657037d96cbf7ec3228ff6105306827061e8dd99
[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_data(struct rds_ib_connection *ic,
71                                    struct rm_data_op *op,
72                                    int wc_status)
73 {
74         if (op->op_nents)
75                 ib_dma_unmap_sg(ic->i_cm_id->device,
76                                 op->op_sg, op->op_nents,
77                                 DMA_TO_DEVICE);
78 }
79
80 static void rds_ib_send_unmap_rdma(struct rds_ib_connection *ic,
81                                    struct rm_rdma_op *op,
82                                    int wc_status)
83 {
84         if (op->op_mapped) {
85                 ib_dma_unmap_sg(ic->i_cm_id->device,
86                                 op->op_sg, op->op_nents,
87                                 op->op_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
88                 op->op_mapped = 0;
89         }
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          * we would need to take an event for the rdma WR. 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_complete(container_of(op, struct rds_message, rdma),
112                              wc_status, rds_rdma_send_complete);
113
114         if (op->op_write)
115                 rds_stats_add(s_send_rdma_bytes, op->op_bytes);
116         else
117                 rds_stats_add(s_recv_rdma_bytes, op->op_bytes);
118 }
119
120 static void rds_ib_send_unmap_atomic(struct rds_ib_connection *ic,
121                                      struct rm_atomic_op *op,
122                                      int wc_status)
123 {
124         /* unmap atomic recvbuf */
125         if (op->op_mapped) {
126                 ib_dma_unmap_sg(ic->i_cm_id->device, op->op_sg, 1,
127                                 DMA_FROM_DEVICE);
128                 op->op_mapped = 0;
129         }
130
131         rds_ib_send_complete(container_of(op, struct rds_message, atomic),
132                              wc_status, rds_atomic_send_complete);
133
134         if (op->op_type == RDS_ATOMIC_TYPE_CSWP)
135                 rds_ib_stats_inc(s_ib_atomic_cswp);
136         else
137                 rds_ib_stats_inc(s_ib_atomic_fadd);
138 }
139
140 /*
141  * Unmap the resources associated with a struct send_work.
142  *
143  * Returns the rm for no good reason other than it is unobtainable
144  * other than by switching on wr.opcode, currently, and the caller,
145  * the event handler, needs it.
146  */
147 static struct rds_message *rds_ib_send_unmap_op(struct rds_ib_connection *ic,
148                                                 struct rds_ib_send_work *send,
149                                                 int wc_status)
150 {
151         struct rds_message *rm = NULL;
152
153         /* In the error case, wc.opcode sometimes contains garbage */
154         switch (send->s_wr.opcode) {
155         case IB_WR_SEND:
156                 if (send->s_op) {
157                         rm = container_of(send->s_op, struct rds_message, data);
158                         rds_ib_send_unmap_data(ic, send->s_op, wc_status);
159                 }
160                 break;
161         case IB_WR_RDMA_WRITE:
162         case IB_WR_RDMA_READ:
163                 if (send->s_op) {
164                         rm = container_of(send->s_op, struct rds_message, rdma);
165                         rds_ib_send_unmap_rdma(ic, send->s_op, wc_status);
166                 }
167                 break;
168         case IB_WR_ATOMIC_FETCH_AND_ADD:
169         case IB_WR_ATOMIC_CMP_AND_SWP:
170                 if (send->s_op) {
171                         rm = container_of(send->s_op, struct rds_message, atomic);
172                         rds_ib_send_unmap_atomic(ic, send->s_op, wc_status);
173                 }
174                 break;
175         default:
176                 if (printk_ratelimit())
177                         printk(KERN_NOTICE
178                                "RDS/IB: %s: unexpected opcode 0x%x in WR!\n",
179                                __func__, send->s_wr.opcode);
180                 break;
181         }
182
183         send->s_wr.opcode = 0xdead;
184
185         return rm;
186 }
187
188 void rds_ib_send_init_ring(struct rds_ib_connection *ic)
189 {
190         struct rds_ib_send_work *send;
191         u32 i;
192
193         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
194                 struct ib_sge *sge;
195
196                 send->s_op = NULL;
197
198                 send->s_wr.wr_id = i;
199                 send->s_wr.sg_list = send->s_sge;
200                 send->s_wr.ex.imm_data = 0;
201
202                 sge = &send->s_sge[0];
203                 sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header));
204                 sge->length = sizeof(struct rds_header);
205                 sge->lkey = ic->i_mr->lkey;
206
207                 send->s_sge[1].lkey = ic->i_mr->lkey;
208         }
209 }
210
211 void rds_ib_send_clear_ring(struct rds_ib_connection *ic)
212 {
213         struct rds_ib_send_work *send;
214         u32 i;
215
216         for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
217                 if (send->s_op && send->s_wr.opcode != 0xdead)
218                         rds_ib_send_unmap_op(ic, send, IB_WC_WR_FLUSH_ERR);
219         }
220 }
221
222 /*
223  * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
224  * operations performed in the send path.  As the sender allocs and potentially
225  * unallocs the next free entry in the ring it doesn't alter which is
226  * the next to be freed, which is what this is concerned with.
