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_thread_kmem;
70 static struct kmem_cache *ccc_req_kmem;
72 static struct lu_kmem_descr ccc_caches[] = {
74 .ckd_cache = &ccc_thread_kmem,
75 .ckd_name = "ccc_thread_kmem",
76 .ckd_size = sizeof (struct ccc_thread_info),
79 .ckd_cache = &ccc_req_kmem,
80 .ckd_name = "ccc_req_kmem",
81 .ckd_size = sizeof (struct ccc_req)
88 /*****************************************************************************
90 * Vvp device and device type functions.
94 void *ccc_key_init(const struct lu_context *ctx, struct lu_context_key *key)
96 struct ccc_thread_info *info;
98 OBD_SLAB_ALLOC_PTR_GFP(info, ccc_thread_kmem, GFP_NOFS);
100 info = ERR_PTR(-ENOMEM);
104 void ccc_key_fini(const struct lu_context *ctx,
105 struct lu_context_key *key, void *data)
107 struct ccc_thread_info *info = data;
108 OBD_SLAB_FREE_PTR(info, ccc_thread_kmem);
111 struct lu_context_key ccc_key = {
112 .lct_tags = LCT_CL_THREAD,
113 .lct_init = ccc_key_init,
114 .lct_fini = ccc_key_fini
117 int ccc_req_init(const struct lu_env *env, struct cl_device *dev,
123 OBD_SLAB_ALLOC_PTR_GFP(vrq, ccc_req_kmem, GFP_NOFS);
125 cl_req_slice_add(req, &vrq->crq_cl, dev, &ccc_req_ops);
133 * An `emergency' environment used by ccc_inode_fini() when cl_env_get()
134 * fails. Access to this environment is serialized by ccc_inode_fini_guard
137 static struct lu_env *ccc_inode_fini_env = NULL;
140 * A mutex serializing calls to slp_inode_fini() under extreme memory
141 * pressure, when environments cannot be allocated.
143 static DEFINE_MUTEX(ccc_inode_fini_guard);
144 static int dummy_refcheck;
146 int ccc_global_init(struct lu_device_type *device_type)
150 result = lu_kmem_init(ccc_caches);
154 result = lu_device_type_init(device_type);
158 ccc_inode_fini_env = cl_env_alloc(&dummy_refcheck,
159 LCT_REMEMBER|LCT_NOREF);
160 if (IS_ERR(ccc_inode_fini_env)) {
161 result = PTR_ERR(ccc_inode_fini_env);
165 ccc_inode_fini_env->le_ctx.lc_cookie = 0x4;
168 lu_device_type_fini(device_type);
170 lu_kmem_fini(ccc_caches);
174 void ccc_global_fini(struct lu_device_type *device_type)
176 if (ccc_inode_fini_env != NULL) {
177 cl_env_put(ccc_inode_fini_env, &dummy_refcheck);
178 ccc_inode_fini_env = NULL;
180 lu_device_type_fini(device_type);
181 lu_kmem_fini(ccc_caches);
184 static void vvp_object_size_lock(struct cl_object *obj)
186 struct inode *inode = vvp_object_inode(obj);
188 ll_inode_size_lock(inode);
189 cl_object_attr_lock(obj);
192 static void vvp_object_size_unlock(struct cl_object *obj)
194 struct inode *inode = vvp_object_inode(obj);
196 cl_object_attr_unlock(obj);
197 ll_inode_size_unlock(inode);
200 /*****************************************************************************
206 int vvp_io_one_lock_index(const struct lu_env *env, struct cl_io *io,
207 __u32 enqflags, enum cl_lock_mode mode,
208 pgoff_t start, pgoff_t end)
210 struct vvp_io *vio = vvp_env_io(env);
211 struct cl_lock_descr *descr = &vio->vui_link.cill_descr;
212 struct cl_object *obj = io->ci_obj;
214 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
217 CDEBUG(D_VFSTRACE, "lock: %d [%lu, %lu]\n", mode, start, end);
219 memset(&vio->vui_link, 0, sizeof vio->vui_link);
221 if (vio->vui_fd && (vio->vui_fd->fd_flags & LL_FILE_GROUP_LOCKED)) {
222 descr->cld_mode = CLM_GROUP;
