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, write to the
18 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
19 * Boston, MA 021110-1307, USA
24 * Copyright (c) 2012 Intel, Inc.
25 * Use is subject to license terms.
27 * Author: Johann Lombardi <johann.lombardi@intel.com>
28 * Author: Niu Yawei <yawei.niu@intel.com>
32 * Management of the device associated with a Quota Master Target (QMT).
34 * The QMT holds the cluster wide quota limits. It stores the quota settings
35 * ({hard,soft} limit & grace time) in a global index file and is in charge
36 * of allocating quota space to slaves while guaranteeing that the overall
37 * limits aren't exceeded. The QMT also maintains one index per slave (in fact,
38 * one per slave per quota type) used to track how much space is allocated
39 * to a given slave. Now that the QMT is aware of the quota space distribution
40 * among slaves, it can afford to rebalance efficiently quota space from one
41 * slave to another. Slaves are asked to release quota space via glimpse
42 * callbacks sent on DLM locks which are granted to slaves when those latters
43 * acquire quota space.
45 * The QMT device is currently set up by the MDT and should probably be moved
46 * to a separate target in the future. Meanwhile, the MDT forwards all quota
47 * requests to the QMT via a list of request handlers (see struct qmt_handlers
48 * in lustre_quota.h). The QMT also borrows the LDLM namespace from the MDT.
50 * To bring up a QMT device, the following steps must be completed:
52 * - call ->ldto_device_alloc to allocate the QMT device and perform basic
53 * initialization like connecting to the backend OSD device or setting up the
54 * default pools and the QMT procfs directory.
56 * - the MDT can then connect to the QMT instance via legacy obd_connect path.
58 * - once the MDT stack has been fully configured, ->ldto_prepare must be called
59 * to configure on-disk objects associated with this master target.
61 * To shutdown a QMT device, the MDT just has to disconnect from the QMT.
63 * The qmt_device_type structure is registered when the lquota module is
64 * loaded and all the steps described above are automatically done when the MDT
65 * set up the Quota Master Target via calls to class_attach/class_setup, see
66 * mdt_quota_init() for more details.
70 # define EXPORT_SYMTAB
73 #define DEBUG_SUBSYSTEM S_LQUOTA
75 #include <obd_class.h>
76 #include <lprocfs_status.h>
77 #include <lustre_disk.h>
78 #include "qmt_internal.h"
80 static const struct lu_device_operations qmt_lu_ops;
83 * Release quota master target and all data structure associated with this
85 * Called on MDT0 cleanup.
87 * \param env - is the environment passed by the caller
88 * \param ld - is the lu_device associated with the qmt device to be released
90 * \retval - NULL on success (backend OSD device is managed by the main stack),
91 * appropriate error on failure
93 static struct lu_device *qmt_device_fini(const struct lu_env *env,
96 struct qmt_device *qmt = lu2qmt_dev(ld);
101 CDEBUG(D_QUOTA, "%s: initiating QMT shutdown\n", qmt->qmt_svname);
102 qmt->qmt_stopping = true;
104 /* kill pool instances, if any */
105 qmt_pool_fini(env, qmt);
107 /* remove qmt proc entry */
108 if (qmt->qmt_proc != NULL && !IS_ERR(qmt->qmt_proc)) {
109 lprocfs_remove(&qmt->qmt_proc);
110 qmt->qmt_proc = NULL;
113 /* stop rebalance thread */
114 qmt_stop_reba_thread(qmt);
116 /* disconnect from OSD */
117 if (qmt->qmt_child_exp != NULL) {
118 obd_disconnect(qmt->qmt_child_exp);
119 qmt->qmt_child_exp = NULL;
120 qmt->qmt_child = NULL;
123 /* clear references to MDT namespace */
124 ld->ld_obd->obd_namespace = NULL;
131 * Connect a quota master to the backend OSD device.
133 * \param env - is the environment passed by the caller
134 * \param qmt - is the quota master target to be connected
135 * \param cfg - is the configuration log record from which we need to extract
136 * the service name of the backend OSD device to connect to.
138 * \retval - 0 on success, appropriate error on failure
140 static int qmt_connect_to_osd(const struct lu_env *env, struct qmt_device *qmt,
141 struct lustre_cfg *cfg)
143 struct obd_connect_data *data = NULL;
144 struct obd_device *obd;
145 struct lu_device *ld = qmt2lu_dev(qmt);
149 LASSERT(qmt->qmt_child_exp == NULL);
153 GOTO(out, rc = -ENOMEM);
155 /* look-up OBD device associated with the backend OSD device.
