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 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 lu_object_init(obj, &hdr->coh_lu, dev);
355 lu_object_add_top(&hdr->coh_lu, obj);
357 vob->cob_cl.co_ops = clops;
364 int ccc_object_init0(const struct lu_env *env,
365 struct ccc_object *vob,
366 const struct cl_object_conf *conf)
368 vob->cob_inode = conf->coc_inode;
369 vob->cob_transient_pages = 0;
370 cl_object_page_init(&vob->cob_cl, sizeof(struct ccc_page));
374 int ccc_object_init(const struct lu_env *env, struct lu_object *obj,
375 const struct lu_object_conf *conf)
377 struct ccc_device *dev = lu2ccc_dev(obj->lo_dev);
378 struct ccc_object *vob = lu2ccc(obj);
379 struct lu_object *below;
380 struct lu_device *under;
383 under = &dev->cdv_next->cd_lu_dev;
384 below = under->ld_ops->ldo_object_alloc(env, obj->lo_header, under);
386 const struct cl_object_conf *cconf;
388 cconf = lu2cl_conf(conf);
389 CFS_INIT_LIST_HEAD(&vob->cob_pending_list);
390 lu_object_add(obj, below);
391 result = ccc_object_init0(env, vob, cconf);
397 void ccc_object_free(const struct lu_env *env, struct lu_object *obj)
399 struct ccc_object *vob = lu2ccc(obj);
402 lu_object_header_fini(obj->lo_header);
403 OBD_SLAB_FREE_PTR(vob, ccc_object_kmem);
406 int ccc_lock_init(const struct lu_env *env,
407 struct cl_object *obj, struct cl_lock *lock,
408 const struct cl_io *unused,
409 const struct cl_lock_operations *lkops)
411 struct ccc_lock *clk;
414 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
416 OBD_SLAB_ALLOC_PTR_GFP(clk, ccc_lock_kmem, GFP_NOFS);
418 cl_lock_slice_add(lock, &clk->clk_cl, obj, lkops);
425 int ccc_attr_set(const struct lu_env *env, struct cl_object *obj,
426 const struct cl_attr *attr, unsigned valid)
431 int ccc_object_glimpse(const struct lu_env *env,
432 const struct cl_object *obj, struct ost_lvb *lvb)
434 struct inode *inode = ccc_object_inode(obj);
437 lvb->lvb_mtime = cl_inode_mtime(inode);
438 lvb->lvb_atime = cl_inode_atime(inode);
439 lvb->lvb_ctime = cl_inode_ctime(inode);
441 * LU-417: Add dirty pages block count lest i_blocks reports 0, some
442 * "cp" or "tar" on remote node may think it's a completely sparse file
445 if (lvb->lvb_size > 0 && lvb->lvb_blocks == 0)
446 lvb->lvb_blocks = dirty_cnt(inode);
452 int ccc_conf_set(const struct lu_env *env, struct cl_object *obj,
453 const struct cl_object_conf *conf)
455 /* TODO: destroy all pages attached to this object. */
459 static void ccc_object_size_lock(struct cl_object *obj)
461 struct inode *inode = ccc_object_inode(obj);
463 cl_isize_lock(inode);
464 cl_object_attr_lock(obj);
467 static void ccc_object_size_unlock(struct cl_object *obj)
469 struct inode *inode = ccc_object_inode(obj);
471 cl_object_attr_unlock(obj);
472 cl_isize_unlock(inode);
475 /*****************************************************************************
481 struct page *ccc_page_vmpage(const struct lu_env *env,
482 const struct cl_page_slice *slice)
484 return cl2vm_page(slice);
487 int ccc_page_is_under_lock(const struct lu_env *env,
488 const struct cl_page_slice *slice,
491 struct ccc_io *cio = ccc_env_io(env);
492 struct cl_lock_descr *desc = &ccc_env_info(env)->cti_descr;
493 struct cl_page *page = slice->cpl_page;
499 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE ||
500 io->ci_type == CIT_FAULT) {
501 if (cio->cui_fd->fd_flags & LL_FILE_GROUP_LOCKED)
504 desc->cld_start = page->cp_index;
505 desc->cld_end = page->cp_index;
506 desc->cld_obj = page->cp_obj;
507 desc->cld_mode = CLM_READ;
508 result = cl_queue_match(&io->ci_lockset.cls_done,
516 int ccc_fail(const struct lu_env *env, const struct cl_page_slice *slice)
