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"
44 #include <libcfs/linux/linux-misc.h>
46 static struct vvp_io *cl2vvp_io(const struct lu_env *env,
47 const struct cl_io_slice *slice)
51 vio = container_of(slice, struct vvp_io, vui_cl);
52 LASSERT(vio == vvp_env_io(env));
58 * True, if \a io is a normal io, False for splice_{read,write}
60 static 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->vui_io_subtype == IO_NORMAL;
70 * For swapping layout. The file's layout may have changed.
71 * To avoid populating pages to a wrong stripe, we have to verify the
72 * correctness of layout. It works because swapping layout processes
73 * have to acquire group lock.
75 static bool can_populate_pages(const struct lu_env *env, struct cl_io *io,
78 struct ll_inode_info *lli = ll_i2info(inode);
79 struct vvp_io *vio = vvp_env_io(env);
82 switch (io->ci_type) {
85 /* don't need lock here to check lli_layout_gen as we have held
86 * extent lock and GROUP lock has to hold to swap layout */
87 if (ll_layout_version_get(lli) != vio->vui_layout_gen ||
88 OBD_FAIL_CHECK_RESET(OBD_FAIL_LLITE_LOST_LAYOUT, 0)) {
89 io->ci_need_restart = 1;
90 /* this will cause a short read/write */
95 /* fault is okay because we've already had a page. */
103 static void vvp_object_size_lock(struct cl_object *obj)
105 struct inode *inode = vvp_object_inode(obj);
107 ll_inode_size_lock(inode);
108 cl_object_attr_lock(obj);
111 static void vvp_object_size_unlock(struct cl_object *obj)
113 struct inode *inode = vvp_object_inode(obj);
115 cl_object_attr_unlock(obj);
116 ll_inode_size_unlock(inode);
120 * Helper function that if necessary adjusts file size (inode->i_size), when
121 * position at the offset \a pos is accessed. File size can be arbitrary stale
122 * on a Lustre client, but client at least knows KMS. If accessed area is
123 * inside [0, KMS], set file size to KMS, otherwise glimpse file size.
125 * Locking: i_size_lock is used to serialize changes to inode size and to
126 * protect consistency between inode size and cl_object
127 * attributes. cl_object_size_lock() protects consistency between cl_attr's of
128 * top-object and sub-objects.
130 static int vvp_prep_size(const struct lu_env *env, struct cl_object *obj,
131 struct cl_io *io, loff_t start, size_t count,
134 struct cl_attr *attr = vvp_env_thread_attr(env);
135 struct inode *inode = vvp_object_inode(obj);
136 loff_t pos = start + count - 1;
141 * Consistency guarantees: following possibilities exist for the
142 * relation between region being accessed and real file size at this
145 * (A): the region is completely inside of the file;
147 * (B-x): x bytes of region are inside of the file, the rest is
150 * (C): the region is completely outside of the file.
152 * This classification is stable under DLM lock already acquired by
153 * the caller, because to change the class, other client has to take
154 * DLM lock conflicting with our lock. Also, any updates to ->i_size
155 * by other threads on this client are serialized by
156 * ll_inode_size_lock(). This guarantees that short reads are handled
157 * correctly in the face of concurrent writes and truncates.
159 vvp_object_size_lock(obj);
160 result = cl_object_attr_get(env, obj, attr);
165 * A glimpse is necessary to determine whether we
166 * return a short read (B) or some zeroes at the end
169 vvp_object_size_unlock(obj);
170 result = cl_glimpse_lock(env, io, inode, obj, 0);
171 if (result == 0 && exceed != NULL) {
172 /* If objective page index exceed end-of-file
173 * page index, return directly. Do not expect
174 * kernel will check such case correctly.
175 * linux-2.6.18-128.1.1 miss to do that.
177 loff_t size = i_size_read(inode);
178 unsigned long cur_index = start >>
181 if ((size == 0 && cur_index != 0) ||
182 (((size - 1) >> PAGE_SHIFT) <
190 * region is within kms and, hence, within real file
191 * size (A). We need to increase i_size to cover the
192 * read region so that generic_file_read() will do its
193 * job, but that doesn't mean the kms size is
194 * _correct_, it is only the _minimum_ size. If
195 * someone does a stat they will get the correct size
196 * which will always be >= the kms value here.
