1 /*
2  * Copyright (c) 2005 Cisco Systems.  All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32 
33 #include <linux/module.h>
34 #include <linux/init.h>
35 #include <linux/slab.h>
36 #include <linux/err.h>
37 #include <linux/string.h>
38 #include <linux/parser.h>
39 #include <linux/random.h>
40 #include <linux/jiffies.h>
41 
42 #include <asm/atomic.h>
43 
44 #include <scsi/scsi.h>
45 #include <scsi/scsi_device.h>
46 #include <scsi/scsi_dbg.h>
47 #include <scsi/srp.h>
48 #include <scsi/scsi_transport_srp.h>
49 
50 #include "ib_srp.h"
51 
52 #define DRV_NAME	"ib_srp"
53 #define PFX		DRV_NAME ": "
54 #define DRV_VERSION	"0.2"
55 #define DRV_RELDATE	"November 1, 2005"
56 
57 MODULE_AUTHOR("Roland Dreier");
58 MODULE_DESCRIPTION("InfiniBand SCSI RDMA Protocol initiator "
59 		   "v" DRV_VERSION " (" DRV_RELDATE ")");
60 MODULE_LICENSE("Dual BSD/GPL");
61 
62 static int srp_sg_tablesize = SRP_DEF_SG_TABLESIZE;
63 static int srp_max_iu_len;
64 
65 module_param(srp_sg_tablesize, int, 0444);
66 MODULE_PARM_DESC(srp_sg_tablesize,
67 		 "Max number of gather/scatter entries per I/O (default is 12, max 255)");
68 
69 static int topspin_workarounds = 1;
70 
71 module_param(topspin_workarounds, int, 0444);
72 MODULE_PARM_DESC(topspin_workarounds,
73 		 "Enable workarounds for Topspin/Cisco SRP target bugs if != 0");
74 
75 static int mellanox_workarounds = 1;
76 
77 module_param(mellanox_workarounds, int, 0444);
78 MODULE_PARM_DESC(mellanox_workarounds,
79 		 "Enable workarounds for Mellanox SRP target bugs if != 0");
80 
81 static void srp_add_one(struct ib_device *device);
82 static void srp_remove_one(struct ib_device *device);
83 static void srp_recv_completion(struct ib_cq *cq, void *target_ptr);
84 static void srp_send_completion(struct ib_cq *cq, void *target_ptr);
85 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event);
86 
87 static struct scsi_transport_template *ib_srp_transport_template;
88 
89 static struct ib_client srp_client = {
90 	.name   = "srp",
91 	.add    = srp_add_one,
92 	.remove = srp_remove_one
93 };
94 
95 static struct ib_sa_client srp_sa_client;
96 
97 static inline struct srp_target_port *host_to_target(struct Scsi_Host *host)
98 {
99 	return (struct srp_target_port *) host->hostdata;
100 }
101 
102 static const char *srp_target_info(struct Scsi_Host *host)
103 {
104 	return host_to_target(host)->target_name;
105 }
106 
107 static int srp_target_is_topspin(struct srp_target_port *target)
108 {
109 	static const u8 topspin_oui[3] = { 0x00, 0x05, 0xad };
110 	static const u8 cisco_oui[3]   = { 0x00, 0x1b, 0x0d };
111 
112 	return topspin_workarounds &&
113 		(!memcmp(&target->ioc_guid, topspin_oui, sizeof topspin_oui) ||
114 		 !memcmp(&target->ioc_guid, cisco_oui, sizeof cisco_oui));
115 }
116 
117 static int srp_target_is_mellanox(struct srp_target_port *target)
118 {
119 	static const u8 mellanox_oui[3] = { 0x00, 0x02, 0xc9 };
120 
121 	return mellanox_workarounds &&
122 		!memcmp(&target->ioc_guid, mellanox_oui, sizeof mellanox_oui);
123 }
124 
125 static struct srp_iu *srp_alloc_iu(struct srp_host *host, size_t size,
126 				   gfp_t gfp_mask,
127 				   enum dma_data_direction direction)
128 {
129 	struct srp_iu *iu;
130 
131 	iu = kmalloc(sizeof *iu, gfp_mask);
132 	if (!iu)
133 		goto out;
134 
135 	iu->buf = kzalloc(size, gfp_mask);
136 	if (!iu->buf)
137 		goto out_free_iu;
138 
139 	iu->dma = ib_dma_map_single(host->srp_dev->dev, iu->buf, size,
140 				    direction);
141 	if (ib_dma_mapping_error(host->srp_dev->dev, iu->dma))
142 		goto out_free_buf;
143 
144 	iu->size      = size;
145 	iu->direction = direction;
146 
147 	return iu;
148 
149 out_free_buf:
150 	kfree(iu->buf);
151 out_free_iu:
152 	kfree(iu);
153 out:
154 	return NULL;
155 }
156 
157 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
158 {
159 	if (!iu)
160 		return;
161 
162 	ib_dma_unmap_single(host->srp_dev->dev, iu->dma, iu->size,
163 			    iu->direction);
164 	kfree(iu->buf);
165 	kfree(iu);
166 }
167 
168 static void srp_qp_event(struct ib_event *event, void *context)
169 {
170 	printk(KERN_ERR PFX "QP event %d\n", event->event);
171 }
172 
173 static int srp_init_qp(struct srp_target_port *target,
174 		       struct ib_qp *qp)
175 {
176 	struct ib_qp_attr *attr;
177 	int ret;
178 
179 	attr = kmalloc(sizeof *attr, GFP_KERNEL);
180 	if (!attr)
181 		return -ENOMEM;
182 
183 	ret = ib_find_pkey(target->srp_host->srp_dev->dev,
184 			   target->srp_host->port,
185 			   be16_to_cpu(target->path.pkey),
186 			   &attr->pkey_index);
187 	if (ret)
188 		goto out;
189 
190 	attr->qp_state        = IB_QPS_INIT;
191 	attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
192 				    IB_ACCESS_REMOTE_WRITE);
193 	attr->port_num        = target->srp_host->port;
194 
195 	ret = ib_modify_qp(qp, attr,
196 			   IB_QP_STATE		|
197 			   IB_QP_PKEY_INDEX	|
198 			   IB_QP_ACCESS_FLAGS	|
199 			   IB_QP_PORT);
200 
201 out:
202 	kfree(attr);
203 	return ret;
204 }
205 
206 static int srp_new_cm_id(struct srp_target_port *target)
207 {
208 	struct ib_cm_id *new_cm_id;
209 
210 	new_cm_id = ib_create_cm_id(target->srp_host->srp_dev->dev,
211 				    srp_cm_handler, target);
212 	if (IS_ERR(new_cm_id))
213 		return PTR_ERR(new_cm_id);
214 
215 	if (target->cm_id)
216 		ib_destroy_cm_id(target->cm_id);
217 	target->cm_id = new_cm_id;
218 
219 	return 0;
220 }
221 
222 static int srp_create_target_ib(struct srp_target_port *target)
223 {
224 	struct ib_qp_init_attr *init_attr;
225 	int ret;
226 
227 	init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
228 	if (!init_attr)
229 		return -ENOMEM;
230 
231 	target->recv_cq = ib_create_cq(target->srp_host->srp_dev->dev,
232 				       srp_recv_completion, NULL, target, SRP_RQ_SIZE, 0);
233 	if (IS_ERR(target->recv_cq)) {
234 		ret = PTR_ERR(target->recv_cq);
235 		goto err;
236 	}
237 
238 	target->send_cq = ib_create_cq(target->srp_host->srp_dev->dev,
239 				       srp_send_completion, NULL, target, SRP_SQ_SIZE, 0);
240 	if (IS_ERR(target->send_cq)) {
241 		ret = PTR_ERR(target->send_cq);
242 		goto err_recv_cq;
243 	}
244 
245 	ib_req_notify_cq(target->recv_cq, IB_CQ_NEXT_COMP);
246 
247 	init_attr->event_handler       = srp_qp_event;
248 	init_attr->cap.max_send_wr     = SRP_SQ_SIZE;
249 	init_attr->cap.max_recv_wr     = SRP_RQ_SIZE;
250 	init_attr->cap.max_recv_sge    = 1;
251 	init_attr->cap.max_send_sge    = 1;
252 	init_attr->sq_sig_type         = IB_SIGNAL_ALL_WR;
253 	init_attr->qp_type             = IB_QPT_RC;
254 	init_attr->send_cq             = target->send_cq;
255 	init_attr->recv_cq             = target->recv_cq;
256 
257 	target->qp = ib_create_qp(target->srp_host->srp_dev->pd, init_attr);
258 	if (IS_ERR(target->qp)) {
259 		ret = PTR_ERR(target->qp);
260 		goto err_send_cq;
261 	}
262 
263 	ret = srp_init_qp(target, target->qp);
264 	if (ret)
265 		goto err_qp;
266 
267 	kfree(init_attr);
268 	return 0;
269 
270 err_qp:
271 	ib_destroy_qp(target->qp);
272 
273 err_send_cq:
274 	ib_destroy_cq(target->send_cq);
275 
276 err_recv_cq:
277 	ib_destroy_cq(target->recv_cq);
278 
279 err:
280 	kfree(init_attr);
281 	return ret;
282 }
283 
284 static void srp_free_target_ib(struct srp_target_port *target)
285 {
286 	int i;
287 
288 	ib_destroy_qp(target->qp);
289 	ib_destroy_cq(target->send_cq);
290 	ib_destroy_cq(target->recv_cq);
291 
292 	for (i = 0; i < SRP_RQ_SIZE; ++i)
293 		srp_free_iu(target->srp_host, target->rx_ring[i]);
294 	for (i = 0; i < SRP_SQ_SIZE; ++i)
295 		srp_free_iu(target->srp_host, target->tx_ring[i]);
296 }
297 
298 static void srp_path_rec_completion(int status,
299 				    struct ib_sa_path_rec *pathrec,
300 				    void *target_ptr)
301 {
302 	struct srp_target_port *target = target_ptr;
303 
304 	target->status = status;
305 	if (status)
306 		shost_printk(KERN_ERR, target->scsi_host,
307 			     PFX "Got failed path rec status %d\n", status);
308 	else
309 		target->path = *pathrec;
310 	complete(&target->done);
311 }
312 
313 static int srp_lookup_path(struct srp_target_port *target)
314 {
315 	target->path.numb_path = 1;
316 
317 	init_completion(&target->done);
318 
319 	target->path_query_id = ib_sa_path_rec_get(&srp_sa_client,
320 						   target->srp_host->srp_dev->dev,
321 						   target->srp_host->port,
322 						   &target->path,
323 						   IB_SA_PATH_REC_SERVICE_ID	|
324 						   IB_SA_PATH_REC_DGID		|
325 						   IB_SA_PATH_REC_SGID		|
326 						   IB_SA_PATH_REC_NUMB_PATH	|
327 						   IB_SA_PATH_REC_PKEY,
328 						   SRP_PATH_REC_TIMEOUT_MS,
329 						   GFP_KERNEL,
330 						   srp_path_rec_completion,
331 						   target, &target->path_query);
332 	if (target->path_query_id < 0)
333 		return target->path_query_id;
334 
335 	wait_for_completion(&target->done);
336 
337 	if (target->status < 0)
338 		shost_printk(KERN_WARNING, target->scsi_host,
339 			     PFX "Path record query failed\n");
340 
341 	return target->status;
342 }
343 
344 static int srp_send_req(struct srp_target_port *target)
345 {
346 	struct {
347 		struct ib_cm_req_param param;
348 		struct srp_login_req   priv;
349 	} *req = NULL;
350 	int status;
351 
352 	req = kzalloc(sizeof *req, GFP_KERNEL);
353 	if (!req)
354 		return -ENOMEM;
355 
356 	req->param.primary_path 	      = &target->path;
357 	req->param.alternate_path 	      = NULL;
358 	req->param.service_id 		      = target->service_id;
359 	req->param.qp_num 		      = target->qp->qp_num;
360 	req->param.qp_type 		      = target->qp->qp_type;
361 	req->param.private_data 	      = &req->priv;
362 	req->param.private_data_len 	      = sizeof req->priv;
363 	req->param.flow_control 	      = 1;
364 
365 	get_random_bytes(&req->param.starting_psn, 4);
366 	req->param.starting_psn 	     &= 0xffffff;
367 
368 	/*
369 	 * Pick some arbitrary defaults here; we could make these
370 	 * module parameters if anyone cared about setting them.
