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, 2013, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * cl code shared between vvp and liblustre (and other Lustre clients in the
39 * Author: Nikita Danilov <nikita.danilov@sun.com>
42 #define DEBUG_SUBSYSTEM S_LLITE
44 #include <libcfs/libcfs.h>
46 #include <linux/sched.h>
48 #include <linux/quotaops.h>
49 #include <linux/highmem.h>
50 #include <linux/pagemap.h>
51 #include <linux/rbtree.h>
54 #include <obd_support.h>
55 #include <lustre_fid.h>
56 #include <lustre_lite.h>
57 #include <lustre_dlm.h>
58 #include <lustre_ver.h>
59 #include <lustre_mdc.h>
60 #include <cl_object.h>
64 #include "llite_internal.h"
66 static const struct cl_req_operations ccc_req_ops;
69 * ccc_ prefix stands for "Common Client Code".
72 static struct kmem_cache *ccc_lock_kmem;
73 static struct kmem_cache *ccc_object_kmem;
74 static struct kmem_cache *ccc_thread_kmem;
75 static struct kmem_cache *ccc_session_kmem;
76 static struct kmem_cache *ccc_req_kmem;
78 static struct lu_kmem_descr ccc_caches[] = {
80 .ckd_cache = &ccc_lock_kmem,
81 .ckd_name = "ccc_lock_kmem",
82 .ckd_size = sizeof (struct ccc_lock)
85 .ckd_cache = &ccc_object_kmem,
86 .ckd_name = "ccc_object_kmem",
87 .ckd_size = sizeof (struct ccc_object)
90 .ckd_cache = &ccc_thread_kmem,
91 .ckd_name = "ccc_thread_kmem",
92 .ckd_size = sizeof (struct ccc_thread_info),
95 .ckd_cache = &ccc_session_kmem,
96 .ckd_name = "ccc_session_kmem",
97 .ckd_size = sizeof (struct ccc_session)
100 .ckd_cache = &ccc_req_kmem,
101 .ckd_name = "ccc_req_kmem",
102 .ckd_size = sizeof (struct ccc_req)
109 /*****************************************************************************
111 * Vvp device and device type functions.
115 void *ccc_key_init(const struct lu_context *ctx, struct lu_context_key *key)
117 struct ccc_thread_info *info;
119 OBD_SLAB_ALLOC_PTR_GFP(info, ccc_thread_kmem, GFP_NOFS);
121 info = ERR_PTR(-ENOMEM);
125 void ccc_key_fini(const struct lu_context *ctx,
126 struct lu_context_key *key, void *data)
128 struct ccc_thread_info *info = data;
129 OBD_SLAB_FREE_PTR(info, ccc_thread_kmem);
132 void *ccc_session_key_init(const struct lu_context *ctx,
133 struct lu_context_key *key)
135 struct ccc_session *session;
137 OBD_SLAB_ALLOC_PTR_GFP(session, ccc_session_kmem, GFP_NOFS);
139 session = ERR_PTR(-ENOMEM);
143 void ccc_session_key_fini(const struct lu_context *ctx,
144 struct lu_context_key *key, void *data)
146 struct ccc_session *session = data;
147 OBD_SLAB_FREE_PTR(session, ccc_session_kmem);
150 struct lu_context_key ccc_key = {
151 .lct_tags = LCT_CL_THREAD,
152 .lct_init = ccc_key_init,
153 .lct_fini = ccc_key_fini
156 struct lu_context_key ccc_session_key = {
157 .lct_tags = LCT_SESSION,
158 .lct_init = ccc_session_key_init,
159 .lct_fini = ccc_session_key_fini
163 /* type constructor/destructor: ccc_type_{init,fini,start,stop}(). */
164 // LU_TYPE_INIT_FINI(ccc, &ccc_key, &ccc_session_key);
166 int ccc_device_init(const struct lu_env *env, struct lu_device *d,
167 const char *name, struct lu_device *next)
169 struct ccc_device *vdv;
174 vdv->cdv_next = lu2cl_dev(next);
176 LASSERT(d->ld_site != NULL && next->ld_type != NULL);
177 next->ld_site = d->ld_site;
178 rc = next->ld_type->ldt_ops->ldto_device_init(
179 env, next, next->ld_type->ldt_name, NULL);
182 lu_ref_add(&next->ld_reference, "lu-stack", &lu_site_init);
187 struct lu_device *ccc_device_fini(const struct lu_env *env,
190 return cl2lu_dev(lu2ccc_dev(d)->cdv_next);
193 struct lu_device *ccc_device_alloc(const struct lu_env *env,
194 struct lu_device_type *t,
195 struct lustre_cfg *cfg,
196 const struct lu_device_operations *luops,
197 const struct cl_device_operations *clops)
199 struct ccc_device *vdv;
200 struct lu_device *lud;
201 struct cl_site *site;
207 RETURN(ERR_PTR(-ENOMEM));
209 lud = &vdv->cdv_cl.cd_lu_dev;
210 cl_device_init(&vdv->cdv_cl, t);
