xref: /linux/net/rds/ib_send.c (revision a3a4a816b4b194c45d0217e8b9e08b2639802cda)
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 #include <linux/ratelimit.h>
38 
39 #include "rds_single_path.h"
40 #include "rds.h"
41 #include "ib.h"
42 
43 /*
44  * Convert IB-specific error message to RDS error message and call core
45  * completion handler.
46  */
47 static void rds_ib_send_complete(struct rds_message *rm,
48 				 int wc_status,
49 				 void (*complete)(struct rds_message *rm, int status))
50 {
51 	int notify_status;
52 
53 	switch (wc_status) {
54 	case IB_WC_WR_FLUSH_ERR:
55 		return;
56 
57 	case IB_WC_SUCCESS:
58 		notify_status = RDS_RDMA_SUCCESS;
59 		break;
60 
61 	case IB_WC_REM_ACCESS_ERR:
62 		notify_status = RDS_RDMA_REMOTE_ERROR;
63 		break;
64 
65 	default:
66 		notify_status = RDS_RDMA_OTHER_ERROR;
67 		break;
68 	}
69 	complete(rm, notify_status);
70 }
71 
72 static void rds_ib_send_unmap_rdma(struct rds_ib_connection *ic,
73 				   struct rm_rdma_op *op,
74 				   int wc_status)
75 {
76 	if (op->op_mapped) {
77 		ib_dma_unmap_sg(ic->i_cm_id->device,
78 				op->op_sg, op->op_nents,
79 				op->op_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
80 		op->op_mapped = 0;
81 	}
82 
83 	/* If the user asked for a completion notification on this
84 	 * message, we can implement three different semantics:
85 	 *  1.	Notify when we received the ACK on the RDS message
86 	 *	that was queued with the RDMA. This provides reliable
87 	 *	notification of RDMA status at the expense of a one-way
88 	 *	packet delay.
89 	 *  2.	Notify when the IB stack gives us the completion event for
90 	 *	the RDMA operation.
91 	 *  3.	Notify when the IB stack gives us the completion event for
92 	 *	the accompanying RDS messages.
93 	 * Here, we implement approach #3. To implement approach #2,
94 	 * we would need to take an event for the rdma WR. To implement #1,
95 	 * don't call rds_rdma_send_complete at all, and fall back to the notify
96 	 * handling in the ACK processing code.
97 	 *
98 	 * Note: There's no need to explicitly sync any RDMA buffers using
99 	 * ib_dma_sync_sg_for_cpu - the completion for the RDMA
100 	 * operation itself unmapped the RDMA buffers, which takes care
101 	 * of synching.
102 	 */
103 	rds_ib_send_complete(container_of(op, struct rds_message, rdma),
104 			     wc_status, rds_rdma_send_complete);
105 
106 	if (op->op_write)
107 		rds_stats_add(s_send_rdma_bytes, op->op_bytes);
108 	else
109 		rds_stats_add(s_recv_rdma_bytes, op->op_bytes);
110 }
111 
112 static void rds_ib_send_unmap_atomic(struct rds_ib_connection *ic,
113 				     struct rm_atomic_op *op,
114 				     int wc_status)
115 {
116 	/* unmap atomic recvbuf */
117 	if (op->op_mapped) {
118 		ib_dma_unmap_sg(ic->i_cm_id->device, op->op_sg, 1,
119 				DMA_FROM_DEVICE);
120 		op->op_mapped = 0;
121 	}
122 
123 	rds_ib_send_complete(container_of(op, struct rds_message, atomic),
124 			     wc_status, rds_atomic_send_complete);
125 
126 	if (op->op_type == RDS_ATOMIC_TYPE_CSWP)
127 		rds_ib_stats_inc(s_ib_atomic_cswp);
128 	else
129 		rds_ib_stats_inc(s_ib_atomic_fadd);
130 }
131 
132 static void rds_ib_send_unmap_data(struct rds_ib_connection *ic,
133 				   struct rm_data_op *op,
134 				   int wc_status)
135 {
136 	struct rds_message *rm = container_of(op, struct rds_message, data);
137 
138 	if (op->op_nents)
139 		ib_dma_unmap_sg(ic->i_cm_id->device,
140 				op->op_sg, op->op_nents,
141 				DMA_TO_DEVICE);
142 
143 	if (rm->rdma.op_active && rm->data.op_notify)
144 		rds_ib_send_unmap_rdma(ic, &rm->rdma, wc_status);
145 }
146 
147 /*
148  * Unmap the resources associated with a struct send_work.
149  *
150  * Returns the rm for no good reason other than it is unobtainable
151  * other than by switching on wr.opcode, currently, and the caller,
152  * the event handler, needs it.