371 	 */
372 	req->param.responder_resources	      = 4;
373 	req->param.remote_cm_response_timeout = 20;
374 	req->param.local_cm_response_timeout  = 20;
375 	req->param.retry_count 		      = 7;
376 	req->param.rnr_retry_count 	      = 7;
377 	req->param.max_cm_retries 	      = 15;
378 
379 	req->priv.opcode     	= SRP_LOGIN_REQ;
380 	req->priv.tag        	= 0;
381 	req->priv.req_it_iu_len = cpu_to_be32(srp_max_iu_len);
382 	req->priv.req_buf_fmt 	= cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
383 					      SRP_BUF_FORMAT_INDIRECT);
384 	/*
385 	 * In the published SRP specification (draft rev. 16a), the
386 	 * port identifier format is 8 bytes of ID extension followed
387 	 * by 8 bytes of GUID.  Older drafts put the two halves in the
388 	 * opposite order, so that the GUID comes first.
389 	 *
390 	 * Targets conforming to these obsolete drafts can be
391 	 * recognized by the I/O Class they report.
392 	 */
393 	if (target->io_class == SRP_REV10_IB_IO_CLASS) {
394 		memcpy(req->priv.initiator_port_id,
395 		       &target->path.sgid.global.interface_id, 8);
396 		memcpy(req->priv.initiator_port_id + 8,
397 		       &target->initiator_ext, 8);
398 		memcpy(req->priv.target_port_id,     &target->ioc_guid, 8);
399 		memcpy(req->priv.target_port_id + 8, &target->id_ext, 8);
400 	} else {
401 		memcpy(req->priv.initiator_port_id,
402 		       &target->initiator_ext, 8);
403 		memcpy(req->priv.initiator_port_id + 8,
404 		       &target->path.sgid.global.interface_id, 8);
405 		memcpy(req->priv.target_port_id,     &target->id_ext, 8);
406 		memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
407 	}
408 
409 	/*
410 	 * Topspin/Cisco SRP targets will reject our login unless we
411 	 * zero out the first 8 bytes of our initiator port ID and set
412 	 * the second 8 bytes to the local node GUID.
413 	 */
414 	if (srp_target_is_topspin(target)) {
415 		shost_printk(KERN_DEBUG, target->scsi_host,
416 			     PFX "Topspin/Cisco initiator port ID workaround "
417 			     "activated for target GUID %016llx\n",
418 			     (unsigned long long) be64_to_cpu(target->ioc_guid));
419 		memset(req->priv.initiator_port_id, 0, 8);
420 		memcpy(req->priv.initiator_port_id + 8,
421 		       &target->srp_host->srp_dev->dev->node_guid, 8);
422 	}
423 
424 	status = ib_send_cm_req(target->cm_id, &req->param);
425 
426 	kfree(req);
427 
428 	return status;
429 }
430 
431 static void srp_disconnect_target(struct srp_target_port *target)
432 {
433 	/* XXX should send SRP_I_LOGOUT request */
434 
435 	init_completion(&target->done);
436 	if (ib_send_cm_dreq(target->cm_id, NULL, 0)) {
437 		shost_printk(KERN_DEBUG, target->scsi_host,
438 			     PFX "Sending CM DREQ failed\n");
439 		return;
440 	}
441 	wait_for_completion(&target->done);
442 }
443 
444 static bool srp_change_state(struct srp_target_port *target,
445 			    enum srp_target_state old,
446 			    enum srp_target_state new)
447 {
448 	bool changed = false;
449 
450 	spin_lock_irq(&target->lock);
451 	if (target->state == old) {
452 		target->state = new;
453 		changed = true;
454 	}
455 	spin_unlock_irq(&target->lock);
456 	return changed;
457 }
458 
459 static void srp_remove_work(struct work_struct *work)
460 {
461 	struct srp_target_port *target =
462 		container_of(work, struct srp_target_port, work);
463 
464 	if (!srp_change_state(target, SRP_TARGET_DEAD, SRP_TARGET_REMOVED))
465 		return;
466 
467 	spin_lock(&target->srp_host->target_lock);
468 	list_del(&target->list);
469 	spin_unlock(&target->srp_host->target_lock);
470 
471 	srp_remove_host(target->scsi_host);
472 	scsi_remove_host(target->scsi_host);
473 	ib_destroy_cm_id(target->cm_id);
474 	srp_free_target_ib(target);
475 	scsi_host_put(target->scsi_host);
476 }
477 
478 static int srp_connect_target(struct srp_target_port *target)
479 {
480 	int retries = 3;
481 	int ret;
482 
483 	ret = srp_lookup_path(target);
484 	if (ret)
485 		return ret;
486 
487 	while (1) {
488 		init_completion(&target->done);
489 		ret = srp_send_req(target);
490 		if (ret)
491 			return ret;
492 		wait_for_completion(&target->done);
493 
494 		/*
495 		 * The CM event handling code will set status to
496 		 * SRP_PORT_REDIRECT if we get a port redirect REJ
497 		 * back, or SRP_DLID_REDIRECT if we get a lid/qp
498 		 * redirect REJ back.
499 		 */
500 		switch (target->status) {
501 		case 0:
502 			return 0;
503 
504 		case SRP_PORT_REDIRECT:
505 			ret = srp_lookup_path(target);
506 			if (ret)
507 				return ret;
508 			break;
509 
510 		case SRP_DLID_REDIRECT:
511 			break;
512 
513 		case SRP_STALE_CONN:
514 			/* Our current CM id was stale, and is now in timewait.
515 			 * Try to reconnect with a new one.
516 			 */
517 			if (!retries-- || srp_new_cm_id(target)) {
518 				shost_printk(KERN_ERR, target->scsi_host, PFX
519 					     "giving up on stale connection\n");
520 				target->status = -ECONNRESET;
521 				return target->status;
522 			}
523 
524 			shost_printk(KERN_ERR, target->scsi_host, PFX
525 				     "retrying stale connection\n");
526 			break;
527 
528 		default:
529 			return target->status;
530 		}
531 	}
532 }
533 
534 static void srp_unmap_data(struct scsi_cmnd *scmnd,
535 			   struct srp_target_port *target,
536 			   struct srp_request *req)
537 {
538 	if (!scsi_sglist(scmnd) ||
539 	    (scmnd->sc_data_direction != DMA_TO_DEVICE &&
540 	     scmnd->sc_data_direction != DMA_FROM_DEVICE))
541 		return;
542 
543 	if (req->fmr) {
544 		ib_fmr_pool_unmap(req->fmr);
545 		req->fmr = NULL;
546 	}
547 
548 	ib_dma_unmap_sg(target->srp_host->srp_dev->dev, scsi_sglist(scmnd),
549 			scsi_sg_count(scmnd), scmnd->sc_data_direction);
550 }
551 
552 static void srp_remove_req(struct srp_target_port *target,
553 			   struct srp_request *req, s32 req_lim_delta)
554 {
555 	unsigned long flags;
556 
557 	srp_unmap_data(req->scmnd, target, req);
558 	spin_lock_irqsave(&target->lock, flags);
559 	target->req_lim += req_lim_delta;
560 	req->scmnd = NULL;
561 	list_add_tail(&req->list, &target->free_reqs);
562 	spin_unlock_irqrestore(&target->lock, flags);
563 }
564 
565 static void srp_reset_req(struct srp_target_port *target, struct srp_request *req)
566 {
567 	req->scmnd->result = DID_RESET << 16;
568 	req->scmnd->scsi_done(req->scmnd);
569 	srp_remove_req(target, req, 0);
570 }
571 
572 static int srp_reconnect_target(struct srp_target_port *target)
573 {
574 	struct ib_qp_attr qp_attr;
575 	struct ib_wc wc;
576 	int i, ret;
577 
578 	if (!srp_change_state(target, SRP_TARGET_LIVE, SRP_TARGET_CONNECTING))
579 		return -EAGAIN;
580 
581 	srp_disconnect_target(target);
582 	/*
583 	 * Now get a new local CM ID so that we avoid confusing the
584 	 * target in case things are really fouled up.
585 	 */
586 	ret = srp_new_cm_id(target);
587 	if (ret)
588 		goto err;
589 
590 	qp_attr.qp_state = IB_QPS_RESET;
591 	ret = ib_modify_qp(target->qp, &qp_attr, IB_QP_STATE);
592 	if (ret)
593 		goto err;
594 
595 	ret = srp_init_qp(target, target->qp);
596 	if (ret)
597 		goto err;
598 
599 	while (ib_poll_cq(target->recv_cq, 1, &wc) > 0)
600 		; /* nothing */
601 	while (ib_poll_cq(target->send_cq, 1, &wc) > 0)
602 		; /* nothing */
603 
604 	for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
605 		struct srp_request *req = &target->req_ring[i];
606 		if (req->scmnd)
607 			srp_reset_req(target, req);
608 	}
609 
610 	INIT_LIST_HEAD(&target->free_tx);
611 	for (i = 0; i < SRP_SQ_SIZE; ++i)
612 		list_add(&target->tx_ring[i]->list, &target->free_tx);
613 
614 	target->qp_in_error = 0;
615 	ret = srp_connect_target(target);
616 	if (ret)
617 		goto err;
618 
619 	if (!srp_change_state(target, SRP_TARGET_CONNECTING, SRP_TARGET_LIVE))
620 		ret = -EAGAIN;
621 
622 	return ret;
623 
624 err:
625 	shost_printk(KERN_ERR, target->scsi_host,
626 		     PFX "reconnect failed (%d), removing target port.\n", ret);
627 
628 	/*
629 	 * We couldn't reconnect, so kill our target port off.
