xref: /titanic_51/usr/src/uts/common/io/ib/clients/rdsv3/ib_send.c (revision fe817b6022080da0a98b5d2d8cd179f594d6ca5e)
1 /*
2  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
3  */
4 
5 /*
6  * This file contains code imported from the OFED rds source file ib_send.c
7  * Oracle elects to have and use the contents of ib_send.c under and governed
8  * by the OpenIB.org BSD license (see below for full license text). However,
9  * the following notice accompanied the original version of this file:
10  */
11 
12 /*
13  * Copyright (c) 2006 Oracle.  All rights reserved.
14  *
15  * This software is available to you under a choice of one of two
16  * licenses.  You may choose to be licensed under the terms of the GNU
17  * General Public License (GPL) Version 2, available from the file
18  * COPYING in the main directory of this source tree, or the
19  * OpenIB.org BSD license below:
20  *
21  *     Redistribution and use in source and binary forms, with or
22  *     without modification, are permitted provided that the following
23  *     conditions are met:
24  *
25  *      - Redistributions of source code must retain the above
26  *        copyright notice, this list of conditions and the following
27  *        disclaimer.
28  *
29  *      - Redistributions in binary form must reproduce the above
30  *        copyright notice, this list of conditions and the following
31  *        disclaimer in the documentation and/or other materials
32  *        provided with the distribution.
33  *
34  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
35  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
36  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
37  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
38  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
39  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
40  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
41  * SOFTWARE.
42  *
43  */
44 #include <sys/rds.h>
45 
46 #include <sys/ib/clients/rdsv3/rdsv3.h>
47 #include <sys/ib/clients/rdsv3/rdma.h>
48 #include <sys/ib/clients/rdsv3/ib.h>
49 #include <sys/ib/clients/rdsv3/rdsv3_debug.h>
50 
51 static void
52 rdsv3_ib_send_rdma_complete(struct rdsv3_message *rm,
53     int wc_status)
54 {
55 	int notify_status;
56 
57 	RDSV3_DPRINTF4("rdsv3_ib_send_rdma_complete", "rm: %p, wc_status: %d",
58 	    rm, wc_status);
59 
60 	switch (wc_status) {
61 	case IBT_WC_WR_FLUSHED_ERR:
62 		return;
63 
64 	case IBT_WC_SUCCESS:
65 		notify_status = RDS_RDMA_SUCCESS;
66 		break;
67 
68 	case IBT_WC_REMOTE_ACCESS_ERR:
69 		notify_status = RDS_RDMA_REMOTE_ERROR;
70 		break;
71 
72 	default:
73 		notify_status = RDS_RDMA_OTHER_ERROR;
74 		break;
75 	}
76 	rdsv3_rdma_send_complete(rm, notify_status);
77 
78 	RDSV3_DPRINTF4("rdsv3_ib_send_rdma_complete", "rm: %p, wc_status: %d",
79 	    rm, wc_status);
80 }
81 
82 static void rdsv3_ib_dma_unmap_sg_rdma(struct ib_device *dev,
83     uint_t num, struct rdsv3_rdma_sg scat[]);
84 
85 void
86 rdsv3_ib_send_unmap_rdma(struct rdsv3_ib_connection *ic,
87     struct rdsv3_rdma_op *op)
88 {
89 	RDSV3_DPRINTF4("rdsv3_ib_send_unmap_rdma", "ic: %p, op: %p", ic, op);
90 	if (op->r_mapped) {
91 		op->r_mapped = 0;
92 		if (ic->i_cm_id) {
93 			rdsv3_ib_dma_unmap_sg_rdma(ic->i_cm_id->device,
94 			    op->r_nents, op->r_rdma_sg);
95 		} else {
96 			rdsv3_ib_dma_unmap_sg_rdma((struct ib_device *)NULL,
97 			    op->r_nents, op->r_rdma_sg);
98 		}
99 	}
100 }
101 
102 static void
103 rdsv3_ib_send_unmap_rm(struct rdsv3_ib_connection *ic,
104     struct rdsv3_ib_send_work *send,
105     int wc_status)
106 {
107 	struct rdsv3_message *rm = send->s_rm;
108 
109 	RDSV3_DPRINTF4("rdsv3_ib_send_unmap_rm", "ic %p send %p rm %p\n",
110 	    ic, send, rm);
111 
112 	mutex_enter(&rm->m_rs_lock);
113 	if (rm->m_count) {
114 		rdsv3_ib_dma_unmap_sg(ic->i_cm_id->device,
115 		    rm->m_sg, rm->m_count);
116 		rm->m_count = 0;
117 	}
118 	mutex_exit(&rm->m_rs_lock);
119 
120 	if (rm->m_rdma_op != NULL) {
121 		rdsv3_ib_send_unmap_rdma(ic, rm->m_rdma_op);
122 
123 		/*
124 		 * If the user asked for a completion notification on this
125 		 * message, we can implement three different semantics:
126 		 *  1.	Notify when we received the ACK on the RDS message
127 		 *	that was queued with the RDMA. This provides reliable
128 		 *	notification of RDMA status at the expense of a one-way
129 		 *	packet delay.
130 		 *  2.	Notify when the IB stack gives us the completion
131 		 *	event for the RDMA operation.
132 		 *  3.	Notify when the IB stack gives us the completion
133 		 *	event for the accompanying RDS messages.
134 		 * Here, we implement approach #3. To implement approach #2,
135 		 * call rdsv3_rdma_send_complete from the cq_handler.
136 		 * To implement #1,
137 		 * don't call rdsv3_rdma_send_complete at all, and fall back to
138 		 * the notify
139 		 * handling in the ACK processing code.
140 		 *
141 		 * Note: There's no need to explicitly sync any RDMA buffers
142 		 * using
143 		 * ib_dma_sync_sg_for_cpu - the completion for the RDMA
144 		 * operation itself unmapped the RDMA buffers, which takes care
145 		 * of synching.
