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 #define pr_fmt(fmt) PFX fmt
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 <linux/atomic.h>
45 
46 #include <scsi/scsi.h>
47 #include <scsi/scsi_device.h>
48 #include <scsi/scsi_dbg.h>
49 #include <scsi/scsi_tcq.h>
50 #include <scsi/srp.h>
51 #include <scsi/scsi_transport_srp.h>
52 
53 #include "ib_srp.h"
54 
55 #define DRV_NAME	"ib_srp"
56 #define PFX		DRV_NAME ": "
57 #define DRV_VERSION	"1.0"
58 #define DRV_RELDATE	"July 1, 2013"
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 unsigned int srp_sg_tablesize;
66 static unsigned int cmd_sg_entries;
67 static unsigned int indirect_sg_entries;
68 static bool allow_ext_sg;
69 static int topspin_workarounds = 1;
70 
71 module_param(srp_sg_tablesize, uint, 0444);
72 MODULE_PARM_DESC(srp_sg_tablesize, "Deprecated name for cmd_sg_entries");
73 
74 module_param(cmd_sg_entries, uint, 0444);
75 MODULE_PARM_DESC(cmd_sg_entries,
76 		 "Default number of gather/scatter entries in the SRP command (default is 12, max 255)");
77 
78 module_param(indirect_sg_entries, uint, 0444);
79 MODULE_PARM_DESC(indirect_sg_entries,
80 		 "Default max number of gather/scatter entries (default is 12, max is " __stringify(SCSI_MAX_SG_CHAIN_SEGMENTS) ")");
81 
82 module_param(allow_ext_sg, bool, 0444);
83 MODULE_PARM_DESC(allow_ext_sg,
84 		  "Default behavior when there are more than cmd_sg_entries S/G entries after mapping; fails the request when false (default false)");
85 
86 module_param(topspin_workarounds, int, 0444);
87 MODULE_PARM_DESC(topspin_workarounds,
88 		 "Enable workarounds for Topspin/Cisco SRP target bugs if != 0");
89 
90 static struct kernel_param_ops srp_tmo_ops;
91 
92 static int srp_reconnect_delay = 10;
93 module_param_cb(reconnect_delay, &srp_tmo_ops, &srp_reconnect_delay,
94 		S_IRUGO | S_IWUSR);
95 MODULE_PARM_DESC(reconnect_delay, "Time between successive reconnect attempts");
96 
97 static int srp_fast_io_fail_tmo = 15;
98 module_param_cb(fast_io_fail_tmo, &srp_tmo_ops, &srp_fast_io_fail_tmo,
99 		S_IRUGO | S_IWUSR);
100 MODULE_PARM_DESC(fast_io_fail_tmo,
101 		 "Number of seconds between the observation of a transport"
102 		 " layer error and failing all I/O. \"off\" means that this"
103 		 " functionality is disabled.");
104 
105 static int srp_dev_loss_tmo = 600;
106 module_param_cb(dev_loss_tmo, &srp_tmo_ops, &srp_dev_loss_tmo,
107 		S_IRUGO | S_IWUSR);
108 MODULE_PARM_DESC(dev_loss_tmo,
109 		 "Maximum number of seconds that the SRP transport should"
110 		 " insulate transport layer errors. After this time has been"
111 		 " exceeded the SCSI host is removed. Should be"
112 		 " between 1 and " __stringify(SCSI_DEVICE_BLOCK_MAX_TIMEOUT)
113 		 " if fast_io_fail_tmo has not been set. \"off\" means that"
114 		 " this functionality is disabled.");
115 
116 static void srp_add_one(struct ib_device *device);
117 static void srp_remove_one(struct ib_device *device);
118 static void srp_recv_completion(struct ib_cq *cq, void *target_ptr);
119 static void srp_send_completion(struct ib_cq *cq, void *target_ptr);
120 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event);
121 
122 static struct scsi_transport_template *ib_srp_transport_template;
