net: unset IFF_XMIT_DST_RELEASE in ipgre_tunnel_setup()
[linux-2.6] / net / rds / ib_send.c
1 /*
2  * Copyright (c) 2006 Oracle.  All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  *
32  */
33 #include <linux/kernel.h>
34 #include <linux/in.h>
35 #include <linux/device.h>
36 #include <linux/dmapool.h>
37
38 #include "rds.h"
39 #include "rdma.h"
40 #include "ib.h"
41
42 static void rds_ib_send_rdma_complete(struct rds_message *rm,
43                                       int wc_status)
44 {
45         int notify_status;
46
47         switch (wc_status) {
48         case IB_WC_WR_FLUSH_ERR:
49                 return;
50
51         case IB_WC_SUCCESS:
52                 notify_status = RDS_RDMA_SUCCESS;
53                 break;
54
55         case IB_WC_REM_ACCESS_ERR:
56                 notify_status = RDS_RDMA_REMOTE_ERROR;
57                 break;
58
59         default:
60                 notify_status = RDS_RDMA_OTHER_ERROR;
61                 break;
62         }
63         rds_rdma_send_complete(rm, notify_status);
64 }
65
66 static void rds_ib_send_unmap_rdma(struct rds_ib_connection *ic,
67                                    struct rds_rdma_op *op)
68 {
69         if (op->r_mapped) {
70                 ib_dma_unmap_sg(ic->i_cm_id->device,
71                         op->r_sg, op->r_nents,
72                         op->r_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
73                 op->r_mapped = 0;
74         }
75 }
76
77 static void rds_ib_send_unmap_rm(struct rds_ib_connection *ic,
78                           struct rds_ib_send_work *send,
79                           int wc_status)
80 {
81         struct rds_message *rm = send->s_rm;
82
83         rdsdebug("ic %p send %p rm %p\n", ic, send, rm);
84
85         ib_dma_unmap_sg(ic->i_cm_id->device,
86                      rm->m_sg, rm->m_nents,
87                      DMA_TO_DEVICE);
88
89         if (rm->m_rdma_op != NULL) {
90                 rds_ib_send_unmap_rdma(ic, rm->m_rdma_op);
91
92                 /* If the user asked for a completion notification on this
93                  * message, we can implement three different semantics:
94                  *  1.  Notify when we received the ACK on the RDS message
95                  *      that was queued with the RDMA. This provides reliable
96                  *      notification of RDMA status at the expense of a one-way
97                  *      packet delay.
98                  *  2.  Notify when the IB stack gives us the completion event for
99                  *      the RDMA operation.
100                  *  3.  Notify when the IB stack gives us the completion event for
101                  *      the accompanying RDS messages.
102                  * Here, we implement approach #3. To implement approach #2,
103                  * call rds_rdma_send_complete from the cq_handler. To implement #1,
104                  * don't call rds_rdma_send_complete at all, and fall back to the notify
105                  * handling in the ACK processing code.
106                  *
107                  * Note: There's no need to explicitly sync any RDMA buffers using
108                  * ib_dma_sync_sg_for_cpu - the completion for the RDMA
109                  * operation itself unmapped the RDMA buffers, which takes care
110                  * of synching.