153  */
154 static struct rds_message *rds_ib_send_unmap_op(struct rds_ib_connection *ic,
155 						struct rds_ib_send_work *send,
156 						int wc_status)
157 {
158 	struct rds_message *rm = NULL;
159 
160 	/* In the error case, wc.opcode sometimes contains garbage */
161 	switch (send->s_wr.opcode) {
162 	case IB_WR_SEND:
163 		if (send->s_op) {
164 			rm = container_of(send->s_op, struct rds_message, data);
165 			rds_ib_send_unmap_data(ic, send->s_op, wc_status);
166 		}
167 		break;
168 	case IB_WR_RDMA_WRITE:
169 	case IB_WR_RDMA_READ:
170 		if (send->s_op) {
171 			rm = container_of(send->s_op, struct rds_message, rdma);
172 			rds_ib_send_unmap_rdma(ic, send->s_op, wc_status);
173 		}
174 		break;
175 	case IB_WR_ATOMIC_FETCH_AND_ADD:
176 	case IB_WR_ATOMIC_CMP_AND_SWP:
177 		if (send->s_op) {
178 			rm = container_of(send->s_op, struct rds_message, atomic);
179 			rds_ib_send_unmap_atomic(ic, send->s_op, wc_status);
180 		}
181 		break;
182 	default:
183 		printk_ratelimited(KERN_NOTICE
184 			       "RDS/IB: %s: unexpected opcode 0x%x in WR!\n",
185 			       __func__, send->s_wr.opcode);
186 		break;
187 	}
188 
189 	send->s_wr.opcode = 0xdead;
190 
191 	return rm;
192 }
193 
194 void rds_ib_send_init_ring(struct rds_ib_connection *ic)
195 {
196 	struct rds_ib_send_work *send;
197 	u32 i;
198 
199 	for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
200 		struct ib_sge *sge;
201 
202 		send->s_op = NULL;
203 
204 		send->s_wr.wr_id = i;
205 		send->s_wr.sg_list = send->s_sge;
206 		send->s_wr.ex.imm_data = 0;
207 
208 		sge = &send->s_sge[0];
209 		sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header));
210 		sge->length = sizeof(struct rds_header);
211 		sge->lkey = ic->i_pd->local_dma_lkey;
212 
213 		send->s_sge[1].lkey = ic->i_pd->local_dma_lkey;
214 	}
215 }
216 
217 void rds_ib_send_clear_ring(struct rds_ib_connection *ic)
218 {
219 	struct rds_ib_send_work *send;
220 	u32 i;
221 
222 	for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
223 		if (send->s_op && send->s_wr.opcode != 0xdead)
224 			rds_ib_send_unmap_op(ic, send, IB_WC_WR_FLUSH_ERR);
225 	}
226 }
227 
228 /*
229  * The only fast path caller always has a non-zero nr, so we don't
230  * bother testing nr before performing the atomic sub.
231  */
232 static void rds_ib_sub_signaled(struct rds_ib_connection *ic, int nr)
233 {
234 	if ((atomic_sub_return(nr, &ic->i_signaled_sends) == 0) &&
235 	    waitqueue_active(&rds_ib_ring_empty_wait))
236 		wake_up(&rds_ib_ring_empty_wait);
237 	BUG_ON(atomic_read(&ic->i_signaled_sends) < 0);
238 }
239 
240 /*
241  * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
242  * operations performed in the send path.  As the sender allocs and potentially
243  * unallocs the next free entry in the ring it doesn't alter which is
244  * the next to be freed, which is what this is concerned with.
