xref: /openbmc/linux/net/rds/ib_send.c (revision ca55b2fe)
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.h"
40 #include "ib.h"
41 
42 /*
43  * Convert IB-specific error message to RDS error message and call core
44  * completion handler.
45  */
46 static void rds_ib_send_complete(struct rds_message *rm,
47 				 int wc_status,
48 				 void (*complete)(struct rds_message *rm, int status))
49 {
50 	int notify_status;
51 
52 	switch (wc_status) {
53 	case IB_WC_WR_FLUSH_ERR:
54 		return;
55 
56 	case IB_WC_SUCCESS:
57 		notify_status = RDS_RDMA_SUCCESS;
58 		break;
59 
60 	case IB_WC_REM_ACCESS_ERR:
61 		notify_status = RDS_RDMA_REMOTE_ERROR;
62 		break;
63 
64 	default:
65 		notify_status = RDS_RDMA_OTHER_ERROR;
66 		break;
67 	}
68 	complete(rm, notify_status);
69 }
70 
71 static void rds_ib_send_unmap_data(struct rds_ib_connection *ic,
72 				   struct rm_data_op *op,
73 				   int wc_status)
74 {
75 	if (op->op_nents)
76 		ib_dma_unmap_sg(ic->i_cm_id->device,
77 				op->op_sg, op->op_nents,
78 				DMA_TO_DEVICE);
79 }
80 
81 static void rds_ib_send_unmap_rdma(struct rds_ib_connection *ic,
82 				   struct rm_rdma_op *op,
83 				   int wc_status)
84 {
85 	if (op->op_mapped) {
86 		ib_dma_unmap_sg(ic->i_cm_id->device,
87 				op->op_sg, op->op_nents,
88 				op->op_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
89 		op->op_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 	 * we would need to take an event for the rdma WR. 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(container_of(op, struct rds_message, rdma),
113 			     wc_status, rds_rdma_send_complete);
114 
115 	if (op->op_write)
116 		rds_stats_add(s_send_rdma_bytes, op->op_bytes);
117 	else
118 		rds_stats_add(s_recv_rdma_bytes, op->op_bytes);
119 }
120 
121 static void rds_ib_send_unmap_atomic(struct rds_ib_connection *ic,
122 				     struct rm_atomic_op *op,
123 				     int wc_status)
124 {
125 	/* unmap atomic recvbuf */
126 	if (op->op_mapped) {
127 		ib_dma_unmap_sg(ic->i_cm_id->device, op->op_sg, 1,
128 				DMA_FROM_DEVICE);
129 		op->op_mapped = 0;
130 	}
131 
132 	rds_ib_send_complete(container_of(op, struct rds_message, atomic),
133 			     wc_status, rds_atomic_send_complete);
134 
135 	if (op->op_type == RDS_ATOMIC_TYPE_CSWP)
136 		rds_ib_stats_inc(s_ib_atomic_cswp);
137 	else
138 		rds_ib_stats_inc(s_ib_atomic_fadd);
139 }
140 
141 /*
142  * Unmap the resources associated with a struct send_work.
143  *
144  * Returns the rm for no good reason other than it is unobtainable
145  * other than by switching on wr.opcode, currently, and the caller,
146  * the event handler, needs it.
147  */
148 static struct rds_message *rds_ib_send_unmap_op(struct rds_ib_connection *ic,
149 						struct rds_ib_send_work *send,
150 						int wc_status)
151 {
152 	struct rds_message *rm = NULL;
153 
154 	/* In the error case, wc.opcode sometimes contains garbage */
155 	switch (send->s_wr.opcode) {
156 	case IB_WR_SEND:
157 		if (send->s_op) {
158 			rm = container_of(send->s_op, struct rds_message, data);
159 			rds_ib_send_unmap_data(ic, send->s_op, wc_status);
160 		}
161 		break;
162 	case IB_WR_RDMA_WRITE:
163 	case IB_WR_RDMA_READ:
164 		if (send->s_op) {
165 			rm = container_of(send->s_op, struct rds_message, rdma);
166 			rds_ib_send_unmap_rdma(ic, send->s_op, wc_status);
167 		}
168 		break;
169 	case IB_WR_ATOMIC_FETCH_AND_ADD:
170 	case IB_WR_ATOMIC_CMP_AND_SWP:
171 		if (send->s_op) {
172 			rm = container_of(send->s_op, struct rds_message, atomic);
173 			rds_ib_send_unmap_atomic(ic, send->s_op, wc_status);
174 		}
175 		break;
176 	default:
177 		printk_ratelimited(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_pd->local_dma_lkey;
206 
207 		send->s_sge[1].lkey = ic->i_pd->local_dma_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 only fast path caller always has a non-zero nr, so we don't
224  * bother testing nr before performing the atomic sub.
