xref: /openbmc/linux/net/smc/smc_wr.c (revision f9834f18)
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 /* returns true if at least one tx work request is pending on the given link */
54 static inline bool smc_wr_is_tx_pend(struct smc_link *link)
55 {
56 	if (find_first_bit(link->wr_tx_mask, link->wr_tx_cnt) !=
57 							link->wr_tx_cnt) {
58 		return true;
59 	}
60 	return false;
61 }
62 
63 /* wait till all pending tx work requests on the given link are completed */
64 static inline int smc_wr_tx_wait_no_pending_sends(struct smc_link *link)
65 {
66 	if (wait_event_timeout(link->wr_tx_wait, !smc_wr_is_tx_pend(link),
67 			       SMC_WR_TX_WAIT_PENDING_TIME))
68 		return 0;
69 	else /* timeout */
70 		return -EPIPE;
71 }
72 
73 static inline int smc_wr_tx_find_pending_index(struct smc_link *link, u64 wr_id)
74 {
75 	u32 i;
76 
77 	for (i = 0; i < link->wr_tx_cnt; i++) {
78 		if (link->wr_tx_pends[i].wr_id == wr_id)
79 			return i;
80 	}
81 	return link->wr_tx_cnt;
82 }
83 
84 static inline void smc_wr_tx_process_cqe(struct ib_wc *wc)
85 {
86 	struct smc_wr_tx_pend pnd_snd;
87 	struct smc_link *link;
88 	u32 pnd_snd_idx;
89 	int i;
90 
91 	link = wc->qp->qp_context;
92 
93 	if (wc->opcode == IB_WC_REG_MR) {
94 		if (wc->status)
95 			link->wr_reg_state = FAILED;
96 		else
97 			link->wr_reg_state = CONFIRMED;
98 		smc_wr_wakeup_reg_wait(link);
99 		return;
100 	}
101 
102 	pnd_snd_idx = smc_wr_tx_find_pending_index(link, wc->wr_id);
103 	if (pnd_snd_idx == link->wr_tx_cnt)
104 		return;
105 	link->wr_tx_pends[pnd_snd_idx].wc_status = wc->status;
106 	memcpy(&pnd_snd, &link->wr_tx_pends[pnd_snd_idx], sizeof(pnd_snd));
107 	/* clear the full struct smc_wr_tx_pend including .priv */
108 	memset(&link->wr_tx_pends[pnd_snd_idx], 0,
109 	       sizeof(link->wr_tx_pends[pnd_snd_idx]));
110 	memset(&link->wr_tx_bufs[pnd_snd_idx], 0,
111 	       sizeof(link->wr_tx_bufs[pnd_snd_idx]));
112 	if (!test_and_clear_bit(pnd_snd_idx, link->wr_tx_mask))
113 		return;
114 	if (wc->status) {
115 		for_each_set_bit(i, link->wr_tx_mask, link->wr_tx_cnt) {
116 			/* clear full struct smc_wr_tx_pend including .priv */
117 			memset(&link->wr_tx_pends[i], 0,
118 			       sizeof(link->wr_tx_pends[i]));
119 			memset(&link->wr_tx_bufs[i], 0,
120 			       sizeof(link->wr_tx_bufs[i]));
121 			clear_bit(i, link->wr_tx_mask);
122 		}
123 		/* terminate connections of this link group abnormally */
124 		smc_lgr_terminate_sched(smc_get_lgr(link));
125 	}
126 	if (pnd_snd.handler)
127 		pnd_snd.handler(&pnd_snd.priv, link, wc->status);
