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.gnu.org/licenses/gpl-2.0.html
23 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
24 * Use is subject to license terms.
26 * Copyright (c) 2011, 2017, Intel Corporation.
29 * This file is part of Lustre, http://www.lustre.org/
30 * Lustre is a trademark of Sun Microsystems, Inc.
32 * Implementation of cl_io for VVP layer.
34 * Author: Nikita Danilov <nikita.danilov@sun.com>
35 * Author: Jinshan Xiong <jinshan.xiong@whamcloud.com>
38 #define DEBUG_SUBSYSTEM S_LLITE
42 #include "llite_internal.h"
43 #include "vvp_internal.h"
45 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
46 const struct cl_io_slice *slice)
50 vio = container_of(slice, struct vvp_io, vui_cl);
51 LASSERT(vio == vvp_env_io(env));
57 * True, if \a io is a normal io, False for splice_{read,write}
59 static int cl_is_normalio(const struct lu_env *env, const struct cl_io *io)
61 struct vvp_io *vio = vvp_env_io(env);
63 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
65 return vio->vui_io_subtype == IO_NORMAL;
69 * For swapping layout. The file's layout may have changed.
70 * To avoid populating pages to a wrong stripe, we have to verify the
71 * correctness of layout. It works because swapping layout processes
72 * have to acquire group lock.
74 static bool can_populate_pages(const struct lu_env *env, struct cl_io *io,
77 struct ll_inode_info *lli = ll_i2info(inode);
78 struct vvp_io *vio = vvp_env_io(env);
81 switch (io->ci_type) {
84 /* don't need lock here to check lli_layout_gen as we have held
85 * extent lock and GROUP lock has to hold to swap layout */
86 if (ll_layout_version_get(lli) != vio->vui_layout_gen ||
87 OBD_FAIL_CHECK_RESET(OBD_FAIL_LLITE_LOST_LAYOUT, 0)) {
88 io->ci_need_restart = 1;
89 /* this will cause a short read/write */
94 /* fault is okay because we've already had a page. */
102 static void vvp_object_size_lock(struct cl_object *obj)
104 struct inode *inode = vvp_object_inode(obj);
106 ll_inode_size_lock(inode);
107 cl_object_attr_lock(obj);
110 static void vvp_object_size_unlock(struct cl_object *obj)
112 struct inode *inode = vvp_object_inode(obj);
114 cl_object_attr_unlock(obj);
115 ll_inode_size_unlock(inode);
119 * Helper function that if necessary adjusts file size (inode->i_size), when
120 * position at the offset \a pos is accessed. File size can be arbitrary stale
121 * on a Lustre client, but client at least knows KMS. If accessed area is
122 * inside [0, KMS], set file size to KMS, otherwise glimpse file size.
124 * Locking: i_size_lock is used to serialize changes to inode size and to
125 * protect consistency between inode size and cl_object
126 * attributes. cl_object_size_lock() protects consistency between cl_attr's of
127 * top-object and sub-objects.
129 static int vvp_prep_size(const struct lu_env *env, struct cl_object *obj,
130 struct cl_io *io, loff_t start, size_t count,
133 struct cl_attr *attr = vvp_env_thread_attr(env);
134 struct inode *inode = vvp_object_inode(obj);
135 loff_t pos = start + count - 1;
140 * Consistency guarantees: following possibilities exist for the
141 * relation between region being accessed and real file size at this
144 * (A): the region is completely inside of the file;
146 * (B-x): x bytes of region are inside of the file, the rest is
149 * (C): the region is completely outside of the file.
151 * This classification is stable under DLM lock already acquired by
152 * the caller, because to change the class, other client has to take
153 * DLM lock conflicting with our lock. Also, any updates to ->i_size
154 * by other threads on this client are serialized by
155 * ll_inode_size_lock(). This guarantees that short reads are handled
156 * correctly in the face of concurrent writes and truncates.
158 vvp_object_size_lock(obj);
159 result = cl_object_attr_get(env, obj, attr);
164 * A glimpse is necessary to determine whether we
165 * return a short read (B) or some zeroes at the end
168 vvp_object_size_unlock(obj);
169 result = cl_glimpse_lock(env, io, inode, obj, 0);
170 if (result == 0 && exceed != NULL) {
171 /* If objective page index exceed end-of-file
172 * page index, return directly. Do not expect
173 * kernel will check such case correctly.
174 * linux-2.6.18-128.1.1 miss to do that.
176 loff_t size = i_size_read(inode);
177 unsigned long cur_index = start >>
180 if ((size == 0 && cur_index != 0) ||
181 (((size - 1) >> PAGE_SHIFT) <
189 * region is within kms and, hence, within real file
190 * size (A). We need to increase i_size to cover the
191 * read region so that generic_file_read() will do its
192 * job, but that doesn't mean the kms size is
193 * _correct_, it is only the _minimum_ size. If
194 * someone does a stat they will get the correct size
195 * which will always be >= the kms value here.
