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  * $Id: ib_srp.c 3932 2005-11-01 17:19:29Z roland $
33  */
34 
35 #include <linux/module.h>
36 #include <linux/init.h>
37 #include <linux/slab.h>
38 #include <linux/err.h>
39 #include <linux/string.h>
40 #include <linux/parser.h>
41 #include <linux/random.h>
42 #include <linux/jiffies.h>
43 
44 #include <asm/atomic.h>
45 
46 #include <scsi/scsi.h>
47 #include <scsi/scsi_device.h>
48 #include <scsi/scsi_dbg.h>
49 #include <scsi/srp.h>
50 
51 #include <rdma/ib_cache.h>
52 
53 #include "ib_srp.h"
54 
55 #define DRV_NAME	"ib_srp"
56 #define PFX		DRV_NAME ": "
57 #define DRV_VERSION	"0.2"
58 #define DRV_RELDATE	"November 1, 2005"
59 
60 MODULE_AUTHOR("Roland Dreier");
61 MODULE_DESCRIPTION("InfiniBand SCSI RDMA Protocol initiator "
62 		   "v" DRV_VERSION " (" DRV_RELDATE ")");
63 MODULE_LICENSE("Dual BSD/GPL");
64 
65 static int topspin_workarounds = 1;
66 
67 module_param(topspin_workarounds, int, 0444);
68 MODULE_PARM_DESC(topspin_workarounds,
69 		 "Enable workarounds for Topspin/Cisco SRP target bugs if != 0");
70 
71 static const u8 topspin_oui[3] = { 0x00, 0x05, 0xad };
72 
73 static void srp_add_one(struct ib_device *device);
74 static void srp_remove_one(struct ib_device *device);
75 static void srp_completion(struct ib_cq *cq, void *target_ptr);
76 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event);
77 
78 static struct ib_client srp_client = {
79 	.name   = "srp",
80 	.add    = srp_add_one,
81 	.remove = srp_remove_one
82 };
83 
84 static inline struct srp_target_port *host_to_target(struct Scsi_Host *host)
85 {
86 	return (struct srp_target_port *) host->hostdata;
87 }
88 
89 static const char *srp_target_info(struct Scsi_Host *host)
90 {
91 	return host_to_target(host)->target_name;
92 }
93 
94 static struct srp_iu *srp_alloc_iu(struct srp_host *host, size_t size,
95 				   gfp_t gfp_mask,
96 				   enum dma_data_direction direction)
97 {
98 	struct srp_iu *iu;
99 
100 	iu = kmalloc(sizeof *iu, gfp_mask);
101 	if (!iu)
102 		goto out;
103 
104 	iu->buf = kzalloc(size, gfp_mask);
105 	if (!iu->buf)
106 		goto out_free_iu;
107 
108 	iu->dma = dma_map_single(host->dev->dma_device, iu->buf, size, direction);
109 	if (dma_mapping_error(iu->dma))
110 		goto out_free_buf;
111 
112 	iu->size      = size;
113 	iu->direction = direction;
114 
115 	return iu;
116 
117 out_free_buf:
118 	kfree(iu->buf);
119 out_free_iu:
120 	kfree(iu);
121 out:
122 	return NULL;
123 }
124 
125 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
126 {
127 	if (!iu)
128 		return;
129 
130 	dma_unmap_single(host->dev->dma_device, iu->dma, iu->size, iu->direction);
131 	kfree(iu->buf);
132 	kfree(iu);
133 }
134 
135 static void srp_qp_event(struct ib_event *event, void *context)
136 {
137 	printk(KERN_ERR PFX "QP event %d\n", event->event);
138 }
139 
140 static int srp_init_qp(struct srp_target_port *target,
141 		       struct ib_qp *qp)
142 {
143 	struct ib_qp_attr *attr;
144 	int ret;
145 
146 	attr = kmalloc(sizeof *attr, GFP_KERNEL);
147 	if (!attr)
148 		return -ENOMEM;
149 
150 	ret = ib_find_cached_pkey(target->srp_host->dev,
151 				  target->srp_host->port,
152 				  be16_to_cpu(target->path.pkey),
153 				  &attr->pkey_index);
154 	if (ret)
155 		goto out;
156 
157 	attr->qp_state        = IB_QPS_INIT;
158 	attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
159 				    IB_ACCESS_REMOTE_WRITE);
160 	attr->port_num        = target->srp_host->port;
161 
162 	ret = ib_modify_qp(qp, attr,
163 			   IB_QP_STATE		|
164 			   IB_QP_PKEY_INDEX	|
165 			   IB_QP_ACCESS_FLAGS	|
166 			   IB_QP_PORT);
167 
168 out:
169 	kfree(attr);
170 	return ret;
171 }
172 
173 static int srp_create_target_ib(struct srp_target_port *target)
174 {
175 	struct ib_qp_init_attr *init_attr;
176 	int ret;
177 
178 	init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
179 	if (!init_attr)
180 		return -ENOMEM;
181 
182 	target->cq = ib_create_cq(target->srp_host->dev, srp_completion,
183 				  NULL, target, SRP_CQ_SIZE);
184 	if (IS_ERR(target->cq)) {
185 		ret = PTR_ERR(target->cq);
186 		goto out;
187 	}
188 
189 	ib_req_notify_cq(target->cq, IB_CQ_NEXT_COMP);
190 
191 	init_attr->event_handler       = srp_qp_event;
192 	init_attr->cap.max_send_wr     = SRP_SQ_SIZE;
193 	init_attr->cap.max_recv_wr     = SRP_RQ_SIZE;
194 	init_attr->cap.max_recv_sge    = 1;
195 	init_attr->cap.max_send_sge    = 1;
196 	init_attr->sq_sig_type         = IB_SIGNAL_ALL_WR;
197 	init_attr->qp_type             = IB_QPT_RC;
198 	init_attr->send_cq             = target->cq;
199 	init_attr->recv_cq             = target->cq;
200 
201 	target->qp = ib_create_qp(target->srp_host->pd, init_attr);
202 	if (IS_ERR(target->qp)) {
203 		ret = PTR_ERR(target->qp);
204 		ib_destroy_cq(target->cq);
205 		goto out;
206 	}
207 
208 	ret = srp_init_qp(target, target->qp);
209 	if (ret) {
210 		ib_destroy_qp(target->qp);
211 		ib_destroy_cq(target->cq);
212 		goto out;
213 	}
214 
215 out:
216 	kfree(init_attr);
217 	return ret;
218 }
219 
220 static void srp_free_target_ib(struct srp_target_port *target)
221 {
222 	int i;
223 
224 	ib_destroy_qp(target->qp);
225 	ib_destroy_cq(target->cq);
226 
227 	for (i = 0; i < SRP_RQ_SIZE; ++i)
228 		srp_free_iu(target->srp_host, target->rx_ring[i]);
229 	for (i = 0; i < SRP_SQ_SIZE + 1; ++i)
230 		srp_free_iu(target->srp_host, target->tx_ring[i]);
231 }
232 
233 static void srp_path_rec_completion(int status,
234 				    struct ib_sa_path_rec *pathrec,
235 				    void *target_ptr)
236 {
237 	struct srp_target_port *target = target_ptr;
238 
239 	target->status = status;
240 	if (status)
241 		printk(KERN_ERR PFX "Got failed path rec status %d\n", status);
242 	else
243 		target->path = *pathrec;
244 	complete(&target->done);
245 }
246 
247 static int srp_lookup_path(struct srp_target_port *target)
248 {
249 	target->path.numb_path = 1;
250 
251 	init_completion(&target->done);
252 
253 	target->path_query_id = ib_sa_path_rec_get(target->srp_host->dev,
254 						   target->srp_host->port,
255 						   &target->path,
256 						   IB_SA_PATH_REC_DGID		|
257 						   IB_SA_PATH_REC_SGID		|
258 						   IB_SA_PATH_REC_NUMB_PATH	|
259 						   IB_SA_PATH_REC_PKEY,
260 						   SRP_PATH_REC_TIMEOUT_MS,
261 						   GFP_KERNEL,
262 						   srp_path_rec_completion,
263 						   target, &target->path_query);
264 	if (target->path_query_id < 0)
265 		return target->path_query_id;
266 
267 	wait_for_completion(&target->done);
268 
269 	if (target->status < 0)
270 		printk(KERN_WARNING PFX "Path record query failed\n");
271 
272 	return target->status;
273 }
274 
275 static int srp_send_req(struct srp_target_port *target)
276 {
277 	struct {
278 		struct ib_cm_req_param param;
279 		struct srp_login_req   priv;
280 	} *req = NULL;
281 	int status;
282 
283 	req = kzalloc(sizeof *req, GFP_KERNEL);
284 	if (!req)
285 		return -ENOMEM;
286 
287 	req->param.primary_path 	      = &target->path;
288 	req->param.alternate_path 	      = NULL;
289 	req->param.service_id 		      = target->service_id;
290 	req->param.qp_num 		      = target->qp->qp_num;
291 	req->param.qp_type 		      = target->qp->qp_type;
292 	req->param.private_data 	      = &req->priv;
293 	req->param.private_data_len 	      = sizeof req->priv;
294 	req->param.flow_control 	      = 1;
295 
296 	get_random_bytes(&req->param.starting_psn, 4);
297 	req->param.starting_psn 	     &= 0xffffff;
298 
299 	/*
300 	 * Pick some arbitrary defaults here; we could make these
301 	 * module parameters if anyone cared about setting them.