225  */
226 static void rds_ib_sub_signaled(struct rds_ib_connection *ic, int nr)
227 {
228 	if ((atomic_sub_return(nr, &ic->i_signaled_sends) == 0) &&
229 	    waitqueue_active(&rds_ib_ring_empty_wait))
230 		wake_up(&rds_ib_ring_empty_wait);
231 	BUG_ON(atomic_read(&ic->i_signaled_sends) < 0);
232 }
233 
234 /*
235  * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
236  * operations performed in the send path.  As the sender allocs and potentially
237  * unallocs the next free entry in the ring it doesn't alter which is
238  * the next to be freed, which is what this is concerned with.
239  */
240 void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
241 {
242 	struct rds_connection *conn = context;
243 	struct rds_ib_connection *ic = conn->c_transport_data;
244 	struct rds_message *rm = NULL;
245 	struct ib_wc wc;
246 	struct rds_ib_send_work *send;
247 	u32 completed;
248 	u32 oldest;
249 	u32 i = 0;
250 	int ret;
251 	int nr_sig = 0;
252 
253 	rdsdebug("cq %p conn %p\n", cq, conn);
254 	rds_ib_stats_inc(s_ib_tx_cq_call);
255 	ret = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
256 	if (ret)
257 		rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
258 
259 	while (ib_poll_cq(cq, 1, &wc) > 0) {
260 		rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
261 			 (unsigned long long)wc.wr_id, wc.status,
262 			 ib_wc_status_msg(wc.status), wc.byte_len,
263 			 be32_to_cpu(wc.ex.imm_data));
264 		rds_ib_stats_inc(s_ib_tx_cq_event);
265 
266 		if (wc.wr_id == RDS_IB_ACK_WR_ID) {
267 			if (time_after(jiffies, ic->i_ack_queued + HZ/2))
268 				rds_ib_stats_inc(s_ib_tx_stalled);
269 			rds_ib_ack_send_complete(ic);
270 			continue;
271 		}
272 
273 		oldest = rds_ib_ring_oldest(&ic->i_send_ring);
274 
275 		completed = rds_ib_ring_completed(&ic->i_send_ring, wc.wr_id, oldest);
276 
277 		for (i = 0; i < completed; i++) {
278 			send = &ic->i_sends[oldest];
279 			if (send->s_wr.send_flags & IB_SEND_SIGNALED)
280 				nr_sig++;
281 
282 			rm = rds_ib_send_unmap_op(ic, send, wc.status);
283 
284 			if (time_after(jiffies, send->s_queued + HZ/2))
285 				rds_ib_stats_inc(s_ib_tx_stalled);
286 
287 			if (send->s_op) {
288 				if (send->s_op == rm->m_final_op) {
289 					/* If anyone waited for this message to get flushed out, wake
290 					 * them up now */
291 					rds_message_unmapped(rm);
292 				}
293 				rds_message_put(rm);
294 				send->s_op = NULL;
295 			}
296 
297 			oldest = (oldest + 1) % ic->i_send_ring.w_nr;
298 		}
299 
300 		rds_ib_ring_free(&ic->i_send_ring, completed);
301 		rds_ib_sub_signaled(ic, nr_sig);
302 		nr_sig = 0;
303 
304 		if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
305 		    test_bit(0, &conn->c_map_queued))
306 			queue_delayed_work(rds_wq, &conn->c_send_w, 0);
307 
308 		/* We expect errors as the qp is drained during shutdown */
309 		if (wc.status != IB_WC_SUCCESS && rds_conn_up(conn)) {
310 			rds_ib_conn_error(conn, "send completion on %pI4 had status "
311 					  "%u (%s), disconnecting and reconnecting\n",
312 					  &conn->c_faddr, wc.status,
313 					  ib_wc_status_msg(wc.status));
314 		}
315 	}
316 }
317 
318 /*
319  * This is the main function for allocating credits when sending
320  * messages.