198 if (i_size_read(inode) < kms) {
199 i_size_write(inode, kms);
201 DFID" updating i_size %llu\n",
202 PFID(lu_object_fid(&obj->co_lu)),
203 (__u64)i_size_read(inode));
208 vvp_object_size_unlock(obj);
213 /*****************************************************************************
219 static int vvp_io_one_lock_index(const struct lu_env *env, struct cl_io *io,
220 __u32 enqflags, enum cl_lock_mode mode,
221 pgoff_t start, pgoff_t end)
223 struct vvp_io *vio = vvp_env_io(env);
224 struct cl_lock_descr *descr = &vio->vui_link.cill_descr;
225 struct cl_object *obj = io->ci_obj;
227 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
230 CDEBUG(D_VFSTRACE, "lock: %d [%lu, %lu]\n", mode, start, end);
232 memset(&vio->vui_link, 0, sizeof vio->vui_link);
234 if (vio->vui_fd && (vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
235 descr->cld_mode = CLM_GROUP;
236 descr->cld_gid = vio->vui_fd->fd_grouplock.lg_gid;
237 enqflags |= CEF_LOCK_MATCH;
239 descr->cld_mode = mode;
242 descr->cld_obj = obj;
243 descr->cld_start = start;
244 descr->cld_end = end;
245 descr->cld_enq_flags = enqflags;
247 cl_io_lock_add(env, io, &vio->vui_link);
252 static int vvp_io_one_lock(const struct lu_env *env, struct cl_io *io,
253 __u32 enqflags, enum cl_lock_mode mode,
254 loff_t start, loff_t end)
256 struct cl_object *obj = io->ci_obj;
258 return vvp_io_one_lock_index(env, io, enqflags, mode,
259 cl_index(obj, start), cl_index(obj, end));
262 static int vvp_io_write_iter_init(const struct lu_env *env,
263 const struct cl_io_slice *ios)
265 struct vvp_io *vio = cl2vvp_io(env, ios);
267 cl_page_list_init(&vio->u.write.vui_queue);
268 vio->u.write.vui_written = 0;
269 vio->u.write.vui_from = 0;
270 vio->u.write.vui_to = PAGE_SIZE;
275 static void vvp_io_write_iter_fini(const struct lu_env *env,
276 const struct cl_io_slice *ios)
278 struct vvp_io *vio = cl2vvp_io(env, ios);
280 LASSERT(vio->u.write.vui_queue.pl_nr == 0);
283 static int vvp_io_fault_iter_init(const struct lu_env *env,
284 const struct cl_io_slice *ios)
286 struct vvp_io *vio = cl2vvp_io(env, ios);
287 struct inode *inode = vvp_object_inode(ios->cis_obj);
289 LASSERT(inode == file_inode(vio->vui_fd->fd_file));
290 vio->u.fault.ft_mtime = inode->i_mtime.tv_sec;
295 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
297 struct cl_io *io = ios->cis_io;
298 struct cl_object *obj = io->ci_obj;
299 struct vvp_io *vio = cl2vvp_io(env, ios);
300 struct inode *inode = vvp_object_inode(obj);
305 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
307 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
308 "need write layout %d, restore needed %d\n",
309 PFID(lu_object_fid(&obj->co_lu)),
310 io->ci_ignore_layout, io->ci_verify_layout,
311 vio->vui_layout_gen, io->ci_need_write_intent,
312 io->ci_restore_needed);
314 if (io->ci_restore_needed) {
315 /* file was detected release, we need to restore it
316 * before finishing the io
318 rc = ll_layout_restore(inode, 0, OBD_OBJECT_EOF);
319 /* if restore registration failed, no restart,
320 * we will return -ENODATA */
321 /* The layout will change after restore, so we need to
322 * block on layout lock held by the MDT
323 * as MDT will not send new layout in lvb (see LU-3124)
324 * we have to explicitly fetch it, all this will be done
325 * by ll_layout_refresh().
326 * Even if ll_layout_restore() returns zero, it doesn't mean
327 * that restore has been successful. Therefore it sets
328 * ci_verify_layout so that it will check layout at the end
332 io->ci_restore_needed = 1;
333 io->ci_need_restart = 0;
334 io->ci_verify_layout = 0;
339 io->ci_restore_needed = 0;
341 /* Even if ll_layout_restore() returns zero, it doesn't mean
342 * that restore has been successful. Therefore it should verify
343 * if there was layout change and restart I/O correspondingly.
