Whamcloud - gitweb
b3614f8d562a35e4777658e316d5b393c60f5215
[fs/lustre-release.git] / lustre / osc / osc_object.c
1 /*
2  * GPL HEADER START
3  *
4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5  *
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.
9  *
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).
15  *
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
19  *
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
22  * have any questions.
23  *
24  * GPL HEADER END
25  */
26 /*
27  * Copyright (c) 2008, 2010, Oracle and/or its affiliates. All rights reserved.
28  * Use is subject to license terms.
29  *
30  * Copyright (c) 2011, 2014, Intel Corporation.
31  */
32 /*
33  * This file is part of Lustre, http://www.lustre.org/
34  * Lustre is a trademark of Sun Microsystems, Inc.
35  *
36  * Implementation of cl_object for OSC layer.
37  *
38  *   Author: Nikita Danilov <nikita.danilov@sun.com>
39  *   Author: Jinshan Xiong <jinshan.xiong@intel.com>
40  */
41
42 #define DEBUG_SUBSYSTEM S_OSC
43
44 #include "osc_cl_internal.h"
45
46 /** \addtogroup osc
47  *  @{
48  */
49
50 /*****************************************************************************
51  *
52  * Type conversions.
53  *
54  */
55
56 static struct lu_object *osc2lu(struct osc_object *osc)
57 {
58         return &osc->oo_cl.co_lu;
59 }
60
61 static struct osc_object *lu2osc(const struct lu_object *obj)
62 {
63         LINVRNT(osc_is_object(obj));
64         return container_of0(obj, struct osc_object, oo_cl.co_lu);
65 }
66
67 /*****************************************************************************
68  *
69  * Object operations.
70  *
71  */
72
73 static int osc_object_init(const struct lu_env *env, struct lu_object *obj,
74                            const struct lu_object_conf *conf)
75 {
76         struct osc_object           *osc   = lu2osc(obj);
77         const struct cl_object_conf *cconf = lu2cl_conf(conf);
78         int i;
79
80         osc->oo_oinfo = cconf->u.coc_oinfo;
81 #ifdef CONFIG_LUSTRE_DEBUG_EXPENSIVE_CHECK
82         mutex_init(&osc->oo_debug_mutex);
83 #endif
84         spin_lock_init(&osc->oo_seatbelt);
85         for (i = 0; i < CRT_NR; ++i)
86                 INIT_LIST_HEAD(&osc->oo_inflight[i]);
87
88         INIT_LIST_HEAD(&osc->oo_ready_item);
89         INIT_LIST_HEAD(&osc->oo_hp_ready_item);
90         INIT_LIST_HEAD(&osc->oo_write_item);
91         INIT_LIST_HEAD(&osc->oo_read_item);
92
93         osc->oo_root.rb_node = NULL;
94         INIT_LIST_HEAD(&osc->oo_hp_exts);
95         INIT_LIST_HEAD(&osc->oo_urgent_exts);
96         INIT_LIST_HEAD(&osc->oo_rpc_exts);
97         INIT_LIST_HEAD(&osc->oo_reading_exts);
98         atomic_set(&osc->oo_nr_reads, 0);
99         atomic_set(&osc->oo_nr_writes, 0);
100         spin_lock_init(&osc->oo_lock);
101         spin_lock_init(&osc->oo_tree_lock);
102         spin_lock_init(&osc->oo_ol_spin);
103         INIT_LIST_HEAD(&osc->oo_ol_list);
104
105         cl_object_page_init(lu2cl(obj), sizeof(struct osc_page));
106
107         return 0;
108 }
109
110 static void osc_object_free(const struct lu_env *env, struct lu_object *obj)
111 {
112         struct osc_object *osc = lu2osc(obj);
113         int i;
114
115         for (i = 0; i < CRT_NR; ++i)
116                 LASSERT(list_empty(&osc->oo_inflight[i]));
117
118         LASSERT(list_empty(&osc->oo_ready_item));
119         LASSERT(list_empty(&osc->oo_hp_ready_item));
120         LASSERT(list_empty(&osc->oo_write_item));
121         LASSERT(list_empty(&osc->oo_read_item));
122
123         LASSERT(osc->oo_root.rb_node == NULL);
124         LASSERT(list_empty(&osc->oo_hp_exts));
125         LASSERT(list_empty(&osc->oo_urgent_exts));
126         LASSERT(list_empty(&osc->oo_rpc_exts));
127         LASSERT(list_empty(&osc->oo_reading_exts));
128         LASSERT(atomic_read(&osc->oo_nr_reads) == 0);
129         LASSERT(atomic_read(&osc->oo_nr_writes) == 0);
130         LASSERT(list_empty(&osc->oo_ol_list));
131
132         lu_object_fini(obj);
133         OBD_SLAB_FREE_PTR(osc, osc_object_kmem);
