1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2 * vim:expandtab:shiftwidth=8:tabstop=8:
6 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 only,
10 * as published by the Free Software Foundation.
12 * This program is distributed in the hope that it will be useful, but
13 * WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * General Public License version 2 for more details (a copy is included
16 * in the LICENSE file that accompanied this code).
18 * You should have received a copy of the GNU General Public License
19 * version 2 along with this program; If not, see
20 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
22 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
23 * CA 95054 USA or visit www.sun.com if you need additional information or
29 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
30 * Use is subject to license terms.
32 * Copyright (c) 2011, 2012, Whamcloud, Inc.
35 * This file is part of Lustre, http://www.lustre.org/
36 * Lustre is a trademark of Sun Microsystems, Inc.
38 * cl code shared between vvp and liblustre (and other Lustre clients in the
41 * Author: Nikita Danilov <nikita.danilov@sun.com>
44 #define DEBUG_SUBSYSTEM S_LLITE
47 # include <libcfs/libcfs.h>
48 # include <linux/fs.h>
49 # include <linux/sched.h>
50 # include <linux/mm.h>
51 # include <linux/smp_lock.h>
52 # include <linux/quotaops.h>
53 # include <linux/highmem.h>
54 # include <linux/pagemap.h>
55 # include <linux/rbtree.h>
56 #else /* __KERNEL__ */
61 #include <sys/types.h>
63 #include <sys/queue.h>
65 # include <liblustre.h>
69 #include <obd_support.h>
70 #include <lustre_fid.h>
71 #include <lustre_lite.h>
72 #include <lustre_dlm.h>
73 #include <lustre_ver.h>
74 #include <lustre_mdc.h>
75 #include <cl_object.h>
80 #include "../llite/llite_internal.h"
82 #include "../liblustre/llite_lib.h"
85 const struct cl_req_operations ccc_req_ops;
88 * ccc_ prefix stands for "Common Client Code".
91 static cfs_mem_cache_t *ccc_lock_kmem;
92 static cfs_mem_cache_t *ccc_object_kmem;
93 static cfs_mem_cache_t *ccc_thread_kmem;
94 static cfs_mem_cache_t *ccc_session_kmem;
95 static cfs_mem_cache_t *ccc_req_kmem;
97 static struct lu_kmem_descr ccc_caches[] = {
99 .ckd_cache = &ccc_lock_kmem,
100 .ckd_name = "ccc_lock_kmem",
101 .ckd_size = sizeof (struct ccc_lock)
104 .ckd_cache = &ccc_object_kmem,
105 .ckd_name = "ccc_object_kmem",
106 .ckd_size = sizeof (struct ccc_object)
109 .ckd_cache = &ccc_thread_kmem,
110 .ckd_name = "ccc_thread_kmem",
111 .ckd_size = sizeof (struct ccc_thread_info),
114 .ckd_cache = &ccc_session_kmem,
115 .ckd_name = "ccc_session_kmem",
116 .ckd_size = sizeof (struct ccc_session)
119 .ckd_cache = &ccc_req_kmem,
120 .ckd_name = "ccc_req_kmem",
121 .ckd_size = sizeof (struct ccc_req)
128 /*****************************************************************************
130 * Vvp device and device type functions.
134 void *ccc_key_init(const struct lu_context *ctx,
135 struct lu_context_key *key)
137 struct ccc_thread_info *info;
139 OBD_SLAB_ALLOC_PTR_GFP(info, ccc_thread_kmem, CFS_ALLOC_IO);
141 info = ERR_PTR(-ENOMEM);
145 void ccc_key_fini(const struct lu_context *ctx,
146 struct lu_context_key *key, void *data)
148 struct ccc_thread_info *info = data;
149 OBD_SLAB_FREE_PTR(info, ccc_thread_kmem);
152 void *ccc_session_key_init(const struct lu_context *ctx,
153 struct lu_context_key *key)
155 struct ccc_session *session;
157 OBD_SLAB_ALLOC_PTR_GFP(session, ccc_session_kmem, CFS_ALLOC_IO);
159 session = ERR_PTR(-ENOMEM);
163 void ccc_session_key_fini(const struct lu_context *ctx,
164 struct lu_context_key *key, void *data)
166 struct ccc_session *session = data;
167 OBD_SLAB_FREE_PTR(session, ccc_session_kmem);
170 struct lu_context_key ccc_key = {
171 .lct_tags = LCT_CL_THREAD,
172 .lct_init = ccc_key_init,
173 .lct_fini = ccc_key_fini
176 struct lu_context_key ccc_session_key = {