321  *
322  * Conceptually, we have two counters:
323  *  -	send credits: this tells us how many WRs we're allowed
324  *	to submit without overruning the receiver's queue. For
325  *	each SEND WR we post, we decrement this by one.
326  *
327  *  -	posted credits: this tells us how many WRs we recently
328  *	posted to the receive queue. This value is transferred
329  *	to the peer as a "credit update" in a RDS header field.
330  *	Every time we transmit credits to the peer, we subtract
331  *	the amount of transferred credits from this counter.
332  *
333  * It is essential that we avoid situations where both sides have
334  * exhausted their send credits, and are unable to send new credits
335  * to the peer. We achieve this by requiring that we send at least
336  * one credit update to the peer before exhausting our credits.
337  * When new credits arrive, we subtract one credit that is withheld
338  * until we've posted new buffers and are ready to transmit these
339  * credits (see rds_ib_send_add_credits below).
340  *
341  * The RDS send code is essentially single-threaded; rds_send_xmit
342  * sets RDS_IN_XMIT to ensure exclusive access to the send ring.
343  * However, the ACK sending code is independent and can race with
344  * message SENDs.
345  *
346  * In the send path, we need to update the counters for send credits
347  * and the counter of posted buffers atomically - when we use the
348  * last available credit, we cannot allow another thread to race us
349  * and grab the posted credits counter.  Hence, we have to use a
350  * spinlock to protect the credit counter, or use atomics.
351  *
352  * Spinlocks shared between the send and the receive path are bad,
353  * because they create unnecessary delays. An early implementation
354  * using a spinlock showed a 5% degradation in throughput at some
355  * loads.
356  *
357  * This implementation avoids spinlocks completely, putting both
358  * counters into a single atomic, and updating that atomic using
359  * atomic_add (in the receive path, when receiving fresh credits),
360  * and using atomic_cmpxchg when updating the two counters.
361  */
362 int rds_ib_send_grab_credits(struct rds_ib_connection *ic,
363 			     u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
364 {
365 	unsigned int avail, posted, got = 0, advertise;
366 	long oldval, newval;
367 
368 	*adv_credits = 0;
369 	if (!ic->i_flowctl)
370 		return wanted;
371 
372 try_again:
373 	advertise = 0;
374 	oldval = newval = atomic_read(&ic->i_credits);
375 	posted = IB_GET_POST_CREDITS(oldval);
376 	avail = IB_GET_SEND_CREDITS(oldval);
377 
378 	rdsdebug("wanted=%u credits=%u posted=%u\n",
379 			wanted, avail, posted);
380 
381 	/* The last credit must be used to send a credit update. */
382 	if (avail && !posted)
383 		avail--;
384 
385 	if (avail < wanted) {
386 		struct rds_connection *conn = ic->i_cm_id->context;
387 
388 		/* Oops, there aren't that many credits left! */
389 		set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
390 		got = avail;
391 	} else {
392 		/* Sometimes you get what you want, lalala. */
393 		got = wanted;
394 	}
395 	newval -= IB_SET_SEND_CREDITS(got);
396 
397 	/*
398 	 * If need_posted is non-zero, then the caller wants
399 	 * the posted regardless of whether any send credits are
400 	 * available.
401 	 */
402 	if (posted && (got || need_posted)) {
403 		advertise = min_t(unsigned int, posted, max_posted);
404 		newval -= IB_SET_POST_CREDITS(advertise);
405 	}
406 
407 	/* Finally bill everything */
408 	if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
409 		goto try_again;
410 
411 	*adv_credits = advertise;
412 	return got;
413 }
414 
415 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits)
416 {
417 	struct rds_ib_connection *ic = conn->c_transport_data;
418 
419 	if (credits == 0)
420 		return;
421 
422 	rdsdebug("credits=%u current=%u%s\n",
423 			credits,
424 			IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
425 			test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
426 
427 	atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
428 	if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
429 		queue_delayed_work(rds_wq, &conn->c_send_w, 0);
430 
431 	WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
432 
433 	rds_ib_stats_inc(s_ib_rx_credit_updates);
434 }
435 
436 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted)
437 {
438 	struct rds_ib_connection *ic = conn->c_transport_data;
439 
440 	if (posted == 0)
441 		return;
442 
443 	atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
444 
445 	/* Decide whether to send an update to the peer now.