211 ccc2lu_dev(vdv)->ld_ops = luops;
212 vdv->cdv_cl.cd_ops = clops;
216 rc = cl_site_init(site, &vdv->cdv_cl);
218 rc = lu_site_init_finish(&site->cs_lu);
220 LASSERT(lud->ld_site == NULL);
221 CERROR("Cannot init lu_site, rc %d.\n", rc);
227 ccc_device_free(env, lud);
233 struct lu_device *ccc_device_free(const struct lu_env *env,
236 struct ccc_device *vdv = lu2ccc_dev(d);
237 struct cl_site *site = lu2cl_site(d->ld_site);
238 struct lu_device *next = cl2lu_dev(vdv->cdv_next);
240 if (d->ld_site != NULL) {
244 cl_device_fini(lu2cl_dev(d));
249 int ccc_req_init(const struct lu_env *env, struct cl_device *dev,
255 OBD_SLAB_ALLOC_PTR_GFP(vrq, ccc_req_kmem, GFP_NOFS);
257 cl_req_slice_add(req, &vrq->crq_cl, dev, &ccc_req_ops);
265 * An `emergency' environment used by ccc_inode_fini() when cl_env_get()
266 * fails. Access to this environment is serialized by ccc_inode_fini_guard
269 static struct lu_env *ccc_inode_fini_env = NULL;
272 * A mutex serializing calls to slp_inode_fini() under extreme memory
273 * pressure, when environments cannot be allocated.
275 static DEFINE_MUTEX(ccc_inode_fini_guard);
276 static int dummy_refcheck;
278 int ccc_global_init(struct lu_device_type *device_type)
282 result = lu_kmem_init(ccc_caches);
286 result = lu_device_type_init(device_type);
290 ccc_inode_fini_env = cl_env_alloc(&dummy_refcheck,
291 LCT_REMEMBER|LCT_NOREF);
292 if (IS_ERR(ccc_inode_fini_env)) {
293 result = PTR_ERR(ccc_inode_fini_env);
297 ccc_inode_fini_env->le_ctx.lc_cookie = 0x4;
300 lu_device_type_fini(device_type);
302 lu_kmem_fini(ccc_caches);
306 void ccc_global_fini(struct lu_device_type *device_type)
308 if (ccc_inode_fini_env != NULL) {
309 cl_env_put(ccc_inode_fini_env, &dummy_refcheck);
310 ccc_inode_fini_env = NULL;
312 lu_device_type_fini(device_type);
313 lu_kmem_fini(ccc_caches);
316 /*****************************************************************************
322 struct lu_object *ccc_object_alloc(const struct lu_env *env,
323 const struct lu_object_header *unused,
324 struct lu_device *dev,
325 const struct cl_object_operations *clops,
326 const struct lu_object_operations *luops)
328 struct ccc_object *vob;
329 struct lu_object *obj;
331 OBD_SLAB_ALLOC_PTR_GFP(vob, ccc_object_kmem, GFP_NOFS);
333 struct cl_object_header *hdr;
336 hdr = &vob->cob_header;
337 cl_object_header_init(hdr);
338 hdr->coh_page_bufsize = cfs_size_round(sizeof(struct cl_page));
340 lu_object_init(obj, &hdr->coh_lu, dev);
341 lu_object_add_top(&hdr->coh_lu, obj);
343 vob->cob_cl.co_ops = clops;
350 int ccc_object_init0(const struct lu_env *env,
351 struct ccc_object *vob,
352 const struct cl_object_conf *conf)
354 vob->cob_inode = conf->coc_inode;
355 vob->cob_transient_pages = 0;
356 cl_object_page_init(&vob->cob_cl, sizeof(struct ccc_page));
360 int ccc_object_init(const struct lu_env *env, struct lu_object *obj,
361 const struct lu_object_conf *conf)
363 struct ccc_device *dev = lu2ccc_dev(obj->lo_dev);
364 struct ccc_object *vob = lu2ccc(obj);
365 struct lu_object *below;
366 struct lu_device *under;
369 under = &dev->cdv_next->cd_lu_dev;
370 below = under->ld_ops->ldo_object_alloc(env, obj->lo_header, under);
372 const struct cl_object_conf *cconf;
374 cconf = lu2cl_conf(conf);
375 CFS_INIT_LIST_HEAD(&vob->cob_pending_list);
376 lu_object_add(obj, below);
377 result = ccc_object_init0(env, vob, cconf);
383 void ccc_object_free(const struct lu_env *env, struct lu_object *obj)
385 struct ccc_object *vob = lu2ccc(obj);
388 lu_object_header_fini(obj->lo_header);
389 OBD_SLAB_FREE_PTR(vob, ccc_object_kmem);
392 int ccc_lock_init(const struct lu_env *env,
393 struct cl_object *obj, struct cl_lock *lock,
394 const struct cl_io *unused,
395 const struct cl_lock_operations *lkops)
397 struct ccc_lock *clk;
400 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
402 OBD_SLAB_ALLOC_PTR_GFP(clk, ccc_lock_kmem, GFP_NOFS);
