RDS/IB: Make all flow control code conditional on i_flowctl
[linux-2.6.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 /*
42  * Convert IB-specific error message to RDS error message and call core
43  * completion handler.
44  */
45 static void rds_ib_send_complete(struct rds_message *rm,
46                                  int wc_status,
47                                  void (*complete)(struct rds_message *rm, int status))
48 {
49         int notify_status;
50
51         switch (wc_status) {
52         case IB_WC_WR_FLUSH_ERR:
53                 return;
54
55         case IB_WC_SUCCESS:
56                 notify_status = RDS_RDMA_SUCCESS;
57                 break;
58
59         case IB_WC_REM_ACCESS_ERR:
60                 notify_status = RDS_RDMA_REMOTE_ERROR;
61                 break;
62
63         default:
64                 notify_status = RDS_RDMA_OTHER_ERROR;
65                 break;
66         }
67         complete(rm, notify_status);
68 }
69
70 static void rds_ib_send_unmap_rm(struct rds_ib_connection *ic,
71                           struct rds_ib_send_work *send,
72                           int wc_status)
73 {
74         struct rds_message *rm = send->s_rm;
75
76         rdsdebug("ic %p send %p rm %p\n", ic, send, rm);
77
78         ib_dma_unmap_sg(ic->i_cm_id->device,
79                         rm->data.m_sg, rm->data.m_nents,
80                         DMA_TO_DEVICE);
81
82         if (rm->rdma.m_rdma_op.r_active) {
83                 struct rds_rdma_op *op = &rm->rdma.m_rdma_op;
84
85                 if (op->r_mapped) {
86                         ib_dma_unmap_sg(ic->i_cm_id->device,
87                                         op->r_sg, op->r_nents,
88                                         op->r_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
89                         op->r_mapped = 0;
90                 }
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_complete(rm, wc_status, rds_rdma_send_complete);
113
114                 if (rm->rdma.m_rdma_op.r_write)
115                         rds_stats_add(s_send_rdma_bytes, rm->rdma.m_rdma_op.r_bytes);
116                 else
117                         rds_stats_add(s_recv_rdma_bytes, rm->rdma.m_rdma_op.r_bytes);
118         }
119
120         if (rm->atomic.op_active) {
121                 struct rm_atomic_op *op = &rm->atomic;
122
123                 /* unmap atomic recvbuf */
124                 if (op->op_mapped) {
125                         ib_dma_unmap_sg(ic->i_cm_id->device, op->op_sg, 1,
126                                         DMA_FROM_DEVICE);
127                         op->op_mapped = 0;
128                 }
129
130                 rds_ib_send_complete(rm, wc_status, rds_atomic_send_complete);
131
132                 if (rm->atomic.op_type == RDS_ATOMIC_TYPE_CSWP)
133                         rds_stats_inc(s_atomic_cswp);
134                 else
135                         rds_stats_inc(s_atomic_fadd);
136         }
137
138         /* If anyone waited for this message to get flushed out, wake
139          * them up now */
140         rds_message_unmapped(rm);
141
142         rds_message_put(rm);
143         send->s_rm = NULL;
144 }
145
146 void rds_ib_send_init_ring(struct rds_ib_connection *ic)
147 {
148         struct rds_ib_send_work *send;
149         u32 i;
150
151         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
152                 struct ib_sge *sge;
153
154                 send->s_rm = NULL;
155                 send->s_op = NULL;
156
157                 send->s_wr.wr_id = i;
158                 send->s_wr.sg_list = send->s_sge;
159                 send->s_wr.ex.imm_data = 0;
160
161                 sge = &send->s_sge[0];
162                 sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header));
163                 sge->length = sizeof(struct rds_header);
164                 sge->lkey = ic->i_mr->lkey;
165
166                 send->s_sge[1].lkey = ic->i_mr->lkey;
167         }
168 }
169
170 void rds_ib_send_clear_ring(struct rds_ib_connection *ic)
171 {
172         struct rds_ib_send_work *send;
173         u32 i;
174
175         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
176                 if (!send->s_rm || send->s_wr.opcode == 0xdead)
177                         continue;
178                 rds_ib_send_unmap_rm(ic, send, IB_WC_WR_FLUSH_ERR);
179         }
180 }
181
182 /*
183  * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
184  * operations performed in the send path.  As the sender allocs and potentially
185  * unallocs the next free entry in the ring it doesn't alter which is
186  * the next to be freed, which is what this is concerned with.