199 if (i_size_read(inode) < kms) {
200 i_size_write(inode, kms);
202 DFID" updating i_size %llu\n",
203 PFID(lu_object_fid(&obj->co_lu)),
204 (__u64)i_size_read(inode));
209 vvp_object_size_unlock(obj);
214 /*****************************************************************************
220 static int vvp_io_one_lock_index(const struct lu_env *env, struct cl_io *io,
221 __u32 enqflags, enum cl_lock_mode mode,
222 pgoff_t start, pgoff_t end)
224 struct vvp_io *vio = vvp_env_io(env);
225 struct cl_lock_descr *descr = &vio->vui_link.cill_descr;
226 struct cl_object *obj = io->ci_obj;
228 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
231 CDEBUG(D_VFSTRACE, "lock: %d [%lu, %lu]\n", mode, start, end);
233 memset(&vio->vui_link, 0, sizeof vio->vui_link);
235 if (vio->vui_fd && (vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
236 descr->cld_mode = CLM_GROUP;
237 descr->cld_gid = vio->vui_fd->fd_grouplock.lg_gid;
238 enqflags |= CEF_LOCK_MATCH;
240 descr->cld_mode = mode;
243 descr->cld_obj = obj;
244 descr->cld_start = start;
245 descr->cld_end = end;
246 descr->cld_enq_flags = enqflags;
248 cl_io_lock_add(env, io, &vio->vui_link);
253 static int vvp_io_one_lock(const struct lu_env *env, struct cl_io *io,
254 __u32 enqflags, enum cl_lock_mode mode,
255 loff_t start, loff_t end)
257 struct cl_object *obj = io->ci_obj;
259 return vvp_io_one_lock_index(env, io, enqflags, mode,
260 cl_index(obj, start), cl_index(obj, end));
263 static int vvp_io_write_iter_init(const struct lu_env *env,
264 const struct cl_io_slice *ios)
266 struct vvp_io *vio = cl2vvp_io(env, ios);
268 cl_page_list_init(&vio->u.write.vui_queue);
269 vio->u.write.vui_written = 0;
270 vio->u.write.vui_from = 0;
271 vio->u.write.vui_to = PAGE_SIZE;
276 static void vvp_io_write_iter_fini(const struct lu_env *env,
277 const struct cl_io_slice *ios)
279 struct vvp_io *vio = cl2vvp_io(env, ios);
281 LASSERT(vio->u.write.vui_queue.pl_nr == 0);
284 static int vvp_io_fault_iter_init(const struct lu_env *env,
285 const struct cl_io_slice *ios)
287 struct vvp_io *vio = cl2vvp_io(env, ios);
288 struct inode *inode = vvp_object_inode(ios->cis_obj);
290 LASSERT(inode == file_inode(vio->vui_fd->fd_file));
291 vio->u.fault.ft_mtime = inode->i_mtime.tv_sec;
296 static void vvp_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
298 struct cl_io *io = ios->cis_io;
299 struct cl_object *obj = io->ci_obj;
300 struct vvp_io *vio = cl2vvp_io(env, ios);
301 struct inode *inode = vvp_object_inode(obj);
306 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
308 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
309 "need write layout %d, restore needed %d\n",
310 PFID(lu_object_fid(&obj->co_lu)),
311 io->ci_ignore_layout, io->ci_verify_layout,
312 vio->vui_layout_gen, io->ci_need_write_intent,
313 io->ci_restore_needed);
315 if (io->ci_restore_needed) {
316 /* file was detected release, we need to restore it
317 * before finishing the io
319 rc = ll_layout_restore(inode, 0, OBD_OBJECT_EOF);
320 /* if restore registration failed, no restart,
321 * we will return -ENODATA */
322 /* The layout will change after restore, so we need to
323 * block on layout lock held by the MDT
324 * as MDT will not send new layout in lvb (see LU-3124)
325 * we have to explicitly fetch it, all this will be done
326 * by ll_layout_refresh().
327 * Even if ll_layout_restore() returns zero, it doesn't mean
328 * that restore has been successful. Therefore it sets
329 * ci_verify_layout so that it will check layout at the end
333 io->ci_restore_needed = 1;
334 io->ci_need_restart = 0;
335 io->ci_verify_layout = 0;
340 io->ci_restore_needed = 0;
342 /* Even if ll_layout_restore() returns zero, it doesn't mean
343 * that restore has been successful. Therefore it should verify
344 * if there was layout change and restart I/O correspondingly.
346 ll_layout_refresh(inode, &gen);
347 io->ci_need_restart = vio->vui_layout_gen != gen;
348 if (io->ci_need_restart) {
350 DFID" layout changed from %d to %d.\n",
351 PFID(lu_object_fid(&obj->co_lu)),
352 vio->vui_layout_gen, gen);
353 /* today successful restore is the only possible
355 /* restore was done, clear restoring state */
356 ll_file_clear_flag(ll_i2info(vvp_object_inode(obj)),
357 LLIF_FILE_RESTORING);
363 * dynamic layout change needed, send layout intent
366 if (io->ci_need_write_intent) {
367 enum layout_intent_opc opc = LAYOUT_INTENT_WRITE;
369 io->ci_need_write_intent = 0;
371 LASSERT(io->ci_type == CIT_WRITE ||
372 cl_io_is_trunc(io) || cl_io_is_mkwrite(io));
374 CDEBUG(D_VFSTRACE, DFID" write layout, type %u "DEXT"\n",
375 PFID(lu_object_fid(&obj->co_lu)), io->ci_type,
376 PEXT(&io->ci_write_intent));
378 if (cl_io_is_trunc(io))
379 opc = LAYOUT_INTENT_TRUNC;
381 rc = ll_layout_write_intent(inode, opc, &io->ci_write_intent);
384 io->ci_need_restart = 1;
388 if (!io->ci_need_restart &&
389 !io->ci_ignore_layout && io->ci_verify_layout) {
390 /* check layout version */
391 ll_layout_refresh(inode, &gen);
