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,
  93                    struct quota_adjust_qunit *oqaq)
  94 {
  95         LASSERT(qdata);
  96         LASSERT(oqaq);
  97
  98         oqaq->qaq_flags = qdata->qd_flags;
  99         oqaq->qaq_id    = qdata->qd_id;
 100         if (QDATA_IS_ADJBLK(qdata))
 101                 oqaq->qaq_bunit_sz = qdata->qd_qunit;
 102         if (QDATA_IS_ADJINO(qdata))
 103                 oqaq->qaq_iunit_sz = qdata->qd_qunit;
 104 }
 105
 106 int quota_search_lqs(struct qunit_data *qdata,
 107                   struct quota_adjust_qunit *oqaq,
 108                   struct lustre_quota_ctxt *qctxt,
 109                   struct lustre_qunit_size **lqs_return)
 110 {
 111         struct quota_adjust_qunit *oqaq_tmp = NULL;
 112         ENTRY;
 113
 114         LASSERT(*lqs_return == NULL);
 115         LASSERT(oqaq || qdata);
 116
 117         if (!oqaq) {
 118                 OBD_ALLOC_PTR(oqaq_tmp);
 119                 if (!oqaq_tmp)
 120                         RETURN(-ENOMEM);
 121                 qdata_to_oqaq(qdata, oqaq_tmp);
 122         } else {
 123                 oqaq_tmp = oqaq;
 124         }
 125
 126         *lqs_return = lustre_hash_get_object_by_key(LQC_HASH_BODY(qctxt),
 127                                                     oqaq_tmp);
 128         if (*lqs_return)
 129                 LQS_DEBUG((*lqs_return), "show lqs\n");
 130
 131         if (!oqaq)
 132                 OBD_FREE_PTR(oqaq_tmp);
 133         RETURN(0);
 134 }
 135
 136 int quota_create_lqs(struct qunit_data *qdata,
 137                      struct quota_adjust_qunit *oqaq,
 138                      struct lustre_quota_ctxt *qctxt,
 139                      struct lustre_qunit_size **lqs_return)
 140 {
 141         int rc = 0;
 142         struct quota_adjust_qunit *oqaq_tmp = NULL;
 143         struct lustre_qunit_size *lqs = NULL;
 144         ENTRY;
 145
 146         LASSERT(*lqs_return == NULL);
 147         LASSERT(oqaq || qdata);
 148
 149         if (!oqaq) {
 150                 OBD_ALLOC_PTR(oqaq_tmp);
 151                 if (!oqaq_tmp)
 152                         RETURN(-ENOMEM);
 153                 qdata_to_oqaq(qdata, oqaq_tmp);
 154         } else {
 155                 oqaq_tmp = oqaq;
 156         }
 157
 158         OBD_ALLOC_PTR(lqs);
 159         if (!lqs)
 160                 GOTO(out, rc = -ENOMEM);
 161
 162         spin_lock_init(&lqs->lqs_lock);
 163         lqs->lqs_bwrite_pending = 0;
 164         lqs->lqs_iwrite_pending = 0;
 165         lqs->lqs_ino_rec = 0;
 166         lqs->lqs_blk_rec = 0;
 167         lqs->lqs_id = oqaq_tmp->qaq_id;
 168         lqs->lqs_flags = QAQ_IS_GRP(oqaq_tmp);
 169         lqs->lqs_bunit_sz = qctxt->lqc_bunit_sz;
 170         lqs->lqs_iunit_sz = qctxt->lqc_iunit_sz;
 171         lqs->lqs_btune_sz = qctxt->lqc_btune_sz;
 172         lqs->lqs_itune_sz = qctxt->lqc_itune_sz;
 173         if (qctxt->lqc_handler) {
 174                 lqs->lqs_last_bshrink  = 0;
 175                 lqs->lqs_last_ishrink  = 0;
 176         }
 177         lqs_initref(lqs);
 178         rc = lustre_hash_additem_unique(LQC_HASH_BODY(qctxt),
 179                                         oqaq_tmp, &lqs->lqs_hash);
 180         LQS_DEBUG(lqs, "create lqs\n");
 181         if (!rc) {
 182                 lqs_getref(lqs);
 183                 *lqs_return = lqs;
 184         }
 185  out:
 186         if (rc && lqs)
 187                 OBD_FREE_PTR(lqs);
 188         if (!oqaq)
 189                 OBD_FREE_PTR(oqaq_tmp);
 190         RETURN(rc);
 191 }
 192
 193 int quota_adjust_slave_lqs(struct quota_adjust_qunit *oqaq, struct
 194                           lustre_quota_ctxt *qctxt)
 195 {
 196         struct lustre_qunit_size *lqs = NULL;
 197         unsigned long *lbunit, *liunit, *lbtune, *litune;
 198         signed long b_tmp = 0, i_tmp = 0;
 199         static cfs_time_t time_limit = 0;
 200         int rc = 0;
 201         ENTRY;