187  */
188 void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
189 {
190         struct rds_connection *conn = context;
191         struct rds_ib_connection *ic = conn->c_transport_data;
192         struct ib_wc wc;
193         struct rds_ib_send_work *send;
194         u32 completed;
195         u32 oldest;
196         u32 i = 0;
197         int ret;
198
199         rdsdebug("cq %p conn %p\n", cq, conn);
200         rds_ib_stats_inc(s_ib_tx_cq_call);
201         ret = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
202         if (ret)
203                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
204
205         while (ib_poll_cq(cq, 1, &wc) > 0) {
206                 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
207                          (unsigned long long)wc.wr_id, wc.status, wc.byte_len,
208                          be32_to_cpu(wc.ex.imm_data));
209                 rds_ib_stats_inc(s_ib_tx_cq_event);
210
211                 if (wc.wr_id == RDS_IB_ACK_WR_ID) {
212                         if (ic->i_ack_queued + HZ/2 < jiffies)
213                                 rds_ib_stats_inc(s_ib_tx_stalled);
214                         rds_ib_ack_send_complete(ic);
215                         continue;
216                 }
217
218                 oldest = rds_ib_ring_oldest(&ic->i_send_ring);
219
220                 completed = rds_ib_ring_completed(&ic->i_send_ring, wc.wr_id, oldest);
221
222                 for (i = 0; i < completed; i++) {
223                         send = &ic->i_sends[oldest];
224
225                         /* In the error case, wc.opcode sometimes contains garbage */
226                         switch (send->s_wr.opcode) {
227                         case IB_WR_SEND:
228                                 if (send->s_rm)
229                                         rds_ib_send_unmap_rm(ic, send, wc.status);
230                                 break;
231                         case IB_WR_RDMA_WRITE:
232                         case IB_WR_RDMA_READ:
233                         case IB_WR_ATOMIC_FETCH_AND_ADD:
234                         case IB_WR_ATOMIC_CMP_AND_SWP:
235                                 /* Nothing to be done - the SG list will be unmapped
236                                  * when the SEND completes. */
237                                 break;
238                         default:
239                                 if (printk_ratelimit())
240                                         printk(KERN_NOTICE
241                                                 "RDS/IB: %s: unexpected opcode 0x%x in WR!\n",
242                                                 __func__, send->s_wr.opcode);
243                                 break;
244                         }
245
246                         send->s_wr.opcode = 0xdead;
247                         send->s_wr.num_sge = 1;
248                         if (send->s_queued + HZ/2 < jiffies)
249                                 rds_ib_stats_inc(s_ib_tx_stalled);
250
251                         /* If a RDMA operation produced an error, signal this right
252                          * away. If we don't, the subsequent SEND that goes with this
253                          * RDMA will be canceled with ERR_WFLUSH, and the application
254                          * never learn that the RDMA failed. */
255                         if (unlikely(wc.status == IB_WC_REM_ACCESS_ERR && send->s_op)) {
256                                 struct rds_message *rm;
257
258                                 rm = rds_send_get_message(conn, send->s_op);
259                                 if (rm) {
260                                         rds_ib_send_unmap_rm(ic, send, wc.status);
261                                         rds_ib_send_complete(rm, wc.status, rds_rdma_send_complete);
262                                         rds_message_put(rm);
263                                 }
264                         }
265
266                         oldest = (oldest + 1) % ic->i_send_ring.w_nr;
267                 }
268
269                 rds_ib_ring_free(&ic->i_send_ring, completed);
270
271                 if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
272                     test_bit(0, &conn->c_map_queued))
273                         queue_delayed_work(rds_wq, &conn->c_send_w, 0);
274
275                 /* We expect errors as the qp is drained during shutdown */
276                 if (wc.status != IB_WC_SUCCESS && rds_conn_up(conn)) {
277                         rds_ib_conn_error(conn,
278                                 "send completion on %pI4 "
279                                 "had status %u, disconnecting and reconnecting\n",
280                                 &conn->c_faddr, wc.status);
281                 }
282         }
283 }
284
285 /*
286  * This is the main function for allocating credits when sending
287  * messages.