111                  */
112                 rds_ib_send_rdma_complete(rm, wc_status);
113
114                 if (rm->m_rdma_op->r_write)
115                         rds_stats_add(s_send_rdma_bytes, rm->m_rdma_op->r_bytes);
116                 else
117                         rds_stats_add(s_recv_rdma_bytes, rm->m_rdma_op->r_bytes);
118         }
119
120         /* If anyone waited for this message to get flushed out, wake
121          * them up now */
122         rds_message_unmapped(rm);
123
124         rds_message_put(rm);
125         send->s_rm = NULL;
126 }
127
128 void rds_ib_send_init_ring(struct rds_ib_connection *ic)
129 {
130         struct rds_ib_send_work *send;
131         u32 i;
132
133         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
134                 struct ib_sge *sge;
135
136                 send->s_rm = NULL;
137                 send->s_op = NULL;
138
139                 send->s_wr.wr_id = i;
140                 send->s_wr.sg_list = send->s_sge;
141                 send->s_wr.num_sge = 1;
142                 send->s_wr.opcode = IB_WR_SEND;
143                 send->s_wr.send_flags = 0;
144                 send->s_wr.ex.imm_data = 0;
145
146                 sge = rds_ib_data_sge(ic, send->s_sge);
147                 sge->lkey = ic->i_mr->lkey;
148
149                 sge = rds_ib_header_sge(ic, send->s_sge);
150                 sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header));
151                 sge->length = sizeof(struct rds_header);
152                 sge->lkey = ic->i_mr->lkey;
153         }
154 }
155
156 void rds_ib_send_clear_ring(struct rds_ib_connection *ic)
157 {
158         struct rds_ib_send_work *send;
159         u32 i;
160
161         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
162                 if (send->s_wr.opcode == 0xdead)
163                         continue;
164                 if (send->s_rm)
165                         rds_ib_send_unmap_rm(ic, send, IB_WC_WR_FLUSH_ERR);
166                 if (send->s_op)
167                         rds_ib_send_unmap_rdma(ic, send->s_op);
168         }
169 }
170
171 /*
172  * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
173  * operations performed in the send path.  As the sender allocs and potentially
174  * unallocs the next free entry in the ring it doesn't alter which is
175  * the next to be freed, which is what this is concerned with.
176  */
177 void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
178 {
179         struct rds_connection *conn = context;
180         struct rds_ib_connection *ic = conn->c_transport_data;
181         struct ib_wc wc;
182         struct rds_ib_send_work *send;
183         u32 completed;
184         u32 oldest;
185         u32 i = 0;
186         int ret;
187
188         rdsdebug("cq %p conn %p\n", cq, conn);
189         rds_ib_stats_inc(s_ib_tx_cq_call);
190         ret = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
191         if (ret)
192                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
193
194         while (ib_poll_cq(cq, 1, &wc) > 0) {
195                 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
196                          (unsigned long long)wc.wr_id, wc.status, wc.byte_len,
197                          be32_to_cpu(wc.ex.imm_data));
198                 rds_ib_stats_inc(s_ib_tx_cq_event);
199
200                 if (wc.wr_id == RDS_IB_ACK_WR_ID) {
201                         if (ic->i_ack_queued + HZ/2 < jiffies)
202                                 rds_ib_stats_inc(s_ib_tx_stalled);
203                         rds_ib_ack_send_complete(ic);
204                         continue;
205                 }
206
207                 oldest = rds_ib_ring_oldest(&ic->i_send_ring);
208
209                 completed = rds_ib_ring_completed(&ic->i_send_ring, wc.wr_id, oldest);
210
211                 for (i = 0; i < completed; i++) {
212                         send = &ic->i_sends[oldest];
213
214                         /* In the error case, wc.opcode sometimes contains garbage */
215                         switch (send->s_wr.opcode) {
216                         case IB_WR_SEND:
217                                 if (send->s_rm)
218                                         rds_ib_send_unmap_rm(ic, send, wc.status);
219                                 break;
220                         case IB_WR_RDMA_WRITE:
221                         case IB_WR_RDMA_READ:
222                                 /* Nothing to be done - the SG list will be unmapped
223                                  * when the SEND completes. */
224                                 break;
225                         default:
226                                 if (printk_ratelimit())
227                                         printk(KERN_NOTICE
228                                                 "RDS/IB: %s: unexpected opcode 0x%x in WR!\n",
229                                                 __func__, send->s_wr.opcode);
230                                 break;
231                         }
232
233                         send->s_wr.opcode = 0xdead;
234                         send->s_wr.num_sge = 1;
235                         if (send->s_queued + HZ/2 < jiffies)
236                                 rds_ib_stats_inc(s_ib_tx_stalled);
237
238                         /* If a RDMA operation produced an error, signal this right
239                          * away. If we don't, the subsequent SEND that goes with this
240                          * RDMA will be canceled with ERR_WFLUSH, and the application
241                          * never learn that the RDMA failed. */
242                         if (unlikely(wc.status == IB_WC_REM_ACCESS_ERR && send->s_op)) {
243                                 struct rds_message *rm;
244
245                                 rm = rds_send_get_message(conn, send->s_op);
246                                 if (rm)
247                                         rds_ib_send_rdma_complete(rm, wc.status);
248                         }
249
250                         oldest = (oldest + 1) % ic->i_send_ring.w_nr;
251                 }
252
253                 rds_ib_ring_free(&ic->i_send_ring, completed);
254
255                 if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags)
256                  || test_bit(0, &conn->c_map_queued))
257                         queue_delayed_work(rds_wq, &conn->c_send_w, 0);
258
259                 /* We expect errors as the qp is drained during shutdown */
260                 if (wc.status != IB_WC_SUCCESS && rds_conn_up(conn)) {
261                         rds_ib_conn_error(conn,
262                                 "send completion on %pI4 "
263                                 "had status %u, disconnecting and reconnecting\n",
264                                 &conn->c_faddr, wc.status);
265                 }
266         }
267 }
268
269 /*
270  * This is the main function for allocating credits when sending
271  * messages.