146 		 */
147 		rdsv3_ib_send_rdma_complete(rm, wc_status);
148 
149 		if (rm->m_rdma_op->r_write)
150 			rdsv3_stats_add(s_send_rdma_bytes,
151 			    rm->m_rdma_op->r_bytes);
152 		else
153 			rdsv3_stats_add(s_recv_rdma_bytes,
154 			    rm->m_rdma_op->r_bytes);
155 	}
156 
157 	/*
158 	 * If anyone waited for this message to get flushed out, wake
159 	 * them up now
160 	 */
161 	rdsv3_message_unmapped(rm);
162 
163 	rdsv3_message_put(rm);
164 	send->s_rm = NULL;
165 }
166 
167 void
168 rdsv3_ib_send_init_ring(struct rdsv3_ib_connection *ic)
169 {
170 	struct rdsv3_ib_send_work *send;
171 	uint32_t i;
172 
173 	RDSV3_DPRINTF4("rdsv3_ib_send_init_ring", "ic: %p", ic);
174 
175 	for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
176 		send->s_rm = NULL;
177 		send->s_op = NULL;
178 	}
179 }
180 
181 void
182 rdsv3_ib_send_clear_ring(struct rdsv3_ib_connection *ic)
183 {
184 	struct rdsv3_ib_send_work *send;
185 	uint32_t i;
186 
187 	RDSV3_DPRINTF4("rdsv3_ib_send_clear_ring", "ic: %p", ic);
188 
189 	for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
190 		if (send->s_opcode == 0xdd)
191 			continue;
192 		if (send->s_rm)
193 			rdsv3_ib_send_unmap_rm(ic, send, IBT_WC_WR_FLUSHED_ERR);
194 		if (send->s_op)
195 			rdsv3_ib_send_unmap_rdma(ic, send->s_op);
196 	}
197 
198 	RDSV3_DPRINTF4("rdsv3_ib_send_clear_ring", "Return: ic: %p", ic);
199 }
200 
201 /*
202  * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
203  * operations performed in the send path.  As the sender allocs and potentially
204  * unallocs the next free entry in the ring it doesn't alter which is
205  * the next to be freed, which is what this is concerned with.
206  */
207 void
208 rdsv3_ib_send_cqe_handler(struct rdsv3_ib_connection *ic, ibt_wc_t *wc)
209 {
210 	struct rdsv3_connection *conn = ic->conn;
211 	struct rdsv3_ib_send_work *send;
212 	uint32_t completed, polled;
213 	uint32_t oldest;
214 	uint32_t i = 0;
215 	int ret;
216 
217 	RDSV3_DPRINTF4("rdsv3_ib_send_cqe_handler",
218 	    "wc wc_id 0x%llx status %u byte_len %u imm_data %u\n",
219 	    (unsigned long long)wc->wc_id, wc->wc_status,
220 	    wc->wc_bytes_xfer, ntohl(wc->wc_immed_data));
221 
222 	rdsv3_ib_stats_inc(s_ib_tx_cq_event);
223 
224 	if (wc->wc_id == RDSV3_IB_ACK_WR_ID) {
225 		if (ic->i_ack_queued + HZ/2 < jiffies)
226 			rdsv3_ib_stats_inc(s_ib_tx_stalled);
227 		rdsv3_ib_ack_send_complete(ic);
228 		return;
229 	}
230 
231 	oldest = rdsv3_ib_ring_oldest(&ic->i_send_ring);
232 
233 	completed = rdsv3_ib_ring_completed(&ic->i_send_ring,
234 	    (wc->wc_id & ~RDSV3_IB_SEND_OP), oldest);
235 
236 	for (i = 0; i < completed; i++) {
237 		send = &ic->i_sends[oldest];
238 
239 		/*
240 		 * In the error case, wc->opcode sometimes contains
241 		 * garbage
242 		 */
243 		switch (send->s_opcode) {
244 		case IBT_WRC_SEND:
245 			if (send->s_rm)
246 				rdsv3_ib_send_unmap_rm(ic, send,
247 				    wc->wc_status);
248 			break;
249 		case IBT_WRC_RDMAW:
250 		case IBT_WRC_RDMAR:
251 			/*
252 			 * Nothing to be done - the SG list will
253 			 * be unmapped
254 			 * when the SEND completes.
255 			 */
256 			break;
257 		default:
258 #ifndef __lock_lint
259 			RDSV3_DPRINTF2("rdsv3_ib_send_cq_comp_handler",
260 			    "RDS/IB: %s: unexpected opcode "
261 			    "0x%x in WR!",
262 			    __func__, send->s_opcode);
263 #endif
264 			break;
265 		}
266 
267 		send->s_opcode = 0xdd;
268 		if (send->s_queued + HZ/2 < jiffies)
269 			rdsv3_ib_stats_inc(s_ib_tx_stalled);
270 
271 		/*
272 		 * If a RDMA operation produced an error, signal
273 		 * this right
274 		 * away. If we don't, the subsequent SEND that goes
275 		 * with this
276 		 * RDMA will be canceled with ERR_WFLUSH, and the
277 		 * application
278 		 * never learn that the RDMA failed.
