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  *  lustre/quota/quota_adjust_qunit.c
   5  *
   6  *  Copyright (c) 2005 Cluster File Systems, Inc.
   7  *
   8  *   This file is part of Lustre, http://www.lustre.org.
   9  *
  10  *   No redistribution or use is permitted outside of Cluster File Systems, Inc.
  11  *
  12  */
  13 #ifndef EXPORT_SYMTAB
  14 # define EXPORT_SYMTAB
  15 #endif
  16 #define DEBUG_SUBSYSTEM S_MDS
  17
  18 #ifdef __KERNEL__
  19 # include <linux/version.h>
  20 # include <linux/module.h>
  21 # include <linux/init.h>
  22 # include <linux/fs.h>
  23 # include <linux/jbd.h>
  24 # include <linux/ext3_fs.h>
  25 # include <linux/quota.h>
  26 # if (LINUX_VERSION_CODE >= KERNEL_VERSION(2,5,0))
  27 #  include <linux/smp_lock.h>
  28 #  include <linux/buffer_head.h>
  29 #  include <linux/workqueue.h>
  30 #  include <linux/mount.h>
  31 # else
  32 #  include <linux/locks.h>
  33 # endif
  34 #else /* __KERNEL__ */
  35 # include <liblustre.h>
  36 #endif
  37
  38 #include <obd_class.h>
  39 #include <lustre_mds.h>
  40 #include <lustre_dlm.h>
  41 #include <lustre_cfg.h>
  42 #include <obd_ost.h>
  43 #include <lustre_fsfilt.h>
  44 #include <linux/lustre_quota.h>
  45 #include <class_hash.h>
  46 #include "quota_internal.h"
  47
  48 #ifdef __KERNEL__
  49 /* this function is charge of recording lqs_ino_rec and
  50  * lqs_blk_rec. when a lquota slave checks a quota
  51  * request(check_cur_qunit) and finishes a quota
  52  * request(dqacq_completion), it will be called.
  53  * is_chk: whether it is checking quota; otherwise, it is finishing
  54  * is_acq: whether it is acquiring; otherwise, it is releasing
  55  */
  56 void quota_compute_lqs(struct qunit_data *qdata, struct lustre_qunit_size *lqs,
  57                       int is_chk, int is_acq)
  58 {
  59         int is_blk;
  60
  61         LASSERT(qdata && lqs);
  62         LASSERT_SPIN_LOCKED(&lqs->lqs_lock);
  63         is_blk = QDATA_IS_BLK(qdata);
  64
  65         if (is_chk) {
  66                 if (is_acq) {
  67                         if (is_blk)
  68                                 lqs->lqs_blk_rec += qdata->qd_count;
  69                         else
  70                                 lqs->lqs_ino_rec += qdata->qd_count;
  71                 } else {
  72                         if (is_blk)
  73                                 lqs->lqs_blk_rec -= qdata->qd_count;
  74                         else
  75                                 lqs->lqs_ino_rec -= qdata->qd_count;
  76                 }
  77         } else {
  78                 if (is_acq) {
  79                         if (is_blk)
  80                                 lqs->lqs_blk_rec -= qdata->qd_count;
  81                         else
  82                                 lqs->lqs_ino_rec -= qdata->qd_count;
  83                 } else {
  84                         if (is_blk)
  85                                 lqs->lqs_blk_rec += qdata->qd_count;
  86                         else
  87                                 lqs->lqs_ino_rec += qdata->qd_count;
  88                 }
  89         }
  90 }
  91
  92 void qdata_to_oqaq(struct qunit_data *qdata, struct quota_adjust_qunit *oqaq)
  93 {
  94         LASSERT(qdata);
  95         LASSERT(oqaq);
  96
  97         oqaq->qaq_flags = qdata->qd_flags;
  98         oqaq->qaq_id    = qdata->qd_id;
  99         if (QDATA_IS_ADJBLK(qdata))
 100                 oqaq->qaq_bunit_sz = qdata->qd_qunit;
 101         if (QDATA_IS_ADJINO(qdata))
 102                 oqaq->qaq_iunit_sz = qdata->qd_qunit;
 103 }
 104
 105 int quota_search_lqs(struct qunit_data *qdata, struct quota_adjust_qunit *oqaq,
 106                      struct lustre_quota_ctxt *qctxt,
 107                      struct lustre_qunit_size **lqs_return)
