xref: /openbmc/linux/net/smc/smc_wr.c (revision 242cdad8)
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 };
48 
49 /******************************** send queue *********************************/
50 
51 /*------------------------------- completion --------------------------------*/
52 
53 static inline int smc_wr_tx_find_pending_index(struct smc_link *link, u64 wr_id)
54 {
55 	u32 i;
56 
57 	for (i = 0; i < link->wr_tx_cnt; i++) {
58 		if (link->wr_tx_pends[i].wr_id == wr_id)
59 			return i;
60 	}
61 	return link->wr_tx_cnt;
62 }
63 
64 static inline void smc_wr_tx_process_cqe(struct ib_wc *wc)
65 {
66 	struct smc_wr_tx_pend pnd_snd;
67 	struct smc_link *link;
68 	u32 pnd_snd_idx;
69 	int i;
70 
71 	link = wc->qp->qp_context;
72 
73 	if (wc->opcode == IB_WC_REG_MR) {
74 		if (wc->status)
75 			link->wr_reg_state = FAILED;
76 		else
77 			link->wr_reg_state = CONFIRMED;
78 		wake_up(&link->wr_reg_wait);
79 		return;
80 	}
81 
82 	pnd_snd_idx = smc_wr_tx_find_pending_index(link, wc->wr_id);
83 	if (pnd_snd_idx == link->wr_tx_cnt)
84 		return;
85 	link->wr_tx_pends[pnd_snd_idx].wc_status = wc->status;
86 	memcpy(&pnd_snd, &link->wr_tx_pends[pnd_snd_idx], sizeof(pnd_snd));
87 	/* clear the full struct smc_wr_tx_pend including .priv */
88 	memset(&link->wr_tx_pends[pnd_snd_idx], 0,
89 	       sizeof(link->wr_tx_pends[pnd_snd_idx]));
90 	memset(&link->wr_tx_bufs[pnd_snd_idx], 0,
91 	       sizeof(link->wr_tx_bufs[pnd_snd_idx]));
92 	if (!test_and_clear_bit(pnd_snd_idx, link->wr_tx_mask))
93 		return;
94 	if (wc->status) {
95 		for_each_set_bit(i, link->wr_tx_mask, link->wr_tx_cnt) {
96 			/* clear full struct smc_wr_tx_pend including .priv */
97 			memset(&link->wr_tx_pends[i], 0,
98 			       sizeof(link->wr_tx_pends[i]));
99 			memset(&link->wr_tx_bufs[i], 0,
100 			       sizeof(link->wr_tx_bufs[i]));
101 			clear_bit(i, link->wr_tx_mask);
102 		}
103 		/* terminate connections of this link group abnormally */
104 		smc_lgr_terminate(smc_get_lgr(link));
105 	}
106 	if (pnd_snd.handler)
107 		pnd_snd.handler(&pnd_snd.priv, link, wc->status);
108 	wake_up(&link->wr_tx_wait);
109 }
110 
111 static void smc_wr_tx_tasklet_fn(unsigned long data)
112 {
113 	struct smc_ib_device *dev = (struct smc_ib_device *)data;
114 	struct ib_wc wc[SMC_WR_MAX_POLL_CQE];
115 	int i = 0, rc;
116 	int polled = 0;
117 
118 again:
119 	polled++;
120 	do {
121 		memset(&wc, 0, sizeof(wc));
122 		rc = ib_poll_cq(dev->roce_cq_send, SMC_WR_MAX_POLL_CQE, wc);
123 		if (polled == 1) {
124 			ib_req_notify_cq(dev->roce_cq_send,
125 					 IB_CQ_NEXT_COMP |
126 					 IB_CQ_REPORT_MISSED_EVENTS);
127 		}
128 		if (!rc)
129 			break;
130 		for (i = 0; i < rc; i++)
131 			smc_wr_tx_process_cqe(&wc[i]);
132 	} while (rc > 0);
133 	if (polled == 1)
134 		goto again;
135 }
136 
137 void smc_wr_tx_cq_handler(struct ib_cq *ib_cq, void *cq_context)
138 {
139 	struct smc_ib_device *dev = (struct smc_ib_device *)cq_context;
140 
141 	tasklet_schedule(&dev->send_tasklet);
142 }
143 
144 /*---------------------------- request submission ---------------------------*/
145 
146 static inline int smc_wr_tx_get_free_slot_index(struct smc_link *link, u32 *idx)
147 {
148 	*idx = link->wr_tx_cnt;
149 	for_each_clear_bit(*idx, link->wr_tx_mask, link->wr_tx_cnt) {
150 		if (!test_and_set_bit(*idx, link->wr_tx_mask))
151 			return 0;
152 	}
153 	*idx = link->wr_tx_cnt;
154 	return -EBUSY;
155 }
156 
157 /**
158  * smc_wr_tx_get_free_slot() - returns buffer for message assembly,
159  *			and sets info for pending transmit tracking
160  * @link:		Pointer to smc_link used to later send the message.
