lustre/quota/quota_adjust_qunit.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  * GPL HEADER START
   5  *
   6  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   7  *
   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.
  11  *
  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).
  17  *
  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
  21  *
  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
  24  * have any questions.
  25  *
  26  * GPL HEADER END
  27  */
  28 /*
  29  * Copyright  2008 Sun Microsystems, Inc. All rights reserved
  30  * Use is subject to license terms.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  */
  36 #ifndef EXPORT_SYMTAB
  37 # define EXPORT_SYMTAB
  38 #endif
  39 #define DEBUG_SUBSYSTEM S_LQUOTA
  40
  41 #ifdef __KERNEL__
  42 # include <linux/version.h>
  43 # include <linux/module.h>
  44 # include <linux/init.h>
  45 # include <linux/fs.h>
  46 # include <linux/jbd.h>
  47 # include <linux/quota.h>
  48 # if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  49 #  include <linux/smp_lock.h>
  50 #  include <linux/buffer_head.h>
  51 #  include <linux/workqueue.h>
  52 #  include <linux/mount.h>
  53 # else
  54 #  include <linux/locks.h>
  55 # endif
  56 #else /* __KERNEL__ */
  57 # include <liblustre.h>
  58 #endif
  59
  60 #include <obd_class.h>
  61 #include <lustre_mds.h>
  62 #include <lustre_dlm.h>
  63 #include <lustre_cfg.h>
  64 #include <obd_ost.h>
  65 #include <lustre_fsfilt.h>
  66 #include <linux/lustre_quota.h>
  67 #include <class_hash.h>
  68 #include "quota_internal.h"
  69
  70 #ifdef HAVE_QUOTA_SUPPORT
  71
  72 #ifdef __KERNEL__
  73 /**
  74  * This function is charge of recording lqs_ino_rec and
  75  * lqs_blk_rec. when a lquota slave checks a quota
  76  * request(check_cur_qunit) and finishes a quota
  77  * request(dqacq_completion), it will be called.
  78  * is_chk: whether it is checking quota; otherwise, it is finishing
  79  * is_acq: whether it is acquiring; otherwise, it is releasing
  80  */
  81 void quota_compute_lqs(struct qunit_data *qdata, struct lustre_qunit_size *lqs,
  82                       int is_chk, int is_acq)
  83 {
  84         int is_blk;
  85
  86         LASSERT(qdata && lqs);
  87         LASSERT_SPIN_LOCKED(&lqs->lqs_lock);
  88         is_blk = QDATA_IS_BLK(qdata);
  89
  90         if (is_chk) {
  91                 if (is_acq) {
  92                         if (is_blk)
  93                                 lqs->lqs_blk_rec += qdata->qd_count;
  94                         else
  95                                 lqs->lqs_ino_rec += qdata->qd_count;
  96                 } else {
  97                         if (is_blk)
  98                                 lqs->lqs_blk_rec -= qdata->qd_count;
  99                         else
 100                                 lqs->lqs_ino_rec -= qdata->qd_count;
 101                 }
 102         } else {
 103                 if (is_acq) {
 104                         if (is_blk)
 105                                 lqs->lqs_blk_rec -= qdata->qd_count;
 106                         else
 107                                 lqs->lqs_ino_rec -= qdata->qd_count;
 108                 } else {
 109                         if (is_blk)
 110                                 lqs->lqs_blk_rec += qdata->qd_count;
 111                         else
 112                                 lqs->lqs_ino_rec += qdata->qd_count;
 113                 }
 114         }
 115 }
 116
 117 void qdata_to_oqaq(struct qunit_data *qdata, struct quota_adjust_qunit *oqaq)
 118 {
 119         LASSERT(qdata);
 120         LASSERT(oqaq);
 121
 122         oqaq->qaq_flags = qdata->qd_flags;
 123         oqaq->qaq_id    = qdata->qd_id;
 124         if (QDATA_IS_ADJBLK(qdata))
