xref: /openbmc/linux/net/smc/smc_wr.c (revision 869b6ca39c08c5b10eeb29d4b3c4bc433bf8ba5e)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Shared Memory Communications over RDMA (SMC-R) and RoCE
4  *
5  * Work Requests exploiting Infiniband API
6  *
7  * Work requests (WR) of type ib_post_send or ib_post_recv respectively
8  * are submitted to either RC SQ or RC RQ respectively
9  * (reliably connected send/receive queue)
10  * and become work queue entries (WQEs).
11  * While an SQ WR/WQE is pending, we track it until transmission completion.
12  * Through a send or receive completion queue (CQ) respectively,
13  * we get completion queue entries (CQEs) [aka work completions (WCs)].
14  * Since the CQ callback is called from IRQ context, we split work by using
15  * bottom halves implemented by tasklets.
16  *
17  * SMC uses this to exchange LLC (link layer control)
18  * and CDC (connection data control) messages.
19  *
20  * Copyright IBM Corp. 2016
21  *
22  * Author(s):  Steffen Maier <maier@linux.vnet.ibm.com>
23  */
24 
25 #include <linux/atomic.h>
26 #include <linux/hashtable.h>
27 #include <linux/wait.h>
28 #include <rdma/ib_verbs.h>
29 #include <asm/div64.h>
30 
31 #include "smc.h"
32 #include "smc_wr.h"
33 
34 #define SMC_WR_MAX_POLL_CQE 10	/* max. # of compl. queue elements in 1 poll */
35 
36 #define SMC_WR_RX_HASH_BITS 4
37 static DEFINE_HASHTABLE(smc_wr_rx_hash, SMC_WR_RX_HASH_BITS);
38 static DEFINE_SPINLOCK(smc_wr_rx_hash_lock);
39 
40 struct smc_wr_tx_pend {	/* control data for a pending send request */
41 	u64			wr_id;		/* work request id sent */
42 	smc_wr_tx_handler	handler;
43 	enum ib_wc_status	wc_status;	/* CQE status */
44 	struct smc_link		*link;
45 	u32			idx;
46 	struct smc_wr_tx_pend_priv priv;
47 	u8			compl_requested;
48 };
49 
50 /******************************** send queue *********************************/
51 
52 /*------------------------------- completion --------------------------------*/
53 
54 /* returns true if at least one tx work request is pending on the given link */
55 static inline bool smc_wr_is_tx_pend(struct smc_link *link)
56 {
57 	if (find_first_bit(link->wr_tx_mask, link->wr_tx_cnt) !=
58 							link->wr_tx_cnt) {
59 		return true;
60 	}
61 	return false;
62 }
63 
64 /* wait till all pending tx work requests on the given link are completed */
65 void smc_wr_tx_wait_no_pending_sends(struct smc_link *link)
66 {
67 	wait_event(link->wr_tx_wait, !smc_wr_is_tx_pend(link));
68 }
69 
70 static inline int smc_wr_tx_find_pending_index(struct smc_link *link, u64 wr_id)
71 {
72 	u32 i;
73 
74 	for (i = 0; i < link->wr_tx_cnt; i++) {
75 		if (link->wr_tx_pends[i].wr_id == wr_id)
76 			return i;
77 	}
78 	return link->wr_tx_cnt;
79 }
80 
81 static inline void smc_wr_tx_process_cqe(struct ib_wc *wc)
82 {
83 	struct smc_wr_tx_pend pnd_snd;
84 	struct smc_link *link;
85 	u32 pnd_snd_idx;
86 
87 	link = wc->qp->qp_context;
88 
89 	if (wc->opcode == IB_WC_REG_MR) {
90 		if (wc->status)
91 			link->wr_reg_state = FAILED;
92 		else
93 			link->wr_reg_state = CONFIRMED;
94 		smc_wr_wakeup_reg_wait(link);
95 		return;
96 	}
97 
98 	pnd_snd_idx = smc_wr_tx_find_pending_index(link, wc->wr_id);
99 	if (pnd_snd_idx == link->wr_tx_cnt) {
100 		if (link->lgr->smc_version != SMC_V2 ||
101 		    link->wr_tx_v2_pend->wr_id != wc->wr_id)
102 			return;
103 		link->wr_tx_v2_pend->wc_status = wc->status;
104 		memcpy(&pnd_snd, link->wr_tx_v2_pend, sizeof(pnd_snd));
105 		/* clear the full struct smc_wr_tx_pend including .priv */
106 		memset(link->wr_tx_v2_pend, 0,
107 		       sizeof(*link->wr_tx_v2_pend));
108 		memset(link->lgr->wr_tx_buf_v2, 0,
109 		       sizeof(*link->lgr->wr_tx_buf_v2));
110 	} else {