177 .lct_tags = LCT_SESSION,
178 .lct_init = ccc_session_key_init,
179 .lct_fini = ccc_session_key_fini
183 /* type constructor/destructor: ccc_type_{init,fini,start,stop}(). */
184 // LU_TYPE_INIT_FINI(ccc, &ccc_key, &ccc_session_key);
186 int ccc_device_init(const struct lu_env *env, struct lu_device *d,
187 const char *name, struct lu_device *next)
189 struct ccc_device *vdv;
194 vdv->cdv_next = lu2cl_dev(next);
196 LASSERT(d->ld_site != NULL && next->ld_type != NULL);
197 next->ld_site = d->ld_site;
198 rc = next->ld_type->ldt_ops->ldto_device_init(
199 env, next, next->ld_type->ldt_name, NULL);
202 lu_ref_add(&next->ld_reference, "lu-stack", &lu_site_init);
207 struct lu_device *ccc_device_fini(const struct lu_env *env,
210 return cl2lu_dev(lu2ccc_dev(d)->cdv_next);
213 struct lu_device *ccc_device_alloc(const struct lu_env *env,
214 struct lu_device_type *t,
215 struct lustre_cfg *cfg,
216 const struct lu_device_operations *luops,
217 const struct cl_device_operations *clops)
219 struct ccc_device *vdv;
220 struct lu_device *lud;
221 struct cl_site *site;
227 RETURN(ERR_PTR(-ENOMEM));
229 lud = &vdv->cdv_cl.cd_lu_dev;
230 cl_device_init(&vdv->cdv_cl, t);
231 ccc2lu_dev(vdv)->ld_ops = luops;
232 vdv->cdv_cl.cd_ops = clops;
236 rc = cl_site_init(site, &vdv->cdv_cl);
238 rc = lu_site_init_finish(&site->cs_lu);
240 LASSERT(lud->ld_site == NULL);
241 CERROR("Cannot init lu_site, rc %d.\n", rc);
247 ccc_device_free(env, lud);
253 struct lu_device *ccc_device_free(const struct lu_env *env,
256 struct ccc_device *vdv = lu2ccc_dev(d);
257 struct cl_site *site = lu2cl_site(d->ld_site);
258 struct lu_device *next = cl2lu_dev(vdv->cdv_next);
260 if (d->ld_site != NULL) {
264 cl_device_fini(lu2cl_dev(d));
269 int ccc_req_init(const struct lu_env *env, struct cl_device *dev,
275 OBD_SLAB_ALLOC_PTR_GFP(vrq, ccc_req_kmem, CFS_ALLOC_IO);
277 cl_req_slice_add(req, &vrq->crq_cl, dev, &ccc_req_ops);
285 * An `emergency' environment used by ccc_inode_fini() when cl_env_get()
286 * fails. Access to this environment is serialized by ccc_inode_fini_guard
289 static struct lu_env *ccc_inode_fini_env = NULL;
292 * A mutex serializing calls to slp_inode_fini() under extreme memory
293 * pressure, when environments cannot be allocated.
295 static CFS_DEFINE_MUTEX(ccc_inode_fini_guard);
296 static int dummy_refcheck;
298 int ccc_global_init(struct lu_device_type *device_type)
302 result = lu_kmem_init(ccc_caches);
306 result = lu_device_type_init(device_type);
310 ccc_inode_fini_env = cl_env_alloc(&dummy_refcheck,
311 LCT_REMEMBER|LCT_NOREF);
312 if (IS_ERR(ccc_inode_fini_env)) {
313 result = PTR_ERR(ccc_inode_fini_env);
317 ccc_inode_fini_env->le_ctx.lc_cookie = 0x4;
320 lu_device_type_fini(device_type);
322 lu_kmem_fini(ccc_caches);
326 void ccc_global_fini(struct lu_device_type *device_type)
328 if (ccc_inode_fini_env != NULL) {
329 cl_env_put(ccc_inode_fini_env, &dummy_refcheck);
330 ccc_inode_fini_env = NULL;
332 lu_device_type_fini(device_type);
333 lu_kmem_fini(ccc_caches);
336 /*****************************************************************************
342 struct lu_object *ccc_object_alloc(const struct lu_env *env,
343 const struct lu_object_header *unused,
344 struct lu_device *dev,
345 const struct cl_object_operations *clops,
346 const struct lu_object_operations *luops)
348 struct ccc_object *vob;
349 struct lu_object *obj;
351 OBD_SLAB_ALLOC_PTR_GFP(vob, ccc_object_kmem, CFS_ALLOC_IO);
353 struct cl_object_header *hdr;
356 hdr = &vob->cob_header;
357 cl_object_header_init(hdr);
358 lu_object_init(obj, &hdr->coh_lu, dev);
359 lu_object_add_top(&hdr->coh_lu, obj);
361 vob->cob_cl.co_ops = clops;
368 int ccc_object_init0(const struct lu_env *env,
369 struct ccc_object *vob,
370 const struct cl_object_conf *conf)
372 vob->cob_inode = conf->coc_inode;
373 vob->cob_transient_pages = 0;