245  */
246 void rds_ib_send_cqe_handler(struct rds_ib_connection *ic, struct ib_wc *wc)
247 {
248 	struct rds_message *rm = NULL;
249 	struct rds_connection *conn = ic->conn;
250 	struct rds_ib_send_work *send;
251 	u32 completed;
252 	u32 oldest;
253 	u32 i = 0;
254 	int nr_sig = 0;
255 
256 
257 	rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
258 		 (unsigned long long)wc->wr_id, wc->status,
259 		 ib_wc_status_msg(wc->status), wc->byte_len,
260 		 be32_to_cpu(wc->ex.imm_data));
261 	rds_ib_stats_inc(s_ib_tx_cq_event);
262 
263 	if (wc->wr_id == RDS_IB_ACK_WR_ID) {
264 		if (time_after(jiffies, ic->i_ack_queued + HZ / 2))
265 			rds_ib_stats_inc(s_ib_tx_stalled);
266 		rds_ib_ack_send_complete(ic);
267 		return;
268 	}
269 
270 	oldest = rds_ib_ring_oldest(&ic->i_send_ring);
271 
272 	completed = rds_ib_ring_completed(&ic->i_send_ring, wc->wr_id, oldest);
273 
274 	for (i = 0; i < completed; i++) {
275 		send = &ic->i_sends[oldest];
276 		if (send->s_wr.send_flags & IB_SEND_SIGNALED)
277 			nr_sig++;
278 
279 		rm = rds_ib_send_unmap_op(ic, send, wc->status);
280 
281 		if (time_after(jiffies, send->s_queued + HZ / 2))
282 			rds_ib_stats_inc(s_ib_tx_stalled);
283 
284 		if (send->s_op) {
285 			if (send->s_op == rm->m_final_op) {
286 				/* If anyone waited for this message to get
287 				 * flushed out, wake them up now
288 				 */
289 				rds_message_unmapped(rm);
290 			}
291 			rds_message_put(rm);
292 			send->s_op = NULL;
293 		}
294 
295 		oldest = (oldest + 1) % ic->i_send_ring.w_nr;
296 	}
297 
298 	rds_ib_ring_free(&ic->i_send_ring, completed);
299 	rds_ib_sub_signaled(ic, nr_sig);
300 	nr_sig = 0;
301 
302 	if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
303 	    test_bit(0, &conn->c_map_queued))
304 		queue_delayed_work(rds_wq, &conn->c_send_w, 0);
305 
306 	/* We expect errors as the qp is drained during shutdown */
307 	if (wc->status != IB_WC_SUCCESS && rds_conn_up(conn)) {
308 		rds_ib_conn_error(conn, "send completion on <%pI4,%pI4> had status %u (%s), disconnecting and reconnecting\n",
309 				  &conn->c_laddr, &conn->c_faddr, wc->status,
310 				  ib_wc_status_msg(wc->status));
311 	}
312 }
313 
314 /*
315  * This is the main function for allocating credits when sending
316  * messages.
317  *
318  * Conceptually, we have two counters:
319  *  -	send credits: this tells us how many WRs we're allowed
320  *	to submit without overruning the receiver's queue. For
321  *	each SEND WR we post, we decrement this by one.
322  *
323  *  -	posted credits: this tells us how many WRs we recently
324  *	posted to the receive queue. This value is transferred
325  *	to the peer as a "credit update" in a RDS header field.
326  *	Every time we transmit credits to the peer, we subtract
327  *	the amount of transferred credits from this counter.
328  *
329  * It is essential that we avoid situations where both sides have
330  * exhausted their send credits, and are unable to send new credits
331  * to the peer. We achieve this by requiring that we send at least
332  * one credit update to the peer before exhausting our credits.
333  * When new credits arrive, we subtract one credit that is withheld
334  * until we've posted new buffers and are ready to transmit these
335  * credits (see rds_ib_send_add_credits below).
336  *
337  * The RDS send code is essentially single-threaded; rds_send_xmit
338  * sets RDS_IN_XMIT to ensure exclusive access to the send ring.
339  * However, the ACK sending code is independent and can race with
340  * message SENDs.
341  *
342  * In the send path, we need to update the counters for send credits
343  * and the counter of posted buffers atomically - when we use the
344  * last available credit, we cannot allow another thread to race us
345  * and grab the posted credits counter.  Hence, we have to use a
346  * spinlock to protect the credit counter, or use atomics.
347  *
348  * Spinlocks shared between the send and the receive path are bad,
349  * because they create unnecessary delays. An early implementation
350  * using a spinlock showed a 5% degradation in throughput at some
351  * loads.
352  *
353  * This implementation avoids spinlocks completely, putting both
354  * counters into a single atomic, and updating that atomic using
355  * atomic_add (in the receive path, when receiving fresh credits),
356  * and using atomic_cmpxchg when updating the two counters.
357  */
358 int rds_ib_send_grab_credits(struct rds_ib_connection *ic,
359 			     u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
360 {
361 	unsigned int avail, posted, got = 0, advertise;
362 	long oldval, newval;
363 
364 	*adv_credits = 0;
365 	if (!ic->i_flowctl)
366 		return wanted;
367 
368 try_again:
369 	advertise = 0;
370 	oldval = newval = atomic_read(&ic->i_credits);
371 	posted = IB_GET_POST_CREDITS(oldval);
372 	avail = IB_GET_SEND_CREDITS(oldval);
373 
374 	rdsdebug("wanted=%u credits=%u posted=%u\n",
375 			wanted, avail, posted);
376 
377 	/* The last credit must be used to send a credit update. */
378 	if (avail && !posted)
379 		avail--;
380 
381 	if (avail < wanted) {
382 		struct rds_connection *conn = ic->i_cm_id->context;
383 
384 		/* Oops, there aren't that many credits left! */
385 		set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
386 		got = avail;
387 	} else {
388 		/* Sometimes you get what you want, lalala. */
389 		got = wanted;
390 	}
391 	newval -= IB_SET_SEND_CREDITS(got);
392 
393 	/*
394 	 * If need_posted is non-zero, then the caller wants
395 	 * the posted regardless of whether any send credits are
396 	 * available.