 125                 oqaq->qaq_bunit_sz = qdata->qd_qunit;
 126         if (QDATA_IS_ADJINO(qdata))
 127                 oqaq->qaq_iunit_sz = qdata->qd_qunit;
 128 }
 129
 130 int quota_search_lqs(struct qunit_data *qdata, struct quota_adjust_qunit *oqaq,
 131                      struct lustre_quota_ctxt *qctxt,
 132                      struct lustre_qunit_size **lqs_return)
 133 {
 134         struct quota_adjust_qunit *oqaq_tmp = NULL;
 135         ENTRY;
 136
 137         LASSERT(*lqs_return == NULL);
 138         LASSERT(oqaq || qdata);
 139
 140         if (!oqaq) {
 141                 OBD_ALLOC_PTR(oqaq_tmp);
 142                 if (!oqaq_tmp)
 143                         RETURN(-ENOMEM);
 144                 qdata_to_oqaq(qdata, oqaq_tmp);
 145         } else {
 146                 oqaq_tmp = oqaq;
 147         }
 148
 149         *lqs_return = lustre_hash_lookup(qctxt->lqc_lqs_hash, oqaq_tmp);
 150         if (*lqs_return)
 151                 LQS_DEBUG((*lqs_return), "show lqs\n");
 152
 153         if (!oqaq)
 154                 OBD_FREE_PTR(oqaq_tmp);
 155         RETURN(0);
 156 }
 157
 158 int quota_create_lqs(struct qunit_data *qdata, struct quota_adjust_qunit *oqaq,
 159                      struct lustre_quota_ctxt *qctxt,
 160                      struct lustre_qunit_size **lqs_return)
 161 {
 162         struct lustre_qunit_size *lqs = NULL;
 163         int rc = 0;
 164         ENTRY;
 165
 166         LASSERT(*lqs_return == NULL);
 167         LASSERT(oqaq || qdata);
 168
 169         OBD_ALLOC_PTR(lqs);
 170         if (!lqs)
 171                 GOTO(out, rc = -ENOMEM);
 172
 173         if (!oqaq)
 174                 qdata_to_oqaq(qdata, &lqs->lqs_key);
 175         else
 176                 lqs->lqs_key = *oqaq;
 177
 178         spin_lock_init(&lqs->lqs_lock);
 179         lqs->lqs_bwrite_pending = 0;
 180         lqs->lqs_iwrite_pending = 0;
 181         lqs->lqs_ino_rec = 0;
 182         lqs->lqs_blk_rec = 0;
 183         lqs->lqs_id = lqs->lqs_key.qaq_id;
 184         lqs->lqs_flags = QAQ_IS_GRP(&lqs->lqs_key);
 185         lqs->lqs_bunit_sz = qctxt->lqc_bunit_sz;
 186         lqs->lqs_iunit_sz = qctxt->lqc_iunit_sz;
 187         lqs->lqs_btune_sz = qctxt->lqc_btune_sz;
 188         lqs->lqs_itune_sz = qctxt->lqc_itune_sz;
 189         lqs->lqs_ctxt = qctxt;
 190         if (qctxt->lqc_handler) {
 191                 lqs->lqs_last_bshrink  = 0;
 192                 lqs->lqs_last_ishrink  = 0;
 193         }
 194         lqs_initref(lqs);
 195         rc = lustre_hash_add_unique(qctxt->lqc_lqs_hash,
 196                                     &lqs->lqs_key, &lqs->lqs_hash);
 197         LQS_DEBUG(lqs, "create lqs\n");
 198         if (!rc) {
 199                 lqs_getref(lqs);
 200                 *lqs_return = lqs;
 201         }
 202 out:
 203         if (rc && lqs)
 204                 OBD_FREE_PTR(lqs);
 205         RETURN(rc);
 206 }
 207
 208 int quota_adjust_slave_lqs(struct quota_adjust_qunit *oqaq,
 209                            struct lustre_quota_ctxt *qctxt)
 210 {
 211         struct lustre_qunit_size *lqs = NULL;
 212         unsigned long *lbunit, *liunit, *lbtune, *litune;
 213         signed long b_tmp = 0, i_tmp = 0;
 214         cfs_time_t time_limit = 0;
 215         int rc = 0;
 216         ENTRY;
 217
 218         LASSERT(qctxt);
 219 search_lqs:
 220         rc = quota_search_lqs(NULL, oqaq, qctxt, &lqs);
 221
 222         /* deleting the lqs, because a user sets lfs quota 0 0 0 0  */
 223         if (!oqaq->qaq_bunit_sz && !oqaq->qaq_iunit_sz && QAQ_IS_ADJBLK(oqaq) &&
 224             QAQ_IS_ADJINO(oqaq)) {
 225                 if (lqs) {
 226                         LQS_DEBUG(lqs, "release lqs\n");
 227                         /* this is for quota_search_lqs */
 228                         lqs_putref(lqs);
