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
45 # include <libcfs/libcfs.h>
46 # include <linux/fs.h>
47 # include <linux/sched.h>
48 # include <linux/mm.h>
49 # include <linux/quotaops.h>
50 # include <linux/highmem.h>
51 # include <linux/pagemap.h>
52 # include <linux/rbtree.h>
53 #else /* __KERNEL__ */
58 #include <sys/types.h>
60 #include <sys/queue.h>
62 # include <liblustre.h>
66 #include <obd_support.h>
67 #include <lustre_fid.h>
68 #include <lustre_lite.h>
69 #include <lustre_dlm.h>
70 #include <lustre_ver.h>
71 #include <lustre_mdc.h>
72 #include <cl_object.h>
77 #include "../llite/llite_internal.h"
79 #include "../liblustre/llite_lib.h"
82 static const struct cl_req_operations ccc_req_ops;
85 * ccc_ prefix stands for "Common Client Code".
88 static struct kmem_cache *ccc_lock_kmem;
89 static struct kmem_cache *ccc_object_kmem;
90 static struct kmem_cache *ccc_thread_kmem;
91 static struct kmem_cache *ccc_session_kmem;
92 static struct kmem_cache *ccc_req_kmem;
94 static struct lu_kmem_descr ccc_caches[] = {
96 .ckd_cache = &ccc_lock_kmem,
97 .ckd_name = "ccc_lock_kmem",
98 .ckd_size = sizeof (struct ccc_lock)
101 .ckd_cache = &ccc_object_kmem,
102 .ckd_name = "ccc_object_kmem",
103 .ckd_size = sizeof (struct ccc_object)
106 .ckd_cache = &ccc_thread_kmem,
107 .ckd_name = "ccc_thread_kmem",
108 .ckd_size = sizeof (struct ccc_thread_info),
111 .ckd_cache = &ccc_session_kmem,
112 .ckd_name = "ccc_session_kmem",
113 .ckd_size = sizeof (struct ccc_session)
116 .ckd_cache = &ccc_req_kmem,
117 .ckd_name = "ccc_req_kmem",
118 .ckd_size = sizeof (struct ccc_req)
125 /*****************************************************************************
127 * Vvp device and device type functions.
131 void *ccc_key_init(const struct lu_context *ctx, struct lu_context_key *key)
133 struct ccc_thread_info *info;
135 OBD_SLAB_ALLOC_PTR_GFP(info, ccc_thread_kmem, GFP_NOFS);
137 info = ERR_PTR(-ENOMEM);
141 void ccc_key_fini(const struct lu_context *ctx,
142 struct lu_context_key *key, void *data)
144 struct ccc_thread_info *info = data;
145 OBD_SLAB_FREE_PTR(info, ccc_thread_kmem);
148 void *ccc_session_key_init(const struct lu_context *ctx,
149 struct lu_context_key *key)
151 struct ccc_session *session;
153 OBD_SLAB_ALLOC_PTR_GFP(session, ccc_session_kmem, GFP_NOFS);
155 session = ERR_PTR(-ENOMEM);
159 void ccc_session_key_fini(const struct lu_context *ctx,
160 struct lu_context_key *key, void *data)
162 struct ccc_session *session = data;
163 OBD_SLAB_FREE_PTR(session, ccc_session_kmem);
166 struct lu_context_key ccc_key = {
167 .lct_tags = LCT_CL_THREAD,
168 .lct_init = ccc_key_init,
169 .lct_fini = ccc_key_fini
172 struct lu_context_key ccc_session_key = {
173 .lct_tags = LCT_SESSION,
174 .lct_init = ccc_session_key_init,
175 .lct_fini = ccc_session_key_fini
179 /* type constructor/destructor: ccc_type_{init,fini,start,stop}(). */
180 // LU_TYPE_INIT_FINI(ccc, &ccc_key, &ccc_session_key);
182 int ccc_device_init(const struct lu_env *env, struct lu_device *d,
183 const char *name, struct lu_device *next)
185 struct ccc_device *vdv;
190 vdv->cdv_next = lu2cl_dev(next);
192 LASSERT(d->ld_site != NULL && next->ld_type != NULL);
193 next->ld_site = d->ld_site;
194 rc = next->ld_type->ldt_ops->ldto_device_init(
195 env, next, next->ld_type->ldt_name, NULL);
198 lu_ref_add(&next->ld_reference, "lu-stack", &lu_site_init);
203 struct lu_device *ccc_device_fini(const struct lu_env *env,
206 return cl2lu_dev(lu2ccc_dev(d)->cdv_next);
209 struct lu_device *ccc_device_alloc(const struct lu_env *env,
210 struct lu_device_type *t,
211 struct lustre_cfg *cfg,
212 const struct lu_device_operations *luops,
213 const struct cl_device_operations *clops)
215 struct ccc_device *vdv;
216 struct lu_device *lud;
217 struct cl_site *site;
223 RETURN(ERR_PTR(-ENOMEM));
225 lud = &vdv->cdv_cl.cd_lu_dev;
226 cl_device_init(&vdv->cdv_cl, t);