377 int ccc_object_init(const struct lu_env *env, struct lu_object *obj,
378 const struct lu_object_conf *conf)
380 struct ccc_device *dev = lu2ccc_dev(obj->lo_dev);
381 struct ccc_object *vob = lu2ccc(obj);
382 struct lu_object *below;
383 struct lu_device *under;
386 under = &dev->cdv_next->cd_lu_dev;
387 below = under->ld_ops->ldo_object_alloc(env, obj->lo_header, under);
389 const struct cl_object_conf *cconf;
391 cconf = lu2cl_conf(conf);
392 CFS_INIT_LIST_HEAD(&vob->cob_pending_list);
393 lu_object_add(obj, below);
394 result = ccc_object_init0(env, vob, cconf);
400 void ccc_object_free(const struct lu_env *env, struct lu_object *obj)
402 struct ccc_object *vob = lu2ccc(obj);
405 lu_object_header_fini(obj->lo_header);
406 OBD_SLAB_FREE_PTR(vob, ccc_object_kmem);
409 int ccc_lock_init(const struct lu_env *env,
410 struct cl_object *obj, struct cl_lock *lock,
411 const struct cl_io *unused,
412 const struct cl_lock_operations *lkops)
414 struct ccc_lock *clk;
417 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
419 OBD_SLAB_ALLOC_PTR_GFP(clk, ccc_lock_kmem, CFS_ALLOC_IO);
421 cl_lock_slice_add(lock, &clk->clk_cl, obj, lkops);
428 int ccc_attr_set(const struct lu_env *env, struct cl_object *obj,
429 const struct cl_attr *attr, unsigned valid)
434 int ccc_object_glimpse(const struct lu_env *env,
435 const struct cl_object *obj, struct ost_lvb *lvb)
437 struct inode *inode = ccc_object_inode(obj);
440 lvb->lvb_mtime = cl_inode_mtime(inode);
441 lvb->lvb_atime = cl_inode_atime(inode);
442 lvb->lvb_ctime = cl_inode_ctime(inode);
444 * LU-417: Add dirty pages block count lest i_blocks reports 0, some
445 * "cp" or "tar" on remote node may think it's a completely sparse file
448 if (lvb->lvb_size > 0 && lvb->lvb_blocks == 0)
449 lvb->lvb_blocks = dirty_cnt(inode);
455 int ccc_conf_set(const struct lu_env *env, struct cl_object *obj,
456 const struct cl_object_conf *conf)
458 /* TODO: destroy all pages attached to this object. */
462 /*****************************************************************************
468 cfs_page_t *ccc_page_vmpage(const struct lu_env *env,
469 const struct cl_page_slice *slice)
471 return cl2vm_page(slice);
474 int ccc_page_is_under_lock(const struct lu_env *env,
475 const struct cl_page_slice *slice,
478 struct ccc_io *cio = ccc_env_io(env);
479 struct cl_lock_descr *desc = &ccc_env_info(env)->cti_descr;
480 struct cl_page *page = slice->cpl_page;
486 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE ||
487 io->ci_type == CIT_FAULT) {
488 if (cio->cui_fd->fd_flags & LL_FILE_GROUP_LOCKED)
491 desc->cld_start = page->cp_index;
492 desc->cld_end = page->cp_index;
493 desc->cld_obj = page->cp_obj;
494 desc->cld_mode = CLM_READ;
495 result = cl_queue_match(&io->ci_lockset.cls_done,
503 int ccc_fail(const struct lu_env *env, const struct cl_page_slice *slice)
512 void ccc_transient_page_verify(const struct cl_page *page)
516 int ccc_transient_page_own(const struct lu_env *env,
517 const struct cl_page_slice *slice,
518 struct cl_io *unused,
521 ccc_transient_page_verify(slice->cpl_page);
525 void ccc_transient_page_assume(const struct lu_env *env,
526 const struct cl_page_slice *slice,
527 struct cl_io *unused)
529 ccc_transient_page_verify(slice->cpl_page);
532 void ccc_transient_page_unassume(const struct lu_env *env,
533 const struct cl_page_slice *slice,
534 struct cl_io *unused)
536 ccc_transient_page_verify(slice->cpl_page);
539 void ccc_transient_page_disown(const struct lu_env *env,
540 const struct cl_page_slice *slice,
541 struct cl_io *unused)
543 ccc_transient_page_verify(slice->cpl_page);
546 void ccc_transient_page_discard(const struct lu_env *env,
547 const struct cl_page_slice *slice,
548 struct cl_io *unused)
550 struct cl_page *page = slice->cpl_page;
552 ccc_transient_page_verify(slice->cpl_page);
555 * For transient pages, remove it from the radix tree.
