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 2008 Sun Microsystems, Inc. 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 * Implementation of cl_object for LOV layer.
38 * Author: Nikita Danilov <nikita.danilov@sun.com>
41 #define DEBUG_SUBSYSTEM S_LOV
43 #include "lov_cl_internal.h"
49 /*****************************************************************************
55 struct lov_layout_operations {
56 int (*llo_init)(const struct lu_env *env, struct lov_device *dev,
57 struct lov_object *lov,
58 const struct cl_object_conf *conf,
59 union lov_layout_state *state);
60 void (*llo_delete)(const struct lu_env *env, struct lov_object *lov,
61 union lov_layout_state *state);
62 void (*llo_fini)(const struct lu_env *env, struct lov_object *lov,
63 union lov_layout_state *state);
64 void (*llo_install)(const struct lu_env *env, struct lov_object *lov,
65 union lov_layout_state *state);
66 int (*llo_print)(const struct lu_env *env, void *cookie,
67 lu_printer_t p, const struct lu_object *o);
68 struct cl_page *(*llo_page_init)(const struct lu_env *env,
69 struct cl_object *obj,
72 int (*llo_lock_init)(const struct lu_env *env,
73 struct cl_object *obj, struct cl_lock *lock,
74 const struct cl_io *io);
75 int (*llo_io_init)(const struct lu_env *env,
76 struct cl_object *obj, struct cl_io *io);
77 int (*llo_getattr)(const struct lu_env *env, struct cl_object *obj,
78 struct cl_attr *attr);
81 /*****************************************************************************
83 * Lov object layout operations.
87 static void lov_install_empty(const struct lu_env *env,
88 struct lov_object *lov,
89 union lov_layout_state *state)
92 * File without objects.
96 static int lov_init_empty(const struct lu_env *env,
97 struct lov_device *dev, struct lov_object *lov,
98 const struct cl_object_conf *conf,
99 union lov_layout_state *state)
104 static void lov_install_raid0(const struct lu_env *env,
105 struct lov_object *lov,
106 union lov_layout_state *state)
111 static void oinfo_get_fid(const struct lov_oinfo *oinfo, struct lu_fid *fid)
113 __u64 idx = oinfo->loi_id;
115 /* See idif definition in wiki:CMD3_interoperability_architecture */
117 LASSERT(oinfo->loi_gr < 1ULL << 16);
118 LASSERT(oinfo->loi_id < 1ULL << 49);
122 * Now that the fid of stripe is not unique now, ost_idx have to
123 * be used to make it unique. This is ok because the stripe fids
124 * are just used in client side(to locate the objects). -jay
126 fid->f_seq = ((__u64)oinfo->loi_ost_idx) << 32 |
127 oinfo->loi_gr << 16 | idx >> 32;
128 fid->f_oid = idx; /* truncated to 32 bits by assignment */
133 static struct cl_object *lov_sub_find(const struct lu_env *env,
134 struct cl_device *dev,
135 const struct lu_fid *fid,
136 const struct cl_object_conf *conf)
141 o = lu_object_find_at(env, cl2lu_dev(dev), fid, &conf->coc_lu);
142 LASSERT(ergo(!IS_ERR(o), o->lo_dev->ld_type == &lovsub_device_type));
146 static int lov_init_sub(const struct lu_env *env, struct lov_object *lov,
147 struct cl_object *stripe,
148 struct lov_layout_raid0 *r0, int idx)
150 struct cl_object_header *hdr;
151 struct cl_object_header *subhdr;
152 struct cl_object_header *parent;
153 struct lov_oinfo *oinfo;
156 hdr = cl_object_header(lov2cl(lov));
157 subhdr = cl_object_header(stripe);
158 parent = subhdr->coh_parent;
160 oinfo = r0->lo_lsm->lsm_oinfo[idx];
161 CDEBUG(D_INODE, DFID"@%p[%d] -> "DFID"@%p: id: "LPU64" gr: "LPU64
162 " idx: %d gen: %d\n",
163 PFID(&subhdr->coh_lu.loh_fid), subhdr, idx,
164 PFID(&hdr->coh_lu.loh_fid), hdr,
