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_dlm.h>
57 #include <lustre_ver.h>
58 #include <lustre_mdc.h>
59 #include <cl_object.h>
63 #include "llite_internal.h"
65 static const struct cl_req_operations ccc_req_ops;
68 * ccc_ prefix stands for "Common Client Code".
71 static struct kmem_cache *ccc_lock_kmem;
72 static struct kmem_cache *ccc_object_kmem;
73 static struct kmem_cache *ccc_thread_kmem;
74 static struct kmem_cache *ccc_session_kmem;
75 static struct kmem_cache *ccc_req_kmem;
77 static struct lu_kmem_descr ccc_caches[] = {
79 .ckd_cache = &ccc_lock_kmem,
80 .ckd_name = "ccc_lock_kmem",
81 .ckd_size = sizeof (struct ccc_lock)
84 .ckd_cache = &ccc_object_kmem,
85 .ckd_name = "ccc_object_kmem",
86 .ckd_size = sizeof (struct ccc_object)
89 .ckd_cache = &ccc_thread_kmem,
90 .ckd_name = "ccc_thread_kmem",
91 .ckd_size = sizeof (struct ccc_thread_info),
94 .ckd_cache = &ccc_session_kmem,
95 .ckd_name = "ccc_session_kmem",
96 .ckd_size = sizeof (struct ccc_session)
99 .ckd_cache = &ccc_req_kmem,
100 .ckd_name = "ccc_req_kmem",
101 .ckd_size = sizeof (struct ccc_req)
108 /*****************************************************************************
110 * Vvp device and device type functions.
114 void *ccc_key_init(const struct lu_context *ctx, struct lu_context_key *key)
116 struct ccc_thread_info *info;
118 OBD_SLAB_ALLOC_PTR_GFP(info, ccc_thread_kmem, GFP_NOFS);
120 info = ERR_PTR(-ENOMEM);
124 void ccc_key_fini(const struct lu_context *ctx,
125 struct lu_context_key *key, void *data)
127 struct ccc_thread_info *info = data;
128 OBD_SLAB_FREE_PTR(info, ccc_thread_kmem);
131 void *ccc_session_key_init(const struct lu_context *ctx,
132 struct lu_context_key *key)
134 struct ccc_session *session;
136 OBD_SLAB_ALLOC_PTR_GFP(session, ccc_session_kmem, GFP_NOFS);
138 session = ERR_PTR(-ENOMEM);
142 void ccc_session_key_fini(const struct lu_context *ctx,
143 struct lu_context_key *key, void *data)
145 struct ccc_session *session = data;
146 OBD_SLAB_FREE_PTR(session, ccc_session_kmem);
149 struct lu_context_key ccc_key = {
150 .lct_tags = LCT_CL_THREAD,
151 .lct_init = ccc_key_init,
152 .lct_fini = ccc_key_fini
155 struct lu_context_key ccc_session_key = {
156 .lct_tags = LCT_SESSION,
157 .lct_init = ccc_session_key_init,
158 .lct_fini = ccc_session_key_fini
162 /* type constructor/destructor: ccc_type_{init,fini,start,stop}(). */
163 // LU_TYPE_INIT_FINI(ccc, &ccc_key, &ccc_session_key);
165 int ccc_device_init(const struct lu_env *env, struct lu_device *d,
166 const char *name, struct lu_device *next)
168 struct ccc_device *vdv;
173 vdv->cdv_next = lu2cl_dev(next);
175 LASSERT(d->ld_site != NULL && next->ld_type != NULL);
176 next->ld_site = d->ld_site;
177 rc = next->ld_type->ldt_ops->ldto_device_init(
178 env, next, next->ld_type->ldt_name, NULL);
181 lu_ref_add(&next->ld_reference, "lu-stack", &lu_site_init);
186 struct lu_device *ccc_device_fini(const struct lu_env *env,
189 return cl2lu_dev(lu2ccc_dev(d)->cdv_next);
192 struct lu_device *ccc_device_alloc(const struct lu_env *env,
193 struct lu_device_type *t,
194 struct lustre_cfg *cfg,
195 const struct lu_device_operations *luops,
196 const struct cl_device_operations *clops)
198 struct ccc_device *vdv;
199 struct lu_device *lud;
200 struct cl_site *site;
206 RETURN(ERR_PTR(-ENOMEM));
208 lud = &vdv->cdv_cl.cd_lu_dev;
209 cl_device_init(&vdv->cdv_cl, t);
210 ccc2lu_dev(vdv)->ld_ops = luops;
211 vdv->cdv_cl.cd_ops = clops;
215 rc = cl_site_init(site, &vdv->cdv_cl);
217 rc = lu_site_init_finish(&site->cs_lu);
219 LASSERT(lud->ld_site == NULL);
220 CERROR("Cannot init lu_site, rc %d.\n", rc);
226 ccc_device_free(env, lud);
232 struct lu_device *ccc_device_free(const struct lu_env *env,