446 	 * If we would send a credit update for every single buffer we
447 	 * post, we would end up with an ACK storm (ACK arrives,
448 	 * consumes buffer, we refill the ring, send ACK to remote
449 	 * advertising the newly posted buffer... ad inf)
450 	 *
451 	 * Performance pretty much depends on how often we send
452 	 * credit updates - too frequent updates mean lots of ACKs.
453 	 * Too infrequent updates, and the peer will run out of
454 	 * credits and has to throttle.
455 	 * For the time being, 16 seems to be a good compromise.
456 	 */
457 	if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
458 		set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
459 }
460 
461 static inline int rds_ib_set_wr_signal_state(struct rds_ib_connection *ic,
462 					     struct rds_ib_send_work *send,
463 					     bool notify)
464 {
465 	/*
466 	 * We want to delay signaling completions just enough to get
467 	 * the batching benefits but not so much that we create dead time
468 	 * on the wire.
469 	 */
470 	if (ic->i_unsignaled_wrs-- == 0 || notify) {
471 		ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
472 		send->s_wr.send_flags |= IB_SEND_SIGNALED;
473 		return 1;
474 	}
475 	return 0;
476 }
477 
478 /*
479  * This can be called multiple times for a given message.  The first time
480  * we see a message we map its scatterlist into the IB device so that
481  * we can provide that mapped address to the IB scatter gather entries
482  * in the IB work requests.  We translate the scatterlist into a series
483  * of work requests that fragment the message.  These work requests complete
484  * in order so we pass ownership of the message to the completion handler
485  * once we send the final fragment.
486  *
487  * The RDS core uses the c_send_lock to only enter this function once
488  * per connection.  This makes sure that the tx ring alloc/unalloc pairs
489  * don't get out of sync and confuse the ring.
490  */
491 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
492 		unsigned int hdr_off, unsigned int sg, unsigned int off)
493 {
494 	struct rds_ib_connection *ic = conn->c_transport_data;
495 	struct ib_device *dev = ic->i_cm_id->device;
496 	struct rds_ib_send_work *send = NULL;
497 	struct rds_ib_send_work *first;
498 	struct rds_ib_send_work *prev;
499 	struct ib_send_wr *failed_wr;
500 	struct scatterlist *scat;
501 	u32 pos;
502 	u32 i;
503 	u32 work_alloc;
504 	u32 credit_alloc = 0;
505 	u32 posted;
506 	u32 adv_credits = 0;
507 	int send_flags = 0;
508 	int bytes_sent = 0;
509 	int ret;
510 	int flow_controlled = 0;
511 	int nr_sig = 0;
512 
513 	BUG_ON(off % RDS_FRAG_SIZE);
514 	BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
515 
516 	/* Do not send cong updates to IB loopback */
517 	if (conn->c_loopback
518 	    && rm->m_inc.i_hdr.h_flags & RDS_FLAG_CONG_BITMAP) {
519 		rds_cong_map_updated(conn->c_fcong, ~(u64) 0);
520 		scat = &rm->data.op_sg[sg];
521 		ret = max_t(int, RDS_CONG_MAP_BYTES, scat->length);
522 		return sizeof(struct rds_header) + ret;
523 	}
524 
525 	/* FIXME we may overallocate here */
526 	if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
527 		i = 1;
528 	else
529 		i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
530 
531 	work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
532 	if (work_alloc == 0) {
533 		set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
534 		rds_ib_stats_inc(s_ib_tx_ring_full);
535 		ret = -ENOMEM;
536 		goto out;
537 	}
538 
539 	if (ic->i_flowctl) {
540 		credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
541 		adv_credits += posted;
542 		if (credit_alloc < work_alloc) {
543 			rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
544 			work_alloc = credit_alloc;
545 			flow_controlled = 1;
546 		}
547 		if (work_alloc == 0) {
548 			set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
549 			rds_ib_stats_inc(s_ib_tx_throttle);
550 			ret = -ENOMEM;
551 			goto out;
552 		}
553 	}
554 
555 	/* map the message the first time we see it */