227  */
228 void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
229 {
230         struct rds_connection *conn = context;
231         struct rds_ib_connection *ic = conn->c_transport_data;
232         struct rds_message *rm = NULL;
233         struct ib_wc wc;
234         struct rds_ib_send_work *send;
235         u32 completed;
236         u32 oldest;
237         u32 i = 0;
238         int ret;
239
240         rdsdebug("cq %p conn %p\n", cq, conn);
241         rds_ib_stats_inc(s_ib_tx_cq_call);
242         ret = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
243         if (ret)
244                 rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
245
246         while (ib_poll_cq(cq, 1, &wc) > 0) {
247                 rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
248                          (unsigned long long)wc.wr_id, wc.status, wc.byte_len,
249                          be32_to_cpu(wc.ex.imm_data));
250                 rds_ib_stats_inc(s_ib_tx_cq_event);
251
252                 if (wc.wr_id == RDS_IB_ACK_WR_ID) {
253                         if (ic->i_ack_queued + HZ/2 < jiffies)
254                                 rds_ib_stats_inc(s_ib_tx_stalled);
255                         rds_ib_ack_send_complete(ic);
256                         continue;
257                 }
258
259                 oldest = rds_ib_ring_oldest(&ic->i_send_ring);
260
261                 completed = rds_ib_ring_completed(&ic->i_send_ring, wc.wr_id, oldest);
262
263                 for (i = 0; i < completed; i++) {
264                         send = &ic->i_sends[oldest];
265
266                         rm = rds_ib_send_unmap_op(ic, send, wc.status);
267
268                         if (send->s_queued + HZ/2 < jiffies)
269                                 rds_ib_stats_inc(s_ib_tx_stalled);
270
271                         if (&send->s_op == &rm->m_final_op) {
272                                 /* If anyone waited for this message to get flushed out, wake
273                                  * them up now */
274                                 rds_message_unmapped(rm);
275
276                                 rds_message_put(rm);
277                                 send->s_op = NULL;
278                         }
279
280                         oldest = (oldest + 1) % ic->i_send_ring.w_nr;
281                 }
282
283                 rds_ib_ring_free(&ic->i_send_ring, completed);
284
285                 if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
286                     test_bit(0, &conn->c_map_queued))
287                         queue_delayed_work(rds_wq, &conn->c_send_w, 0);
288
289                 /* We expect errors as the qp is drained during shutdown */
290                 if (wc.status != IB_WC_SUCCESS && rds_conn_up(conn)) {
291                         rds_ib_conn_error(conn,
292                                 "send completion on %pI4 "
293                                 "had status %u, disconnecting and reconnecting\n",
294                                 &conn->c_faddr, wc.status);
295                 }
296         }
297 }
298
299 /*
300  * This is the main function for allocating credits when sending
301  * messages.
302  *
303  * Conceptually, we have two counters:
304  *  -   send credits: this tells us how many WRs we're allowed
305  *      to submit without overruning the reciever's queue. For
306  *      each SEND WR we post, we decrement this by one.
307  *
308  *  -   posted credits: this tells us how many WRs we recently
309  *      posted to the receive queue. This value is transferred
310  *      to the peer as a "credit update" in a RDS header field.
311  *      Every time we transmit credits to the peer, we subtract
312  *      the amount of transferred credits from this counter.
313  *
314  * It is essential that we avoid situations where both sides have
315  * exhausted their send credits, and are unable to send new credits
316  * to the peer. We achieve this by requiring that we send at least
317  * one credit update to the peer before exhausting our credits.