 108 {
 109         struct quota_adjust_qunit *oqaq_tmp = NULL;
 110         ENTRY;
 111
 112         LASSERT(*lqs_return == NULL);
 113         LASSERT(oqaq || qdata);
 114
 115         if (!oqaq) {
 116                 OBD_ALLOC_PTR(oqaq_tmp);
 117                 if (!oqaq_tmp)
 118                         RETURN(-ENOMEM);
 119                 qdata_to_oqaq(qdata, oqaq_tmp);
 120         } else {
 121                 oqaq_tmp = oqaq;
 122         }
 123
 124         *lqs_return = lustre_hash_get_object_by_key(LQC_HASH_BODY(qctxt),
 125                                                     oqaq_tmp);
 126         if (*lqs_return)
 127                 LQS_DEBUG((*lqs_return), "show lqs\n");
 128
 129         if (!oqaq)
 130                 OBD_FREE_PTR(oqaq_tmp);
 131         RETURN(0);
 132 }
 133
 134 int quota_create_lqs(struct qunit_data *qdata, struct quota_adjust_qunit *oqaq,
 135                      struct lustre_quota_ctxt *qctxt,
 136                      struct lustre_qunit_size **lqs_return)
 137 {
 138         int rc = 0;
 139         struct quota_adjust_qunit *oqaq_tmp = NULL;
 140         struct lustre_qunit_size *lqs = NULL;
 141         ENTRY;
 142
 143         LASSERT(*lqs_return == NULL);
 144         LASSERT(oqaq || qdata);
 145
 146         if (!oqaq) {
 147                 OBD_ALLOC_PTR(oqaq_tmp);
 148                 if (!oqaq_tmp)
 149                         RETURN(-ENOMEM);
 150                 qdata_to_oqaq(qdata, oqaq_tmp);
 151         } else {
 152                 oqaq_tmp = oqaq;
 153         }
 154
 155         OBD_ALLOC_PTR(lqs);
 156         if (!lqs)
 157                 GOTO(out, rc = -ENOMEM);
 158
 159         spin_lock_init(&lqs->lqs_lock);
 160         lqs->lqs_bwrite_pending = 0;
 161         lqs->lqs_iwrite_pending = 0;
 162         lqs->lqs_ino_rec = 0;
 163         lqs->lqs_blk_rec = 0;
 164         lqs->lqs_id = oqaq_tmp->qaq_id;
 165         lqs->lqs_flags = QAQ_IS_GRP(oqaq_tmp);
 166         lqs->lqs_bunit_sz = qctxt->lqc_bunit_sz;
 167         lqs->lqs_iunit_sz = qctxt->lqc_iunit_sz;
 168         lqs->lqs_btune_sz = qctxt->lqc_btune_sz;
 169         lqs->lqs_itune_sz = qctxt->lqc_itune_sz;
 170         if (qctxt->lqc_handler) {
 171                 lqs->lqs_last_bshrink  = 0;
 172                 lqs->lqs_last_ishrink  = 0;
 173         }
 174         lqs_initref(lqs);
 175         rc = lustre_hash_additem_unique(LQC_HASH_BODY(qctxt),
 176                                         oqaq_tmp, &lqs->lqs_hash);
 177         LQS_DEBUG(lqs, "create lqs\n");
 178         if (!rc) {
 179                 lqs_getref(lqs);
 180                 *lqs_return = lqs;
 181         }
 182  out:
 183         if (rc && lqs)
 184                 OBD_FREE_PTR(lqs);
 185         if (!oqaq)
 186                 OBD_FREE_PTR(oqaq_tmp);
 187         RETURN(rc);
 188 }
 189
 190 int quota_adjust_slave_lqs(struct quota_adjust_qunit *oqaq,
 191                            struct lustre_quota_ctxt *qctxt)
 192 {
 193         struct lustre_qunit_size *lqs = NULL;
 194         unsigned long *lbunit, *liunit, *lbtune, *litune;
 195         signed long b_tmp = 0, i_tmp = 0;
 196         cfs_time_t time_limit = 0;
 197         int rc = 0;
 198         ENTRY;
 199
 200         if (OBD_FAIL_CHECK(OBD_FAIL_QUOTA_WITHOUT_CHANGE_QS))
 201                 RETURN(0);
 202
 203         LASSERT(qctxt);
 204 search_lqs:
 205         rc = quota_search_lqs(NULL, oqaq, qctxt, &lqs);
 206
 207         /* deleting the lqs, because a user sets lfs quota 0 0 0 0  */
 208         if (!oqaq->qaq_bunit_sz && !oqaq->qaq_iunit_sz && QAQ_IS_ADJBLK(oqaq) &&
 209             QAQ_IS_ADJINO(oqaq)) {
 210                 if (lqs) {
 211                         LQS_DEBUG(lqs, "release lqs\n");
