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