288  *
289  * Conceptually, we have two counters:
290  *  -   send credits: this tells us how many WRs we're allowed
291  *      to submit without overruning the reciever's queue. For
292  *      each SEND WR we post, we decrement this by one.
293  *
294  *  -   posted credits: this tells us how many WRs we recently
295  *      posted to the receive queue. This value is transferred
296  *      to the peer as a "credit update" in a RDS header field.
297  *      Every time we transmit credits to the peer, we subtract
298  *      the amount of transferred credits from this counter.
299  *
300  * It is essential that we avoid situations where both sides have
301  * exhausted their send credits, and are unable to send new credits
302  * to the peer. We achieve this by requiring that we send at least
303  * one credit update to the peer before exhausting our credits.
304  * When new credits arrive, we subtract one credit that is withheld
305  * until we've posted new buffers and are ready to transmit these
306  * credits (see rds_ib_send_add_credits below).
307  *
308  * The RDS send code is essentially single-threaded; rds_send_xmit
309  * grabs c_send_lock to ensure exclusive access to the send ring.
310  * However, the ACK sending code is independent and can race with
311  * message SENDs.
312  *
313  * In the send path, we need to update the counters for send credits
314  * and the counter of posted buffers atomically - when we use the
315  * last available credit, we cannot allow another thread to race us
316  * and grab the posted credits counter.  Hence, we have to use a
317  * spinlock to protect the credit counter, or use atomics.
318  *
319  * Spinlocks shared between the send and the receive path are bad,
320  * because they create unnecessary delays. An early implementation
321  * using a spinlock showed a 5% degradation in throughput at some
322  * loads.
323  *
324  * This implementation avoids spinlocks completely, putting both
325  * counters into a single atomic, and updating that atomic using
326  * atomic_add (in the receive path, when receiving fresh credits),
327  * and using atomic_cmpxchg when updating the two counters.
328  */
329 int rds_ib_send_grab_credits(struct rds_ib_connection *ic,
330                              u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
331 {
332         unsigned int avail, posted, got = 0, advertise;
333         long oldval, newval;
334
335         *adv_credits = 0;
336         if (!ic->i_flowctl)
337                 return wanted;
338
339 try_again:
340         advertise = 0;
341         oldval = newval = atomic_read(&ic->i_credits);
342         posted = IB_GET_POST_CREDITS(oldval);
343         avail = IB_GET_SEND_CREDITS(oldval);
344
345         rdsdebug("rds_ib_send_grab_credits(%u): credits=%u posted=%u\n",
346                         wanted, avail, posted);
347
348         /* The last credit must be used to send a credit update. */
349         if (avail && !posted)
350                 avail--;
351
352         if (avail < wanted) {
353                 struct rds_connection *conn = ic->i_cm_id->context;
354
355                 /* Oops, there aren't that many credits left! */
356                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
357                 got = avail;
358         } else {
359                 /* Sometimes you get what you want, lalala. */
360                 got = wanted;
361         }
362         newval -= IB_SET_SEND_CREDITS(got);
363
364         /*
365          * If need_posted is non-zero, then the caller wants
366          * the posted regardless of whether any send credits are
367          * available.
