Whamcloud - gitweb
land v0.9.1 on HEAD, in preparation for a 1.0.x branch
[fs/lustre-release.git] / lnet / klnds / gmlnd / gmlnd_comm.c
1 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
2  * vim:expandtab:shiftwidth=8:tabstop=8:
3  *
4  *  Copyright (c) 2003 Los Alamos National Laboratory (LANL)
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  *      This file contains all gmnal send and receive functions
24  */
25
26 #include "gmnal.h"
27
28 /*
29  *      The caretaker thread
30  *      This is main thread of execution for the NAL side
31  *      This guy waits in gm_blocking_recvive and gets
32  *      woken up when the myrinet adaptor gets an interrupt.
33  *      Hands off receive operations to the receive thread 
34  *      This thread Looks after gm_callbacks etc inline.
35  */
36 int
37 gmnal_ct_thread(void *arg)
38 {
39         gmnal_data_t            *nal_data;
40         gm_recv_event_t         *rxevent = NULL;
41         gm_recv_t               *recv = NULL;
42
43         if (!arg) {
44                 CDEBUG(D_TRACE, "NO nal_data. Exiting\n");
45                 return(-1);
46         }
47
48         nal_data = (gmnal_data_t*)arg;
49         CDEBUG(D_TRACE, "nal_data is [%p]\n", arg);
50
51         daemonize();
52
53         nal_data->ctthread_flag = GMNAL_CTTHREAD_STARTED;
54
55         GMNAL_GM_LOCK(nal_data);
56         while(nal_data->ctthread_flag == GMNAL_CTTHREAD_STARTED) {
57                 CDEBUG(D_NET, "waiting\n");
58                 rxevent = gm_blocking_receive_no_spin(nal_data->gm_port);
59                 if (nal_data->ctthread_flag == GMNAL_THREAD_STOP) {
60                         CDEBUG(D_INFO, "time to exit\n");
61                         break;
62                 }
63                 CDEBUG(D_INFO, "got [%s]\n", gmnal_rxevent(rxevent));
64                 switch (GM_RECV_EVENT_TYPE(rxevent)) {
65
66                         case(GM_RECV_EVENT):
67                                 CDEBUG(D_NET, "CTTHREAD:: GM_RECV_EVENT\n");
68                                 recv = (gm_recv_t*)&rxevent->recv;
69                                 GMNAL_GM_UNLOCK(nal_data);
70                                 gmnal_add_rxtwe(nal_data, recv);
71                                 GMNAL_GM_LOCK(nal_data);
72                                 CDEBUG(D_NET, "CTTHREAD:: Added event to Q\n");
73                         break;
74                         case(_GM_SLEEP_EVENT):
75                                 /*
76                                  *      Blocking receive above just returns
77                                  *      immediatly with _GM_SLEEP_EVENT
78                                  *      Don't know what this is
79                                  */
80                                 CDEBUG(D_NET, "Sleeping in gm_unknown\n");
81                                 GMNAL_GM_UNLOCK(nal_data);
82                                 gm_unknown(nal_data->gm_port, rxevent);
83                                 GMNAL_GM_LOCK(nal_data);
84                                 CDEBUG(D_INFO, "Awake from gm_unknown\n");
85                                 break;
86                                 
87                         default:
88                                 /*
89                                  *      Don't know what this is
90                                  *      gm_unknown will make sense of it
91                                  *      Should be able to do something with
92                                  *      FAST_RECV_EVENTS here.
93                                  */
94                                 CDEBUG(D_NET, "Passing event to gm_unknown\n");
95                                 GMNAL_GM_UNLOCK(nal_data);
96                                 gm_unknown(nal_data->gm_port, rxevent);
97                                 GMNAL_GM_LOCK(nal_data);
98                                 CDEBUG(D_INFO, "Processed unknown event\n");
99                 }
100         }
101         GMNAL_GM_UNLOCK(nal_data);
102         nal_data->ctthread_flag = GMNAL_THREAD_RESET;
103         CDEBUG(D_INFO, "thread nal_data [%p] is exiting\n", nal_data);
104         return(GMNAL_STATUS_OK);
105 }
106
107
108 /*
109  *      process a receive event
110  */
111 int gmnal_rx_thread(void *arg)
112 {
113         gmnal_data_t            *nal_data;
114         void                    *buffer;
115         gmnal_rxtwe_t           *we = NULL;
116
117         if (!arg) {
118                 CDEBUG(D_TRACE, "NO nal_data. Exiting\n");
119                 return(-1);
120         }
121
122         nal_data = (gmnal_data_t*)arg;
123         CDEBUG(D_TRACE, "nal_data is [%p]\n", arg);
124
125         daemonize();
126         /*
127          *      set 1 bit for each thread started
128          *      doesn't matter which bit
129          */
130         spin_lock(&nal_data->rxthread_flag_lock);
131         if (nal_data->rxthread_flag)
132                 nal_data->rxthread_flag=nal_data->rxthread_flag*2 + 1;
133         else
134                 nal_data->rxthread_flag = 1;
135         CDEBUG(D_INFO, "rxthread flag is [%ld]\n", nal_data->rxthread_flag);
136         spin_unlock(&nal_data->rxthread_flag_lock);
137
138         while(nal_data->rxthread_stop_flag != GMNAL_THREAD_STOP) {
139                 CDEBUG(D_NET, "RXTHREAD:: Receive thread waiting\n");
140                 we = gmnal_get_rxtwe(nal_data);
141                 if (!we) {
142                         CDEBUG(D_INFO, "Receive thread time to exit\n");
143                         break;
144                 }
145
146                 buffer = we->buffer;
147                 switch(((gmnal_msghdr_t*)buffer)->type) {
148                 case(GMNAL_SMALL_MESSAGE):
149                         gmnal_pre_receive(nal_data, we, 
150                                            GMNAL_SMALL_MESSAGE);
151                 break;  
152                 case(GMNAL_LARGE_MESSAGE_INIT):
153                         gmnal_pre_receive(nal_data, we, 
154                                            GMNAL_LARGE_MESSAGE_INIT);
155                 break;  
156                 case(GMNAL_LARGE_MESSAGE_ACK):
157                         gmnal_pre_receive(nal_data, we, 
158                                            GMNAL_LARGE_MESSAGE_ACK);
159                 break;  
160                 default:
161                         CDEBUG(D_ERROR, "Unsupported message type\n");
162                         gmnal_rx_bad(nal_data, we, NULL);
163                 }
164                 PORTAL_FREE(we, sizeof(gmnal_rxtwe_t));
165         }
166
167         spin_lock(&nal_data->rxthread_flag_lock);
168         nal_data->rxthread_flag/=2;
169         CDEBUG(D_INFO, "rxthread flag is [%ld]\n", nal_data->rxthread_flag);
170         spin_unlock(&nal_data->rxthread_flag_lock);
171         CDEBUG(D_INFO, "thread nal_data [%p] is exiting\n", nal_data);
172         return(GMNAL_STATUS_OK);
173 }
174
175
176
177 /*
178  *      Start processing a small message receive
179  *      Get here from gmnal_receive_thread
180  *      Hand off to lib_parse, which calls cb_recv
181  *      which hands back to gmnal_small_receive
182  *      Deal with all endian stuff here.