128 	wake_up(&link->wr_tx_wait);
129 }
130 
131 static void smc_wr_tx_tasklet_fn(unsigned long data)
132 {
133 	struct smc_ib_device *dev = (struct smc_ib_device *)data;
134 	struct ib_wc wc[SMC_WR_MAX_POLL_CQE];
135 	int i = 0, rc;
136 	int polled = 0;
137 
138 again:
139 	polled++;
140 	do {
141 		memset(&wc, 0, sizeof(wc));
142 		rc = ib_poll_cq(dev->roce_cq_send, SMC_WR_MAX_POLL_CQE, wc);
143 		if (polled == 1) {
144 			ib_req_notify_cq(dev->roce_cq_send,
145 					 IB_CQ_NEXT_COMP |
146 					 IB_CQ_REPORT_MISSED_EVENTS);
147 		}
148 		if (!rc)
149 			break;
150 		for (i = 0; i < rc; i++)
151 			smc_wr_tx_process_cqe(&wc[i]);
152 	} while (rc > 0);
153 	if (polled == 1)
154 		goto again;
155 }
156 
157 void smc_wr_tx_cq_handler(struct ib_cq *ib_cq, void *cq_context)
158 {
159 	struct smc_ib_device *dev = (struct smc_ib_device *)cq_context;
160 
161 	tasklet_schedule(&dev->send_tasklet);
162 }
163 
164 /*---------------------------- request submission ---------------------------*/
165 
166 static inline int smc_wr_tx_get_free_slot_index(struct smc_link *link, u32 *idx)
167 {
168 	*idx = link->wr_tx_cnt;
169 	for_each_clear_bit(*idx, link->wr_tx_mask, link->wr_tx_cnt) {
170 		if (!test_and_set_bit(*idx, link->wr_tx_mask))
171 			return 0;
172 	}
173 	*idx = link->wr_tx_cnt;
174 	return -EBUSY;
175 }
176 
177 /**
178  * smc_wr_tx_get_free_slot() - returns buffer for message assembly,
179  *			and sets info for pending transmit tracking
180  * @link:		Pointer to smc_link used to later send the message.
181  * @handler:		Send completion handler function pointer.
182  * @wr_buf:		Out value returns pointer to message buffer.
183  * @wr_rdma_buf:	Out value returns pointer to rdma work request.
184  * @wr_pend_priv:	Out value returns pointer serving as handler context.
185  *
186  * Return: 0 on success, or -errno on error.
187  */
188 int smc_wr_tx_get_free_slot(struct smc_link *link,
189 			    smc_wr_tx_handler handler,
190 			    struct smc_wr_buf **wr_buf,
191 			    struct smc_rdma_wr **wr_rdma_buf,
192 			    struct smc_wr_tx_pend_priv **wr_pend_priv)
193 {
194 	struct smc_link_group *lgr = smc_get_lgr(link);
195 	struct smc_wr_tx_pend *wr_pend;
196 	u32 idx = link->wr_tx_cnt;
197 	struct ib_send_wr *wr_ib;
198 	u64 wr_id;
199 	int rc;
200 
201 	*wr_buf = NULL;
202 	*wr_pend_priv = NULL;
203 	if (in_softirq() || lgr->terminating) {
204 		rc = smc_wr_tx_get_free_slot_index(link, &idx);
205 		if (rc)
206 			return rc;
207 	} else {
208 		rc = wait_event_interruptible_timeout(
209 			link->wr_tx_wait,
210 			link->state == SMC_LNK_INACTIVE ||
