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 static char *rds_ib_wc_status_strings[] = {
43 #define RDS_IB_WC_STATUS_STR(foo) \
44 		[IB_WC_##foo] = __stringify(IB_WC_##foo)
45 	RDS_IB_WC_STATUS_STR(SUCCESS),
46 	RDS_IB_WC_STATUS_STR(LOC_LEN_ERR),
47 	RDS_IB_WC_STATUS_STR(LOC_QP_OP_ERR),
48 	RDS_IB_WC_STATUS_STR(LOC_EEC_OP_ERR),
49 	RDS_IB_WC_STATUS_STR(LOC_PROT_ERR),
50 	RDS_IB_WC_STATUS_STR(WR_FLUSH_ERR),
51 	RDS_IB_WC_STATUS_STR(MW_BIND_ERR),
52 	RDS_IB_WC_STATUS_STR(BAD_RESP_ERR),
53 	RDS_IB_WC_STATUS_STR(LOC_ACCESS_ERR),
54 	RDS_IB_WC_STATUS_STR(REM_INV_REQ_ERR),
55 	RDS_IB_WC_STATUS_STR(REM_ACCESS_ERR),
56 	RDS_IB_WC_STATUS_STR(REM_OP_ERR),
57 	RDS_IB_WC_STATUS_STR(RETRY_EXC_ERR),
58 	RDS_IB_WC_STATUS_STR(RNR_RETRY_EXC_ERR),
59 	RDS_IB_WC_STATUS_STR(LOC_RDD_VIOL_ERR),
60 	RDS_IB_WC_STATUS_STR(REM_INV_RD_REQ_ERR),
61 	RDS_IB_WC_STATUS_STR(REM_ABORT_ERR),
62 	RDS_IB_WC_STATUS_STR(INV_EECN_ERR),
63 	RDS_IB_WC_STATUS_STR(INV_EEC_STATE_ERR),
64 	RDS_IB_WC_STATUS_STR(FATAL_ERR),
65 	RDS_IB_WC_STATUS_STR(RESP_TIMEOUT_ERR),
66 	RDS_IB_WC_STATUS_STR(GENERAL_ERR),
67 #undef RDS_IB_WC_STATUS_STR
68 };
69 
rds_ib_wc_status_str(enum ib_wc_status status)70 char *rds_ib_wc_status_str(enum ib_wc_status status)
71 {
72 	return rds_str_array(rds_ib_wc_status_strings,
73 			     ARRAY_SIZE(rds_ib_wc_status_strings), status);
74 }
75 
76 /*
77  * Convert IB-specific error message to RDS error message and call core
78  * completion handler.
79  */
rds_ib_send_complete(struct rds_message * rm,int wc_status,void (* complete)(struct rds_message * rm,int status))80 static void rds_ib_send_complete(struct rds_message *rm,
81 				 int wc_status,
82 				 void (*complete)(struct rds_message *rm, int status))
83 {
84 	int notify_status;
85 
86 	switch (wc_status) {
87 	case IB_WC_WR_FLUSH_ERR:
88 		return;
89 
90 	case IB_WC_SUCCESS:
91 		notify_status = RDS_RDMA_SUCCESS;
92 		break;
93 
94 	case IB_WC_REM_ACCESS_ERR:
95 		notify_status = RDS_RDMA_REMOTE_ERROR;
96 		break;
97 
98 	default:
99 		notify_status = RDS_RDMA_OTHER_ERROR;
100 		break;
101 	}
102 	complete(rm, notify_status);
103 }
104 
rds_ib_send_unmap_data(struct rds_ib_connection * ic,struct rm_data_op * op,int wc_status)105 static void rds_ib_send_unmap_data(struct rds_ib_connection *ic,
106 				   struct rm_data_op *op,
107 				   int wc_status)
108 {
109 	if (op->op_nents)
110 		ib_dma_unmap_sg(ic->i_cm_id->device,
111 				op->op_sg, op->op_nents,
112 				DMA_TO_DEVICE);
113 }
114 
rds_ib_send_unmap_rdma(struct rds_ib_connection * ic,struct rm_rdma_op * op,int wc_status)115 static void rds_ib_send_unmap_rdma(struct rds_ib_connection *ic,
116 				   struct rm_rdma_op *op,
117 				   int wc_status)
118 {
119 	if (op->op_mapped) {
120 		ib_dma_unmap_sg(ic->i_cm_id->device,
121 				op->op_sg, op->op_nents,
122 				op->op_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
123 		op->op_mapped = 0;
124 	}
125 
126 	/* If the user asked for a completion notification on this
127 	 * message, we can implement three different semantics:
128 	 *  1.	Notify when we received the ACK on the RDS message
129 	 *	that was queued with the RDMA. This provides reliable
130 	 *	notification of RDMA status at the expense of a one-way
131 	 *	packet delay.
132 	 *  2.	Notify when the IB stack gives us the completion event for
133 	 *	the RDMA operation.
134 	 *  3.	Notify when the IB stack gives us the completion event for
135 	 *	the accompanying RDS messages.
136 	 * Here, we implement approach #3. To implement approach #2,
137 	 * we would need to take an event for the rdma WR. To implement #1,
138 	 * don't call rds_rdma_send_complete at all, and fall back to the notify
139 	 * handling in the ACK processing code.