525 void ccc_transient_page_verify(const struct cl_page *page)
529 int ccc_transient_page_own(const struct lu_env *env,
530 const struct cl_page_slice *slice,
531 struct cl_io *unused,
534 ccc_transient_page_verify(slice->cpl_page);
538 void ccc_transient_page_assume(const struct lu_env *env,
539 const struct cl_page_slice *slice,
540 struct cl_io *unused)
542 ccc_transient_page_verify(slice->cpl_page);
545 void ccc_transient_page_unassume(const struct lu_env *env,
546 const struct cl_page_slice *slice,
547 struct cl_io *unused)
549 ccc_transient_page_verify(slice->cpl_page);
552 void ccc_transient_page_disown(const struct lu_env *env,
553 const struct cl_page_slice *slice,
554 struct cl_io *unused)
556 ccc_transient_page_verify(slice->cpl_page);
559 void ccc_transient_page_discard(const struct lu_env *env,
560 const struct cl_page_slice *slice,
561 struct cl_io *unused)
563 struct cl_page *page = slice->cpl_page;
565 ccc_transient_page_verify(slice->cpl_page);
568 * For transient pages, remove it from the radix tree.
570 cl_page_delete(env, page);
573 int ccc_transient_page_prep(const struct lu_env *env,
574 const struct cl_page_slice *slice,
575 struct cl_io *unused)
578 /* transient page should always be sent. */
582 /*****************************************************************************
588 void ccc_lock_delete(const struct lu_env *env,
589 const struct cl_lock_slice *slice)
591 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
594 void ccc_lock_fini(const struct lu_env *env, struct cl_lock_slice *slice)
596 struct ccc_lock *clk = cl2ccc_lock(slice);
597 OBD_SLAB_FREE_PTR(clk, ccc_lock_kmem);
600 int ccc_lock_enqueue(const struct lu_env *env,
601 const struct cl_lock_slice *slice,
602 struct cl_io *unused, __u32 enqflags)
604 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
608 int ccc_lock_unuse(const struct lu_env *env, const struct cl_lock_slice *slice)
610 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
614 int ccc_lock_wait(const struct lu_env *env, const struct cl_lock_slice *slice)
616 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
621 * Implementation of cl_lock_operations::clo_fits_into() methods for ccc
622 * layer. This function is executed every time io finds an existing lock in
623 * the lock cache while creating new lock. This function has to decide whether
624 * cached lock "fits" into io.
626 * \param slice lock to be checked
627 * \param io IO that wants a lock.
629 * \see lov_lock_fits_into().
631 int ccc_lock_fits_into(const struct lu_env *env,
632 const struct cl_lock_slice *slice,
633 const struct cl_lock_descr *need,
634 const struct cl_io *io)
636 const struct cl_lock *lock = slice->cls_lock;
637 const struct cl_lock_descr *descr = &lock->cll_descr;
638 const struct ccc_io *cio = ccc_env_io(env);
643 * Work around DLM peculiarity: it assumes that glimpse
644 * (LDLM_FL_HAS_INTENT) lock is always LCK_PR, and returns reads lock
645 * when asked for LCK_PW lock with LDLM_FL_HAS_INTENT flag set. Make
646 * sure that glimpse doesn't get CLM_WRITE top-lock, so that it
647 * doesn't enqueue CLM_WRITE sub-locks.
649 if (cio->cui_glimpse)
650 result = descr->cld_mode != CLM_WRITE;
653 * Also, don't match incomplete write locks for read, otherwise read
654 * would enqueue missing sub-locks in the write mode.
656 else if (need->cld_mode != descr->cld_mode)
657 result = lock->cll_state >= CLS_ENQUEUED;
664 * Implements cl_lock_operations::clo_state() method for ccc layer, invoked
665 * whenever lock state changes. Transfers object attributes, that might be
666 * updated as a result of lock acquiring into inode.