404 cl_lock_slice_add(lock, &clk->clk_cl, obj, lkops);
411 int ccc_attr_set(const struct lu_env *env, struct cl_object *obj,
412 const struct cl_attr *attr, unsigned valid)
417 int ccc_object_glimpse(const struct lu_env *env,
418 const struct cl_object *obj, struct ost_lvb *lvb)
420 struct inode *inode = ccc_object_inode(obj);
423 lvb->lvb_mtime = cl_inode_mtime(inode);
424 lvb->lvb_atime = cl_inode_atime(inode);
425 lvb->lvb_ctime = cl_inode_ctime(inode);
427 * LU-417: Add dirty pages block count lest i_blocks reports 0, some
428 * "cp" or "tar" on remote node may think it's a completely sparse file
431 if (lvb->lvb_size > 0 && lvb->lvb_blocks == 0)
432 lvb->lvb_blocks = dirty_cnt(inode);
438 int ccc_conf_set(const struct lu_env *env, struct cl_object *obj,
439 const struct cl_object_conf *conf)
441 /* TODO: destroy all pages attached to this object. */
445 static void ccc_object_size_lock(struct cl_object *obj)
447 struct inode *inode = ccc_object_inode(obj);
449 cl_isize_lock(inode);
450 cl_object_attr_lock(obj);
453 static void ccc_object_size_unlock(struct cl_object *obj)
455 struct inode *inode = ccc_object_inode(obj);
457 cl_object_attr_unlock(obj);
458 cl_isize_unlock(inode);
461 /*****************************************************************************
467 int ccc_fail(const struct lu_env *env, const struct cl_page_slice *slice)
476 void ccc_transient_page_verify(const struct cl_page *page)
480 int ccc_transient_page_own(const struct lu_env *env,
481 const struct cl_page_slice *slice,
482 struct cl_io *unused,
485 ccc_transient_page_verify(slice->cpl_page);
489 void ccc_transient_page_assume(const struct lu_env *env,
490 const struct cl_page_slice *slice,
491 struct cl_io *unused)
493 ccc_transient_page_verify(slice->cpl_page);
496 void ccc_transient_page_unassume(const struct lu_env *env,
497 const struct cl_page_slice *slice,
498 struct cl_io *unused)
500 ccc_transient_page_verify(slice->cpl_page);
503 void ccc_transient_page_disown(const struct lu_env *env,
504 const struct cl_page_slice *slice,
505 struct cl_io *unused)
507 ccc_transient_page_verify(slice->cpl_page);
510 void ccc_transient_page_discard(const struct lu_env *env,
511 const struct cl_page_slice *slice,
512 struct cl_io *unused)
514 struct cl_page *page = slice->cpl_page;
516 ccc_transient_page_verify(slice->cpl_page);
519 * For transient pages, remove it from the radix tree.
521 cl_page_delete(env, page);
524 int ccc_transient_page_prep(const struct lu_env *env,
525 const struct cl_page_slice *slice,
526 struct cl_io *unused)
529 /* transient page should always be sent. */
533 /*****************************************************************************
539 void ccc_lock_delete(const struct lu_env *env,
540 const struct cl_lock_slice *slice)
542 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
545 void ccc_lock_fini(const struct lu_env *env, struct cl_lock_slice *slice)
547 struct ccc_lock *clk = cl2ccc_lock(slice);
548 OBD_SLAB_FREE_PTR(clk, ccc_lock_kmem);
551 int ccc_lock_enqueue(const struct lu_env *env,
552 const struct cl_lock_slice *slice,
553 struct cl_io *unused, __u32 enqflags)
555 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
559 int ccc_lock_use(const struct lu_env *env, const struct cl_lock_slice *slice)
561 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
565 int ccc_lock_unuse(const struct lu_env *env, const struct cl_lock_slice *slice)
567 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
571 int ccc_lock_wait(const struct lu_env *env, const struct cl_lock_slice *slice)
573 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
578 * Implementation of cl_lock_operations::clo_fits_into() methods for ccc
579 * layer. This function is executed every time io finds an existing lock in
580 * the lock cache while creating new lock. This function has to decide whether
581 * cached lock "fits" into io.