140 	 *
141 	 * Note: There's no need to explicitly sync any RDMA buffers using
142 	 * ib_dma_sync_sg_for_cpu - the completion for the RDMA
143 	 * operation itself unmapped the RDMA buffers, which takes care
144 	 * of synching.
145 	 */
146 	rds_ib_send_complete(container_of(op, struct rds_message, rdma),
147 			     wc_status, rds_rdma_send_complete);
148 
149 	if (op->op_write)
150 		rds_stats_add(s_send_rdma_bytes, op->op_bytes);
151 	else
152 		rds_stats_add(s_recv_rdma_bytes, op->op_bytes);
153 }
154 
rds_ib_send_unmap_atomic(struct rds_ib_connection * ic,struct rm_atomic_op * op,int wc_status)155 static void rds_ib_send_unmap_atomic(struct rds_ib_connection *ic,
156 				     struct rm_atomic_op *op,
157 				     int wc_status)
158 {
159 	/* unmap atomic recvbuf */
160 	if (op->op_mapped) {
161 		ib_dma_unmap_sg(ic->i_cm_id->device, op->op_sg, 1,
162 				DMA_FROM_DEVICE);
163 		op->op_mapped = 0;
164 	}
165 
166 	rds_ib_send_complete(container_of(op, struct rds_message, atomic),
167 			     wc_status, rds_atomic_send_complete);
168 
169 	if (op->op_type == RDS_ATOMIC_TYPE_CSWP)
170 		rds_ib_stats_inc(s_ib_atomic_cswp);
171 	else
172 		rds_ib_stats_inc(s_ib_atomic_fadd);
173 }
174 
175 /*
176  * Unmap the resources associated with a struct send_work.
177  *
178  * Returns the rm for no good reason other than it is unobtainable
179  * other than by switching on wr.opcode, currently, and the caller,
180  * the event handler, needs it.
181  */
rds_ib_send_unmap_op(struct rds_ib_connection * ic,struct rds_ib_send_work * send,int wc_status)182 static struct rds_message *rds_ib_send_unmap_op(struct rds_ib_connection *ic,
183 						struct rds_ib_send_work *send,
184 						int wc_status)
185 {
186 	struct rds_message *rm = NULL;
187 
188 	/* In the error case, wc.opcode sometimes contains garbage */
189 	switch (send->s_wr.opcode) {
190 	case IB_WR_SEND:
191 		if (send->s_op) {
192 			rm = container_of(send->s_op, struct rds_message, data);
193 			rds_ib_send_unmap_data(ic, send->s_op, wc_status);
194 		}
195 		break;
196 	case IB_WR_RDMA_WRITE:
197 	case IB_WR_RDMA_READ:
198 		if (send->s_op) {
199 			rm = container_of(send->s_op, struct rds_message, rdma);
200 			rds_ib_send_unmap_rdma(ic, send->s_op, wc_status);
201 		}
202 		break;
203 	case IB_WR_ATOMIC_FETCH_AND_ADD:
204 	case IB_WR_ATOMIC_CMP_AND_SWP:
205 		if (send->s_op) {
206 			rm = container_of(send->s_op, struct rds_message, atomic);
207 			rds_ib_send_unmap_atomic(ic, send->s_op, wc_status);
208 		}
209 		break;
210 	default:
211 		printk_ratelimited(KERN_NOTICE
212 			       "RDS/IB: %s: unexpected opcode 0x%x in WR!\n",
213 			       __func__, send->s_wr.opcode);
214 		break;
215 	}
216 
217 	send->s_wr.opcode = 0xdead;
218 
219 	return rm;
220 }
221 
rds_ib_send_init_ring(struct rds_ib_connection * ic)222 void rds_ib_send_init_ring(struct rds_ib_connection *ic)
223 {
224 	struct rds_ib_send_work *send;
225 	u32 i;
226 
227 	for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
228 		struct ib_sge *sge;
229 
230 		send->s_op = NULL;
231 
232 		send->s_wr.wr_id = i;
233 		send->s_wr.sg_list = send->s_sge;
234 		send->s_wr.ex.imm_data = 0;
235 
236 		sge = &send->s_sge[0];
237 		sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header));
238 		sge->length = sizeof(struct rds_header);
239 		sge->lkey = ic->i_mr->lkey;
240 
241 		send->s_sge[1].lkey = ic->i_mr->lkey;
242 	}
243 }
244 
rds_ib_send_clear_ring(struct rds_ib_connection * ic)245 void rds_ib_send_clear_ring(struct rds_ib_connection *ic)
246 {
247 	struct rds_ib_send_work *send;
248 	u32 i;
249 
250 	for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
251 		if (send->s_op && send->s_wr.opcode != 0xdead)
252 			rds_ib_send_unmap_op(ic, send, IB_WC_WR_FLUSH_ERR);
253 	}
254 }
255 
256 /*
257  * The only fast path caller always has a non-zero nr, so we don't
258  * bother testing nr before performing the atomic sub.
