lustre/ptlrpc/events.c

   1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
   2  * vim:expandtab:shiftwidth=8:tabstop=8:
   3  *
   4  *  Copyright (c) 2002, 2003 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 #ifdef __KERNEL__
  26 #include <linux/module.h>
  27 #else
  28 #include <liblustre.h>
  29 #endif
  30 #include <linux/obd_class.h>
  31 #include <linux/lustre_net.h>
  32
  33 struct ptlrpc_ni  ptlrpc_interfaces[NAL_MAX_NR];
  34 int               ptlrpc_ninterfaces;
  35
  36 /*
  37  *  Free the packet when it has gone out
  38  */
  39 static int request_out_callback(ptl_event_t *ev)
  40 {
  41         struct ptlrpc_request *req = ev->mem_desc.user_ptr;
  42         ENTRY;
  43
  44         /* requests always contiguous */
  45         LASSERT((ev->mem_desc.options & PTL_MD_IOV) == 0);
  46
  47         if (ev->type != PTL_EVENT_SENT) {
  48                 // XXX make sure we understand all events, including ACK's
  49                 CERROR("Unknown event %d\n", ev->type);
  50                 LBUG();
  51         }
  52
  53         /* this balances the atomic_inc in ptl_send_rpc */
  54         ptlrpc_req_finished(req);
  55         RETURN(1);
  56 }
  57
  58
  59 /*
  60  *  Free the packet when it has gone out
  61  */
  62 static int reply_out_callback(ptl_event_t *ev)
  63 {
  64         ENTRY;
  65
  66         /* replies always contiguous */
  67         LASSERT((ev->mem_desc.options & PTL_MD_IOV) == 0);
  68
  69         if (ev->type == PTL_EVENT_SENT) {
  70                 OBD_FREE(ev->mem_desc.start, ev->mem_desc.length);
  71         } else if (ev->type == PTL_EVENT_ACK) {
  72                 struct ptlrpc_request *req = ev->mem_desc.user_ptr;
  73                 if (req->rq_flags & PTL_RPC_FL_WANT_ACK) {
  74                         req->rq_flags &= ~PTL_RPC_FL_WANT_ACK;
  75                         wake_up(&req->rq_wait_for_rep);
  76                 } else {
  77                         DEBUG_REQ(D_ERROR, req,
  78                                   "ack received for reply, not wanted");
  79                 }
  80         } else {
  81                 // XXX make sure we understand all events
  82                 CERROR("Unknown event %d\n", ev->type);
  83                 LBUG();
  84         }
  85
  86         RETURN(1);
  87 }
  88
  89 /*
  90  * Wake up the thread waiting for the reply once it comes in.
  91  */
  92 int reply_in_callback(ptl_event_t *ev)
  93 {
  94         struct ptlrpc_request *req = ev->mem_desc.user_ptr;
  95         ENTRY;
  96
  97         /* replies always contiguous */
  98         LASSERT((ev->mem_desc.options & PTL_MD_IOV) == 0);
  99
 100         if (req->rq_xid == 0x5a5a5a5a5a5a5a5a) {
 101                 CERROR("Reply received for freed request!  Probably a missing "
 102                        "ptlrpc_abort()\n");
 103                 LBUG();
 104         }
 105
 106         if (req->rq_xid != ev->match_bits) {
 107                 CERROR("Reply packet for wrong request\n");
 108                 LBUG();
 109         }
 110
 111         if (ev->type == PTL_EVENT_PUT) {
 112                 req->rq_repmsg = ev->mem_desc.start + ev->offset;
 113                 barrier();
 114                 wake_up(&req->rq_wait_for_rep);
 115         } else {
 116                 // XXX make sure we understand all events, including ACK's
 117                 CERROR("Unknown event %d\n", ev->type);
 118                 LBUG();
 119         }
 120
 121         RETURN(1);
 122 }
 123
 124 int request_in_callback(ptl_event_t *ev)
 125 {
 126         struct ptlrpc_request_buffer_desc *rqbd = ev->mem_desc.user_ptr;
 127         struct ptlrpc_srv_ni  *srv_ni = rqbd->rqbd_srv_ni;
 128         struct ptlrpc_service *service = srv_ni->sni_service;
 129
 130         /* requests always contiguous */
 131         LASSERT((ev->mem_desc.options & PTL_MD_IOV) == 0);
 132         /* we only enable puts */
 133         LASSERT(ev->type == PTL_EVENT_PUT);
 134         LASSERT(atomic_read(&srv_ni->sni_nrqbds_receiving) > 0);
 135         LASSERT(atomic_read(&rqbd->rqbd_refcount) > 0);
 136
 137         if (ev->rlength != ev->mlength)
 138                 CERROR("Warning: Possibly truncated rpc (%d/%d)\n",
 139                        ev->mlength, ev->rlength);
 140
 141         if (ptl_is_valid_handle(&ev->unlinked_me)) {
 142                 /* This is the last request to be received into this
 143                  * request buffer.  We don't bump the refcount, since the
 144                  * thread servicing this event is effectively taking over
 145                  * portals' reference.