583 * \param slice lock to be checked
584 * \param io IO that wants a lock.
586 * \see lov_lock_fits_into().
588 int ccc_lock_fits_into(const struct lu_env *env,
589 const struct cl_lock_slice *slice,
590 const struct cl_lock_descr *need,
591 const struct cl_io *io)
593 const struct cl_lock *lock = slice->cls_lock;
594 const struct cl_lock_descr *descr = &lock->cll_descr;
595 const struct ccc_io *cio = ccc_env_io(env);
600 * Work around DLM peculiarity: it assumes that glimpse
601 * (LDLM_FL_HAS_INTENT) lock is always LCK_PR, and returns reads lock
602 * when asked for LCK_PW lock with LDLM_FL_HAS_INTENT flag set. Make
603 * sure that glimpse doesn't get CLM_WRITE top-lock, so that it
604 * doesn't enqueue CLM_WRITE sub-locks.
606 if (cio->cui_glimpse)
607 result = descr->cld_mode != CLM_WRITE;
610 * Also, don't match incomplete write locks for read, otherwise read
611 * would enqueue missing sub-locks in the write mode.
613 else if (need->cld_mode != descr->cld_mode)
614 result = lock->cll_state >= CLS_ENQUEUED;
621 * Implements cl_lock_operations::clo_state() method for ccc layer, invoked
622 * whenever lock state changes. Transfers object attributes, that might be
623 * updated as a result of lock acquiring into inode.
625 void ccc_lock_state(const struct lu_env *env,
626 const struct cl_lock_slice *slice,
627 enum cl_lock_state state)
629 struct cl_lock *lock = slice->cls_lock;
633 * Refresh inode attributes when the lock is moving into CLS_HELD
634 * state, and only when this is a result of real enqueue, rather than
635 * of finding lock in the cache.
637 if (state == CLS_HELD && lock->cll_state < CLS_HELD) {
638 struct cl_object *obj;
641 obj = slice->cls_obj;
642 inode = ccc_object_inode(obj);
644 /* vmtruncate() sets the i_size
645 * under both a DLM lock and the
646 * ll_inode_size_lock(). If we don't get the
647 * ll_inode_size_lock() here we can match the DLM lock and
648 * reset i_size. generic_file_write can then trust the
649 * stale i_size when doing appending writes and effectively
650 * cancel the result of the truncate. Getting the
651 * ll_inode_size_lock() after the enqueue maintains the DLM
652 * -> ll_inode_size_lock() acquiring order. */
653 if (lock->cll_descr.cld_start == 0 &&
654 lock->cll_descr.cld_end == CL_PAGE_EOF)
655 cl_merge_lvb(env, inode);
660 /*****************************************************************************
666 void ccc_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
668 struct cl_io *io = ios->cis_io;
670 CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj));
673 int ccc_io_one_lock_index(const struct lu_env *env, struct cl_io *io,
674 __u32 enqflags, enum cl_lock_mode mode,
675 pgoff_t start, pgoff_t end)
677 struct ccc_io *cio = ccc_env_io(env);
678 struct cl_lock_descr *descr = &cio->cui_link.cill_descr;
679 struct cl_object *obj = io->ci_obj;
681 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
684 CDEBUG(D_VFSTRACE, "lock: %d [%lu, %lu]\n", mode, start, end);
686 memset(&cio->cui_link, 0, sizeof cio->cui_link);
688 if (cio->cui_fd && (cio->cui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
689 descr->cld_mode = CLM_GROUP;
690 descr->cld_gid = cio->cui_fd->fd_grouplock.cg_gid;
692 descr->cld_mode = mode;
694 descr->cld_obj = obj;
695 descr->cld_start = start;
696 descr->cld_end = end;
697 descr->cld_enq_flags = enqflags;
699 cl_io_lock_add(env, io, &cio->cui_link);
703 void ccc_io_update_iov(const struct lu_env *env,
704 struct ccc_io *cio, struct cl_io *io)
707 size_t size = io->u.ci_rw.crw_count;
709 cio->cui_iov_olen = 0;
710 if (!cl_is_normalio(env, io) || cio->cui_tot_nrsegs == 0)
713 for (i = 0; i < cio->cui_tot_nrsegs; i++) {
