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 EXPORT_SYMTAB
  24
  25 #define DEBUG_SUBSYSTEM S_RPC
  26
  27 #include <linux/lustre_net.h>
  28
  29 extern ptl_handle_eq_t request_out_eq,
  30         reply_in_eq,
  31         reply_out_eq,
  32         bulk_source_eq,
  33         bulk_sink_eq;
  34 static ptl_process_id_t local_id = {PTL_ID_ANY, PTL_ID_ANY};
  35
  36
  37 int ptlrpc_check_bulk_sent(struct ptlrpc_bulk_desc *bulk)
  38 {
  39         ENTRY;
  40
  41         if (bulk->b_flags == PTL_BULK_SENT)
  42                 RETURN(1);
  43
  44         if (sigismember(&(current->pending.signal), SIGKILL) ||
  45             sigismember(&(current->pending.signal), SIGINT)) {
  46                 bulk->b_flags = PTL_RPC_INTR;
  47                 RETURN(1);
  48         }
  49
  50         CDEBUG(D_NET, "no event yet\n");
  51         RETURN(0);
  52 }
  53
  54 int ptl_send_buf(struct ptlrpc_request *request, struct ptlrpc_connection *conn,
  55                  int portal)
  56 {
  57         int rc;
  58         ptl_process_id_t remote_id;
  59         ptl_handle_md_t md_h;
  60         ptl_ack_req_t ack;
  61
  62         request->rq_req_md.user_ptr = request;
  63
  64         switch (request->rq_type) {
  65         case PTL_RPC_BULK:
  66                 request->rq_req_md.start = request->rq_bulkbuf;
  67                 request->rq_req_md.length = request->rq_bulklen;
  68                 request->rq_req_md.eventq = bulk_source_eq;
  69                 request->rq_req_md.threshold = 2; /* SENT and ACK events */
  70                 ack = PTL_ACK_REQ;
  71                 break;
  72         case PTL_RPC_REQUEST:
  73                 request->rq_req_md.start = request->rq_reqmsg;
  74                 request->rq_req_md.length = request->rq_reqlen;
  75                 request->rq_req_md.eventq = request_out_eq;
  76                 request->rq_req_md.threshold = 1;
  77                 ack = PTL_NOACK_REQ;
  78                 break;
  79         case PTL_RPC_REPLY:
  80                 request->rq_req_md.start = request->rq_repmsg;
  81                 request->rq_req_md.length = request->rq_replen;
  82                 request->rq_req_md.eventq = reply_out_eq;
  83                 request->rq_req_md.threshold = 1;
  84                 ack = PTL_NOACK_REQ;
  85                 break;
  86         default:
  87                 LBUG();
  88                 return -1; /* notreached */
  89         }
  90         request->rq_req_md.options = PTL_MD_OP_PUT;
  91         request->rq_req_md.user_ptr = request;
  92
  93         rc = PtlMDBind(conn->c_peer.peer_ni, request->rq_req_md, &md_h);
  94         //CERROR("MDBind (outgoing req/rep/bulk): %Lu\n", (__u64)md_h);
  95         if (rc != 0) {
  96                 CERROR("PtlMDBind failed: %d\n", rc);
  97                 LBUG();
  98                 return rc;
  99         }
 100
 101         remote_id.nid = conn->c_peer.peer_nid;
 102         remote_id.pid = 0;
 103
 104         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %d\n",
 105                request->rq_req_md.length, portal, request->rq_reqmsg->xid);
 106
 107         rc = PtlPut(md_h, ack, remote_id, portal, 0, request->rq_reqmsg->xid,
 108                     0, 0);
 109         if (rc != PTL_OK) {
 110                 CERROR("PtlPut(%d, %d, %d) failed: %d\n", remote_id.nid,
 111                        portal, request->rq_reqmsg->xid, rc);
 112                 PtlMDUnlink(md_h);
 113         }
 114
 115         return rc;
 116 }
 117
 118 int ptlrpc_send_bulk(struct ptlrpc_bulk_desc *bulk, int portal)
 119 {
 120         int rc;
 121         ptl_process_id_t remote_id;
 122
 123         bulk->b_md.start = bulk->b_buf;
 124         bulk->b_md.length = bulk->b_buflen;
 125         bulk->b_md.eventq = bulk_source_eq;
 126         bulk->b_md.threshold = 2; /* SENT and ACK events */
 127         bulk->b_md.options = PTL_MD_OP_PUT;
 128         bulk->b_md.user_ptr = bulk;
 129
 130         rc = PtlMDBind(bulk->b_connection->c_peer.peer_ni, bulk->b_md,
 131                        &bulk->b_md_h);
 132         if (rc != 0) {
 133                 CERROR("PtlMDBind failed: %d\n", rc);
 134                 LBUG();
 135                 return rc;
 136         }
 137
 138         remote_id.nid = bulk->b_connection->c_peer.peer_nid;
 139         remote_id.pid = 0;
 140
 141         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %d\n",