345 ll_layout_refresh(inode, &gen);
346 io->ci_need_restart = vio->vui_layout_gen != gen;
347 if (io->ci_need_restart) {
349 DFID" layout changed from %d to %d.\n",
350 PFID(lu_object_fid(&obj->co_lu)),
351 vio->vui_layout_gen, gen);
352 /* today successful restore is the only possible
354 /* restore was done, clear restoring state */
355 ll_file_clear_flag(ll_i2info(vvp_object_inode(obj)),
356 LLIF_FILE_RESTORING);
362 * dynamic layout change needed, send layout intent
365 if (io->ci_need_write_intent) {
366 enum layout_intent_opc opc = LAYOUT_INTENT_WRITE;
368 io->ci_need_write_intent = 0;
370 LASSERT(io->ci_type == CIT_WRITE ||
371 cl_io_is_trunc(io) || cl_io_is_mkwrite(io));
373 CDEBUG(D_VFSTRACE, DFID" write layout, type %u "DEXT"\n",
374 PFID(lu_object_fid(&obj->co_lu)), io->ci_type,
375 PEXT(&io->ci_write_intent));
377 if (cl_io_is_trunc(io))
378 opc = LAYOUT_INTENT_TRUNC;
380 rc = ll_layout_write_intent(inode, opc, &io->ci_write_intent);
383 io->ci_need_restart = 1;
387 if (!io->ci_need_restart &&
388 !io->ci_ignore_layout && io->ci_verify_layout) {
389 /* check layout version */
390 ll_layout_refresh(inode, &gen);
391 io->ci_need_restart = vio->vui_layout_gen != gen;
392 if (io->ci_need_restart) {
394 DFID" layout changed from %d to %d.\n",
395 PFID(lu_object_fid(&obj->co_lu)),
396 vio->vui_layout_gen, gen);
404 static void vvp_io_fault_fini(const struct lu_env *env,
405 const struct cl_io_slice *ios)
407 struct cl_io *io = ios->cis_io;
408 struct cl_page *page = io->u.ci_fault.ft_page;
410 CLOBINVRNT(env, io->ci_obj, vvp_object_invariant(io->ci_obj));
413 lu_ref_del(&page->cp_reference, "fault", io);
414 cl_page_put(env, page);
415 io->u.ci_fault.ft_page = NULL;
417 vvp_io_fini(env, ios);
420 static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
423 * we only want to hold PW locks if the mmap() can generate
424 * writes back to the file and that only happens in shared
427 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
432 static int vvp_mmap_locks(const struct lu_env *env, struct cl_io *io)
434 struct vvp_thread_info *vti = vvp_env_info(env);
435 struct mm_struct *mm = current->mm;
436 struct vm_area_struct *vma;
437 struct cl_lock_descr *descr = &vti->vti_descr;
438 union ldlm_policy_data policy;
444 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
446 if (!cl_is_normalio(env, io))
449 /* No MM (e.g. NFS)? No vmas too. */
453 iov_for_each(iov, i, io->u.ci_rw.rw_iter) {
454 unsigned long addr = (unsigned long)iov.iov_base;
455 size_t count = iov.iov_len;
460 count += addr & ~PAGE_MASK;
463 down_read(&mm->mmap_sem);
464 while((vma = our_vma(mm, addr, count)) != NULL) {
465 struct dentry *de = file_dentry(vma->vm_file);
466 struct inode *inode = de->d_inode;
467 int flags = CEF_MUST;
469 if (ll_file_nolock(vma->vm_file)) {
471 * For no lock case is not allowed for mmap
478 * XXX: Required lock mode can be weakened: CIT_WRITE
479 * io only ever reads user level buffer, and CIT_READ
482 policy_from_vma(&policy, vma, addr, count);
483 descr->cld_mode = vvp_mode_from_vma(vma);
484 descr->cld_obj = ll_i2info(inode)->lli_clob;
485 descr->cld_start = cl_index(descr->cld_obj,
486 policy.l_extent.start);
487 descr->cld_end = cl_index(descr->cld_obj,
488 policy.l_extent.end);
489 descr->cld_enq_flags = flags;
490 result = cl_io_lock_alloc_add(env, io, descr);
492 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
493 descr->cld_mode, descr->cld_start,
499 if (vma->vm_end - addr >= count)
502 count -= vma->vm_end - addr;
505 up_read(&mm->mmap_sem);
512 static void vvp_io_advance(const struct lu_env *env,
513 const struct cl_io_slice *ios,
516 struct vvp_io *vio = cl2vvp_io(env, ios);
517 struct cl_io *io = ios->cis_io;
518 struct cl_object *obj = ios->cis_io->ci_obj;
520 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
522 if (!cl_is_normalio(env, io))
525 vio->vui_tot_count -= nob;
527 iov_iter_advance(&io->u.ci_rw.rw_iter, nob);
528 io->u.ci_rw.rw_iocb.ki_pos = io->u.ci_rw.rw_range.cir_pos;
529 #ifdef HAVE_KIOCB_KI_LEFT
530 io->u.ci_rw.rw_iocb.ki_left = vio->vui_tot_count;
531 #elif defined(HAVE_KI_NBYTES)
532 io->u.ci_rw.rw_iocb.ki_nbytes = vio->vui_tot_count;
535 /* It was truncated to stripe size in vvp_io_rw_lock() */
536 iov_iter_reexpand(&io->u.ci_rw.rw_iter, vio->vui_tot_count);
540 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
541 enum cl_lock_mode mode, loff_t start, loff_t end)
546 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
549 if (cl_is_normalio(env, io))
550 iov_iter_truncate(&io->u.ci_rw.rw_iter,
551 io->u.ci_rw.rw_range.cir_count);
553 if (io->u.ci_rw.rw_nonblock)
554 ast_flags |= CEF_NONBLOCK;
555 if (io->ci_lock_no_expand)
556 ast_flags |= CEF_LOCK_NO_EXPAND;
558 result = vvp_mmap_locks(env, io);
560 result = vvp_io_one_lock(env, io, ast_flags, mode, start, end);
565 static int vvp_io_read_lock(const struct lu_env *env,
566 const struct cl_io_slice *ios)
568 struct cl_io *io = ios->cis_io;
569 struct cl_io_range *range = &io->u.ci_rw.rw_range;
573 rc = vvp_io_rw_lock(env, io, CLM_READ, range->cir_pos,
574 range->cir_pos + range->cir_count - 1);
578 static int vvp_io_fault_lock(const struct lu_env *env,
579 const struct cl_io_slice *ios)
581 struct cl_io *io = ios->cis_io;
582 struct vvp_io *vio = cl2vvp_io(env, ios);
584 * XXX LDLM_FL_CBPENDING
586 return vvp_io_one_lock_index(env,
588 vvp_mode_from_vma(vio->u.fault.ft_vma),
589 io->u.ci_fault.ft_index,
590 io->u.ci_fault.ft_index);
593 static int vvp_io_write_lock(const struct lu_env *env,
594 const struct cl_io_slice *ios)
596 struct cl_io *io = ios->cis_io;
602 if (io->u.ci_rw.rw_append) {
604 end = OBD_OBJECT_EOF;
606 start = io->u.ci_rw.rw_range.cir_pos;
607 end = start + io->u.ci_rw.rw_range.cir_count - 1;
609 rc = vvp_io_rw_lock(env, io, CLM_WRITE, start, end);
613 static int vvp_io_setattr_iter_init(const struct lu_env *env,
614 const struct cl_io_slice *ios)
620 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
622 * Handles "lockless io" mode when extent locking is done by server.
