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 #define DEBUG_SUBSYSTEM S_RPC
  25
  26 #include <linux/lustre_net.h>
  27
  28 extern ptl_handle_eq_t request_out_eq, reply_in_eq, reply_out_eq,
  29         bulk_source_eq, bulk_sink_eq;
  30 static ptl_process_id_t local_id = {PTL_ID_ANY, PTL_ID_ANY};
  31
  32 static int ptlrpc_check_bulk_sent(struct ptlrpc_bulk_desc *bulk)
  33 {
  34         ENTRY;
  35
  36         if (bulk->b_flags & PTL_BULK_FL_SENT)
  37                 RETURN(1);
  38
  39         if (sigismember(&(current->pending.signal), SIGKILL) ||
  40             sigismember(&(current->pending.signal), SIGINT)) {
  41                 bulk->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 static int ptl_send_buf(struct ptlrpc_request *request,
  50                         struct ptlrpc_connection *conn, int portal)
  51 {
  52         int rc;
  53         ptl_process_id_t remote_id;
  54         ptl_handle_md_t md_h;
  55         ptl_ack_req_t ack;
  56
  57         request->rq_req_md.user_ptr = request;
  58
  59         switch (request->rq_type) {
  60         case PTL_RPC_TYPE_REQUEST:
  61                 request->rq_req_md.start = request->rq_reqmsg;
  62                 request->rq_req_md.length = request->rq_reqlen;
  63                 request->rq_req_md.eventq = request_out_eq;
  64                 request->rq_req_md.threshold = 1;
  65                 ack = PTL_NOACK_REQ;
  66                 break;
  67         case PTL_RPC_TYPE_REPLY:
  68                 request->rq_req_md.start = request->rq_repmsg;
  69                 request->rq_req_md.length = request->rq_replen;
  70                 request->rq_req_md.eventq = reply_out_eq;
  71                 request->rq_req_md.threshold = 1;
  72                 ack = PTL_NOACK_REQ;
  73                 break;
  74         default:
  75                 LBUG();
  76                 return -1; /* notreached */
  77         }
  78         request->rq_req_md.options = PTL_MD_OP_PUT;
  79         request->rq_req_md.user_ptr = request;
  80
  81         rc = PtlMDBind(conn->c_peer.peer_ni, request->rq_req_md, &md_h);
  82         //CERROR("MDBind (outgoing req/rep/bulk): %Lu\n", (__u64)md_h);
  83         if (rc != 0) {
  84                 CERROR("PtlMDBind failed: %d\n", rc);
  85                 LBUG();
  86                 return rc;
  87         }
  88
  89         remote_id.nid = conn->c_peer.peer_nid;
  90         remote_id.pid = 0;
  91
  92         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %Ld\n",
  93                request->rq_req_md.length, portal, request->rq_xid);
  94
  95         rc = PtlPut(md_h, ack, remote_id, portal, 0, request->rq_reqmsg->xid,
  96                     0, 0);
  97         if (rc != PTL_OK) {
  98                 CERROR("PtlPut(%Lu, %d, %Ld) failed: %d\n", remote_id.nid,
  99                        portal, request->rq_xid, rc);
 100                 PtlMDUnlink(md_h);
 101         }
 102
 103         return rc;
 104 }
 105
 106 int ptlrpc_send_bulk(struct ptlrpc_bulk_desc *desc)
 107 {
 108         int rc;
 109         struct list_head *tmp, *next;
 110         ptl_process_id_t remote_id;
 111         ENTRY;
 112
 113         list_for_each_safe(tmp, next, &desc->b_page_list) {
 114                 /* only request an ACK for the last page */
 115                 int ack = (next == &desc->b_page_list);
 116                 struct ptlrpc_bulk_page *bulk;
 117                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, b_link);
 118
 119                 bulk->b_md.start = bulk->b_buf;
 120                 bulk->b_md.length = bulk->b_buflen;