235 struct ccc_device *vdv = lu2ccc_dev(d);
236 struct cl_site *site = lu2cl_site(d->ld_site);
237 struct lu_device *next = cl2lu_dev(vdv->cdv_next);
239 if (d->ld_site != NULL) {
243 cl_device_fini(lu2cl_dev(d));
248 int ccc_req_init(const struct lu_env *env, struct cl_device *dev,
254 OBD_SLAB_ALLOC_PTR_GFP(vrq, ccc_req_kmem, GFP_NOFS);
256 cl_req_slice_add(req, &vrq->crq_cl, dev, &ccc_req_ops);
264 * An `emergency' environment used by ccc_inode_fini() when cl_env_get()
265 * fails. Access to this environment is serialized by ccc_inode_fini_guard
268 static struct lu_env *ccc_inode_fini_env = NULL;
271 * A mutex serializing calls to slp_inode_fini() under extreme memory
272 * pressure, when environments cannot be allocated.
274 static DEFINE_MUTEX(ccc_inode_fini_guard);
275 static int dummy_refcheck;
277 int ccc_global_init(struct lu_device_type *device_type)
281 result = lu_kmem_init(ccc_caches);
285 result = lu_device_type_init(device_type);
289 ccc_inode_fini_env = cl_env_alloc(&dummy_refcheck,
290 LCT_REMEMBER|LCT_NOREF);
291 if (IS_ERR(ccc_inode_fini_env)) {
292 result = PTR_ERR(ccc_inode_fini_env);
296 ccc_inode_fini_env->le_ctx.lc_cookie = 0x4;
299 lu_device_type_fini(device_type);
301 lu_kmem_fini(ccc_caches);
305 void ccc_global_fini(struct lu_device_type *device_type)
307 if (ccc_inode_fini_env != NULL) {
308 cl_env_put(ccc_inode_fini_env, &dummy_refcheck);
309 ccc_inode_fini_env = NULL;
311 lu_device_type_fini(device_type);
312 lu_kmem_fini(ccc_caches);
315 /*****************************************************************************
321 struct lu_object *ccc_object_alloc(const struct lu_env *env,
322 const struct lu_object_header *unused,
323 struct lu_device *dev,
324 const struct cl_object_operations *clops,
325 const struct lu_object_operations *luops)
327 struct ccc_object *vob;
328 struct lu_object *obj;
330 OBD_SLAB_ALLOC_PTR_GFP(vob, ccc_object_kmem, GFP_NOFS);
332 struct cl_object_header *hdr;
335 hdr = &vob->cob_header;
336 cl_object_header_init(hdr);
337 hdr->coh_page_bufsize = cfs_size_round(sizeof(struct cl_page));
339 lu_object_init(obj, &hdr->coh_lu, dev);
340 lu_object_add_top(&hdr->coh_lu, obj);
342 vob->cob_cl.co_ops = clops;
349 int ccc_object_init0(const struct lu_env *env,
350 struct ccc_object *vob,
351 const struct cl_object_conf *conf)
353 vob->cob_inode = conf->coc_inode;
354 vob->cob_transient_pages = 0;
355 cl_object_page_init(&vob->cob_cl, sizeof(struct ccc_page));
359 int ccc_object_init(const struct lu_env *env, struct lu_object *obj,
360 const struct lu_object_conf *conf)
362 struct ccc_device *dev = lu2ccc_dev(obj->lo_dev);
363 struct ccc_object *vob = lu2ccc(obj);
364 struct lu_object *below;
365 struct lu_device *under;
368 under = &dev->cdv_next->cd_lu_dev;
369 below = under->ld_ops->ldo_object_alloc(env, obj->lo_header, under);
371 const struct cl_object_conf *cconf;
373 cconf = lu2cl_conf(conf);
374 INIT_LIST_HEAD(&vob->cob_pending_list);
375 lu_object_add(obj, below);
376 result = ccc_object_init0(env, vob, cconf);
382 void ccc_object_free(const struct lu_env *env, struct lu_object *obj)
384 struct ccc_object *vob = lu2ccc(obj);
387 lu_object_header_fini(obj->lo_header);
388 OBD_SLAB_FREE_PTR(vob, ccc_object_kmem);
391 int ccc_lock_init(const struct lu_env *env,
392 struct cl_object *obj, struct cl_lock *lock,
393 const struct cl_io *unused,
394 const struct cl_lock_operations *lkops)
396 struct ccc_lock *clk;
399 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
401 OBD_SLAB_ALLOC_PTR_GFP(clk, ccc_lock_kmem, GFP_NOFS);
403 cl_lock_slice_add(lock, &clk->clk_cl, obj, lkops);
410 int ccc_attr_set(const struct lu_env *env, struct cl_object *obj,
411 const struct cl_attr *attr, unsigned valid)