392 io->ci_need_restart = vio->vui_layout_gen != gen;
393 if (io->ci_need_restart) {
395 DFID" layout changed from %d to %d.\n",
396 PFID(lu_object_fid(&obj->co_lu)),
397 vio->vui_layout_gen, gen);
405 static void vvp_io_fault_fini(const struct lu_env *env,
406 const struct cl_io_slice *ios)
408 struct cl_io *io = ios->cis_io;
409 struct cl_page *page = io->u.ci_fault.ft_page;
411 CLOBINVRNT(env, io->ci_obj, vvp_object_invariant(io->ci_obj));
414 lu_ref_del(&page->cp_reference, "fault", io);
415 cl_page_put(env, page);
416 io->u.ci_fault.ft_page = NULL;
418 vvp_io_fini(env, ios);
421 static enum cl_lock_mode vvp_mode_from_vma(struct vm_area_struct *vma)
424 * we only want to hold PW locks if the mmap() can generate
425 * writes back to the file and that only happens in shared
428 if ((vma->vm_flags & VM_SHARED) && (vma->vm_flags & VM_WRITE))
433 static int vvp_mmap_locks(const struct lu_env *env,
434 struct vvp_io *vio, struct cl_io *io)
436 struct vvp_thread_info *vti = vvp_env_info(env);
437 struct mm_struct *mm = current->mm;
438 struct vm_area_struct *vma;
439 struct cl_lock_descr *descr = &vti->vti_descr;
440 union ldlm_policy_data policy;
448 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
450 if (!cl_is_normalio(env, io))
453 /* nfs or loop back device write */
454 if (vio->vui_iter == NULL)
457 /* No MM (e.g. NFS)? No vmas too. */
461 if (!iter_is_iovec(vio->vui_iter) && !iov_iter_is_kvec(vio->vui_iter))
464 for (i = *vio->vui_iter;
466 iov_iter_advance(&i, iov.iov_len)) {
467 iov = iov_iter_iovec(&i);
468 addr = (unsigned long)iov.iov_base;
474 count += addr & ~PAGE_MASK;
477 down_read(&mm->mmap_sem);
478 while((vma = our_vma(mm, addr, count)) != NULL) {
479 struct dentry *de = file_dentry(vma->vm_file);
480 struct inode *inode = de->d_inode;
481 int flags = CEF_MUST;
483 if (ll_file_nolock(vma->vm_file)) {
485 * For no lock case is not allowed for mmap
492 * XXX: Required lock mode can be weakened: CIT_WRITE
493 * io only ever reads user level buffer, and CIT_READ
496 policy_from_vma(&policy, vma, addr, count);
497 descr->cld_mode = vvp_mode_from_vma(vma);
498 descr->cld_obj = ll_i2info(inode)->lli_clob;
499 descr->cld_start = cl_index(descr->cld_obj,
500 policy.l_extent.start);
501 descr->cld_end = cl_index(descr->cld_obj,
502 policy.l_extent.end);
503 descr->cld_enq_flags = flags;
504 result = cl_io_lock_alloc_add(env, io, descr);
506 CDEBUG(D_VFSTRACE, "lock: %d: [%lu, %lu]\n",
507 descr->cld_mode, descr->cld_start,
513 if (vma->vm_end - addr >= count)
516 count -= vma->vm_end - addr;
519 up_read(&mm->mmap_sem);
526 static void vvp_io_advance(const struct lu_env *env,
527 const struct cl_io_slice *ios,
530 struct vvp_io *vio = cl2vvp_io(env, ios);
531 struct cl_io *io = ios->cis_io;
532 struct cl_object *obj = ios->cis_io->ci_obj;
534 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
536 if (!cl_is_normalio(env, io))
539 vio->vui_tot_count -= nob;
540 iov_iter_reexpand(vio->vui_iter, vio->vui_tot_count);
543 static void vvp_io_update_iov(const struct lu_env *env,
544 struct vvp_io *vio, struct cl_io *io)
546 size_t size = io->u.ci_rw.crw_count;
548 if (!cl_is_normalio(env, io) || vio->vui_iter == NULL)
551 iov_iter_truncate(vio->vui_iter, size);
554 static int vvp_io_rw_lock(const struct lu_env *env, struct cl_io *io,
555 enum cl_lock_mode mode, loff_t start, loff_t end)
557 struct vvp_io *vio = vvp_env_io(env);
561 LASSERT(io->ci_type == CIT_READ || io->ci_type == CIT_WRITE);
564 vvp_io_update_iov(env, vio, io);
566 if (io->u.ci_rw.crw_nonblock)
567 ast_flags |= CEF_NONBLOCK;
568 if (io->ci_lock_no_expand)
569 ast_flags |= CEF_LOCK_NO_EXPAND;
571 result = vvp_mmap_locks(env, vio, io);
573 result = vvp_io_one_lock(env, io, ast_flags, mode, start, end);
578 static int vvp_io_read_lock(const struct lu_env *env,
579 const struct cl_io_slice *ios)
581 struct cl_io *io = ios->cis_io;
582 struct cl_io_rw_common *rd = &io->u.ci_rd.rd;
586 result = vvp_io_rw_lock(env, io, CLM_READ, rd->crw_pos,
587 rd->crw_pos + rd->crw_count - 1);
591 static int vvp_io_fault_lock(const struct lu_env *env,
592 const struct cl_io_slice *ios)
594 struct cl_io *io = ios->cis_io;
595 struct vvp_io *vio = cl2vvp_io(env, ios);
597 * XXX LDLM_FL_CBPENDING
599 return vvp_io_one_lock_index(env,
601 vvp_mode_from_vma(vio->u.fault.ft_vma),
602 io->u.ci_fault.ft_index,
603 io->u.ci_fault.ft_index);
606 static int vvp_io_write_lock(const struct lu_env *env,
607 const struct cl_io_slice *ios)
609 struct cl_io *io = ios->cis_io;
613 if (io->u.ci_wr.wr_append) {
615 end = OBD_OBJECT_EOF;
617 start = io->u.ci_wr.wr.crw_pos;
618 end = start + io->u.ci_wr.wr.crw_count - 1;
621 RETURN(vvp_io_rw_lock(env, io, CLM_WRITE, start, end));
624 static int vvp_io_setattr_iter_init(const struct lu_env *env,
625 const struct cl_io_slice *ios)
632 * Implementation of cl_io_operations::cio_lock() method for CIT_SETATTR io.
634 * Handles "lockless io" mode when extent locking is done by server.