134 }
135
136 int osc_lvb_print(const struct lu_env *env, void *cookie,
137                   lu_printer_t p, const struct ost_lvb *lvb)
138 {
139         return (*p)(env, cookie, "size: "LPU64" mtime: "LPU64" atime: "LPU64" "
140                     "ctime: "LPU64" blocks: "LPU64,
141                     lvb->lvb_size, lvb->lvb_mtime, lvb->lvb_atime,
142                     lvb->lvb_ctime, lvb->lvb_blocks);
143 }
144
145 static int osc_object_print(const struct lu_env *env, void *cookie,
146                             lu_printer_t p, const struct lu_object *obj)
147 {
148         struct osc_object   *osc   = lu2osc(obj);
149         struct lov_oinfo    *oinfo = osc->oo_oinfo;
150         struct osc_async_rc *ar    = &oinfo->loi_ar;
151
152         (*p)(env, cookie, "id: "DOSTID" "
153              "idx: %d gen: %d kms_valid: %u kms "LPU64" "
154              "rc: %d force_sync: %d min_xid: "LPU64" ",
155              POSTID(&oinfo->loi_oi), oinfo->loi_ost_idx,
156              oinfo->loi_ost_gen, oinfo->loi_kms_valid, oinfo->loi_kms,
157              ar->ar_rc, ar->ar_force_sync, ar->ar_min_xid);
158         osc_lvb_print(env, cookie, p, &oinfo->loi_lvb);
159         return 0;
160 }
161
162
163 static int osc_attr_get(const struct lu_env *env, struct cl_object *obj,
164                         struct cl_attr *attr)
165 {
166         struct lov_oinfo *oinfo = cl2osc(obj)->oo_oinfo;
167
168         cl_lvb2attr(attr, &oinfo->loi_lvb);
169         attr->cat_kms = oinfo->loi_kms_valid ? oinfo->loi_kms : 0;
170         return 0;
171 }
172
173 static int osc_attr_update(const struct lu_env *env, struct cl_object *obj,
174                            const struct cl_attr *attr, unsigned valid)
175 {
176         struct lov_oinfo *oinfo = cl2osc(obj)->oo_oinfo;
177         struct ost_lvb   *lvb   = &oinfo->loi_lvb;
178
179         if (valid & CAT_SIZE)
180                 lvb->lvb_size = attr->cat_size;
181         if (valid & CAT_MTIME)
182                 lvb->lvb_mtime = attr->cat_mtime;
183         if (valid & CAT_ATIME)
184                 lvb->lvb_atime = attr->cat_atime;
185         if (valid & CAT_CTIME)
186                 lvb->lvb_ctime = attr->cat_ctime;
187         if (valid & CAT_BLOCKS)
188                 lvb->lvb_blocks = attr->cat_blocks;
189         if (valid & CAT_KMS) {
190                 CDEBUG(D_CACHE, "set kms from "LPU64"to "LPU64"\n",
191                        oinfo->loi_kms, (__u64)attr->cat_kms);
192                 loi_kms_set(oinfo, attr->cat_kms);
193         }
194         return 0;
195 }
196
197 static int osc_object_glimpse(const struct lu_env *env,
198                               const struct cl_object *obj, struct ost_lvb *lvb)
199 {
200         struct lov_oinfo *oinfo = cl2osc(obj)->oo_oinfo;
201
202         ENTRY;
203         lvb->lvb_size   = oinfo->loi_kms;
204         lvb->lvb_blocks = oinfo->loi_lvb.lvb_blocks;
205         RETURN(0);
206 }
207
208 static int osc_object_ast_clear(struct ldlm_lock *lock, void *data)
209 {
210         ENTRY;
211
212         if (lock->l_ast_data == data)
213                 lock->l_ast_data = NULL;
214         RETURN(LDLM_ITER_CONTINUE);
215 }
216
217 static int osc_object_prune(const struct lu_env *env, struct cl_object *obj)
218 {
219         struct osc_object       *osc = cl2osc(obj);
220         struct ldlm_res_id      *resname = &osc_env_info(env)->oti_resname;
221
222         LASSERTF(osc->oo_npages == 0,
223                  DFID "still have %lu pages, obj: %p, osc: %p\n",
224                  PFID(lu_object_fid(&obj->co_lu)), osc->oo_npages, obj, osc);
225
226         /* DLM locks don't hold a reference of osc_object so we have to
227          * clear it before the object is being destroyed. */
228         ostid_build_res_name(&osc->oo_oinfo->loi_oi, resname);
229         ldlm_resource_iterate(osc_export(osc)->exp_obd->obd_namespace, resname,
230                               osc_object_ast_clear, osc);
231         return 0;
232 }
233 /**
234  * Find any ldlm lock covers the osc object.
