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 * Copyright (c) 2011 Whamcloud, Inc.
36 * This file is part of Lustre, http://www.lustre.org/
37 * Lustre is a trademark of Sun Microsystems, Inc.
39 * cl code shared between vvp and liblustre (and other Lustre clients in the
42 * Author: Nikita Danilov <nikita.danilov@sun.com>
45 #define DEBUG_SUBSYSTEM S_LLITE
48 # include <libcfs/libcfs.h>
49 # include <linux/fs.h>
50 # include <linux/sched.h>
51 # include <linux/mm.h>
52 # include <linux/smp_lock.h>
53 # include <linux/quotaops.h>
54 # include <linux/highmem.h>
55 # include <linux/pagemap.h>
56 # include <linux/rbtree.h>
57 #else /* __KERNEL__ */
62 #include <sys/types.h>
64 #include <sys/queue.h>
66 # include <liblustre.h>
70 #include <obd_support.h>
71 #include <lustre_fid.h>
72 #include <lustre_lite.h>
73 #include <lustre_dlm.h>
74 #include <lustre_ver.h>
75 #include <lustre_mdc.h>
76 #include <cl_object.h>
81 #include "../llite/llite_internal.h"
83 #include "../liblustre/llite_lib.h"
86 const struct cl_req_operations ccc_req_ops;
89 * ccc_ prefix stands for "Common Client Code".
92 static cfs_mem_cache_t *ccc_lock_kmem;
93 static cfs_mem_cache_t *ccc_object_kmem;
94 static cfs_mem_cache_t *ccc_thread_kmem;
95 static cfs_mem_cache_t *ccc_session_kmem;
96 static cfs_mem_cache_t *ccc_req_kmem;
98 static struct lu_kmem_descr ccc_caches[] = {
100 .ckd_cache = &ccc_lock_kmem,
101 .ckd_name = "ccc_lock_kmem",
102 .ckd_size = sizeof (struct ccc_lock)
105 .ckd_cache = &ccc_object_kmem,
106 .ckd_name = "ccc_object_kmem",
107 .ckd_size = sizeof (struct ccc_object)
110 .ckd_cache = &ccc_thread_kmem,
111 .ckd_name = "ccc_thread_kmem",
112 .ckd_size = sizeof (struct ccc_thread_info),
115 .ckd_cache = &ccc_session_kmem,
116 .ckd_name = "ccc_session_kmem",
117 .ckd_size = sizeof (struct ccc_session)
120 .ckd_cache = &ccc_req_kmem,
121 .ckd_name = "ccc_req_kmem",
122 .ckd_size = sizeof (struct ccc_req)
129 /*****************************************************************************
131 * Vvp device and device type functions.
135 void *ccc_key_init(const struct lu_context *ctx,
136 struct lu_context_key *key)
138 struct ccc_thread_info *info;
140 OBD_SLAB_ALLOC_PTR_GFP(info, ccc_thread_kmem, CFS_ALLOC_IO);
142 info = ERR_PTR(-ENOMEM);
146 void ccc_key_fini(const struct lu_context *ctx,
147 struct lu_context_key *key, void *data)
149 struct ccc_thread_info *info = data;
150 OBD_SLAB_FREE_PTR(info, ccc_thread_kmem);
153 void *ccc_session_key_init(const struct lu_context *ctx,
154 struct lu_context_key *key)
156 struct ccc_session *session;
158 OBD_SLAB_ALLOC_PTR_GFP(session, ccc_session_kmem, CFS_ALLOC_IO);
160 session = ERR_PTR(-ENOMEM);
164 void ccc_session_key_fini(const struct lu_context *ctx,
165 struct lu_context_key *key, void *data)
167 struct ccc_session *session = data;
168 OBD_SLAB_FREE_PTR(session, ccc_session_kmem);
171 struct lu_context_key ccc_key = {
172 .lct_tags = LCT_CL_THREAD,
173 .lct_init = ccc_key_init,
174 .lct_fini = ccc_key_fini
177 struct lu_context_key ccc_session_key = {
178 .lct_tags = LCT_SESSION,
179 .lct_init = ccc_session_key_init,
180 .lct_fini = ccc_session_key_fini
184 /* type constructor/destructor: ccc_type_{init,fini,start,stop}(). */
185 // LU_TYPE_INIT_FINI(ccc, &ccc_key, &ccc_session_key);
187 int ccc_device_init(const struct lu_env *env, struct lu_device *d,
188 const char *name, struct lu_device *next)
190 struct ccc_device *vdv;
195 vdv->cdv_next = lu2cl_dev(next);
