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 static 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         list_for_each_safe(tmp, next, &desc->b_page_list) {
 113                 /* only request an ACK for the last page */
 114                 int ack = (next == &desc->b_page_list);
 115                 struct ptlrpc_bulk_page *bulk;
 116                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, b_link);
 117
 118                 bulk->b_md.start = bulk->b_buf;
 119                 bulk->b_md.length = bulk->b_buflen;
 120                 bulk->b_md.eventq = bulk_source_eq;
 121                 bulk->b_md.threshold = 1 + ack; /* SENT and (if last) ACK */
 122                 bulk->b_md.options = PTL_MD_OP_PUT;
 123                 bulk->b_md.user_ptr = bulk;
 124
 125                 rc = PtlMDBind(desc->b_connection->c_peer.peer_ni, bulk->b_md,
 126                                &bulk->b_md_h);
 127                 if (rc != 0) {
 128                         CERROR("PtlMDBind failed: %d\n", rc);
 129                         LBUG();
 130                         RETURN(rc);
 131                 }
 132
 133                 remote_id.nid = desc->b_connection->c_peer.peer_nid;
 134                 remote_id.pid = 0;
 135
 136                 CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %d\n",
 137                        bulk->b_md.length, desc->b_portal, bulk->b_xid);
 138
 139                 rc = PtlPut(bulk->b_md_h, (ack ? PTL_ACK_REQ : PTL_NOACK_REQ),
 140                             remote_id, desc->b_portal, 0, bulk->b_xid, 0, 0);
 141                 if (rc != PTL_OK) {
 142                         CERROR("PtlPut(%Lu, %d, %d) failed: %d\n", remote_id.nid,
 143                                desc->b_portal, bulk->b_xid, rc);
 144                         PtlMDUnlink(bulk->b_md_h);
 145                         LBUG();
 146                         RETURN(rc);
 147                 }
 148         }
 149
 150         wait_event_interruptible(desc->b_waitq, ptlrpc_check_bulk_sent(desc));
 151
 152         if (desc->b_flags & PTL_RPC_FL_INTR)
 153                 RETURN(-EINTR);
 154
 155         RETURN(0);
 156 }
 157
 158 int ptlrpc_register_bulk(struct ptlrpc_bulk_desc *desc)
 159 {
 160         struct list_head *tmp, *next;
 161         int rc;
 162         ENTRY;
 163
 164         list_for_each_safe(tmp, next, &desc->b_page_list) {
 165                 struct ptlrpc_bulk_page *bulk;
 166                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, b_link);
 167
 168                 rc = PtlMEAttach(desc->b_connection->c_peer.peer_ni,
 169                                  desc->b_portal, local_id, bulk->b_xid, 0,
 170                                  PTL_UNLINK, PTL_INS_AFTER, &bulk->b_me_h);
 171                 if (rc != PTL_OK) {
 172                         CERROR("PtlMEAttach failed: %d\n", rc);
 173                         LBUG();
 174                         GOTO(cleanup, rc);
 175                 }
 176
 177                 bulk->b_md.start = bulk->b_buf;
 178                 bulk->b_md.length = bulk->b_buflen;
 179                 bulk->b_md.threshold = 1;
 180                 bulk->b_md.options = PTL_MD_OP_PUT;
 181                 bulk->b_md.user_ptr = bulk;
 182                 bulk->b_md.eventq = bulk_sink_eq;
 183
 184                 rc = PtlMDAttach(bulk->b_me_h, bulk->b_md, PTL_UNLINK,
 185                                  &bulk->b_md_h);
 186                 if (rc != PTL_OK) {
 187                         CERROR("PtlMDAttach failed: %d\n", rc);
 188                         LBUG();
 189                         GOTO(cleanup, rc);
 190                 }
 191
 192                 CDEBUG(D_NET, "Setup bulk sink buffer: %u bytes, xid %u, "
 193                        "portal %u\n", bulk->b_buflen, bulk->b_xid,