272  *
273  * Conceptually, we have two counters:
274  *  -   send credits: this tells us how many WRs we're allowed
275  *      to submit without overruning the reciever's queue. For
276  *      each SEND WR we post, we decrement this by one.
277  *
278  *  -   posted credits: this tells us how many WRs we recently
279  *      posted to the receive queue. This value is transferred
280  *      to the peer as a "credit update" in a RDS header field.
281  *      Every time we transmit credits to the peer, we subtract
282  *      the amount of transferred credits from this counter.
283  *
284  * It is essential that we avoid situations where both sides have
285  * exhausted their send credits, and are unable to send new credits
286  * to the peer. We achieve this by requiring that we send at least
287  * one credit update to the peer before exhausting our credits.
288  * When new credits arrive, we subtract one credit that is withheld
289  * until we've posted new buffers and are ready to transmit these
290  * credits (see rds_ib_send_add_credits below).
291  *
292  * The RDS send code is essentially single-threaded; rds_send_xmit
293  * grabs c_send_lock to ensure exclusive access to the send ring.
294  * However, the ACK sending code is independent and can race with
295  * message SENDs.
296  *
297  * In the send path, we need to update the counters for send credits
298  * and the counter of posted buffers atomically - when we use the
299  * last available credit, we cannot allow another thread to race us
300  * and grab the posted credits counter.  Hence, we have to use a
301  * spinlock to protect the credit counter, or use atomics.
302  *
303  * Spinlocks shared between the send and the receive path are bad,
304  * because they create unnecessary delays. An early implementation
305  * using a spinlock showed a 5% degradation in throughput at some
306  * loads.
307  *
308  * This implementation avoids spinlocks completely, putting both
309  * counters into a single atomic, and updating that atomic using
310  * atomic_add (in the receive path, when receiving fresh credits),
311  * and using atomic_cmpxchg when updating the two counters.
312  */
313 int rds_ib_send_grab_credits(struct rds_ib_connection *ic,
314                              u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
315 {
316         unsigned int avail, posted, got = 0, advertise;
317         long oldval, newval;
318
319         *adv_credits = 0;
320         if (!ic->i_flowctl)
321                 return wanted;
322
323 try_again:
324         advertise = 0;
325         oldval = newval = atomic_read(&ic->i_credits);
326         posted = IB_GET_POST_CREDITS(oldval);
327         avail = IB_GET_SEND_CREDITS(oldval);
328
329         rdsdebug("rds_ib_send_grab_credits(%u): credits=%u posted=%u\n",
330                         wanted, avail, posted);
331
332         /* The last credit must be used to send a credit update. */
333         if (avail && !posted)
334                 avail--;
335
336         if (avail < wanted) {
337                 struct rds_connection *conn = ic->i_cm_id->context;
338
339                 /* Oops, there aren't that many credits left! */
340                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
341                 got = avail;
342         } else {
343                 /* Sometimes you get what you want, lalala. */
344                 got = wanted;
345         }
346         newval -= IB_SET_SEND_CREDITS(got);
347
348         /*
349          * If need_posted is non-zero, then the caller wants
350          * the posted regardless of whether any send credits are
351          * available.