235  * \retval 0    not found
236  * \retval 1    find one
237  * \retval < 0  error
238  */
239 static int osc_object_find_cbdata(const struct lu_env *env,
240                                   struct cl_object *obj, ldlm_iterator_t iter,
241                                   void *data)
242 {
243         struct ldlm_res_id              res_id;
244         struct obd_device               *obd;
245         int                             rc = 0;
246
247         ostid_build_res_name(&cl2osc(obj)->oo_oinfo->loi_oi, &res_id);
248         obd = obj->co_lu.lo_dev->ld_obd;
249         rc = ldlm_resource_iterate(obd->obd_namespace, &res_id, iter, data);
250         if (rc == LDLM_ITER_STOP)
251                 return 1;
252
253         if (rc == LDLM_ITER_CONTINUE)
254                 return 0;
255
256         return rc;
257 }
258
259 static int osc_object_fiemap(const struct lu_env *env, struct cl_object *obj,
260                              struct ll_fiemap_info_key *fmkey,
261                              struct fiemap *fiemap, size_t *buflen)
262 {
263         struct obd_export               *exp = osc_export(cl2osc(obj));
264         struct ldlm_res_id              resid;
265         ldlm_policy_data_t              policy;
266         struct lustre_handle            lockh;
267         ldlm_mode_t                     mode = 0;
268         struct ptlrpc_request           *req;
269         struct fiemap                   *reply;
270         char                            *tmp;
271         int                             rc;
272         ENTRY;
273
274         fmkey->oa.o_oi = cl2osc(obj)->oo_oinfo->loi_oi;
275         if (!(fmkey->fiemap.fm_flags & FIEMAP_FLAG_SYNC))
276                 goto skip_locking;
277
278         policy.l_extent.start = fmkey->fiemap.fm_start & PAGE_CACHE_MASK;
279
280         if (OBD_OBJECT_EOF - fmkey->fiemap.fm_length <=
281             fmkey->fiemap.fm_start + PAGE_CACHE_SIZE - 1)
282                 policy.l_extent.end = OBD_OBJECT_EOF;
283         else
284                 policy.l_extent.end = (fmkey->fiemap.fm_start +
285                                        fmkey->fiemap.fm_length +
286                                        PAGE_CACHE_SIZE - 1) & PAGE_CACHE_MASK;
287
288         ostid_build_res_name(&fmkey->oa.o_oi, &resid);
289         mode = ldlm_lock_match(exp->exp_obd->obd_namespace,
290                                LDLM_FL_BLOCK_GRANTED | LDLM_FL_LVB_READY,
291                                &resid, LDLM_EXTENT, &policy,
292                                LCK_PR | LCK_PW, &lockh, 0);
293         if (mode) { /* lock is cached on client */
294                 if (mode != LCK_PR) {
295                         ldlm_lock_addref(&lockh, LCK_PR);
296                         ldlm_lock_decref(&lockh, LCK_PW);
297                 }
298         } else { /* no cached lock, needs acquire lock on server side */
299                 fmkey->oa.o_valid |= OBD_MD_FLFLAGS;
300                 fmkey->oa.o_flags |= OBD_FL_SRVLOCK;
301         }
302
303 skip_locking:
304         req = ptlrpc_request_alloc(class_exp2cliimp(exp),
305                                    &RQF_OST_GET_INFO_FIEMAP);
306         if (req == NULL)
307                 GOTO(drop_lock, rc = -ENOMEM);
308
309         req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_KEY, RCL_CLIENT,
310                              sizeof(*fmkey));
311         req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL, RCL_CLIENT,
312                              *buflen);
313         req_capsule_set_size(&req->rq_pill, &RMF_FIEMAP_VAL, RCL_SERVER,
314                              *buflen);
315