624 static int vvp_io_setattr_lock(const struct lu_env *env,
625 const struct cl_io_slice *ios)
627 struct cl_io *io = ios->cis_io;
631 if (cl_io_is_trunc(io)) {
632 new_size = io->u.ci_setattr.sa_attr.lvb_size;
634 enqflags = CEF_DISCARD_DATA;
636 unsigned int valid = io->u.ci_setattr.sa_avalid;
638 if (!(valid & TIMES_SET_FLAGS))
641 if ((!(valid & ATTR_MTIME) ||
642 io->u.ci_setattr.sa_attr.lvb_mtime >=
643 io->u.ci_setattr.sa_attr.lvb_ctime) &&
644 (!(valid & ATTR_ATIME) ||
645 io->u.ci_setattr.sa_attr.lvb_atime >=
646 io->u.ci_setattr.sa_attr.lvb_ctime))
652 return vvp_io_one_lock(env, io, enqflags, CLM_WRITE,
653 new_size, OBD_OBJECT_EOF);
656 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
661 * Only ll_inode_size_lock is taken at this level.
663 ll_inode_size_lock(inode);
664 result = inode_newsize_ok(inode, size);
666 ll_inode_size_unlock(inode);
669 i_size_write(inode, size);
671 ll_truncate_pagecache(inode, size);
672 ll_inode_size_unlock(inode);
676 static int vvp_io_setattr_time(const struct lu_env *env,
677 const struct cl_io_slice *ios)
679 struct cl_io *io = ios->cis_io;
680 struct cl_object *obj = io->ci_obj;
681 struct cl_attr *attr = vvp_env_thread_attr(env);
683 unsigned valid = CAT_CTIME;
685 cl_object_attr_lock(obj);
686 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
687 if (io->u.ci_setattr.sa_avalid & ATTR_ATIME_SET) {
688 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
691 if (io->u.ci_setattr.sa_avalid & ATTR_MTIME_SET) {
692 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
695 result = cl_object_attr_update(env, obj, attr, valid);
696 cl_object_attr_unlock(obj);
701 static int vvp_io_setattr_start(const struct lu_env *env,
702 const struct cl_io_slice *ios)
704 struct cl_io *io = ios->cis_io;
705 struct inode *inode = vvp_object_inode(io->ci_obj);
706 struct ll_inode_info *lli = ll_i2info(inode);
708 if (cl_io_is_trunc(io)) {
709 down_write(&lli->lli_trunc_sem);
711 inode_dio_wait(inode);
716 if (io->u.ci_setattr.sa_avalid & TIMES_SET_FLAGS)
717 return vvp_io_setattr_time(env, ios);
722 static void vvp_io_setattr_end(const struct lu_env *env,
723 const struct cl_io_slice *ios)
725 struct cl_io *io = ios->cis_io;
726 struct inode *inode = vvp_object_inode(io->ci_obj);
727 struct ll_inode_info *lli = ll_i2info(inode);
729 if (cl_io_is_trunc(io)) {
730 /* Truncate in memory pages - they must be clean pages
731 * because osc has already notified to destroy osc_extents. */
732 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
733 inode_dio_write_done(inode);
735 up_write(&lli->lli_trunc_sem);
741 static void vvp_io_setattr_fini(const struct lu_env *env,
742 const struct cl_io_slice *ios)
744 bool restore_needed = ios->cis_io->ci_restore_needed;
745 struct inode *inode = vvp_object_inode(ios->cis_obj);
747 vvp_io_fini(env, ios);
749 if (restore_needed && !ios->cis_io->ci_restore_needed) {
750 /* restore finished, set data modified flag for HSM */
751 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
755 static int vvp_io_read_start(const struct lu_env *env,
756 const struct cl_io_slice *ios)
758 struct vvp_io *vio = cl2vvp_io(env, ios);
759 struct cl_io *io = ios->cis_io;
760 struct cl_object *obj = io->ci_obj;
761 struct inode *inode = vvp_object_inode(obj);
762 struct ll_inode_info *lli = ll_i2info(inode);
763 struct file *file = vio->vui_fd->fd_file;
764 struct cl_io_range *range = &io->u.ci_rw.rw_range;
765 loff_t pos = range->cir_pos; /* for generic_file_splice_read() only */
766 size_t tot = vio->vui_tot_count;
771 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
773 CDEBUG(D_VFSTRACE, "%s: read [%llu, %llu)\n",
774 file_dentry(file)->d_name.name,
775 range->cir_pos, range->cir_pos + range->cir_count);
777 if (vio->vui_io_subtype == IO_NORMAL)
778 down_read(&lli->lli_trunc_sem);
780 if (!can_populate_pages(env, io, inode))
783 /* Unless this is reading a sparse file, otherwise the lock has already