 121                 bulk->b_md.eventq = bulk_source_eq;
 122                 bulk->b_md.threshold = 1 + ack; /* SENT and (if last) ACK */
 123                 bulk->b_md.options = PTL_MD_OP_PUT;
 124                 bulk->b_md.user_ptr = bulk;
 125
 126                 rc = PtlMDBind(desc->b_connection->c_peer.peer_ni, bulk->b_md,
 127                                &bulk->b_md_h);
 128                 if (rc != 0) {
 129                         CERROR("PtlMDBind failed: %d\n", rc);
 130                         LBUG();
 131                         RETURN(rc);
 132                 }
 133
 134                 remote_id.nid = desc->b_connection->c_peer.peer_nid;
 135                 remote_id.pid = 0;
 136
 137                 CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %d\n",
 138                        bulk->b_md.length, desc->b_portal, bulk->b_xid);
 139
 140                 rc = PtlPut(bulk->b_md_h, (ack ? PTL_ACK_REQ : PTL_NOACK_REQ),
 141                             remote_id, desc->b_portal, 0, bulk->b_xid, 0, 0);
 142                 if (rc != PTL_OK) {
 143                         CERROR("PtlPut(%Lu, %d, %d) failed: %d\n", remote_id.nid,
 144                                desc->b_portal, bulk->b_xid, rc);
 145                         PtlMDUnlink(bulk->b_md_h);
 146                         LBUG();
 147                         RETURN(rc);
 148                 }
 149         }
 150
 151         wait_event_interruptible(desc->b_waitq, ptlrpc_check_bulk_sent(desc));
 152
 153         if (desc->b_flags & PTL_RPC_FL_INTR)
 154                 RETURN(-EINTR);
 155
 156         RETURN(0);
 157 }
 158
 159 int ptlrpc_register_bulk(struct ptlrpc_bulk_desc *desc)
 160 {
 161         struct list_head *tmp, *next;
 162         int rc;
 163         ENTRY;
 164
 165         list_for_each_safe(tmp, next, &desc->b_page_list) {
 166                 struct ptlrpc_bulk_page *bulk;
 167                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, b_link);
 168
 169                 rc = PtlMEAttach(desc->b_connection->c_peer.peer_ni,
 170                                  desc->b_portal, local_id, bulk->b_xid, 0,
 171                                  PTL_UNLINK, PTL_INS_AFTER, &bulk->b_me_h);
 172                 if (rc != PTL_OK) {
 173                         CERROR("PtlMEAttach failed: %d\n", rc);
 174                         LBUG();
 175                         GOTO(cleanup, rc);
 176                 }
 177
 178                 bulk->b_md.start = bulk->b_buf;
 179                 bulk->b_md.length = bulk->b_buflen;
 180                 bulk->b_md.threshold = 1;
 181                 bulk->b_md.options = PTL_MD_OP_PUT;
 182                 bulk->b_md.user_ptr = bulk;
 183                 bulk->b_md.eventq = bulk_sink_eq;
 184
 185                 rc = PtlMDAttach(bulk->b_me_h, bulk->b_md, PTL_UNLINK,
 186                                  &bulk->b_md_h);
 187                 if (rc != PTL_OK) {
 188                         CERROR("PtlMDAttach failed: %d\n", rc);
 189                         LBUG();
 190                         GOTO(cleanup, rc);
 191                 }
 192
 193                 CDEBUG(D_NET, "Setup bulk sink buffer: %u bytes, xid %u, "
 194                        "portal %u\n", bulk->b_buflen, bulk->b_xid,
 195                        desc->b_portal);
 196         }
 197
 198         RETURN(0);
 199
 200  cleanup:
 201         ptlrpc_abort_bulk(desc);
 202
 203         return rc;
 204 }
 205
 206 int ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
 207 {
 208         struct list_head *tmp, *next;
 209
