4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License version 2 only,
8 * as published by the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but
11 * WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
13 * General Public License version 2 for more details (a copy is included
14 * in the LICENSE file that accompanied this code).
16 * You should have received a copy of the GNU General Public License
17 * version 2 along with this program; If not, see
18 * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
20 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
21 * CA 95054 USA or visit www.sun.com if you need additional information or
27 * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28 * Use is subject to license terms.
30 * Copyright (c) 2011, 2014, Intel Corporation.
33 * This file is part of Lustre, http://www.lustre.org/
34 * Lustre is a trademark of Sun Microsystems, Inc.
36 * cl code shared between vvp and liblustre (and other Lustre clients in the
39 * Author: Nikita Danilov <nikita.danilov@sun.com>
42 #define DEBUG_SUBSYSTEM S_LLITE
44 #include <libcfs/libcfs.h>
46 #include <linux/sched.h>
48 #include <linux/quotaops.h>
49 #include <linux/highmem.h>
50 #include <linux/pagemap.h>
51 #include <linux/rbtree.h>
54 #include <obd_support.h>
55 #include <lustre_fid.h>
56 #include <lustre_dlm.h>
57 #include <lustre_ver.h>
58 #include <lustre_mdc.h>
59 #include <cl_object.h>
61 #include "llite_internal.h"
63 static const struct cl_req_operations ccc_req_ops;
66 * ccc_ prefix stands for "Common Client Code".
69 static struct kmem_cache *ccc_lock_kmem;
70 static struct kmem_cache *ccc_thread_kmem;
71 static struct kmem_cache *ccc_session_kmem;
72 static struct kmem_cache *ccc_req_kmem;
74 static struct lu_kmem_descr ccc_caches[] = {
76 .ckd_cache = &ccc_lock_kmem,
77 .ckd_name = "ccc_lock_kmem",
78 .ckd_size = sizeof (struct ccc_lock)
81 .ckd_cache = &ccc_thread_kmem,
82 .ckd_name = "ccc_thread_kmem",
83 .ckd_size = sizeof (struct ccc_thread_info),
86 .ckd_cache = &ccc_session_kmem,
87 .ckd_name = "ccc_session_kmem",
88 .ckd_size = sizeof (struct ccc_session)
91 .ckd_cache = &ccc_req_kmem,
92 .ckd_name = "ccc_req_kmem",
93 .ckd_size = sizeof (struct ccc_req)
100 /*****************************************************************************
102 * Vvp device and device type functions.
106 void *ccc_key_init(const struct lu_context *ctx, struct lu_context_key *key)
108 struct ccc_thread_info *info;
110 OBD_SLAB_ALLOC_PTR_GFP(info, ccc_thread_kmem, GFP_NOFS);
112 info = ERR_PTR(-ENOMEM);
116 void ccc_key_fini(const struct lu_context *ctx,
117 struct lu_context_key *key, void *data)
119 struct ccc_thread_info *info = data;
120 OBD_SLAB_FREE_PTR(info, ccc_thread_kmem);
123 void *ccc_session_key_init(const struct lu_context *ctx,
124 struct lu_context_key *key)
126 struct ccc_session *session;
128 OBD_SLAB_ALLOC_PTR_GFP(session, ccc_session_kmem, GFP_NOFS);
130 session = ERR_PTR(-ENOMEM);
134 void ccc_session_key_fini(const struct lu_context *ctx,
135 struct lu_context_key *key, void *data)
137 struct ccc_session *session = data;
138 OBD_SLAB_FREE_PTR(session, ccc_session_kmem);
141 struct lu_context_key ccc_key = {
142 .lct_tags = LCT_CL_THREAD,
143 .lct_init = ccc_key_init,
144 .lct_fini = ccc_key_fini
147 struct lu_context_key ccc_session_key = {
148 .lct_tags = LCT_SESSION,
149 .lct_init = ccc_session_key_init,
150 .lct_fini = ccc_session_key_fini
153 int ccc_req_init(const struct lu_env *env, struct cl_device *dev,
159 OBD_SLAB_ALLOC_PTR_GFP(vrq, ccc_req_kmem, GFP_NOFS);
161 cl_req_slice_add(req, &vrq->crq_cl, dev, &ccc_req_ops);
169 * An `emergency' environment used by ccc_inode_fini() when cl_env_get()
170 * fails. Access to this environment is serialized by ccc_inode_fini_guard
173 static struct lu_env *ccc_inode_fini_env = NULL;
176 * A mutex serializing calls to slp_inode_fini() under extreme memory
177 * pressure, when environments cannot be allocated.
