lustre/ost/ost_handler.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  * lustre/ost/ost_handler.c
  37  *
  38  * Author: Peter J. Braam <braam@clusterfs.com>
  39  * Author: Phil Schwan <phil@clusterfs.com>
  40  */
  41
  42 #ifndef EXPORT_SYMTAB
  43 # define EXPORT_SYMTAB
  44 #endif
  45 #define DEBUG_SUBSYSTEM S_OST
  46
  47 #include <linux/module.h>
  48 #include <obd_cksum.h>
  49 #include <obd_ost.h>
  50 #include <lustre_net.h>
  51 #include <lustre_dlm.h>
  52 #include <lustre_export.h>
  53 #include <lustre_debug.h>
  54 #include <linux/init.h>
  55 #include <lprocfs_status.h>
  56 #include <libcfs/list.h>
  57 #include <lustre_quota.h>
  58 #include "ost_internal.h"
  59
  60 static int oss_num_threads;
  61 CFS_MODULE_PARM(oss_num_threads, "i", int, 0444,
  62                 "number of OSS service threads to start");
  63
  64 static int ost_num_threads;
  65 CFS_MODULE_PARM(ost_num_threads, "i", int, 0444,
  66                 "number of OST service threads to start (deprecated)");
  67
  68 static int oss_num_create_threads;
  69 CFS_MODULE_PARM(oss_num_create_threads, "i", int, 0444,
  70                 "number of OSS create threads to start");
  71
  72 void oti_to_request(struct obd_trans_info *oti, struct ptlrpc_request *req)
  73 {
  74         struct oti_req_ack_lock *ack_lock;
  75         int i;
  76
  77         if (oti == NULL)
  78                 return;
  79
  80         if (req->rq_repmsg)
  81                 lustre_msg_set_transno(req->rq_repmsg, oti->oti_transno);
  82         req->rq_transno = oti->oti_transno;
  83
  84         /* XXX 4 == entries in oti_ack_locks??? */
  85         for (ack_lock = oti->oti_ack_locks, i = 0; i < 4; i++, ack_lock++) {
  86                 if (!ack_lock->mode)
  87                         break;
  88                 /* XXX not even calling target_send_reply in some cases... */
  89                 ptlrpc_save_lock (req, &ack_lock->lock, ack_lock->mode);
  90         }
  91 }
  92
  93 static int ost_destroy(struct obd_export *exp, struct ptlrpc_request *req,
  94                        struct obd_trans_info *oti)
  95 {
  96         struct ost_body *body, *repbody;
  97         __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
  98         int rc;
  99         ENTRY;
 100
 101         body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
 102                                   lustre_swab_ost_body);
 103         if (body == NULL)
 104                 RETURN(-EFAULT);
 105
 106         if (body->oa.o_id == 0)
 107                 RETURN(-EPROTO);
 108
 109         if (lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF + 1)) {
 110                 struct ldlm_request *dlm;
 111                 dlm = lustre_swab_reqbuf(req, REQ_REC_OFF + 1, sizeof(*dlm),
 112                                          lustre_swab_ldlm_request);
 113                 if (dlm == NULL)
 114                         RETURN (-EFAULT);
 115                 ldlm_request_cancel(req, dlm, 0);
 116         }
 117
 118         rc = lustre_pack_reply(req, 2, size, NULL);
 119         if (rc)
 120                 RETURN(rc);
 121
 122         if (body->oa.o_valid & OBD_MD_FLCOOKIE)
 123                 oti->oti_logcookies = obdo_logcookie(&body->oa);
 124         repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 125                                  sizeof(*repbody));
 126         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
 127         req->rq_status = obd_destroy(exp, &body->oa, NULL, oti, NULL);
 128         RETURN(0);
 129 }
 130
 131 static int ost_getattr(struct obd_export *exp, struct ptlrpc_request *req)
 132 {
 133         struct ost_body *body, *repbody;
 134         struct obd_info oinfo = { { { 0 } } };
 135         __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
 136         int rc;
 137         ENTRY;
 138
 139         body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
 140                                   lustre_swab_ost_body);
 141         if (body == NULL)
 142                 RETURN(-EFAULT);
 143
 144         rc = lustre_pack_reply(req, 2, size, NULL);
 145         if (rc)
 146                 RETURN(rc);
 147
 148         repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 149                                  sizeof(*repbody));
 150         repbody->oa = body->oa;
 151
 152         oinfo.oi_oa = &repbody->oa;
 153         if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA)
 154                 oinfo.oi_capa = lustre_unpack_capa(req->rq_reqmsg,
 155                                                    REQ_REC_OFF + 1);
 156         req->rq_status = obd_getattr(exp, &oinfo);
 157         RETURN(0);
 158 }
 159
 160 static int ost_statfs(struct ptlrpc_request *req)
 161 {
 162         struct obd_statfs *osfs;
 163         __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*osfs) };
 164         int rc;
 165         ENTRY;
 166
 167         rc = lustre_pack_reply(req, 2, size, NULL);
 168         if (rc)
 169                 RETURN(rc);
 170
 171         osfs = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*osfs));
 172
 173         req->rq_status = obd_statfs(req->rq_export->exp_obd, osfs,
 174                                     cfs_time_current_64() - HZ, 0);
 175         if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
 176                 osfs->os_bfree = osfs->os_bavail = 64;
 177         if (req->rq_status != 0)
 178                 CERROR("ost: statfs failed: rc %d\n", req->rq_status);
 179
 180         RETURN(0);
 181 }
 182
 183 static int ost_create(struct obd_export *exp, struct ptlrpc_request *req,
 184                       struct obd_trans_info *oti)
 185 {
 186         struct ost_body *body, *repbody;
 187         __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
 188         int rc;
 189         ENTRY;
 190
 191         body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
 192                                   lustre_swab_ost_body);
 193         if (body == NULL)
 194                 RETURN(-EFAULT);
 195
 196         rc = lustre_pack_reply(req, 2, size, NULL);
 197         if (rc)
 198                 RETURN(rc);
 199
 200         repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 201                                  sizeof(*repbody));
 202         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
 203         oti->oti_logcookies = obdo_logcookie(&repbody->oa);
 204         req->rq_status = obd_create(exp, &repbody->oa, NULL, oti);
 205         //obd_log_cancel(conn, NULL, 1, oti->oti_logcookies, 0);
 206         RETURN(0);
 207 }
 208
 209 /*
 210  * Helper function for ost_punch(): if asked by client, acquire [size, EOF]
 211  * lock on the file being truncated.
 212  */
 213 static int ost_punch_lock_get(struct obd_export *exp, struct obdo *oa,
 214                               struct lustre_handle *lh)
 215 {
 216         int flags;
 217         struct ldlm_res_id res_id;
 218         ldlm_policy_data_t policy;
 219         __u64 start;
 220         __u64 finis;
 221
 222         ENTRY;
 223
 224         osc_build_res_name(oa->o_id, oa->o_gr, &res_id);
 225         LASSERT(!lustre_handle_is_used(lh));
 226
 227         if (!(oa->o_valid & OBD_MD_FLFLAGS) ||
 228             !(oa->o_flags & OBD_FL_TRUNCLOCK))
 229                 RETURN(0);
 230
 231         CDEBUG(D_INODE, "OST-side truncate lock.\n");
 232
 233         start = oa->o_size;
 234         finis = start + oa->o_blocks;
 235
 236         /*
 237          * standard truncate optimization: if file body is completely
 238          * destroyed, don't send data back to the server.
 239          */
 240         flags = (start == 0) ? LDLM_AST_DISCARD_DATA : 0;
 241
 242         policy.l_extent.start = start & CFS_PAGE_MASK;
 243
 244         /*
 245          * If ->o_blocks is EOF it means "lock till the end of the
 246          * file". Otherwise, it's size of a hole being punched (in bytes)
 247          */
 248         if (oa->o_blocks == OBD_OBJECT_EOF || finis < start)
 249                 policy.l_extent.end = OBD_OBJECT_EOF;
 250         else
 251                 policy.l_extent.end = finis | ~CFS_PAGE_MASK;
 252
 253         RETURN(ldlm_cli_enqueue_local(exp->exp_obd->obd_namespace, &res_id,
 254                                       LDLM_EXTENT, &policy, LCK_PW, &flags,
 255                                       ldlm_blocking_ast, ldlm_completion_ast,
 256                                       ldlm_glimpse_ast, NULL, 0, NULL, lh));
 257 }
 258
 259 /*
 260  * Helper function for ost_punch(): release lock acquired by
 261  * ost_punch_lock_get(), if any.
 262  */
 263 static void ost_punch_lock_put(struct obd_export *exp, struct obdo *oa,
 264                                struct lustre_handle *lh)
 265 {
 266         ENTRY;
 267         if (lustre_handle_is_used(lh))
 268                 ldlm_lock_decref(lh, LCK_PW);
 269         EXIT;
 270 }
 271
 272 static int ost_punch(struct obd_export *exp, struct ptlrpc_request *req,
 273                      struct obd_trans_info *oti)
 274 {
 275         struct obd_info oinfo = { { { 0 } } };
 276         struct ost_body *body, *repbody;
 277         __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
 278         int rc;
 279         struct lustre_handle lh = {0,};
 280         ENTRY;
 281
 282         /* check that we do support OBD_CONNECT_TRUNCLOCK. */
 283         CLASSERT(OST_CONNECT_SUPPORTED & OBD_CONNECT_TRUNCLOCK);
 284
 285         body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
 286                                   lustre_swab_ost_body);
 287         if (body == NULL)
 288                 RETURN(-EFAULT);
 289
 290         oinfo.oi_oa = &body->oa;
 291         oinfo.oi_policy.l_extent.start = oinfo.oi_oa->o_size;
 292         oinfo.oi_policy.l_extent.end = oinfo.oi_oa->o_blocks;
 293
 294         if ((oinfo.oi_oa->o_valid & (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS)) !=
 295             (OBD_MD_FLSIZE | OBD_MD_FLBLOCKS))
 296                 RETURN(-EINVAL);
 297
 298         rc = lustre_pack_reply(req, 2, size, NULL);
 299         if (rc)
 300                 RETURN(rc);
 301
 302         repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 303                                  sizeof(*repbody));
 304         rc = ost_punch_lock_get(exp, oinfo.oi_oa, &lh);
 305         if (rc == 0) {
 306                 if (oinfo.oi_oa->o_valid & OBD_MD_FLFLAGS &&
 307                     oinfo.oi_oa->o_flags == OBD_FL_TRUNCLOCK)
 308                         /*
 309                          * If OBD_FL_TRUNCLOCK is the only bit set in
 310                          * ->o_flags, clear OBD_MD_FLFLAGS to avoid falling
 311                          * through filter_setattr() to filter_iocontrol().
