528 cio->cui_fd, pos, result, 0);
534 static void vvp_io_read_fini(const struct lu_env *env, const struct cl_io_slice *ios)
536 struct vvp_io *vio = cl2vvp_io(env, ios);
537 struct ccc_io *cio = cl2ccc_io(env, ios);
539 if (vio->cui_ra_window_set)
540 ll_ra_read_ex(cio->cui_fd->fd_file, &vio->cui_bead);
542 vvp_io_fini(env, ios);
545 static int vvp_io_write_start(const struct lu_env *env,
546 const struct cl_io_slice *ios)
548 struct ccc_io *cio = cl2ccc_io(env, ios);
549 struct cl_io *io = ios->cis_io;
550 struct cl_object *obj = io->ci_obj;
551 struct inode *inode = ccc_object_inode(obj);
552 struct file *file = cio->cui_fd->fd_file;
554 loff_t pos = io->u.ci_wr.wr.crw_pos;
555 size_t cnt = io->u.ci_wr.wr.crw_count;
559 if (!can_populate_pages(env, io, inode))
562 if (cl_io_is_append(io)) {
564 * PARALLEL IO This has to be changed for parallel IO doing
565 * out-of-order writes.
567 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
568 cio->cui_iocb->ki_pos = pos;
571 CDEBUG(D_VFSTRACE, "write: [%lli, %lli)\n", pos, pos + (long long)cnt);
573 if (cio->cui_iov == NULL) /* from a temp io in ll_cl_init(). */
576 result = lustre_generic_file_write(file, cio, &pos);
581 io->ci_nob += result;
582 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
583 cio->cui_fd, pos, result, 0);
589 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
591 struct vm_fault *vmf = cfio->fault.ft_vmf;
593 cfio->fault.ft_flags = filemap_fault(cfio->ft_vma, vmf);
596 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
597 vmf->virtual_address);
598 if (unlikely(!(cfio->fault.ft_flags & VM_FAULT_LOCKED))) {
599 lock_page(vmf->page);
600 cfio->fault.ft_flags &= VM_FAULT_LOCKED;
603 cfio->ft_vmpage = vmf->page;
607 if (cfio->fault.ft_flags & VM_FAULT_SIGBUS) {
608 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", vmf->virtual_address);
612 if (cfio->fault.ft_flags & VM_FAULT_OOM) {
613 CDEBUG(D_PAGE, "got addr %p - OOM\n", vmf->virtual_address);
617 if (cfio->fault.ft_flags & VM_FAULT_RETRY)
620 CERROR("unknow error in page fault %d!\n", cfio->fault.ft_flags);
625 static int vvp_io_fault_start(const struct lu_env *env,
626 const struct cl_io_slice *ios)
628 struct vvp_io *vio = cl2vvp_io(env, ios);
629 struct cl_io *io = ios->cis_io;
630 struct cl_object *obj = io->ci_obj;
631 struct inode *inode = ccc_object_inode(obj);
632 struct cl_fault_io *fio = &io->u.ci_fault;
633 struct vvp_fault_io *cfio = &vio->u.fault;
636 struct page *vmpage = NULL;
637 struct cl_page *page;
639 pgoff_t last; /* last page in a file data region */
641 if (fio->ft_executable &&
642 LTIME_S(inode->i_mtime) != vio->u.fault.ft_mtime)
644 " changed while waiting for the page fault lock\n",
645 PFID(lu_object_fid(&obj->co_lu)));
647 /* offset of the last byte on the page */
648 offset = cl_offset(obj, fio->ft_index + 1) - 1;
649 LASSERT(cl_index(obj, offset) == fio->ft_index);
650 result = ccc_prep_size(env, obj, io, 0, offset + 1, NULL);
654 /* must return locked page */
655 if (fio->ft_mkwrite) {
656 LASSERT(cfio->ft_vmpage != NULL);
657 lock_page(cfio->ft_vmpage);
659 result = vvp_io_kernel_fault(cfio);
664 vmpage = cfio->ft_vmpage;
665 LASSERT(PageLocked(vmpage));
667 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
668 ll_invalidate_page(vmpage);
670 size = i_size_read(inode);
671 /* Though we have already held a cl_lock upon this page, but
672 * it still can be truncated locally. */
673 if (unlikely((vmpage->mapping != inode->i_mapping) ||
674 (page_offset(vmpage) > size))) {
675 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
677 /* return +1 to stop cl_io_loop() and ll_fault() will catch
679 GOTO(out, result = +1);
683 if (fio->ft_mkwrite ) {
686 * Capture the size while holding the lli_trunc_sem from above
687 * we want to make sure that we complete the mkwrite action
688 * while holding this lock. We need to make sure that we are
689 * not past the end of the file.