397 	 */
398 	if (posted && (got || need_posted)) {
399 		advertise = min_t(unsigned int, posted, max_posted);
400 		newval -= IB_SET_POST_CREDITS(advertise);
401 	}
402 
403 	/* Finally bill everything */
404 	if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
405 		goto try_again;
406 
407 	*adv_credits = advertise;
408 	return got;
409 }
410 
411 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits)
412 {
413 	struct rds_ib_connection *ic = conn->c_transport_data;
414 
415 	if (credits == 0)
416 		return;
417 
418 	rdsdebug("credits=%u current=%u%s\n",
419 			credits,
420 			IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
421 			test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
422 
423 	atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
424 	if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
425 		queue_delayed_work(rds_wq, &conn->c_send_w, 0);
426 
427 	WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
428 
429 	rds_ib_stats_inc(s_ib_rx_credit_updates);
430 }
431 
432 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted)
433 {
434 	struct rds_ib_connection *ic = conn->c_transport_data;
435 
436 	if (posted == 0)
437 		return;
438 
439 	atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
440 
441 	/* Decide whether to send an update to the peer now.
442 	 * If we would send a credit update for every single buffer we
443 	 * post, we would end up with an ACK storm (ACK arrives,
444 	 * consumes buffer, we refill the ring, send ACK to remote
445 	 * advertising the newly posted buffer... ad inf)
446 	 *
447 	 * Performance pretty much depends on how often we send
448 	 * credit updates - too frequent updates mean lots of ACKs.
449 	 * Too infrequent updates, and the peer will run out of
450 	 * credits and has to throttle.
451 	 * For the time being, 16 seems to be a good compromise.
452 	 */
453 	if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
454 		set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
455 }
456 
457 static inline int rds_ib_set_wr_signal_state(struct rds_ib_connection *ic,
458 					     struct rds_ib_send_work *send,
459 					     bool notify)
460 {
461 	/*
462 	 * We want to delay signaling completions just enough to get
463 	 * the batching benefits but not so much that we create dead time
464 	 * on the wire.
465 	 */
466 	if (ic->i_unsignaled_wrs-- == 0 || notify) {
467 		ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
468 		send->s_wr.send_flags |= IB_SEND_SIGNALED;
469 		return 1;
470 	}
471 	return 0;
472 }
473 
474 /*
475  * This can be called multiple times for a given message.  The first time
476  * we see a message we map its scatterlist into the IB device so that
477  * we can provide that mapped address to the IB scatter gather entries
478  * in the IB work requests.  We translate the scatterlist into a series
479  * of work requests that fragment the message.  These work requests complete
480  * in order so we pass ownership of the message to the completion handler
481  * once we send the final fragment.
482  *
483  * The RDS core uses the c_send_lock to only enter this function once
484  * per connection.  This makes sure that the tx ring alloc/unalloc pairs
485  * don't get out of sync and confuse the ring.