 312                          */
 313                         oinfo.oi_oa->o_valid &= ~OBD_MD_FLFLAGS;
 314
 315                 if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA)
 316                         oinfo.oi_capa = lustre_unpack_capa(req->rq_reqmsg,
 317                                                            REQ_REC_OFF + 1);
 318                 req->rq_status = obd_punch(exp, &oinfo, oti, NULL);
 319                 ost_punch_lock_put(exp, oinfo.oi_oa, &lh);
 320         }
 321         repbody->oa = *oinfo.oi_oa;
 322         RETURN(rc);
 323 }
 324
 325 static int ost_sync(struct obd_export *exp, struct ptlrpc_request *req)
 326 {
 327         struct ost_body *body, *repbody;
 328         struct lustre_capa *capa = NULL;
 329         __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
 330         int rc;
 331         ENTRY;
 332
 333         body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
 334                                   lustre_swab_ost_body);
 335         if (body == NULL)
 336                 RETURN(-EFAULT);
 337
 338         if (body->oa.o_valid & OBD_MD_FLOSSCAPA)
 339                 capa = lustre_unpack_capa(req->rq_reqmsg, REQ_REC_OFF + 1);
 340
 341         rc = lustre_pack_reply(req, 2, size, NULL);
 342         if (rc)
 343                 RETURN(rc);
 344
 345         repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 346                                  sizeof(*repbody));
 347         memcpy(&repbody->oa, &body->oa, sizeof(body->oa));
 348         req->rq_status = obd_sync(exp, &repbody->oa, NULL, repbody->oa.o_size,
 349                                   repbody->oa.o_blocks, capa);
 350         RETURN(0);
 351 }
 352
 353 static int ost_setattr(struct obd_export *exp, struct ptlrpc_request *req,
 354                        struct obd_trans_info *oti)
 355 {
 356         struct ost_body *body, *repbody;
 357         __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*repbody) };
 358         int rc;
 359         struct obd_info oinfo = { { { 0 } } };
 360         ENTRY;
 361
 362         body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
 363                                   lustre_swab_ost_body);
 364         if (body == NULL)
 365                 RETURN(-EFAULT);
 366
 367         rc = lustre_pack_reply(req, 2, size, NULL);
 368         if (rc)
 369                 RETURN(rc);
 370
 371         repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 372                                  sizeof(*repbody));
 373         repbody->oa = body->oa;
 374
 375         oinfo.oi_oa = &repbody->oa;
 376         if (oinfo.oi_oa->o_valid & OBD_MD_FLOSSCAPA)
 377                 oinfo.oi_capa = lustre_unpack_capa(req->rq_reqmsg,
 378                                                    REQ_REC_OFF + 1);
 379         req->rq_status = obd_setattr(exp, &oinfo, oti);
 380         RETURN(0);
 381 }
 382
 383 static int ost_bulk_timeout(void *data)
 384 {
 385         ENTRY;
 386         /* We don't fail the connection here, because having the export
 387          * killed makes the (vital) call to commitrw very sad.
 388          */
 389         RETURN(1);
 390 }
 391
 392 static int get_per_page_niobufs(struct obd_ioobj *ioo, int nioo,
 393                                 struct niobuf_remote *rnb, int nrnb,
 394                                 struct niobuf_remote **pp_rnbp)
 395 {
 396         /* Copy a remote niobuf, splitting it into page-sized chunks
 397          * and setting ioo[i].ioo_bufcnt accordingly */
 398         struct niobuf_remote *pp_rnb;
 399         int   i;
 400         int   j;
 401         int   page;
 402         int   rnbidx = 0;
 403         int   npages = 0;
 404
 405         /*
 406          * array of sufficient size already preallocated by caller
 407          */
 408         LASSERT(pp_rnbp != NULL);
 409         LASSERT(*pp_rnbp != NULL);
 410
 411         /* first count and check the number of pages required */
 412         for (i = 0; i < nioo; i++)
 413                 for (j = 0; j < ioo->ioo_bufcnt; j++, rnbidx++) {
 414                         obd_off offset = rnb[rnbidx].offset;
 415                         obd_off p0 = offset >> CFS_PAGE_SHIFT;
 416                         obd_off pn = (offset + rnb[rnbidx].len - 1) >>
 417                                      CFS_PAGE_SHIFT;
 418
 419                         LASSERT(rnbidx < nrnb);
 420
 421                         npages += (pn + 1 - p0);
 422
 423                         if (rnb[rnbidx].len == 0) {
 424                                 CERROR("zero len BRW: obj %d objid "LPX64
 425                                        " buf %u\n", i, ioo[i].ioo_id, j);
 426                                 return -EINVAL;
 427                         }
 428                         if (j > 0 &&
 429                             rnb[rnbidx].offset <= rnb[rnbidx-1].offset) {
 430                                 CERROR("unordered BRW: obj %d objid "LPX64
 431                                        " buf %u offset "LPX64" <= "LPX64"\n",
 432                                        i, ioo[i].ioo_id, j, rnb[rnbidx].offset,
 433                                        rnb[rnbidx].offset);
 434                                 return -EINVAL;
 435                         }
 436                 }
 437
 438         LASSERT(rnbidx == nrnb);
 439
 440         if (npages == nrnb) {       /* all niobufs are for single pages */
 441                 *pp_rnbp = rnb;
 442                 return npages;
 443         }
 444
 445         pp_rnb = *pp_rnbp;
 446
 447         /* now do the actual split */
 448         page = rnbidx = 0;
 449         for (i = 0; i < nioo; i++) {
 450                 int  obj_pages = 0;
 451
 452                 for (j = 0; j < ioo[i].ioo_bufcnt; j++, rnbidx++) {
 453                         obd_off off = rnb[rnbidx].offset;
 454                         int     nob = rnb[rnbidx].len;
 455
 456                         LASSERT(rnbidx < nrnb);
 457                         do {
 458                                 obd_off  poff = off & ~CFS_PAGE_MASK;
 459                                 int      pnob = (poff + nob > CFS_PAGE_SIZE) ?
 460                                                 CFS_PAGE_SIZE - poff : nob;
 461
 462                                 LASSERT(page < npages);
 463                                 pp_rnb[page].len = pnob;
 464                                 pp_rnb[page].offset = off;
 465                                 pp_rnb[page].flags = rnb[rnbidx].flags;
 466
 467                                 CDEBUG(0, "   obj %d id "LPX64
 468                                        "page %d(%d) "LPX64" for %d, flg %x\n",
 469                                        i, ioo[i].ioo_id, obj_pages, page,
 470                                        pp_rnb[page].offset, pp_rnb[page].len,
 471                                        pp_rnb[page].flags);
 472                                 page++;
 473                                 obj_pages++;
 474
 475                                 off += pnob;
 476                                 nob -= pnob;
 477                         } while (nob > 0);
 478                         LASSERT(nob == 0);
 479                 }
 480                 ioo[i].ioo_bufcnt = obj_pages;
 481         }
 482         LASSERT(page == npages);
 483
 484         return npages;
 485 }
 486
 487 static __u32 ost_checksum_bulk(struct ptlrpc_bulk_desc *desc, int opc,
 488                                cksum_type_t cksum_type)
 489 {
 490         __u32 cksum;
 491         int i;
 492
 493         cksum = init_checksum(cksum_type);
 494         for (i = 0; i < desc->bd_iov_count; i++) {
 495                 struct page *page = desc->bd_iov[i].kiov_page;
 496                 int off = desc->bd_iov[i].kiov_offset & ~CFS_PAGE_MASK;
 497                 char *ptr = kmap(page) + off;
 498                 int len = desc->bd_iov[i].kiov_len;
 499
 500                 /* corrupt the data before we compute the checksum, to
 501                  * simulate a client->OST data error */
 502                 if (i == 0 && opc == OST_WRITE &&
 503                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_RECEIVE))
 504                         memcpy(ptr, "bad3", min(4, len));
 505                 cksum = compute_checksum(cksum, ptr, len, cksum_type);
 506                 /* corrupt the data after we compute the checksum, to
 507                  * simulate an OST->client data error */
 508                 if (i == 0 && opc == OST_READ &&
 509                     OBD_FAIL_CHECK(OBD_FAIL_OST_CHECKSUM_SEND))
 510                         memcpy(ptr, "bad4", min(4, len));
 511                 kunmap(page);
 512         }
 513
 514         return cksum;
 515 }
 516
 517 /*
 518  * populate @nio by @nrpages pages from per-thread page pool
 519  */
 520 static void ost_nio_pages_get(struct ptlrpc_request *req,
 521                               struct niobuf_local *nio, int nrpages)
 522 {
 523         int i;
 524         struct ost_thread_local_cache *tls;
 525
 526         ENTRY;
 527
 528         LASSERT(nrpages <= OST_THREAD_POOL_SIZE);
 529         LASSERT(req != NULL);
 530         LASSERT(req->rq_svc_thread != NULL);
 531
 532         tls = ost_tls(req);
 533         LASSERT(tls != NULL);
 534
 535         memset(nio, 0, nrpages * sizeof *nio);
 536         for (i = 0; i < nrpages; ++ i) {
 537                 struct page *page;
 538
 539                 page = tls->page[i];
 540                 LASSERT(page != NULL);
 541                 POISON_PAGE(page, 0xf1);
 542                 nio[i].page = page;
 543                 LL_CDEBUG_PAGE(D_INFO, page, "%d\n", i);
 544         }
 545         EXIT;
 546 }
 547
 548 /*
 549  * Dual for ost_nio_pages_get(). Poison pages in pool for debugging
 550  */
 551 static void ost_nio_pages_put(struct ptlrpc_request *req,
 552                               struct niobuf_local *nio, int nrpages)
 553 {
 554         int i;
 555
 556         ENTRY;
 557
 558         LASSERT(nrpages <= OST_THREAD_POOL_SIZE);
 559
 560         for (i = 0; i < nrpages; ++ i)
 561                 POISON_PAGE(nio[i].page, 0xf2);
 562         EXIT;
 563 }
 564
 565 static int ost_brw_lock_get(int mode, struct obd_export *exp,
 566                             struct obd_ioobj *obj, struct niobuf_remote *nb,
 567                             struct lustre_handle *lh)
 568 {
 569         int flags                 = 0;
 570         int nrbufs                = obj->ioo_bufcnt;
 571         struct ldlm_res_id res_id;
 572         ldlm_policy_data_t policy;
 573         int i;
 574         ENTRY;
 575
 576         osc_build_res_name(obj->ioo_id, obj->ioo_gr, &res_id);
 577         LASSERT(mode == LCK_PR || mode == LCK_PW);
 578         LASSERT(!lustre_handle_is_used(lh));
 579
 580         if (nrbufs == 0 || !(nb[0].flags & OBD_BRW_SRVLOCK))
 581                 RETURN(0);
 582
 583         /* EXPENSIVE ASSERTION */
 584         for (i = 1; i < nrbufs; i ++)
 585                 LASSERT((nb[0].flags & OBD_BRW_SRVLOCK) ==
 586                         (nb[i].flags & OBD_BRW_SRVLOCK));
 587
 588         policy.l_extent.start = nb[0].offset & CFS_PAGE_MASK;
 589         policy.l_extent.end   = (nb[nrbufs - 1].offset +
 590                                  nb[nrbufs - 1].len - 1) | ~CFS_PAGE_MASK;
 591
 592         RETURN(ldlm_cli_enqueue_local(exp->exp_obd->obd_namespace, &res_id,
 593                                       LDLM_EXTENT, &policy, mode, &flags,
 594                                       ldlm_blocking_ast, ldlm_completion_ast,
 595                                       ldlm_glimpse_ast, NULL, 0, NULL, lh));
 596 }
 597
 598 static void ost_brw_lock_put(int mode,
 599                              struct obd_ioobj *obj, struct niobuf_remote *niob,
 600                              struct lustre_handle *lh)