 212                         /* this is for quota_search_lqs */
 213                         lqs_putref(lqs);
 214                         /* this is for deleting this lqs */
 215                         lqs_putref(lqs);
 216                 }
 217                 RETURN(rc);
 218         }
 219
 220         if (!lqs) {
 221                 rc = quota_create_lqs(NULL, oqaq, qctxt, &lqs);
 222                 if (rc == -EALREADY)
 223                         goto search_lqs;
 224                 if (rc < 0)
 225                         RETURN(rc);
 226         }
 227
 228         lbunit = &lqs->lqs_bunit_sz;
 229         liunit = &lqs->lqs_iunit_sz;
 230         lbtune = &lqs->lqs_btune_sz;
 231         litune = &lqs->lqs_itune_sz;
 232
 233         spin_lock(&lqs->lqs_lock);
 234         CDEBUG(D_QUOTA, "before: bunit: %lu, iunit: %lu.\n", *lbunit, *liunit);
 235         /* adjust the slave's block qunit size */
 236         if (QAQ_IS_ADJBLK(oqaq)) {
 237                 cfs_duration_t sec = cfs_time_seconds(qctxt->lqc_switch_seconds);
 238
 239                 b_tmp = *lbunit - oqaq->qaq_bunit_sz;
 240
 241                 if (qctxt->lqc_handler && b_tmp > 0)
 242                         lqs->lqs_last_bshrink = cfs_time_current();
 243
 244                 if (qctxt->lqc_handler && b_tmp < 0) {
 245                         time_limit = cfs_time_add(lqs->lqs_last_bshrink, sec);
 246                         if (!lqs->lqs_last_bshrink ||
 247                             cfs_time_after(cfs_time_current(), time_limit)) {
 248                                 *lbunit = oqaq->qaq_bunit_sz;
 249                                 *lbtune = (*lbunit) / 2;
 250                         } else {
 251                                 b_tmp = 0;
 252                         }
 253                 } else {
 254                         *lbunit = oqaq->qaq_bunit_sz;
 255                         *lbtune = (*lbunit) / 2;
 256                 }
 257         }
 258
 259         /* adjust the slave's file qunit size */
 260         if (QAQ_IS_ADJINO(oqaq)) {
 261                 i_tmp = *liunit - oqaq->qaq_iunit_sz;
 262
 263                 if (qctxt->lqc_handler && i_tmp > 0)
 264                         lqs->lqs_last_ishrink  = cfs_time_current();
 265
 266                 if (qctxt->lqc_handler && i_tmp < 0) {
 267                         time_limit = cfs_time_add(lqs->lqs_last_ishrink,
 268                                                   cfs_time_seconds(qctxt->
 269                                                   lqc_switch_seconds));
 270                         if (!lqs->lqs_last_ishrink ||
 271                             cfs_time_after(cfs_time_current(), time_limit)) {
 272                                 *liunit = oqaq->qaq_iunit_sz;
 273                                 *litune = (*liunit) / 2;
 274                         } else {
 275                                 i_tmp = 0;
 276                         }
 277                 } else {
 278                         *liunit = oqaq->qaq_iunit_sz;
 279                         *litune = (*liunit) / 2;
 280                 }
 281         }
 282         CDEBUG(D_QUOTA, "after: bunit: %lu, iunit: %lu.\n", *lbunit, *liunit);
 283         spin_unlock(&lqs->lqs_lock);
 284
 285         lqs_putref(lqs);
 286
 287         if (b_tmp > 0)
 288                 rc |= LQS_BLK_DECREASE;
 289         else if (b_tmp < 0)
 290                 rc |= LQS_BLK_INCREASE;
 291
 292         if (i_tmp > 0)
 293                 rc |= LQS_INO_DECREASE;
 294         else if (i_tmp < 0)
 295                 rc |= LQS_INO_INCREASE;
 296
 297         RETURN(rc);
 298 }
 299
 300 int filter_quota_adjust_qunit(struct obd_export *exp,
 301                               struct quota_adjust_qunit *oqaq)
 302 {
 303         struct obd_device *obd = exp->exp_obd;