784 * been acquired so vvp_prep_size() is an empty op. */
785 result = vvp_prep_size(env, obj, io, range->cir_pos, range->cir_count,
789 else if (exceed != 0)
792 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
793 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
794 inode->i_ino, range->cir_count, range->cir_pos,
797 /* turn off the kernel's read-ahead */
798 vio->vui_fd->fd_file->f_ra.ra_pages = 0;
800 /* initialize read-ahead window once per syscall */
801 if (!vio->vui_ra_valid) {
802 vio->vui_ra_valid = true;
803 vio->vui_ra_start = cl_index(obj, range->cir_pos);
804 vio->vui_ra_count = cl_index(obj, tot + PAGE_SIZE - 1);
810 switch (vio->vui_io_subtype) {
812 LASSERTF(io->u.ci_rw.rw_iocb.ki_pos == range->cir_pos,
813 "ki_pos %lld [%lld, %lld)\n",
814 io->u.ci_rw.rw_iocb.ki_pos,
815 range->cir_pos, range->cir_pos + range->cir_count);
816 result = generic_file_read_iter(&io->u.ci_rw.rw_iocb,
817 &io->u.ci_rw.rw_iter);
820 result = generic_file_splice_read(file, &pos,
821 vio->u.splice.vui_pipe,
823 vio->u.splice.vui_flags);
824 /* LU-1109: do splice read stripe by stripe otherwise if it
825 * may make nfsd stuck if this read occupied all internal pipe
830 CERROR("Wrong IO type %u\n", vio->vui_io_subtype);
837 if (result < range->cir_count)
839 io->ci_nob += result;
840 ll_rw_stats_tally(ll_i2sbi(inode), current->pid, vio->vui_fd,
841 range->cir_pos, result, READ);
848 static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io,
849 struct cl_page_list *plist, int from, int to)
851 struct cl_2queue *queue = &io->ci_queue;
852 struct cl_page *page;
853 unsigned int bytes = 0;
857 if (plist->pl_nr == 0)
860 if (from > 0 || to != PAGE_SIZE) {
861 page = cl_page_list_first(plist);
862 if (plist->pl_nr == 1) {
863 cl_page_clip(env, page, from, to);
866 cl_page_clip(env, page, from, PAGE_SIZE);
867 if (to != PAGE_SIZE) {
868 page = cl_page_list_last(plist);
869 cl_page_clip(env, page, 0, to);
874 cl_2queue_init(queue);
875 cl_page_list_splice(plist, &queue->c2_qin);
876 rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0);
878 /* plist is not sorted any more */
879 cl_page_list_splice(&queue->c2_qin, plist);
880 cl_page_list_splice(&queue->c2_qout, plist);
881 cl_2queue_fini(env, queue);
884 /* calculate bytes */
885 bytes = plist->pl_nr << PAGE_SHIFT;
886 bytes -= from + PAGE_SIZE - to;
888 while (plist->pl_nr > 0) {
889 page = cl_page_list_first(plist);
890 cl_page_list_del(env, plist, page);
892 cl_page_clip(env, page, 0, PAGE_SIZE);
894 SetPageUptodate(cl_page_vmpage(page));
895 cl_page_disown(env, io, page);
897 lu_ref_del(&page->cp_reference, "cl_io", io);
898 cl_page_put(env, page);
902 RETURN(bytes > 0 ? bytes : rc);
905 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
906 struct cl_page *page)
908 struct page *vmpage = page->cp_vmpage;
910 SetPageUptodate(vmpage);
911 set_page_dirty(vmpage);
913 cl_page_disown(env, io, page);
915 lu_ref_del(&page->cp_reference, "cl_io", cl_io_top(io));
916 cl_page_put(env, page);
919 /* make sure the page list is contiguous */
920 static bool page_list_sanity_check(struct cl_object *obj,
921 struct cl_page_list *plist)
923 struct cl_page *page;
924 pgoff_t index = CL_PAGE_EOF;
926 cl_page_list_for_each(page, plist) {
927 struct vvp_page *vpg = cl_object_page_slice(obj, page);
929 if (index == CL_PAGE_EOF) {
930 index = vvp_index(vpg);
935 if (index == vvp_index(vpg))
943 /* Return how many bytes have queued or written */
944 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
946 struct cl_object *obj = io->ci_obj;
947 struct inode *inode = vvp_object_inode(obj);
948 struct vvp_io *vio = vvp_env_io(env);
949 struct cl_page_list *queue = &vio->u.write.vui_queue;
950 struct cl_page *page;
953 unsigned int npages = vio->u.write.vui_queue.pl_nr;
959 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
960 npages, vio->u.write.vui_from, vio->u.write.vui_to);