 601 {
 602         ENTRY;
 603         LASSERT(mode == LCK_PR || mode == LCK_PW);
 604         LASSERT((obj->ioo_bufcnt > 0 && (niob[0].flags & OBD_BRW_SRVLOCK)) ==
 605                 lustre_handle_is_used(lh));
 606         if (lustre_handle_is_used(lh))
 607                 ldlm_lock_decref(lh, mode);
 608         EXIT;
 609 }
 610
 611 struct ost_prolong_data {
 612         struct obd_export *opd_exp;
 613         ldlm_policy_data_t opd_policy;
 614         ldlm_mode_t opd_mode;
 615 };
 616
 617 static int ost_prolong_locks_iter(struct ldlm_lock *lock, void *data)
 618 {
 619         struct ost_prolong_data *opd = data;
 620
 621         LASSERT(lock->l_resource->lr_type == LDLM_EXTENT);
 622
 623         if (lock->l_req_mode != lock->l_granted_mode) {
 624                 /* scan granted locks only */
 625                 return LDLM_ITER_STOP;
 626         }
 627
 628         if (lock->l_export != opd->opd_exp) {
 629                 /* prolong locks only for given client */
 630                 return LDLM_ITER_CONTINUE;
 631         }
 632
 633         if (!(lock->l_granted_mode & opd->opd_mode)) {
 634                 /* we aren't interesting in all type of locks */
 635                 return LDLM_ITER_CONTINUE;
 636         }
 637
 638         if (lock->l_policy_data.l_extent.end < opd->opd_policy.l_extent.start ||
 639             lock->l_policy_data.l_extent.start > opd->opd_policy.l_extent.end) {
 640                 /* the request doesn't cross the lock, skip it */
 641                 return LDLM_ITER_CONTINUE;
 642         }
 643
 644         if (!(lock->l_flags & LDLM_FL_AST_SENT)) {
 645                 /* ignore locks not being cancelled */
 646                 return LDLM_ITER_CONTINUE;
 647         }
 648
 649         /* OK. this is a possible lock the user holds doing I/O
 650          * let's refresh eviction timer for it */
 651         ldlm_refresh_waiting_lock(lock);
 652
 653         return LDLM_ITER_CONTINUE;
 654 }
 655
 656 static void ost_prolong_locks(struct obd_export *exp, struct obd_ioobj *obj,
 657                               struct niobuf_remote *nb, ldlm_mode_t mode)
 658 {
 659         struct ldlm_res_id res_id;
 660         int nrbufs = obj->ioo_bufcnt;
 661         struct ost_prolong_data opd;
 662         ENTRY;
 663
 664         osc_build_res_name(obj->ioo_id, obj->ioo_gr, &res_id);
 665
 666         opd.opd_mode = mode;
 667         opd.opd_exp = exp;
 668         opd.opd_policy.l_extent.start = nb[0].offset & CFS_PAGE_MASK;
 669         opd.opd_policy.l_extent.end = (nb[nrbufs - 1].offset +
 670                                        nb[nrbufs - 1].len - 1) | ~CFS_PAGE_MASK;
 671
 672         CDEBUG(D_DLMTRACE,"refresh locks: "LPU64"/"LPU64" ("LPU64"->"LPU64")\n",
 673                res_id.name[0], res_id.name[1], opd.opd_policy.l_extent.start,
 674                opd.opd_policy.l_extent.end);
 675         ldlm_resource_iterate(exp->exp_obd->obd_namespace, &res_id,
 676                               ost_prolong_locks_iter, &opd);
 677 }
 678
 679 static int ost_brw_read(struct ptlrpc_request *req, struct obd_trans_info *oti)
 680 {
 681         struct ptlrpc_bulk_desc *desc;
 682         struct obd_export *exp = req->rq_export;
 683         struct niobuf_remote *remote_nb;
 684         struct niobuf_remote *pp_rnb = NULL;
 685         struct niobuf_local *local_nb;
 686         struct obd_ioobj *ioo;
 687         struct ost_body *body, *repbody;
 688         struct lustre_capa *capa = NULL;
 689         struct l_wait_info lwi;
 690         struct lustre_handle lockh = { 0 };
 691         __u32  size[2] = { sizeof(struct ptlrpc_body), sizeof(*body) };
 692         int niocount, npages, nob = 0, rc, i;
 693         int no_reply = 0;
 694         ENTRY;
 695
 696         req->rq_bulk_read = 1;
 697
 698         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_READ_BULK))
 699                 GOTO(out, rc = -EIO);
 700
 701         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, (obd_timeout + 1) / 4);
 702
 703         /* Check if there is eviction in progress, and if so, wait for it to
 704          * finish */
 705         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
 706                 lwi = LWI_INTR(NULL, NULL); // We do not care how long it takes
 707                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
 708                         !atomic_read(&exp->exp_obd->obd_evict_inprogress),
 709                         &lwi);
 710         }
 711         if (exp->exp_failed)
 712                 GOTO(out, rc = -ENOTCONN);
 713
 714         body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
 715                                   lustre_swab_ost_body);
 716         if (body == NULL) {
 717                 CERROR("Missing/short ost_body\n");
 718                 GOTO(out, rc = -EFAULT);
 719         }
 720
 721         ioo = lustre_swab_reqbuf(req, REQ_REC_OFF + 1, sizeof(*ioo),
 722                                  lustre_swab_obd_ioobj);
 723         if (ioo == NULL) {
 724                 CERROR("Missing/short ioobj\n");
 725                 GOTO(out, rc = -EFAULT);
 726         }
 727
 728         niocount = ioo->ioo_bufcnt;
 729         if (niocount > PTLRPC_MAX_BRW_PAGES) {
 730                 DEBUG_REQ(D_ERROR, req, "bulk has too many pages (%d)",
 731                           niocount);
 732                 GOTO(out, rc = -EFAULT);
 733         }
 734
 735         remote_nb = lustre_swab_reqbuf(req, REQ_REC_OFF + 2,
 736                                        niocount * sizeof(*remote_nb),
 737                                        lustre_swab_niobuf_remote);
 738         if (remote_nb == NULL) {
 739                 CERROR("Missing/short niobuf\n");
 740                 GOTO(out, rc = -EFAULT);
 741         }
 742         if (lustre_msg_swabbed(req->rq_reqmsg)) { /* swab remaining niobufs */
 743                 for (i = 1; i < niocount; i++)
 744                         lustre_swab_niobuf_remote (&remote_nb[i]);
 745         }
 746
 747         if (body->oa.o_valid & OBD_MD_FLOSSCAPA)
 748                 capa = lustre_unpack_capa(req->rq_reqmsg, REQ_REC_OFF + 3);
 749
 750         rc = lustre_pack_reply(req, 2, size, NULL);
 751         if (rc)
 752                 GOTO(out, rc);
 753
 754         /*
 755          * Per-thread array of struct niobuf_{local,remote}'s was allocated by
 756          * ost_thread_init().
 757          */
 758         local_nb = ost_tls(req)->local;
 759         pp_rnb   = ost_tls(req)->remote;
 760
 761         /* FIXME all niobuf splitting should be done in obdfilter if needed */
 762         /* CAVEAT EMPTOR this sets ioo->ioo_bufcnt to # pages */
 763         npages = get_per_page_niobufs(ioo, 1, remote_nb, niocount, &pp_rnb);
 764         if (npages < 0)
 765                 GOTO(out, rc = npages);
 766
 767         LASSERT(npages <= OST_THREAD_POOL_SIZE);
 768
 769         ost_nio_pages_get(req, local_nb, npages);
 770
 771         desc = ptlrpc_prep_bulk_exp(req, npages,
 772                                      BULK_PUT_SOURCE, OST_BULK_PORTAL);
 773         if (desc == NULL)
 774                 GOTO(out, rc = -ENOMEM);
 775
 776         rc = ost_brw_lock_get(LCK_PR, exp, ioo, pp_rnb, &lockh);
 777         if (rc != 0)
 778                 GOTO(out_bulk, rc);
 779
 780         /*
 781          * If getting the lock took more time than
 782          * client was willing to wait, drop it. b=11330
 783          */
 784         if (cfs_time_current_sec() > req->rq_deadline ||
 785             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
 786                 no_reply = 1;
 787                 CERROR("Dropping timed-out read from %s because locking"
 788                        "object "LPX64" took %ld seconds (limit was %ld).\n",
 789                        libcfs_id2str(req->rq_peer), ioo->ioo_id,
 790                        cfs_time_current_sec() - req->rq_arrival_time.tv_sec,
 791                        req->rq_deadline - req->rq_arrival_time.tv_sec);
 792                 GOTO(out_lock, rc = -ETIMEDOUT);
 793         }
 794
 795         rc = obd_preprw(OBD_BRW_READ, exp, &body->oa, 1,
 796                         ioo, npages, pp_rnb, local_nb, oti, capa);
 797         if (rc != 0)
 798                 GOTO(out_lock, rc);
 799
 800         ost_prolong_locks(exp, ioo, pp_rnb, LCK_PW | LCK_PR);
 801
 802         nob = 0;
 803         for (i = 0; i < npages; i++) {
 804                 int page_rc = local_nb[i].rc;
 805
 806                 if (page_rc < 0) {              /* error */
 807                         rc = page_rc;
 808                         break;
 809                 }
 810
 811                 LASSERTF(page_rc <= pp_rnb[i].len, "page_rc (%d) > "
 812                          "pp_rnb[%d].len (%d)\n", page_rc, i, pp_rnb[i].len);
 813                 nob += page_rc;
 814                 if (page_rc != 0) {             /* some data! */
 815                         LASSERT (local_nb[i].page != NULL);
 816                         ptlrpc_prep_bulk_page(desc, local_nb[i].page,
 817                                               pp_rnb[i].offset & ~CFS_PAGE_MASK,
 818                                               page_rc);
 819                 }
 820
 821                 if (page_rc != pp_rnb[i].len) { /* short read */
 822                         /* All subsequent pages should be 0 */
 823                         while(++i < npages)
 824                                 LASSERT(local_nb[i].rc == 0);
 825                         break;
 826                 }
 827         }
 828
 829         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
 830                 cksum_type_t cksum_type = OBD_CKSUM_CRC32;
 831
 832                 if (body->oa.o_valid & OBD_MD_FLFLAGS)
 833                         cksum_type = cksum_type_unpack(body->oa.o_flags);
 834                 body->oa.o_flags = cksum_type_pack(cksum_type);
 835                 body->oa.o_valid = OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
 836                 body->oa.o_cksum = ost_checksum_bulk(desc, OST_READ, cksum_type);
 837                 CDEBUG(D_PAGE,"checksum at read origin: %x\n",body->oa.o_cksum);
 838         } else {
 839                 body->oa.o_valid = 0;
 840         }
 841         /* We're finishing using body->oa as an input variable */
 842
 843         /* Check if client was evicted while we were doing i/o before touching
 844            network */
 845         if (rc == 0) {
 846                 /* Check if there is eviction in progress, and if so, wait for
 847                  * it to finish */
 848                 if (unlikely(atomic_read(&exp->exp_obd->
 849                                                 obd_evict_inprogress))) {
 850                         lwi = LWI_INTR(NULL, NULL);
 851                         rc = l_wait_event(exp->exp_obd->
 852                                                 obd_evict_inprogress_waitq,
 853                                           !atomic_read(&exp->exp_obd->
 854                                                         obd_evict_inprogress),
 855                                           &lwi);
 856                 }
 857                 if (exp->exp_failed)
 858                         rc = -ENOTCONN;
 859                 else {
 860                         sptlrpc_svc_wrap_bulk(req, desc);
 861
 862                         rc = ptlrpc_start_bulk_transfer(desc);
 863                 }
 864
 865                 if (rc == 0) {
 866                         time_t start = cfs_time_current_sec();
 867                         do {
 868                                 long timeoutl = req->rq_deadline -
 869                                         cfs_time_current_sec();
 870                                 cfs_duration_t timeout = (timeoutl <= 0 || rc) ?
