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/lustre_net.h>
  26
  27 extern ptl_handle_eq_t request_out_eq, reply_in_eq, reply_out_eq,
  28         bulk_source_eq, bulk_sink_eq;
  29 static ptl_process_id_t local_id = {PTL_ID_ANY, PTL_ID_ANY};
  30
  31 int ptlrpc_check_bulk_sent(struct ptlrpc_bulk_desc *bulk)
  32 {
  33         ENTRY;
  34
  35         if (bulk->b_flags & PTL_BULK_FL_SENT)
  36                 RETURN(1);
  37
  38         if (sigismember(&(current->pending.signal), SIGKILL) ||
  39             sigismember(&(current->pending.signal), SIGINT)) {
  40                 bulk->b_flags |= PTL_RPC_FL_INTR;
  41                 RETURN(1);
  42         }
  43
  44         CDEBUG(D_NET, "no event yet\n");
  45         RETURN(0);
  46 }
  47
  48 static int ptl_send_buf(struct ptlrpc_request *request,
  49                         struct ptlrpc_connection *conn, int portal)
  50 {
  51         int rc;
  52         ptl_process_id_t remote_id;
  53         ptl_handle_md_t md_h;
  54         ptl_ack_req_t ack;
  55
  56         request->rq_req_md.user_ptr = request;
  57
  58         switch (request->rq_type) {
  59         case PTL_RPC_TYPE_REQUEST:
  60                 request->rq_req_md.start = request->rq_reqmsg;
  61                 request->rq_req_md.length = request->rq_reqlen;
  62                 request->rq_req_md.eventq = request_out_eq;
  63                 request->rq_req_md.threshold = 1;
  64                 ack = PTL_NOACK_REQ;
  65                 break;
  66         case PTL_RPC_TYPE_REPLY:
  67                 request->rq_req_md.start = request->rq_repmsg;
  68                 request->rq_req_md.length = request->rq_replen;
  69                 request->rq_req_md.eventq = reply_out_eq;
  70                 request->rq_req_md.threshold = 1;
  71                 ack = PTL_NOACK_REQ;
  72                 break;
  73         default:
  74                 LBUG();
  75                 return -1; /* notreached */
  76         }
  77         request->rq_req_md.options = PTL_MD_OP_PUT;
  78         request->rq_req_md.user_ptr = request;
  79
  80         rc = PtlMDBind(conn->c_peer.peer_ni, request->rq_req_md, &md_h);
  81         //CERROR("MDBind (outgoing req/rep/bulk): %Lu\n", (__u64)md_h);
  82         if (rc != 0) {
  83                 CERROR("PtlMDBind failed: %d\n", rc);
  84                 LBUG();
  85                 return rc;
  86         }
  87
  88         remote_id.nid = conn->c_peer.peer_nid;
  89         remote_id.pid = 0;
  90
  91         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %Ld\n",
  92                request->rq_req_md.length, portal, request->rq_xid);
  93
  94         rc = PtlPut(md_h, ack, remote_id, portal, 0, request->rq_reqmsg->xid,
  95                     0, 0);
  96         if (rc != PTL_OK) {
  97                 CERROR("PtlPut(%Lu, %d, %Ld) failed: %d\n", remote_id.nid,
  98                        portal, request->rq_xid, rc);
  99                 PtlMDUnlink(md_h);
 100         }
 101
 102         return rc;
 103 }
 104
 105 int ptlrpc_send_bulk(struct ptlrpc_bulk_desc *desc)
 106 {
 107         int rc;
 108         struct list_head *tmp, *next;
 109         ptl_process_id_t remote_id;
 110         ENTRY;
 111
 112         atomic_set(&desc->b_finished_count, desc->b_page_count);
 113
 114         list_for_each_safe(tmp, next, &desc->b_page_list) {
 115                 /* only request an ACK for the last page */
 116                 int ack = (next == &desc->b_page_list);
 117                 struct ptlrpc_bulk_page *bulk;
 118                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, b_link);
 119
 120                 bulk->b_md.start = bulk->b_buf;
 121                 bulk->b_md.length = bulk->b_buflen;
 122                 bulk->b_md.eventq = bulk_source_eq;
 123                 bulk->b_md.threshold = 1 + ack; /* SENT and (if last) ACK */
 124                 bulk->b_md.options = PTL_MD_OP_PUT;
 125                 bulk->b_md.user_ptr = bulk;
 126
 127                 rc = PtlMDBind(desc->b_connection->c_peer.peer_ni, bulk->b_md,
 128                                &bulk->b_md_h);
 129                 if (rc != 0) {
 130                         CERROR("PtlMDBind failed: %d\n", rc);
 131                         LBUG();
 132                         RETURN(rc);
 133                 }
 134
 135                 remote_id.nid = desc->b_connection->c_peer.peer_nid;
 136                 remote_id.pid = 0;
 137
 138                 CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %d\n",
 139                        bulk->b_md.length, desc->b_portal, bulk->b_xid);
 140
 141                 rc = PtlPut(bulk->b_md_h, (ack ? PTL_ACK_REQ : PTL_NOACK_REQ),
 142                             remote_id, desc->b_portal, 0, bulk->b_xid, 0, 0);