416 int ccc_object_glimpse(const struct lu_env *env,
417 const struct cl_object *obj, struct ost_lvb *lvb)
419 struct inode *inode = ccc_object_inode(obj);
422 lvb->lvb_mtime = cl_inode_mtime(inode);
423 lvb->lvb_atime = cl_inode_atime(inode);
424 lvb->lvb_ctime = cl_inode_ctime(inode);
426 * LU-417: Add dirty pages block count lest i_blocks reports 0, some
427 * "cp" or "tar" on remote node may think it's a completely sparse file
430 if (lvb->lvb_size > 0 && lvb->lvb_blocks == 0)
431 lvb->lvb_blocks = dirty_cnt(inode);
437 int ccc_conf_set(const struct lu_env *env, struct cl_object *obj,
438 const struct cl_object_conf *conf)
440 /* TODO: destroy all pages attached to this object. */
444 static void ccc_object_size_lock(struct cl_object *obj)
446 struct inode *inode = ccc_object_inode(obj);
448 cl_isize_lock(inode);
449 cl_object_attr_lock(obj);
452 static void ccc_object_size_unlock(struct cl_object *obj)
454 struct inode *inode = ccc_object_inode(obj);
456 cl_object_attr_unlock(obj);
457 cl_isize_unlock(inode);
460 /*****************************************************************************
466 int ccc_fail(const struct lu_env *env, const struct cl_page_slice *slice)
475 void ccc_transient_page_verify(const struct cl_page *page)
479 int ccc_transient_page_own(const struct lu_env *env,
480 const struct cl_page_slice *slice,
481 struct cl_io *unused,
484 ccc_transient_page_verify(slice->cpl_page);
488 void ccc_transient_page_assume(const struct lu_env *env,
489 const struct cl_page_slice *slice,
490 struct cl_io *unused)
492 ccc_transient_page_verify(slice->cpl_page);
495 void ccc_transient_page_unassume(const struct lu_env *env,
496 const struct cl_page_slice *slice,
497 struct cl_io *unused)
499 ccc_transient_page_verify(slice->cpl_page);
502 void ccc_transient_page_disown(const struct lu_env *env,
503 const struct cl_page_slice *slice,
504 struct cl_io *unused)
506 ccc_transient_page_verify(slice->cpl_page);
509 void ccc_transient_page_discard(const struct lu_env *env,
510 const struct cl_page_slice *slice,
511 struct cl_io *unused)
513 struct cl_page *page = slice->cpl_page;
515 ccc_transient_page_verify(slice->cpl_page);
518 * For transient pages, remove it from the radix tree.
520 cl_page_delete(env, page);
523 int ccc_transient_page_prep(const struct lu_env *env,
524 const struct cl_page_slice *slice,
525 struct cl_io *unused)
528 /* transient page should always be sent. */
532 /*****************************************************************************
538 void ccc_lock_delete(const struct lu_env *env,
539 const struct cl_lock_slice *slice)
541 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
544 void ccc_lock_fini(const struct lu_env *env, struct cl_lock_slice *slice)
546 struct ccc_lock *clk = cl2ccc_lock(slice);
547 OBD_SLAB_FREE_PTR(clk, ccc_lock_kmem);
550 int ccc_lock_enqueue(const struct lu_env *env,
551 const struct cl_lock_slice *slice,
552 struct cl_io *unused, __u32 enqflags)
554 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
558 int ccc_lock_use(const struct lu_env *env, const struct cl_lock_slice *slice)
560 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
564 int ccc_lock_unuse(const struct lu_env *env, const struct cl_lock_slice *slice)
566 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
570 int ccc_lock_wait(const struct lu_env *env, const struct cl_lock_slice *slice)
572 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
577 * Implementation of cl_lock_operations::clo_fits_into() methods for ccc
578 * layer. This function is executed every time io finds an existing lock in
579 * the lock cache while creating new lock. This function has to decide whether
580 * cached lock "fits" into io.