123 
124 static struct ib_client srp_client = {
125 	.name   = "srp",
126 	.add    = srp_add_one,
127 	.remove = srp_remove_one
128 };
129 
130 static struct ib_sa_client srp_sa_client;
131 
132 static int srp_tmo_get(char *buffer, const struct kernel_param *kp)
133 {
134 	int tmo = *(int *)kp->arg;
135 
136 	if (tmo >= 0)
137 		return sprintf(buffer, "%d", tmo);
138 	else
139 		return sprintf(buffer, "off");
140 }
141 
142 static int srp_tmo_set(const char *val, const struct kernel_param *kp)
143 {
144 	int tmo, res;
145 
146 	if (strncmp(val, "off", 3) != 0) {
147 		res = kstrtoint(val, 0, &tmo);
148 		if (res)
149 			goto out;
150 	} else {
151 		tmo = -1;
152 	}
153 	if (kp->arg == &srp_reconnect_delay)
154 		res = srp_tmo_valid(tmo, srp_fast_io_fail_tmo,
155 				    srp_dev_loss_tmo);
156 	else if (kp->arg == &srp_fast_io_fail_tmo)
157 		res = srp_tmo_valid(srp_reconnect_delay, tmo, srp_dev_loss_tmo);
158 	else
159 		res = srp_tmo_valid(srp_reconnect_delay, srp_fast_io_fail_tmo,
160 				    tmo);
161 	if (res)
162 		goto out;
163 	*(int *)kp->arg = tmo;
164 
165 out:
166 	return res;
167 }
168 
169 static struct kernel_param_ops srp_tmo_ops = {
170 	.get = srp_tmo_get,
171 	.set = srp_tmo_set,
172 };
173 
174 static inline struct srp_target_port *host_to_target(struct Scsi_Host *host)
175 {
176 	return (struct srp_target_port *) host->hostdata;
177 }
178 
179 static const char *srp_target_info(struct Scsi_Host *host)
180 {
181 	return host_to_target(host)->target_name;
182 }
183 
184 static int srp_target_is_topspin(struct srp_target_port *target)
185 {
186 	static const u8 topspin_oui[3] = { 0x00, 0x05, 0xad };
187 	static const u8 cisco_oui[3]   = { 0x00, 0x1b, 0x0d };
188 
189 	return topspin_workarounds &&
190 		(!memcmp(&target->ioc_guid, topspin_oui, sizeof topspin_oui) ||
191 		 !memcmp(&target->ioc_guid, cisco_oui, sizeof cisco_oui));
192 }
193 
194 static struct srp_iu *srp_alloc_iu(struct srp_host *host, size_t size,
195 				   gfp_t gfp_mask,
196 				   enum dma_data_direction direction)
197 {
198 	struct srp_iu *iu;
199 
200 	iu = kmalloc(sizeof *iu, gfp_mask);
201 	if (!iu)
202 		goto out;
203 
204 	iu->buf = kzalloc(size, gfp_mask);
205 	if (!iu->buf)
206 		goto out_free_iu;
207 
208 	iu->dma = ib_dma_map_single(host->srp_dev->dev, iu->buf, size,
209 				    direction);
210 	if (ib_dma_mapping_error(host->srp_dev->dev, iu->dma))
211 		goto out_free_buf;
212 
213 	iu->size      = size;
214 	iu->direction = direction;
215 
216 	return iu;
217 
218 out_free_buf:
219 	kfree(iu->buf);
220 out_free_iu:
221 	kfree(iu);
222 out:
223 	return NULL;
224 }
225 
226 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
227 {
228 	if (!iu)
229 		return;
230 
231 	ib_dma_unmap_single(host->srp_dev->dev, iu->dma, iu->size,
232 			    iu->direction);
233 	kfree(iu->buf);
234 	kfree(iu);
235 }
236 
237 static void srp_qp_event(struct ib_event *event, void *context)
238 {
239 	pr_debug("QP event %d\n", event->event);
240 }
241 
242 static int srp_init_qp(struct srp_target_port *target,
243 		       struct ib_qp *qp)
244 {
245 	struct ib_qp_attr *attr;
246 	int ret;
247 
248 	attr = kmalloc(sizeof *attr, GFP_KERNEL);
249 	if (!attr)
250 		return -ENOMEM;