318  * When new credits arrive, we subtract one credit that is withheld
319  * until we've posted new buffers and are ready to transmit these
320  * credits (see rds_ib_send_add_credits below).
321  *
322  * The RDS send code is essentially single-threaded; rds_send_xmit
323  * grabs c_send_lock to ensure exclusive access to the send ring.
324  * However, the ACK sending code is independent and can race with
325  * message SENDs.
326  *
327  * In the send path, we need to update the counters for send credits
328  * and the counter of posted buffers atomically - when we use the
329  * last available credit, we cannot allow another thread to race us
330  * and grab the posted credits counter.  Hence, we have to use a
331  * spinlock to protect the credit counter, or use atomics.
332  *
333  * Spinlocks shared between the send and the receive path are bad,
334  * because they create unnecessary delays. An early implementation
335  * using a spinlock showed a 5% degradation in throughput at some
336  * loads.
337  *
338  * This implementation avoids spinlocks completely, putting both
339  * counters into a single atomic, and updating that atomic using
340  * atomic_add (in the receive path, when receiving fresh credits),
341  * and using atomic_cmpxchg when updating the two counters.
342  */
343 int rds_ib_send_grab_credits(struct rds_ib_connection *ic,
344                              u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
345 {
346         unsigned int avail, posted, got = 0, advertise;
347         long oldval, newval;
348
349         *adv_credits = 0;
350         if (!ic->i_flowctl)
351                 return wanted;
352
353 try_again:
354         advertise = 0;
355         oldval = newval = atomic_read(&ic->i_credits);
356         posted = IB_GET_POST_CREDITS(oldval);
357         avail = IB_GET_SEND_CREDITS(oldval);
358
359         rdsdebug("rds_ib_send_grab_credits(%u): credits=%u posted=%u\n",
360                         wanted, avail, posted);
361
362         /* The last credit must be used to send a credit update. */
363         if (avail && !posted)
364                 avail--;
365
366         if (avail < wanted) {
367                 struct rds_connection *conn = ic->i_cm_id->context;
368
369                 /* Oops, there aren't that many credits left! */
370                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
371                 got = avail;
372         } else {
373                 /* Sometimes you get what you want, lalala. */
374                 got = wanted;
375         }
376         newval -= IB_SET_SEND_CREDITS(got);
377
378         /*
379          * If need_posted is non-zero, then the caller wants
380          * the posted regardless of whether any send credits are
381          * available.
382          */
383         if (posted && (got || need_posted)) {
384                 advertise = min_t(unsigned int, posted, max_posted);
385                 newval -= IB_SET_POST_CREDITS(advertise);
386         }
387
388         /* Finally bill everything */
389         if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
390                 goto try_again;
391
392         *adv_credits = advertise;
393         return got;
394 }
395
396 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits)
397 {
398         struct rds_ib_connection *ic = conn->c_transport_data;
399
400         if (credits == 0)
401                 return;
402
403         rdsdebug("rds_ib_send_add_credits(%u): current=%u%s\n",
404                         credits,
405                         IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
406                         test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
407
408         atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
409         if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
410                 queue_delayed_work(rds_wq, &conn->c_send_w, 0);
411
412         WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
413
414         rds_ib_stats_inc(s_ib_rx_credit_updates);
415 }
416
417 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted)
418 {
419         struct rds_ib_connection *ic = conn->c_transport_data;
420
421         if (posted == 0)
422                 return;
423
424         atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
425
426         /* Decide whether to send an update to the peer now.
427          * If we would send a credit update for every single buffer we
428          * post, we would end up with an ACK storm (ACK arrives,
429          * consumes buffer, we refill the ring, send ACK to remote
430          * advertising the newly posted buffer... ad inf)
431          *
432          * Performance pretty much depends on how often we send
433          * credit updates - too frequent updates mean lots of ACKs.
434          * Too infrequent updates, and the peer will run out of
435          * credits and has to throttle.
436          * For the time being, 16 seems to be a good compromise.
437          */
438         if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
439                 set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
440 }
441
442 static inline void rds_ib_set_wr_signal_state(struct rds_ib_connection *ic,
443                                               struct rds_ib_send_work *send,
444                                               bool notify)
445 {
446         /*
447          * We want to delay signaling completions just enough to get
448          * the batching benefits but not so much that we create dead time
449          * on the wire.
