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 int ptlrpc_send_bulk(struct ptlrpc_bulk_desc *desc)
 123 {
 124         int rc;
 125         struct list_head *tmp, *next;
 126         ptl_process_id_t remote_id;
 127         ENTRY;
 128
 129         list_for_each_safe(tmp, next, &desc->b_page_list) {
 130                 struct ptlrpc_bulk_page *bulk;
 131                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, b_link);
 132
 133                 bulk->b_md.start = bulk->b_buf;
 134                 bulk->b_md.length = bulk->b_buflen;
 135                 bulk->b_md.eventq = bulk_source_eq;
 136                 bulk->b_md.threshold = 2; /* SENT and ACK */
 137                 bulk->b_md.options = PTL_MD_OP_PUT;
 138                 bulk->b_md.user_ptr = bulk;
 139
 140                 rc = PtlMDBind(desc->b_connection->c_peer.peer_ni, bulk->b_md,
 141                                &bulk->b_md_h);
 142                 if (rc != 0) {
 143                         CERROR("PtlMDBind failed: %d\n", rc);
 144                         LBUG();
 145                         RETURN(rc);
 146                 }
 147
 148                 remote_id.nid = desc->b_connection->c_peer.peer_nid;
 149                 remote_id.pid = 0;
 150
 151                 CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %d\n",
 152                        bulk->b_md.length, desc->b_portal, bulk->b_xid);
 153
 154                 rc = PtlPut(bulk->b_md_h, PTL_ACK_REQ, remote_id,
 155                             desc->b_portal, 0, bulk->b_xid, 0, 0);
 156                 if (rc != PTL_OK) {
 157                         CERROR("PtlPut(%Lu, %d, %d) failed: %d\n",
 158                                remote_id.nid, desc->b_portal, bulk->b_xid, rc);
 159                         PtlMDUnlink(bulk->b_md_h);
 160                         LBUG();
 161                         RETURN(rc);
 162                 }
 163         }
 164
 165         RETURN(0);
 166 }
 167
 168 int ptlrpc_register_bulk(struct ptlrpc_bulk_desc *desc)
 169 {
 170         struct list_head *tmp, *next;
 171         int rc;
 172         ENTRY;
 173
 174         list_for_each_safe(tmp, next, &desc->b_page_list) {
 175                 struct ptlrpc_bulk_page *bulk;
 176                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, b_link);
 177
 178                 rc = PtlMEAttach(desc->b_connection->c_peer.peer_ni,
 179                                  desc->b_portal, local_id, bulk->b_xid, 0,
 180                                  PTL_UNLINK, PTL_INS_AFTER, &bulk->b_me_h);
 181                 if (rc != PTL_OK) {
 182                         CERROR("PtlMEAttach failed: %d\n", rc);
 183                         LBUG();
 184                         GOTO(cleanup, rc);
 185                 }
 186
 187                 bulk->b_md.start = bulk->b_buf;
 188                 bulk->b_md.length = bulk->b_buflen;
 189                 bulk->b_md.threshold = 1;
 190                 bulk->b_md.options = PTL_MD_OP_PUT;
 191                 bulk->b_md.user_ptr = bulk;
 192                 bulk->b_md.eventq = bulk_sink_eq;
 193
 194                 rc = PtlMDAttach(bulk->b_me_h, bulk->b_md, PTL_UNLINK,
 195                                  &bulk->b_md_h);
 196                 if (rc != PTL_OK) {
 197                         CERROR("PtlMDAttach failed: %d\n", rc);
 198                         LBUG();
 199                         GOTO(cleanup, rc);
 200                 }
 201
 202                 CDEBUG(D_NET, "Setup bulk sink buffer: %u bytes, xid %u, "
 203                        "portal %u\n", bulk->b_buflen, bulk->b_xid,
 204                        desc->b_portal);
 205         }
 206
 207         RETURN(0);
 208
 209  cleanup:
 210         ptlrpc_abort_bulk(desc);
 211
 212         return rc;
 213 }
 214
 215 int ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
 216 {
 217         struct list_head *tmp, *next;
 218
 219         list_for_each_safe(tmp, next, &desc->b_page_list) {
 220                 struct ptlrpc_bulk_page *bulk;
 221                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, b_link);
 222
 223                 /* This should be safe: these handles are initialized to be
 224                  * invalid in ptlrpc_prep_bulk_page() */
 225                 PtlMDUnlink(bulk->b_md_h);
 226                 PtlMEUnlink(bulk->b_me_h);
 227         }
 228
 229         return 0;
 230 }
 231
 232 int ptlrpc_reply(struct ptlrpc_service *svc, struct ptlrpc_request *req)
 233 {
 234         if (req->rq_repmsg == NULL) {
 235                 CERROR("bad: someone called ptlrpc_reply when they meant "
 236                        "ptlrpc_error\n");
 237                 return -EINVAL;
 238         }
 239
 240         /* FIXME: we need to increment the count of handled events */
 241         req->rq_type = PTL_RPC_TYPE_REPLY;
 242         //req->rq_repmsg->conn = req->rq_connection->c_remote_conn;
 243         //req->rq_repmsg->token = req->rq_connection->c_remote_token;
 244         req->rq_repmsg->status = HTON__u32(req->rq_status);
 245         req->rq_reqmsg->type = HTON__u32(req->rq_type);
 246         return ptl_send_buf(req, req->rq_connection, svc->srv_rep_portal);
 247 }
 248
 249 int ptlrpc_error(struct ptlrpc_service *svc, struct ptlrpc_request *req)
 250 {
 251         int rc;
 252         ENTRY;
 253
 254         if (req->rq_repmsg) {
 255                 CERROR("req already has repmsg\n");
 256                 LBUG();
 257         }
 258
 259         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 260         if (rc)
 261                 RETURN(rc);
 262
 263         req->rq_repmsg->type = HTON__u32(PTL_RPC_MSG_ERR);
 264
 265         rc = ptlrpc_reply(svc, req);
 266         RETURN(rc);
 267 }
 268
 269
 270 int ptl_send_rpc(struct ptlrpc_request *request)
 271 {
 272         int rc;
 273         char *repbuf;
 274
 275         ENTRY;
 276
 277         if (NTOH__u32(request->rq_reqmsg->type) != PTL_RPC_MSG_REQUEST) {
 278                 CERROR("wrong packet type sent %d\n",
 279                        NTOH__u32(request->rq_reqmsg->type));
 280                 LBUG();
 281                 RETURN(EINVAL);
 282         }
 283         if (request->rq_replen == 0) {
 284                 CERROR("request->rq_replen is 0!\n");
 285                 RETURN(EINVAL);
 286         }
 287
 288         /* request->rq_repmsg is set only when the reply comes in, in
 289          * client_packet_callback() */
 290         if (request->rq_reply_md.start)
 291                 OBD_FREE(request->rq_reply_md.start, request->rq_replen);
 292
 293         OBD_ALLOC(repbuf, request->rq_replen);
 294         if (!repbuf) {
 295                 LBUG();
 296                 RETURN(ENOMEM);
 297         }
 298
 299         down(&request->rq_client->cli_rpc_sem);
 300
 301         rc = PtlMEAttach(request->rq_connection->c_peer.peer_ni,
 302                          request->rq_client->cli_reply_portal,
 303                          local_id, request->rq_xid, 0, PTL_UNLINK,
 304                          PTL_INS_AFTER, &request->rq_reply_me_h);
 305         if (rc != PTL_OK) {
 306                 CERROR("PtlMEAttach failed: %d\n", rc);
 307                 LBUG();
 308                 GOTO(cleanup, rc);
 309         }
 310
 311         request->rq_type = PTL_RPC_TYPE_REQUEST;
 312         request->rq_reply_md.start = repbuf;
 313         request->rq_reply_md.length = request->rq_replen;