582 * \param slice lock to be checked
583 * \param io IO that wants a lock.
585 * \see lov_lock_fits_into().
587 int ccc_lock_fits_into(const struct lu_env *env,
588 const struct cl_lock_slice *slice,
589 const struct cl_lock_descr *need,
590 const struct cl_io *io)
592 const struct cl_lock *lock = slice->cls_lock;
593 const struct cl_lock_descr *descr = &lock->cll_descr;
594 const struct ccc_io *cio = ccc_env_io(env);
599 * Work around DLM peculiarity: it assumes that glimpse
600 * (LDLM_FL_HAS_INTENT) lock is always LCK_PR, and returns reads lock
601 * when asked for LCK_PW lock with LDLM_FL_HAS_INTENT flag set. Make
602 * sure that glimpse doesn't get CLM_WRITE top-lock, so that it
603 * doesn't enqueue CLM_WRITE sub-locks.
605 if (cio->cui_glimpse)
606 result = descr->cld_mode != CLM_WRITE;
609 * Also, don't match incomplete write locks for read, otherwise read
610 * would enqueue missing sub-locks in the write mode.
612 else if (need->cld_mode != descr->cld_mode)
613 result = lock->cll_state >= CLS_ENQUEUED;
620 * Implements cl_lock_operations::clo_state() method for ccc layer, invoked
621 * whenever lock state changes. Transfers object attributes, that might be
622 * updated as a result of lock acquiring into inode.
624 void ccc_lock_state(const struct lu_env *env,
625 const struct cl_lock_slice *slice,
626 enum cl_lock_state state)
628 struct cl_lock *lock = slice->cls_lock;
632 * Refresh inode attributes when the lock is moving into CLS_HELD
633 * state, and only when this is a result of real enqueue, rather than
634 * of finding lock in the cache.
636 if (state == CLS_HELD && lock->cll_state < CLS_HELD) {
637 struct cl_object *obj;
640 obj = slice->cls_obj;
641 inode = ccc_object_inode(obj);
643 /* vmtruncate() sets the i_size
644 * under both a DLM lock and the
645 * ll_inode_size_lock(). If we don't get the
646 * ll_inode_size_lock() here we can match the DLM lock and
647 * reset i_size. generic_file_write can then trust the
648 * stale i_size when doing appending writes and effectively
649 * cancel the result of the truncate. Getting the
650 * ll_inode_size_lock() after the enqueue maintains the DLM
651 * -> ll_inode_size_lock() acquiring order. */
652 if (lock->cll_descr.cld_start == 0 &&
653 lock->cll_descr.cld_end == CL_PAGE_EOF)
654 cl_merge_lvb(env, inode);
659 /*****************************************************************************
665 void ccc_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
667 struct cl_io *io = ios->cis_io;
669 CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj));
672 int ccc_io_one_lock_index(const struct lu_env *env, struct cl_io *io,
673 __u32 enqflags, enum cl_lock_mode mode,
674 pgoff_t start, pgoff_t end)
676 struct ccc_io *cio = ccc_env_io(env);
677 struct cl_lock_descr *descr = &cio->cui_link.cill_descr;
678 struct cl_object *obj = io->ci_obj;
680 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
683 CDEBUG(D_VFSTRACE, "lock: %d [%lu, %lu]\n", mode, start, end);
685 memset(&cio->cui_link, 0, sizeof cio->cui_link);
687 if (cio->cui_fd && (cio->cui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
688 descr->cld_mode = CLM_GROUP;
689 descr->cld_gid = cio->cui_fd->fd_grouplock.cg_gid;
691 descr->cld_mode = mode;
693 descr->cld_obj = obj;
694 descr->cld_start = start;
695 descr->cld_end = end;
696 descr->cld_enq_flags = enqflags;
698 cl_io_lock_add(env, io, &cio->cui_link);
702 void ccc_io_update_iov(const struct lu_env *env,
703 struct ccc_io *cio, struct cl_io *io)
706 size_t size = io->u.ci_rw.crw_count;
708 cio->cui_iov_olen = 0;
709 if (!cl_is_normalio(env, io) || cio->cui_tot_nrsegs == 0)
712 for (i = 0; i < cio->cui_tot_nrsegs; i++) {
713 struct iovec *iv = &cio->cui_iov[i];
715 if (iv->iov_len < size)
718 if (iv->iov_len > size) {
719 cio->cui_iov_olen = iv->iov_len;
726 cio->cui_nrsegs = i + 1;
727 LASSERTF(cio->cui_tot_nrsegs >= cio->cui_nrsegs,
728 "tot_nrsegs: %lu, nrsegs: %lu\n",
729 cio->cui_tot_nrsegs, cio->cui_nrsegs);