223 descr->cld_gid = vio->vui_fd->fd_grouplock.cg_gid;
225 descr->cld_mode = mode;
228 descr->cld_obj = obj;
229 descr->cld_start = start;
230 descr->cld_end = end;
231 descr->cld_enq_flags = enqflags;
233 cl_io_lock_add(env, io, &vio->vui_link);
238 void vvp_io_update_iov(const struct lu_env *env,
239 struct vvp_io *vio, struct cl_io *io)
242 size_t size = io->u.ci_rw.crw_count;
244 vio->vui_iov_olen = 0;
245 if (!cl_is_normalio(env, io) || vio->vui_tot_nrsegs == 0)
248 for (i = 0; i < vio->vui_tot_nrsegs; i++) {
249 struct iovec *iv = &vio->vui_iov[i];
251 if (iv->iov_len < size) {
254 if (iv->iov_len > size) {
255 vio->vui_iov_olen = iv->iov_len;
262 vio->vui_nrsegs = i + 1;
263 LASSERTF(vio->vui_tot_nrsegs >= vio->vui_nrsegs,
264 "tot_nrsegs: %lu, nrsegs: %lu\n",
265 vio->vui_tot_nrsegs, vio->vui_nrsegs);
268 int vvp_io_one_lock(const struct lu_env *env, struct cl_io *io,
269 __u32 enqflags, enum cl_lock_mode mode,
270 loff_t start, loff_t end)
272 struct cl_object *obj = io->ci_obj;
274 return vvp_io_one_lock_index(env, io, enqflags, mode,
275 cl_index(obj, start), cl_index(obj, end));
278 void vvp_io_end(const struct lu_env *env, const struct cl_io_slice *ios)
280 CLOBINVRNT(env, ios->cis_io->ci_obj,
281 vvp_object_invariant(ios->cis_io->ci_obj));
284 void vvp_io_advance(const struct lu_env *env,
285 const struct cl_io_slice *ios,
288 struct vvp_io *vio = cl2vvp_io(env, ios);
289 struct cl_io *io = ios->cis_io;
290 struct cl_object *obj = ios->cis_io->ci_obj;
292 CLOBINVRNT(env, obj, vvp_object_invariant(obj));
294 if (!cl_is_normalio(env, io))
297 LASSERT(vio->vui_tot_nrsegs >= vio->vui_nrsegs);
298 LASSERT(vio->vui_tot_count >= nob);
300 vio->vui_iov += vio->vui_nrsegs;
301 vio->vui_tot_nrsegs -= vio->vui_nrsegs;
302 vio->vui_tot_count -= nob;
305 if (vio->vui_iov_olen > 0) {
309 vio->vui_tot_nrsegs++;
310 iv = &vio->vui_iov[0];
311 if (io->ci_continue) {
312 iv->iov_base += iv->iov_len;
313 LASSERT(vio->vui_iov_olen > iv->iov_len);
314 iv->iov_len = vio->vui_iov_olen - iv->iov_len;
316 /* restore the iov_len, in case of restart io. */
317 iv->iov_len = vio->vui_iov_olen;
319 vio->vui_iov_olen = 0;
324 * Helper function that if necessary adjusts file size (inode->i_size), when
325 * position at the offset \a pos is accessed. File size can be arbitrary stale
326 * on a Lustre client, but client at least knows KMS. If accessed area is
327 * inside [0, KMS], set file size to KMS, otherwise glimpse file size.
329 * Locking: cl_isize_lock is used to serialize changes to inode size and to
330 * protect consistency between inode size and cl_object
331 * attributes. cl_object_size_lock() protects consistency between cl_attr's of
332 * top-object and sub-objects.
334 int ccc_prep_size(const struct lu_env *env, struct cl_object *obj,
335 struct cl_io *io, loff_t start, size_t count, int *exceed)
337 struct cl_attr *attr = ccc_env_thread_attr(env);
338 struct inode *inode = vvp_object_inode(obj);
339 loff_t pos = start + count - 1;
344 * Consistency guarantees: following possibilities exist for the
345 * relation between region being accessed and real file size at this
348 * (A): the region is completely inside of the file;
350 * (B-x): x bytes of region are inside of the file, the rest is
353 * (C): the region is completely outside of the file.
