RDS: fold rdma.h into rds.h
[linux-3.10.git] / 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 "ib.h"
40
41 static void rds_ib_send_rdma_complete(struct rds_message *rm,
42                                       int wc_status)
43 {
44         int notify_status;
45
46         switch (wc_status) {
47         case IB_WC_WR_FLUSH_ERR:
48                 return;
49
50         case IB_WC_SUCCESS:
51                 notify_status = RDS_RDMA_SUCCESS;
52                 break;
53
54         case IB_WC_REM_ACCESS_ERR:
55                 notify_status = RDS_RDMA_REMOTE_ERROR;
56                 break;
57
58         default:
59                 notify_status = RDS_RDMA_OTHER_ERROR;
60                 break;
61         }
62         rds_rdma_send_complete(rm, notify_status);
63 }
64
65 static void rds_ib_send_unmap_rdma(struct rds_ib_connection *ic,
66                                    struct rds_rdma_op *op)
67 {
68         if (op->r_mapped) {
69                 ib_dma_unmap_sg(ic->i_cm_id->device,
70                         op->r_sg, op->r_nents,
71                         op->r_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
72                 op->r_mapped = 0;
73         }
74 }
75
76 static void rds_ib_send_unmap_rm(struct rds_ib_connection *ic,
77                           struct rds_ib_send_work *send,
78                           int wc_status)
79 {
80         struct rds_message *rm = send->s_rm;
81
82         rdsdebug("ic %p send %p rm %p\n", ic, send, rm);
83
84         ib_dma_unmap_sg(ic->i_cm_id->device,
85                         rm->data.m_sg, rm->data.m_nents,
86                         DMA_TO_DEVICE);
87
88         if (rm->rdma.m_rdma_op) {
89                 rds_ib_send_unmap_rdma(ic, rm->rdma.m_rdma_op);
90
91                 /* If the user asked for a completion notification on this
92                  * message, we can implement three different semantics:
93                  *  1.  Notify when we received the ACK on the RDS message
94                  *      that was queued with the RDMA. This provides reliable
95                  *      notification of RDMA status at the expense of a one-way
96                  *      packet delay.
97                  *  2.  Notify when the IB stack gives us the completion event for
98                  *      the RDMA operation.
99                  *  3.  Notify when the IB stack gives us the completion event for
100                  *      the accompanying RDS messages.
101                  * Here, we implement approach #3. To implement approach #2,
102                  * call rds_rdma_send_complete from the cq_handler. To implement #1,
103                  * don't call rds_rdma_send_complete at all, and fall back to the notify
104                  * handling in the ACK processing code.
105                  *
106                  * Note: There's no need to explicitly sync any RDMA buffers using
107                  * ib_dma_sync_sg_for_cpu - the completion for the RDMA
108                  * operation itself unmapped the RDMA buffers, which takes care
109                  * of synching.
110                  */
111                 rds_ib_send_rdma_complete(rm, wc_status);
112
113                 if (rm->rdma.m_rdma_op->r_write)
114                         rds_stats_add(s_send_rdma_bytes, rm->rdma.m_rdma_op->r_bytes);
115                 else
116                         rds_stats_add(s_recv_rdma_bytes, rm->rdma.m_rdma_op->r_bytes);
117         }
118
119         /* If anyone waited for this message to get flushed out, wake
120          * them up now */
121         rds_message_unmapped(rm);
122
123         rds_message_put(rm);
124         send->s_rm = NULL;
125 }
126
127 void rds_ib_send_init_ring(struct rds_ib_connection *ic)
128 {
129         struct rds_ib_send_work *send;
130         u32 i;
131
132         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
133                 struct ib_sge *sge;
134
135                 send->s_rm = NULL;
136                 send->s_op = NULL;
137
138                 send->s_wr.wr_id = i;
139                 send->s_wr.sg_list = send->s_sge;
140                 send->s_wr.num_sge = 1;
141                 send->s_wr.opcode = IB_WR_SEND;
142                 send->s_wr.send_flags = 0;
143                 send->s_wr.ex.imm_data = 0;
144
145                 sge = rds_ib_data_sge(ic, send->s_sge);
146                 sge->lkey = ic->i_mr->lkey;
147
148                 sge = rds_ib_header_sge(ic, send->s_sge);
149                 sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header));
150                 sge->length = sizeof(struct rds_header);
151                 sge->lkey = ic->i_mr->lkey;
152         }
153 }
154
155 void rds_ib_send_clear_ring(struct rds_ib_connection *ic)
156 {
157         struct rds_ib_send_work *send;
158         u32 i;
159
160         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
161                 if (send->s_wr.opcode == 0xdead)
162                         continue;
163                 if (send->s_rm)
164                         rds_ib_send_unmap_rm(ic, send, IB_WC_WR_FLUSH_ERR);
165                 if (send->s_op)
166                         rds_ib_send_unmap_rdma(ic, send->s_op);
167         }
168 }
169
170 /*
171  * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
172  * operations performed in the send path.  As the sender allocs and potentially
173  * unallocs the next free entry in the ring it doesn't alter which is
174  * the next to be freed, which is what this is concerned with.