486  */
487 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
488 		unsigned int hdr_off, unsigned int sg, unsigned int off)
489 {
490 	struct rds_ib_connection *ic = conn->c_transport_data;
491 	struct ib_device *dev = ic->i_cm_id->device;
492 	struct rds_ib_send_work *send = NULL;
493 	struct rds_ib_send_work *first;
494 	struct rds_ib_send_work *prev;
495 	struct ib_send_wr *failed_wr;
496 	struct scatterlist *scat;
497 	u32 pos;
498 	u32 i;
499 	u32 work_alloc;
500 	u32 credit_alloc = 0;
501 	u32 posted;
502 	u32 adv_credits = 0;
503 	int send_flags = 0;
504 	int bytes_sent = 0;
505 	int ret;
506 	int flow_controlled = 0;
507 	int nr_sig = 0;
508 
509 	BUG_ON(off % RDS_FRAG_SIZE);
510 	BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
511 
512 	/* Do not send cong updates to IB loopback */
513 	if (conn->c_loopback
514 	    && rm->m_inc.i_hdr.h_flags & RDS_FLAG_CONG_BITMAP) {
515 		rds_cong_map_updated(conn->c_fcong, ~(u64) 0);
516 		scat = &rm->data.op_sg[sg];
517 		ret = max_t(int, RDS_CONG_MAP_BYTES, scat->length);
518 		return sizeof(struct rds_header) + ret;
519 	}
520 
521 	/* FIXME we may overallocate here */
522 	if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
523 		i = 1;
524 	else
525 		i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
526 
527 	work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
528 	if (work_alloc == 0) {
529 		set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
530 		rds_ib_stats_inc(s_ib_tx_ring_full);
531 		ret = -ENOMEM;
532 		goto out;
533 	}
534 
535 	if (ic->i_flowctl) {
536 		credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
537 		adv_credits += posted;
538 		if (credit_alloc < work_alloc) {
539 			rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
540 			work_alloc = credit_alloc;
541 			flow_controlled = 1;
542 		}
543 		if (work_alloc == 0) {
544 			set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
545 			rds_ib_stats_inc(s_ib_tx_throttle);
546 			ret = -ENOMEM;
547 			goto out;
548 		}
549 	}
550 
551 	/* map the message the first time we see it */
552 	if (!ic->i_data_op) {
553 		if (rm->data.op_nents) {
554 			rm->data.op_count = ib_dma_map_sg(dev,
555 							  rm->data.op_sg,
556 							  rm->data.op_nents,
557 							  DMA_TO_DEVICE);
558 			rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.op_count);
559 			if (rm->data.op_count == 0) {
560 				rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
561 				rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
562 				ret = -ENOMEM; /* XXX ? */
563 				goto out;
564 			}
565 		} else {
566 			rm->data.op_count = 0;
567 		}
568 
569 		rds_message_addref(rm);
570 		rm->data.op_dmasg = 0;
571 		rm->data.op_dmaoff = 0;
572 		ic->i_data_op = &rm->data;
573 
574 		/* Finalize the header */
575 		if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
576 			rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
577 		if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
578 			rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
579 
580 		/* If it has a RDMA op, tell the peer we did it. This is
581 		 * used by the peer to release use-once RDMA MRs. */
582 		if (rm->rdma.op_active) {
583 			struct rds_ext_header_rdma ext_hdr;
584 
585 			ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.op_rkey);
586 			rds_message_add_extension(&rm->m_inc.i_hdr,
587 					RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
588 		}
589 		if (rm->m_rdma_cookie) {
590 			rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
591 					rds_rdma_cookie_key(rm->m_rdma_cookie),
592 					rds_rdma_cookie_offset(rm->m_rdma_cookie));
593 		}
594 
595 		/* Note - rds_ib_piggyb_ack clears the ACK_REQUIRED bit, so
596 		 * we should not do this unless we have a chance of at least
597 		 * sticking the header into the send ring. Which is why we
598 		 * should call rds_ib_ring_alloc first. */
599 		rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_ib_piggyb_ack(ic));
600 		rds_message_make_checksum(&rm->m_inc.i_hdr);
601 
602 		/*
603 		 * Update adv_credits since we reset the ACK_REQUIRED bit.
604 		 */
605 		if (ic->i_flowctl) {
606 			rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
607 			adv_credits += posted;
608 			BUG_ON(adv_credits > 255);
609 		}
610 	}
611 
612 	/* Sometimes you want to put a fence between an RDMA
613 	 * READ and the following SEND.
614 	 * We could either do this all the time
615 	 * or when requested by the user. Right now, we let
616 	 * the application choose.
617 	 */
618 	if (rm->rdma.op_active && rm->rdma.op_fence)
619 		send_flags = IB_SEND_FENCE;
620 
621 	/* Each frag gets a header. Msgs may be 0 bytes */
622 	send = &ic->i_sends[pos];
623 	first = send;
624 	prev = NULL;
625 	scat = &ic->i_data_op->op_sg[rm->data.op_dmasg];
626 	i = 0;
627 	do {
628 		unsigned int len = 0;
629 
630 		/* Set up the header */
631 		send->s_wr.send_flags = send_flags;
632 		send->s_wr.opcode = IB_WR_SEND;
633 		send->s_wr.num_sge = 1;
634 		send->s_wr.next = NULL;
635 		send->s_queued = jiffies;
636 		send->s_op = NULL;
637 
638 		send->s_sge[0].addr = ic->i_send_hdrs_dma
639 			+ (pos * sizeof(struct rds_header));
640 		send->s_sge[0].length = sizeof(struct rds_header);
641 
642 		memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
643 
644 		/* Set up the data, if present */
645 		if (i < work_alloc
646 		    && scat != &rm->data.op_sg[rm->data.op_count]) {
647 			len = min(RDS_FRAG_SIZE,
648 				ib_sg_dma_len(dev, scat) - rm->data.op_dmaoff);
649 			send->s_wr.num_sge = 2;
650 
651 			send->s_sge[1].addr = ib_sg_dma_address(dev, scat);
652 			send->s_sge[1].addr += rm->data.op_dmaoff;
653 			send->s_sge[1].length = len;
654 
655 			bytes_sent += len;
656 			rm->data.op_dmaoff += len;
657 			if (rm->data.op_dmaoff == ib_sg_dma_len(dev, scat)) {
658 				scat++;
659 				rm->data.op_dmasg++;
660 				rm->data.op_dmaoff = 0;
661 			}
662 		}
663 
664 		rds_ib_set_wr_signal_state(ic, send, 0);
665 
666 		/*
667 		 * Always signal the last one if we're stopping due to flow control.