691 last_index = cl_index(obj, size - 1);
692 if (last_index < fio->ft_index) {
694 "llite: mkwrite and truncate race happened: "
696 vmpage->mapping,fio->ft_index,last_index);
698 * We need to return if we are
699 * passed the end of the file. This will propagate
700 * up the call stack to ll_page_mkwrite where
701 * we will return VM_FAULT_NOPAGE. Any non-negative
702 * value returned here will be silently
703 * converted to 0. If the vmpage->mapping is null
704 * the error code would be converted back to ENODATA
705 * in ll_page_mkwrite0. Thus we return -ENODATA
706 * to handle both cases
708 GOTO(out, result = -ENODATA);
712 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
714 GOTO(out, result = PTR_ERR(page));
716 /* if page is going to be written, we should add this page into cache
718 if (fio->ft_mkwrite) {
719 wait_on_page_writeback(vmpage);
720 if (set_page_dirty(vmpage)) {
723 /* vvp_page_assume() calls wait_on_page_writeback(). */
724 cl_page_assume(env, io, page);
726 cp = cl2ccc_page(cl_page_at(page, &vvp_device_type));
727 vvp_write_pending(cl2ccc(obj), cp);
729 /* Do not set Dirty bit here so that in case IO is
730 * started before the page is really made dirty, we
731 * still have chance to detect it. */
732 result = cl_page_cache_add(env, io, page, CRT_WRITE);
733 LASSERT(cl_page_is_owned(page, io));
737 cl_page_unmap(env, io, page);
738 cl_page_discard(env, io, page);
739 cl_page_disown(env, io, page);
741 cl_page_put(env, page);
743 /* we're in big trouble, what can we do now? */
744 if (result == -EDQUOT)
748 cl_page_disown(env, io, page);
752 last = cl_index(obj, size - 1);
754 * The ft_index is only used in the case of
755 * a mkwrite action. We need to check
756 * our assertions are correct, since
757 * we should have caught this above
759 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last);
760 if (fio->ft_index == last)
762 * Last page is mapped partially.
764 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
766 fio->ft_nob = cl_page_size(obj);
768 lu_ref_add(&page->cp_reference, "fault", io);
773 /* return unlocked vmpage to avoid deadlocking */
776 cfio->fault.ft_flags &= ~VM_FAULT_LOCKED;
780 static int vvp_io_fsync_start(const struct lu_env *env,
781 const struct cl_io_slice *ios)
783 /* we should mark TOWRITE bit to each dirty page in radix tree to
784 * verify pages have been written, but this is difficult because of
789 static int vvp_io_read_page(const struct lu_env *env,
790 const struct cl_io_slice *ios,
791 const struct cl_page_slice *slice)
793 struct cl_io *io = ios->cis_io;
794 struct cl_object *obj = slice->cpl_obj;
795 struct ccc_page *cp = cl2ccc_page(slice);
796 struct cl_page *page = slice->cpl_page;
797 struct inode *inode = ccc_object_inode(obj);
798 struct ll_sb_info *sbi = ll_i2sbi(inode);
799 struct ll_file_data *fd = cl2ccc_io(env, ios)->cui_fd;
800 struct ll_readahead_state *ras = &fd->fd_ras;
801 struct page *vmpage = cp->cpg_page;
802 struct cl_2queue *queue = &io->ci_queue;
805 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
806 LASSERT(slice->cpl_obj == obj);
810 if (sbi->ll_ra_info.ra_max_pages_per_file &&
811 sbi->ll_ra_info.ra_max_pages)
812 ras_update(sbi, inode, ras, page->cp_index,
813 cp->cpg_defer_uptodate);
815 /* Sanity check whether the page is protected by a lock. */
816 rc = cl_page_is_under_lock(env, io, page);
818 CL_PAGE_HEADER(D_WARNING, env, page, "%s: %d\n",
819 rc == -ENODATA ? "without a lock" :
825 if (cp->cpg_defer_uptodate) {
827 cl_page_export(env, page, 1);
830 * Add page into the queue even when it is marked uptodate above.