352          */
353         if (posted && (got || need_posted)) {
354                 advertise = min_t(unsigned int, posted, max_posted);
355                 newval -= IB_SET_POST_CREDITS(advertise);
356         }
357
358         /* Finally bill everything */
359         if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
360                 goto try_again;
361
362         *adv_credits = advertise;
363         return got;
364 }
365
366 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits)
367 {
368         struct rds_ib_connection *ic = conn->c_transport_data;
369
370         if (credits == 0)
371                 return;
372
373         rdsdebug("rds_ib_send_add_credits(%u): current=%u%s\n",
374                         credits,
375                         IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
376                         test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
377
378         atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
379         if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
380                 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
381
382         WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
383
384         rds_ib_stats_inc(s_ib_rx_credit_updates);
385 }
386
387 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted)
388 {
389         struct rds_ib_connection *ic = conn->c_transport_data;
390
391         if (posted == 0)
392                 return;
393
394         atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
395
396         /* Decide whether to send an update to the peer now.
397          * If we would send a credit update for every single buffer we
398          * post, we would end up with an ACK storm (ACK arrives,
399          * consumes buffer, we refill the ring, send ACK to remote
400          * advertising the newly posted buffer... ad inf)
401          *
402          * Performance pretty much depends on how often we send
403          * credit updates - too frequent updates mean lots of ACKs.
404          * Too infrequent updates, and the peer will run out of
405          * credits and has to throttle.
406          * For the time being, 16 seems to be a good compromise.
407          */
408         if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
409                 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
410 }
411
412 static inline void
413 rds_ib_xmit_populate_wr(struct rds_ib_connection *ic,
414                 struct rds_ib_send_work *send, unsigned int pos,
415                 unsigned long buffer, unsigned int length,
416                 int send_flags)
417 {
418         struct ib_sge *sge;
419
420         WARN_ON(pos != send - ic->i_sends);
421
422         send->s_wr.send_flags = send_flags;
423         send->s_wr.opcode = IB_WR_SEND;
424         send->s_wr.num_sge = 2;
425         send->s_wr.next = NULL;
426         send->s_queued = jiffies;
427         send->s_op = NULL;
428
429         if (length != 0) {
430                 sge = rds_ib_data_sge(ic, send->s_sge);
431                 sge->addr = buffer;
432                 sge->length = length;
433                 sge->lkey = ic->i_mr->lkey;
434
435                 sge = rds_ib_header_sge(ic, send->s_sge);
436         } else {
437                 /* We're sending a packet with no payload. There is only
438                  * one SGE */
439                 send->s_wr.num_sge = 1;
440                 sge = &send->s_sge[0];
441         }
442
443         sge->addr = ic->i_send_hdrs_dma + (pos * sizeof(struct rds_header));
444         sge->length = sizeof(struct rds_header);
445         sge->lkey = ic->i_mr->lkey;
446 }
447
448 /*
449  * This can be called multiple times for a given message.  The first time
450  * we see a message we map its scatterlist into the IB device so that
451  * we can provide that mapped address to the IB scatter gather entries
452  * in the IB work requests.  We translate the scatterlist into a series
453  * of work requests that fragment the message.  These work requests complete
454  * in order so we pass ownership of the message to the completion handler
455  * once we send the final fragment.
456  *
457  * The RDS core uses the c_send_lock to only enter this function once
458  * per connection.  This makes sure that the tx ring alloc/unalloc pairs
459  * don't get out of sync and confuse the ring.