251 
252 	ret = ib_find_pkey(target->srp_host->srp_dev->dev,
253 			   target->srp_host->port,
254 			   be16_to_cpu(target->path.pkey),
255 			   &attr->pkey_index);
256 	if (ret)
257 		goto out;
258 
259 	attr->qp_state        = IB_QPS_INIT;
260 	attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
261 				    IB_ACCESS_REMOTE_WRITE);
262 	attr->port_num        = target->srp_host->port;
263 
264 	ret = ib_modify_qp(qp, attr,
265 			   IB_QP_STATE		|
266 			   IB_QP_PKEY_INDEX	|
267 			   IB_QP_ACCESS_FLAGS	|
268 			   IB_QP_PORT);
269 
270 out:
271 	kfree(attr);
272 	return ret;
273 }
274 
275 static int srp_new_cm_id(struct srp_target_port *target)
276 {
277 	struct ib_cm_id *new_cm_id;
278 
279 	new_cm_id = ib_create_cm_id(target->srp_host->srp_dev->dev,
280 				    srp_cm_handler, target);
281 	if (IS_ERR(new_cm_id))
282 		return PTR_ERR(new_cm_id);
283 
284 	if (target->cm_id)
285 		ib_destroy_cm_id(target->cm_id);
286 	target->cm_id = new_cm_id;
287 
288 	return 0;
289 }
290 
291 static int srp_create_target_ib(struct srp_target_port *target)
292 {
293 	struct ib_qp_init_attr *init_attr;
294 	struct ib_cq *recv_cq, *send_cq;
295 	struct ib_qp *qp;
296 	int ret;
297 
298 	init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
299 	if (!init_attr)
300 		return -ENOMEM;
301 
302 	recv_cq = ib_create_cq(target->srp_host->srp_dev->dev,
303 			       srp_recv_completion, NULL, target,
304 			       target->queue_size, target->comp_vector);
305 	if (IS_ERR(recv_cq)) {
306 		ret = PTR_ERR(recv_cq);
307 		goto err;
308 	}
309 
310 	send_cq = ib_create_cq(target->srp_host->srp_dev->dev,
311 			       srp_send_completion, NULL, target,
312 			       target->queue_size, target->comp_vector);
313 	if (IS_ERR(send_cq)) {
314 		ret = PTR_ERR(send_cq);
315 		goto err_recv_cq;
316 	}
317 
318 	ib_req_notify_cq(recv_cq, IB_CQ_NEXT_COMP);
319 
320 	init_attr->event_handler       = srp_qp_event;
321 	init_attr->cap.max_send_wr     = target->queue_size;
322 	init_attr->cap.max_recv_wr     = target->queue_size;
323 	init_attr->cap.max_recv_sge    = 1;
324 	init_attr->cap.max_send_sge    = 1;
325 	init_attr->sq_sig_type         = IB_SIGNAL_ALL_WR;
326 	init_attr->qp_type             = IB_QPT_RC;
327 	init_attr->send_cq             = send_cq;
328 	init_attr->recv_cq             = recv_cq;
329 
330 	qp = ib_create_qp(target->srp_host->srp_dev->pd, init_attr);
331 	if (IS_ERR(qp)) {
332 		ret = PTR_ERR(qp);
333 		goto err_send_cq;
334 	}
335 
336 	ret = srp_init_qp(target, qp);
337 	if (ret)
338 		goto err_qp;
339 
340 	if (target->qp)
341 		ib_destroy_qp(target->qp);
342 	if (target->recv_cq)
343 		ib_destroy_cq(target->recv_cq);
344 	if (target->send_cq)
345 		ib_destroy_cq(target->send_cq);
346 
347 	target->qp = qp;
348 	target->recv_cq = recv_cq;
349 	target->send_cq = send_cq;
350 
351 	kfree(init_attr);
352 	return 0;
353 
354 err_qp:
355 	ib_destroy_qp(qp);
356 
357 err_send_cq:
358 	ib_destroy_cq(send_cq);
359 
360 err_recv_cq:
361 	ib_destroy_cq(recv_cq);
362 
363 err:
364 	kfree(init_attr);
365 	return ret;
366 }
367 
368 /*
369  * Note: this function may be called without srp_alloc_iu_bufs() having been
370  * invoked. Hence the target->[rt]x_ring checks.
