xref: /openbmc/linux/net/smc/smc_wr.c (revision b8265621)
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 int smc_wr_tx_wait_no_pending_sends(struct smc_link *link)
66 {
67 	if (wait_event_timeout(link->wr_tx_wait, !smc_wr_is_tx_pend(link),
68 			       SMC_WR_TX_WAIT_PENDING_TIME))
69 		return 0;
70 	else /* timeout */
71 		return -EPIPE;
72 }
73 
74 static inline int smc_wr_tx_find_pending_index(struct smc_link *link, u64 wr_id)
75 {
76 	u32 i;
77 
78 	for (i = 0; i < link->wr_tx_cnt; i++) {
79 		if (link->wr_tx_pends[i].wr_id == wr_id)
80 			return i;
81 	}
82 	return link->wr_tx_cnt;
83 }
84 
85 static inline void smc_wr_tx_process_cqe(struct ib_wc *wc)
86 {
87 	struct smc_wr_tx_pend pnd_snd;
88 	struct smc_link *link;
89 	u32 pnd_snd_idx;
90 	int i;
91 
92 	link = wc->qp->qp_context;
93 
94 	if (wc->opcode == IB_WC_REG_MR) {
95 		if (wc->status)
96 			link->wr_reg_state = FAILED;
97 		else
98 			link->wr_reg_state = CONFIRMED;
99 		smc_wr_wakeup_reg_wait(link);
100 		return;
101 	}
102 
103 	pnd_snd_idx = smc_wr_tx_find_pending_index(link, wc->wr_id);
104 	if (pnd_snd_idx == link->wr_tx_cnt)
105 		return;
106 	link->wr_tx_pends[pnd_snd_idx].wc_status = wc->status;
107 	if (link->wr_tx_pends[pnd_snd_idx].compl_requested)
108 		complete(&link->wr_tx_compl[pnd_snd_idx]);
109 	memcpy(&pnd_snd, &link->wr_tx_pends[pnd_snd_idx], sizeof(pnd_snd));
110 	/* clear the full struct smc_wr_tx_pend including .priv */
111 	memset(&link->wr_tx_pends[pnd_snd_idx], 0,
112 	       sizeof(link->wr_tx_pends[pnd_snd_idx]));
113 	memset(&link->wr_tx_bufs[pnd_snd_idx], 0,
114 	       sizeof(link->wr_tx_bufs[pnd_snd_idx]));
115 	if (!test_and_clear_bit(pnd_snd_idx, link->wr_tx_mask))
116 		return;
117 	if (wc->status) {
118 		for_each_set_bit(i, link->wr_tx_mask, link->wr_tx_cnt) {
119 			/* clear full struct smc_wr_tx_pend including .priv */
120 			memset(&link->wr_tx_pends[i], 0,
121 			       sizeof(link->wr_tx_pends[i]));
122 			memset(&link->wr_tx_bufs[i], 0,
123 			       sizeof(link->wr_tx_bufs[i]));
124 			clear_bit(i, link->wr_tx_mask);
125 		}
126 		/* terminate link */
127 		smcr_link_down_cond_sched(link);
128 	}
129 	if (pnd_snd.handler)
130 		pnd_snd.handler(&pnd_snd.priv, link, wc->status);
131 	wake_up(&link->wr_tx_wait);
132 }
133 
134 static void smc_wr_tx_tasklet_fn(unsigned long data)
135 {
136 	struct smc_ib_device *dev = (struct smc_ib_device *)data;
137 	struct ib_wc wc[SMC_WR_MAX_POLL_CQE];
138 	int i = 0, rc;
139 	int polled = 0;
140 
141 again:
142 	polled++;
143 	do {
144 		memset(&wc, 0, sizeof(wc));
145 		rc = ib_poll_cq(dev->roce_cq_send, SMC_WR_MAX_POLL_CQE, wc);
146 		if (polled == 1) {
147 			ib_req_notify_cq(dev->roce_cq_send,
148 					 IB_CQ_NEXT_COMP |
149 					 IB_CQ_REPORT_MISSED_EVENTS);
150 		}
151 		if (!rc)
152 			break;
153 		for (i = 0; i < rc; i++)
154 			smc_wr_tx_process_cqe(&wc[i]);
155 	} while (rc > 0);
156 	if (polled == 1)
157 		goto again;
158 }
159 
160 void smc_wr_tx_cq_handler(struct ib_cq *ib_cq, void *cq_context)
161 {
162 	struct smc_ib_device *dev = (struct smc_ib_device *)cq_context;
163 
164 	tasklet_schedule(&dev->send_tasklet);
165 }
166 