668 void ccc_lock_state(const struct lu_env *env,
669 const struct cl_lock_slice *slice,
670 enum cl_lock_state state)
672 struct cl_lock *lock = slice->cls_lock;
676 * Refresh inode attributes when the lock is moving into CLS_HELD
677 * state, and only when this is a result of real enqueue, rather than
678 * of finding lock in the cache.
680 if (state == CLS_HELD && lock->cll_state < CLS_HELD) {
681 struct cl_object *obj;
684 obj = slice->cls_obj;
685 inode = ccc_object_inode(obj);
687 /* vmtruncate() sets the i_size
688 * under both a DLM lock and the
689 * ll_inode_size_lock(). If we don't get the
690 * ll_inode_size_lock() here we can match the DLM lock and
691 * reset i_size. generic_file_write can then trust the
692 * stale i_size when doing appending writes and effectively
693 * cancel the result of the truncate. Getting the
694 * ll_inode_size_lock() after the enqueue maintains the DLM
695 * -> ll_inode_size_lock() acquiring order. */
696 if (lock->cll_descr.cld_start == 0 &&
697 lock->cll_descr.cld_end == CL_PAGE_EOF)
698 cl_merge_lvb(env, inode);
703 /*****************************************************************************
709 void ccc_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
711 struct cl_io *io = ios->cis_io;
713 CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj));
716 int ccc_io_one_lock_index(const struct lu_env *env, struct cl_io *io,
717 __u32 enqflags, enum cl_lock_mode mode,
718 pgoff_t start, pgoff_t end)
720 struct ccc_io *cio = ccc_env_io(env);
721 struct cl_lock_descr *descr = &cio->cui_link.cill_descr;
722 struct cl_object *obj = io->ci_obj;
724 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
727 CDEBUG(D_VFSTRACE, "lock: %d [%lu, %lu]\n", mode, start, end);
729 memset(&cio->cui_link, 0, sizeof cio->cui_link);
731 if (cio->cui_fd && (cio->cui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
732 descr->cld_mode = CLM_GROUP;
733 descr->cld_gid = cio->cui_fd->fd_grouplock.cg_gid;
735 descr->cld_mode = mode;
737 descr->cld_obj = obj;
738 descr->cld_start = start;
739 descr->cld_end = end;
740 descr->cld_enq_flags = enqflags;
742 cl_io_lock_add(env, io, &cio->cui_link);
746 void ccc_io_update_iov(const struct lu_env *env,
747 struct ccc_io *cio, struct cl_io *io)
750 size_t size = io->u.ci_rw.crw_count;
752 cio->cui_iov_olen = 0;
753 if (!cl_is_normalio(env, io) || cio->cui_tot_nrsegs == 0)
756 for (i = 0; i < cio->cui_tot_nrsegs; i++) {
757 struct iovec *iv = &cio->cui_iov[i];
759 if (iv->iov_len < size)
762 if (iv->iov_len > size) {
763 cio->cui_iov_olen = iv->iov_len;
770 cio->cui_nrsegs = i + 1;
771 LASSERTF(cio->cui_tot_nrsegs >= cio->cui_nrsegs,
772 "tot_nrsegs: %lu, nrsegs: %lu\n",
773 cio->cui_tot_nrsegs, cio->cui_nrsegs);
776 int ccc_io_one_lock(const struct lu_env *env, struct cl_io *io,
777 __u32 enqflags, enum cl_lock_mode mode,
778 loff_t start, loff_t end)
780 struct cl_object *obj = io->ci_obj;
781 return ccc_io_one_lock_index(env, io, enqflags, mode,
782 cl_index(obj, start), cl_index(obj, end));
785 void ccc_io_end(const struct lu_env *env, const struct cl_io_slice *ios)
787 CLOBINVRNT(env, ios->cis_io->ci_obj,
788 ccc_object_invariant(ios->cis_io->ci_obj));
791 void ccc_io_advance(const struct lu_env *env,
792 const struct cl_io_slice *ios,
795 struct ccc_io *cio = cl2ccc_io(env, ios);
796 struct cl_io *io = ios->cis_io;
797 struct cl_object *obj = ios->cis_io->ci_obj;
799 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
801 if (!cl_is_normalio(env, io))
804 LASSERT(cio->cui_tot_nrsegs >= cio->cui_nrsegs);
805 LASSERT(cio->cui_tot_count >= nob);
807 cio->cui_iov += cio->cui_nrsegs;
808 cio->cui_tot_nrsegs -= cio->cui_nrsegs;
809 cio->cui_tot_count -= nob;
812 if (cio->cui_iov_olen > 0) {
816 cio->cui_tot_nrsegs++;
817 iv = &cio->cui_iov[0];
818 if (io->ci_continue) {
819 iv->iov_base += iv->iov_len;
820 LASSERT(cio->cui_iov_olen > iv->iov_len);
821 iv->iov_len = cio->cui_iov_olen - iv->iov_len;
823 /* restore the iov_len, in case of restart io. */
824 iv->iov_len = cio->cui_iov_olen;
826 cio->cui_iov_olen = 0;
831 * Helper function that if necessary adjusts file size (inode->i_size), when
832 * position at the offset \a pos is accessed. File size can be arbitrary stale
833 * on a Lustre client, but client at least knows KMS. If accessed area is
834 * inside [0, KMS], set file size to KMS, otherwise glimpse file size.