714 struct iovec *iv = &cio->cui_iov[i];
716 if (iv->iov_len < size)
719 if (iv->iov_len > size) {
720 cio->cui_iov_olen = iv->iov_len;
727 cio->cui_nrsegs = i + 1;
728 LASSERTF(cio->cui_tot_nrsegs >= cio->cui_nrsegs,
729 "tot_nrsegs: %lu, nrsegs: %lu\n",
730 cio->cui_tot_nrsegs, cio->cui_nrsegs);
733 int ccc_io_one_lock(const struct lu_env *env, struct cl_io *io,
734 __u32 enqflags, enum cl_lock_mode mode,
735 loff_t start, loff_t end)
737 struct cl_object *obj = io->ci_obj;
738 return ccc_io_one_lock_index(env, io, enqflags, mode,
739 cl_index(obj, start), cl_index(obj, end));
742 void ccc_io_end(const struct lu_env *env, const struct cl_io_slice *ios)
744 CLOBINVRNT(env, ios->cis_io->ci_obj,
745 ccc_object_invariant(ios->cis_io->ci_obj));
748 void ccc_io_advance(const struct lu_env *env,
749 const struct cl_io_slice *ios,
752 struct ccc_io *cio = cl2ccc_io(env, ios);
753 struct cl_io *io = ios->cis_io;
754 struct cl_object *obj = ios->cis_io->ci_obj;
756 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
758 if (!cl_is_normalio(env, io))
761 LASSERT(cio->cui_tot_nrsegs >= cio->cui_nrsegs);
762 LASSERT(cio->cui_tot_count >= nob);
764 cio->cui_iov += cio->cui_nrsegs;
765 cio->cui_tot_nrsegs -= cio->cui_nrsegs;
766 cio->cui_tot_count -= nob;
769 if (cio->cui_iov_olen > 0) {
773 cio->cui_tot_nrsegs++;
774 iv = &cio->cui_iov[0];
775 if (io->ci_continue) {
776 iv->iov_base += iv->iov_len;
777 LASSERT(cio->cui_iov_olen > iv->iov_len);
778 iv->iov_len = cio->cui_iov_olen - iv->iov_len;
780 /* restore the iov_len, in case of restart io. */
781 iv->iov_len = cio->cui_iov_olen;
783 cio->cui_iov_olen = 0;
788 * Helper function that if necessary adjusts file size (inode->i_size), when
789 * position at the offset \a pos is accessed. File size can be arbitrary stale
790 * on a Lustre client, but client at least knows KMS. If accessed area is
791 * inside [0, KMS], set file size to KMS, otherwise glimpse file size.
793 * Locking: cl_isize_lock is used to serialize changes to inode size and to
794 * protect consistency between inode size and cl_object
795 * attributes. cl_object_size_lock() protects consistency between cl_attr's of
796 * top-object and sub-objects.
798 int ccc_prep_size(const struct lu_env *env, struct cl_object *obj,
799 struct cl_io *io, loff_t start, size_t count, int *exceed)
801 struct cl_attr *attr = ccc_env_thread_attr(env);
802 struct inode *inode = ccc_object_inode(obj);
803 loff_t pos = start + count - 1;
808 * Consistency guarantees: following possibilities exist for the
809 * relation between region being accessed and real file size at this
812 * (A): the region is completely inside of the file;
814 * (B-x): x bytes of region are inside of the file, the rest is
817 * (C): the region is completely outside of the file.
819 * This classification is stable under DLM lock already acquired by
820 * the caller, because to change the class, other client has to take
821 * DLM lock conflicting with our lock. Also, any updates to ->i_size
822 * by other threads on this client are serialized by
823 * ll_inode_size_lock(). This guarantees that short reads are handled
824 * correctly in the face of concurrent writes and truncates.
826 ccc_object_size_lock(obj);
827 result = cl_object_attr_get(env, obj, attr);
832 * A glimpse is necessary to determine whether we
833 * return a short read (B) or some zeroes at the end
836 ccc_object_size_unlock(obj);
837 result = cl_glimpse_lock(env, io, inode, obj, 0);
838 if (result == 0 && exceed != NULL) {
839 /* If objective page index exceed end-of-file
840 * page index, return directly. Do not expect
841 * kernel will check such case correctly.
842 * linux-2.6.18-128.1.1 miss to do that.