 143                 if (rc != PTL_OK) {
 144                         CERROR("PtlPut(%Lu, %d, %d) failed: %d\n",
 145                                remote_id.nid, desc->b_portal, bulk->b_xid, rc);
 146                         PtlMDUnlink(bulk->b_md_h);
 147                         LBUG();
 148                         RETURN(rc);
 149                 }
 150         }
 151
 152         RETURN(0);
 153 }
 154
 155 int ptlrpc_register_bulk(struct ptlrpc_bulk_desc *desc)
 156 {
 157         struct list_head *tmp, *next;
 158         int rc;
 159         ENTRY;
 160
 161         list_for_each_safe(tmp, next, &desc->b_page_list) {
 162                 struct ptlrpc_bulk_page *bulk;
 163                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, b_link);
 164
 165                 rc = PtlMEAttach(desc->b_connection->c_peer.peer_ni,
 166                                  desc->b_portal, local_id, bulk->b_xid, 0,
 167                                  PTL_UNLINK, PTL_INS_AFTER, &bulk->b_me_h);
 168                 if (rc != PTL_OK) {
 169                         CERROR("PtlMEAttach failed: %d\n", rc);
 170                         LBUG();
 171                         GOTO(cleanup, rc);
 172                 }
 173
 174                 bulk->b_md.start = bulk->b_buf;
 175                 bulk->b_md.length = bulk->b_buflen;
 176                 bulk->b_md.threshold = 1;
 177                 bulk->b_md.options = PTL_MD_OP_PUT;
 178                 bulk->b_md.user_ptr = bulk;
 179                 bulk->b_md.eventq = bulk_sink_eq;
 180
 181                 rc = PtlMDAttach(bulk->b_me_h, bulk->b_md, PTL_UNLINK,
 182                                  &bulk->b_md_h);
 183                 if (rc != PTL_OK) {
 184                         CERROR("PtlMDAttach failed: %d\n", rc);
 185                         LBUG();
 186                         GOTO(cleanup, rc);
 187                 }
 188
 189                 CDEBUG(D_NET, "Setup bulk sink buffer: %u bytes, xid %u, "
 190                        "portal %u\n", bulk->b_buflen, bulk->b_xid,
 191                        desc->b_portal);
 192         }
 193
 194         RETURN(0);
 195
 196  cleanup:
 197         ptlrpc_abort_bulk(desc);
 198
 199         return rc;
 200 }
 201
 202 int ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
 203 {
 204         struct list_head *tmp, *next;
 205
 206         list_for_each_safe(tmp, next, &desc->b_page_list) {
 207                 struct ptlrpc_bulk_page *bulk;
 208                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, b_link);
 209
 210                 /* This should be safe: these handles are initialized to be
 211                  * invalid in ptlrpc_prep_bulk_page() */
 212                 PtlMDUnlink(bulk->b_md_h);
 213                 PtlMEUnlink(bulk->b_me_h);
 214         }
 215
 216         return 0;
 217 }
 218
 219 int ptlrpc_reply(struct ptlrpc_service *svc, struct ptlrpc_request *req)
 220 {
 221         /* FIXME: we need to increment the count of handled events */
 222         req->rq_type = PTL_RPC_TYPE_REPLY;
 223         req->rq_repmsg->conn = req->rq_connection->c_remote_conn;
 224         req->rq_repmsg->token = req->rq_connection->c_remote_token;
 225         req->rq_repmsg->xid = HTON__u32(req->rq_reqmsg->xid);
 226         req->rq_repmsg->status = HTON__u32(req->rq_status);
 227         req->rq_reqmsg->type = HTON__u32(req->rq_type);
 228         return ptl_send_buf(req, req->rq_connection, svc->srv_rep_portal);
 229 }
 230
 231 int ptlrpc_error(struct ptlrpc_service *svc, struct ptlrpc_request *req)
 232 {
 233         int rc;
 234         ENTRY;
 235
 236         if (req->rq_repmsg) {
 237                 CERROR("req already has repmsg\n");
 238                 LBUG();
 239         }
 240
 241         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 242         if (rc)
 243                 RETURN(rc);
 244
 245         req->rq_repmsg->type = HTON__u32(PTL_RPC_MSG_ERR);
 246
 247         rc = ptlrpc_reply(svc, req);
 248         RETURN(rc);
 249 }
 250
 251
 252 int ptl_send_rpc(struct ptlrpc_request *request)
 253 {
 254         int rc;
 255         char *repbuf;
 256
 257         ENTRY;
 258
 259         if (NTOH__u32(request->rq_reqmsg->type) != PTL_RPC_MSG_REQUEST) {
 260                 CERROR("wrong packet type sent %d\n",
 261                        NTOH__u32(request->rq_reqmsg->type));
 262                 LBUG();
 263                 RETURN(EINVAL);
 264         }
 265         if (request->rq_replen == 0) {
 266                 CERROR("request->rq_replen is 0!\n");
 267                 RETURN(EINVAL);
 268         }
 269
 270         /* request->rq_repmsg is set only when the reply comes in, in
 271          * client_packet_callback() */