183  */
184 int
185 gmnal_pre_receive(gmnal_data_t *nal_data, gmnal_rxtwe_t *we, int gmnal_type)
186 {
187         gmnal_srxd_t    *srxd = NULL;
188         void            *buffer = NULL;
189         unsigned int snode, sport, type, length;
190         gmnal_msghdr_t  *gmnal_msghdr;
191         ptl_hdr_t       *portals_hdr;
192
193         CDEBUG(D_INFO, "nal_data [%p], we[%p] type [%d]\n", 
194                nal_data, we, gmnal_type);
195
196         buffer = we->buffer;
197         snode = we->snode;
198         sport = we->sport;
199         type = we->type;
200         buffer = we->buffer;
201         length = we->length;
202
203         gmnal_msghdr = (gmnal_msghdr_t*)buffer;
204         portals_hdr = (ptl_hdr_t*)(buffer+GMNAL_MSGHDR_SIZE);
205
206         CDEBUG(D_INFO, "rx_event:: Sender node [%d], Sender Port [%d], 
207                type [%d], length [%d], buffer [%p]\n",
208                snode, sport, type, length, buffer);
209         CDEBUG(D_INFO, "gmnal_msghdr:: Sender node [%u], magic [%d], 
210                gmnal_type [%d]\n", gmnal_msghdr->sender_node_id, 
211                gmnal_msghdr->magic, gmnal_msghdr->type);
212         CDEBUG(D_INFO, "portals_hdr:: Sender node ["LPD64"], 
213                dest_node ["LPD64"]\n", portals_hdr->src_nid, 
214                portals_hdr->dest_nid);
215
216         
217         /*
218          *      Get a receive descriptor for this message
219          */
220         srxd = gmnal_rxbuffer_to_srxd(nal_data, buffer);
221         CDEBUG(D_INFO, "Back from gmnal_rxbuffer_to_srxd\n");
222         srxd->nal_data = nal_data;
223         if (!srxd) {
224                 CDEBUG(D_ERROR, "Failed to get receive descriptor\n");
225                 lib_parse(nal_data->nal_cb, portals_hdr, srxd);
226                 return(GMNAL_STATUS_FAIL);
227         }
228
229         /*
230          *      no need to bother portals library with this
231          */
232         if (gmnal_type == GMNAL_LARGE_MESSAGE_ACK) {
233                 gmnal_large_tx_ack_received(nal_data, srxd);
234                 return(GMNAL_STATUS_OK);
235         }
236
237         srxd->type = gmnal_type;
238         srxd->nsiov = gmnal_msghdr->niov;
239         srxd->gm_source_node = gmnal_msghdr->sender_node_id;
240         
241         CDEBUG(D_PORTALS, "Calling lib_parse buffer is [%p]\n", 
242                buffer+GMNAL_MSGHDR_SIZE);
243         /*
244          *      control passes to lib, which calls cb_recv 
245          *      cb_recv is responsible for returning the buffer 
246          *      for future receive
247          */
248         lib_parse(nal_data->nal_cb, portals_hdr, srxd);
249
250         return(GMNAL_STATUS_OK);
251 }
252
253
254
255 /*
256  *      After a receive has been processed, 
257  *      hang out the receive buffer again.
258  *      This implicitly returns a receive token.
259  */
260 int
261 gmnal_rx_requeue_buffer(gmnal_data_t *nal_data, gmnal_srxd_t *srxd)
262 {
263         CDEBUG(D_TRACE, "gmnal_rx_requeue_buffer\n");
264
265         CDEBUG(D_NET, "requeueing srxd[%p] nal_data[%p]\n", srxd, nal_data);
266
267         GMNAL_GM_LOCK(nal_data);
268         gm_provide_receive_buffer_with_tag(nal_data->gm_port, srxd->buffer,
269                                         srxd->gmsize, GM_LOW_PRIORITY, 0 );
270         GMNAL_GM_UNLOCK(nal_data);
271
272         return(GMNAL_STATUS_OK);
273 }
274
275
276 /*
277  *      Handle a bad message
278  *      A bad message is one we don't expect or can't interpret
279  */
280 int
281 gmnal_rx_bad(gmnal_data_t *nal_data, gmnal_rxtwe_t *we, gmnal_srxd_t *srxd)
282 {
283         CDEBUG(D_TRACE, "Can't handle message\n");
284
285         if (!srxd)
286                 srxd = gmnal_rxbuffer_to_srxd(nal_data, 
287                                                we->buffer);
288         if (srxd) {
289                 gmnal_rx_requeue_buffer(nal_data, srxd);
290         } else {
291                 CDEBUG(D_ERROR, "Can't find a descriptor for this buffer\n");
292                 /*
293                  *      get rid of it ?
294                  */
295                 return(GMNAL_STATUS_FAIL);
296         }
297
298         return(GMNAL_STATUS_OK);
299 }
300
301
302
303 /*
304  *      Process a small message receive.
305  *      Get here from gmnal_receive_thread, gmnal_pre_receive
306  *      lib_parse, cb_recv
307  *      Put data from prewired receive buffer into users buffer(s)
308  *      Hang out the receive buffer again for another receive
309  *      Call lib_finalize
310  */
311 int
312 gmnal_small_rx(nal_cb_t *nal_cb, void *private, lib_msg_t *cookie, 
313                 unsigned int niov, struct iovec *iov, size_t mlen, size_t rlen)
314 {
315         gmnal_srxd_t    *srxd = NULL;
316         void    *buffer = NULL;
317         gmnal_data_t    *nal_data = (gmnal_data_t*)nal_cb->nal_data;
318
319
320         CDEBUG(D_TRACE, "niov [%d] mlen["LPSZ"]\n", niov, mlen);
321
322         if (!private) {
323                 CDEBUG(D_ERROR, "gmnal_small_rx no context\n");
324                 lib_finalize(nal_cb, private, cookie);
325                 return(PTL_FAIL);
326         }
327
328         srxd = (gmnal_srxd_t*)private;
329         buffer = srxd->buffer;
330         buffer += sizeof(gmnal_msghdr_t);
331         buffer += sizeof(ptl_hdr_t);
332
333         while(niov--) {
334                 CDEBUG(D_INFO, "processing [%p] len ["LPSZ"]\n", iov, 
335                        iov->iov_len);
336                 gm_bcopy(buffer, iov->iov_base, iov->iov_len);                  
337                 buffer += iov->iov_len;
338                 iov++;
339         }
340
341
342         /*
343          *      let portals library know receive is complete
344          */
345         CDEBUG(D_PORTALS, "calling lib_finalize\n");
346         if (lib_finalize(nal_cb, private, cookie) != PTL_OK) {
347                 /* TO DO what to do with failed lib_finalise? */
348                 CDEBUG(D_INFO, "lib_finalize failed\n");
349         }
350         /*
351          *      return buffer so it can be used again
352          */
353         CDEBUG(D_NET, "calling gm_provide_receive_buffer\n");
354         GMNAL_GM_LOCK(nal_data);
355         gm_provide_receive_buffer_with_tag(nal_data->gm_port, srxd->buffer, 
356                                            srxd->gmsize, GM_LOW_PRIORITY, 0);   
357         GMNAL_GM_UNLOCK(nal_data);
358
359         return(PTL_OK);
360 }
361
362
363 /*
364  *      Start a small transmit. 