179 static DEFINE_MUTEX(ccc_inode_fini_guard);
180 static int dummy_refcheck;
182 int ccc_global_init(struct lu_device_type *device_type)
186 result = lu_kmem_init(ccc_caches);
190 result = lu_device_type_init(device_type);
194 ccc_inode_fini_env = cl_env_alloc(&dummy_refcheck,
195 LCT_REMEMBER|LCT_NOREF);
196 if (IS_ERR(ccc_inode_fini_env)) {
197 result = PTR_ERR(ccc_inode_fini_env);
201 ccc_inode_fini_env->le_ctx.lc_cookie = 0x4;
204 lu_device_type_fini(device_type);
206 lu_kmem_fini(ccc_caches);
210 void ccc_global_fini(struct lu_device_type *device_type)
212 if (ccc_inode_fini_env != NULL) {
213 cl_env_put(ccc_inode_fini_env, &dummy_refcheck);
214 ccc_inode_fini_env = NULL;
216 lu_device_type_fini(device_type);
217 lu_kmem_fini(ccc_caches);
220 int ccc_lock_init(const struct lu_env *env,
221 struct cl_object *obj, struct cl_lock *lock,
222 const struct cl_io *unused,
223 const struct cl_lock_operations *lkops)
225 struct ccc_lock *clk;
228 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
230 OBD_SLAB_ALLOC_PTR_GFP(clk, ccc_lock_kmem, GFP_NOFS);
232 cl_lock_slice_add(lock, &clk->clk_cl, obj, lkops);
239 static void vvp_object_size_lock(struct cl_object *obj)
241 struct inode *inode = vvp_object_inode(obj);
243 ll_inode_size_lock(inode);
244 cl_object_attr_lock(obj);
247 static void vvp_object_size_unlock(struct cl_object *obj)
249 struct inode *inode = vvp_object_inode(obj);
251 cl_object_attr_unlock(obj);
252 ll_inode_size_unlock(inode);
255 /*****************************************************************************
261 void ccc_lock_fini(const struct lu_env *env, struct cl_lock_slice *slice)
263 struct ccc_lock *clk = cl2ccc_lock(slice);
264 OBD_SLAB_FREE_PTR(clk, ccc_lock_kmem);
267 int ccc_lock_enqueue(const struct lu_env *env,
268 const struct cl_lock_slice *slice,
269 struct cl_io *unused, struct cl_sync_io *anchor)
271 CLOBINVRNT(env, slice->cls_obj, vvp_object_invariant(slice->cls_obj));
275 /*****************************************************************************
281 void ccc_io_fini(const struct lu_env *env, const struct cl_io_slice *ios)
283 struct cl_io *io = ios->cis_io;
285 CLOBINVRNT(env, io->ci_obj, vvp_object_invariant(io->ci_obj));
288 int ccc_io_one_lock_index(const struct lu_env *env, struct cl_io *io,
289 __u32 enqflags, enum cl_lock_mode mode,
290 pgoff_t start, pgoff_t end)
292 struct ccc_io *cio = ccc_env_io(env);
293 struct cl_lock_descr *descr = &cio->cui_link.cill_descr;
294 struct cl_object *obj = io->ci_obj;
296 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
299 CDEBUG(D_VFSTRACE, "lock: %d [%lu, %lu]\n", mode, start, end);
301 memset(&cio->cui_link, 0, sizeof cio->cui_link);
303 if (cio->cui_fd && (cio->cui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
304 descr->cld_mode = CLM_GROUP;
305 descr->cld_gid = cio->cui_fd->fd_grouplock.cg_gid;
307 descr->cld_mode = mode;
309 descr->cld_obj = obj;
310 descr->cld_start = start;
311 descr->cld_end = end;
312 descr->cld_enq_flags = enqflags;
314 cl_io_lock_add(env, io, &cio->cui_link);