 210         list_for_each_safe(tmp, next, &desc->b_page_list) {
 211                 struct ptlrpc_bulk_page *bulk;
 212                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, b_link);
 213
 214                 /* This should be safe: these handles are initialized to be
 215                  * invalid in ptlrpc_prep_bulk_page() */
 216                 PtlMDUnlink(bulk->b_md_h);
 217                 PtlMEUnlink(bulk->b_me_h);
 218         }
 219
 220         return 0;
 221 }
 222
 223 int ptlrpc_reply(struct ptlrpc_service *svc, struct ptlrpc_request *req)
 224 {
 225         /* FIXME: we need to increment the count of handled events */
 226         req->rq_type = PTL_RPC_TYPE_REPLY;
 227         req->rq_repmsg->conn = req->rq_connection->c_remote_conn;
 228         req->rq_repmsg->token = req->rq_connection->c_remote_token;
 229         req->rq_repmsg->xid = HTON__u32(req->rq_reqmsg->xid);
 230         req->rq_repmsg->status = HTON__u32(req->rq_status);
 231         req->rq_reqmsg->type = HTON__u32(req->rq_type);
 232         return ptl_send_buf(req, req->rq_connection, svc->srv_rep_portal);
 233 }
 234
 235 int ptlrpc_error(struct ptlrpc_service *svc, struct ptlrpc_request *req)
 236 {
 237         int rc;
 238         ENTRY;
 239
 240         if (req->rq_repmsg) {
 241                 CERROR("req already has repmsg\n");
 242                 LBUG();
 243         }
 244
 245         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 246         if (rc)
 247                 RETURN(rc);
 248
 249         req->rq_repmsg->type = HTON__u32(PTL_RPC_MSG_ERR);
 250
 251         rc = ptlrpc_reply(svc, req);
 252         RETURN(rc);
 253 }
 254
 255
 256 int ptl_send_rpc(struct ptlrpc_request *request)
 257 {
 258         ptl_process_id_t local_id;
 259         int rc;
 260         char *repbuf;
 261
 262         ENTRY;
 263
 264         if (NTOH__u32(request->rq_reqmsg->type) != PTL_RPC_MSG_REQUEST) {
 265                 CERROR("wrong packet type sent %d\n",
 266                        NTOH__u32(request->rq_reqmsg->type));
 267                 LBUG();
 268                 RETURN(EINVAL);
 269         }
 270         if (request->rq_replen == 0) {
 271                 CERROR("request->rq_replen is 0!\n");
 272                 RETURN(EINVAL);
 273         }
 274
 275         /* request->rq_repmsg is set only when the reply comes in, in
 276          * client_packet_callback() */
 277         if (request->rq_reply_md.start)
 278                 OBD_FREE(request->rq_reply_md.start, request->rq_replen);
 279
 280         OBD_ALLOC(repbuf, request->rq_replen);
 281         if (!repbuf)
 282                 RETURN(ENOMEM);
 283
 284         local_id.nid = PTL_ID_ANY;
 285         local_id.pid = PTL_ID_ANY;
 286
 287         down(&request->rq_client->cli_rpc_sem);
 288
 289         rc = PtlMEAttach(request->rq_connection->c_peer.peer_ni,
 290                          request->rq_client->cli_reply_portal,
 291                          local_id, request->rq_reqmsg->xid, 0, PTL_UNLINK,
 292                          PTL_INS_AFTER, &request->rq_reply_me_h);
 293         if (rc != PTL_OK) {
 294                 CERROR("PtlMEAttach failed: %d\n", rc);
 295                 LBUG();
 296                 GOTO(cleanup, rc);
 297         }
 298
 299         request->rq_type = PTL_RPC_TYPE_REQUEST;
 300         request->rq_reply_md.start = repbuf;
 301         request->rq_reply_md.length = request->rq_replen;
 302         request->rq_reply_md.threshold = 1;
 303         request->rq_reply_md.options = PTL_MD_OP_PUT;
 304         request->rq_reply_md.user_ptr = request;
