lustre/ptlrpc/niobuf.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (C) 2002 Cluster File Systems, Inc.
   5  *
   6  *   This file is part of Lustre, http://www.lustre.org.
   7  *
   8  *   Lustre is free software; you can redistribute it and/or
   9  *   modify it under the terms of version 2 of the GNU General Public
  10  *   License as published by the Free Software Foundation.
  11  *
  12  *   Lustre is distributed in the hope that it will be useful,
  13  *   but WITHOUT ANY WARRANTY; without even the implied warranty of
  14  *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  15  *   GNU General Public License for more details.
  16  *
  17  *   You should have received a copy of the GNU General Public License
  18  *   along with Lustre; if not, write to the Free Software
  19  *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  20  *
  21  */
  22
  23 #define DEBUG_SUBSYSTEM S_RPC
  24
  25 #include <linux/obd_support.h>
  26 #include <linux/lustre_net.h>
  27 #include <linux/lustre_lib.h>
  28 #include <linux/obd.h>
  29
  30 extern ptl_handle_eq_t request_out_eq, reply_in_eq, reply_out_eq,
  31         bulk_source_eq, bulk_sink_eq;
  32
  33 static int ptl_send_buf(struct ptlrpc_request *request,
  34                         struct ptlrpc_connection *conn, int portal)
  35 {
  36         int rc;
  37         ptl_process_id_t remote_id;
  38         ptl_handle_md_t md_h;
  39
  40         LASSERT(conn);
  41
  42         request->rq_req_md.user_ptr = request;
  43
  44         switch (request->rq_type) {
  45         case PTL_RPC_MSG_REQUEST:
  46                 request->rq_reqmsg->type = HTON__u32(request->rq_type);
  47                 request->rq_req_md.start = request->rq_reqmsg;
  48                 request->rq_req_md.length = request->rq_reqlen;
  49                 request->rq_req_md.eventq = request_out_eq;
  50                 break;
  51         case PTL_RPC_MSG_ERR:
  52         case PTL_RPC_MSG_REPLY:
  53                 request->rq_repmsg->type = HTON__u32(request->rq_type);
  54                 request->rq_req_md.start = request->rq_repmsg;
  55                 request->rq_req_md.length = request->rq_replen;
  56                 request->rq_req_md.eventq = reply_out_eq;
  57                 break;
  58         default:
  59                 LBUG();
  60                 return -1; /* notreached */
  61         }
  62         request->rq_req_md.threshold = 1;
  63         request->rq_req_md.options = PTL_MD_OP_PUT;
  64         request->rq_req_md.user_ptr = request;
  65
  66         rc = PtlMDBind(conn->c_peer.peer_ni, request->rq_req_md, &md_h);
  67         if (rc != 0) {
  68                 CERROR("PtlMDBind failed: %d\n", rc);
  69                 LBUG();
  70                 return rc;
  71         }
  72
  73         remote_id.nid = conn->c_peer.peer_nid;
  74         remote_id.pid = 0;
  75
  76         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid "LPD64"\n",
  77                request->rq_req_md.length, portal, request->rq_xid);
  78
  79         if (!portal)
  80                 LBUG();
  81         rc = PtlPut(md_h, PTL_NOACK_REQ, remote_id, portal, 0, request->rq_xid,
  82                     0, 0);
  83         if (rc != PTL_OK) {
  84                 CERROR("PtlPut("LPU64", %d, "LPD64") failed: %d\n",
  85                        remote_id.nid, portal, request->rq_xid, rc);
  86                 PtlMDUnlink(md_h);
  87         }
  88
  89         return rc;
  90 }
  91
  92 static inline struct iovec *
  93 ptlrpc_get_bulk_iov (struct ptlrpc_bulk_desc *desc)
  94 {