175  */
176 void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
177 {
178         struct rds_connection *conn = context;
179         struct rds_ib_connection *ic = conn->c_transport_data;
180         struct ib_wc wc;
181         struct rds_ib_send_work *send;
182         u32 completed;
183         u32 oldest;
184         u32 i = 0;
185         int ret;
186
187         rdsdebug("cq %p conn %p\n", cq, conn);
188         rds_ib_stats_inc(s_ib_tx_cq_call);
189         ret = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
190         if (ret)
191                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
192
193         while (ib_poll_cq(cq, 1, &wc) > 0) {
194                 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
195                          (unsigned long long)wc.wr_id, wc.status, wc.byte_len,
196                          be32_to_cpu(wc.ex.imm_data));
197                 rds_ib_stats_inc(s_ib_tx_cq_event);
198
199                 if (wc.wr_id == RDS_IB_ACK_WR_ID) {
200                         if (ic->i_ack_queued + HZ/2 < jiffies)
201                                 rds_ib_stats_inc(s_ib_tx_stalled);
202                         rds_ib_ack_send_complete(ic);
203                         continue;
204                 }
205
206                 oldest = rds_ib_ring_oldest(&ic->i_send_ring);
207
208                 completed = rds_ib_ring_completed(&ic->i_send_ring, wc.wr_id, oldest);
209
210                 for (i = 0; i < completed; i++) {
211                         send = &ic->i_sends[oldest];
212
213                         /* In the error case, wc.opcode sometimes contains garbage */
214                         switch (send->s_wr.opcode) {
215                         case IB_WR_SEND:
216                                 if (send->s_rm)
217                                         rds_ib_send_unmap_rm(ic, send, wc.status);
218                                 break;
219                         case IB_WR_RDMA_WRITE:
220                         case IB_WR_RDMA_READ:
221                                 /* Nothing to be done - the SG list will be unmapped
222                                  * when the SEND completes. */
223                                 break;
224                         default:
225                                 if (printk_ratelimit())
226                                         printk(KERN_NOTICE
227                                                 "RDS/IB: %s: unexpected opcode 0x%x in WR!\n",
228                                                 __func__, send->s_wr.opcode);
229                                 break;
230                         }
231
232                         send->s_wr.opcode = 0xdead;
233                         send->s_wr.num_sge = 1;
234                         if (send->s_queued + HZ/2 < jiffies)
235                                 rds_ib_stats_inc(s_ib_tx_stalled);
236
237                         /* If a RDMA operation produced an error, signal this right
238                          * away. If we don't, the subsequent SEND that goes with this
239                          * RDMA will be canceled with ERR_WFLUSH, and the application
240                          * never learn that the RDMA failed. */
241                         if (unlikely(wc.status == IB_WC_REM_ACCESS_ERR && send->s_op)) {
242                                 struct rds_message *rm;
243
244                                 rm = rds_send_get_message(conn, send->s_op);
245                                 if (rm) {
246                                         if (rm->rdma.m_rdma_op)
247                                                 rds_ib_send_unmap_rdma(ic, rm->rdma.m_rdma_op);
248                                         rds_ib_send_rdma_complete(rm, wc.status);
249                                         rds_message_put(rm);
250                                 }
251                         }
252
253                         oldest = (oldest + 1) % ic->i_send_ring.w_nr;
254                 }
255
256                 rds_ib_ring_free(&ic->i_send_ring, completed);
257
258                 if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
259                     test_bit(0, &conn->c_map_queued))
260                         queue_delayed_work(rds_wq, &conn->c_send_w, 0);
261
262                 /* We expect errors as the qp is drained during shutdown */
263                 if (wc.status != IB_WC_SUCCESS && rds_conn_up(conn)) {
264                         rds_ib_conn_error(conn,
265                                 "send completion on %pI4 "
266                                 "had status %u, disconnecting and reconnecting\n",
267                                 &conn->c_faddr, wc.status);
268                 }
269         }
270 }
271
272 /*
273  * This is the main function for allocating credits when sending
274  * messages.
