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
  29 extern ptl_handle_eq_t request_out_eq, reply_in_eq, reply_out_eq,
  30         bulk_source_eq, bulk_sink_eq;
  31 static ptl_process_id_t local_id = {PTL_NID_ANY, PTL_PID_ANY};
  32
  33 int ptlrpc_check_bulk_sent(struct ptlrpc_bulk_desc *desc)
  34 {
  35         ENTRY;
  36
  37         if (desc->b_flags & PTL_BULK_FL_SENT)
  38                 RETURN(1);
  39
  40         if (l_killable_pending(current)) {
  41                 desc->b_flags |= PTL_RPC_FL_INTR;
  42                 RETURN(1);
  43         }
  44
  45         CDEBUG(D_NET, "no event yet\n");
  46         RETURN(0);
  47 }
  48
  49 int ptlrpc_check_bulk_received(struct ptlrpc_bulk_desc *desc)
  50 {
  51         ENTRY;
  52
  53         if (desc->b_flags & PTL_BULK_FL_RCVD)
  54                 RETURN(1);
  55
  56         if (l_killable_pending(current)) {
  57                 desc->b_flags |= PTL_RPC_FL_INTR;
  58                 RETURN(1);
  59         }
  60
  61         CDEBUG(D_NET, "no event yet\n");
  62         RETURN(0);
  63 }
  64
  65 static int ptl_send_buf(struct ptlrpc_request *request,
  66                         struct ptlrpc_connection *conn, int portal)
  67 {
  68         int rc;
  69         ptl_process_id_t remote_id;
  70         ptl_handle_md_t md_h;
  71         ptl_ack_req_t ack;
  72
  73         request->rq_req_md.user_ptr = request;
  74
  75         switch (request->rq_type) {
  76         case PTL_RPC_TYPE_REQUEST:
  77                 request->rq_req_md.start = request->rq_reqmsg;
  78                 request->rq_req_md.length = request->rq_reqlen;
  79                 request->rq_req_md.eventq = request_out_eq;
  80                 request->rq_req_md.threshold = 1;
  81                 ack = PTL_NOACK_REQ;
  82                 break;
  83         case PTL_RPC_TYPE_REPLY:
  84                 request->rq_req_md.start = request->rq_repmsg;
  85                 request->rq_req_md.length = request->rq_replen;
  86                 request->rq_req_md.eventq = reply_out_eq;
  87                 request->rq_req_md.threshold = 1;
  88                 ack = PTL_NOACK_REQ;
  89                 break;
  90         default:
  91                 LBUG();
  92                 return -1; /* notreached */
  93         }
  94         request->rq_req_md.options = PTL_MD_OP_PUT;
  95         request->rq_req_md.user_ptr = request;
  96
  97         rc = PtlMDBind(conn->c_peer.peer_ni, request->rq_req_md, &md_h);
  98         //CERROR("MDBind (outgoing req/rep/bulk): %Lu\n", (__u64)md_h);
  99         if (rc != 0) {
 100                 CERROR("PtlMDBind failed: %d\n", rc);
 101                 LBUG();
 102                 return rc;
 103         }
 104
 105         remote_id.nid = conn->c_peer.peer_nid;
 106         remote_id.pid = 0;
 107
 108         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %Ld\n",
 109                request->rq_req_md.length, portal, request->rq_xid);
 110
 111         rc = PtlPut(md_h, ack, remote_id, portal, 0, request->rq_xid,
 112                     0, 0);
 113         if (rc != PTL_OK) {
 114                 CERROR("PtlPut(%Lu, %d, %Ld) failed: %d\n", remote_id.nid,
 115                        portal, request->rq_xid, rc);
 116                 PtlMDUnlink(md_h);
 117         }
 118
 119         return rc;
 120 }
 121
 122 static inline struct iovec *
 123 ptlrpc_get_bulk_iov (struct ptlrpc_bulk_desc *desc)
 124 {
 125         struct iovec *iov;
 126
 127         if (desc->b_page_count <= sizeof (desc->b_iov)/sizeof (struct iovec))
 128                 return (desc->b_iov);
 129
 130         OBD_ALLOC (iov, desc->b_page_count * sizeof (struct iovec));
 131         if (iov == NULL)
 132                 LBUG();
 133
 134         return (iov);
 135 }
 136
 137 static inline void
 138 ptlrpc_put_bulk_iov (struct ptlrpc_bulk_desc *desc, struct iovec *iov)
 139 {
 140         if (desc->b_page_count <= sizeof (desc->b_iov)/sizeof (struct iovec))
 141                 return;
 142
 143         OBD_FREE (iov, desc->b_page_count * sizeof (struct iovec));
 144 }
 145
