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.
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/smp_lock.h>
50 # include <linux/quotaops.h>
51 # include <linux/highmem.h>
52 # include <linux/pagemap.h>
53 # include <linux/rbtree.h>
54 #else /* __KERNEL__ */
59 #include <sys/types.h>
61 #include <sys/queue.h>
63 # include <liblustre.h>
67 #include <obd_support.h>
68 #include <lustre_fid.h>
69 #include <lustre_lite.h>
70 #include <lustre_dlm.h>
71 #include <lustre_ver.h>
72 #include <lustre_mdc.h>
73 #include <cl_object.h>
78 #include "../llite/llite_internal.h"
80 #include "../liblustre/llite_lib.h"
83 const struct cl_req_operations ccc_req_ops;
86 * ccc_ prefix stands for "Common Client Code".
89 static cfs_mem_cache_t *ccc_lock_kmem;
90 static cfs_mem_cache_t *ccc_object_kmem;
91 static cfs_mem_cache_t *ccc_thread_kmem;
92 static cfs_mem_cache_t *ccc_session_kmem;
93 static cfs_mem_cache_t *ccc_req_kmem;
95 static struct lu_kmem_descr ccc_caches[] = {
97 .ckd_cache = &ccc_lock_kmem,
98 .ckd_name = "ccc_lock_kmem",
99 .ckd_size = sizeof (struct ccc_lock)
102 .ckd_cache = &ccc_object_kmem,
103 .ckd_name = "ccc_object_kmem",
104 .ckd_size = sizeof (struct ccc_object)
107 .ckd_cache = &ccc_thread_kmem,
108 .ckd_name = "ccc_thread_kmem",
109 .ckd_size = sizeof (struct ccc_thread_info),
112 .ckd_cache = &ccc_session_kmem,
113 .ckd_name = "ccc_session_kmem",
114 .ckd_size = sizeof (struct ccc_session)
117 .ckd_cache = &ccc_req_kmem,
118 .ckd_name = "ccc_req_kmem",
119 .ckd_size = sizeof (struct ccc_req)
126 /*****************************************************************************
128 * Vvp device and device type functions.
132 void *ccc_key_init(const struct lu_context *ctx,
133 struct lu_context_key *key)
135 struct ccc_thread_info *info;
137 OBD_SLAB_ALLOC_PTR_GFP(info, ccc_thread_kmem, CFS_ALLOC_IO);
139 info = ERR_PTR(-ENOMEM);
143 void ccc_key_fini(const struct lu_context *ctx,
144 struct lu_context_key *key, void *data)
146 struct ccc_thread_info *info = data;
147 OBD_SLAB_FREE_PTR(info, ccc_thread_kmem);
150 void *ccc_session_key_init(const struct lu_context *ctx,
151 struct lu_context_key *key)
153 struct ccc_session *session;
155 OBD_SLAB_ALLOC_PTR_GFP(session, ccc_session_kmem, CFS_ALLOC_IO);
157 session = ERR_PTR(-ENOMEM);
161 void ccc_session_key_fini(const struct lu_context *ctx,
162 struct lu_context_key *key, void *data)
164 struct ccc_session *session = data;
165 OBD_SLAB_FREE_PTR(session, ccc_session_kmem);
168 struct lu_context_key ccc_key = {
169 .lct_tags = LCT_CL_THREAD,
170 .lct_init = ccc_key_init,
171 .lct_fini = ccc_key_fini
174 struct lu_context_key ccc_session_key = {
175 .lct_tags = LCT_SESSION,
176 .lct_init = ccc_session_key_init,
177 .lct_fini = ccc_session_key_fini
181 /* type constructor/destructor: ccc_type_{init,fini,start,stop}(). */
182 // LU_TYPE_INIT_FINI(ccc, &ccc_key, &ccc_session_key);
184 int ccc_device_init(const struct lu_env *env, struct lu_device *d,
185 const char *name, struct lu_device *next)
187 struct ccc_device *vdv;
192 vdv->cdv_next = lu2cl_dev(next);
194 LASSERT(d->ld_site != NULL && next->ld_type != NULL);
195 next->ld_site = d->ld_site;
196 rc = next->ld_type->ldt_ops->ldto_device_init(
197 env, next, next->ld_type->ldt_name, NULL);
200 lu_ref_add(&next->ld_reference, "lu-stack", &lu_site_init);
205 struct lu_device *ccc_device_fini(const struct lu_env *env,