557 cl_page_delete(env, page);
560 int ccc_transient_page_prep(const struct lu_env *env,
561 const struct cl_page_slice *slice,
562 struct cl_io *unused)
565 /* transient page should always be sent. */
569 /*****************************************************************************
575 void ccc_lock_delete(const struct lu_env *env,
576 const struct cl_lock_slice *slice)
578 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
581 void ccc_lock_fini(const struct lu_env *env, struct cl_lock_slice *slice)
583 struct ccc_lock *clk = cl2ccc_lock(slice);
584 OBD_SLAB_FREE_PTR(clk, ccc_lock_kmem);
587 int ccc_lock_enqueue(const struct lu_env *env,
588 const struct cl_lock_slice *slice,
589 struct cl_io *unused, __u32 enqflags)
591 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
595 int ccc_lock_unuse(const struct lu_env *env, const struct cl_lock_slice *slice)
597 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
601 int ccc_lock_wait(const struct lu_env *env, const struct cl_lock_slice *slice)
603 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
608 * Implementation of cl_lock_operations::clo_fits_into() methods for ccc
609 * layer. This function is executed every time io finds an existing lock in
610 * the lock cache while creating new lock. This function has to decide whether
611 * cached lock "fits" into io.
613 * \param slice lock to be checked
614 * \param io IO that wants a lock.
616 * \see lov_lock_fits_into().
618 int ccc_lock_fits_into(const struct lu_env *env,
619 const struct cl_lock_slice *slice,
620 const struct cl_lock_descr *need,
621 const struct cl_io *io)
623 const struct cl_lock *lock = slice->cls_lock;
624 const struct cl_lock_descr *descr = &lock->cll_descr;
625 const struct ccc_io *cio = ccc_env_io(env);
630 * Work around DLM peculiarity: it assumes that glimpse
631 * (LDLM_FL_HAS_INTENT) lock is always LCK_PR, and returns reads lock
632 * when asked for LCK_PW lock with LDLM_FL_HAS_INTENT flag set. Make
633 * sure that glimpse doesn't get CLM_WRITE top-lock, so that it
634 * doesn't enqueue CLM_WRITE sub-locks.
636 if (cio->cui_glimpse)
637 result = descr->cld_mode != CLM_WRITE;
640 * Also, don't match incomplete write locks for read, otherwise read
641 * would enqueue missing sub-locks in the write mode.
643 else if (need->cld_mode != descr->cld_mode)
644 result = lock->cll_state >= CLS_ENQUEUED;
651 * Implements cl_lock_operations::clo_state() method for ccc layer, invoked
652 * whenever lock state changes. Transfers object attributes, that might be
653 * updated as a result of lock acquiring into inode.
655 void ccc_lock_state(const struct lu_env *env,
656 const struct cl_lock_slice *slice,
657 enum cl_lock_state state)
659 struct cl_lock *lock;
660 struct cl_object *obj;
662 struct cl_attr *attr;
665 lock = slice->cls_lock;
668 * Refresh inode attributes when the lock is moving into CLS_HELD
669 * state, and only when this is a result of real enqueue, rather than
670 * of finding lock in the cache.
672 if (state == CLS_HELD && lock->cll_state < CLS_HELD) {
675 obj = slice->cls_obj;
676 inode = ccc_object_inode(obj);
677 attr = ccc_env_thread_attr(env);
679 /* vmtruncate()->ll_truncate() first sets the i_size and then
680 * the kms under both a DLM lock and the
681 * ll_inode_size_lock(). If we don't get the
682 * ll_inode_size_lock() here we can match the DLM lock and
683 * reset i_size from the kms before the truncating path has
684 * updated the kms. generic_file_write can then trust the
685 * stale i_size when doing appending writes and effectively
686 * cancel the result of the truncate. Getting the
687 * ll_inode_size_lock() after the enqueue maintains the DLM
688 * -> ll_inode_size_lock() acquiring order. */
689 cl_isize_lock(inode, 0);
690 cl_object_attr_lock(obj);
691 rc = cl_object_attr_get(env, obj, attr);
693 if (lock->cll_descr.cld_start == 0 &&
694 lock->cll_descr.cld_end == CL_PAGE_EOF) {
695 cl_isize_write_nolock(inode, attr->cat_kms);
696 CDEBUG(D_INODE|D_VFSTRACE,