259  */
rds_ib_sub_signaled(struct rds_ib_connection * ic,int nr)260 static void rds_ib_sub_signaled(struct rds_ib_connection *ic, int nr)
261 {
262 	if ((atomic_sub_return(nr, &ic->i_signaled_sends) == 0) &&
263 	    waitqueue_active(&rds_ib_ring_empty_wait))
264 		wake_up(&rds_ib_ring_empty_wait);
265 	BUG_ON(atomic_read(&ic->i_signaled_sends) < 0);
266 }
267 
268 /*
269  * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
270  * operations performed in the send path.  As the sender allocs and potentially
271  * unallocs the next free entry in the ring it doesn't alter which is
272  * the next to be freed, which is what this is concerned with.
273  */
rds_ib_send_cq_comp_handler(struct ib_cq * cq,void * context)274 void rds_ib_send_cq_comp_handler(struct ib_cq *cq, void *context)
275 {
276 	struct rds_connection *conn = context;
277 	struct rds_ib_connection *ic = conn->c_transport_data;
278 	struct rds_message *rm = NULL;
279 	struct ib_wc wc;
280 	struct rds_ib_send_work *send;
281 	u32 completed;
282 	u32 oldest;
283 	u32 i = 0;
284 	int ret;
285 	int nr_sig = 0;
286 
287 	rdsdebug("cq %p conn %p\n", cq, conn);
288 	rds_ib_stats_inc(s_ib_tx_cq_call);
289 	ret = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
290 	if (ret)
291 		rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
292 
293 	while (ib_poll_cq(cq, 1, &wc) > 0) {
294 		rdsdebug("wc wr_id 0x%llx status %u (%s) byte_len %u imm_data %u\n",
295 			 (unsigned long long)wc.wr_id, wc.status,
296 			 rds_ib_wc_status_str(wc.status), wc.byte_len,
297 			 be32_to_cpu(wc.ex.imm_data));
298 		rds_ib_stats_inc(s_ib_tx_cq_event);
299 
300 		if (wc.wr_id == RDS_IB_ACK_WR_ID) {
301 			if (ic->i_ack_queued + HZ/2 < jiffies)
302 				rds_ib_stats_inc(s_ib_tx_stalled);
303 			rds_ib_ack_send_complete(ic);
304 			continue;
305 		}
306 
307 		oldest = rds_ib_ring_oldest(&ic->i_send_ring);
308 
309 		completed = rds_ib_ring_completed(&ic->i_send_ring, wc.wr_id, oldest);
310 
311 		for (i = 0; i < completed; i++) {
312 			send = &ic->i_sends[oldest];
313 			if (send->s_wr.send_flags & IB_SEND_SIGNALED)
314 				nr_sig++;
315 
316 			rm = rds_ib_send_unmap_op(ic, send, wc.status);
317 
318 			if (send->s_queued + HZ/2 < jiffies)
319 				rds_ib_stats_inc(s_ib_tx_stalled);
320 
321 			if (send->s_op) {
322 				if (send->s_op == rm->m_final_op) {
323 					/* If anyone waited for this message to get flushed out, wake
324 					 * them up now */
325 					rds_message_unmapped(rm);
326 				}
327 				rds_message_put(rm);
328 				send->s_op = NULL;
329 			}
330 
331 			oldest = (oldest + 1) % ic->i_send_ring.w_nr;
332 		}
333 
334 		rds_ib_ring_free(&ic->i_send_ring, completed);
335 		rds_ib_sub_signaled(ic, nr_sig);
336 		nr_sig = 0;
337 
338 		if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
339 		    test_bit(0, &conn->c_map_queued))
340 			queue_delayed_work(rds_wq, &conn->c_send_w, 0);
341 
342 		/* We expect errors as the qp is drained during shutdown */
343 		if (wc.status != IB_WC_SUCCESS && rds_conn_up(conn)) {
344 			rds_ib_conn_error(conn, "send completion on %pI4 had status "
345 					  "%u (%s), disconnecting and reconnecting\n",
346 					  &conn->c_faddr, wc.status,
347 					  rds_ib_wc_status_str(wc.status));
348 		}
349 	}
350 }
351 
352 /*
353  * This is the main function for allocating credits when sending
354  * messages.
355  *
356  * Conceptually, we have two counters:
357  *  -	send credits: this tells us how many WRs we're allowed
358  *	to submit without overruning the receiver's queue. For
359  *	each SEND WR we post, we decrement this by one.
360  *
361  *  -	posted credits: this tells us how many WRs we recently
362  *	posted to the receive queue. This value is transferred
363  *	to the peer as a "credit update" in a RDS header field.
364  *	Every time we transmit credits to the peer, we subtract
365  *	the amount of transferred credits from this counter.
366  *
367  * It is essential that we avoid situations where both sides have
368  * exhausted their send credits, and are unable to send new credits
369  * to the peer. We achieve this by requiring that we send at least
370  * one credit update to the peer before exhausting our credits.
371  * When new credits arrive, we subtract one credit that is withheld
372  * until we've posted new buffers and are ready to transmit these
373  * credits (see rds_ib_send_add_credits below).
374  *
375  * The RDS send code is essentially single-threaded; rds_send_xmit
376  * sets RDS_IN_XMIT to ensure exclusive access to the send ring.
377  * However, the ACK sending code is independent and can race with
378  * message SENDs.