302 	 */
303 	req->param.responder_resources	      = 4;
304 	req->param.remote_cm_response_timeout = 20;
305 	req->param.local_cm_response_timeout  = 20;
306 	req->param.retry_count 		      = 7;
307 	req->param.rnr_retry_count 	      = 7;
308 	req->param.max_cm_retries 	      = 15;
309 
310 	req->priv.opcode     	= SRP_LOGIN_REQ;
311 	req->priv.tag        	= 0;
312 	req->priv.req_it_iu_len = cpu_to_be32(SRP_MAX_IU_LEN);
313 	req->priv.req_buf_fmt 	= cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
314 					      SRP_BUF_FORMAT_INDIRECT);
315 	memcpy(req->priv.initiator_port_id, target->srp_host->initiator_port_id, 16);
316 	/*
317 	 * Topspin/Cisco SRP targets will reject our login unless we
318 	 * zero out the first 8 bytes of our initiator port ID.  The
319 	 * second 8 bytes must be our local node GUID, but we always
320 	 * use that anyway.
321 	 */
322 	if (topspin_workarounds && !memcmp(&target->ioc_guid, topspin_oui, 3)) {
323 		printk(KERN_DEBUG PFX "Topspin/Cisco initiator port ID workaround "
324 		       "activated for target GUID %016llx\n",
325 		       (unsigned long long) be64_to_cpu(target->ioc_guid));
326 		memset(req->priv.initiator_port_id, 0, 8);
327 	}
328 	memcpy(req->priv.target_port_id,     &target->id_ext, 8);
329 	memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
330 
331 	status = ib_send_cm_req(target->cm_id, &req->param);
332 
333 	kfree(req);
334 
335 	return status;
336 }
337 
338 static void srp_disconnect_target(struct srp_target_port *target)
339 {
340 	/* XXX should send SRP_I_LOGOUT request */
341 
342 	init_completion(&target->done);
343 	ib_send_cm_dreq(target->cm_id, NULL, 0);
344 	wait_for_completion(&target->done);
345 }
346 
347 static void srp_remove_work(void *target_ptr)
348 {
349 	struct srp_target_port *target = target_ptr;
350 
351 	spin_lock_irq(target->scsi_host->host_lock);
352 	if (target->state != SRP_TARGET_DEAD) {
353 		spin_unlock_irq(target->scsi_host->host_lock);
354 		scsi_host_put(target->scsi_host);
355 		return;
356 	}
357 	target->state = SRP_TARGET_REMOVED;
358 	spin_unlock_irq(target->scsi_host->host_lock);
359 
360 	down(&target->srp_host->target_mutex);
361 	list_del(&target->list);
362 	up(&target->srp_host->target_mutex);
363 
364 	scsi_remove_host(target->scsi_host);
365 	ib_destroy_cm_id(target->cm_id);
366 	srp_free_target_ib(target);
367 	scsi_host_put(target->scsi_host);
368 	/* And another put to really free the target port... */
369 	scsi_host_put(target->scsi_host);
370 }
371 
372 static int srp_connect_target(struct srp_target_port *target)
373 {
374 	int ret;
375 
376 	ret = srp_lookup_path(target);
377 	if (ret)
378 		return ret;
379 
380 	while (1) {
381 		init_completion(&target->done);
382 		ret = srp_send_req(target);
383 		if (ret)
384 			return ret;
385 		wait_for_completion(&target->done);
386 
387 		/*
388 		 * The CM event handling code will set status to
389 		 * SRP_PORT_REDIRECT if we get a port redirect REJ
390 		 * back, or SRP_DLID_REDIRECT if we get a lid/qp
391 		 * redirect REJ back.
392 		 */
393 		switch (target->status) {
394 		case 0:
395 			return 0;
396 
397 		case SRP_PORT_REDIRECT:
398 			ret = srp_lookup_path(target);
399 			if (ret)
400 				return ret;
401 			break;
402 
403 		case SRP_DLID_REDIRECT:
404 			break;
405 
406 		default:
407 			return target->status;
408 		}
409 	}
410 }
411 
412 static int srp_reconnect_target(struct srp_target_port *target)
413 {
414 	struct ib_cm_id *new_cm_id;
415 	struct ib_qp_attr qp_attr;
416 	struct srp_request *req;
417 	struct ib_wc wc;
418 	int ret;
419 	int i;
420 
421 	spin_lock_irq(target->scsi_host->host_lock);
422 	if (target->state != SRP_TARGET_LIVE) {
423 		spin_unlock_irq(target->scsi_host->host_lock);
424 		return -EAGAIN;
425 	}
426 	target->state = SRP_TARGET_CONNECTING;
427 	spin_unlock_irq(target->scsi_host->host_lock);
428 
429 	srp_disconnect_target(target);
430 	/*
431 	 * Now get a new local CM ID so that we avoid confusing the
432 	 * target in case things are really fouled up.