630 	 * However, we have to defer the real removal because we
631 	 * are in the context of the SCSI error handler now, which
632 	 * will deadlock if we call scsi_remove_host().
633 	 *
634 	 * Schedule our work inside the lock to avoid a race with
635 	 * the flush_scheduled_work() in srp_remove_one().
636 	 */
637 	spin_lock_irq(&target->lock);
638 	if (target->state == SRP_TARGET_CONNECTING) {
639 		target->state = SRP_TARGET_DEAD;
640 		INIT_WORK(&target->work, srp_remove_work);
641 		queue_work(ib_wq, &target->work);
642 	}
643 	spin_unlock_irq(&target->lock);
644 
645 	return ret;
646 }
647 
648 static int srp_map_fmr(struct srp_target_port *target, struct scatterlist *scat,
649 		       int sg_cnt, struct srp_request *req,
650 		       struct srp_direct_buf *buf)
651 {
652 	u64 io_addr = 0;
653 	u64 *dma_pages;
654 	u32 len;
655 	int page_cnt;
656 	int i, j;
657 	int ret;
658 	struct srp_device *dev = target->srp_host->srp_dev;
659 	struct ib_device *ibdev = dev->dev;
660 	struct scatterlist *sg;
661 
662 	if (!dev->fmr_pool)
663 		return -ENODEV;
664 
665 	if (srp_target_is_mellanox(target) &&
666 	    (ib_sg_dma_address(ibdev, &scat[0]) & ~dev->fmr_page_mask))
667 		return -EINVAL;
668 
669 	len = page_cnt = 0;
670 	scsi_for_each_sg(req->scmnd, sg, sg_cnt, i) {
671 		unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
672 
673 		if (ib_sg_dma_address(ibdev, sg) & ~dev->fmr_page_mask) {
674 			if (i > 0)
675 				return -EINVAL;
676 			else
677 				++page_cnt;
678 		}
679 		if ((ib_sg_dma_address(ibdev, sg) + dma_len) &
680 		    ~dev->fmr_page_mask) {
681 			if (i < sg_cnt - 1)
682 				return -EINVAL;
683 			else
684 				++page_cnt;
685 		}
686 
687 		len += dma_len;
688 	}
689 
690 	page_cnt += len >> dev->fmr_page_shift;
691 	if (page_cnt > SRP_FMR_SIZE)
692 		return -ENOMEM;
693 
694 	dma_pages = kmalloc(sizeof (u64) * page_cnt, GFP_ATOMIC);
695 	if (!dma_pages)
696 		return -ENOMEM;
697 
698 	page_cnt = 0;
699 	scsi_for_each_sg(req->scmnd, sg, sg_cnt, i) {
700 		unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
701 
702 		for (j = 0; j < dma_len; j += dev->fmr_page_size)
703 			dma_pages[page_cnt++] =
704 				(ib_sg_dma_address(ibdev, sg) &
705 				 dev->fmr_page_mask) + j;
706 	}
707 
708 	req->fmr = ib_fmr_pool_map_phys(dev->fmr_pool,
709 					dma_pages, page_cnt, io_addr);
710 	if (IS_ERR(req->fmr)) {
711 		ret = PTR_ERR(req->fmr);
712 		req->fmr = NULL;
713 		goto out;
714 	}
715 
716 	buf->va  = cpu_to_be64(ib_sg_dma_address(ibdev, &scat[0]) &
717 			       ~dev->fmr_page_mask);
718 	buf->key = cpu_to_be32(req->fmr->fmr->rkey);
719 	buf->len = cpu_to_be32(len);
720 
721 	ret = 0;
722 
723 out:
724 	kfree(dma_pages);
725 
726 	return ret;
727 }
728 
729 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_target_port *target,
730 			struct srp_request *req)
731 {
732 	struct scatterlist *scat;
733 	struct srp_cmd *cmd = req->cmd->buf;
734 	int len, nents, count;
735 	u8 fmt = SRP_DATA_DESC_DIRECT;
736 	struct srp_device *dev;
737 	struct ib_device *ibdev;
738 
739 	if (!scsi_sglist(scmnd) || scmnd->sc_data_direction == DMA_NONE)
740 		return sizeof (struct srp_cmd);
741 
742 	if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
743 	    scmnd->sc_data_direction != DMA_TO_DEVICE) {
744 		shost_printk(KERN_WARNING, target->scsi_host,
745 			     PFX "Unhandled data direction %d\n",
746 			     scmnd->sc_data_direction);
747 		return -EINVAL;
748 	}
749 
750 	nents = scsi_sg_count(scmnd);
751 	scat  = scsi_sglist(scmnd);
752 
753 	dev = target->srp_host->srp_dev;
754 	ibdev = dev->dev;
755 
756 	count = ib_dma_map_sg(ibdev, scat, nents, scmnd->sc_data_direction);
757 
758 	fmt = SRP_DATA_DESC_DIRECT;
759 	len = sizeof (struct srp_cmd) +	sizeof (struct srp_direct_buf);
760 
761 	if (count == 1) {
762 		/*
763 		 * The midlayer only generated a single gather/scatter
764 		 * entry, or DMA mapping coalesced everything to a
765 		 * single entry.  So a direct descriptor along with
766 		 * the DMA MR suffices.
767 		 */
768 		struct srp_direct_buf *buf = (void *) cmd->add_data;
769 
770 		buf->va  = cpu_to_be64(ib_sg_dma_address(ibdev, scat));
771 		buf->key = cpu_to_be32(target->rkey);
772 		buf->len = cpu_to_be32(ib_sg_dma_len(ibdev, scat));
773 	} else if (srp_map_fmr(target, scat, count, req,
774 			       (void *) cmd->add_data)) {
775 		/*
776 		 * FMR mapping failed, and the scatterlist has more
777 		 * than one entry.  Generate an indirect memory
778 		 * descriptor.
779 		 */
780 		struct srp_indirect_buf *buf = (void *) cmd->add_data;
781 		struct scatterlist *sg;
782 		u32 datalen = 0;
783 		int i;
784 
785 		fmt = SRP_DATA_DESC_INDIRECT;
786 		len = sizeof (struct srp_cmd) +
787 			sizeof (struct srp_indirect_buf) +
788 			count * sizeof (struct srp_direct_buf);
789 
790 		scsi_for_each_sg(scmnd, sg, count, i) {
791 			unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
792 
793 			buf->desc_list[i].va  =
794 				cpu_to_be64(ib_sg_dma_address(ibdev, sg));
795 			buf->desc_list[i].key =
796 				cpu_to_be32(target->rkey);
797 			buf->desc_list[i].len = cpu_to_be32(dma_len);
798 			datalen += dma_len;
799 		}
800 
801 		if (scmnd->sc_data_direction == DMA_TO_DEVICE)
802 			cmd->data_out_desc_cnt = count;
803 		else
804 			cmd->data_in_desc_cnt = count;
805 
806 		buf->table_desc.va  =
807 			cpu_to_be64(req->cmd->dma + sizeof *cmd + sizeof *buf);
808 		buf->table_desc.key =
809 			cpu_to_be32(target->rkey);
810 		buf->table_desc.len =
811 			cpu_to_be32(count * sizeof (struct srp_direct_buf));
812 
813 		buf->len = cpu_to_be32(datalen);
814 	}
815 
816 	if (scmnd->sc_data_direction == DMA_TO_DEVICE)
817 		cmd->buf_fmt = fmt << 4;
818 	else
819 		cmd->buf_fmt = fmt;
820 
821 	return len;
822 }
823 
824 /*
825  * Return an IU and possible credit to the free pool
826  */
827 static void srp_put_tx_iu(struct srp_target_port *target, struct srp_iu *iu,
828 			  enum srp_iu_type iu_type)
829 {
830 	unsigned long flags;
831 
832 	spin_lock_irqsave(&target->lock, flags);
833 	list_add(&iu->list, &target->free_tx);
834 	if (iu_type != SRP_IU_RSP)
835 		++target->req_lim;
836 	spin_unlock_irqrestore(&target->lock, flags);
837 }
838 
839 /*
840  * Must be called with target->lock held to protect req_lim and free_tx.
841  * If IU is not sent, it must be returned using srp_put_tx_iu().
842  *
843  * Note:
844  * An upper limit for the number of allocated information units for each
845  * request type is:
846  * - SRP_IU_CMD: SRP_CMD_SQ_SIZE, since the SCSI mid-layer never queues
847  *   more than Scsi_Host.can_queue requests.
848  * - SRP_IU_TSK_MGMT: SRP_TSK_MGMT_SQ_SIZE.
849  * - SRP_IU_RSP: 1, since a conforming SRP target never sends more than
850  *   one unanswered SRP request to an initiator.