 229                         /* kill lqs */
 230                         lqs_putref(lqs);
 231                 }
 232                 RETURN(rc);
 233         }
 234
 235         if (!lqs) {
 236                 rc = quota_create_lqs(NULL, oqaq, qctxt, &lqs);
 237                 if (rc == -EALREADY)
 238                         goto search_lqs;
 239                 if (rc < 0)
 240                         RETURN(rc);
 241         }
 242
 243         lbunit = &lqs->lqs_bunit_sz;
 244         liunit = &lqs->lqs_iunit_sz;
 245         lbtune = &lqs->lqs_btune_sz;
 246         litune = &lqs->lqs_itune_sz;
 247
 248         CDEBUG(D_QUOTA, "before: bunit: %lu, iunit: %lu.\n", *lbunit, *liunit);
 249         spin_lock(&lqs->lqs_lock);
 250         /* adjust the slave's block qunit size */
 251         if (QAQ_IS_ADJBLK(oqaq)) {
 252                 cfs_duration_t sec = cfs_time_seconds(qctxt->lqc_switch_seconds);
 253
 254                 b_tmp = *lbunit - oqaq->qaq_bunit_sz;
 255
 256                 if (qctxt->lqc_handler && b_tmp > 0)
 257                         lqs->lqs_last_bshrink = cfs_time_current();
 258
 259                 if (qctxt->lqc_handler && b_tmp < 0) {
 260                         time_limit = cfs_time_add(lqs->lqs_last_bshrink, sec);
 261                         if (!lqs->lqs_last_bshrink ||
 262                             cfs_time_after(cfs_time_current(), time_limit)) {
 263                                 *lbunit = oqaq->qaq_bunit_sz;
 264                                 *lbtune = (*lbunit) / 2;
 265                         } else {
 266                                 b_tmp = 0;
 267                         }
 268                 } else {
 269                         *lbunit = oqaq->qaq_bunit_sz;
 270                         *lbtune = (*lbunit) / 2;
 271                 }
 272         }
 273
 274         /* adjust the slave's file qunit size */
 275         if (QAQ_IS_ADJINO(oqaq)) {
 276                 i_tmp = *liunit - oqaq->qaq_iunit_sz;
 277
 278                 if (qctxt->lqc_handler && i_tmp > 0)
 279                         lqs->lqs_last_ishrink  = cfs_time_current();
 280
 281                 if (qctxt->lqc_handler && i_tmp < 0) {
 282                         time_limit = cfs_time_add(lqs->lqs_last_ishrink,
 283                                                   cfs_time_seconds(qctxt->
 284                                                   lqc_switch_seconds));
 285                         if (!lqs->lqs_last_ishrink ||
 286                             cfs_time_after(cfs_time_current(), time_limit)) {
 287                                 *liunit = oqaq->qaq_iunit_sz;
 288                                 *litune = (*liunit) / 2;
 289                         } else {
 290                                 i_tmp = 0;
 291                         }
 292                 } else {
 293                         *liunit = oqaq->qaq_iunit_sz;
 294                         *litune = (*liunit) / 2;
 295                 }
 296         }
 297         spin_unlock(&lqs->lqs_lock);
 298         CDEBUG(D_QUOTA, "after: bunit: %lu, iunit: %lu.\n", *lbunit, *liunit);
 299
 300         lqs_putref(lqs);
 301
 302         if (b_tmp > 0)
 303                 rc |= LQS_BLK_DECREASE;
 304         else if (b_tmp < 0)
 305                 rc |= LQS_BLK_INCREASE;
 306
 307         if (i_tmp > 0)
 308                 rc |= LQS_INO_DECREASE;
 309         else if (i_tmp < 0)
 310                 rc |= LQS_INO_INCREASE;
 311
 312         RETURN(rc);
 313 }
 314
 315 int filter_quota_adjust_qunit(struct obd_export *exp,
 316                               struct quota_adjust_qunit *oqaq,
 317                               struct lustre_quota_ctxt *qctxt)
 318 {
 319         struct obd_device *obd = exp->exp_obd;