318 void ccc_io_update_iov(const struct lu_env *env,
319 struct ccc_io *cio, struct cl_io *io)
322 size_t size = io->u.ci_rw.crw_count;
324 cio->cui_iov_olen = 0;
325 if (!cl_is_normalio(env, io) || cio->cui_tot_nrsegs == 0)
328 for (i = 0; i < cio->cui_tot_nrsegs; i++) {
329 struct iovec *iv = &cio->cui_iov[i];
331 if (iv->iov_len < size)
334 if (iv->iov_len > size) {
335 cio->cui_iov_olen = iv->iov_len;
342 cio->cui_nrsegs = i + 1;
343 LASSERTF(cio->cui_tot_nrsegs >= cio->cui_nrsegs,
344 "tot_nrsegs: %lu, nrsegs: %lu\n",
345 cio->cui_tot_nrsegs, cio->cui_nrsegs);
348 int ccc_io_one_lock(const struct lu_env *env, struct cl_io *io,
349 __u32 enqflags, enum cl_lock_mode mode,
350 loff_t start, loff_t end)
352 struct cl_object *obj = io->ci_obj;
353 return ccc_io_one_lock_index(env, io, enqflags, mode,
354 cl_index(obj, start), cl_index(obj, end));
357 void ccc_io_end(const struct lu_env *env, const struct cl_io_slice *ios)
359 CLOBINVRNT(env, ios->cis_io->ci_obj,
360 vvp_object_invariant(ios->cis_io->ci_obj));
363 void ccc_io_advance(const struct lu_env *env,
364 const struct cl_io_slice *ios,
367 struct ccc_io *cio = cl2ccc_io(env, ios);
368 struct cl_io *io = ios->cis_io;
369 struct cl_object *obj = ios->cis_io->ci_obj;
371 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
373 if (!cl_is_normalio(env, io))
376 LASSERT(cio->cui_tot_nrsegs >= cio->cui_nrsegs);
377 LASSERT(cio->cui_tot_count >= nob);
379 cio->cui_iov += cio->cui_nrsegs;
380 cio->cui_tot_nrsegs -= cio->cui_nrsegs;
381 cio->cui_tot_count -= nob;
384 if (cio->cui_iov_olen > 0) {
388 cio->cui_tot_nrsegs++;
389 iv = &cio->cui_iov[0];
390 if (io->ci_continue) {
391 iv->iov_base += iv->iov_len;
392 LASSERT(cio->cui_iov_olen > iv->iov_len);
393 iv->iov_len = cio->cui_iov_olen - iv->iov_len;
395 /* restore the iov_len, in case of restart io. */
396 iv->iov_len = cio->cui_iov_olen;
398 cio->cui_iov_olen = 0;
403 * Helper function that if necessary adjusts file size (inode->i_size), when
404 * position at the offset \a pos is accessed. File size can be arbitrary stale
405 * on a Lustre client, but client at least knows KMS. If accessed area is
406 * inside [0, KMS], set file size to KMS, otherwise glimpse file size.
408 * Locking: cl_isize_lock is used to serialize changes to inode size and to
409 * protect consistency between inode size and cl_object
410 * attributes. cl_object_size_lock() protects consistency between cl_attr's of
411 * top-object and sub-objects.
413 int ccc_prep_size(const struct lu_env *env, struct cl_object *obj,
414 struct cl_io *io, loff_t start, size_t count, int *exceed)
416 struct cl_attr *attr = ccc_env_thread_attr(env);
417 struct inode *inode = vvp_object_inode(obj);
418 loff_t pos = start + count - 1;
423 * Consistency guarantees: following possibilities exist for the
424 * relation between region being accessed and real file size at this
427 * (A): the region is completely inside of the file;
429 * (B-x): x bytes of region are inside of the file, the rest is
432 * (C): the region is completely outside of the file.