 194                        desc->b_portal);
 195         }
 196
 197         RETURN(0);
 198
 199  cleanup:
 200         ptlrpc_abort_bulk(desc);
 201
 202         return rc;
 203 }
 204
 205 int ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
 206 {
 207         struct list_head *tmp, *next;
 208
 209         list_for_each_safe(tmp, next, &desc->b_page_list) {
 210                 struct ptlrpc_bulk_page *bulk;
 211                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, b_link);
 212
 213                 /* This should be safe: these handles are initialized to be
 214                  * invalid in ptlrpc_prep_bulk_page() */
 215                 PtlMDUnlink(bulk->b_md_h);
 216                 PtlMEUnlink(bulk->b_me_h);
 217         }
 218
 219         return 0;
 220 }
 221
 222 int ptlrpc_reply(struct ptlrpc_service *svc, struct ptlrpc_request *req)
 223 {
 224         /* FIXME: we need to increment the count of handled events */
 225         req->rq_type = PTL_RPC_TYPE_REPLY;
 226         req->rq_repmsg->conn = req->rq_connection->c_remote_conn;
 227         req->rq_repmsg->token = req->rq_connection->c_remote_token;
 228         req->rq_repmsg->xid = HTON__u32(req->rq_reqmsg->xid);
 229         req->rq_repmsg->status = HTON__u32(req->rq_status);
 230         req->rq_reqmsg->type = HTON__u32(req->rq_type);
 231         return ptl_send_buf(req, req->rq_connection, svc->srv_rep_portal);
 232 }
 233
 234 int ptlrpc_error(struct ptlrpc_service *svc, struct ptlrpc_request *req)
 235 {
 236         int rc;
 237         ENTRY;
 238
 239         if (req->rq_repmsg) {
 240                 CERROR("req already has repmsg\n");
 241                 LBUG();
 242         }
 243
 244         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 245         if (rc)
 246                 RETURN(rc);
 247
 248         req->rq_repmsg->type = HTON__u32(PTL_RPC_MSG_ERR);
 249
 250         rc = ptlrpc_reply(svc, req);
 251         RETURN(rc);
 252 }
 253
 254
 255 int ptl_send_rpc(struct ptlrpc_request *request)
 256 {
 257         ptl_process_id_t local_id;
 258         int rc;
 259         char *repbuf;
 260
 261         ENTRY;
 262
 263         if (NTOH__u32(request->rq_reqmsg->type) != PTL_RPC_MSG_REQUEST) {
 264                 CERROR("wrong packet type sent %d\n",
 265                        NTOH__u32(request->rq_reqmsg->type));
 266                 LBUG();
 267                 RETURN(EINVAL);
 268         }
 269         if (request->rq_replen == 0) {
 270                 CERROR("request->rq_replen is 0!\n");
 271                 RETURN(EINVAL);
 272         }
 273
 274         /* request->rq_repmsg is set only when the reply comes in, in
 275          * client_packet_callback() */
 276         if (request->rq_reply_md.start)
 277                 OBD_FREE(request->rq_reply_md.start, request->rq_replen);
 278
 279         OBD_ALLOC(repbuf, request->rq_replen);
 280         if (!repbuf) {
 281                 LBUG();
 282                 RETURN(ENOMEM);
 283         }
 284
 285         local_id.nid = PTL_ID_ANY;
 286         local_id.pid = PTL_ID_ANY;
 287
 288         down(&request->rq_client->cli_rpc_sem);
 289
 290         rc = PtlMEAttach(request->rq_connection->c_peer.peer_ni,
 291                          request->rq_client->cli_reply_portal,
 292                          local_id, request->rq_reqmsg->xid, 0, PTL_UNLINK,
 293                          PTL_INS_AFTER, &request->rq_reply_me_h);
 294         if (rc != PTL_OK) {
 295                 CERROR("PtlMEAttach failed: %d\n", rc);
 296                 LBUG();
 297                 GOTO(cleanup, rc);
 298         }
 299
 300         request->rq_type = PTL_RPC_TYPE_REQUEST;
 301         request->rq_reply_md.start = repbuf;