636 static int vvp_io_setattr_lock(const struct lu_env *env,
637 const struct cl_io_slice *ios)
639 struct cl_io *io = ios->cis_io;
643 if (cl_io_is_trunc(io)) {
644 new_size = io->u.ci_setattr.sa_attr.lvb_size;
646 enqflags = CEF_DISCARD_DATA;
648 unsigned int valid = io->u.ci_setattr.sa_avalid;
650 if (!(valid & TIMES_SET_FLAGS))
653 if ((!(valid & ATTR_MTIME) ||
654 io->u.ci_setattr.sa_attr.lvb_mtime >=
655 io->u.ci_setattr.sa_attr.lvb_ctime) &&
656 (!(valid & ATTR_ATIME) ||
657 io->u.ci_setattr.sa_attr.lvb_atime >=
658 io->u.ci_setattr.sa_attr.lvb_ctime))
664 return vvp_io_one_lock(env, io, enqflags, CLM_WRITE,
665 new_size, OBD_OBJECT_EOF);
668 static int vvp_do_vmtruncate(struct inode *inode, size_t size)
673 * Only ll_inode_size_lock is taken at this level.
675 ll_inode_size_lock(inode);
676 result = inode_newsize_ok(inode, size);
678 ll_inode_size_unlock(inode);
681 i_size_write(inode, size);
683 ll_truncate_pagecache(inode, size);
684 ll_inode_size_unlock(inode);
688 static int vvp_io_setattr_time(const struct lu_env *env,
689 const struct cl_io_slice *ios)
691 struct cl_io *io = ios->cis_io;
692 struct cl_object *obj = io->ci_obj;
693 struct cl_attr *attr = vvp_env_thread_attr(env);
695 unsigned valid = CAT_CTIME;
697 cl_object_attr_lock(obj);
698 attr->cat_ctime = io->u.ci_setattr.sa_attr.lvb_ctime;
699 if (io->u.ci_setattr.sa_avalid & ATTR_ATIME_SET) {
700 attr->cat_atime = io->u.ci_setattr.sa_attr.lvb_atime;
703 if (io->u.ci_setattr.sa_avalid & ATTR_MTIME_SET) {
704 attr->cat_mtime = io->u.ci_setattr.sa_attr.lvb_mtime;
707 result = cl_object_attr_update(env, obj, attr, valid);
708 cl_object_attr_unlock(obj);
713 static int vvp_io_setattr_start(const struct lu_env *env,
714 const struct cl_io_slice *ios)
716 struct cl_io *io = ios->cis_io;
717 struct inode *inode = vvp_object_inode(io->ci_obj);
718 struct ll_inode_info *lli = ll_i2info(inode);
720 if (cl_io_is_trunc(io)) {
721 down_write(&lli->lli_trunc_sem);
723 inode_dio_wait(inode);
728 if (io->u.ci_setattr.sa_avalid & TIMES_SET_FLAGS)
729 return vvp_io_setattr_time(env, ios);
734 static void vvp_io_setattr_end(const struct lu_env *env,
735 const struct cl_io_slice *ios)
737 struct cl_io *io = ios->cis_io;
738 struct inode *inode = vvp_object_inode(io->ci_obj);
739 struct ll_inode_info *lli = ll_i2info(inode);
741 if (cl_io_is_trunc(io)) {
742 /* Truncate in memory pages - they must be clean pages
743 * because osc has already notified to destroy osc_extents. */
744 vvp_do_vmtruncate(inode, io->u.ci_setattr.sa_attr.lvb_size);
745 inode_dio_write_done(inode);
747 up_write(&lli->lli_trunc_sem);
753 static void vvp_io_setattr_fini(const struct lu_env *env,
754 const struct cl_io_slice *ios)
756 bool restore_needed = ios->cis_io->ci_restore_needed;
757 struct inode *inode = vvp_object_inode(ios->cis_obj);
759 vvp_io_fini(env, ios);
761 if (restore_needed && !ios->cis_io->ci_restore_needed) {
762 /* restore finished, set data modified flag for HSM */
763 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
767 static int vvp_io_read_start(const struct lu_env *env,
768 const struct cl_io_slice *ios)
770 struct vvp_io *vio = cl2vvp_io(env, ios);
771 struct cl_io *io = ios->cis_io;
772 struct cl_object *obj = io->ci_obj;
773 struct inode *inode = vvp_object_inode(obj);
774 struct ll_inode_info *lli = ll_i2info(inode);
775 struct file *file = vio->vui_fd->fd_file;
776 loff_t pos = io->u.ci_rd.rd.crw_pos;
777 long cnt = io->u.ci_rd.rd.crw_count;
778 long tot = vio->vui_tot_count;
783 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
785 CDEBUG(D_VFSTRACE, "%s: read [%llu, %llu)\n",
786 file_dentry(file)->d_name.name,
789 if (vio->vui_io_subtype == IO_NORMAL)
790 down_read(&lli->lli_trunc_sem);
792 if (io->ci_async_readahead) {
797 if (!can_populate_pages(env, io, inode))
800 /* Unless this is reading a sparse file, otherwise the lock has already
801 * been acquired so vvp_prep_size() is an empty op. */
802 result = vvp_prep_size(env, obj, io, pos, cnt, &exceed);
805 else if (exceed != 0)
808 LU_OBJECT_HEADER(D_INODE, env, &obj->co_lu,
809 "Read ino %lu, %lu bytes, offset %lld, size %llu\n",
810 inode->i_ino, cnt, pos, i_size_read(inode));
812 /* turn off the kernel's read-ahead */
813 vio->vui_fd->fd_file->f_ra.ra_pages = 0;
815 /* initialize read-ahead window once per syscall */
816 if (!vio->vui_ra_valid) {
817 vio->vui_ra_valid = true;
818 vio->vui_ra_start = cl_index(obj, pos);
819 vio->vui_ra_count = cl_index(obj, tot + PAGE_SIZE - 1);
824 switch (vio->vui_io_subtype) {
826 LASSERT(vio->vui_iocb->ki_pos == pos);
827 result = generic_file_read_iter(vio->vui_iocb, vio->vui_iter);
830 result = generic_file_splice_read(file, &pos,
831 vio->u.splice.vui_pipe, cnt,
832 vio->u.splice.vui_flags);
833 /* LU-1109: do splice read stripe by stripe otherwise if it
834 * may make nfsd stuck if this read occupied all internal pipe
839 CERROR("Wrong IO type %u\n", vio->vui_io_subtype);
848 io->ci_nob += result;
849 ll_rw_stats_tally(ll_i2sbi(inode), current->pid, vio->vui_fd,
