1 /*
2  * Copyright (c) 2004, 2005, 2006 Voltaire, Inc. All rights reserved.
3  * Copyright (c) 2005, 2006 Cisco Systems.  All rights reserved.
4  * Copyright (c) 2013 Mellanox Technologies. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *	- Redistributions of source code must retain the above
17  *	  copyright notice, this list of conditions and the following
18  *	  disclaimer.
19  *
20  *	- Redistributions in binary form must reproduce the above
21  *	  copyright notice, this list of conditions and the following
22  *	  disclaimer in the documentation and/or other materials
23  *	  provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  */
34 #include <linux/kernel.h>
35 #include <linux/module.h>
36 #include <linux/slab.h>
37 #include <linux/delay.h>
38 
39 #include "iscsi_iser.h"
40 
41 #define ISCSI_ISER_MAX_CONN	8
42 #define ISER_MAX_RX_CQ_LEN	(ISER_QP_MAX_RECV_DTOS * ISCSI_ISER_MAX_CONN)
43 #define ISER_MAX_TX_CQ_LEN	(ISER_QP_MAX_REQ_DTOS  * ISCSI_ISER_MAX_CONN)
44 
45 static void iser_cq_tasklet_fn(unsigned long data);
46 static void iser_cq_callback(struct ib_cq *cq, void *cq_context);
47 
48 static void iser_cq_event_callback(struct ib_event *cause, void *context)
49 {
50 	iser_err("got cq event %d \n", cause->event);
51 }
52 
53 static void iser_qp_event_callback(struct ib_event *cause, void *context)
54 {
55 	iser_err("got qp event %d\n",cause->event);
56 }
57 
58 static void iser_event_handler(struct ib_event_handler *handler,
59 				struct ib_event *event)
60 {
61 	iser_err("async event %d on device %s port %d\n", event->event,
62 		event->device->name, event->element.port_num);
63 }
64 
65 /**
66  * iser_create_device_ib_res - creates Protection Domain (PD), Completion
67  * Queue (CQ), DMA Memory Region (DMA MR) with the device associated with
68  * the adapator.
69  *
70  * returns 0 on success, -1 on failure
71  */
72 static int iser_create_device_ib_res(struct iser_device *device)
73 {
74 	int i, j;
75 	struct iser_cq_desc *cq_desc;
76 
77 	device->cqs_used = min(ISER_MAX_CQ, device->ib_device->num_comp_vectors);
78 	iser_info("using %d CQs, device %s supports %d vectors\n",
79 		  device->cqs_used, device->ib_device->name,
80 		  device->ib_device->num_comp_vectors);
81 
82 	device->cq_desc = kmalloc(sizeof(struct iser_cq_desc) * device->cqs_used,
83 				  GFP_KERNEL);
84 	if (device->cq_desc == NULL)
85 		goto cq_desc_err;
86 	cq_desc = device->cq_desc;
87 
88 	device->pd = ib_alloc_pd(device->ib_device);
89 	if (IS_ERR(device->pd))
90 		goto pd_err;
91 
92 	for (i = 0; i < device->cqs_used; i++) {
93 		cq_desc[i].device   = device;
94 		cq_desc[i].cq_index = i;
95 
96 		device->rx_cq[i] = ib_create_cq(device->ib_device,
97 					  iser_cq_callback,
98 					  iser_cq_event_callback,
99 					  (void *)&cq_desc[i],
100 					  ISER_MAX_RX_CQ_LEN, i);
101 		if (IS_ERR(device->rx_cq[i]))
102 			goto cq_err;
103 
104 		device->tx_cq[i] = ib_create_cq(device->ib_device,
105 					  NULL, iser_cq_event_callback,
106 					  (void *)&cq_desc[i],
107 					  ISER_MAX_TX_CQ_LEN, i);
108 
109 		if (IS_ERR(device->tx_cq[i]))
110 			goto cq_err;
111 
112 		if (ib_req_notify_cq(device->rx_cq[i], IB_CQ_NEXT_COMP))
113 			goto cq_err;
114 
115 		tasklet_init(&device->cq_tasklet[i],
116 			     iser_cq_tasklet_fn,
117 			(unsigned long)&cq_desc[i]);
118 	}
119 
120 	device->mr = ib_get_dma_mr(device->pd, IB_ACCESS_LOCAL_WRITE |
121 				   IB_ACCESS_REMOTE_WRITE |
122 				   IB_ACCESS_REMOTE_READ);
123 	if (IS_ERR(device->mr))
124 		goto dma_mr_err;
125 
126 	INIT_IB_EVENT_HANDLER(&device->event_handler, device->ib_device,
127 				iser_event_handler);
128 	if (ib_register_event_handler(&device->event_handler))
129 		goto handler_err;
130 
131 	return 0;
132 
133 handler_err:
134 	ib_dereg_mr(device->mr);
135 dma_mr_err:
136 	for (j = 0; j < device->cqs_used; j++)
137 		tasklet_kill(&device->cq_tasklet[j]);
138 cq_err:
139 	for (j = 0; j < i; j++) {
140 		if (device->tx_cq[j])
141 			ib_destroy_cq(device->tx_cq[j]);
142 		if (device->rx_cq[j])
143 			ib_destroy_cq(device->rx_cq[j]);
144 	}
145 	ib_dealloc_pd(device->pd);
146 pd_err:
147 	kfree(device->cq_desc);
148 cq_desc_err:
149 	iser_err("failed to allocate an IB resource\n");
150 	return -1;
151 }
152 
153 /**
154  * iser_free_device_ib_res - destroy/dealloc/dereg the DMA MR,
155  * CQ and PD created with the device associated with the adapator.