167 /*---------------------------- request submission ---------------------------*/
168 
169 static inline int smc_wr_tx_get_free_slot_index(struct smc_link *link, u32 *idx)
170 {
171 	*idx = link->wr_tx_cnt;
172 	if (!smc_link_usable(link))
173 		return -ENOLINK;
174 	for_each_clear_bit(*idx, link->wr_tx_mask, link->wr_tx_cnt) {
175 		if (!test_and_set_bit(*idx, link->wr_tx_mask))
176 			return 0;
177 	}
178 	*idx = link->wr_tx_cnt;
179 	return -EBUSY;
180 }
181 
182 /**
183  * smc_wr_tx_get_free_slot() - returns buffer for message assembly,
184  *			and sets info for pending transmit tracking
185  * @link:		Pointer to smc_link used to later send the message.
186  * @handler:		Send completion handler function pointer.
187  * @wr_buf:		Out value returns pointer to message buffer.
188  * @wr_rdma_buf:	Out value returns pointer to rdma work request.
189  * @wr_pend_priv:	Out value returns pointer serving as handler context.
190  *
191  * Return: 0 on success, or -errno on error.
192  */
193 int smc_wr_tx_get_free_slot(struct smc_link *link,
194 			    smc_wr_tx_handler handler,
195 			    struct smc_wr_buf **wr_buf,
196 			    struct smc_rdma_wr **wr_rdma_buf,
197 			    struct smc_wr_tx_pend_priv **wr_pend_priv)
198 {
199 	struct smc_link_group *lgr = smc_get_lgr(link);
200 	struct smc_wr_tx_pend *wr_pend;
201 	u32 idx = link->wr_tx_cnt;
202 	struct ib_send_wr *wr_ib;
203 	u64 wr_id;
204 	int rc;
205 
206 	*wr_buf = NULL;
207 	*wr_pend_priv = NULL;
208 	if (in_softirq() || lgr->terminating) {
209 		rc = smc_wr_tx_get_free_slot_index(link, &idx);
210 		if (rc)
211 			return rc;
212 	} else {
213 		rc = wait_event_interruptible_timeout(
214 			link->wr_tx_wait,
215 			!smc_link_usable(link) ||
216 			lgr->terminating ||
217 			(smc_wr_tx_get_free_slot_index(link, &idx) != -EBUSY),
218 			SMC_WR_TX_WAIT_FREE_SLOT_TIME);
219 		if (!rc) {
220 			/* timeout - terminate link */
221 			smcr_link_down_cond_sched(link);
222 			return -EPIPE;
223 		}
224 		if (idx == link->wr_tx_cnt)
225 			return -EPIPE;
226 	}
227 	wr_id = smc_wr_tx_get_next_wr_id(link);
228 	wr_pend = &link->wr_tx_pends[idx];
229 	wr_pend->wr_id = wr_id;
230 	wr_pend->handler = handler;
231 	wr_pend->link = link;
232 	wr_pend->idx = idx;
233 	wr_ib = &link->wr_tx_ibs[idx];
234 	wr_ib->wr_id = wr_id;
235 	*wr_buf = &link->wr_tx_bufs[idx];
236 	if (wr_rdma_buf)
237 		*wr_rdma_buf = &link->wr_tx_rdmas[idx];
238 	*wr_pend_priv = &wr_pend->priv;
239 	return 0;
240 }
241 
242 int smc_wr_tx_put_slot(struct smc_link *link,
243 		       struct smc_wr_tx_pend_priv *wr_pend_priv)
244 {
245 	struct smc_wr_tx_pend *pend;
246 
247 	pend = container_of(wr_pend_priv, struct smc_wr_tx_pend, priv);
248 	if (pend->idx < link->wr_tx_cnt) {
249 		u32 idx = pend->idx;
250 
251 		/* clear the full struct smc_wr_tx_pend including .priv */
252 		memset(&link->wr_tx_pends[idx], 0,
253 		       sizeof(link->wr_tx_pends[idx]));
254 		memset(&link->wr_tx_bufs[idx], 0,
255 		       sizeof(link->wr_tx_bufs[idx]));
256 		test_and_clear_bit(idx, link->wr_tx_mask);
257 		wake_up(&link->wr_tx_wait);
258 		return 1;
259 	}
260 
261 	return 0;
262 }
263 
264 /* Send prepared WR slot via ib_post_send.