368          */
369         if (posted && (got || need_posted)) {
370                 advertise = min_t(unsigned int, posted, max_posted);
371                 newval -= IB_SET_POST_CREDITS(advertise);
372         }
373
374         /* Finally bill everything */
375         if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
376                 goto try_again;
377
378         *adv_credits = advertise;
379         return got;
380 }
381
382 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits)
383 {
384         struct rds_ib_connection *ic = conn->c_transport_data;
385
386         if (credits == 0)
387                 return;
388
389         rdsdebug("rds_ib_send_add_credits(%u): current=%u%s\n",
390                         credits,
391                         IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
392                         test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
393
394         atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
395         if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
396                 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
397
398         WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
399
400         rds_ib_stats_inc(s_ib_rx_credit_updates);
401 }
402
403 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted)
404 {
405         struct rds_ib_connection *ic = conn->c_transport_data;
406
407         if (posted == 0)
408                 return;
409
410         atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
411
412         /* Decide whether to send an update to the peer now.
413          * If we would send a credit update for every single buffer we
414          * post, we would end up with an ACK storm (ACK arrives,
415          * consumes buffer, we refill the ring, send ACK to remote
416          * advertising the newly posted buffer... ad inf)
417          *
418          * Performance pretty much depends on how often we send
419          * credit updates - too frequent updates mean lots of ACKs.
420          * Too infrequent updates, and the peer will run out of
421          * credits and has to throttle.
422          * For the time being, 16 seems to be a good compromise.
423          */
424         if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
425                 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
426 }
427
428 /*
429  * This can be called multiple times for a given message.  The first time
430  * we see a message we map its scatterlist into the IB device so that
431  * we can provide that mapped address to the IB scatter gather entries
432  * in the IB work requests.  We translate the scatterlist into a series
433  * of work requests that fragment the message.  These work requests complete
434  * in order so we pass ownership of the message to the completion handler
435  * once we send the final fragment.
436  *
437  * The RDS core uses the c_send_lock to only enter this function once
438  * per connection.  This makes sure that the tx ring alloc/unalloc pairs
439  * don't get out of sync and confuse the ring.
440  */
441 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
442                 unsigned int hdr_off, unsigned int sg, unsigned int off)
443 {
444         struct rds_ib_connection *ic = conn->c_transport_data;
445         struct ib_device *dev = ic->i_cm_id->device;
446         struct rds_ib_send_work *send = NULL;
447         struct rds_ib_send_work *first;
448         struct rds_ib_send_work *prev;
449         struct ib_send_wr *failed_wr;
450         struct scatterlist *scat;
451         u32 pos;
452         u32 i;
453         u32 work_alloc;
454         u32 credit_alloc = 0;
455         u32 posted;
456         u32 adv_credits = 0;
457         int send_flags = 0;
458         int bytes_sent = 0;
459         int ret;
460         int flow_controlled = 0;
461
462         BUG_ON(off % RDS_FRAG_SIZE);
463         BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
464
465         /* Do not send cong updates to IB loopback */
466         if (conn->c_loopback
467             && rm->m_inc.i_hdr.h_flags & RDS_FLAG_CONG_BITMAP) {
468                 rds_cong_map_updated(conn->c_fcong, ~(u64) 0);
469                 return sizeof(struct rds_header) + RDS_CONG_MAP_BYTES;
470         }
471
472         /* FIXME we may overallocate here */
473         if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
474                 i = 1;
475         else
476                 i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
477
478         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
479         if (work_alloc == 0) {
480                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
481                 rds_ib_stats_inc(s_ib_tx_ring_full);
482                 ret = -ENOMEM;
483                 goto out;
484         }
485
486         if (ic->i_flowctl) {
487                 credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
488                 adv_credits += posted;
489                 if (credit_alloc < work_alloc) {
490                         rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
491                         work_alloc = credit_alloc;
492                         flow_controlled = 1;
493                 }
494                 if (work_alloc == 0) {
495                         set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
496                         rds_ib_stats_inc(s_ib_tx_throttle);
497                         ret = -ENOMEM;
498                         goto out;
499                 }
