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 static int ptl_send_buf(struct ptlrpc_request *request,
  34                         struct ptlrpc_connection *conn, int portal)
  35 {
  36         int rc;
  37         ptl_process_id_t remote_id;
  38         ptl_handle_md_t md_h;
  39
  40         request->rq_req_md.user_ptr = request;
  41
  42         switch (request->rq_type) {
  43         case PTL_RPC_MSG_REQUEST:
  44                 request->rq_reqmsg->type = HTON__u32(request->rq_type);
  45                 request->rq_req_md.start = request->rq_reqmsg;
  46                 request->rq_req_md.length = request->rq_reqlen;
  47                 request->rq_req_md.eventq = request_out_eq;
  48                 break;
  49         case PTL_RPC_MSG_REPLY:
  50                 request->rq_repmsg->type = HTON__u32(request->rq_type);
  51                 request->rq_req_md.start = request->rq_repmsg;
  52                 request->rq_req_md.length = request->rq_replen;
  53                 request->rq_req_md.eventq = reply_out_eq;
  54                 break;
  55         default:
  56                 LBUG();
  57                 return -1; /* notreached */
  58         }
  59         request->rq_req_md.threshold = 1;
  60         request->rq_req_md.options = PTL_MD_OP_PUT;
  61         request->rq_req_md.user_ptr = request;
  62
  63         rc = PtlMDBind(conn->c_peer.peer_ni, request->rq_req_md, &md_h);
  64         if (rc != 0) {
  65                 CERROR("PtlMDBind failed: %d\n", rc);
  66                 LBUG();
  67                 return rc;
  68         }
  69
  70         remote_id.nid = conn->c_peer.peer_nid;
  71         remote_id.pid = 0;
  72
  73         CDEBUG(D_NET, "Sending %d bytes to portal %d, xid %Ld\n",
  74                request->rq_req_md.length, portal, request->rq_xid);
  75
  76         if (!portal)
  77                 LBUG();
  78         rc = PtlPut(md_h, PTL_NOACK_REQ, remote_id, portal, 0, request->rq_xid,
  79                     0, 0);
  80         if (rc != PTL_OK) {
  81                 CERROR("PtlPut(%Lu, %d, %Ld) failed: %d\n", remote_id.nid,
  82                        portal, request->rq_xid, rc);
  83                 PtlMDUnlink(md_h);
  84         }
  85
  86         return rc;
  87 }
  88
  89 static inline struct iovec *
  90 ptlrpc_get_bulk_iov (struct ptlrpc_bulk_desc *desc)
  91 {
  92         struct iovec *iov;
  93
  94         if (desc->bd_page_count <= sizeof (desc->bd_iov)/sizeof (struct iovec))
  95                 return (desc->bd_iov);
  96
  97         OBD_ALLOC (iov, desc->bd_page_count * sizeof (struct iovec));
  98         if (iov == NULL)
  99                 LBUG();
 100
 101         return (iov);
 102 }
 103
 104 static inline void
 105 ptlrpc_put_bulk_iov (struct ptlrpc_bulk_desc *desc, struct iovec *iov)
 106 {
 107         if (desc->bd_page_count <= sizeof (desc->bd_iov)/sizeof (struct iovec))
 108                 return;
 109
 110         OBD_FREE (iov, desc->bd_page_count * sizeof (struct iovec));
 111 }
 112
 113 int ptlrpc_send_bulk(struct ptlrpc_bulk_desc *desc)
 114 {
 115         int rc;
 116         struct list_head *tmp, *next;
 117         ptl_process_id_t remote_id;
 118         __u32 xid = 0;
 119         struct iovec *iov;
 120         ENTRY;
 121
 122         iov = ptlrpc_get_bulk_iov (desc);
 123         if (iov == NULL)
 124                 RETURN (-ENOMEM);
 125
 126         desc->bd_md.start = iov;
 127         desc->bd_md.niov = 0;
 128         desc->bd_md.length = 0;
 129         desc->bd_md.eventq = bulk_source_eq;
 130         desc->bd_md.threshold = 2; /* SENT and ACK */
 131         desc->bd_md.options = PTL_MD_OP_PUT | PTL_MD_IOV;
 132         desc->bd_md.user_ptr = desc;
 133
 134         list_for_each_safe(tmp, next, &desc->bd_page_list) {
 135                 struct ptlrpc_bulk_page *bulk;
 136                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