732 int ccc_io_one_lock(const struct lu_env *env, struct cl_io *io,
733 __u32 enqflags, enum cl_lock_mode mode,
734 loff_t start, loff_t end)
736 struct cl_object *obj = io->ci_obj;
737 return ccc_io_one_lock_index(env, io, enqflags, mode,
738 cl_index(obj, start), cl_index(obj, end));
741 void ccc_io_end(const struct lu_env *env, const struct cl_io_slice *ios)
743 CLOBINVRNT(env, ios->cis_io->ci_obj,
744 ccc_object_invariant(ios->cis_io->ci_obj));
747 void ccc_io_advance(const struct lu_env *env,
748 const struct cl_io_slice *ios,
751 struct ccc_io *cio = cl2ccc_io(env, ios);
752 struct cl_io *io = ios->cis_io;
753 struct cl_object *obj = ios->cis_io->ci_obj;
755 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
757 if (!cl_is_normalio(env, io))
760 LASSERT(cio->cui_tot_nrsegs >= cio->cui_nrsegs);
761 LASSERT(cio->cui_tot_count >= nob);
763 cio->cui_iov += cio->cui_nrsegs;
764 cio->cui_tot_nrsegs -= cio->cui_nrsegs;
765 cio->cui_tot_count -= nob;
768 if (cio->cui_iov_olen > 0) {
772 cio->cui_tot_nrsegs++;
773 iv = &cio->cui_iov[0];
774 if (io->ci_continue) {
775 iv->iov_base += iv->iov_len;
776 LASSERT(cio->cui_iov_olen > iv->iov_len);
777 iv->iov_len = cio->cui_iov_olen - iv->iov_len;
779 /* restore the iov_len, in case of restart io. */
780 iv->iov_len = cio->cui_iov_olen;
782 cio->cui_iov_olen = 0;
787 * Helper function that if necessary adjusts file size (inode->i_size), when
788 * position at the offset \a pos is accessed. File size can be arbitrary stale
789 * on a Lustre client, but client at least knows KMS. If accessed area is
790 * inside [0, KMS], set file size to KMS, otherwise glimpse file size.
792 * Locking: cl_isize_lock is used to serialize changes to inode size and to
793 * protect consistency between inode size and cl_object
794 * attributes. cl_object_size_lock() protects consistency between cl_attr's of
795 * top-object and sub-objects.
797 int ccc_prep_size(const struct lu_env *env, struct cl_object *obj,
798 struct cl_io *io, loff_t start, size_t count, int *exceed)
800 struct cl_attr *attr = ccc_env_thread_attr(env);
801 struct inode *inode = ccc_object_inode(obj);
802 loff_t pos = start + count - 1;
807 * Consistency guarantees: following possibilities exist for the
808 * relation between region being accessed and real file size at this
811 * (A): the region is completely inside of the file;
813 * (B-x): x bytes of region are inside of the file, the rest is
816 * (C): the region is completely outside of the file.
818 * This classification is stable under DLM lock already acquired by
819 * the caller, because to change the class, other client has to take
820 * DLM lock conflicting with our lock. Also, any updates to ->i_size
821 * by other threads on this client are serialized by
822 * ll_inode_size_lock(). This guarantees that short reads are handled
823 * correctly in the face of concurrent writes and truncates.
825 ccc_object_size_lock(obj);
826 result = cl_object_attr_get(env, obj, attr);
831 * A glimpse is necessary to determine whether we
832 * return a short read (B) or some zeroes at the end
835 ccc_object_size_unlock(obj);
836 result = cl_glimpse_lock(env, io, inode, obj, 0);
837 if (result == 0 && exceed != NULL) {
838 /* If objective page index exceed end-of-file
839 * page index, return directly. Do not expect
840 * kernel will check such case correctly.
841 * linux-2.6.18-128.1.1 miss to do that.
843 loff_t size = cl_isize_read(inode);
844 unsigned long cur_index = start >>
847 if ((size == 0 && cur_index != 0) ||
848 (((size - 1) >> PAGE_CACHE_SHIFT) <
855 * region is within kms and, hence, within real file
856 * size (A). We need to increase i_size to cover the
857 * read region so that generic_file_read() will do its
858 * job, but that doesn't mean the kms size is
859 * _correct_, it is only the _minimum_ size. If
860 * someone does a stat they will get the correct size
861 * which will always be >= the kms value here.