 146 int ptlrpc_send_bulk(struct ptlrpc_bulk_desc *desc)
 147 {
 148         int rc;
 149         struct list_head *tmp, *next;
 150         ptl_process_id_t remote_id;
 151         __u32 xid = 0;
 152         struct iovec *iov;
 153         ENTRY;
 154
 155         iov = ptlrpc_get_bulk_iov (desc);
 156         if (iov == NULL)
 157                 RETURN (-ENOMEM);
 158
 159         desc->b_md.start = iov;
 160         desc->b_md.niov = 0;
 161         desc->b_md.length = 0;
 162         desc->b_md.eventq = bulk_source_eq;
 163         desc->b_md.threshold = 2; /* SENT and ACK */
 164         desc->b_md.options = PTL_MD_OP_PUT | PTL_MD_IOV;
 165         desc->b_md.user_ptr = desc;
 166
 167         list_for_each_safe(tmp, next, &desc->b_page_list) {
 168                 struct ptlrpc_bulk_page *bulk;
 169                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, b_link);
 170
 171                 LASSERT (desc->b_md.niov < desc->b_page_count);
 172
 173                 if (desc->b_md.niov == 0)
 174                         xid = bulk->b_xid;
 175                 LASSERT (xid == bulk->b_xid);   /* should all be the same */
 176
 177                 iov[desc->b_md.niov].iov_base = bulk->b_buf;
 178                 iov[desc->b_md.niov].iov_len = bulk->b_buflen;
 179                 desc->b_md.niov++;
 180                 desc->b_md.length += bulk->b_buflen;
 181         }
 182
 183         LASSERT (desc->b_md.niov == desc->b_page_count);
 184         LASSERT (desc->b_md.niov != 0);
 185
 186         rc = PtlMDBind(desc->b_connection->c_peer.peer_ni, desc->b_md,
 187                        &desc->b_md_h);
 188
 189         ptlrpc_put_bulk_iov (desc, iov);        /* move down to reduce latency to send */
 190
 191         if (rc != PTL_OK) {
 192                 CERROR("PtlMDBind failed: %d\n", rc);
 193                 LBUG();
 194                 RETURN(rc);
 195         }
 196
 197         remote_id.nid = desc->b_connection->c_peer.peer_nid;
 198         remote_id.pid = 0;
 199
 200         CDEBUG(D_NET, "Sending %u pages %u bytes to portal %d nid %Lx pid %d xid %d\n",
 201                desc->b_md.niov, desc->b_md.length,
 202                desc->b_portal, remote_id.nid, remote_id.pid, xid);
 203
 204         rc = PtlPut(desc->b_md_h, PTL_ACK_REQ, remote_id,
 205                     desc->b_portal, 0, xid, 0, 0);
 206         if (rc != PTL_OK) {
 207                 CERROR("PtlPut(%Lu, %d, %d) failed: %d\n",
 208                        remote_id.nid, desc->b_portal, xid, rc);
 209                 PtlMDUnlink(desc->b_md_h);
 210                 LBUG();
 211                 RETURN(rc);
 212         }
 213
 214         RETURN(0);
 215 }
 216
 217 int ptlrpc_register_bulk(struct ptlrpc_bulk_desc *desc)
 218 {
 219         struct list_head *tmp, *next;
 220         int rc;
 221         __u32 xid = 0;
 222         struct iovec *iov;
 223         ENTRY;
 224
 225         iov = ptlrpc_get_bulk_iov (desc);
 226         if (iov == NULL)
 227                 return (-ENOMEM);
 228
 229         desc->b_md.start = iov;
 230         desc->b_md.niov = 0;
 231         desc->b_md.length = 0;
 232         desc->b_md.threshold = 1;
 233         desc->b_md.options = PTL_MD_OP_PUT | PTL_MD_IOV;
 234         desc->b_md.user_ptr = desc;
 235         desc->b_md.eventq = bulk_sink_eq;
 236
 237         list_for_each_safe(tmp, next, &desc->b_page_list) {
 238                 struct ptlrpc_bulk_page *bulk;
 239                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, b_link);
 240
 241                 LASSERT (desc->b_md.niov < desc->b_page_count);
 242
 243                 if (desc->b_md.niov == 0)
 244                         xid = bulk->b_xid;
 245                 LASSERT (xid == bulk->b_xid);   /* should all be the same */
 246
 247                 iov[desc->b_md.niov].iov_base = bulk->b_buf;
 248                 iov[desc->b_md.niov].iov_len = bulk->b_buflen;
 249                 desc->b_md.niov++;
 250                 desc->b_md.length += bulk->b_buflen;
 251         }
 252
 253         LASSERT (desc->b_md.niov == desc->b_page_count);