 137
 138                 LASSERT (desc->bd_md.niov < desc->bd_page_count);
 139
 140                 if (desc->bd_md.niov == 0)
 141                         xid = bulk->bp_xid;
 142                 LASSERT (xid == bulk->bp_xid);   /* should all be the same */
 143
 144                 iov[desc->bd_md.niov].iov_base = bulk->bp_buf;
 145                 iov[desc->bd_md.niov].iov_len = bulk->bp_buflen;
 146                 desc->bd_md.niov++;
 147                 desc->bd_md.length += bulk->bp_buflen;
 148         }
 149
 150         LASSERT (desc->bd_md.niov == desc->bd_page_count);
 151         LASSERT (desc->bd_md.niov != 0);
 152
 153         rc = PtlMDBind(desc->bd_connection->c_peer.peer_ni, desc->bd_md,
 154                        &desc->bd_md_h);
 155
 156         ptlrpc_put_bulk_iov (desc, iov);        /* move down to reduce latency to send */
 157
 158         if (rc != PTL_OK) {
 159                 CERROR("PtlMDBind failed: %d\n", rc);
 160                 LBUG();
 161                 RETURN(rc);
 162         }
 163
 164         remote_id.nid = desc->bd_connection->c_peer.peer_nid;
 165         remote_id.pid = 0;
 166
 167         CDEBUG(D_NET, "Sending %u pages %u bytes to portal %d nid %Lx pid %d xid %d\n",
 168                desc->bd_md.niov, desc->bd_md.length,
 169                desc->bd_portal, remote_id.nid, remote_id.pid, xid);
 170
 171         rc = PtlPut(desc->bd_md_h, PTL_ACK_REQ, remote_id,
 172                     desc->bd_portal, 0, xid, 0, 0);
 173         if (rc != PTL_OK) {
 174                 CERROR("PtlPut(%Lu, %d, %d) failed: %d\n",
 175                        remote_id.nid, desc->bd_portal, xid, rc);
 176                 PtlMDUnlink(desc->bd_md_h);
 177                 LBUG();
 178                 RETURN(rc);
 179         }
 180
 181         RETURN(0);
 182 }
 183
 184 int ptlrpc_register_bulk(struct ptlrpc_bulk_desc *desc)
 185 {
 186         struct list_head *tmp, *next;
 187         int rc;
 188         __u32 xid = 0;
 189         struct iovec *iov;
 190         ptl_process_id_t source_id;
 191         ENTRY;
 192
 193         iov = ptlrpc_get_bulk_iov (desc);
 194         if (iov == NULL)
 195                 return (-ENOMEM);
 196
 197         desc->bd_md.start = iov;
 198         desc->bd_md.niov = 0;
 199         desc->bd_md.length = 0;
 200         desc->bd_md.threshold = 1;
 201         desc->bd_md.options = PTL_MD_OP_PUT | PTL_MD_IOV;
 202         desc->bd_md.user_ptr = desc;
 203         desc->bd_md.eventq = bulk_sink_eq;
 204
 205         list_for_each_safe(tmp, next, &desc->bd_page_list) {
 206                 struct ptlrpc_bulk_page *bulk;
 207                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
 208
 209                 LASSERT (desc->bd_md.niov < desc->bd_page_count);
 210
 211                 if (desc->bd_md.niov == 0)
 212                         xid = bulk->bp_xid;
 213                 LASSERT (xid == bulk->bp_xid);   /* should all be the same */
 214
 215                 iov[desc->bd_md.niov].iov_base = bulk->bp_buf;
 216                 iov[desc->bd_md.niov].iov_len = bulk->bp_buflen;
 217                 desc->bd_md.niov++;
 218                 desc->bd_md.length += bulk->bp_buflen;
 219         }
 220
 221         LASSERT (desc->bd_md.niov == desc->bd_page_count);
 222         LASSERT (desc->bd_md.niov != 0);
 223
 224         source_id.nid = desc->bd_connection->c_peer.peer_nid;
 225         source_id.pid = PTL_PID_ANY;
 226
 227         rc = PtlMEAttach(desc->bd_connection->c_peer.peer_ni,
 228                          desc->bd_portal, source_id, xid, 0,
 229                          PTL_UNLINK, PTL_INS_AFTER, &desc->bd_me_h);
 230
 231         ptlrpc_put_bulk_iov (desc, iov);
 232
 233         if (rc != PTL_OK) {
 234                 CERROR("PtlMEAttach failed: %d\n", rc);
 235                 LBUG();
 236                 GOTO(cleanup, rc);
 237         }
 238