208 return cl2lu_dev(lu2ccc_dev(d)->cdv_next);
211 struct lu_device *ccc_device_alloc(const struct lu_env *env,
212 struct lu_device_type *t,
213 struct lustre_cfg *cfg,
214 const struct lu_device_operations *luops,
215 const struct cl_device_operations *clops)
217 struct ccc_device *vdv;
218 struct lu_device *lud;
219 struct cl_site *site;
225 RETURN(ERR_PTR(-ENOMEM));
227 lud = &vdv->cdv_cl.cd_lu_dev;
228 cl_device_init(&vdv->cdv_cl, t);
229 ccc2lu_dev(vdv)->ld_ops = luops;
230 vdv->cdv_cl.cd_ops = clops;
234 rc = cl_site_init(site, &vdv->cdv_cl);
236 rc = lu_site_init_finish(&site->cs_lu);
238 LASSERT(lud->ld_site == NULL);
239 CERROR("Cannot init lu_site, rc %d.\n", rc);
245 ccc_device_free(env, lud);
251 struct lu_device *ccc_device_free(const struct lu_env *env,
254 struct ccc_device *vdv = lu2ccc_dev(d);
255 struct cl_site *site = lu2cl_site(d->ld_site);
256 struct lu_device *next = cl2lu_dev(vdv->cdv_next);
258 if (d->ld_site != NULL) {
262 cl_device_fini(lu2cl_dev(d));
267 int ccc_req_init(const struct lu_env *env, struct cl_device *dev,
273 OBD_SLAB_ALLOC_PTR_GFP(vrq, ccc_req_kmem, CFS_ALLOC_IO);
275 cl_req_slice_add(req, &vrq->crq_cl, dev, &ccc_req_ops);
283 * An `emergency' environment used by ccc_inode_fini() when cl_env_get()
284 * fails. Access to this environment is serialized by ccc_inode_fini_guard
287 static struct lu_env *ccc_inode_fini_env = NULL;
290 * A mutex serializing calls to slp_inode_fini() under extreme memory
291 * pressure, when environments cannot be allocated.
293 static CFS_DEFINE_MUTEX(ccc_inode_fini_guard);
294 static int dummy_refcheck;
296 int ccc_global_init(struct lu_device_type *device_type)
300 result = lu_kmem_init(ccc_caches);
302 result = lu_device_type_init(device_type);
303 ccc_inode_fini_env = cl_env_alloc(&dummy_refcheck,
304 LCT_REMEMBER|LCT_NOREF);
305 if (IS_ERR(ccc_inode_fini_env))
306 result = PTR_ERR(ccc_inode_fini_env);
308 ccc_inode_fini_env->le_ctx.lc_cookie = 0x4;
313 void ccc_global_fini(struct lu_device_type *device_type)
315 if (ccc_inode_fini_env != NULL) {
316 cl_env_put(ccc_inode_fini_env, &dummy_refcheck);
317 ccc_inode_fini_env = NULL;
319 lu_device_type_fini(device_type);
320 lu_kmem_fini(ccc_caches);
323 /*****************************************************************************
329 struct lu_object *ccc_object_alloc(const struct lu_env *env,
330 const struct lu_object_header *unused,
331 struct lu_device *dev,
332 const struct cl_object_operations *clops,
333 const struct lu_object_operations *luops)
335 struct ccc_object *vob;
336 struct lu_object *obj;
338 OBD_SLAB_ALLOC_PTR_GFP(vob, ccc_object_kmem, CFS_ALLOC_IO);
340 struct cl_object_header *hdr;
343 hdr = &vob->cob_header;
344 cl_object_header_init(hdr);
345 lu_object_init(obj, &hdr->coh_lu, dev);
346 lu_object_add_top(&hdr->coh_lu, obj);
348 vob->cob_cl.co_ops = clops;
355 int ccc_object_init0(const struct lu_env *env,
356 struct ccc_object *vob,
357 const struct cl_object_conf *conf)
359 vob->cob_inode = conf->coc_inode;
360 vob->cob_transient_pages = 0;
364 int ccc_object_init(const struct lu_env *env, struct lu_object *obj,
365 const struct lu_object_conf *conf)
367 struct ccc_device *dev = lu2ccc_dev(obj->lo_dev);
368 struct ccc_object *vob = lu2ccc(obj);
369 struct lu_object *below;
370 struct lu_device *under;
373 under = &dev->cdv_next->cd_lu_dev;
374 below = under->ld_ops->ldo_object_alloc(env, obj->lo_header, under);