156  */
157 static void iser_free_device_ib_res(struct iser_device *device)
158 {
159 	int i;
160 	BUG_ON(device->mr == NULL);
161 
162 	for (i = 0; i < device->cqs_used; i++) {
163 		tasklet_kill(&device->cq_tasklet[i]);
164 		(void)ib_destroy_cq(device->tx_cq[i]);
165 		(void)ib_destroy_cq(device->rx_cq[i]);
166 		device->tx_cq[i] = NULL;
167 		device->rx_cq[i] = NULL;
168 	}
169 
170 	(void)ib_unregister_event_handler(&device->event_handler);
171 	(void)ib_dereg_mr(device->mr);
172 	(void)ib_dealloc_pd(device->pd);
173 
174 	kfree(device->cq_desc);
175 
176 	device->mr = NULL;
177 	device->pd = NULL;
178 }
179 
180 /**
181  * iser_create_ib_conn_res - Creates FMR pool and Queue-Pair (QP)
182  *
183  * returns 0 on success, -1 on failure
184  */
185 static int iser_create_ib_conn_res(struct iser_conn *ib_conn)
186 {
187 	struct iser_device	*device;
188 	struct ib_qp_init_attr	init_attr;
189 	int			req_err, resp_err, ret = -ENOMEM;
190 	struct ib_fmr_pool_param params;
191 	int index, min_index = 0;
192 
193 	BUG_ON(ib_conn->device == NULL);
194 
195 	device = ib_conn->device;
196 
197 	ib_conn->login_buf = kmalloc(ISCSI_DEF_MAX_RECV_SEG_LEN +
198 					ISER_RX_LOGIN_SIZE, GFP_KERNEL);
199 	if (!ib_conn->login_buf)
200 		goto out_err;
201 
202 	ib_conn->login_req_buf  = ib_conn->login_buf;
203 	ib_conn->login_resp_buf = ib_conn->login_buf + ISCSI_DEF_MAX_RECV_SEG_LEN;
204 
205 	ib_conn->login_req_dma = ib_dma_map_single(ib_conn->device->ib_device,
206 				(void *)ib_conn->login_req_buf,
207 				ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_TO_DEVICE);
208 
209 	ib_conn->login_resp_dma = ib_dma_map_single(ib_conn->device->ib_device,
210 				(void *)ib_conn->login_resp_buf,
211 				ISER_RX_LOGIN_SIZE, DMA_FROM_DEVICE);
212 
213 	req_err  = ib_dma_mapping_error(device->ib_device, ib_conn->login_req_dma);
214 	resp_err = ib_dma_mapping_error(device->ib_device, ib_conn->login_resp_dma);
215 
216 	if (req_err || resp_err) {
217 		if (req_err)
218 			ib_conn->login_req_dma = 0;
219 		if (resp_err)
220 			ib_conn->login_resp_dma = 0;
221 		goto out_err;
222 	}
223 
224 	ib_conn->page_vec = kmalloc(sizeof(struct iser_page_vec) +
225 				    (sizeof(u64) * (ISCSI_ISER_SG_TABLESIZE +1)),
226 				    GFP_KERNEL);
227 	if (!ib_conn->page_vec)
228 		goto out_err;
229 
230 	ib_conn->page_vec->pages = (u64 *) (ib_conn->page_vec + 1);
231 
232 	params.page_shift        = SHIFT_4K;
233 	/* when the first/last SG element are not start/end *
234 	 * page aligned, the map whould be of N+1 pages     */
235 	params.max_pages_per_fmr = ISCSI_ISER_SG_TABLESIZE + 1;
236 	/* make the pool size twice the max number of SCSI commands *
237 	 * the ML is expected to queue, watermark for unmap at 50%  */
238 	params.pool_size	 = ISCSI_DEF_XMIT_CMDS_MAX * 2;
239 	params.dirty_watermark	 = ISCSI_DEF_XMIT_CMDS_MAX;
240 	params.cache		 = 0;
241 	params.flush_function	 = NULL;
242 	params.access		 = (IB_ACCESS_LOCAL_WRITE  |
243 				    IB_ACCESS_REMOTE_WRITE |
244 				    IB_ACCESS_REMOTE_READ);
245 
246 	ib_conn->fmr_pool = ib_create_fmr_pool(device->pd, &params);