831 * this will unlock it automatically as part of cl_page_list_disown().
833 cl_2queue_add(queue, page);
834 if (sbi->ll_ra_info.ra_max_pages_per_file &&
835 sbi->ll_ra_info.ra_max_pages)
836 ll_readahead(env, io, ras,
837 vmpage->mapping, &queue->c2_qin, fd->fd_flags);
842 static int vvp_page_sync_io(const struct lu_env *env, struct cl_io *io,
843 struct cl_page *page, struct ccc_page *cp,
844 enum cl_req_type crt)
846 struct cl_2queue *queue;
849 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
851 queue = &io->ci_queue;
852 cl_2queue_init_page(queue, page);
854 result = cl_io_submit_sync(env, io, crt, queue, 0);
855 LASSERT(cl_page_is_owned(page, io));
859 * in CRT_WRITE case page is left locked even in case of
862 cl_page_list_disown(env, io, &queue->c2_qin);
863 cl_2queue_fini(env, queue);
869 * Prepare partially written-to page for a write.
871 static int vvp_io_prepare_partial(const struct lu_env *env, struct cl_io *io,
872 struct cl_object *obj, struct cl_page *pg,
874 unsigned from, unsigned to)
876 struct cl_attr *attr = ccc_env_thread_attr(env);
877 loff_t offset = cl_offset(obj, pg->cp_index);
880 cl_object_attr_lock(obj);
881 result = cl_object_attr_get(env, obj, attr);
882 cl_object_attr_unlock(obj);
885 * If are writing to a new page, no need to read old data.
886 * The extent locking will have updated the KMS, and for our
887 * purposes here we can treat it like i_size.
889 if (attr->cat_kms <= offset) {
890 char *kaddr = ll_kmap_atomic(cp->cpg_page, KM_USER0);
892 memset(kaddr, 0, cl_page_size(obj));
893 ll_kunmap_atomic(kaddr, KM_USER0);
894 } else if (cp->cpg_defer_uptodate)
897 result = vvp_page_sync_io(env, io, pg, cp, CRT_READ);
899 * In older implementations, obdo_refresh_inode is called here
900 * to update the inode because the write might modify the
901 * object info at OST. However, this has been proven useless,
902 * since LVB functions will be called when user space program
903 * tries to retrieve inode attribute. Also, see bug 15909 for
907 cl_page_export(env, pg, 1);
912 static int vvp_io_prepare_write(const struct lu_env *env,
913 const struct cl_io_slice *ios,
914 const struct cl_page_slice *slice,
915 unsigned from, unsigned to)
917 struct cl_object *obj = slice->cpl_obj;
918 struct ccc_page *cp = cl2ccc_page(slice);
919 struct cl_page *pg = slice->cpl_page;
920 struct page *vmpage = cp->cpg_page;
926 LINVRNT(cl_page_is_vmlocked(env, pg));
927 LASSERT(vmpage->mapping->host == ccc_object_inode(obj));
931 CL_PAGE_HEADER(D_PAGE, env, pg, "preparing: [%d, %d]\n", from, to);
932 if (!PageUptodate(vmpage)) {
934 * We're completely overwriting an existing page, so _don't_
935 * set it up to date until commit_write
937 if (from == 0 && to == PAGE_CACHE_SIZE) {
938 CL_PAGE_HEADER(D_PAGE, env, pg, "full page write\n");
939 POISON_PAGE(page, 0x11);
941 result = vvp_io_prepare_partial(env, ios->cis_io, obj,
944 CL_PAGE_HEADER(D_PAGE, env, pg, "uptodate\n");
948 static int vvp_io_commit_write(const struct lu_env *env,