697 DFID" updating i_size "LPU64"\n",
698 PFID(lu_object_fid(&obj->co_lu)),
699 (__u64)cl_isize_read(inode));
701 cl_inode_mtime(inode) = attr->cat_mtime;
702 cl_inode_atime(inode) = attr->cat_atime;
703 cl_inode_ctime(inode) = attr->cat_ctime;
705 CL_LOCK_DEBUG(D_INFO, env, lock, "attr_get: %d\n", rc);
707 cl_object_attr_unlock(obj);
708 cl_isize_unlock(inode, 0);
713 /*****************************************************************************
719 void ccc_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
721 struct cl_io *io = ios->cis_io;
723 CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj));
726 int ccc_io_one_lock_index(const struct lu_env *env, struct cl_io *io,
727 __u32 enqflags, enum cl_lock_mode mode,
728 pgoff_t start, pgoff_t end)
730 struct ccc_io *cio = ccc_env_io(env);
731 struct cl_lock_descr *descr = &cio->cui_link.cill_descr;
732 struct cl_object *obj = io->ci_obj;
734 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
737 CDEBUG(D_VFSTRACE, "lock: %d [%lu, %lu]\n", mode, start, end);
739 memset(&cio->cui_link, 0, sizeof cio->cui_link);
741 if (cio->cui_fd && (cio->cui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
742 descr->cld_mode = CLM_GROUP;
743 descr->cld_gid = cio->cui_fd->fd_grouplock.cg_gid;
745 descr->cld_mode = mode;
747 descr->cld_obj = obj;
748 descr->cld_start = start;
749 descr->cld_end = end;
750 descr->cld_enq_flags = enqflags;
752 cl_io_lock_add(env, io, &cio->cui_link);
756 void ccc_io_update_iov(const struct lu_env *env,
757 struct ccc_io *cio, struct cl_io *io)
760 size_t size = io->u.ci_rw.crw_count;
762 cio->cui_iov_olen = 0;
763 if (!cl_is_normalio(env, io))
766 for (i = 0; i < cio->cui_tot_nrsegs; i++) {
767 struct iovec *iv = &cio->cui_iov[i];
769 if (iv->iov_len < size)
772 if (iv->iov_len > size) {
773 cio->cui_iov_olen = iv->iov_len;
780 cio->cui_nrsegs = i + 1;
783 int ccc_io_one_lock(const struct lu_env *env, struct cl_io *io,
784 __u32 enqflags, enum cl_lock_mode mode,
785 loff_t start, loff_t end)
787 struct cl_object *obj = io->ci_obj;
788 return ccc_io_one_lock_index(env, io, enqflags, mode,
789 cl_index(obj, start), cl_index(obj, end));
792 void ccc_io_end(const struct lu_env *env, const struct cl_io_slice *ios)
794 CLOBINVRNT(env, ios->cis_io->ci_obj,
795 ccc_object_invariant(ios->cis_io->ci_obj));
798 void ccc_io_advance(const struct lu_env *env,
799 const struct cl_io_slice *ios,
802 struct ccc_io *cio = cl2ccc_io(env, ios);
803 struct cl_io *io = ios->cis_io;
804 struct cl_object *obj = ios->cis_io->ci_obj;
806 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
808 if (cl_is_normalio(env, io) && io->ci_continue) {
810 LASSERT(cio->cui_tot_nrsegs >= cio->cui_nrsegs);
811 LASSERT(cio->cui_tot_count >= nob);
813 cio->cui_iov += cio->cui_nrsegs;
814 cio->cui_tot_nrsegs -= cio->cui_nrsegs;
815 cio->cui_tot_count -= nob;
817 if (cio->cui_iov_olen) {
821 cio->cui_tot_nrsegs++;
822 iv = &cio->cui_iov[0];
823 iv->iov_base += iv->iov_len;
824 LASSERT(cio->cui_iov_olen > iv->iov_len);
825 iv->iov_len = cio->cui_iov_olen - iv->iov_len;
830 static void ccc_object_size_lock(struct cl_object *obj)
832 struct inode *inode = ccc_object_inode(obj);
834 cl_isize_lock(inode, 0);
835 cl_object_attr_lock(obj);
838 static void ccc_object_size_unlock(struct cl_object *obj)
840 struct inode *inode = ccc_object_inode(obj);
842 cl_object_attr_unlock(obj);
843 cl_isize_unlock(inode, 0);
847 * Helper function that if necessary adjusts file size (inode->i_size), when
848 * position at the offset \a pos is accessed. File size can be arbitrary stale
849 * on a Lustre client, but client at least knows KMS. If accessed area is
850 * inside [0, KMS], set file size to KMS, otherwise glimpse file size.
852 * Locking: cl_isize_lock is used to serialize changes to inode size and to
853 * protect consistency between inode size and cl_object
854 * attributes. cl_object_size_lock() protects consistency between cl_attr's of
855 * top-object and sub-objects.