279 		 */
280 		if (wc->wc_status ==
281 		    IBT_WC_REMOTE_ACCESS_ERR && send->s_op) {
282 			struct rdsv3_message *rm;
283 
284 			rm = rdsv3_send_get_message(conn, send->s_op);
285 			if (rm) {
286 				if (rm->m_rdma_op != NULL)
287 					rdsv3_ib_send_unmap_rdma(ic,
288 					    rm->m_rdma_op);
289 				rdsv3_ib_send_rdma_complete(rm,
290 				    wc->wc_status);
291 				rdsv3_message_put(rm);
292 			}
293 		}
294 
295 		oldest = (oldest + 1) % ic->i_send_ring.w_nr;
296 	}
297 
298 	rdsv3_ib_ring_free(&ic->i_send_ring, completed);
299 
300 	clear_bit(RDSV3_LL_SEND_FULL, &conn->c_flags);
301 
302 	/* We expect errors as the qp is drained during shutdown */
303 	if (wc->wc_status != IBT_WC_SUCCESS && rdsv3_conn_up(conn)) {
304 		RDSV3_DPRINTF2("rdsv3_ib_send_cqe_handler",
305 		    "send completion on %u.%u.%u.%u "
306 		    "had status %u, disconnecting and reconnecting\n",
307 		    NIPQUAD(conn->c_faddr), wc->wc_status);
308 		rdsv3_conn_drop(conn);
309 	}
310 
311 	RDSV3_DPRINTF4("rdsv3_ib_send_cqe_handler", "Return: conn: %p", ic);
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 reciever'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 rdsv3_ib_send_add_credits below).
336  *
337  * The RDS send code is essentially single-threaded; rdsv3_send_xmit
338  * grabs c_send_lock 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
359 rdsv3_ib_send_grab_credits(struct rdsv3_ib_connection *ic,
360     uint32_t wanted, uint32_t *adv_credits, int need_posted)
361 {
362 	unsigned int avail, posted, got = 0, advertise;
363 	long oldval, newval;
364 
365 	RDSV3_DPRINTF4("rdsv3_ib_send_grab_credits", "ic: %p, %d %d %d",
366 	    ic, wanted, *adv_credits, need_posted);
367 
368 	*adv_credits = 0;
369 	if (!ic->i_flowctl)
370 		return (wanted);
371 
372 try_again:
373 	advertise = 0;
374 	oldval = newval = atomic_get(&ic->i_credits);
375 	posted = IB_GET_POST_CREDITS(oldval);
376 	avail = IB_GET_SEND_CREDITS(oldval);
377 
378 	RDSV3_DPRINTF5("rdsv3_ib_send_grab_credits",
379 	    "wanted (%u): credits=%u posted=%u\n", 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 rdsv3_connection *conn = ic->i_cm_id->context;
387 
388 		/* Oops, there aren't that many credits left! */
389 		set_bit(RDSV3_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(posted, RDSV3_MAX_ADV_CREDIT);
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 
413 	RDSV3_DPRINTF4("rdsv3_ib_send_grab_credits", "ic: %p, %d %d %d",
414 	    ic, got, *adv_credits, need_posted);
415 
416 	return (got);
417 }
418 
419 void
420 rdsv3_ib_send_add_credits(struct rdsv3_connection *conn, unsigned int credits)
421 {
422 	struct rdsv3_ib_connection *ic = conn->c_transport_data;
423 
424 	if (credits == 0)
425 		return;
426 
427 	RDSV3_DPRINTF5("rdsv3_ib_send_add_credits",
428 	    "credits (%u): current=%u%s\n",
429 	    credits,
430 	    IB_GET_SEND_CREDITS(atomic_get(&ic->i_credits)),
431 	    test_bit(RDSV3_LL_SEND_FULL, &conn->c_flags) ?
432 	    ", ll_send_full" : "");
433 
434 	atomic_add_32(&ic->i_credits, IB_SET_SEND_CREDITS(credits));
435 	if (test_and_clear_bit(RDSV3_LL_SEND_FULL, &conn->c_flags))
436 		rdsv3_queue_delayed_work(rdsv3_wq, &conn->c_send_w, 0);
437 
438 	ASSERT(!(IB_GET_SEND_CREDITS(credits) >= 16384));
439 
440 	rdsv3_ib_stats_inc(s_ib_rx_credit_updates);
441 
442 	RDSV3_DPRINTF4("rdsv3_ib_send_add_credits",
443 	    "Return: conn: %p, credits: %d",
444 	    conn, credits);
445 }
446 
447 void
448 rdsv3_ib_advertise_credits(struct rdsv3_connection *conn, unsigned int posted)
449 {
450 	struct rdsv3_ib_connection *ic = conn->c_transport_data;
451 
452 	RDSV3_DPRINTF4("rdsv3_ib_advertise_credits", "conn: %p, posted: %d",
453 	    conn, posted);
454 
455 	if (posted == 0)
456 		return;
457 
458 	atomic_add_32(&ic->i_credits, IB_SET_POST_CREDITS(posted));
459 
460 	/*
461 	 * Decide whether to send an update to the peer now.
462 	 * If we would send a credit update for every single buffer we
463 	 * post, we would end up with an ACK storm (ACK arrives,
464 	 * consumes buffer, we refill the ring, send ACK to remote
465 	 * advertising the newly posted buffer... ad inf)
466 	 *
467 	 * Performance pretty much depends on how often we send
468 	 * credit updates - too frequent updates mean lots of ACKs.
469 	 * Too infrequent updates, and the peer will run out of
470 	 * credits and has to throttle.
471 	 * For the time being, 16 seems to be a good compromise.