949 const struct cl_io_slice *ios,
950 const struct cl_page_slice *slice,
951 unsigned from, unsigned to)
953 struct cl_object *obj = slice->cpl_obj;
954 struct cl_io *io = ios->cis_io;
955 struct ccc_page *cp = cl2ccc_page(slice);
956 struct cl_page *pg = slice->cpl_page;
957 struct inode *inode = ccc_object_inode(obj);
958 struct ll_sb_info *sbi = ll_i2sbi(inode);
959 struct ll_inode_info *lli = ll_i2info(inode);
960 struct page *vmpage = cp->cpg_page;
968 LINVRNT(cl_page_is_vmlocked(env, pg));
969 LASSERT(vmpage->mapping->host == inode);
971 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu, "commiting page write\n");
972 CL_PAGE_HEADER(D_PAGE, env, pg, "committing: [%d, %d]\n", from, to);
975 * queue a write for some time in the future the first time we
978 * This is different from what other file systems do: they usually
979 * just mark page (and some of its buffers) dirty and rely on
980 * balance_dirty_pages() to start a write-back. Lustre wants write-back
981 * to be started earlier for the following reasons:
983 * (1) with a large number of clients we need to limit the amount
984 * of cached data on the clients a lot;
986 * (2) large compute jobs generally want compute-only then io-only
987 * and the IO should complete as quickly as possible;
989 * (3) IO is batched up to the RPC size and is async until the
990 * client max cache is hit
991 * (/proc/fs/lustre/osc/OSC.../max_dirty_mb)
994 if (!PageDirty(vmpage)) {
995 tallyop = LPROC_LL_DIRTY_MISSES;
996 result = cl_page_cache_add(env, io, pg, CRT_WRITE);
998 /* page was added into cache successfully. */
999 set_page_dirty(vmpage);
1000 vvp_write_pending(cl2ccc(obj), cp);
1001 } else if (result == -EDQUOT) {
1002 pgoff_t last_index = i_size_read(inode) >> PAGE_CACHE_SHIFT;
1003 bool need_clip = true;
1006 * Client ran out of disk space grant. Possible
1009 * (a) do a sync write, renewing grant;
1011 * (b) stop writing on this stripe, switch to the
1014 * (b) is a part of "parallel io" design that is the
1015 * ultimate goal. (a) is what "old" client did, and
1016 * what the new code continues to do for the time
1019 if (last_index > pg->cp_index) {
1020 to = PAGE_CACHE_SIZE;
1022 } else if (last_index == pg->cp_index) {
1023 int size_to = i_size_read(inode) & ~CFS_PAGE_MASK;
1028 cl_page_clip(env, pg, 0, to);
1029 result = vvp_page_sync_io(env, io, pg, cp, CRT_WRITE);
1031 CERROR("Write page %lu of inode %p failed %d\n",
1032 pg->cp_index, inode, result);
1035 tallyop = LPROC_LL_DIRTY_HITS;
1038 ll_stats_ops_tally(sbi, tallyop, 1);
1040 /* Inode should be marked DIRTY even if no new page was marked DIRTY
1041 * because page could have been not flushed between 2 modifications.
1042 * It is important the file is marked DIRTY as soon as the I/O is done
1043 * Indeed, when cache is flushed, file could be already closed and it
1044 * is too late to warn the MDT.
1045 * It is acceptable that file is marked DIRTY even if I/O is dropped
1046 * for some reasons before being flushed to OST.
