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