111 		link->wr_tx_pends[pnd_snd_idx].wc_status = wc->status;
112 		if (link->wr_tx_pends[pnd_snd_idx].compl_requested)
113 			complete(&link->wr_tx_compl[pnd_snd_idx]);
114 		memcpy(&pnd_snd, &link->wr_tx_pends[pnd_snd_idx],
115 		       sizeof(pnd_snd));
116 		/* clear the full struct smc_wr_tx_pend including .priv */
117 		memset(&link->wr_tx_pends[pnd_snd_idx], 0,
118 		       sizeof(link->wr_tx_pends[pnd_snd_idx]));
119 		memset(&link->wr_tx_bufs[pnd_snd_idx], 0,
120 		       sizeof(link->wr_tx_bufs[pnd_snd_idx]));
121 		if (!test_and_clear_bit(pnd_snd_idx, link->wr_tx_mask))
122 			return;
123 	}
124 
125 	if (wc->status) {
126 		if (link->lgr->smc_version == SMC_V2) {
127 			memset(link->wr_tx_v2_pend, 0,
128 			       sizeof(*link->wr_tx_v2_pend));
129 			memset(link->lgr->wr_tx_buf_v2, 0,
130 			       sizeof(*link->lgr->wr_tx_buf_v2));
131 		}
132 		/* terminate link */
133 		smcr_link_down_cond_sched(link);
134 	}
135 	if (pnd_snd.handler)
136 		pnd_snd.handler(&pnd_snd.priv, link, wc->status);
137 	wake_up(&link->wr_tx_wait);
138 }
139 
140 static void smc_wr_tx_tasklet_fn(struct tasklet_struct *t)
141 {
142 	struct smc_ib_device *dev = from_tasklet(dev, t, send_tasklet);
143 	struct ib_wc wc[SMC_WR_MAX_POLL_CQE];
144 	int i = 0, rc;
145 	int polled = 0;
146 
147 again:
148 	polled++;
149 	do {
150 		memset(&wc, 0, sizeof(wc));
151 		rc = ib_poll_cq(dev->roce_cq_send, SMC_WR_MAX_POLL_CQE, wc);
152 		if (polled == 1) {
153 			ib_req_notify_cq(dev->roce_cq_send,
154 					 IB_CQ_NEXT_COMP |
155 					 IB_CQ_REPORT_MISSED_EVENTS);
156 		}
157 		if (!rc)
158 			break;
159 		for (i = 0; i < rc; i++)
160 			smc_wr_tx_process_cqe(&wc[i]);
161 	} while (rc > 0);
162 	if (polled == 1)
163 		goto again;
164 }
165 
166 void smc_wr_tx_cq_handler(struct ib_cq *ib_cq, void *cq_context)
167 {
168 	struct smc_ib_device *dev = (struct smc_ib_device *)cq_context;
169 
170 	tasklet_schedule(&dev->send_tasklet);
171 }
172 
173 /*---------------------------- request submission ---------------------------*/
174 
175 static inline int smc_wr_tx_get_free_slot_index(struct smc_link *link, u32 *idx)
176 {
177 	*idx = link->wr_tx_cnt;
178 	if (!smc_link_sendable(link))
179 		return -ENOLINK;
180 	for_each_clear_bit(*idx, link->wr_tx_mask, link->wr_tx_cnt) {
181 		if (!test_and_set_bit(*idx, link->wr_tx_mask))
182 			return 0;
183 	}
184 	*idx = link->wr_tx_cnt;
185 	return -EBUSY;
186 }
187 
188 /**
189  * smc_wr_tx_get_free_slot() - returns buffer for message assembly,
190  *			and sets info for pending transmit tracking
191  * @link:		Pointer to smc_link used to later send the message.
192  * @handler:		Send completion handler function pointer.
193  * @wr_buf:		Out value returns pointer to message buffer.
194  * @wr_rdma_buf:	Out value returns pointer to rdma work request.
195  * @wr_pend_priv:	Out value returns pointer serving as handler context.
196  *
197  * Return: 0 on success, or -errno on error.
198  */
199 int smc_wr_tx_get_free_slot(struct smc_link *link,
200 			    smc_wr_tx_handler handler,
201 			    struct smc_wr_buf **wr_buf,
202 			    struct smc_rdma_wr **wr_rdma_buf,
203 			    struct smc_wr_tx_pend_priv **wr_pend_priv)
204 {
205 	struct smc_link_group *lgr = smc_get_lgr(link);
206 	struct smc_wr_tx_pend *wr_pend;
207 	u32 idx = link->wr_tx_cnt;
208 	struct ib_send_wr *wr_ib;
209 	u64 wr_id;
210 	int rc;
211 
212 	*wr_buf = NULL;
213 	*wr_pend_priv = NULL;
214 	if (in_softirq() || lgr->terminating) {
215 		rc = smc_wr_tx_get_free_slot_index(link, &idx);
216 		if (rc)
217 			return rc;
218 	} else {
219 		rc = wait_event_interruptible_timeout(
220 			link->wr_tx_wait,
221 			!smc_link_sendable(link) ||
222 			lgr->terminating ||