365  *      Get a send token (and wired transmit buffer).
366  *      Copy data from senders buffer to wired buffer and
367  *      initiate gm_send from the wired buffer.
368  *      The callback function informs when the send is complete.
369  */
370 int
371 gmnal_small_tx(nal_cb_t *nal_cb, void *private, lib_msg_t *cookie, 
372                 ptl_hdr_t *hdr, int type, ptl_nid_t global_nid, ptl_pid_t pid, 
373                 unsigned int niov, struct iovec *iov, int size)
374 {
375         gmnal_data_t    *nal_data = (gmnal_data_t*)nal_cb->nal_data;
376         gmnal_stxd_t    *stxd = NULL;
377         void            *buffer = NULL;
378         gmnal_msghdr_t  *msghdr = NULL;
379         int             tot_size = 0;
380         unsigned int    local_nid;
381         gm_status_t     gm_status = GM_SUCCESS;
382
383         CDEBUG(D_TRACE, "gmnal_small_tx nal_cb [%p] private [%p] cookie [%p] 
384                hdr [%p] type [%d] global_nid ["LPU64"] pid [%d] niov [%d] 
385                iov [%p] size [%d]\n", nal_cb, private, cookie, hdr, type, 
386                global_nid, pid, niov, iov, size);
387
388         CDEBUG(D_INFO, "portals_hdr:: dest_nid ["LPU64"], src_nid ["LPU64"]\n",
389                hdr->dest_nid, hdr->src_nid);
390
391         if (!nal_data) {
392                 CDEBUG(D_ERROR, "no nal_data\n");
393                 return(GMNAL_STATUS_FAIL);
394         } else {
395                 CDEBUG(D_INFO, "nal_data [%p]\n", nal_data);
396         }
397
398         GMNAL_GM_LOCK(nal_data);
399         gm_status = gm_global_id_to_node_id(nal_data->gm_port, global_nid, 
400                                             &local_nid);
401         GMNAL_GM_UNLOCK(nal_data);
402         if (gm_status != GM_SUCCESS) {
403                 CDEBUG(D_ERROR, "Failed to obtain local id\n");
404                 return(GMNAL_STATUS_FAIL);
405         }
406         CDEBUG(D_INFO, "Local Node_id is [%u][%x]\n", local_nid, local_nid);
407
408         stxd = gmnal_get_stxd(nal_data, 1);
409         CDEBUG(D_INFO, "stxd [%p]\n", stxd);
410
411         stxd->type = GMNAL_SMALL_MESSAGE;
412         stxd->cookie = cookie;
413
414         /*
415          *      Copy gmnal_msg_hdr and portals header to the transmit buffer
416          *      Then copy the data in
417          */
418         buffer = stxd->buffer;
419         msghdr = (gmnal_msghdr_t*)buffer;
420
421         msghdr->magic = GMNAL_MAGIC;
422         msghdr->type = GMNAL_SMALL_MESSAGE;
423         msghdr->sender_node_id = nal_data->gm_global_nid;
424         CDEBUG(D_INFO, "processing msghdr at [%p]\n", buffer);
425
426         buffer += sizeof(gmnal_msghdr_t);
427
428         CDEBUG(D_INFO, "processing  portals hdr at [%p]\n", buffer);
429         gm_bcopy(hdr, buffer, sizeof(ptl_hdr_t));
430
431         buffer += sizeof(ptl_hdr_t);
432
433         while(niov--) {
434                 CDEBUG(D_INFO, "processing iov [%p] len ["LPSZ"] to [%p]\n", 
435                        iov, iov->iov_len, buffer);
436                 gm_bcopy(iov->iov_base, buffer, iov->iov_len);
437                 buffer+= iov->iov_len;
438                 iov++;
439         }
440
441         CDEBUG(D_INFO, "sending\n");
442         tot_size = size+sizeof(ptl_hdr_t)+sizeof(gmnal_msghdr_t);
443         stxd->msg_size = tot_size;
444
445
446         CDEBUG(D_NET, "Calling gm_send_to_peer port [%p] buffer [%p] 
447                gmsize [%lu] msize [%d] global_nid ["LPU64"] local_nid[%d] 
448                stxd [%p]\n", nal_data->gm_port, stxd->buffer, stxd->gm_size, 
449                stxd->msg_size, global_nid, local_nid, stxd);
450
451         GMNAL_GM_LOCK(nal_data);
452         stxd->gm_priority = GM_LOW_PRIORITY;
453         stxd->gm_target_node = local_nid;
454         gm_send_to_peer_with_callback(nal_data->gm_port, stxd->buffer, 
455                                       stxd->gm_size, stxd->msg_size, 
456                                       GM_LOW_PRIORITY, local_nid, 
457                                       gmnal_small_tx_callback, (void*)stxd);
458         GMNAL_GM_UNLOCK(nal_data);
459         CDEBUG(D_INFO, "done\n");
460                 
461         return(PTL_OK);
462 }
463
464
465 /*
466  *      A callback to indicate the small transmit operation is compete
467  *      Check for erros and try to deal with them.
468  *      Call lib_finalise to inform the client application that the send 
469  *      is complete and the memory can be reused.
470  *      Return the stxd when finished with it (returns a send token)
471  */
472 void 
473 gmnal_small_tx_callback(gm_port_t *gm_port, void *context, gm_status_t status)
474 {
475         gmnal_stxd_t    *stxd = (gmnal_stxd_t*)context;
476         lib_msg_t       *cookie = stxd->cookie;
477         gmnal_data_t    *nal_data = (gmnal_data_t*)stxd->nal_data;
478         nal_cb_t        *nal_cb = nal_data->nal_cb;
479
480         if (!stxd) {
481                 CDEBUG(D_TRACE, "send completion event for unknown stxd\n");
482                 return;
483         }
484         if (status != GM_SUCCESS) {
485                 CDEBUG(D_ERROR, "Result of send stxd [%p] is [%s]\n", 
486                        stxd, gmnal_gm_error(status));
487         }
488
489         switch(status) {
490                 case(GM_SUCCESS):
491                 break;
492
493
494
495                 case(GM_SEND_DROPPED):
496                 /*
497                  *      do a resend on the dropped ones
498                  */
499                         CDEBUG(D_ERROR, "send stxd [%p] was dropped 
500                                resending\n", context);
501                         GMNAL_GM_LOCK(nal_data);
502                         gm_send_to_peer_with_callback(nal_data->gm_port, 
503                                                       stxd->buffer, 
504                                                       stxd->gm_size, 
505                                                       stxd->msg_size, 
506                                                       stxd->gm_priority, 
507                                                       stxd->gm_target_node, 
508                                                       gmnal_small_tx_callback,
509                                                       context);
510                         GMNAL_GM_UNLOCK(nal_data);
511                 
512                 return;
513                 case(GM_TIMED_OUT):
514                 case(GM_SEND_TIMED_OUT):
515                 /*
516                  *      drop these ones
517                  */
518                         CDEBUG(D_INFO, "calling gm_drop_sends\n");
519                         GMNAL_GM_LOCK(nal_data);
520                         gm_drop_sends(nal_data->gm_port, stxd->gm_priority, 
521                                       stxd->gm_target_node, GMNAL_GM_PORT, 
522                                       gmnal_drop_sends_callback, context);
523                         GMNAL_GM_UNLOCK(nal_data);
524
525                 return;
526
527
528                 /*
529                  *      abort on these ?