275  *
276  * Conceptually, we have two counters:
277  *  -   send credits: this tells us how many WRs we're allowed
278  *      to submit without overruning the reciever's queue. For
279  *      each SEND WR we post, we decrement this by one.
280  *
281  *  -   posted credits: this tells us how many WRs we recently
282  *      posted to the receive queue. This value is transferred
283  *      to the peer as a "credit update" in a RDS header field.
284  *      Every time we transmit credits to the peer, we subtract
285  *      the amount of transferred credits from this counter.
286  *
287  * It is essential that we avoid situations where both sides have
288  * exhausted their send credits, and are unable to send new credits
289  * to the peer. We achieve this by requiring that we send at least
290  * one credit update to the peer before exhausting our credits.
291  * When new credits arrive, we subtract one credit that is withheld
292  * until we've posted new buffers and are ready to transmit these
293  * credits (see rds_ib_send_add_credits below).
294  *
295  * The RDS send code is essentially single-threaded; rds_send_xmit
296  * grabs c_send_lock to ensure exclusive access to the send ring.
297  * However, the ACK sending code is independent and can race with
298  * message SENDs.
299  *
300  * In the send path, we need to update the counters for send credits
301  * and the counter of posted buffers atomically - when we use the
302  * last available credit, we cannot allow another thread to race us
303  * and grab the posted credits counter.  Hence, we have to use a
304  * spinlock to protect the credit counter, or use atomics.
305  *
306  * Spinlocks shared between the send and the receive path are bad,
307  * because they create unnecessary delays. An early implementation
308  * using a spinlock showed a 5% degradation in throughput at some
309  * loads.
310  *
311  * This implementation avoids spinlocks completely, putting both
312  * counters into a single atomic, and updating that atomic using
313  * atomic_add (in the receive path, when receiving fresh credits),
314  * and using atomic_cmpxchg when updating the two counters.
315  */
316 int rds_ib_send_grab_credits(struct rds_ib_connection *ic,
317                              u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
318 {
319         unsigned int avail, posted, got = 0, advertise;
320         long oldval, newval;
321
322         *adv_credits = 0;
323         if (!ic->i_flowctl)
324                 return wanted;
325
326 try_again:
327         advertise = 0;
328         oldval = newval = atomic_read(&ic->i_credits);
329         posted = IB_GET_POST_CREDITS(oldval);
330         avail = IB_GET_SEND_CREDITS(oldval);
331
332         rdsdebug("rds_ib_send_grab_credits(%u): credits=%u posted=%u\n",
333                         wanted, avail, posted);
334
335         /* The last credit must be used to send a credit update. */
336         if (avail && !posted)
337                 avail--;
338
339         if (avail < wanted) {
340                 struct rds_connection *conn = ic->i_cm_id->context;
341
342                 /* Oops, there aren't that many credits left! */
343                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
344                 got = avail;
345         } else {
346                 /* Sometimes you get what you want, lalala. */
347                 got = wanted;
348         }
349         newval -= IB_SET_SEND_CREDITS(got);
350
351         /*
352          * If need_posted is non-zero, then the caller wants
353          * the posted regardless of whether any send credits are
354          * available.
355          */
356         if (posted && (got || need_posted)) {
357                 advertise = min_t(unsigned int, posted, max_posted);
358                 newval -= IB_SET_POST_CREDITS(advertise);
359         }
360
361         /* Finally bill everything */
362         if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
363                 goto try_again;
364
365         *adv_credits = advertise;
366         return got;
367 }
368
369 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits)
370 {
371         struct rds_ib_connection *ic = conn->c_transport_data;
372
373         if (credits == 0)
374                 return;
375
376         rdsdebug("rds_ib_send_add_credits(%u): current=%u%s\n",
377                         credits,
378                         IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
379                         test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
380
381         atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
382         if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
383                 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
384
385         WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
386
387         rds_ib_stats_inc(s_ib_rx_credit_updates);
388 }
389
390 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted)
391 {
392         struct rds_ib_connection *ic = conn->c_transport_data;
393
394         if (posted == 0)
395                 return;
396
397         atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
398
399         /* Decide whether to send an update to the peer now.