161  * @handler:		Send completion handler function pointer.
162  * @wr_buf:		Out value returns pointer to message buffer.
163  * @wr_pend_priv:	Out value returns pointer serving as handler context.
164  *
165  * Return: 0 on success, or -errno on error.
166  */
167 int smc_wr_tx_get_free_slot(struct smc_link *link,
168 			    smc_wr_tx_handler handler,
169 			    struct smc_wr_buf **wr_buf,
170 			    struct smc_wr_tx_pend_priv **wr_pend_priv)
171 {
172 	struct smc_wr_tx_pend *wr_pend;
173 	u32 idx = link->wr_tx_cnt;
174 	struct ib_send_wr *wr_ib;
175 	u64 wr_id;
176 	int rc;
177 
178 	*wr_buf = NULL;
179 	*wr_pend_priv = NULL;
180 	if (in_softirq()) {
181 		rc = smc_wr_tx_get_free_slot_index(link, &idx);
182 		if (rc)
183 			return rc;
184 	} else {
185 		rc = wait_event_timeout(
186 			link->wr_tx_wait,
187 			link->state == SMC_LNK_INACTIVE ||
188 			(smc_wr_tx_get_free_slot_index(link, &idx) != -EBUSY),
189 			SMC_WR_TX_WAIT_FREE_SLOT_TIME);
190 		if (!rc) {
191 			/* timeout - terminate connections */
192 			smc_lgr_terminate(smc_get_lgr(link));
193 			return -EPIPE;
194 		}
195 		if (idx == link->wr_tx_cnt)
196 			return -EPIPE;
197 	}
198 	wr_id = smc_wr_tx_get_next_wr_id(link);
199 	wr_pend = &link->wr_tx_pends[idx];
200 	wr_pend->wr_id = wr_id;
201 	wr_pend->handler = handler;
202 	wr_pend->link = link;
203 	wr_pend->idx = idx;
204 	wr_ib = &link->wr_tx_ibs[idx];
205 	wr_ib->wr_id = wr_id;
206 	*wr_buf = &link->wr_tx_bufs[idx];
207 	*wr_pend_priv = &wr_pend->priv;
208 	return 0;
209 }
210 
211 int smc_wr_tx_put_slot(struct smc_link *link,
212 		       struct smc_wr_tx_pend_priv *wr_pend_priv)
213 {
214 	struct smc_wr_tx_pend *pend;
215 
216 	pend = container_of(wr_pend_priv, struct smc_wr_tx_pend, priv);
217 	if (pend->idx < link->wr_tx_cnt) {
218 		/* clear the full struct smc_wr_tx_pend including .priv */
219 		memset(&link->wr_tx_pends[pend->idx], 0,
220 		       sizeof(link->wr_tx_pends[pend->idx]));
221 		memset(&link->wr_tx_bufs[pend->idx], 0,
222 		       sizeof(link->wr_tx_bufs[pend->idx]));
223 		test_and_clear_bit(pend->idx, link->wr_tx_mask);
224 		return 1;
225 	}
226 
227 	return 0;
228 }
229 
230 /* Send prepared WR slot via ib_post_send.