247 	ret = PTR_ERR(ib_conn->fmr_pool);
248 	if (IS_ERR(ib_conn->fmr_pool) && ret != -ENOSYS) {
249 		ib_conn->fmr_pool = NULL;
250 		goto out_err;
251 	} else if (ret == -ENOSYS) {
252 		ib_conn->fmr_pool = NULL;
253 		iser_warn("FMRs are not supported, using unaligned mode\n");
254 		ret = 0;
255 	}
256 
257 	memset(&init_attr, 0, sizeof init_attr);
258 
259 	mutex_lock(&ig.connlist_mutex);
260 	/* select the CQ with the minimal number of usages */
261 	for (index = 0; index < device->cqs_used; index++)
262 		if (device->cq_active_qps[index] <
263 		    device->cq_active_qps[min_index])
264 			min_index = index;
265 	device->cq_active_qps[min_index]++;
266 	mutex_unlock(&ig.connlist_mutex);
267 	iser_info("cq index %d used for ib_conn %p\n", min_index, ib_conn);
268 
269 	init_attr.event_handler = iser_qp_event_callback;
270 	init_attr.qp_context	= (void *)ib_conn;
271 	init_attr.send_cq	= device->tx_cq[min_index];
272 	init_attr.recv_cq	= device->rx_cq[min_index];
273 	init_attr.cap.max_send_wr  = ISER_QP_MAX_REQ_DTOS;
274 	init_attr.cap.max_recv_wr  = ISER_QP_MAX_RECV_DTOS;
275 	init_attr.cap.max_send_sge = 2;
276 	init_attr.cap.max_recv_sge = 1;
277 	init_attr.sq_sig_type	= IB_SIGNAL_REQ_WR;
278 	init_attr.qp_type	= IB_QPT_RC;
279 
280 	ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
281 	if (ret)
282 		goto out_err;
283 
284 	ib_conn->qp = ib_conn->cma_id->qp;
285 	iser_info("setting conn %p cma_id %p: fmr_pool %p qp %p\n",
286 		  ib_conn, ib_conn->cma_id,
287 		  ib_conn->fmr_pool, ib_conn->cma_id->qp);
288 	return ret;
289 
290 out_err:
291 	iser_err("unable to alloc mem or create resource, err %d\n", ret);
292 	return ret;
293 }
294 
295 /**
296  * releases the FMR pool and QP objects, returns 0 on success,
297  * -1 on failure
298  */
299 static int iser_free_ib_conn_res(struct iser_conn *ib_conn)
300 {
301 	int cq_index;
302 	BUG_ON(ib_conn == NULL);
303 
304 	iser_info("freeing conn %p cma_id %p fmr pool %p qp %p\n",
305 		  ib_conn, ib_conn->cma_id,
306 		  ib_conn->fmr_pool, ib_conn->qp);
307 
308 	/* qp is created only once both addr & route are resolved */
309 	if (ib_conn->fmr_pool != NULL)
310 		ib_destroy_fmr_pool(ib_conn->fmr_pool);
311 
312 	if (ib_conn->qp != NULL) {
313 		cq_index = ((struct iser_cq_desc *)ib_conn->qp->recv_cq->cq_context)->cq_index;
314 		ib_conn->device->cq_active_qps[cq_index]--;
315 
316 		rdma_destroy_qp(ib_conn->cma_id);
317 	}
318 
319 	ib_conn->fmr_pool = NULL;
320 	ib_conn->qp	  = NULL;
321 	kfree(ib_conn->page_vec);
322 
323 	if (ib_conn->login_buf) {
324 		if (ib_conn->login_req_dma)
325 			ib_dma_unmap_single(ib_conn->device->ib_device,
326 				ib_conn->login_req_dma,
327 				ISCSI_DEF_MAX_RECV_SEG_LEN, DMA_TO_DEVICE);
328 		if (ib_conn->login_resp_dma)
329 			ib_dma_unmap_single(ib_conn->device->ib_device,
330 				ib_conn->login_resp_dma,
331 				ISER_RX_LOGIN_SIZE, DMA_FROM_DEVICE);
332 		kfree(ib_conn->login_buf);
333 	}
334 
335 	return 0;
336 }
337 
338 /**
339  * based on the resolved device node GUID see if there already allocated
340  * device for this device. If there's no such, create one.