433 	 */
434 	new_cm_id = ib_create_cm_id(target->srp_host->dev,
435 				    srp_cm_handler, target);
436 	if (IS_ERR(new_cm_id)) {
437 		ret = PTR_ERR(new_cm_id);
438 		goto err;
439 	}
440 	ib_destroy_cm_id(target->cm_id);
441 	target->cm_id = new_cm_id;
442 
443 	qp_attr.qp_state = IB_QPS_RESET;
444 	ret = ib_modify_qp(target->qp, &qp_attr, IB_QP_STATE);
445 	if (ret)
446 		goto err;
447 
448 	ret = srp_init_qp(target, target->qp);
449 	if (ret)
450 		goto err;
451 
452 	while (ib_poll_cq(target->cq, 1, &wc) > 0)
453 		; /* nothing */
454 
455 	list_for_each_entry(req, &target->req_queue, list) {
456 		req->scmnd->result = DID_RESET << 16;
457 		req->scmnd->scsi_done(req->scmnd);
458 	}
459 
460 	target->rx_head	 = 0;
461 	target->tx_head	 = 0;
462 	target->tx_tail  = 0;
463 	target->req_head = 0;
464 	for (i = 0; i < SRP_SQ_SIZE - 1; ++i)
465 		target->req_ring[i].next = i + 1;
466 	target->req_ring[SRP_SQ_SIZE - 1].next = -1;
467 	INIT_LIST_HEAD(&target->req_queue);
468 
469 	ret = srp_connect_target(target);
470 	if (ret)
471 		goto err;
472 
473 	spin_lock_irq(target->scsi_host->host_lock);
474 	if (target->state == SRP_TARGET_CONNECTING) {
475 		ret = 0;
476 		target->state = SRP_TARGET_LIVE;
477 	} else
478 		ret = -EAGAIN;
479 	spin_unlock_irq(target->scsi_host->host_lock);
480 
481 	return ret;
482 
483 err:
484 	printk(KERN_ERR PFX "reconnect failed (%d), removing target port.\n", ret);
485 
486 	/*
487 	 * We couldn't reconnect, so kill our target port off.
488 	 * However, we have to defer the real removal because we might
489 	 * be in the context of the SCSI error handler now, which
490 	 * would deadlock if we call scsi_remove_host().
491 	 */
492 	spin_lock_irq(target->scsi_host->host_lock);
493 	if (target->state == SRP_TARGET_CONNECTING) {
494 		target->state = SRP_TARGET_DEAD;
495 		INIT_WORK(&target->work, srp_remove_work, target);
496 		schedule_work(&target->work);
497 	}
498 	spin_unlock_irq(target->scsi_host->host_lock);
499 
500 	return ret;
501 }
502 
503 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_target_port *target,
504 			struct srp_request *req)
505 {
506 	struct srp_cmd *cmd = req->cmd->buf;
507 	int len;
508 	u8 fmt;
509 
510 	if (!scmnd->request_buffer || scmnd->sc_data_direction == DMA_NONE)
511 		return sizeof (struct srp_cmd);
512 
513 	if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
514 	    scmnd->sc_data_direction != DMA_TO_DEVICE) {
515 		printk(KERN_WARNING PFX "Unhandled data direction %d\n",
516 		       scmnd->sc_data_direction);
517 		return -EINVAL;
518 	}
519 
520 	if (scmnd->use_sg) {
521 		struct scatterlist *scat = scmnd->request_buffer;
522 		int n;
523 		int i;
524 
525 		n = dma_map_sg(target->srp_host->dev->dma_device,
526 			       scat, scmnd->use_sg, scmnd->sc_data_direction);
527 
528 		if (n == 1) {
529 			struct srp_direct_buf *buf = (void *) cmd->add_data;
530 
531 			fmt = SRP_DATA_DESC_DIRECT;
532 
533 			buf->va  = cpu_to_be64(sg_dma_address(scat));
534 			buf->key = cpu_to_be32(target->srp_host->mr->rkey);
535 			buf->len = cpu_to_be32(sg_dma_len(scat));
536 
537 			len = sizeof (struct srp_cmd) +
538 				sizeof (struct srp_direct_buf);
539 		} else {
540 			struct srp_indirect_buf *buf = (void *) cmd->add_data;
541 			u32 datalen = 0;
542 
543 			fmt = SRP_DATA_DESC_INDIRECT;
544 
545 			if (scmnd->sc_data_direction == DMA_TO_DEVICE)
546 				cmd->data_out_desc_cnt = n;
547 			else
548 				cmd->data_in_desc_cnt = n;
549 
550 			buf->table_desc.va  = cpu_to_be64(req->cmd->dma +
551 							  sizeof *cmd +
552 							  sizeof *buf);
553 			buf->table_desc.key =
554 				cpu_to_be32(target->srp_host->mr->rkey);
555 			buf->table_desc.len =
556 				cpu_to_be32(n * sizeof (struct srp_direct_buf));
557 
558 			for (i = 0; i < n; ++i) {
559 				buf->desc_list[i].va  = cpu_to_be64(sg_dma_address(&scat[i]));
560 				buf->desc_list[i].key =
561 					cpu_to_be32(target->srp_host->mr->rkey);
562 				buf->desc_list[i].len = cpu_to_be32(sg_dma_len(&scat[i]));
563 
564 				datalen += sg_dma_len(&scat[i]);
565 			}
566 
567 			buf->len = cpu_to_be32(datalen);
568 
569 			len = sizeof (struct srp_cmd) +
570 				sizeof (struct srp_indirect_buf) +
571 				n * sizeof (struct srp_direct_buf);
572 		}
573 	} else {
574 		struct srp_direct_buf *buf = (void *) cmd->add_data;
575 		dma_addr_t dma;
576 
577 		dma = dma_map_single(target->srp_host->dev->dma_device,
578 				     scmnd->request_buffer, scmnd->request_bufflen,
579 				     scmnd->sc_data_direction);
580 		if (dma_mapping_error(dma)) {
581 			printk(KERN_WARNING PFX "unable to map %p/%d (dir %d)\n",
582 			       scmnd->request_buffer, (int) scmnd->request_bufflen,
583 			       scmnd->sc_data_direction);
584 			return -EINVAL;
585 		}
586 
587 		pci_unmap_addr_set(req, direct_mapping, dma);
588 
589 		buf->va  = cpu_to_be64(dma);
590 		buf->key = cpu_to_be32(target->srp_host->mr->rkey);
591 		buf->len = cpu_to_be32(scmnd->request_bufflen);
592 
593 		fmt = SRP_DATA_DESC_DIRECT;
594 
595 		len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
596 	}
597 
598 	if (scmnd->sc_data_direction == DMA_TO_DEVICE)
599 		cmd->buf_fmt = fmt << 4;
600 	else
601 		cmd->buf_fmt = fmt;
602 
603 
604 	return len;
605 }
606 
607 static void srp_unmap_data(struct scsi_cmnd *scmnd,
608 			   struct srp_target_port *target,
609 			   struct srp_request *req)
610 {
611 	if (!scmnd->request_buffer ||
612 	    (scmnd->sc_data_direction != DMA_TO_DEVICE &&
613 	     scmnd->sc_data_direction != DMA_FROM_DEVICE))
614 	    return;
615 
616 	if (scmnd->use_sg)
617 		dma_unmap_sg(target->srp_host->dev->dma_device,
618 			     (struct scatterlist *) scmnd->request_buffer,
619 			     scmnd->use_sg, scmnd->sc_data_direction);
620 	else
621 		dma_unmap_single(target->srp_host->dev->dma_device,