472 	 */
473 	if (IB_GET_POST_CREDITS(atomic_get(&ic->i_credits)) >= 16)
474 		set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
475 }
476 
477 static inline void
478 rdsv3_ib_xmit_populate_wr(struct rdsv3_ib_connection *ic,
479     ibt_send_wr_t *wr, unsigned int pos,
480     struct rdsv3_scatterlist *scat, unsigned int off, unsigned int length,
481     int send_flags)
482 {
483 	ibt_wr_ds_t *sge;
484 
485 	RDSV3_DPRINTF4("rdsv3_ib_xmit_populate_wr",
486 	    "ic: %p, wr: %p scat: %p %d %d %d %d",
487 	    ic, wr, scat, pos, off, length, send_flags);
488 
489 	wr->wr_id = pos | RDSV3_IB_SEND_OP;
490 	wr->wr_trans = IBT_RC_SRV;
491 	wr->wr_flags = send_flags;
492 	wr->wr_opcode = IBT_WRC_SEND;
493 
494 	if (length != 0) {
495 		int	ix, len, assigned;
496 		ibt_wr_ds_t *sgl;
497 
498 		ASSERT(length <= scat->length - off);
499 
500 		sgl = scat->sgl;
501 		if (off != 0) {
502 			/* find the right sgl to begin with */
503 			while (sgl->ds_len <= off) {
504 				off -= sgl->ds_len;
505 				sgl++;
506 			}
507 		}
508 
509 		ix = 1; /* first data sgl is at 1 */
510 		assigned = 0;
511 		len = length;
512 		do {
513 			sge = &wr->wr_sgl[ix++];
514 			sge->ds_va = sgl->ds_va + off;
515 			assigned = min(len, sgl->ds_len - off);
516 			sge->ds_len = assigned;
517 			sge->ds_key = sgl->ds_key;
518 			len -= assigned;
519 			if (len != 0) {
520 				sgl++;
521 				off = 0;
522 			}
523 		} while (len > 0);
524 
525 		wr->wr_nds = ix;
526 	} else {
527 		/*
528 		 * We're sending a packet with no payload. There is only
529 		 * one SGE
530 		 */
531 		wr->wr_nds = 1;
532 	}
533 
534 	sge = &wr->wr_sgl[0];
535 	sge->ds_va = ic->i_send_hdrs_dma + (pos * sizeof (struct rdsv3_header));
536 	sge->ds_len = sizeof (struct rdsv3_header);
537 	sge->ds_key = ic->i_mr->lkey;
538 
539 	RDSV3_DPRINTF4("rdsv3_ib_xmit_populate_wr",
540 	    "Return: ic: %p, wr: %p scat: %p", ic, wr, scat);
541 }
542 
543 /*
544  * This can be called multiple times for a given message.  The first time
545  * we see a message we map its scatterlist into the IB device so that
546  * we can provide that mapped address to the IB scatter gather entries
547  * in the IB work requests.  We translate the scatterlist into a series
548  * of work requests that fragment the message.  These work requests complete
549  * in order so we pass ownership of the message to the completion handler
550  * once we send the final fragment.
551  *
552  * The RDS core uses the c_send_lock to only enter this function once
553  * per connection.  This makes sure that the tx ring alloc/unalloc pairs
554  * don't get out of sync and confuse the ring.
555  */
556 int
557 rdsv3_ib_xmit(struct rdsv3_connection *conn, struct rdsv3_message *rm,
558     unsigned int hdr_off, unsigned int sg, unsigned int off)
559 {
560 	struct rdsv3_ib_connection *ic = conn->c_transport_data;
561 	struct ib_device *dev = ic->i_cm_id->device;
562 	struct rdsv3_ib_send_work *send = NULL;
563 	struct rdsv3_ib_send_work *first;
564 	struct rdsv3_ib_send_work *prev;
565 	ibt_send_wr_t *wr;
566 	struct rdsv3_scatterlist *scat;
567 	uint32_t pos;
568 	uint32_t i;
569 	uint32_t work_alloc;
570 	uint32_t credit_alloc;
571 	uint32_t posted;
572 	uint32_t adv_credits = 0;
573 	int send_flags = 0;
574 	int sent;
575 	int ret;
576 	int flow_controlled = 0;
577 
578 	RDSV3_DPRINTF4("rdsv3_ib_xmit", "conn: %p, rm: %p", conn, rm);
579 
580 	ASSERT(!(off % RDSV3_FRAG_SIZE));
581 	ASSERT(!(hdr_off != 0 && hdr_off != sizeof (struct rdsv3_header)));
582 
583 	/* Do not send cong updates to IB loopback */
584 	if (conn->c_loopback &&
585 	    rm->m_inc.i_hdr.h_flags & RDSV3_FLAG_CONG_BITMAP) {
586 		rdsv3_cong_map_updated(conn->c_fcong, ~(uint64_t)0);
587 		return (sizeof (struct rdsv3_header) + RDSV3_CONG_MAP_BYTES);
588 	}
589 
590 #ifndef __lock_lint
591 	/* FIXME we may overallocate here */
592 	if (ntohl(rm->m_inc.i_hdr.h_len) == 0)
593 		i = 1;
594 	else
595 		i = ceil(ntohl(rm->m_inc.i_hdr.h_len), RDSV3_FRAG_SIZE);
596 #endif
597 
598 	work_alloc = rdsv3_ib_ring_alloc(&ic->i_send_ring, i, &pos);
599 	if (work_alloc != i) {
600 		rdsv3_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
601 		set_bit(RDSV3_LL_SEND_FULL, &conn->c_flags);
602 		rdsv3_ib_stats_inc(s_ib_tx_ring_full);
603 		ret = -ENOMEM;
604 		goto out;
605 	}
606 
607 	credit_alloc = work_alloc;
608 	if (ic->i_flowctl) {
609 		credit_alloc = rdsv3_ib_send_grab_credits(ic, work_alloc,
610 		    &posted, 0);
611 		adv_credits += posted;
612 		if (credit_alloc < work_alloc) {
613 			rdsv3_ib_ring_unalloc(&ic->i_send_ring,
614 			    work_alloc - credit_alloc);
615 			work_alloc = credit_alloc;
616 			flow_controlled++;
617 		}
618 		if (work_alloc == 0) {