 871                                         CFS_TICK : cfs_time_seconds(timeoutl);
 872                                 lwi = LWI_TIMEOUT_INTERVAL(timeout,
 873                                                            cfs_time_seconds(1),
 874                                                            ost_bulk_timeout,
 875                                                            desc);
 876                                 rc = l_wait_event(desc->bd_waitq,
 877                                                   !ptlrpc_bulk_active(desc) ||
 878                                                   exp->exp_failed, &lwi);
 879                                 LASSERT(rc == 0 || rc == -ETIMEDOUT);
 880                                 /* Wait again if we changed deadline */
 881                         } while ((rc == -ETIMEDOUT) &&
 882                                  (req->rq_deadline > cfs_time_current_sec()));
 883
 884                         if (rc == -ETIMEDOUT) {
 885                                 DEBUG_REQ(D_ERROR, req,
 886                                           "timeout on bulk PUT after %ld%+lds",
 887                                           req->rq_deadline - start,
 888                                           cfs_time_current_sec() -
 889                                           req->rq_deadline);
 890                                 ptlrpc_abort_bulk(desc);
 891                         } else if (exp->exp_failed) {
 892                                 DEBUG_REQ(D_ERROR, req, "Eviction on bulk PUT");
 893                                 rc = -ENOTCONN;
 894                                 ptlrpc_abort_bulk(desc);
 895                         } else if (!desc->bd_success ||
 896                                    desc->bd_nob_transferred != desc->bd_nob) {
 897                                 DEBUG_REQ(D_ERROR, req, "%s bulk PUT %d(%d)",
 898                                           desc->bd_success ?
 899                                           "truncated" : "network error on",
 900                                           desc->bd_nob_transferred,
 901                                           desc->bd_nob);
 902                                 /* XXX should this be a different errno? */
 903                                 rc = -ETIMEDOUT;
 904                         }
 905                 } else {
 906                         DEBUG_REQ(D_ERROR, req, "bulk PUT failed: rc %d", rc);
 907                 }
 908                 no_reply = rc != 0;
 909         }
 910
 911         /* Must commit after prep above in all cases */
 912         rc = obd_commitrw(OBD_BRW_READ, exp, &body->oa, 1,
 913                           ioo, npages, local_nb, oti, rc);
 914
 915         ost_nio_pages_put(req, local_nb, npages);
 916
 917         if (rc == 0) {
 918                 repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
 919                                          sizeof(*repbody));
 920                 memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
 921         }
 922
 923 out_lock:
 924         ost_brw_lock_put(LCK_PR, ioo, pp_rnb, &lockh);
 925 out_bulk:
 926         ptlrpc_free_bulk(desc);
 927 out:
 928         LASSERT(rc <= 0);
 929         if (rc == 0) {
 930                 req->rq_status = nob;
 931                 target_committed_to_req(req);
 932                 ptlrpc_reply(req);
 933         } else if (!no_reply) {
 934                 /* Only reply if there was no comms problem with bulk */
 935                 target_committed_to_req(req);
 936                 req->rq_status = rc;
 937                 ptlrpc_error(req);
 938         } else {
 939                 /* reply out callback would free */
 940                 ptlrpc_req_drop_rs(req);
 941                 CWARN("%s: ignoring bulk IO comm error with %s@%s id %s - "
 942                       "client will retry\n",
 943                       exp->exp_obd->obd_name,
 944                       exp->exp_client_uuid.uuid,
 945                       exp->exp_connection->c_remote_uuid.uuid,
 946                       libcfs_id2str(req->rq_peer));
 947         }
 948
 949         RETURN(rc);
 950 }
 951
 952 static int ost_brw_write(struct ptlrpc_request *req, struct obd_trans_info *oti)
 953 {
 954         struct ptlrpc_bulk_desc *desc;
 955         struct obd_export       *exp = req->rq_export;
 956         struct niobuf_remote    *remote_nb;
 957         struct niobuf_remote    *pp_rnb;
 958         struct niobuf_local     *local_nb;
 959         struct obd_ioobj        *ioo;
 960         struct ost_body         *body, *repbody;
 961         struct l_wait_info       lwi;
 962         struct lustre_handle     lockh = {0};
 963         struct lustre_capa      *capa = NULL;
 964         __u32                   *rcs;
 965         __u32 size[3] = { sizeof(struct ptlrpc_body), sizeof(*body) };
 966         int objcount, niocount, npages;
 967         int rc, swab, i, j;
 968         obd_count                client_cksum = 0, server_cksum = 0;
 969         cksum_type_t             cksum_type = OBD_CKSUM_CRC32;
 970         int                      no_reply = 0;
 971         ENTRY;
 972
 973         req->rq_bulk_write = 1;
 974
 975         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK))
 976                 GOTO(out, rc = -EIO);
 977         if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_WRITE_BULK2))
 978                 GOTO(out, rc = -EFAULT);
 979
 980         /* pause before transaction has been started */
 981         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_BULK, (obd_timeout + 1) / 4);
 982
 983         /* Check if there is eviction in progress, and if so, wait for it to
 984          * finish */
 985         if (unlikely(atomic_read(&exp->exp_obd->obd_evict_inprogress))) {
 986                 lwi = LWI_INTR(NULL, NULL); // We do not care how long it takes
 987                 rc = l_wait_event(exp->exp_obd->obd_evict_inprogress_waitq,
 988                         !atomic_read(&exp->exp_obd->obd_evict_inprogress),
 989                         &lwi);
 990         }
 991         if (exp->exp_failed)
 992                 GOTO(out, rc = -ENOTCONN);
 993
 994         swab = lustre_msg_swabbed(req->rq_reqmsg);
 995         body = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*body),
 996                                   lustre_swab_ost_body);
 997         if (body == NULL) {
 998                 CERROR("Missing/short ost_body\n");
 999                 GOTO(out, rc = -EFAULT);
1000         }
1001
1002         lustre_set_req_swabbed(req, REQ_REC_OFF + 1);
1003         objcount = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF + 1) /
1004                    sizeof(*ioo);
1005         if (objcount == 0) {
1006                 CERROR("Missing/short ioobj\n");
1007                 GOTO(out, rc = -EFAULT);
1008         }
1009         if (objcount > 1) {
1010                 CERROR("too many ioobjs (%d)\n", objcount);
1011                 GOTO(out, rc = -EFAULT);
1012         }
1013
1014         ioo = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 1,
1015                              objcount * sizeof(*ioo));
1016         LASSERT (ioo != NULL);
1017         for (niocount = i = 0; i < objcount; i++) {
1018                 if (swab)
1019                         lustre_swab_obd_ioobj(&ioo[i]);
1020                 if (ioo[i].ioo_bufcnt == 0) {
1021                         CERROR("ioo[%d] has zero bufcnt\n", i);
1022                         GOTO(out, rc = -EFAULT);
1023                 }
1024                 niocount += ioo[i].ioo_bufcnt;
1025         }
1026
1027         if (niocount > PTLRPC_MAX_BRW_PAGES) {
1028                 DEBUG_REQ(D_ERROR, req, "bulk has too many pages (%d)",
1029                           niocount);
1030                 GOTO(out, rc = -EFAULT);
1031         }
1032
1033         remote_nb = lustre_swab_reqbuf(req, REQ_REC_OFF + 2,
1034                                        niocount * sizeof(*remote_nb),
1035                                        lustre_swab_niobuf_remote);
1036         if (remote_nb == NULL) {
1037                 CERROR("Missing/short niobuf\n");
1038                 GOTO(out, rc = -EFAULT);
1039         }
1040         if (swab) {                             /* swab the remaining niobufs */
1041                 for (i = 1; i < niocount; i++)
1042                         lustre_swab_niobuf_remote (&remote_nb[i]);
1043         }
1044
1045         if (body->oa.o_valid & OBD_MD_FLOSSCAPA)
1046                 capa = lustre_unpack_capa(req->rq_reqmsg, REQ_REC_OFF + 3);
1047
1048         size[REPLY_REC_OFF + 1] = niocount * sizeof(*rcs);
1049         rc = lustre_pack_reply(req, 3, size, NULL);
1050         if (rc != 0)
1051                 GOTO(out, rc);
1052         OBD_FAIL_TIMEOUT(OBD_FAIL_OST_BRW_PAUSE_PACK, obd_fail_val);
1053         rcs = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF + 1,
1054                              niocount * sizeof(*rcs));
1055
1056         /*
1057          * Per-thread array of struct niobuf_{local,remote}'s was allocated by
1058          * ost_thread_init().
