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 #  include <linux/smp_lock.h>
  49 #  include <linux/buffer_head.h>
  50 #  include <linux/workqueue.h>
  51 #  include <linux/mount.h>
  52 #else /* __KERNEL__ */
  53 # include <liblustre.h>
  54 #endif
  55
  56 #include <obd_class.h>
  57 #include <lustre_mds.h>
  58 #include <lustre_dlm.h>
  59 #include <lustre_cfg.h>
  60 #include <obd_ost.h>
  61 #include <lustre_fsfilt.h>
  62 #include <linux/lustre_quota.h>
  63 #include <class_hash.h>
  64 #include "quota_internal.h"
  65
  66 #ifdef HAVE_QUOTA_SUPPORT
  67
  68 #ifdef __KERNEL__
  69 /* this function is charge of recording lqs_ino_rec and
  70  * lqs_blk_rec. when a lquota slave checks a quota
  71  * request(check_cur_qunit) and finishes a quota
  72  * request(dqacq_completion), it will be called.
  73  * is_chk: whether it is checking quota; otherwise, it is finishing
  74  * is_acq: whether it is acquiring; otherwise, it is releasing
  75  */
  76 void quota_compute_lqs(struct qunit_data *qdata, struct lustre_qunit_size *lqs,
  77                       int is_chk, int is_acq)
  78 {
  79         int is_blk;
  80
  81         LASSERT(qdata && lqs);
  82         LASSERT_SPIN_LOCKED(&lqs->lqs_lock);
  83         is_blk = QDATA_IS_BLK(qdata);
  84
  85         if (is_chk) {
  86                 if (is_acq) {
  87                         if (is_blk)
  88                                 lqs->lqs_blk_rec += qdata->qd_count;
  89                         else
  90                                 lqs->lqs_ino_rec += qdata->qd_count;
  91                 } else {
  92                         if (is_blk)
  93                                 lqs->lqs_blk_rec -= qdata->qd_count;
  94                         else
  95                                 lqs->lqs_ino_rec -= qdata->qd_count;
  96                 }
  97         } else {
  98                 if (is_acq) {
  99                         if (is_blk)
 100                                 lqs->lqs_blk_rec -= qdata->qd_count;
 101                         else
 102                                 lqs->lqs_ino_rec -= qdata->qd_count;
 103                 } else {
 104                         if (is_blk)
 105                                 lqs->lqs_blk_rec += qdata->qd_count;
 106                         else
 107                                 lqs->lqs_ino_rec += qdata->qd_count;
 108                 }
 109         }
 110 }
 111
 112 void qdata_to_oqaq(struct qunit_data *qdata, struct quota_adjust_qunit *oqaq)
 113 {
 114         LASSERT(qdata);
 115         LASSERT(oqaq);
 116
 117         oqaq->qaq_flags = qdata->qd_flags;
 118         oqaq->qaq_id    = qdata->qd_id;
 119         if (QDATA_IS_ADJBLK(qdata))
 120                 oqaq->qaq_bunit_sz = qdata->qd_qunit;
 121         if (QDATA_IS_ADJINO(qdata))
 122                 oqaq->qaq_iunit_sz = qdata->qd_qunit;
 123 }
 124
 125 int quota_search_lqs(struct qunit_data *qdata, struct quota_adjust_qunit *oqaq,
 126                      struct lustre_quota_ctxt *qctxt,
 127                      struct lustre_qunit_size **lqs_return)
 128 {
 129         struct quota_adjust_qunit *oqaq_tmp = NULL;
 130         ENTRY;
 131
 132         LASSERT(*lqs_return == NULL);
 133         LASSERT(oqaq || qdata);
 134
 135         if (!oqaq) {
 136                 OBD_ALLOC_PTR(oqaq_tmp);
 137                 if (!oqaq_tmp)
 138                         RETURN(-ENOMEM);
 139                 qdata_to_oqaq(qdata, oqaq_tmp);
 140         } else {
 141                 oqaq_tmp = oqaq;
 142         }
 143
