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_FL_SENT)
  42                 RETURN(1);
  43
  44         if (sigismember(&(current->pending.signal), SIGKILL) ||
  45             sigismember(&(current->pending.signal), SIGINT)) {
  46                 bulk->b_flags |= PTL_RPC_FL_INTR;
  47                 RETURN(1);
  48         }
  49
  50         CDEBUG(D_NET, "no event yet\n");
  51         RETURN(0);
  52 }
  53
  54 static int ptl_send_buf(struct ptlrpc_request *request,
  55                         struct ptlrpc_connection *conn, 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_TYPE_REQUEST:
  66                 request->rq_req_md.start = request->rq_reqmsg;
  67                 request->rq_req_md.length = request->rq_reqlen;
  68                 request->rq_req_md.eventq = request_out_eq;
  69                 request->rq_req_md.threshold = 1;
  70                 ack = PTL_NOACK_REQ;
  71                 break;
  72         case PTL_RPC_TYPE_REPLY:
  73                 request->rq_req_md.start = request->rq_repmsg;
  74                 request->rq_req_md.length = request->rq_replen;
  75                 request->rq_req_md.eventq = reply_out_eq;
  76                 request->rq_req_md.threshold = 1;
  77                 ack = PTL_NOACK_REQ;
  78                 break;
  79         default:
  80                 LBUG();
  81                 return -1; /* notreached */
  82         }
  83         request->rq_req_md.options = PTL_MD_OP_PUT;
  84         request->rq_req_md.user_ptr = request;
  85
  86         rc = PtlMDBind(conn->c_peer.peer_ni, request->rq_req_md, &md_h);
  87         //CERROR("MDBind (outgoing req/rep/bulk): %Lu\n", (__u64)md_h);
  88         if (rc != 0) {
  89                 CERROR("PtlMDBind failed: %d\n", rc);
  90                 LBUG();
  91                 return rc;
  92         }
  93
  94         remote_id.nid = conn->c_peer.peer_nid;
  95         remote_id.pid = 0;
  96
  97         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %d\n",
  98                request->rq_req_md.length, portal, request->rq_reqmsg->xid);
  99
 100         rc = PtlPut(md_h, ack, remote_id, portal, 0, request->rq_reqmsg->xid,
 101                     0, 0);
 102         if (rc != PTL_OK) {
 103                 CERROR("PtlPut(%d, %d, %d) failed: %d\n", remote_id.nid,
 104                        portal, request->rq_reqmsg->xid, rc);
 105                 PtlMDUnlink(md_h);
 106         }
 107
 108         return rc;
 109 }
 110
 111 int ptlrpc_send_bulk(struct ptlrpc_bulk_desc *bulk, int portal)
 112 {
 113         int rc;
 114         ptl_process_id_t remote_id;
 115
 116         bulk->b_md.start = bulk->b_buf;
 117         bulk->b_md.length = bulk->b_buflen;
 118         bulk->b_md.eventq = bulk_source_eq;
 119         bulk->b_md.threshold = 2; /* SENT and ACK events */
 120         bulk->b_md.options = PTL_MD_OP_PUT;
 121         bulk->b_md.user_ptr = bulk;
 122
 123         rc = PtlMDBind(bulk->b_connection->c_peer.peer_ni, bulk->b_md,
 124                        &bulk->b_md_h);
 125         if (rc != 0) {
 126                 CERROR("PtlMDBind failed: %d\n", rc);
 127                 LBUG();
 128                 return rc;
 129         }
 130
 131         remote_id.nid = bulk->b_connection->c_peer.peer_nid;
 132         remote_id.pid = 0;
 133
 134         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %d\n",
 135                bulk->b_md.length, portal, bulk->b_xid);
 136
 137         rc = PtlPut(bulk->b_md_h, PTL_ACK_REQ, remote_id, portal, 0,
 138                     bulk->b_xid, 0, 0);
 139         if (rc != PTL_OK) {
 140                 CERROR("PtlPut(%d, %d, %d) failed: %d\n", remote_id.nid,
 141                        portal, bulk->b_xid, rc);
 142                 PtlMDUnlink(bulk->b_md_h);
 143                 LBUG();
 144         }
 145
 146         return rc;
 147 }
 148
 149 int ptlrpc_register_bulk(struct ptlrpc_bulk_desc *bulk)
 150 {
 151         int rc;
 152         ENTRY;
 153
 154         rc = PtlMEAttach(bulk->b_connection->c_peer.peer_ni, bulk->b_portal,
 155                          local_id, bulk->b_xid, 0, PTL_UNLINK, PTL_INS_AFTER,
 156                          &bulk->b_me_h);
 157         if (rc != PTL_OK) {
 158                 CERROR("PtlMEAttach failed: %d\n", rc);
 159                 LBUG();
 160                 GOTO(cleanup, rc);
 161         }
 162
 163         bulk->b_md.start = bulk->b_buf;
 164         bulk->b_md.length = bulk->b_buflen;
 165         bulk->b_md.threshold = 1;
 166         bulk->b_md.options = PTL_MD_OP_PUT;
 167         bulk->b_md.user_ptr = bulk;
 168         bulk->b_md.eventq = bulk_sink_eq;
 169
 170         rc = PtlMDAttach(bulk->b_me_h, bulk->b_md, PTL_UNLINK, &bulk->b_md_h);
 171         //CERROR("MDAttach (bulk sink): %Lu\n", (__u64)bulk->b_md_h);
 172         if (rc != PTL_OK) {
 173                 CERROR("PtlMDAttach failed: %d\n", rc);
 174                 LBUG();
 175                 GOTO(cleanup, rc);
 176         }
 177
 178         CDEBUG(D_NET, "Setup bulk sink buffer: %u bytes, xid %u, portal %u\n",
 179                bulk->b_buflen, bulk->b_xid, bulk->b_portal);
 180         RETURN(0);
 181
 182         // XXX Confirm that this is safe!