371  */
372 static void srp_free_target_ib(struct srp_target_port *target)
373 {
374 	int i;
375 
376 	ib_destroy_qp(target->qp);
377 	ib_destroy_cq(target->send_cq);
378 	ib_destroy_cq(target->recv_cq);
379 
380 	target->qp = NULL;
381 	target->send_cq = target->recv_cq = NULL;
382 
383 	if (target->rx_ring) {
384 		for (i = 0; i < target->queue_size; ++i)
385 			srp_free_iu(target->srp_host, target->rx_ring[i]);
386 		kfree(target->rx_ring);
387 		target->rx_ring = NULL;
388 	}
389 	if (target->tx_ring) {
390 		for (i = 0; i < target->queue_size; ++i)
391 			srp_free_iu(target->srp_host, target->tx_ring[i]);
392 		kfree(target->tx_ring);
393 		target->tx_ring = NULL;
394 	}
395 }
396 
397 static void srp_path_rec_completion(int status,
398 				    struct ib_sa_path_rec *pathrec,
399 				    void *target_ptr)
400 {
401 	struct srp_target_port *target = target_ptr;
402 
403 	target->status = status;
404 	if (status)
405 		shost_printk(KERN_ERR, target->scsi_host,
406 			     PFX "Got failed path rec status %d\n", status);
407 	else
408 		target->path = *pathrec;
409 	complete(&target->done);
410 }
411 
412 static int srp_lookup_path(struct srp_target_port *target)
413 {
414 	target->path.numb_path = 1;
415 
416 	init_completion(&target->done);
417 
418 	target->path_query_id = ib_sa_path_rec_get(&srp_sa_client,
419 						   target->srp_host->srp_dev->dev,
420 						   target->srp_host->port,
421 						   &target->path,
422 						   IB_SA_PATH_REC_SERVICE_ID	|
423 						   IB_SA_PATH_REC_DGID		|
424 						   IB_SA_PATH_REC_SGID		|
425 						   IB_SA_PATH_REC_NUMB_PATH	|
426 						   IB_SA_PATH_REC_PKEY,
427 						   SRP_PATH_REC_TIMEOUT_MS,
428 						   GFP_KERNEL,
429 						   srp_path_rec_completion,
430 						   target, &target->path_query);
431 	if (target->path_query_id < 0)
432 		return target->path_query_id;
433 
434 	wait_for_completion(&target->done);
435 
436 	if (target->status < 0)
437 		shost_printk(KERN_WARNING, target->scsi_host,
438 			     PFX "Path record query failed\n");
439 
440 	return target->status;
441 }
442 
443 static int srp_send_req(struct srp_target_port *target)
444 {
445 	struct {
446 		struct ib_cm_req_param param;
447 		struct srp_login_req   priv;
448 	} *req = NULL;
449 	int status;
450 
451 	req = kzalloc(sizeof *req, GFP_KERNEL);
452 	if (!req)
453 		return -ENOMEM;
454 
455 	req->param.primary_path 	      = &target->path;
456 	req->param.alternate_path 	      = NULL;
457 	req->param.service_id 		      = target->service_id;
458 	req->param.qp_num 		      = target->qp->qp_num;
459 	req->param.qp_type 		      = target->qp->qp_type;
460 	req->param.private_data 	      = &req->priv;
461 	req->param.private_data_len 	      = sizeof req->priv;
462 	req->param.flow_control 	      = 1;
463 
464 	get_random_bytes(&req->param.starting_psn, 4);
465 	req->param.starting_psn 	     &= 0xffffff;
466 
467 	/*
468 	 * Pick some arbitrary defaults here; we could make these
469 	 * module parameters if anyone cared about setting them.