223 			(smc_wr_tx_get_free_slot_index(link, &idx) != -EBUSY),
224 			SMC_WR_TX_WAIT_FREE_SLOT_TIME);
225 		if (!rc) {
226 			/* timeout - terminate link */
227 			smcr_link_down_cond_sched(link);
228 			return -EPIPE;
229 		}
230 		if (idx == link->wr_tx_cnt)
231 			return -EPIPE;
232 	}
233 	wr_id = smc_wr_tx_get_next_wr_id(link);
234 	wr_pend = &link->wr_tx_pends[idx];
235 	wr_pend->wr_id = wr_id;
236 	wr_pend->handler = handler;
237 	wr_pend->link = link;
238 	wr_pend->idx = idx;
239 	wr_ib = &link->wr_tx_ibs[idx];
240 	wr_ib->wr_id = wr_id;
241 	*wr_buf = &link->wr_tx_bufs[idx];
242 	if (wr_rdma_buf)
243 		*wr_rdma_buf = &link->wr_tx_rdmas[idx];
244 	*wr_pend_priv = &wr_pend->priv;
245 	return 0;
246 }
247 
248 int smc_wr_tx_get_v2_slot(struct smc_link *link,
249 			  smc_wr_tx_handler handler,
250 			  struct smc_wr_v2_buf **wr_buf,
251 			  struct smc_wr_tx_pend_priv **wr_pend_priv)
252 {
253 	struct smc_wr_tx_pend *wr_pend;
254 	struct ib_send_wr *wr_ib;
255 	u64 wr_id;
256 
257 	if (link->wr_tx_v2_pend->idx == link->wr_tx_cnt)
258 		return -EBUSY;
259 
260 	*wr_buf = NULL;
261 	*wr_pend_priv = NULL;
262 	wr_id = smc_wr_tx_get_next_wr_id(link);
263 	wr_pend = link->wr_tx_v2_pend;
264 	wr_pend->wr_id = wr_id;
265 	wr_pend->handler = handler;
266 	wr_pend->link = link;
267 	wr_pend->idx = link->wr_tx_cnt;
268 	wr_ib = link->wr_tx_v2_ib;
269 	wr_ib->wr_id = wr_id;
270 	*wr_buf = link->lgr->wr_tx_buf_v2;
271 	*wr_pend_priv = &wr_pend->priv;
272 	return 0;
273 }
274 
275 int smc_wr_tx_put_slot(struct smc_link *link,
276 		       struct smc_wr_tx_pend_priv *wr_pend_priv)
277 {
278 	struct smc_wr_tx_pend *pend;
279 
280 	pend = container_of(wr_pend_priv, struct smc_wr_tx_pend, priv);
281 	if (pend->idx < link->wr_tx_cnt) {
282 		u32 idx = pend->idx;
283 
284 		/* clear the full struct smc_wr_tx_pend including .priv */
285 		memset(&link->wr_tx_pends[idx], 0,
286 		       sizeof(link->wr_tx_pends[idx]));
287 		memset(&link->wr_tx_bufs[idx], 0,
288 		       sizeof(link->wr_tx_bufs[idx]));
289 		test_and_clear_bit(idx, link->wr_tx_mask);
290 		wake_up(&link->wr_tx_wait);
291 		return 1;
292 	} else if (link->lgr->smc_version == SMC_V2 &&
293 		   pend->idx == link->wr_tx_cnt) {
294 		/* Large v2 buffer */
295 		memset(&link->wr_tx_v2_pend, 0,
296 		       sizeof(link->wr_tx_v2_pend));
297 		memset(&link->lgr->wr_tx_buf_v2, 0,
298 		       sizeof(link->lgr->wr_tx_buf_v2));
299 		return 1;
300 	}
301 
302 	return 0;
303 }
304 
305 /* Send prepared WR slot via ib_post_send.
306  * @priv: pointer to smc_wr_tx_pend_priv identifying prepared message buffer
307  */
308 int smc_wr_tx_send(struct smc_link *link, struct smc_wr_tx_pend_priv *priv)
309 {
310 	struct smc_wr_tx_pend *pend;
311 	int rc;
312 
313 	ib_req_notify_cq(link->smcibdev->roce_cq_send,
314 			 IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS);
315 	pend = container_of(priv, struct smc_wr_tx_pend, priv);
316 	rc = ib_post_send(link->roce_qp, &link->wr_tx_ibs[pend->idx], NULL);
317 	if (rc) {
318 		smc_wr_tx_put_slot(link, priv);
319 		smcr_link_down_cond_sched(link);
320 	}
321 	return rc;
322 }
323 
324 int smc_wr_tx_v2_send(struct smc_link *link, struct smc_wr_tx_pend_priv *priv,
325 		      int len)
326 {
327 	int rc;
328 
329 	link->wr_tx_v2_ib->sg_list[0].length = len;
330 	ib_req_notify_cq(link->smcibdev->roce_cq_send,
331 			 IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS);
332 	rc = ib_post_send(link->roce_qp, link->wr_tx_v2_ib, NULL);
333 	if (rc) {
334 		smc_wr_tx_put_slot(link, priv);
335 		smcr_link_down_cond_sched(link);
336 	}
337 	return rc;
338 }
339 
340 /* Send prepared WR slot via ib_post_send and wait for send completion
341  * notification.