 314         request->rq_reply_md.threshold = 1;
 315         request->rq_reply_md.options = PTL_MD_OP_PUT;
 316         request->rq_reply_md.user_ptr = request;
 317         request->rq_reply_md.eventq = reply_in_eq;
 318
 319         rc = PtlMDAttach(request->rq_reply_me_h, request->rq_reply_md,
 320                          PTL_UNLINK, &request->rq_reply_md_h);
 321         if (rc != PTL_OK) {
 322                 CERROR("PtlMDAttach failed: %d\n", rc);
 323                 LBUG();
 324                 GOTO(cleanup2, rc);
 325         }
 326
 327         CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %Lu, portal %u\n",
 328                request->rq_replen, request->rq_xid,
 329                request->rq_client->cli_reply_portal);
 330
 331         rc = ptl_send_buf(request, request->rq_connection,
 332                           request->rq_client->cli_request_portal);
 333         RETURN(rc);
 334
 335  cleanup2:
 336         PtlMEUnlink(request->rq_reply_me_h);
 337  cleanup:
 338         OBD_FREE(repbuf, request->rq_replen);
 339         up(&request->rq_client->cli_rpc_sem);
 340
 341         return rc;
 342 }
 343
 344 void ptlrpc_link_svc_me(struct ptlrpc_service *service, int i)
 345 {
 346         int rc;
 347         ptl_md_t dummy;
 348         ptl_handle_md_t md_h;
 349
 350         /* Attach the leading ME on which we build the ring */
 351         rc = PtlMEAttach(service->srv_self.peer_ni, service->srv_req_portal,
 352                          local_id, 0, ~0, PTL_RETAIN, PTL_INS_BEFORE,
 353                          &(service->srv_me_h[i]));
 354         if (rc != PTL_OK) {
 355                 CERROR("PtlMEAttach failed: %d\n", rc);
 356                 LBUG();
 357         }
 358
 359         if (service->srv_ref_count[i])
 360                 LBUG();
 361
 362         dummy.start         = service->srv_buf[i];
 363         dummy.length        = service->srv_buf_size;
 364         dummy.max_offset    = service->srv_buf_size;
 365         dummy.threshold     = PTL_MD_THRESH_INF;
 366         dummy.options       = PTL_MD_OP_PUT | PTL_MD_AUTO_UNLINK;
 367         dummy.user_ptr      = service;
 368         dummy.eventq        = service->srv_eq_h;
 369         dummy.max_offset    = service->srv_buf_size;
 370
 371         rc = PtlMDAttach(service->srv_me_h[i], dummy, PTL_UNLINK, &md_h);
 372         if (rc != PTL_OK) {
 373                 /* cleanup */
 374                 CERROR("PtlMDAttach failed: %d\n", rc);
 375                 LBUG();
 376         }
 377 }
 378
 379 /* ptl_handled_rpc() should be called by the sleeping process once
 380  * it finishes processing an event.  This ensures the ref count is
 381  * decremented and that the rpc ring buffer cycles properly.
 382  */
 383 int ptl_handled_rpc(struct ptlrpc_service *service, void *start)
 384 {
 385         int index;
 386
 387         spin_lock(&service->srv_lock);
 388         for (index = 0; index < service->srv_ring_length; index++)
 389                 if (service->srv_buf[index] == start)
 390                         break;
 391
 392         if (index == service->srv_ring_length)
 393                 LBUG();
 394
 395         CDEBUG(D_INFO, "MD index=%d Ref Count=%d\n", index,
 396                service->srv_ref_count[index]);
 397         service->srv_ref_count[index]--;
 398
 399         if (service->srv_ref_count[index] < 0)
 400                 LBUG();
 401
 402         if (service->srv_ref_count[index] == 0 &&
 403             !ptl_is_valid_handle(&(service->srv_me_h[index]))) {
 404                 CDEBUG(D_NET, "relinking %d\n", index);
 405                 ptlrpc_link_svc_me(service, index);
 406         }
 407
 408         spin_unlock(&service->srv_lock);
 409         return 0;
 410 }