470 	 */
471 	req->param.responder_resources	      = 4;
472 	req->param.remote_cm_response_timeout = 20;
473 	req->param.local_cm_response_timeout  = 20;
474 	req->param.retry_count                = target->tl_retry_count;
475 	req->param.rnr_retry_count 	      = 7;
476 	req->param.max_cm_retries 	      = 15;
477 
478 	req->priv.opcode     	= SRP_LOGIN_REQ;
479 	req->priv.tag        	= 0;
480 	req->priv.req_it_iu_len = cpu_to_be32(target->max_iu_len);
481 	req->priv.req_buf_fmt 	= cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
482 					      SRP_BUF_FORMAT_INDIRECT);
483 	/*
484 	 * In the published SRP specification (draft rev. 16a), the
485 	 * port identifier format is 8 bytes of ID extension followed
486 	 * by 8 bytes of GUID.  Older drafts put the two halves in the
487 	 * opposite order, so that the GUID comes first.
488 	 *
489 	 * Targets conforming to these obsolete drafts can be
490 	 * recognized by the I/O Class they report.
491 	 */
492 	if (target->io_class == SRP_REV10_IB_IO_CLASS) {
493 		memcpy(req->priv.initiator_port_id,
494 		       &target->path.sgid.global.interface_id, 8);
495 		memcpy(req->priv.initiator_port_id + 8,
496 		       &target->initiator_ext, 8);
497 		memcpy(req->priv.target_port_id,     &target->ioc_guid, 8);
498 		memcpy(req->priv.target_port_id + 8, &target->id_ext, 8);
499 	} else {
500 		memcpy(req->priv.initiator_port_id,
501 		       &target->initiator_ext, 8);
502 		memcpy(req->priv.initiator_port_id + 8,
503 		       &target->path.sgid.global.interface_id, 8);
504 		memcpy(req->priv.target_port_id,     &target->id_ext, 8);
505 		memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
506 	}
507 
508 	/*
509 	 * Topspin/Cisco SRP targets will reject our login unless we
510 	 * zero out the first 8 bytes of our initiator port ID and set
511 	 * the second 8 bytes to the local node GUID.
512 	 */
513 	if (srp_target_is_topspin(target)) {
514 		shost_printk(KERN_DEBUG, target->scsi_host,
515 			     PFX "Topspin/Cisco initiator port ID workaround "
516 			     "activated for target GUID %016llx\n",
517 			     (unsigned long long) be64_to_cpu(target->ioc_guid));
518 		memset(req->priv.initiator_port_id, 0, 8);
519 		memcpy(req->priv.initiator_port_id + 8,
520 		       &target->srp_host->srp_dev->dev->node_guid, 8);
521 	}
522 
523 	status = ib_send_cm_req(target->cm_id, &req->param);
524 
525 	kfree(req);
526 
527 	return status;
528 }
529 
530 static bool srp_queue_remove_work(struct srp_target_port *target)
531 {
532 	bool changed = false;
533 
534 	spin_lock_irq(&target->lock);
535 	if (target->state != SRP_TARGET_REMOVED) {
536 		target->state = SRP_TARGET_REMOVED;
537 		changed = true;
538 	}
539 	spin_unlock_irq(&target->lock);
540 
541 	if (changed)
542 		queue_work(system_long_wq, &target->remove_work);
543 
544 	return changed;
545 }
546 
547 static bool srp_change_conn_state(struct srp_target_port *target,