211 			lgr->terminating ||
212 			(smc_wr_tx_get_free_slot_index(link, &idx) != -EBUSY),
213 			SMC_WR_TX_WAIT_FREE_SLOT_TIME);
214 		if (!rc) {
215 			/* timeout - terminate connections */
216 			smc_lgr_terminate_sched(lgr);
217 			return -EPIPE;
218 		}
219 		if (idx == link->wr_tx_cnt)
220 			return -EPIPE;
221 	}
222 	wr_id = smc_wr_tx_get_next_wr_id(link);
223 	wr_pend = &link->wr_tx_pends[idx];
224 	wr_pend->wr_id = wr_id;
225 	wr_pend->handler = handler;
226 	wr_pend->link = link;
227 	wr_pend->idx = idx;
228 	wr_ib = &link->wr_tx_ibs[idx];
229 	wr_ib->wr_id = wr_id;
230 	*wr_buf = &link->wr_tx_bufs[idx];
231 	if (wr_rdma_buf)
232 		*wr_rdma_buf = &link->wr_tx_rdmas[idx];
233 	*wr_pend_priv = &wr_pend->priv;
234 	return 0;
235 }
236 
237 int smc_wr_tx_put_slot(struct smc_link *link,
238 		       struct smc_wr_tx_pend_priv *wr_pend_priv)
239 {
240 	struct smc_wr_tx_pend *pend;
241 
242 	pend = container_of(wr_pend_priv, struct smc_wr_tx_pend, priv);
243 	if (pend->idx < link->wr_tx_cnt) {
244 		u32 idx = pend->idx;
245 
246 		/* clear the full struct smc_wr_tx_pend including .priv */
247 		memset(&link->wr_tx_pends[idx], 0,
248 		       sizeof(link->wr_tx_pends[idx]));
249 		memset(&link->wr_tx_bufs[idx], 0,
250 		       sizeof(link->wr_tx_bufs[idx]));
251 		test_and_clear_bit(idx, link->wr_tx_mask);
252 		wake_up(&link->wr_tx_wait);
253 		return 1;
254 	}
255 
256 	return 0;
257 }
258 
259 /* Send prepared WR slot via ib_post_send.
260  * @priv: pointer to smc_wr_tx_pend_priv identifying prepared message buffer
261  */
262 int smc_wr_tx_send(struct smc_link *link, struct smc_wr_tx_pend_priv *priv)
263 {
264 	struct smc_wr_tx_pend *pend;
265 	int rc;
266 
267 	ib_req_notify_cq(link->smcibdev->roce_cq_send,
268 			 IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS);
269 	pend = container_of(priv, struct smc_wr_tx_pend, priv);
270 	rc = ib_post_send(link->roce_qp, &link->wr_tx_ibs[pend->idx], NULL);
271 	if (rc) {
272 		smc_wr_tx_put_slot(link, priv);
273 		smc_lgr_terminate_sched(smc_get_lgr(link));
274 	}
275 	return rc;
276 }
277 
278 /* Register a memory region and wait for result. */
279 int smc_wr_reg_send(struct smc_link *link, struct ib_mr *mr)
280 {
281 	int rc;
282 
283 	ib_req_notify_cq(link->smcibdev->roce_cq_send,
284 			 IB_CQ_NEXT_COMP | IB_CQ_REPORT_MISSED_EVENTS);
285 	link->wr_reg_state = POSTED;
286 	link->wr_reg.wr.wr_id = (u64)(uintptr_t)mr;
287 	link->wr_reg.mr = mr;
288 	link->wr_reg.key = mr->rkey;
289 	rc = ib_post_send(link->roce_qp, &link->wr_reg.wr, NULL);
290 	if (rc)