962 LASSERT(page_list_sanity_check(obj, queue));
964 /* submit IO with async write */
965 rc = cl_io_commit_async(env, io, queue,
966 vio->u.write.vui_from, vio->u.write.vui_to,
967 write_commit_callback);
968 npages -= queue->pl_nr; /* already committed pages */
970 /* calculate how many bytes were written */
971 bytes = npages << PAGE_SHIFT;
974 bytes -= vio->u.write.vui_from;
975 if (queue->pl_nr == 0) /* last page */
976 bytes -= PAGE_SIZE - vio->u.write.vui_to;
977 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
979 vio->u.write.vui_written += bytes;
981 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
982 npages, bytes, vio->u.write.vui_written);
984 /* the first page must have been written. */
985 vio->u.write.vui_from = 0;
987 LASSERT(page_list_sanity_check(obj, queue));
988 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
990 /* out of quota, try sync write */
991 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
992 rc = vvp_io_commit_sync(env, io, queue,
993 vio->u.write.vui_from,
994 vio->u.write.vui_to);
996 vio->u.write.vui_written += rc;
1001 /* update inode size */
1002 ll_merge_attr(env, inode);
1004 /* Now the pages in queue were failed to commit, discard them
1005 * unless they were dirtied before. */
1006 while (queue->pl_nr > 0) {
1007 page = cl_page_list_first(queue);
1008 cl_page_list_del(env, queue, page);
1010 if (!PageDirty(cl_page_vmpage(page)))
1011 cl_page_discard(env, io, page);
1013 cl_page_disown(env, io, page);
1015 lu_ref_del(&page->cp_reference, "cl_io", io);
1016 cl_page_put(env, page);
1018 cl_page_list_fini(env, queue);
1023 static int vvp_io_write_start(const struct lu_env *env,
1024 const struct cl_io_slice *ios)
1026 struct vvp_io *vio = cl2vvp_io(env, ios);
1027 struct cl_io *io = ios->cis_io;
1028 struct cl_object *obj = io->ci_obj;
1029 struct inode *inode = vvp_object_inode(obj);
1030 struct ll_inode_info *lli = ll_i2info(inode);
1031 struct file *file = vio->vui_fd->fd_file;
1032 struct cl_io_range *range = &io->u.ci_rw.rw_range;
1033 bool lock_inode = !lli->lli_inode_locked &&
1038 if (vio->vui_io_subtype == IO_NORMAL)
1039 down_read(&lli->lli_trunc_sem);
1041 if (!can_populate_pages(env, io, inode))
1044 if (cl_io_is_append(io)) {
1046 * PARALLEL IO This has to be changed for parallel IO doing
1047 * out-of-order writes.
1049 ll_merge_attr(env, inode);
1050 range->cir_pos = i_size_read(inode);
1051 io->u.ci_rw.rw_iocb.ki_pos = range->cir_pos;
1053 LASSERTF(io->u.ci_rw.rw_iocb.ki_pos == range->cir_pos,
1054 "ki_pos %lld [%lld, %lld)\n",
1055 io->u.ci_rw.rw_iocb.ki_pos,
1056 range->cir_pos, range->cir_pos + range->cir_count);
1059 CDEBUG(D_VFSTRACE, "%s: write [%llu, %llu)\n",
1060 file_dentry(file)->d_name.name,
1061 range->cir_pos, range->cir_pos + range->cir_count);
1063 /* The maximum Lustre file size is variable, based on the OST maximum
1064 * object size and number of stripes. This needs another check in
1065 * addition to the VFS checks earlier. */
1066 if (range->cir_pos + range->cir_count > ll_file_maxbytes(inode)) {
1068 "%s: file %s ("DFID") offset %llu > maxbytes %llu\n",
1069 ll_get_fsname(inode->i_sb, NULL, 0),
1070 file_dentry(file)->d_name.name,
1071 PFID(ll_inode2fid(inode)),
1072 range->cir_pos + range->cir_count,
1073 ll_file_maxbytes(inode));
1077 /* Tests to verify we take the i_mutex correctly */
1078 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_IMUTEX_SEC) && !lock_inode)
1081 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_IMUTEX_NOSEC) && lock_inode)
1085 * When using the locked AIO function (generic_file_aio_write())
1086 * testing has shown the inode mutex to be a limiting factor
1087 * with multi-threaded single shared file performance. To get
1088 * around this, we now use the lockless version. To maintain
1089 * consistency, proper locking to protect against writes,
1090 * trucates, etc. is handled in the higher layers of lustre.