864 if (cl_isize_read(inode) < kms) {
865 cl_isize_write_nolock(inode, kms);
867 DFID" updating i_size "LPU64"\n",
868 PFID(lu_object_fid(&obj->co_lu)),
869 (__u64)cl_isize_read(inode));
874 ccc_object_size_unlock(obj);
878 /*****************************************************************************
880 * Transfer operations.
884 void ccc_req_completion(const struct lu_env *env,
885 const struct cl_req_slice *slice, int ioret)
890 cl_stats_tally(slice->crs_dev, slice->crs_req->crq_type, ioret);
892 vrq = cl2ccc_req(slice);
893 OBD_SLAB_FREE_PTR(vrq, ccc_req_kmem);
897 * Implementation of struct cl_req_operations::cro_attr_set() for ccc
898 * layer. ccc is responsible for
916 void ccc_req_attr_set(const struct lu_env *env,
917 const struct cl_req_slice *slice,
918 const struct cl_object *obj,
919 struct cl_req_attr *attr, obd_valid flags)
923 obd_flag valid_flags;
926 inode = ccc_object_inode(obj);
927 valid_flags = OBD_MD_FLTYPE;
929 if ((flags & OBD_MD_FLOSSCAPA) != 0) {
930 LASSERT(attr->cra_capa == NULL);
931 attr->cra_capa = cl_capa_lookup(inode,
932 slice->crs_req->crq_type);
935 if (slice->crs_req->crq_type == CRT_WRITE) {
936 if (flags & OBD_MD_FLEPOCH) {
937 oa->o_valid |= OBD_MD_FLEPOCH;
938 oa->o_ioepoch = cl_i2info(inode)->lli_ioepoch;
939 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME |
940 OBD_MD_FLUID | OBD_MD_FLGID;
943 obdo_from_inode(oa, inode, valid_flags & flags);
944 obdo_set_parent_fid(oa, &cl_i2info(inode)->lli_fid);
945 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_INVALID_PFID))
947 memcpy(attr->cra_jobid, cl_i2info(inode)->lli_jobid,
948 JOBSTATS_JOBID_SIZE);
951 static const struct cl_req_operations ccc_req_ops = {
952 .cro_attr_set = ccc_req_attr_set,
953 .cro_completion = ccc_req_completion
956 int cl_setattr_ost(struct inode *inode, const struct iattr *attr,
957 struct obd_capa *capa)
966 env = cl_env_get(&refcheck);
968 RETURN(PTR_ERR(env));
970 io = ccc_env_thread_io(env);
971 io->ci_obj = cl_i2info(inode)->lli_clob;
973 io->u.ci_setattr.sa_attr.lvb_atime = LTIME_S(attr->ia_atime);
974 io->u.ci_setattr.sa_attr.lvb_mtime = LTIME_S(attr->ia_mtime);
975 io->u.ci_setattr.sa_attr.lvb_ctime = LTIME_S(attr->ia_ctime);
976 io->u.ci_setattr.sa_attr.lvb_size = attr->ia_size;
977 io->u.ci_setattr.sa_valid = attr->ia_valid;
978 io->u.ci_setattr.sa_capa = capa;
981 if (cl_io_init(env, io, CIT_SETATTR, io->ci_obj) == 0) {
982 struct ccc_io *cio = ccc_env_io(env);
984 if (attr->ia_valid & ATTR_FILE)
985 /* populate the file descriptor for ftruncate to honor
986 * group lock - see LU-787 */
987 cio->cui_fd = cl_iattr2fd(inode, attr);
989 result = cl_io_loop(env, io);
991 result = io->ci_result;
994 if (unlikely(io->ci_need_restart))
996 /* HSM import case: file is released, cannot be restored
997 * no need to fail except if restore registration failed
999 if (result == -ENODATA && io->ci_restore_needed &&
1000 io->ci_result != -ENODATA)
1002 cl_env_put(env, &refcheck);
1006 /*****************************************************************************
1012 struct lu_device *ccc2lu_dev(struct ccc_device *vdv)
1014 return &vdv->cdv_cl.cd_lu_dev;
1017 struct ccc_device *lu2ccc_dev(const struct lu_device *d)
1019 return container_of0(d, struct ccc_device, cdv_cl.cd_lu_dev);
1022 struct ccc_device *cl2ccc_dev(const struct cl_device *d)
1024 return container_of0(d, struct ccc_device, cdv_cl);
1027 struct lu_object *ccc2lu(struct ccc_object *vob)
1029 return &vob->cob_cl.co_lu;