227 ccc2lu_dev(vdv)->ld_ops = luops;
228 vdv->cdv_cl.cd_ops = clops;
232 rc = cl_site_init(site, &vdv->cdv_cl);
234 rc = lu_site_init_finish(&site->cs_lu);
236 LASSERT(lud->ld_site == NULL);
237 CERROR("Cannot init lu_site, rc %d.\n", rc);
243 ccc_device_free(env, lud);
249 struct lu_device *ccc_device_free(const struct lu_env *env,
252 struct ccc_device *vdv = lu2ccc_dev(d);
253 struct cl_site *site = lu2cl_site(d->ld_site);
254 struct lu_device *next = cl2lu_dev(vdv->cdv_next);
256 if (d->ld_site != NULL) {
260 cl_device_fini(lu2cl_dev(d));
265 int ccc_req_init(const struct lu_env *env, struct cl_device *dev,
271 OBD_SLAB_ALLOC_PTR_GFP(vrq, ccc_req_kmem, GFP_NOFS);
273 cl_req_slice_add(req, &vrq->crq_cl, dev, &ccc_req_ops);
281 * An `emergency' environment used by ccc_inode_fini() when cl_env_get()
282 * fails. Access to this environment is serialized by ccc_inode_fini_guard
285 static struct lu_env *ccc_inode_fini_env = NULL;
288 * A mutex serializing calls to slp_inode_fini() under extreme memory
289 * pressure, when environments cannot be allocated.
291 static DEFINE_MUTEX(ccc_inode_fini_guard);
292 static int dummy_refcheck;
294 int ccc_global_init(struct lu_device_type *device_type)
298 result = lu_kmem_init(ccc_caches);
302 result = lu_device_type_init(device_type);
306 ccc_inode_fini_env = cl_env_alloc(&dummy_refcheck,
307 LCT_REMEMBER|LCT_NOREF);
308 if (IS_ERR(ccc_inode_fini_env)) {
309 result = PTR_ERR(ccc_inode_fini_env);
313 ccc_inode_fini_env->le_ctx.lc_cookie = 0x4;
316 lu_device_type_fini(device_type);
318 lu_kmem_fini(ccc_caches);
322 void ccc_global_fini(struct lu_device_type *device_type)
324 if (ccc_inode_fini_env != NULL) {
325 cl_env_put(ccc_inode_fini_env, &dummy_refcheck);
326 ccc_inode_fini_env = NULL;
328 lu_device_type_fini(device_type);
329 lu_kmem_fini(ccc_caches);
332 /*****************************************************************************
338 struct lu_object *ccc_object_alloc(const struct lu_env *env,
339 const struct lu_object_header *unused,
340 struct lu_device *dev,
341 const struct cl_object_operations *clops,
342 const struct lu_object_operations *luops)
344 struct ccc_object *vob;
345 struct lu_object *obj;
347 OBD_SLAB_ALLOC_PTR_GFP(vob, ccc_object_kmem, GFP_NOFS);
349 struct cl_object_header *hdr;
352 hdr = &vob->cob_header;
353 cl_object_header_init(hdr);
354 hdr->coh_page_bufsize = cfs_size_round(sizeof(struct cl_page));
356 lu_object_init(obj, &hdr->coh_lu, dev);
357 lu_object_add_top(&hdr->coh_lu, obj);
359 vob->cob_cl.co_ops = clops;
366 int ccc_object_init0(const struct lu_env *env,
367 struct ccc_object *vob,
368 const struct cl_object_conf *conf)
370 vob->cob_inode = conf->coc_inode;
371 vob->cob_transient_pages = 0;
372 cl_object_page_init(&vob->cob_cl, sizeof(struct ccc_page));
376 int ccc_object_init(const struct lu_env *env, struct lu_object *obj,
377 const struct lu_object_conf *conf)
379 struct ccc_device *dev = lu2ccc_dev(obj->lo_dev);
380 struct ccc_object *vob = lu2ccc(obj);
381 struct lu_object *below;
382 struct lu_device *under;
385 under = &dev->cdv_next->cd_lu_dev;
386 below = under->ld_ops->ldo_object_alloc(env, obj->lo_header, under);
388 const struct cl_object_conf *cconf;
390 cconf = lu2cl_conf(conf);
391 CFS_INIT_LIST_HEAD(&vob->cob_pending_list);
392 lu_object_add(obj, below);
393 result = ccc_object_init0(env, vob, cconf);
399 void ccc_object_free(const struct lu_env *env, struct lu_object *obj)
401 struct ccc_object *vob = lu2ccc(obj);
404 lu_object_header_fini(obj->lo_header);
405 OBD_SLAB_FREE_PTR(vob, ccc_object_kmem);
408 int ccc_lock_init(const struct lu_env *env,
409 struct cl_object *obj, struct cl_lock *lock,
410 const struct cl_io *unused,
411 const struct cl_lock_operations *lkops)
413 struct ccc_lock *clk;
416 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
418 OBD_SLAB_ALLOC_PTR_GFP(clk, ccc_lock_kmem, GFP_NOFS);