1059          */
1060         local_nb = ost_tls(req)->local;
1061         pp_rnb   = ost_tls(req)->remote;
1062
1063         /* FIXME all niobuf splitting should be done in obdfilter if needed */
1064         /* CAVEAT EMPTOR this sets ioo->ioo_bufcnt to # pages */
1065         npages = get_per_page_niobufs(ioo, objcount,remote_nb,niocount,&pp_rnb);
1066         if (npages < 0)
1067                 GOTO(out, rc = npages);
1068
1069         LASSERT(npages <= OST_THREAD_POOL_SIZE);
1070
1071         ost_nio_pages_get(req, local_nb, npages);
1072
1073         desc = ptlrpc_prep_bulk_exp(req, npages,
1074                                      BULK_GET_SINK, OST_BULK_PORTAL);
1075         if (desc == NULL)
1076                 GOTO(out, rc = -ENOMEM);
1077
1078         rc = ost_brw_lock_get(LCK_PW, exp, ioo, pp_rnb, &lockh);
1079         if (rc != 0)
1080                 GOTO(out_bulk, rc);
1081
1082         /*
1083          * If getting the lock took more time than
1084          * client was willing to wait, drop it. b=11330
1085          */
1086         if (cfs_time_current_sec() > req->rq_deadline ||
1087             OBD_FAIL_CHECK(OBD_FAIL_OST_DROP_REQ)) {
1088                 no_reply = 1;
1089                 CERROR("Dropping timed-out write from %s because locking "
1090                        "object "LPX64" took %ld seconds (limit was %ld).\n",
1091                        libcfs_id2str(req->rq_peer), ioo->ioo_id,
1092                        cfs_time_current_sec() - req->rq_arrival_time.tv_sec,
1093                        req->rq_deadline - req->rq_arrival_time.tv_sec);
1094                 GOTO(out_lock, rc = -ETIMEDOUT);
1095         }
1096
1097         ost_prolong_locks(exp, ioo, pp_rnb, LCK_PW);
1098
1099         /* obd_preprw clobbers oa->valid, so save what we need */
1100         if (body->oa.o_valid & OBD_MD_FLCKSUM) {
1101                 client_cksum = body->oa.o_cksum;
1102                 if (body->oa.o_valid & OBD_MD_FLFLAGS)
1103                         cksum_type = cksum_type_unpack(body->oa.o_flags);
1104         }
1105
1106         /* Because we already sync grant info with client when reconnect,
1107          * grant info will be cleared for resent req, then fed_grant and
1108          * total_grant will not be modified in following preprw_write */
1109         if (lustre_msg_get_flags(req->rq_reqmsg) & (MSG_RESENT | MSG_REPLAY)) {
1110                 DEBUG_REQ(D_CACHE, req, "clear resent/replay req grant info");
1111                 body->oa.o_valid &= ~OBD_MD_FLGRANT;
1112         }
1113
1114         rc = obd_preprw(OBD_BRW_WRITE, exp, &body->oa, objcount,
1115                         ioo, npages, pp_rnb, local_nb, oti, capa);
1116         if (rc != 0)
1117                 GOTO(out_lock, rc);
1118
1119         /* NB Having prepped, we must commit... */
1120
1121         for (i = 0; i < npages; i++)
1122                 ptlrpc_prep_bulk_page(desc, local_nb[i].page,
1123                                       pp_rnb[i].offset & ~CFS_PAGE_MASK,
1124                                       pp_rnb[i].len);
1125
1126         /* Check if client was evicted while we were doing i/o before touching
1127            network */
1128         if (desc->bd_export->exp_failed)
1129                 rc = -ENOTCONN;
1130         else
1131                 rc = ptlrpc_start_bulk_transfer (desc);
1132         if (rc == 0) {
1133                 time_t start = cfs_time_current_sec();
1134                 do {
1135                         long timeoutl = req->rq_deadline -
1136                                 cfs_time_current_sec();
1137                         cfs_duration_t timeout = (timeoutl <= 0 || rc) ?
1138                                 CFS_TICK : cfs_time_seconds(timeoutl);
1139                         lwi = LWI_TIMEOUT_INTERVAL(timeout, cfs_time_seconds(1),
1140                                                    ost_bulk_timeout, desc);
1141                         rc = l_wait_event(desc->bd_waitq,
1142                                           !ptlrpc_bulk_active(desc) ||
1143                                           desc->bd_export->exp_failed, &lwi);
1144                         LASSERT(rc == 0 || rc == -ETIMEDOUT);
1145                         /* Wait again if we changed deadline */
1146                 } while ((rc == -ETIMEDOUT) &&
1147                          (req->rq_deadline > cfs_time_current_sec()));
1148
1149                 if (rc == -ETIMEDOUT) {
1150                         DEBUG_REQ(D_ERROR, req,
1151                                   "timeout on bulk GET after %ld%+lds",
1152                                   req->rq_deadline - start,
1153                                   cfs_time_current_sec() -
1154                                   req->rq_deadline);
1155                         ptlrpc_abort_bulk(desc);
1156                 } else if (desc->bd_export->exp_failed) {
1157                         DEBUG_REQ(D_ERROR, req, "Eviction on bulk GET");
1158                         rc = -ENOTCONN;
1159                         ptlrpc_abort_bulk(desc);
1160                 } else if (!desc->bd_success ||
1161                            desc->bd_nob_transferred != desc->bd_nob) {
1162                         DEBUG_REQ(D_ERROR, req, "%s bulk GET %d(%d)",
1163                                   desc->bd_success ?
1164                                   "truncated" : "network error on",
1165                                   desc->bd_nob_transferred, desc->bd_nob);
1166                         /* XXX should this be a different errno? */
1167                         rc = -ETIMEDOUT;
1168                 }
1169         } else {
1170                 DEBUG_REQ(D_ERROR, req, "ptlrpc_bulk_get failed: rc %d", rc);
1171         }
1172         no_reply = rc != 0;
1173
1174         if (rc == 0)
1175                 sptlrpc_svc_unwrap_bulk(req, desc);
1176
1177         repbody = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF,
1178                                  sizeof(*repbody));
1179         memcpy(&repbody->oa, &body->oa, sizeof(repbody->oa));
1180
1181         if (unlikely(client_cksum != 0 && rc == 0)) {
1182                 static int cksum_counter;
1183                 repbody->oa.o_valid |= OBD_MD_FLCKSUM | OBD_MD_FLFLAGS;
1184                 repbody->oa.o_flags &= ~OBD_FL_CKSUM_ALL;
1185                 repbody->oa.o_flags |= cksum_type_pack(cksum_type);
1186                 server_cksum = ost_checksum_bulk(desc, OST_WRITE, cksum_type);
1187                 repbody->oa.o_cksum = server_cksum;
1188                 cksum_counter++;
1189                 if (unlikely(client_cksum != server_cksum)) {
1190                         CERROR("client csum %x, server csum %x\n",
1191                                client_cksum, server_cksum);
1192                         cksum_counter = 0;
1193                 } else if ((cksum_counter & (-cksum_counter)) == cksum_counter){
1194                         CDEBUG(D_INFO, "Checksum %u from %s OK: %x\n",
1195                                cksum_counter, libcfs_id2str(req->rq_peer),
1196                                server_cksum);
1197                 }
1198         }
1199
1200         /* Must commit after prep above in all cases */
1201         rc = obd_commitrw(OBD_BRW_WRITE, exp, &repbody->oa,
1202                            objcount, ioo, npages, local_nb, oti, rc);
1203
1204         if (unlikely(client_cksum != server_cksum && rc == 0)) {
1205                 int  new_cksum = ost_checksum_bulk(desc, OST_WRITE, cksum_type);
1206                 char *msg;
1207                 char *via;
1208                 char *router;
1209
1210                 if (new_cksum == server_cksum)
1211                         msg = "changed in transit before arrival at OST";
1212                 else if (new_cksum == client_cksum)
1213                         msg = "initial checksum before message complete";
1214                 else
1215                         msg = "changed in transit AND after initial checksum";
1216
1217                 if (req->rq_peer.nid == desc->bd_sender) {
1218                         via = router = "";
1219                 } else {
1220                         via = " via ";
1221                         router = libcfs_nid2str(desc->bd_sender);
1222                 }
1223
1224                 LCONSOLE_ERROR_MSG(0x168, "%s: BAD WRITE CHECKSUM: %s from "
1225                                    "%s%s%s inum "LPU64"/"LPU64" object "
1226                                    LPU64"/"LPU64" extent ["LPU64"-"LPU64"]\n",
1227                                    exp->exp_obd->obd_name, msg,
1228                                    libcfs_id2str(req->rq_peer),
1229                                    via, router,
1230                                    body->oa.o_valid & OBD_MD_FLFID ?
1231                                                 body->oa.o_fid : (__u64)0,
1232                                    body->oa.o_valid & OBD_MD_FLFID ?
1233                                                 body->oa.o_generation :(__u64)0,
1234                                    body->oa.o_id,
1235                                    body->oa.o_valid & OBD_MD_FLGROUP ?