 183  cleanup:
 184         PtlMEUnlink(bulk->b_me_h);
 185         return rc;
 186 }
 187
 188 int ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *bulk)
 189 {
 190         int rc;
 191
 192         rc = PtlMEUnlink(bulk->b_me_h);
 193         if (rc != PTL_OK)
 194                 CERROR("PtlMEUnlink failed: %d\n", rc);
 195
 196         return rc;
 197 }
 198
 199 int ptlrpc_reply(struct ptlrpc_service *svc, struct ptlrpc_request *req)
 200 {
 201         /* FIXME: we need to increment the count of handled events */
 202         req->rq_type = PTL_RPC_TYPE_REPLY;
 203         req->rq_repmsg->conn = req->rq_connection->c_remote_conn;
 204         req->rq_repmsg->token = req->rq_connection->c_remote_token;
 205         req->rq_repmsg->xid = HTON__u32(req->rq_reqmsg->xid);
 206         req->rq_repmsg->status = HTON__u32(req->rq_status);
 207         req->rq_reqmsg->type = HTON__u32(req->rq_type);
 208         return ptl_send_buf(req, req->rq_connection, svc->srv_rep_portal);
 209 }
 210
 211 int ptlrpc_error(struct ptlrpc_service *svc, struct ptlrpc_request *req)
 212 {
 213         int rc;
 214         ENTRY;
 215
 216         if (req->rq_repmsg) {
 217                 CERROR("req already has repmsg\n");
 218                 LBUG();
 219         }
 220
 221         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 222         if (rc)
 223                 RETURN(rc);
 224
 225         req->rq_repmsg->type = HTON__u32(PTL_RPC_MSG_ERR);
 226
 227         rc = ptlrpc_reply(svc, req);
 228         RETURN(rc);
 229 }
 230
 231 void ptlrpc_resend_req(struct ptlrpc_request *req)
 232 {
 233         ENTRY;
 234         req->rq_flags |= PTL_RPC_FL_RESEND;
 235         req->rq_flags &= ~PTL_RPC_FL_TIMEOUT;
 236         wake_up_interruptible(&req->rq_wait_for_rep);
 237         EXIT;
 238         return;
 239 }
 240
 241 int ptl_send_rpc(struct ptlrpc_request *request)
 242 {
 243         ptl_process_id_t local_id;
 244         int rc;
 245         char *repbuf;
 246
 247         ENTRY;
 248
 249         if (NTOH__u32(request->rq_reqmsg->type) != PTL_RPC_MSG_REQUEST) {
 250                 CERROR("wrong packet type sent %d\n",
 251                        NTOH__u32(request->rq_reqmsg->type));
 252                 LBUG();
 253                 RETURN(-EINVAL);
 254         }
 255         if (request->rq_replen == 0) {
 256                 CERROR("request->rq_replen is 0!\n");
 257                 RETURN(-EINVAL);
 258         }
 259
 260         /* request->rq_repmsg is set only when the reply comes in, in
 261          * client_packet_callback() */
 262         OBD_ALLOC(repbuf, request->rq_replen);
 263         if (!repbuf)
 264                 RETURN(-ENOMEM);
 265
 266         local_id.nid = PTL_ID_ANY;
 267         local_id.pid = PTL_ID_ANY;
 268
 269         down(&request->rq_client->cli_rpc_sem);
 270
 271         rc = PtlMEAttach(request->rq_connection->c_peer.peer_ni,
 272                          request->rq_client->cli_reply_portal,
 273                          local_id, request->rq_reqmsg->xid, 0, PTL_UNLINK,
 274                          PTL_INS_AFTER, &request->rq_reply_me_h);
 275         if (rc != PTL_OK) {
 276                 CERROR("PtlMEAttach failed: %d\n", rc);
 277                 LBUG();
 278                 GOTO(cleanup, rc);
 279         }
 280
 281         request->rq_type = PTL_RPC_TYPE_REQUEST;
 282         request->rq_reply_md.start = repbuf;
 283         request->rq_reply_md.length = request->rq_replen;
 284         request->rq_reply_md.threshold = 1;
 285         request->rq_reply_md.options = PTL_MD_OP_PUT;
 286         request->rq_reply_md.user_ptr = request;
 287         request->rq_reply_md.eventq = reply_in_eq;
 288