197 LASSERT(d->ld_site != NULL && next->ld_type != NULL);
198 next->ld_site = d->ld_site;
199 rc = next->ld_type->ldt_ops->ldto_device_init(
200 env, next, next->ld_type->ldt_name, NULL);
203 lu_ref_add(&next->ld_reference, "lu-stack", &lu_site_init);
208 struct lu_device *ccc_device_fini(const struct lu_env *env,
211 return cl2lu_dev(lu2ccc_dev(d)->cdv_next);
214 struct lu_device *ccc_device_alloc(const struct lu_env *env,
215 struct lu_device_type *t,
216 struct lustre_cfg *cfg,
217 const struct lu_device_operations *luops,
218 const struct cl_device_operations *clops)
220 struct ccc_device *vdv;
221 struct lu_device *lud;
222 struct cl_site *site;
228 RETURN(ERR_PTR(-ENOMEM));
230 lud = &vdv->cdv_cl.cd_lu_dev;
231 cl_device_init(&vdv->cdv_cl, t);
232 ccc2lu_dev(vdv)->ld_ops = luops;
233 vdv->cdv_cl.cd_ops = clops;
237 rc = cl_site_init(site, &vdv->cdv_cl);
239 rc = lu_site_init_finish(&site->cs_lu);
241 LASSERT(lud->ld_site == NULL);
242 CERROR("Cannot init lu_site, rc %d.\n", rc);
248 ccc_device_free(env, lud);
254 struct lu_device *ccc_device_free(const struct lu_env *env,
257 struct ccc_device *vdv = lu2ccc_dev(d);
258 struct cl_site *site = lu2cl_site(d->ld_site);
259 struct lu_device *next = cl2lu_dev(vdv->cdv_next);
261 if (d->ld_site != NULL) {
265 cl_device_fini(lu2cl_dev(d));
270 int ccc_req_init(const struct lu_env *env, struct cl_device *dev,
276 OBD_SLAB_ALLOC_PTR_GFP(vrq, ccc_req_kmem, CFS_ALLOC_IO);
278 cl_req_slice_add(req, &vrq->crq_cl, dev, &ccc_req_ops);
286 * An `emergency' environment used by ccc_inode_fini() when cl_env_get()
287 * fails. Access to this environment is serialized by ccc_inode_fini_guard
290 static struct lu_env *ccc_inode_fini_env = NULL;
293 * A mutex serializing calls to slp_inode_fini() under extreme memory
294 * pressure, when environments cannot be allocated.
296 static CFS_DEFINE_MUTEX(ccc_inode_fini_guard);
297 static int dummy_refcheck;
299 int ccc_global_init(struct lu_device_type *device_type)
303 result = lu_kmem_init(ccc_caches);
307 result = lu_device_type_init(device_type);
311 ccc_inode_fini_env = cl_env_alloc(&dummy_refcheck,
312 LCT_REMEMBER|LCT_NOREF);
313 if (IS_ERR(ccc_inode_fini_env)) {
314 result = PTR_ERR(ccc_inode_fini_env);
318 ccc_inode_fini_env->le_ctx.lc_cookie = 0x4;
321 lu_device_type_fini(device_type);
323 lu_kmem_fini(ccc_caches);
327 void ccc_global_fini(struct lu_device_type *device_type)
329 if (ccc_inode_fini_env != NULL) {
330 cl_env_put(ccc_inode_fini_env, &dummy_refcheck);
331 ccc_inode_fini_env = NULL;
333 lu_device_type_fini(device_type);
334 lu_kmem_fini(ccc_caches);
337 /*****************************************************************************
343 struct lu_object *ccc_object_alloc(const struct lu_env *env,
344 const struct lu_object_header *unused,
345 struct lu_device *dev,
346 const struct cl_object_operations *clops,
347 const struct lu_object_operations *luops)
349 struct ccc_object *vob;
350 struct lu_object *obj;
352 OBD_SLAB_ALLOC_PTR_GFP(vob, ccc_object_kmem, CFS_ALLOC_IO);
354 struct cl_object_header *hdr;
357 hdr = &vob->cob_header;
358 cl_object_header_init(hdr);
359 lu_object_init(obj, &hdr->coh_lu, dev);
360 lu_object_add_top(&hdr->coh_lu, obj);
362 vob->cob_cl.co_ops = clops;
369 int ccc_object_init0(const struct lu_env *env,
370 struct ccc_object *vob,
371 const struct cl_object_conf *conf)
373 vob->cob_inode = conf->coc_inode;
374 vob->cob_transient_pages = 0;
378 int ccc_object_init(const struct lu_env *env, struct lu_object *obj,