1236                                                 body->oa.o_gr : (__u64)0,
1237                                    pp_rnb[0].offset,
1238                                    pp_rnb[npages-1].offset+pp_rnb[npages-1].len
1239                                    - 1 );
1240                 CERROR("client csum %x, original server csum %x, "
1241                        "server csum now %x\n",
1242                        client_cksum, server_cksum, new_cksum);
1243         }
1244
1245         ost_nio_pages_put(req, local_nb, npages);
1246
1247         if (rc == 0) {
1248                 /* set per-requested niobuf return codes */
1249                 for (i = j = 0; i < niocount; i++) {
1250                         int nob = remote_nb[i].len;
1251
1252                         rcs[i] = 0;
1253                         do {
1254                                 LASSERT(j < npages);
1255                                 if (local_nb[j].rc < 0)
1256                                         rcs[i] = local_nb[j].rc;
1257                                 nob -= pp_rnb[j].len;
1258                                 j++;
1259                         } while (nob > 0);
1260                         LASSERT(nob == 0);
1261                 }
1262                 LASSERT(j == npages);
1263         }
1264
1265 out_lock:
1266         ost_brw_lock_put(LCK_PW, ioo, pp_rnb, &lockh);
1267 out_bulk:
1268         ptlrpc_free_bulk(desc);
1269 out:
1270         if (rc == 0) {
1271                 oti_to_request(oti, req);
1272                 target_committed_to_req(req);
1273                 rc = ptlrpc_reply(req);
1274         } else if (!no_reply) {
1275                 /* Only reply if there was no comms problem with bulk */
1276                 target_committed_to_req(req);
1277                 req->rq_status = rc;
1278                 ptlrpc_error(req);
1279         } else {
1280                 /* reply out callback would free */
1281                 ptlrpc_req_drop_rs(req);
1282                 CWARN("%s: ignoring bulk IO comm error with %s@%s id %s - "
1283                       "client will retry\n",
1284                       exp->exp_obd->obd_name,
1285                       exp->exp_client_uuid.uuid,
1286                       exp->exp_connection->c_remote_uuid.uuid,
1287                       libcfs_id2str(req->rq_peer));
1288         }
1289         RETURN(rc);
1290 }
1291
1292 static int ost_set_info(struct obd_export *exp, struct ptlrpc_request *req)
1293 {
1294         char *key, *val = NULL;
1295         int keylen, vallen, rc = 0;
1296         ENTRY;
1297
1298         key = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF, 1);
1299         if (key == NULL) {
1300                 DEBUG_REQ(D_HA, req, "no set_info key");
1301                 RETURN(-EFAULT);
1302         }
1303         keylen = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF);
1304
1305         rc = lustre_pack_reply(req, 1, NULL, NULL);
1306         if (rc)
1307                 RETURN(rc);
1308
1309         vallen = lustre_msg_buflen(req->rq_reqmsg, REQ_REC_OFF + 1);
1310         if (vallen)
1311                 val = lustre_msg_buf(req->rq_reqmsg, REQ_REC_OFF + 1, 0);
1312
1313         if (KEY_IS(KEY_EVICT_BY_NID)) {
1314                 if (val && vallen)
1315                         obd_export_evict_by_nid(exp->exp_obd, val);
1316
1317                 GOTO(out, rc = 0);
1318         }
1319
1320         rc = obd_set_info_async(exp, keylen, key, vallen, val, NULL);
1321 out:
1322         lustre_msg_set_status(req->rq_repmsg, 0);
1323         RETURN(rc);
1324 }
1325
1326 static int ost_get_info(struct obd_export *exp, struct ptlrpc_request *req)
1327 {
1328         void *key, *reply;
1329         int keylen, replylen, rc = 0;
1330         struct req_capsule *pill = &req->rq_pill;
1331         ENTRY;
1332
1333         req_capsule_set(&req->rq_pill, &RQF_OST_GET_INFO_GENERIC);
1334
1335         /* this common part for get_info rpc */
1336         key = req_capsule_client_get(pill, &RMF_SETINFO_KEY);
1337         if (key == NULL) {
1338                 DEBUG_REQ(D_HA, req, "no get_info key");
1339                 RETURN(-EFAULT);
1340         }
1341         keylen = req_capsule_get_size(pill, &RMF_SETINFO_KEY, RCL_CLIENT);
1342
1343         rc = obd_get_info(exp, keylen, key, &replylen, NULL, NULL);
1344         if (rc)
1345                 RETURN(rc);
1346
1347         req_capsule_set_size(pill, &RMF_GENERIC_DATA,
1348                              RCL_SERVER, replylen);
1349
1350         rc = req_capsule_server_pack(pill);
1351         if (rc)
1352                 RETURN(rc);
1353
1354         reply = req_capsule_server_get(pill, &RMF_GENERIC_DATA);
1355         if (reply == NULL)
1356                 RETURN(-ENOMEM);
1357
1358         /* call again to fill in the reply buffer */
1359         rc = obd_get_info(exp, keylen, key, &replylen, reply, NULL);
1360
1361         lustre_msg_set_status(req->rq_repmsg, 0);
1362         RETURN(rc);
1363 }
1364
1365 static int ost_handle_quotactl(struct ptlrpc_request *req)
1366 {
1367         struct obd_quotactl *oqctl, *repoqc;
1368         __u32 size[2] = { sizeof(struct ptlrpc_body), sizeof(*repoqc) };
1369         int rc;
1370         ENTRY;
1371
1372         oqctl = lustre_swab_reqbuf(req, REQ_REC_OFF, sizeof(*oqctl),
1373                                    lustre_swab_obd_quotactl);
1374         if (oqctl == NULL)
1375                 GOTO(out, rc = -EPROTO);
1376
1377         rc = lustre_pack_reply(req, 2, size, NULL);
1378         if (rc)
1379                 GOTO(out, rc);
1380
1381         repoqc = lustre_msg_buf(req->rq_repmsg, REPLY_REC_OFF, sizeof(*repoqc));
1382
1383         req->rq_status = obd_quotactl(req->rq_export, oqctl);
1384         *repoqc = *oqctl;
1385 out:
1386         RETURN(rc);
1387 }
1388
1389 static int ost_handle_quotacheck(struct ptlrpc_request *req)
1390 {
1391         struct obd_quotactl *oqctl;
1392         int rc;
1393         ENTRY;
1394
1395         oqctl = req_capsule_client_get(&req->rq_pill, &RMF_OBD_QUOTACTL);
1396         if (oqctl == NULL)
1397                 RETURN(-EPROTO);
1398
1399         rc = req_capsule_server_pack(&req->rq_pill);
1400         if (rc) {
1401                 CERROR("ost: out of memory while packing quotacheck reply\n");
1402                 RETURN(-ENOMEM);
1403         }
1404
1405         req->rq_status = obd_quotacheck(req->rq_export, oqctl);
1406         RETURN(0);
1407 }
1408
1409 static int ost_llog_handle_connect(struct obd_export *exp,
1410                                    struct ptlrpc_request *req)
1411 {
1412         struct llogd_conn_body *body;
1413         int rc;
1414         ENTRY;
1415
1416         body = lustre_msg_buf(req->rq_reqmsg, 1, sizeof(*body));
1417         rc = obd_llog_connect(exp, body);
1418         RETURN(rc);
1419 }
1420
1421 static int filter_export_check_flavor(struct filter_obd *filter,
1422                                       struct obd_export *exp,
1423                                       struct ptlrpc_request *req)
1424 {
1425         int     rc = 0;
1426
1427         /* FIXME
1428          * this should be done in filter_connect()/filter_reconnect(), but
1429          * we can't obtain information like NID, which stored in incoming
1430          * request, thus can't decide what flavor to use. so we do it here.
1431          *
1432          * This hack should be removed after the OST stack be rewritten, just
1433          * like what we are doing in mdt_obd_connect()/mdt_obd_reconnect().
1434          */
1435         if (exp->exp_flvr.sf_rpc != SPTLRPC_FLVR_INVALID)
1436                 return 0;
1437
1438         CDEBUG(D_SEC, "from %s\n", sptlrpc_part2name(req->rq_sp_from));
1439         spin_lock(&exp->exp_lock);
1440         exp->exp_sp_peer = req->rq_sp_from;
1441
1442         read_lock(&filter->fo_sptlrpc_lock);
1443         sptlrpc_rule_set_choose(&filter->fo_sptlrpc_rset, exp->exp_sp_peer,
1444                                 req->rq_peer.nid, &exp->exp_flvr);
1445         read_unlock(&filter->fo_sptlrpc_lock);
1446
1447         if (exp->exp_flvr.sf_rpc != req->rq_flvr.sf_rpc) {
1448                 CERROR("invalid rpc flavor %x, expect %x, from %s\n",
1449                        req->rq_flvr.sf_rpc, exp->exp_flvr.sf_rpc,
1450                        libcfs_nid2str(req->rq_peer.nid));
1451                 exp->exp_flvr.sf_rpc = SPTLRPC_FLVR_INVALID;
1452                 rc = -EACCES;
1453         }
1454
1455         spin_unlock(&exp->exp_lock);
1456
1457         return rc;
1458 }
1459
1460 static int ost_filter_recovery_request(struct ptlrpc_request *req,
1461                                        struct obd_device *obd, int *process)
1462 {
1463         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1464         case OST_CONNECT: /* This will never get here, but for completeness. */
1465         case OST_DISCONNECT:
1466                *process = 1;
1467                RETURN(0);
1468
1469         case OBD_PING:
1470         case OST_CREATE:
1471         case OST_DESTROY:
1472         case OST_PUNCH:
1473         case OST_SETATTR:
1474         case OST_SYNC:
1475         case OST_WRITE:
1476         case OBD_LOG_CANCEL:
1477         case LDLM_ENQUEUE:
1478                 *process = target_queue_recovery_request(req, obd);
1479                 RETURN(0);
1480
1481         default:
1482                 DEBUG_REQ(D_ERROR, req, "not permitted during recovery");
1483                 *process = -EAGAIN;
1484                 RETURN(0);
1485         }
1486 }
1487
1488 int ost_msg_check_version(struct lustre_msg *msg)
1489 {
1490         int rc;
1491
1492         switch(lustre_msg_get_opc(msg)) {
1493         case OST_CONNECT:
1494         case OST_DISCONNECT:
1495         case OBD_PING:
1496         case SEC_CTX_INIT:
1497         case SEC_CTX_INIT_CONT:
1498         case SEC_CTX_FINI:
1499                 rc = lustre_msg_check_version(msg, LUSTRE_OBD_VERSION);
1500                 if (rc)
1501                         CERROR("bad opc %u version %08x, expecting %08x\n",
1502                                lustre_msg_get_opc(msg),
1503                                lustre_msg_get_version(msg),
1504                                LUSTRE_OBD_VERSION);
1505                 break;
1506         case OST_CREATE:
1507         case OST_DESTROY:
1508         case OST_GETATTR:
1509         case OST_SETATTR:
1510         case OST_WRITE:
1511         case OST_READ:
1512         case OST_PUNCH:
1513         case OST_STATFS:
1514         case OST_SYNC:
1515         case OST_SET_INFO:
1516         case OST_GET_INFO:
1517         case OST_QUOTACHECK:
1518         case OST_QUOTACTL:
1519                 rc = lustre_msg_check_version(msg, LUSTRE_OST_VERSION);
1520                 if (rc)
1521                         CERROR("bad opc %u version %08x, expecting %08x\n",
1522                                lustre_msg_get_opc(msg),
1523                                lustre_msg_get_version(msg),
1524                                LUSTRE_OST_VERSION);
1525                 break;
1526         case LDLM_ENQUEUE:
1527         case LDLM_CONVERT:
1528         case LDLM_CANCEL:
1529         case LDLM_BL_CALLBACK:
1530         case LDLM_CP_CALLBACK:
1531                 rc = lustre_msg_check_version(msg, LUSTRE_DLM_VERSION);
1532                 if (rc)
1533                         CERROR("bad opc %u version %08x, expecting %08x\n",
1534                                lustre_msg_get_opc(msg),
1535                                lustre_msg_get_version(msg),
1536                                LUSTRE_DLM_VERSION);
1537                 break;
1538         case LLOG_ORIGIN_CONNECT:
1539         case OBD_LOG_CANCEL:
1540                 rc = lustre_msg_check_version(msg, LUSTRE_LOG_VERSION);
1541                 if (rc)
1542                         CERROR("bad opc %u version %08x, expecting %08x\n",
1543                                lustre_msg_get_opc(msg),
1544                                lustre_msg_get_version(msg),
1545                                LUSTRE_LOG_VERSION);
1546                 break;
1547         default:
1548                 CERROR("Unexpected opcode %d\n", lustre_msg_get_opc(msg));
1549                 rc = -ENOTSUPP;
1550         }
1551         return rc;
1552 }
1553
1554 /* TODO: handle requests in a similar way as MDT: see mdt_handle_common() */
1555 int ost_handle(struct ptlrpc_request *req)
1556 {
1557         struct obd_trans_info trans_info = { 0, };
1558         struct obd_trans_info *oti = &trans_info;
1559         int should_process, fail = OBD_FAIL_OST_ALL_REPLY_NET, rc = 0;
1560         struct obd_device *obd = NULL;
1561         ENTRY;
1562
1563         LASSERT(current->journal_info == NULL);
1564
1565         /* primordial rpcs don't affect server recovery */
1566         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1567         case SEC_CTX_INIT:
1568         case SEC_CTX_INIT_CONT:
1569         case SEC_CTX_FINI:
1570                 GOTO(out, rc = 0);