1094 result = __generic_file_write_iter(&io->u.ci_rw.rw_iocb,
1095 &io->u.ci_rw.rw_iter);
1097 inode_unlock(inode);
1099 if (result > 0 || result == -EIOCBQUEUED)
1100 #ifdef HAVE_GENERIC_WRITE_SYNC_2ARGS
1101 result = generic_write_sync(&io->u.ci_rw.rw_iocb, result);
1106 err = generic_write_sync(io->u.ci_rw.rw_iocb.ki_filp,
1107 range->cir_pos, result);
1108 if (err < 0 && result > 0)
1114 result = vvp_io_write_commit(env, io);
1115 if (vio->u.write.vui_written > 0) {
1116 result = vio->u.write.vui_written;
1117 CDEBUG(D_VFSTRACE, "%s: write nob %zd, result: %zd\n",
1118 file_dentry(file)->d_name.name,
1119 io->ci_nob, result);
1120 io->ci_nob += result;
1124 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
1126 if (result < range->cir_count)
1127 io->ci_continue = 0;
1128 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
1129 vio->vui_fd, range->cir_pos, result, WRITE);
1136 static void vvp_io_rw_end(const struct lu_env *env,
1137 const struct cl_io_slice *ios)
1139 struct vvp_io *vio = cl2vvp_io(env, ios);
1140 struct inode *inode = vvp_object_inode(ios->cis_obj);
1141 struct ll_inode_info *lli = ll_i2info(inode);
1143 if (vio->vui_io_subtype == IO_NORMAL)
1144 up_read(&lli->lli_trunc_sem);
1147 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
1149 struct vm_fault *vmf = cfio->ft_vmf;
1151 cfio->ft_flags = ll_filemap_fault(cfio->ft_vma, vmf);
1152 cfio->ft_flags_valid = 1;
1155 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
1156 get_vmf_address(vmf));
1157 if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) {
1158 lock_page(vmf->page);
1159 cfio->ft_flags |= VM_FAULT_LOCKED;
1162 cfio->ft_vmpage = vmf->page;
1167 if (cfio->ft_flags & VM_FAULT_SIGBUS) {
1168 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", get_vmf_address(vmf));
1172 if (cfio->ft_flags & VM_FAULT_OOM) {
1173 CDEBUG(D_PAGE, "got addr %p - OOM\n", get_vmf_address(vmf));
1177 if (cfio->ft_flags & VM_FAULT_RETRY)
1180 CERROR("unknown error in page fault %d\n", cfio->ft_flags);
1185 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
1186 struct cl_page *page)
1188 set_page_dirty(page->cp_vmpage);
1191 static int vvp_io_fault_start(const struct lu_env *env,
1192 const struct cl_io_slice *ios)
1194 struct vvp_io *vio = cl2vvp_io(env, ios);
1195 struct cl_io *io = ios->cis_io;
1196 struct cl_object *obj = io->ci_obj;
1197 struct inode *inode = vvp_object_inode(obj);
1198 struct ll_inode_info *lli = ll_i2info(inode);
1199 struct cl_fault_io *fio = &io->u.ci_fault;
1200 struct vvp_fault_io *cfio = &vio->u.fault;
1203 struct page *vmpage = NULL;
1204 struct cl_page *page;
1209 down_read(&lli->lli_trunc_sem);
1211 /* offset of the last byte on the page */
1212 offset = cl_offset(obj, fio->ft_index + 1) - 1;
1213 LASSERT(cl_index(obj, offset) == fio->ft_index);
1214 result = vvp_prep_size(env, obj, io, 0, offset + 1, NULL);
1218 /* must return locked page */
1219 if (fio->ft_mkwrite) {
1220 LASSERT(cfio->ft_vmpage != NULL);
1221 lock_page(cfio->ft_vmpage);
1223 result = vvp_io_kernel_fault(cfio);
1228 vmpage = cfio->ft_vmpage;
1229 LASSERT(PageLocked(vmpage));
1231 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
1232 ll_invalidate_page(vmpage);
1234 size = i_size_read(inode);
1235 /* Though we have already held a cl_lock upon this page, but
1236 * it still can be truncated locally. */
1237 if (unlikely((vmpage->mapping != inode->i_mapping) ||
1238 (page_offset(vmpage) > size))) {
1239 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
1241 /* return +1 to stop cl_io_loop() and ll_fault() will catch
1243 GOTO(out, result = +1);
1246 last_index = cl_index(obj, size - 1);
1248 if (fio->ft_mkwrite ) {
1250 * Capture the size while holding the lli_trunc_sem from above
1251 * we want to make sure that we complete the mkwrite action
1252 * while holding this lock. We need to make sure that we are
1253 * not past the end of the file.
1255 if (last_index < fio->ft_index) {
1257 "llite: mkwrite and truncate race happened: "
1258 "%p: 0x%lx 0x%lx\n",
1259 vmpage->mapping,fio->ft_index,last_index);
1261 * We need to return if we are
1262 * passed the end of the file. This will propagate
1263 * up the call stack to ll_page_mkwrite where
1264 * we will return VM_FAULT_NOPAGE. Any non-negative
1265 * value returned here will be silently
1266 * converted to 0. If the vmpage->mapping is null
1267 * the error code would be converted back to ENODATA
1268 * in ll_page_mkwrite0. Thus we return -ENODATA
1269 * to handle both cases
1271 GOTO(out, result = -ENODATA);
1275 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
1277 GOTO(out, result = PTR_ERR(page));
1279 /* if page is going to be written, we should add this page into cache
1281 if (fio->ft_mkwrite) {
1282 wait_on_page_writeback(vmpage);
1283 if (!PageDirty(vmpage)) {
1284 struct cl_page_list *plist = &io->ci_queue.c2_qin;
1285 struct vvp_page *vpg = cl_object_page_slice(obj, page);
1288 /* vvp_page_assume() calls wait_on_page_writeback(). */
1289 cl_page_assume(env, io, page);
1291 cl_page_list_init(plist);
1292 cl_page_list_add(plist, page);
1295 if (last_index == vvp_index(vpg))
1296 to = size & ~PAGE_MASK;
1298 /* Do not set Dirty bit here so that in case IO is
1299 * started before the page is really made dirty, we
1300 * still have chance to detect it. */
1301 result = cl_io_commit_async(env, io, plist, 0, to,
1302 mkwrite_commit_callback);
1303 LASSERT(cl_page_is_owned(page, io));
1304 cl_page_list_fini(env, plist);
1308 cl_page_discard(env, io, page);
1309 cl_page_disown(env, io, page);
1311 cl_page_put(env, page);
1313 /* we're in big trouble, what can we do now? */
1314 if (result == -EDQUOT)
1318 cl_page_disown(env, io, page);
1323 * The ft_index is only used in the case of
1324 * a mkwrite action. We need to check
1325 * our assertions are correct, since
1326 * we should have caught this above
1328 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1329 if (fio->ft_index == last_index)
1331 * Last page is mapped partially.