 254         LASSERT (desc->b_md.niov != 0);
 255
 256         rc = PtlMEAttach(desc->b_connection->c_peer.peer_ni,
 257                          desc->b_portal, local_id, xid, 0,
 258                          PTL_UNLINK, PTL_INS_AFTER, &desc->b_me_h);
 259
 260         ptlrpc_put_bulk_iov (desc, iov);
 261
 262         if (rc != PTL_OK) {
 263                 CERROR("PtlMEAttach failed: %d\n", rc);
 264                 LBUG();
 265                 GOTO(cleanup, rc);
 266         }
 267
 268         rc = PtlMDAttach(desc->b_me_h, desc->b_md, PTL_UNLINK,
 269                          &desc->b_md_h);
 270         if (rc != PTL_OK) {
 271                 CERROR("PtlMDAttach failed: %d\n", rc);
 272                 LBUG();
 273                 GOTO(cleanup, rc);
 274         }
 275
 276         CDEBUG(D_NET, "Setup bulk sink buffers: %u pages %u bytes, xid %u, "
 277                "portal %u\n", desc->b_md.niov, desc->b_md.length,
 278                xid, desc->b_portal);
 279
 280         RETURN(0);
 281
 282  cleanup:
 283         ptlrpc_abort_bulk(desc);
 284
 285         return rc;
 286 }
 287
 288 int ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
 289 {
 290         /* This should be safe: these handles are initialized to be
 291          * invalid in ptlrpc_prep_bulk() */
 292         PtlMDUnlink(desc->b_md_h);
 293         PtlMEUnlink(desc->b_me_h);
 294
 295         return 0;
 296 }
 297
 298 int ptlrpc_reply(struct ptlrpc_service *svc, struct ptlrpc_request *req)
 299 {
 300         if (req->rq_repmsg == NULL) {
 301                 CERROR("bad: someone called ptlrpc_reply when they meant "
 302                        "ptlrpc_error\n");
 303                 return -EINVAL;
 304         }
 305
 306         /* FIXME: we need to increment the count of handled events */
 307         req->rq_type = PTL_RPC_TYPE_REPLY;
 308         //req->rq_repmsg->conn = req->rq_connection->c_remote_conn;
 309         //req->rq_repmsg->token = req->rq_connection->c_remote_token;
 310         req->rq_repmsg->status = HTON__u32(req->rq_status);
 311         req->rq_reqmsg->type = HTON__u32(req->rq_type);
 312         return ptl_send_buf(req, req->rq_connection, svc->srv_rep_portal);
 313 }
 314
 315 int ptlrpc_error(struct ptlrpc_service *svc, struct ptlrpc_request *req)
 316 {
 317         int rc;
 318         ENTRY;
 319
 320         if (req->rq_repmsg) {
 321                 CERROR("req already has repmsg\n");
 322                 LBUG();
 323         }
 324
 325         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 326         if (rc)
 327                 RETURN(rc);
 328
 329         req->rq_repmsg->type = HTON__u32(PTL_RPC_MSG_ERR);
 330
 331         rc = ptlrpc_reply(svc, req);
 332         RETURN(rc);
 333 }
 334
 335
 336 int ptl_send_rpc(struct ptlrpc_request *request)
 337 {
 338         int rc;
 339         char *repbuf;
 340
 341         ENTRY;
 342
 343         if (NTOH__u32(request->rq_reqmsg->type) != PTL_RPC_MSG_REQUEST) {
 344                 CERROR("wrong packet type sent %d\n",
 345                        NTOH__u32(request->rq_reqmsg->type));
 346                 LBUG();
 347                 RETURN(EINVAL);
 348         }
 349         if (request->rq_replen == 0) {
 350                 CERROR("request->rq_replen is 0!\n");
 351                 RETURN(EINVAL);
 352         }
 353
 354         /* request->rq_repmsg is set only when the reply comes in, in
 355          * client_packet_callback() */
 356         if (request->rq_reply_md.start)
 357                 OBD_FREE(request->rq_reply_md.start, request->rq_replen);
 358
 359         OBD_ALLOC(repbuf, request->rq_replen);
 360         if (!repbuf) {
 361                 LBUG();
 362                 RETURN(ENOMEM);
 363         }
 364
 365         down(&request->rq_client->cli_rpc_sem);
 366
 367         rc = PtlMEAttach(request->rq_connection->c_peer.peer_ni,
 368                          request->rq_client->cli_reply_portal,
 369                          local_id, request->rq_xid, 0, PTL_UNLINK,
 370                          PTL_INS_AFTER, &request->rq_reply_me_h);