851  */
852 static struct srp_iu *__srp_get_tx_iu(struct srp_target_port *target,
853 				      enum srp_iu_type iu_type)
854 {
855 	s32 rsv = (iu_type == SRP_IU_TSK_MGMT) ? 0 : SRP_TSK_MGMT_SQ_SIZE;
856 	struct srp_iu *iu;
857 
858 	srp_send_completion(target->send_cq, target);
859 
860 	if (list_empty(&target->free_tx))
861 		return NULL;
862 
863 	/* Initiator responses to target requests do not consume credits */
864 	if (iu_type != SRP_IU_RSP) {
865 		if (target->req_lim <= rsv) {
866 			++target->zero_req_lim;
867 			return NULL;
868 		}
869 
870 		--target->req_lim;
871 	}
872 
873 	iu = list_first_entry(&target->free_tx, struct srp_iu, list);
874 	list_del(&iu->list);
875 	return iu;
876 }
877 
878 static int srp_post_send(struct srp_target_port *target,
879 			 struct srp_iu *iu, int len)
880 {
881 	struct ib_sge list;
882 	struct ib_send_wr wr, *bad_wr;
883 
884 	list.addr   = iu->dma;
885 	list.length = len;
886 	list.lkey   = target->lkey;
887 
888 	wr.next       = NULL;
889 	wr.wr_id      = (uintptr_t) iu;
890 	wr.sg_list    = &list;
891 	wr.num_sge    = 1;
892 	wr.opcode     = IB_WR_SEND;
893 	wr.send_flags = IB_SEND_SIGNALED;
894 
895 	return ib_post_send(target->qp, &wr, &bad_wr);
896 }
897 
898 static int srp_post_recv(struct srp_target_port *target, struct srp_iu *iu)
899 {
900 	struct ib_recv_wr wr, *bad_wr;
901 	struct ib_sge list;
902 
903 	list.addr   = iu->dma;
904 	list.length = iu->size;
905 	list.lkey   = target->lkey;
906 
907 	wr.next     = NULL;
908 	wr.wr_id    = (uintptr_t) iu;
909 	wr.sg_list  = &list;
910 	wr.num_sge  = 1;
911 
912 	return ib_post_recv(target->qp, &wr, &bad_wr);
913 }
914 
915 static void srp_process_rsp(struct srp_target_port *target, struct srp_rsp *rsp)
916 {
917 	struct srp_request *req;
918 	struct scsi_cmnd *scmnd;
919 	unsigned long flags;
920 
921 	if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
922 		spin_lock_irqsave(&target->lock, flags);
923 		target->req_lim += be32_to_cpu(rsp->req_lim_delta);
924 		spin_unlock_irqrestore(&target->lock, flags);
925 
926 		target->tsk_mgmt_status = -1;
927 		if (be32_to_cpu(rsp->resp_data_len) >= 4)
928 			target->tsk_mgmt_status = rsp->data[3];
929 		complete(&target->tsk_mgmt_done);
930 	} else {
931 		req = &target->req_ring[rsp->tag];
932 		scmnd = req->scmnd;
933 		if (!scmnd)
934 			shost_printk(KERN_ERR, target->scsi_host,
935 				     "Null scmnd for RSP w/tag %016llx\n",
936 				     (unsigned long long) rsp->tag);
937 		scmnd->result = rsp->status;
938 
939 		if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
940 			memcpy(scmnd->sense_buffer, rsp->data +
941 			       be32_to_cpu(rsp->resp_data_len),
942 			       min_t(int, be32_to_cpu(rsp->sense_data_len),
943 				     SCSI_SENSE_BUFFERSIZE));
944 		}
945 
946 		if (rsp->flags & (SRP_RSP_FLAG_DOOVER | SRP_RSP_FLAG_DOUNDER))
947 			scsi_set_resid(scmnd, be32_to_cpu(rsp->data_out_res_cnt));
948 		else if (rsp->flags & (SRP_RSP_FLAG_DIOVER | SRP_RSP_FLAG_DIUNDER))
949 			scsi_set_resid(scmnd, be32_to_cpu(rsp->data_in_res_cnt));
950 
951 		srp_remove_req(target, req, be32_to_cpu(rsp->req_lim_delta));
952 		scmnd->host_scribble = NULL;
953 		scmnd->scsi_done(scmnd);
954 	}
955 }
956 
957 static int srp_response_common(struct srp_target_port *target, s32 req_delta,
958 			       void *rsp, int len)
959 {
960 	struct ib_device *dev = target->srp_host->srp_dev->dev;
961 	unsigned long flags;
962 	struct srp_iu *iu;
963 	int err;
964 
965 	spin_lock_irqsave(&target->lock, flags);
966 	target->req_lim += req_delta;
967 	iu = __srp_get_tx_iu(target, SRP_IU_RSP);
968 	spin_unlock_irqrestore(&target->lock, flags);
969 
970 	if (!iu) {
971 		shost_printk(KERN_ERR, target->scsi_host, PFX
972 			     "no IU available to send response\n");
973 		return 1;
974 	}
975 
976 	ib_dma_sync_single_for_cpu(dev, iu->dma, len, DMA_TO_DEVICE);
977 	memcpy(iu->buf, rsp, len);
978 	ib_dma_sync_single_for_device(dev, iu->dma, len, DMA_TO_DEVICE);
979 
980 	err = srp_post_send(target, iu, len);
981 	if (err) {
982 		shost_printk(KERN_ERR, target->scsi_host, PFX
983 			     "unable to post response: %d\n", err);
984 		srp_put_tx_iu(target, iu, SRP_IU_RSP);
985 	}
986 
987 	return err;
988 }
989 
990 static void srp_process_cred_req(struct srp_target_port *target,
991 				 struct srp_cred_req *req)
992 {
993 	struct srp_cred_rsp rsp = {
994 		.opcode = SRP_CRED_RSP,
995 		.tag = req->tag,
996 	};
997 	s32 delta = be32_to_cpu(req->req_lim_delta);
998 
999 	if (srp_response_common(target, delta, &rsp, sizeof rsp))
1000 		shost_printk(KERN_ERR, target->scsi_host, PFX
1001 			     "problems processing SRP_CRED_REQ\n");
1002 }
1003 
1004 static void srp_process_aer_req(struct srp_target_port *target,
1005 				struct srp_aer_req *req)
1006 {
1007 	struct srp_aer_rsp rsp = {
1008 		.opcode = SRP_AER_RSP,
1009 		.tag = req->tag,
1010 	};
1011 	s32 delta = be32_to_cpu(req->req_lim_delta);
1012 
1013 	shost_printk(KERN_ERR, target->scsi_host, PFX
1014 		     "ignoring AER for LUN %llu\n", be64_to_cpu(req->lun));
1015 
1016 	if (srp_response_common(target, delta, &rsp, sizeof rsp))
1017 		shost_printk(KERN_ERR, target->scsi_host, PFX
1018 			     "problems processing SRP_AER_REQ\n");
1019 }
1020 
1021 static void srp_handle_recv(struct srp_target_port *target, struct ib_wc *wc)
1022 {
1023 	struct ib_device *dev = target->srp_host->srp_dev->dev;
1024 	struct srp_iu *iu = (struct srp_iu *) wc->wr_id;
1025 	int res;
1026 	u8 opcode;
1027 
1028 	ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_ti_iu_len,
1029 				   DMA_FROM_DEVICE);
1030 
1031 	opcode = *(u8 *) iu->buf;
1032 
1033 	if (0) {
1034 		shost_printk(KERN_ERR, target->scsi_host,
1035 			     PFX "recv completion, opcode 0x%02x\n", opcode);
1036 		print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 8, 1,
1037 			       iu->buf, wc->byte_len, true);
1038 	}
1039 
1040 	switch (opcode) {
1041 	case SRP_RSP:
1042 		srp_process_rsp(target, iu->buf);
1043 		break;
1044 
1045 	case SRP_CRED_REQ:
1046 		srp_process_cred_req(target, iu->buf);
1047 		break;
1048 
1049 	case SRP_AER_REQ:
1050 		srp_process_aer_req(target, iu->buf);
1051 		break;
1052 
1053 	case SRP_T_LOGOUT:
1054 		/* XXX Handle target logout */
1055 		shost_printk(KERN_WARNING, target->scsi_host,
1056 			     PFX "Got target logout request\n");
1057 		break;
1058 
1059 	default:
1060 		shost_printk(KERN_WARNING, target->scsi_host,
1061 			     PFX "Unhandled SRP opcode 0x%02x\n", opcode);
1062 		break;
1063 	}
1064 
1065 	ib_dma_sync_single_for_device(dev, iu->dma, target->max_ti_iu_len,
1066 				      DMA_FROM_DEVICE);
1067 
1068 	res = srp_post_recv(target, iu);
1069 	if (res != 0)
1070 		shost_printk(KERN_ERR, target->scsi_host,
1071 			     PFX "Recv failed with error code %d\n", res);
1072 }
1073 
1074 static void srp_recv_completion(struct ib_cq *cq, void *target_ptr)
1075 {
1076 	struct srp_target_port *target = target_ptr;
1077 	struct ib_wc wc;
1078 
1079 	ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1080 	while (ib_poll_cq(cq, 1, &wc) > 0) {
1081 		if (wc.status) {
1082 			shost_printk(KERN_ERR, target->scsi_host,
1083 				     PFX "failed receive status %d\n",
1084 				     wc.status);
1085 			target->qp_in_error = 1;
1086 			break;
1087 		}
1088 
1089 		srp_handle_recv(target, &wc);
1090 	}
1091 }
1092 
1093 static void srp_send_completion(struct ib_cq *cq, void *target_ptr)
1094 {
1095 	struct srp_target_port *target = target_ptr;
1096 	struct ib_wc wc;
1097 	struct srp_iu *iu;
1098 
1099 	while (ib_poll_cq(cq, 1, &wc) > 0) {
1100 		if (wc.status) {
1101 			shost_printk(KERN_ERR, target->scsi_host,
1102 				     PFX "failed send status %d\n",
1103 				     wc.status);
1104 			target->qp_in_error = 1;
1105 			break;
1106 		}
1107 
1108 		iu = (struct srp_iu *) wc.wr_id;
1109 		list_add(&iu->list, &target->free_tx);
1110 	}
1111 }
1112 
1113 static int srp_queuecommand(struct Scsi_Host *shost, struct scsi_cmnd *scmnd)
1114 {
1115 	struct srp_target_port *target = host_to_target(shost);
1116 	struct srp_request *req;
1117 	struct srp_iu *iu;
1118 	struct srp_cmd *cmd;
1119 	struct ib_device *dev;
1120 	unsigned long flags;
1121 	int len;
1122 
1123 	if (target->state == SRP_TARGET_CONNECTING)
1124 		goto err;
1125 
1126 	if (target->state == SRP_TARGET_DEAD ||
1127 	    target->state == SRP_TARGET_REMOVED) {
1128 		scmnd->result = DID_BAD_TARGET << 16;
1129 		scmnd->scsi_done(scmnd);
1130 		return 0;
1131 	}
1132 
1133 	spin_lock_irqsave(&target->lock, flags);
1134 	iu = __srp_get_tx_iu(target, SRP_IU_CMD);
1135 	if (!iu)
1136 		goto err_unlock;
1137 
1138 	req = list_first_entry(&target->free_reqs, struct srp_request, list);
1139 	list_del(&req->list);
1140 	spin_unlock_irqrestore(&target->lock, flags);
1141 
1142 	dev = target->srp_host->srp_dev->dev;
1143 	ib_dma_sync_single_for_cpu(dev, iu->dma, srp_max_iu_len,
1144 				   DMA_TO_DEVICE);
1145 
1146 	scmnd->result        = 0;
1147 	scmnd->host_scribble = (void *) req;
1148 
1149 	cmd = iu->buf;
1150 	memset(cmd, 0, sizeof *cmd);
1151 
1152 	cmd->opcode = SRP_CMD;
1153 	cmd->lun    = cpu_to_be64((u64) scmnd->device->lun << 48);
1154 	cmd->tag    = req->index;
1155 	memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
1156 
1157 	req->scmnd    = scmnd;
1158 	req->cmd      = iu;
1159 
1160 	len = srp_map_data(scmnd, target, req);
1161 	if (len < 0) {
1162 		shost_printk(KERN_ERR, target->scsi_host,
1163 			     PFX "Failed to map data\n");
1164 		goto err_iu;
1165 	}
1166 
1167 	ib_dma_sync_single_for_device(dev, iu->dma, srp_max_iu_len,
1168 				      DMA_TO_DEVICE);
1169 
1170 	if (srp_post_send(target, iu, len)) {
1171 		shost_printk(KERN_ERR, target->scsi_host, PFX "Send failed\n");
1172 		goto err_unmap;
1173 	}
1174 
1175 	return 0;
1176 
1177 err_unmap:
1178 	srp_unmap_data(scmnd, target, req);
1179 
1180 err_iu:
1181 	srp_put_tx_iu(target, iu, SRP_IU_CMD);
1182 
1183 	spin_lock_irqsave(&target->lock, flags);
1184 	list_add(&req->list, &target->free_reqs);
1185 
1186 err_unlock:
1187 	spin_unlock_irqrestore(&target->lock, flags);
1188 
1189 err:
1190 	return SCSI_MLQUEUE_HOST_BUSY;
1191 }
1192 
1193 static int srp_alloc_iu_bufs(struct srp_target_port *target)
1194 {
1195 	int i;
1196 
1197 	for (i = 0; i < SRP_RQ_SIZE; ++i) {
1198 		target->rx_ring[i] = srp_alloc_iu(target->srp_host,
1199 						  target->max_ti_iu_len,
1200 						  GFP_KERNEL, DMA_FROM_DEVICE);
1201 		if (!target->rx_ring[i])
1202 			goto err;
1203 	}
1204 
1205 	for (i = 0; i < SRP_SQ_SIZE; ++i) {
1206 		target->tx_ring[i] = srp_alloc_iu(target->srp_host,
1207 						  srp_max_iu_len,
1208 						  GFP_KERNEL, DMA_TO_DEVICE);
1209 		if (!target->tx_ring[i])
1210 			goto err;
1211 
1212 		list_add(&target->tx_ring[i]->list, &target->free_tx);
1213 	}
1214 
1215 	return 0;
1216 
1217 err:
1218 	for (i = 0; i < SRP_RQ_SIZE; ++i) {
1219 		srp_free_iu(target->srp_host, target->rx_ring[i]);
1220 		target->rx_ring[i] = NULL;
1221 	}
1222 
1223 	for (i = 0; i < SRP_SQ_SIZE; ++i) {
1224 		srp_free_iu(target->srp_host, target->tx_ring[i]);
1225 		target->tx_ring[i] = NULL;
1226 	}
1227 
1228 	return -ENOMEM;
1229 }
1230 
1231 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
1232 			       struct ib_cm_event *event,
1233 			       struct srp_target_port *target)
1234 {
1235 	struct Scsi_Host *shost = target->scsi_host;
1236 	struct ib_class_port_info *cpi;
1237 	int opcode;
1238 
1239 	switch (event->param.rej_rcvd.reason) {
1240 	case IB_CM_REJ_PORT_CM_REDIRECT:
1241 		cpi = event->param.rej_rcvd.ari;
1242 		target->path.dlid = cpi->redirect_lid;
1243 		target->path.pkey = cpi->redirect_pkey;
1244 		cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
1245 		memcpy(target->path.dgid.raw, cpi->redirect_gid, 16);
1246 
1247 		target->status = target->path.dlid ?