  95         struct iovec *iov;
  96
  97         if (desc->bd_page_count <= sizeof (desc->bd_iov)/sizeof (struct iovec))
  98                 return (desc->bd_iov);
  99
 100         OBD_ALLOC (iov, desc->bd_page_count * sizeof (struct iovec));
 101         if (iov == NULL)
 102                 LBUG();
 103
 104         return (iov);
 105 }
 106
 107 static inline void
 108 ptlrpc_put_bulk_iov (struct ptlrpc_bulk_desc *desc, struct iovec *iov)
 109 {
 110         if (desc->bd_page_count <= sizeof (desc->bd_iov)/sizeof (struct iovec))
 111                 return;
 112
 113         OBD_FREE (iov, desc->bd_page_count * sizeof (struct iovec));
 114 }
 115
 116 int ptlrpc_send_bulk(struct ptlrpc_bulk_desc *desc)
 117 {
 118         int rc;
 119         struct list_head *tmp, *next;
 120         ptl_process_id_t remote_id;
 121         __u32 xid = 0;
 122         struct iovec *iov;
 123         ENTRY;
 124
 125         iov = ptlrpc_get_bulk_iov (desc);
 126         if (iov == NULL)
 127                 RETURN (-ENOMEM);
 128
 129         desc->bd_md.start = iov;
 130         desc->bd_md.niov = 0;
 131         desc->bd_md.length = 0;
 132         desc->bd_md.eventq = bulk_source_eq;
 133         desc->bd_md.threshold = 2; /* SENT and ACK */
 134         desc->bd_md.options = PTL_MD_OP_PUT | PTL_MD_IOV;
 135         desc->bd_md.user_ptr = desc;
 136
 137         atomic_set (&desc->bd_source_callback_count, 2);
 138
 139         list_for_each_safe(tmp, next, &desc->bd_page_list) {
 140                 struct ptlrpc_bulk_page *bulk;
 141                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
 142
 143                 LASSERT (desc->bd_md.niov < desc->bd_page_count);
 144
 145                 if (desc->bd_md.niov == 0)
 146                         xid = bulk->bp_xid;
 147                 LASSERT (xid == bulk->bp_xid);   /* should all be the same */
 148
 149                 iov[desc->bd_md.niov].iov_base = bulk->bp_buf;
 150                 iov[desc->bd_md.niov].iov_len = bulk->bp_buflen;
 151                 desc->bd_md.niov++;
 152                 desc->bd_md.length += bulk->bp_buflen;
 153         }
 154
 155         LASSERT (desc->bd_md.niov == desc->bd_page_count);
 156         LASSERT (desc->bd_md.niov != 0);
 157
 158         rc = PtlMDBind(desc->bd_connection->c_peer.peer_ni, desc->bd_md,
 159                        &desc->bd_md_h);
 160
 161         ptlrpc_put_bulk_iov (desc, iov); /*move down to reduce latency to send*/
 162
 163         if (rc != PTL_OK) {
 164                 CERROR("PtlMDBind failed: %d\n", rc);
 165                 LBUG();
 166                 RETURN(rc);
 167         }
 168
 169         remote_id.nid = desc->bd_connection->c_peer.peer_nid;
 170         remote_id.pid = 0;
 171
 172         CDEBUG(D_NET, "Sending %u pages %u bytes to portal %d nid "LPX64" pid "
 173                "%d xid %d\n", desc->bd_md.niov, desc->bd_md.length,
 174                desc->bd_portal, remote_id.nid, remote_id.pid, xid);
 175
 176         rc = PtlPut(desc->bd_md_h, PTL_ACK_REQ, remote_id,
 177                     desc->bd_portal, 0, xid, 0, 0);
 178         if (rc != PTL_OK) {
 179                 CERROR("PtlPut("LPU64", %d, %d) failed: %d\n",
 180                        remote_id.nid, desc->bd_portal, xid, rc);
 181                 PtlMDUnlink(desc->bd_md_h);
 182                 LBUG();
 183                 RETURN(rc);
 184         }
 185
 186         RETURN(0);
 187 }
 188
 189 int ptlrpc_register_bulk(struct ptlrpc_bulk_desc *desc)
 190 {
 191         struct list_head *tmp, *next;
 192         int rc;