379 const struct lu_object_conf *conf)
381 struct ccc_device *dev = lu2ccc_dev(obj->lo_dev);
382 struct ccc_object *vob = lu2ccc(obj);
383 struct lu_object *below;
384 struct lu_device *under;
387 under = &dev->cdv_next->cd_lu_dev;
388 below = under->ld_ops->ldo_object_alloc(env, obj->lo_header, under);
390 const struct cl_object_conf *cconf;
392 cconf = lu2cl_conf(conf);
393 CFS_INIT_LIST_HEAD(&vob->cob_pending_list);
394 lu_object_add(obj, below);
395 result = ccc_object_init0(env, vob, cconf);
401 void ccc_object_free(const struct lu_env *env, struct lu_object *obj)
403 struct ccc_object *vob = lu2ccc(obj);
406 lu_object_header_fini(obj->lo_header);
407 OBD_SLAB_FREE_PTR(vob, ccc_object_kmem);
410 int ccc_lock_init(const struct lu_env *env,
411 struct cl_object *obj, struct cl_lock *lock,
412 const struct cl_io *unused,
413 const struct cl_lock_operations *lkops)
415 struct ccc_lock *clk;
418 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
420 OBD_SLAB_ALLOC_PTR_GFP(clk, ccc_lock_kmem, CFS_ALLOC_IO);
422 cl_lock_slice_add(lock, &clk->clk_cl, obj, lkops);
429 int ccc_attr_set(const struct lu_env *env, struct cl_object *obj,
430 const struct cl_attr *attr, unsigned valid)
435 int ccc_object_glimpse(const struct lu_env *env,
436 const struct cl_object *obj, struct ost_lvb *lvb)
438 struct inode *inode = ccc_object_inode(obj);
441 lvb->lvb_mtime = cl_inode_mtime(inode);
442 lvb->lvb_atime = cl_inode_atime(inode);
443 lvb->lvb_ctime = cl_inode_ctime(inode);
449 int ccc_conf_set(const struct lu_env *env, struct cl_object *obj,
450 const struct cl_object_conf *conf)
452 /* TODO: destroy all pages attached to this object. */
456 /*****************************************************************************
462 cfs_page_t *ccc_page_vmpage(const struct lu_env *env,
463 const struct cl_page_slice *slice)
465 return cl2vm_page(slice);
468 int ccc_page_is_under_lock(const struct lu_env *env,
469 const struct cl_page_slice *slice,
472 struct ccc_io *cio = ccc_env_io(env);
473 struct cl_lock_descr *desc = &ccc_env_info(env)->cti_descr;
474 struct cl_page *page = slice->cpl_page;
480 if (io->ci_type == CIT_READ || io->ci_type == CIT_WRITE ||
481 io->ci_type == CIT_FAULT) {
482 if (cio->cui_fd->fd_flags & LL_FILE_GROUP_LOCKED)
485 desc->cld_start = page->cp_index;
486 desc->cld_end = page->cp_index;
487 desc->cld_obj = page->cp_obj;
488 desc->cld_mode = CLM_READ;
489 result = cl_queue_match(&io->ci_lockset.cls_done,
497 int ccc_fail(const struct lu_env *env, const struct cl_page_slice *slice)
506 void ccc_transient_page_verify(const struct cl_page *page)
510 int ccc_transient_page_own(const struct lu_env *env,
511 const struct cl_page_slice *slice,
512 struct cl_io *unused,
515 ccc_transient_page_verify(slice->cpl_page);
519 void ccc_transient_page_assume(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_unassume(const struct lu_env *env,
527 const struct cl_page_slice *slice,
528 struct cl_io *unused)
530 ccc_transient_page_verify(slice->cpl_page);
533 void ccc_transient_page_disown(const struct lu_env *env,
534 const struct cl_page_slice *slice,
535 struct cl_io *unused)
537 ccc_transient_page_verify(slice->cpl_page);
540 void ccc_transient_page_discard(const struct lu_env *env,
541 const struct cl_page_slice *slice,
542 struct cl_io *unused)
544 struct cl_page *page = slice->cpl_page;
546 ccc_transient_page_verify(slice->cpl_page);
549 * For transient pages, remove it from the radix tree.