1032 struct ccc_object *lu2ccc(const struct lu_object *obj)
1034 return container_of0(obj, struct ccc_object, cob_cl.co_lu);
1037 struct ccc_object *cl2ccc(const struct cl_object *obj)
1039 return container_of0(obj, struct ccc_object, cob_cl);
1042 struct ccc_lock *cl2ccc_lock(const struct cl_lock_slice *slice)
1044 return container_of(slice, struct ccc_lock, clk_cl);
1047 struct ccc_io *cl2ccc_io(const struct lu_env *env,
1048 const struct cl_io_slice *slice)
1052 cio = container_of(slice, struct ccc_io, cui_cl);
1053 LASSERT(cio == ccc_env_io(env));
1057 struct ccc_req *cl2ccc_req(const struct cl_req_slice *slice)
1059 return container_of0(slice, struct ccc_req, crq_cl);
1062 struct page *cl2vm_page(const struct cl_page_slice *slice)
1064 return cl2ccc_page(slice)->cpg_page;
1067 /*****************************************************************************
1072 int ccc_object_invariant(const struct cl_object *obj)
1074 struct inode *inode = ccc_object_inode(obj);
1075 struct cl_inode_info *lli = cl_i2info(inode);
1077 return (S_ISREG(cl_inode_mode(inode)) ||
1078 /* i_mode of unlinked inode is zeroed. */
1079 cl_inode_mode(inode) == 0) && lli->lli_clob == obj;
1082 struct inode *ccc_object_inode(const struct cl_object *obj)
1084 return cl2ccc(obj)->cob_inode;
1088 * Returns a pointer to cl_page associated with \a vmpage, without acquiring
1089 * additional reference to the resulting page. This is an unsafe version of
1090 * cl_vmpage_page() that can only be used under vmpage lock.
1092 struct cl_page *ccc_vmpage_page_transient(struct page *vmpage)
1094 KLASSERT(PageLocked(vmpage));
1095 return (struct cl_page *)vmpage->private;
1099 * Initialize or update CLIO structures for regular files when new
1100 * meta-data arrives from the server.
1102 * \param inode regular file inode
1103 * \param md new file metadata from MDS
1104 * - allocates cl_object if necessary,
1105 * - updated layout, if object was already here.
1107 int cl_file_inode_init(struct inode *inode, struct lustre_md *md)
1110 struct cl_inode_info *lli;
1111 struct cl_object *clob;
1112 struct lu_site *site;
1114 struct cl_object_conf conf = {
1123 LASSERT(md->body->mbo_valid & OBD_MD_FLID);
1124 LASSERT(S_ISREG(cl_inode_mode(inode)));
1126 env = cl_env_get(&refcheck);
1128 return PTR_ERR(env);
1130 site = cl_i2sbi(inode)->ll_site;
1131 lli = cl_i2info(inode);
1132 fid = &lli->lli_fid;
1133 LASSERT(fid_is_sane(fid));
1135 if (lli->lli_clob == NULL) {
1136 /* clob is slave of inode, empty lli_clob means for new inode,
1137 * there is no clob in cache with the given fid, so it is
1138 * unnecessary to perform lookup-alloc-lookup-insert, just
1139 * alloc and insert directly. */
1140 LASSERT(inode->i_state & I_NEW);
1141 conf.coc_lu.loc_flags = LOC_F_NEW;
1142 clob = cl_object_find(env, lu2cl_dev(site->ls_top_dev),
1144 if (!IS_ERR(clob)) {
1146 * No locking is necessary, as new inode is
1147 * locked by I_NEW bit.
1149 lli->lli_clob = clob;
1150 lli->lli_has_smd = lsm_has_objects(md->lsm);
1151 lu_object_ref_add(&clob->co_lu, "inode", inode);
1153 result = PTR_ERR(clob);
1155 result = cl_conf_set(env, lli->lli_clob, &conf);
1158 cl_env_put(env, &refcheck);
1161 CERROR("Failure to initialize cl object "DFID": %d\n",
1167 * Wait for others drop their references of the object at first, then we drop
1168 * the last one, which will lead to the object be destroyed immediately.
1169 * Must be called after cl_object_kill() against this object.