844 loff_t size = cl_isize_read(inode);
845 unsigned long cur_index = start >>
848 if ((size == 0 && cur_index != 0) ||
849 (((size - 1) >> PAGE_CACHE_SHIFT) <
856 * region is within kms and, hence, within real file
857 * size (A). We need to increase i_size to cover the
858 * read region so that generic_file_read() will do its
859 * job, but that doesn't mean the kms size is
860 * _correct_, it is only the _minimum_ size. If
861 * someone does a stat they will get the correct size
862 * which will always be >= the kms value here.
865 if (cl_isize_read(inode) < kms) {
866 cl_isize_write_nolock(inode, kms);
868 DFID" updating i_size "LPU64"\n",
869 PFID(lu_object_fid(&obj->co_lu)),
870 (__u64)cl_isize_read(inode));
875 ccc_object_size_unlock(obj);
879 /*****************************************************************************
881 * Transfer operations.
885 void ccc_req_completion(const struct lu_env *env,
886 const struct cl_req_slice *slice, int ioret)
891 cl_stats_tally(slice->crs_dev, slice->crs_req->crq_type, ioret);
893 vrq = cl2ccc_req(slice);
894 OBD_SLAB_FREE_PTR(vrq, ccc_req_kmem);
898 * Implementation of struct cl_req_operations::cro_attr_set() for ccc
899 * layer. ccc is responsible for
917 void ccc_req_attr_set(const struct lu_env *env,
918 const struct cl_req_slice *slice,
919 const struct cl_object *obj,
920 struct cl_req_attr *attr, obd_valid flags)
924 obd_flag valid_flags;
927 inode = ccc_object_inode(obj);
928 valid_flags = OBD_MD_FLTYPE;
930 if ((flags & OBD_MD_FLOSSCAPA) != 0) {
931 LASSERT(attr->cra_capa == NULL);
932 attr->cra_capa = cl_capa_lookup(inode,
933 slice->crs_req->crq_type);
936 if (slice->crs_req->crq_type == CRT_WRITE) {
937 if (flags & OBD_MD_FLEPOCH) {
938 oa->o_valid |= OBD_MD_FLEPOCH;
939 oa->o_ioepoch = cl_i2info(inode)->lli_ioepoch;
940 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME |
941 OBD_MD_FLUID | OBD_MD_FLGID;
944 obdo_from_inode(oa, inode, valid_flags & flags);
945 obdo_set_parent_fid(oa, &cl_i2info(inode)->lli_fid);
946 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_INVALID_PFID))
948 memcpy(attr->cra_jobid, cl_i2info(inode)->lli_jobid,
949 JOBSTATS_JOBID_SIZE);
952 static const struct cl_req_operations ccc_req_ops = {
953 .cro_attr_set = ccc_req_attr_set,
954 .cro_completion = ccc_req_completion
957 int cl_setattr_ost(struct inode *inode, const struct iattr *attr,
958 struct obd_capa *capa)
967 env = cl_env_get(&refcheck);
969 RETURN(PTR_ERR(env));
971 io = ccc_env_thread_io(env);
972 io->ci_obj = cl_i2info(inode)->lli_clob;
974 io->u.ci_setattr.sa_attr.lvb_atime = LTIME_S(attr->ia_atime);
975 io->u.ci_setattr.sa_attr.lvb_mtime = LTIME_S(attr->ia_mtime);
976 io->u.ci_setattr.sa_attr.lvb_ctime = LTIME_S(attr->ia_ctime);
977 io->u.ci_setattr.sa_attr.lvb_size = attr->ia_size;
978 io->u.ci_setattr.sa_valid = attr->ia_valid;
979 io->u.ci_setattr.sa_capa = capa;
982 if (cl_io_init(env, io, CIT_SETATTR, io->ci_obj) == 0) {
983 struct ccc_io *cio = ccc_env_io(env);
985 if (attr->ia_valid & ATTR_FILE)
986 /* populate the file descriptor for ftruncate to honor
987 * group lock - see LU-787 */
988 cio->cui_fd = cl_iattr2fd(inode, attr);
990 result = cl_io_loop(env, io);
992 result = io->ci_result;
995 if (unlikely(io->ci_need_restart))
997 /* HSM import case: file is released, cannot be restored
998 * no need to fail except if restore registration failed
1000 if (result == -ENODATA && io->ci_restore_needed &&
1001 io->ci_result != -ENODATA)
1003 cl_env_put(env, &refcheck);
1007 /*****************************************************************************
1013 struct lu_device *ccc2lu_dev(struct ccc_device *vdv)
1015 return &vdv->cdv_cl.cd_lu_dev;
1018 struct ccc_device *lu2ccc_dev(const struct lu_device *d)