 305         request->rq_reply_md.eventq = reply_in_eq;
 306
 307         rc = PtlMDAttach(request->rq_reply_me_h, request->rq_reply_md,
 308                          PTL_UNLINK, &request->rq_reply_md_h);
 309         if (rc != PTL_OK) {
 310                 CERROR("PtlMDAttach failed: %d\n", rc);
 311                 LBUG();
 312                 GOTO(cleanup2, rc);
 313         }
 314
 315         CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %Lu, portal %u\n",
 316                request->rq_replen, request->rq_xid,
 317                request->rq_client->cli_reply_portal);
 318
 319         rc = ptl_send_buf(request, request->rq_connection,
 320                           request->rq_client->cli_request_portal);
 321         RETURN(rc);
 322
 323  cleanup2:
 324         PtlMEUnlink(request->rq_reply_me_h);
 325  cleanup:
 326         OBD_FREE(repbuf, request->rq_replen);
 327         up(&request->rq_client->cli_rpc_sem);
 328
 329         return rc;
 330 }
 331
 332 void ptlrpc_link_svc_me(struct ptlrpc_service *service, int i)
 333 {
 334         int rc;
 335         ptl_md_t dummy;
 336         ptl_handle_md_t md_h;
 337
 338         /* Attach the leading ME on which we build the ring */
 339         rc = PtlMEAttach(service->srv_self.peer_ni, service->srv_req_portal,
 340                          local_id, 0, ~0, PTL_RETAIN, PTL_INS_BEFORE,
 341                          &(service->srv_me_h[i]));
 342         if (rc != PTL_OK) {
 343                 CERROR("PtlMEAttach failed: %d\n", rc);
 344                 LBUG();
 345         }
 346
 347         if (service->srv_ref_count[i])
 348                 LBUG();
 349
 350         dummy.start         = service->srv_buf[i];
 351         dummy.length        = service->srv_buf_size;
 352         dummy.max_offset    = service->srv_buf_size;
 353         dummy.threshold     = PTL_MD_THRESH_INF;
 354         dummy.options       = PTL_MD_OP_PUT | PTL_MD_AUTO_UNLINK;
 355         dummy.user_ptr      = service;
 356         dummy.eventq        = service->srv_eq_h;
 357         dummy.max_offset    = service->srv_buf_size;
 358
 359         rc = PtlMDAttach(service->srv_me_h[i], dummy, PTL_UNLINK, &md_h);
 360         if (rc != PTL_OK) {
 361                 /* cleanup */
 362                 CERROR("PtlMDAttach failed: %d\n", rc);
 363                 LBUG();
 364         }
 365 }
 366
 367 /* ptl_handled_rpc() should be called by the sleeping process once
 368  * it finishes processing an event.  This ensures the ref count is
 369  * decremented and that the rpc ring buffer cycles properly.
 370  */
 371 int ptl_handled_rpc(struct ptlrpc_service *service, void *start)
 372 {
 373         int index;
 374
 375         spin_lock(&service->srv_lock);
 376
 377         for (index = 0; index < service->srv_ring_length; index++)
 378                 if (service->srv_buf[index] == start)
 379                         break;
 380
 381         if (index == service->srv_ring_length)
 382                 LBUG();
 383
 384         CDEBUG(D_INFO, "MD index=%d Ref Count=%d\n", index,
 385                service->srv_ref_count[index]);
 386         service->srv_ref_count[index]--;
 387
 388         if (service->srv_ref_count[index] < 0)
 389                 LBUG();
 390
 391         if (service->srv_ref_count[index] == 0 &&
 392             !ptl_is_valid_handle(&(service->srv_me_h[index]))) {
 393                 CDEBUG(D_NET, "relinking %d\n", index);
 394                 ptlrpc_link_svc_me(service, index);
 395         }
 396
 397         spin_unlock(&service->srv_lock);
 398         return 0;
 399 }