668 		 */
669 		if (ic->i_flowctl && flow_controlled && i == (work_alloc-1))
670 			send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
671 
672 		if (send->s_wr.send_flags & IB_SEND_SIGNALED)
673 			nr_sig++;
674 
675 		rdsdebug("send %p wr %p num_sge %u next %p\n", send,
676 			 &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
677 
678 		if (ic->i_flowctl && 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 		send = &ic->i_sends[pos];
694 		i++;
695 
696 	} while (i < work_alloc
697 		 && scat != &rm->data.op_sg[rm->data.op_count]);
698 
699 	/* Account the RDS header in the number of bytes we sent, but just once.
700 	 * The caller has no concept of fragmentation. */
701 	if (hdr_off == 0)
702 		bytes_sent += sizeof(struct rds_header);
703 
704 	/* if we finished the message then send completion owns it */
705 	if (scat == &rm->data.op_sg[rm->data.op_count]) {
706 		prev->s_op = ic->i_data_op;
707 		prev->s_wr.send_flags |= IB_SEND_SOLICITED;
708 		if (!(prev->s_wr.send_flags & IB_SEND_SIGNALED)) {
709 			ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
710 			prev->s_wr.send_flags |= IB_SEND_SIGNALED;
711 			nr_sig++;
712 		}
713 		ic->i_data_op = NULL;
714 	}
715 
716 	/* Put back wrs & credits we didn't use */
717 	if (i < work_alloc) {
718 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
719 		work_alloc = i;
720 	}
721 	if (ic->i_flowctl && i < credit_alloc)
722 		rds_ib_send_add_credits(conn, credit_alloc - i);
723 
724 	if (nr_sig)
725 		atomic_add(nr_sig, &ic->i_signaled_sends);
726 
727 	/* XXX need to worry about failed_wr and partial sends. */
728 	failed_wr = &first->s_wr;
729 	ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
730 	rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
731 		 first, &first->s_wr, ret, failed_wr);
732 	BUG_ON(failed_wr != &first->s_wr);
733 	if (ret) {
734 		printk(KERN_WARNING "RDS/IB: ib_post_send to %pI4 "
735 		       "returned %d\n", &conn->c_faddr, ret);
736 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
737 		rds_ib_sub_signaled(ic, nr_sig);
738 		if (prev->s_op) {
739 			ic->i_data_op = prev->s_op;
740 			prev->s_op = NULL;
741 		}
742 
743 		rds_ib_conn_error(ic->conn, "ib_post_send failed\n");
744 		goto out;
745 	}
746 
747 	ret = bytes_sent;
748 out:
749 	BUG_ON(adv_credits);
750 	return ret;
751 }
752 
753 /*
754  * Issue atomic operation.
755  * A simplified version of the rdma case, we always map 1 SG, and
756  * only 8 bytes, for the return value from the atomic operation.