342  * @priv: pointer to smc_wr_tx_pend_priv identifying prepared message buffer
343  */
344 int smc_wr_tx_send_wait(struct smc_link *link, struct smc_wr_tx_pend_priv *priv,
345 			unsigned long timeout)
346 {
347 	struct smc_wr_tx_pend *pend;
348 	u32 pnd_idx;
349 	int rc;
350 
351 	pend = container_of(priv, struct smc_wr_tx_pend, priv);
352 	pend->compl_requested = 1;
353 	pnd_idx = pend->idx;
354 	init_completion(&link->wr_tx_compl[pnd_idx]);
355 
356 	rc = smc_wr_tx_send(link, priv);
357 	if (rc)
358 		return rc;
359 	/* wait for completion by smc_wr_tx_process_cqe() */
360 	rc = wait_for_completion_interruptible_timeout(
361 					&link->wr_tx_compl[pnd_idx], timeout);
362 	if (rc <= 0)
363 		rc = -ENODATA;
364 	if (rc > 0)
365 		rc = 0;
366 	return rc;
367 }
368 
369 /* Register a memory region and wait for result. */
370 int smc_wr_reg_send(struct smc_link *link, struct ib_mr *mr)
371 {
372 	int rc;
373 
374 	ib_req_notify_cq(link->smcibdev->roce_cq_send,
375 			 IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS);
376 	link->wr_reg_state = POSTED;
377 	link->wr_reg.wr.wr_id = (u64)(uintptr_t)mr;
378 	link->wr_reg.mr = mr;
379 	link->wr_reg.key = mr->rkey;
380 	rc = ib_post_send(link->roce_qp, &link->wr_reg.wr, NULL);
381 	if (rc)
382 		return rc;
383 
384 	atomic_inc(&link->wr_reg_refcnt);
385 	rc = wait_event_interruptible_timeout(link->wr_reg_wait,
386 					      (link->wr_reg_state != POSTED),
387 					      SMC_WR_REG_MR_WAIT_TIME);
388 	if (atomic_dec_and_test(&link->wr_reg_refcnt))
389 		wake_up_all(&link->wr_reg_wait);
390 	if (!rc) {
391 		/* timeout - terminate link */
392 		smcr_link_down_cond_sched(link);
393 		return -EPIPE;
394 	}
395 	if (rc == -ERESTARTSYS)
396 		return -EINTR;
397 	switch (link->wr_reg_state) {
398 	case CONFIRMED:
399 		rc = 0;
400 		break;
401 	case FAILED:
402 		rc = -EIO;
403 		break;
404 	case POSTED:
405 		rc = -EPIPE;
406 		break;
407 	}
408 	return rc;
409 }
410 
411 /****************************** receive queue ********************************/
412 
413 int smc_wr_rx_register_handler(struct smc_wr_rx_handler *handler)
414 {
415 	struct smc_wr_rx_handler *h_iter;
416 	int rc = 0;
417 
418 	spin_lock(&smc_wr_rx_hash_lock);
419 	hash_for_each_possible(smc_wr_rx_hash, h_iter, list, handler->type) {
420 		if (h_iter->type == handler->type) {
421 			rc = -EEXIST;
422 			goto out_unlock;
423 		}
424 	}
425 	hash_add(smc_wr_rx_hash, &handler->list, handler->type);
426 out_unlock:
427 	spin_unlock(&smc_wr_rx_hash_lock);
428 	return rc;
429 }
430 
431 /* Demultiplex a received work request based on the message type to its handler.
432  * Relies on smc_wr_rx_hash having been completely filled before any IB WRs,
433  * and not being modified any more afterwards so we don't need to lock it.
434  */
435 static inline void smc_wr_rx_demultiplex(struct ib_wc *wc)
436 {
437 	struct smc_link *link = (struct smc_link *)wc->qp->qp_context;
438 	struct smc_wr_rx_handler *handler;
439 	struct smc_wr_rx_hdr *wr_rx;
440 	u64 temp_wr_id;
441 	u32 index;
442 
443 	if (wc->byte_len < sizeof(*wr_rx))
444 		return; /* short message */
445 	temp_wr_id = wc->wr_id;
446 	index = do_div(temp_wr_id, link->wr_rx_cnt);
447 	wr_rx = (struct smc_wr_rx_hdr *)&link->wr_rx_bufs[index];
448 	hash_for_each_possible(smc_wr_rx_hash, handler, list, wr_rx->type) {
449 		if (handler->type == wr_rx->type)
450 			handler->handler(wc, wr_rx);
451 	}
452 }
453 
454 static inline void smc_wr_rx_process_cqes(struct ib_wc wc[], int num)
455 {
456 	struct smc_link *link;
457 	int i;
458 
459 	for (i = 0; i < num; i++) {
460 		link = wc[i].qp->qp_context;
461 		if (wc[i].status == IB_WC_SUCCESS) {
462 			link->wr_rx_tstamp = jiffies;
463 			smc_wr_rx_demultiplex(&wc[i]);
464 			smc_wr_rx_post(link); /* refill WR RX */
465 		} else {
466 			/* handle status errors */
467 			switch (wc[i].status) {
468 			case IB_WC_RETRY_EXC_ERR:
469 			case IB_WC_RNR_RETRY_EXC_ERR:
470 			case IB_WC_WR_FLUSH_ERR:
471 				smcr_link_down_cond_sched(link);
472 				break;
473 			default:
474 				smc_wr_rx_post(link); /* refill WR RX */
475 				break;
476 			}
477 		}
478 	}
479 }
480 
481 static void smc_wr_rx_tasklet_fn(struct tasklet_struct *t)
482 {
483 	struct smc_ib_device *dev = from_tasklet(dev, t, recv_tasklet);