420 cl_lock_slice_add(lock, &clk->clk_cl, obj, lkops);
427 int ccc_attr_set(const struct lu_env *env, struct cl_object *obj,
428 const struct cl_attr *attr, unsigned valid)
433 int ccc_object_glimpse(const struct lu_env *env,
434 const struct cl_object *obj, struct ost_lvb *lvb)
436 struct inode *inode = ccc_object_inode(obj);
439 lvb->lvb_mtime = cl_inode_mtime(inode);
440 lvb->lvb_atime = cl_inode_atime(inode);
441 lvb->lvb_ctime = cl_inode_ctime(inode);
443 * LU-417: Add dirty pages block count lest i_blocks reports 0, some
444 * "cp" or "tar" on remote node may think it's a completely sparse file
447 if (lvb->lvb_size > 0 && lvb->lvb_blocks == 0)
448 lvb->lvb_blocks = dirty_cnt(inode);
454 int ccc_conf_set(const struct lu_env *env, struct cl_object *obj,
455 const struct cl_object_conf *conf)
457 /* TODO: destroy all pages attached to this object. */
461 static void ccc_object_size_lock(struct cl_object *obj)
463 struct inode *inode = ccc_object_inode(obj);
465 cl_isize_lock(inode);
466 cl_object_attr_lock(obj);
469 static void ccc_object_size_unlock(struct cl_object *obj)
471 struct inode *inode = ccc_object_inode(obj);
473 cl_object_attr_unlock(obj);
474 cl_isize_unlock(inode);
477 /*****************************************************************************
483 int ccc_fail(const struct lu_env *env, const struct cl_page_slice *slice)
492 void ccc_transient_page_verify(const struct cl_page *page)
496 int ccc_transient_page_own(const struct lu_env *env,
497 const struct cl_page_slice *slice,
498 struct cl_io *unused,
501 ccc_transient_page_verify(slice->cpl_page);
505 void ccc_transient_page_assume(const struct lu_env *env,
506 const struct cl_page_slice *slice,
507 struct cl_io *unused)
509 ccc_transient_page_verify(slice->cpl_page);
512 void ccc_transient_page_unassume(const struct lu_env *env,
513 const struct cl_page_slice *slice,
514 struct cl_io *unused)
516 ccc_transient_page_verify(slice->cpl_page);
519 void ccc_transient_page_disown(const struct lu_env *env,
520 const struct cl_page_slice *slice,
521 struct cl_io *unused)
523 ccc_transient_page_verify(slice->cpl_page);
526 void ccc_transient_page_discard(const struct lu_env *env,
527 const struct cl_page_slice *slice,
528 struct cl_io *unused)
530 struct cl_page *page = slice->cpl_page;
532 ccc_transient_page_verify(slice->cpl_page);
535 * For transient pages, remove it from the radix tree.
537 cl_page_delete(env, page);
540 int ccc_transient_page_prep(const struct lu_env *env,
541 const struct cl_page_slice *slice,
542 struct cl_io *unused)
545 /* transient page should always be sent. */
549 /*****************************************************************************
555 void ccc_lock_delete(const struct lu_env *env,
556 const struct cl_lock_slice *slice)
558 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
561 void ccc_lock_fini(const struct lu_env *env, struct cl_lock_slice *slice)
563 struct ccc_lock *clk = cl2ccc_lock(slice);
564 OBD_SLAB_FREE_PTR(clk, ccc_lock_kmem);
567 int ccc_lock_enqueue(const struct lu_env *env,
568 const struct cl_lock_slice *slice,
569 struct cl_io *unused, __u32 enqflags)
571 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
575 int ccc_lock_use(const struct lu_env *env, const struct cl_lock_slice *slice)
577 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
581 int ccc_lock_unuse(const struct lu_env *env, const struct cl_lock_slice *slice)
583 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
587 int ccc_lock_wait(const struct lu_env *env, const struct cl_lock_slice *slice)
589 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
594 * Implementation of cl_lock_operations::clo_fits_into() methods for ccc
595 * layer. This function is executed every time io finds an existing lock in
596 * the lock cache while creating new lock. This function has to decide whether
597 * cached lock "fits" into io.