379  *
380  * In the send path, we need to update the counters for send credits
381  * and the counter of posted buffers atomically - when we use the
382  * last available credit, we cannot allow another thread to race us
383  * and grab the posted credits counter.  Hence, we have to use a
384  * spinlock to protect the credit counter, or use atomics.
385  *
386  * Spinlocks shared between the send and the receive path are bad,
387  * because they create unnecessary delays. An early implementation
388  * using a spinlock showed a 5% degradation in throughput at some
389  * loads.
390  *
391  * This implementation avoids spinlocks completely, putting both
392  * counters into a single atomic, and updating that atomic using
393  * atomic_add (in the receive path, when receiving fresh credits),
394  * and using atomic_cmpxchg when updating the two counters.
395  */
rds_ib_send_grab_credits(struct rds_ib_connection * ic,u32 wanted,u32 * adv_credits,int need_posted,int max_posted)396 int rds_ib_send_grab_credits(struct rds_ib_connection *ic,
397 			     u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
398 {
399 	unsigned int avail, posted, got = 0, advertise;
400 	long oldval, newval;
401 
402 	*adv_credits = 0;
403 	if (!ic->i_flowctl)
404 		return wanted;
405 
406 try_again:
407 	advertise = 0;
408 	oldval = newval = atomic_read(&ic->i_credits);
409 	posted = IB_GET_POST_CREDITS(oldval);
410 	avail = IB_GET_SEND_CREDITS(oldval);
411 
412 	rdsdebug("rds_ib_send_grab_credits(%u): credits=%u posted=%u\n",
413 			wanted, avail, posted);
414 
415 	/* The last credit must be used to send a credit update. */
416 	if (avail && !posted)
417 		avail--;
418 
419 	if (avail < wanted) {
420 		struct rds_connection *conn = ic->i_cm_id->context;
421 
422 		/* Oops, there aren't that many credits left! */
423 		set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
424 		got = avail;
425 	} else {
426 		/* Sometimes you get what you want, lalala. */
427 		got = wanted;
428 	}
429 	newval -= IB_SET_SEND_CREDITS(got);
430 
431 	/*
432 	 * If need_posted is non-zero, then the caller wants
433 	 * the posted regardless of whether any send credits are
434 	 * available.
435 	 */
436 	if (posted && (got || need_posted)) {
437 		advertise = min_t(unsigned int, posted, max_posted);
438 		newval -= IB_SET_POST_CREDITS(advertise);
439 	}
440 
441 	/* Finally bill everything */
442 	if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
443 		goto try_again;
444 
445 	*adv_credits = advertise;
446 	return got;
447 }
448 
rds_ib_send_add_credits(struct rds_connection * conn,unsigned int credits)449 void rds_ib_send_add_credits(struct rds_connection *conn, unsigned int credits)
450 {
451 	struct rds_ib_connection *ic = conn->c_transport_data;
452 
453 	if (credits == 0)
454 		return;
455 
456 	rdsdebug("rds_ib_send_add_credits(%u): current=%u%s\n",
457 			credits,
458 			IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
459 			test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
460 
461 	atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
462 	if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
463 		queue_delayed_work(rds_wq, &conn->c_send_w, 0);
464 
465 	WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
466 
467 	rds_ib_stats_inc(s_ib_rx_credit_updates);
468 }
469 
rds_ib_advertise_credits(struct rds_connection * conn,unsigned int posted)470 void rds_ib_advertise_credits(struct rds_connection *conn, unsigned int posted)
471 {
472 	struct rds_ib_connection *ic = conn->c_transport_data;
473 
474 	if (posted == 0)
475 		return;
476 
477 	atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
478 
479 	/* Decide whether to send an update to the peer now.
480 	 * If we would send a credit update for every single buffer we
481 	 * post, we would end up with an ACK storm (ACK arrives,
482 	 * consumes buffer, we refill the ring, send ACK to remote
483 	 * advertising the newly posted buffer... ad inf)
484 	 *
485 	 * Performance pretty much depends on how often we send
486 	 * credit updates - too frequent updates mean lots of ACKs.
487 	 * Too infrequent updates, and the peer will run out of
488 	 * credits and has to throttle.
489 	 * For the time being, 16 seems to be a good compromise.
490 	 */
491 	if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
492 		set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
493 }
494 
rds_ib_set_wr_signal_state(struct rds_ib_connection * ic,struct rds_ib_send_work * send,bool notify)495 static inline int rds_ib_set_wr_signal_state(struct rds_ib_connection *ic,
496 					     struct rds_ib_send_work *send,
497 					     bool notify)
498 {
499 	/*
500 	 * We want to delay signaling completions just enough to get
501 	 * the batching benefits but not so much that we create dead time
502 	 * on the wire.
503 	 */
504 	if (ic->i_unsignaled_wrs-- == 0 || notify) {
505 		ic->i_unsignaled_wrs = rds_ib_sysctl_max_unsig_wrs;
506 		send->s_wr.send_flags |= IB_SEND_SIGNALED;
507 		return 1;
508 	}
509 	return 0;
510 }
511 
512 /*
513  * This can be called multiple times for a given message.  The first time
514  * we see a message we map its scatterlist into the IB device so that
515  * we can provide that mapped address to the IB scatter gather entries
516  * in the IB work requests.  We translate the scatterlist into a series
517  * of work requests that fragment the message.  These work requests complete
518  * in order so we pass ownership of the message to the completion handler
519  * once we send the final fragment.
