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