 193         __u32 xid = 0;
 194         struct iovec *iov;
 195         ptl_process_id_t source_id;
 196         ENTRY;
 197
 198         if (desc->bd_page_count > PTL_MD_MAX_IOV) {
 199                 CERROR("iov longer than %d pages not supported (count=%d)\n",
 200                        PTL_MD_MAX_IOV, desc->bd_page_count);
 201                 RETURN(-EINVAL);
 202         }
 203
 204         iov = ptlrpc_get_bulk_iov (desc);
 205         if (iov == NULL)
 206                 return (-ENOMEM);
 207
 208         desc->bd_md.start = iov;
 209         desc->bd_md.niov = 0;
 210         desc->bd_md.length = 0;
 211         desc->bd_md.threshold = 1;
 212         desc->bd_md.options = PTL_MD_OP_PUT | PTL_MD_IOV;
 213         desc->bd_md.user_ptr = desc;
 214         desc->bd_md.eventq = bulk_sink_eq;
 215
 216         list_for_each_safe(tmp, next, &desc->bd_page_list) {
 217                 struct ptlrpc_bulk_page *bulk;
 218                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
 219
 220                 LASSERT (desc->bd_md.niov < desc->bd_page_count);
 221
 222                 if (desc->bd_md.niov == 0)
 223                         xid = bulk->bp_xid;
 224                 LASSERT (xid == bulk->bp_xid);   /* should all be the same */
 225
 226                 iov[desc->bd_md.niov].iov_base = bulk->bp_buf;
 227                 iov[desc->bd_md.niov].iov_len = bulk->bp_buflen;
 228                 desc->bd_md.niov++;
 229                 desc->bd_md.length += bulk->bp_buflen;
 230         }
 231
 232         LASSERT (desc->bd_md.niov == desc->bd_page_count);
 233         LASSERT (desc->bd_md.niov != 0);
 234
 235         source_id.nid = desc->bd_connection->c_peer.peer_nid;
 236         source_id.pid = PTL_PID_ANY;
 237
 238         rc = PtlMEAttach(desc->bd_connection->c_peer.peer_ni,
 239                          desc->bd_portal, source_id, xid, 0,
 240                          PTL_UNLINK, PTL_INS_AFTER, &desc->bd_me_h);
 241
 242         if (rc != PTL_OK) {
 243                 CERROR("PtlMEAttach failed: %d\n", rc);
 244                 LBUG();
 245                 GOTO(cleanup, rc);
 246         }
 247
 248         rc = PtlMDAttach(desc->bd_me_h, desc->bd_md, PTL_UNLINK,
 249                          &desc->bd_md_h);
 250         if (rc != PTL_OK) {
 251                 CERROR("PtlMDAttach failed: %d\n", rc);
 252                 LBUG();
 253                 GOTO(cleanup, rc);
 254         }
 255
 256         ptlrpc_put_bulk_iov (desc, iov);
 257
 258         CDEBUG(D_NET, "Setup bulk sink buffers: %u pages %u bytes, xid %u, "
 259                "portal %u\n", desc->bd_md.niov, desc->bd_md.length,
 260                xid, desc->bd_portal);
 261
 262         RETURN(0);
 263
 264  cleanup:
 265         ptlrpc_put_bulk_iov (desc, iov);
 266         ptlrpc_abort_bulk(desc);
 267
 268         return rc;
 269 }
 270
 271 int ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
 272 {
 273         /* This should be safe: these handles are initialized to be
 274          * invalid in ptlrpc_prep_bulk() */
 275         PtlMDUnlink(desc->bd_md_h);
 276         PtlMEUnlink(desc->bd_me_h);
 277
 278         return 0;
 279 }
 280
 281 void obd_brw_set_add(struct obd_brw_set *set, struct ptlrpc_bulk_desc *desc)
 282 {
 283         atomic_inc(&desc->bd_refcount);
 284         atomic_inc(&set->brw_refcount);
 285         desc->bd_brw_set = set;
 286         list_add(&desc->bd_set_chain, &set->brw_desc_head);
 287 }
 288
 289 struct obd_brw_set *obd_brw_set_new(void)
 290 {
 291         struct obd_brw_set *set;
 292