341  */
342 static
343 struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
344 {
345 	struct iser_device *device;
346 
347 	mutex_lock(&ig.device_list_mutex);
348 
349 	list_for_each_entry(device, &ig.device_list, ig_list)
350 		/* find if there's a match using the node GUID */
351 		if (device->ib_device->node_guid == cma_id->device->node_guid)
352 			goto inc_refcnt;
353 
354 	device = kzalloc(sizeof *device, GFP_KERNEL);
355 	if (device == NULL)
356 		goto out;
357 
358 	/* assign this device to the device */
359 	device->ib_device = cma_id->device;
360 	/* init the device and link it into ig device list */
361 	if (iser_create_device_ib_res(device)) {
362 		kfree(device);
363 		device = NULL;
364 		goto out;
365 	}
366 	list_add(&device->ig_list, &ig.device_list);
367 
368 inc_refcnt:
369 	device->refcount++;
370 out:
371 	mutex_unlock(&ig.device_list_mutex);
372 	return device;
373 }
374 
375 /* if there's no demand for this device, release it */
376 static void iser_device_try_release(struct iser_device *device)
377 {
378 	mutex_lock(&ig.device_list_mutex);
379 	device->refcount--;
380 	iser_info("device %p refcount %d\n", device, device->refcount);
381 	if (!device->refcount) {
382 		iser_free_device_ib_res(device);
383 		list_del(&device->ig_list);
384 		kfree(device);
385 	}
386 	mutex_unlock(&ig.device_list_mutex);
387 }
388 
389 static int iser_conn_state_comp_exch(struct iser_conn *ib_conn,
390 				     enum iser_ib_conn_state comp,
391 				     enum iser_ib_conn_state exch)
392 {
393 	int ret;
394 
395 	spin_lock_bh(&ib_conn->lock);
396 	if ((ret = (ib_conn->state == comp)))
397 		ib_conn->state = exch;
398 	spin_unlock_bh(&ib_conn->lock);
399 	return ret;
400 }
401 
402 /**
403  * Frees all conn objects and deallocs conn descriptor
404  */
405 static void iser_conn_release(struct iser_conn *ib_conn, int can_destroy_id)
406 {
407 	struct iser_device  *device = ib_conn->device;
408 
409 	BUG_ON(ib_conn->state != ISER_CONN_DOWN);
410 
411 	mutex_lock(&ig.connlist_mutex);
412 	list_del(&ib_conn->conn_list);
413 	mutex_unlock(&ig.connlist_mutex);
414 	iser_free_rx_descriptors(ib_conn);
415 	iser_free_ib_conn_res(ib_conn);
416 	ib_conn->device = NULL;
417 	/* on EVENT_ADDR_ERROR there's no device yet for this conn */
418 	if (device != NULL)
419 		iser_device_try_release(device);
420 	/* if cma handler context, the caller actually destroy the id */
421 	if (ib_conn->cma_id != NULL && can_destroy_id) {
422 		rdma_destroy_id(ib_conn->cma_id);
423 		ib_conn->cma_id = NULL;
424 	}
425 	iscsi_destroy_endpoint(ib_conn->ep);
426 }
427 
428 void iser_conn_get(struct iser_conn *ib_conn)
429 {
430 	atomic_inc(&ib_conn->refcount);
431 }
432 
433 int iser_conn_put(struct iser_conn *ib_conn, int can_destroy_id)
434 {
435 	if (atomic_dec_and_test(&ib_conn->refcount)) {
436 		iser_conn_release(ib_conn, can_destroy_id);
437 		return 1;
438 	}
439 	return 0;
440 }
441 
442 /**
443  * triggers start of the disconnect procedures and wait for them to be done
444  */
445 void iser_conn_terminate(struct iser_conn *ib_conn)
446 {
447 	int err = 0;
448 
449 	/* change the ib conn state only if the conn is UP, however always call
450 	 * rdma_disconnect since this is the only way to cause the CMA to change
451 	 * the QP state to ERROR
452 	 */
453 
454 	iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP, ISER_CONN_TERMINATING);
455 	err = rdma_disconnect(ib_conn->cma_id);
456 	if (err)
457 		iser_err("Failed to disconnect, conn: 0x%p err %d\n",
458 			 ib_conn,err);
459 
460 	wait_event_interruptible(ib_conn->wait,
461 				 ib_conn->state == ISER_CONN_DOWN);
462 
463 	iser_conn_put(ib_conn, 1); /* deref ib conn deallocate */
464 }
465 
466 static int iser_connect_error(struct rdma_cm_id *cma_id)
467 {
468 	struct iser_conn *ib_conn;
469 	ib_conn = (struct iser_conn *)cma_id->context;
470 
471 	ib_conn->state = ISER_CONN_DOWN;