1248 			SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
1249 		break;
1250 
1251 	case IB_CM_REJ_PORT_REDIRECT:
1252 		if (srp_target_is_topspin(target)) {
1253 			/*
1254 			 * Topspin/Cisco SRP gateways incorrectly send
1255 			 * reject reason code 25 when they mean 24
1256 			 * (port redirect).
1257 			 */
1258 			memcpy(target->path.dgid.raw,
1259 			       event->param.rej_rcvd.ari, 16);
1260 
1261 			shost_printk(KERN_DEBUG, shost,
1262 				     PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
1263 				     (unsigned long long) be64_to_cpu(target->path.dgid.global.subnet_prefix),
1264 				     (unsigned long long) be64_to_cpu(target->path.dgid.global.interface_id));
1265 
1266 			target->status = SRP_PORT_REDIRECT;
1267 		} else {
1268 			shost_printk(KERN_WARNING, shost,
1269 				     "  REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
1270 			target->status = -ECONNRESET;
1271 		}
1272 		break;
1273 
1274 	case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
1275 		shost_printk(KERN_WARNING, shost,
1276 			    "  REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
1277 		target->status = -ECONNRESET;
1278 		break;
1279 
1280 	case IB_CM_REJ_CONSUMER_DEFINED:
1281 		opcode = *(u8 *) event->private_data;
1282 		if (opcode == SRP_LOGIN_REJ) {
1283 			struct srp_login_rej *rej = event->private_data;
1284 			u32 reason = be32_to_cpu(rej->reason);
1285 
1286 			if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
1287 				shost_printk(KERN_WARNING, shost,
1288 					     PFX "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
1289 			else
1290 				shost_printk(KERN_WARNING, shost,
1291 					    PFX "SRP LOGIN REJECTED, reason 0x%08x\n", reason);
1292 		} else
1293 			shost_printk(KERN_WARNING, shost,
1294 				     "  REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
1295 				     " opcode 0x%02x\n", opcode);
1296 		target->status = -ECONNRESET;
1297 		break;
1298 
1299 	case IB_CM_REJ_STALE_CONN:
1300 		shost_printk(KERN_WARNING, shost, "  REJ reason: stale connection\n");
1301 		target->status = SRP_STALE_CONN;
1302 		break;
1303 
1304 	default:
1305 		shost_printk(KERN_WARNING, shost, "  REJ reason 0x%x\n",
1306 			     event->param.rej_rcvd.reason);
1307 		target->status = -ECONNRESET;
1308 	}
1309 }
1310 
1311 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
1312 {
1313 	struct srp_target_port *target = cm_id->context;
1314 	struct ib_qp_attr *qp_attr = NULL;
1315 	int attr_mask = 0;
1316 	int comp = 0;
1317 	int opcode = 0;
1318 	int i;
1319 
1320 	switch (event->event) {
1321 	case IB_CM_REQ_ERROR:
1322 		shost_printk(KERN_DEBUG, target->scsi_host,
1323 			     PFX "Sending CM REQ failed\n");
1324 		comp = 1;
1325 		target->status = -ECONNRESET;
1326 		break;
1327 
1328 	case IB_CM_REP_RECEIVED:
1329 		comp = 1;
1330 		opcode = *(u8 *) event->private_data;
1331 
1332 		if (opcode == SRP_LOGIN_RSP) {
1333 			struct srp_login_rsp *rsp = event->private_data;
1334 
1335 			target->max_ti_iu_len = be32_to_cpu(rsp->max_ti_iu_len);
1336 			target->req_lim       = be32_to_cpu(rsp->req_lim_delta);
1337 
1338 			/*
1339 			 * Reserve credits for task management so we don't
1340 			 * bounce requests back to the SCSI mid-layer.
1341 			 */
1342 			target->scsi_host->can_queue
1343 				= min(target->req_lim - SRP_TSK_MGMT_SQ_SIZE,
1344 				      target->scsi_host->can_queue);
1345 		} else {
1346 			shost_printk(KERN_WARNING, target->scsi_host,
1347 				    PFX "Unhandled RSP opcode %#x\n", opcode);
1348 			target->status = -ECONNRESET;
1349 			break;
1350 		}
1351 
1352 		if (!target->rx_ring[0]) {
1353 			target->status = srp_alloc_iu_bufs(target);
1354 			if (target->status)
1355 				break;
1356 		}
1357 
1358 		qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
1359 		if (!qp_attr) {
1360 			target->status = -ENOMEM;
1361 			break;
1362 		}
1363 
1364 		qp_attr->qp_state = IB_QPS_RTR;
1365 		target->status = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1366 		if (target->status)
1367 			break;
1368 
1369 		target->status = ib_modify_qp(target->qp, qp_attr, attr_mask);
1370 		if (target->status)
1371 			break;
1372 
1373 		for (i = 0; i < SRP_RQ_SIZE; i++) {
1374 			struct srp_iu *iu = target->rx_ring[i];
1375 			target->status = srp_post_recv(target, iu);
1376 			if (target->status)
1377 				break;
1378 		}
1379 		if (target->status)
1380 			break;
1381 
1382 		qp_attr->qp_state = IB_QPS_RTS;
1383 		target->status = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1384 		if (target->status)
1385 			break;
1386 
1387 		target->status = ib_modify_qp(target->qp, qp_attr, attr_mask);
1388 		if (target->status)
1389 			break;
1390 
1391 		target->status = ib_send_cm_rtu(cm_id, NULL, 0);
1392 		if (target->status)
1393 			break;
1394 
1395 		break;
1396 
1397 	case IB_CM_REJ_RECEIVED:
1398 		shost_printk(KERN_DEBUG, target->scsi_host, PFX "REJ received\n");
1399 		comp = 1;
1400 
1401 		srp_cm_rej_handler(cm_id, event, target);
1402 		break;
1403 
1404 	case IB_CM_DREQ_RECEIVED:
1405 		shost_printk(KERN_WARNING, target->scsi_host,
1406 			     PFX "DREQ received - connection closed\n");
1407 		if (ib_send_cm_drep(cm_id, NULL, 0))
1408 			shost_printk(KERN_ERR, target->scsi_host,
1409 				     PFX "Sending CM DREP failed\n");
1410 		break;
1411 
1412 	case IB_CM_TIMEWAIT_EXIT:
1413 		shost_printk(KERN_ERR, target->scsi_host,
1414 			     PFX "connection closed\n");
1415 
1416 		comp = 1;
1417 		target->status = 0;
1418 		break;
1419 
1420 	case IB_CM_MRA_RECEIVED:
1421 	case IB_CM_DREQ_ERROR:
1422 	case IB_CM_DREP_RECEIVED:
1423 		break;
1424 
1425 	default:
1426 		shost_printk(KERN_WARNING, target->scsi_host,
1427 			     PFX "Unhandled CM event %d\n", event->event);
1428 		break;
1429 	}
1430 
1431 	if (comp)
1432 		complete(&target->done);
1433 
1434 	kfree(qp_attr);
1435 
1436 	return 0;
1437 }
1438 
1439 static int srp_send_tsk_mgmt(struct srp_target_port *target,
1440 			     u64 req_tag, unsigned int lun, u8 func)
1441 {
1442 	struct ib_device *dev = target->srp_host->srp_dev->dev;
1443 	struct srp_iu *iu;
1444 	struct srp_tsk_mgmt *tsk_mgmt;
1445 
1446 	if (target->state == SRP_TARGET_DEAD ||
1447 	    target->state == SRP_TARGET_REMOVED)
1448 		return -1;
1449 
1450 	init_completion(&target->tsk_mgmt_done);
1451 
1452 	spin_lock_irq(&target->lock);
1453 	iu = __srp_get_tx_iu(target, SRP_IU_TSK_MGMT);
1454 	spin_unlock_irq(&target->lock);
1455 
1456 	if (!iu)
1457 		return -1;
1458 
1459 	ib_dma_sync_single_for_cpu(dev, iu->dma, sizeof *tsk_mgmt,
1460 				   DMA_TO_DEVICE);
1461 	tsk_mgmt = iu->buf;
1462 	memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
1463 
1464 	tsk_mgmt->opcode 	= SRP_TSK_MGMT;
1465 	tsk_mgmt->lun		= cpu_to_be64((u64) lun << 48);
1466 	tsk_mgmt->tag		= req_tag | SRP_TAG_TSK_MGMT;
1467 	tsk_mgmt->tsk_mgmt_func = func;
1468 	tsk_mgmt->task_tag	= req_tag;
1469 
1470 	ib_dma_sync_single_for_device(dev, iu->dma, sizeof *tsk_mgmt,
1471 				      DMA_TO_DEVICE);
1472 	if (srp_post_send(target, iu, sizeof *tsk_mgmt)) {
1473 		srp_put_tx_iu(target, iu, SRP_IU_TSK_MGMT);
1474 		return -1;
1475 	}
1476 
1477 	if (!wait_for_completion_timeout(&target->tsk_mgmt_done,
1478 					 msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
1479 		return -1;
1480 
1481 	return 0;
1482 }
1483 
1484 static int srp_abort(struct scsi_cmnd *scmnd)
1485 {
1486 	struct srp_target_port *target = host_to_target(scmnd->device->host);
1487 	struct srp_request *req = (struct srp_request *) scmnd->host_scribble;
1488 	int ret = SUCCESS;
1489 
1490 	shost_printk(KERN_ERR, target->scsi_host, "SRP abort called\n");
1491 
1492 	if (!req || target->qp_in_error)
1493 		return FAILED;
1494 	if (srp_send_tsk_mgmt(target, req->index, scmnd->device->lun,
1495 			      SRP_TSK_ABORT_TASK))
1496 		return FAILED;
1497 
1498 	if (req->scmnd) {
1499 		if (!target->tsk_mgmt_status) {
1500 			srp_remove_req(target, req, 0);
1501 			scmnd->result = DID_ABORT << 16;
1502 		} else
1503 			ret = FAILED;
1504 	}
1505 