450          */
451         if (ic->i_unsignaled_wrs-- == 0 || notify) {
452                 ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
453                 send->s_wr.send_flags |= IB_SEND_SIGNALED;
454         }
455 }
456
457 /*
458  * This can be called multiple times for a given message.  The first time
459  * we see a message we map its scatterlist into the IB device so that
460  * we can provide that mapped address to the IB scatter gather entries
461  * in the IB work requests.  We translate the scatterlist into a series
462  * of work requests that fragment the message.  These work requests complete
463  * in order so we pass ownership of the message to the completion handler
464  * once we send the final fragment.
465  *
466  * The RDS core uses the c_send_lock to only enter this function once
467  * per connection.  This makes sure that the tx ring alloc/unalloc pairs
468  * don't get out of sync and confuse the ring.
469  */
470 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
471                 unsigned int hdr_off, unsigned int sg, unsigned int off)
472 {
473         struct rds_ib_connection *ic = conn->c_transport_data;
474         struct ib_device *dev = ic->i_cm_id->device;
475         struct rds_ib_send_work *send = NULL;
476         struct rds_ib_send_work *first;
477         struct rds_ib_send_work *prev;
478         struct ib_send_wr *failed_wr;
479         struct scatterlist *scat;
480         u32 pos;
481         u32 i;
482         u32 work_alloc;
483         u32 credit_alloc = 0;
484         u32 posted;
485         u32 adv_credits = 0;
486         int send_flags = 0;
487         int bytes_sent = 0;
488         int ret;
489         int flow_controlled = 0;
490
491         BUG_ON(off % RDS_FRAG_SIZE);
492         BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
493
494         /* Do not send cong updates to IB loopback */
495         if (conn->c_loopback
496             && rm->m_inc.i_hdr.h_flags & RDS_FLAG_CONG_BITMAP) {
497                 rds_cong_map_updated(conn->c_fcong, ~(u64) 0);
498                 return sizeof(struct rds_header) + RDS_CONG_MAP_BYTES;
499         }
500
501         /* FIXME we may overallocate here */
502         if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
503                 i = 1;
504         else
505                 i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
506
507         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
508         if (work_alloc == 0) {
509                 set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
510                 rds_ib_stats_inc(s_ib_tx_ring_full);
511                 ret = -ENOMEM;
512                 goto out;
513         }
514
515         if (ic->i_flowctl) {
516                 credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
517                 adv_credits += posted;
518                 if (credit_alloc < work_alloc) {
519                         rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
520                         work_alloc = credit_alloc;
521                         flow_controlled = 1;
522                 }
523                 if (work_alloc == 0) {
524                         set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
525                         rds_ib_stats_inc(s_ib_tx_throttle);
526                         ret = -ENOMEM;
527                         goto out;
528                 }
529         }
530
531         /* map the message the first time we see it */
532         if (!ic->i_data_op) {
533                 if (rm->data.op_nents) {
534                         rm->data.op_count = ib_dma_map_sg(dev,
535                                                           rm->data.op_sg,
536                                                           rm->data.op_nents,
537                                                           DMA_TO_DEVICE);
538                         rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.op_count);
539                         if (rm->data.op_count == 0) {
540                                 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
541                                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
542                                 ret = -ENOMEM; /* XXX ? */
543                                 goto out;
544                         }
545                 } else {
546                         rm->data.op_count = 0;
547                 }
548
549                 rds_message_addref(rm);
550                 ic->i_data_op = &rm->data;
551
552                 /* Finalize the header */
553                 if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
554                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
555                 if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
556                         rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
557
558                 /* If it has a RDMA op, tell the peer we did it. This is
559                  * used by the peer to release use-once RDMA MRs. */
560                 if (rm->rdma.op_active) {
561                         struct rds_ext_header_rdma ext_hdr;
562
563                         ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.op_rkey);
564                         rds_message_add_extension(&rm->m_inc.i_hdr,
565                                         RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
566                 }
567                 if (rm->m_rdma_cookie) {
568                         rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
569                                         rds_rdma_cookie_key(rm->m_rdma_cookie),
570                                         rds_rdma_cookie_offset(rm->m_rdma_cookie));
571                 }
572
573                 /* Note - rds_ib_piggyb_ack clears the ACK_REQUIRED bit, so
574                  * we should not do this unless we have a chance of at least
575                  * sticking the header into the send ring. Which is why we
576                  * should call rds_ib_ring_alloc first. */
577                 rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_ib_piggyb_ack(ic));
578                 rds_message_make_checksum(&rm->m_inc.i_hdr);
579
580                 /*
581                  * Update adv_credits since we reset the ACK_REQUIRED bit.