355 * This classification is stable under DLM lock already acquired by
356 * the caller, because to change the class, other client has to take
357 * DLM lock conflicting with our lock. Also, any updates to ->i_size
358 * by other threads on this client are serialized by
359 * ll_inode_size_lock(). This guarantees that short reads are handled
360 * correctly in the face of concurrent writes and truncates.
362 vvp_object_size_lock(obj);
363 result = cl_object_attr_get(env, obj, attr);
368 * A glimpse is necessary to determine whether we
369 * return a short read (B) or some zeroes at the end
372 vvp_object_size_unlock(obj);
373 result = cl_glimpse_lock(env, io, inode, obj, 0);
374 if (result == 0 && exceed != NULL) {
375 /* If objective page index exceed end-of-file
376 * page index, return directly. Do not expect
377 * kernel will check such case correctly.
378 * linux-2.6.18-128.1.1 miss to do that.
380 loff_t size = i_size_read(inode);
381 unsigned long cur_index = start >>
384 if ((size == 0 && cur_index != 0) ||
385 (((size - 1) >> PAGE_CACHE_SHIFT) <
392 * region is within kms and, hence, within real file
393 * size (A). We need to increase i_size to cover the
394 * read region so that generic_file_read() will do its
395 * job, but that doesn't mean the kms size is
396 * _correct_, it is only the _minimum_ size. If
397 * someone does a stat they will get the correct size
398 * which will always be >= the kms value here.
401 if (i_size_read(inode) < kms) {
402 i_size_write(inode, kms);
404 DFID" updating i_size "LPU64"\n",
405 PFID(lu_object_fid(&obj->co_lu)),
406 (__u64)i_size_read(inode));
412 vvp_object_size_unlock(obj);
417 /*****************************************************************************
419 * Transfer operations.
423 void ccc_req_completion(const struct lu_env *env,
424 const struct cl_req_slice *slice, int ioret)
429 cl_stats_tally(slice->crs_dev, slice->crs_req->crq_type, ioret);
431 vrq = cl2ccc_req(slice);
432 OBD_SLAB_FREE_PTR(vrq, ccc_req_kmem);
436 * Implementation of struct cl_req_operations::cro_attr_set() for ccc
437 * layer. ccc is responsible for
455 void ccc_req_attr_set(const struct lu_env *env,
456 const struct cl_req_slice *slice,
457 const struct cl_object *obj,
458 struct cl_req_attr *attr, u64 flags)
465 inode = vvp_object_inode(obj);
466 valid_flags = OBD_MD_FLTYPE;
468 if ((flags & OBD_MD_FLOSSCAPA) != 0) {
469 LASSERT(attr->cra_capa == NULL);
470 attr->cra_capa = cl_capa_lookup(inode,
471 slice->crs_req->crq_type);
474 if (slice->crs_req->crq_type == CRT_WRITE) {
475 if (flags & OBD_MD_FLEPOCH) {
476 oa->o_valid |= OBD_MD_FLEPOCH;
477 oa->o_ioepoch = ll_i2info(inode)->lli_ioepoch;
478 valid_flags |= OBD_MD_FLMTIME | OBD_MD_FLCTIME |
479 OBD_MD_FLUID | OBD_MD_FLGID;
482 obdo_from_inode(oa, inode, valid_flags & flags);
483 obdo_set_parent_fid(oa, &ll_i2info(inode)->lli_fid);
484 if (OBD_FAIL_CHECK(OBD_FAIL_LFSCK_INVALID_PFID))
486 memcpy(attr->cra_jobid, ll_i2info(inode)->lli_jobid,
490 static const struct cl_req_operations ccc_req_ops = {
491 .cro_attr_set = ccc_req_attr_set,
492 .cro_completion = ccc_req_completion
495 int cl_setattr_ost(struct inode *inode, const struct iattr *attr,
496 struct obd_capa *capa)
505 env = cl_env_get(&refcheck);
507 RETURN(PTR_ERR(env));
509 io = ccc_env_thread_io(env);
510 io->ci_obj = ll_i2info(inode)->lli_clob;