619 			rdsv3_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
620 			rdsv3_ib_stats_inc(s_ib_tx_throttle);
621 			ret = -ENOMEM;
622 			goto out;
623 		}
624 	}
625 
626 	/* map the message the first time we see it */
627 	if (ic->i_rm == NULL) {
628 		/*
629 		 * printk(KERN_NOTICE
630 		 * "rdsv3_ib_xmit prep msg dport=%u flags=0x%x len=%d\n",
631 		 * be16_to_cpu(rm->m_inc.i_hdr.h_dport),
632 		 * rm->m_inc.i_hdr.h_flags,
633 		 * be32_to_cpu(rm->m_inc.i_hdr.h_len));
634 		 */
635 		if (rm->m_nents) {
636 			rm->m_count = rdsv3_ib_dma_map_sg(dev,
637 			    rm->m_sg, rm->m_nents);
638 			RDSV3_DPRINTF5("rdsv3_ib_xmit",
639 			    "ic %p mapping rm %p: %d\n", ic, rm, rm->m_count);
640 			if (rm->m_count == 0) {
641 				rdsv3_ib_stats_inc(s_ib_tx_sg_mapping_failure);
642 				rdsv3_ib_ring_unalloc(&ic->i_send_ring,
643 				    work_alloc);
644 				ret = -ENOMEM; /* XXX ? */
645 				RDSV3_DPRINTF2("rdsv3_ib_xmit",
646 				    "fail: ic %p mapping rm %p: %d\n",
647 				    ic, rm, rm->m_count);
648 				goto out;
649 			}
650 		} else {
651 			rm->m_count = 0;
652 		}
653 
654 		ic->i_unsignaled_wrs = rdsv3_ib_sysctl_max_unsig_wrs;
655 		ic->i_unsignaled_bytes = rdsv3_ib_sysctl_max_unsig_bytes;
656 		rdsv3_message_addref(rm);
657 		ic->i_rm = rm;
658 
659 		/* Finalize the header */
660 		if (test_bit(RDSV3_MSG_ACK_REQUIRED, &rm->m_flags))
661 			rm->m_inc.i_hdr.h_flags |= RDSV3_FLAG_ACK_REQUIRED;
662 		if (test_bit(RDSV3_MSG_RETRANSMITTED, &rm->m_flags))
663 			rm->m_inc.i_hdr.h_flags |= RDSV3_FLAG_RETRANSMITTED;
664 
665 		/*
666 		 * If it has a RDMA op, tell the peer we did it. This is
667 		 * used by the peer to release use-once RDMA MRs.
668 		 */
669 		if (rm->m_rdma_op) {
670 			struct rdsv3_ext_header_rdma ext_hdr;
671 
672 			ext_hdr.h_rdma_rkey = htonl(rm->m_rdma_op->r_key);
673 			(void) rdsv3_message_add_extension(&rm->m_inc.i_hdr,
674 			    RDSV3_EXTHDR_RDMA, &ext_hdr,
675 			    sizeof (ext_hdr));
676 		}
677 		if (rm->m_rdma_cookie) {
678 			(void) rdsv3_message_add_rdma_dest_extension(
679 			    &rm->m_inc.i_hdr,
680 			    rdsv3_rdma_cookie_key(rm->m_rdma_cookie),
681 			    rdsv3_rdma_cookie_offset(rm->m_rdma_cookie));
682 		}
683 
684 		/*
685 		 * Note - rdsv3_ib_piggyb_ack clears the ACK_REQUIRED bit, so
686 		 * we should not do this unless we have a chance of at least
687 		 * sticking the header into the send ring. Which is why we
688 		 * should call rdsv3_ib_ring_alloc first.
689 		 */
690 		rm->m_inc.i_hdr.h_ack = htonll(rdsv3_ib_piggyb_ack(ic));
691 		rdsv3_message_make_checksum(&rm->m_inc.i_hdr);
692 
693 		/*
694 		 * Update adv_credits since we reset the ACK_REQUIRED bit.
695 		 */
696 		(void) rdsv3_ib_send_grab_credits(ic, 0, &posted, 1);
697 		adv_credits += posted;
698 		ASSERT(adv_credits <= 255);
699 	}
700 
701 	send = &ic->i_sends[pos];
702 	first = send;
703 	prev = NULL;
704 	scat = &rm->m_sg[sg];
705 	sent = 0;
706 	i = 0;
707 
708 	/*
709 	 * Sometimes you want to put a fence between an RDMA
710 	 * READ and the following SEND.
711 	 * We could either do this all the time
712 	 * or when requested by the user. Right now, we let
713 	 * the application choose.
714 	 */
715 	if (rm->m_rdma_op && rm->m_rdma_op->r_fence)
716 		send_flags = IBT_WR_SEND_FENCE;
717 
718 	/*
719 	 * We could be copying the header into the unused tail of the page.
720 	 * That would need to be changed in the future when those pages might
721 	 * be mapped userspace pages or page cache pages.  So instead we always
722 	 * use a second sge and our long-lived ring of mapped headers.  We send
723 	 * the header after the data so that the data payload can be aligned on
724 	 * the receiver.
725 	 */
726 
727 	/* handle a 0-len message */
728 	if (ntohl(rm->m_inc.i_hdr.h_len) == 0) {
729 		wr = &ic->i_send_wrs[0];
730 		rdsv3_ib_xmit_populate_wr(ic, wr, pos, NULL, 0, 0, send_flags);
731 		send->s_queued = jiffies;
732 		send->s_op = NULL;
733 		send->s_opcode = wr->wr_opcode;
734 		goto add_header;
735 	}
736 
737 	/* if there's data reference it with a chain of work reqs */
738 	for (; i < work_alloc && scat != &rm->m_sg[rm->m_count]; i++) {
739 		unsigned int len;
740 
741 		send = &ic->i_sends[pos];
742 
743 		wr = &ic->i_send_wrs[i];
744 		len = min(RDSV3_FRAG_SIZE,
745 		    rdsv3_ib_sg_dma_len(dev, scat) - off);
746 		rdsv3_ib_xmit_populate_wr(ic, wr, pos, scat, off, len,
747 		    send_flags);
748 		send->s_queued = jiffies;
749 		send->s_op = NULL;
750 		send->s_opcode = wr->wr_opcode;
751 
752 		/*
753 		 * We want to delay signaling completions just enough to get
754 		 * the batching benefits but not so much that we create dead
755 		 * time
756 		 * on the wire.