582                  */
583                 if (ic->i_flowctl) {
584                         rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
585                         adv_credits += posted;
586                         BUG_ON(adv_credits > 255);
587                 }
588         }
589
590         /* Sometimes you want to put a fence between an RDMA
591          * READ and the following SEND.
592          * We could either do this all the time
593          * or when requested by the user. Right now, we let
594          * the application choose.
595          */
596         if (rm->rdma.op_active && rm->rdma.op_fence)
597                 send_flags = IB_SEND_FENCE;
598
599         /* Each frag gets a header. Msgs may be 0 bytes */
600         send = &ic->i_sends[pos];
601         first = send;
602         prev = NULL;
603         scat = &ic->i_data_op->op_sg[sg];
604         i = 0;
605         do {
606                 unsigned int len = 0;
607
608                 /* Set up the header */
609                 send->s_wr.send_flags = send_flags;
610                 send->s_wr.opcode = IB_WR_SEND;
611                 send->s_wr.num_sge = 1;
612                 send->s_wr.next = NULL;
613                 send->s_queued = jiffies;
614                 send->s_op = NULL;
615
616                 send->s_sge[0].addr = ic->i_send_hdrs_dma
617                         + (pos * sizeof(struct rds_header));
618                 send->s_sge[0].length = sizeof(struct rds_header);
619
620                 memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
621
622                 /* Set up the data, if present */
623                 if (i < work_alloc
624                     && scat != &rm->data.op_sg[rm->data.op_count]) {
625                         len = min(RDS_FRAG_SIZE, ib_sg_dma_len(dev, scat) - off);
626                         send->s_wr.num_sge = 2;
627
628                         send->s_sge[1].addr = ib_sg_dma_address(dev, scat) + off;
629                         send->s_sge[1].length = len;
630
631                         bytes_sent += len;
632                         off += len;
633                         if (off == ib_sg_dma_len(dev, scat)) {
634                                 scat++;
635                                 off = 0;
636                         }
637                 }
638
639                 rds_ib_set_wr_signal_state(ic, send, 0);
640
641                 /*
642                  * Always signal the last one if we're stopping due to flow control.
643                  */
644                 if (ic->i_flowctl && flow_controlled && i == (work_alloc-1))
645                         send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
646
647                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
648                          &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
649
650                 if (ic->i_flowctl && adv_credits) {
651                         struct rds_header *hdr = &ic->i_send_hdrs[pos];
652
653                         /* add credit and redo the header checksum */
654                         hdr->h_credit = adv_credits;
655                         rds_message_make_checksum(hdr);
656                         adv_credits = 0;
657                         rds_ib_stats_inc(s_ib_tx_credit_updates);
658                 }
659
660                 if (prev)
661                         prev->s_wr.next = &send->s_wr;
662                 prev = send;
663
664                 pos = (pos + 1) % ic->i_send_ring.w_nr;
665                 send = &ic->i_sends[pos];
666                 i++;
667
668         } while (i < work_alloc
669                  && scat != &rm->data.op_sg[rm->data.op_count]);
670
671         /* Account the RDS header in the number of bytes we sent, but just once.
672          * The caller has no concept of fragmentation. */
673         if (hdr_off == 0)
674                 bytes_sent += sizeof(struct rds_header);
675
676         /* if we finished the message then send completion owns it */
677         if (scat == &rm->data.op_sg[rm->data.op_count]) {
678                 prev->s_op = ic->i_data_op;
679                 prev->s_wr.send_flags |= IB_SEND_SOLICITED;
680                 ic->i_data_op = NULL;
681         }
682
683         /* Put back wrs & credits we didn't use */
684         if (i < work_alloc) {
685                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
686                 work_alloc = i;
687         }
688         if (ic->i_flowctl && i < credit_alloc)
689                 rds_ib_send_add_credits(conn, credit_alloc - i);
690
691         /* XXX need to worry about failed_wr and partial sends. */
692         failed_wr = &first->s_wr;
693         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
694         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
695                  first, &first->s_wr, ret, failed_wr);
696         BUG_ON(failed_wr != &first->s_wr);
697         if (ret) {
698                 printk(KERN_WARNING "RDS/IB: ib_post_send to %pI4 "
699                        "returned %d\n", &conn->c_faddr, ret);
700                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
701                 if (prev->s_op) {
702                         ic->i_data_op = prev->s_op;
703                         prev->s_op = NULL;
704                 }
705
706                 rds_ib_conn_error(ic->conn, "ib_post_send failed\n");
707                 goto out;
708         }
709
710         ret = bytes_sent;
711 out:
712         BUG_ON(adv_credits);
713         return ret;
714 }
715
716 /*
717  * Issue atomic operation.