836 * Locking: cl_isize_lock is used to serialize changes to inode size and to
837 * protect consistency between inode size and cl_object
838 * attributes. cl_object_size_lock() protects consistency between cl_attr's of
839 * top-object and sub-objects.
841 int ccc_prep_size(const struct lu_env *env, struct cl_object *obj,
842 struct cl_io *io, loff_t start, size_t count, int *exceed)
844 struct cl_attr *attr = ccc_env_thread_attr(env);
845 struct inode *inode = ccc_object_inode(obj);
846 loff_t pos = start + count - 1;
851 * Consistency guarantees: following possibilities exist for the
852 * relation between region being accessed and real file size at this
855 * (A): the region is completely inside of the file;
857 * (B-x): x bytes of region are inside of the file, the rest is
860 * (C): the region is completely outside of the file.
862 * This classification is stable under DLM lock already acquired by
863 * the caller, because to change the class, other client has to take
864 * DLM lock conflicting with our lock. Also, any updates to ->i_size
865 * by other threads on this client are serialized by
866 * ll_inode_size_lock(). This guarantees that short reads are handled
867 * correctly in the face of concurrent writes and truncates.
869 ccc_object_size_lock(obj);
870 result = cl_object_attr_get(env, obj, attr);
875 * A glimpse is necessary to determine whether we
876 * return a short read (B) or some zeroes at the end
879 ccc_object_size_unlock(obj);
880 result = cl_glimpse_lock(env, io, inode, obj, 0);
881 if (result == 0 && exceed != NULL) {
882 /* If objective page index exceed end-of-file
883 * page index, return directly. Do not expect
884 * kernel will check such case correctly.
885 * linux-2.6.18-128.1.1 miss to do that.
887 loff_t size = cl_isize_read(inode);
888 unsigned long cur_index = start >>
891 if ((size == 0 && cur_index != 0) ||
892 (((size - 1) >> PAGE_CACHE_SHIFT) <
899 * region is within kms and, hence, within real file
900 * size (A). We need to increase i_size to cover the
901 * read region so that generic_file_read() will do its
902 * job, but that doesn't mean the kms size is
903 * _correct_, it is only the _minimum_ size. If
904 * someone does a stat they will get the correct size
905 * which will always be >= the kms value here.
908 if (cl_isize_read(inode) < kms) {
909 cl_isize_write_nolock(inode, kms);
911 DFID" updating i_size "LPU64"\n",
912 PFID(lu_object_fid(&obj->co_lu)),
913 (__u64)cl_isize_read(inode));
918 ccc_object_size_unlock(obj);
922 /*****************************************************************************
924 * Transfer operations.