757 		 */
758 		if (ic->i_unsignaled_wrs-- == 0) {
759 			ic->i_unsignaled_wrs = rdsv3_ib_sysctl_max_unsig_wrs;
760 			wr->wr_flags |=
761 			    IBT_WR_SEND_SIGNAL | IBT_WR_SEND_SOLICIT;
762 		}
763 
764 		ic->i_unsignaled_bytes -= len;
765 		if (ic->i_unsignaled_bytes <= 0) {
766 			ic->i_unsignaled_bytes =
767 			    rdsv3_ib_sysctl_max_unsig_bytes;
768 			wr->wr_flags |=
769 			    IBT_WR_SEND_SIGNAL | IBT_WR_SEND_SOLICIT;
770 		}
771 
772 		/*
773 		 * Always signal the last one if we're stopping due to flow
774 		 * control.
775 		 */
776 		if (flow_controlled && i == (work_alloc-1)) {
777 			wr->wr_flags |=
778 			    IBT_WR_SEND_SIGNAL | IBT_WR_SEND_SOLICIT;
779 		}
780 
781 		RDSV3_DPRINTF5("rdsv3_ib_xmit", "send %p wr %p num_sge %u \n",
782 		    send, wr, wr->wr_nds);
783 
784 		sent += len;
785 		off += len;
786 		if (off == rdsv3_ib_sg_dma_len(dev, scat)) {
787 			scat++;
788 			off = 0;
789 		}
790 
791 add_header:
792 		/*
793 		 * Tack on the header after the data. The header SGE
794 		 * should already
795 		 * have been set up to point to the right header buffer.
796 		 */
797 		(void) memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr,
798 		    sizeof (struct rdsv3_header));
799 
800 		if (0) {
801 			struct rdsv3_header *hdr = &ic->i_send_hdrs[pos];
802 
803 			RDSV3_DPRINTF2("rdsv3_ib_xmit",
804 			    "send WR dport=%u flags=0x%x len=%d",
805 			    ntohs(hdr->h_dport),
806 			    hdr->h_flags,
807 			    ntohl(hdr->h_len));
808 		}
809 		if (adv_credits) {
810 			struct rdsv3_header *hdr = &ic->i_send_hdrs[pos];
811 
812 			/* add credit and redo the header checksum */
813 			hdr->h_credit = adv_credits;
814 			rdsv3_message_make_checksum(hdr);
815 			adv_credits = 0;
816 			rdsv3_ib_stats_inc(s_ib_tx_credit_updates);
817 		}
818 
819 		prev = send;
820 
821 		pos = (pos + 1) % ic->i_send_ring.w_nr;
822 	}
823 
824 	/*
825 	 * Account the RDS header in the number of bytes we sent, but just once.
826 	 * The caller has no concept of fragmentation.
827 	 */
828 	if (hdr_off == 0)
829 		sent += sizeof (struct rdsv3_header);
830 
831 	/* if we finished the message then send completion owns it */
832 	if (scat == &rm->m_sg[rm->m_count]) {
833 		prev->s_rm = ic->i_rm;
834 		wr->wr_flags |= IBT_WR_SEND_SIGNAL | IBT_WR_SEND_SOLICIT;
835 		ic->i_rm = NULL;
836 	}
837 
838 	if (i < work_alloc) {
839 		rdsv3_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
840 		work_alloc = i;
841 	}
842 	if (ic->i_flowctl && i < credit_alloc)
843 		rdsv3_ib_send_add_credits(conn, credit_alloc - i);
844 
845 	/* XXX need to worry about failed_wr and partial sends. */
846 	ret = ibt_post_send(ib_get_ibt_channel_hdl(ic->i_cm_id),
847 	    ic->i_send_wrs, i, &posted);
848 	if (posted != i) {
849 		RDSV3_DPRINTF2("rdsv3_ib_xmit",
850 		    "ic %p first %p nwr: %d ret %d:%d",
851 		    ic, first, i, ret, posted);
852 	}
853 	if (ret) {
854 		RDSV3_DPRINTF2("rdsv3_ib_xmit",
855 		    "RDS/IB: ib_post_send to %u.%u.%u.%u "
856 		    "returned %d\n", NIPQUAD(conn->c_faddr), ret);
857 		rdsv3_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
858 		if (prev->s_rm) {
859 			ic->i_rm = prev->s_rm;
860 			prev->s_rm = NULL;
861 		}
862 		RDSV3_DPRINTF2("rdsv3_ib_xmit", "ibt_post_send failed\n");
863 		rdsv3_conn_drop(ic->conn);
864 		ret = -EAGAIN;
865 		goto out;
866 	}
867 
868 	ret = sent;
869 
870 	RDSV3_DPRINTF4("rdsv3_ib_xmit", "Return: conn: %p, rm: %p", conn, rm);
871 out:
872 	ASSERT(!adv_credits);
873 	return (ret);
874 }
875 
876 static void
877 rdsv3_ib_dma_unmap_sg_rdma(struct ib_device *dev, uint_t num,
878 	struct rdsv3_rdma_sg scat[])
879 {
880 	ibt_hca_hdl_t hca_hdl;
881 	int i;
882 	int num_sgl;
883 
884 	RDSV3_DPRINTF4("rdsv3_ib_dma_unmap_sg", "rdma_sg: %p", scat);
885 
886 	if (dev) {
887 		hca_hdl = ib_get_ibt_hca_hdl(dev);
888 	} else {
889 		hca_hdl = scat[0].hca_hdl;
890 		RDSV3_DPRINTF2("rdsv3_ib_dma_unmap_sg_rdma",
891 		    "NULL dev use cached hca_hdl %p", hca_hdl);
892 	}
893 
894 	if (hca_hdl == NULL)
895 		return;
896 	scat[0].hca_hdl = NULL;
897 