460  */
461 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
462                 unsigned int hdr_off, unsigned int sg, unsigned int off)
463 {
464         struct rds_ib_connection *ic = conn->c_transport_data;
465         struct ib_device *dev = ic->i_cm_id->device;
466         struct rds_ib_send_work *send = NULL;
467         struct rds_ib_send_work *first;
468         struct rds_ib_send_work *prev;
469         struct ib_send_wr *failed_wr;
470         struct scatterlist *scat;
471         u32 pos;
472         u32 i;
473         u32 work_alloc;
474         u32 credit_alloc;
475         u32 posted;
476         u32 adv_credits = 0;
477         int send_flags = 0;
478         int sent;
479         int ret;
480         int flow_controlled = 0;
481
482         BUG_ON(off % RDS_FRAG_SIZE);
483         BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
484
485         /* FIXME we may overallocate here */
486         if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
487                 i = 1;
488         else
489                 i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
490
491         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
492         if (work_alloc == 0) {
493                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
494                 rds_ib_stats_inc(s_ib_tx_ring_full);
495                 ret = -ENOMEM;
496                 goto out;
497         }
498
499         credit_alloc = work_alloc;
500         if (ic->i_flowctl) {
501                 credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
502                 adv_credits += posted;
503                 if (credit_alloc < work_alloc) {
504                         rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
505                         work_alloc = credit_alloc;
506                         flow_controlled++;
507                 }
508                 if (work_alloc == 0) {
509                         set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
510                         rds_ib_stats_inc(s_ib_tx_throttle);
511                         ret = -ENOMEM;
512                         goto out;
513                 }
514         }
515
516         /* map the message the first time we see it */
517         if (ic->i_rm == NULL) {
518                 /*
519                 printk(KERN_NOTICE "rds_ib_xmit prep msg dport=%u flags=0x%x len=%d\n",
520                                 be16_to_cpu(rm->m_inc.i_hdr.h_dport),
521                                 rm->m_inc.i_hdr.h_flags,
522                                 be32_to_cpu(rm->m_inc.i_hdr.h_len));
523                    */
524                 if (rm->m_nents) {
525                         rm->m_count = ib_dma_map_sg(dev,
526                                          rm->m_sg, rm->m_nents, DMA_TO_DEVICE);
527                         rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->m_count);
528                         if (rm->m_count == 0) {
529                                 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
530                                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
531                                 ret = -ENOMEM; /* XXX ? */
532                                 goto out;
533                         }
534                 } else {
535                         rm->m_count = 0;
536                 }
537
538                 ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
539                 ic->i_unsignaled_bytes = rds_ib_sysctl_max_unsig_bytes;
540                 rds_message_addref(rm);
541                 ic->i_rm = rm;
542
543                 /* Finalize the header */
544                 if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
545                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
546                 if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
547                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
548
549                 /* If it has a RDMA op, tell the peer we did it. This is
550                  * used by the peer to release use-once RDMA MRs. */
551                 if (rm->m_rdma_op) {
552                         struct rds_ext_header_rdma ext_hdr;
553
554                         ext_hdr.h_rdma_rkey = cpu_to_be32(rm->m_rdma_op->r_key);
555                         rds_message_add_extension(&rm->m_inc.i_hdr,
556                                         RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
557                 }
558                 if (rm->m_rdma_cookie) {
559                         rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
560                                         rds_rdma_cookie_key(rm->m_rdma_cookie),
561                                         rds_rdma_cookie_offset(rm->m_rdma_cookie));
562                 }
563
564                 /* Note - rds_ib_piggyb_ack clears the ACK_REQUIRED bit, so
565                  * we should not do this unless we have a chance of at least
566                  * sticking the header into the send ring. Which is why we
567                  * should call rds_ib_ring_alloc first. */
568                 rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_ib_piggyb_ack(ic));
569                 rds_message_make_checksum(&rm->m_inc.i_hdr);
570
571                 /*
572                  * Update adv_credits since we reset the ACK_REQUIRED bit.
573                  */
574                 rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
575                 adv_credits += posted;
576                 BUG_ON(adv_credits > 255);
577         } else if (ic->i_rm != rm)
578                 BUG();
579
580         send = &ic->i_sends[pos];
581         first = send;
582         prev = NULL;
583         scat = &rm->m_sg[sg];
584         sent = 0;
585         i = 0;
586
587         /* Sometimes you want to put a fence between an RDMA
588          * READ and the following SEND.
589          * We could either do this all the time
590          * or when requested by the user. Right now, we let
591          * the application choose.
592          */
593         if (rm->m_rdma_op && rm->m_rdma_op->r_fence)
594                 send_flags = IB_SEND_FENCE;
595
596         /*
597          * We could be copying the header into the unused tail of the page.
598          * That would need to be changed in the future when those pages might
599          * be mapped userspace pages or page cache pages.  So instead we always
600          * use a second sge and our long-lived ring of mapped headers.  We send
601          * the header after the data so that the data payload can be aligned on
602          * the receiver.