556 	if (!ic->i_data_op) {
557 		if (rm->data.op_nents) {
558 			rm->data.op_count = ib_dma_map_sg(dev,
559 							  rm->data.op_sg,
560 							  rm->data.op_nents,
561 							  DMA_TO_DEVICE);
562 			rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.op_count);
563 			if (rm->data.op_count == 0) {
564 				rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
565 				rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
566 				ret = -ENOMEM; /* XXX ? */
567 				goto out;
568 			}
569 		} else {
570 			rm->data.op_count = 0;
571 		}
572 
573 		rds_message_addref(rm);
574 		rm->data.op_dmasg = 0;
575 		rm->data.op_dmaoff = 0;
576 		ic->i_data_op = &rm->data;
577 
578 		/* Finalize the header */
579 		if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
580 			rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
581 		if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
582 			rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
583 
584 		/* If it has a RDMA op, tell the peer we did it. This is
585 		 * used by the peer to release use-once RDMA MRs. */
586 		if (rm->rdma.op_active) {
587 			struct rds_ext_header_rdma ext_hdr;
588 
589 			ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.op_rkey);
590 			rds_message_add_extension(&rm->m_inc.i_hdr,
591 					RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
592 		}
593 		if (rm->m_rdma_cookie) {
594 			rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
595 					rds_rdma_cookie_key(rm->m_rdma_cookie),
596 					rds_rdma_cookie_offset(rm->m_rdma_cookie));
597 		}
598 
599 		/* Note - rds_ib_piggyb_ack clears the ACK_REQUIRED bit, so
600 		 * we should not do this unless we have a chance of at least
601 		 * sticking the header into the send ring. Which is why we
602 		 * should call rds_ib_ring_alloc first. */
603 		rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_ib_piggyb_ack(ic));
604 		rds_message_make_checksum(&rm->m_inc.i_hdr);
605 
606 		/*
607 		 * Update adv_credits since we reset the ACK_REQUIRED bit.
608 		 */
609 		if (ic->i_flowctl) {
610 			rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
611 			adv_credits += posted;
612 			BUG_ON(adv_credits > 255);
613 		}
614 	}
615 
616 	/* Sometimes you want to put a fence between an RDMA
617 	 * READ and the following SEND.
618 	 * We could either do this all the time
619 	 * or when requested by the user. Right now, we let
620 	 * the application choose.
621 	 */
622 	if (rm->rdma.op_active && rm->rdma.op_fence)
623 		send_flags = IB_SEND_FENCE;
624 
625 	/* Each frag gets a header. Msgs may be 0 bytes */
626 	send = &ic->i_sends[pos];
627 	first = send;
628 	prev = NULL;
629 	scat = &ic->i_data_op->op_sg[rm->data.op_dmasg];
630 	i = 0;
631 	do {
632 		unsigned int len = 0;
633 
634 		/* Set up the header */
635 		send->s_wr.send_flags = send_flags;
636 		send->s_wr.opcode = IB_WR_SEND;
637 		send->s_wr.num_sge = 1;
638 		send->s_wr.next = NULL;
639 		send->s_queued = jiffies;
640 		send->s_op = NULL;
641 
642 		send->s_sge[0].addr = ic->i_send_hdrs_dma
643 			+ (pos * sizeof(struct rds_header));
644 		send->s_sge[0].length = sizeof(struct rds_header);
645 
646 		memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
647 
648 		/* Set up the data, if present */
649 		if (i < work_alloc
650 		    && scat != &rm->data.op_sg[rm->data.op_count]) {
651 			len = min(RDS_FRAG_SIZE,
652 				ib_sg_dma_len(dev, scat) - rm->data.op_dmaoff);
653 			send->s_wr.num_sge = 2;
654 
655 			send->s_sge[1].addr = ib_sg_dma_address(dev, scat);
656 			send->s_sge[1].addr += rm->data.op_dmaoff;
657 			send->s_sge[1].length = len;
658 
659 			bytes_sent += len;
660 			rm->data.op_dmaoff += len;
661 			if (rm->data.op_dmaoff == ib_sg_dma_len(dev, scat)) {
662 				scat++;
663 				rm->data.op_dmasg++;
664 				rm->data.op_dmaoff = 0;
665 			}
666 		}
667 
668 		rds_ib_set_wr_signal_state(ic, send, 0);
669 
670 		/*
671 		 * Always signal the last one if we're stopping due to flow control.