231  * @priv: pointer to smc_wr_tx_pend_priv identifying prepared message buffer
232  */
233 int smc_wr_tx_send(struct smc_link *link, struct smc_wr_tx_pend_priv *priv)
234 {
235 	struct smc_wr_tx_pend *pend;
236 	int rc;
237 
238 	ib_req_notify_cq(link->smcibdev->roce_cq_send,
239 			 IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS);
240 	pend = container_of(priv, struct smc_wr_tx_pend, priv);
241 	rc = ib_post_send(link->roce_qp, &link->wr_tx_ibs[pend->idx], NULL);
242 	if (rc) {
243 		smc_wr_tx_put_slot(link, priv);
244 		smc_lgr_terminate(smc_get_lgr(link));
245 	}
246 	return rc;
247 }
248 
249 /* Register a memory region and wait for result. */
250 int smc_wr_reg_send(struct smc_link *link, struct ib_mr *mr)
251 {
252 	int rc;
253 
254 	ib_req_notify_cq(link->smcibdev->roce_cq_send,
255 			 IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS);
256 	link->wr_reg_state = POSTED;
257 	link->wr_reg.wr.wr_id = (u64)(uintptr_t)mr;
258 	link->wr_reg.mr = mr;
259 	link->wr_reg.key = mr->rkey;
260 	rc = ib_post_send(link->roce_qp, &link->wr_reg.wr, NULL);
261 	if (rc)
262 		return rc;
263 
264 	rc = wait_event_interruptible_timeout(link->wr_reg_wait,
265 					      (link->wr_reg_state != POSTED),
266 					      SMC_WR_REG_MR_WAIT_TIME);
267 	if (!rc) {
268 		/* timeout - terminate connections */
269 		smc_lgr_terminate(smc_get_lgr(link));
270 		return -EPIPE;
271 	}
272 	if (rc == -ERESTARTSYS)
273 		return -EINTR;
274 	switch (link->wr_reg_state) {
275 	case CONFIRMED:
276 		rc = 0;
277 		break;
278 	case FAILED:
279 		rc = -EIO;
280 		break;
281 	case POSTED:
282 		rc = -EPIPE;
283 		break;
284 	}
285 	return rc;
286 }
287 
288 void smc_wr_tx_dismiss_slots(struct smc_link *link, u8 wr_tx_hdr_type,
289 			     smc_wr_tx_filter filter,
290 			     smc_wr_tx_dismisser dismisser,
291 			     unsigned long data)
292 {
293 	struct smc_wr_tx_pend_priv *tx_pend;
294 	struct smc_wr_rx_hdr *wr_tx;
295 	int i;
296 
297 	for_each_set_bit(i, link->wr_tx_mask, link->wr_tx_cnt) {
298 		wr_tx = (struct smc_wr_rx_hdr *)&link->wr_tx_bufs[i];
299 		if (wr_tx->type != wr_tx_hdr_type)
300 			continue;
301 		tx_pend = &link->wr_tx_pends[i].priv;
302 		if (filter(tx_pend, data))
303 			dismisser(tx_pend);
304 	}
305 }
306 
307 /****************************** receive queue ********************************/
308 
309 int smc_wr_rx_register_handler(struct smc_wr_rx_handler *handler)
310 {
311 	struct smc_wr_rx_handler *h_iter;
312 	int rc = 0;
313 
314 	spin_lock(&smc_wr_rx_hash_lock);
315 	hash_for_each_possible(smc_wr_rx_hash, h_iter, list, handler->type) {
316 		if (h_iter->type == handler->type) {
317 			rc = -EEXIST;
318 			goto out_unlock;
319 		}
320 	}
321 	hash_add(smc_wr_rx_hash, &handler->list, handler->type);
322 out_unlock:
323 	spin_unlock(&smc_wr_rx_hash_lock);
324 	return rc;
325 }
326 
327 /* Demultiplex a received work request based on the message type to its handler.