718  * A simplified version of the rdma case, we always map 1 SG, and
719  * only 8 bytes, for the return value from the atomic operation.
720  */
721 int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op)
722 {
723         struct rds_ib_connection *ic = conn->c_transport_data;
724         struct rds_ib_send_work *send = NULL;
725         struct ib_send_wr *failed_wr;
726         struct rds_ib_device *rds_ibdev;
727         u32 pos;
728         u32 work_alloc;
729         int ret;
730
731         rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
732
733         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, 1, &pos);
734         if (work_alloc != 1) {
735                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
736                 rds_ib_stats_inc(s_ib_tx_ring_full);
737                 ret = -ENOMEM;
738                 goto out;
739         }
740
741         /* address of send request in ring */
742         send = &ic->i_sends[pos];
743         send->s_queued = jiffies;
744
745         if (op->op_type == RDS_ATOMIC_TYPE_CSWP) {
746                 send->s_wr.opcode = IB_WR_ATOMIC_CMP_AND_SWP;
747                 send->s_wr.wr.atomic.compare_add = op->op_compare;
748                 send->s_wr.wr.atomic.swap = op->op_swap_add;
749         } else { /* FADD */
750                 send->s_wr.opcode = IB_WR_ATOMIC_FETCH_AND_ADD;
751                 send->s_wr.wr.atomic.compare_add = op->op_swap_add;
752                 send->s_wr.wr.atomic.swap = 0;
753         }
754         rds_ib_set_wr_signal_state(ic, send, op->op_notify);
755         send->s_wr.num_sge = 1;
756         send->s_wr.next = NULL;
757         send->s_wr.wr.atomic.remote_addr = op->op_remote_addr;
758         send->s_wr.wr.atomic.rkey = op->op_rkey;
759
760         /* map 8 byte retval buffer to the device */
761         ret = ib_dma_map_sg(ic->i_cm_id->device, op->op_sg, 1, DMA_FROM_DEVICE);
762         rdsdebug("ic %p mapping atomic op %p. mapped %d pg\n", ic, op, ret);
763         if (ret != 1) {
764                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
765                 rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
766                 ret = -ENOMEM; /* XXX ? */
767                 goto out;
768         }
769
770         /* Convert our struct scatterlist to struct ib_sge */
771         send->s_sge[0].addr = ib_sg_dma_address(ic->i_cm_id->device, op->op_sg);
772         send->s_sge[0].length = ib_sg_dma_len(ic->i_cm_id->device, op->op_sg);
773         send->s_sge[0].lkey = ic->i_mr->lkey;
774
775         rdsdebug("rva %Lx rpa %Lx len %u\n", op->op_remote_addr,
776                  send->s_sge[0].addr, send->s_sge[0].length);
777
778         failed_wr = &send->s_wr;
779         ret = ib_post_send(ic->i_cm_id->qp, &send->s_wr, &failed_wr);
780         rdsdebug("ic %p send %p (wr %p) ret %d wr %p\n", ic,
781                  send, &send->s_wr, ret, failed_wr);
782         BUG_ON(failed_wr != &send->s_wr);
783         if (ret) {
784                 printk(KERN_WARNING "RDS/IB: atomic ib_post_send to %pI4 "
785                        "returned %d\n", &conn->c_faddr, ret);
786                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
787                 goto out;
788         }
789
790         if (unlikely(failed_wr != &send->s_wr)) {
791                 printk(KERN_WARNING "RDS/IB: atomic ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
792                 BUG_ON(failed_wr != &send->s_wr);
793         }
794
795 out:
796         return ret;
797 }
798
799 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op)
800 {
801         struct rds_ib_connection *ic = conn->c_transport_data;
802         struct rds_ib_send_work *send = NULL;
803         struct rds_ib_send_work *first;
804         struct rds_ib_send_work *prev;
805         struct ib_send_wr *failed_wr;
806         struct rds_ib_device *rds_ibdev;
807         struct scatterlist *scat;
808         unsigned long len;
809         u64 remote_addr = op->op_remote_addr;
810         u32 pos;
811         u32 work_alloc;
812         u32 i;
813         u32 j;
814         int sent;
815         int ret;
816         int num_sge;
817
818         rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
819
820         /* map the op the first time we see it */
821         if (!op->op_mapped) {
822                 op->op_count = ib_dma_map_sg(ic->i_cm_id->device,
823                                              op->op_sg, op->op_nents, (op->op_write) ?