1506 	return ret;
1507 }
1508 
1509 static int srp_reset_device(struct scsi_cmnd *scmnd)
1510 {
1511 	struct srp_target_port *target = host_to_target(scmnd->device->host);
1512 	int i;
1513 
1514 	shost_printk(KERN_ERR, target->scsi_host, "SRP reset_device called\n");
1515 
1516 	if (target->qp_in_error)
1517 		return FAILED;
1518 	if (srp_send_tsk_mgmt(target, SRP_TAG_NO_REQ, scmnd->device->lun,
1519 			      SRP_TSK_LUN_RESET))
1520 		return FAILED;
1521 	if (target->tsk_mgmt_status)
1522 		return FAILED;
1523 
1524 	for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
1525 		struct srp_request *req = &target->req_ring[i];
1526 		if (req->scmnd && req->scmnd->device == scmnd->device)
1527 			srp_reset_req(target, req);
1528 	}
1529 
1530 	return SUCCESS;
1531 }
1532 
1533 static int srp_reset_host(struct scsi_cmnd *scmnd)
1534 {
1535 	struct srp_target_port *target = host_to_target(scmnd->device->host);
1536 	int ret = FAILED;
1537 
1538 	shost_printk(KERN_ERR, target->scsi_host, PFX "SRP reset_host called\n");
1539 
1540 	if (!srp_reconnect_target(target))
1541 		ret = SUCCESS;
1542 
1543 	return ret;
1544 }
1545 
1546 static ssize_t show_id_ext(struct device *dev, struct device_attribute *attr,
1547 			   char *buf)
1548 {
1549 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
1550 
1551 	if (target->state == SRP_TARGET_DEAD ||
1552 	    target->state == SRP_TARGET_REMOVED)
1553 		return -ENODEV;
1554 
1555 	return sprintf(buf, "0x%016llx\n",
1556 		       (unsigned long long) be64_to_cpu(target->id_ext));
1557 }
1558 
1559 static ssize_t show_ioc_guid(struct device *dev, struct device_attribute *attr,
1560 			     char *buf)
1561 {
1562 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
1563 
1564 	if (target->state == SRP_TARGET_DEAD ||
1565 	    target->state == SRP_TARGET_REMOVED)
1566 		return -ENODEV;
1567 
1568 	return sprintf(buf, "0x%016llx\n",
1569 		       (unsigned long long) be64_to_cpu(target->ioc_guid));
1570 }
1571 
1572 static ssize_t show_service_id(struct device *dev,
1573 			       struct device_attribute *attr, char *buf)
1574 {
1575 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
1576 
1577 	if (target->state == SRP_TARGET_DEAD ||
1578 	    target->state == SRP_TARGET_REMOVED)
1579 		return -ENODEV;
1580 
1581 	return sprintf(buf, "0x%016llx\n",
1582 		       (unsigned long long) be64_to_cpu(target->service_id));
1583 }
1584 
1585 static ssize_t show_pkey(struct device *dev, struct device_attribute *attr,
1586 			 char *buf)
1587 {
1588 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
1589 
1590 	if (target->state == SRP_TARGET_DEAD ||
1591 	    target->state == SRP_TARGET_REMOVED)
1592 		return -ENODEV;
1593 
1594 	return sprintf(buf, "0x%04x\n", be16_to_cpu(target->path.pkey));
1595 }
1596 
1597 static ssize_t show_dgid(struct device *dev, struct device_attribute *attr,
1598 			 char *buf)
1599 {
1600 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
1601 
1602 	if (target->state == SRP_TARGET_DEAD ||
1603 	    target->state == SRP_TARGET_REMOVED)
1604 		return -ENODEV;
1605 
1606 	return sprintf(buf, "%pI6\n", target->path.dgid.raw);
1607 }
1608 
1609 static ssize_t show_orig_dgid(struct device *dev,
1610 			      struct device_attribute *attr, char *buf)
1611 {
1612 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
1613 
1614 	if (target->state == SRP_TARGET_DEAD ||
1615 	    target->state == SRP_TARGET_REMOVED)
1616 		return -ENODEV;
1617 
1618 	return sprintf(buf, "%pI6\n", target->orig_dgid);
1619 }
1620 
1621 static ssize_t show_req_lim(struct device *dev,
1622 			    struct device_attribute *attr, char *buf)
1623 {
1624 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
1625 
1626 	if (target->state == SRP_TARGET_DEAD ||
1627 	    target->state == SRP_TARGET_REMOVED)
1628 		return -ENODEV;
1629 
1630 	return sprintf(buf, "%d\n", target->req_lim);
1631 }
1632 
1633 static ssize_t show_zero_req_lim(struct device *dev,
1634 				 struct device_attribute *attr, char *buf)
1635 {
1636 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
1637 
1638 	if (target->state == SRP_TARGET_DEAD ||
1639 	    target->state == SRP_TARGET_REMOVED)
1640 		return -ENODEV;
1641 
1642 	return sprintf(buf, "%d\n", target->zero_req_lim);
1643 }
1644 
1645 static ssize_t show_local_ib_port(struct device *dev,
1646 				  struct device_attribute *attr, char *buf)
1647 {
1648 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
1649 
1650 	return sprintf(buf, "%d\n", target->srp_host->port);
1651 }
1652 
1653 static ssize_t show_local_ib_device(struct device *dev,
1654 				    struct device_attribute *attr, char *buf)
1655 {
1656 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
1657 
1658 	return sprintf(buf, "%s\n", target->srp_host->srp_dev->dev->name);
1659 }
1660 
1661 static DEVICE_ATTR(id_ext,	    S_IRUGO, show_id_ext,	   NULL);
1662 static DEVICE_ATTR(ioc_guid,	    S_IRUGO, show_ioc_guid,	   NULL);
1663 static DEVICE_ATTR(service_id,	    S_IRUGO, show_service_id,	   NULL);
1664 static DEVICE_ATTR(pkey,	    S_IRUGO, show_pkey,		   NULL);
1665 static DEVICE_ATTR(dgid,	    S_IRUGO, show_dgid,		   NULL);
1666 static DEVICE_ATTR(orig_dgid,	    S_IRUGO, show_orig_dgid,	   NULL);
1667 static DEVICE_ATTR(req_lim,         S_IRUGO, show_req_lim,         NULL);
1668 static DEVICE_ATTR(zero_req_lim,    S_IRUGO, show_zero_req_lim,	   NULL);
1669 static DEVICE_ATTR(local_ib_port,   S_IRUGO, show_local_ib_port,   NULL);
1670 static DEVICE_ATTR(local_ib_device, S_IRUGO, show_local_ib_device, NULL);
1671 
1672 static struct device_attribute *srp_host_attrs[] = {
1673 	&dev_attr_id_ext,
1674 	&dev_attr_ioc_guid,
1675 	&dev_attr_service_id,
1676 	&dev_attr_pkey,
1677 	&dev_attr_dgid,
1678 	&dev_attr_orig_dgid,
1679 	&dev_attr_req_lim,
1680 	&dev_attr_zero_req_lim,
1681 	&dev_attr_local_ib_port,
1682 	&dev_attr_local_ib_device,
1683 	NULL
1684 };
1685 
1686 static struct scsi_host_template srp_template = {
1687 	.module				= THIS_MODULE,
1688 	.name				= "InfiniBand SRP initiator",
1689 	.proc_name			= DRV_NAME,
1690 	.info				= srp_target_info,
1691 	.queuecommand			= srp_queuecommand,
1692 	.eh_abort_handler		= srp_abort,
1693 	.eh_device_reset_handler	= srp_reset_device,
1694 	.eh_host_reset_handler		= srp_reset_host,
1695 	.can_queue			= SRP_CMD_SQ_SIZE,
1696 	.this_id			= -1,
1697 	.cmd_per_lun			= SRP_CMD_SQ_SIZE,
1698 	.use_clustering			= ENABLE_CLUSTERING,
1699 	.shost_attrs			= srp_host_attrs
1700 };
1701 
1702 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
1703 {
1704 	struct srp_rport_identifiers ids;
1705 	struct srp_rport *rport;
1706 
1707 	sprintf(target->target_name, "SRP.T10:%016llX",
1708 		 (unsigned long long) be64_to_cpu(target->id_ext));
1709 
1710 	if (scsi_add_host(target->scsi_host, host->srp_dev->dev->dma_device))
1711 		return -ENODEV;
1712 
1713 	memcpy(ids.port_id, &target->id_ext, 8);
1714 	memcpy(ids.port_id + 8, &target->ioc_guid, 8);
1715 	ids.roles = SRP_RPORT_ROLE_TARGET;
1716 	rport = srp_rport_add(target->scsi_host, &ids);
1717 	if (IS_ERR(rport)) {
1718 		scsi_remove_host(target->scsi_host);
1719 		return PTR_ERR(rport);
1720 	}
1721 
1722 	spin_lock(&host->target_lock);
1723 	list_add_tail(&target->list, &host->target_list);
1724 	spin_unlock(&host->target_lock);
1725 
1726 	target->state = SRP_TARGET_LIVE;
1727 
1728 	scsi_scan_target(&target->scsi_host->shost_gendev,
1729 			 0, target->scsi_id, SCAN_WILD_CARD, 0);
1730 
1731 	return 0;
1732 }
1733 
1734 static void srp_release_dev(struct device *dev)
1735 {
1736 	struct srp_host *host =
1737 		container_of(dev, struct srp_host, dev);
1738 
1739 	complete(&host->released);
1740 }
1741 
1742 static struct class srp_class = {
1743 	.name    = "infiniband_srp",
1744 	.dev_release = srp_release_dev
1745 };
1746 
1747 /*
1748  * Target ports are added by writing
1749  *
1750  *     id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
1751  *     pkey=<P_Key>,service_id=<service ID>
1752  *
1753  * to the add_target sysfs attribute.