548 				  bool connected)
549 {
550 	bool changed = false;
551 
552 	spin_lock_irq(&target->lock);
553 	if (target->connected != connected) {
554 		target->connected = connected;
555 		changed = true;
556 	}
557 	spin_unlock_irq(&target->lock);
558 
559 	return changed;
560 }
561 
562 static void srp_disconnect_target(struct srp_target_port *target)
563 {
564 	if (srp_change_conn_state(target, false)) {
565 		/* XXX should send SRP_I_LOGOUT request */
566 
567 		if (ib_send_cm_dreq(target->cm_id, NULL, 0)) {
568 			shost_printk(KERN_DEBUG, target->scsi_host,
569 				     PFX "Sending CM DREQ failed\n");
570 		}
571 	}
572 }
573 
574 static void srp_free_req_data(struct srp_target_port *target)
575 {
576 	struct ib_device *ibdev = target->srp_host->srp_dev->dev;
577 	struct srp_request *req;
578 	int i;
579 
580 	if (!target->req_ring)
581 		return;
582 
583 	for (i = 0; i < target->req_ring_size; ++i) {
584 		req = &target->req_ring[i];
585 		kfree(req->fmr_list);
586 		kfree(req->map_page);
587 		if (req->indirect_dma_addr) {
588 			ib_dma_unmap_single(ibdev, req->indirect_dma_addr,
589 					    target->indirect_size,
590 					    DMA_TO_DEVICE);
591 		}
592 		kfree(req->indirect_desc);
593 	}
594 
595 	kfree(target->req_ring);
596 	target->req_ring = NULL;
597 }
598 
599 static int srp_alloc_req_data(struct srp_target_port *target)
600 {
601 	struct srp_device *srp_dev = target->srp_host->srp_dev;
602 	struct ib_device *ibdev = srp_dev->dev;
603 	struct srp_request *req;
604 	dma_addr_t dma_addr;
605 	int i, ret = -ENOMEM;
606 
607 	INIT_LIST_HEAD(&target->free_reqs);
608 
609 	target->req_ring = kzalloc(target->req_ring_size *
610 				   sizeof(*target->req_ring), GFP_KERNEL);
611 	if (!target->req_ring)
612 		goto out;
613 
614 	for (i = 0; i < target->req_ring_size; ++i) {
615 		req = &target->req_ring[i];
616 		req->fmr_list = kmalloc(target->cmd_sg_cnt * sizeof(void *),
617 					GFP_KERNEL);
618 		req->map_page = kmalloc(SRP_FMR_SIZE * sizeof(void *),
619 					GFP_KERNEL);
620 		req->indirect_desc = kmalloc(target->indirect_size, GFP_KERNEL);
621 		if (!req->fmr_list || !req->map_page || !req->indirect_desc)
622 			goto out;
623 
624 		dma_addr = ib_dma_map_single(ibdev, req->indirect_desc,
625 					     target->indirect_size,
626 					     DMA_TO_DEVICE);
627 		if (ib_dma_mapping_error(ibdev, dma_addr))
628 			goto out;
629 
630 		req->indirect_dma_addr = dma_addr;
631 		req->index = i;
632 		list_add_tail(&req->list, &target->free_reqs);
633 	}
634 	ret = 0;
635 
636 out:
637 	return ret;
638 }
639 
640 /**
641  * srp_del_scsi_host_attr() - Remove attributes defined in the host template.
642  * @shost: SCSI host whose attributes to remove from sysfs.
643  *
644  * Note: Any attributes defined in the host template and that did not exist
645  * before invocation of this function will be ignored.