551 cl_page_delete(env, page);
554 int ccc_transient_page_prep(const struct lu_env *env,
555 const struct cl_page_slice *slice,
556 struct cl_io *unused)
559 /* transient page should always be sent. */
563 /*****************************************************************************
569 void ccc_lock_delete(const struct lu_env *env,
570 const struct cl_lock_slice *slice)
572 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
575 void ccc_lock_fini(const struct lu_env *env, struct cl_lock_slice *slice)
577 struct ccc_lock *clk = cl2ccc_lock(slice);
578 OBD_SLAB_FREE_PTR(clk, ccc_lock_kmem);
581 int ccc_lock_enqueue(const struct lu_env *env,
582 const struct cl_lock_slice *slice,
583 struct cl_io *unused, __u32 enqflags)
585 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
589 int ccc_lock_unuse(const struct lu_env *env, const struct cl_lock_slice *slice)
591 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
595 int ccc_lock_wait(const struct lu_env *env, const struct cl_lock_slice *slice)
597 CLOBINVRNT(env, slice->cls_obj, ccc_object_invariant(slice->cls_obj));
602 * Implementation of cl_lock_operations::clo_fits_into() methods for ccc
603 * layer. This function is executed every time io finds an existing lock in
604 * the lock cache while creating new lock. This function has to decide whether
605 * cached lock "fits" into io.
607 * \param slice lock to be checked
608 * \param io IO that wants a lock.
610 * \see lov_lock_fits_into().
612 int ccc_lock_fits_into(const struct lu_env *env,
613 const struct cl_lock_slice *slice,
614 const struct cl_lock_descr *need,
615 const struct cl_io *io)
617 const struct cl_lock *lock = slice->cls_lock;
618 const struct cl_lock_descr *descr = &lock->cll_descr;
619 const struct ccc_io *cio = ccc_env_io(env);
624 * Work around DLM peculiarity: it assumes that glimpse
625 * (LDLM_FL_HAS_INTENT) lock is always LCK_PR, and returns reads lock
626 * when asked for LCK_PW lock with LDLM_FL_HAS_INTENT flag set. Make
627 * sure that glimpse doesn't get CLM_WRITE top-lock, so that it
628 * doesn't enqueue CLM_WRITE sub-locks.
630 if (cio->cui_glimpse)
631 result = descr->cld_mode != CLM_WRITE;
634 * Also, don't match incomplete write locks for read, otherwise read
635 * would enqueue missing sub-locks in the write mode.
637 else if (need->cld_mode != descr->cld_mode)
638 result = lock->cll_state >= CLS_ENQUEUED;
645 * Implements cl_lock_operations::clo_state() method for ccc layer, invoked
646 * whenever lock state changes. Transfers object attributes, that might be
647 * updated as a result of lock acquiring into inode.
649 void ccc_lock_state(const struct lu_env *env,
650 const struct cl_lock_slice *slice,
651 enum cl_lock_state state)
653 struct cl_lock *lock;
654 struct cl_object *obj;
656 struct cl_attr *attr;
659 lock = slice->cls_lock;
662 * Refresh inode attributes when the lock is moving into CLS_HELD
663 * state, and only when this is a result of real enqueue, rather than
664 * of finding lock in the cache.