603          */
604
605         /* handle a 0-len message */
606         if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0) {
607                 rds_ib_xmit_populate_wr(ic, send, pos, 0, 0, send_flags);
608                 goto add_header;
609         }
610
611         /* if there's data reference it with a chain of work reqs */
612         for (; i < work_alloc && scat != &rm->m_sg[rm->m_count]; i++) {
613                 unsigned int len;
614
615                 send = &ic->i_sends[pos];
616
617                 len = min(RDS_FRAG_SIZE, ib_sg_dma_len(dev, scat) - off);
618                 rds_ib_xmit_populate_wr(ic, send, pos,
619                                 ib_sg_dma_address(dev, scat) + off, len,
620                                 send_flags);
621
622                 /*
623                  * We want to delay signaling completions just enough to get
624                  * the batching benefits but not so much that we create dead time
625                  * on the wire.
626                  */
627                 if (ic->i_unsignaled_wrs-- == 0) {
628                         ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
629                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
630                 }
631
632                 ic->i_unsignaled_bytes -= len;
633                 if (ic->i_unsignaled_bytes <= 0) {
634                         ic->i_unsignaled_bytes = rds_ib_sysctl_max_unsig_bytes;
635                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
636                 }
637
638                 /*
639                  * Always signal the last one if we're stopping due to flow control.
640                  */
641                 if (flow_controlled && i == (work_alloc-1))
642                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
643
644                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
645                          &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
646
647                 sent += len;
648                 off += len;
649                 if (off == ib_sg_dma_len(dev, scat)) {
650                         scat++;
651                         off = 0;
652                 }
653
654 add_header:
655                 /* Tack on the header after the data. The header SGE should already
656                  * have been set up to point to the right header buffer. */
657                 memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
658
659                 if (0) {
660                         struct rds_header *hdr = &ic->i_send_hdrs[pos];
661
662                         printk(KERN_NOTICE "send WR dport=%u flags=0x%x len=%d\n",
663                                 be16_to_cpu(hdr->h_dport),
664                                 hdr->h_flags,
665                                 be32_to_cpu(hdr->h_len));
666                 }
667                 if (adv_credits) {
668                         struct rds_header *hdr = &ic->i_send_hdrs[pos];
669
670                         /* add credit and redo the header checksum */
671                         hdr->h_credit = adv_credits;
672                         rds_message_make_checksum(hdr);
673                         adv_credits = 0;
674                         rds_ib_stats_inc(s_ib_tx_credit_updates);
675                 }
676
677                 if (prev)
678                         prev->s_wr.next = &send->s_wr;
679                 prev = send;
680
681                 pos = (pos + 1) % ic->i_send_ring.w_nr;
682         }
683
684         /* Account the RDS header in the number of bytes we sent, but just once.
685          * The caller has no concept of fragmentation. */
686         if (hdr_off == 0)
687                 sent += sizeof(struct rds_header);
688
689         /* if we finished the message then send completion owns it */
690         if (scat == &rm->m_sg[rm->m_count]) {
691                 prev->s_rm = ic->i_rm;
692                 prev->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
693                 ic->i_rm = NULL;
694         }
695
696         if (i < work_alloc) {
697                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
698                 work_alloc = i;
699         }
700         if (ic->i_flowctl && i < credit_alloc)
701                 rds_ib_send_add_credits(conn, credit_alloc - i);
702
703         /* XXX need to worry about failed_wr and partial sends. */
704         failed_wr = &first->s_wr;
705         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
706         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
707                  first, &first->s_wr, ret, failed_wr);
708         BUG_ON(failed_wr != &first->s_wr);
709         if (ret) {
710                 printk(KERN_WARNING "RDS/IB: ib_post_send to %pI4 "
711                        "returned %d\n", &conn->c_faddr, ret);
712                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
713                 if (prev->s_rm) {
714                         ic->i_rm = prev->s_rm;
715                         prev->s_rm = NULL;
716                 }
717                 /* Finesse this later */
718                 BUG();
719                 goto out;
720         }
721
722         ret = sent;
723 out:
724         BUG_ON(adv_credits);
725         return ret;
726 }
727
728 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
729 {
730         struct rds_ib_connection *ic = conn->c_transport_data;
731         struct rds_ib_send_work *send = NULL;
732         struct rds_ib_send_work *first;
733         struct rds_ib_send_work *prev;
734         struct ib_send_wr *failed_wr;
735         struct rds_ib_device *rds_ibdev;
736         struct scatterlist *scat;
737         unsigned long len;
738         u64 remote_addr = op->r_remote_addr;
739         u32 pos;
740         u32 work_alloc;
741         u32 i;
742         u32 j;
743         int sent;
744         int ret;
745         int num_sge;
746
747         rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
748
749         /* map the message the first time we see it */
750         if (!op->r_mapped) {
751                 op->r_count = ib_dma_map_sg(ic->i_cm_id->device,
752                                         op->r_sg, op->r_nents, (op->r_write) ?