291 		return rc;
292 
293 	rc = wait_event_interruptible_timeout(link->wr_reg_wait,
294 					      (link->wr_reg_state != POSTED),
295 					      SMC_WR_REG_MR_WAIT_TIME);
296 	if (!rc) {
297 		/* timeout - terminate connections */
298 		smc_lgr_terminate_sched(smc_get_lgr(link));
299 		return -EPIPE;
300 	}
301 	if (rc == -ERESTARTSYS)
302 		return -EINTR;
303 	switch (link->wr_reg_state) {
304 	case CONFIRMED:
305 		rc = 0;
306 		break;
307 	case FAILED:
308 		rc = -EIO;
309 		break;
310 	case POSTED:
311 		rc = -EPIPE;
312 		break;
313 	}
314 	return rc;
315 }
316 
317 void smc_wr_tx_dismiss_slots(struct smc_link *link, u8 wr_tx_hdr_type,
318 			     smc_wr_tx_filter filter,
319 			     smc_wr_tx_dismisser dismisser,
320 			     unsigned long data)
321 {
322 	struct smc_wr_tx_pend_priv *tx_pend;
323 	struct smc_wr_rx_hdr *wr_tx;
324 	int i;
325 
326 	for_each_set_bit(i, link->wr_tx_mask, link->wr_tx_cnt) {
327 		wr_tx = (struct smc_wr_rx_hdr *)&link->wr_tx_bufs[i];
328 		if (wr_tx->type != wr_tx_hdr_type)
329 			continue;
330 		tx_pend = &link->wr_tx_pends[i].priv;
331 		if (filter(tx_pend, data))
332 			dismisser(tx_pend);
333 	}
334 }
335 
336 /****************************** receive queue ********************************/
337 
338 int smc_wr_rx_register_handler(struct smc_wr_rx_handler *handler)
339 {
340 	struct smc_wr_rx_handler *h_iter;
341 	int rc = 0;
342 
343 	spin_lock(&smc_wr_rx_hash_lock);
344 	hash_for_each_possible(smc_wr_rx_hash, h_iter, list, handler->type) {
345 		if (h_iter->type == handler->type) {
346 			rc = -EEXIST;
347 			goto out_unlock;
348 		}
349 	}
350 	hash_add(smc_wr_rx_hash, &handler->list, handler->type);
351 out_unlock:
352 	spin_unlock(&smc_wr_rx_hash_lock);
353 	return rc;
354 }
355 
356 /* Demultiplex a received work request based on the message type to its handler.
357  * Relies on smc_wr_rx_hash having been completely filled before any IB WRs,
358  * and not being modified any more afterwards so we don't need to lock it.
359  */
360 static inline void smc_wr_rx_demultiplex(struct ib_wc *wc)
361 {
362 	struct smc_link *link = (struct smc_link *)wc->qp->qp_context;
363 	struct smc_wr_rx_handler *handler;
364 	struct smc_wr_rx_hdr *wr_rx;
365 	u64 temp_wr_id;
366 	u32 index;
367 
368 	if (wc->byte_len < sizeof(*wr_rx))
369 		return; /* short message */
370 	temp_wr_id = wc->wr_id;
371 	index = do_div(temp_wr_id, link->wr_rx_cnt);
372 	wr_rx = (struct smc_wr_rx_hdr *)&link->wr_rx_bufs[index];
373 	hash_for_each_possible(smc_wr_rx_hash, handler, list, wr_rx->type) {
374 		if (handler->type == wr_rx->type)