857 static int vvp_io_commit_sync(const struct lu_env *env, struct cl_io *io,
858 struct cl_page_list *plist, int from, int to)
860 struct cl_2queue *queue = &io->ci_queue;
861 struct cl_page *page;
862 unsigned int bytes = 0;
866 if (plist->pl_nr == 0)
869 if (from > 0 || to != PAGE_SIZE) {
870 page = cl_page_list_first(plist);
871 if (plist->pl_nr == 1) {
872 cl_page_clip(env, page, from, to);
875 cl_page_clip(env, page, from, PAGE_SIZE);
876 if (to != PAGE_SIZE) {
877 page = cl_page_list_last(plist);
878 cl_page_clip(env, page, 0, to);
883 cl_2queue_init(queue);
884 cl_page_list_splice(plist, &queue->c2_qin);
885 rc = cl_io_submit_sync(env, io, CRT_WRITE, queue, 0);
887 /* plist is not sorted any more */
888 cl_page_list_splice(&queue->c2_qin, plist);
889 cl_page_list_splice(&queue->c2_qout, plist);
890 cl_2queue_fini(env, queue);
893 /* calculate bytes */
894 bytes = plist->pl_nr << PAGE_SHIFT;
895 bytes -= from + PAGE_SIZE - to;
897 while (plist->pl_nr > 0) {
898 page = cl_page_list_first(plist);
899 cl_page_list_del(env, plist, page);
901 cl_page_clip(env, page, 0, PAGE_SIZE);
903 SetPageUptodate(cl_page_vmpage(page));
904 cl_page_disown(env, io, page);
906 /* held in ll_cl_init() */
907 lu_ref_del(&page->cp_reference, "cl_io", io);
908 cl_page_put(env, page);
912 RETURN(bytes > 0 ? bytes : rc);
915 static void write_commit_callback(const struct lu_env *env, struct cl_io *io,
916 struct cl_page *page)
918 struct page *vmpage = page->cp_vmpage;
920 SetPageUptodate(vmpage);
921 set_page_dirty(vmpage);
923 cl_page_disown(env, io, page);
925 /* held in ll_cl_init() */
926 lu_ref_del(&page->cp_reference, "cl_io", cl_io_top(io));
927 cl_page_put(env, page);
930 /* make sure the page list is contiguous */
931 static bool page_list_sanity_check(struct cl_object *obj,
932 struct cl_page_list *plist)
934 struct cl_page *page;
935 pgoff_t index = CL_PAGE_EOF;
937 cl_page_list_for_each(page, plist) {
938 struct vvp_page *vpg = cl_object_page_slice(obj, page);
940 if (index == CL_PAGE_EOF) {
941 index = vvp_index(vpg);
946 if (index == vvp_index(vpg))
954 /* Return how many bytes have queued or written */
955 int vvp_io_write_commit(const struct lu_env *env, struct cl_io *io)
957 struct cl_object *obj = io->ci_obj;
958 struct inode *inode = vvp_object_inode(obj);
959 struct vvp_io *vio = vvp_env_io(env);
960 struct cl_page_list *queue = &vio->u.write.vui_queue;
961 struct cl_page *page;
964 unsigned int npages = vio->u.write.vui_queue.pl_nr;
970 CDEBUG(D_VFSTRACE, "commit async pages: %d, from %d, to %d\n",
971 npages, vio->u.write.vui_from, vio->u.write.vui_to);
973 LASSERT(page_list_sanity_check(obj, queue));
975 /* submit IO with async write */
976 rc = cl_io_commit_async(env, io, queue,
977 vio->u.write.vui_from, vio->u.write.vui_to,
978 write_commit_callback);
979 npages -= queue->pl_nr; /* already committed pages */
981 /* calculate how many bytes were written */
982 bytes = npages << PAGE_SHIFT;
985 bytes -= vio->u.write.vui_from;
986 if (queue->pl_nr == 0) /* last page */
987 bytes -= PAGE_SIZE - vio->u.write.vui_to;
988 LASSERTF(bytes > 0, "bytes = %d, pages = %d\n", bytes, npages);
990 vio->u.write.vui_written += bytes;
992 CDEBUG(D_VFSTRACE, "Committed %d pages %d bytes, tot: %ld\n",
993 npages, bytes, vio->u.write.vui_written);
995 /* the first page must have been written. */
996 vio->u.write.vui_from = 0;
998 LASSERT(page_list_sanity_check(obj, queue));
999 LASSERT(ergo(rc == 0, queue->pl_nr == 0));
1001 /* out of quota, try sync write */
1002 if (rc == -EDQUOT && !cl_io_is_mkwrite(io)) {
1003 rc = vvp_io_commit_sync(env, io, queue,
1004 vio->u.write.vui_from,
1005 vio->u.write.vui_to);
1007 vio->u.write.vui_written += rc;
1012 /* update inode size */
1013 ll_merge_attr(env, inode);
1015 /* Now the pages in queue were failed to commit, discard them
1016 * unless they were dirtied before. */
1017 while (queue->pl_nr > 0) {
1018 page = cl_page_list_first(queue);
1019 cl_page_list_del(env, queue, page);
1021 if (!PageDirty(cl_page_vmpage(page)))
1022 cl_page_discard(env, io, page);
1024 cl_page_disown(env, io, page);
1026 /* held in ll_cl_init() */
1027 lu_ref_del(&page->cp_reference, "cl_io", io);
1028 cl_page_put(env, page);
1030 cl_page_list_fini(env, queue);
1035 static int vvp_io_write_start(const struct lu_env *env,
1036 const struct cl_io_slice *ios)
1038 struct vvp_io *vio = cl2vvp_io(env, ios);
1039 struct cl_io *io = ios->cis_io;
1040 struct cl_object *obj = io->ci_obj;
1041 struct inode *inode = vvp_object_inode(obj);
1042 struct ll_inode_info *lli = ll_i2info(inode);
1043 struct file *file = vio->vui_fd->fd_file;
1045 loff_t pos = io->u.ci_wr.wr.crw_pos;
1046 size_t cnt = io->u.ci_wr.wr.crw_count;
1047 bool lock_inode = !IS_NOSEC(inode);
1051 if (vio->vui_io_subtype == IO_NORMAL)
1052 down_read(&lli->lli_trunc_sem);
1054 if (!can_populate_pages(env, io, inode))
1057 if (cl_io_is_append(io)) {
1059 * PARALLEL IO This has to be changed for parallel IO doing
1060 * out-of-order writes.