512 io->u.ci_setattr.sa_attr.lvb_atime = LTIME_S(attr->ia_atime);
513 io->u.ci_setattr.sa_attr.lvb_mtime = LTIME_S(attr->ia_mtime);
514 io->u.ci_setattr.sa_attr.lvb_ctime = LTIME_S(attr->ia_ctime);
515 io->u.ci_setattr.sa_attr.lvb_size = attr->ia_size;
516 io->u.ci_setattr.sa_valid = attr->ia_valid;
517 io->u.ci_setattr.sa_parent_fid = ll_inode2fid(inode);
518 io->u.ci_setattr.sa_capa = capa;
521 if (cl_io_init(env, io, CIT_SETATTR, io->ci_obj) == 0) {
522 struct vvp_io *vio = vvp_env_io(env);
524 if (attr->ia_valid & ATTR_FILE)
525 /* populate the file descriptor for ftruncate to honor
526 * group lock - see LU-787 */
527 vio->vui_fd = LUSTRE_FPRIVATE(attr->ia_file);
529 result = cl_io_loop(env, io);
531 result = io->ci_result;
534 if (unlikely(io->ci_need_restart))
536 /* HSM import case: file is released, cannot be restored
537 * no need to fail except if restore registration failed
539 if (result == -ENODATA && io->ci_restore_needed &&
540 io->ci_result != -ENODATA)
542 cl_env_put(env, &refcheck);
546 /*****************************************************************************
552 struct vvp_io *cl2vvp_io(const struct lu_env *env,
553 const struct cl_io_slice *slice)
557 vio = container_of(slice, struct vvp_io, vui_cl);
558 LASSERT(vio == vvp_env_io(env));
563 struct ccc_req *cl2ccc_req(const struct cl_req_slice *slice)
565 return container_of0(slice, struct ccc_req, crq_cl);
569 * Initialize or update CLIO structures for regular files when new
570 * meta-data arrives from the server.
572 * \param inode regular file inode
573 * \param md new file metadata from MDS
574 * - allocates cl_object if necessary,
575 * - updated layout, if object was already here.
577 int cl_file_inode_init(struct inode *inode, struct lustre_md *md)
580 struct ll_inode_info *lli;
581 struct cl_object *clob;
582 struct lu_site *site;
584 struct cl_object_conf conf = {
593 LASSERT(md->body->mbo_valid & OBD_MD_FLID);
594 LASSERT(S_ISREG(inode->i_mode));
596 env = cl_env_get(&refcheck);
600 site = ll_i2sbi(inode)->ll_site;
601 lli = ll_i2info(inode);
603 LASSERT(fid_is_sane(fid));
605 if (lli->lli_clob == NULL) {
606 /* clob is slave of inode, empty lli_clob means for new inode,
607 * there is no clob in cache with the given fid, so it is
608 * unnecessary to perform lookup-alloc-lookup-insert, just
609 * alloc and insert directly. */
610 LASSERT(inode->i_state & I_NEW);
611 conf.coc_lu.loc_flags = LOC_F_NEW;
612 clob = cl_object_find(env, lu2cl_dev(site->ls_top_dev),
616 * No locking is necessary, as new inode is
617 * locked by I_NEW bit.
619 lli->lli_clob = clob;
620 lli->lli_has_smd = lsm_has_objects(md->lsm);
621 lu_object_ref_add(&clob->co_lu, "inode", inode);
623 result = PTR_ERR(clob);
625 result = cl_conf_set(env, lli->lli_clob, &conf);
628 cl_env_put(env, &refcheck);
631 CERROR("Failure to initialize cl object "DFID": %d\n",
637 * Wait for others drop their references of the object at first, then we drop
638 * the last one, which will lead to the object be destroyed immediately.
639 * Must be called after cl_object_kill() against this object.
641 * The reason we want to do this is: destroying top object will wait for sub
642 * objects being destroyed first, so we can't let bottom layer (e.g. from ASTs)
643 * to initiate top object destroying which may deadlock. See bz22520.