 239         rc = PtlMDAttach(desc->bd_me_h, desc->bd_md, PTL_UNLINK,
 240                          &desc->bd_md_h);
 241         if (rc != PTL_OK) {
 242                 CERROR("PtlMDAttach failed: %d\n", rc);
 243                 LBUG();
 244                 GOTO(cleanup, rc);
 245         }
 246
 247         CDEBUG(D_NET, "Setup bulk sink buffers: %u pages %u bytes, xid %u, "
 248                "portal %u\n", desc->bd_md.niov, desc->bd_md.length,
 249                xid, desc->bd_portal);
 250
 251         RETURN(0);
 252
 253  cleanup:
 254         ptlrpc_abort_bulk(desc);
 255
 256         return rc;
 257 }
 258
 259 int ptlrpc_abort_bulk(struct ptlrpc_bulk_desc *desc)
 260 {
 261         /* This should be safe: these handles are initialized to be
 262          * invalid in ptlrpc_prep_bulk() */
 263         PtlMDUnlink(desc->bd_md_h);
 264         PtlMEUnlink(desc->bd_me_h);
 265
 266         return 0;
 267 }
 268
 269 int ptlrpc_reply(struct ptlrpc_service *svc, struct ptlrpc_request *req)
 270 {
 271         if (req->rq_repmsg == NULL) {
 272                 CERROR("bad: someone called ptlrpc_reply when they meant "
 273                        "ptlrpc_error\n");
 274                 return -EINVAL;
 275         }
 276
 277         /* FIXME: we need to increment the count of handled events */
 278         if (req->rq_type != PTL_RPC_MSG_ERR)
 279                 req->rq_type = PTL_RPC_MSG_REPLY;
 280         //req->rq_repmsg->conn = req->rq_connection->c_remote_conn;
 281         //req->rq_repmsg->token = req->rq_connection->c_remote_token;
 282         req->rq_repmsg->status = HTON__u32(req->rq_status);
 283         return ptl_send_buf(req, req->rq_connection, svc->srv_rep_portal);
 284 }
 285
 286 int ptlrpc_error(struct ptlrpc_service *svc, struct ptlrpc_request *req)
 287 {
 288         int rc;
 289         ENTRY;
 290
 291         if (req->rq_repmsg) {
 292                 CERROR("req already has repmsg\n");
 293                 LBUG();
 294         }
 295
 296         rc = lustre_pack_msg(0, NULL, NULL, &req->rq_replen, &req->rq_repmsg);
 297         if (rc)
 298                 RETURN(rc);
 299
 300         req->rq_repmsg->type = HTON__u32(PTL_RPC_MSG_ERR);
 301
 302         rc = ptlrpc_reply(svc, req);
 303         RETURN(rc);
 304 }
 305
 306 int ptl_send_rpc(struct ptlrpc_request *request)
 307 {
 308         int rc;
 309         char *repbuf;
 310         ptl_process_id_t source_id;
 311
 312         ENTRY;
 313
 314         if (request->rq_type != PTL_RPC_MSG_REQUEST) {
 315                 CERROR("wrong packet type sent %d\n",
 316                        NTOH__u32(request->rq_reqmsg->type));
 317                 LBUG();
 318                 RETURN(EINVAL);
 319         }
 320         if (request->rq_replen == 0) {
 321                 CERROR("request->rq_replen is 0!\n");
 322                 RETURN(EINVAL);
 323         }
 324
 325         /* request->rq_repmsg is set only when the reply comes in, in
 326          * client_packet_callback() */
 327         if (request->rq_reply_md.start)
 328                 OBD_FREE(request->rq_reply_md.start, request->rq_replen);
 329
 330         OBD_ALLOC(repbuf, request->rq_replen);
 331         if (!repbuf) {
 332                 LBUG();
 333                 RETURN(ENOMEM);
 334         }
 335
 336         // down(&request->rq_client->cli_rpc_sem);
 337
 338         source_id.nid = request->rq_connection->c_peer.peer_nid;
 339         source_id.pid = PTL_PID_ANY;
 340
 341         rc = PtlMEAttach(request->rq_connection->c_peer.peer_ni,
 342                          request->rq_import->imp_client->cli_reply_portal,
 343                          source_id, request->rq_xid, 0, PTL_UNLINK,
 344                          PTL_INS_AFTER, &request->rq_reply_me_h);
 345         if (rc != PTL_OK) {
 346                 CERROR("PtlMEAttach failed: %d\n", rc);
 347                 LBUG();
 348                 GOTO(cleanup, rc);