165 oinfo->loi_id, oinfo->loi_gr,
166 oinfo->loi_ost_idx, oinfo->loi_ost_gen);
168 if (parent == NULL) {
169 subhdr->coh_parent = hdr;
170 subhdr->coh_nesting = hdr->coh_nesting + 1;
171 lu_object_ref_add(&stripe->co_lu, "lov-parent", lov);
172 r0->lo_sub[idx] = cl2lovsub(stripe);
173 r0->lo_sub[idx]->lso_super = lov;
174 r0->lo_sub[idx]->lso_index = idx;
177 CERROR("Stripe is already owned by other file (%i).\n", idx);
178 LU_OBJECT_DEBUG(D_ERROR, env, &stripe->co_lu, "\n");
179 LU_OBJECT_DEBUG(D_ERROR, env, lu_object_top(&parent->coh_lu),
181 LU_OBJECT_HEADER(D_ERROR, env, lov2lu(lov), "new\n");
182 cl_object_put(env, stripe);
188 static int lov_init_raid0(const struct lu_env *env,
189 struct lov_device *dev, struct lov_object *lov,
190 const struct cl_object_conf *conf,
191 union lov_layout_state *state)
196 struct cl_object *stripe;
197 struct lov_thread_info *lti = lov_env_info(env);
198 struct cl_object_conf *subconf = <i->lti_stripe_conf;
199 struct lov_stripe_md *lsm = conf->u.coc_md->lsm;
200 struct lu_fid *ofid = <i->lti_fid;
201 struct lov_layout_raid0 *r0 = &state->raid0;
204 r0->lo_nr = conf->u.coc_md->lsm->lsm_stripe_count;
205 r0->lo_lsm = conf->u.coc_md->lsm;
206 LASSERT(r0->lo_nr <= lov_targets_nr(dev));
208 OBD_ALLOC(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
209 if (r0->lo_sub != NULL) {
211 subconf->coc_inode = conf->coc_inode;
213 * Create stripe cl_objects.
215 for (i = 0; i < r0->lo_nr && result == 0; ++i) {
216 struct cl_device *subdev;
217 struct lov_oinfo *oinfo = lsm->lsm_oinfo[i];
218 int ost_idx = oinfo->loi_ost_idx;
220 oinfo_get_fid(oinfo, ofid);
221 subdev = lovsub2cl_dev(dev->ld_target[ost_idx]);
222 subconf->u.coc_oinfo = oinfo;
223 LASSERTF(subdev != NULL, "not init ost %d\n", ost_idx);
224 stripe = lov_sub_find(env, subdev, ofid, subconf);
226 result = lov_init_sub(env, lov, stripe, r0, i);
228 result = PTR_ERR(stripe);
235 static void lov_delete_empty(const struct lu_env *env, struct lov_object *lov,
236 union lov_layout_state *state)
238 LASSERT(lov->lo_type == LLT_EMPTY);
241 static void lov_subobject_kill(const struct lu_env *env, struct lov_object *lov,
242 struct lovsub_object *los, int idx)
244 struct cl_object *sub;
245 struct lov_layout_raid0 *r0;
246 struct lu_site *site;
247 cfs_waitlink_t *waiter;
250 LASSERT(r0->lo_sub[idx] == los);
252 sub = lovsub2cl(los);
253 site = sub->co_lu.lo_dev->ld_site;
255 cl_object_kill(env, sub);
256 /* release a reference to the sub-object and ... */
257 lu_object_ref_del(&sub->co_lu, "lov-parent", lov);
258 cl_object_put(env, sub);
260 /* ... wait until it is actually destroyed---sub-object clears its
261 * ->lo_sub[] slot in lovsub_object_fini() */
262 if (r0->lo_sub[idx] == los) {
263 waiter = &lov_env_info(env)->lti_waiter;
264 cfs_waitlink_init(waiter);
265 cfs_waitq_add(&site->ls_marche_funebre, waiter);
266 set_current_state(CFS_TASK_UNINT);
268 while (r0->lo_sub[idx] == los)
269 /* this wait-queue is signaled at the end of
270 * lu_object_free(). */
271 cfs_waitq_wait(waiter, CFS_TASK_UNINT);
272 cfs_waitq_del(&site->ls_marche_funebre, waiter);
274 LASSERT(r0->lo_sub[idx] == NULL);
277 static void lov_delete_raid0(const struct lu_env *env, struct lov_object *lov,
278 union lov_layout_state *state)
280 struct lov_layout_raid0 *r0 = &state->raid0;
284 if (r0->lo_sub != NULL) {
285 for (i = 0; i < r0->lo_nr; ++i) {
286 struct lovsub_object *los = r0->lo_sub[i];
290 * If top-level object is to be evicted from
291 * the cache, so are its sub-objects.