757  */
758 int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op)
759 {
760 	struct rds_ib_connection *ic = conn->c_transport_data;
761 	struct rds_ib_send_work *send = NULL;
762 	struct ib_send_wr *failed_wr;
763 	struct rds_ib_device *rds_ibdev;
764 	u32 pos;
765 	u32 work_alloc;
766 	int ret;
767 	int nr_sig = 0;
768 
769 	rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
770 
771 	work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, 1, &pos);
772 	if (work_alloc != 1) {
773 		rds_ib_stats_inc(s_ib_tx_ring_full);
774 		ret = -ENOMEM;
775 		goto out;
776 	}
777 
778 	/* address of send request in ring */
779 	send = &ic->i_sends[pos];
780 	send->s_queued = jiffies;
781 
782 	if (op->op_type == RDS_ATOMIC_TYPE_CSWP) {
783 		send->s_atomic_wr.wr.opcode = IB_WR_MASKED_ATOMIC_CMP_AND_SWP;
784 		send->s_atomic_wr.compare_add = op->op_m_cswp.compare;
785 		send->s_atomic_wr.swap = op->op_m_cswp.swap;
786 		send->s_atomic_wr.compare_add_mask = op->op_m_cswp.compare_mask;
787 		send->s_atomic_wr.swap_mask = op->op_m_cswp.swap_mask;
788 	} else { /* FADD */
789 		send->s_atomic_wr.wr.opcode = IB_WR_MASKED_ATOMIC_FETCH_AND_ADD;
790 		send->s_atomic_wr.compare_add = op->op_m_fadd.add;
791 		send->s_atomic_wr.swap = 0;
792 		send->s_atomic_wr.compare_add_mask = op->op_m_fadd.nocarry_mask;
793 		send->s_atomic_wr.swap_mask = 0;
794 	}
795 	nr_sig = rds_ib_set_wr_signal_state(ic, send, op->op_notify);
796 	send->s_atomic_wr.wr.num_sge = 1;
797 	send->s_atomic_wr.wr.next = NULL;
798 	send->s_atomic_wr.remote_addr = op->op_remote_addr;
799 	send->s_atomic_wr.rkey = op->op_rkey;
800 	send->s_op = op;
801 	rds_message_addref(container_of(send->s_op, struct rds_message, atomic));
802 
803 	/* map 8 byte retval buffer to the device */
804 	ret = ib_dma_map_sg(ic->i_cm_id->device, op->op_sg, 1, DMA_FROM_DEVICE);
805 	rdsdebug("ic %p mapping atomic op %p. mapped %d pg\n", ic, op, ret);
806 	if (ret != 1) {
807 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
808 		rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
809 		ret = -ENOMEM; /* XXX ? */
810 		goto out;
811 	}
812 
813 	/* Convert our struct scatterlist to struct ib_sge */
814 	send->s_sge[0].addr = ib_sg_dma_address(ic->i_cm_id->device, op->op_sg);
815 	send->s_sge[0].length = ib_sg_dma_len(ic->i_cm_id->device, op->op_sg);
816 	send->s_sge[0].lkey = ic->i_pd->local_dma_lkey;
817 
818 	rdsdebug("rva %Lx rpa %Lx len %u\n", op->op_remote_addr,
819 		 send->s_sge[0].addr, send->s_sge[0].length);
820 
821 	if (nr_sig)
822 		atomic_add(nr_sig, &ic->i_signaled_sends);
823 
824 	failed_wr = &send->s_atomic_wr.wr;
825 	ret = ib_post_send(ic->i_cm_id->qp, &send->s_atomic_wr.wr, &failed_wr);
826 	rdsdebug("ic %p send %p (wr %p) ret %d wr %p\n", ic,
827 		 send, &send->s_atomic_wr, ret, failed_wr);
828 	BUG_ON(failed_wr != &send->s_atomic_wr.wr);
829 	if (ret) {
830 		printk(KERN_WARNING "RDS/IB: atomic ib_post_send to %pI4 "
831 		       "returned %d\n", &conn->c_faddr, ret);
832 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
833 		rds_ib_sub_signaled(ic, nr_sig);
834 		goto out;
835 	}
836 
837 	if (unlikely(failed_wr != &send->s_atomic_wr.wr)) {
838 		printk(KERN_WARNING "RDS/IB: atomic ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
839 		BUG_ON(failed_wr != &send->s_atomic_wr.wr);
840 	}
841 
842 out:
843 	return ret;
844 }
845 
846 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op)
847 {
848 	struct rds_ib_connection *ic = conn->c_transport_data;
849 	struct rds_ib_send_work *send = NULL;
850 	struct rds_ib_send_work *first;
851 	struct rds_ib_send_work *prev;
852 	struct ib_send_wr *failed_wr;
853 	struct scatterlist *scat;
854 	unsigned long len;
855 	u64 remote_addr = op->op_remote_addr;
856 	u32 max_sge = ic->rds_ibdev->max_sge;
857 	u32 pos;
858 	u32 work_alloc;
859 	u32 i;
860 	u32 j;
861 	int sent;
862 	int ret;
863 	int num_sge;
864 	int nr_sig = 0;
865 
866 	/* map the op the first time we see it */
867 	if (!op->op_mapped) {
868 		op->op_count = ib_dma_map_sg(ic->i_cm_id->device,
869 					     op->op_sg, op->op_nents, (op->op_write) ?