530                  */
531                 case(GM_TRY_AGAIN):
532                 case(GM_INTERRUPTED):
533                 case(GM_FAILURE):
534                 case(GM_INPUT_BUFFER_TOO_SMALL):
535                 case(GM_OUTPUT_BUFFER_TOO_SMALL):
536                 case(GM_BUSY):
537                 case(GM_MEMORY_FAULT):
538                 case(GM_INVALID_PARAMETER):
539                 case(GM_OUT_OF_MEMORY):
540                 case(GM_INVALID_COMMAND):
541                 case(GM_PERMISSION_DENIED):
542                 case(GM_INTERNAL_ERROR):
543                 case(GM_UNATTACHED):
544                 case(GM_UNSUPPORTED_DEVICE):
545                 case(GM_SEND_REJECTED):
546                 case(GM_SEND_TARGET_PORT_CLOSED):
547                 case(GM_SEND_TARGET_NODE_UNREACHABLE):
548                 case(GM_SEND_PORT_CLOSED):
549                 case(GM_NODE_ID_NOT_YET_SET):
550                 case(GM_STILL_SHUTTING_DOWN):
551                 case(GM_CLONE_BUSY):
552                 case(GM_NO_SUCH_DEVICE):
553                 case(GM_ABORTED):
554                 case(GM_INCOMPATIBLE_LIB_AND_DRIVER):
555                 case(GM_UNTRANSLATED_SYSTEM_ERROR):
556                 case(GM_ACCESS_DENIED):
557                 case(GM_NO_DRIVER_SUPPORT):
558                 case(GM_PTE_REF_CNT_OVERFLOW):
559                 case(GM_NOT_SUPPORTED_IN_KERNEL):
560                 case(GM_NOT_SUPPORTED_ON_ARCH):
561                 case(GM_NO_MATCH):
562                 case(GM_USER_ERROR):
563                 case(GM_DATA_CORRUPTED):
564                 case(GM_HARDWARE_FAULT):
565                 case(GM_SEND_ORPHANED):
566                 case(GM_MINOR_OVERFLOW):
567                 case(GM_PAGE_TABLE_FULL):
568                 case(GM_UC_ERROR):
569                 case(GM_INVALID_PORT_NUMBER):
570                 case(GM_DEV_NOT_FOUND):
571                 case(GM_FIRMWARE_NOT_RUNNING):
572                 case(GM_YP_NO_MATCH):
573                 default:
574                         CDEBUG(D_ERROR, "Unknown send error\n");
575         }
576
577         /*
578          *      TO DO
579          *      If this is a large message init,
580          *      we're not finished with the data yet,
581          *      so can't call lib_finalise.
582          *      However, we're also holding on to a 
583          *      stxd here (to keep track of the source
584          *      iovec only). Should use another structure
585          *      to keep track of iovec and return stxd to 
586          *      free list earlier.
587          */
588         if (stxd->type == GMNAL_LARGE_MESSAGE_INIT) {
589                 CDEBUG(D_INFO, "large transmit done\n");
590                 return;
591         }
592         gmnal_return_stxd(nal_data, stxd);
593         if (lib_finalize(nal_cb, stxd, cookie) != PTL_OK) {
594                 CDEBUG(D_INFO, "Call to lib_finalize failed for stxd [%p]\n", 
595                        stxd);
596         }
597         return;
598 }
599
600
601
602 void gmnal_drop_sends_callback(struct gm_port *gm_port, void *context, 
603                                 gm_status_t status)
604 {
605         gmnal_stxd_t    *stxd = (gmnal_stxd_t*)context;
606         gmnal_data_t    *nal_data = stxd->nal_data;
607
608         CDEBUG(D_TRACE, "status is [%d] context is [%p]\n", status, context);
609         if (status == GM_SUCCESS) {
610                 GMNAL_GM_LOCK(nal_data);
611                 gm_send_to_peer_with_callback(gm_port, stxd->buffer, 
612                                               stxd->gm_size, stxd->msg_size, 
613                                               stxd->gm_priority, 
614                                               stxd->gm_target_node, 
615                                               gmnal_small_tx_callback, 
616                                               context);
617                 GMNAL_GM_LOCK(nal_data);
618         } else {
619                 CDEBUG(D_ERROR, "send_to_peer status for stxd [%p] is 
620                        [%d][%s]\n", stxd, status, gmnal_gm_error(status));
621         }
622
623
624         return;
625 }
626
627
628 /*
629  *      Begine a large transmit.
630  *      Do a gm_register of the memory pointed to by the iovec 
631  *      and send details to the receiver. The receiver does a gm_get
632  *      to pull the data and sends and ack when finished. Upon receipt of
633  *      this ack, deregister the memory. Only 1 send token is required here.
634  */
635 int
636 gmnal_large_tx(nal_cb_t *nal_cb, void *private, lib_msg_t *cookie, 
637                 ptl_hdr_t *hdr, int type, ptl_nid_t global_nid, ptl_pid_t pid, 
638                 unsigned int niov, struct iovec *iov, int size)
639 {
640
641         gmnal_data_t    *nal_data;
642         gmnal_stxd_t    *stxd = NULL;
643         void            *buffer = NULL;
644         gmnal_msghdr_t  *msghdr = NULL;
645         unsigned int    local_nid;
646         int             mlen = 0;       /* the size of the init message data */
647         struct iovec    *iov_dup = NULL;
648         gm_status_t     gm_status;
649         int             niov_dup;
650
651
652         CDEBUG(D_TRACE, "gmnal_large_tx nal_cb [%p] private [%p], cookie [%p] 
653                hdr [%p], type [%d] global_nid ["LPU64"], pid [%d], niov [%d], 
654                iov [%p], size [%d]\n", nal_cb, private, cookie, hdr, type, 
655                global_nid, pid, niov, iov, size);
656
657         if (nal_cb)
658                 nal_data = (gmnal_data_t*)nal_cb->nal_data;
659         else  {
660                 CDEBUG(D_ERROR, "no nal_cb.\n");
661                 return(GMNAL_STATUS_FAIL);
662         }
663         
664
665         /*
666          *      Get stxd and buffer. Put local address of data in buffer, 
667          *      send local addresses to target, 
668          *      wait for the target node to suck the data over.