520  *
521  * The RDS core uses the c_send_lock to only enter this function once
522  * per connection.  This makes sure that the tx ring alloc/unalloc pairs
523  * don't get out of sync and confuse the ring.
524  */
rds_ib_xmit(struct rds_connection * conn,struct rds_message * rm,unsigned int hdr_off,unsigned int sg,unsigned int off)525 int rds_ib_xmit(struct rds_connection *conn, struct rds_message *rm,
526 		unsigned int hdr_off, unsigned int sg, unsigned int off)
527 {
528 	struct rds_ib_connection *ic = conn->c_transport_data;
529 	struct ib_device *dev = ic->i_cm_id->device;
530 	struct rds_ib_send_work *send = NULL;
531 	struct rds_ib_send_work *first;
532 	struct rds_ib_send_work *prev;
533 	struct ib_send_wr *failed_wr;
534 	struct scatterlist *scat;
535 	u32 pos;
536 	u32 i;
537 	u32 work_alloc;
538 	u32 credit_alloc = 0;
539 	u32 posted;
540 	u32 adv_credits = 0;
541 	int send_flags = 0;
542 	int bytes_sent = 0;
543 	int ret;
544 	int flow_controlled = 0;
545 	int nr_sig = 0;
546 
547 	BUG_ON(off % RDS_FRAG_SIZE);
548 	BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
549 
550 	/* Do not send cong updates to IB loopback */
551 	if (conn->c_loopback
552 	    && rm->m_inc.i_hdr.h_flags & RDS_FLAG_CONG_BITMAP) {
553 		rds_cong_map_updated(conn->c_fcong, ~(u64) 0);
554 		scat = &rm->data.op_sg[sg];
555 		ret = max_t(int, RDS_CONG_MAP_BYTES, scat->length);
556 		return sizeof(struct rds_header) + ret;
557 	}
558 
559 	/* FIXME we may overallocate here */
560 	if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
561 		i = 1;
562 	else
563 		i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
564 
565 	work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
566 	if (work_alloc == 0) {
567 		set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
568 		rds_ib_stats_inc(s_ib_tx_ring_full);
569 		ret = -ENOMEM;
570 		goto out;
571 	}
572 
573 	if (ic->i_flowctl) {
574 		credit_alloc = rds_ib_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
575 		adv_credits += posted;
576 		if (credit_alloc < work_alloc) {
577 			rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
578 			work_alloc = credit_alloc;
579 			flow_controlled = 1;
580 		}
581 		if (work_alloc == 0) {
582 			set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
583 			rds_ib_stats_inc(s_ib_tx_throttle);
584 			ret = -ENOMEM;
585 			goto out;
586 		}
587 	}
588 
589 	/* map the message the first time we see it */
590 	if (!ic->i_data_op) {
591 		if (rm->data.op_nents) {
592 			rm->data.op_count = ib_dma_map_sg(dev,
593 							  rm->data.op_sg,
594 							  rm->data.op_nents,
595 							  DMA_TO_DEVICE);
596 			rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->data.op_count);
597 			if (rm->data.op_count == 0) {
598 				rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
599 				rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
600 				ret = -ENOMEM; /* XXX ? */
601 				goto out;
602 			}
603 		} else {
604 			rm->data.op_count = 0;
605 		}
606 
607 		rds_message_addref(rm);
608 		ic->i_data_op = &rm->data;
609 
610 		/* Finalize the header */
611 		if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
612 			rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
613 		if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
614 			rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
615 
616 		/* If it has a RDMA op, tell the peer we did it. This is
617 		 * used by the peer to release use-once RDMA MRs. */
618 		if (rm->rdma.op_active) {
619 			struct rds_ext_header_rdma ext_hdr;
620 
621 			ext_hdr.h_rdma_rkey = cpu_to_be32(rm->rdma.op_rkey);
622 			rds_message_add_extension(&rm->m_inc.i_hdr,
623 					RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
624 		}
625 		if (rm->m_rdma_cookie) {
626 			rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
627 					rds_rdma_cookie_key(rm->m_rdma_cookie),
628 					rds_rdma_cookie_offset(rm->m_rdma_cookie));
629 		}
630 
631 		/* Note - rds_ib_piggyb_ack clears the ACK_REQUIRED bit, so
632 		 * we should not do this unless we have a chance of at least
633 		 * sticking the header into the send ring. Which is why we
634 		 * should call rds_ib_ring_alloc first. */
635 		rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_ib_piggyb_ack(ic));
636 		rds_message_make_checksum(&rm->m_inc.i_hdr);
637 
638 		/*
639 		 * Update adv_credits since we reset the ACK_REQUIRED bit.
640 		 */
641 		if (ic->i_flowctl) {
642 			rds_ib_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
643 			adv_credits += posted;
644 			BUG_ON(adv_credits > 255);
645 		}
646 	}
647 
648 	/* Sometimes you want to put a fence between an RDMA
649 	 * READ and the following SEND.