 293         OBD_ALLOC(set, sizeof(*set));
 294
 295         if (set != NULL) {
 296                 init_waitqueue_head(&set->brw_waitq);
 297                 INIT_LIST_HEAD(&set->brw_desc_head);
 298                 atomic_set(&set->brw_refcount, 0);
 299         }
 300
 301         return set;
 302 }
 303
 304 void obd_brw_set_free(struct obd_brw_set *set)
 305 {
 306         struct list_head *tmp, *next;
 307         ENTRY;
 308
 309         if (!list_empty(&set->brw_desc_head)) {
 310                 EXIT;
 311                 return;
 312         }
 313
 314         list_for_each_safe(tmp, next, &set->brw_desc_head) {
 315                 struct ptlrpc_bulk_desc *desc =
 316                         list_entry(tmp, struct ptlrpc_bulk_desc, bd_set_chain);
 317
 318                 CERROR("Unfinished bulk descriptor: %p\n", desc);
 319
 320                 ptlrpc_abort_bulk(desc);
 321         }
 322         OBD_FREE(set, sizeof(*set));
 323         EXIT;
 324         return;
 325 }
 326
 327 int ptlrpc_reply(struct ptlrpc_service *svc, struct ptlrpc_request *req)
 328 {
 329         if (req->rq_repmsg == NULL) {
 330                 CERROR("bad: someone called ptlrpc_reply when they meant "
 331                        "ptlrpc_error\n");
 332                 return -EINVAL;
 333         }
 334
 335         /* FIXME: we need to increment the count of handled events */
 336         if (req->rq_type != PTL_RPC_MSG_ERR)
 337                 req->rq_type = PTL_RPC_MSG_REPLY;
 338         //req->rq_repmsg->conn = req->rq_connection->c_remote_conn;
 339         //req->rq_repmsg->token = req->rq_connection->c_remote_token;
 340         req->rq_repmsg->status = HTON__u32(req->rq_status);
 341         return ptl_send_buf(req, req->rq_connection, svc->srv_rep_portal);
 342 }
 343
 344 int ptlrpc_error(struct ptlrpc_service *svc, struct ptlrpc_request *req)
 345 {
 346         int rc;
 347         ENTRY;
 348
 349         if (req->rq_repmsg) {
 350                 CERROR("req already has repmsg\n");
 351                 LBUG();
 352         }
 353
 354         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 355         if (rc)
 356                 RETURN(rc);
 357
 358         req->rq_type = PTL_RPC_MSG_ERR;
 359
 360         rc = ptlrpc_reply(svc, req);
 361         RETURN(rc);
 362 }
 363
 364 int ptl_send_rpc(struct ptlrpc_request *request)
 365 {
 366         int rc;
 367         char *repbuf;
 368         ptl_process_id_t source_id;
 369
 370         ENTRY;
 371
 372         if (request->rq_type != PTL_RPC_MSG_REQUEST) {
 373                 CERROR("wrong packet type sent %d\n",
 374                        NTOH__u32(request->rq_reqmsg->type));
 375                 LBUG();
 376                 RETURN(EINVAL);
 377         }
 378
 379         source_id.nid = request->rq_connection->c_peer.peer_nid;
 380         source_id.pid = PTL_PID_ANY;
 381
 382         /* add a ref, which will be balanced in request_out_callback */
 383         atomic_inc(&request->rq_refcount);
 384         if (request->rq_replen != 0) {
 385                 /* request->rq_repmsg is set only when the reply comes in, in
 386                  * client_packet_callback() */
 387                 if (request->rq_reply_md.start) {
 388                         PtlMEUnlink(request->rq_reply_me_h);
 389                         OBD_FREE(request->rq_reply_md.start,
 390                                  request->rq_replen);
 391                         /* If we're resending, rq_repmsg needs to be NULLed out
 392                          * again so that ptlrpc_check_reply doesn't trip early.