1171 * The reason we want to do this is: destroying top object will wait for sub
1172 * objects being destroyed first, so we can't let bottom layer (e.g. from ASTs)
1173 * to initiate top object destroying which may deadlock. See bz22520.
1175 static void cl_object_put_last(struct lu_env *env, struct cl_object *obj)
1177 struct lu_object_header *header = obj->co_lu.lo_header;
1178 wait_queue_t waiter;
1180 if (unlikely(atomic_read(&header->loh_ref) != 1)) {
1181 struct lu_site *site = obj->co_lu.lo_dev->ld_site;
1182 struct lu_site_bkt_data *bkt;
1184 bkt = lu_site_bkt_from_fid(site, &header->loh_fid);
1186 init_waitqueue_entry_current(&waiter);
1187 add_wait_queue(&bkt->lsb_marche_funebre, &waiter);
1190 set_current_state(TASK_UNINTERRUPTIBLE);
1191 if (atomic_read(&header->loh_ref) == 1)
1193 waitq_wait(&waiter, TASK_UNINTERRUPTIBLE);
1196 set_current_state(TASK_RUNNING);
1197 remove_wait_queue(&bkt->lsb_marche_funebre, &waiter);
1200 cl_object_put(env, obj);
1203 void cl_inode_fini(struct inode *inode)
1206 struct cl_inode_info *lli = cl_i2info(inode);
1207 struct cl_object *clob = lli->lli_clob;
1214 cookie = cl_env_reenter();
1215 env = cl_env_get(&refcheck);
1216 emergency = IS_ERR(env);
1218 mutex_lock(&ccc_inode_fini_guard);
1219 LASSERT(ccc_inode_fini_env != NULL);
1220 cl_env_implant(ccc_inode_fini_env, &refcheck);
1221 env = ccc_inode_fini_env;
1224 * cl_object cache is a slave to inode cache (which, in turn
1225 * is a slave to dentry cache), don't keep cl_object in memory
1226 * when its master is evicted.
1228 cl_object_kill(env, clob);
1229 lu_object_ref_del(&clob->co_lu, "inode", inode);
1230 cl_object_put_last(env, clob);
1231 lli->lli_clob = NULL;
1233 cl_env_unplant(ccc_inode_fini_env, &refcheck);
1234 mutex_unlock(&ccc_inode_fini_guard);
1236 cl_env_put(env, &refcheck);
1237 cl_env_reexit(cookie);
1242 * return IF_* type for given lu_dirent entry.
1243 * IF_* flag shld be converted to particular OS file type in
1244 * platform llite module.
1246 __u16 ll_dirent_type_get(struct lu_dirent *ent)
1249 struct luda_type *lt;
1252 if (le32_to_cpu(ent->lde_attrs) & LUDA_TYPE) {
1253 const unsigned align = sizeof(struct luda_type) - 1;
1255 len = le16_to_cpu(ent->lde_namelen);
1256 len = (len + align) & ~align;
1257 lt = (void *)ent->lde_name + len;
1258 type = IFTODT(le16_to_cpu(lt->lt_type));
1264 * build inode number from passed @fid */
1265 __u64 cl_fid_build_ino(const struct lu_fid *fid, int api32)
1267 if (BITS_PER_LONG == 32 || api32)
1268 RETURN(fid_flatten32(fid));
1270 RETURN(fid_flatten(fid));
1274 * build inode generation from passed @fid. If our FID overflows the 32-bit
1275 * inode number then return a non-zero generation to distinguish them. */
1276 __u32 cl_fid_build_gen(const struct lu_fid *fid)
1281 if (fid_is_igif(fid)) {
1282 gen = lu_igif_gen(fid);
1286 gen = (fid_flatten(fid) >> 32);
1290 /* lsm is unreliable after hsm implementation as layout can be changed at
1291 * any time. This is only to support old, non-clio-ized interfaces. It will
1292 * cause deadlock if clio operations are called with this extra layout refcount
1293 * because in case the layout changed during the IO, ll_layout_refresh() will
1294 * have to wait for the refcount to become zero to destroy the older layout.
1296 * Notice that the lsm returned by this function may not be valid unless called
1297 * inside layout lock - MDS_INODELOCK_LAYOUT. */
1298 struct lov_stripe_md *ccc_inode_lsm_get(struct inode *inode)
1300 return lov_lsm_get(cl_i2info(inode)->lli_clob);
1303 void inline ccc_inode_lsm_put(struct inode *inode, struct lov_stripe_md *lsm)
1305 lov_lsm_put(cl_i2info(inode)->lli_clob, lsm);