666 if (state == CLS_HELD && lock->cll_state < CLS_HELD) {
669 obj = slice->cls_obj;
670 inode = ccc_object_inode(obj);
671 attr = ccc_env_thread_attr(env);
673 /* vmtruncate()->ll_truncate() first sets the i_size and then
674 * the kms under both a DLM lock and the
675 * ll_inode_size_lock(). If we don't get the
676 * ll_inode_size_lock() here we can match the DLM lock and
677 * reset i_size from the kms before the truncating path has
678 * updated the kms. generic_file_write can then trust the
679 * stale i_size when doing appending writes and effectively
680 * cancel the result of the truncate. Getting the
681 * ll_inode_size_lock() after the enqueue maintains the DLM
682 * -> ll_inode_size_lock() acquiring order. */
683 cl_isize_lock(inode, 0);
684 cl_object_attr_lock(obj);
685 rc = cl_object_attr_get(env, obj, attr);
687 if (lock->cll_descr.cld_start == 0 &&
688 lock->cll_descr.cld_end == CL_PAGE_EOF) {
689 cl_isize_write_nolock(inode, attr->cat_kms);
690 CDEBUG(D_INODE|D_VFSTRACE,
691 DFID" updating i_size "LPU64"\n",
692 PFID(lu_object_fid(&obj->co_lu)),
693 (__u64)cl_isize_read(inode));
695 cl_inode_mtime(inode) = attr->cat_mtime;
696 cl_inode_atime(inode) = attr->cat_atime;
697 cl_inode_ctime(inode) = attr->cat_ctime;
699 CL_LOCK_DEBUG(D_INFO, env, lock, "attr_get: %d\n", rc);
701 cl_object_attr_unlock(obj);
702 cl_isize_unlock(inode, 0);
707 /*****************************************************************************
713 void ccc_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
715 struct cl_io *io = ios->cis_io;
717 CLOBINVRNT(env, io->ci_obj, ccc_object_invariant(io->ci_obj));
720 int ccc_io_one_lock_index(const struct lu_env *env, struct cl_io *io,
721 __u32 enqflags, enum cl_lock_mode mode,
722 pgoff_t start, pgoff_t end)
724 struct ccc_io *cio = ccc_env_io(env);
725 struct cl_lock_descr *descr = &cio->cui_link.cill_descr;
726 struct cl_object *obj = io->ci_obj;
728 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
731 CDEBUG(D_VFSTRACE, "lock: %d [%lu, %lu]\n", mode, start, end);
733 memset(&cio->cui_link, 0, sizeof cio->cui_link);
735 if (cio->cui_fd && (cio->cui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
736 descr->cld_mode = CLM_GROUP;
737 descr->cld_gid = cio->cui_fd->fd_grouplock.cg_gid;
739 descr->cld_mode = mode;
741 descr->cld_obj = obj;
742 descr->cld_start = start;
743 descr->cld_end = end;
744 descr->cld_enq_flags = enqflags;
746 cl_io_lock_add(env, io, &cio->cui_link);
750 void ccc_io_update_iov(const struct lu_env *env,
751 struct ccc_io *cio, struct cl_io *io)
754 size_t size = io->u.ci_rw.crw_count;
756 cio->cui_iov_olen = 0;
757 if (!cl_is_normalio(env, io))
760 for (i = 0; i < cio->cui_tot_nrsegs; i++) {
761 struct iovec *iv = &cio->cui_iov[i];
763 if (iv->iov_len < size)
766 if (iv->iov_len > size) {
767 cio->cui_iov_olen = iv->iov_len;
774 cio->cui_nrsegs = i + 1;
777 int ccc_io_one_lock(const struct lu_env *env, struct cl_io *io,
778 __u32 enqflags, enum cl_lock_mode mode,
779 loff_t start, loff_t end)
781 struct cl_object *obj = io->ci_obj;
782 return ccc_io_one_lock_index(env, io, enqflags, mode,
783 cl_index(obj, start), cl_index(obj, end));
786 void ccc_io_end(const struct lu_env *env, const struct cl_io_slice *ios)
788 CLOBINVRNT(env, ios->cis_io->ci_obj,
789 ccc_object_invariant(ios->cis_io->ci_obj));
792 void ccc_io_advance(const struct lu_env *env,
793 const struct cl_io_slice *ios,
796 struct ccc_io *cio = cl2ccc_io(env, ios);
797 struct cl_io *io = ios->cis_io;
798 struct cl_object *obj = ios->cis_io->ci_obj;
800 CLOBINVRNT(env, obj, ccc_object_invariant(obj));
802 if (cl_is_normalio(env, io) && io->ci_continue) {
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;
811 if (cio->cui_iov_olen) {
815 cio->cui_tot_nrsegs++;
816 iv = &cio->cui_iov[0];
817 iv->iov_base += iv->iov_len;
818 LASSERT(cio->cui_iov_olen > iv->iov_len);
819 iv->iov_len = cio->cui_iov_olen - iv->iov_len;
824 static void ccc_object_size_lock(struct cl_object *obj, int vfslock)
826 struct inode *inode = ccc_object_inode(obj);
829 cl_isize_lock(inode, 0);
830 cl_object_attr_lock(obj);
833 static void ccc_object_size_unlock(struct cl_object *obj, int vfslock)
835 struct inode *inode = ccc_object_inode(obj);
837 cl_object_attr_unlock(obj);
839 cl_isize_unlock(inode, 0);
843 * Helper function that if necessary adjusts file size (inode->i_size), when
844 * position at the offset \a pos is accessed. File size can be arbitrary stale
845 * on a Lustre client, but client at least knows KMS. If accessed area is
846 * inside [0, KMS], set file size to KMS, otherwise glimpse file size.