857 int ccc_prep_size(const struct lu_env *env, struct cl_object *obj,
858 struct cl_io *io, loff_t start, size_t count, int *exceed)
860 struct cl_attr *attr = ccc_env_thread_attr(env);
861 struct inode *inode = ccc_object_inode(obj);
862 loff_t pos = start + count - 1;
867 * Consistency guarantees: following possibilities exist for the
868 * relation between region being accessed and real file size at this
871 * (A): the region is completely inside of the file;
873 * (B-x): x bytes of region are inside of the file, the rest is
876 * (C): the region is completely outside of the file.
878 * This classification is stable under DLM lock already acquired by
879 * the caller, because to change the class, other client has to take
880 * DLM lock conflicting with our lock. Also, any updates to ->i_size
881 * by other threads on this client are serialized by
882 * ll_inode_size_lock(). This guarantees that short reads are handled
883 * correctly in the face of concurrent writes and truncates.
885 ccc_object_size_lock(obj);
886 result = cl_object_attr_get(env, obj, attr);
891 * A glimpse is necessary to determine whether we
892 * return a short read (B) or some zeroes at the end
895 ccc_object_size_unlock(obj);
896 result = cl_glimpse_lock(env, io, inode, obj, 0);
897 if (result == 0 && exceed != NULL) {
898 /* If objective page index exceed end-of-file
899 * page index, return directly. Do not expect
900 * kernel will check such case correctly.
901 * linux-2.6.18-128.1.1 miss to do that.
903 loff_t size = cl_isize_read(inode);
904 unsigned long cur_index = start >> CFS_PAGE_SHIFT;
906 if ((size == 0 && cur_index != 0) ||
907 (((size - 1) >> CFS_PAGE_SHIFT) < cur_index))
913 * region is within kms and, hence, within real file
914 * size (A). We need to increase i_size to cover the
915 * read region so that generic_file_read() will do its
916 * job, but that doesn't mean the kms size is
917 * _correct_, it is only the _minimum_ size. If
918 * someone does a stat they will get the correct size
919 * which will always be >= the kms value here.
922 if (cl_isize_read(inode) < kms) {
923 cl_isize_write_nolock(inode, kms);
925 DFID" updating i_size "LPU64"\n",
926 PFID(lu_object_fid(&obj->co_lu)),
927 (__u64)cl_isize_read(inode));
932 ccc_object_size_unlock(obj);
936 /*****************************************************************************
938 * Transfer operations.
942 void ccc_req_completion(const struct lu_env *env,
943 const struct cl_req_slice *slice, int ioret)
948 cl_stats_tally(slice->crs_dev, slice->crs_req->crq_type, ioret);
950 vrq = cl2ccc_req(slice);
951 OBD_SLAB_FREE_PTR(vrq, ccc_req_kmem);
955 * Implementation of struct cl_req_operations::cro_attr_set() for ccc
956 * layer. ccc is responsible for
974 void ccc_req_attr_set(const struct lu_env *env,
975 const struct cl_req_slice *slice,
976 const struct cl_object *obj,
977 struct cl_req_attr *attr, obd_valid flags)
981 obd_flag valid_flags;
984 inode = ccc_object_inode(obj);
985 valid_flags = OBD_MD_FLTYPE|OBD_MD_FLATIME;
987 if (flags != (obd_valid)~0ULL)
988 valid_flags |= OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME;
990 LASSERT(attr->cra_capa == NULL);
991 attr->cra_capa = cl_capa_lookup(inode,
992 slice->crs_req->crq_type);
995 if (slice->crs_req->crq_type == CRT_WRITE) {
996 if (flags & OBD_MD_FLEPOCH) {
997 oa->o_valid |= OBD_MD_FLEPOCH;
998 oa->o_ioepoch = cl_i2info(inode)->lli_ioepoch;
999 valid_flags |= OBD_MD_FLMTIME|OBD_MD_FLCTIME|
1000 OBD_MD_FLUID|OBD_MD_FLGID;
1003 obdo_from_inode(oa, inode, valid_flags & flags);
1004 obdo_set_parent_fid(oa, &cl_i2info(inode)->lli_fid);
1007 const struct cl_req_operations ccc_req_ops = {
1008 .cro_attr_set = ccc_req_attr_set,
1009 .cro_completion = ccc_req_completion
1012 int cl_setattr_ost(struct inode *inode, const struct iattr *attr,
1013 struct obd_capa *capa)
1022 env = cl_env_get(&refcheck);
1024 RETURN(PTR_ERR(env));
1026 io = ccc_env_thread_io(env);
1027 io->ci_obj = cl_i2info(inode)->lli_clob;
1029 io->u.ci_setattr.sa_attr.lvb_atime = LTIME_S(attr->ia_atime);