599 * \param slice lock to be checked
600 * \param io IO that wants a lock.
602 * \see lov_lock_fits_into().
604 int ccc_lock_fits_into(const struct lu_env *env,
605 const struct cl_lock_slice *slice,
606 const struct cl_lock_descr *need,
607 const struct cl_io *io)
609 const struct cl_lock *lock = slice->cls_lock;
610 const struct cl_lock_descr *descr = &lock->cll_descr;
611 const struct ccc_io *cio = ccc_env_io(env);
616 * Work around DLM peculiarity: it assumes that glimpse
617 * (LDLM_FL_HAS_INTENT) lock is always LCK_PR, and returns reads lock
618 * when asked for LCK_PW lock with LDLM_FL_HAS_INTENT flag set. Make
619 * sure that glimpse doesn't get CLM_WRITE top-lock, so that it
620 * doesn't enqueue CLM_WRITE sub-locks.
622 if (cio->cui_glimpse)
623 result = descr->cld_mode != CLM_WRITE;
626 * Also, don't match incomplete write locks for read, otherwise read
627 * would enqueue missing sub-locks in the write mode.
629 else if (need->cld_mode != descr->cld_mode)
630 result = lock->cll_state >= CLS_ENQUEUED;
637 * Implements cl_lock_operations::clo_state() method for ccc layer, invoked
638 * whenever lock state changes. Transfers object attributes, that might be
639 * updated as a result of lock acquiring into inode.
641 void ccc_lock_state(const struct lu_env *env,
642 const struct cl_lock_slice *slice,
643 enum cl_lock_state state)
645 struct cl_lock *lock = slice->cls_lock;
649 * Refresh inode attributes when the lock is moving into CLS_HELD
650 * state, and only when this is a result of real enqueue, rather than
651 * of finding lock in the cache.
653 if (state == CLS_HELD && lock->cll_state < CLS_HELD) {
654 struct cl_object *obj;
657 obj = slice->cls_obj;
658 inode = ccc_object_inode(obj);
660 /* vmtruncate() sets the i_size
661 * under both a DLM lock and the
662 * ll_inode_size_lock(). If we don't get the
663 * ll_inode_size_lock() here we can match the DLM lock and
664 * reset i_size. generic_file_write can then trust the
665 * stale i_size when doing appending writes and effectively
666 * cancel the result of the truncate. Getting the
667 * ll_inode_size_lock() after the enqueue maintains the DLM
668 * -> ll_inode_size_lock() acquiring order. */
669 if (lock->cll_descr.cld_start == 0 &&
670 lock->cll_descr.cld_end == CL_PAGE_EOF)
671 cl_merge_lvb(env, inode);
676 /*****************************************************************************
682 void ccc_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
684 struct cl_io *io = ios->cis_io;
686 CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj));
689 int ccc_io_one_lock_index(const struct lu_env *env, struct cl_io *io,
690 __u32 enqflags, enum cl_lock_mode mode,
691 pgoff_t start, pgoff_t end)
693 struct ccc_io *cio = ccc_env_io(env);
694 struct cl_lock_descr *descr = &cio->cui_link.cill_descr;
695 struct cl_object *obj = io->ci_obj;
697 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
700 CDEBUG(D_VFSTRACE, "lock: %d [%lu, %lu]\n", mode, start, end);
702 memset(&cio->cui_link, 0, sizeof cio->cui_link);
704 if (cio->cui_fd && (cio->cui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
705 descr->cld_mode = CLM_GROUP;
706 descr->cld_gid = cio->cui_fd->fd_grouplock.cg_gid;
708 descr->cld_mode = mode;
710 descr->cld_obj = obj;
711 descr->cld_start = start;
712 descr->cld_end = end;
713 descr->cld_enq_flags = enqflags;
715 cl_io_lock_add(env, io, &cio->cui_link);
719 void ccc_io_update_iov(const struct lu_env *env,
720 struct ccc_io *cio, struct cl_io *io)
723 size_t size = io->u.ci_rw.crw_count;
725 cio->cui_iov_olen = 0;
726 if (!cl_is_normalio(env, io) || cio->cui_tot_nrsegs == 0)
729 for (i = 0; i < cio->cui_tot_nrsegs; i++) {