1062 ll_merge_attr(env, inode);
1063 pos = io->u.ci_wr.wr.crw_pos = i_size_read(inode);
1064 vio->vui_iocb->ki_pos = pos;
1066 LASSERTF(vio->vui_iocb->ki_pos == pos,
1067 "ki_pos %lld [%lld, %lld)\n",
1068 vio->vui_iocb->ki_pos,
1072 CDEBUG(D_VFSTRACE, "%s: write [%llu, %llu)\n",
1073 file_dentry(file)->d_name.name,
1076 /* The maximum Lustre file size is variable, based on the OST maximum
1077 * object size and number of stripes. This needs another check in
1078 * addition to the VFS checks earlier. */
1079 if (pos + cnt > ll_file_maxbytes(inode)) {
1081 "%s: file %s ("DFID") offset %llu > maxbytes %llu\n",
1082 ll_i2sbi(inode)->ll_fsname,
1083 file_dentry(file)->d_name.name,
1084 PFID(ll_inode2fid(inode)), pos + cnt,
1085 ll_file_maxbytes(inode));
1089 /* Tests to verify we take the i_mutex correctly */
1090 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_IMUTEX_SEC) && !lock_inode)
1093 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_IMUTEX_NOSEC) && lock_inode)
1096 if (vio->vui_iter == NULL) {
1097 /* from a temp io in ll_cl_init(). */
1101 * When using the locked AIO function (generic_file_aio_write())
1102 * testing has shown the inode mutex to be a limiting factor
1103 * with multi-threaded single shared file performance. To get
1104 * around this, we now use the lockless version. To maintain
1105 * consistency, proper locking to protect against writes,
1106 * trucates, etc. is handled in the higher layers of lustre.
1108 lock_inode = !IS_NOSEC(inode);
1110 if (unlikely(lock_inode))
1112 result = __generic_file_write_iter(vio->vui_iocb,
1114 if (unlikely(lock_inode))
1115 inode_unlock(inode);
1117 if (result > 0 || result == -EIOCBQUEUED)
1118 #ifdef HAVE_GENERIC_WRITE_SYNC_2ARGS
1119 result = generic_write_sync(vio->vui_iocb, result);
1124 err = generic_write_sync(vio->vui_iocb->ki_filp, pos,
1126 if (err < 0 && result > 0)
1133 result = vvp_io_write_commit(env, io);
1134 if (vio->u.write.vui_written > 0) {
1135 result = vio->u.write.vui_written;
1136 CDEBUG(D_VFSTRACE, "%s: write nob %zd, result: %zd\n",
1137 file_dentry(file)->d_name.name,
1138 io->ci_nob, result);
1139 io->ci_nob += result;
1143 ll_file_set_flag(ll_i2info(inode), LLIF_DATA_MODIFIED);
1146 io->ci_continue = 0;
1147 ll_rw_stats_tally(ll_i2sbi(inode), current->pid,
1148 vio->vui_fd, pos, result, WRITE);
1155 static void vvp_io_rw_end(const struct lu_env *env,
1156 const struct cl_io_slice *ios)
1158 struct vvp_io *vio = cl2vvp_io(env, ios);
1159 struct inode *inode = vvp_object_inode(ios->cis_obj);
1160 struct ll_inode_info *lli = ll_i2info(inode);
1162 if (vio->vui_io_subtype == IO_NORMAL)
1163 up_read(&lli->lli_trunc_sem);
1166 static int vvp_io_kernel_fault(struct vvp_fault_io *cfio)
1168 struct vm_fault *vmf = cfio->ft_vmf;
1170 cfio->ft_flags = ll_filemap_fault(cfio->ft_vma, vmf);
1171 cfio->ft_flags_valid = 1;
1174 LL_CDEBUG_PAGE(D_PAGE, vmf->page, "got addr %p type NOPAGE\n",
1175 get_vmf_address(vmf));
1176 if (unlikely(!(cfio->ft_flags & VM_FAULT_LOCKED))) {
1177 lock_page(vmf->page);
1178 cfio->ft_flags |= VM_FAULT_LOCKED;
1181 cfio->ft_vmpage = vmf->page;
1186 if (cfio->ft_flags & VM_FAULT_SIGBUS) {
1187 CDEBUG(D_PAGE, "got addr %p - SIGBUS\n", get_vmf_address(vmf));
1191 if (cfio->ft_flags & VM_FAULT_OOM) {
1192 CDEBUG(D_PAGE, "got addr %p - OOM\n", get_vmf_address(vmf));
1196 if (cfio->ft_flags & VM_FAULT_RETRY)
1199 CERROR("unknown error in page fault %d\n", cfio->ft_flags);
1204 static void mkwrite_commit_callback(const struct lu_env *env, struct cl_io *io,
1205 struct cl_page *page)
1207 set_page_dirty(page->cp_vmpage);
1210 static int vvp_io_fault_start(const struct lu_env *env,
1211 const struct cl_io_slice *ios)
1213 struct vvp_io *vio = cl2vvp_io(env, ios);