472 	wake_up_interruptible(&ib_conn->wait);
473 	return iser_conn_put(ib_conn, 0); /* deref ib conn's cma id */
474 }
475 
476 static int iser_addr_handler(struct rdma_cm_id *cma_id)
477 {
478 	struct iser_device *device;
479 	struct iser_conn   *ib_conn;
480 	int    ret;
481 
482 	device = iser_device_find_by_ib_device(cma_id);
483 	if (!device) {
484 		iser_err("device lookup/creation failed\n");
485 		return iser_connect_error(cma_id);
486 	}
487 
488 	ib_conn = (struct iser_conn *)cma_id->context;
489 	ib_conn->device = device;
490 
491 	ret = rdma_resolve_route(cma_id, 1000);
492 	if (ret) {
493 		iser_err("resolve route failed: %d\n", ret);
494 		return iser_connect_error(cma_id);
495 	}
496 
497 	return 0;
498 }
499 
500 static int iser_route_handler(struct rdma_cm_id *cma_id)
501 {
502 	struct rdma_conn_param conn_param;
503 	int    ret;
504 	struct iser_cm_hdr req_hdr;
505 
506 	ret = iser_create_ib_conn_res((struct iser_conn *)cma_id->context);
507 	if (ret)
508 		goto failure;
509 
510 	memset(&conn_param, 0, sizeof conn_param);
511 	conn_param.responder_resources = 4;
512 	conn_param.initiator_depth     = 1;
513 	conn_param.retry_count	       = 7;
514 	conn_param.rnr_retry_count     = 6;
515 
516 	memset(&req_hdr, 0, sizeof(req_hdr));
517 	req_hdr.flags = (ISER_ZBVA_NOT_SUPPORTED |
518 			ISER_SEND_W_INV_NOT_SUPPORTED);
519 	conn_param.private_data		= (void *)&req_hdr;
520 	conn_param.private_data_len	= sizeof(struct iser_cm_hdr);
521 
522 	ret = rdma_connect(cma_id, &conn_param);
523 	if (ret) {
524 		iser_err("failure connecting: %d\n", ret);
525 		goto failure;
526 	}
527 
528 	return 0;
529 failure:
530 	return iser_connect_error(cma_id);
531 }
532 
533 static void iser_connected_handler(struct rdma_cm_id *cma_id)
534 {
535 	struct iser_conn *ib_conn;
536 
537 	ib_conn = (struct iser_conn *)cma_id->context;
538 	ib_conn->state = ISER_CONN_UP;
539 	wake_up_interruptible(&ib_conn->wait);
540 }
541 
542 static int iser_disconnected_handler(struct rdma_cm_id *cma_id)
543 {
544 	struct iser_conn *ib_conn;
545 	int ret;
546 
547 	ib_conn = (struct iser_conn *)cma_id->context;
548 
549 	/* getting here when the state is UP means that the conn is being *
550 	 * terminated asynchronously from the iSCSI layer's perspective.  */
551 	if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
552 				      ISER_CONN_TERMINATING))
553 		iscsi_conn_failure(ib_conn->iser_conn->iscsi_conn,
554 				   ISCSI_ERR_CONN_FAILED);
555 
556 	/* Complete the termination process if no posts are pending */
557 	if (ib_conn->post_recv_buf_count == 0 &&
558 	    (atomic_read(&ib_conn->post_send_buf_count) == 0)) {
559 		ib_conn->state = ISER_CONN_DOWN;
560 		wake_up_interruptible(&ib_conn->wait);
561 	}
562 
563 	ret = iser_conn_put(ib_conn, 0); /* deref ib conn's cma id */
564 	return ret;
565 }
566 
567 static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
568 {
569 	int ret = 0;
570 
571 	iser_info("event %d status %d conn %p id %p\n",
572 		  event->event, event->status, cma_id->context, cma_id);
573 
574 	switch (event->event) {
575 	case RDMA_CM_EVENT_ADDR_RESOLVED:
576 		ret = iser_addr_handler(cma_id);
577 		break;
578 	case RDMA_CM_EVENT_ROUTE_RESOLVED:
579 		ret = iser_route_handler(cma_id);
580 		break;
581 	case RDMA_CM_EVENT_ESTABLISHED:
582 		iser_connected_handler(cma_id);
583 		break;
584 	case RDMA_CM_EVENT_ADDR_ERROR:
585 	case RDMA_CM_EVENT_ROUTE_ERROR:
586 	case RDMA_CM_EVENT_CONNECT_ERROR:
587 	case RDMA_CM_EVENT_UNREACHABLE:
588 	case RDMA_CM_EVENT_REJECTED:
589 		ret = iser_connect_error(cma_id);
590 		break;
591 	case RDMA_CM_EVENT_DISCONNECTED:
592 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
593 	case RDMA_CM_EVENT_ADDR_CHANGE:
594 		ret = iser_disconnected_handler(cma_id);
595 		break;
596 	default:
597 		iser_err("Unexpected RDMA CM event (%d)\n", event->event);
598 		break;
599 	}
600 	return ret;