156 * The MDT is kind enough to pass the OBD name in QMT configuration */
157 obd = class_name2obd(lustre_cfg_string(cfg, 3));
159 CERROR("%s: can't locate backend osd device: %s\n",
160 qmt->qmt_svname, lustre_cfg_string(cfg, 3));
161 GOTO(out, rc = -ENOTCONN);
164 data->ocd_connect_flags = OBD_CONNECT_VERSION;
165 data->ocd_version = LUSTRE_VERSION_CODE;
167 /* connect to OSD device */
168 rc = obd_connect(NULL, &qmt->qmt_child_exp, obd, &obd->obd_uuid, data,
171 CERROR("%s: cannot connect to osd dev %s (%d)\n",
172 qmt->qmt_svname, obd->obd_name, rc);
176 /* initialize site (although it isn't used anywhere) and lu_device
177 * pointer to next device */
178 qmt->qmt_child = lu2dt_dev(qmt->qmt_child_exp->exp_obd->obd_lu_dev);
179 ld->ld_site = qmt->qmt_child_exp->exp_obd->obd_lu_dev->ld_site;
188 * Initialize quota master target device. This includers connecting to
189 * the backend OSD device, initializing the pool configuration and creating the
190 * root procfs directory dedicated to this quota target.
191 * The rest of the initialization is done when the stack is fully configured
192 * (i.e. when ->ldo_start is called across the stack).
194 * This function is called on MDT0 setup.
196 * \param env - is the environment passed by the caller
197 * \param qmt - is the quota master target to be initialized
198 * \param ldt - is the device type structure associated with the qmt device
199 * \param cfg - is the configuration record used to configure the qmt device
201 * \retval - 0 on success, appropriate error on failure
203 static int qmt_device_init0(const struct lu_env *env, struct qmt_device *qmt,
204 struct lu_device_type *ldt, struct lustre_cfg *cfg)
206 struct lu_device *ld = qmt2lu_dev(qmt);
207 struct obd_device *obd, *mdt_obd;
208 struct obd_type *type;
212 /* record who i am, it might be useful ... */
213 strncpy(qmt->qmt_svname, lustre_cfg_string(cfg, 0),
214 sizeof(qmt->qmt_svname) - 1);
216 /* look-up the obd_device associated with the qmt */
217 obd = class_name2obd(qmt->qmt_svname);
221 /* reference each other */
222 obd->obd_lu_dev = ld;
225 /* look-up the parent MDT to steal its ldlm namespace ... */
226 mdt_obd = class_name2obd(lustre_cfg_string(cfg, 2));
230 /* borrow MDT namespace. kind of a hack until we have our own namespace
231 * & service threads */
232 LASSERT(mdt_obd->obd_namespace != NULL);
233 obd->obd_namespace = mdt_obd->obd_namespace;
234 qmt->qmt_ns = obd->obd_namespace;
236 /* connect to backend osd device */
237 rc = qmt_connect_to_osd(env, qmt, cfg);
241 /* set up and start rebalance thread */
242 thread_set_flags(&qmt->qmt_reba_thread, SVC_STOPPED);
243 cfs_waitq_init(&qmt->qmt_reba_thread.t_ctl_waitq);
244 CFS_INIT_LIST_HEAD(&qmt->qmt_reba_list);
245 spin_lock_init(&qmt->qmt_reba_lock);
246 rc = qmt_start_reba_thread(qmt);
248 CERROR("%s: failed to start rebalance thread (%d)\n",
249 qmt->qmt_svname, rc);
253 /* at the moment there is no linkage between lu_type and obd_type, so
254 * we lookup obd_type this way */
255 type = class_search_type(LUSTRE_QMT_NAME);
256 LASSERT(type != NULL);
258 /* register proc directory associated with this qmt */
259 qmt->qmt_proc = lprocfs_register(qmt->qmt_svname, type->typ_procroot,
261 if (IS_ERR(qmt->qmt_proc)) {
262 rc = PTR_ERR(qmt->qmt_proc);
263 CERROR("%s: failed to create qmt proc entry (%d)\n",
264 qmt->qmt_svname, rc);
268 /* initialize pool configuration */
269 rc = qmt_pool_init(env, qmt);
275 qmt_device_fini(env, ld);
280 * Free quota master target device. Companion of qmt_device_alloc()
282 * \param env - is the environment passed by the caller
283 * \param ld - is the lu_device associated with the qmt dev to be freed
285 * \retval - NULL on success (backend OSD device is managed by the main stack),
286 * appropriate error on failure
288 static struct lu_device *qmt_device_free(const struct lu_env *env,
289 struct lu_device *ld)
291 struct qmt_device *qmt = lu2qmt_dev(ld);
294 LASSERT(qmt != NULL);
302 * Allocate quota master target and initialize it.