 304         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
 305         unsigned int uid = 0, gid = 0;
 306         int rc = 0;
 307         ENTRY;
 308
 309         LASSERT(oqaq);
 310         LASSERT(QAQ_IS_ADJBLK(oqaq));
 311         rc = quota_adjust_slave_lqs(oqaq, qctxt);
 312         if (rc < 0) {
 313                 CERROR("adjust mds slave's qunit size failed!(rc:%d)\n", rc);
 314                 RETURN(rc);
 315         }
 316         if (QAQ_IS_GRP(oqaq))
 317                 gid = oqaq->qaq_id;
 318         else
 319                 uid = oqaq->qaq_id;
 320
 321         if (rc > 0) {
 322                 rc = qctxt_adjust_qunit(obd, qctxt, uid, gid, 1, 0);
 323                 if (rc == -EDQUOT || rc == -EBUSY) {
 324                         CDEBUG(D_QUOTA, "rc: %d.\n", rc);
 325                         rc = 0;
 326                 }
 327                 if (rc)
 328                         CERROR("slave adjust block quota failed!(rc:%d)\n", rc);
 329         }
 330         RETURN(rc);
 331 }
 332 #endif /* __KERNEL__ */
 333
 334 int client_quota_adjust_qunit(struct obd_export *exp,
 335                               struct quota_adjust_qunit *oqaq)
 336 {
 337         struct ptlrpc_request *req;
 338         struct quota_adjust_qunit *oqa;
 339         int size[2] = { sizeof(struct ptlrpc_body), sizeof(*oqaq) };
 340         int rc = 0;
 341         ENTRY;
 342
 343         /* client don't support this kind of operation, abort it */
 344         if (!(exp->exp_connect_flags & OBD_CONNECT_CHANGE_QS)||
 345             OBD_FAIL_CHECK(OBD_FAIL_QUOTA_WITHOUT_CHANGE_QS)) {
 346                 CDEBUG(D_QUOTA, "osc: %s don't support change qunit size\n",
 347                        exp->exp_obd->obd_name);
 348                 RETURN(rc);
 349         }
 350         if (strcmp(exp->exp_obd->obd_type->typ_name, LUSTRE_OSC_NAME))
 351                 RETURN(-EINVAL);
 352
 353         req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OST_VERSION,
 354                               OST_QUOTA_ADJUST_QUNIT, 2, size, NULL);
 355         if (!req)
 356                 GOTO(out, rc = -ENOMEM);
 357
 358         oqa = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, sizeof(*oqaq));
 359         *oqa = *oqaq;
 360
 361         ptlrpc_req_set_repsize(req, 2, size);
 362
 363         rc = ptlrpc_queue_wait(req);
 364         if (rc) {
 365                 CERROR("%s: %s failed: rc = %d\n", exp->exp_obd->obd_name,
 366                        __FUNCTION__, rc);
 367                 GOTO(out, rc);
 368         }
 369         ptlrpc_req_finished(req);
 370 out:
 371         RETURN (rc);
 372 }
 373
 374 int lov_quota_adjust_qunit(struct obd_export *exp,
 375                            struct quota_adjust_qunit *oqaq)
 376 {
 377         struct obd_device *obd = class_exp2obd(exp);
 378         struct lov_obd *lov = &obd->u.lov;
 379         int i, rc = 0;
 380         ENTRY;
 381
 382         if (!QAQ_IS_ADJBLK(oqaq)) {
 383                 CERROR("bad qaq_flags %x for lov obd.\n", oqaq->qaq_flags);
 384                 RETURN(-EFAULT);
 385         }
 386
 387         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 388                 int err;
 389
 390                 if (!lov->lov_tgts[i]->ltd_active) {
 391                         CDEBUG(D_HA, "ost %d is inactive\n", i);
 392                         continue;
 393                 }
 394
 395                 err = obd_quota_adjust_qunit(lov->lov_tgts[i]->ltd_exp, oqaq);
 396                 if (err) {
 397                         if (lov->lov_tgts[i]->ltd_active && !rc)
 398                                 rc = err;
 399                         continue;
 400                 }
 401         }
 402         RETURN(rc);
 403 }