928 void ccc_req_completion(const struct lu_env *env,
929 const struct cl_req_slice *slice, int ioret)
934 cl_stats_tally(slice->crs_dev, slice->crs_req->crq_type, ioret);
936 vrq = cl2ccc_req(slice);
937 OBD_SLAB_FREE_PTR(vrq, ccc_req_kmem);
941 * Implementation of struct cl_req_operations::cro_attr_set() for ccc
942 * layer. ccc is responsible for
960 void ccc_req_attr_set(const struct lu_env *env,
961 const struct cl_req_slice *slice,
962 const struct cl_object *obj,
963 struct cl_req_attr *attr, obd_valid flags)
967 obd_flag valid_flags;
970 inode = ccc_object_inode(obj);
971 valid_flags = OBD_MD_FLTYPE;
973 if ((flags & OBD_MD_FLOSSCAPA) != 0) {
974 LASSERT(attr->cra_capa == NULL);
975 attr->cra_capa = cl_capa_lookup(inode,
976 slice->crs_req->crq_type);
979 if (slice->crs_req->crq_type == CRT_WRITE) {
980 if (flags & OBD_MD_FLEPOCH) {
981 oa->o_valid |= OBD_MD_FLEPOCH;
982 oa->o_ioepoch = cl_i2info(inode)->lli_ioepoch;
983 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME |
984 OBD_MD_FLUID | OBD_MD_FLGID;
987 obdo_from_inode(oa, inode, valid_flags & flags);
988 obdo_set_parent_fid(oa, &cl_i2info(inode)->lli_fid);
990 memcpy(attr->cra_jobid, cl_i2info(inode)->lli_jobid,
991 JOBSTATS_JOBID_SIZE);
995 const struct cl_req_operations ccc_req_ops = {
996 .cro_attr_set = ccc_req_attr_set,
997 .cro_completion = ccc_req_completion
1000 int cl_setattr_ost(struct inode *inode, const struct iattr *attr,
1001 struct obd_capa *capa)
1010 env = cl_env_get(&refcheck);
1012 RETURN(PTR_ERR(env));
1014 io = ccc_env_thread_io(env);
1015 io->ci_obj = cl_i2info(inode)->lli_clob;
1017 io->u.ci_setattr.sa_attr.lvb_atime = LTIME_S(attr->ia_atime);
1018 io->u.ci_setattr.sa_attr.lvb_mtime = LTIME_S(attr->ia_mtime);
1019 io->u.ci_setattr.sa_attr.lvb_ctime = LTIME_S(attr->ia_ctime);
1020 io->u.ci_setattr.sa_attr.lvb_size = attr->ia_size;
1021 io->u.ci_setattr.sa_valid = attr->ia_valid;
1022 io->u.ci_setattr.sa_capa = capa;
1025 if (cl_io_init(env, io, CIT_SETATTR, io->ci_obj) == 0) {
1026 struct ccc_io *cio = ccc_env_io(env);
1028 if (attr->ia_valid & ATTR_FILE)
1029 /* populate the file descriptor for ftruncate to honor
1030 * group lock - see LU-787 */
1031 cio->cui_fd = cl_iattr2fd(inode, attr);
1033 result = cl_io_loop(env, io);
1035 result = io->ci_result;
1037 cl_io_fini(env, io);
1038 if (unlikely(io->ci_need_restart))
1040 /* HSM import case: file is released, cannot be restored
1041 * no need to fail except if restore registration failed
1043 if (result == -ENODATA && io->ci_restore_needed &&
1044 io->ci_result != -ENODATA)
1046 cl_env_put(env, &refcheck);
1050 /*****************************************************************************
1056 struct lu_device *ccc2lu_dev(struct ccc_device *vdv)
1058 return &vdv->cdv_cl.cd_lu_dev;
1061 struct ccc_device *lu2ccc_dev(const struct lu_device *d)
1063 return container_of0(d, struct ccc_device, cdv_cl.cd_lu_dev);
1066 struct ccc_device *cl2ccc_dev(const struct cl_device *d)
1068 return container_of0(d, struct ccc_device, cdv_cl);
1071 struct lu_object *ccc2lu(struct ccc_object *vob)
1073 return &vob->cob_cl.co_lu;
1076 struct ccc_object *lu2ccc(const struct lu_object *obj)
1078 return container_of0(obj, struct ccc_object, cob_cl.co_lu);
1081 struct ccc_object *cl2ccc(const struct cl_object *obj)
1083 return container_of0(obj, struct ccc_object, cob_cl);
1086 struct ccc_lock *cl2ccc_lock(const struct cl_lock_slice *slice)