265  * @priv: pointer to smc_wr_tx_pend_priv identifying prepared message buffer
266  */
267 int smc_wr_tx_send(struct smc_link *link, struct smc_wr_tx_pend_priv *priv)
268 {
269 	struct smc_wr_tx_pend *pend;
270 	int rc;
271 
272 	ib_req_notify_cq(link->smcibdev->roce_cq_send,
273 			 IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS);
274 	pend = container_of(priv, struct smc_wr_tx_pend, priv);
275 	rc = ib_post_send(link->roce_qp, &link->wr_tx_ibs[pend->idx], NULL);
276 	if (rc) {
277 		smc_wr_tx_put_slot(link, priv);
278 		smcr_link_down_cond_sched(link);
279 	}
280 	return rc;
281 }
282 
283 /* Send prepared WR slot via ib_post_send and wait for send completion
284  * notification.
285  * @priv: pointer to smc_wr_tx_pend_priv identifying prepared message buffer
286  */
287 int smc_wr_tx_send_wait(struct smc_link *link, struct smc_wr_tx_pend_priv *priv,
288 			unsigned long timeout)
289 {
290 	struct smc_wr_tx_pend *pend;
291 	int rc;
292 
293 	pend = container_of(priv, struct smc_wr_tx_pend, priv);
294 	pend->compl_requested = 1;
295 	init_completion(&link->wr_tx_compl[pend->idx]);
296 
297 	rc = smc_wr_tx_send(link, priv);
298 	if (rc)
299 		return rc;
300 	/* wait for completion by smc_wr_tx_process_cqe() */
301 	rc = wait_for_completion_interruptible_timeout(
302 					&link->wr_tx_compl[pend->idx], timeout);
303 	if (rc <= 0)
304 		rc = -ENODATA;
305 	if (rc > 0)
306 		rc = 0;
307 	return rc;
308 }
309 
310 /* Register a memory region and wait for result. */
311 int smc_wr_reg_send(struct smc_link *link, struct ib_mr *mr)
312 {
313 	int rc;
314 
315 	ib_req_notify_cq(link->smcibdev->roce_cq_send,
316 			 IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS);
317 	link->wr_reg_state = POSTED;
318 	link->wr_reg.wr.wr_id = (u64)(uintptr_t)mr;
319 	link->wr_reg.mr = mr;
320 	link->wr_reg.key = mr->rkey;
321 	rc = ib_post_send(link->roce_qp, &link->wr_reg.wr, NULL);
322 	if (rc)
323 		return rc;
324 
325 	rc = wait_event_interruptible_timeout(link->wr_reg_wait,
326 					      (link->wr_reg_state != POSTED),
327 					      SMC_WR_REG_MR_WAIT_TIME);
328 	if (!rc) {
329 		/* timeout - terminate link */
330 		smcr_link_down_cond_sched(link);
331 		return -EPIPE;
332 	}
333 	if (rc == -ERESTARTSYS)
334 		return -EINTR;
335 	switch (link->wr_reg_state) {
336 	case CONFIRMED:
337 		rc = 0;
338 		break;
339 	case FAILED:
340 		rc = -EIO;
341 		break;
342 	case POSTED:
343 		rc = -EPIPE;
344 		break;
345 	}
346 	return rc;
347 }
348 