650 	 * We could either do this all the time
651 	 * or when requested by the user. Right now, we let
652 	 * the application choose.
653 	 */
654 	if (rm->rdma.op_active && rm->rdma.op_fence)
655 		send_flags = IB_SEND_FENCE;
656 
657 	/* Each frag gets a header. Msgs may be 0 bytes */
658 	send = &ic->i_sends[pos];
659 	first = send;
660 	prev = NULL;
661 	scat = &ic->i_data_op->op_sg[sg];
662 	i = 0;
663 	do {
664 		unsigned int len = 0;
665 
666 		/* Set up the header */
667 		send->s_wr.send_flags = send_flags;
668 		send->s_wr.opcode = IB_WR_SEND;
669 		send->s_wr.num_sge = 1;
670 		send->s_wr.next = NULL;
671 		send->s_queued = jiffies;
672 		send->s_op = NULL;
673 
674 		send->s_sge[0].addr = ic->i_send_hdrs_dma
675 			+ (pos * sizeof(struct rds_header));
676 		send->s_sge[0].length = sizeof(struct rds_header);
677 
678 		memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
679 
680 		/* Set up the data, if present */
681 		if (i < work_alloc
682 		    && scat != &rm->data.op_sg[rm->data.op_count]) {
683 			len = min(RDS_FRAG_SIZE, ib_sg_dma_len(dev, scat) - off);
684 			send->s_wr.num_sge = 2;
685 
686 			send->s_sge[1].addr = ib_sg_dma_address(dev, scat) + off;
687 			send->s_sge[1].length = len;
688 
689 			bytes_sent += len;
690 			off += len;
691 			if (off == ib_sg_dma_len(dev, scat)) {
692 				scat++;
693 				off = 0;
694 			}
695 		}
696 
697 		rds_ib_set_wr_signal_state(ic, send, 0);
698 
699 		/*
700 		 * Always signal the last one if we're stopping due to flow control.
701 		 */
702 		if (ic->i_flowctl && flow_controlled && i == (work_alloc-1))
703 			send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
704 
705 		if (send->s_wr.send_flags & IB_SEND_SIGNALED)
706 			nr_sig++;
707 
708 		rdsdebug("send %p wr %p num_sge %u next %p\n", send,
709 			 &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
710 
711 		if (ic->i_flowctl && adv_credits) {
712 			struct rds_header *hdr = &ic->i_send_hdrs[pos];
713 
714 			/* add credit and redo the header checksum */
715 			hdr->h_credit = adv_credits;
716 			rds_message_make_checksum(hdr);
717 			adv_credits = 0;
718 			rds_ib_stats_inc(s_ib_tx_credit_updates);
719 		}
720 
721 		if (prev)
722 			prev->s_wr.next = &send->s_wr;
723 		prev = send;
724 
725 		pos = (pos + 1) % ic->i_send_ring.w_nr;
726 		send = &ic->i_sends[pos];
727 		i++;
728 
729 	} while (i < work_alloc
730 		 && scat != &rm->data.op_sg[rm->data.op_count]);
731 
732 	/* Account the RDS header in the number of bytes we sent, but just once.
733 	 * The caller has no concept of fragmentation. */
734 	if (hdr_off == 0)
735 		bytes_sent += sizeof(struct rds_header);
736 
737 	/* if we finished the message then send completion owns it */
738 	if (scat == &rm->data.op_sg[rm->data.op_count]) {
739 		prev->s_op = ic->i_data_op;
740 		prev->s_wr.send_flags |= IB_SEND_SOLICITED;
741 		ic->i_data_op = NULL;
742 	}
743 
744 	/* Put back wrs & credits we didn't use */
745 	if (i < work_alloc) {
746 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
747 		work_alloc = i;
748 	}
749 	if (ic->i_flowctl && i < credit_alloc)
750 		rds_ib_send_add_credits(conn, credit_alloc - i);
751 
752 	if (nr_sig)
753 		atomic_add(nr_sig, &ic->i_signaled_sends);
754 
755 	/* XXX need to worry about failed_wr and partial sends. */
756 	failed_wr = &first->s_wr;
757 	ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
758 	rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
759 		 first, &first->s_wr, ret, failed_wr);
760 	BUG_ON(failed_wr != &first->s_wr);
761 	if (ret) {
762 		printk(KERN_WARNING "RDS/IB: ib_post_send to %pI4 "
763 		       "returned %d\n", &conn->c_faddr, ret);
764 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
765 		rds_ib_sub_signaled(ic, nr_sig);
766 		if (prev->s_op) {
767 			ic->i_data_op = prev->s_op;
768 			prev->s_op = NULL;
769 		}
770 
771 		rds_ib_conn_error(ic->conn, "ib_post_send failed\n");
772 		goto out;
773 	}
774 
775 	ret = bytes_sent;
776 out:
777 	BUG_ON(adv_credits);
778 	return ret;
779 }
780 
781 /*
782  * Issue atomic operation.
783  * A simplified version of the rdma case, we always map 1 SG, and
784  * only 8 bytes, for the return value from the atomic operation.