434 * This classification is stable under DLM lock already acquired by
435 * the caller, because to change the class, other client has to take
436 * DLM lock conflicting with our lock. Also, any updates to ->i_size
437 * by other threads on this client are serialized by
438 * ll_inode_size_lock(). This guarantees that short reads are handled
439 * correctly in the face of concurrent writes and truncates.
441 vvp_object_size_lock(obj);
442 result = cl_object_attr_get(env, obj, attr);
447 * A glimpse is necessary to determine whether we
448 * return a short read (B) or some zeroes at the end
451 vvp_object_size_unlock(obj);
452 result = cl_glimpse_lock(env, io, inode, obj, 0);
453 if (result == 0 && exceed != NULL) {
454 /* If objective page index exceed end-of-file
455 * page index, return directly. Do not expect
456 * kernel will check such case correctly.
457 * linux-2.6.18-128.1.1 miss to do that.
459 loff_t size = i_size_read(inode);
460 unsigned long cur_index = start >>
463 if ((size == 0 && cur_index != 0) ||
464 (((size - 1) >> PAGE_CACHE_SHIFT) <
471 * region is within kms and, hence, within real file
472 * size (A). We need to increase i_size to cover the
473 * read region so that generic_file_read() will do its
474 * job, but that doesn't mean the kms size is
475 * _correct_, it is only the _minimum_ size. If
476 * someone does a stat they will get the correct size
477 * which will always be >= the kms value here.
480 if (i_size_read(inode) < kms) {
481 i_size_write(inode, kms);
483 DFID" updating i_size "LPU64"\n",
484 PFID(lu_object_fid(&obj->co_lu)),
485 (__u64)i_size_read(inode));
491 vvp_object_size_unlock(obj);
496 /*****************************************************************************
498 * Transfer operations.
502 void ccc_req_completion(const struct lu_env *env,
503 const struct cl_req_slice *slice, int ioret)
508 cl_stats_tally(slice->crs_dev, slice->crs_req->crq_type, ioret);
510 vrq = cl2ccc_req(slice);
511 OBD_SLAB_FREE_PTR(vrq, ccc_req_kmem);
515 * Implementation of struct cl_req_operations::cro_attr_set() for ccc
516 * layer. ccc is responsible for
534 void ccc_req_attr_set(const struct lu_env *env,
535 const struct cl_req_slice *slice,
536 const struct cl_object *obj,
537 struct cl_req_attr *attr, obd_valid flags)
541 obd_flag valid_flags;
544 inode = vvp_object_inode(obj);
545 valid_flags = OBD_MD_FLTYPE;
547 if ((flags & OBD_MD_FLOSSCAPA) != 0) {
548 LASSERT(attr->cra_capa == NULL);
549 attr->cra_capa = cl_capa_lookup(inode,
550 slice->crs_req->crq_type);
553 if (slice->crs_req->crq_type == CRT_WRITE) {
554 if (flags & OBD_MD_FLEPOCH) {
555 oa->o_valid |= OBD_MD_FLEPOCH;
556 oa->o_ioepoch = ll_i2info(inode)->lli_ioepoch;
557 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME |
558 OBD_MD_FLUID | OBD_MD_FLGID;
561 obdo_from_inode(oa, inode, valid_flags & flags);
562 obdo_set_parent_fid(oa, &ll_i2info(inode)->lli_fid);
563 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_INVALID_PFID))
565 memcpy(attr->cra_jobid, ll_i2info(inode)->lli_jobid,
569 static const struct cl_req_operations ccc_req_ops = {
570 .cro_attr_set = ccc_req_attr_set,
571 .cro_completion = ccc_req_completion
574 int cl_setattr_ost(struct inode *inode, const struct iattr *attr,
575 struct obd_capa *capa)
584 env = cl_env_get(&refcheck);
586 RETURN(PTR_ERR(env));
588 io = ccc_env_thread_io(env);
589 io->ci_obj = ll_i2info(inode)->lli_clob;