 144         *lqs_return = lustre_hash_lookup(qctxt->lqc_lqs_hash, oqaq_tmp);
 145         if (*lqs_return)
 146                 LQS_DEBUG((*lqs_return), "show lqs\n");
 147
 148         if (!oqaq)
 149                 OBD_FREE_PTR(oqaq_tmp);
 150         RETURN(0);
 151 }
 152
 153 int quota_create_lqs(struct qunit_data *qdata, struct quota_adjust_qunit *oqaq,
 154                      struct lustre_quota_ctxt *qctxt,
 155                      struct lustre_qunit_size **lqs_return)
 156 {
 157         struct lustre_qunit_size *lqs = NULL;
 158         int rc = 0;
 159         ENTRY;
 160
 161         LASSERT(*lqs_return == NULL);
 162         LASSERT(oqaq || qdata);
 163
 164         OBD_ALLOC_PTR(lqs);
 165         if (!lqs)
 166                 GOTO(out, rc = -ENOMEM);
 167
 168         if (!oqaq) {
 169                 qdata_to_oqaq(qdata, &lqs->lqs_key);
 170         } else {
 171                 lqs->lqs_key = *oqaq;
 172         }
 173
 174         spin_lock_init(&lqs->lqs_lock);
 175         lqs->lqs_bwrite_pending = 0;
 176         lqs->lqs_iwrite_pending = 0;
 177         lqs->lqs_ino_rec = 0;
 178         lqs->lqs_blk_rec = 0;
 179         lqs->lqs_id = lqs->lqs_key.qaq_id;
 180         lqs->lqs_flags = QAQ_IS_GRP(&lqs->lqs_key);
 181         lqs->lqs_bunit_sz = qctxt->lqc_bunit_sz;
 182         lqs->lqs_iunit_sz = qctxt->lqc_iunit_sz;
 183         lqs->lqs_btune_sz = qctxt->lqc_btune_sz;
 184         lqs->lqs_itune_sz = qctxt->lqc_itune_sz;
 185         lqs->lqs_ctxt = qctxt;
 186         if (qctxt->lqc_handler) {
 187                 lqs->lqs_last_bshrink  = 0;
 188                 lqs->lqs_last_ishrink  = 0;
 189         }
 190         lqs_initref(lqs);
 191         rc = lustre_hash_add_unique(qctxt->lqc_lqs_hash,
 192                                     &lqs->lqs_key, &lqs->lqs_hash);
 193         LQS_DEBUG(lqs, "create lqs\n");
 194         if (!rc) {
 195                 lqs_getref(lqs);
 196                 *lqs_return = lqs;
 197         }
 198  out:
 199         if (rc && lqs)
 200                 OBD_FREE_PTR(lqs);
 201         RETURN(rc);
 202 }
 203
 204 int quota_adjust_slave_lqs(struct quota_adjust_qunit *oqaq,
 205                            struct lustre_quota_ctxt *qctxt)
 206 {
 207         struct lustre_qunit_size *lqs = NULL;
 208         unsigned long *lbunit, *liunit, *lbtune, *litune;
 209         signed long b_tmp = 0, i_tmp = 0;
 210         cfs_time_t time_limit = 0;
 211         int rc = 0;
 212         ENTRY;
 213
 214         if (OBD_FAIL_CHECK(OBD_FAIL_QUOTA_WITHOUT_CHANGE_QS))
 215                 RETURN(0);
 216
 217         LASSERT(qctxt);
 218 search_lqs:
 219         rc = quota_search_lqs(NULL, oqaq, qctxt, &lqs);
 220
 221         /* deleting the lqs, because a user sets lfs quota 0 0 0 0  */
 222         if (!oqaq->qaq_bunit_sz && !oqaq->qaq_iunit_sz && QAQ_IS_ADJBLK(oqaq) &&
 223             QAQ_IS_ADJINO(oqaq)) {
 224                 if (lqs) {
 225                         LQS_DEBUG(lqs, "release lqs\n");
 226                         /* this is for quota_search_lqs */
 227                         lqs_putref(lqs);
 228                         /* kill lqs */
 229                         lqs_putref(lqs);
 230                 }
 231                 RETURN(rc);
 232         }
 233
 234         if (!lqs) {
 235                 rc = quota_create_lqs(NULL, oqaq, qctxt, &lqs);
 236                 if (rc == -EALREADY)
 237                         goto search_lqs;
 238                 if (rc < 0)
 239                         RETURN(rc);
 240         }
 241