375 			handler->handler(wc, wr_rx);
376 	}
377 }
378 
379 static inline void smc_wr_rx_process_cqes(struct ib_wc wc[], int num)
380 {
381 	struct smc_link *link;
382 	int i;
383 
384 	for (i = 0; i < num; i++) {
385 		link = wc[i].qp->qp_context;
386 		if (wc[i].status == IB_WC_SUCCESS) {
387 			link->wr_rx_tstamp = jiffies;
388 			smc_wr_rx_demultiplex(&wc[i]);
389 			smc_wr_rx_post(link); /* refill WR RX */
390 		} else {
391 			/* handle status errors */
392 			switch (wc[i].status) {
393 			case IB_WC_RETRY_EXC_ERR:
394 			case IB_WC_RNR_RETRY_EXC_ERR:
395 			case IB_WC_WR_FLUSH_ERR:
396 				/* terminate connections of this link group
397 				 * abnormally
398 				 */
399 				smc_lgr_terminate_sched(smc_get_lgr(link));
400 				break;
401 			default:
402 				smc_wr_rx_post(link); /* refill WR RX */
403 				break;
404 			}
405 		}
406 	}
407 }
408 
409 static void smc_wr_rx_tasklet_fn(unsigned long data)
410 {
411 	struct smc_ib_device *dev = (struct smc_ib_device *)data;
412 	struct ib_wc wc[SMC_WR_MAX_POLL_CQE];
413 	int polled = 0;
414 	int rc;
415 
416 again:
417 	polled++;
418 	do {
419 		memset(&wc, 0, sizeof(wc));
420 		rc = ib_poll_cq(dev->roce_cq_recv, SMC_WR_MAX_POLL_CQE, wc);
421 		if (polled == 1) {
422 			ib_req_notify_cq(dev->roce_cq_recv,
423 					 IB_CQ_SOLICITED_MASK
424 					 | IB_CQ_REPORT_MISSED_EVENTS);
425 		}
426 		if (!rc)
427 			break;
428 		smc_wr_rx_process_cqes(&wc[0], rc);
429 	} while (rc > 0);
430 	if (polled == 1)
431 		goto again;
432 }
433 
434 void smc_wr_rx_cq_handler(struct ib_cq *ib_cq, void *cq_context)
435 {
436 	struct smc_ib_device *dev = (struct smc_ib_device *)cq_context;
437 
438 	tasklet_schedule(&dev->recv_tasklet);
439 }
440 
441 int smc_wr_rx_post_init(struct smc_link *link)
442 {
443 	u32 i;
444 	int rc = 0;
445 
446 	for (i = 0; i < link->wr_rx_cnt; i++)
447 		rc = smc_wr_rx_post(link);
448 	return rc;
449 }
450 
451 /***************************** init, exit, misc ******************************/
452 
453 void smc_wr_remember_qp_attr(struct smc_link *lnk)
454 {
455 	struct ib_qp_attr *attr = &lnk->qp_attr;
456 	struct ib_qp_init_attr init_attr;
457 
458 	memset(attr, 0, sizeof(*attr));
459 	memset(&init_attr, 0, sizeof(init_attr));
460 	ib_query_qp(lnk->roce_qp, attr,
461 		    IB_QP_STATE |
462 		    IB_QP_CUR_STATE |
463 		    IB_QP_PKEY_INDEX |
464 		    IB_QP_PORT |
465 		    IB_QP_QKEY |
466 		    IB_QP_AV |
467 		    IB_QP_PATH_MTU |
468 		    IB_QP_TIMEOUT |
469 		    IB_QP_RETRY_CNT |
470 		    IB_QP_RNR_RETRY |
471 		    IB_QP_RQ_PSN |