349 void smc_wr_tx_dismiss_slots(struct smc_link *link, u8 wr_tx_hdr_type,
350 			     smc_wr_tx_filter filter,
351 			     smc_wr_tx_dismisser dismisser,
352 			     unsigned long data)
353 {
354 	struct smc_wr_tx_pend_priv *tx_pend;
355 	struct smc_wr_rx_hdr *wr_tx;
356 	int i;
357 
358 	for_each_set_bit(i, link->wr_tx_mask, link->wr_tx_cnt) {
359 		wr_tx = (struct smc_wr_rx_hdr *)&link->wr_tx_bufs[i];
360 		if (wr_tx->type != wr_tx_hdr_type)
361 			continue;
362 		tx_pend = &link->wr_tx_pends[i].priv;
363 		if (filter(tx_pend, data))
364 			dismisser(tx_pend);
365 	}
366 }
367 
368 /****************************** receive queue ********************************/
369 
370 int smc_wr_rx_register_handler(struct smc_wr_rx_handler *handler)
371 {
372 	struct smc_wr_rx_handler *h_iter;
373 	int rc = 0;
374 
375 	spin_lock(&smc_wr_rx_hash_lock);
376 	hash_for_each_possible(smc_wr_rx_hash, h_iter, list, handler->type) {
377 		if (h_iter->type == handler->type) {
378 			rc = -EEXIST;
379 			goto out_unlock;
380 		}
381 	}
382 	hash_add(smc_wr_rx_hash, &handler->list, handler->type);
383 out_unlock:
384 	spin_unlock(&smc_wr_rx_hash_lock);
385 	return rc;
386 }
387 
388 /* Demultiplex a received work request based on the message type to its handler.
389  * Relies on smc_wr_rx_hash having been completely filled before any IB WRs,
390  * and not being modified any more afterwards so we don't need to lock it.
391  */
392 static inline void smc_wr_rx_demultiplex(struct ib_wc *wc)
393 {
394 	struct smc_link *link = (struct smc_link *)wc->qp->qp_context;
395 	struct smc_wr_rx_handler *handler;
396 	struct smc_wr_rx_hdr *wr_rx;
397 	u64 temp_wr_id;
398 	u32 index;
399 
400 	if (wc->byte_len < sizeof(*wr_rx))
401 		return; /* short message */
402 	temp_wr_id = wc->wr_id;
403 	index = do_div(temp_wr_id, link->wr_rx_cnt);
404 	wr_rx = (struct smc_wr_rx_hdr *)&link->wr_rx_bufs[index];
405 	hash_for_each_possible(smc_wr_rx_hash, handler, list, wr_rx->type) {
406 		if (handler->type == wr_rx->type)
407 			handler->handler(wc, wr_rx);
408 	}
409 }
410 
411 static inline void smc_wr_rx_process_cqes(struct ib_wc wc[], int num)
412 {
413 	struct smc_link *link;
414 	int i;
415 
416 	for (i = 0; i < num; i++) {
417 		link = wc[i].qp->qp_context;
418 		if (wc[i].status == IB_WC_SUCCESS) {
419 			link->wr_rx_tstamp = jiffies;
420 			smc_wr_rx_demultiplex(&wc[i]);
421 			smc_wr_rx_post(link); /* refill WR RX */
422 		} else {
423 			/* handle status errors */
424 			switch (wc[i].status) {