753                                         DMA_TO_DEVICE : DMA_FROM_DEVICE);
754                 rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->r_count);
755                 if (op->r_count == 0) {
756                         rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
757                         ret = -ENOMEM; /* XXX ? */
758                         goto out;
759                 }
760
761                 op->r_mapped = 1;
762         }
763
764         /*
765          * Instead of knowing how to return a partial rdma read/write we insist that there
766          * be enough work requests to send the entire message.
767          */
768         i = ceil(op->r_count, rds_ibdev->max_sge);
769
770         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
771         if (work_alloc != i) {
772                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
773                 rds_ib_stats_inc(s_ib_tx_ring_full);
774                 ret = -ENOMEM;
775                 goto out;
776         }
777
778         send = &ic->i_sends[pos];
779         first = send;
780         prev = NULL;
781         scat = &op->r_sg[0];
782         sent = 0;
783         num_sge = op->r_count;
784
785         for (i = 0; i < work_alloc && scat != &op->r_sg[op->r_count]; i++) {
786                 send->s_wr.send_flags = 0;
787                 send->s_queued = jiffies;
788                 /*
789                  * We want to delay signaling completions just enough to get
790                  * the batching benefits but not so much that we create dead time on the wire.
791                  */
792                 if (ic->i_unsignaled_wrs-- == 0) {
793                         ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
794                         send->s_wr.send_flags = IB_SEND_SIGNALED;
795                 }
796
797                 send->s_wr.opcode = op->r_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ;
798                 send->s_wr.wr.rdma.remote_addr = remote_addr;
799                 send->s_wr.wr.rdma.rkey = op->r_key;
800                 send->s_op = op;
801
802                 if (num_sge > rds_ibdev->max_sge) {
803                         send->s_wr.num_sge = rds_ibdev->max_sge;
804                         num_sge -= rds_ibdev->max_sge;
805                 } else {
806                         send->s_wr.num_sge = num_sge;
807                 }
808
809                 send->s_wr.next = NULL;
810
811                 if (prev)
812                         prev->s_wr.next = &send->s_wr;
813
814                 for (j = 0; j < send->s_wr.num_sge && scat != &op->r_sg[op->r_count]; j++) {
815                         len = ib_sg_dma_len(ic->i_cm_id->device, scat);
816                         send->s_sge[j].addr =
817                                  ib_sg_dma_address(ic->i_cm_id->device, scat);
818                         send->s_sge[j].length = len;
819                         send->s_sge[j].lkey = ic->i_mr->lkey;
820
821                         sent += len;
822                         rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
823
824                         remote_addr += len;
825                         scat++;
826                 }
827
828                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
829                         &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
830
831                 prev = send;
832                 if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
833                         send = ic->i_sends;
834         }
835
836         /* if we finished the message then send completion owns it */
837         if (scat == &op->r_sg[op->r_count])
838                 prev->s_wr.send_flags = IB_SEND_SIGNALED;
839
840         if (i < work_alloc) {
841                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
842                 work_alloc = i;
843         }
844
845         failed_wr = &first->s_wr;
846         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
847         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
848                  first, &first->s_wr, ret, failed_wr);
849         BUG_ON(failed_wr != &first->s_wr);
850         if (ret) {
851                 printk(KERN_WARNING "RDS/IB: rdma ib_post_send to %pI4 "
852                        "returned %d\n", &conn->c_faddr, ret);
853                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
854                 goto out;
855         }
856
857         if (unlikely(failed_wr != &first->s_wr)) {
858                 printk(KERN_WARNING "RDS/IB: ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
859                 BUG_ON(failed_wr != &first->s_wr);
860         }
861
862
863 out:
864         return ret;
865 }
866
867 void rds_ib_xmit_complete(struct rds_connection *conn)
868 {
869         struct rds_ib_connection *ic = conn->c_transport_data;
870
871         /* We may have a pending ACK or window update we were unable
872          * to send previously (due to flow control). Try again. */
873         rds_ib_attempt_ack(ic);
874 }