484 	struct ib_wc wc[SMC_WR_MAX_POLL_CQE];
485 	int polled = 0;
486 	int rc;
487 
488 again:
489 	polled++;
490 	do {
491 		memset(&wc, 0, sizeof(wc));
492 		rc = ib_poll_cq(dev->roce_cq_recv, SMC_WR_MAX_POLL_CQE, wc);
493 		if (polled == 1) {
494 			ib_req_notify_cq(dev->roce_cq_recv,
495 					 IB_CQ_SOLICITED_MASK
496 					 | IB_CQ_REPORT_MISSED_EVENTS);
497 		}
498 		if (!rc)
499 			break;
500 		smc_wr_rx_process_cqes(&wc[0], rc);
501 	} while (rc > 0);
502 	if (polled == 1)
503 		goto again;
504 }
505 
506 void smc_wr_rx_cq_handler(struct ib_cq *ib_cq, void *cq_context)
507 {
508 	struct smc_ib_device *dev = (struct smc_ib_device *)cq_context;
509 
510 	tasklet_schedule(&dev->recv_tasklet);
511 }
512 
513 int smc_wr_rx_post_init(struct smc_link *link)
514 {
515 	u32 i;
516 	int rc = 0;
517 
518 	for (i = 0; i < link->wr_rx_cnt; i++)
519 		rc = smc_wr_rx_post(link);
520 	return rc;
521 }
522 
523 /***************************** init, exit, misc ******************************/
524 
525 void smc_wr_remember_qp_attr(struct smc_link *lnk)
526 {
527 	struct ib_qp_attr *attr = &lnk->qp_attr;
528 	struct ib_qp_init_attr init_attr;
529 
530 	memset(attr, 0, sizeof(*attr));
531 	memset(&init_attr, 0, sizeof(init_attr));
532 	ib_query_qp(lnk->roce_qp, attr,
533 		    IB_QP_STATE |
534 		    IB_QP_CUR_STATE |
535 		    IB_QP_PKEY_INDEX |
536 		    IB_QP_PORT |
537 		    IB_QP_QKEY |
538 		    IB_QP_AV |
539 		    IB_QP_PATH_MTU |
540 		    IB_QP_TIMEOUT |
541 		    IB_QP_RETRY_CNT |
542 		    IB_QP_RNR_RETRY |
543 		    IB_QP_RQ_PSN |
544 		    IB_QP_ALT_PATH |
545 		    IB_QP_MIN_RNR_TIMER |
546 		    IB_QP_SQ_PSN |
547 		    IB_QP_PATH_MIG_STATE |
548 		    IB_QP_CAP |
549 		    IB_QP_DEST_QPN,
550 		    &init_attr);
551 
552 	lnk->wr_tx_cnt = min_t(size_t, SMC_WR_BUF_CNT,
553 			       lnk->qp_attr.cap.max_send_wr);
554 	lnk->wr_rx_cnt = min_t(size_t, SMC_WR_BUF_CNT * 3,
555 			       lnk->qp_attr.cap.max_recv_wr);
556 }
557 
558 static void smc_wr_init_sge(struct smc_link *lnk)
559 {
560 	int sges_per_buf = (lnk->lgr->smc_version == SMC_V2) ? 2 : 1;
561 	u32 i;
562 
563 	for (i = 0; i < lnk->wr_tx_cnt; i++) {
564 		lnk->wr_tx_sges[i].addr =
565 			lnk->wr_tx_dma_addr + i * SMC_WR_BUF_SIZE;
566 		lnk->wr_tx_sges[i].length = SMC_WR_TX_SIZE;
567 		lnk->wr_tx_sges[i].lkey = lnk->roce_pd->local_dma_lkey;
568 		lnk->wr_tx_rdma_sges[i].tx_rdma_sge[0].wr_tx_rdma_sge[0].lkey =
569 			lnk->roce_pd->local_dma_lkey;
570 		lnk->wr_tx_rdma_sges[i].tx_rdma_sge[0].wr_tx_rdma_sge[1].lkey =
571 			lnk->roce_pd->local_dma_lkey;
572 		lnk->wr_tx_rdma_sges[i].tx_rdma_sge[1].wr_tx_rdma_sge[0].lkey =
573 			lnk->roce_pd->local_dma_lkey;
574 		lnk->wr_tx_rdma_sges[i].tx_rdma_sge[1].wr_tx_rdma_sge[1].lkey =
575 			lnk->roce_pd->local_dma_lkey;
576 		lnk->wr_tx_ibs[i].next = NULL;
577 		lnk->wr_tx_ibs[i].sg_list = &lnk->wr_tx_sges[i];
578 		lnk->wr_tx_ibs[i].num_sge = 1;
579 		lnk->wr_tx_ibs[i].opcode = IB_WR_SEND;
580 		lnk->wr_tx_ibs[i].send_flags =
581 			IB_SEND_SIGNALED | IB_SEND_SOLICITED;
582 		lnk->wr_tx_rdmas[i].wr_tx_rdma[0].wr.opcode = IB_WR_RDMA_WRITE;
583 		lnk->wr_tx_rdmas[i].wr_tx_rdma[1].wr.opcode = IB_WR_RDMA_WRITE;
584 		lnk->wr_tx_rdmas[i].wr_tx_rdma[0].wr.sg_list =
585 			lnk->wr_tx_rdma_sges[i].tx_rdma_sge[0].wr_tx_rdma_sge;
586 		lnk->wr_tx_rdmas[i].wr_tx_rdma[1].wr.sg_list =
587 			lnk->wr_tx_rdma_sges[i].tx_rdma_sge[1].wr_tx_rdma_sge;
588 	}
589 
590 	if (lnk->lgr->smc_version == SMC_V2) {
591 		lnk->wr_tx_v2_sge->addr = lnk->wr_tx_v2_dma_addr;
592 		lnk->wr_tx_v2_sge->length = SMC_WR_BUF_V2_SIZE;
593 		lnk->wr_tx_v2_sge->lkey = lnk->roce_pd->local_dma_lkey;
594 
595 		lnk->wr_tx_v2_ib->next = NULL;
596 		lnk->wr_tx_v2_ib->sg_list = lnk->wr_tx_v2_sge;
597 		lnk->wr_tx_v2_ib->num_sge = 1;
598 		lnk->wr_tx_v2_ib->opcode = IB_WR_SEND;
599 		lnk->wr_tx_v2_ib->send_flags =
600 			IB_SEND_SIGNALED | IB_SEND_SOLICITED;
601 	}
602 
603 	/* With SMC-Rv2 there can be messages larger than SMC_WR_TX_SIZE.
604 	 * Each ib_recv_wr gets 2 sges, the second one is a spillover buffer
605 	 * and the same buffer for all sges. When a larger message arrived then
606 	 * the content of the first small sge is copied to the beginning of
607 	 * the larger spillover buffer, allowing easy data mapping.