669          *      The stxd is used to ren
670          */
671         stxd = gmnal_get_stxd(nal_data, 1);
672         CDEBUG(D_INFO, "stxd [%p]\n", stxd);
673
674         stxd->type = GMNAL_LARGE_MESSAGE_INIT;
675         stxd->cookie = cookie;
676
677         /*
678          *      Copy gmnal_msg_hdr and portals header to the transmit buffer
679          *      Then copy the iov in
680          */
681         buffer = stxd->buffer;
682         msghdr = (gmnal_msghdr_t*)buffer;
683
684         CDEBUG(D_INFO, "processing msghdr at [%p]\n", buffer);
685
686         msghdr->magic = GMNAL_MAGIC;
687         msghdr->type = GMNAL_LARGE_MESSAGE_INIT;
688         msghdr->sender_node_id = nal_data->gm_global_nid;
689         msghdr->stxd = stxd;
690         msghdr->niov = niov ;
691         buffer += sizeof(gmnal_msghdr_t);
692         mlen = sizeof(gmnal_msghdr_t);
693         CDEBUG(D_INFO, "mlen is [%d]\n", mlen);
694
695
696         CDEBUG(D_INFO, "processing  portals hdr at [%p]\n", buffer);
697
698         gm_bcopy(hdr, buffer, sizeof(ptl_hdr_t));
699         buffer += sizeof(ptl_hdr_t);
700         mlen += sizeof(ptl_hdr_t); 
701         CDEBUG(D_INFO, "mlen is [%d]\n", mlen);
702
703         /*
704          *      copy the iov to the buffer so target knows 
705          *      where to get the data from
706          */
707         CDEBUG(D_INFO, "processing iov to [%p]\n", buffer);
708         gm_bcopy(iov, buffer, niov*sizeof(struct iovec));
709         mlen += niov*(sizeof(struct iovec));
710         CDEBUG(D_INFO, "mlen is [%d]\n", mlen);
711
712
713         /*
714          *      Store the iovs in the stxd for we can get 
715          *      them later if we need them
716          */
717         CDEBUG(D_NET, "Copying iov [%p] to [%p]\n", iov, stxd->iov);
718         gm_bcopy(iov, stxd->iov, niov*sizeof(struct iovec));
719         stxd->niov = niov;
720         
721
722         /*
723          *      register the memory so the NIC can get hold of the data
724          *      This is a slow process. it'd be good to overlap it 
725          *      with something else.
726          */
727         iov_dup = iov;
728         niov_dup = niov;
729         while(niov--) {
730                 CDEBUG(D_INFO, "Registering memory [%p] len ["LPSZ"] \n", 
731                        iov->iov_base, iov->iov_len);
732                 GMNAL_GM_LOCK(nal_data);
733                 gm_status = gm_register_memory(nal_data->gm_port, 
734                                                iov->iov_base, iov->iov_len);
735                 if (gm_status != GM_SUCCESS) {
736                         GMNAL_GM_UNLOCK(nal_data);
737                         CDEBUG(D_ERROR, "gm_register_memory returns [%d][%s] 
738                                for memory [%p] len ["LPSZ"]\n", 
739                                gm_status, gmnal_gm_error(gm_status), 
740                                iov->iov_base, iov->iov_len);
741                         GMNAL_GM_LOCK(nal_data);
742                         while (iov_dup != iov) {
743                                 gm_deregister_memory(nal_data->gm_port, 
744                                                      iov_dup->iov_base, 
745                                                      iov_dup->iov_len);
746                                 iov_dup++;
747                         }
748                         GMNAL_GM_UNLOCK(nal_data);
749                         gmnal_return_stxd(nal_data, stxd);
750                         return(PTL_FAIL);
751                 }
752
753                 GMNAL_GM_UNLOCK(nal_data);
754                 iov++;
755         }
756
757         /*
758          *      Send the init message to the target
759          */
760         CDEBUG(D_INFO, "sending mlen [%d]\n", mlen);
761         GMNAL_GM_LOCK(nal_data);
762         gm_status = gm_global_id_to_node_id(nal_data->gm_port, global_nid, 
763                                             &local_nid);
764         if (gm_status != GM_SUCCESS) {
765                 GMNAL_GM_UNLOCK(nal_data);
766                 CDEBUG(D_ERROR, "Failed to obtain local id\n");
767                 gmnal_return_stxd(nal_data, stxd);
768                 /* TO DO deregister memory on failure */
769                 return(GMNAL_STATUS_FAIL);
770         }
771         CDEBUG(D_INFO, "Local Node_id is [%d]\n", local_nid);
772         gm_send_to_peer_with_callback(nal_data->gm_port, stxd->buffer, 
773                                       stxd->gm_size, mlen, GM_LOW_PRIORITY, 
774                                       local_nid, gmnal_large_tx_callback, 
775                                       (void*)stxd);
776         GMNAL_GM_UNLOCK(nal_data);
777         
778         CDEBUG(D_INFO, "done\n");
779                 
780         return(PTL_OK);
781 }
782
783 /*
784  *      Callback function indicates that send of buffer with 
785  *      large message iovec has completed (or failed).
786  */
787 void 
788 gmnal_large_tx_callback(gm_port_t *gm_port, void *context, gm_status_t status)
789 {
790         gmnal_small_tx_callback(gm_port, context, status);
791
792 }
793
794
795
796 /*
797  *      Have received a buffer that contains an iovec of the sender. 
798  *      Do a gm_register_memory of the receivers buffer and then do a get
799  *      data from the sender.
