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