400          * If we would send a credit update for every single buffer we
401          * post, we would end up with an ACK storm (ACK arrives,
402          * consumes buffer, we refill the ring, send ACK to remote
403          * advertising the newly posted buffer... ad inf)
404          *
405          * Performance pretty much depends on how often we send
406          * credit updates - too frequent updates mean lots of ACKs.
407          * Too infrequent updates, and the peer will run out of
408          * credits and has to throttle.
409          * For the time being, 16 seems to be a good compromise.
410          */
411         if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
412                 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
413 }
414
415 static inline void
416 rds_ib_xmit_populate_wr(struct rds_ib_connection *ic,
417                 struct rds_ib_send_work *send, unsigned int pos,
418                 unsigned long buffer, unsigned int length,
419                 int send_flags)
420 {
421         struct ib_sge *sge;
422
423         WARN_ON(pos != send - ic->i_sends);
424
425         send->s_wr.send_flags = send_flags;
426         send->s_wr.opcode = IB_WR_SEND;
427         send->s_wr.num_sge = 2;
428         send->s_wr.next = NULL;
429         send->s_queued = jiffies;
430         send->s_op = NULL;
431
432         if (length != 0) {
433                 sge = rds_ib_data_sge(ic, send->s_sge);
434                 sge->addr = buffer;
435                 sge->length = length;
436                 sge->lkey = ic->i_mr->lkey;
437
438                 sge = rds_ib_header_sge(ic, send->s_sge);
439         } else {
440                 /* We're sending a packet with no payload. There is only
441                  * one SGE */
442                 send->s_wr.num_sge = 1;
443                 sge = &send->s_sge[0];
444         }
445
446         sge->addr = ic->i_send_hdrs_dma + (pos * sizeof(struct rds_header));
447         sge->length = sizeof(struct rds_header);
448         sge->lkey = ic->i_mr->lkey;
449 }
450
451 /*
452  * This can be called multiple times for a given message.  The first time
453  * we see a message we map its scatterlist into the IB device so that
454  * we can provide that mapped address to the IB scatter gather entries
455  * in the IB work requests.  We translate the scatterlist into a series
456  * of work requests that fragment the message.  These work requests complete
457  * in order so we pass ownership of the message to the completion handler
458  * once we send the final fragment.
459  *
460  * The RDS core uses the c_send_lock to only enter this function once
461  * per connection.  This makes sure that the tx ring alloc/unalloc pairs
462  * don't get out of sync and confuse the ring.
463  */
464 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
465                 unsigned int hdr_off, unsigned int sg, unsigned int off)
466 {
467         struct rds_ib_connection *ic = conn->c_transport_data;
468         struct ib_device *dev = ic->i_cm_id->device;
469         struct rds_ib_send_work *send = NULL;
470         struct rds_ib_send_work *first;
471         struct rds_ib_send_work *prev;
472         struct ib_send_wr *failed_wr;
473         struct scatterlist *scat;
474         u32 pos;
475         u32 i;
476         u32 work_alloc;
477         u32 credit_alloc;
478         u32 posted;
479         u32 adv_credits = 0;
480         int send_flags = 0;
481         int sent;
482         int ret;
483         int flow_controlled = 0;
484
485         BUG_ON(off % RDS_FRAG_SIZE);
486         BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
487
488         /* Do not send cong updates to IB loopback */
489         if (conn->c_loopback
490             && rm->m_inc.i_hdr.h_flags & RDS_FLAG_CONG_BITMAP) {
491                 rds_cong_map_updated(conn->c_fcong, ~(u64) 0);
492                 return sizeof(struct rds_header) + RDS_CONG_MAP_BYTES;
493         }
494
495         /* FIXME we may overallocate here */
496         if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
497                 i = 1;
498         else
499                 i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
500
501         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
502         if (work_alloc == 0) {
503                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
504                 rds_ib_stats_inc(s_ib_tx_ring_full);
505                 ret = -ENOMEM;
506                 goto out;
507         }
508
509         credit_alloc = work_alloc;
510         if (ic->i_flowctl) {
511                 credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
512                 adv_credits += posted;
513                 if (credit_alloc < work_alloc) {
514                         rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
515                         work_alloc = credit_alloc;
516                         flow_controlled++;