328  * Relies on smc_wr_rx_hash having been completely filled before any IB WRs,
329  * and not being modified any more afterwards so we don't need to lock it.
330  */
331 static inline void smc_wr_rx_demultiplex(struct ib_wc *wc)
332 {
333 	struct smc_link *link = (struct smc_link *)wc->qp->qp_context;
334 	struct smc_wr_rx_handler *handler;
335 	struct smc_wr_rx_hdr *wr_rx;
336 	u64 temp_wr_id;
337 	u32 index;
338 
339 	if (wc->byte_len < sizeof(*wr_rx))
340 		return; /* short message */
341 	temp_wr_id = wc->wr_id;
342 	index = do_div(temp_wr_id, link->wr_rx_cnt);
343 	wr_rx = (struct smc_wr_rx_hdr *)&link->wr_rx_bufs[index];
344 	hash_for_each_possible(smc_wr_rx_hash, handler, list, wr_rx->type) {
345 		if (handler->type == wr_rx->type)
346 			handler->handler(wc, wr_rx);
347 	}
348 }
349 
350 static inline void smc_wr_rx_process_cqes(struct ib_wc wc[], int num)
351 {
352 	struct smc_link *link;
353 	int i;
354 
355 	for (i = 0; i < num; i++) {
356 		link = wc[i].qp->qp_context;
357 		if (wc[i].status == IB_WC_SUCCESS) {
358 			link->wr_rx_tstamp = jiffies;
359 			smc_wr_rx_demultiplex(&wc[i]);
360 			smc_wr_rx_post(link); /* refill WR RX */
361 		} else {
362 			/* handle status errors */
363 			switch (wc[i].status) {
364 			case IB_WC_RETRY_EXC_ERR:
365 			case IB_WC_RNR_RETRY_EXC_ERR:
366 			case IB_WC_WR_FLUSH_ERR:
367 				/* terminate connections of this link group
368 				 * abnormally
369 				 */
370 				smc_lgr_terminate(smc_get_lgr(link));
371 				break;
372 			default:
373 				smc_wr_rx_post(link); /* refill WR RX */
374 				break;
375 			}
376 		}
377 	}
378 }
379 
380 static void smc_wr_rx_tasklet_fn(unsigned long data)
381 {
382 	struct smc_ib_device *dev = (struct smc_ib_device *)data;
383 	struct ib_wc wc[SMC_WR_MAX_POLL_CQE];
384 	int polled = 0;
385 	int rc;
386 
387 again:
388 	polled++;
389 	do {
390 		memset(&wc, 0, sizeof(wc));
391 		rc = ib_poll_cq(dev->roce_cq_recv, SMC_WR_MAX_POLL_CQE, wc);
392 		if (polled == 1) {
393 			ib_req_notify_cq(dev->roce_cq_recv,
394 					 IB_CQ_SOLICITED_MASK
395 					 | IB_CQ_REPORT_MISSED_EVENTS);
396 		}
397 		if (!rc)
398 			break;
399 		smc_wr_rx_process_cqes(&wc[0], rc);
400 	} while (rc > 0);
401 	if (polled == 1)
402 		goto again;
403 }
404 
405 void smc_wr_rx_cq_handler(struct ib_cq *ib_cq, void *cq_context)
406 {
407 	struct smc_ib_device *dev = (struct smc_ib_device *)cq_context;