622 				 pci_unmap_addr(req, direct_mapping),
623 				 scmnd->request_bufflen,
624 				 scmnd->sc_data_direction);
625 }
626 
627 static void srp_process_rsp(struct srp_target_port *target, struct srp_rsp *rsp)
628 {
629 	struct srp_request *req;
630 	struct scsi_cmnd *scmnd;
631 	unsigned long flags;
632 	s32 delta;
633 
634 	delta = (s32) be32_to_cpu(rsp->req_lim_delta);
635 
636 	spin_lock_irqsave(target->scsi_host->host_lock, flags);
637 
638 	target->req_lim += delta;
639 
640 	req = &target->req_ring[rsp->tag & ~SRP_TAG_TSK_MGMT];
641 
642 	if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
643 		if (be32_to_cpu(rsp->resp_data_len) < 4)
644 			req->tsk_status = -1;
645 		else
646 			req->tsk_status = rsp->data[3];
647 		complete(&req->done);
648 	} else {
649 		scmnd 	      = req->scmnd;
650 		if (!scmnd)
651 			printk(KERN_ERR "Null scmnd for RSP w/tag %016llx\n",
652 			       (unsigned long long) rsp->tag);
653 		scmnd->result = rsp->status;
654 
655 		if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
656 			memcpy(scmnd->sense_buffer, rsp->data +
657 			       be32_to_cpu(rsp->resp_data_len),
658 			       min_t(int, be32_to_cpu(rsp->sense_data_len),
659 				     SCSI_SENSE_BUFFERSIZE));
660 		}
661 
662 		if (rsp->flags & (SRP_RSP_FLAG_DOOVER | SRP_RSP_FLAG_DOUNDER))
663 			scmnd->resid = be32_to_cpu(rsp->data_out_res_cnt);
664 		else if (rsp->flags & (SRP_RSP_FLAG_DIOVER | SRP_RSP_FLAG_DIUNDER))
665 			scmnd->resid = be32_to_cpu(rsp->data_in_res_cnt);
666 
667 		srp_unmap_data(scmnd, target, req);
668 
669 		if (!req->tsk_mgmt) {
670 			req->scmnd = NULL;
671 			scmnd->host_scribble = (void *) -1L;
672 			scmnd->scsi_done(scmnd);
673 
674 			list_del(&req->list);
675 			req->next = target->req_head;
676 			target->req_head = rsp->tag & ~SRP_TAG_TSK_MGMT;
677 		} else
678 			req->cmd_done = 1;
679 	}
680 
681 	spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
682 }
683 
684 static void srp_reconnect_work(void *target_ptr)
685 {
686 	struct srp_target_port *target = target_ptr;
687 
688 	srp_reconnect_target(target);
689 }
690 
691 static void srp_handle_recv(struct srp_target_port *target, struct ib_wc *wc)
692 {
693 	struct srp_iu *iu;
694 	u8 opcode;
695 
696 	iu = target->rx_ring[wc->wr_id & ~SRP_OP_RECV];
697 
698 	dma_sync_single_for_cpu(target->srp_host->dev->dma_device, iu->dma,
699 				target->max_ti_iu_len, DMA_FROM_DEVICE);
700 
701 	opcode = *(u8 *) iu->buf;
702 
703 	if (0) {
704 		int i;
705 
706 		printk(KERN_ERR PFX "recv completion, opcode 0x%02x\n", opcode);
707 
708 		for (i = 0; i < wc->byte_len; ++i) {
709 			if (i % 8 == 0)
710 				printk(KERN_ERR "  [%02x] ", i);
711 			printk(" %02x", ((u8 *) iu->buf)[i]);
712 			if ((i + 1) % 8 == 0)
713 				printk("\n");
714 		}
715 
716 		if (wc->byte_len % 8)
717 			printk("\n");
718 	}
719 
720 	switch (opcode) {
721 	case SRP_RSP:
722 		srp_process_rsp(target, iu->buf);
723 		break;
724 
725 	case SRP_T_LOGOUT:
726 		/* XXX Handle target logout */
727 		printk(KERN_WARNING PFX "Got target logout request\n");
728 		break;
729 
730 	default:
731 		printk(KERN_WARNING PFX "Unhandled SRP opcode 0x%02x\n", opcode);
732 		break;
733 	}
734 
735 	dma_sync_single_for_device(target->srp_host->dev->dma_device, iu->dma,
736 				   target->max_ti_iu_len, DMA_FROM_DEVICE);
737 }
738 
739 static void srp_completion(struct ib_cq *cq, void *target_ptr)
740 {
741 	struct srp_target_port *target = target_ptr;
742 	struct ib_wc wc;
743 	unsigned long flags;
744 
745 	ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
746 	while (ib_poll_cq(cq, 1, &wc) > 0) {
747 		if (wc.status) {
748 			printk(KERN_ERR PFX "failed %s status %d\n",
749 			       wc.wr_id & SRP_OP_RECV ? "receive" : "send",
750 			       wc.status);
751 			spin_lock_irqsave(target->scsi_host->host_lock, flags);
752 			if (target->state == SRP_TARGET_LIVE)
753 				schedule_work(&target->work);
754 			spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
755 			break;
756 		}
757 
758 		if (wc.wr_id & SRP_OP_RECV)
759 			srp_handle_recv(target, &wc);
760 		else
761 			++target->tx_tail;
762 	}
763 }
764 
765 static int __srp_post_recv(struct srp_target_port *target)
766 {
767 	struct srp_iu *iu;
768 	struct ib_sge list;
769 	struct ib_recv_wr wr, *bad_wr;
770 	unsigned int next;
771 	int ret;
772 
773 	next 	 = target->rx_head & (SRP_RQ_SIZE - 1);
774 	wr.wr_id = next | SRP_OP_RECV;
775 	iu 	 = target->rx_ring[next];
776 
777 	list.addr   = iu->dma;
778 	list.length = iu->size;
779 	list.lkey   = target->srp_host->mr->lkey;
780 
781 	wr.next     = NULL;
782 	wr.sg_list  = &list;
783 	wr.num_sge  = 1;
784 
785 	ret = ib_post_recv(target->qp, &wr, &bad_wr);
786 	if (!ret)
787 		++target->rx_head;
788 
789 	return ret;
790 }
791 
792 static int srp_post_recv(struct srp_target_port *target)
793 {
794 	unsigned long flags;
795 	int ret;
796 
797 	spin_lock_irqsave(target->scsi_host->host_lock, flags);
798 	ret = __srp_post_recv(target);
799 	spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
800 
801 	return ret;
802 }
803 
804 /*
805  * Must be called with target->scsi_host->host_lock held to protect
806  * req_lim and tx_head.  Lock cannot be dropped between call here and
807  * call to __srp_post_send().
808  */
809 static struct srp_iu *__srp_get_tx_iu(struct srp_target_port *target)
810 {
811 	if (target->tx_head - target->tx_tail >= SRP_SQ_SIZE)
812 		return NULL;
813 
814 	if (unlikely(target->req_lim < 1)) {
815 		if (printk_ratelimit())
816 			printk(KERN_DEBUG PFX "Target has req_lim %d\n",
817 			       target->req_lim);
818 		return NULL;
819 	}
820 
821 	return target->tx_ring[target->tx_head & SRP_SQ_SIZE];
822 }
823 
824 /*
825  * Must be called with target->scsi_host->host_lock held to protect
826  * req_lim and tx_head.