517                 }
518                 if (work_alloc == 0) {
519                         set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
520                         rds_ib_stats_inc(s_ib_tx_throttle);
521                         ret = -ENOMEM;
522                         goto out;
523                 }
524         }
525
526         /* map the message the first time we see it */
527         if (!ic->i_rm) {
528                 /*
529                 printk(KERN_NOTICE "rds_ib_xmit prep msg dport=%u flags=0x%x len=%d\n",
530                                 be16_to_cpu(rm->m_inc.i_hdr.h_dport),
531                                 rm->m_inc.i_hdr.h_flags,
532                                 be32_to_cpu(rm->m_inc.i_hdr.h_len));
533                    */
534                 if (rm->data.m_nents) {
535                         rm->data.m_count = ib_dma_map_sg(dev,
536                                                             rm->data.m_sg,
537                                                             rm->data.m_nents,
538                                                             DMA_TO_DEVICE);
539                         rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.m_count);
540                         if (rm->data.m_count == 0) {
541                                 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
542                                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
543                                 ret = -ENOMEM; /* XXX ? */
544                                 goto out;
545                         }
546                 } else {
547                         rm->data.m_count = 0;
548                 }
549
550                 ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
551                 ic->i_unsignaled_bytes = rds_ib_sysctl_max_unsig_bytes;
552                 rds_message_addref(rm);
553                 ic->i_rm = rm;
554
555                 /* Finalize the header */
556                 if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
557                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
558                 if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
559                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
560
561                 /* If it has a RDMA op, tell the peer we did it. This is
562                  * used by the peer to release use-once RDMA MRs. */
563                 if (rm->rdma.m_rdma_op) {
564                         struct rds_ext_header_rdma ext_hdr;
565
566                         ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.m_rdma_op->r_key);
567                         rds_message_add_extension(&rm->m_inc.i_hdr,
568                                         RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
569                 }
570                 if (rm->m_rdma_cookie) {
571                         rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
572                                         rds_rdma_cookie_key(rm->m_rdma_cookie),
573                                         rds_rdma_cookie_offset(rm->m_rdma_cookie));
574                 }
575
576                 /* Note - rds_ib_piggyb_ack clears the ACK_REQUIRED bit, so
577                  * we should not do this unless we have a chance of at least
578                  * sticking the header into the send ring. Which is why we
579                  * should call rds_ib_ring_alloc first. */
580                 rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_ib_piggyb_ack(ic));
581                 rds_message_make_checksum(&rm->m_inc.i_hdr);
582
583                 /*
584                  * Update adv_credits since we reset the ACK_REQUIRED bit.
585                  */
586                 rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
587                 adv_credits += posted;
588                 BUG_ON(adv_credits > 255);
589         }
590
591         send = &ic->i_sends[pos];
592         first = send;
593         prev = NULL;
594         scat = &rm->data.m_sg[sg];
595         sent = 0;
596         i = 0;
597
598         /* Sometimes you want to put a fence between an RDMA
599          * READ and the following SEND.
600          * We could either do this all the time
601          * or when requested by the user. Right now, we let
602          * the application choose.
603          */
604         if (rm->rdma.m_rdma_op && rm->rdma.m_rdma_op->r_fence)
605                 send_flags = IB_SEND_FENCE;
606
607         /*
608          * We could be copying the header into the unused tail of the page.
609          * That would need to be changed in the future when those pages might
610          * be mapped userspace pages or page cache pages.  So instead we always
611          * use a second sge and our long-lived ring of mapped headers.  We send
612          * the header after the data so that the data payload can be aligned on
613          * the receiver.