408 
409 	tasklet_schedule(&dev->recv_tasklet);
410 }
411 
412 int smc_wr_rx_post_init(struct smc_link *link)
413 {
414 	u32 i;
415 	int rc = 0;
416 
417 	for (i = 0; i < link->wr_rx_cnt; i++)
418 		rc = smc_wr_rx_post(link);
419 	return rc;
420 }
421 
422 /***************************** init, exit, misc ******************************/
423 
424 void smc_wr_remember_qp_attr(struct smc_link *lnk)
425 {
426 	struct ib_qp_attr *attr = &lnk->qp_attr;
427 	struct ib_qp_init_attr init_attr;
428 
429 	memset(attr, 0, sizeof(*attr));
430 	memset(&init_attr, 0, sizeof(init_attr));
431 	ib_query_qp(lnk->roce_qp, attr,
432 		    IB_QP_STATE |
433 		    IB_QP_CUR_STATE |
434 		    IB_QP_PKEY_INDEX |
435 		    IB_QP_PORT |
436 		    IB_QP_QKEY |
437 		    IB_QP_AV |
438 		    IB_QP_PATH_MTU |
439 		    IB_QP_TIMEOUT |
440 		    IB_QP_RETRY_CNT |
441 		    IB_QP_RNR_RETRY |
442 		    IB_QP_RQ_PSN |
443 		    IB_QP_ALT_PATH |
444 		    IB_QP_MIN_RNR_TIMER |
445 		    IB_QP_SQ_PSN |
446 		    IB_QP_PATH_MIG_STATE |
447 		    IB_QP_CAP |
448 		    IB_QP_DEST_QPN,
449 		    &init_attr);
450 
451 	lnk->wr_tx_cnt = min_t(size_t, SMC_WR_BUF_CNT,
452 			       lnk->qp_attr.cap.max_send_wr);
453 	lnk->wr_rx_cnt = min_t(size_t, SMC_WR_BUF_CNT * 3,
454 			       lnk->qp_attr.cap.max_recv_wr);
455 }
456 
457 static void smc_wr_init_sge(struct smc_link *lnk)
458 {
459 	u32 i;
460 
461 	for (i = 0; i < lnk->wr_tx_cnt; i++) {
462 		lnk->wr_tx_sges[i].addr =
463 			lnk->wr_tx_dma_addr + i * SMC_WR_BUF_SIZE;
464 		lnk->wr_tx_sges[i].length = SMC_WR_TX_SIZE;
465 		lnk->wr_tx_sges[i].lkey = lnk->roce_pd->local_dma_lkey;
466 		lnk->wr_tx_ibs[i].next = NULL;
467 		lnk->wr_tx_ibs[i].sg_list = &lnk->wr_tx_sges[i];
468 		lnk->wr_tx_ibs[i].num_sge = 1;
469 		lnk->wr_tx_ibs[i].opcode = IB_WR_SEND;
470 		lnk->wr_tx_ibs[i].send_flags =
471 			IB_SEND_SIGNALED | IB_SEND_SOLICITED;
472 	}
473 	for (i = 0; i < lnk->wr_rx_cnt; i++) {
474 		lnk->wr_rx_sges[i].addr =
475 			lnk->wr_rx_dma_addr + i * SMC_WR_BUF_SIZE;
476 		lnk->wr_rx_sges[i].length = SMC_WR_BUF_SIZE;
477 		lnk->wr_rx_sges[i].lkey = lnk->roce_pd->local_dma_lkey;
478 		lnk->wr_rx_ibs[i].next = NULL;
479 		lnk->wr_rx_ibs[i].sg_list = &lnk->wr_rx_sges[i];
480 		lnk->wr_rx_ibs[i].num_sge = 1;
481 	}
482 	lnk->wr_reg.wr.next = NULL;
483 	lnk->wr_reg.wr.num_sge = 0;
484 	lnk->wr_reg.wr.send_flags = IB_SEND_SIGNALED;