827  */
828 static int __srp_post_send(struct srp_target_port *target,
829 			   struct srp_iu *iu, int len)
830 {
831 	struct ib_sge list;
832 	struct ib_send_wr wr, *bad_wr;
833 	int ret = 0;
834 
835 	list.addr   = iu->dma;
836 	list.length = len;
837 	list.lkey   = target->srp_host->mr->lkey;
838 
839 	wr.next       = NULL;
840 	wr.wr_id      = target->tx_head & SRP_SQ_SIZE;
841 	wr.sg_list    = &list;
842 	wr.num_sge    = 1;
843 	wr.opcode     = IB_WR_SEND;
844 	wr.send_flags = IB_SEND_SIGNALED;
845 
846 	ret = ib_post_send(target->qp, &wr, &bad_wr);
847 
848 	if (!ret) {
849 		++target->tx_head;
850 		--target->req_lim;
851 	}
852 
853 	return ret;
854 }
855 
856 static int srp_queuecommand(struct scsi_cmnd *scmnd,
857 			    void (*done)(struct scsi_cmnd *))
858 {
859 	struct srp_target_port *target = host_to_target(scmnd->device->host);
860 	struct srp_request *req;
861 	struct srp_iu *iu;
862 	struct srp_cmd *cmd;
863 	long req_index;
864 	int len;
865 
866 	if (target->state == SRP_TARGET_CONNECTING)
867 		goto err;
868 
869 	if (target->state == SRP_TARGET_DEAD ||
870 	    target->state == SRP_TARGET_REMOVED) {
871 		scmnd->result = DID_BAD_TARGET << 16;
872 		done(scmnd);
873 		return 0;
874 	}
875 
876 	iu = __srp_get_tx_iu(target);
877 	if (!iu)
878 		goto err;
879 
880 	dma_sync_single_for_cpu(target->srp_host->dev->dma_device, iu->dma,
881 				SRP_MAX_IU_LEN, DMA_TO_DEVICE);
882 
883 	req_index = target->req_head;
884 
885 	scmnd->scsi_done     = done;
886 	scmnd->result        = 0;
887 	scmnd->host_scribble = (void *) req_index;
888 
889 	cmd = iu->buf;
890 	memset(cmd, 0, sizeof *cmd);
891 
892 	cmd->opcode = SRP_CMD;
893 	cmd->lun    = cpu_to_be64((u64) scmnd->device->lun << 48);
894 	cmd->tag    = req_index;
895 	memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
896 
897 	req = &target->req_ring[req_index];
898 
899 	req->scmnd    = scmnd;
900 	req->cmd      = iu;
901 	req->cmd_done = 0;
902 	req->tsk_mgmt = NULL;
903 
904 	len = srp_map_data(scmnd, target, req);
905 	if (len < 0) {
906 		printk(KERN_ERR PFX "Failed to map data\n");
907 		goto err;
908 	}
909 
910 	if (__srp_post_recv(target)) {
911 		printk(KERN_ERR PFX "Recv failed\n");
912 		goto err_unmap;
913 	}
914 
915 	dma_sync_single_for_device(target->srp_host->dev->dma_device, iu->dma,
916 				   SRP_MAX_IU_LEN, DMA_TO_DEVICE);
917 
918 	if (__srp_post_send(target, iu, len)) {
919 		printk(KERN_ERR PFX "Send failed\n");
920 		goto err_unmap;
921 	}
922 
923 	target->req_head = req->next;
924 	list_add_tail(&req->list, &target->req_queue);
925 
926 	return 0;
927 
928 err_unmap:
929 	srp_unmap_data(scmnd, target, req);
930 
931 err:
932 	return SCSI_MLQUEUE_HOST_BUSY;
933 }
934 
935 static int srp_alloc_iu_bufs(struct srp_target_port *target)
936 {
937 	int i;
938 
939 	for (i = 0; i < SRP_RQ_SIZE; ++i) {
940 		target->rx_ring[i] = srp_alloc_iu(target->srp_host,
941 						  target->max_ti_iu_len,
942 						  GFP_KERNEL, DMA_FROM_DEVICE);
943 		if (!target->rx_ring[i])
944 			goto err;
945 	}
946 
947 	for (i = 0; i < SRP_SQ_SIZE + 1; ++i) {
948 		target->tx_ring[i] = srp_alloc_iu(target->srp_host,
949 						  SRP_MAX_IU_LEN,
950 						  GFP_KERNEL, DMA_TO_DEVICE);
951 		if (!target->tx_ring[i])
952 			goto err;
953 	}
954 
955 	return 0;
956 
957 err:
958 	for (i = 0; i < SRP_RQ_SIZE; ++i) {
959 		srp_free_iu(target->srp_host, target->rx_ring[i]);
960 		target->rx_ring[i] = NULL;
961 	}
962 
963 	for (i = 0; i < SRP_SQ_SIZE + 1; ++i) {
964 		srp_free_iu(target->srp_host, target->tx_ring[i]);
965 		target->tx_ring[i] = NULL;
966 	}
967 
968 	return -ENOMEM;
969 }
970 
971 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
972 			       struct ib_cm_event *event,
973 			       struct srp_target_port *target)
974 {
975 	struct ib_class_port_info *cpi;
976 	int opcode;
977 
978 	switch (event->param.rej_rcvd.reason) {
979 	case IB_CM_REJ_PORT_CM_REDIRECT:
980 		cpi = event->param.rej_rcvd.ari;
981 		target->path.dlid = cpi->redirect_lid;
982 		target->path.pkey = cpi->redirect_pkey;
983 		cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
984 		memcpy(target->path.dgid.raw, cpi->redirect_gid, 16);
985 
986 		target->status = target->path.dlid ?
987 			SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
988 		break;
989 
990 	case IB_CM_REJ_PORT_REDIRECT:
991 		if (topspin_workarounds &&
992 		    !memcmp(&target->ioc_guid, topspin_oui, 3)) {
993 			/*
994 			 * Topspin/Cisco SRP gateways incorrectly send
995 			 * reject reason code 25 when they mean 24
996 			 * (port redirect).
997 			 */
998 			memcpy(target->path.dgid.raw,
999 			       event->param.rej_rcvd.ari, 16);
1000 
1001 			printk(KERN_DEBUG PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
1002 			       (unsigned long long) be64_to_cpu(target->path.dgid.global.subnet_prefix),
1003 			       (unsigned long long) be64_to_cpu(target->path.dgid.global.interface_id));
1004 
1005 			target->status = SRP_PORT_REDIRECT;
1006 		} else {
1007 			printk(KERN_WARNING "  REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
1008 			target->status = -ECONNRESET;
1009 		}
1010 		break;
1011 
1012 	case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
1013 		printk(KERN_WARNING "  REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
1014 		target->status = -ECONNRESET;
1015 		break;
1016 
1017 	case IB_CM_REJ_CONSUMER_DEFINED:
1018 		opcode = *(u8 *) event->private_data;
1019 		if (opcode == SRP_LOGIN_REJ) {
1020 			struct srp_login_rej *rej = event->private_data;
1021 			u32 reason = be32_to_cpu(rej->reason);
1022 
1023 			if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
1024 				printk(KERN_WARNING PFX
1025 				       "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
1026 			else
1027 				printk(KERN_WARNING PFX
1028 				       "SRP LOGIN REJECTED, reason 0x%08x\n", reason);
1029 		} else
1030 			printk(KERN_WARNING "  REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
1031 			       " opcode 0x%02x\n", opcode);
1032 		target->status = -ECONNRESET;
1033 		break;
1034 
1035 	default:
1036 		printk(KERN_WARNING "  REJ reason 0x%x\n",
1037 		       event->param.rej_rcvd.reason);
1038 		target->status = -ECONNRESET;
1039 	}
1040 }
1041 
1042 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
1043 {
1044 	struct srp_target_port *target = cm_id->context;
1045 	struct ib_qp_attr *qp_attr = NULL;
1046 	int attr_mask = 0;
1047 	int comp = 0;
1048 	int opcode = 0;
1049 
1050 	switch (event->event) {
1051 	case IB_CM_REQ_ERROR:
1052 		printk(KERN_DEBUG PFX "Sending CM REQ failed\n");
1053 		comp = 1;