1030 io->u.ci_setattr.sa_attr.lvb_mtime = LTIME_S(attr->ia_mtime);
1031 io->u.ci_setattr.sa_attr.lvb_ctime = LTIME_S(attr->ia_ctime);
1032 io->u.ci_setattr.sa_attr.lvb_size = attr->ia_size;
1033 io->u.ci_setattr.sa_valid = attr->ia_valid;
1034 io->u.ci_setattr.sa_capa = capa;
1036 if (cl_io_init(env, io, CIT_SETATTR, io->ci_obj) == 0) {
1037 struct ccc_io *cio = ccc_env_io(env);
1039 if (attr->ia_valid & ATTR_FILE)
1040 /* populate the file descriptor for ftruncate to honor
1041 * group lock - see LU-787 */
1042 cio->cui_fd = cl_iattr2fd(inode, attr);
1044 result = cl_io_loop(env, io);
1046 result = io->ci_result;
1048 cl_io_fini(env, io);
1049 cl_env_put(env, &refcheck);
1053 /*****************************************************************************
1059 struct lu_device *ccc2lu_dev(struct ccc_device *vdv)
1061 return &vdv->cdv_cl.cd_lu_dev;
1064 struct ccc_device *lu2ccc_dev(const struct lu_device *d)
1066 return container_of0(d, struct ccc_device, cdv_cl.cd_lu_dev);
1069 struct ccc_device *cl2ccc_dev(const struct cl_device *d)
1071 return container_of0(d, struct ccc_device, cdv_cl);
1074 struct lu_object *ccc2lu(struct ccc_object *vob)
1076 return &vob->cob_cl.co_lu;
1079 struct ccc_object *lu2ccc(const struct lu_object *obj)
1081 return container_of0(obj, struct ccc_object, cob_cl.co_lu);
1084 struct ccc_object *cl2ccc(const struct cl_object *obj)
1086 return container_of0(obj, struct ccc_object, cob_cl);
1089 struct ccc_lock *cl2ccc_lock(const struct cl_lock_slice *slice)
1091 return container_of(slice, struct ccc_lock, clk_cl);
1094 struct ccc_io *cl2ccc_io(const struct lu_env *env,
1095 const struct cl_io_slice *slice)
1099 cio = container_of(slice, struct ccc_io, cui_cl);
1100 LASSERT(cio == ccc_env_io(env));
1104 struct ccc_req *cl2ccc_req(const struct cl_req_slice *slice)
1106 return container_of0(slice, struct ccc_req, crq_cl);
1109 cfs_page_t *cl2vm_page(const struct cl_page_slice *slice)
1111 return cl2ccc_page(slice)->cpg_page;
1114 /*****************************************************************************
1119 int ccc_object_invariant(const struct cl_object *obj)
1121 struct inode *inode = ccc_object_inode(obj);
1122 struct cl_inode_info *lli = cl_i2info(inode);
1124 return (S_ISREG(cl_inode_mode(inode)) ||
1125 /* i_mode of unlinked inode is zeroed. */
1126 cl_inode_mode(inode) == 0) && lli->lli_clob == obj;
1129 struct inode *ccc_object_inode(const struct cl_object *obj)
1131 return cl2ccc(obj)->cob_inode;
1135 * Returns a pointer to cl_page associated with \a vmpage, without acquiring
1136 * additional reference to the resulting page. This is an unsafe version of
1137 * cl_vmpage_page() that can only be used under vmpage lock.
1139 struct cl_page *ccc_vmpage_page_transient(cfs_page_t *vmpage)
1141 KLASSERT(PageLocked(vmpage));
1142 return (struct cl_page *)vmpage->private;
1146 * Initialize or update CLIO structures for regular files when new
1147 * meta-data arrives from the server.
1149 * \param inode regular file inode
1150 * \param md new file metadata from MDS
1151 * - allocates cl_object if necessary,
1152 * - updated layout, if object was already here.
1154 int cl_file_inode_init(struct inode *inode, struct lustre_md *md)
1157 struct cl_inode_info *lli;
1158 struct cl_object *clob;
1159 struct lu_site *site;
1161 struct cl_object_conf conf = {
1170 LASSERT(md->body->valid & OBD_MD_FLID);
1171 LASSERT(S_ISREG(cl_inode_mode(inode)));
1173 env = cl_env_get(&refcheck);
1175 return PTR_ERR(env);
1177 site = cl_i2sbi(inode)->ll_site;
1178 lli = cl_i2info(inode);
1179 fid = &lli->lli_fid;
1180 LASSERT(fid_is_sane(fid));
1182 if (lli->lli_clob == NULL) {
1183 /* clob is slave of inode, empty lli_clob means for new inode,
1184 * there is no clob in cache with the given fid, so it is
1185 * unnecessary to perform lookup-alloc-lookup-insert, just
1186 * alloc and insert directly. */
1188 LASSERT(inode->i_state & I_NEW);
1190 conf.coc_lu.loc_flags = LOC_F_NEW;
1191 clob = cl_object_find(env, lu2cl_dev(site->ls_top_dev),
1193 if (!IS_ERR(clob)) {
1195 * No locking is necessary, as new inode is
1196 * locked by I_NEW bit.