472 		    IB_QP_ALT_PATH |
473 		    IB_QP_MIN_RNR_TIMER |
474 		    IB_QP_SQ_PSN |
475 		    IB_QP_PATH_MIG_STATE |
476 		    IB_QP_CAP |
477 		    IB_QP_DEST_QPN,
478 		    &init_attr);
479 
480 	lnk->wr_tx_cnt = min_t(size_t, SMC_WR_BUF_CNT,
481 			       lnk->qp_attr.cap.max_send_wr);
482 	lnk->wr_rx_cnt = min_t(size_t, SMC_WR_BUF_CNT * 3,
483 			       lnk->qp_attr.cap.max_recv_wr);
484 }
485 
486 static void smc_wr_init_sge(struct smc_link *lnk)
487 {
488 	u32 i;
489 
490 	for (i = 0; i < lnk->wr_tx_cnt; i++) {
491 		lnk->wr_tx_sges[i].addr =
492 			lnk->wr_tx_dma_addr + i * SMC_WR_BUF_SIZE;
493 		lnk->wr_tx_sges[i].length = SMC_WR_TX_SIZE;
494 		lnk->wr_tx_sges[i].lkey = lnk->roce_pd->local_dma_lkey;
495 		lnk->wr_tx_rdma_sges[i].tx_rdma_sge[0].wr_tx_rdma_sge[0].lkey =
496 			lnk->roce_pd->local_dma_lkey;
497 		lnk->wr_tx_rdma_sges[i].tx_rdma_sge[0].wr_tx_rdma_sge[1].lkey =
498 			lnk->roce_pd->local_dma_lkey;
499 		lnk->wr_tx_rdma_sges[i].tx_rdma_sge[1].wr_tx_rdma_sge[0].lkey =
500 			lnk->roce_pd->local_dma_lkey;
501 		lnk->wr_tx_rdma_sges[i].tx_rdma_sge[1].wr_tx_rdma_sge[1].lkey =
502 			lnk->roce_pd->local_dma_lkey;
503 		lnk->wr_tx_ibs[i].next = NULL;
504 		lnk->wr_tx_ibs[i].sg_list = &lnk->wr_tx_sges[i];
505 		lnk->wr_tx_ibs[i].num_sge = 1;
506 		lnk->wr_tx_ibs[i].opcode = IB_WR_SEND;
507 		lnk->wr_tx_ibs[i].send_flags =
508 			IB_SEND_SIGNALED | IB_SEND_SOLICITED;
509 		lnk->wr_tx_rdmas[i].wr_tx_rdma[0].wr.opcode = IB_WR_RDMA_WRITE;
510 		lnk->wr_tx_rdmas[i].wr_tx_rdma[1].wr.opcode = IB_WR_RDMA_WRITE;
511 		lnk->wr_tx_rdmas[i].wr_tx_rdma[0].wr.sg_list =
512 			lnk->wr_tx_rdma_sges[i].tx_rdma_sge[0].wr_tx_rdma_sge;
513 		lnk->wr_tx_rdmas[i].wr_tx_rdma[1].wr.sg_list =
514 			lnk->wr_tx_rdma_sges[i].tx_rdma_sge[1].wr_tx_rdma_sge;
515 	}
516 	for (i = 0; i < lnk->wr_rx_cnt; i++) {
517 		lnk->wr_rx_sges[i].addr =
518 			lnk->wr_rx_dma_addr + i * SMC_WR_BUF_SIZE;
519 		lnk->wr_rx_sges[i].length = SMC_WR_BUF_SIZE;
520 		lnk->wr_rx_sges[i].lkey = lnk->roce_pd->local_dma_lkey;
521 		lnk->wr_rx_ibs[i].next = NULL;
522 		lnk->wr_rx_ibs[i].sg_list = &lnk->wr_rx_sges[i];
523 		lnk->wr_rx_ibs[i].num_sge = 1;
524 	}
525 	lnk->wr_reg.wr.next = NULL;
526 	lnk->wr_reg.wr.num_sge = 0;
527 	lnk->wr_reg.wr.send_flags = IB_SEND_SIGNALED;
528 	lnk->wr_reg.wr.opcode = IB_WR_REG_MR;
529 	lnk->wr_reg.access = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_WRITE;
530 }
531 
532 void smc_wr_free_link(struct smc_link *lnk)
533 {
534 	struct ib_device *ibdev;
535 
536 	if (smc_wr_tx_wait_no_pending_sends(lnk))
537 		memset(lnk->wr_tx_mask, 0,