1054 		target->status = -ECONNRESET;
1055 		break;
1056 
1057 	case IB_CM_REP_RECEIVED:
1058 		comp = 1;
1059 		opcode = *(u8 *) event->private_data;
1060 
1061 		if (opcode == SRP_LOGIN_RSP) {
1062 			struct srp_login_rsp *rsp = event->private_data;
1063 
1064 			target->max_ti_iu_len = be32_to_cpu(rsp->max_ti_iu_len);
1065 			target->req_lim       = be32_to_cpu(rsp->req_lim_delta);
1066 
1067 			target->scsi_host->can_queue = min(target->req_lim,
1068 							   target->scsi_host->can_queue);
1069 		} else {
1070 			printk(KERN_WARNING PFX "Unhandled RSP opcode %#x\n", opcode);
1071 			target->status = -ECONNRESET;
1072 			break;
1073 		}
1074 
1075 		target->status = srp_alloc_iu_bufs(target);
1076 		if (target->status)
1077 			break;
1078 
1079 		qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
1080 		if (!qp_attr) {
1081 			target->status = -ENOMEM;
1082 			break;
1083 		}
1084 
1085 		qp_attr->qp_state = IB_QPS_RTR;
1086 		target->status = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1087 		if (target->status)
1088 			break;
1089 
1090 		target->status = ib_modify_qp(target->qp, qp_attr, attr_mask);
1091 		if (target->status)
1092 			break;
1093 
1094 		target->status = srp_post_recv(target);
1095 		if (target->status)
1096 			break;
1097 
1098 		qp_attr->qp_state = IB_QPS_RTS;
1099 		target->status = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1100 		if (target->status)
1101 			break;
1102 
1103 		target->status = ib_modify_qp(target->qp, qp_attr, attr_mask);
1104 		if (target->status)
1105 			break;
1106 
1107 		target->status = ib_send_cm_rtu(cm_id, NULL, 0);
1108 		if (target->status)
1109 			break;
1110 
1111 		break;
1112 
1113 	case IB_CM_REJ_RECEIVED:
1114 		printk(KERN_DEBUG PFX "REJ received\n");
1115 		comp = 1;
1116 
1117 		srp_cm_rej_handler(cm_id, event, target);
1118 		break;
1119 
1120 	case IB_CM_MRA_RECEIVED:
1121 		printk(KERN_ERR PFX "MRA received\n");
1122 		break;
1123 
1124 	case IB_CM_DREP_RECEIVED:
1125 		break;
1126 
1127 	case IB_CM_TIMEWAIT_EXIT:
1128 		printk(KERN_ERR PFX "connection closed\n");
1129 
1130 		comp = 1;
1131 		target->status = 0;
1132 		break;
1133 
1134 	default:
1135 		printk(KERN_WARNING PFX "Unhandled CM event %d\n", event->event);
1136 		break;
1137 	}
1138 
1139 	if (comp)
1140 		complete(&target->done);
1141 
1142 	kfree(qp_attr);
1143 
1144 	return 0;
1145 }
1146 
1147 static int srp_send_tsk_mgmt(struct scsi_cmnd *scmnd, u8 func)
1148 {
1149 	struct srp_target_port *target = host_to_target(scmnd->device->host);
1150 	struct srp_request *req;
1151 	struct srp_iu *iu;
1152 	struct srp_tsk_mgmt *tsk_mgmt;
1153 	int req_index;
1154 	int ret = FAILED;
1155 
1156 	spin_lock_irq(target->scsi_host->host_lock);
1157 
1158 	if (scmnd->host_scribble == (void *) -1L)
1159 		goto out;
1160 
1161 	req_index = (long) scmnd->host_scribble;
1162 	printk(KERN_ERR "Abort for req_index %d\n", req_index);
1163 
1164 	req = &target->req_ring[req_index];
1165 	init_completion(&req->done);
1166 
1167 	iu = __srp_get_tx_iu(target);
1168 	if (!iu)
1169 		goto out;
1170 
1171 	tsk_mgmt = iu->buf;
1172 	memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
1173 
1174 	tsk_mgmt->opcode 	= SRP_TSK_MGMT;
1175 	tsk_mgmt->lun 		= cpu_to_be64((u64) scmnd->device->lun << 48);
1176 	tsk_mgmt->tag 		= req_index | SRP_TAG_TSK_MGMT;
1177 	tsk_mgmt->tsk_mgmt_func = func;
1178 	tsk_mgmt->task_tag 	= req_index;
1179 
1180 	if (__srp_post_send(target, iu, sizeof *tsk_mgmt))
1181 		goto out;
1182 
1183 	req->tsk_mgmt = iu;
1184 
1185 	spin_unlock_irq(target->scsi_host->host_lock);
1186 	if (!wait_for_completion_timeout(&req->done,
1187 					 msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
1188 		return FAILED;
1189 	spin_lock_irq(target->scsi_host->host_lock);
1190 
1191 	if (req->cmd_done) {
1192 		list_del(&req->list);
1193 		req->next = target->req_head;
1194 		target->req_head = req_index;
1195 
1196 		scmnd->scsi_done(scmnd);
1197 	} else if (!req->tsk_status) {
1198 		scmnd->result = DID_ABORT << 16;
1199 		ret = SUCCESS;
1200 	}
1201 
1202 out:
1203 	spin_unlock_irq(target->scsi_host->host_lock);
1204 	return ret;
1205 }
1206 
1207 static int srp_abort(struct scsi_cmnd *scmnd)
1208 {
1209 	printk(KERN_ERR "SRP abort called\n");
1210 
1211 	return srp_send_tsk_mgmt(scmnd, SRP_TSK_ABORT_TASK);
1212 }
1213 
1214 static int srp_reset_device(struct scsi_cmnd *scmnd)
1215 {
1216 	printk(KERN_ERR "SRP reset_device called\n");
1217 
1218 	return srp_send_tsk_mgmt(scmnd, SRP_TSK_LUN_RESET);
1219 }
1220 
1221 static int srp_reset_host(struct scsi_cmnd *scmnd)
1222 {
1223 	struct srp_target_port *target = host_to_target(scmnd->device->host);
1224 	int ret = FAILED;
1225 
1226 	printk(KERN_ERR PFX "SRP reset_host called\n");
1227 
1228 	if (!srp_reconnect_target(target))
1229 		ret = SUCCESS;
1230 
1231 	return ret;
1232 }
1233 
1234 static struct scsi_host_template srp_template = {
1235 	.module				= THIS_MODULE,
1236 	.name				= DRV_NAME,
1237 	.info				= srp_target_info,
1238 	.queuecommand			= srp_queuecommand,
1239 	.eh_abort_handler		= srp_abort,
1240 	.eh_device_reset_handler	= srp_reset_device,
1241 	.eh_host_reset_handler		= srp_reset_host,
1242 	.can_queue			= SRP_SQ_SIZE,
1243 	.this_id			= -1,
1244 	.sg_tablesize			= SRP_MAX_INDIRECT,
1245 	.cmd_per_lun			= SRP_SQ_SIZE,
1246 	.use_clustering			= ENABLE_CLUSTERING
1247 };
1248 
1249 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
1250 {
1251 	sprintf(target->target_name, "SRP.T10:%016llX",
1252 		 (unsigned long long) be64_to_cpu(target->id_ext));
1253 
1254 	if (scsi_add_host(target->scsi_host, host->dev->dma_device))
1255 		return -ENODEV;
1256 
1257 	down(&host->target_mutex);
1258 	list_add_tail(&target->list, &host->target_list);
1259 	up(&host->target_mutex);
1260 
1261 	target->state = SRP_TARGET_LIVE;
1262 
1263 	/* XXX: are we supposed to have a definition of SCAN_WILD_CARD ?? */
1264 	scsi_scan_target(&target->scsi_host->shost_gendev,
1265 			 0, target->scsi_id, ~0, 0);
1266 
1267 	return 0;
1268 }
1269 
1270 static void srp_release_class_dev(struct class_device *class_dev)
1271 {
1272 	struct srp_host *host =
1273 		container_of(class_dev, struct srp_host, class_dev);
1274 
1275 	complete(&host->released);
1276 }
1277 
1278 static struct class srp_class = {
1279 	.name    = "infiniband_srp",
1280 	.release = srp_release_class_dev
1281 };
1282 
1283 /*
1284  * Target ports are added by writing
1285  *
1286  *     id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
1287  *     pkey=<P_Key>,service_id=<service ID>
1288  *
1289  * to the add_target sysfs attribute.