376 const struct cl_object_conf *cconf;
378 cconf = lu2cl_conf(conf);
379 CFS_INIT_LIST_HEAD(&vob->cob_pending_list);
380 lu_object_add(obj, below);
381 result = ccc_object_init0(env, vob, cconf);
387 void ccc_object_free(const struct lu_env *env, struct lu_object *obj)
389 struct ccc_object *vob = lu2ccc(obj);
392 lu_object_header_fini(obj->lo_header);
393 OBD_SLAB_FREE_PTR(vob, ccc_object_kmem);
396 int ccc_lock_init(const struct lu_env *env,
397 struct cl_object *obj, struct cl_lock *lock,
398 const struct cl_io *unused,
399 const struct cl_lock_operations *lkops)
401 struct ccc_lock *clk;
404 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
406 OBD_SLAB_ALLOC_PTR_GFP(clk, ccc_lock_kmem, CFS_ALLOC_IO);
408 cl_lock_slice_add(lock, &clk->clk_cl, obj, lkops);
415 int ccc_attr_set(const struct lu_env *env, struct cl_object *obj,
416 const struct cl_attr *attr, unsigned valid)
421 int ccc_object_glimpse(const struct lu_env *env,
422 const struct cl_object *obj, struct ost_lvb *lvb)
424 struct inode *inode = ccc_object_inode(obj);
427 lvb->lvb_mtime = cl_inode_mtime(inode);
428 lvb->lvb_atime = cl_inode_atime(inode);
429 lvb->lvb_ctime = cl_inode_ctime(inode);
435 int ccc_conf_set(const struct lu_env *env, struct cl_object *obj,
436 const struct cl_object_conf *conf)
438 /* TODO: destroy all pages attached to this object. */
442 /*****************************************************************************
448 cfs_page_t *ccc_page_vmpage(const struct lu_env *env,
449 const struct cl_page_slice *slice)
451 return cl2vm_page(slice);
454 int ccc_page_is_under_lock(const struct lu_env *env,
455 const struct cl_page_slice *slice,
458 struct ccc_io *cio = ccc_env_io(env);
459 struct cl_lock_descr *desc = &ccc_env_info(env)->cti_descr;
460 struct cl_page *page = slice->cpl_page;
466 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE ||
467 io->ci_type == CIT_FAULT) {
468 if (cio->cui_fd->fd_flags & LL_FILE_GROUP_LOCKED)
471 desc->cld_start = page->cp_index;
472 desc->cld_end = page->cp_index;
473 desc->cld_obj = page->cp_obj;
474 desc->cld_mode = CLM_READ;
475 result = cl_queue_match(&io->ci_lockset.cls_done,
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_unuse(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_wait(const struct lu_env *env, const struct cl_lock_slice *slice)
583 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
588 * Implementation of cl_lock_operations::clo_fits_into() methods for ccc
589 * layer. This function is executed every time io finds an existing lock in
590 * the lock cache while creating new lock. This function has to decide whether
591 * cached lock "fits" into io.
593 * \param slice lock to be checked
594 * \param io IO that wants a lock.
596 * \see lov_lock_fits_into().
598 int ccc_lock_fits_into(const struct lu_env *env,
599 const struct cl_lock_slice *slice,
600 const struct cl_lock_descr *need,
601 const struct cl_io *io)
603 const struct cl_lock *lock = slice->cls_lock;
604 const struct cl_lock_descr *descr = &lock->cll_descr;
605 const struct ccc_io *cio = ccc_env_io(env);
610 * Work around DLM peculiarity: it assumes that glimpse
611 * (LDLM_FL_HAS_INTENT) lock is always LCK_PR, and returns reads lock
612 * when asked for LCK_PW lock with LDLM_FL_HAS_INTENT flag set. Make
613 * sure that glimpse doesn't get CLM_WRITE top-lock, so that it
614 * doesn't enqueue CLM_WRITE sub-locks.