848 * Locking: cl_isize_lock is used to serialize changes to inode size and to
849 * protect consistency between inode size and cl_object
850 * attributes. cl_object_size_lock() protects consistency between cl_attr's of
851 * top-object and sub-objects.
853 * In page fault path cl_isize_lock cannot be taken, client has to live with
854 * the resulting races.
856 int ccc_prep_size(const struct lu_env *env, struct cl_object *obj,
857 struct cl_io *io, loff_t start, size_t count, int vfslock,
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, vfslock);
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, vfslock);
896 result = cl_glimpse_lock(env, io, inode, obj);
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.
923 * XXX in a page fault path, change inode size without
924 * ll_inode_size_lock() held! there is a race
925 * condition with truncate path. (see ll_extent_lock)
928 * XXX i_size_write() is not used because it is not
929 * safe to take the ll_inode_size_lock() due to a
930 * potential lock inversion (bug 6077). And since
931 * it's not safe to use i_size_write() without a
932 * covering mutex we do the assignment directly. It
933 * is not critical that the size be correct.
935 if (cl_isize_read(inode) < kms) {
937 cl_isize_write_nolock(inode, kms);
939 cl_isize_write(inode, kms);
941 DFID" updating i_size "LPU64"\n",
942 PFID(lu_object_fid(&obj->co_lu)),
943 (__u64)cl_isize_read(inode));
948 ccc_object_size_unlock(obj, vfslock);
952 /*****************************************************************************
954 * Transfer operations.
958 void ccc_req_completion(const struct lu_env *env,
959 const struct cl_req_slice *slice, int ioret)
964 cl_stats_tally(slice->crs_dev, slice->crs_req->crq_type, ioret);
966 vrq = cl2ccc_req(slice);
967 OBD_SLAB_FREE_PTR(vrq, ccc_req_kmem);
971 * Implementation of struct cl_req_operations::cro_attr_set() for ccc
972 * layer. ccc is responsible for
990 void ccc_req_attr_set(const struct lu_env *env,
991 const struct cl_req_slice *slice,
992 const struct cl_object *obj,
993 struct cl_req_attr *attr, obd_valid flags)
997 obd_flag valid_flags;
1000 inode = ccc_object_inode(obj);
1001 valid_flags = OBD_MD_FLTYPE|OBD_MD_FLATIME;
1003 if (flags != (obd_valid)~0ULL)
1004 valid_flags |= OBD_MD_FLMTIME|OBD_MD_FLCTIME|OBD_MD_FLATIME;
1006 LASSERT(attr->cra_capa == NULL);
1007 attr->cra_capa = cl_capa_lookup(inode,
1008 slice->crs_req->crq_type);
1011 if (slice->crs_req->crq_type == CRT_WRITE) {
1012 if (flags & OBD_MD_FLEPOCH) {
1013 oa->o_valid |= OBD_MD_FLEPOCH;
1014 oa->o_ioepoch = cl_i2info(inode)->lli_ioepoch;
1015 valid_flags |= OBD_MD_FLMTIME|OBD_MD_FLCTIME|
1016 OBD_MD_FLUID|OBD_MD_FLGID;
1019 obdo_from_inode(oa, inode, &cl_i2info(inode)->lli_fid,
1020 valid_flags & flags);
1022 /* Bug11742 - set the OBD_FL_MMAP flag for memory mapped files */
1023 if (cfs_atomic_read(&(cl_inode2ccc(inode)->cob_mmap_cnt)) != 0) {
1024 if (!(oa->o_valid & OBD_MD_FLFLAGS)) {
1025 oa->o_valid |= OBD_MD_FLFLAGS;
1026 oa->o_flags = OBD_FL_MMAP;
1028 oa->o_flags |= OBD_FL_MMAP;
1034 const struct cl_req_operations ccc_req_ops = {
1035 .cro_attr_set = ccc_req_attr_set,
1036 .cro_completion = ccc_req_completion
1039 int cl_setattr_ost(struct inode *inode, const struct iattr *attr,
1040 struct obd_capa *capa)
1049 env = cl_env_get(&refcheck);
1051 RETURN(PTR_ERR(env));
1053 io = ccc_env_thread_io(env);
1054 io->ci_obj = cl_i2info(inode)->lli_clob;
1056 io->u.ci_setattr.sa_attr.lvb_atime = LTIME_S(attr->ia_atime);
1057 io->u.ci_setattr.sa_attr.lvb_mtime = LTIME_S(attr->ia_mtime);
1058 io->u.ci_setattr.sa_attr.lvb_ctime = LTIME_S(attr->ia_ctime);