730 struct iovec *iv = &cio->cui_iov[i];
732 if (iv->iov_len < size)
735 if (iv->iov_len > size) {
736 cio->cui_iov_olen = iv->iov_len;
743 cio->cui_nrsegs = i + 1;
744 LASSERTF(cio->cui_tot_nrsegs >= cio->cui_nrsegs,
745 "tot_nrsegs: %lu, nrsegs: %lu\n",
746 cio->cui_tot_nrsegs, cio->cui_nrsegs);
749 int ccc_io_one_lock(const struct lu_env *env, struct cl_io *io,
750 __u32 enqflags, enum cl_lock_mode mode,
751 loff_t start, loff_t end)
753 struct cl_object *obj = io->ci_obj;
754 return ccc_io_one_lock_index(env, io, enqflags, mode,
755 cl_index(obj, start), cl_index(obj, end));
758 void ccc_io_end(const struct lu_env *env, const struct cl_io_slice *ios)
760 CLOBINVRNT(env, ios->cis_io->ci_obj,
761 ccc_object_invariant(ios->cis_io->ci_obj));
764 void ccc_io_advance(const struct lu_env *env,
765 const struct cl_io_slice *ios,
768 struct ccc_io *cio = cl2ccc_io(env, ios);
769 struct cl_io *io = ios->cis_io;
770 struct cl_object *obj = ios->cis_io->ci_obj;
772 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
774 if (!cl_is_normalio(env, io))
777 LASSERT(cio->cui_tot_nrsegs >= cio->cui_nrsegs);
778 LASSERT(cio->cui_tot_count >= nob);
780 cio->cui_iov += cio->cui_nrsegs;
781 cio->cui_tot_nrsegs -= cio->cui_nrsegs;
782 cio->cui_tot_count -= nob;
785 if (cio->cui_iov_olen > 0) {
789 cio->cui_tot_nrsegs++;
790 iv = &cio->cui_iov[0];
791 if (io->ci_continue) {
792 iv->iov_base += iv->iov_len;
793 LASSERT(cio->cui_iov_olen > iv->iov_len);
794 iv->iov_len = cio->cui_iov_olen - iv->iov_len;
796 /* restore the iov_len, in case of restart io. */
797 iv->iov_len = cio->cui_iov_olen;
799 cio->cui_iov_olen = 0;
804 * Helper function that if necessary adjusts file size (inode->i_size), when
805 * position at the offset \a pos is accessed. File size can be arbitrary stale
806 * on a Lustre client, but client at least knows KMS. If accessed area is
807 * inside [0, KMS], set file size to KMS, otherwise glimpse file size.
809 * Locking: cl_isize_lock is used to serialize changes to inode size and to
810 * protect consistency between inode size and cl_object
811 * attributes. cl_object_size_lock() protects consistency between cl_attr's of
812 * top-object and sub-objects.
814 int ccc_prep_size(const struct lu_env *env, struct cl_object *obj,
815 struct cl_io *io, loff_t start, size_t count, int *exceed)
817 struct cl_attr *attr = ccc_env_thread_attr(env);
818 struct inode *inode = ccc_object_inode(obj);
819 loff_t pos = start + count - 1;
824 * Consistency guarantees: following possibilities exist for the
825 * relation between region being accessed and real file size at this
828 * (A): the region is completely inside of the file;
830 * (B-x): x bytes of region are inside of the file, the rest is
833 * (C): the region is completely outside of the file.
835 * This classification is stable under DLM lock already acquired by
836 * the caller, because to change the class, other client has to take
837 * DLM lock conflicting with our lock. Also, any updates to ->i_size
838 * by other threads on this client are serialized by
839 * ll_inode_size_lock(). This guarantees that short reads are handled
840 * correctly in the face of concurrent writes and truncates.
842 ccc_object_size_lock(obj);
843 result = cl_object_attr_get(env, obj, attr);
848 * A glimpse is necessary to determine whether we
849 * return a short read (B) or some zeroes at the end
852 ccc_object_size_unlock(obj);
853 result = cl_glimpse_lock(env, io, inode, obj, 0);
854 if (result == 0 && exceed != NULL) {
855 /* If objective page index exceed end-of-file
856 * page index, return directly. Do not expect
857 * kernel will check such case correctly.
858 * linux-2.6.18-128.1.1 miss to do that.