616 if (cio->cui_glimpse)
617 result = descr->cld_mode != CLM_WRITE;
620 * Also, don't match incomplete write locks for read, otherwise read
621 * would enqueue missing sub-locks in the write mode.
623 else if (need->cld_mode != descr->cld_mode)
624 result = lock->cll_state >= CLS_ENQUEUED;
631 * Implements cl_lock_operations::clo_state() method for ccc layer, invoked
632 * whenever lock state changes. Transfers object attributes, that might be
633 * updated as a result of lock acquiring into inode.
635 void ccc_lock_state(const struct lu_env *env,
636 const struct cl_lock_slice *slice,
637 enum cl_lock_state state)
639 struct cl_lock *lock;
640 struct cl_object *obj;
642 struct cl_attr *attr;
645 lock = slice->cls_lock;
648 * Refresh inode attributes when the lock is moving into CLS_HELD
649 * state, and only when this is a result of real enqueue, rather than
650 * of finding lock in the cache.
652 if (state == CLS_HELD && lock->cll_state < CLS_HELD) {
655 obj = slice->cls_obj;
656 inode = ccc_object_inode(obj);
657 attr = ccc_env_thread_attr(env);
659 /* vmtruncate()->ll_truncate() first sets the i_size and then
660 * the kms under both a DLM lock and the
661 * ll_inode_size_lock(). If we don't get the
662 * ll_inode_size_lock() here we can match the DLM lock and
663 * reset i_size from the kms before the truncating path has
664 * updated the kms. 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 cl_isize_lock(inode, 0);
670 cl_object_attr_lock(obj);
671 rc = cl_object_attr_get(env, obj, attr);
673 if (lock->cll_descr.cld_start == 0 &&
674 lock->cll_descr.cld_end == CL_PAGE_EOF) {
675 cl_isize_write_nolock(inode, attr->cat_kms);
676 CDEBUG(D_INODE|D_VFSTRACE,
677 DFID" updating i_size "LPU64"\n",
678 PFID(lu_object_fid(&obj->co_lu)),
679 (__u64)cl_isize_read(inode));
681 cl_inode_mtime(inode) = attr->cat_mtime;
682 cl_inode_atime(inode) = attr->cat_atime;
683 cl_inode_ctime(inode) = attr->cat_ctime;
685 CL_LOCK_DEBUG(D_INFO, env, lock, "attr_get: %d\n", rc);
687 cl_object_attr_unlock(obj);
688 cl_isize_unlock(inode, 0);
693 /*****************************************************************************
699 void ccc_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
701 struct cl_io *io = ios->cis_io;
703 CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj));
706 int ccc_io_one_lock_index(const struct lu_env *env, struct cl_io *io,
707 __u32 enqflags, enum cl_lock_mode mode,
708 pgoff_t start, pgoff_t end)
710 struct ccc_io *cio = ccc_env_io(env);
711 struct cl_lock_descr *descr = &cio->cui_link.cill_descr;
712 struct cl_object *obj = io->ci_obj;
714 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
717 CDEBUG(D_VFSTRACE, "lock: %d [%lu, %lu]\n", mode, start, end);
719 memset(&cio->cui_link, 0, sizeof cio->cui_link);
721 if (cio->cui_fd && (cio->cui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
722 descr->cld_mode = CLM_GROUP;
723 descr->cld_gid = cio->cui_fd->fd_grouplock.cg_gid;
725 descr->cld_mode = mode;
727 descr->cld_obj = obj;
728 descr->cld_start = start;
729 descr->cld_end = end;
730 descr->cld_enq_flags = enqflags;
732 cl_io_lock_add(env, io, &cio->cui_link);
736 void ccc_io_update_iov(const struct lu_env *env,
737 struct ccc_io *cio, struct cl_io *io)
740 size_t size = io->u.ci_rw.crw_count;
742 cio->cui_iov_olen = 0;
743 if (!cl_is_normalio(env, io))
746 for (i = 0; i < cio->cui_tot_nrsegs; i++) {
747 struct iovec *iv = &cio->cui_iov[i];
749 if (iv->iov_len < size)
752 if (iv->iov_len > size) {
753 cio->cui_iov_olen = iv->iov_len;
760 cio->cui_nrsegs = i + 1;
763 int ccc_io_one_lock(const struct lu_env *env, struct cl_io *io,