 142                bulk->b_md.length, portal, bulk->b_xid);
 143
 144         rc = PtlPut(bulk->b_md_h, PTL_ACK_REQ, remote_id, portal, 0,
 145                     bulk->b_xid, 0, 0);
 146         if (rc != PTL_OK) {
 147                 CERROR("PtlPut(%d, %d, %d) failed: %d\n", remote_id.nid,
 148                        portal, bulk->b_xid, rc);
 149                 PtlMDUnlink(bulk->b_md_h);
 150                 LBUG();
 151         }
 152
 153         return rc;
 154 }
 155
 156 int ptlrpc_register_bulk(struct ptlrpc_bulk_desc *bulk)
 157 {
 158         int rc;
 159         ENTRY;
 160
 161         rc = PtlMEAttach(bulk->b_connection->c_peer.peer_ni, bulk->b_portal,
 162                          local_id, bulk->b_xid, 0, PTL_UNLINK, PTL_INS_AFTER,
 163                          &bulk->b_me_h);
 164         if (rc != PTL_OK) {
 165                 CERROR("PtlMEAttach failed: %d\n", rc);
 166                 LBUG();
 167                 GOTO(cleanup, rc);
 168         }
 169
 170         bulk->b_md.start = bulk->b_buf;
 171         bulk->b_md.length = bulk->b_buflen;
 172         bulk->b_md.threshold = 1;
 173         bulk->b_md.options = PTL_MD_OP_PUT;
 174         bulk->b_md.user_ptr = bulk;
 175         bulk->b_md.eventq = bulk_sink_eq;
 176
 177         rc = PtlMDAttach(bulk->b_me_h, bulk->b_md, PTL_UNLINK, &bulk->b_md_h);
 178         //CERROR("MDAttach (bulk sink): %Lu\n", (__u64)bulk->b_md_h);
 179         if (rc != PTL_OK) {
 180                 CERROR("PtlMDAttach failed: %d\n", rc);
 181                 LBUG();
 182                 GOTO(cleanup, rc);
 183         }
 184
 185         CDEBUG(D_NET, "Setup bulk sink buffer: %u bytes, xid %u, portal %u\n",
 186                bulk->b_buflen, bulk->b_xid, bulk->b_portal);
 187         RETURN(0);
 188
 189         // XXX Confirm that this is safe!
 190  cleanup:
 191         PtlMEUnlink(bulk->b_me_h);
 192         return rc;
 193 }
 194
 195 int ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *bulk)
 196 {
 197         int rc;
 198
 199         rc = PtlMEUnlink(bulk->b_me_h);
 200         if (rc != PTL_OK)
 201                 CERROR("PtlMEUnlink failed: %d\n", rc);
 202
 203         return rc;
 204 }
 205
 206 int ptlrpc_reply(struct ptlrpc_service *svc, struct ptlrpc_request *req)
 207 {
 208         /* FIXME: we need to increment the count of handled events */
 209         req->rq_type = PTL_RPC_REPLY;
 210         req->rq_repmsg->conn = req->rq_reqmsg->conn;
 211         req->rq_repmsg->token = req->rq_reqmsg->token;
 212         req->rq_repmsg->xid = HTON__u32(req->rq_reqmsg->xid);
 213         req->rq_repmsg->status = HTON__u32(req->rq_status);
 214         req->rq_reqmsg->type = HTON__u32(req->rq_type);
 215         return ptl_send_buf(req, req->rq_connection, svc->srv_rep_portal);
 216 }
 217
 218 int ptlrpc_error(struct ptlrpc_service *svc, struct ptlrpc_request *req)
 219 {
 220         int rc;
 221         ENTRY;
 222
 223         if (req->rq_repmsg) {
 224                 CERROR("req already has repmsg\n");
 225                 LBUG();
 226         }
 227
 228         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 229         if (rc)
 230                 RETURN(rc);
 231
 232         req->rq_repmsg->type = HTON__u32(PTL_RPC_ERR);
 233
 234         rc = ptlrpc_reply(svc, req);
 235         RETURN(rc);
 236 }
 237
 238 int ptl_send_rpc(struct ptlrpc_request *request)
 239 {
 240         ptl_process_id_t local_id;
 241         int rc;
 242         char *repbuf;
 243
 244         ENTRY;
 245
 246         if (NTOH__u32(request->rq_reqmsg->type) != PTL_RPC_REQUEST) {
 247                 CERROR("wrong packet type sent %d\n",
 248                        NTOH__u32(request->rq_reqmsg->type));
 249                 LBUG();
 250                 RETURN(-EINVAL);
 251         }
 252         if (request->rq_replen == 0) {
 253                 CERROR("request->rq_replen is 0!\n");
 254                 RETURN(-EINVAL);
 255         }
 256
 257         /* request->rq_repmsg is set only when the reply comes in, in
 258          * client_packet_callback() */
 259         OBD_ALLOC(repbuf, request->rq_replen);
 260         if (!repbuf)
 261                 RETURN(-ENOMEM);
 262
 263         local_id.nid = PTL_ID_ANY;
 264         local_id.pid = PTL_ID_ANY;
 265
 266         down(&request->rq_client->cli_rpc_sem);
 267
 268         rc = PtlMEAttach(request->rq_connection->c_peer.peer_ni,