 242         lbunit = &lqs->lqs_bunit_sz;
 243         liunit = &lqs->lqs_iunit_sz;
 244         lbtune = &lqs->lqs_btune_sz;
 245         litune = &lqs->lqs_itune_sz;
 246
 247         spin_lock(&lqs->lqs_lock);
 248         CDEBUG(D_QUOTA, "before: bunit: %lu, iunit: %lu.\n", *lbunit, *liunit);
 249         /* adjust the slave's block qunit size */
 250         if (QAQ_IS_ADJBLK(oqaq)) {
 251                 cfs_duration_t sec = cfs_time_seconds(qctxt->lqc_switch_seconds);
 252
 253                 b_tmp = *lbunit - oqaq->qaq_bunit_sz;
 254
 255                 if (qctxt->lqc_handler && b_tmp > 0)
 256                         lqs->lqs_last_bshrink = cfs_time_current();
 257
 258                 if (qctxt->lqc_handler && b_tmp < 0) {
 259                         time_limit = cfs_time_add(lqs->lqs_last_bshrink, sec);
 260                         if (!lqs->lqs_last_bshrink ||
 261                             cfs_time_after(cfs_time_current(), time_limit)) {
 262                                 *lbunit = oqaq->qaq_bunit_sz;
 263                                 *lbtune = (*lbunit) / 2;
 264                         } else {
 265                                 b_tmp = 0;
 266                         }
 267                 } else {
 268                         *lbunit = oqaq->qaq_bunit_sz;
 269                         *lbtune = (*lbunit) / 2;
 270                 }
 271         }
 272
 273         /* adjust the slave's file qunit size */
 274         if (QAQ_IS_ADJINO(oqaq)) {
 275                 i_tmp = *liunit - oqaq->qaq_iunit_sz;
 276
 277                 if (qctxt->lqc_handler && i_tmp > 0)
 278                         lqs->lqs_last_ishrink  = cfs_time_current();
 279
 280                 if (qctxt->lqc_handler && i_tmp < 0) {
 281                         time_limit = cfs_time_add(lqs->lqs_last_ishrink,
 282                                                   cfs_time_seconds(qctxt->
 283                                                   lqc_switch_seconds));
 284                         if (!lqs->lqs_last_ishrink ||
 285                             cfs_time_after(cfs_time_current(), time_limit)) {
 286                                 *liunit = oqaq->qaq_iunit_sz;
 287                                 *litune = (*liunit) / 2;
 288                         } else {
 289                                 i_tmp = 0;
 290                         }
 291                 } else {
 292                         *liunit = oqaq->qaq_iunit_sz;
 293                         *litune = (*liunit) / 2;
 294                 }
 295         }
 296         CDEBUG(D_QUOTA, "after: bunit: %lu, iunit: %lu.\n", *lbunit, *liunit);
 297         spin_unlock(&lqs->lqs_lock);
 298
 299         lqs_putref(lqs);
 300
 301         if (b_tmp > 0)
 302                 rc |= LQS_BLK_DECREASE;
 303         else if (b_tmp < 0)
 304                 rc |= LQS_BLK_INCREASE;
 305
 306         if (i_tmp > 0)
 307                 rc |= LQS_INO_DECREASE;
 308         else if (i_tmp < 0)
 309                 rc |= LQS_INO_INCREASE;
 310
 311         RETURN(rc);
 312 }
 313
 314 int filter_quota_adjust_qunit(struct obd_export *exp,
 315                               struct quota_adjust_qunit *oqaq,
 316                               struct lustre_quota_ctxt *qctxt)
 317 {
 318         struct obd_device *obd = exp->exp_obd;
 319         unsigned int uid = 0, gid = 0;
 320         int rc = 0;
 321         ENTRY;
 322
 323         LASSERT(oqaq);
 324         LASSERT(QAQ_IS_ADJBLK(oqaq));
 325         rc = quota_adjust_slave_lqs(oqaq, qctxt);