293 lov_subobject_kill(env, lov, los, i);
299 static void lov_fini_empty(const struct lu_env *env, struct lov_object *lov,
300 union lov_layout_state *state)
302 LASSERT(lov->lo_type == LLT_EMPTY);
305 static void lov_fini_raid0(const struct lu_env *env, struct lov_object *lov,
306 union lov_layout_state *state)
308 struct lov_layout_raid0 *r0 = &state->raid0;
311 if (r0->lo_sub != NULL) {
312 OBD_FREE(r0->lo_sub, r0->lo_nr * sizeof r0->lo_sub[0]);
318 static int lov_print_empty(const struct lu_env *env, void *cookie,
319 lu_printer_t p, const struct lu_object *o)
321 (*p)(env, cookie, "empty\n");
325 static int lov_print_raid0(const struct lu_env *env, void *cookie,
326 lu_printer_t p, const struct lu_object *o)
328 struct lov_object *lov = lu2lov(o);
329 struct lov_layout_raid0 *r0 = lov_r0(lov);
332 (*p)(env, cookie, "stripes: %d:\n", r0->lo_nr);
333 for (i = 0; i < r0->lo_nr; ++i) {
334 struct lu_object *sub;
336 if (r0->lo_sub[i] != NULL) {
337 sub = lovsub2lu(r0->lo_sub[i]);
338 lu_object_print(env, cookie, p, sub);
340 (*p)(env, cookie, "sub %d absent\n", i);
346 * Implements cl_object_operations::coo_attr_get() method for an object
347 * without stripes (LLT_EMPTY layout type).
349 * The only attributes this layer is authoritative in this case is
350 * cl_attr::cat_blocks---it's 0.
352 static int lov_attr_get_empty(const struct lu_env *env, struct cl_object *obj,
353 struct cl_attr *attr)
355 attr->cat_blocks = 0;
359 static int lov_attr_get_raid0(const struct lu_env *env, struct cl_object *obj,
360 struct cl_attr *attr)
362 struct lov_object *lov = cl2lov(obj);
363 struct lov_layout_raid0 *r0 = lov_r0(lov);
364 struct lov_stripe_md *lsm = lov->u.raid0.lo_lsm;
365 struct ost_lvb *lvb = &lov_env_info(env)->lti_lvb;
370 if (!r0->lo_attr_valid) {
372 * Fill LVB with attributes already initialized by the upper
375 cl_attr2lvb(lvb, attr);
379 * XXX that should be replaced with a loop over sub-objects,
380 * doing cl_object_attr_get() on them. But for now, let's
381 * reuse old lov code.
385 * XXX take lsm spin-lock to keep lov_merge_lvb_kms()
386 * happy. It's not needed, because new code uses
387 * ->coh_attr_guard spin-lock to protect consistency of
388 * sub-object attributes.
390 lov_stripe_lock(lsm);
391 result = lov_merge_lvb_kms(lsm, lvb, &kms);
392 lov_stripe_unlock(lsm);
394 cl_lvb2attr(attr, lvb);
396 r0->lo_attr_valid = 1;
404 const static struct lov_layout_operations lov_dispatch[] = {
406 .llo_init = lov_init_empty,
407 .llo_delete = lov_delete_empty,
408 .llo_fini = lov_fini_empty,
409 .llo_install = lov_install_empty,
410 .llo_print = lov_print_empty,
411 .llo_page_init = lov_page_init_empty,
412 .llo_lock_init = NULL,
413 .llo_io_init = lov_io_init_empty,
414 .llo_getattr = lov_attr_get_empty
417 .llo_init = lov_init_raid0,
418 .llo_delete = lov_delete_raid0,
419 .llo_fini = lov_fini_raid0,
420 .llo_install = lov_install_raid0,
421 .llo_print = lov_print_raid0,
422 .llo_page_init = lov_page_init_raid0,
423 .llo_lock_init = lov_lock_init_raid0,
424 .llo_io_init = lov_io_init_raid0,
425 .llo_getattr = lov_attr_get_raid0
431 * Performs a double-dispatch based on the layout type of an object.