500         }
501
502         /* map the message the first time we see it */
503         if (!ic->i_rm) {
504                 if (rm->data.m_nents) {
505                         rm->data.m_count = ib_dma_map_sg(dev,
506                                                             rm->data.m_sg,
507                                                             rm->data.m_nents,
508                                                             DMA_TO_DEVICE);
509                         rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.m_count);
510                         if (rm->data.m_count == 0) {
511                                 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
512                                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
513                                 ret = -ENOMEM; /* XXX ? */
514                                 goto out;
515                         }
516                 } else {
517                         rm->data.m_count = 0;
518                 }
519
520                 ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
521                 rds_message_addref(rm);
522                 ic->i_rm = rm;
523
524                 /* Finalize the header */
525                 if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
526                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
527                 if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
528                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
529
530                 /* If it has a RDMA op, tell the peer we did it. This is
531                  * used by the peer to release use-once RDMA MRs. */
532                 if (rm->rdma.m_rdma_op.r_active) {
533                         struct rds_ext_header_rdma ext_hdr;
534
535                         ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.m_rdma_op.r_key);
536                         rds_message_add_extension(&rm->m_inc.i_hdr,
537                                         RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
538                 }
539                 if (rm->m_rdma_cookie) {
540                         rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
541                                         rds_rdma_cookie_key(rm->m_rdma_cookie),
542                                         rds_rdma_cookie_offset(rm->m_rdma_cookie));
543                 }
544
545                 /* Note - rds_ib_piggyb_ack clears the ACK_REQUIRED bit, so
546                  * we should not do this unless we have a chance of at least
547                  * sticking the header into the send ring. Which is why we
548                  * should call rds_ib_ring_alloc first. */
549                 rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_ib_piggyb_ack(ic));
550                 rds_message_make_checksum(&rm->m_inc.i_hdr);
551
552                 /*
553                  * Update adv_credits since we reset the ACK_REQUIRED bit.
554                  */
555                 if (ic->i_flowctl) {
556                         rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
557                         adv_credits += posted;
558                         BUG_ON(adv_credits > 255);
559                 }
560         }
561
562         /* Sometimes you want to put a fence between an RDMA
563          * READ and the following SEND.
564          * We could either do this all the time
565          * or when requested by the user. Right now, we let
566          * the application choose.
567          */
568         if (rm->rdma.m_rdma_op.r_active && rm->rdma.m_rdma_op.r_fence)
569                 send_flags = IB_SEND_FENCE;
570
571         /* Each frag gets a header. Msgs may be 0 bytes */
572         send = &ic->i_sends[pos];
573         first = send;
574         prev = NULL;
575         scat = &rm->data.m_sg[sg];
576         i = 0;
577         do {
578                 unsigned int len = 0;
579
580                 /* Set up the header */
581                 send->s_wr.send_flags = send_flags;
582                 send->s_wr.opcode = IB_WR_SEND;
583                 send->s_wr.num_sge = 1;
584                 send->s_wr.next = NULL;
585                 send->s_queued = jiffies;
586                 send->s_op = NULL;
587
588                 send->s_sge[0].addr = ic->i_send_hdrs_dma
589                         + (pos * sizeof(struct rds_header));
590                 send->s_sge[0].length = sizeof(struct rds_header);
591
592                 memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
593
594                 /* Set up the data, if present */
595                 if (i < work_alloc
596                     && scat != &rm->data.m_sg[rm->data.m_count]) {
597                         len = min(RDS_FRAG_SIZE, ib_sg_dma_len(dev, scat) - off);
598                         send->s_wr.num_sge = 2;
599
600                         send->s_sge[1].addr = ib_sg_dma_address(dev, scat) + off;
601                         send->s_sge[1].length = len;
602
603                         bytes_sent += len;
604                         off += len;
605                         if (off == ib_sg_dma_len(dev, scat)) {
606                                 scat++;
607                                 off = 0;
608                         }
609                 }
610
611                 /*
612                  * We want to delay signaling completions just enough to get
613                  * the batching benefits but not so much that we create dead time
614                  * on the wire.