 146                  */
 147 #warning ev->unlinked_me.nal_idx is not set properly in a callback
 148                 LASSERT(ev->unlinked_me.handle_idx==rqbd->rqbd_me_h.handle_idx);
 149
 150                 /* we're off the air */
 151                 /* we'll probably start dropping packets in portals soon */
 152                 if (atomic_dec_and_test(&srv_ni->sni_nrqbds_receiving))
 153                         CERROR("All request buffers busy\n");
 154         } else {
 155                 /* +1 ref for service thread */
 156                 atomic_inc(&rqbd->rqbd_refcount);
 157         }
 158
 159         wake_up(&service->srv_waitq);
 160
 161         return 0;
 162 }
 163
 164 static int bulk_put_source_callback(ptl_event_t *ev)
 165 {
 166         struct ptlrpc_bulk_desc *desc = ev->mem_desc.user_ptr;
 167         struct ptlrpc_bulk_page *bulk;
 168         struct list_head        *tmp;
 169         struct list_head        *next;
 170         ENTRY;
 171
 172         CDEBUG(D_NET, "got %s event %d\n",
 173                (ev->type == PTL_EVENT_SENT) ? "SENT" :
 174                (ev->type == PTL_EVENT_ACK)  ? "ACK"  : "UNEXPECTED", ev->type);
 175
 176         LASSERT(ev->type == PTL_EVENT_SENT || ev->type == PTL_EVENT_ACK);
 177
 178         LASSERT(atomic_read(&desc->bd_source_callback_count) > 0 &&
 179                 atomic_read(&desc->bd_source_callback_count) <= 2);
 180
 181         /* 1 fragment for each page always */
 182         LASSERT(ev->mem_desc.niov == desc->bd_page_count);
 183
 184         if (atomic_dec_and_test(&desc->bd_source_callback_count)) {
 185                 void (*event_handler)(struct ptlrpc_bulk_desc *);
 186
 187                 list_for_each_safe(tmp, next, &desc->bd_page_list) {
 188                         bulk = list_entry(tmp, struct ptlrpc_bulk_page,
 189                                           bp_link);
 190
 191                         if (bulk->bp_cb != NULL)
 192                                 bulk->bp_cb(bulk);
 193                 }
 194
 195                 /* We need to make a note of whether there's an event handler
 196                  * before we call wake_up, because if there is no event handler,
 197                  * 'desc' might be freed before we're scheduled again. */
 198                 event_handler = desc->bd_ptl_ev_hdlr;
 199
 200                 desc->bd_flags |= PTL_BULK_FL_SENT;
 201                 wake_up(&desc->bd_waitq);
 202                 if (event_handler) {
 203                         LASSERT(desc->bd_ptl_ev_hdlr == event_handler);
 204                         event_handler(desc);
 205                 }
 206         }
 207
 208         RETURN(0);
 209 }
 210
 211 static int bulk_put_sink_callback(ptl_event_t *ev)
 212 {
 213         struct ptlrpc_bulk_desc *desc = ev->mem_desc.user_ptr;
 214         struct ptlrpc_bulk_page *bulk;
 215         struct list_head        *tmp;
 216         struct list_head        *next;
 217         ptl_size_t               total = 0;
 218         void                   (*event_handler)(struct ptlrpc_bulk_desc *);
 219         ENTRY;
 220
 221         LASSERT(ev->type == PTL_EVENT_PUT);
 222
 223         /* put with zero offset */
 224         LASSERT(ev->offset == 0);
 225         /* used iovs */
 226         LASSERT((ev->mem_desc.options & PTL_MD_IOV) != 0);
 227         /* 1 fragment for each page always */
 228         LASSERT(ev->mem_desc.niov == desc->bd_page_count);
 229
 230         list_for_each_safe (tmp, next, &desc->bd_page_list) {