304 * \param env - is the environment passed by the caller
305 * \param ldt - is the device type structure associated with the qmt
306 * \param cfg - is the configuration record used to configure the qmt
308 * \retval - lu_device structure associated with the qmt on success,
309 * appropriate error on failure
311 static struct lu_device *qmt_device_alloc(const struct lu_env *env,
312 struct lu_device_type *ldt,
313 struct lustre_cfg *cfg)
315 struct qmt_device *qmt;
316 struct lu_device *ld;
320 /* allocate qmt device */
323 RETURN(ERR_PTR(-ENOMEM));
325 /* configure lu/dt_device */
326 ld = qmt2lu_dev(qmt);
327 dt_device_init(&qmt->qmt_dt_dev, ldt);
328 ld->ld_ops = &qmt_lu_ops;
330 /* initialize qmt device */
331 rc = qmt_device_init0(env, qmt, ldt, cfg);
333 qmt_device_free(env, ld);
340 LU_KEY_INIT_FINI(qmt, struct qmt_thread_info);
341 LU_TYPE_INIT_FINI(qmt, &qmt_thread_key);
342 LU_CONTEXT_KEY_DEFINE(qmt, LCT_MD_THREAD);
345 * lu device type operations associated with the master target.
347 static struct lu_device_type_operations qmt_device_type_ops = {
348 .ldto_init = qmt_type_init,
349 .ldto_fini = qmt_type_fini,
351 .ldto_start = qmt_type_start,
352 .ldto_stop = qmt_type_stop,
354 .ldto_device_alloc = qmt_device_alloc,
355 .ldto_device_free = qmt_device_free,
357 .ldto_device_fini = qmt_device_fini,
361 * lu device type structure associated with the master target.
362 * MDT0 uses this structure to configure the qmt.
364 static struct lu_device_type qmt_device_type = {
365 .ldt_tags = LU_DEVICE_DT,
366 .ldt_name = LUSTRE_QMT_NAME,
367 .ldt_ops = &qmt_device_type_ops,
368 .ldt_ctx_tags = LCT_MD_THREAD,
372 * obd_connect handler used by the MDT to connect to the master target.
374 static int qmt_device_obd_connect(const struct lu_env *env,
375 struct obd_export **exp,
376 struct obd_device *obd,
377 struct obd_uuid *cluuid,
378 struct obd_connect_data *data,
381 struct lustre_handle conn;
385 rc = class_connect(&conn, obd, cluuid);
389 *exp = class_conn2export(&conn);
394 * obd_disconnect handler used by the MDT to disconnect from the master target.
395 * We trigger cleanup on disconnect since it means that the MDT is about to
398 static int qmt_device_obd_disconnect(struct obd_export *exp)
400 struct obd_device *obd = exp->exp_obd;
404 rc = class_disconnect(exp);
408 rc = class_manual_cleanup(obd);
413 * obd device operations associated with the master target.
415 struct obd_ops qmt_obd_ops = {
416 .o_owner = THIS_MODULE,
417 .o_connect = qmt_device_obd_connect,
418 .o_disconnect = qmt_device_obd_disconnect,
422 * Called when the MDS is fully configured. We use it to set up local objects
423 * associated with the quota master target.
425 * \param env - is the environment passed by the caller
426 * \param parent - is the lu_device of the parent, that's to say the mdt
427 * \param ld - is the lu_device associated with the master target
429 * \retval - 0 on success, appropriate error on failure
431 static int qmt_device_prepare(const struct lu_env *env,
432 struct lu_device *parent,
433 struct lu_device *ld)
435 struct qmt_device *qmt = lu2qmt_dev(ld);
436 struct dt_object *qmt_root;
440 /* initialize quota master root directory where all index files will be
442 qmt_root = lquota_disk_dir_find_create(env, qmt->qmt_child, NULL,
444 if (IS_ERR(qmt_root)) {
445 rc = PTR_ERR(qmt_root);
446 CERROR("%s: failed to create master quota directory (%d)\n",
447 qmt->qmt_svname, rc);
451 /* initialize on-disk indexes associated with each pool */
452 rc = qmt_pool_prepare(env, qmt, qmt_root);
454 lu_object_put(env, &qmt_root->do_lu);
459 * lu device operations for the quota master target
461 static const struct lu_device_operations qmt_lu_ops = {
462 .ldo_prepare = qmt_device_prepare,
463 .ldo_process_config = NULL, /* to be defined for dynamic pool
467 /* global variable initialization called when the lquota module is loaded */
468 int qmt_glb_init(void)
473 rc = class_register_type(&qmt_obd_ops, NULL, NULL, LUSTRE_QMT_NAME,
478 /* called when the lquota module is about to be unloaded */
479 void qmt_glb_fini(void)
481 class_unregister_type(LUSTRE_QMT_NAME);