800  */
801 int
802 gmnal_large_rx(nal_cb_t *nal_cb, void *private, lib_msg_t *cookie, 
803                 unsigned int nriov, struct iovec *riov, size_t mlen, 
804                 size_t rlen)
805 {
806         gmnal_data_t    *nal_data = nal_cb->nal_data;
807         gmnal_srxd_t    *srxd = (gmnal_srxd_t*)private;
808         void            *buffer = NULL;
809         struct  iovec   *riov_dup;
810         int             nriov_dup;
811         gmnal_msghdr_t  *msghdr = NULL;
812         gm_status_t     gm_status;
813
814         CDEBUG(D_TRACE, "gmnal_large_rx :: nal_cb[%p], private[%p], 
815                cookie[%p], niov[%d], iov[%p], mlen["LPSZ"], rlen["LPSZ"]\n",
816                 nal_cb, private, cookie, nriov, riov, mlen, rlen);
817
818         if (!srxd) {
819                 CDEBUG(D_ERROR, "gmnal_large_rx no context\n");
820                 lib_finalize(nal_cb, private, cookie);
821                 return(PTL_FAIL);
822         }
823
824         buffer = srxd->buffer;
825         msghdr = (gmnal_msghdr_t*)buffer;
826         buffer += sizeof(gmnal_msghdr_t);
827         buffer += sizeof(ptl_hdr_t);
828
829         /*
830          *      Store the senders stxd address in the srxd for this message
831          *      The gmnal_large_message_ack needs it to notify the sender
832          *      the pull of data is complete
833          */
834         srxd->source_stxd = msghdr->stxd;
835
836         /*
837          *      Register the receivers memory
838          *      get the data,
839          *      tell the sender that we got the data
840          *      then tell the receiver we got the data
841          *      TO DO
842          *      If the iovecs match, could interleave 
843          *      gm_registers and gm_gets for each element
844          */
845         nriov_dup = nriov;
846         riov_dup = riov;
847         while(nriov--) {
848                 CDEBUG(D_INFO, "Registering memory [%p] len ["LPSZ"] \n", 
849                        riov->iov_base, riov->iov_len);
850                 GMNAL_GM_LOCK(nal_data);
851                 gm_status = gm_register_memory(nal_data->gm_port, 
852                                                riov->iov_base, riov->iov_len);
853                 if (gm_status != GM_SUCCESS) {
854                         GMNAL_GM_UNLOCK(nal_data);
855                         CDEBUG(D_ERROR, "gm_register_memory returns [%d][%s] 
856                                for memory [%p] len ["LPSZ"]\n", 
857                                gm_status, gmnal_gm_error(gm_status), 
858                                riov->iov_base, riov->iov_len);
859                         GMNAL_GM_LOCK(nal_data);
860                         while (riov_dup != riov) {
861                                 gm_deregister_memory(nal_data->gm_port, 
862                                                      riov_dup->iov_base, 
863                                                      riov_dup->iov_len);
864                                 riov_dup++;
865                         }
866                         GMNAL_GM_LOCK(nal_data);
867                         /*
868                          *      give back srxd and buffer. Send NACK to sender
869                          */
870                         return(PTL_FAIL);
871                 }
872                 GMNAL_GM_UNLOCK(nal_data);
873                 riov++;
874         }
875         /*
876          *      do this so the final gm_get callback can deregister the memory
877          */
878         PORTAL_ALLOC(srxd->riov, nriov_dup*(sizeof(struct iovec)));
879         gm_bcopy(riov_dup, srxd->riov, nriov_dup*(sizeof(struct iovec)));
880         srxd->nriov = nriov_dup;
881
882         /*
883          *      now do gm_get to get the data
884          */
885         srxd->cookie = cookie;
886         if (gmnal_remote_get(srxd, srxd->nsiov, (struct iovec*)buffer, 
887                               nriov_dup, riov_dup) != GMNAL_STATUS_OK) {
888                 CDEBUG(D_ERROR, "can't get the data");
889         }
890
891         CDEBUG(D_INFO, "lgmanl_large_rx done\n");
892
893         return(PTL_OK);
894 }
895
896
897 /*
898  *      Perform a number of remote gets as part of receiving 
899  *      a large message.
900  *      The final one to complete (i.e. the last callback to get called)
901  *      tidies up.
902  *      gm_get requires a send token.
903  */
904 int
905 gmnal_remote_get(gmnal_srxd_t *srxd, int nsiov, struct iovec *siov, 
906                   int nriov, struct iovec *riov)
907 {
908
909         int     ncalls = 0;
910
911         CDEBUG(D_TRACE, "gmnal_remote_get srxd[%p], nriov[%d], riov[%p], 
912                nsiov[%d], siov[%p]\n", srxd, nriov, riov, nsiov, siov);
913
914
915         ncalls = gmnal_copyiov(0, srxd, nsiov, siov, nriov, riov);
916         if (ncalls < 0) {
917                 CDEBUG(D_ERROR, "there's something wrong with the iovecs\n");
918                 return(GMNAL_STATUS_FAIL);
919         }
920         CDEBUG(D_INFO, "gmnal_remote_get ncalls [%d]\n", ncalls);
921         spin_lock_init(&srxd->callback_lock);
922         srxd->ncallbacks = ncalls;
923         srxd->callback_status = 0;
924
925         ncalls = gmnal_copyiov(1, srxd, nsiov, siov, nriov, riov);
926         if (ncalls < 0) {
927                 CDEBUG(D_ERROR, "there's something wrong with the iovecs\n");
928                 return(GMNAL_STATUS_FAIL);
929         }
930
931         return(GMNAL_STATUS_OK);
932
933 }
934
935
936 /*
937  *      pull data from source node (source iovec) to a local iovec.
938  *      The iovecs may not match which adds the complications below.
939  *      Count the number of gm_gets that will be required to the callbacks
940  *      can determine who is the last one.
941  */     
942 int
943 gmnal_copyiov(int do_copy, gmnal_srxd_t *srxd, int nsiov, 
944                struct iovec *siov, int nriov, struct iovec *riov)
945 {
946
947         int     ncalls = 0;
948         int     slen = siov->iov_len, rlen = riov->iov_len;
949         char    *sbuf = siov->iov_base, *rbuf = riov->iov_base; 
950         unsigned long   sbuf_long;
951         gm_remote_ptr_t remote_ptr = 0;
952         unsigned int    source_node;
953         gmnal_ltxd_t    *ltxd = NULL;
954         gmnal_data_t    *nal_data = srxd->nal_data;
955
956         CDEBUG(D_TRACE, "copy[%d] nal_data[%p]\n", do_copy, nal_data);
957         if (do_copy) {
958                 if (!nal_data) {
959                         CDEBUG(D_ERROR, "Bad args No nal_data\n");
960                         return(GMNAL_STATUS_FAIL);
961                 }
962                 GMNAL_GM_LOCK(nal_data);
963                 if (gm_global_id_to_node_id(nal_data->gm_port, 
964                                             srxd->gm_source_node, 
965                                             &source_node) != GM_SUCCESS) {
966
967                         CDEBUG(D_ERROR, "cannot resolve global_id [%u] 
968                                to local node_id\n", srxd->gm_source_node);
969                         GMNAL_GM_UNLOCK(nal_data);
970                         return(GMNAL_STATUS_FAIL);
971                 }
972                 GMNAL_GM_UNLOCK(nal_data);
973                 /*
974                  *      We need a send token to use gm_get
975                  *      getting an stxd gets us a send token.
976                  *      the stxd is used as the context to the
977                  *      callback function (so stxd can be returned).