425 			case IB_WC_RETRY_EXC_ERR:
426 			case IB_WC_RNR_RETRY_EXC_ERR:
427 			case IB_WC_WR_FLUSH_ERR:
428 				smcr_link_down_cond_sched(link);
429 				break;
430 			default:
431 				smc_wr_rx_post(link); /* refill WR RX */
432 				break;
433 			}
434 		}
435 	}
436 }
437 
438 static void smc_wr_rx_tasklet_fn(unsigned long data)
439 {
440 	struct smc_ib_device *dev = (struct smc_ib_device *)data;
441 	struct ib_wc wc[SMC_WR_MAX_POLL_CQE];
442 	int polled = 0;
443 	int rc;
444 
445 again:
446 	polled++;
447 	do {
448 		memset(&wc, 0, sizeof(wc));
449 		rc = ib_poll_cq(dev->roce_cq_recv, SMC_WR_MAX_POLL_CQE, wc);
450 		if (polled == 1) {
451 			ib_req_notify_cq(dev->roce_cq_recv,
452 					 IB_CQ_SOLICITED_MASK
453 					 | IB_CQ_REPORT_MISSED_EVENTS);
454 		}
455 		if (!rc)
456 			break;
457 		smc_wr_rx_process_cqes(&wc[0], rc);
458 	} while (rc > 0);
459 	if (polled == 1)
460 		goto again;
461 }
462 
463 void smc_wr_rx_cq_handler(struct ib_cq *ib_cq, void *cq_context)
464 {
465 	struct smc_ib_device *dev = (struct smc_ib_device *)cq_context;
466 
467 	tasklet_schedule(&dev->recv_tasklet);
468 }
469 
470 int smc_wr_rx_post_init(struct smc_link *link)
471 {
472 	u32 i;
473 	int rc = 0;
474 
475 	for (i = 0; i < link->wr_rx_cnt; i++)
476 		rc = smc_wr_rx_post(link);
477 	return rc;
478 }
479 
480 /***************************** init, exit, misc ******************************/
481 
482 void smc_wr_remember_qp_attr(struct smc_link *lnk)
483 {
484 	struct ib_qp_attr *attr = &lnk->qp_attr;
485 	struct ib_qp_init_attr init_attr;
486 
487 	memset(attr, 0, sizeof(*attr));
488 	memset(&init_attr, 0, sizeof(init_attr));
489 	ib_query_qp(lnk->roce_qp, attr,
490 		    IB_QP_STATE |
491 		    IB_QP_CUR_STATE |
492 		    IB_QP_PKEY_INDEX |
493 		    IB_QP_PORT |
494 		    IB_QP_QKEY |
495 		    IB_QP_AV |
496 		    IB_QP_PATH_MTU |
497 		    IB_QP_TIMEOUT |
498 		    IB_QP_RETRY_CNT |
499 		    IB_QP_RNR_RETRY |
500 		    IB_QP_RQ_PSN |
501 		    IB_QP_ALT_PATH |
502 		    IB_QP_MIN_RNR_TIMER |
503 		    IB_QP_SQ_PSN |
504 		    IB_QP_PATH_MIG_STATE |
505 		    IB_QP_CAP |
506 		    IB_QP_DEST_QPN,
507 		    &init_attr);
508 
509 	lnk->wr_tx_cnt = min_t(size_t, SMC_WR_BUF_CNT,
510 			       lnk->qp_attr.cap.max_send_wr);
511 	lnk->wr_rx_cnt = min_t(size_t, SMC_WR_BUF_CNT * 3,
512 			       lnk->qp_attr.cap.max_recv_wr);
513 }
514 
515 static void smc_wr_init_sge(struct smc_link *lnk)
516 {
517 	u32 i;
518 
519 	for (i = 0; i < lnk->wr_tx_cnt; i++) {