1571         }
1572
1573         req_capsule_init(&req->rq_pill, req, RCL_SERVER);
1574
1575         /* XXX identical to MDS */
1576         if (lustre_msg_get_opc(req->rq_reqmsg) != OST_CONNECT) {
1577                 int recovering;
1578
1579                 if (req->rq_export == NULL) {
1580                         CDEBUG(D_HA,"operation %d on unconnected OST from %s\n",
1581                                lustre_msg_get_opc(req->rq_reqmsg),
1582                                libcfs_id2str(req->rq_peer));
1583                         req->rq_status = -ENOTCONN;
1584                         GOTO(out, rc = -ENOTCONN);
1585                 }
1586
1587                 obd = req->rq_export->exp_obd;
1588
1589                 /* Check for aborted recovery. */
1590                 spin_lock_bh(&obd->obd_processing_task_lock);
1591                 recovering = obd->obd_recovering;
1592                 spin_unlock_bh(&obd->obd_processing_task_lock);
1593                 if (recovering) {
1594                         rc = ost_filter_recovery_request(req, obd,
1595                                                          &should_process);
1596                         if (rc || !should_process)
1597                                 RETURN(rc);
1598                         else if (should_process < 0) {
1599                                 req->rq_status = should_process;
1600                                 rc = ptlrpc_error(req);
1601                                 RETURN(rc);
1602                         }
1603                 }
1604         }
1605
1606         oti_init(oti, req);
1607
1608         rc = ost_msg_check_version(req->rq_reqmsg);
1609         if (rc)
1610                 RETURN(rc);
1611
1612         switch (lustre_msg_get_opc(req->rq_reqmsg)) {
1613         case OST_CONNECT: {
1614                 CDEBUG(D_INODE, "connect\n");
1615                 req_capsule_set(&req->rq_pill, &RQF_OST_CONNECT);
1616                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET))
1617                         RETURN(0);
1618                 rc = target_handle_connect(req);
1619                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_CONNECT_NET2))
1620                         RETURN(0);
1621                 if (!rc) {
1622                         struct obd_export *exp = req->rq_export;
1623
1624                         obd = exp->exp_obd;
1625
1626                         rc = filter_export_check_flavor(&obd->u.filter,
1627                                                         exp, req);
1628                 }
1629                 break;
1630         }
1631         case OST_DISCONNECT:
1632                 CDEBUG(D_INODE, "disconnect\n");
1633                 req_capsule_set(&req->rq_pill, &RQF_OST_DISCONNECT);
1634                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DISCONNECT_NET))
1635                         RETURN(0);
1636                 rc = target_handle_disconnect(req);
1637                 break;
1638         case OST_CREATE:
1639                 CDEBUG(D_INODE, "create\n");
1640                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_CREATE_NET))
1641                         RETURN(0);
1642                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
1643                         GOTO(out, rc = -ENOSPC);
1644                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
1645                         GOTO(out, rc = -EROFS);
1646                 rc = ost_create(req->rq_export, req, oti);
1647                 break;
1648         case OST_DESTROY:
1649                 CDEBUG(D_INODE, "destroy\n");
1650                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_DESTROY_NET))
1651                         RETURN(0);
1652                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
1653                         GOTO(out, rc = -EROFS);
1654                 rc = ost_destroy(req->rq_export, req, oti);
1655                 break;
1656         case OST_GETATTR:
1657                 CDEBUG(D_INODE, "getattr\n");
1658                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_GETATTR_NET))
1659                         RETURN(0);
1660                 rc = ost_getattr(req->rq_export, req);
1661                 break;
1662         case OST_SETATTR:
1663                 CDEBUG(D_INODE, "setattr\n");
1664                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SETATTR_NET))
1665                         RETURN(0);
1666                 rc = ost_setattr(req->rq_export, req, oti);
1667                 break;
1668         case OST_WRITE:
1669                 CDEBUG(D_INODE, "write\n");
1670                 /* req->rq_request_portal would be nice, if it was set */
1671                 if (req->rq_rqbd->rqbd_service->srv_req_portal !=OST_IO_PORTAL){
1672                         CERROR("%s: deny write request from %s to portal %u\n",
1673                                req->rq_export->exp_obd->obd_name,
1674                                obd_export_nid2str(req->rq_export),
1675                                req->rq_rqbd->rqbd_service->srv_req_portal);
1676                         GOTO(out, rc = -EPROTO);
1677                 }
1678                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_NET))
1679                         RETURN(0);
1680                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_ENOSPC))
1681                         GOTO(out, rc = -ENOSPC);
1682                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
1683                         GOTO(out, rc = -EROFS);
1684                 rc = ost_brw_write(req, oti);
1685                 LASSERT(current->journal_info == NULL);
1686                 /* ost_brw_write sends its own replies */
1687                 RETURN(rc);
1688         case OST_READ:
1689                 CDEBUG(D_INODE, "read\n");
1690                 /* req->rq_request_portal would be nice, if it was set */
1691                 if (req->rq_rqbd->rqbd_service->srv_req_portal !=OST_IO_PORTAL){
1692                         CERROR("%s: deny read request from %s to portal %u\n",
1693                                req->rq_export->exp_obd->obd_name,
1694                                obd_export_nid2str(req->rq_export),
1695                                req->rq_rqbd->rqbd_service->srv_req_portal);
1696                         GOTO(out, rc = -EPROTO);
1697                 }
1698                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_BRW_NET))
1699                         RETURN(0);
1700                 rc = ost_brw_read(req, oti);
1701                 LASSERT(current->journal_info == NULL);
1702                 /* ost_brw_read sends its own replies */
1703                 RETURN(rc);
1704         case OST_PUNCH:
1705                 CDEBUG(D_INODE, "punch\n");
1706                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_PUNCH_NET))
1707                         RETURN(0);
1708                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_EROFS))
1709                         GOTO(out, rc = -EROFS);
1710                 rc = ost_punch(req->rq_export, req, oti);
1711                 break;
1712         case OST_STATFS:
1713                 CDEBUG(D_INODE, "statfs\n");
1714                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_STATFS_NET))
1715                         RETURN(0);
1716                 rc = ost_statfs(req);
1717                 break;
1718         case OST_SYNC:
1719                 CDEBUG(D_INODE, "sync\n");
1720                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_SYNC_NET))
1721                         RETURN(0);
1722                 rc = ost_sync(req->rq_export, req);
1723                 break;
1724         case OST_SET_INFO:
1725                 DEBUG_REQ(D_INODE, req, "set_info");
1726                 req_capsule_set(&req->rq_pill, &RQF_OST_SET_INFO);
1727                 rc = ost_set_info(req->rq_export, req);
1728                 break;
1729         case OST_GET_INFO:
1730                 DEBUG_REQ(D_INODE, req, "get_info");
1731                 rc = ost_get_info(req->rq_export, req);
1732                 break;
1733         case OST_QUOTACHECK:
1734                 CDEBUG(D_INODE, "quotacheck\n");
1735                 req_capsule_set(&req->rq_pill, &RQF_OST_QUOTACHECK);
1736                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_QUOTACHECK_NET))
1737                         RETURN(0);
1738                 rc = ost_handle_quotacheck(req);
1739                 break;
1740         case OST_QUOTACTL:
1741                 CDEBUG(D_INODE, "quotactl\n");
1742                 req_capsule_set(&req->rq_pill, &RQF_OST_QUOTACTL);
1743                 if (OBD_FAIL_CHECK(OBD_FAIL_OST_QUOTACTL_NET))
1744                         RETURN(0);
1745                 rc = ost_handle_quotactl(req);
1746                 break;
1747         case OBD_PING:
1748                 DEBUG_REQ(D_INODE, req, "ping");
1749                 req_capsule_set(&req->rq_pill, &RQF_OBD_PING);
1750                 rc = target_handle_ping(req);
1751                 break;
1752         /* FIXME - just reply status */
1753         case LLOG_ORIGIN_CONNECT:
1754                 DEBUG_REQ(D_INODE, req, "log connect");
1755                 req_capsule_set(&req->rq_pill, &RQF_LLOG_ORIGIN_CONNECT);
1756                 rc = ost_llog_handle_connect(req->rq_export, req);
1757                 req->rq_status = rc;
1758                 rc = req_capsule_server_pack(&req->rq_pill);
1759                 if (rc)
1760                         RETURN(rc);
1761                 RETURN(ptlrpc_reply(req));
1762         case OBD_LOG_CANCEL:
1763                 CDEBUG(D_INODE, "log cancel\n");
1764                 req_capsule_set(&req->rq_pill, &RQF_LOG_CANCEL);
1765                 if (OBD_FAIL_CHECK(OBD_FAIL_OBD_LOG_CANCEL_NET))
1766                         RETURN(0);
1767                 rc = llog_origin_handle_cancel(req);
1768                 req->rq_status = rc;
1769                 rc = req_capsule_server_pack(&req->rq_pill);
1770                 if (rc)
1771                         RETURN(rc);
1772                 RETURN(ptlrpc_reply(req));
1773         case LDLM_ENQUEUE:
1774                 CDEBUG(D_INODE, "enqueue\n");
1775                 req_capsule_set(&req->rq_pill, &RQF_LDLM_ENQUEUE);
1776                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_ENQUEUE))
1777                         RETURN(0);
1778                 rc = ldlm_handle_enqueue(req, ldlm_server_completion_ast,
1779                                          ldlm_server_blocking_ast,
1780                                          ldlm_server_glimpse_ast);
1781                 fail = OBD_FAIL_OST_LDLM_REPLY_NET;
1782                 break;
1783         case LDLM_CONVERT:
1784                 CDEBUG(D_INODE, "convert\n");
1785                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CONVERT);
1786                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CONVERT))
1787                         RETURN(0);
1788                 rc = ldlm_handle_convert(req);
1789                 break;
1790         case LDLM_CANCEL:
1791                 CDEBUG(D_INODE, "cancel\n");
1792                 req_capsule_set(&req->rq_pill, &RQF_LDLM_CANCEL);
1793                 if (OBD_FAIL_CHECK(OBD_FAIL_LDLM_CANCEL))
1794                         RETURN(0);
1795                 rc = ldlm_handle_cancel(req);
1796                 break;
1797         case LDLM_BL_CALLBACK:
1798         case LDLM_CP_CALLBACK:
1799                 CDEBUG(D_INODE, "callback\n");
1800                 CERROR("callbacks should not happen on OST\n");
1801                 /* fall through */
1802         default:
1803                 CERROR("Unexpected opcode %d\n",
1804                        lustre_msg_get_opc(req->rq_reqmsg));
1805                 req->rq_status = -ENOTSUPP;
1806                 rc = ptlrpc_error(req);
1807                 RETURN(rc);
1808         }
1809
1810         LASSERT(current->journal_info == NULL);
1811
1812         EXIT;
1813         /* If we're DISCONNECTing, the export_data is already freed */
1814         if (!rc && lustre_msg_get_opc(req->rq_reqmsg) != OST_DISCONNECT)
1815                 target_committed_to_req(req);
1816
1817 out:
1818         if (!rc)
1819                 oti_to_request(oti, req);
1820
1821         target_send_reply(req, rc, fail);
1822         return 0;
1823 }
1824 EXPORT_SYMBOL(ost_handle);
1825 /*
1826  * free per-thread pool created by ost_thread_init().
1827  */
1828 static void ost_thread_done(struct ptlrpc_thread *thread)
1829 {
1830         int i;
1831         struct ost_thread_local_cache *tls; /* TLS stands for Thread-Local
1832                                              * Storage */
1833
1834         ENTRY;
1835
1836         LASSERT(thread != NULL);
1837
1838         /*
1839          * be prepared to handle partially-initialized pools (because this is
1840          * called from ost_thread_init() for cleanup.