1754  */
1755 enum {
1756 	SRP_OPT_ERR		= 0,
1757 	SRP_OPT_ID_EXT		= 1 << 0,
1758 	SRP_OPT_IOC_GUID	= 1 << 1,
1759 	SRP_OPT_DGID		= 1 << 2,
1760 	SRP_OPT_PKEY		= 1 << 3,
1761 	SRP_OPT_SERVICE_ID	= 1 << 4,
1762 	SRP_OPT_MAX_SECT	= 1 << 5,
1763 	SRP_OPT_MAX_CMD_PER_LUN	= 1 << 6,
1764 	SRP_OPT_IO_CLASS	= 1 << 7,
1765 	SRP_OPT_INITIATOR_EXT	= 1 << 8,
1766 	SRP_OPT_ALL		= (SRP_OPT_ID_EXT	|
1767 				   SRP_OPT_IOC_GUID	|
1768 				   SRP_OPT_DGID		|
1769 				   SRP_OPT_PKEY		|
1770 				   SRP_OPT_SERVICE_ID),
1771 };
1772 
1773 static const match_table_t srp_opt_tokens = {
1774 	{ SRP_OPT_ID_EXT,		"id_ext=%s" 		},
1775 	{ SRP_OPT_IOC_GUID,		"ioc_guid=%s" 		},
1776 	{ SRP_OPT_DGID,			"dgid=%s" 		},
1777 	{ SRP_OPT_PKEY,			"pkey=%x" 		},
1778 	{ SRP_OPT_SERVICE_ID,		"service_id=%s"		},
1779 	{ SRP_OPT_MAX_SECT,		"max_sect=%d" 		},
1780 	{ SRP_OPT_MAX_CMD_PER_LUN,	"max_cmd_per_lun=%d" 	},
1781 	{ SRP_OPT_IO_CLASS,		"io_class=%x"		},
1782 	{ SRP_OPT_INITIATOR_EXT,	"initiator_ext=%s"	},
1783 	{ SRP_OPT_ERR,			NULL 			}
1784 };
1785 
1786 static int srp_parse_options(const char *buf, struct srp_target_port *target)
1787 {
1788 	char *options, *sep_opt;
1789 	char *p;
1790 	char dgid[3];
1791 	substring_t args[MAX_OPT_ARGS];
1792 	int opt_mask = 0;
1793 	int token;
1794 	int ret = -EINVAL;
1795 	int i;
1796 
1797 	options = kstrdup(buf, GFP_KERNEL);
1798 	if (!options)
1799 		return -ENOMEM;
1800 
1801 	sep_opt = options;
1802 	while ((p = strsep(&sep_opt, ",")) != NULL) {
1803 		if (!*p)
1804 			continue;
1805 
1806 		token = match_token(p, srp_opt_tokens, args);
1807 		opt_mask |= token;
1808 
1809 		switch (token) {
1810 		case SRP_OPT_ID_EXT:
1811 			p = match_strdup(args);
1812 			if (!p) {
1813 				ret = -ENOMEM;
1814 				goto out;
1815 			}
1816 			target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
1817 			kfree(p);
1818 			break;
1819 
1820 		case SRP_OPT_IOC_GUID:
1821 			p = match_strdup(args);
1822 			if (!p) {
1823 				ret = -ENOMEM;
1824 				goto out;
1825 			}
1826 			target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
1827 			kfree(p);
1828 			break;
1829 
1830 		case SRP_OPT_DGID:
1831 			p = match_strdup(args);
1832 			if (!p) {
1833 				ret = -ENOMEM;
1834 				goto out;
1835 			}
1836 			if (strlen(p) != 32) {
1837 				printk(KERN_WARNING PFX "bad dest GID parameter '%s'\n", p);
1838 				kfree(p);
1839 				goto out;
1840 			}
1841 
1842 			for (i = 0; i < 16; ++i) {
1843 				strlcpy(dgid, p + i * 2, 3);
1844 				target->path.dgid.raw[i] = simple_strtoul(dgid, NULL, 16);
1845 			}
1846 			kfree(p);
1847 			memcpy(target->orig_dgid, target->path.dgid.raw, 16);
1848 			break;
1849 
1850 		case SRP_OPT_PKEY:
1851 			if (match_hex(args, &token)) {
1852 				printk(KERN_WARNING PFX "bad P_Key parameter '%s'\n", p);
1853 				goto out;
1854 			}
1855 			target->path.pkey = cpu_to_be16(token);
1856 			break;
1857 
1858 		case SRP_OPT_SERVICE_ID:
1859 			p = match_strdup(args);
1860 			if (!p) {
1861 				ret = -ENOMEM;
1862 				goto out;
1863 			}
1864 			target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
1865 			target->path.service_id = target->service_id;
1866 			kfree(p);
1867 			break;
1868 
1869 		case SRP_OPT_MAX_SECT:
1870 			if (match_int(args, &token)) {
1871 				printk(KERN_WARNING PFX "bad max sect parameter '%s'\n", p);
1872 				goto out;
1873 			}
1874 			target->scsi_host->max_sectors = token;
1875 			break;
1876 
1877 		case SRP_OPT_MAX_CMD_PER_LUN:
1878 			if (match_int(args, &token)) {
1879 				printk(KERN_WARNING PFX "bad max cmd_per_lun parameter '%s'\n", p);
1880 				goto out;
1881 			}
1882 			target->scsi_host->cmd_per_lun = min(token, SRP_CMD_SQ_SIZE);
1883 			break;
1884 
1885 		case SRP_OPT_IO_CLASS:
1886 			if (match_hex(args, &token)) {
1887 				printk(KERN_WARNING PFX "bad  IO class parameter '%s' \n", p);
1888 				goto out;
1889 			}
1890 			if (token != SRP_REV10_IB_IO_CLASS &&
1891 			    token != SRP_REV16A_IB_IO_CLASS) {
1892 				printk(KERN_WARNING PFX "unknown IO class parameter value"
1893 				       " %x specified (use %x or %x).\n",
1894 				       token, SRP_REV10_IB_IO_CLASS, SRP_REV16A_IB_IO_CLASS);
1895 				goto out;
1896 			}
1897 			target->io_class = token;
1898 			break;
1899 
1900 		case SRP_OPT_INITIATOR_EXT:
1901 			p = match_strdup(args);
1902 			if (!p) {
1903 				ret = -ENOMEM;
1904 				goto out;
1905 			}
1906 			target->initiator_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
1907 			kfree(p);
1908 			break;
1909 
1910 		default:
1911 			printk(KERN_WARNING PFX "unknown parameter or missing value "
1912 			       "'%s' in target creation request\n", p);
1913 			goto out;
1914 		}
1915 	}
1916 
1917 	if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
1918 		ret = 0;
1919 	else
1920 		for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
1921 			if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
1922 			    !(srp_opt_tokens[i].token & opt_mask))
1923 				printk(KERN_WARNING PFX "target creation request is "
1924 				       "missing parameter '%s'\n",
1925 				       srp_opt_tokens[i].pattern);
1926 
1927 out:
1928 	kfree(options);
1929 	return ret;
1930 }
1931 
1932 static ssize_t srp_create_target(struct device *dev,
1933 				 struct device_attribute *attr,
1934 				 const char *buf, size_t count)
1935 {
1936 	struct srp_host *host =
1937 		container_of(dev, struct srp_host, dev);
1938 	struct Scsi_Host *target_host;
1939 	struct srp_target_port *target;
1940 	int ret;
1941 	int i;
1942 
1943 	target_host = scsi_host_alloc(&srp_template,
1944 				      sizeof (struct srp_target_port));
1945 	if (!target_host)
1946 		return -ENOMEM;
1947 
1948 	target_host->transportt = ib_srp_transport_template;
1949 	target_host->max_lun     = SRP_MAX_LUN;
1950 	target_host->max_cmd_len = sizeof ((struct srp_cmd *) (void *) 0L)->cdb;
1951 
1952 	target = host_to_target(target_host);
1953 
1954 	target->io_class   = SRP_REV16A_IB_IO_CLASS;
1955 	target->scsi_host  = target_host;
1956 	target->srp_host   = host;
1957 	target->lkey	   = host->srp_dev->mr->lkey;
1958 	target->rkey	   = host->srp_dev->mr->rkey;
1959 
1960 	spin_lock_init(&target->lock);
1961 	INIT_LIST_HEAD(&target->free_tx);
1962 	INIT_LIST_HEAD(&target->free_reqs);
1963 	for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
1964 		target->req_ring[i].index = i;
1965 		list_add_tail(&target->req_ring[i].list, &target->free_reqs);
1966 	}
1967 
1968 	ret = srp_parse_options(buf, target);
1969 	if (ret)
1970 		goto err;
1971 
1972 	ib_query_gid(host->srp_dev->dev, host->port, 0, &target->path.sgid);
1973 
1974 	shost_printk(KERN_DEBUG, target->scsi_host, PFX
1975 		     "new target: id_ext %016llx ioc_guid %016llx pkey %04x "
1976 		     "service_id %016llx dgid %pI6\n",
1977 	       (unsigned long long) be64_to_cpu(target->id_ext),
1978 	       (unsigned long long) be64_to_cpu(target->ioc_guid),
1979 	       be16_to_cpu(target->path.pkey),
1980 	       (unsigned long long) be64_to_cpu(target->service_id),
1981 	       target->path.dgid.raw);
1982 
1983 	ret = srp_create_target_ib(target);
1984 	if (ret)
1985 		goto err;
1986 
1987 	ret = srp_new_cm_id(target);
1988 	if (ret)
1989 		goto err_free;
1990 
1991 	target->qp_in_error = 0;
1992 	ret = srp_connect_target(target);
1993 	if (ret) {
1994 		shost_printk(KERN_ERR, target->scsi_host,
1995 			     PFX "Connection failed\n");
1996 		goto err_cm_id;
1997 	}
1998 
1999 	ret = srp_add_target(host, target);
2000 	if (ret)
2001 		goto err_disconnect;
2002 
2003 	return count;
2004 
2005 err_disconnect:
2006 	srp_disconnect_target(target);
2007 
2008 err_cm_id:
2009 	ib_destroy_cm_id(target->cm_id);
2010 
2011 err_free:
2012 	srp_free_target_ib(target);
2013 
2014 err:
2015 	scsi_host_put(target_host);
2016 