 202
 203         if (OBD_FAIL_CHECK(OBD_FAIL_QUOTA_WITHOUT_CHANGE_QS))
 204                 RETURN(0);
 205
 206         LASSERT(qctxt);
 207  search_lqs:
 208         rc = quota_search_lqs(NULL, oqaq, qctxt, &lqs);
 209
 210         /* deleting the lqs, because a user sets lfs quota 0 0 0 0  */
 211         if (!oqaq->qaq_bunit_sz && !oqaq->qaq_iunit_sz && QAQ_IS_ADJBLK(oqaq) &&
 212             QAQ_IS_ADJINO(oqaq)) {
 213                 if (lqs) {
 214                         LQS_DEBUG(lqs, "release lqs\n");
 215                         /* this is for quota_search_lqs */
 216                         lqs_putref(lqs);
 217                         /* this is for deleting this lqs */
 218                         lqs_putref(lqs);
 219                 }
 220                 RETURN(rc);
 221         }
 222
 223         if (!lqs) {
 224                 rc = quota_create_lqs(NULL, oqaq, qctxt, &lqs);
 225                 if (rc == -EALREADY)
 226                         goto search_lqs;
 227                 if (rc < 0)
 228                         RETURN(rc);
 229         }
 230
 231         lbunit = &lqs->lqs_bunit_sz;
 232         liunit = &lqs->lqs_iunit_sz;
 233         lbtune = &lqs->lqs_btune_sz;
 234         litune = &lqs->lqs_itune_sz;
 235
 236         spin_lock(&lqs->lqs_lock);
 237         CDEBUG(D_QUOTA, "before: bunit: %lu, iunit: %lu.\n", *lbunit, *liunit);
 238         /* adjust the slave's block qunit size */
 239         if (QAQ_IS_ADJBLK(oqaq)) {
 240                 cfs_duration_t sec = cfs_time_seconds(qctxt->lqc_switch_seconds);
 241
 242                 b_tmp = *lbunit - oqaq->qaq_bunit_sz;
 243
 244                 if (qctxt->lqc_handler && b_tmp > 0)
 245                         lqs->lqs_last_bshrink = cfs_time_current();
 246
 247                 if (qctxt->lqc_handler && b_tmp < 0) {
 248                         time_limit = cfs_time_add(lqs->lqs_last_bshrink, sec);
 249                         if (!lqs->lqs_last_bshrink ||
 250                             cfs_time_after(cfs_time_current(), time_limit)) {
 251                                 *lbunit = oqaq->qaq_bunit_sz;
 252                                 *lbtune = (*lbunit) / 2;
 253                         } else {
 254                                 b_tmp = 0;
 255                         }
 256                 } else {
 257                         *lbunit = oqaq->qaq_bunit_sz;
 258                         *lbtune = (*lbunit) / 2;
 259                 }
 260         }
 261
 262         /* adjust the slave's file qunit size */
 263         if (QAQ_IS_ADJINO(oqaq)) {
 264                 i_tmp = *liunit - oqaq->qaq_iunit_sz;
 265
 266                 if (qctxt->lqc_handler && i_tmp > 0)
 267                         lqs->lqs_last_ishrink  = cfs_time_current();
 268
 269                 if (qctxt->lqc_handler && i_tmp < 0) {
 270                         time_limit = cfs_time_add(lqs->lqs_last_ishrink,
 271                                                   cfs_time_seconds(qctxt->
 272                                                   lqc_switch_seconds));
 273                         if (!lqs->lqs_last_ishrink ||
 274                             cfs_time_after(cfs_time_current(), time_limit)) {
 275                                 *liunit = oqaq->qaq_iunit_sz;
 276                                 *litune = (*liunit) / 2;
 277                         } else {
 278                                 i_tmp = 0;
 279                         }
 280                 } else {
 281                         *liunit = oqaq->qaq_iunit_sz;
 282                         *litune = (*liunit) / 2;
 283                 }
 284         }
 285         CDEBUG(D_QUOTA, "after: bunit: %lu, iunit: %lu.\n", *lbunit, *liunit);
 286         spin_unlock(&lqs->lqs_lock);
 287
 288         lqs_putref(lqs);
 289
 290         if (b_tmp > 0)
 291                 rc |= LQS_BLK_DECREASE;
 292         else if (b_tmp < 0)
 293                 rc |= LQS_BLK_INCREASE;