672 		 */
673 		if (ic->i_flowctl && flow_controlled && i == (work_alloc-1))
674 			send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
675 
676 		if (send->s_wr.send_flags & IB_SEND_SIGNALED)
677 			nr_sig++;
678 
679 		rdsdebug("send %p wr %p num_sge %u next %p\n", send,
680 			 &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
681 
682 		if (ic->i_flowctl && adv_credits) {
683 			struct rds_header *hdr = &ic->i_send_hdrs[pos];
684 
685 			/* add credit and redo the header checksum */
686 			hdr->h_credit = adv_credits;
687 			rds_message_make_checksum(hdr);
688 			adv_credits = 0;
689 			rds_ib_stats_inc(s_ib_tx_credit_updates);
690 		}
691 
692 		if (prev)
693 			prev->s_wr.next = &send->s_wr;
694 		prev = send;
695 
696 		pos = (pos + 1) % ic->i_send_ring.w_nr;
697 		send = &ic->i_sends[pos];
698 		i++;
699 
700 	} while (i < work_alloc
701 		 && scat != &rm->data.op_sg[rm->data.op_count]);
702 
703 	/* Account the RDS header in the number of bytes we sent, but just once.
704 	 * The caller has no concept of fragmentation. */
705 	if (hdr_off == 0)
706 		bytes_sent += sizeof(struct rds_header);
707 
708 	/* if we finished the message then send completion owns it */
709 	if (scat == &rm->data.op_sg[rm->data.op_count]) {
710 		prev->s_op = ic->i_data_op;
711 		prev->s_wr.send_flags |= IB_SEND_SOLICITED;
712 		if (!(prev->s_wr.send_flags & IB_SEND_SIGNALED)) {
713 			ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
714 			prev->s_wr.send_flags |= IB_SEND_SIGNALED;
715 			nr_sig++;
716 		}
717 		ic->i_data_op = NULL;
718 	}
719 
720 	/* Put back wrs & credits we didn't use */
721 	if (i < work_alloc) {
722 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
723 		work_alloc = i;
724 	}
725 	if (ic->i_flowctl && i < credit_alloc)
726 		rds_ib_send_add_credits(conn, credit_alloc - i);
727 
728 	if (nr_sig)
729 		atomic_add(nr_sig, &ic->i_signaled_sends);
730 
731 	/* XXX need to worry about failed_wr and partial sends. */
732 	failed_wr = &first->s_wr;
733 	ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
734 	rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
735 		 first, &first->s_wr, ret, failed_wr);
736 	BUG_ON(failed_wr != &first->s_wr);
737 	if (ret) {
738 		printk(KERN_WARNING "RDS/IB: ib_post_send to %pI4 "
739 		       "returned %d\n", &conn->c_faddr, ret);
740 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
741 		rds_ib_sub_signaled(ic, nr_sig);
742 		if (prev->s_op) {
743 			ic->i_data_op = prev->s_op;
744 			prev->s_op = NULL;
745 		}
746 
747 		rds_ib_conn_error(ic->conn, "ib_post_send failed\n");
748 		goto out;
749 	}
750 
751 	ret = bytes_sent;
752 out:
753 	BUG_ON(adv_credits);
754 	return ret;
755 }
756 
757 /*
758  * Issue atomic operation.
759  * A simplified version of the rdma case, we always map 1 SG, and
760  * only 8 bytes, for the return value from the atomic operation.