978                  *      Set pointer in stxd to srxd so callback count in srxd
979                  *      can be decremented to find last callback to complete
980                  */
981                 CDEBUG(D_INFO, "gmnal_copyiov source node is G[%u]L[%d]\n", 
982                        srxd->gm_source_node, source_node);
983         }
984
985         do {
986                 CDEBUG(D_INFO, "sbuf[%p] slen[%d] rbuf[%p], rlen[%d]\n",
987                                 sbuf, slen, rbuf, rlen);
988                 if (slen > rlen) {
989                         ncalls++;
990                         if (do_copy) {
991                                 CDEBUG(D_INFO, "slen>rlen\n");
992                                 ltxd = gmnal_get_ltxd(nal_data);
993                                 ltxd->srxd = srxd;
994                                 GMNAL_GM_LOCK(nal_data);
995                                 /* 
996                                  *      funny business to get rid 
997                                  *      of compiler warning 
998                                  */
999                                 sbuf_long = (unsigned long) sbuf;
1000                                 remote_ptr = (gm_remote_ptr_t)sbuf_long;
1001                                 gm_get(nal_data->gm_port, remote_ptr, rbuf, 
1002                                        rlen, GM_LOW_PRIORITY, source_node, 
1003                                        GMNAL_GM_PORT, 
1004                                        gmnal_remote_get_callback, ltxd);
1005                                 GMNAL_GM_UNLOCK(nal_data);
1006                         }
1007                         /*
1008                          *      at the end of 1 iov element
1009                          */
1010                         sbuf+=rlen;
1011                         slen-=rlen;
1012                         riov++;
1013                         nriov--;
1014                         rbuf = riov->iov_base;
1015                         rlen = riov->iov_len;
1016                 } else if (rlen > slen) {
1017                         ncalls++;
1018                         if (do_copy) {
1019                                 CDEBUG(D_INFO, "slen<rlen\n");
1020                                 ltxd = gmnal_get_ltxd(nal_data);
1021                                 ltxd->srxd = srxd;
1022                                 GMNAL_GM_LOCK(nal_data);
1023                                 sbuf_long = (unsigned long) sbuf;
1024                                 remote_ptr = (gm_remote_ptr_t)sbuf_long;
1025                                 gm_get(nal_data->gm_port, remote_ptr, rbuf, 
1026                                        slen, GM_LOW_PRIORITY, source_node, 
1027                                        GMNAL_GM_PORT, 
1028                                        gmnal_remote_get_callback, ltxd);
1029                                 GMNAL_GM_UNLOCK(nal_data);
1030                         }
1031                         /*
1032                          *      at end of siov element
1033                          */
1034                         rbuf+=slen;
1035                         rlen-=slen;
1036                         siov++;
1037                         sbuf = siov->iov_base;
1038                         slen = siov->iov_len;
1039                 } else {
1040                         ncalls++;
1041                         if (do_copy) {
1042                                 CDEBUG(D_INFO, "rlen=slen\n");
1043                                 ltxd = gmnal_get_ltxd(nal_data);
1044                                 ltxd->srxd = srxd;
1045                                 GMNAL_GM_LOCK(nal_data);
1046                                 sbuf_long = (unsigned long) sbuf;
1047                                 remote_ptr = (gm_remote_ptr_t)sbuf_long;
1048                                 gm_get(nal_data->gm_port, remote_ptr, rbuf, 
1049                                        rlen, GM_LOW_PRIORITY, source_node, 
1050                                        GMNAL_GM_PORT, 
1051                                        gmnal_remote_get_callback, ltxd);
1052                                 GMNAL_GM_UNLOCK(nal_data);
1053                         }
1054                         /*
1055                          *      at end of siov and riov element
1056                          */
1057                         siov++;
1058                         sbuf = siov->iov_base;
1059                         slen = siov->iov_len;
1060                         riov++;
1061                         nriov--;
1062                         rbuf = riov->iov_base;
1063                         rlen = riov->iov_len;
1064                 }
1065
1066         } while (nriov);
1067         return(ncalls);
1068 }
1069
1070
1071 /*
1072  *      The callback function that is invoked after each gm_get call completes.
1073  *      Multiple callbacks may be invoked for 1 transaction, only the final
1074  *      callback has work to do.
1075  */
1076 void
1077 gmnal_remote_get_callback(gm_port_t *gm_port, void *context, 
1078                            gm_status_t status)
1079 {
1080
1081         gmnal_ltxd_t    *ltxd = (gmnal_ltxd_t*)context;
1082         gmnal_srxd_t    *srxd = ltxd->srxd;
1083         nal_cb_t        *nal_cb = srxd->nal_data->nal_cb;
1084         int             lastone;
1085         struct  iovec   *riov;
1086         int             nriov;
1087         gmnal_data_t    *nal_data;
1088
1089         CDEBUG(D_TRACE, "called for context [%p]\n", context);
1090
1091         if (status != GM_SUCCESS) {
1092                 CDEBUG(D_ERROR, "reports error [%d][%s]\n", status, 
1093                        gmnal_gm_error(status));
1094         }
1095
1096         spin_lock(&srxd->callback_lock);
1097         srxd->ncallbacks--;
1098         srxd->callback_status |= status;
1099         lastone = srxd->ncallbacks?0:1;
1100         spin_unlock(&srxd->callback_lock);
1101         nal_data = srxd->nal_data;
1102
1103         /*
1104          *      everyone returns a send token
1105          */
1106         gmnal_return_ltxd(nal_data, ltxd);
1107
1108         if (!lastone) {
1109                 CDEBUG(D_ERROR, "NOT final callback context[%p]\n", srxd);
1110                 return;
1111         }
1112         
1113         /*
1114          *      Let our client application proceed
1115          */     
1116         CDEBUG(D_ERROR, "final callback context[%p]\n", srxd);
1117         if (lib_finalize(nal_cb, srxd, srxd->cookie) != PTL_OK) {
1118                 CDEBUG(D_INFO, "Call to lib_finalize failed for srxd [%p]\n", 
1119                        srxd);
1120         }
1121
1122         /*
1123          *      send an ack to the sender to let him know we got the data
1124          */
1125         gmnal_large_tx_ack(nal_data, srxd);
1126
1127         /*
1128          *      Unregister the memory that was used
1129          *      This is a very slow business (slower then register)
1130          */
1131         nriov = srxd->nriov;
1132         riov = srxd->riov;
1133         GMNAL_GM_LOCK(nal_data);
1134         while (nriov--) {
1135                 CDEBUG(D_ERROR, "deregister memory [%p]\n", riov->iov_base);
1136                 if (gm_deregister_memory(srxd->nal_data->gm_port, 
1137                                          riov->iov_base, riov->iov_len)) {
1138                         CDEBUG(D_ERROR, "failed to deregister memory [%p]\n", 
1139                                riov->iov_base);
1140                 }
1141                 riov++;
1142         }
1143         GMNAL_GM_UNLOCK(nal_data);
1144         PORTAL_FREE(srxd->riov, sizeof(struct iovec)*nriov);
1145
1146         /*
1147          *      repost the receive buffer (return receive token)
1148          */
1149         GMNAL_GM_LOCK(nal_data);
1150         gm_provide_receive_buffer_with_tag(nal_data->gm_port, srxd->buffer, 
1151                                            srxd->gmsize, GM_LOW_PRIORITY, 0);   
1152         GMNAL_GM_UNLOCK(nal_data);
1153         
1154         return;
1155 }
1156
1157
1158 /*
1159  *      Called on target node.
1160  *      After pulling data from a source node
1161  *      send an ack message to indicate the large transmit is complete.
