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