614          */
615
616         /* handle a 0-len message */
617         if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0) {
618                 rds_ib_xmit_populate_wr(ic, send, pos, 0, 0, send_flags);
619                 goto add_header;
620         }
621
622         /* if there's data reference it with a chain of work reqs */
623         for (; i < work_alloc && scat != &rm->data.m_sg[rm->data.m_count]; i++) {
624                 unsigned int len;
625
626                 send = &ic->i_sends[pos];
627
628                 len = min(RDS_FRAG_SIZE, ib_sg_dma_len(dev, scat) - off);
629                 rds_ib_xmit_populate_wr(ic, send, pos,
630                                 ib_sg_dma_address(dev, scat) + off, len,
631                                 send_flags);
632
633                 /*
634                  * We want to delay signaling completions just enough to get
635                  * the batching benefits but not so much that we create dead time
636                  * on the wire.
637                  */
638                 if (ic->i_unsignaled_wrs-- == 0) {
639                         ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
640                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
641                 }
642
643                 ic->i_unsignaled_bytes -= len;
644                 if (ic->i_unsignaled_bytes <= 0) {
645                         ic->i_unsignaled_bytes = rds_ib_sysctl_max_unsig_bytes;
646                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
647                 }
648
649                 /*
650                  * Always signal the last one if we're stopping due to flow control.
651                  */
652                 if (flow_controlled && i == (work_alloc-1))
653                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
654
655                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
656                          &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
657
658                 sent += len;
659                 off += len;
660                 if (off == ib_sg_dma_len(dev, scat)) {
661                         scat++;
662                         off = 0;
663                 }
664
665 add_header:
666                 /* Tack on the header after the data. The header SGE should already
667                  * have been set up to point to the right header buffer. */
668                 memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
669
670                 if (0) {
671                         struct rds_header *hdr = &ic->i_send_hdrs[pos];
672
673                         printk(KERN_NOTICE "send WR dport=%u flags=0x%x len=%d\n",
674                                 be16_to_cpu(hdr->h_dport),
675                                 hdr->h_flags,
676                                 be32_to_cpu(hdr->h_len));
677                 }
678                 if (adv_credits) {
679                         struct rds_header *hdr = &ic->i_send_hdrs[pos];
680
681                         /* add credit and redo the header checksum */
682                         hdr->h_credit = adv_credits;
683                         rds_message_make_checksum(hdr);
684                         adv_credits = 0;
685                         rds_ib_stats_inc(s_ib_tx_credit_updates);
686                 }
687
688                 if (prev)
689                         prev->s_wr.next = &send->s_wr;
690                 prev = send;
691
692                 pos = (pos + 1) % ic->i_send_ring.w_nr;
693         }
694
695         /* Account the RDS header in the number of bytes we sent, but just once.
696          * The caller has no concept of fragmentation. */
697         if (hdr_off == 0)
698                 sent += sizeof(struct rds_header);
699
700         /* if we finished the message then send completion owns it */
701         if (scat == &rm->data.m_sg[rm->data.m_count]) {
702                 prev->s_rm = ic->i_rm;
703                 prev->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
704                 ic->i_rm = NULL;
705         }
706
707         if (i < work_alloc) {
708                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
709                 work_alloc = i;
710         }
711         if (ic->i_flowctl && i < credit_alloc)
712                 rds_ib_send_add_credits(conn, credit_alloc - i);
713
714         /* XXX need to worry about failed_wr and partial sends. */
715         failed_wr = &first->s_wr;
716         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
717         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
718                  first, &first->s_wr, ret, failed_wr);
719         BUG_ON(failed_wr != &first->s_wr);
720         if (ret) {
721                 printk(KERN_WARNING "RDS/IB: ib_post_send to %pI4 "
722                        "returned %d\n", &conn->c_faddr, ret);
723                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
724                 if (prev->s_rm) {
725                         ic->i_rm = prev->s_rm;
726                         prev->s_rm = NULL;
727                 }
728
729                 rds_ib_conn_error(ic->conn, "ib_post_send failed\n");
730                 goto out;
731         }
732
733         ret = sent;
734 out:
735         BUG_ON(adv_credits);
736         return ret;
737 }
738
739 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
740 {
741         struct rds_ib_connection *ic = conn->c_transport_data;
742         struct rds_ib_send_work *send = NULL;
743         struct rds_ib_send_work *first;
744         struct rds_ib_send_work *prev;
745         struct ib_send_wr *failed_wr;
746         struct rds_ib_device *rds_ibdev;
747         struct scatterlist *scat;
748         unsigned long len;
749         u64 remote_addr = op->r_remote_addr;
750         u32 pos;
751         u32 work_alloc;
752         u32 i;
753         u32 j;
754         int sent;
755         int ret;
756         int num_sge;
757
758         rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
759
760         /* map the message the first time we see it */
761         if (!op->r_mapped) {
762                 op->r_count = ib_dma_map_sg(ic->i_cm_id->device,
763                                         op->r_sg, op->r_nents, (op->r_write) ?