485 	lnk->wr_reg.wr.opcode = IB_WR_REG_MR;
486 	lnk->wr_reg.access = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE;
487 }
488 
489 void smc_wr_free_link(struct smc_link *lnk)
490 {
491 	struct ib_device *ibdev;
492 
493 	memset(lnk->wr_tx_mask, 0,
494 	       BITS_TO_LONGS(SMC_WR_BUF_CNT) * sizeof(*lnk->wr_tx_mask));
495 
496 	if (!lnk->smcibdev)
497 		return;
498 	ibdev = lnk->smcibdev->ibdev;
499 
500 	if (lnk->wr_rx_dma_addr) {
501 		ib_dma_unmap_single(ibdev, lnk->wr_rx_dma_addr,
502 				    SMC_WR_BUF_SIZE * lnk->wr_rx_cnt,
503 				    DMA_FROM_DEVICE);
504 		lnk->wr_rx_dma_addr = 0;
505 	}
506 	if (lnk->wr_tx_dma_addr) {
507 		ib_dma_unmap_single(ibdev, lnk->wr_tx_dma_addr,
508 				    SMC_WR_BUF_SIZE * lnk->wr_tx_cnt,
509 				    DMA_TO_DEVICE);
510 		lnk->wr_tx_dma_addr = 0;
511 	}
512 }
513 
514 void smc_wr_free_link_mem(struct smc_link *lnk)
515 {
516 	kfree(lnk->wr_tx_pends);
517 	lnk->wr_tx_pends = NULL;
518 	kfree(lnk->wr_tx_mask);
519 	lnk->wr_tx_mask = NULL;
520 	kfree(lnk->wr_tx_sges);
521 	lnk->wr_tx_sges = NULL;
522 	kfree(lnk->wr_rx_sges);
523 	lnk->wr_rx_sges = NULL;
524 	kfree(lnk->wr_rx_ibs);
525 	lnk->wr_rx_ibs = NULL;
526 	kfree(lnk->wr_tx_ibs);
527 	lnk->wr_tx_ibs = NULL;
528 	kfree(lnk->wr_tx_bufs);
529 	lnk->wr_tx_bufs = NULL;
530 	kfree(lnk->wr_rx_bufs);
531 	lnk->wr_rx_bufs = NULL;
532 }
533 
534 int smc_wr_alloc_link_mem(struct smc_link *link)
535 {
536 	/* allocate link related memory */
537 	link->wr_tx_bufs = kcalloc(SMC_WR_BUF_CNT, SMC_WR_BUF_SIZE, GFP_KERNEL);
538 	if (!link->wr_tx_bufs)
539 		goto no_mem;
540 	link->wr_rx_bufs = kcalloc(SMC_WR_BUF_CNT * 3, SMC_WR_BUF_SIZE,
541 				   GFP_KERNEL);
542 	if (!link->wr_rx_bufs)
543 		goto no_mem_wr_tx_bufs;
544 	link->wr_tx_ibs = kcalloc(SMC_WR_BUF_CNT, sizeof(link->wr_tx_ibs[0]),
545 				  GFP_KERNEL);
546 	if (!link->wr_tx_ibs)
547 		goto no_mem_wr_rx_bufs;
548 	link->wr_rx_ibs = kcalloc(SMC_WR_BUF_CNT * 3,
549 				  sizeof(link->wr_rx_ibs[0]),
550 				  GFP_KERNEL);
551 	if (!link->wr_rx_ibs)
552 		goto no_mem_wr_tx_ibs;
553 	link->wr_tx_sges = kcalloc(SMC_WR_BUF_CNT, sizeof(link->wr_tx_sges[0]),
554 				   GFP_KERNEL);
555 	if (!link->wr_tx_sges)
556 		goto no_mem_wr_rx_ibs;
557 	link->wr_rx_sges = kcalloc(SMC_WR_BUF_CNT * 3,
558 				   sizeof(link->wr_rx_sges[0]),
559 				   GFP_KERNEL);
560 	if (!link->wr_rx_sges)
561 		goto no_mem_wr_tx_sges;