1020 return container_of0(d, struct ccc_device, cdv_cl.cd_lu_dev);
1023 struct ccc_device *cl2ccc_dev(const struct cl_device *d)
1025 return container_of0(d, struct ccc_device, cdv_cl);
1028 struct lu_object *ccc2lu(struct ccc_object *vob)
1030 return &vob->cob_cl.co_lu;
1033 struct ccc_object *lu2ccc(const struct lu_object *obj)
1035 return container_of0(obj, struct ccc_object, cob_cl.co_lu);
1038 struct ccc_object *cl2ccc(const struct cl_object *obj)
1040 return container_of0(obj, struct ccc_object, cob_cl);
1043 struct ccc_lock *cl2ccc_lock(const struct cl_lock_slice *slice)
1045 return container_of(slice, struct ccc_lock, clk_cl);
1048 struct ccc_io *cl2ccc_io(const struct lu_env *env,
1049 const struct cl_io_slice *slice)
1053 cio = container_of(slice, struct ccc_io, cui_cl);
1054 LASSERT(cio == ccc_env_io(env));
1058 struct ccc_req *cl2ccc_req(const struct cl_req_slice *slice)
1060 return container_of0(slice, struct ccc_req, crq_cl);
1063 struct page *cl2vm_page(const struct cl_page_slice *slice)
1065 return cl2ccc_page(slice)->cpg_page;
1068 /*****************************************************************************
1073 int ccc_object_invariant(const struct cl_object *obj)
1075 struct inode *inode = ccc_object_inode(obj);
1076 struct cl_inode_info *lli = cl_i2info(inode);
1078 return (S_ISREG(cl_inode_mode(inode)) ||
1079 /* i_mode of unlinked inode is zeroed. */
1080 cl_inode_mode(inode) == 0) && lli->lli_clob == obj;
1083 struct inode *ccc_object_inode(const struct cl_object *obj)
1085 return cl2ccc(obj)->cob_inode;
1089 * Returns a pointer to cl_page associated with \a vmpage, without acquiring
1090 * additional reference to the resulting page. This is an unsafe version of
1091 * cl_vmpage_page() that can only be used under vmpage lock.
1093 struct cl_page *ccc_vmpage_page_transient(struct page *vmpage)
1095 KLASSERT(PageLocked(vmpage));
1096 return (struct cl_page *)vmpage->private;
1100 * Initialize or update CLIO structures for regular files when new
1101 * meta-data arrives from the server.
1103 * \param inode regular file inode
1104 * \param md new file metadata from MDS
1105 * - allocates cl_object if necessary,
1106 * - updated layout, if object was already here.
1108 int cl_file_inode_init(struct inode *inode, struct lustre_md *md)
1111 struct cl_inode_info *lli;
1112 struct cl_object *clob;
1113 struct lu_site *site;
1115 struct cl_object_conf conf = {
1124 LASSERT(md->body->mbo_valid & OBD_MD_FLID);
1125 LASSERT(S_ISREG(cl_inode_mode(inode)));
1127 env = cl_env_get(&refcheck);
1129 return PTR_ERR(env);
1131 site = cl_i2sbi(inode)->ll_site;
1132 lli = cl_i2info(inode);
1133 fid = &lli->lli_fid;
1134 LASSERT(fid_is_sane(fid));
1136 if (lli->lli_clob == NULL) {
1137 /* clob is slave of inode, empty lli_clob means for new inode,
1138 * there is no clob in cache with the given fid, so it is
1139 * unnecessary to perform lookup-alloc-lookup-insert, just
1140 * alloc and insert directly. */
1141 LASSERT(inode->i_state & I_NEW);
1142 conf.coc_lu.loc_flags = LOC_F_NEW;
1143 clob = cl_object_find(env, lu2cl_dev(site->ls_top_dev),
1145 if (!IS_ERR(clob)) {
1147 * No locking is necessary, as new inode is
1148 * locked by I_NEW bit.
1150 lli->lli_clob = clob;
1151 lli->lli_has_smd = lsm_has_objects(md->lsm);
1152 lu_object_ref_add(&clob->co_lu, "inode", inode);
1154 result = PTR_ERR(clob);
1156 result = cl_conf_set(env, lli->lli_clob, &conf);
1159 cl_env_put(env, &refcheck);
1162 CERROR("Failure to initialize cl object "DFID": %d\n",
1168 * Wait for others drop their references of the object at first, then we drop
1169 * the last one, which will lead to the object be destroyed immediately.
1170 * Must be called after cl_object_kill() against this object.