1214 struct cl_io *io = ios->cis_io;
1215 struct cl_object *obj = io->ci_obj;
1216 struct inode *inode = vvp_object_inode(obj);
1217 struct ll_inode_info *lli = ll_i2info(inode);
1218 struct cl_fault_io *fio = &io->u.ci_fault;
1219 struct vvp_fault_io *cfio = &vio->u.fault;
1222 struct page *vmpage = NULL;
1223 struct cl_page *page;
1228 down_read(&lli->lli_trunc_sem);
1230 /* offset of the last byte on the page */
1231 offset = cl_offset(obj, fio->ft_index + 1) - 1;
1232 LASSERT(cl_index(obj, offset) == fio->ft_index);
1233 result = vvp_prep_size(env, obj, io, 0, offset + 1, NULL);
1237 /* must return locked page */
1238 if (fio->ft_mkwrite) {
1239 LASSERT(cfio->ft_vmpage != NULL);
1240 lock_page(cfio->ft_vmpage);
1242 result = vvp_io_kernel_fault(cfio);
1247 vmpage = cfio->ft_vmpage;
1248 LASSERT(PageLocked(vmpage));
1250 if (OBD_FAIL_CHECK(OBD_FAIL_LLITE_FAULT_TRUNC_RACE))
1251 generic_error_remove_page(vmpage->mapping, vmpage);
1253 size = i_size_read(inode);
1254 /* Though we have already held a cl_lock upon this page, but
1255 * it still can be truncated locally. */
1256 if (unlikely((vmpage->mapping != inode->i_mapping) ||
1257 (page_offset(vmpage) > size))) {
1258 CDEBUG(D_PAGE, "llite: fault and truncate race happened!\n");
1260 /* return +1 to stop cl_io_loop() and ll_fault() will catch
1262 GOTO(out, result = +1);
1265 last_index = cl_index(obj, size - 1);
1267 if (fio->ft_mkwrite ) {
1269 * Capture the size while holding the lli_trunc_sem from above
1270 * we want to make sure that we complete the mkwrite action
1271 * while holding this lock. We need to make sure that we are
1272 * not past the end of the file.
1274 if (last_index < fio->ft_index) {
1276 "llite: mkwrite and truncate race happened: "
1277 "%p: 0x%lx 0x%lx\n",
1278 vmpage->mapping,fio->ft_index,last_index);
1280 * We need to return if we are
1281 * passed the end of the file. This will propagate
1282 * up the call stack to ll_page_mkwrite where
1283 * we will return VM_FAULT_NOPAGE. Any non-negative
1284 * value returned here will be silently
1285 * converted to 0. If the vmpage->mapping is null
1286 * the error code would be converted back to ENODATA
1287 * in ll_page_mkwrite0. Thus we return -ENODATA
1288 * to handle both cases
1290 GOTO(out, result = -ENODATA);
1294 page = cl_page_find(env, obj, fio->ft_index, vmpage, CPT_CACHEABLE);
1296 GOTO(out, result = PTR_ERR(page));
1298 /* if page is going to be written, we should add this page into cache
1300 if (fio->ft_mkwrite) {
1301 wait_on_page_writeback(vmpage);
1302 if (!PageDirty(vmpage)) {
1303 struct cl_page_list *plist = &io->ci_queue.c2_qin;
1304 struct vvp_page *vpg = cl_object_page_slice(obj, page);
1307 /* vvp_page_assume() calls wait_on_page_writeback(). */
1308 cl_page_assume(env, io, page);
1310 cl_page_list_init(plist);
1311 cl_page_list_add(plist, page);
1314 if (last_index == vvp_index(vpg))
1315 to = size & ~PAGE_MASK;
1317 /* Do not set Dirty bit here so that in case IO is
1318 * started before the page is really made dirty, we
1319 * still have chance to detect it. */
1320 result = cl_io_commit_async(env, io, plist, 0, to,
1321 mkwrite_commit_callback);
1322 LASSERT(cl_page_is_owned(page, io));
1323 cl_page_list_fini(env, plist);
1327 cl_page_discard(env, io, page);
1328 cl_page_disown(env, io, page);
1330 cl_page_put(env, page);
1332 /* we're in big trouble, what can we do now? */
1333 if (result == -EDQUOT)
1337 cl_page_disown(env, io, page);
1342 * The ft_index is only used in the case of
1343 * a mkwrite action. We need to check
1344 * our assertions are correct, since
1345 * we should have caught this above
1347 LASSERT(!fio->ft_mkwrite || fio->ft_index <= last_index);
1348 if (fio->ft_index == last_index)
1350 * Last page is mapped partially.