 269                          request->rq_client->cli_reply_portal,
 270                          local_id, request->rq_reqmsg->xid, 0, PTL_UNLINK,
 271                          PTL_INS_AFTER, &request->rq_reply_me_h);
 272         if (rc != PTL_OK) {
 273                 CERROR("PtlMEAttach failed: %d\n", rc);
 274                 LBUG();
 275                 GOTO(cleanup, rc);
 276         }
 277
 278         request->rq_type = PTL_RPC_REQUEST;
 279         request->rq_reply_md.start = repbuf;
 280         request->rq_reply_md.length = request->rq_replen;
 281         request->rq_reply_md.threshold = 1;
 282         request->rq_reply_md.options = PTL_MD_OP_PUT;
 283         request->rq_reply_md.user_ptr = request;
 284         request->rq_reply_md.eventq = reply_in_eq;
 285
 286         rc = PtlMDAttach(request->rq_reply_me_h, request->rq_reply_md,
 287                          PTL_UNLINK, &request->rq_reply_md_h);
 288         if (rc != PTL_OK) {
 289                 CERROR("PtlMDAttach failed: %d\n", rc);
 290                 LBUG();
 291                 GOTO(cleanup2, rc);
 292         }
 293
 294         CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %u, portal %u\n",
 295                request->rq_replen, request->rq_reqmsg->xid,
 296                request->rq_client->cli_request_portal);
 297
 298         list_add(&request->rq_list, &request->rq_client->cli_sending_head);
 299         rc = ptl_send_buf(request, request->rq_connection,
 300                           request->rq_client->cli_request_portal);
 301         RETURN(rc);
 302
 303  cleanup2:
 304         PtlMEUnlink(request->rq_reply_me_h);
 305  cleanup:
 306         OBD_FREE(repbuf, request->rq_replen);
 307         up(&request->rq_client->cli_rpc_sem);
 308
 309         return rc;
 310 }
 311
 312 void ptlrpc_link_svc_me(struct ptlrpc_service *service, int i)
 313 {
 314         int rc;
 315         ptl_md_t dummy;
 316         ptl_handle_md_t md_h;
 317
 318         /* Attach the leading ME on which we build the ring */
 319         rc = PtlMEAttach(service->srv_self.peer_ni, service->srv_req_portal,
 320                          service->srv_id, 0, ~0, PTL_RETAIN, PTL_INS_BEFORE,
 321                          &(service->srv_me_h[i]));
 322         if (rc != PTL_OK) {
 323                 CERROR("PtlMEAttach failed: %d\n", rc);
 324                 LBUG();
 325         }
 326
 327         if (service->srv_ref_count[i])
 328                 LBUG();
 329
 330         dummy.start         = service->srv_buf[i];
 331         dummy.length        = service->srv_buf_size;
 332         dummy.max_offset    = service->srv_buf_size;
 333         dummy.threshold     = PTL_MD_THRESH_INF;
 334         dummy.options       = PTL_MD_OP_PUT | PTL_MD_AUTO_UNLINK;
 335         dummy.user_ptr      = service;
 336         dummy.eventq        = service->srv_eq_h;
 337         dummy.max_offset    = service->srv_buf_size;
 338
 339         rc = PtlMDAttach(service->srv_me_h[i], dummy, PTL_UNLINK, &md_h);
 340         if (rc != PTL_OK) {
 341                 /* cleanup */
 342                 CERROR("PtlMDAttach failed: %d\n", rc);
 343                 LBUG();
 344         }
 345 }
 346
 347 /* ptl_handled_rpc() should be called by the sleeping process once
 348  * it finishes processing an event.  This ensures the ref count is
 349  * decremented and that the rpc ring buffer cycles properly.
 350  */
 351 int ptl_handled_rpc(struct ptlrpc_service *service, void *start)
 352 {
 353         int index;
 354
 355         spin_lock(&service->srv_lock);
 356
 357         for (index = 0; index < service->srv_ring_length; index++)
 358                 if (service->srv_buf[index] == start)
 359                         break;
 360
 361         if (index == service->srv_ring_length)
 362                 LBUG();
 363
 364         CDEBUG(D_INFO, "MD index=%d Ref Count=%d\n", index,
 365                service->srv_ref_count[index]);
 366         service->srv_ref_count[index]--;
 367
 368         if (service->srv_ref_count[index] < 0)
 369                 LBUG();
 370
 371         if (service->srv_ref_count[index] == 0 &&
 372             !ptl_is_valid_handle(&(service->srv_me_h[index]))) {
 373                 CDEBUG(D_NET, "relinking %d\n", index);
 374                 ptlrpc_link_svc_me(service, index);
 375         }
 376
 377         spin_unlock(&service->srv_lock);
 378         return 0;
 379 }