764 __u32 enqflags, enum cl_lock_mode mode,
765 loff_t start, loff_t end)
767 struct cl_object *obj = io->ci_obj;
768 return ccc_io_one_lock_index(env, io, enqflags, mode,
769 cl_index(obj, start), cl_index(obj, end));
772 void ccc_io_end(const struct lu_env *env, const struct cl_io_slice *ios)
774 CLOBINVRNT(env, ios->cis_io->ci_obj,
775 ccc_object_invariant(ios->cis_io->ci_obj));
778 void ccc_io_advance(const struct lu_env *env,
779 const struct cl_io_slice *ios,
782 struct ccc_io *cio = cl2ccc_io(env, ios);
783 struct cl_io *io = ios->cis_io;
784 struct cl_object *obj = ios->cis_io->ci_obj;
786 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
788 if (cl_is_normalio(env, io) && io->ci_continue) {
790 LASSERT(cio->cui_tot_nrsegs >= cio->cui_nrsegs);
791 LASSERT(cio->cui_tot_count >= nob);
793 cio->cui_iov += cio->cui_nrsegs;
794 cio->cui_tot_nrsegs -= cio->cui_nrsegs;
795 cio->cui_tot_count -= nob;
797 if (cio->cui_iov_olen) {
801 cio->cui_tot_nrsegs++;
802 iv = &cio->cui_iov[0];
803 iv->iov_base += iv->iov_len;
804 LASSERT(cio->cui_iov_olen > iv->iov_len);
805 iv->iov_len = cio->cui_iov_olen - iv->iov_len;
810 static void ccc_object_size_lock(struct cl_object *obj, int vfslock)
812 struct inode *inode = ccc_object_inode(obj);
815 cl_isize_lock(inode, 0);
816 cl_object_attr_lock(obj);
819 static void ccc_object_size_unlock(struct cl_object *obj, int vfslock)
821 struct inode *inode = ccc_object_inode(obj);
823 cl_object_attr_unlock(obj);
825 cl_isize_unlock(inode, 0);
829 * Helper function that if necessary adjusts file size (inode->i_size), when
830 * position at the offset \a pos is accessed. File size can be arbitrary stale
831 * on a Lustre client, but client at least knows KMS. If accessed area is
832 * inside [0, KMS], set file size to KMS, otherwise glimpse file size.
834 * Locking: cl_isize_lock is used to serialize changes to inode size and to
835 * protect consistency between inode size and cl_object
836 * attributes. cl_object_size_lock() protects consistency between cl_attr's of
837 * top-object and sub-objects.
839 * In page fault path cl_isize_lock cannot be taken, client has to live with
840 * the resulting races.
842 int ccc_prep_size(const struct lu_env *env, struct cl_object *obj,
843 struct cl_io *io, loff_t start, size_t count, int vfslock,
846 struct cl_attr *attr = ccc_env_thread_attr(env);
847 struct inode *inode = ccc_object_inode(obj);
848 loff_t pos = start + count - 1;
853 * Consistency guarantees: following possibilities exist for the
854 * relation between region being accessed and real file size at this
857 * (A): the region is completely inside of the file;
859 * (B-x): x bytes of region are inside of the file, the rest is
862 * (C): the region is completely outside of the file.
864 * This classification is stable under DLM lock already acquired by
865 * the caller, because to change the class, other client has to take
866 * DLM lock conflicting with our lock. Also, any updates to ->i_size
867 * by other threads on this client are serialized by
868 * ll_inode_size_lock(). This guarantees that short reads are handled
869 * correctly in the face of concurrent writes and truncates.
871 ccc_object_size_lock(obj, vfslock);
872 result = cl_object_attr_get(env, obj, attr);
877 * A glimpse is necessary to determine whether we
878 * return a short read (B) or some zeroes at the end
881 ccc_object_size_unlock(obj, vfslock);
882 result = cl_glimpse_lock(env, io, inode, obj);
883 if (result == 0 && exceed != NULL) {
884 /* If objective page index exceed end-of-file
885 * page index, return directly. Do not expect
886 * kernel will check such case correctly.