1049 spin_lock(&lli->lli_lock);
1050 lli->lli_flags |= LLIF_DATA_MODIFIED;
1051 spin_unlock(&lli->lli_lock);
1054 size = cl_offset(obj, pg->cp_index) + to;
1056 ll_inode_size_lock(inode);
1058 if (size > i_size_read(inode)) {
1059 cl_isize_write_nolock(inode, size);
1060 CDEBUG(D_VFSTRACE, DFID" updating i_size %lu\n",
1061 PFID(lu_object_fid(&obj->co_lu)),
1062 (unsigned long)size);
1064 cl_page_export(env, pg, 1);
1066 if (size > i_size_read(inode))
1067 cl_page_discard(env, io, pg);
1069 ll_inode_size_unlock(inode);
1073 static const struct cl_io_operations vvp_io_ops = {
1076 .cio_fini = vvp_io_read_fini,
1077 .cio_lock = vvp_io_read_lock,
1078 .cio_start = vvp_io_read_start,
1079 .cio_advance = ccc_io_advance
1082 .cio_fini = vvp_io_fini,
1083 .cio_lock = vvp_io_write_lock,
1084 .cio_start = vvp_io_write_start,
1085 .cio_advance = ccc_io_advance
1088 .cio_fini = vvp_io_setattr_fini,
1089 .cio_iter_init = vvp_io_setattr_iter_init,
1090 .cio_lock = vvp_io_setattr_lock,
1091 .cio_start = vvp_io_setattr_start,
1092 .cio_end = vvp_io_setattr_end
1095 .cio_fini = vvp_io_fault_fini,
1096 .cio_iter_init = vvp_io_fault_iter_init,
1097 .cio_lock = vvp_io_fault_lock,
1098 .cio_start = vvp_io_fault_start,
1099 .cio_end = ccc_io_end
1102 .cio_start = vvp_io_fsync_start,
1103 .cio_fini = vvp_io_fini
1106 .cio_fini = vvp_io_fini
1109 .cio_read_page = vvp_io_read_page,
1110 .cio_prepare_write = vvp_io_prepare_write,
1111 .cio_commit_write = vvp_io_commit_write
1114 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1117 struct vvp_io *vio = vvp_env_io(env);
1118 struct ccc_io *cio = ccc_env_io(env);
1119 struct inode *inode = ccc_object_inode(obj);
1122 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
1125 CL_IO_SLICE_CLEAN(cio, cui_cl);
1126 cl_io_slice_add(io, &cio->cui_cl, obj, &vvp_io_ops);
1127 vio->cui_ra_window_set = 0;
1129 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1131 struct ll_inode_info *lli = ll_i2info(inode);
1133 count = io->u.ci_rw.crw_count;
1134 /* "If nbyte is 0, read() will return 0 and have no other
1135 * results." -- Single Unix Spec */
1139 cio->cui_tot_count = count;
1140 cio->cui_tot_nrsegs = 0;
1142 /* for read/write, we store the jobid in the inode, and
1143 * it'll be fetched by osc when building RPC.
1145 * it's not accurate if the file is shared by different
1148 lustre_get_jobid(lli->lli_jobid);
1149 } else if (io->ci_type == CIT_SETATTR) {
1150 if (!cl_io_is_trunc(io))
1151 io->ci_lockreq = CILR_MANDATORY;
1154 /* ignore layout change for generic CIT_MISC but not for glimpse.
1155 * io context for glimpse must set ci_verify_layout to true,
1156 * see cl_glimpse_size0() for details. */
1157 if (io->ci_type == CIT_MISC && !io->ci_verify_layout)
1158 io->ci_ignore_layout = 1;
1160 /* Enqueue layout lock and get layout version. We need to do this
1161 * even for operations requiring to open file, such as read and write,
1162 * because it might not grant layout lock in IT_OPEN. */
1163 if (result == 0 && !io->ci_ignore_layout)
1164 result = ll_layout_refresh(inode, &cio->cui_layout_gen);
1169 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
1170 const struct cl_io_slice *slice)
1172 /* Caling just for assertion */
1173 cl2ccc_io(env, slice);
1174 return vvp_env_io(env);