764                                         DMA_TO_DEVICE : DMA_FROM_DEVICE);
765                 rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->r_count);
766                 if (op->r_count == 0) {
767                         rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
768                         ret = -ENOMEM; /* XXX ? */
769                         goto out;
770                 }
771
772                 op->r_mapped = 1;
773         }
774
775         /*
776          * Instead of knowing how to return a partial rdma read/write we insist that there
777          * be enough work requests to send the entire message.
778          */
779         i = ceil(op->r_count, rds_ibdev->max_sge);
780
781         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
782         if (work_alloc != i) {
783                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
784                 rds_ib_stats_inc(s_ib_tx_ring_full);
785                 ret = -ENOMEM;
786                 goto out;
787         }
788
789         send = &ic->i_sends[pos];
790         first = send;
791         prev = NULL;
792         scat = &op->r_sg[0];
793         sent = 0;
794         num_sge = op->r_count;
795
796         for (i = 0; i < work_alloc && scat != &op->r_sg[op->r_count]; i++) {
797                 send->s_wr.send_flags = 0;
798                 send->s_queued = jiffies;
799                 /*
800                  * We want to delay signaling completions just enough to get
801                  * the batching benefits but not so much that we create dead time on the wire.
802                  */
803                 if (ic->i_unsignaled_wrs-- == 0) {
804                         ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
805                         send->s_wr.send_flags = IB_SEND_SIGNALED;
806                 }
807
808                 send->s_wr.opcode = op->r_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ;
809                 send->s_wr.wr.rdma.remote_addr = remote_addr;
810                 send->s_wr.wr.rdma.rkey = op->r_key;
811                 send->s_op = op;
812
813                 if (num_sge > rds_ibdev->max_sge) {
814                         send->s_wr.num_sge = rds_ibdev->max_sge;
815                         num_sge -= rds_ibdev->max_sge;
816                 } else {
817                         send->s_wr.num_sge = num_sge;
818                 }
819
820                 send->s_wr.next = NULL;
821
822                 if (prev)
823                         prev->s_wr.next = &send->s_wr;
824
825                 for (j = 0; j < send->s_wr.num_sge && scat != &op->r_sg[op->r_count]; j++) {
826                         len = ib_sg_dma_len(ic->i_cm_id->device, scat);
827                         send->s_sge[j].addr =
828                                  ib_sg_dma_address(ic->i_cm_id->device, scat);
829                         send->s_sge[j].length = len;
830                         send->s_sge[j].lkey = ic->i_mr->lkey;
831
832                         sent += len;
833                         rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
834
835                         remote_addr += len;
836                         scat++;
837                 }
838
839                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
840                         &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
841
842                 prev = send;
843                 if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
844                         send = ic->i_sends;
845         }
846
847         /* if we finished the message then send completion owns it */
848         if (scat == &op->r_sg[op->r_count])
849                 prev->s_wr.send_flags = IB_SEND_SIGNALED;
850
851         if (i < work_alloc) {
852                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
853                 work_alloc = i;
854         }
855
856         failed_wr = &first->s_wr;
857         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
858         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
859                  first, &first->s_wr, ret, failed_wr);
860         BUG_ON(failed_wr != &first->s_wr);
861         if (ret) {
862                 printk(KERN_WARNING "RDS/IB: rdma ib_post_send to %pI4 "
863                        "returned %d\n", &conn->c_faddr, ret);
864                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
865                 goto out;
866         }
867
868         if (unlikely(failed_wr != &first->s_wr)) {
869                 printk(KERN_WARNING "RDS/IB: ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
870                 BUG_ON(failed_wr != &first->s_wr);
871         }
872
873
874 out:
875         return ret;
876 }
877
878 void rds_ib_xmit_complete(struct rds_connection *conn)
879 {
880         struct rds_ib_connection *ic = conn->c_transport_data;
881
882         /* We may have a pending ACK or window update we were unable
883          * to send previously (due to flow control). Try again. */
884         rds_ib_attempt_ack(ic);
885 }