520 		lnk->wr_tx_sges[i].addr =
521 			lnk->wr_tx_dma_addr + i * SMC_WR_BUF_SIZE;
522 		lnk->wr_tx_sges[i].length = SMC_WR_TX_SIZE;
523 		lnk->wr_tx_sges[i].lkey = lnk->roce_pd->local_dma_lkey;
524 		lnk->wr_tx_rdma_sges[i].tx_rdma_sge[0].wr_tx_rdma_sge[0].lkey =
525 			lnk->roce_pd->local_dma_lkey;
526 		lnk->wr_tx_rdma_sges[i].tx_rdma_sge[0].wr_tx_rdma_sge[1].lkey =
527 			lnk->roce_pd->local_dma_lkey;
528 		lnk->wr_tx_rdma_sges[i].tx_rdma_sge[1].wr_tx_rdma_sge[0].lkey =
529 			lnk->roce_pd->local_dma_lkey;
530 		lnk->wr_tx_rdma_sges[i].tx_rdma_sge[1].wr_tx_rdma_sge[1].lkey =
531 			lnk->roce_pd->local_dma_lkey;
532 		lnk->wr_tx_ibs[i].next = NULL;
533 		lnk->wr_tx_ibs[i].sg_list = &lnk->wr_tx_sges[i];
534 		lnk->wr_tx_ibs[i].num_sge = 1;
535 		lnk->wr_tx_ibs[i].opcode = IB_WR_SEND;
536 		lnk->wr_tx_ibs[i].send_flags =
537 			IB_SEND_SIGNALED | IB_SEND_SOLICITED;
538 		lnk->wr_tx_rdmas[i].wr_tx_rdma[0].wr.opcode = IB_WR_RDMA_WRITE;
539 		lnk->wr_tx_rdmas[i].wr_tx_rdma[1].wr.opcode = IB_WR_RDMA_WRITE;
540 		lnk->wr_tx_rdmas[i].wr_tx_rdma[0].wr.sg_list =
541 			lnk->wr_tx_rdma_sges[i].tx_rdma_sge[0].wr_tx_rdma_sge;
542 		lnk->wr_tx_rdmas[i].wr_tx_rdma[1].wr.sg_list =
543 			lnk->wr_tx_rdma_sges[i].tx_rdma_sge[1].wr_tx_rdma_sge;
544 	}
545 	for (i = 0; i < lnk->wr_rx_cnt; i++) {
546 		lnk->wr_rx_sges[i].addr =
547 			lnk->wr_rx_dma_addr + i * SMC_WR_BUF_SIZE;
548 		lnk->wr_rx_sges[i].length = SMC_WR_BUF_SIZE;
549 		lnk->wr_rx_sges[i].lkey = lnk->roce_pd->local_dma_lkey;
550 		lnk->wr_rx_ibs[i].next = NULL;
551 		lnk->wr_rx_ibs[i].sg_list = &lnk->wr_rx_sges[i];
552 		lnk->wr_rx_ibs[i].num_sge = 1;
553 	}
554 	lnk->wr_reg.wr.next = NULL;
555 	lnk->wr_reg.wr.num_sge = 0;
556 	lnk->wr_reg.wr.send_flags = IB_SEND_SIGNALED;
557 	lnk->wr_reg.wr.opcode = IB_WR_REG_MR;
558 	lnk->wr_reg.access = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE;
559 }
560 
561 void smc_wr_free_link(struct smc_link *lnk)
562 {
563 	struct ib_device *ibdev;
564 
565 	if (!lnk->smcibdev)
566 		return;
567 	ibdev = lnk->smcibdev->ibdev;
568 
569 	if (smc_wr_tx_wait_no_pending_sends(lnk))
570 		memset(lnk->wr_tx_mask, 0,
571 		       BITS_TO_LONGS(SMC_WR_BUF_CNT) *
572 						sizeof(*lnk->wr_tx_mask));
573 
574 	if (lnk->wr_rx_dma_addr) {
575 		ib_dma_unmap_single(ibdev, lnk->wr_rx_dma_addr,
576 				    SMC_WR_BUF_SIZE * lnk->wr_rx_cnt,
577 				    DMA_FROM_DEVICE);
578 		lnk->wr_rx_dma_addr = 0;
579 	}
580 	if (lnk->wr_tx_dma_addr) {
581 		ib_dma_unmap_single(ibdev, lnk->wr_tx_dma_addr,