 349         }
 350
 351         request->rq_reply_md.start = repbuf;
 352         request->rq_reply_md.length = request->rq_replen;
 353         request->rq_reply_md.threshold = 1;
 354         request->rq_reply_md.options = PTL_MD_OP_PUT;
 355         request->rq_reply_md.user_ptr = request;
 356         request->rq_reply_md.eventq = reply_in_eq;
 357
 358         rc = PtlMDAttach(request->rq_reply_me_h, request->rq_reply_md,
 359                          PTL_UNLINK, &request->rq_reply_md_h);
 360         if (rc != PTL_OK) {
 361                 CERROR("PtlMDAttach failed: %d\n", rc);
 362                 LBUG();
 363                 GOTO(cleanup2, rc);
 364         }
 365
 366         CDEBUG(D_NET, "Setup reply buffer: %u bytes, xid %Lu, portal %u\n",
 367                request->rq_replen, request->rq_xid,
 368                request->rq_import->imp_client->cli_reply_portal);
 369
 370         rc = ptl_send_buf(request, request->rq_connection,
 371                           request->rq_import->imp_client->cli_request_portal);
 372         RETURN(rc);
 373
 374  cleanup2:
 375         PtlMEUnlink(request->rq_reply_me_h);
 376  cleanup:
 377         OBD_FREE(repbuf, request->rq_replen);
 378         // up(&request->rq_client->cli_rpc_sem);
 379
 380         return rc;
 381 }
 382
 383 void ptlrpc_link_svc_me(struct ptlrpc_service *service, int i)
 384 {
 385         int rc;
 386         ptl_md_t dummy;
 387         ptl_handle_md_t md_h;
 388
 389         /* Attach the leading ME on which we build the ring */
 390         rc = PtlMEAttach(service->srv_self.peer_ni, service->srv_req_portal,
 391                          local_id, 0, ~0, PTL_RETAIN, PTL_INS_BEFORE,
 392                          &(service->srv_me_h[i]));
 393         if (rc != PTL_OK) {
 394                 CERROR("PtlMEAttach failed: %d\n", rc);
 395                 LBUG();
 396         }
 397
 398         if (service->srv_ref_count[i])
 399                 LBUG();
 400
 401         dummy.start         = service->srv_buf[i];
 402         dummy.length        = service->srv_buf_size;
 403         dummy.max_offset    = service->srv_buf_size;
 404         dummy.threshold     = PTL_MD_THRESH_INF;
 405         dummy.options       = PTL_MD_OP_PUT | PTL_MD_AUTO_UNLINK;
 406         dummy.user_ptr      = service;
 407         dummy.eventq        = service->srv_eq_h;
 408         dummy.max_offset    = service->srv_buf_size;
 409
 410         rc = PtlMDAttach(service->srv_me_h[i], dummy, PTL_UNLINK, &md_h);
 411         if (rc != PTL_OK) {
 412                 /* cleanup */
 413                 CERROR("PtlMDAttach failed: %d\n", rc);
 414                 LBUG();
 415         }
 416 }
 417
 418 /* ptl_handled_rpc() should be called by the sleeping process once
 419  * it finishes processing an event.  This ensures the ref count is
 420  * decremented and that the rpc ring buffer cycles properly.
 421  */
 422 int ptl_handled_rpc(struct ptlrpc_service *service, void *start)
 423 {
 424         int index;
 425
 426         spin_lock(&service->srv_lock);
 427         for (index = 0; index < service->srv_ring_length; index++)
 428                 if (service->srv_buf[index] == start)
 429                         break;
 430
 431         if (index == service->srv_ring_length)
 432                 LBUG();
 433
 434         CDEBUG(D_INFO, "MD index=%d Ref Count=%d\n", index,
 435                service->srv_ref_count[index]);
 436         service->srv_ref_count[index]--;
 437
 438         if (service->srv_ref_count[index] < 0)
 439                 LBUG();
 440
 441         if (service->srv_ref_count[index] == 0 &&
 442             !ptl_is_valid_handle(&(service->srv_me_h[index]))) {
 443                 CDEBUG(D_NET, "relinking %d\n", index);
 444                 ptlrpc_link_svc_me(service, index);
 445         }
 446
 447         spin_unlock(&service->srv_lock);
 448         return 0;
 449 }