538 		       BITS_TO_LONGS(SMC_WR_BUF_CNT) *
539 						sizeof(*lnk->wr_tx_mask));
540 
541 	if (!lnk->smcibdev)
542 		return;
543 	ibdev = lnk->smcibdev->ibdev;
544 
545 	if (lnk->wr_rx_dma_addr) {
546 		ib_dma_unmap_single(ibdev, lnk->wr_rx_dma_addr,
547 				    SMC_WR_BUF_SIZE * lnk->wr_rx_cnt,
548 				    DMA_FROM_DEVICE);
549 		lnk->wr_rx_dma_addr = 0;
550 	}
551 	if (lnk->wr_tx_dma_addr) {
552 		ib_dma_unmap_single(ibdev, lnk->wr_tx_dma_addr,
553 				    SMC_WR_BUF_SIZE * lnk->wr_tx_cnt,
554 				    DMA_TO_DEVICE);
555 		lnk->wr_tx_dma_addr = 0;
556 	}
557 }
558 
559 void smc_wr_free_link_mem(struct smc_link *lnk)
560 {
561 	kfree(lnk->wr_tx_pends);
562 	lnk->wr_tx_pends = NULL;
563 	kfree(lnk->wr_tx_mask);
564 	lnk->wr_tx_mask = NULL;
565 	kfree(lnk->wr_tx_sges);
566 	lnk->wr_tx_sges = NULL;
567 	kfree(lnk->wr_tx_rdma_sges);
568 	lnk->wr_tx_rdma_sges = NULL;
569 	kfree(lnk->wr_rx_sges);
570 	lnk->wr_rx_sges = NULL;
571 	kfree(lnk->wr_tx_rdmas);
572 	lnk->wr_tx_rdmas = NULL;
573 	kfree(lnk->wr_rx_ibs);
574 	lnk->wr_rx_ibs = NULL;
575 	kfree(lnk->wr_tx_ibs);
576 	lnk->wr_tx_ibs = NULL;
577 	kfree(lnk->wr_tx_bufs);
578 	lnk->wr_tx_bufs = NULL;
579 	kfree(lnk->wr_rx_bufs);
580 	lnk->wr_rx_bufs = NULL;
581 }
582 
583 int smc_wr_alloc_link_mem(struct smc_link *link)
584 {
585 	/* allocate link related memory */
586 	link->wr_tx_bufs = kcalloc(SMC_WR_BUF_CNT, SMC_WR_BUF_SIZE, GFP_KERNEL);
587 	if (!link->wr_tx_bufs)
588 		goto no_mem;
589 	link->wr_rx_bufs = kcalloc(SMC_WR_BUF_CNT * 3, SMC_WR_BUF_SIZE,
590 				   GFP_KERNEL);
591 	if (!link->wr_rx_bufs)
592 		goto no_mem_wr_tx_bufs;
593 	link->wr_tx_ibs = kcalloc(SMC_WR_BUF_CNT, sizeof(link->wr_tx_ibs[0]),
594 				  GFP_KERNEL);
595 	if (!link->wr_tx_ibs)
596 		goto no_mem_wr_rx_bufs;
597 	link->wr_rx_ibs = kcalloc(SMC_WR_BUF_CNT * 3,
598 				  sizeof(link->wr_rx_ibs[0]),
599 				  GFP_KERNEL);
600 	if (!link->wr_rx_ibs)
601 		goto no_mem_wr_tx_ibs;
602 	link->wr_tx_rdmas = kcalloc(SMC_WR_BUF_CNT,
603 				    sizeof(link->wr_tx_rdmas[0]),
604 				    GFP_KERNEL);
605 	if (!link->wr_tx_rdmas)
606 		goto no_mem_wr_rx_ibs;
607 	link->wr_tx_rdma_sges = kcalloc(SMC_WR_BUF_CNT,
608 					sizeof(link->wr_tx_rdma_sges[0]),
609 					GFP_KERNEL);
610 	if (!link->wr_tx_rdma_sges)
611 		goto no_mem_wr_tx_rdmas;
612 	link->wr_tx_sges = kcalloc(SMC_WR_BUF_CNT, sizeof(link->wr_tx_sges[0]),
613 				   GFP_KERNEL);
614 	if (!link->wr_tx_sges)
615 		goto no_mem_wr_tx_rdma_sges;
616 	link->wr_rx_sges = kcalloc(SMC_WR_BUF_CNT * 3,
617 				   sizeof(link->wr_rx_sges[0]),
618 				   GFP_KERNEL);