1088 return container_of(slice, struct ccc_lock, clk_cl);
1091 struct ccc_io *cl2ccc_io(const struct lu_env *env,
1092 const struct cl_io_slice *slice)
1096 cio = container_of(slice, struct ccc_io, cui_cl);
1097 LASSERT(cio == ccc_env_io(env));
1101 struct ccc_req *cl2ccc_req(const struct cl_req_slice *slice)
1103 return container_of0(slice, struct ccc_req, crq_cl);
1106 struct page *cl2vm_page(const struct cl_page_slice *slice)
1108 return cl2ccc_page(slice)->cpg_page;
1111 /*****************************************************************************
1116 int ccc_object_invariant(const struct cl_object *obj)
1118 struct inode *inode = ccc_object_inode(obj);
1119 struct cl_inode_info *lli = cl_i2info(inode);
1121 return (S_ISREG(cl_inode_mode(inode)) ||
1122 /* i_mode of unlinked inode is zeroed. */
1123 cl_inode_mode(inode) == 0) && lli->lli_clob == obj;
1126 struct inode *ccc_object_inode(const struct cl_object *obj)
1128 return cl2ccc(obj)->cob_inode;
1132 * Returns a pointer to cl_page associated with \a vmpage, without acquiring
1133 * additional reference to the resulting page. This is an unsafe version of
1134 * cl_vmpage_page() that can only be used under vmpage lock.
1136 struct cl_page *ccc_vmpage_page_transient(struct page *vmpage)
1138 KLASSERT(PageLocked(vmpage));
1139 return (struct cl_page *)vmpage->private;
1143 * Initialize or update CLIO structures for regular files when new
1144 * meta-data arrives from the server.
1146 * \param inode regular file inode
1147 * \param md new file metadata from MDS
1148 * - allocates cl_object if necessary,
1149 * - updated layout, if object was already here.
1151 int cl_file_inode_init(struct inode *inode, struct lustre_md *md)
1154 struct cl_inode_info *lli;
1155 struct cl_object *clob;
1156 struct lu_site *site;
1158 struct cl_object_conf conf = {
1167 LASSERT(md->body->valid & OBD_MD_FLID);
1168 LASSERT(S_ISREG(cl_inode_mode(inode)));
1170 env = cl_env_get(&refcheck);
1172 return PTR_ERR(env);
1174 site = cl_i2sbi(inode)->ll_site;
1175 lli = cl_i2info(inode);
1176 fid = &lli->lli_fid;
1177 LASSERT(fid_is_sane(fid));
1179 if (lli->lli_clob == NULL) {
1180 /* clob is slave of inode, empty lli_clob means for new inode,
1181 * there is no clob in cache with the given fid, so it is
1182 * unnecessary to perform lookup-alloc-lookup-insert, just
1183 * alloc and insert directly. */
1185 LASSERT(inode->i_state & I_NEW);
1187 conf.coc_lu.loc_flags = LOC_F_NEW;
1188 clob = cl_object_find(env, lu2cl_dev(site->ls_top_dev),
1190 if (!IS_ERR(clob)) {
1192 * No locking is necessary, as new inode is
1193 * locked by I_NEW bit.
1195 lli->lli_clob = clob;
1196 lli->lli_has_smd = lsm_has_objects(md->lsm);
1197 lu_object_ref_add(&clob->co_lu, "inode", inode);
1199 result = PTR_ERR(clob);
1201 result = cl_conf_set(env, lli->lli_clob, &conf);
1204 cl_env_put(env, &refcheck);
1207 CERROR("Failure to initialize cl object "DFID": %d\n",
1213 * Wait for others drop their references of the object at first, then we drop
1214 * the last one, which will lead to the object be destroyed immediately.
1215 * Must be called after cl_object_kill() against this object.
1217 * The reason we want to do this is: destroying top object will wait for sub
1218 * objects being destroyed first, so we can't let bottom layer (e.g. from ASTs)
1219 * to initiate top object destroying which may deadlock. See bz22520.