1059 io->u.ci_setattr.sa_attr.lvb_size = attr->ia_size;
1060 io->u.ci_setattr.sa_valid = attr->ia_valid;
1061 io->u.ci_setattr.sa_capa = capa;
1063 if (cl_io_init(env, io, CIT_SETATTR, io->ci_obj) == 0)
1064 result = cl_io_loop(env, io);
1066 result = io->ci_result;
1067 cl_io_fini(env, io);
1068 cl_env_put(env, &refcheck);
1072 /*****************************************************************************
1078 struct lu_device *ccc2lu_dev(struct ccc_device *vdv)
1080 return &vdv->cdv_cl.cd_lu_dev;
1083 struct ccc_device *lu2ccc_dev(const struct lu_device *d)
1085 return container_of0(d, struct ccc_device, cdv_cl.cd_lu_dev);
1088 struct ccc_device *cl2ccc_dev(const struct cl_device *d)
1090 return container_of0(d, struct ccc_device, cdv_cl);
1093 struct lu_object *ccc2lu(struct ccc_object *vob)
1095 return &vob->cob_cl.co_lu;
1098 struct ccc_object *lu2ccc(const struct lu_object *obj)
1100 return container_of0(obj, struct ccc_object, cob_cl.co_lu);
1103 struct ccc_object *cl2ccc(const struct cl_object *obj)
1105 return container_of0(obj, struct ccc_object, cob_cl);
1108 struct ccc_lock *cl2ccc_lock(const struct cl_lock_slice *slice)
1110 return container_of(slice, struct ccc_lock, clk_cl);
1113 struct ccc_io *cl2ccc_io(const struct lu_env *env,
1114 const struct cl_io_slice *slice)
1118 cio = container_of(slice, struct ccc_io, cui_cl);
1119 LASSERT(cio == ccc_env_io(env));
1123 struct ccc_req *cl2ccc_req(const struct cl_req_slice *slice)
1125 return container_of0(slice, struct ccc_req, crq_cl);
1128 cfs_page_t *cl2vm_page(const struct cl_page_slice *slice)
1130 return cl2ccc_page(slice)->cpg_page;
1133 /*****************************************************************************
1138 int ccc_object_invariant(const struct cl_object *obj)
1140 struct inode *inode = ccc_object_inode(obj);
1141 struct cl_inode_info *lli = cl_i2info(inode);
1143 return (S_ISREG(cl_inode_mode(inode)) ||
1144 /* i_mode of unlinked inode is zeroed. */
1145 cl_inode_mode(inode) == 0) && lli->lli_clob == obj;
1148 struct inode *ccc_object_inode(const struct cl_object *obj)
1150 return cl2ccc(obj)->cob_inode;
1154 * Returns a pointer to cl_page associated with \a vmpage, without acquiring
1155 * additional reference to the resulting page. This is an unsafe version of
1156 * cl_vmpage_page() that can only be used under vmpage lock.
1158 struct cl_page *ccc_vmpage_page_transient(cfs_page_t *vmpage)
1160 KLASSERT(PageLocked(vmpage));
1161 return (struct cl_page *)vmpage->private;
1165 * Initializes or updates CLIO part when new meta-data arrives from the
1168 * - allocates cl_object if necessary,
1169 * - updated layout, if object was already here.
1171 int cl_inode_init(struct inode *inode, struct lustre_md *md)
1174 struct cl_inode_info *lli;
1175 struct cl_object *clob;
1176 struct lu_site *site;
1178 struct cl_object_conf conf = {
1187 LASSERT(md->body->valid & OBD_MD_FLID);
1189 if (!S_ISREG(cl_inode_mode(inode)))
1192 env = cl_env_get(&refcheck);
1194 return PTR_ERR(env);
1196 site = cl_i2sbi(inode)->ll_site;
1197 lli = cl_i2info(inode);
1198 fid = &lli->lli_fid;
1199 LASSERT(fid_is_sane(fid));
1201 if (lli->lli_clob == NULL) {
1202 /* clob is slave of inode, empty lli_clob means for new inode,
1203 * there is no clob in cache with the given fid, so it is
1204 * unnecessary to perform lookup-alloc-lookup-insert, just
1205 * alloc and insert directly. */
1207 LASSERT(inode->i_state & I_NEW);
1209 conf.coc_lu.loc_flags = LOC_F_NEW;
1210 clob = cl_object_find(env, lu2cl_dev(site->ls_top_dev),
1212 if (!IS_ERR(clob)) {
1214 * No locking is necessary, as new inode is
1215 * locked by I_NEW bit.