562 	link->wr_tx_mask = kcalloc(BITS_TO_LONGS(SMC_WR_BUF_CNT),
563 				   sizeof(*link->wr_tx_mask),
564 				   GFP_KERNEL);
565 	if (!link->wr_tx_mask)
566 		goto no_mem_wr_rx_sges;
567 	link->wr_tx_pends = kcalloc(SMC_WR_BUF_CNT,
568 				    sizeof(link->wr_tx_pends[0]),
569 				    GFP_KERNEL);
570 	if (!link->wr_tx_pends)
571 		goto no_mem_wr_tx_mask;
572 	return 0;
573 
574 no_mem_wr_tx_mask:
575 	kfree(link->wr_tx_mask);
576 no_mem_wr_rx_sges:
577 	kfree(link->wr_rx_sges);
578 no_mem_wr_tx_sges:
579 	kfree(link->wr_tx_sges);
580 no_mem_wr_rx_ibs:
581 	kfree(link->wr_rx_ibs);
582 no_mem_wr_tx_ibs:
583 	kfree(link->wr_tx_ibs);
584 no_mem_wr_rx_bufs:
585 	kfree(link->wr_rx_bufs);
586 no_mem_wr_tx_bufs:
587 	kfree(link->wr_tx_bufs);
588 no_mem:
589 	return -ENOMEM;
590 }
591 
592 void smc_wr_remove_dev(struct smc_ib_device *smcibdev)
593 {
594 	tasklet_kill(&smcibdev->recv_tasklet);
595 	tasklet_kill(&smcibdev->send_tasklet);
596 }
597 
598 void smc_wr_add_dev(struct smc_ib_device *smcibdev)
599 {
600 	tasklet_init(&smcibdev->recv_tasklet, smc_wr_rx_tasklet_fn,
601 		     (unsigned long)smcibdev);
602 	tasklet_init(&smcibdev->send_tasklet, smc_wr_tx_tasklet_fn,
603 		     (unsigned long)smcibdev);
604 }
605 
606 int smc_wr_create_link(struct smc_link *lnk)
607 {
608 	struct ib_device *ibdev = lnk->smcibdev->ibdev;
609 	int rc = 0;
610 
611 	smc_wr_tx_set_wr_id(&lnk->wr_tx_id, 0);
612 	lnk->wr_rx_id = 0;
613 	lnk->wr_rx_dma_addr = ib_dma_map_single(
614 		ibdev, lnk->wr_rx_bufs,	SMC_WR_BUF_SIZE * lnk->wr_rx_cnt,
615 		DMA_FROM_DEVICE);
616 	if (ib_dma_mapping_error(ibdev, lnk->wr_rx_dma_addr)) {
617 		lnk->wr_rx_dma_addr = 0;
618 		rc = -EIO;
619 		goto out;
620 	}
621 	lnk->wr_tx_dma_addr = ib_dma_map_single(
622 		ibdev, lnk->wr_tx_bufs,	SMC_WR_BUF_SIZE * lnk->wr_tx_cnt,
623 		DMA_TO_DEVICE);
624 	if (ib_dma_mapping_error(ibdev, lnk->wr_tx_dma_addr)) {
625 		rc = -EIO;
626 		goto dma_unmap;
627 	}
628 	smc_wr_init_sge(lnk);
629 	memset(lnk->wr_tx_mask, 0,
630 	       BITS_TO_LONGS(SMC_WR_BUF_CNT) * sizeof(*lnk->wr_tx_mask));
631 	init_waitqueue_head(&lnk->wr_tx_wait);
632 	init_waitqueue_head(&lnk->wr_reg_wait);
633 	return rc;
634 
635 dma_unmap:
636 	ib_dma_unmap_single(ibdev, lnk->wr_rx_dma_addr,
637 			    SMC_WR_BUF_SIZE * lnk->wr_rx_cnt,
638 			    DMA_FROM_DEVICE);
639 	lnk->wr_rx_dma_addr = 0;
640 out:
641 	return rc;
642 }
643