1221 static void cl_object_put_last(struct lu_env *env, struct cl_object *obj)
1223 struct lu_object_header *header = obj->co_lu.lo_header;
1224 wait_queue_t waiter;
1226 if (unlikely(cfs_atomic_read(&header->loh_ref) != 1)) {
1227 struct lu_site *site = obj->co_lu.lo_dev->ld_site;
1228 struct lu_site_bkt_data *bkt;
1230 bkt = lu_site_bkt_from_fid(site, &header->loh_fid);
1232 init_waitqueue_entry_current(&waiter);
1233 add_wait_queue(&bkt->lsb_marche_funebre, &waiter);
1236 set_current_state(TASK_UNINTERRUPTIBLE);
1237 if (cfs_atomic_read(&header->loh_ref) == 1)
1239 waitq_wait(&waiter, TASK_UNINTERRUPTIBLE);
1242 set_current_state(TASK_RUNNING);
1243 remove_wait_queue(&bkt->lsb_marche_funebre, &waiter);
1246 cl_object_put(env, obj);
1249 void cl_inode_fini(struct inode *inode)
1252 struct cl_inode_info *lli = cl_i2info(inode);
1253 struct cl_object *clob = lli->lli_clob;
1260 cookie = cl_env_reenter();
1261 env = cl_env_get(&refcheck);
1262 emergency = IS_ERR(env);
1264 mutex_lock(&ccc_inode_fini_guard);
1265 LASSERT(ccc_inode_fini_env != NULL);
1266 cl_env_implant(ccc_inode_fini_env, &refcheck);
1267 env = ccc_inode_fini_env;
1270 * cl_object cache is a slave to inode cache (which, in turn
1271 * is a slave to dentry cache), don't keep cl_object in memory
1272 * when its master is evicted.
1274 cl_object_kill(env, clob);
1275 lu_object_ref_del(&clob->co_lu, "inode", inode);
1276 cl_object_put_last(env, clob);
1277 lli->lli_clob = NULL;
1279 cl_env_unplant(ccc_inode_fini_env, &refcheck);
1280 mutex_unlock(&ccc_inode_fini_guard);
1282 cl_env_put(env, &refcheck);
1283 cl_env_reexit(cookie);
1288 * return IF_* type for given lu_dirent entry.
1289 * IF_* flag shld be converted to particular OS file type in
1290 * platform llite module.
1292 __u16 ll_dirent_type_get(struct lu_dirent *ent)
1295 struct luda_type *lt;
1298 if (le32_to_cpu(ent->lde_attrs) & LUDA_TYPE) {
1299 const unsigned align = sizeof(struct luda_type) - 1;
1301 len = le16_to_cpu(ent->lde_namelen);
1302 len = (len + align) & ~align;
1303 lt = (void *)ent->lde_name + len;
1304 type = IFTODT(le16_to_cpu(lt->lt_type));
1310 * build inode number from passed @fid */
1311 __u64 cl_fid_build_ino(const struct lu_fid *fid, int api32)
1313 if (BITS_PER_LONG == 32 || api32)
1314 RETURN(fid_flatten32(fid));
1316 RETURN(fid_flatten(fid));
1320 * build inode generation from passed @fid. If our FID overflows the 32-bit
1321 * inode number then return a non-zero generation to distinguish them. */
1322 __u32 cl_fid_build_gen(const struct lu_fid *fid)
1327 if (fid_is_igif(fid)) {
1328 gen = lu_igif_gen(fid);
1332 gen = (fid_flatten(fid) >> 32);
1336 /* lsm is unreliable after hsm implementation as layout can be changed at
1337 * any time. This is only to support old, non-clio-ized interfaces. It will
1338 * cause deadlock if clio operations are called with this extra layout refcount
1339 * because in case the layout changed during the IO, ll_layout_refresh() will
1340 * have to wait for the refcount to become zero to destroy the older layout.
1342 * Notice that the lsm returned by this function may not be valid unless called
1343 * inside layout lock - MDS_INODELOCK_LAYOUT. */
1344 struct lov_stripe_md *ccc_inode_lsm_get(struct inode *inode)
1346 return lov_lsm_get(cl_i2info(inode)->lli_clob);
1349 void inline ccc_inode_lsm_put(struct inode *inode, struct lov_stripe_md *lsm)
1351 lov_lsm_put(cl_i2info(inode)->lli_clob, lsm);