615                  */
616                 if (ic->i_unsignaled_wrs-- == 0) {
617                         ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
618                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
619                 }
620
621                 /*
622                  * Always signal the last one if we're stopping due to flow control.
623                  */
624                 if (ic->i_flowctl && flow_controlled && i == (work_alloc-1))
625                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
626
627                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
628                          &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
629
630                 if (ic->i_flowctl && adv_credits) {
631                         struct rds_header *hdr = &ic->i_send_hdrs[pos];
632
633                         /* add credit and redo the header checksum */
634                         hdr->h_credit = adv_credits;
635                         rds_message_make_checksum(hdr);
636                         adv_credits = 0;
637                         rds_ib_stats_inc(s_ib_tx_credit_updates);
638                 }
639
640                 if (prev)
641                         prev->s_wr.next = &send->s_wr;
642                 prev = send;
643
644                 pos = (pos + 1) % ic->i_send_ring.w_nr;
645                 send = &ic->i_sends[pos];
646                 i++;
647
648         } while (i < work_alloc
649                  && scat != &rm->data.m_sg[rm->data.m_count]);
650
651         /* Account the RDS header in the number of bytes we sent, but just once.
652          * The caller has no concept of fragmentation. */
653         if (hdr_off == 0)
654                 bytes_sent += sizeof(struct rds_header);
655
656         /* if we finished the message then send completion owns it */
657         if (scat == &rm->data.m_sg[rm->data.m_count]) {
658                 prev->s_rm = ic->i_rm;
659                 prev->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
660                 ic->i_rm = NULL;
661         }
662
663         /* Put back wrs & credits we didn't use */
664         if (i < work_alloc) {
665                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
666                 work_alloc = i;
667         }
668         if (ic->i_flowctl && i < credit_alloc)
669                 rds_ib_send_add_credits(conn, credit_alloc - i);
670
671         /* XXX need to worry about failed_wr and partial sends. */
672         failed_wr = &first->s_wr;
673         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
674         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
675                  first, &first->s_wr, ret, failed_wr);
676         BUG_ON(failed_wr != &first->s_wr);
677         if (ret) {
678                 printk(KERN_WARNING "RDS/IB: ib_post_send to %pI4 "
679                        "returned %d\n", &conn->c_faddr, ret);
680                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
681                 if (prev->s_rm) {
682                         ic->i_rm = prev->s_rm;
683                         prev->s_rm = NULL;
684                 }
685
686                 rds_ib_conn_error(ic->conn, "ib_post_send failed\n");
687                 goto out;
688         }
689
690         ret = bytes_sent;
691 out:
692         BUG_ON(adv_credits);
693         return ret;
694 }
695
696 /*
697  * Issue atomic operation.
698  * A simplified version of the rdma case, we always map 1 SG, and
699  * only 8 bytes, for the return value from the atomic operation.