 320         unsigned int uid = 0, gid = 0;
 321         int rc = 0;
 322         ENTRY;
 323
 324         LASSERT(oqaq);
 325         LASSERT(QAQ_IS_ADJBLK(oqaq));
 326         rc = quota_adjust_slave_lqs(oqaq, qctxt);
 327         if (rc < 0) {
 328                 CERROR("adjust mds slave's qunit size failed!(rc:%d)\n", rc);
 329                 RETURN(rc);
 330         }
 331         if (QAQ_IS_GRP(oqaq))
 332                 gid = oqaq->qaq_id;
 333         else
 334                 uid = oqaq->qaq_id;
 335
 336         if (rc > 0) {
 337                 rc = qctxt_adjust_qunit(obd, qctxt, uid, gid, 1, 0, NULL);
 338                 if (rc == -EDQUOT || rc == -EBUSY || rc == -EAGAIN) {
 339                         CDEBUG(D_QUOTA, "rc: %d.\n", rc);
 340                         rc = 0;
 341                 }
 342                 if (rc)
 343                         CERROR("slave adjust block quota failed!(rc:%d)\n", rc);
 344         }
 345         RETURN(rc);
 346 }
 347 #endif /* __KERNEL__ */
 348 #endif
 349
 350 int client_quota_adjust_qunit(struct obd_export *exp,
 351                               struct quota_adjust_qunit *oqaq,
 352                               struct lustre_quota_ctxt *qctxt)
 353 {
 354         struct ptlrpc_request *req;
 355         struct quota_adjust_qunit *oqa;
 356         int rc = 0;
 357         ENTRY;
 358
 359         /* client don't support this kind of operation, abort it */
 360         if (!(exp->exp_connect_flags & OBD_CONNECT_CHANGE_QS)) {
 361                 CDEBUG(D_QUOTA, "osc: %s don't support change qunit size\n",
 362                        exp->exp_obd->obd_name);
 363                 RETURN(rc);
 364         }
 365         if (strcmp(exp->exp_obd->obd_type->typ_name, LUSTRE_OSC_NAME))
 366                 RETURN(-EINVAL);
 367
 368         req = ptlrpc_request_alloc_pack(class_exp2cliimp(exp),
 369                                         &RQF_OST_QUOTA_ADJUST_QUNIT,
 370                                         LUSTRE_OST_VERSION,
 371                                         OST_QUOTA_ADJUST_QUNIT);
 372         if (req == NULL)
 373                 RETURN(-ENOMEM);
 374
 375         oqa = req_capsule_client_get(&req->rq_pill, &RMF_QUOTA_ADJUST_QUNIT);
 376         *oqa = *oqaq;
 377
 378         ptlrpc_request_set_replen(req);
 379
 380         rc = ptlrpc_queue_wait(req);
 381         if (rc)
 382                 CERROR("%s: %s failed: rc = %d\n", exp->exp_obd->obd_name,
 383                        __FUNCTION__, rc);
 384         ptlrpc_req_finished(req);
 385         RETURN (rc);
 386 }
 387
 388 int lov_quota_adjust_qunit(struct obd_export *exp,
 389                            struct quota_adjust_qunit *oqaq,
 390                            struct lustre_quota_ctxt *qctxt)
 391 {
 392         struct obd_device *obd = class_exp2obd(exp);
 393         struct lov_obd *lov = &obd->u.lov;
 394         int i, rc = 0;
 395         ENTRY;
 396
 397         if (!QAQ_IS_ADJBLK(oqaq)) {
 398                 CERROR("bad qaq_flags %x for lov obd.\n", oqaq->qaq_flags);
 399                 RETURN(-EFAULT);
 400         }
 401
 402         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 403                 int err;
 404
 405                 if (!lov->lov_tgts[i]->ltd_active) {
 406                         CDEBUG(D_HA, "ost %d is inactive\n", i);
 407                         continue;
 408                 }
 409
 410                 err = obd_quota_adjust_qunit(lov->lov_tgts[i]->ltd_exp, oqaq,
 411                                              NULL);
 412                 if (err) {
 413                         if (lov->lov_tgts[i]->ltd_active && !rc)
 414                                 rc = err;
 415                         continue;
 416                 }
 417         }
 418         RETURN(rc);
 419 }