1841          */
1842         tls = thread->t_data;
1843         if (tls != NULL) {
1844                 for (i = 0; i < OST_THREAD_POOL_SIZE; ++ i) {
1845                         if (tls->page[i] != NULL)
1846                                 OBD_PAGE_FREE(tls->page[i]);
1847                 }
1848                 OBD_FREE_PTR(tls);
1849                 thread->t_data = NULL;
1850         }
1851         EXIT;
1852 }
1853
1854 /*
1855  * initialize per-thread page pool (bug 5137).
1856  */
1857 static int ost_thread_init(struct ptlrpc_thread *thread)
1858 {
1859         int result;
1860         int i;
1861         struct ost_thread_local_cache *tls;
1862
1863         ENTRY;
1864
1865         LASSERT(thread != NULL);
1866         LASSERT(thread->t_data == NULL);
1867         LASSERTF(thread->t_id <= OSS_THREADS_MAX, "%u\n", thread->t_id);
1868
1869         OBD_ALLOC_PTR(tls);
1870         if (tls != NULL) {
1871                 result = 0;
1872                 thread->t_data = tls;
1873                 /*
1874                  * populate pool
1875                  */
1876                 for (i = 0; i < OST_THREAD_POOL_SIZE; ++ i) {
1877                         OBD_PAGE_ALLOC(tls->page[i], OST_THREAD_POOL_GFP);
1878                         if (tls->page[i] == NULL) {
1879                                 ost_thread_done(thread);
1880                                 result = -ENOMEM;
1881                                 break;
1882                         }
1883                 }
1884         } else
1885                 result = -ENOMEM;
1886         RETURN(result);
1887 }
1888
1889 #define OST_WATCHDOG_TIMEOUT (obd_timeout * 1000)
1890
1891 /* Sigh - really, this is an OSS, the _server_, not the _target_ */
1892 static int ost_setup(struct obd_device *obd, struct lustre_cfg* lcfg)
1893 {
1894         struct ost_obd *ost = &obd->u.ost;
1895         struct lprocfs_static_vars lvars;
1896         int oss_min_threads;
1897         int oss_max_threads;
1898         int oss_min_create_threads;
1899         int oss_max_create_threads;
1900         int rc;
1901         ENTRY;
1902
1903         rc = cleanup_group_info();
1904         if (rc)
1905                 RETURN(rc);
1906
1907         lprocfs_ost_init_vars(&lvars);
1908         lprocfs_obd_setup(obd, lvars.obd_vars);
1909
1910         sema_init(&ost->ost_health_sem, 1);
1911
1912         if (oss_num_threads) {
1913                 /* If oss_num_threads is set, it is the min and the max. */
1914                 if (oss_num_threads > OSS_THREADS_MAX)
1915                         oss_num_threads = OSS_THREADS_MAX;
1916                 if (oss_num_threads < OSS_THREADS_MIN)
1917                         oss_num_threads = OSS_THREADS_MIN;
1918                 oss_max_threads = oss_min_threads = oss_num_threads;
1919         } else {
1920                 /* Base min threads on memory and cpus */
1921                 oss_min_threads = num_possible_cpus() * num_physpages >>
1922                         (27 - CFS_PAGE_SHIFT);
1923                 if (oss_min_threads < OSS_THREADS_MIN)
1924                         oss_min_threads = OSS_THREADS_MIN;
1925                 /* Insure a 4x range for dynamic threads */
1926                 if (oss_min_threads > OSS_THREADS_MAX / 4)
1927                         oss_min_threads = OSS_THREADS_MAX / 4;
1928                 oss_max_threads = min(OSS_THREADS_MAX, oss_min_threads * 4);
1929         }
1930
1931         ost->ost_service =
1932                 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
1933                                 OST_MAXREPSIZE, OST_REQUEST_PORTAL,
1934                                 OSC_REPLY_PORTAL, OSS_SERVICE_WATCHDOG_FACTOR,
1935                                 ost_handle, LUSTRE_OSS_NAME,
1936                                 obd->obd_proc_entry, target_print_req,
1937                                 oss_min_threads, oss_max_threads,
1938                                 "ll_ost", LCT_DT_THREAD);
1939         if (ost->ost_service == NULL) {
1940                 CERROR("failed to start service\n");
1941                 GOTO(out_lprocfs, rc = -ENOMEM);
1942         }
1943
1944         rc = ptlrpc_start_threads(obd, ost->ost_service);
1945         if (rc)
1946                 GOTO(out_service, rc = -EINVAL);
1947
1948         if (oss_num_create_threads) {
1949                 if (oss_num_create_threads > OSS_MAX_CREATE_THREADS)
1950                         oss_num_create_threads = OSS_MAX_CREATE_THREADS;
1951                 if (oss_num_create_threads < OSS_MIN_CREATE_THREADS)
1952                         oss_num_create_threads = OSS_MIN_CREATE_THREADS;
1953                 oss_min_create_threads = oss_max_create_threads =
1954                         oss_num_create_threads;
1955         } else {
1956                 oss_min_create_threads = OSS_MIN_CREATE_THREADS;
1957                 oss_max_create_threads = OSS_MAX_CREATE_THREADS;
1958         }
1959
1960         ost->ost_create_service =
1961                 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
1962                                 OST_MAXREPSIZE, OST_CREATE_PORTAL,
1963                                 OSC_REPLY_PORTAL, OSS_SERVICE_WATCHDOG_FACTOR,
1964                                 ost_handle, "ost_create",
1965                                 obd->obd_proc_entry, target_print_req,
1966                                 oss_min_create_threads, oss_max_create_threads,
1967                                 "ll_ost_creat", LCT_DT_THREAD);
1968         if (ost->ost_create_service == NULL) {
1969                 CERROR("failed to start OST create service\n");
1970                 GOTO(out_service, rc = -ENOMEM);
1971         }
1972
1973         rc = ptlrpc_start_threads(obd, ost->ost_create_service);
1974         if (rc)
1975                 GOTO(out_create, rc = -EINVAL);
1976
1977         ost->ost_io_service =
1978                 ptlrpc_init_svc(OST_NBUFS, OST_BUFSIZE, OST_MAXREQSIZE,
1979                                 OST_MAXREPSIZE, OST_IO_PORTAL,
1980                                 OSC_REPLY_PORTAL, OSS_SERVICE_WATCHDOG_FACTOR,
1981                                 ost_handle, "ost_io",
1982                                 obd->obd_proc_entry, target_print_req,
1983                                 oss_min_threads, oss_max_threads,
1984                                 "ll_ost_io", LCT_DT_THREAD);
1985         if (ost->ost_io_service == NULL) {
1986                 CERROR("failed to start OST I/O service\n");
1987                 GOTO(out_create, rc = -ENOMEM);
1988         }
1989
1990         ost->ost_io_service->srv_init = ost_thread_init;
1991         ost->ost_io_service->srv_done = ost_thread_done;
1992         ost->ost_io_service->srv_cpu_affinity = 1;
1993         rc = ptlrpc_start_threads(obd, ost->ost_io_service);
1994         if (rc)
1995                 GOTO(out_io, rc = -EINVAL);
1996
1997         ping_evictor_start();
1998
1999         RETURN(0);
2000
2001 out_io:
2002         ptlrpc_unregister_service(ost->ost_io_service);
2003         ost->ost_io_service = NULL;
2004 out_create:
2005         ptlrpc_unregister_service(ost->ost_create_service);
2006         ost->ost_create_service = NULL;
2007 out_service:
2008         ptlrpc_unregister_service(ost->ost_service);
2009         ost->ost_service = NULL;
2010 out_lprocfs:
2011         lprocfs_obd_cleanup(obd);
2012         RETURN(rc);
2013 }
2014
2015 static int ost_cleanup(struct obd_device *obd)
2016 {
2017         struct ost_obd *ost = &obd->u.ost;
2018         int err = 0;
2019         ENTRY;
2020
2021         ping_evictor_stop();
2022
2023         spin_lock_bh(&obd->obd_processing_task_lock);
2024         if (obd->obd_recovering) {
2025                 target_cancel_recovery_timer(obd);
2026                 obd->obd_recovering = 0;
2027         }
2028         spin_unlock_bh(&obd->obd_processing_task_lock);
2029
2030         down(&ost->ost_health_sem);
2031         ptlrpc_unregister_service(ost->ost_service);
2032         ptlrpc_unregister_service(ost->ost_create_service);
2033         ptlrpc_unregister_service(ost->ost_io_service);
2034         ost->ost_service = NULL;
2035         ost->ost_create_service = NULL;
2036         up(&ost->ost_health_sem);
2037
2038         lprocfs_obd_cleanup(obd);
2039
2040         RETURN(err);
2041 }
2042
2043 static int ost_health_check(struct obd_device *obd)
2044 {
2045         struct ost_obd *ost = &obd->u.ost;
2046         int rc = 0;
2047
2048         down(&ost->ost_health_sem);
2049         rc |= ptlrpc_service_health_check(ost->ost_service);
2050         rc |= ptlrpc_service_health_check(ost->ost_create_service);
2051         rc |= ptlrpc_service_health_check(ost->ost_io_service);
2052         up(&ost->ost_health_sem);
2053
2054         /*
2055          * health_check to return 0 on healthy
2056          * and 1 on unhealthy.
2057          */
2058         if( rc != 0)
2059                 rc = 1;
2060
2061         return rc;
2062 }
2063
2064 struct ost_thread_local_cache *ost_tls(struct ptlrpc_request *r)
2065 {
2066         return (struct ost_thread_local_cache *)(r->rq_svc_thread->t_data);
2067 }
2068
2069 /* use obd ops to offer management infrastructure */
2070 static struct obd_ops ost_obd_ops = {
2071         .o_owner        = THIS_MODULE,
2072         .o_setup        = ost_setup,
2073         .o_cleanup      = ost_cleanup,
2074         .o_health_check = ost_health_check,
2075 };
2076
2077
2078 static int __init ost_init(void)
2079 {
2080         struct lprocfs_static_vars lvars;
2081         int rc;
2082         ENTRY;
2083
2084         lprocfs_ost_init_vars(&lvars);
2085         rc = class_register_type(&ost_obd_ops, NULL, lvars.module_vars,
2086                                  LUSTRE_OSS_NAME, NULL);
2087
2088         if (ost_num_threads != 0 && oss_num_threads == 0) {
2089                 LCONSOLE_INFO("ost_num_threads module parameter is deprecated, "
2090                               "use oss_num_threads instead or unset both for "
2091                               "dynamic thread startup\n");
2092                 oss_num_threads = ost_num_threads;
2093         }
2094
2095         RETURN(rc);
2096 }
2097
2098 static void /*__exit*/ ost_exit(void)
2099 {
2100         class_unregister_type(LUSTRE_OSS_NAME);
2101 }
2102
2103 MODULE_AUTHOR("Sun Microsystems, Inc. <http://www.lustre.org/>");
2104 MODULE_DESCRIPTION("Lustre Object Storage Target (OST) v0.01");
2105 MODULE_LICENSE("GPL");
2106
2107 module_init(ost_init);
2108 module_exit(ost_exit);