761  */
762 int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op)
763 {
764 	struct rds_ib_connection *ic = conn->c_transport_data;
765 	struct rds_ib_send_work *send = NULL;
766 	struct ib_send_wr *failed_wr;
767 	struct rds_ib_device *rds_ibdev;
768 	u32 pos;
769 	u32 work_alloc;
770 	int ret;
771 	int nr_sig = 0;
772 
773 	rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
774 
775 	work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, 1, &pos);
776 	if (work_alloc != 1) {
777 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
778 		rds_ib_stats_inc(s_ib_tx_ring_full);
779 		ret = -ENOMEM;
780 		goto out;
781 	}
782 
783 	/* address of send request in ring */
784 	send = &ic->i_sends[pos];
785 	send->s_queued = jiffies;
786 
787 	if (op->op_type == RDS_ATOMIC_TYPE_CSWP) {
788 		send->s_wr.opcode = IB_WR_MASKED_ATOMIC_CMP_AND_SWP;
789 		send->s_wr.wr.atomic.compare_add = op->op_m_cswp.compare;
790 		send->s_wr.wr.atomic.swap = op->op_m_cswp.swap;
791 		send->s_wr.wr.atomic.compare_add_mask = op->op_m_cswp.compare_mask;
792 		send->s_wr.wr.atomic.swap_mask = op->op_m_cswp.swap_mask;
793 	} else { /* FADD */
794 		send->s_wr.opcode = IB_WR_MASKED_ATOMIC_FETCH_AND_ADD;
795 		send->s_wr.wr.atomic.compare_add = op->op_m_fadd.add;
796 		send->s_wr.wr.atomic.swap = 0;
797 		send->s_wr.wr.atomic.compare_add_mask = op->op_m_fadd.nocarry_mask;
798 		send->s_wr.wr.atomic.swap_mask = 0;
799 	}
800 	nr_sig = rds_ib_set_wr_signal_state(ic, send, op->op_notify);
801 	send->s_wr.num_sge = 1;
802 	send->s_wr.next = NULL;
803 	send->s_wr.wr.atomic.remote_addr = op->op_remote_addr;
804 	send->s_wr.wr.atomic.rkey = op->op_rkey;
805 	send->s_op = op;
806 	rds_message_addref(container_of(send->s_op, struct rds_message, atomic));
807 
808 	/* map 8 byte retval buffer to the device */
809 	ret = ib_dma_map_sg(ic->i_cm_id->device, op->op_sg, 1, DMA_FROM_DEVICE);
810 	rdsdebug("ic %p mapping atomic op %p. mapped %d pg\n", ic, op, ret);
811 	if (ret != 1) {
812 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
813 		rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
814 		ret = -ENOMEM; /* XXX ? */
815 		goto out;
816 	}
817 
818 	/* Convert our struct scatterlist to struct ib_sge */
819 	send->s_sge[0].addr = ib_sg_dma_address(ic->i_cm_id->device, op->op_sg);
820 	send->s_sge[0].length = ib_sg_dma_len(ic->i_cm_id->device, op->op_sg);
821 	send->s_sge[0].lkey = ic->i_pd->local_dma_lkey;
822 
823 	rdsdebug("rva %Lx rpa %Lx len %u\n", op->op_remote_addr,
824 		 send->s_sge[0].addr, send->s_sge[0].length);
825 
826 	if (nr_sig)
827 		atomic_add(nr_sig, &ic->i_signaled_sends);
828 
829 	failed_wr = &send->s_wr;
830 	ret = ib_post_send(ic->i_cm_id->qp, &send->s_wr, &failed_wr);
831 	rdsdebug("ic %p send %p (wr %p) ret %d wr %p\n", ic,
832 		 send, &send->s_wr, ret, failed_wr);
833 	BUG_ON(failed_wr != &send->s_wr);
834 	if (ret) {
835 		printk(KERN_WARNING "RDS/IB: atomic ib_post_send to %pI4 "
836 		       "returned %d\n", &conn->c_faddr, ret);
837 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
838 		rds_ib_sub_signaled(ic, nr_sig);
839 		goto out;
840 	}
841 
842 	if (unlikely(failed_wr != &send->s_wr)) {
843 		printk(KERN_WARNING "RDS/IB: atomic ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
844 		BUG_ON(failed_wr != &send->s_wr);
845 	}
846 
847 out:
848 	return ret;
849 }
850 
851 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op)
852 {
853 	struct rds_ib_connection *ic = conn->c_transport_data;
854 	struct rds_ib_send_work *send = NULL;
855 	struct rds_ib_send_work *first;
856 	struct rds_ib_send_work *prev;
857 	struct ib_send_wr *failed_wr;
858 	struct scatterlist *scat;
859 	unsigned long len;
860 	u64 remote_addr = op->op_remote_addr;
861 	u32 max_sge = ic->rds_ibdev->max_sge;
862 	u32 pos;
863 	u32 work_alloc;
864 	u32 i;
865 	u32 j;
866 	int sent;
867 	int ret;
868 	int num_sge;
869 	int nr_sig = 0;
870 
871 	/* map the op the first time we see it */
872 	if (!op->op_mapped) {
873 		op->op_count = ib_dma_map_sg(ic->i_cm_id->device,
874 					     op->op_sg, op->op_nents, (op->op_write) ?