 393                          */
 394                         request->rq_repmsg = NULL;
 395                 }
 396                 OBD_ALLOC(repbuf, request->rq_replen);
 397                 if (!repbuf) {
 398                         LBUG();
 399                         RETURN(ENOMEM);
 400                 }
 401
 402                 rc = PtlMEAttach(request->rq_connection->c_peer.peer_ni,
 403                              request->rq_reply_portal,/* XXX FIXME bug 625069 */
 404                                  source_id, request->rq_xid, 0, PTL_UNLINK,
 405                                  PTL_INS_AFTER, &request->rq_reply_me_h);
 406                 if (rc != PTL_OK) {
 407                         CERROR("PtlMEAttach failed: %d\n", rc);
 408                         LBUG();
 409                         GOTO(cleanup, rc);
 410                 }
 411
 412                 request->rq_reply_md.start = repbuf;
 413                 request->rq_reply_md.length = request->rq_replen;
 414                 request->rq_reply_md.threshold = 1;
 415                 request->rq_reply_md.options = PTL_MD_OP_PUT;
 416                 request->rq_reply_md.user_ptr = request;
 417                 request->rq_reply_md.eventq = reply_in_eq;
 418
 419                 rc = PtlMDAttach(request->rq_reply_me_h, request->rq_reply_md,
 420                                  PTL_UNLINK, NULL);
 421                 if (rc != PTL_OK) {
 422                         CERROR("PtlMDAttach failed: %d\n", rc);
 423                         LBUG();
 424                         GOTO(cleanup2, rc);
 425                 }
 426
 427                 CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid "LPU64
 428                        ", portal %u\n",
 429                        request->rq_replen, request->rq_xid,
 430                        request->rq_reply_portal);
 431         }
 432
 433         /* Clear any flags that may be present from previous sends,
 434          * except for REPLAY. */
 435         request->rq_flags &= PTL_RPC_FL_REPLAY;
 436         rc = ptl_send_buf(request, request->rq_connection,
 437                           request->rq_request_portal);
 438         RETURN(rc);
 439
 440  cleanup2:
 441         PtlMEUnlink(request->rq_reply_me_h);
 442  cleanup:
 443         OBD_FREE(repbuf, request->rq_replen);
 444         // up(&request->rq_client->cli_rpc_sem);
 445
 446         return rc;
 447 }
 448
 449 void ptlrpc_link_svc_me(struct ptlrpc_request_buffer_desc *rqbd)
 450 {
 451         struct ptlrpc_service *service = rqbd->rqbd_service;
 452         static ptl_process_id_t match_id = {PTL_NID_ANY, PTL_PID_ANY};
 453         int rc;
 454         ptl_md_t dummy;
 455         ptl_handle_md_t md_h;
 456
 457         LASSERT (atomic_read (&rqbd->rqbd_refcount) == 0);
 458
 459         /* Attach the leading ME on which we build the ring */
 460         rc = PtlMEAttach(service->srv_self.peer_ni, service->srv_req_portal,
 461                          match_id, 0, ~0,
 462                          PTL_UNLINK, PTL_INS_AFTER, &rqbd->rqbd_me_h);
 463         if (rc != PTL_OK) {
 464                 CERROR("PtlMEAttach failed: %d\n", rc);
 465                 LBUG();
 466         }
 467
 468         dummy.start      = rqbd->rqbd_buffer;
 469         dummy.length     = service->srv_buf_size;
 470         dummy.max_size   = service->srv_max_req_size;
 471         dummy.threshold  = PTL_MD_THRESH_INF;
 472         dummy.options    = PTL_MD_OP_PUT | PTL_MD_MAX_SIZE | PTL_MD_AUTO_UNLINK;
 473         dummy.user_ptr   = rqbd;
 474         dummy.eventq     = service->srv_eq_h;
 475
 476         atomic_inc (&service->srv_nrqbds_receiving);
 477         atomic_set (&rqbd->rqbd_refcount, 1);   /* 1 ref for portals */
 478
 479         rc = PtlMDAttach(rqbd->rqbd_me_h, dummy, PTL_UNLINK, &md_h);
 480         if (rc != PTL_OK) {
 481                 CERROR("PtlMDAttach failed: %d\n", rc);
 482                 LBUG();
 483 #warning proper cleanup required
 484                 PtlMEUnlink (rqbd->rqbd_me_h);
 485                 atomic_set (&rqbd->rqbd_refcount, 0);
 486                 atomic_dec (&service->srv_nrqbds_receiving);
 487         }
 488 }