1198 * XXX not true for call from ll_update_inode().
1200 lli->lli_clob = clob;
1201 lu_object_ref_add(&clob->co_lu, "inode", inode);
1203 result = PTR_ERR(clob);
1205 result = cl_conf_set(env, lli->lli_clob, &conf);
1206 cl_env_put(env, &refcheck);
1209 CERROR("Failure to initialize cl object "DFID": %d\n",
1215 * Wait for others drop their references of the object at first, then we drop
1216 * the last one, which will lead to the object be destroyed immediately.
1217 * Must be called after cl_object_kill() against this object.
1219 * The reason we want to do this is: destroying top object will wait for sub
1220 * objects being destroyed first, so we can't let bottom layer (e.g. from ASTs)
1221 * to initiate top object destroying which may deadlock. See bz22520.
1223 static void cl_object_put_last(struct lu_env *env, struct cl_object *obj)
1225 struct lu_object_header *header = obj->co_lu.lo_header;
1226 cfs_waitlink_t waiter;
1228 if (unlikely(cfs_atomic_read(&header->loh_ref) != 1)) {
1229 struct lu_site *site = obj->co_lu.lo_dev->ld_site;
1230 struct lu_site_bkt_data *bkt;
1232 bkt = lu_site_bkt_from_fid(site, &header->loh_fid);
1234 cfs_waitlink_init(&waiter);
1235 cfs_waitq_add(&bkt->lsb_marche_funebre, &waiter);
1238 cfs_set_current_state(CFS_TASK_UNINT);
1239 if (cfs_atomic_read(&header->loh_ref) == 1)
1241 cfs_waitq_wait(&waiter, CFS_TASK_UNINT);
1244 cfs_set_current_state(CFS_TASK_RUNNING);
1245 cfs_waitq_del(&bkt->lsb_marche_funebre, &waiter);
1248 cl_object_put(env, obj);
1251 void cl_inode_fini(struct inode *inode)
1254 struct cl_inode_info *lli = cl_i2info(inode);
1255 struct cl_object *clob = lli->lli_clob;
1262 cookie = cl_env_reenter();
1263 env = cl_env_get(&refcheck);
1264 emergency = IS_ERR(env);
1266 cfs_mutex_lock(&ccc_inode_fini_guard);
1267 LASSERT(ccc_inode_fini_env != NULL);
1268 cl_env_implant(ccc_inode_fini_env, &refcheck);
1269 env = ccc_inode_fini_env;
1272 * cl_object cache is a slave to inode cache (which, in turn
1273 * is a slave to dentry cache), don't keep cl_object in memory
1274 * when its master is evicted.
1276 cl_object_kill(env, clob);
1277 lu_object_ref_del(&clob->co_lu, "inode", inode);
1278 cl_object_put_last(env, clob);
1279 lli->lli_clob = NULL;
1281 cl_env_unplant(ccc_inode_fini_env, &refcheck);
1282 cfs_mutex_unlock(&ccc_inode_fini_guard);
1284 cl_env_put(env, &refcheck);
1285 cl_env_reexit(cookie);
1290 * return IF_* type for given lu_dirent entry.
1291 * IF_* flag shld be converted to particular OS file type in
1292 * platform llite module.
1294 __u16 ll_dirent_type_get(struct lu_dirent *ent)
1297 struct luda_type *lt;
1300 if (le32_to_cpu(ent->lde_attrs) & LUDA_TYPE) {
1301 const unsigned align = sizeof(struct luda_type) - 1;
1303 len = le16_to_cpu(ent->lde_namelen);
1304 len = (len + align) & ~align;
1305 lt = (void *) ent->lde_name + len;
1306 type = CFS_IFTODT(le16_to_cpu(lt->lt_type));
1312 * build inode number from passed @fid */
1313 __u64 cl_fid_build_ino(const struct lu_fid *fid, int api32)
1315 if (BITS_PER_LONG == 32 || api32)
1316 RETURN(fid_flatten32(fid));
1318 RETURN(fid_flatten(fid));
1322 * build inode generation from passed @fid. If our FID overflows the 32-bit
1323 * inode number then return a non-zero generation to distinguish them. */
1324 __u32 cl_fid_build_gen(const struct lu_fid *fid)
1329 if (fid_is_igif(fid)) {
1330 gen = lu_igif_gen(fid);
1334 gen = (fid_flatten(fid) >> 32);