887 * linux-2.6.18-128.1.1 miss to do that.
889 loff_t size = cl_isize_read(inode);
890 unsigned long cur_index = start >> CFS_PAGE_SHIFT;
892 if ((size == 0 && cur_index != 0) ||
893 (((size - 1) >> CFS_PAGE_SHIFT) < cur_index))
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.
909 * XXX in a page fault path, change inode size without
910 * ll_inode_size_lock() held! there is a race
911 * condition with truncate path. (see ll_extent_lock)
914 * XXX i_size_write() is not used because it is not
915 * safe to take the ll_inode_size_lock() due to a
916 * potential lock inversion (bug 6077). And since
917 * it's not safe to use i_size_write() without a
918 * covering mutex we do the assignment directly. It
919 * is not critical that the size be correct.
921 if (cl_isize_read(inode) < kms) {
923 cl_isize_write_nolock(inode, kms);
925 cl_isize_write(inode, kms);
927 DFID" updating i_size "LPU64"\n",
928 PFID(lu_object_fid(&obj->co_lu)),
929 (__u64)cl_isize_read(inode));
934 ccc_object_size_unlock(obj, vfslock);
938 /*****************************************************************************
940 * Transfer operations.
944 void ccc_req_completion(const struct lu_env *env,
945 const struct cl_req_slice *slice, int ioret)
949 vrq = cl2ccc_req(slice);
950 OBD_SLAB_FREE_PTR(vrq, ccc_req_kmem);
954 * Implementation of struct cl_req_operations::cro_attr_set() for ccc
955 * layer. ccc is responsible for
973 void ccc_req_attr_set(const struct lu_env *env,
974 const struct cl_req_slice *slice,
975 const struct cl_object *obj,
976 struct cl_req_attr *attr, obd_valid flags)
980 obd_flag valid_flags;
983 inode = ccc_object_inode(obj);
984 valid_flags = OBD_MD_FLTYPE|OBD_MD_FLATIME;
986 if (flags != (obd_valid)~0ULL)
987 valid_flags |= OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME;
989 LASSERT(attr->cra_capa == NULL);
990 attr->cra_capa = cl_capa_lookup(inode,
991 slice->crs_req->crq_type);
994 if (slice->crs_req->crq_type == CRT_WRITE) {
995 if (flags & OBD_MD_FLEPOCH) {
996 oa->o_valid |= OBD_MD_FLEPOCH;
997 oa->o_ioepoch = cl_i2info(inode)->lli_ioepoch;
998 valid_flags |= OBD_MD_FLMTIME|OBD_MD_FLCTIME|
999 OBD_MD_FLUID|OBD_MD_FLGID;
1002 obdo_from_inode(oa, inode, &cl_i2info(inode)->lli_fid,
1003 valid_flags & flags);
1005 /* Bug11742 - set the OBD_FL_MMAP flag for memory mapped files */
1006 if (cfs_atomic_read(&(cl_inode2ccc(inode)->cob_mmap_cnt)) != 0) {
1007 if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
1008 oa->o_valid |= OBD_MD_FLFLAGS;
1009 oa->o_flags = OBD_FL_MMAP;
1011 oa->o_flags |= OBD_FL_MMAP;
1017 const struct cl_req_operations ccc_req_ops = {
1018 .cro_attr_set = ccc_req_attr_set,
1019 .cro_completion = ccc_req_completion
1022 int cl_setattr_ost(struct inode *inode, const struct iattr *attr,
1023 struct obd_capa *capa)
1032 env = cl_env_get(&refcheck);
1034 RETURN(PTR_ERR(env));
1036 io = ccc_env_thread_io(env);
1037 io->ci_obj = cl_i2info(inode)->lli_clob;
1039 io->u.ci_setattr.sa_attr.lvb_atime = LTIME_S(attr->ia_atime);
1040 io->u.ci_setattr.sa_attr.lvb_mtime = LTIME_S(attr->ia_mtime);
1041 io->u.ci_setattr.sa_attr.lvb_ctime = LTIME_S(attr->ia_ctime);
1042 io->u.ci_setattr.sa_attr.lvb_size = attr->ia_size;
1043 io->u.ci_setattr.sa_valid = attr->ia_valid;