 231                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
 232
 233                 total += bulk->bp_buflen;
 234
 235                 if (bulk->bp_cb != NULL)
 236                         bulk->bp_cb(bulk);
 237         }
 238
 239         LASSERT(ev->mem_desc.length == total);
 240
 241         /* We need to make a note of whether there's an event handler
 242          * before we call wake_up, because if there is no event
 243          * handler, 'desc' might be freed before we're scheduled again. */
 244         event_handler = desc->bd_ptl_ev_hdlr;
 245
 246         desc->bd_flags |= PTL_BULK_FL_RCVD;
 247         wake_up(&desc->bd_waitq);
 248         if (event_handler) {
 249                 LASSERT(desc->bd_ptl_ev_hdlr == event_handler);
 250                 event_handler(desc);
 251         }
 252
 253         RETURN(1);
 254 }
 255
 256 static int bulk_get_source_callback(ptl_event_t *ev)
 257 {
 258         struct ptlrpc_bulk_desc *desc = ev->mem_desc.user_ptr;
 259         struct ptlrpc_bulk_page *bulk;
 260         struct list_head        *tmp;
 261         struct list_head        *next;
 262         ptl_size_t               total = 0;
 263         void                   (*event_handler)(struct ptlrpc_bulk_desc *);
 264         ENTRY;
 265
 266         LASSERT(ev->type == PTL_EVENT_GET);
 267
 268         /* put with zero offset */
 269         LASSERT(ev->offset == 0);
 270         /* used iovs */
 271         LASSERT((ev->mem_desc.options & PTL_MD_IOV) != 0);
 272         /* 1 fragment for each page always */
 273         LASSERT(ev->mem_desc.niov == desc->bd_page_count);
 274
 275         list_for_each_safe (tmp, next, &desc->bd_page_list) {
 276                 bulk = list_entry(tmp, struct ptlrpc_bulk_page, bp_link);
 277
 278                 total += bulk->bp_buflen;
 279
 280                 if (bulk->bp_cb != NULL)
 281                         bulk->bp_cb(bulk);
 282         }
 283
 284         LASSERT(ev->mem_desc.length == total);
 285
 286         /* We need to make a note of whether there's an event handler
 287          * before we call wake_up, because if there is no event
 288          * handler, 'desc' might be freed before we're scheduled again. */
 289         event_handler = desc->bd_ptl_ev_hdlr;
 290
 291         desc->bd_flags |= PTL_BULK_FL_SENT;
 292         wake_up(&desc->bd_waitq);
 293         if (event_handler) {
 294                 LASSERT(desc->bd_ptl_ev_hdlr == event_handler);
 295                 event_handler(desc);
 296         }
 297
 298         RETURN(1);
 299 }
 300
 301
 302 static int bulk_get_sink_callback(ptl_event_t *ev)
 303 {
 304         struct ptlrpc_bulk_desc *desc = ev->mem_desc.user_ptr;
 305         struct ptlrpc_bulk_page *bulk;
 306         struct list_head        *tmp;
 307         struct list_head        *next;
 308         ENTRY;
 309
 310         CDEBUG(D_NET, "got %s event %d\n",
 311                (ev->type == PTL_EVENT_SENT) ? "SENT" :
 312                (ev->type == PTL_EVENT_REPLY)  ? "REPLY"  : "UNEXPECTED",
 313                ev->type);
 314
 315         LASSERT(ev->type == PTL_EVENT_SENT || ev->type == PTL_EVENT_REPLY);
 316
 317         LASSERT(atomic_read(&desc->bd_source_callback_count) > 0 &&
 318                 atomic_read(&desc->bd_source_callback_count) <= 2);
 319
 320         /* 1 fragment for each page always */
 321         LASSERT(ev->mem_desc.niov == desc->bd_page_count);
 322
 323         if (atomic_dec_and_test(&desc->bd_source_callback_count)) {