1162  */
1163 void 
1164 gmnal_large_tx_ack(gmnal_data_t *nal_data, gmnal_srxd_t *srxd)
1165 {
1166
1167         gmnal_stxd_t    *stxd;
1168         gmnal_msghdr_t *msghdr;
1169         void            *buffer = NULL;
1170         unsigned int    local_nid;
1171         gm_status_t     gm_status = GM_SUCCESS;
1172
1173         CDEBUG(D_TRACE, "srxd[%p] target_node [%u]\n", srxd, 
1174                srxd->gm_source_node);
1175
1176         GMNAL_GM_LOCK(nal_data);
1177         gm_status = gm_global_id_to_node_id(nal_data->gm_port, 
1178                                             srxd->gm_source_node, &local_nid);
1179         GMNAL_GM_UNLOCK(nal_data);
1180         if (gm_status != GM_SUCCESS) {
1181                 CDEBUG(D_ERROR, "Failed to obtain local id\n");
1182                 return;
1183         }
1184         CDEBUG(D_INFO, "Local Node_id is [%u][%x]\n", local_nid, local_nid);
1185
1186         stxd = gmnal_get_stxd(nal_data, 1);
1187         CDEBUG(D_TRACE, "gmnal_large_tx_ack got stxd[%p]\n", stxd);
1188
1189         stxd->nal_data = nal_data;
1190         stxd->type = GMNAL_LARGE_MESSAGE_ACK;
1191
1192         /*
1193          *      Copy gmnal_msg_hdr and portals header to the transmit buffer
1194          *      Then copy the data in
1195          */
1196         buffer = stxd->buffer;
1197         msghdr = (gmnal_msghdr_t*)buffer;
1198
1199         /*
1200          *      Add in the address of the original stxd from the sender node
1201          *      so it knows which thread to notify.
1202          */
1203         msghdr->magic = GMNAL_MAGIC;
1204         msghdr->type = GMNAL_LARGE_MESSAGE_ACK;
1205         msghdr->sender_node_id = nal_data->gm_global_nid;
1206         msghdr->stxd = srxd->source_stxd;
1207         CDEBUG(D_INFO, "processing msghdr at [%p]\n", buffer);
1208
1209         CDEBUG(D_INFO, "sending\n");
1210         stxd->msg_size= sizeof(gmnal_msghdr_t);
1211
1212
1213         CDEBUG(D_NET, "Calling gm_send_to_peer port [%p] buffer [%p] 
1214                gmsize [%lu] msize [%d] global_nid [%u] local_nid[%d] 
1215                stxd [%p]\n", nal_data->gm_port, stxd->buffer, stxd->gm_size, 
1216                stxd->msg_size, srxd->gm_source_node, local_nid, stxd);
1217         GMNAL_GM_LOCK(nal_data);
1218         stxd->gm_priority = GM_LOW_PRIORITY;
1219         stxd->gm_target_node = local_nid;
1220         gm_send_to_peer_with_callback(nal_data->gm_port, stxd->buffer, 
1221                                       stxd->gm_size, stxd->msg_size, 
1222                                       GM_LOW_PRIORITY, local_nid, 
1223                                       gmnal_large_tx_ack_callback, 
1224                                       (void*)stxd);
1225         
1226         GMNAL_GM_UNLOCK(nal_data);
1227         CDEBUG(D_INFO, "gmnal_large_tx_ack :: done\n");
1228                 
1229         return;
1230 }
1231
1232
1233 /*
1234  *      A callback to indicate the small transmit operation is compete
1235  *      Check for errors and try to deal with them.
1236  *      Call lib_finalise to inform the client application that the 
1237  *      send is complete and the memory can be reused.
1238  *      Return the stxd when finished with it (returns a send token)
1239  */
1240 void 
1241 gmnal_large_tx_ack_callback(gm_port_t *gm_port, void *context, 
1242                              gm_status_t status)
1243 {
1244         gmnal_stxd_t    *stxd = (gmnal_stxd_t*)context;
1245         gmnal_data_t    *nal_data = (gmnal_data_t*)stxd->nal_data;
1246
1247         if (!stxd) {
1248                 CDEBUG(D_ERROR, "send completion event for unknown stxd\n");
1249                 return;
1250         }
1251         CDEBUG(D_TRACE, "send completion event for stxd [%p] status is [%d]\n",
1252                stxd, status);
1253         gmnal_return_stxd(stxd->nal_data, stxd);
1254
1255         GMNAL_GM_UNLOCK(nal_data);
1256         return;
1257 }
1258
1259 /*
1260  *      Indicates the large transmit operation is compete.
1261  *      Called on transmit side (means data has been pulled  by receiver 
1262  *      or failed).
1263  *      Call lib_finalise to inform the client application that the send 
1264  *      is complete, deregister the memory and return the stxd. 
1265  *      Finally, report the rx buffer that the ack message was delivered in.
1266  */
1267 void 
1268 gmnal_large_tx_ack_received(gmnal_data_t *nal_data, gmnal_srxd_t *srxd)
1269 {
1270         nal_cb_t        *nal_cb = nal_data->nal_cb;
1271         gmnal_stxd_t    *stxd = NULL;
1272         gmnal_msghdr_t  *msghdr = NULL;
1273         void            *buffer = NULL;
1274         struct  iovec   *iov;
1275
1276
1277         CDEBUG(D_TRACE, "gmnal_large_tx_ack_received buffer [%p]\n", buffer);
1278
1279         buffer = srxd->buffer;
1280         msghdr = (gmnal_msghdr_t*)buffer;
1281         stxd = msghdr->stxd;
1282
1283         CDEBUG(D_INFO, "gmnal_large_tx_ack_received stxd [%p]\n", stxd);
1284
1285         if (lib_finalize(nal_cb, stxd, stxd->cookie) != PTL_OK) {
1286                 CDEBUG(D_INFO, "Call to lib_finalize failed for stxd [%p]\n", 
1287                        stxd);
1288         }
1289
1290         /*
1291          *      extract the iovec from the stxd, deregister the memory.
1292          *      free the space used to store the iovec
1293          */
1294         iov = stxd->iov;
1295         while(stxd->niov--) {
1296                 CDEBUG(D_INFO, "deregister memory [%p] size ["LPSZ"]\n",
1297                        iov->iov_base, iov->iov_len);
1298                 GMNAL_GM_LOCK(nal_data);
1299                 gm_deregister_memory(nal_data->gm_port, iov->iov_base, 
1300                                      iov->iov_len);
1301                 GMNAL_GM_UNLOCK(nal_data);
1302                 iov++;
1303         }
1304
1305         /*
1306          *      return the send token
1307          *      TO DO It is bad to hold onto the send token so long?
1308          */
1309         gmnal_return_stxd(nal_data, stxd);
1310
1311
1312         /*
1313          *      requeue the receive buffer 
1314          */
1315         gmnal_rx_requeue_buffer(nal_data, srxd);
1316         
1317
1318         return;
1319 }