1044 io->u.ci_setattr.sa_capa = capa;
1046 if (cl_io_init(env, io, CIT_SETATTR, io->ci_obj) == 0)
1047 result = cl_io_loop(env, io);
1049 result = io->ci_result;
1050 cl_io_fini(env, io);
1051 cl_env_put(env, &refcheck);
1055 /*****************************************************************************
1061 struct lu_device *ccc2lu_dev(struct ccc_device *vdv)
1063 return &vdv->cdv_cl.cd_lu_dev;
1066 struct ccc_device *lu2ccc_dev(const struct lu_device *d)
1068 return container_of0(d, struct ccc_device, cdv_cl.cd_lu_dev);
1071 struct ccc_device *cl2ccc_dev(const struct cl_device *d)
1073 return container_of0(d, struct ccc_device, cdv_cl);
1076 struct lu_object *ccc2lu(struct ccc_object *vob)
1078 return &vob->cob_cl.co_lu;
1081 struct ccc_object *lu2ccc(const struct lu_object *obj)
1083 return container_of0(obj, struct ccc_object, cob_cl.co_lu);
1086 struct ccc_object *cl2ccc(const struct cl_object *obj)
1088 return container_of0(obj, struct ccc_object, cob_cl);
1091 struct ccc_lock *cl2ccc_lock(const struct cl_lock_slice *slice)
1093 return container_of(slice, struct ccc_lock, clk_cl);
1096 struct ccc_io *cl2ccc_io(const struct lu_env *env,
1097 const struct cl_io_slice *slice)
1101 cio = container_of(slice, struct ccc_io, cui_cl);
1102 LASSERT(cio == ccc_env_io(env));
1106 struct ccc_req *cl2ccc_req(const struct cl_req_slice *slice)
1108 return container_of0(slice, struct ccc_req, crq_cl);
1111 cfs_page_t *cl2vm_page(const struct cl_page_slice *slice)
1113 return cl2ccc_page(slice)->cpg_page;
1116 /*****************************************************************************
1121 int ccc_object_invariant(const struct cl_object *obj)
1123 struct inode *inode = ccc_object_inode(obj);
1124 struct cl_inode_info *lli = cl_i2info(inode);
1126 return (S_ISREG(cl_inode_mode(inode)) ||
1127 /* i_mode of unlinked inode is zeroed. */
1128 cl_inode_mode(inode) == 0) && lli->lli_clob == obj;
1131 struct inode *ccc_object_inode(const struct cl_object *obj)
1133 return cl2ccc(obj)->cob_inode;
1137 * Returns a pointer to cl_page associated with \a vmpage, without acquiring
1138 * additional reference to the resulting page. This is an unsafe version of
1139 * cl_vmpage_page() that can only be used under vmpage lock.
1141 struct cl_page *ccc_vmpage_page_transient(cfs_page_t *vmpage)
1143 KLASSERT(PageLocked(vmpage));
1144 return (struct cl_page *)vmpage->private;
1148 * Initializes or updates CLIO part when new meta-data arrives from the
1151 * - allocates cl_object if necessary,
1152 * - updated layout, if object was already here.
1154 int cl_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 const struct cl_object_conf conf = {
1170 /* LASSERT(inode->i_state & I_NEW); */
1171 LASSERT(md->body->valid & OBD_MD_FLID);
1173 if (!S_ISREG(cl_inode_mode(inode)))
1176 env = cl_env_get(&refcheck);
1178 return PTR_ERR(env);
1180 site = cl_i2sbi(inode)->ll_site;
1181 lli = cl_i2info(inode);
1182 fid = &lli->lli_fid;
1183 LASSERT(fid_is_sane(fid));
1185 if (lli->lli_clob == NULL) {
1186 clob = cl_object_find(env, lu2cl_dev(site->ls_top_dev),
1188 if (!IS_ERR(clob)) {
1190 * No locking is necessary, as new inode is
1191 * locked by I_NEW bit.
1193 * XXX not true for call from ll_update_inode().