860 loff_t size = cl_isize_read(inode);
861 unsigned long cur_index = start >>
864 if ((size == 0 && cur_index != 0) ||
865 (((size - 1) >> PAGE_CACHE_SHIFT) <
872 * region is within kms and, hence, within real file
873 * size (A). We need to increase i_size to cover the
874 * read region so that generic_file_read() will do its
875 * job, but that doesn't mean the kms size is
876 * _correct_, it is only the _minimum_ size. If
877 * someone does a stat they will get the correct size
878 * which will always be >= the kms value here.
881 if (cl_isize_read(inode) < kms) {
882 cl_isize_write_nolock(inode, kms);
884 DFID" updating i_size "LPU64"\n",
885 PFID(lu_object_fid(&obj->co_lu)),
886 (__u64)cl_isize_read(inode));
891 ccc_object_size_unlock(obj);
895 /*****************************************************************************
897 * Transfer operations.
901 void ccc_req_completion(const struct lu_env *env,
902 const struct cl_req_slice *slice, int ioret)
907 cl_stats_tally(slice->crs_dev, slice->crs_req->crq_type, ioret);
909 vrq = cl2ccc_req(slice);
910 OBD_SLAB_FREE_PTR(vrq, ccc_req_kmem);
914 * Implementation of struct cl_req_operations::cro_attr_set() for ccc
915 * layer. ccc is responsible for
933 void ccc_req_attr_set(const struct lu_env *env,
934 const struct cl_req_slice *slice,
935 const struct cl_object *obj,
936 struct cl_req_attr *attr, obd_valid flags)
940 obd_flag valid_flags;
943 inode = ccc_object_inode(obj);
944 valid_flags = OBD_MD_FLTYPE;
946 if ((flags & OBD_MD_FLOSSCAPA) != 0) {
947 LASSERT(attr->cra_capa == NULL);
948 attr->cra_capa = cl_capa_lookup(inode,
949 slice->crs_req->crq_type);
952 if (slice->crs_req->crq_type == CRT_WRITE) {
953 if (flags & OBD_MD_FLEPOCH) {
954 oa->o_valid |= OBD_MD_FLEPOCH;
955 oa->o_ioepoch = cl_i2info(inode)->lli_ioepoch;
956 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME |
957 OBD_MD_FLUID | OBD_MD_FLGID;
960 obdo_from_inode(oa, inode, valid_flags & flags);
961 obdo_set_parent_fid(oa, &cl_i2info(inode)->lli_fid);
962 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_INVALID_PFID))
965 memcpy(attr->cra_jobid, cl_i2info(inode)->lli_jobid,
966 JOBSTATS_JOBID_SIZE);
970 static const struct cl_req_operations ccc_req_ops = {
971 .cro_attr_set = ccc_req_attr_set,
972 .cro_completion = ccc_req_completion
975 int cl_setattr_ost(struct inode *inode, const struct iattr *attr,
976 struct obd_capa *capa)
985 env = cl_env_get(&refcheck);
987 RETURN(PTR_ERR(env));
989 io = ccc_env_thread_io(env);
990 io->ci_obj = cl_i2info(inode)->lli_clob;
992 io->u.ci_setattr.sa_attr.lvb_atime = LTIME_S(attr->ia_atime);
993 io->u.ci_setattr.sa_attr.lvb_mtime = LTIME_S(attr->ia_mtime);
994 io->u.ci_setattr.sa_attr.lvb_ctime = LTIME_S(attr->ia_ctime);
995 io->u.ci_setattr.sa_attr.lvb_size = attr->ia_size;
996 io->u.ci_setattr.sa_valid = attr->ia_valid;
997 io->u.ci_setattr.sa_capa = capa;
1000 if (cl_io_init(env, io, CIT_SETATTR, io->ci_obj) == 0) {
1001 struct ccc_io *cio = ccc_env_io(env);
1003 if (attr->ia_valid & ATTR_FILE)
1004 /* populate the file descriptor for ftruncate to honor
1005 * group lock - see LU-787 */
1006 cio->cui_fd = cl_iattr2fd(inode, attr);
1008 result = cl_io_loop(env, io);
1010 result = io->ci_result;
1012 cl_io_fini(env, io);
1013 if (unlikely(io->ci_need_restart))
1015 /* HSM import case: file is released, cannot be restored
1016 * no need to fail except if restore registration failed
1018 if (result == -ENODATA && io->ci_restore_needed &&
1019 io->ci_result != -ENODATA)
1021 cl_env_put(env, &refcheck);
1025 /*****************************************************************************
1031 struct lu_device *ccc2lu_dev(struct ccc_device *vdv)
1033 return &vdv->cdv_cl.cd_lu_dev;
1036 struct ccc_device *lu2ccc_dev(const struct lu_device *d)