582 				    SMC_WR_BUF_SIZE * lnk->wr_tx_cnt,
583 				    DMA_TO_DEVICE);
584 		lnk->wr_tx_dma_addr = 0;
585 	}
586 }
587 
588 void smc_wr_free_link_mem(struct smc_link *lnk)
589 {
590 	kfree(lnk->wr_tx_compl);
591 	lnk->wr_tx_compl = NULL;
592 	kfree(lnk->wr_tx_pends);
593 	lnk->wr_tx_pends = NULL;
594 	kfree(lnk->wr_tx_mask);
595 	lnk->wr_tx_mask = NULL;
596 	kfree(lnk->wr_tx_sges);
597 	lnk->wr_tx_sges = NULL;
598 	kfree(lnk->wr_tx_rdma_sges);
599 	lnk->wr_tx_rdma_sges = NULL;
600 	kfree(lnk->wr_rx_sges);
601 	lnk->wr_rx_sges = NULL;
602 	kfree(lnk->wr_tx_rdmas);
603 	lnk->wr_tx_rdmas = NULL;
604 	kfree(lnk->wr_rx_ibs);
605 	lnk->wr_rx_ibs = NULL;
606 	kfree(lnk->wr_tx_ibs);
607 	lnk->wr_tx_ibs = NULL;
608 	kfree(lnk->wr_tx_bufs);
609 	lnk->wr_tx_bufs = NULL;
610 	kfree(lnk->wr_rx_bufs);
611 	lnk->wr_rx_bufs = NULL;
612 }
613 
614 int smc_wr_alloc_link_mem(struct smc_link *link)
615 {
616 	/* allocate link related memory */
617 	link->wr_tx_bufs = kcalloc(SMC_WR_BUF_CNT, SMC_WR_BUF_SIZE, GFP_KERNEL);
618 	if (!link->wr_tx_bufs)
619 		goto no_mem;
620 	link->wr_rx_bufs = kcalloc(SMC_WR_BUF_CNT * 3, SMC_WR_BUF_SIZE,
621 				   GFP_KERNEL);
622 	if (!link->wr_rx_bufs)
623 		goto no_mem_wr_tx_bufs;
624 	link->wr_tx_ibs = kcalloc(SMC_WR_BUF_CNT, sizeof(link->wr_tx_ibs[0]),
625 				  GFP_KERNEL);
626 	if (!link->wr_tx_ibs)
627 		goto no_mem_wr_rx_bufs;
628 	link->wr_rx_ibs = kcalloc(SMC_WR_BUF_CNT * 3,
629 				  sizeof(link->wr_rx_ibs[0]),
630 				  GFP_KERNEL);
631 	if (!link->wr_rx_ibs)
632 		goto no_mem_wr_tx_ibs;
633 	link->wr_tx_rdmas = kcalloc(SMC_WR_BUF_CNT,
634 				    sizeof(link->wr_tx_rdmas[0]),
635 				    GFP_KERNEL);
636 	if (!link->wr_tx_rdmas)
637 		goto no_mem_wr_rx_ibs;
638 	link->wr_tx_rdma_sges = kcalloc(SMC_WR_BUF_CNT,
639 					sizeof(link->wr_tx_rdma_sges[0]),
640 					GFP_KERNEL);
641 	if (!link->wr_tx_rdma_sges)
642 		goto no_mem_wr_tx_rdmas;
643 	link->wr_tx_sges = kcalloc(SMC_WR_BUF_CNT, sizeof(link->wr_tx_sges[0]),
644 				   GFP_KERNEL);
645 	if (!link->wr_tx_sges)
646 		goto no_mem_wr_tx_rdma_sges;
647 	link->wr_rx_sges = kcalloc(SMC_WR_BUF_CNT * 3,
648 				   sizeof(link->wr_rx_sges[0]),
649 				   GFP_KERNEL);
650 	if (!link->wr_rx_sges)
651 		goto no_mem_wr_tx_sges;
652 	link->wr_tx_mask = kcalloc(BITS_TO_LONGS(SMC_WR_BUF_CNT),
653 				   sizeof(*link->wr_tx_mask),
654 				   GFP_KERNEL);
655 	if (!link->wr_tx_mask)
656 		goto no_mem_wr_rx_sges;
657 	link->wr_tx_pends = kcalloc(SMC_WR_BUF_CNT,