1195 lli->lli_clob = clob;
1196 lu_object_ref_add(&clob->co_lu, "inode", inode);
1198 result = PTR_ERR(clob);
1200 result = cl_conf_set(env, lli->lli_clob, &conf);
1201 cl_env_put(env, &refcheck);
1204 CERROR("Failure to initialize cl object "DFID": %d\n",
1210 * Wait for others drop their references of the object at first, then we drop
1211 * the last one, which will lead to the object be destroyed immediately.
1212 * Must be called after cl_object_kill() against this object.
1214 * The reason we want to do this is: destroying top object will wait for sub
1215 * objects being destroyed first, so we can't let bottom layer (e.g. from ASTs)
1216 * to initiate top object destroying which may deadlock. See bz22520.
1218 static void cl_object_put_last(struct lu_env *env, struct cl_object *obj)
1220 struct lu_object_header *header = obj->co_lu.lo_header;
1221 cfs_waitlink_t waiter;
1223 if (unlikely(cfs_atomic_read(&header->loh_ref) != 1)) {
1224 struct lu_site *site = obj->co_lu.lo_dev->ld_site;
1225 struct lu_site_bkt_data *bkt;
1227 bkt = lu_site_bkt_from_fid(site, &header->loh_fid);
1229 cfs_waitlink_init(&waiter);
1230 cfs_waitq_add(&bkt->lsb_marche_funebre, &waiter);
1233 cfs_set_current_state(CFS_TASK_UNINT);
1234 if (cfs_atomic_read(&header->loh_ref) == 1)
1236 cfs_waitq_wait(&waiter, CFS_TASK_UNINT);
1239 cfs_set_current_state(CFS_TASK_RUNNING);
1240 cfs_waitq_del(&bkt->lsb_marche_funebre, &waiter);
1243 cl_object_put(env, obj);
1246 void cl_inode_fini(struct inode *inode)
1249 struct cl_inode_info *lli = cl_i2info(inode);
1250 struct cl_object *clob = lli->lli_clob;
1257 cookie = cl_env_reenter();
1258 env = cl_env_get(&refcheck);
1259 emergency = IS_ERR(env);
1261 cfs_mutex_lock(&ccc_inode_fini_guard);
1262 LASSERT(ccc_inode_fini_env != NULL);
1263 cl_env_implant(ccc_inode_fini_env, &refcheck);
1264 env = ccc_inode_fini_env;
1267 * cl_object cache is a slave to inode cache (which, in turn
1268 * is a slave to dentry cache), don't keep cl_object in memory
1269 * when its master is evicted.
1271 cl_object_kill(env, clob);
1272 lu_object_ref_del(&clob->co_lu, "inode", inode);
1273 cl_object_put_last(env, clob);
1274 lli->lli_clob = NULL;
1276 cl_env_unplant(ccc_inode_fini_env, &refcheck);
1277 cfs_mutex_unlock(&ccc_inode_fini_guard);
1279 cl_env_put(env, &refcheck);
1280 cl_env_reexit(cookie);
1285 * return IF_* type for given lu_dirent entry.
1286 * IF_* flag shld be converted to particular OS file type in
1287 * platform llite module.
1289 __u16 ll_dirent_type_get(struct lu_dirent *ent)
1292 struct luda_type *lt;
1295 if (le32_to_cpu(ent->lde_attrs) & LUDA_TYPE) {
1296 const unsigned align = sizeof(struct luda_type) - 1;
1298 len = le16_to_cpu(ent->lde_namelen);
1299 len = (len + align) & ~align;
1300 lt = (void *) ent->lde_name + len;
1301 type = CFS_IFTODT(le16_to_cpu(lt->lt_type));
1307 * build inode number from passed @fid */
1308 __u64 cl_fid_build_ino(const struct lu_fid *fid, int api32)
1310 if (BITS_PER_LONG == 32 || api32)
1311 RETURN(fid_flatten32(fid));
1313 RETURN(fid_flatten(fid));
1317 * build inode generation from passed @fid. If our FID overflows the 32-bit
1318 * inode number then return a non-zero generation to distinguish them. */
1319 __u32 cl_fid_build_gen(const struct lu_fid *fid)
1324 if (fid_is_igif(fid)) {
1325 gen = lu_igif_gen(fid);
1329 gen = (fid_flatten(fid) >> 32);