1038 return container_of0(d, struct ccc_device, cdv_cl.cd_lu_dev);
1041 struct ccc_device *cl2ccc_dev(const struct cl_device *d)
1043 return container_of0(d, struct ccc_device, cdv_cl);
1046 struct lu_object *ccc2lu(struct ccc_object *vob)
1048 return &vob->cob_cl.co_lu;
1051 struct ccc_object *lu2ccc(const struct lu_object *obj)
1053 return container_of0(obj, struct ccc_object, cob_cl.co_lu);
1056 struct ccc_object *cl2ccc(const struct cl_object *obj)
1058 return container_of0(obj, struct ccc_object, cob_cl);
1061 struct ccc_lock *cl2ccc_lock(const struct cl_lock_slice *slice)
1063 return container_of(slice, struct ccc_lock, clk_cl);
1066 struct ccc_io *cl2ccc_io(const struct lu_env *env,
1067 const struct cl_io_slice *slice)
1071 cio = container_of(slice, struct ccc_io, cui_cl);
1072 LASSERT(cio == ccc_env_io(env));
1076 struct ccc_req *cl2ccc_req(const struct cl_req_slice *slice)
1078 return container_of0(slice, struct ccc_req, crq_cl);
1081 struct page *cl2vm_page(const struct cl_page_slice *slice)
1083 return cl2ccc_page(slice)->cpg_page;
1086 /*****************************************************************************
1091 int ccc_object_invariant(const struct cl_object *obj)
1093 struct inode *inode = ccc_object_inode(obj);
1094 struct cl_inode_info *lli = cl_i2info(inode);
1096 return (S_ISREG(cl_inode_mode(inode)) ||
1097 /* i_mode of unlinked inode is zeroed. */
1098 cl_inode_mode(inode) == 0) && lli->lli_clob == obj;
1101 struct inode *ccc_object_inode(const struct cl_object *obj)
1103 return cl2ccc(obj)->cob_inode;
1107 * Returns a pointer to cl_page associated with \a vmpage, without acquiring
1108 * additional reference to the resulting page. This is an unsafe version of
1109 * cl_vmpage_page() that can only be used under vmpage lock.
1111 struct cl_page *ccc_vmpage_page_transient(struct page *vmpage)
1113 KLASSERT(PageLocked(vmpage));
1114 return (struct cl_page *)vmpage->private;
1118 * Initialize or update CLIO structures for regular files when new
1119 * meta-data arrives from the server.
1121 * \param inode regular file inode
1122 * \param md new file metadata from MDS
1123 * - allocates cl_object if necessary,
1124 * - updated layout, if object was already here.
1126 int cl_file_inode_init(struct inode *inode, struct lustre_md *md)
1129 struct cl_inode_info *lli;
1130 struct cl_object *clob;
1131 struct lu_site *site;
1133 struct cl_object_conf conf = {
1142 LASSERT(md->body->mbo_valid & OBD_MD_FLID);
1143 LASSERT(S_ISREG(cl_inode_mode(inode)));
1145 env = cl_env_get(&refcheck);
1147 return PTR_ERR(env);
1149 site = cl_i2sbi(inode)->ll_site;
1150 lli = cl_i2info(inode);
1151 fid = &lli->lli_fid;
1152 LASSERT(fid_is_sane(fid));
1154 if (lli->lli_clob == NULL) {
1155 /* clob is slave of inode, empty lli_clob means for new inode,
1156 * there is no clob in cache with the given fid, so it is
1157 * unnecessary to perform lookup-alloc-lookup-insert, just
1158 * alloc and insert directly. */
1160 LASSERT(inode->i_state & I_NEW);
1162 conf.coc_lu.loc_flags = LOC_F_NEW;
1163 clob = cl_object_find(env, lu2cl_dev(site->ls_top_dev),
1165 if (!IS_ERR(clob)) {
1167 * No locking is necessary, as new inode is
1168 * locked by I_NEW bit.
1170 lli->lli_clob = clob;
1171 lli->lli_has_smd = lsm_has_objects(md->lsm);
1172 lu_object_ref_add(&clob->co_lu, "inode", inode);
1174 result = PTR_ERR(clob);
1176 result = cl_conf_set(env, lli->lli_clob, &conf);
1179 cl_env_put(env, &refcheck);
1182 CERROR("Failure to initialize cl object "DFID": %d\n",
1188 * Wait for others drop their references of the object at first, then we drop
1189 * the last one, which will lead to the object be destroyed immediately.
1190 * Must be called after cl_object_kill() against this object.