 294
 295         if (i_tmp > 0)
 296                 rc |= LQS_INO_DECREASE;
 297         else if (i_tmp < 0)
 298                 rc |= LQS_INO_INCREASE;
 299
 300         RETURN(rc);
 301 }
 302
 303 int filter_quota_adjust_qunit(struct obd_export *exp, struct
 304                               quota_adjust_qunit *oqaq)
 305 {
 306         struct obd_device *obd = exp->exp_obd;
 307         struct lustre_quota_ctxt *qctxt = &obd->u.obt.obt_qctxt;
 308         unsigned int uid = 0, gid = 0;
 309         int rc = 0;
 310         ENTRY;
 311
 312         LASSERT(oqaq);
 313         LASSERT(QAQ_IS_ADJBLK(oqaq));
 314         rc = quota_adjust_slave_lqs(oqaq, qctxt);
 315         if (rc < 0) {
 316                 CERROR("adjust mds slave's qunit size failed!(rc:%d)\n", rc);
 317                 RETURN(rc);
 318         }
 319         if (QAQ_IS_GRP(oqaq))
 320                 gid = oqaq->qaq_id;
 321         else
 322                 uid = oqaq->qaq_id;
 323
 324         if (rc > 0) {
 325                 rc = qctxt_adjust_qunit(obd, qctxt, uid, gid, 1, 0);
 326                 if (rc)
 327                         CERROR("slave adjust block quota failed!(rc:%d)\n", rc);
 328         }
 329         RETURN(rc);
 330 }
 331 #endif /* __KERNEL__ */
 332
 333 int client_quota_adjust_qunit(struct obd_export *exp, struct
 334                               quota_adjust_qunit *oqaq)
 335 {
 336         struct ptlrpc_request *req;
 337         struct quota_adjust_qunit *oqa;
 338         int size[2] = { sizeof(struct ptlrpc_body), sizeof(*oqaq) };
 339         int rc = 0;
 340         ENTRY;
 341
 342         /* client don't support this kind of operation, abort it */
 343         if (!(exp->exp_connect_flags & OBD_CONNECT_CHANGE_QS)||
 344             OBD_FAIL_CHECK(OBD_FAIL_QUOTA_WITHOUT_CHANGE_QS)) {
 345                 CDEBUG(D_QUOTA, "osc: %s don't support change qunit size\n",
 346                        exp->exp_obd->obd_name);
 347                 RETURN(rc);
 348         }
 349         if (strcmp(exp->exp_obd->obd_type->typ_name, LUSTRE_OSC_NAME))
 350                 RETURN(-EINVAL);
 351
 352         req = ptlrpc_prep_req(class_exp2cliimp(exp), LUSTRE_OST_VERSION,
 353                               OST_QUOTA_ADJUST_QUNIT, 2, size, NULL);
 354         if (!req)
 355                 GOTO(out, rc = -ENOMEM);
 356
 357         oqa = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, sizeof(*oqaq));
 358         *oqa = *oqaq;
 359
 360         ptlrpc_req_set_repsize(req, 2, size);
 361
 362         rc = ptlrpc_queue_wait(req);
 363         if (rc) {
 364                 CERROR("%s: %s failed: rc = %d\n", exp->exp_obd->obd_name,
 365                        __FUNCTION__, rc);
 366                 GOTO(out, rc);
 367         }
 368         ptlrpc_req_finished(req);
 369 out:
 370         RETURN (rc);
 371 }
 372
 373 int lov_quota_adjust_qunit(struct obd_export *exp, struct
 374                            quota_adjust_qunit *oqaq)
 375 {
 376         struct obd_device *obd = class_exp2obd(exp);
 377         struct lov_obd *lov = &obd->u.lov;
 378         int i, rc = 0;
 379         ENTRY;
 380
 381         if (!QAQ_IS_ADJBLK(oqaq)) {
 382                 CERROR("bad qaq_flags %x for lov obd.\n", oqaq->qaq_flags);
 383                 RETURN(-EFAULT);
 384         }
 385
 386         for (i = 0; i < lov->desc.ld_tgt_count; i++) {
 387                 int err;
 388
 389                 if (!lov->lov_tgts[i]->ltd_active) {
 390                         CDEBUG(D_HA, "ost %d is inactive\n", i);
 391                         continue;
 392                 }
 393
 394                 err = obd_quota_adjust_qunit(lov->lov_tgts[i]->ltd_exp, oqaq);
 395                 if (err) {
 396                         if (lov->lov_tgts[i]->ltd_active && !rc)
 397                                 rc = err;
 398                         continue;
 399                 }
 400         }
 401         RETURN(rc);
 402 }