 324                 void (*event_handler)(struct ptlrpc_bulk_desc *);
 325
 326                 list_for_each_safe(tmp, next, &desc->bd_page_list) {
 327                         bulk = list_entry(tmp, struct ptlrpc_bulk_page,
 328                                           bp_link);
 329
 330                         if (bulk->bp_cb != NULL)
 331                                 bulk->bp_cb(bulk);
 332                 }
 333
 334                 /* We need to make a note of whether there's an event handler
 335                  * before we call wake_up, because if there is no event handler,
 336                  * 'desc' might be freed before we're scheduled again. */
 337                 event_handler = desc->bd_ptl_ev_hdlr;
 338
 339                 desc->bd_flags |= PTL_BULK_FL_RCVD;
 340                 wake_up(&desc->bd_waitq);
 341                 if (event_handler) {
 342                         LASSERT(desc->bd_ptl_ev_hdlr == event_handler);
 343                         event_handler(desc);
 344                 }
 345         }
 346
 347         RETURN(0);
 348 }
 349
 350 int ptlrpc_uuid_to_peer (struct obd_uuid *uuid, struct ptlrpc_peer *peer)
 351 {
 352         struct ptlrpc_ni   *pni;
 353         struct lustre_peer  lpeer;
 354         int                 i;
 355         int                 rc = lustre_uuid_to_peer (uuid->uuid, &lpeer);
 356
 357         if (rc != 0)
 358                 RETURN (rc);
 359
 360         for (i = 0; i < ptlrpc_ninterfaces; i++) {
 361                 pni = &ptlrpc_interfaces[i];
 362
 363                 if (!memcmp (&lpeer.peer_ni, &pni->pni_ni_h,
 364                              sizeof (lpeer.peer_ni))) {
 365                         peer->peer_nid = lpeer.peer_nid;
 366                         peer->peer_ni = pni;
 367                         return (0);
 368                 }
 369         }
 370
 371         CERROR ("Can't find ptlrpc interface for "LPX64" ni handle %08lx %08lx\n",
 372                 lpeer.peer_nid, lpeer.peer_ni.nal_idx, lpeer.peer_ni.handle_idx);
 373         return (-ENOENT);
 374 }
 375
 376 void ptlrpc_ni_fini (struct ptlrpc_ni *pni)
 377 {
 378         PtlEQFree(pni->pni_request_out_eq_h);
 379         PtlEQFree(pni->pni_reply_out_eq_h);
 380         PtlEQFree(pni->pni_reply_in_eq_h);
 381         PtlEQFree(pni->pni_bulk_put_source_eq_h);
 382         PtlEQFree(pni->pni_bulk_put_sink_eq_h);
 383         PtlEQFree(pni->pni_bulk_get_source_eq_h);
 384         PtlEQFree(pni->pni_bulk_get_sink_eq_h);
 385
 386         inter_module_put(pni->pni_name);
 387 }
 388
 389 int ptlrpc_ni_init (char *name, struct ptlrpc_ni *pni)
 390 {
 391         int              rc;
 392         ptl_handle_ni_t *nip;
 393
 394         nip = (ptl_handle_ni_t *)inter_module_get (name);
 395         if (nip == NULL) {
 396                 CDEBUG (D_NET, "Network interface %s not loaded\n", name);
 397                 return (-ENOENT);
 398         }
 399
 400         CDEBUG (D_NET, "init %s: nal_idx %ld\n", name, nip->nal_idx);
 401
 402         pni->pni_name = name;
 403         pni->pni_ni_h = *nip;
 404
 405         ptl_set_inv_handle (&pni->pni_request_out_eq_h);
 406         ptl_set_inv_handle (&pni->pni_reply_out_eq_h);
 407         ptl_set_inv_handle (&pni->pni_reply_in_eq_h);
 408         ptl_set_inv_handle (&pni->pni_bulk_put_source_eq_h);
 409         ptl_set_inv_handle (&pni->pni_bulk_put_sink_eq_h);
 410         ptl_set_inv_handle (&pni->pni_bulk_get_source_eq_h);