875 					     DMA_TO_DEVICE : DMA_FROM_DEVICE);
876 		rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->op_count);
877 		if (op->op_count == 0) {
878 			rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
879 			ret = -ENOMEM; /* XXX ? */
880 			goto out;
881 		}
882 
883 		op->op_mapped = 1;
884 	}
885 
886 	/*
887 	 * Instead of knowing how to return a partial rdma read/write we insist that there
888 	 * be enough work requests to send the entire message.
889 	 */
890 	i = ceil(op->op_count, max_sge);
891 
892 	work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
893 	if (work_alloc != i) {
894 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
895 		rds_ib_stats_inc(s_ib_tx_ring_full);
896 		ret = -ENOMEM;
897 		goto out;
898 	}
899 
900 	send = &ic->i_sends[pos];
901 	first = send;
902 	prev = NULL;
903 	scat = &op->op_sg[0];
904 	sent = 0;
905 	num_sge = op->op_count;
906 
907 	for (i = 0; i < work_alloc && scat != &op->op_sg[op->op_count]; i++) {
908 		send->s_wr.send_flags = 0;
909 		send->s_queued = jiffies;
910 		send->s_op = NULL;
911 
912 		nr_sig += rds_ib_set_wr_signal_state(ic, send, op->op_notify);
913 
914 		send->s_wr.opcode = op->op_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ;
915 		send->s_wr.wr.rdma.remote_addr = remote_addr;
916 		send->s_wr.wr.rdma.rkey = op->op_rkey;
917 
918 		if (num_sge > max_sge) {
919 			send->s_wr.num_sge = max_sge;
920 			num_sge -= max_sge;
921 		} else {
922 			send->s_wr.num_sge = num_sge;
923 		}
924 
925 		send->s_wr.next = NULL;
926 
927 		if (prev)
928 			prev->s_wr.next = &send->s_wr;
929 
930 		for (j = 0; j < send->s_wr.num_sge && scat != &op->op_sg[op->op_count]; j++) {
931 			len = ib_sg_dma_len(ic->i_cm_id->device, scat);
932 			send->s_sge[j].addr =
933 				 ib_sg_dma_address(ic->i_cm_id->device, scat);
934 			send->s_sge[j].length = len;
935 			send->s_sge[j].lkey = ic->i_pd->local_dma_lkey;
936 
937 			sent += len;
938 			rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
939 
940 			remote_addr += len;
941 			scat++;
942 		}
943 
944 		rdsdebug("send %p wr %p num_sge %u next %p\n", send,
945 			&send->s_wr, send->s_wr.num_sge, send->s_wr.next);
946 
947 		prev = send;
948 		if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
949 			send = ic->i_sends;
950 	}
951 
952 	/* give a reference to the last op */
953 	if (scat == &op->op_sg[op->op_count]) {
954 		prev->s_op = op;
955 		rds_message_addref(container_of(op, struct rds_message, rdma));
956 	}
957 
958 	if (i < work_alloc) {
959 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
960 		work_alloc = i;
961 	}
962 
963 	if (nr_sig)
964 		atomic_add(nr_sig, &ic->i_signaled_sends);
965 
966 	failed_wr = &first->s_wr;
967 	ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
968 	rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
969 		 first, &first->s_wr, ret, failed_wr);
970 	BUG_ON(failed_wr != &first->s_wr);
971 	if (ret) {
972 		printk(KERN_WARNING "RDS/IB: rdma ib_post_send to %pI4 "
973 		       "returned %d\n", &conn->c_faddr, ret);
974 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
975 		rds_ib_sub_signaled(ic, nr_sig);
976 		goto out;
977 	}
978 
979 	if (unlikely(failed_wr != &first->s_wr)) {
980 		printk(KERN_WARNING "RDS/IB: ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
981 		BUG_ON(failed_wr != &first->s_wr);
982 	}
983 
984 
985 out:
986 	return ret;
987 }
988 
989 void rds_ib_xmit_complete(struct rds_connection *conn)
990 {
991 	struct rds_ib_connection *ic = conn->c_transport_data;
992 
993 	/* We may have a pending ACK or window update we were unable
994 	 * to send previously (due to flow control). Try again. */
995 	rds_ib_attempt_ack(ic);
996 }
997