1352 fio->ft_nob = size - cl_offset(obj, fio->ft_index);
1354 fio->ft_nob = cl_page_size(obj);
1356 lu_ref_add(&page->cp_reference, "fault", io);
1357 fio->ft_page = page;
1361 /* return unlocked vmpage to avoid deadlocking */
1363 unlock_page(vmpage);
1365 cfio->ft_flags &= ~VM_FAULT_LOCKED;
1370 static void vvp_io_fault_end(const struct lu_env *env,
1371 const struct cl_io_slice *ios)
1373 struct inode *inode = vvp_object_inode(ios->cis_obj);
1374 struct ll_inode_info *lli = ll_i2info(inode);
1376 CLOBINVRNT(env, ios->cis_io->ci_obj,
1377 vvp_object_invariant(ios->cis_io->ci_obj));
1378 up_read(&lli->lli_trunc_sem);
1381 static int vvp_io_fsync_start(const struct lu_env *env,
1382 const struct cl_io_slice *ios)
1384 /* we should mark TOWRITE bit to each dirty page in radix tree to
1385 * verify pages have been written, but this is difficult because of
1390 static int vvp_io_read_ahead(const struct lu_env *env,
1391 const struct cl_io_slice *ios,
1392 pgoff_t start, struct cl_read_ahead *ra)
1397 if (ios->cis_io->ci_type == CIT_READ ||
1398 ios->cis_io->ci_type == CIT_FAULT) {
1399 struct vvp_io *vio = cl2vvp_io(env, ios);
1401 if (unlikely(vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
1402 ra->cra_end = CL_PAGE_EOF;
1403 result = +1; /* no need to call down */
1410 static const struct cl_io_operations vvp_io_ops = {
1413 .cio_fini = vvp_io_fini,
1414 .cio_lock = vvp_io_read_lock,
1415 .cio_start = vvp_io_read_start,
1416 .cio_end = vvp_io_rw_end,
1417 .cio_advance = vvp_io_advance,
1420 .cio_fini = vvp_io_fini,
1421 .cio_iter_init = vvp_io_write_iter_init,
1422 .cio_iter_fini = vvp_io_write_iter_fini,
1423 .cio_lock = vvp_io_write_lock,
1424 .cio_start = vvp_io_write_start,
1425 .cio_end = vvp_io_rw_end,
1426 .cio_advance = vvp_io_advance,
1429 .cio_fini = vvp_io_setattr_fini,
1430 .cio_iter_init = vvp_io_setattr_iter_init,
1431 .cio_lock = vvp_io_setattr_lock,
1432 .cio_start = vvp_io_setattr_start,
1433 .cio_end = vvp_io_setattr_end
1436 .cio_fini = vvp_io_fault_fini,
1437 .cio_iter_init = vvp_io_fault_iter_init,
1438 .cio_lock = vvp_io_fault_lock,
1439 .cio_start = vvp_io_fault_start,
1440 .cio_end = vvp_io_fault_end,
1443 .cio_start = vvp_io_fsync_start,
1444 .cio_fini = vvp_io_fini
1447 .cio_fini = vvp_io_fini
1450 .cio_fini = vvp_io_fini
1453 .cio_fini = vvp_io_fini
1456 .cio_read_ahead = vvp_io_read_ahead
1459 int vvp_io_init(const struct lu_env *env, struct cl_object *obj,
1462 struct vvp_io *vio = vvp_env_io(env);
1463 struct inode *inode = vvp_object_inode(obj);
1466 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
1469 CDEBUG(D_VFSTRACE, DFID" ignore/verify layout %d/%d, layout version %d "
1470 "restore needed %d\n",
1471 PFID(lu_object_fid(&obj->co_lu)),
1472 io->ci_ignore_layout, io->ci_verify_layout,
1473 vio->vui_layout_gen, io->ci_restore_needed);
1475 CL_IO_SLICE_CLEAN(vio, vui_cl);
1476 cl_io_slice_add(io, &vio->vui_cl, obj, &vvp_io_ops);
1477 vio->vui_ra_valid = false;
1479 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE) {
1481 struct ll_inode_info *lli = ll_i2info(inode);
1483 count = io->u.ci_rw.crw_count;
1484 /* "If nbyte is 0, read() will return 0 and have no other
1485 * results." -- Single Unix Spec */
1489 vio->vui_tot_count = count;
1491 /* for read/write, we store the jobid in the inode, and
1492 * it'll be fetched by osc when building RPC.
1494 * it's not accurate if the file is shared by different
1497 lustre_get_jobid(lli->lli_jobid, sizeof(lli->lli_jobid));
1498 } else if (io->ci_type == CIT_SETATTR) {
1499 if (!cl_io_is_trunc(io))
1500 io->ci_lockreq = CILR_MANDATORY;
1503 /* Enqueue layout lock and get layout version. We need to do this
1504 * even for operations requiring to open file, such as read and write,
1505 * because it might not grant layout lock in IT_OPEN. */
1506 if (result == 0 && !io->ci_ignore_layout) {
1507 result = ll_layout_refresh(inode, &vio->vui_layout_gen);
1508 if (result == -ENOENT)
1509 /* If the inode on MDS has been removed, but the objects
1510 * on OSTs haven't been destroyed (async unlink), layout
1511 * fetch will return -ENOENT, we'd ingore this error
1512 * and continue with dirty flush. LU-3230. */
1515 CERROR("%s: refresh file layout " DFID " error %d.\n",
1516 ll_i2sbi(inode)->ll_fsname,
1517 PFID(lu_object_fid(&obj->co_lu)), result);
1520 io->ci_result = result < 0 ? result : 0;