 411         ptl_set_inv_handle (&pni->pni_bulk_get_sink_eq_h);
 412
 413         /* NB We never actually PtlEQGet() out of these events queues since
 414          * we're only interested in the event callback, so we can just let
 415          * them wrap.  Their sizes aren't a big deal, apart from providing
 416          * a little history for debugging... */
 417
 418         rc = PtlEQAlloc(pni->pni_ni_h, 1024, request_out_callback,
 419                         &pni->pni_request_out_eq_h);
 420         if (rc != PTL_OK)
 421                 GOTO (fail, rc = -ENOMEM);
 422
 423         rc = PtlEQAlloc(pni->pni_ni_h, 1024, reply_out_callback,
 424                         &pni->pni_reply_out_eq_h);
 425         if (rc != PTL_OK)
 426                 GOTO (fail, rc = -ENOMEM);
 427
 428         rc = PtlEQAlloc(pni->pni_ni_h, 1024, reply_in_callback,
 429                         &pni->pni_reply_in_eq_h);
 430         if (rc != PTL_OK)
 431                 GOTO (fail, rc = -ENOMEM);
 432
 433         rc = PtlEQAlloc(pni->pni_ni_h, 1024, bulk_put_source_callback,
 434                         &pni->pni_bulk_put_source_eq_h);
 435         if (rc != PTL_OK)
 436                 GOTO (fail, rc = -ENOMEM);
 437
 438         rc = PtlEQAlloc(pni->pni_ni_h, 1024, bulk_put_sink_callback,
 439                         &pni->pni_bulk_put_sink_eq_h);
 440         if (rc != PTL_OK)
 441                 GOTO (fail, rc = -ENOMEM);
 442
 443         rc = PtlEQAlloc(pni->pni_ni_h, 1024, bulk_get_source_callback,
 444                         &pni->pni_bulk_get_source_eq_h);
 445         if (rc != PTL_OK)
 446                 GOTO (fail, rc = -ENOMEM);
 447
 448         rc = PtlEQAlloc(pni->pni_ni_h, 1024, bulk_get_sink_callback,
 449                         &pni->pni_bulk_get_sink_eq_h);
 450         if (rc != PTL_OK)
 451                 GOTO (fail, rc = -ENOMEM);
 452
 453         return (0);
 454  fail:
 455         CERROR ("Failed to initialise network interface %s: %d\n",
 456                 name, rc);
 457
 458         /* OK to do complete teardown since we invalidated the handles above... */
 459         ptlrpc_ni_fini (pni);
 460         return (rc);
 461 }
 462
 463 int ptlrpc_init_portals(void)
 464 {
 465         /* Add new portals network interface names here.
 466          * Order is irrelevent! */
 467         char *ni_names[] = { "kqswnal_ni",
 468                              "kgmnal_ni",
 469                              "ksocknal_ni",
 470                              "ktoenal_ni",
 471                              "tcpnal_ni",
 472                              NULL };
 473         int   rc;
 474         int   i;
 475
 476         LASSERT (ptlrpc_ninterfaces == 0);
 477
 478         for (i = 0; ni_names[i] != NULL; i++) {
 479                 LASSERT (ptlrpc_ninterfaces <
 480                          sizeof (ptlrpc_interfaces)/sizeof (ptlrpc_interfaces[0]));
 481
 482                 rc = ptlrpc_ni_init (ni_names[i],
 483                                      &ptlrpc_interfaces[ptlrpc_ninterfaces]);
 484                 if (rc == 0)
 485                         ptlrpc_ninterfaces++;
 486         }
 487
 488         if (ptlrpc_ninterfaces == 0) {
 489                 CERROR("network initialisation failed: is a NAL module loaded?\n");
 490                 return -EIO;
 491         }
 492         return 0;
 493 }
 494
 495 void ptlrpc_exit_portals(void)
 496 {
 497         while (ptlrpc_ninterfaces > 0)
 498                 ptlrpc_ni_fini (&ptlrpc_interfaces[--ptlrpc_ninterfaces]);
 499 }