608 	 */
609 	for (i = 0; i < lnk->wr_rx_cnt; i++) {
610 		int x = i * sges_per_buf;
611 
612 		lnk->wr_rx_sges[x].addr =
613 			lnk->wr_rx_dma_addr + i * SMC_WR_BUF_SIZE;
614 		lnk->wr_rx_sges[x].length = SMC_WR_TX_SIZE;
615 		lnk->wr_rx_sges[x].lkey = lnk->roce_pd->local_dma_lkey;
616 		if (lnk->lgr->smc_version == SMC_V2) {
617 			lnk->wr_rx_sges[x + 1].addr =
618 					lnk->wr_rx_v2_dma_addr + SMC_WR_TX_SIZE;
619 			lnk->wr_rx_sges[x + 1].length =
620 					SMC_WR_BUF_V2_SIZE - SMC_WR_TX_SIZE;
621 			lnk->wr_rx_sges[x + 1].lkey =
622 					lnk->roce_pd->local_dma_lkey;
623 		}
624 		lnk->wr_rx_ibs[i].next = NULL;
625 		lnk->wr_rx_ibs[i].sg_list = &lnk->wr_rx_sges[x];
626 		lnk->wr_rx_ibs[i].num_sge = sges_per_buf;
627 	}
628 	lnk->wr_reg.wr.next = NULL;
629 	lnk->wr_reg.wr.num_sge = 0;
630 	lnk->wr_reg.wr.send_flags = IB_SEND_SIGNALED;
631 	lnk->wr_reg.wr.opcode = IB_WR_REG_MR;
632 	lnk->wr_reg.access = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE;
633 }
634 
635 void smc_wr_free_link(struct smc_link *lnk)
636 {
637 	struct ib_device *ibdev;
638 
639 	if (!lnk->smcibdev)
640 		return;
641 	ibdev = lnk->smcibdev->ibdev;
642 
643 	smc_wr_wakeup_reg_wait(lnk);
644 	smc_wr_wakeup_tx_wait(lnk);
645 
646 	smc_wr_tx_wait_no_pending_sends(lnk);
647 	wait_event(lnk->wr_reg_wait, (!atomic_read(&lnk->wr_reg_refcnt)));
648 	wait_event(lnk->wr_tx_wait, (!atomic_read(&lnk->wr_tx_refcnt)));
649 
650 	if (lnk->wr_rx_dma_addr) {
651 		ib_dma_unmap_single(ibdev, lnk->wr_rx_dma_addr,
652 				    SMC_WR_BUF_SIZE * lnk->wr_rx_cnt,
653 				    DMA_FROM_DEVICE);
654 		lnk->wr_rx_dma_addr = 0;
655 	}
656 	if (lnk->wr_rx_v2_dma_addr) {
657 		ib_dma_unmap_single(ibdev, lnk->wr_rx_v2_dma_addr,
658 				    SMC_WR_BUF_V2_SIZE,
659 				    DMA_FROM_DEVICE);
660 		lnk->wr_rx_v2_dma_addr = 0;
661 	}
662 	if (lnk->wr_tx_dma_addr) {
663 		ib_dma_unmap_single(ibdev, lnk->wr_tx_dma_addr,
664 				    SMC_WR_BUF_SIZE * lnk->wr_tx_cnt,
665 				    DMA_TO_DEVICE);
666 		lnk->wr_tx_dma_addr = 0;
667 	}
668 	if (lnk->wr_tx_v2_dma_addr) {
669 		ib_dma_unmap_single(ibdev, lnk->wr_tx_v2_dma_addr,
670 				    SMC_WR_BUF_V2_SIZE,
671 				    DMA_TO_DEVICE);
672 		lnk->wr_tx_v2_dma_addr = 0;
673 	}
674 }
675 
676 void smc_wr_free_lgr_mem(struct smc_link_group *lgr)
677 {
678 	if (lgr->smc_version < SMC_V2)
679 		return;
680 
681 	kfree(lgr->wr_rx_buf_v2);
682 	lgr->wr_rx_buf_v2 = NULL;
683 	kfree(lgr->wr_tx_buf_v2);
684 	lgr->wr_tx_buf_v2 = NULL;
685 }
686 
687 void smc_wr_free_link_mem(struct smc_link *lnk)
688 {
689 	kfree(lnk->wr_tx_v2_ib);
690 	lnk->wr_tx_v2_ib = NULL;
691 	kfree(lnk->wr_tx_v2_sge);
692 	lnk->wr_tx_v2_sge = NULL;
693 	kfree(lnk->wr_tx_v2_pend);
694 	lnk->wr_tx_v2_pend = NULL;
695 	kfree(lnk->wr_tx_compl);
696 	lnk->wr_tx_compl = NULL;
697 	kfree(lnk->wr_tx_pends);
698 	lnk->wr_tx_pends = NULL;
699 	kfree(lnk->wr_tx_mask);
700 	lnk->wr_tx_mask = NULL;
701 	kfree(lnk->wr_tx_sges);
702 	lnk->wr_tx_sges = NULL;
703 	kfree(lnk->wr_tx_rdma_sges);
704 	lnk->wr_tx_rdma_sges = NULL;
705 	kfree(lnk->wr_rx_sges);
706 	lnk->wr_rx_sges = NULL;
707 	kfree(lnk->wr_tx_rdmas);
708 	lnk->wr_tx_rdmas = NULL;
709 	kfree(lnk->wr_rx_ibs);
710 	lnk->wr_rx_ibs = NULL;
711 	kfree(lnk->wr_tx_ibs);