 302         request->rq_reply_md.length = request->rq_replen;
 303         request->rq_reply_md.threshold = 1;
 304         request->rq_reply_md.options = PTL_MD_OP_PUT;
 305         request->rq_reply_md.user_ptr = request;
 306         request->rq_reply_md.eventq = reply_in_eq;
 307
 308         rc = PtlMDAttach(request->rq_reply_me_h, request->rq_reply_md,
 309                          PTL_UNLINK, &request->rq_reply_md_h);
 310         if (rc != PTL_OK) {
 311                 CERROR("PtlMDAttach failed: %d\n", rc);
 312                 LBUG();
 313                 GOTO(cleanup2, rc);
 314         }
 315
 316         CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %Lu, portal %u\n",
 317                request->rq_replen, request->rq_xid,
 318                request->rq_client->cli_reply_portal);
 319
 320         rc = ptl_send_buf(request, request->rq_connection,
 321                           request->rq_client->cli_request_portal);
 322         RETURN(rc);
 323
 324  cleanup2:
 325         PtlMEUnlink(request->rq_reply_me_h);
 326  cleanup:
 327         OBD_FREE(repbuf, request->rq_replen);
 328         up(&request->rq_client->cli_rpc_sem);
 329
 330         return rc;
 331 }
 332
 333 void ptlrpc_link_svc_me(struct ptlrpc_service *service, int i)
 334 {
 335         int rc;
 336         ptl_md_t dummy;
 337         ptl_handle_md_t md_h;
 338
 339         /* Attach the leading ME on which we build the ring */
 340         rc = PtlMEAttach(service->srv_self.peer_ni, service->srv_req_portal,
 341                          local_id, 0, ~0, PTL_RETAIN, PTL_INS_BEFORE,
 342                          &(service->srv_me_h[i]));
 343         if (rc != PTL_OK) {
 344                 CERROR("PtlMEAttach failed: %d\n", rc);
 345                 LBUG();
 346         }
 347
 348         if (service->srv_ref_count[i])
 349                 LBUG();
 350
 351         dummy.start         = service->srv_buf[i];
 352         dummy.length        = service->srv_buf_size;
 353         dummy.max_offset    = service->srv_buf_size;
 354         dummy.threshold     = PTL_MD_THRESH_INF;
 355         dummy.options       = PTL_MD_OP_PUT | PTL_MD_AUTO_UNLINK;
 356         dummy.user_ptr      = service;
 357         dummy.eventq        = service->srv_eq_h;
 358         dummy.max_offset    = service->srv_buf_size;
 359
 360         rc = PtlMDAttach(service->srv_me_h[i], dummy, PTL_UNLINK, &md_h);
 361         if (rc != PTL_OK) {
 362                 /* cleanup */
 363                 CERROR("PtlMDAttach failed: %d\n", rc);
 364                 LBUG();
 365         }
 366 }
 367
 368 /* ptl_handled_rpc() should be called by the sleeping process once
 369  * it finishes processing an event.  This ensures the ref count is
 370  * decremented and that the rpc ring buffer cycles properly.
 371  */
 372 int ptl_handled_rpc(struct ptlrpc_service *service, void *start)
 373 {
 374         int index;
 375
 376         spin_lock(&service->srv_lock);
 377
 378         for (index = 0; index < service->srv_ring_length; index++)
 379                 if (service->srv_buf[index] == start)
 380                         break;
 381
 382         if (index == service->srv_ring_length)
 383                 LBUG();
 384
 385         CDEBUG(D_INFO, "MD index=%d Ref Count=%d\n", index,
 386                service->srv_ref_count[index]);
 387         service->srv_ref_count[index]--;
 388
 389         if (service->srv_ref_count[index] < 0)
 390                 LBUG();
 391
 392         if (service->srv_ref_count[index] == 0 &&
 393             !ptl_is_valid_handle(&(service->srv_me_h[index]))) {
 394                 CDEBUG(D_NET, "relinking %d\n", index);
 395                 ptlrpc_link_svc_me(service, index);
 396         }
 397
 398         spin_unlock(&service->srv_lock);
 399         return 0;
 400 }