433 #define LOV_2DISPATCH_NOLOCK(obj, op, ...) \
435 struct lov_object *__obj = (obj); \
436 enum lov_layout_type __llt; \
438 __llt = __obj->lo_type; \
439 LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch)); \
440 lov_dispatch[__llt].op(__VA_ARGS__); \
443 #define LOV_2DISPATCH_MAYLOCK(obj, op, lock, ...) \
445 struct lov_object *__obj = (obj); \
446 int __lock = !!(lock); \
447 typeof(lov_dispatch[0].op(__VA_ARGS__)) __result; \
449 __lock &= __obj->lo_owner != cfs_current(); \
451 down_read(&__obj->lo_type_guard); \
452 __result = LOV_2DISPATCH_NOLOCK(obj, op, __VA_ARGS__); \
454 up_read(&__obj->lo_type_guard); \
459 * Performs a locked double-dispatch based on the layout type of an object.
461 #define LOV_2DISPATCH(obj, op, ...) \
462 LOV_2DISPATCH_MAYLOCK(obj, op, 1, __VA_ARGS__)
464 #define LOV_2DISPATCH_VOID(obj, op, ...) \
466 struct lov_object *__obj = (obj); \
467 enum lov_layout_type __llt; \
469 if (__obj->lo_owner != cfs_current()) \
470 down_read(&__obj->lo_type_guard); \
471 __llt = __obj->lo_type; \
472 LASSERT(0 <= __llt && __llt < ARRAY_SIZE(lov_dispatch)); \
473 lov_dispatch[__llt].op(__VA_ARGS__); \
474 if (__obj->lo_owner != cfs_current()) \
475 up_read(&__obj->lo_type_guard); \
478 static int lov_layout_change(const struct lu_env *env,
479 struct lov_object *obj, enum lov_layout_type llt,
480 const struct cl_object_conf *conf)
483 union lov_layout_state *state = &lov_env_info(env)->lti_state;
484 const struct lov_layout_operations *old_ops;
485 const struct lov_layout_operations *new_ops;
487 LASSERT(0 <= obj->lo_type && obj->lo_type < ARRAY_SIZE(lov_dispatch));
488 LASSERT(0 <= llt && llt < ARRAY_SIZE(lov_dispatch));
491 old_ops = &lov_dispatch[obj->lo_type];
492 new_ops = &lov_dispatch[llt];
494 result = new_ops->llo_init(env, lu2lov_dev(obj->lo_cl.co_lu.lo_dev),
497 struct cl_object_header *hdr = cl_object_header(&obj->lo_cl);
499 struct lu_env *nested;
502 cookie = cl_env_reenter();
503 nested = cl_env_get(&refcheck);
505 cl_object_prune(nested, &obj->lo_cl);
507 result = PTR_ERR(nested);
508 cl_env_put(nested, &refcheck);
509 cl_env_reexit(cookie);
511 old_ops->llo_fini(env, obj, &obj->u);
512 LASSERT(list_empty(&hdr->coh_locks));
513 LASSERT(hdr->coh_tree.rnode == NULL);
514 LASSERT(hdr->coh_pages == 0);
516 new_ops->llo_install(env, obj, state);
519 new_ops->llo_fini(env, obj, state);
523 /*****************************************************************************
525 * Lov object operations.
529 int lov_object_init(const struct lu_env *env, struct lu_object *obj,
530 const struct lu_object_conf *conf)
532 struct lov_device *dev = lu2lov_dev(obj->lo_dev);
533 struct lov_object *lov = lu2lov(obj);
534 const struct cl_object_conf *cconf = lu2cl_conf(conf);
535 union lov_layout_state *set = &lov_env_info(env)->lti_state;
536 const struct lov_layout_operations *ops;
540 init_rwsem(&lov->lo_type_guard);
542 /* no locking is necessary, as object is being created */
543 lov->lo_type = cconf->u.coc_md->lsm != NULL ? LLT_RAID0 : LLT_EMPTY;
544 ops = &lov_dispatch[lov->lo_type];
545 result = ops->llo_init(env, dev, lov, cconf, set);
547 ops->llo_install(env, lov, set);
549 ops->llo_fini(env, lov, set);
553 static int lov_conf_set(const struct lu_env *env, struct cl_object *obj,
554 const struct cl_object_conf *conf)
556 struct lov_object *lov = cl2lov(obj);
561 * Currently only LLT_EMPTY -> LLT_RAID0 transition is supported.