591 io->u.ci_setattr.sa_attr.lvb_atime = LTIME_S(attr->ia_atime);
592 io->u.ci_setattr.sa_attr.lvb_mtime = LTIME_S(attr->ia_mtime);
593 io->u.ci_setattr.sa_attr.lvb_ctime = LTIME_S(attr->ia_ctime);
594 io->u.ci_setattr.sa_attr.lvb_size = attr->ia_size;
595 io->u.ci_setattr.sa_valid = attr->ia_valid;
596 io->u.ci_setattr.sa_capa = capa;
599 if (cl_io_init(env, io, CIT_SETATTR, io->ci_obj) == 0) {
600 struct ccc_io *cio = ccc_env_io(env);
602 if (attr->ia_valid & ATTR_FILE)
603 /* populate the file descriptor for ftruncate to honor
604 * group lock - see LU-787 */
605 cio->cui_fd = LUSTRE_FPRIVATE(attr->ia_file);
607 result = cl_io_loop(env, io);
609 result = io->ci_result;
612 if (unlikely(io->ci_need_restart))
614 /* HSM import case: file is released, cannot be restored
615 * no need to fail except if restore registration failed
617 if (result == -ENODATA && io->ci_restore_needed &&
618 io->ci_result != -ENODATA)
620 cl_env_put(env, &refcheck);
624 /*****************************************************************************
630 struct ccc_lock *cl2ccc_lock(const struct cl_lock_slice *slice)
632 return container_of(slice, struct ccc_lock, clk_cl);
635 struct ccc_io *cl2ccc_io(const struct lu_env *env,
636 const struct cl_io_slice *slice)
640 cio = container_of(slice, struct ccc_io, cui_cl);
641 LASSERT(cio == ccc_env_io(env));
645 struct ccc_req *cl2ccc_req(const struct cl_req_slice *slice)
647 return container_of0(slice, struct ccc_req, crq_cl);
651 * Initialize or update CLIO structures for regular files when new
652 * meta-data arrives from the server.
654 * \param inode regular file inode
655 * \param md new file metadata from MDS
656 * - allocates cl_object if necessary,
657 * - updated layout, if object was already here.
659 int cl_file_inode_init(struct inode *inode, struct lustre_md *md)
662 struct ll_inode_info *lli;
663 struct cl_object *clob;
664 struct lu_site *site;
666 struct cl_object_conf conf = {
675 LASSERT(md->body->mbo_valid & OBD_MD_FLID);
676 LASSERT(S_ISREG(inode->i_mode));
678 env = cl_env_get(&refcheck);
682 site = ll_i2sbi(inode)->ll_site;
683 lli = ll_i2info(inode);
685 LASSERT(fid_is_sane(fid));
687 if (lli->lli_clob == NULL) {
688 /* clob is slave of inode, empty lli_clob means for new inode,
689 * there is no clob in cache with the given fid, so it is
690 * unnecessary to perform lookup-alloc-lookup-insert, just
691 * alloc and insert directly. */
692 LASSERT(inode->i_state & I_NEW);
693 conf.coc_lu.loc_flags = LOC_F_NEW;
694 clob = cl_object_find(env, lu2cl_dev(site->ls_top_dev),
698 * No locking is necessary, as new inode is
699 * locked by I_NEW bit.
701 lli->lli_clob = clob;
702 lli->lli_has_smd = lsm_has_objects(md->lsm);
703 lu_object_ref_add(&clob->co_lu, "inode", inode);
705 result = PTR_ERR(clob);
707 result = cl_conf_set(env, lli->lli_clob, &conf);
710 cl_env_put(env, &refcheck);
713 CERROR("Failure to initialize cl object "DFID": %d\n",
719 * Wait for others drop their references of the object at first, then we drop
720 * the last one, which will lead to the object be destroyed immediately.
721 * Must be called after cl_object_kill() against this object.
723 * The reason we want to do this is: destroying top object will wait for sub
724 * objects being destroyed first, so we can't let bottom layer (e.g. from ASTs)
725 * to initiate top object destroying which may deadlock. See bz22520.