898 	for (i = 0; i < num; i++) {
899 		if (scat[i].mihdl != NULL) {
900 			num_sgl = (scat[i].iovec.bytes / PAGESIZE) + 2;
901 			kmem_free(scat[i].swr.wr_sgl,
902 			    (num_sgl * sizeof (ibt_wr_ds_t)));
903 			scat[i].swr.wr_sgl = NULL;
904 			(void) ibt_unmap_mem_iov(hca_hdl, scat[i].mihdl);
905 			scat[i].mihdl = NULL;
906 		} else
907 			break;
908 	}
909 }
910 
911 /* ARGSUSED */
912 uint_t
913 rdsv3_ib_dma_map_sg_rdma(struct ib_device *dev, struct rdsv3_rdma_sg scat[],
914     uint_t num, struct rdsv3_scatterlist **scatl)
915 {
916 	ibt_hca_hdl_t hca_hdl;
917 	ibt_iov_attr_t iov_attr;
918 	struct buf *bp;
919 	uint_t i, j, k;
920 	uint_t count;
921 	struct rdsv3_scatterlist *sg;
922 	int ret;
923 
924 	RDSV3_DPRINTF4("rdsv3_ib_dma_map_sg_rdma", "scat: %p, num: %d",
925 	    scat, num);
926 
927 	hca_hdl = ib_get_ibt_hca_hdl(dev);
928 	scat[0].hca_hdl = hca_hdl;
929 	bzero(&iov_attr, sizeof (ibt_iov_attr_t));
930 	iov_attr.iov_flags = IBT_IOV_BUF;
931 	iov_attr.iov_lso_hdr_sz = 0;
932 
933 	for (i = 0, count = 0; i < num; i++) {
934 		/* transpose umem_cookie  to buf structure */
935 		bp = ddi_umem_iosetup(scat[i].umem_cookie,
936 		    scat[i].iovec.addr & PAGEOFFSET, scat[i].iovec.bytes,
937 		    B_WRITE, 0, 0, NULL, DDI_UMEM_SLEEP);
938 		if (bp == NULL) {
939 			/* free resources  and return error */
940 			goto out;
941 		}
942 		/* setup ibt_map_mem_iov() attributes */
943 		iov_attr.iov_buf = bp;
944 		iov_attr.iov_wr_nds = (scat[i].iovec.bytes / PAGESIZE) + 2;
945 		scat[i].swr.wr_sgl =
946 		    kmem_zalloc(iov_attr.iov_wr_nds * sizeof (ibt_wr_ds_t),
947 		    KM_SLEEP);
948 
949 		ret = ibt_map_mem_iov(hca_hdl, &iov_attr,
950 		    (ibt_all_wr_t *)&scat[i].swr, &scat[i].mihdl);
951 		freerbuf(bp);
952 		if (ret != IBT_SUCCESS) {
953 			RDSV3_DPRINTF2("rdsv3_ib_dma_map_sg_rdma",
954 			    "ibt_map_mem_iov returned: %d", ret);
955 			/* free resources and return error */
956 			kmem_free(scat[i].swr.wr_sgl,
957 			    iov_attr.iov_wr_nds * sizeof (ibt_wr_ds_t));
958 			goto out;
959 		}
960 		count += scat[i].swr.wr_nds;
961 
962 #ifdef  DEBUG
963 		for (j = 0; j < scat[i].swr.wr_nds; j++) {
964 			RDSV3_DPRINTF5("rdsv3_ib_dma_map_sg_rdma",
965 			    "sgl[%d] va %llx len %x", j,
966 			    scat[i].swr.wr_sgl[j].ds_va,
967 			    scat[i].swr.wr_sgl[j].ds_len);
968 		}
969 #endif
970 		RDSV3_DPRINTF4("rdsv3_ib_dma_map_sg_rdma",
971 		    "iovec.bytes: 0x%x scat[%d]swr.wr_nds: %d",
972 		    scat[i].iovec.bytes, i, scat[i].swr.wr_nds);
973 	}
974 
975 	count = ((count - 1) / RDSV3_IB_MAX_SGE) + 1;
976 	RDSV3_DPRINTF4("rdsv3_ib_dma_map_sg_rdma", "Ret: num: %d", count);
977 	return (count);
978 
979 out:
980 	rdsv3_ib_dma_unmap_sg_rdma(dev, num, scat);
981 	return (0);
982 }
983 
984 int
985 rdsv3_ib_xmit_rdma(struct rdsv3_connection *conn, struct rdsv3_rdma_op *op)
986 {
987 	struct rdsv3_ib_connection *ic = conn->c_transport_data;
988 	struct rdsv3_ib_send_work *send = NULL;
989 	struct rdsv3_rdma_sg *scat;
990 	uint64_t remote_addr;
991 	uint32_t pos;
992 	uint32_t work_alloc;
993 	uint32_t i, j, k, idx;
994 	uint32_t left, count;
995 	uint32_t posted;
996 	int sent;
997 	ibt_status_t status;
998 	ibt_send_wr_t *wr;
999 	ibt_wr_ds_t *sge;
1000 
1001 	RDSV3_DPRINTF4("rdsv3_ib_xmit_rdma", "rdsv3_ib_conn: %p", ic);
1002 
1003 	/* map the message the first time we see it */
1004 	if (!op->r_mapped) {
1005 		op->r_count = rdsv3_ib_dma_map_sg_rdma(ic->i_cm_id->device,
1006 		    op->r_rdma_sg, op->r_nents, &op->r_sg);
1007 		RDSV3_DPRINTF5("rdsv3_ib_xmit_rdma", "ic %p mapping op %p: %d",
1008 		    ic, op, op->r_count);
1009 		if (op->r_count == 0) {
1010 			rdsv3_ib_stats_inc(s_ib_tx_sg_mapping_failure);
1011 			RDSV3_DPRINTF2("rdsv3_ib_xmit_rdma",
1012 			    "fail: ic %p mapping op %p: %d",
1013 			    ic, op, op->r_count);
1014 			return (-ENOMEM); /* XXX ? */
1015 		}
1016 		op->r_mapped = 1;
1017 	}
1018 
1019 	/*
1020 	 * Instead of knowing how to return a partial rdma read/write
1021 	 * we insist that there
1022 	 * be enough work requests to send the entire message.