601 }
602 
603 void iser_conn_init(struct iser_conn *ib_conn)
604 {
605 	ib_conn->state = ISER_CONN_INIT;
606 	init_waitqueue_head(&ib_conn->wait);
607 	ib_conn->post_recv_buf_count = 0;
608 	atomic_set(&ib_conn->post_send_buf_count, 0);
609 	atomic_set(&ib_conn->refcount, 1); /* ref ib conn allocation */
610 	INIT_LIST_HEAD(&ib_conn->conn_list);
611 	spin_lock_init(&ib_conn->lock);
612 }
613 
614  /**
615  * starts the process of connecting to the target
616  * sleeps until the connection is established or rejected
617  */
618 int iser_connect(struct iser_conn   *ib_conn,
619 		 struct sockaddr_in *src_addr,
620 		 struct sockaddr_in *dst_addr,
621 		 int                 non_blocking)
622 {
623 	struct sockaddr *src, *dst;
624 	int err = 0;
625 
626 	sprintf(ib_conn->name, "%pI4:%d",
627 		&dst_addr->sin_addr.s_addr, dst_addr->sin_port);
628 
629 	/* the device is known only --after-- address resolution */
630 	ib_conn->device = NULL;
631 
632 	iser_info("connecting to: %pI4, port 0x%x\n",
633 		  &dst_addr->sin_addr, dst_addr->sin_port);
634 
635 	ib_conn->state = ISER_CONN_PENDING;
636 
637 	iser_conn_get(ib_conn); /* ref ib conn's cma id */
638 	ib_conn->cma_id = rdma_create_id(iser_cma_handler,
639 					     (void *)ib_conn,
640 					     RDMA_PS_TCP, IB_QPT_RC);
641 	if (IS_ERR(ib_conn->cma_id)) {
642 		err = PTR_ERR(ib_conn->cma_id);
643 		iser_err("rdma_create_id failed: %d\n", err);
644 		goto id_failure;
645 	}
646 
647 	src = (struct sockaddr *)src_addr;
648 	dst = (struct sockaddr *)dst_addr;
649 	err = rdma_resolve_addr(ib_conn->cma_id, src, dst, 1000);
650 	if (err) {
651 		iser_err("rdma_resolve_addr failed: %d\n", err);
652 		goto addr_failure;
653 	}
654 
655 	if (!non_blocking) {
656 		wait_event_interruptible(ib_conn->wait,
657 					 (ib_conn->state != ISER_CONN_PENDING));
658 
659 		if (ib_conn->state != ISER_CONN_UP) {
660 			err =  -EIO;
661 			goto connect_failure;
662 		}
663 	}
664 
665 	mutex_lock(&ig.connlist_mutex);
666 	list_add(&ib_conn->conn_list, &ig.connlist);
667 	mutex_unlock(&ig.connlist_mutex);
668 	return 0;
669 
670 id_failure:
671 	ib_conn->cma_id = NULL;
672 addr_failure:
673 	ib_conn->state = ISER_CONN_DOWN;
674 	iser_conn_put(ib_conn, 1); /* deref ib conn's cma id */
675 connect_failure:
676 	iser_conn_put(ib_conn, 1); /* deref ib conn deallocate */
677 	return err;
678 }
679 
680 /**
681  * iser_reg_page_vec - Register physical memory
682  *
683  * returns: 0 on success, errno code on failure
684  */
685 int iser_reg_page_vec(struct iser_conn     *ib_conn,
686 		      struct iser_page_vec *page_vec,
687 		      struct iser_mem_reg  *mem_reg)
688 {
689 	struct ib_pool_fmr *mem;
690 	u64		   io_addr;
691 	u64		   *page_list;
692 	int		   status;
693 
694 	page_list = page_vec->pages;
695 	io_addr	  = page_list[0];
696 
697 	mem  = ib_fmr_pool_map_phys(ib_conn->fmr_pool,
698 				    page_list,
699 				    page_vec->length,
700 				    io_addr);
701 
702 	if (IS_ERR(mem)) {
703 		status = (int)PTR_ERR(mem);
704 		iser_err("ib_fmr_pool_map_phys failed: %d\n", status);
705 		return status;
706 	}
707 
708 	mem_reg->lkey  = mem->fmr->lkey;
709 	mem_reg->rkey  = mem->fmr->rkey;
710 	mem_reg->len   = page_vec->length * SIZE_4K;
711 	mem_reg->va    = io_addr;
712 	mem_reg->is_fmr = 1;
713 	mem_reg->mem_h = (void *)mem;
714 
715 	mem_reg->va   += page_vec->offset;
716 	mem_reg->len   = page_vec->data_size;
717 
718 	iser_dbg("PHYSICAL Mem.register, [PHYS p_array: 0x%p, sz: %d, "
719 		 "entry[0]: (0x%08lx,%ld)] -> "
720 		 "[lkey: 0x%08X mem_h: 0x%p va: 0x%08lX sz: %ld]\n",
721 		 page_vec, page_vec->length,
722 		 (unsigned long)page_vec->pages[0],
723 		 (unsigned long)page_vec->data_size,
724 		 (unsigned int)mem_reg->lkey, mem_reg->mem_h,
725 		 (unsigned long)mem_reg->va, (unsigned long)mem_reg->len);
726 	return 0;
727 }
728 
729 /**
730  * Unregister (previosuly registered) memory.