 371         if (rc != PTL_OK) {
 372                 CERROR("PtlMEAttach failed: %d\n", rc);
 373                 LBUG();
 374                 GOTO(cleanup, rc);
 375         }
 376
 377         request->rq_type = PTL_RPC_TYPE_REQUEST;
 378         request->rq_reply_md.start = repbuf;
 379         request->rq_reply_md.length = request->rq_replen;
 380         request->rq_reply_md.threshold = 1;
 381         request->rq_reply_md.options = PTL_MD_OP_PUT;
 382         request->rq_reply_md.user_ptr = request;
 383         request->rq_reply_md.eventq = reply_in_eq;
 384
 385         rc = PtlMDAttach(request->rq_reply_me_h, request->rq_reply_md,
 386                          PTL_UNLINK, &request->rq_reply_md_h);
 387         if (rc != PTL_OK) {
 388                 CERROR("PtlMDAttach failed: %d\n", rc);
 389                 LBUG();
 390                 GOTO(cleanup2, rc);
 391         }
 392
 393         CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %Lu, portal %u\n",
 394                request->rq_replen, request->rq_xid,
 395                request->rq_client->cli_reply_portal);
 396
 397         rc = ptl_send_buf(request, request->rq_connection,
 398                           request->rq_client->cli_request_portal);
 399         RETURN(rc);
 400
 401  cleanup2:
 402         PtlMEUnlink(request->rq_reply_me_h);
 403  cleanup:
 404         OBD_FREE(repbuf, request->rq_replen);
 405         up(&request->rq_client->cli_rpc_sem);
 406
 407         return rc;
 408 }
 409
 410 void ptlrpc_link_svc_me(struct ptlrpc_service *service, int i)
 411 {
 412         int rc;
 413         ptl_md_t dummy;
 414         ptl_handle_md_t md_h;
 415
 416         /* Attach the leading ME on which we build the ring */
 417         rc = PtlMEAttach(service->srv_self.peer_ni, service->srv_req_portal,
 418                          local_id, 0, ~0, PTL_RETAIN, PTL_INS_BEFORE,
 419                          &(service->srv_me_h[i]));
 420         if (rc != PTL_OK) {
 421                 CERROR("PtlMEAttach failed: %d\n", rc);
 422                 LBUG();
 423         }
 424
 425         if (service->srv_ref_count[i])
 426                 LBUG();
 427
 428         dummy.start         = service->srv_buf[i];
 429         dummy.length        = service->srv_buf_size;
 430         dummy.max_offset    = service->srv_buf_size;
 431         dummy.threshold     = PTL_MD_THRESH_INF;
 432         dummy.options       = PTL_MD_OP_PUT | PTL_MD_AUTO_UNLINK;
 433         dummy.user_ptr      = service;
 434         dummy.eventq        = service->srv_eq_h;
 435         dummy.max_offset    = service->srv_buf_size;
 436
 437         rc = PtlMDAttach(service->srv_me_h[i], dummy, PTL_UNLINK, &md_h);
 438         if (rc != PTL_OK) {
 439                 /* cleanup */
 440                 CERROR("PtlMDAttach failed: %d\n", rc);
 441                 LBUG();
 442         }
 443 }
 444
 445 /* ptl_handled_rpc() should be called by the sleeping process once
 446  * it finishes processing an event.  This ensures the ref count is
 447  * decremented and that the rpc ring buffer cycles properly.
 448  */
 449 int ptl_handled_rpc(struct ptlrpc_service *service, void *start)
 450 {
 451         int index;
 452
 453         spin_lock(&service->srv_lock);
 454         for (index = 0; index < service->srv_ring_length; index++)
 455                 if (service->srv_buf[index] == start)
 456                         break;
 457
 458         if (index == service->srv_ring_length)
 459                 LBUG();
 460
 461         CDEBUG(D_INFO, "MD index=%d Ref Count=%d\n", index,
 462                service->srv_ref_count[index]);
 463         service->srv_ref_count[index]--;
 464
 465         if (service->srv_ref_count[index] < 0)
 466                 LBUG();
 467
 468         if (service->srv_ref_count[index] == 0 &&
 469             !ptl_is_valid_handle(&(service->srv_me_h[index]))) {
 470                 CDEBUG(D_NET, "relinking %d\n", index);
 471                 ptlrpc_link_svc_me(service, index);
 472         }
 473
 474         spin_unlock(&service->srv_lock);
 475         return 0;
 476 }