1192 * The reason we want to do this is: destroying top object will wait for sub
1193 * objects being destroyed first, so we can't let bottom layer (e.g. from ASTs)
1194 * to initiate top object destroying which may deadlock. See bz22520.
1196 static void cl_object_put_last(struct lu_env *env, struct cl_object *obj)
1198 struct lu_object_header *header = obj->co_lu.lo_header;
1199 wait_queue_t waiter;
1201 if (unlikely(atomic_read(&header->loh_ref) != 1)) {
1202 struct lu_site *site = obj->co_lu.lo_dev->ld_site;
1203 struct lu_site_bkt_data *bkt;
1205 bkt = lu_site_bkt_from_fid(site, &header->loh_fid);
1207 init_waitqueue_entry_current(&waiter);
1208 add_wait_queue(&bkt->lsb_marche_funebre, &waiter);
1211 set_current_state(TASK_UNINTERRUPTIBLE);
1212 if (atomic_read(&header->loh_ref) == 1)
1214 waitq_wait(&waiter, TASK_UNINTERRUPTIBLE);
1217 set_current_state(TASK_RUNNING);
1218 remove_wait_queue(&bkt->lsb_marche_funebre, &waiter);
1221 cl_object_put(env, obj);
1224 void cl_inode_fini(struct inode *inode)
1227 struct cl_inode_info *lli = cl_i2info(inode);
1228 struct cl_object *clob = lli->lli_clob;
1235 cookie = cl_env_reenter();
1236 env = cl_env_get(&refcheck);
1237 emergency = IS_ERR(env);
1239 mutex_lock(&ccc_inode_fini_guard);
1240 LASSERT(ccc_inode_fini_env != NULL);
1241 cl_env_implant(ccc_inode_fini_env, &refcheck);
1242 env = ccc_inode_fini_env;
1245 * cl_object cache is a slave to inode cache (which, in turn
1246 * is a slave to dentry cache), don't keep cl_object in memory
1247 * when its master is evicted.
1249 cl_object_kill(env, clob);
1250 lu_object_ref_del(&clob->co_lu, "inode", inode);
1251 cl_object_put_last(env, clob);
1252 lli->lli_clob = NULL;
1254 cl_env_unplant(ccc_inode_fini_env, &refcheck);
1255 mutex_unlock(&ccc_inode_fini_guard);
1257 cl_env_put(env, &refcheck);
1258 cl_env_reexit(cookie);
1263 * return IF_* type for given lu_dirent entry.
1264 * IF_* flag shld be converted to particular OS file type in
1265 * platform llite module.
1267 __u16 ll_dirent_type_get(struct lu_dirent *ent)
1270 struct luda_type *lt;
1273 if (le32_to_cpu(ent->lde_attrs) & LUDA_TYPE) {
1274 const unsigned align = sizeof(struct luda_type) - 1;
1276 len = le16_to_cpu(ent->lde_namelen);
1277 len = (len + align) & ~align;
1278 lt = (void *)ent->lde_name + len;
1279 type = IFTODT(le16_to_cpu(lt->lt_type));
1285 * build inode number from passed @fid */
1286 __u64 cl_fid_build_ino(const struct lu_fid *fid, int api32)
1288 if (BITS_PER_LONG == 32 || api32)
1289 RETURN(fid_flatten32(fid));
1291 RETURN(fid_flatten(fid));
1295 * build inode generation from passed @fid. If our FID overflows the 32-bit
1296 * inode number then return a non-zero generation to distinguish them. */
1297 __u32 cl_fid_build_gen(const struct lu_fid *fid)
1302 if (fid_is_igif(fid)) {
1303 gen = lu_igif_gen(fid);
1307 gen = (fid_flatten(fid) >> 32);
1311 /* lsm is unreliable after hsm implementation as layout can be changed at
1312 * any time. This is only to support old, non-clio-ized interfaces. It will
1313 * cause deadlock if clio operations are called with this extra layout refcount
1314 * because in case the layout changed during the IO, ll_layout_refresh() will
1315 * have to wait for the refcount to become zero to destroy the older layout.
1317 * Notice that the lsm returned by this function may not be valid unless called
1318 * inside layout lock - MDS_INODELOCK_LAYOUT. */
1319 struct lov_stripe_md *ccc_inode_lsm_get(struct inode *inode)
1321 return lov_lsm_get(cl_i2info(inode)->lli_clob);
1324 void inline ccc_inode_lsm_put(struct inode *inode, struct lov_stripe_md *lsm)
1326 lov_lsm_put(cl_i2info(inode)->lli_clob, lsm);