 272         if (request->rq_reply_md.start)
 273                 OBD_FREE(request->rq_reply_md.start, request->rq_replen);
 274
 275         OBD_ALLOC(repbuf, request->rq_replen);
 276         if (!repbuf) {
 277                 LBUG();
 278                 RETURN(ENOMEM);
 279         }
 280
 281         down(&request->rq_client->cli_rpc_sem);
 282
 283         rc = PtlMEAttach(request->rq_connection->c_peer.peer_ni,
 284                          request->rq_client->cli_reply_portal,
 285                          local_id, request->rq_reqmsg->xid, 0, PTL_UNLINK,
 286                          PTL_INS_AFTER, &request->rq_reply_me_h);
 287         if (rc != PTL_OK) {
 288                 CERROR("PtlMEAttach failed: %d\n", rc);
 289                 LBUG();
 290                 GOTO(cleanup, rc);
 291         }
 292
 293         request->rq_type = PTL_RPC_TYPE_REQUEST;
 294         request->rq_reply_md.start = repbuf;
 295         request->rq_reply_md.length = request->rq_replen;
 296         request->rq_reply_md.threshold = 1;
 297         request->rq_reply_md.options = PTL_MD_OP_PUT;
 298         request->rq_reply_md.user_ptr = request;
 299         request->rq_reply_md.eventq = reply_in_eq;
 300
 301         rc = PtlMDAttach(request->rq_reply_me_h, request->rq_reply_md,
 302                          PTL_UNLINK, &request->rq_reply_md_h);
 303         if (rc != PTL_OK) {
 304                 CERROR("PtlMDAttach failed: %d\n", rc);
 305                 LBUG();
 306                 GOTO(cleanup2, rc);
 307         }
 308
 309         CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %Lu, portal %u\n",
 310                request->rq_replen, request->rq_xid,
 311                request->rq_client->cli_reply_portal);
 312
 313         rc = ptl_send_buf(request, request->rq_connection,
 314                           request->rq_client->cli_request_portal);
 315         RETURN(rc);
 316
 317  cleanup2:
 318         PtlMEUnlink(request->rq_reply_me_h);
 319  cleanup:
 320         OBD_FREE(repbuf, request->rq_replen);
 321         up(&request->rq_client->cli_rpc_sem);
 322
 323         return rc;
 324 }
 325
 326 void ptlrpc_link_svc_me(struct ptlrpc_service *service, int i)
 327 {
 328         int rc;
 329         ptl_md_t dummy;
 330         ptl_handle_md_t md_h;
 331
 332         /* Attach the leading ME on which we build the ring */
 333         rc = PtlMEAttach(service->srv_self.peer_ni, service->srv_req_portal,
 334                          local_id, 0, ~0, PTL_RETAIN, PTL_INS_BEFORE,
 335                          &(service->srv_me_h[i]));
 336         if (rc != PTL_OK) {
 337                 CERROR("PtlMEAttach failed: %d\n", rc);
 338                 LBUG();
 339         }
 340
 341         if (service->srv_ref_count[i])
 342                 LBUG();
 343
 344         dummy.start         = service->srv_buf[i];
 345         dummy.length        = service->srv_buf_size;
 346         dummy.max_offset    = service->srv_buf_size;
 347         dummy.threshold     = PTL_MD_THRESH_INF;
 348         dummy.options       = PTL_MD_OP_PUT | PTL_MD_AUTO_UNLINK;
 349         dummy.user_ptr      = service;
 350         dummy.eventq        = service->srv_eq_h;
 351         dummy.max_offset    = service->srv_buf_size;
 352
 353         rc = PtlMDAttach(service->srv_me_h[i], dummy, PTL_UNLINK, &md_h);
 354         if (rc != PTL_OK) {
 355                 /* cleanup */
 356                 CERROR("PtlMDAttach failed: %d\n", rc);
 357                 LBUG();
 358         }
 359 }
 360
 361 /* ptl_handled_rpc() should be called by the sleeping process once
 362  * it finishes processing an event.  This ensures the ref count is
 363  * decremented and that the rpc ring buffer cycles properly.
 364  */
 365 int ptl_handled_rpc(struct ptlrpc_service *service, void *start)
 366 {
 367         int index;
 368
 369         spin_lock(&service->srv_lock);
 370
 371         for (index = 0; index < service->srv_ring_length; index++)
 372                 if (service->srv_buf[index] == start)
 373                         break;
 374
 375         if (index == service->srv_ring_length)
 376                 LBUG();
 377
 378         CDEBUG(D_INFO, "MD index=%d Ref Count=%d\n", index,
 379                service->srv_ref_count[index]);
 380         service->srv_ref_count[index]--;
 381
 382         if (service->srv_ref_count[index] < 0)
 383                 LBUG();
 384
 385         if (service->srv_ref_count[index] == 0 &&
 386             !ptl_is_valid_handle(&(service->srv_me_h[index]))) {
 387                 CDEBUG(D_NET, "relinking %d\n", index);
 388                 ptlrpc_link_svc_me(service, index);
 389         }
 390
 391         spin_unlock(&service->srv_lock);
 392         return 0;
 393 }