731  */
732 void iser_unreg_mem(struct iser_mem_reg *reg)
733 {
734 	int ret;
735 
736 	iser_dbg("PHYSICAL Mem.Unregister mem_h %p\n",reg->mem_h);
737 
738 	ret = ib_fmr_pool_unmap((struct ib_pool_fmr *)reg->mem_h);
739 	if (ret)
740 		iser_err("ib_fmr_pool_unmap failed %d\n", ret);
741 
742 	reg->mem_h = NULL;
743 }
744 
745 int iser_post_recvl(struct iser_conn *ib_conn)
746 {
747 	struct ib_recv_wr rx_wr, *rx_wr_failed;
748 	struct ib_sge	  sge;
749 	int ib_ret;
750 
751 	sge.addr   = ib_conn->login_resp_dma;
752 	sge.length = ISER_RX_LOGIN_SIZE;
753 	sge.lkey   = ib_conn->device->mr->lkey;
754 
755 	rx_wr.wr_id   = (unsigned long)ib_conn->login_resp_buf;
756 	rx_wr.sg_list = &sge;
757 	rx_wr.num_sge = 1;
758 	rx_wr.next    = NULL;
759 
760 	ib_conn->post_recv_buf_count++;
761 	ib_ret	= ib_post_recv(ib_conn->qp, &rx_wr, &rx_wr_failed);
762 	if (ib_ret) {
763 		iser_err("ib_post_recv failed ret=%d\n", ib_ret);
764 		ib_conn->post_recv_buf_count--;
765 	}
766 	return ib_ret;
767 }
768 
769 int iser_post_recvm(struct iser_conn *ib_conn, int count)
770 {
771 	struct ib_recv_wr *rx_wr, *rx_wr_failed;
772 	int i, ib_ret;
773 	unsigned int my_rx_head = ib_conn->rx_desc_head;
774 	struct iser_rx_desc *rx_desc;
775 
776 	for (rx_wr = ib_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
777 		rx_desc		= &ib_conn->rx_descs[my_rx_head];
778 		rx_wr->wr_id	= (unsigned long)rx_desc;
779 		rx_wr->sg_list	= &rx_desc->rx_sg;
780 		rx_wr->num_sge	= 1;
781 		rx_wr->next	= rx_wr + 1;
782 		my_rx_head = (my_rx_head + 1) & (ISER_QP_MAX_RECV_DTOS - 1);
783 	}
784 
785 	rx_wr--;
786 	rx_wr->next = NULL; /* mark end of work requests list */
787 
788 	ib_conn->post_recv_buf_count += count;
789 	ib_ret	= ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &rx_wr_failed);
790 	if (ib_ret) {
791 		iser_err("ib_post_recv failed ret=%d\n", ib_ret);
792 		ib_conn->post_recv_buf_count -= count;
793 	} else
794 		ib_conn->rx_desc_head = my_rx_head;
795 	return ib_ret;
796 }
797 
798 
799 /**
800  * iser_start_send - Initiate a Send DTO operation
801  *
802  * returns 0 on success, -1 on failure
803  */
804 int iser_post_send(struct iser_conn *ib_conn, struct iser_tx_desc *tx_desc)
805 {
806 	int		  ib_ret;
807 	struct ib_send_wr send_wr, *send_wr_failed;
808 
809 	ib_dma_sync_single_for_device(ib_conn->device->ib_device,
810 		tx_desc->dma_addr, ISER_HEADERS_LEN, DMA_TO_DEVICE);
811 
812 	send_wr.next	   = NULL;
813 	send_wr.wr_id	   = (unsigned long)tx_desc;
814 	send_wr.sg_list	   = tx_desc->tx_sg;
815 	send_wr.num_sge	   = tx_desc->num_sge;
816 	send_wr.opcode	   = IB_WR_SEND;
817 	send_wr.send_flags = IB_SEND_SIGNALED;
818 
819 	atomic_inc(&ib_conn->post_send_buf_count);
820 
821 	ib_ret = ib_post_send(ib_conn->qp, &send_wr, &send_wr_failed);
822 	if (ib_ret) {
823 		iser_err("ib_post_send failed, ret:%d\n", ib_ret);
824 		atomic_dec(&ib_conn->post_send_buf_count);
825 	}
826 	return ib_ret;
827 }
828 
829 static void iser_handle_comp_error(struct iser_tx_desc *desc,
830 				struct iser_conn *ib_conn)
831 {
832 	if (desc && desc->type == ISCSI_TX_DATAOUT)
833 		kmem_cache_free(ig.desc_cache, desc);
834 
835 	if (ib_conn->post_recv_buf_count == 0 &&
836 	    atomic_read(&ib_conn->post_send_buf_count) == 0) {
837 		/* getting here when the state is UP means that the conn is *
838 		 * being terminated asynchronously from the iSCSI layer's   *