824                                              DMA_TO_DEVICE : DMA_FROM_DEVICE);
825                 rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->op_count);
826                 if (op->op_count == 0) {
827                         rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
828                         ret = -ENOMEM; /* XXX ? */
829                         goto out;
830                 }
831
832                 op->op_mapped = 1;
833         }
834
835         /*
836          * Instead of knowing how to return a partial rdma read/write we insist that there
837          * be enough work requests to send the entire message.
838          */
839         i = ceil(op->op_count, rds_ibdev->max_sge);
840
841         work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
842         if (work_alloc != i) {
843                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
844                 rds_ib_stats_inc(s_ib_tx_ring_full);
845                 ret = -ENOMEM;
846                 goto out;
847         }
848
849         send = &ic->i_sends[pos];
850         first = send;
851         prev = NULL;
852         scat = &op->op_sg[0];
853         sent = 0;
854         num_sge = op->op_count;
855
856         for (i = 0; i < work_alloc && scat != &op->op_sg[op->op_count]; i++) {
857                 send->s_wr.send_flags = 0;
858                 send->s_queued = jiffies;
859
860                 rds_ib_set_wr_signal_state(ic, send, op->op_notify);
861
862                 send->s_wr.opcode = op->op_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ;
863                 send->s_wr.wr.rdma.remote_addr = remote_addr;
864                 send->s_wr.wr.rdma.rkey = op->op_rkey;
865                 send->s_op = op;
866
867                 if (num_sge > rds_ibdev->max_sge) {
868                         send->s_wr.num_sge = rds_ibdev->max_sge;
869                         num_sge -= rds_ibdev->max_sge;
870                 } else {
871                         send->s_wr.num_sge = num_sge;
872                 }
873
874                 send->s_wr.next = NULL;
875
876                 if (prev)
877                         prev->s_wr.next = &send->s_wr;
878
879                 for (j = 0; j < send->s_wr.num_sge && scat != &op->op_sg[op->op_count]; j++) {
880                         len = ib_sg_dma_len(ic->i_cm_id->device, scat);
881                         send->s_sge[j].addr =
882                                  ib_sg_dma_address(ic->i_cm_id->device, scat);
883                         send->s_sge[j].length = len;
884                         send->s_sge[j].lkey = ic->i_mr->lkey;
885
886                         sent += len;
887                         rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
888
889                         remote_addr += len;
890                         scat++;
891                 }
892
893                 rdsdebug("send %p wr %p num_sge %u next %p\n", send,
894                         &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
895
896                 prev = send;
897                 if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
898                         send = ic->i_sends;
899         }
900
901         if (i < work_alloc) {
902                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
903                 work_alloc = i;
904         }
905
906         failed_wr = &first->s_wr;
907         ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
908         rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
909                  first, &first->s_wr, ret, failed_wr);
910         BUG_ON(failed_wr != &first->s_wr);
911         if (ret) {
912                 printk(KERN_WARNING "RDS/IB: rdma ib_post_send to %pI4 "
913                        "returned %d\n", &conn->c_faddr, ret);
914                 rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
915                 goto out;
916         }
917
918         if (unlikely(failed_wr != &first->s_wr)) {
919                 printk(KERN_WARNING "RDS/IB: ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
920                 BUG_ON(failed_wr != &first->s_wr);
921         }
922
923
924 out:
925         return ret;
926 }
927
928 void rds_ib_xmit_complete(struct rds_connection *conn)
929 {
930         struct rds_ib_connection *ic = conn->c_transport_data;
931
932         /* We may have a pending ACK or window update we were unable
933          * to send previously (due to flow control). Try again. */
934         rds_ib_attempt_ack(ic);
935 }