619 	if (!link->wr_rx_sges)
620 		goto no_mem_wr_tx_sges;
621 	link->wr_tx_mask = kcalloc(BITS_TO_LONGS(SMC_WR_BUF_CNT),
622 				   sizeof(*link->wr_tx_mask),
623 				   GFP_KERNEL);
624 	if (!link->wr_tx_mask)
625 		goto no_mem_wr_rx_sges;
626 	link->wr_tx_pends = kcalloc(SMC_WR_BUF_CNT,
627 				    sizeof(link->wr_tx_pends[0]),
628 				    GFP_KERNEL);
629 	if (!link->wr_tx_pends)
630 		goto no_mem_wr_tx_mask;
631 	return 0;
632 
633 no_mem_wr_tx_mask:
634 	kfree(link->wr_tx_mask);
635 no_mem_wr_rx_sges:
636 	kfree(link->wr_rx_sges);
637 no_mem_wr_tx_sges:
638 	kfree(link->wr_tx_sges);
639 no_mem_wr_tx_rdma_sges:
640 	kfree(link->wr_tx_rdma_sges);
641 no_mem_wr_tx_rdmas:
642 	kfree(link->wr_tx_rdmas);
643 no_mem_wr_rx_ibs:
644 	kfree(link->wr_rx_ibs);
645 no_mem_wr_tx_ibs:
646 	kfree(link->wr_tx_ibs);
647 no_mem_wr_rx_bufs:
648 	kfree(link->wr_rx_bufs);
649 no_mem_wr_tx_bufs:
650 	kfree(link->wr_tx_bufs);
651 no_mem:
652 	return -ENOMEM;
653 }
654 
655 void smc_wr_remove_dev(struct smc_ib_device *smcibdev)
656 {
657 	tasklet_kill(&smcibdev->recv_tasklet);
658 	tasklet_kill(&smcibdev->send_tasklet);
659 }
660 
661 void smc_wr_add_dev(struct smc_ib_device *smcibdev)
662 {
663 	tasklet_init(&smcibdev->recv_tasklet, smc_wr_rx_tasklet_fn,
664 		     (unsigned long)smcibdev);
665 	tasklet_init(&smcibdev->send_tasklet, smc_wr_tx_tasklet_fn,
666 		     (unsigned long)smcibdev);
667 }
668 
669 int smc_wr_create_link(struct smc_link *lnk)
670 {
671 	struct ib_device *ibdev = lnk->smcibdev->ibdev;
672 	int rc = 0;
673 
674 	smc_wr_tx_set_wr_id(&lnk->wr_tx_id, 0);
675 	lnk->wr_rx_id = 0;
676 	lnk->wr_rx_dma_addr = ib_dma_map_single(
677 		ibdev, lnk->wr_rx_bufs,	SMC_WR_BUF_SIZE * lnk->wr_rx_cnt,
678 		DMA_FROM_DEVICE);
679 	if (ib_dma_mapping_error(ibdev, lnk->wr_rx_dma_addr)) {
680 		lnk->wr_rx_dma_addr = 0;
681 		rc = -EIO;
682 		goto out;
683 	}
684 	lnk->wr_tx_dma_addr = ib_dma_map_single(
685 		ibdev, lnk->wr_tx_bufs,	SMC_WR_BUF_SIZE * lnk->wr_tx_cnt,
686 		DMA_TO_DEVICE);
687 	if (ib_dma_mapping_error(ibdev, lnk->wr_tx_dma_addr)) {
688 		rc = -EIO;
689 		goto dma_unmap;
690 	}
691 	smc_wr_init_sge(lnk);
692 	memset(lnk->wr_tx_mask, 0,
693 	       BITS_TO_LONGS(SMC_WR_BUF_CNT) * sizeof(*lnk->wr_tx_mask));
694 	init_waitqueue_head(&lnk->wr_tx_wait);
695 	init_waitqueue_head(&lnk->wr_reg_wait);
696 	return rc;
697 
698 dma_unmap:
699 	ib_dma_unmap_single(ibdev, lnk->wr_rx_dma_addr,
700 			    SMC_WR_BUF_SIZE * lnk->wr_rx_cnt,
701 			    DMA_FROM_DEVICE);
702 	lnk->wr_rx_dma_addr = 0;
703 out:
704 	return rc;
705 }
706