1172 * The reason we want to do this is: destroying top object will wait for sub
1173 * objects being destroyed first, so we can't let bottom layer (e.g. from ASTs)
1174 * to initiate top object destroying which may deadlock. See bz22520.
1176 static void cl_object_put_last(struct lu_env *env, struct cl_object *obj)
1178 struct lu_object_header *header = obj->co_lu.lo_header;
1179 wait_queue_t waiter;
1181 if (unlikely(atomic_read(&header->loh_ref) != 1)) {
1182 struct lu_site *site = obj->co_lu.lo_dev->ld_site;
1183 struct lu_site_bkt_data *bkt;
1185 bkt = lu_site_bkt_from_fid(site, &header->loh_fid);
1187 init_waitqueue_entry_current(&waiter);
1188 add_wait_queue(&bkt->lsb_marche_funebre, &waiter);
1191 set_current_state(TASK_UNINTERRUPTIBLE);
1192 if (atomic_read(&header->loh_ref) == 1)
1194 waitq_wait(&waiter, TASK_UNINTERRUPTIBLE);
1197 set_current_state(TASK_RUNNING);
1198 remove_wait_queue(&bkt->lsb_marche_funebre, &waiter);
1201 cl_object_put(env, obj);
1204 void cl_inode_fini(struct inode *inode)
1207 struct cl_inode_info *lli = cl_i2info(inode);
1208 struct cl_object *clob = lli->lli_clob;
1215 cookie = cl_env_reenter();
1216 env = cl_env_get(&refcheck);
1217 emergency = IS_ERR(env);
1219 mutex_lock(&ccc_inode_fini_guard);
1220 LASSERT(ccc_inode_fini_env != NULL);
1221 cl_env_implant(ccc_inode_fini_env, &refcheck);
1222 env = ccc_inode_fini_env;
1225 * cl_object cache is a slave to inode cache (which, in turn
1226 * is a slave to dentry cache), don't keep cl_object in memory
1227 * when its master is evicted.
1229 cl_object_kill(env, clob);
1230 lu_object_ref_del(&clob->co_lu, "inode", inode);
1231 cl_object_put_last(env, clob);
1232 lli->lli_clob = NULL;
1234 cl_env_unplant(ccc_inode_fini_env, &refcheck);
1235 mutex_unlock(&ccc_inode_fini_guard);
1237 cl_env_put(env, &refcheck);
1238 cl_env_reexit(cookie);
1243 * return IF_* type for given lu_dirent entry.
1244 * IF_* flag shld be converted to particular OS file type in
1245 * platform llite module.
1247 __u16 ll_dirent_type_get(struct lu_dirent *ent)
1250 struct luda_type *lt;
1253 if (le32_to_cpu(ent->lde_attrs) & LUDA_TYPE) {
1254 const unsigned align = sizeof(struct luda_type) - 1;
1256 len = le16_to_cpu(ent->lde_namelen);
1257 len = (len + align) & ~align;
1258 lt = (void *)ent->lde_name + len;
1259 type = IFTODT(le16_to_cpu(lt->lt_type));
1265 * build inode number from passed @fid */
1266 __u64 cl_fid_build_ino(const struct lu_fid *fid, int api32)
1268 if (BITS_PER_LONG == 32 || api32)
1269 RETURN(fid_flatten32(fid));
1271 RETURN(fid_flatten(fid));
1275 * build inode generation from passed @fid. If our FID overflows the 32-bit
1276 * inode number then return a non-zero generation to distinguish them. */
1277 __u32 cl_fid_build_gen(const struct lu_fid *fid)
1282 if (fid_is_igif(fid)) {
1283 gen = lu_igif_gen(fid);
1287 gen = (fid_flatten(fid) >> 32);
1291 /* lsm is unreliable after hsm implementation as layout can be changed at
1292 * any time. This is only to support old, non-clio-ized interfaces. It will
1293 * cause deadlock if clio operations are called with this extra layout refcount
1294 * because in case the layout changed during the IO, ll_layout_refresh() will
1295 * have to wait for the refcount to become zero to destroy the older layout.
1297 * Notice that the lsm returned by this function may not be valid unless called
1298 * inside layout lock - MDS_INODELOCK_LAYOUT. */
1299 struct lov_stripe_md *ccc_inode_lsm_get(struct inode *inode)
1301 return lov_lsm_get(cl_i2info(inode)->lli_clob);
1304 void inline ccc_inode_lsm_put(struct inode *inode, struct lov_stripe_md *lsm)
1306 lov_lsm_put(cl_i2info(inode)->lli_clob, lsm);