1023 	 */
1024 	work_alloc = rdsv3_ib_ring_alloc(&ic->i_send_ring, op->r_count, &pos);
1025 	if (work_alloc != op->r_count) {
1026 		rdsv3_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
1027 		rdsv3_ib_stats_inc(s_ib_tx_ring_full);
1028 		return (-ENOMEM);
1029 	}
1030 
1031 	RDSV3_DPRINTF4("rdsv3_ib_xmit_rdma", "pos %u cnt %u", pos, op->r_count);
1032 	/*
1033 	 * take the scatter list and transpose into a list of
1034 	 * send wr's each with a scatter list of RDSV3_IB_MAX_SGE
1035 	 */
1036 	scat = &op->r_rdma_sg[0];
1037 	sent = 0;
1038 	remote_addr = op->r_remote_addr;
1039 
1040 	for (i = 0, k = 0; i < op->r_nents; i++) {
1041 		left = scat[i].swr.wr_nds;
1042 		for (idx = 0; left > 0; k++) {
1043 			send = &ic->i_sends[pos];
1044 			send->s_queued = jiffies;
1045 			send->s_opcode = op->r_write ? IBT_WRC_RDMAW :
1046 			    IBT_WRC_RDMAR;
1047 			send->s_op = op;
1048 
1049 			wr = &ic->i_send_wrs[k];
1050 			wr->wr_flags = 0;
1051 			wr->wr_id = pos | RDSV3_IB_SEND_OP;
1052 			wr->wr_trans = IBT_RC_SRV;
1053 			wr->wr_opcode = op->r_write ? IBT_WRC_RDMAW :
1054 			    IBT_WRC_RDMAR;
1055 			wr->wr.rc.rcwr.rdma.rdma_raddr = remote_addr;
1056 			wr->wr.rc.rcwr.rdma.rdma_rkey = op->r_key;
1057 
1058 			if (left > RDSV3_IB_MAX_SGE) {
1059 				count = RDSV3_IB_MAX_SGE;
1060 				left -= RDSV3_IB_MAX_SGE;
1061 			} else {
1062 				count = left;
1063 				left = 0;
1064 			}
1065 			wr->wr_nds = count;
1066 
1067 			for (j = 0; j < count; j++) {
1068 				sge = &wr->wr_sgl[j];
1069 				*sge = scat[i].swr.wr_sgl[idx];
1070 				remote_addr += scat[i].swr.wr_sgl[idx].ds_len;
1071 				sent += scat[i].swr.wr_sgl[idx].ds_len;
1072 				idx++;
1073 				RDSV3_DPRINTF5("xmit_rdma",
1074 				    "send_wrs[%d]sgl[%d] va %llx len %x",
1075 				    k, j, sge->ds_va, sge->ds_len);
1076 			}
1077 			RDSV3_DPRINTF5("rdsv3_ib_xmit_rdma",
1078 			    "wr[%d] %p key: %x code: %d tlen: %d",
1079 			    k, wr, wr->wr.rc.rcwr.rdma.rdma_rkey,
1080 			    wr->wr_opcode, sent);
1081 
1082 			/*
1083 			 * We want to delay signaling completions just enough
1084 			 * to get the batching benefits but not so much that
1085 			 * we create dead time on the wire.
1086 			 */
1087 			if (ic->i_unsignaled_wrs-- == 0) {
1088 				ic->i_unsignaled_wrs =
1089 				    rdsv3_ib_sysctl_max_unsig_wrs;
1090 				wr->wr_flags = IBT_WR_SEND_SIGNAL;
1091 			}
1092 
1093 			pos = (pos + 1) % ic->i_send_ring.w_nr;
1094 		}
1095 	}
1096 
1097 	status = ibt_post_send(ib_get_ibt_channel_hdl(ic->i_cm_id),
1098 	    ic->i_send_wrs, k, &posted);
1099 	if (status != IBT_SUCCESS) {
1100 		RDSV3_DPRINTF2("rdsv3_ib_xmit_rdma",
1101 		    "RDS/IB: rdma ib_post_send to %u.%u.%u.%u "
1102 		    "returned %d", NIPQUAD(conn->c_faddr), status);
1103 		rdsv3_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
1104 	}
1105 	RDSV3_DPRINTF4("rdsv3_ib_xmit_rdma", "Ret: %p", ic);
1106 	return (status);
1107 }
1108 
1109 void
1110 rdsv3_ib_xmit_complete(struct rdsv3_connection *conn)
1111 {
1112 	struct rdsv3_ib_connection *ic = conn->c_transport_data;
1113 
1114 	RDSV3_DPRINTF4("rdsv3_ib_xmit_complete", "conn: %p", conn);
1115 
1116 	/*
1117 	 * We may have a pending ACK or window update we were unable
1118 	 * to send previously (due to flow control). Try again.
1119 	 */
1120 	rdsv3_ib_attempt_ack(ic);
1121 }
1122