 326         if (rc < 0) {
 327                 CERROR("adjust mds slave's qunit size failed!(rc:%d)\n", rc);
 328                 RETURN(rc);
 329         }
 330         if (QAQ_IS_GRP(oqaq))
 331                 gid = oqaq->qaq_id;
 332         else
 333                 uid = oqaq->qaq_id;
 334
 335         if (rc > 0) {
 336                 rc = qctxt_adjust_qunit(obd, qctxt, uid, gid, 1, 0, NULL);
 337                 if (rc == -EDQUOT || rc == -EBUSY || rc == -EAGAIN) {
 338                         CDEBUG(D_QUOTA, "rc: %d.\n", rc);
 339                         rc = 0;
 340                 }
 341                 if (rc)
 342                         CERROR("slave adjust block quota failed!(rc:%d)\n", rc);
 343         }
 344         RETURN(rc);
 345 }
 346 #endif /* __KERNEL__ */
 347 #endif
 348
 349 int client_quota_adjust_qunit(struct obd_export *exp,
 350                               struct quota_adjust_qunit *oqaq,
 351                               struct lustre_quota_ctxt *qctxt)
 352 {
 353         struct ptlrpc_request *req;
 354         struct quota_adjust_qunit *oqa;
 355         __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*oqaq) };
 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             OBD_FAIL_CHECK(OBD_FAIL_QUOTA_WITHOUT_CHANGE_QS)) {
 362                 CDEBUG(D_QUOTA, "osc: %s don't support change qunit size\n",
 363                        exp->exp_obd->obd_name);
 364                 RETURN(rc);
 365         }
 366         if (strcmp(exp->exp_obd->obd_type->typ_name, LUSTRE_OSC_NAME))
 367                 RETURN(-EINVAL);
 368
 369         req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OST_VERSION,
 370                               OST_QUOTA_ADJUST_QUNIT, 2, size, NULL);
 371         if (!req)
 372                 GOTO(out, rc = -ENOMEM);
 373
 374         oqa = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, sizeof(*oqaq));
 375         *oqa = *oqaq;
 376
 377         ptlrpc_req_set_repsize(req, 2, size);
 378
 379         rc = ptlrpc_queue_wait(req);
 380         if (rc) {
 381                 CERROR("%s: %s failed: rc = %d\n", exp->exp_obd->obd_name,
 382                        __FUNCTION__, rc);
 383                 GOTO(out, rc);
 384         }
 385         ptlrpc_req_finished(req);
 386 out:
 387         RETURN (rc);
 388 }
 389
 390 int lov_quota_adjust_qunit(struct obd_export *exp,
 391                            struct quota_adjust_qunit *oqaq,
 392                            struct lustre_quota_ctxt *qctxt)
 393 {
 394         struct obd_device *obd = class_exp2obd(exp);
 395         struct lov_obd *lov = &obd->u.lov;
 396         int i, rc = 0;
 397         ENTRY;
 398
 399         if (!QAQ_IS_ADJBLK(oqaq)) {
 400                 CERROR("bad qaq_flags %x for lov obd.\n", oqaq->qaq_flags);
 401                 RETURN(-EFAULT);
 402         }
 403
 404         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 405                 int err;
 406
 407                 if (!lov->lov_tgts[i]->ltd_active) {
 408                         CDEBUG(D_HA, "ost %d is inactive\n", i);
 409                         continue;
 410                 }
 411
 412                 err = obd_quota_adjust_qunit(lov->lov_tgts[i]->ltd_exp, oqaq,
 413                                              NULL);
 414                 if (err) {
 415                         if (lov->lov_tgts[i]->ltd_active && !rc)
 416                                 rc = err;
 417                         continue;
 418                 }
 419         }
 420         RETURN(rc);
 421 }