839 		 * perspective.                                             */
840 		if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
841 		    ISER_CONN_TERMINATING))
842 			iscsi_conn_failure(ib_conn->iser_conn->iscsi_conn,
843 					   ISCSI_ERR_CONN_FAILED);
844 
845 		/* no more non completed posts to the QP, complete the
846 		 * termination process w.o worrying on disconnect event */
847 		ib_conn->state = ISER_CONN_DOWN;
848 		wake_up_interruptible(&ib_conn->wait);
849 	}
850 }
851 
852 static int iser_drain_tx_cq(struct iser_device  *device, int cq_index)
853 {
854 	struct ib_cq  *cq = device->tx_cq[cq_index];
855 	struct ib_wc  wc;
856 	struct iser_tx_desc *tx_desc;
857 	struct iser_conn *ib_conn;
858 	int completed_tx = 0;
859 
860 	while (ib_poll_cq(cq, 1, &wc) == 1) {
861 		tx_desc	= (struct iser_tx_desc *) (unsigned long) wc.wr_id;
862 		ib_conn = wc.qp->qp_context;
863 		if (wc.status == IB_WC_SUCCESS) {
864 			if (wc.opcode == IB_WC_SEND)
865 				iser_snd_completion(tx_desc, ib_conn);
866 			else
867 				iser_err("expected opcode %d got %d\n",
868 					IB_WC_SEND, wc.opcode);
869 		} else {
870 			iser_err("tx id %llx status %d vend_err %x\n",
871 				wc.wr_id, wc.status, wc.vendor_err);
872 			atomic_dec(&ib_conn->post_send_buf_count);
873 			iser_handle_comp_error(tx_desc, ib_conn);
874 		}
875 		completed_tx++;
876 	}
877 	return completed_tx;
878 }
879 
880 
881 static void iser_cq_tasklet_fn(unsigned long data)
882 {
883 	struct iser_cq_desc *cq_desc = (struct iser_cq_desc *)data;
884 	struct iser_device  *device = cq_desc->device;
885 	int cq_index = cq_desc->cq_index;
886 	struct ib_cq	     *cq = device->rx_cq[cq_index];
887 	 struct ib_wc	     wc;
888 	 struct iser_rx_desc *desc;
889 	 unsigned long	     xfer_len;
890 	struct iser_conn *ib_conn;
891 	int completed_tx, completed_rx;
892 	completed_tx = completed_rx = 0;
893 
894 	while (ib_poll_cq(cq, 1, &wc) == 1) {
895 		desc	 = (struct iser_rx_desc *) (unsigned long) wc.wr_id;
896 		BUG_ON(desc == NULL);
897 		ib_conn = wc.qp->qp_context;
898 		if (wc.status == IB_WC_SUCCESS) {
899 			if (wc.opcode == IB_WC_RECV) {
900 				xfer_len = (unsigned long)wc.byte_len;
901 				iser_rcv_completion(desc, xfer_len, ib_conn);
902 			} else
903 				iser_err("expected opcode %d got %d\n",
904 					IB_WC_RECV, wc.opcode);
905 		} else {
906 			if (wc.status != IB_WC_WR_FLUSH_ERR)
907 				iser_err("rx id %llx status %d vend_err %x\n",
908 					wc.wr_id, wc.status, wc.vendor_err);
909 			ib_conn->post_recv_buf_count--;
910 			iser_handle_comp_error(NULL, ib_conn);
911 		}
912 		completed_rx++;
913 		if (!(completed_rx & 63))
914 			completed_tx += iser_drain_tx_cq(device, cq_index);
915 	}
916 	/* #warning "it is assumed here that arming CQ only once its empty" *
917 	 * " would not cause interrupts to be missed"                       */
918 	ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
919 
920 	completed_tx += iser_drain_tx_cq(device, cq_index);
921 	iser_dbg("got %d rx %d tx completions\n", completed_rx, completed_tx);
922 }
923 
924 static void iser_cq_callback(struct ib_cq *cq, void *cq_context)
925 {
926 	struct iser_cq_desc *cq_desc = (struct iser_cq_desc *)cq_context;
927 	struct iser_device  *device = cq_desc->device;
928 	int cq_index = cq_desc->cq_index;
929 
930 	tasklet_schedule(&device->cq_tasklet[cq_index]);
931 }
932