700  */
701 int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op)
702 {
703         struct rds_ib_connection *ic = conn->c_transport_data;
704         struct rds_ib_send_work *send = NULL;
705         struct ib_send_wr *failed_wr;
706         struct rds_ib_device *rds_ibdev;
707         u32 pos;
708         u32 work_alloc;
709         int ret;
710
711         rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
712
713         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, 1, &pos);
714         if (work_alloc != 1) {
715                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
716                 rds_ib_stats_inc(s_ib_tx_ring_full);
717                 ret = -ENOMEM;
718                 goto out;
719         }
720
721         /* address of send request in ring */
722         send = &ic->i_sends[pos];
723         send->s_queued = jiffies;
724
725         if (op->op_type == RDS_ATOMIC_TYPE_CSWP) {
726                 send->s_wr.opcode = IB_WR_ATOMIC_CMP_AND_SWP;
727                 send->s_wr.wr.atomic.compare_add = op->op_compare;
728                 send->s_wr.wr.atomic.swap = op->op_swap_add;
729         } else { /* FADD */
730                 send->s_wr.opcode = IB_WR_ATOMIC_FETCH_AND_ADD;
731                 send->s_wr.wr.atomic.compare_add = op->op_swap_add;
732                 send->s_wr.wr.atomic.swap = 0;
733         }
734         send->s_wr.send_flags = IB_SEND_SIGNALED;
735         send->s_wr.num_sge = 1;
736         send->s_wr.next = NULL;
737         send->s_wr.wr.atomic.remote_addr = op->op_remote_addr;
738         send->s_wr.wr.atomic.rkey = op->op_rkey;
739
740         /* map 8 byte retval buffer to the device */
741         ret = ib_dma_map_sg(ic->i_cm_id->device, op->op_sg, 1, DMA_FROM_DEVICE);
742         rdsdebug("ic %p mapping atomic op %p. mapped %d pg\n", ic, op, ret);
743         if (ret != 1) {
744                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
745                 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
746                 ret = -ENOMEM; /* XXX ? */
747                 goto out;
748         }
749
750         /* Convert our struct scatterlist to struct ib_sge */
751         send->s_sge[0].addr = ib_sg_dma_address(ic->i_cm_id->device, op->op_sg);
752         send->s_sge[0].length = ib_sg_dma_len(ic->i_cm_id->device, op->op_sg);
753         send->s_sge[0].lkey = ic->i_mr->lkey;
754
755         rdsdebug("rva %Lx rpa %Lx len %u\n", op->op_remote_addr,
756                  send->s_sge[0].addr, send->s_sge[0].length);
757
758         failed_wr = &send->s_wr;
759         ret = ib_post_send(ic->i_cm_id->qp, &send->s_wr, &failed_wr);
760         rdsdebug("ic %p send %p (wr %p) ret %d wr %p\n", ic,
761                  send, &send->s_wr, ret, failed_wr);
762         BUG_ON(failed_wr != &send->s_wr);
763         if (ret) {
764                 printk(KERN_WARNING "RDS/IB: atomic ib_post_send to %pI4 "
765                        "returned %d\n", &conn->c_faddr, ret);
766                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
767                 goto out;
768         }
769
770         if (unlikely(failed_wr != &send->s_wr)) {
771                 printk(KERN_WARNING "RDS/IB: atomic ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
772                 BUG_ON(failed_wr != &send->s_wr);
773         }
774
775 out:
776         return ret;
777 }
778
779 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
780 {
781         struct rds_ib_connection *ic = conn->c_transport_data;
782         struct rds_ib_send_work *send = NULL;
783         struct rds_ib_send_work *first;
784         struct rds_ib_send_work *prev;
785         struct ib_send_wr *failed_wr;
786         struct rds_ib_device *rds_ibdev;
787         struct scatterlist *scat;
788         unsigned long len;
789         u64 remote_addr = op->r_remote_addr;
790         u32 pos;
791         u32 work_alloc;
792         u32 i;
793         u32 j;
794         int sent;
795         int ret;
796         int num_sge;
797
798         rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
799
800         /* map the message the first time we see it */
801         if (!op->r_mapped) {
802                 op->r_count = ib_dma_map_sg(ic->i_cm_id->device,
803                                         op->r_sg, op->r_nents, (op->r_write) ?