870 					     DMA_TO_DEVICE : DMA_FROM_DEVICE);
871 		rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->op_count);
872 		if (op->op_count == 0) {
873 			rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
874 			ret = -ENOMEM; /* XXX ? */
875 			goto out;
876 		}
877 
878 		op->op_mapped = 1;
879 	}
880 
881 	/*
882 	 * Instead of knowing how to return a partial rdma read/write we insist that there
883 	 * be enough work requests to send the entire message.
884 	 */
885 	i = ceil(op->op_count, max_sge);
886 
887 	work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
888 	if (work_alloc != i) {
889 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
890 		rds_ib_stats_inc(s_ib_tx_ring_full);
891 		ret = -ENOMEM;
892 		goto out;
893 	}
894 
895 	send = &ic->i_sends[pos];
896 	first = send;
897 	prev = NULL;
898 	scat = &op->op_sg[0];
899 	sent = 0;
900 	num_sge = op->op_count;
901 
902 	for (i = 0; i < work_alloc && scat != &op->op_sg[op->op_count]; i++) {
903 		send->s_wr.send_flags = 0;
904 		send->s_queued = jiffies;
905 		send->s_op = NULL;
906 
907 		nr_sig += rds_ib_set_wr_signal_state(ic, send, op->op_notify);
908 
909 		send->s_wr.opcode = op->op_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ;
910 		send->s_rdma_wr.remote_addr = remote_addr;
911 		send->s_rdma_wr.rkey = op->op_rkey;
912 
913 		if (num_sge > max_sge) {
914 			send->s_rdma_wr.wr.num_sge = max_sge;
915 			num_sge -= max_sge;
916 		} else {
917 			send->s_rdma_wr.wr.num_sge = num_sge;
918 		}
919 
920 		send->s_rdma_wr.wr.next = NULL;
921 
922 		if (prev)
923 			prev->s_rdma_wr.wr.next = &send->s_rdma_wr.wr;
924 
925 		for (j = 0; j < send->s_rdma_wr.wr.num_sge &&
926 		     scat != &op->op_sg[op->op_count]; j++) {
927 			len = ib_sg_dma_len(ic->i_cm_id->device, scat);
928 			send->s_sge[j].addr =
929 				 ib_sg_dma_address(ic->i_cm_id->device, scat);
930 			send->s_sge[j].length = len;
931 			send->s_sge[j].lkey = ic->i_pd->local_dma_lkey;
932 
933 			sent += len;
934 			rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
935 
936 			remote_addr += len;
937 			scat++;
938 		}
939 
940 		rdsdebug("send %p wr %p num_sge %u next %p\n", send,
941 			&send->s_rdma_wr.wr,
942 			send->s_rdma_wr.wr.num_sge,
943 			send->s_rdma_wr.wr.next);
944 
945 		prev = send;
946 		if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
947 			send = ic->i_sends;
948 	}
949 
950 	/* give a reference to the last op */
951 	if (scat == &op->op_sg[op->op_count]) {
952 		prev->s_op = op;
953 		rds_message_addref(container_of(op, struct rds_message, rdma));
954 	}
955 
956 	if (i < work_alloc) {
957 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
958 		work_alloc = i;
959 	}
960 
961 	if (nr_sig)
962 		atomic_add(nr_sig, &ic->i_signaled_sends);
963 
964 	failed_wr = &first->s_rdma_wr.wr;
965 	ret = ib_post_send(ic->i_cm_id->qp, &first->s_rdma_wr.wr, &failed_wr);
966 	rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
967 		 first, &first->s_rdma_wr.wr, ret, failed_wr);
968 	BUG_ON(failed_wr != &first->s_rdma_wr.wr);
969 	if (ret) {
970 		printk(KERN_WARNING "RDS/IB: rdma ib_post_send to %pI4 "
971 		       "returned %d\n", &conn->c_faddr, ret);
972 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
973 		rds_ib_sub_signaled(ic, nr_sig);
974 		goto out;
975 	}
976 
977 	if (unlikely(failed_wr != &first->s_rdma_wr.wr)) {
978 		printk(KERN_WARNING "RDS/IB: ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
979 		BUG_ON(failed_wr != &first->s_rdma_wr.wr);
980 	}
981 
982 
983 out:
984 	return ret;
985 }
986 
987 void rds_ib_xmit_path_complete(struct rds_conn_path *cp)
988 {
989 	struct rds_connection *conn = cp->cp_conn;
990 	struct rds_ib_connection *ic = conn->c_transport_data;
991 
992 	/* We may have a pending ACK or window update we were unable
993 	 * to send previously (due to flow control). Try again. */
994 	rds_ib_attempt_ack(ic);
995 }
996