658 				    sizeof(link->wr_tx_pends[0]),
659 				    GFP_KERNEL);
660 	if (!link->wr_tx_pends)
661 		goto no_mem_wr_tx_mask;
662 	link->wr_tx_compl = kcalloc(SMC_WR_BUF_CNT,
663 				    sizeof(link->wr_tx_compl[0]),
664 				    GFP_KERNEL);
665 	if (!link->wr_tx_compl)
666 		goto no_mem_wr_tx_pends;
667 	return 0;
668 
669 no_mem_wr_tx_pends:
670 	kfree(link->wr_tx_pends);
671 no_mem_wr_tx_mask:
672 	kfree(link->wr_tx_mask);
673 no_mem_wr_rx_sges:
674 	kfree(link->wr_rx_sges);
675 no_mem_wr_tx_sges:
676 	kfree(link->wr_tx_sges);
677 no_mem_wr_tx_rdma_sges:
678 	kfree(link->wr_tx_rdma_sges);
679 no_mem_wr_tx_rdmas:
680 	kfree(link->wr_tx_rdmas);
681 no_mem_wr_rx_ibs:
682 	kfree(link->wr_rx_ibs);
683 no_mem_wr_tx_ibs:
684 	kfree(link->wr_tx_ibs);
685 no_mem_wr_rx_bufs:
686 	kfree(link->wr_rx_bufs);
687 no_mem_wr_tx_bufs:
688 	kfree(link->wr_tx_bufs);
689 no_mem:
690 	return -ENOMEM;
691 }
692 
693 void smc_wr_remove_dev(struct smc_ib_device *smcibdev)
694 {
695 	tasklet_kill(&smcibdev->recv_tasklet);
696 	tasklet_kill(&smcibdev->send_tasklet);
697 }
698 
699 void smc_wr_add_dev(struct smc_ib_device *smcibdev)
700 {
701 	tasklet_init(&smcibdev->recv_tasklet, smc_wr_rx_tasklet_fn,
702 		     (unsigned long)smcibdev);
703 	tasklet_init(&smcibdev->send_tasklet, smc_wr_tx_tasklet_fn,
704 		     (unsigned long)smcibdev);
705 }
706 
707 int smc_wr_create_link(struct smc_link *lnk)
708 {
709 	struct ib_device *ibdev = lnk->smcibdev->ibdev;
710 	int rc = 0;
711 
712 	smc_wr_tx_set_wr_id(&lnk->wr_tx_id, 0);
713 	lnk->wr_rx_id = 0;
714 	lnk->wr_rx_dma_addr = ib_dma_map_single(
715 		ibdev, lnk->wr_rx_bufs,	SMC_WR_BUF_SIZE * lnk->wr_rx_cnt,
716 		DMA_FROM_DEVICE);
717 	if (ib_dma_mapping_error(ibdev, lnk->wr_rx_dma_addr)) {
718 		lnk->wr_rx_dma_addr = 0;
719 		rc = -EIO;
720 		goto out;
721 	}
722 	lnk->wr_tx_dma_addr = ib_dma_map_single(
723 		ibdev, lnk->wr_tx_bufs,	SMC_WR_BUF_SIZE * lnk->wr_tx_cnt,
724 		DMA_TO_DEVICE);
725 	if (ib_dma_mapping_error(ibdev, lnk->wr_tx_dma_addr)) {
726 		rc = -EIO;
727 		goto dma_unmap;
728 	}
729 	smc_wr_init_sge(lnk);
730 	memset(lnk->wr_tx_mask, 0,
731 	       BITS_TO_LONGS(SMC_WR_BUF_CNT) * sizeof(*lnk->wr_tx_mask));
732 	init_waitqueue_head(&lnk->wr_tx_wait);
733 	init_waitqueue_head(&lnk->wr_reg_wait);
734 	return rc;
735 
736 dma_unmap:
737 	ib_dma_unmap_single(ibdev, lnk->wr_rx_dma_addr,
738 			    SMC_WR_BUF_SIZE * lnk->wr_rx_cnt,
739 			    DMA_FROM_DEVICE);
740 	lnk->wr_rx_dma_addr = 0;
741 out:
742 	return rc;
743 }
744