2017 	return ret;
2018 }
2019 
2020 static DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
2021 
2022 static ssize_t show_ibdev(struct device *dev, struct device_attribute *attr,
2023 			  char *buf)
2024 {
2025 	struct srp_host *host = container_of(dev, struct srp_host, dev);
2026 
2027 	return sprintf(buf, "%s\n", host->srp_dev->dev->name);
2028 }
2029 
2030 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
2031 
2032 static ssize_t show_port(struct device *dev, struct device_attribute *attr,
2033 			 char *buf)
2034 {
2035 	struct srp_host *host = container_of(dev, struct srp_host, dev);
2036 
2037 	return sprintf(buf, "%d\n", host->port);
2038 }
2039 
2040 static DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
2041 
2042 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
2043 {
2044 	struct srp_host *host;
2045 
2046 	host = kzalloc(sizeof *host, GFP_KERNEL);
2047 	if (!host)
2048 		return NULL;
2049 
2050 	INIT_LIST_HEAD(&host->target_list);
2051 	spin_lock_init(&host->target_lock);
2052 	init_completion(&host->released);
2053 	host->srp_dev = device;
2054 	host->port = port;
2055 
2056 	host->dev.class = &srp_class;
2057 	host->dev.parent = device->dev->dma_device;
2058 	dev_set_name(&host->dev, "srp-%s-%d", device->dev->name, port);
2059 
2060 	if (device_register(&host->dev))
2061 		goto free_host;
2062 	if (device_create_file(&host->dev, &dev_attr_add_target))
2063 		goto err_class;
2064 	if (device_create_file(&host->dev, &dev_attr_ibdev))
2065 		goto err_class;
2066 	if (device_create_file(&host->dev, &dev_attr_port))
2067 		goto err_class;
2068 
2069 	return host;
2070 
2071 err_class:
2072 	device_unregister(&host->dev);
2073 
2074 free_host:
2075 	kfree(host);
2076 
2077 	return NULL;
2078 }
2079 
2080 static void srp_add_one(struct ib_device *device)
2081 {
2082 	struct srp_device *srp_dev;
2083 	struct ib_device_attr *dev_attr;
2084 	struct ib_fmr_pool_param fmr_param;
2085 	struct srp_host *host;
2086 	int s, e, p;
2087 
2088 	dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
2089 	if (!dev_attr)
2090 		return;
2091 
2092 	if (ib_query_device(device, dev_attr)) {
2093 		printk(KERN_WARNING PFX "Query device failed for %s\n",
2094 		       device->name);
2095 		goto free_attr;
2096 	}
2097 
2098 	srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
2099 	if (!srp_dev)
2100 		goto free_attr;
2101 
2102 	/*
2103 	 * Use the smallest page size supported by the HCA, down to a
2104 	 * minimum of 512 bytes (which is the smallest sector that a
2105 	 * SCSI command will ever carry).
2106 	 */
2107 	srp_dev->fmr_page_shift = max(9, ffs(dev_attr->page_size_cap) - 1);
2108 	srp_dev->fmr_page_size  = 1 << srp_dev->fmr_page_shift;
2109 	srp_dev->fmr_page_mask  = ~((u64) srp_dev->fmr_page_size - 1);
2110 
2111 	INIT_LIST_HEAD(&srp_dev->dev_list);
2112 
2113 	srp_dev->dev = device;
2114 	srp_dev->pd  = ib_alloc_pd(device);
2115 	if (IS_ERR(srp_dev->pd))
2116 		goto free_dev;
2117 
2118 	srp_dev->mr = ib_get_dma_mr(srp_dev->pd,
2119 				    IB_ACCESS_LOCAL_WRITE |
2120 				    IB_ACCESS_REMOTE_READ |
2121 				    IB_ACCESS_REMOTE_WRITE);
2122 	if (IS_ERR(srp_dev->mr))
2123 		goto err_pd;
2124 
2125 	memset(&fmr_param, 0, sizeof fmr_param);
2126 	fmr_param.pool_size	    = SRP_FMR_POOL_SIZE;
2127 	fmr_param.dirty_watermark   = SRP_FMR_DIRTY_SIZE;
2128 	fmr_param.cache		    = 1;
2129 	fmr_param.max_pages_per_fmr = SRP_FMR_SIZE;
2130 	fmr_param.page_shift	    = srp_dev->fmr_page_shift;
2131 	fmr_param.access	    = (IB_ACCESS_LOCAL_WRITE |
2132 				       IB_ACCESS_REMOTE_WRITE |
2133 				       IB_ACCESS_REMOTE_READ);
2134 
2135 	srp_dev->fmr_pool = ib_create_fmr_pool(srp_dev->pd, &fmr_param);
2136 	if (IS_ERR(srp_dev->fmr_pool))
2137 		srp_dev->fmr_pool = NULL;
2138 
2139 	if (device->node_type == RDMA_NODE_IB_SWITCH) {
2140 		s = 0;
2141 		e = 0;
2142 	} else {
2143 		s = 1;
2144 		e = device->phys_port_cnt;
2145 	}
2146 
2147 	for (p = s; p <= e; ++p) {
2148 		host = srp_add_port(srp_dev, p);
2149 		if (host)
2150 			list_add_tail(&host->list, &srp_dev->dev_list);
2151 	}
2152 
2153 	ib_set_client_data(device, &srp_client, srp_dev);
2154 
2155 	goto free_attr;
2156 
2157 err_pd:
2158 	ib_dealloc_pd(srp_dev->pd);
2159 
2160 free_dev:
2161 	kfree(srp_dev);
2162 
2163 free_attr:
2164 	kfree(dev_attr);
2165 }
2166 
2167 static void srp_remove_one(struct ib_device *device)
2168 {
2169 	struct srp_device *srp_dev;
2170 	struct srp_host *host, *tmp_host;
2171 	LIST_HEAD(target_list);
2172 	struct srp_target_port *target, *tmp_target;
2173 
2174 	srp_dev = ib_get_client_data(device, &srp_client);
2175 
2176 	list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
2177 		device_unregister(&host->dev);
2178 		/*
2179 		 * Wait for the sysfs entry to go away, so that no new
2180 		 * target ports can be created.
2181 		 */
2182 		wait_for_completion(&host->released);
2183 
2184 		/*
2185 		 * Mark all target ports as removed, so we stop queueing
2186 		 * commands and don't try to reconnect.
2187 		 */
2188 		spin_lock(&host->target_lock);
2189 		list_for_each_entry(target, &host->target_list, list) {
2190 			spin_lock_irq(&target->lock);
2191 			target->state = SRP_TARGET_REMOVED;
2192 			spin_unlock_irq(&target->lock);
2193 		}
2194 		spin_unlock(&host->target_lock);
2195 
2196 		/*
2197 		 * Wait for any reconnection tasks that may have
2198 		 * started before we marked our target ports as
2199 		 * removed, and any target port removal tasks.
2200 		 */
2201 		flush_workqueue(ib_wq);
2202 
2203 		list_for_each_entry_safe(target, tmp_target,
2204 					 &host->target_list, list) {
2205 			srp_remove_host(target->scsi_host);
2206 			scsi_remove_host(target->scsi_host);
2207 			srp_disconnect_target(target);
2208 			ib_destroy_cm_id(target->cm_id);
2209 			srp_free_target_ib(target);
2210 			scsi_host_put(target->scsi_host);
2211 		}
2212 
2213 		kfree(host);
2214 	}
2215 
2216 	if (srp_dev->fmr_pool)
2217 		ib_destroy_fmr_pool(srp_dev->fmr_pool);
2218 	ib_dereg_mr(srp_dev->mr);
2219 	ib_dealloc_pd(srp_dev->pd);
2220 
2221 	kfree(srp_dev);
2222 }
2223 
2224 static struct srp_function_template ib_srp_transport_functions = {
2225 };
2226 
2227 static int __init srp_init_module(void)
2228 {
2229 	int ret;
2230 
2231 	BUILD_BUG_ON(FIELD_SIZEOF(struct ib_wc, wr_id) < sizeof(void *));
2232 
2233 	if (srp_sg_tablesize > 255) {
2234 		printk(KERN_WARNING PFX "Clamping srp_sg_tablesize to 255\n");
2235 		srp_sg_tablesize = 255;
2236 	}
2237 
2238 	ib_srp_transport_template =
2239 		srp_attach_transport(&ib_srp_transport_functions);
2240 	if (!ib_srp_transport_template)
2241 		return -ENOMEM;
2242 
2243 	srp_template.sg_tablesize = srp_sg_tablesize;
2244 	srp_max_iu_len = (sizeof (struct srp_cmd) +
2245 			  sizeof (struct srp_indirect_buf) +
2246 			  srp_sg_tablesize * 16);
2247 
2248 	ret = class_register(&srp_class);
2249 	if (ret) {
2250 		printk(KERN_ERR PFX "couldn't register class infiniband_srp\n");
2251 		srp_release_transport(ib_srp_transport_template);
2252 		return ret;
2253 	}
2254 
2255 	ib_sa_register_client(&srp_sa_client);
2256 
2257 	ret = ib_register_client(&srp_client);
2258 	if (ret) {
2259 		printk(KERN_ERR PFX "couldn't register IB client\n");
2260 		srp_release_transport(ib_srp_transport_template);
2261 		ib_sa_unregister_client(&srp_sa_client);
2262 		class_unregister(&srp_class);
2263 		return ret;
2264 	}
2265 
2266 	return 0;
2267 }
2268 
2269 static void __exit srp_cleanup_module(void)
2270 {
2271 	ib_unregister_client(&srp_client);
2272 	ib_sa_unregister_client(&srp_sa_client);
2273 	class_unregister(&srp_class);
2274 	srp_release_transport(ib_srp_transport_template);
2275 }
2276 
2277 module_init(srp_init_module);
2278 module_exit(srp_cleanup_module);
2279