785  */
rds_ib_xmit_atomic(struct rds_connection * conn,struct rm_atomic_op * op)786 int rds_ib_xmit_atomic(struct rds_connection *conn, struct rm_atomic_op *op)
787 {
788 	struct rds_ib_connection *ic = conn->c_transport_data;
789 	struct rds_ib_send_work *send = NULL;
790 	struct ib_send_wr *failed_wr;
791 	struct rds_ib_device *rds_ibdev;
792 	u32 pos;
793 	u32 work_alloc;
794 	int ret;
795 	int nr_sig = 0;
796 
797 	rds_ibdev = ib_get_client_data(ic->i_cm_id->device, &rds_ib_client);
798 
799 	work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, 1, &pos);
800 	if (work_alloc != 1) {
801 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
802 		rds_ib_stats_inc(s_ib_tx_ring_full);
803 		ret = -ENOMEM;
804 		goto out;
805 	}
806 
807 	/* address of send request in ring */
808 	send = &ic->i_sends[pos];
809 	send->s_queued = jiffies;
810 
811 	if (op->op_type == RDS_ATOMIC_TYPE_CSWP) {
812 		send->s_wr.opcode = IB_WR_MASKED_ATOMIC_CMP_AND_SWP;
813 		send->s_wr.wr.atomic.compare_add = op->op_m_cswp.compare;
814 		send->s_wr.wr.atomic.swap = op->op_m_cswp.swap;
815 		send->s_wr.wr.atomic.compare_add_mask = op->op_m_cswp.compare_mask;
816 		send->s_wr.wr.atomic.swap_mask = op->op_m_cswp.swap_mask;
817 	} else { /* FADD */
818 		send->s_wr.opcode = IB_WR_MASKED_ATOMIC_FETCH_AND_ADD;
819 		send->s_wr.wr.atomic.compare_add = op->op_m_fadd.add;
820 		send->s_wr.wr.atomic.swap = 0;
821 		send->s_wr.wr.atomic.compare_add_mask = op->op_m_fadd.nocarry_mask;
822 		send->s_wr.wr.atomic.swap_mask = 0;
823 	}
824 	nr_sig = rds_ib_set_wr_signal_state(ic, send, op->op_notify);
825 	send->s_wr.num_sge = 1;
826 	send->s_wr.next = NULL;
827 	send->s_wr.wr.atomic.remote_addr = op->op_remote_addr;
828 	send->s_wr.wr.atomic.rkey = op->op_rkey;
829 	send->s_op = op;
830 	rds_message_addref(container_of(send->s_op, struct rds_message, atomic));
831 
832 	/* map 8 byte retval buffer to the device */
833 	ret = ib_dma_map_sg(ic->i_cm_id->device, op->op_sg, 1, DMA_FROM_DEVICE);
834 	rdsdebug("ic %p mapping atomic op %p. mapped %d pg\n", ic, op, ret);
835 	if (ret != 1) {
836 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
837 		rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
838 		ret = -ENOMEM; /* XXX ? */
839 		goto out;
840 	}
841 
842 	/* Convert our struct scatterlist to struct ib_sge */
843 	send->s_sge[0].addr = ib_sg_dma_address(ic->i_cm_id->device, op->op_sg);
844 	send->s_sge[0].length = ib_sg_dma_len(ic->i_cm_id->device, op->op_sg);
845 	send->s_sge[0].lkey = ic->i_mr->lkey;
846 
847 	rdsdebug("rva %Lx rpa %Lx len %u\n", op->op_remote_addr,
848 		 send->s_sge[0].addr, send->s_sge[0].length);
849 
850 	if (nr_sig)
851 		atomic_add(nr_sig, &ic->i_signaled_sends);
852 
853 	failed_wr = &send->s_wr;
854 	ret = ib_post_send(ic->i_cm_id->qp, &send->s_wr, &failed_wr);
855 	rdsdebug("ic %p send %p (wr %p) ret %d wr %p\n", ic,
856 		 send, &send->s_wr, ret, failed_wr);
857 	BUG_ON(failed_wr != &send->s_wr);
858 	if (ret) {
859 		printk(KERN_WARNING "RDS/IB: atomic ib_post_send to %pI4 "
860 		       "returned %d\n", &conn->c_faddr, ret);
861 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
862 		rds_ib_sub_signaled(ic, nr_sig);
863 		goto out;
864 	}
865 
866 	if (unlikely(failed_wr != &send->s_wr)) {
867 		printk(KERN_WARNING "RDS/IB: atomic ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
868 		BUG_ON(failed_wr != &send->s_wr);
869 	}
870 
871 out:
872 	return ret;
873 }
874 
rds_ib_xmit_rdma(struct rds_connection * conn,struct rm_rdma_op * op)875 int rds_ib_xmit_rdma(struct rds_connection *conn, struct rm_rdma_op *op)
876 {
877 	struct rds_ib_connection *ic = conn->c_transport_data;
878 	struct rds_ib_send_work *send = NULL;
879 	struct rds_ib_send_work *first;
880 	struct rds_ib_send_work *prev;
881 	struct ib_send_wr *failed_wr;
882 	struct scatterlist *scat;
883 	unsigned long len;
884 	u64 remote_addr = op->op_remote_addr;
885 	u32 max_sge = ic->rds_ibdev->max_sge;
886 	u32 pos;
887 	u32 work_alloc;
888 	u32 i;
889 	u32 j;
890 	int sent;
891 	int ret;
892 	int num_sge;
893 	int nr_sig = 0;
894 
895 	/* map the op the first time we see it */
896 	if (!op->op_mapped) {
897 		op->op_count = ib_dma_map_sg(ic->i_cm_id->device,
898 					     op->op_sg, op->op_nents, (op->op_write) ?