712 	lnk->wr_tx_ibs = NULL;
713 	kfree(lnk->wr_tx_bufs);
714 	lnk->wr_tx_bufs = NULL;
715 	kfree(lnk->wr_rx_bufs);
716 	lnk->wr_rx_bufs = NULL;
717 }
718 
719 int smc_wr_alloc_lgr_mem(struct smc_link_group *lgr)
720 {
721 	if (lgr->smc_version < SMC_V2)
722 		return 0;
723 
724 	lgr->wr_rx_buf_v2 = kzalloc(SMC_WR_BUF_V2_SIZE, GFP_KERNEL);
725 	if (!lgr->wr_rx_buf_v2)
726 		return -ENOMEM;
727 	lgr->wr_tx_buf_v2 = kzalloc(SMC_WR_BUF_V2_SIZE, GFP_KERNEL);
728 	if (!lgr->wr_tx_buf_v2) {
729 		kfree(lgr->wr_rx_buf_v2);
730 		return -ENOMEM;
731 	}
732 	return 0;
733 }
734 
735 int smc_wr_alloc_link_mem(struct smc_link *link)
736 {
737 	int sges_per_buf = link->lgr->smc_version == SMC_V2 ? 2 : 1;
738 
739 	/* allocate link related memory */
740 	link->wr_tx_bufs = kcalloc(SMC_WR_BUF_CNT, SMC_WR_BUF_SIZE, GFP_KERNEL);
741 	if (!link->wr_tx_bufs)
742 		goto no_mem;
743 	link->wr_rx_bufs = kcalloc(SMC_WR_BUF_CNT * 3, SMC_WR_BUF_SIZE,
744 				   GFP_KERNEL);
745 	if (!link->wr_rx_bufs)
746 		goto no_mem_wr_tx_bufs;
747 	link->wr_tx_ibs = kcalloc(SMC_WR_BUF_CNT, sizeof(link->wr_tx_ibs[0]),
748 				  GFP_KERNEL);
749 	if (!link->wr_tx_ibs)
750 		goto no_mem_wr_rx_bufs;
751 	link->wr_rx_ibs = kcalloc(SMC_WR_BUF_CNT * 3,
752 				  sizeof(link->wr_rx_ibs[0]),
753 				  GFP_KERNEL);
754 	if (!link->wr_rx_ibs)
755 		goto no_mem_wr_tx_ibs;
756 	link->wr_tx_rdmas = kcalloc(SMC_WR_BUF_CNT,
757 				    sizeof(link->wr_tx_rdmas[0]),
758 				    GFP_KERNEL);
759 	if (!link->wr_tx_rdmas)
760 		goto no_mem_wr_rx_ibs;
761 	link->wr_tx_rdma_sges = kcalloc(SMC_WR_BUF_CNT,
762 					sizeof(link->wr_tx_rdma_sges[0]),
763 					GFP_KERNEL);
764 	if (!link->wr_tx_rdma_sges)
765 		goto no_mem_wr_tx_rdmas;
766 	link->wr_tx_sges = kcalloc(SMC_WR_BUF_CNT, sizeof(link->wr_tx_sges[0]),
767 				   GFP_KERNEL);
768 	if (!link->wr_tx_sges)
769 		goto no_mem_wr_tx_rdma_sges;
770 	link->wr_rx_sges = kcalloc(SMC_WR_BUF_CNT * 3,
771 				   sizeof(link->wr_rx_sges[0]) * sges_per_buf,
772 				   GFP_KERNEL);
773 	if (!link->wr_rx_sges)
774 		goto no_mem_wr_tx_sges;
775 	link->wr_tx_mask = kcalloc(BITS_TO_LONGS(SMC_WR_BUF_CNT),
776 				   sizeof(*link->wr_tx_mask),
777 				   GFP_KERNEL);
778 	if (!link->wr_tx_mask)
779 		goto no_mem_wr_rx_sges;
780 	link->wr_tx_pends = kcalloc(SMC_WR_BUF_CNT,
781 				    sizeof(link->wr_tx_pends[0]),
782 				    GFP_KERNEL);
783 	if (!link->wr_tx_pends)
784 		goto no_mem_wr_tx_mask;
785 	link->wr_tx_compl = kcalloc(SMC_WR_BUF_CNT,
786 				    sizeof(link->wr_tx_compl[0]),
787 				    GFP_KERNEL);
788 	if (!link->wr_tx_compl)
789 		goto no_mem_wr_tx_pends;
790 
791 	if (link->lgr->smc_version == SMC_V2) {
792 		link->wr_tx_v2_ib = kzalloc(sizeof(*link->wr_tx_v2_ib),
793 					    GFP_KERNEL);
794 		if (!link->wr_tx_v2_ib)
795 			goto no_mem_tx_compl;
796 		link->wr_tx_v2_sge = kzalloc(sizeof(*link->wr_tx_v2_sge),
797 					     GFP_KERNEL);
798 		if (!link->wr_tx_v2_sge)
799 			goto no_mem_v2_ib;
800 		link->wr_tx_v2_pend = kzalloc(sizeof(*link->wr_tx_v2_pend),
801 					      GFP_KERNEL);
802 		if (!link->wr_tx_v2_pend)
803 			goto no_mem_v2_sge;
804 	}
805 	return 0;
806 
807 no_mem_v2_sge:
808 	kfree(link->wr_tx_v2_sge);
809 no_mem_v2_ib:
810 	kfree(link->wr_tx_v2_ib);
811 no_mem_tx_compl:
812 	kfree(link->wr_tx_compl);
813 no_mem_wr_tx_pends:
814 	kfree(link->wr_tx_pends);
815 no_mem_wr_tx_mask:
816 	kfree(link->wr_tx_mask);
817 no_mem_wr_rx_sges:
818 	kfree(link->wr_rx_sges);
819 no_mem_wr_tx_sges:
820 	kfree(link->wr_tx_sges);
821 no_mem_wr_tx_rdma_sges:
822 	kfree(link->wr_tx_rdma_sges);
823 no_mem_wr_tx_rdmas:
824 	kfree(link->wr_tx_rdmas);
825 no_mem_wr_rx_ibs:
826 	kfree(link->wr_rx_ibs);
827 no_mem_wr_tx_ibs:
828 	kfree(link->wr_tx_ibs);
829 no_mem_wr_rx_bufs:
830 	kfree(link->wr_rx_bufs);
831 no_mem_wr_tx_bufs:
832 	kfree(link->wr_tx_bufs);
833 no_mem:
834 	return -ENOMEM;
835 }
836 
837 void smc_wr_remove_dev(struct smc_ib_device *smcibdev)
838 {
839 	tasklet_kill(&smcibdev->recv_tasklet);
840 	tasklet_kill(&smcibdev->send_tasklet);
841 }
842 
843 void smc_wr_add_dev(struct smc_ib_device *smcibdev)
844 {
845 	tasklet_setup(&smcibdev->recv_tasklet, smc_wr_rx_tasklet_fn);
846 	tasklet_setup(&smcibdev->send_tasklet, smc_wr_tx_tasklet_fn);
847 }
848 
849 int smc_wr_create_link(struct smc_link *lnk)
850 {
851 	struct ib_device *ibdev = lnk->smcibdev->ibdev;
852 	int rc = 0;
853 
854 	smc_wr_tx_set_wr_id(&lnk->wr_tx_id, 0);
855 	lnk->wr_rx_id = 0;
856 	lnk->wr_rx_dma_addr = ib_dma_map_single(
857 		ibdev, lnk->wr_rx_bufs,	SMC_WR_BUF_SIZE * lnk->wr_rx_cnt,
858 		DMA_FROM_DEVICE);
859 	if (ib_dma_mapping_error(ibdev, lnk->wr_rx_dma_addr)) {
860 		lnk->wr_rx_dma_addr = 0;
861 		rc = -EIO;
862 		goto out;
863 	}
864 	if (lnk->lgr->smc_version == SMC_V2) {
865 		lnk->wr_rx_v2_dma_addr = ib_dma_map_single(ibdev,
866 			lnk->lgr->wr_rx_buf_v2, SMC_WR_BUF_V2_SIZE,
867 			DMA_FROM_DEVICE);
868 		if (ib_dma_mapping_error(ibdev, lnk->wr_rx_v2_dma_addr)) {
869 			lnk->wr_rx_v2_dma_addr = 0;
870 			rc = -EIO;
871 			goto dma_unmap;
872 		}
873 		lnk->wr_tx_v2_dma_addr = ib_dma_map_single(ibdev,
874 			lnk->lgr->wr_tx_buf_v2, SMC_WR_BUF_V2_SIZE,
875 			DMA_TO_DEVICE);
876 		if (ib_dma_mapping_error(ibdev, lnk->wr_tx_v2_dma_addr)) {
877 			lnk->wr_tx_v2_dma_addr = 0;
878 			rc = -EIO;
879 			goto dma_unmap;
880 		}
881 	}
882 	lnk->wr_tx_dma_addr = ib_dma_map_single(
883 		ibdev, lnk->wr_tx_bufs,	SMC_WR_BUF_SIZE * lnk->wr_tx_cnt,
884 		DMA_TO_DEVICE);
885 	if (ib_dma_mapping_error(ibdev, lnk->wr_tx_dma_addr)) {
886 		rc = -EIO;
887 		goto dma_unmap;
888 	}
889 	smc_wr_init_sge(lnk);
890 	memset(lnk->wr_tx_mask, 0,
891 	       BITS_TO_LONGS(SMC_WR_BUF_CNT) * sizeof(*lnk->wr_tx_mask));
892 	init_waitqueue_head(&lnk->wr_tx_wait);
893 	atomic_set(&lnk->wr_tx_refcnt, 0);
894 	init_waitqueue_head(&lnk->wr_reg_wait);
895 	atomic_set(&lnk->wr_reg_refcnt, 0);
896 	return rc;
897 
898 dma_unmap:
899 	if (lnk->wr_rx_v2_dma_addr) {
900 		ib_dma_unmap_single(ibdev, lnk->wr_rx_v2_dma_addr,
901 				    SMC_WR_BUF_V2_SIZE,
902 				    DMA_FROM_DEVICE);
903 		lnk->wr_rx_v2_dma_addr = 0;
904 	}
905 	if (lnk->wr_tx_v2_dma_addr) {
906 		ib_dma_unmap_single(ibdev, lnk->wr_tx_v2_dma_addr,
907 				    SMC_WR_BUF_V2_SIZE,
908 				    DMA_TO_DEVICE);
909 		lnk->wr_tx_v2_dma_addr = 0;
910 	}
911 	ib_dma_unmap_single(ibdev, lnk->wr_rx_dma_addr,
912 			    SMC_WR_BUF_SIZE * lnk->wr_rx_cnt,
913 			    DMA_FROM_DEVICE);
914 	lnk->wr_rx_dma_addr = 0;
915 out:
916 	return rc;
917 }
918