1290  */
1291 enum {
1292 	SRP_OPT_ERR		= 0,
1293 	SRP_OPT_ID_EXT		= 1 << 0,
1294 	SRP_OPT_IOC_GUID	= 1 << 1,
1295 	SRP_OPT_DGID		= 1 << 2,
1296 	SRP_OPT_PKEY		= 1 << 3,
1297 	SRP_OPT_SERVICE_ID	= 1 << 4,
1298 	SRP_OPT_MAX_SECT	= 1 << 5,
1299 	SRP_OPT_ALL		= (SRP_OPT_ID_EXT	|
1300 				   SRP_OPT_IOC_GUID	|
1301 				   SRP_OPT_DGID		|
1302 				   SRP_OPT_PKEY		|
1303 				   SRP_OPT_SERVICE_ID),
1304 };
1305 
1306 static match_table_t srp_opt_tokens = {
1307 	{ SRP_OPT_ID_EXT,	"id_ext=%s" 	},
1308 	{ SRP_OPT_IOC_GUID,	"ioc_guid=%s" 	},
1309 	{ SRP_OPT_DGID,		"dgid=%s" 	},
1310 	{ SRP_OPT_PKEY,		"pkey=%x" 	},
1311 	{ SRP_OPT_SERVICE_ID,	"service_id=%s" },
1312 	{ SRP_OPT_MAX_SECT,     "max_sect=%d" 	},
1313 	{ SRP_OPT_ERR,		NULL 		}
1314 };
1315 
1316 static int srp_parse_options(const char *buf, struct srp_target_port *target)
1317 {
1318 	char *options, *sep_opt;
1319 	char *p;
1320 	char dgid[3];
1321 	substring_t args[MAX_OPT_ARGS];
1322 	int opt_mask = 0;
1323 	int token;
1324 	int ret = -EINVAL;
1325 	int i;
1326 
1327 	options = kstrdup(buf, GFP_KERNEL);
1328 	if (!options)
1329 		return -ENOMEM;
1330 
1331 	sep_opt = options;
1332 	while ((p = strsep(&sep_opt, ",")) != NULL) {
1333 		if (!*p)
1334 			continue;
1335 
1336 		token = match_token(p, srp_opt_tokens, args);
1337 		opt_mask |= token;
1338 
1339 		switch (token) {
1340 		case SRP_OPT_ID_EXT:
1341 			p = match_strdup(args);
1342 			target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
1343 			kfree(p);
1344 			break;
1345 
1346 		case SRP_OPT_IOC_GUID:
1347 			p = match_strdup(args);
1348 			target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
1349 			kfree(p);
1350 			break;
1351 
1352 		case SRP_OPT_DGID:
1353 			p = match_strdup(args);
1354 			if (strlen(p) != 32) {
1355 				printk(KERN_WARNING PFX "bad dest GID parameter '%s'\n", p);
1356 				goto out;
1357 			}
1358 
1359 			for (i = 0; i < 16; ++i) {
1360 				strlcpy(dgid, p + i * 2, 3);
1361 				target->path.dgid.raw[i] = simple_strtoul(dgid, NULL, 16);
1362 			}
1363 			break;
1364 
1365 		case SRP_OPT_PKEY:
1366 			if (match_hex(args, &token)) {
1367 				printk(KERN_WARNING PFX "bad P_Key parameter '%s'\n", p);
1368 				goto out;
1369 			}
1370 			target->path.pkey = cpu_to_be16(token);
1371 			break;
1372 
1373 		case SRP_OPT_SERVICE_ID:
1374 			p = match_strdup(args);
1375 			target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
1376 			kfree(p);
1377 			break;
1378 
1379 		case SRP_OPT_MAX_SECT:
1380 			if (match_int(args, &token)) {
1381 				printk(KERN_WARNING PFX "bad max sect parameter '%s'\n", p);
1382 				goto out;
1383 			}
1384 			target->scsi_host->max_sectors = token;
1385 			break;
1386 
1387 		default:
1388 			printk(KERN_WARNING PFX "unknown parameter or missing value "
1389 			       "'%s' in target creation request\n", p);
1390 			goto out;
1391 		}
1392 	}
1393 
1394 	if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
1395 		ret = 0;
1396 	else
1397 		for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
1398 			if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
1399 			    !(srp_opt_tokens[i].token & opt_mask))
1400 				printk(KERN_WARNING PFX "target creation request is "
1401 				       "missing parameter '%s'\n",
1402 				       srp_opt_tokens[i].pattern);
1403 
1404 out:
1405 	kfree(options);
1406 	return ret;
1407 }
1408 
1409 static ssize_t srp_create_target(struct class_device *class_dev,
1410 				 const char *buf, size_t count)
1411 {
1412 	struct srp_host *host =
1413 		container_of(class_dev, struct srp_host, class_dev);
1414 	struct Scsi_Host *target_host;
1415 	struct srp_target_port *target;
1416 	int ret;
1417 	int i;
1418 
1419 	target_host = scsi_host_alloc(&srp_template,
1420 				      sizeof (struct srp_target_port));
1421 	if (!target_host)
1422 		return -ENOMEM;
1423 
1424 	target_host->max_lun = SRP_MAX_LUN;
1425 
1426 	target = host_to_target(target_host);
1427 	memset(target, 0, sizeof *target);
1428 
1429 	target->scsi_host  = target_host;
1430 	target->srp_host   = host;
1431 
1432 	INIT_WORK(&target->work, srp_reconnect_work, target);
1433 
1434 	for (i = 0; i < SRP_SQ_SIZE - 1; ++i)
1435 		target->req_ring[i].next = i + 1;
1436 	target->req_ring[SRP_SQ_SIZE - 1].next = -1;
1437 	INIT_LIST_HEAD(&target->req_queue);
1438 
1439 	ret = srp_parse_options(buf, target);
1440 	if (ret)
1441 		goto err;
1442 
1443 	ib_get_cached_gid(host->dev, host->port, 0, &target->path.sgid);
1444 
1445 	printk(KERN_DEBUG PFX "new target: id_ext %016llx ioc_guid %016llx pkey %04x "
1446 	       "service_id %016llx dgid %04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x\n",
1447 	       (unsigned long long) be64_to_cpu(target->id_ext),
1448 	       (unsigned long long) be64_to_cpu(target->ioc_guid),
1449 	       be16_to_cpu(target->path.pkey),
1450 	       (unsigned long long) be64_to_cpu(target->service_id),
1451 	       (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[0]),
1452 	       (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[2]),
1453 	       (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[4]),
1454 	       (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[6]),
1455 	       (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[8]),
1456 	       (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[10]),
1457 	       (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[12]),
1458 	       (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[14]));
1459 
1460 	ret = srp_create_target_ib(target);
1461 	if (ret)
1462 		goto err;
1463 
1464 	target->cm_id = ib_create_cm_id(host->dev, srp_cm_handler, target);
1465 	if (IS_ERR(target->cm_id)) {
1466 		ret = PTR_ERR(target->cm_id);
1467 		goto err_free;
1468 	}
1469 
1470 	ret = srp_connect_target(target);
1471 	if (ret) {
1472 		printk(KERN_ERR PFX "Connection failed\n");
1473 		goto err_cm_id;
1474 	}
1475 
1476 	ret = srp_add_target(host, target);
1477 	if (ret)
1478 		goto err_disconnect;
1479 
1480 	return count;
1481 
1482 err_disconnect:
1483 	srp_disconnect_target(target);
1484 
1485 err_cm_id:
1486 	ib_destroy_cm_id(target->cm_id);
1487 
1488 err_free:
1489 	srp_free_target_ib(target);
1490 
1491 err:
1492 	scsi_host_put(target_host);
1493 
1494 	return ret;
1495 }
1496 
1497 static CLASS_DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
1498 
1499 static ssize_t show_ibdev(struct class_device *class_dev, char *buf)
1500 {
1501 	struct srp_host *host =
1502 		container_of(class_dev, struct srp_host, class_dev);
1503 
1504 	return sprintf(buf, "%s\n", host->dev->name);
1505 }
1506 
1507 static CLASS_DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
1508 
1509 static ssize_t show_port(struct class_device *class_dev, char *buf)
1510 {
1511 	struct srp_host *host =
1512 		container_of(class_dev, struct srp_host, class_dev);
1513 
1514 	return sprintf(buf, "%d\n", host->port);
1515 }
1516 
1517 static CLASS_DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
1518 
1519 static struct srp_host *srp_add_port(struct ib_device *device,
1520 				     __be64 node_guid, u8 port)
1521 {
1522 	struct srp_host *host;
1523 
1524 	host = kzalloc(sizeof *host, GFP_KERNEL);
1525 	if (!host)
1526 		return NULL;
1527 
1528 	INIT_LIST_HEAD(&host->target_list);
1529 	init_MUTEX(&host->target_mutex);
1530 	init_completion(&host->released);
1531 	host->dev  = device;
1532 	host->port = port;
1533 
1534 	host->initiator_port_id[7] = port;
1535 	memcpy(host->initiator_port_id + 8, &node_guid, 8);
1536 
1537 	host->pd   = ib_alloc_pd(device);
1538 	if (IS_ERR(host->pd))
1539 		goto err_free;
1540 
1541 	host->mr   = ib_get_dma_mr(host->pd,
1542 				   IB_ACCESS_LOCAL_WRITE |
1543 				   IB_ACCESS_REMOTE_READ |
1544 				   IB_ACCESS_REMOTE_WRITE);
1545 	if (IS_ERR(host->mr))
1546 		goto err_pd;
1547 
1548 	host->class_dev.class = &srp_class;
1549 	host->class_dev.dev   = device->dma_device;
1550 	snprintf(host->class_dev.class_id, BUS_ID_SIZE, "srp-%s-%d",
1551 		 device->name, port);
1552 
1553 	if (class_device_register(&host->class_dev))
1554 		goto err_mr;
1555 	if (class_device_create_file(&host->class_dev, &class_device_attr_add_target))
1556 		goto err_class;
1557 	if (class_device_create_file(&host->class_dev, &class_device_attr_ibdev))
1558 		goto err_class;
1559 	if (class_device_create_file(&host->class_dev, &class_device_attr_port))
1560 		goto err_class;
1561 
1562 	return host;
1563 
1564 err_class:
1565 	class_device_unregister(&host->class_dev);
1566 
1567 err_mr:
1568 	ib_dereg_mr(host->mr);
1569 
1570 err_pd:
1571 	ib_dealloc_pd(host->pd);
1572 
1573 err_free:
1574 	kfree(host);
1575 
1576 	return NULL;
1577 }
1578 
1579 static void srp_add_one(struct ib_device *device)
1580 {
1581 	struct list_head *dev_list;
1582 	struct srp_host *host;
1583 	struct ib_device_attr *dev_attr;
1584 	int s, e, p;
1585 
1586 	dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
1587 	if (!dev_attr)
1588 		return;
1589 
1590 	if (ib_query_device(device, dev_attr)) {
1591 		printk(KERN_WARNING PFX "Couldn't query node GUID for %s.\n",
1592 		       device->name);
1593 		goto out;
1594 	}
1595 
1596 	dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1597 	if (!dev_list)
1598 		goto out;
1599 
1600 	INIT_LIST_HEAD(dev_list);
1601 
1602 	if (device->node_type == IB_NODE_SWITCH) {
1603 		s = 0;
1604 		e = 0;
1605 	} else {
1606 		s = 1;
1607 		e = device->phys_port_cnt;
1608 	}
1609 
1610 	for (p = s; p <= e; ++p) {
1611 		host = srp_add_port(device, dev_attr->node_guid, p);
1612 		if (host)
1613 			list_add_tail(&host->list, dev_list);
1614 	}
1615 
1616 	ib_set_client_data(device, &srp_client, dev_list);
1617 
1618 out:
1619 	kfree(dev_attr);
1620 }
1621 
1622 static void srp_remove_one(struct ib_device *device)
1623 {
1624 	struct list_head *dev_list;
1625 	struct srp_host *host, *tmp_host;
1626 	LIST_HEAD(target_list);
1627 	struct srp_target_port *target, *tmp_target;
1628 	unsigned long flags;
1629 
1630 	dev_list = ib_get_client_data(device, &srp_client);
1631 
1632 	list_for_each_entry_safe(host, tmp_host, dev_list, list) {
1633 		class_device_unregister(&host->class_dev);
1634 		/*
1635 		 * Wait for the sysfs entry to go away, so that no new
1636 		 * target ports can be created.
1637 		 */
1638 		wait_for_completion(&host->released);
1639 
1640 		/*
1641 		 * Mark all target ports as removed, so we stop queueing
1642 		 * commands and don't try to reconnect.
1643 		 */
1644 		down(&host->target_mutex);
1645 		list_for_each_entry_safe(target, tmp_target,
1646 					 &host->target_list, list) {
1647 			spin_lock_irqsave(target->scsi_host->host_lock, flags);
1648 			if (target->state != SRP_TARGET_REMOVED)
1649 				target->state = SRP_TARGET_REMOVED;
1650 			spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
1651 		}
1652 		up(&host->target_mutex);
1653 
1654 		/*
1655 		 * Wait for any reconnection tasks that may have
1656 		 * started before we marked our target ports as
1657 		 * removed, and any target port removal tasks.
1658 		 */
1659 		flush_scheduled_work();
1660 
1661 		list_for_each_entry_safe(target, tmp_target,
1662 					 &host->target_list, list) {
1663 			scsi_remove_host(target->scsi_host);
1664 			srp_disconnect_target(target);
1665 			ib_destroy_cm_id(target->cm_id);
1666 			srp_free_target_ib(target);
1667 			scsi_host_put(target->scsi_host);
1668 		}
1669 
1670 		ib_dereg_mr(host->mr);
1671 		ib_dealloc_pd(host->pd);
1672 		kfree(host);
1673 	}
1674 
1675 	kfree(dev_list);
1676 }
1677 
1678 static int __init srp_init_module(void)
1679 {
1680 	int ret;
1681 
1682 	ret = class_register(&srp_class);
1683 	if (ret) {
1684 		printk(KERN_ERR PFX "couldn't register class infiniband_srp\n");
1685 		return ret;
1686 	}
1687 
1688 	ret = ib_register_client(&srp_client);
1689 	if (ret) {
1690 		printk(KERN_ERR PFX "couldn't register IB client\n");
1691 		class_unregister(&srp_class);
1692 		return ret;
1693 	}
1694 
1695 	return 0;
1696 }
1697 
1698 static void __exit srp_cleanup_module(void)
1699 {
1700 	ib_unregister_client(&srp_client);
1701 	class_unregister(&srp_class);
1702 }
1703 
1704 module_init(srp_init_module);
1705 module_exit(srp_cleanup_module);
1706