727 static void cl_object_put_last(struct lu_env *env, struct cl_object *obj)
729 struct lu_object_header *header = obj->co_lu.lo_header;
732 if (unlikely(atomic_read(&header->loh_ref) != 1)) {
733 struct lu_site *site = obj->co_lu.lo_dev->ld_site;
734 struct lu_site_bkt_data *bkt;
736 bkt = lu_site_bkt_from_fid(site, &header->loh_fid);
738 init_waitqueue_entry_current(&waiter);
739 add_wait_queue(&bkt->lsb_marche_funebre, &waiter);
742 set_current_state(TASK_UNINTERRUPTIBLE);
743 if (atomic_read(&header->loh_ref) == 1)
745 waitq_wait(&waiter, TASK_UNINTERRUPTIBLE);
748 set_current_state(TASK_RUNNING);
749 remove_wait_queue(&bkt->lsb_marche_funebre, &waiter);
752 cl_object_put(env, obj);
755 void cl_inode_fini(struct inode *inode)
758 struct ll_inode_info *lli = ll_i2info(inode);
759 struct cl_object *clob = lli->lli_clob;
766 cookie = cl_env_reenter();
767 env = cl_env_get(&refcheck);
768 emergency = IS_ERR(env);
770 mutex_lock(&ccc_inode_fini_guard);
771 LASSERT(ccc_inode_fini_env != NULL);
772 cl_env_implant(ccc_inode_fini_env, &refcheck);
773 env = ccc_inode_fini_env;
776 * cl_object cache is a slave to inode cache (which, in turn
777 * is a slave to dentry cache), don't keep cl_object in memory
778 * when its master is evicted.
780 cl_object_kill(env, clob);
781 lu_object_ref_del(&clob->co_lu, "inode", inode);
782 cl_object_put_last(env, clob);
783 lli->lli_clob = NULL;
785 cl_env_unplant(ccc_inode_fini_env, &refcheck);
786 mutex_unlock(&ccc_inode_fini_guard);
788 cl_env_put(env, &refcheck);
789 cl_env_reexit(cookie);
794 * return IF_* type for given lu_dirent entry.
795 * IF_* flag shld be converted to particular OS file type in
796 * platform llite module.
798 __u16 ll_dirent_type_get(struct lu_dirent *ent)
801 struct luda_type *lt;
804 if (le32_to_cpu(ent->lde_attrs) & LUDA_TYPE) {
805 const unsigned align = sizeof(struct luda_type) - 1;
807 len = le16_to_cpu(ent->lde_namelen);
808 len = (len + align) & ~align;
809 lt = (void *)ent->lde_name + len;
810 type = IFTODT(le16_to_cpu(lt->lt_type));
816 * build inode number from passed @fid */
817 __u64 cl_fid_build_ino(const struct lu_fid *fid, int api32)
819 if (BITS_PER_LONG == 32 || api32)
820 RETURN(fid_flatten32(fid));
822 RETURN(fid_flatten(fid));
826 * build inode generation from passed @fid. If our FID overflows the 32-bit
827 * inode number then return a non-zero generation to distinguish them. */
828 __u32 cl_fid_build_gen(const struct lu_fid *fid)
833 if (fid_is_igif(fid)) {
834 gen = lu_igif_gen(fid);
838 gen = (fid_flatten(fid) >> 32);
842 /* lsm is unreliable after hsm implementation as layout can be changed at
843 * any time. This is only to support old, non-clio-ized interfaces. It will
844 * cause deadlock if clio operations are called with this extra layout refcount
845 * because in case the layout changed during the IO, ll_layout_refresh() will
846 * have to wait for the refcount to become zero to destroy the older layout.
848 * Notice that the lsm returned by this function may not be valid unless called
849 * inside layout lock - MDS_INODELOCK_LAYOUT. */
850 struct lov_stripe_md *ccc_inode_lsm_get(struct inode *inode)
852 return lov_lsm_get(ll_i2info(inode)->lli_clob);
855 void inline ccc_inode_lsm_put(struct inode *inode, struct lov_stripe_md *lsm)
857 lov_lsm_put(ll_i2info(inode)->lli_clob, lsm);