1217 * XXX not true for call from ll_update_inode().
1219 lli->lli_clob = clob;
1220 lu_object_ref_add(&clob->co_lu, "inode", inode);
1222 result = PTR_ERR(clob);
1224 result = cl_conf_set(env, lli->lli_clob, &conf);
1225 cl_env_put(env, &refcheck);
1228 CERROR("Failure to initialize cl object "DFID": %d\n",
1234 * Wait for others drop their references of the object at first, then we drop
1235 * the last one, which will lead to the object be destroyed immediately.
1236 * Must be called after cl_object_kill() against this object.
1238 * The reason we want to do this is: destroying top object will wait for sub
1239 * objects being destroyed first, so we can't let bottom layer (e.g. from ASTs)
1240 * to initiate top object destroying which may deadlock. See bz22520.
1242 static void cl_object_put_last(struct lu_env *env, struct cl_object *obj)
1244 struct lu_object_header *header = obj->co_lu.lo_header;
1245 cfs_waitlink_t waiter;
1247 if (unlikely(cfs_atomic_read(&header->loh_ref) != 1)) {
1248 struct lu_site *site = obj->co_lu.lo_dev->ld_site;
1249 struct lu_site_bkt_data *bkt;
1251 bkt = lu_site_bkt_from_fid(site, &header->loh_fid);
1253 cfs_waitlink_init(&waiter);
1254 cfs_waitq_add(&bkt->lsb_marche_funebre, &waiter);
1257 cfs_set_current_state(CFS_TASK_UNINT);
1258 if (cfs_atomic_read(&header->loh_ref) == 1)
1260 cfs_waitq_wait(&waiter, CFS_TASK_UNINT);
1263 cfs_set_current_state(CFS_TASK_RUNNING);
1264 cfs_waitq_del(&bkt->lsb_marche_funebre, &waiter);
1267 cl_object_put(env, obj);
1270 void cl_inode_fini(struct inode *inode)
1273 struct cl_inode_info *lli = cl_i2info(inode);
1274 struct cl_object *clob = lli->lli_clob;
1281 cookie = cl_env_reenter();
1282 env = cl_env_get(&refcheck);
1283 emergency = IS_ERR(env);
1285 cfs_mutex_lock(&ccc_inode_fini_guard);
1286 LASSERT(ccc_inode_fini_env != NULL);
1287 cl_env_implant(ccc_inode_fini_env, &refcheck);
1288 env = ccc_inode_fini_env;
1291 * cl_object cache is a slave to inode cache (which, in turn
1292 * is a slave to dentry cache), don't keep cl_object in memory
1293 * when its master is evicted.
1295 cl_object_kill(env, clob);
1296 lu_object_ref_del(&clob->co_lu, "inode", inode);
1297 cl_object_put_last(env, clob);
1298 lli->lli_clob = NULL;
1300 cl_env_unplant(ccc_inode_fini_env, &refcheck);
1301 cfs_mutex_unlock(&ccc_inode_fini_guard);
1303 cl_env_put(env, &refcheck);
1304 cl_env_reexit(cookie);
1309 * return IF_* type for given lu_dirent entry.
1310 * IF_* flag shld be converted to particular OS file type in
1311 * platform llite module.
1313 __u16 ll_dirent_type_get(struct lu_dirent *ent)
1316 struct luda_type *lt;
1319 if (le32_to_cpu(ent->lde_attrs) & LUDA_TYPE) {
1320 const unsigned align = sizeof(struct luda_type) - 1;
1322 len = le16_to_cpu(ent->lde_namelen);
1323 len = (len + align) & ~align;
1324 lt = (void *) ent->lde_name + len;
1325 type = CFS_IFTODT(le16_to_cpu(lt->lt_type));
1331 * build inode number from passed @fid */
1332 __u64 cl_fid_build_ino(const struct lu_fid *fid, int api32)
1334 if (BITS_PER_LONG == 32 || api32)
1335 RETURN(fid_flatten32(fid));
1337 RETURN(fid_flatten(fid));
1341 * build inode generation from passed @fid. If our FID overflows the 32-bit
1342 * inode number then return a non-zero generation to distinguish them. */
1343 __u32 cl_fid_build_gen(const struct lu_fid *fid)
1348 if (fid_is_igif(fid)) {
1349 gen = lu_igif_gen(fid);
1353 gen = (fid_flatten(fid) >> 32);