563 LASSERT(lov->lo_owner != cfs_current());
564 down_write(&lov->lo_type_guard);
565 LASSERT(lov->lo_owner == NULL);
566 lov->lo_owner = cfs_current();
567 if (lov->lo_type == LLT_EMPTY && conf->u.coc_md->lsm != NULL)
568 result = lov_layout_change(env, lov, LLT_RAID0, conf);
570 result = -EOPNOTSUPP;
571 lov->lo_owner = NULL;
572 up_write(&lov->lo_type_guard);
576 static void lov_object_delete(const struct lu_env *env, struct lu_object *obj)
578 struct lov_object *lov = lu2lov(obj);
581 LOV_2DISPATCH_VOID(lov, llo_delete, env, lov, &lov->u);
585 static void lov_object_free(const struct lu_env *env, struct lu_object *obj)
587 struct lov_object *lov = lu2lov(obj);
590 LOV_2DISPATCH_VOID(lov, llo_fini, env, lov, &lov->u);
592 OBD_SLAB_FREE_PTR(lov, lov_object_kmem);
596 static int lov_object_print(const struct lu_env *env, void *cookie,
597 lu_printer_t p, const struct lu_object *o)
599 return LOV_2DISPATCH(lu2lov(o), llo_print, env, cookie, p, o);
602 struct cl_page *lov_page_init(const struct lu_env *env, struct cl_object *obj,
603 struct cl_page *page, cfs_page_t *vmpage)
605 return LOV_2DISPATCH(cl2lov(obj),
606 llo_page_init, env, obj, page, vmpage);
610 * Implements cl_object_operations::clo_io_init() method for lov
611 * layer. Dispatches to the appropriate layout io initialization method.
613 int lov_io_init(const struct lu_env *env, struct cl_object *obj,
616 CL_IO_SLICE_CLEAN(lov_env_io(env), lis_cl);
618 * Do not take lock in case of CIT_MISC io, because
620 * - if this is an io for a glimpse, then we don't care;
622 * - if this not a glimpse (writepage or lock cancellation), then
623 * layout change cannot happen because a page or a lock
626 * - lock ordering (lock mutex nests within layout rw-semaphore)
627 * is obeyed in case of lock cancellation.
629 return LOV_2DISPATCH_MAYLOCK(cl2lov(obj), llo_io_init,
630 io->ci_type != CIT_MISC, env, obj, io);
634 * An implementation of cl_object_operations::clo_attr_get() method for lov
635 * layer. For raid0 layout this collects and merges attributes of all
638 static int lov_attr_get(const struct lu_env *env, struct cl_object *obj,
639 struct cl_attr *attr)
641 /* do not take lock, as this function is called under a
642 * spin-lock. Layout is protected from changing by ongoing IO. */
643 return LOV_2DISPATCH_NOLOCK(cl2lov(obj), llo_getattr, env, obj, attr);
646 static int lov_attr_set(const struct lu_env *env, struct cl_object *obj,
647 const struct cl_attr *attr, unsigned valid)
650 * No dispatch is required here, as no layout implements this.
655 int lov_lock_init(const struct lu_env *env, struct cl_object *obj,
656 struct cl_lock *lock, const struct cl_io *io)
658 return LOV_2DISPATCH(cl2lov(obj), llo_lock_init, env, obj, lock, io);
661 static const struct cl_object_operations lov_ops = {
662 .coo_page_init = lov_page_init,
663 .coo_lock_init = lov_lock_init,
664 .coo_io_init = lov_io_init,
665 .coo_attr_get = lov_attr_get,
666 .coo_attr_set = lov_attr_set,
667 .coo_conf_set = lov_conf_set
670 static const struct lu_object_operations lov_lu_obj_ops = {
671 .loo_object_init = lov_object_init,
672 .loo_object_delete = lov_object_delete,
673 .loo_object_release = NULL,
674 .loo_object_free = lov_object_free,
675 .loo_object_print = lov_object_print,
676 .loo_object_invariant = NULL
679 struct lu_object *lov_object_alloc(const struct lu_env *env,
680 const struct lu_object_header *unused,
681 struct lu_device *dev)
683 struct lov_object *lov;
684 struct lu_object *obj;
687 OBD_SLAB_ALLOC_PTR_GFP(lov, lov_object_kmem, CFS_ALLOC_IO);
690 lu_object_init(obj, NULL, dev);
691 lov->lo_cl.co_ops = &lov_ops;
692 lov->lo_type = -1; /* invalid, to catch uninitialized type */
694 * object io operation vector (cl_object::co_iop) is installed
695 * later in lov_object_init(), as different vectors are used
696 * for object with different layouts.
698 obj->lo_ops = &lov_lu_obj_ops;