645 static void cl_object_put_last(struct lu_env *env, struct cl_object *obj)
647 struct lu_object_header *header = obj->co_lu.lo_header;
650 if (unlikely(atomic_read(&header->loh_ref) != 1)) {
651 struct lu_site *site = obj->co_lu.lo_dev->ld_site;
652 struct lu_site_bkt_data *bkt;
654 bkt = lu_site_bkt_from_fid(site, &header->loh_fid);
656 init_waitqueue_entry_current(&waiter);
657 add_wait_queue(&bkt->lsb_marche_funebre, &waiter);
660 set_current_state(TASK_UNINTERRUPTIBLE);
661 if (atomic_read(&header->loh_ref) == 1)
663 waitq_wait(&waiter, TASK_UNINTERRUPTIBLE);
666 set_current_state(TASK_RUNNING);
667 remove_wait_queue(&bkt->lsb_marche_funebre, &waiter);
670 cl_object_put(env, obj);
673 void cl_inode_fini(struct inode *inode)
676 struct ll_inode_info *lli = ll_i2info(inode);
677 struct cl_object *clob = lli->lli_clob;
684 cookie = cl_env_reenter();
685 env = cl_env_get(&refcheck);
686 emergency = IS_ERR(env);
688 mutex_lock(&ccc_inode_fini_guard);
689 LASSERT(ccc_inode_fini_env != NULL);
690 cl_env_implant(ccc_inode_fini_env, &refcheck);
691 env = ccc_inode_fini_env;
694 * cl_object cache is a slave to inode cache (which, in turn
695 * is a slave to dentry cache), don't keep cl_object in memory
696 * when its master is evicted.
698 cl_object_kill(env, clob);
699 lu_object_ref_del(&clob->co_lu, "inode", inode);
700 cl_object_put_last(env, clob);
701 lli->lli_clob = NULL;
703 cl_env_unplant(ccc_inode_fini_env, &refcheck);
704 mutex_unlock(&ccc_inode_fini_guard);
706 cl_env_put(env, &refcheck);
707 cl_env_reexit(cookie);
712 * return IF_* type for given lu_dirent entry.
713 * IF_* flag shld be converted to particular OS file type in
714 * platform llite module.
716 __u16 ll_dirent_type_get(struct lu_dirent *ent)
719 struct luda_type *lt;
722 if (le32_to_cpu(ent->lde_attrs) & LUDA_TYPE) {
723 const unsigned align = sizeof(struct luda_type) - 1;
725 len = le16_to_cpu(ent->lde_namelen);
726 len = (len + align) & ~align;
727 lt = (void *)ent->lde_name + len;
728 type = IFTODT(le16_to_cpu(lt->lt_type));
734 * build inode number from passed @fid */
735 __u64 cl_fid_build_ino(const struct lu_fid *fid, int api32)
737 if (BITS_PER_LONG == 32 || api32)
738 RETURN(fid_flatten32(fid));
740 RETURN(fid_flatten(fid));
744 * build inode generation from passed @fid. If our FID overflows the 32-bit
745 * inode number then return a non-zero generation to distinguish them. */
746 __u32 cl_fid_build_gen(const struct lu_fid *fid)
751 if (fid_is_igif(fid)) {
752 gen = lu_igif_gen(fid);
756 gen = (fid_flatten(fid) >> 32);
760 /* lsm is unreliable after hsm implementation as layout can be changed at
761 * any time. This is only to support old, non-clio-ized interfaces. It will
762 * cause deadlock if clio operations are called with this extra layout refcount
763 * because in case the layout changed during the IO, ll_layout_refresh() will
764 * have to wait for the refcount to become zero to destroy the older layout.
766 * Notice that the lsm returned by this function may not be valid unless called
767 * inside layout lock - MDS_INODELOCK_LAYOUT. */
768 struct lov_stripe_md *ccc_inode_lsm_get(struct inode *inode)
770 return lov_lsm_get(ll_i2info(inode)->lli_clob);
773 void inline ccc_inode_lsm_put(struct inode *inode, struct lov_stripe_md *lsm)
775 lov_lsm_put(ll_i2info(inode)->lli_clob, lsm);