 289         rc = PtlMDAttach(request->rq_reply_me_h, request->rq_reply_md,
 290                          PTL_UNLINK, &request->rq_reply_md_h);
 291         if (rc != PTL_OK) {
 292                 CERROR("PtlMDAttach failed: %d\n", rc);
 293                 LBUG();
 294                 GOTO(cleanup2, rc);
 295         }
 296
 297         CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %u, portal %u\n",
 298                request->rq_replen, request->rq_reqmsg->xid,
 299                request->rq_client->cli_reply_portal);
 300
 301         spin_lock(&request->rq_client->cli_lock);
 302         list_add(&request->rq_list, &request->rq_client->cli_sending_head);
 303         spin_unlock(&request->rq_client->cli_lock);
 304
 305         rc = ptl_send_buf(request, request->rq_connection,
 306                           request->rq_client->cli_request_portal);
 307         RETURN(rc);
 308
 309  cleanup2:
 310         PtlMEUnlink(request->rq_reply_me_h);
 311  cleanup:
 312         OBD_FREE(repbuf, request->rq_replen);
 313         up(&request->rq_client->cli_rpc_sem);
 314
 315         return rc;
 316 }
 317
 318 void ptlrpc_link_svc_me(struct ptlrpc_service *service, int i)
 319 {
 320         int rc;
 321         ptl_md_t dummy;
 322         ptl_handle_md_t md_h;
 323
 324         /* Attach the leading ME on which we build the ring */
 325         rc = PtlMEAttach(service->srv_self.peer_ni, service->srv_req_portal,
 326                          service->srv_id, 0, ~0, PTL_RETAIN, PTL_INS_BEFORE,
 327                          &(service->srv_me_h[i]));
 328         if (rc != PTL_OK) {
 329                 CERROR("PtlMEAttach failed: %d\n", rc);
 330                 LBUG();
 331         }
 332
 333         if (service->srv_ref_count[i])
 334                 LBUG();
 335
 336         dummy.start         = service->srv_buf[i];
 337         dummy.length        = service->srv_buf_size;
 338         dummy.max_offset    = service->srv_buf_size;
 339         dummy.threshold     = PTL_MD_THRESH_INF;
 340         dummy.options       = PTL_MD_OP_PUT | PTL_MD_AUTO_UNLINK;
 341         dummy.user_ptr      = service;
 342         dummy.eventq        = service->srv_eq_h;
 343         dummy.max_offset    = service->srv_buf_size;
 344
 345         rc = PtlMDAttach(service->srv_me_h[i], dummy, PTL_UNLINK, &md_h);
 346         if (rc != PTL_OK) {
 347                 /* cleanup */
 348                 CERROR("PtlMDAttach failed: %d\n", rc);
 349                 LBUG();
 350         }
 351 }
 352
 353 /* ptl_handled_rpc() should be called by the sleeping process once
 354  * it finishes processing an event.  This ensures the ref count is
 355  * decremented and that the rpc ring buffer cycles properly.
 356  */
 357 int ptl_handled_rpc(struct ptlrpc_service *service, void *start)
 358 {
 359         int index;
 360
 361         spin_lock(&service->srv_lock);
 362
 363         for (index = 0; index < service->srv_ring_length; index++)
 364                 if (service->srv_buf[index] == start)
 365                         break;
 366
 367         if (index == service->srv_ring_length)
 368                 LBUG();
 369
 370         CDEBUG(D_INFO, "MD index=%d Ref Count=%d\n", index,
 371                service->srv_ref_count[index]);
 372         service->srv_ref_count[index]--;
 373
 374         if (service->srv_ref_count[index] < 0)
 375                 LBUG();
 376
 377         if (service->srv_ref_count[index] == 0 &&
 378             !ptl_is_valid_handle(&(service->srv_me_h[index]))) {
 379                 CDEBUG(D_NET, "relinking %d\n", index);
 380                 ptlrpc_link_svc_me(service, index);
 381         }
 382
 383         spin_unlock(&service->srv_lock);
 384         return 0;
 385 }