899 					     DMA_TO_DEVICE : DMA_FROM_DEVICE);
900 		rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->op_count);
901 		if (op->op_count == 0) {
902 			rds_ib_stats_inc(s_ib_tx_sg_mapping_failure);
903 			ret = -ENOMEM; /* XXX ? */
904 			goto out;
905 		}
906 
907 		op->op_mapped = 1;
908 	}
909 
910 	/*
911 	 * Instead of knowing how to return a partial rdma read/write we insist that there
912 	 * be enough work requests to send the entire message.
913 	 */
914 	i = ceil(op->op_count, max_sge);
915 
916 	work_alloc = rds_ib_ring_alloc(&ic->i_send_ring, i, &pos);
917 	if (work_alloc != i) {
918 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
919 		rds_ib_stats_inc(s_ib_tx_ring_full);
920 		ret = -ENOMEM;
921 		goto out;
922 	}
923 
924 	send = &ic->i_sends[pos];
925 	first = send;
926 	prev = NULL;
927 	scat = &op->op_sg[0];
928 	sent = 0;
929 	num_sge = op->op_count;
930 
931 	for (i = 0; i < work_alloc && scat != &op->op_sg[op->op_count]; i++) {
932 		send->s_wr.send_flags = 0;
933 		send->s_queued = jiffies;
934 		send->s_op = NULL;
935 
936 		nr_sig += rds_ib_set_wr_signal_state(ic, send, op->op_notify);
937 
938 		send->s_wr.opcode = op->op_write ? IB_WR_RDMA_WRITE : IB_WR_RDMA_READ;
939 		send->s_wr.wr.rdma.remote_addr = remote_addr;
940 		send->s_wr.wr.rdma.rkey = op->op_rkey;
941 
942 		if (num_sge > max_sge) {
943 			send->s_wr.num_sge = max_sge;
944 			num_sge -= max_sge;
945 		} else {
946 			send->s_wr.num_sge = num_sge;
947 		}
948 
949 		send->s_wr.next = NULL;
950 
951 		if (prev)
952 			prev->s_wr.next = &send->s_wr;
953 
954 		for (j = 0; j < send->s_wr.num_sge && scat != &op->op_sg[op->op_count]; j++) {
955 			len = ib_sg_dma_len(ic->i_cm_id->device, scat);
956 			send->s_sge[j].addr =
957 				 ib_sg_dma_address(ic->i_cm_id->device, scat);
958 			send->s_sge[j].length = len;
959 			send->s_sge[j].lkey = ic->i_mr->lkey;
960 
961 			sent += len;
962 			rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
963 
964 			remote_addr += len;
965 			scat++;
966 		}
967 
968 		rdsdebug("send %p wr %p num_sge %u next %p\n", send,
969 			&send->s_wr, send->s_wr.num_sge, send->s_wr.next);
970 
971 		prev = send;
972 		if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
973 			send = ic->i_sends;
974 	}
975 
976 	/* give a reference to the last op */
977 	if (scat == &op->op_sg[op->op_count]) {
978 		prev->s_op = op;
979 		rds_message_addref(container_of(op, struct rds_message, rdma));
980 	}
981 
982 	if (i < work_alloc) {
983 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc - i);
984 		work_alloc = i;
985 	}
986 
987 	if (nr_sig)
988 		atomic_add(nr_sig, &ic->i_signaled_sends);
989 
990 	failed_wr = &first->s_wr;
991 	ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
992 	rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
993 		 first, &first->s_wr, ret, failed_wr);
994 	BUG_ON(failed_wr != &first->s_wr);
995 	if (ret) {
996 		printk(KERN_WARNING "RDS/IB: rdma ib_post_send to %pI4 "
997 		       "returned %d\n", &conn->c_faddr, ret);
998 		rds_ib_ring_unalloc(&ic->i_send_ring, work_alloc);
999 		rds_ib_sub_signaled(ic, nr_sig);
1000 		goto out;
1001 	}
1002 
1003 	if (unlikely(failed_wr != &first->s_wr)) {
1004 		printk(KERN_WARNING "RDS/IB: ib_post_send() rc=%d, but failed_wqe updated!\n", ret);
1005 		BUG_ON(failed_wr != &first->s_wr);
1006 	}
1007 
1008 
1009 out:
1010 	return ret;
1011 }
1012 
rds_ib_xmit_complete(struct rds_connection * conn)1013 void rds_ib_xmit_complete(struct rds_connection *conn)
1014 {
1015 	struct rds_ib_connection *ic = conn->c_transport_data;
1016 
1017 	/* We may have a pending ACK or window update we were unable
1018 	 * to send previously (due to flow control). Try again. */
1019 	rds_ib_attempt_ack(ic);
1020 }
1021