804                                         DMA_TO_DEVICE : DMA_FROM_DEVICE);
805                 rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->r_count);
806                 if (op->r_count == 0) {
807                         rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
808                         ret = -ENOMEM; /* XXX ? */
809                         goto out;
810                 }
811
812                 op->r_mapped = 1;
813         }
814
815         /*
816          * Instead of knowing how to return a partial rdma read/write we insist that there
817          * be enough work requests to send the entire message.
818          */
819         i = ceil(op->r_count, rds_ibdev->max_sge);
820
821         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
822         if (work_alloc != i) {
823                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
824                 rds_ib_stats_inc(s_ib_tx_ring_full);
825                 ret = -ENOMEM;
826                 goto out;
827         }
828
829         send = &ic->i_sends[pos];
830         first = send;
831         prev = NULL;
832         scat = &op->r_sg[0];
833         sent = 0;
834         num_sge = op->r_count;
835
836         for (i = 0; i < work_alloc && scat != &op->r_sg[op->r_count]; i++) {
837                 send->s_wr.send_flags = 0;
838                 send->s_queued = jiffies;
839                 /*
840                  * We want to delay signaling completions just enough to get
841                  * the batching benefits but not so much that we create dead time on the wire.
842                  */
843                 if (ic->i_unsignaled_wrs-- == 0) {
844                         ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
845                         send->s_wr.send_flags = IB_SEND_SIGNALED;
846                 }
847
848                 send->s_wr.opcode = op->r_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ;
849                 send->s_wr.wr.rdma.remote_addr = remote_addr;
850                 send->s_wr.wr.rdma.rkey = op->r_key;
851                 send->s_op = op;
852
853                 if (num_sge > rds_ibdev->max_sge) {
854                         send->s_wr.num_sge = rds_ibdev->max_sge;
855                         num_sge -= rds_ibdev->max_sge;
856                 } else {
857                         send->s_wr.num_sge = num_sge;
858                 }
859
860                 send->s_wr.next = NULL;
861
862                 if (prev)
863                         prev->s_wr.next = &send->s_wr;
864
865                 for (j = 0; j < send->s_wr.num_sge && scat != &op->r_sg[op->r_count]; j++) {
866                         len = ib_sg_dma_len(ic->i_cm_id->device, scat);
867                         send->s_sge[j].addr =
868                                  ib_sg_dma_address(ic->i_cm_id->device, scat);
869                         send->s_sge[j].length = len;
870                         send->s_sge[j].lkey = ic->i_mr->lkey;
871
872                         sent += len;
873                         rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
874
875                         remote_addr += len;
876                         scat++;
877                 }
878
879                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
880                         &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
881
882                 prev = send;
883                 if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
884                         send = ic->i_sends;
885         }
886
887         /* if we finished the message then send completion owns it */
888         if (scat == &op->r_sg[op->r_count])
889                 prev->s_wr.send_flags = IB_SEND_SIGNALED;
890
891         if (i < work_alloc) {
892                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
893                 work_alloc = i;
894         }
895
896         failed_wr = &first->s_wr;
897         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
898         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
899                  first, &first->s_wr, ret, failed_wr);
900         BUG_ON(failed_wr != &first->s_wr);
901         if (ret) {
902                 printk(KERN_WARNING "RDS/IB: rdma ib_post_send to %pI4 "
903                        "returned %d\n", &conn->c_faddr, ret);
904                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
905                 goto out;
906         }
907
908         if (unlikely(failed_wr != &first->s_wr)) {
909                 printk(KERN_WARNING "RDS/IB: ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
910                 BUG_ON(failed_wr != &first->s_wr);
911         }
912
913
914 out:
915         return ret;
916 }
917
918 void rds_ib_xmit_complete(struct rds_connection *conn)
919 {
920         struct rds_ib_connection *ic = conn->c_transport_data;
921
922         /* We may have a pending ACK or window update we were unable
923          * to send previously (due to flow control). Try again. */
924         rds_ib_attempt_ack(ic);
925 }