316         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, OST_GET_INFO);
317         if (rc != 0) {
318                 ptlrpc_request_free(req);
319                 GOTO(drop_lock, rc);
320         }
321         tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_KEY);
322         memcpy(tmp, fmkey, sizeof(*fmkey));
323         tmp = req_capsule_client_get(&req->rq_pill, &RMF_FIEMAP_VAL);
324         memcpy(tmp, fiemap, *buflen);
325         ptlrpc_request_set_replen(req);
326
327         rc = ptlrpc_queue_wait(req);
328         if (rc != 0)
329                 GOTO(fini_req, rc);
330
331         reply = req_capsule_server_get(&req->rq_pill, &RMF_FIEMAP_VAL);
332         if (reply == NULL)
333                 GOTO(fini_req, rc = -EPROTO);
334
335         memcpy(fiemap, reply, *buflen);
336 fini_req:
337         ptlrpc_req_finished(req);
338 drop_lock:
339         if (mode)
340                 ldlm_lock_decref(&lockh, LCK_PR);
341         RETURN(rc);
342 }
343
344 void osc_object_set_contended(struct osc_object *obj)
345 {
346         obj->oo_contention_time = cfs_time_current();
347         /* mb(); */
348         obj->oo_contended = 1;
349 }
350
351 void osc_object_clear_contended(struct osc_object *obj)
352 {
353         obj->oo_contended = 0;
354 }
355
356 int osc_object_is_contended(struct osc_object *obj)
357 {
358         struct osc_device *dev  = lu2osc_dev(obj->oo_cl.co_lu.lo_dev);
359         int osc_contention_time = dev->od_contention_time;
360         cfs_time_t cur_time     = cfs_time_current();
361         cfs_time_t retry_time;
362
363         if (OBD_FAIL_CHECK(OBD_FAIL_OSC_OBJECT_CONTENTION))
364                 return 1;
365
366         if (!obj->oo_contended)
367                 return 0;
368
369         /*
370          * I like copy-paste. the code is copied from
371          * ll_file_is_contended.
372          */
373         retry_time = cfs_time_add(obj->oo_contention_time,
374                                   cfs_time_seconds(osc_contention_time));
375         if (cfs_time_after(cur_time, retry_time)) {
376                 osc_object_clear_contended(obj);
377                 return 0;
378         }
379         return 1;
380 }
381
382 static const struct cl_object_operations osc_ops = {
383         .coo_page_init   = osc_page_init,
384         .coo_lock_init   = osc_lock_init,
385         .coo_io_init     = osc_io_init,
386         .coo_attr_get    = osc_attr_get,
387         .coo_attr_update = osc_attr_update,
388         .coo_glimpse     = osc_object_glimpse,
389         .coo_prune       = osc_object_prune,
390         .coo_find_cbdata = osc_object_find_cbdata,
391         .coo_fiemap      = osc_object_fiemap,
392 };
393
394 static const struct lu_object_operations osc_lu_obj_ops = {
395         .loo_object_init      = osc_object_init,
396         .loo_object_release   = NULL,
397         .loo_object_free      = osc_object_free,
398         .loo_object_print     = osc_object_print,
399         .loo_object_invariant = NULL
400 };
401
402 struct lu_object *osc_object_alloc(const struct lu_env *env,
403                                    const struct lu_object_header *unused,
404                                    struct lu_device *dev)
405 {
406         struct osc_object *osc;
407         struct lu_object  *obj;
408
409         OBD_SLAB_ALLOC_PTR_GFP(osc, osc_object_kmem, GFP_NOFS);
410         if (osc != NULL) {
411                 obj = osc2lu(osc);
412                 lu_object_init(obj, NULL, dev);
413                 osc->oo_cl.co_ops = &osc_ops;
414                 obj->lo_ops = &osc_lu_obj_ops;
415         } else
416                 obj = NULL;
417         return obj;
418 }
419
420 /** @} osc */