1333 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
1335 fio->ft_nob = cl_page_size(obj);
1337 lu_ref_add(&page->cp_reference, "fault", io);
1338 fio->ft_page = page;
1342 /* return unlocked vmpage to avoid deadlocking */
1344 unlock_page(vmpage);
1346 cfio->ft_flags &= ~VM_FAULT_LOCKED;
1351 static void vvp_io_fault_end(const struct lu_env *env,
1352 const struct cl_io_slice *ios)
1354 struct inode *inode = vvp_object_inode(ios->cis_obj);
1355 struct ll_inode_info *lli = ll_i2info(inode);
1357 CLOBINVRNT(env, ios->cis_io->ci_obj,
1358 vvp_object_invariant(ios->cis_io->ci_obj));
1359 up_read(&lli->lli_trunc_sem);
1362 static int vvp_io_fsync_start(const struct lu_env *env,
1363 const struct cl_io_slice *ios)
1365 /* we should mark TOWRITE bit to each dirty page in radix tree to
1366 * verify pages have been written, but this is difficult because of
1371 static int vvp_io_read_ahead(const struct lu_env *env,
1372 const struct cl_io_slice *ios,
1373 pgoff_t start, struct cl_read_ahead *ra)
1378 if (ios->cis_io->ci_type == CIT_READ ||
1379 ios->cis_io->ci_type == CIT_FAULT) {
1380 struct vvp_io *vio = cl2vvp_io(env, ios);
1382 if (unlikely(vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1383 ra->cra_end = CL_PAGE_EOF;
1384 result = +1; /* no need to call down */
1391 static const struct cl_io_operations vvp_io_ops = {
1394 .cio_fini = vvp_io_fini,
1395 .cio_lock = vvp_io_read_lock,
1396 .cio_start = vvp_io_read_start,
1397 .cio_end = vvp_io_rw_end,
1398 .cio_advance = vvp_io_advance,
1401 .cio_fini = vvp_io_fini,
1402 .cio_iter_init = vvp_io_write_iter_init,
1403 .cio_iter_fini = vvp_io_write_iter_fini,
1404 .cio_lock = vvp_io_write_lock,
1405 .cio_start = vvp_io_write_start,
1406 .cio_end = vvp_io_rw_end,
1407 .cio_advance = vvp_io_advance,
1410 .cio_fini = vvp_io_setattr_fini,
1411 .cio_iter_init = vvp_io_setattr_iter_init,
1412 .cio_lock = vvp_io_setattr_lock,
1413 .cio_start = vvp_io_setattr_start,
1414 .cio_end = vvp_io_setattr_end
1417 .cio_fini = vvp_io_fault_fini,
1418 .cio_iter_init = vvp_io_fault_iter_init,
1419 .cio_lock = vvp_io_fault_lock,
1420 .cio_start = vvp_io_fault_start,
1421 .cio_end = vvp_io_fault_end,
1424 .cio_start = vvp_io_fsync_start,
1425 .cio_fini = vvp_io_fini
1428 .cio_fini = vvp_io_fini
1431 .cio_fini = vvp_io_fini
1434 .cio_fini = vvp_io_fini
1437 .cio_read_ahead = vvp_io_read_ahead
1440 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1443 struct vvp_io *vio = vvp_env_io(env);
1444 struct inode *inode = vvp_object_inode(obj);
1447 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
1450 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1451 "restore needed %d\n",
1452 PFID(lu_object_fid(&obj->co_lu)),
1453 io->ci_ignore_layout, io->ci_verify_layout,
1454 vio->vui_layout_gen, io->ci_restore_needed);
1456 CL_IO_SLICE_CLEAN(vio, vui_cl);
1457 cl_io_slice_add(io, &vio->vui_cl, obj, &vvp_io_ops);
1458 vio->vui_ra_valid = false;
1460 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1461 struct ll_inode_info *lli = ll_i2info(inode);
1463 vio->vui_tot_count = io->u.ci_rw.rw_range.cir_count;
1464 /* "If nbyte is 0, read() will return 0 and have no other
1465 * results." -- Single Unix Spec */
1466 if (vio->vui_tot_count == 0)
1469 /* for read/write, we store the jobid in the inode, and
1470 * it'll be fetched by osc when building RPC.
1472 * it's not accurate if the file is shared by different
1475 lustre_get_jobid(lli->lli_jobid, sizeof(lli->lli_jobid));
1476 } else if (io->ci_type == CIT_SETATTR) {
1477 if (!cl_io_is_trunc(io))
1478 io->ci_lockreq = CILR_MANDATORY;
1481 /* Enqueue layout lock and get layout version. We need to do this
1482 * even for operations requiring to open file, such as read and write,
1483 * because it might not grant layout lock in IT_OPEN. */
1484 if (result == 0 && !io->ci_ignore_layout) {
1485 result = ll_layout_refresh(inode, &vio->vui_layout_gen);
1486 if (result == -ENOENT)
1487 /* If the inode on MDS has been removed, but the objects
1488 * on OSTs haven't been destroyed (async unlink), layout
1489 * fetch will return -ENOENT, we'd ingore this error
1490 * and continue with dirty flush. LU-3230. */
1493 CERROR("%s: refresh file layout " DFID " error %d.\n",
1494 ll_get_fsname(inode->i_sb, NULL, 0),
1495 PFID(lu_object_fid(&obj->co_lu)), result);
1498 io->ci_result = result < 0 ? result : 0;