646  */
647 static void srp_del_scsi_host_attr(struct Scsi_Host *shost)
648 {
649 	struct device_attribute **attr;
650 
651 	for (attr = shost->hostt->shost_attrs; attr && *attr; ++attr)
652 		device_remove_file(&shost->shost_dev, *attr);
653 }
654 
655 static void srp_remove_target(struct srp_target_port *target)
656 {
657 	WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
658 
659 	srp_del_scsi_host_attr(target->scsi_host);
660 	srp_rport_get(target->rport);
661 	srp_remove_host(target->scsi_host);
662 	scsi_remove_host(target->scsi_host);
663 	srp_disconnect_target(target);
664 	ib_destroy_cm_id(target->cm_id);
665 	srp_free_target_ib(target);
666 	cancel_work_sync(&target->tl_err_work);
667 	srp_rport_put(target->rport);
668 	srp_free_req_data(target);
669 
670 	spin_lock(&target->srp_host->target_lock);
671 	list_del(&target->list);
672 	spin_unlock(&target->srp_host->target_lock);
673 
674 	scsi_host_put(target->scsi_host);
675 }
676 
677 static void srp_remove_work(struct work_struct *work)
678 {
679 	struct srp_target_port *target =
680 		container_of(work, struct srp_target_port, remove_work);
681 
682 	WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
683 
684 	srp_remove_target(target);
685 }
686 
687 static void srp_rport_delete(struct srp_rport *rport)
688 {
689 	struct srp_target_port *target = rport->lld_data;
690 
691 	srp_queue_remove_work(target);
692 }
693 
694 static int srp_connect_target(struct srp_target_port *target)
695 {
696 	int retries = 3;
697 	int ret;
698 
699 	WARN_ON_ONCE(target->connected);
700 
701 	target->qp_in_error = false;
702 
703 	ret = srp_lookup_path(target);
704 	if (ret)
705 		return ret;
706 
707 	while (1) {
708 		init_completion(&target->done);
709 		ret = srp_send_req(target);
710 		if (ret)
711 			return ret;
712 		wait_for_completion(&target->done);
713 
714 		/*
715 		 * The CM event handling code will set status to
716 		 * SRP_PORT_REDIRECT if we get a port redirect REJ
717 		 * back, or SRP_DLID_REDIRECT if we get a lid/qp
718 		 * redirect REJ back.
719 		 */
720 		switch (target->status) {
721 		case 0:
722 			srp_change_conn_state(target, true);
723 			return 0;
724 
725 		case SRP_PORT_REDIRECT:
726 			ret = srp_lookup_path(target);
727 			if (ret)
728 				return ret;
729 			break;
730 
731 		case SRP_DLID_REDIRECT:
732 			break;
733 
734 		case SRP_STALE_CONN:
735 			/* Our current CM id was stale, and is now in timewait.
736 			 * Try to reconnect with a new one.
737 			 */
738 			if (!retries-- || srp_new_cm_id(target)) {
739 				shost_printk(KERN_ERR, target->scsi_host, PFX
740 					     "giving up on stale connection\n");
741 				target->status = -ECONNRESET;
742 				return target->status;
743 			}
744 
745 			shost_printk(KERN_ERR, target->scsi_host, PFX
746 				     "retrying stale connection\n");
747 			break;
748 
749 		default:
750 			return target->status;
751 		}
752 	}
753 }
754 
755 static void srp_unmap_data(struct scsi_cmnd *scmnd,
756 			   struct srp_target_port *target,
757 			   struct srp_request *req)
758 {
759 	struct ib_device *ibdev = target->srp_host->srp_dev->dev;
760 	struct ib_pool_fmr **pfmr;
761 
762 	if (!scsi_sglist(scmnd) ||
763 	    (scmnd->sc_data_direction != DMA_TO_DEVICE &&
764 	     scmnd->sc_data_direction != DMA_FROM_DEVICE))
765 		return;
766 
767 	pfmr = req->fmr_list;
768 	while (req->nfmr--)
769 		ib_fmr_pool_unmap(*pfmr++);
770 
771 	ib_dma_unmap_sg(ibdev, scsi_sglist(scmnd), scsi_sg_count(scmnd),
772 			scmnd->sc_data_direction);
773 }
774 
775 /**
776  * srp_claim_req - Take ownership of the scmnd associated with a request.
777  * @target: SRP target port.
778  * @req: SRP request.
779  * @scmnd: If NULL, take ownership of @req->scmnd. If not NULL, only take
780  *         ownership of @req->scmnd if it equals @scmnd.
781  *
782  * Return value:
783  * Either NULL or a pointer to the SCSI command the caller became owner of.
784  */
785 static struct scsi_cmnd *srp_claim_req(struct srp_target_port *target,
786 				       struct srp_request *req,
787 				       struct scsi_cmnd *scmnd)
788 {
789 	unsigned long flags;
790 
791 	spin_lock_irqsave(&target->lock, flags);
792 	if (!scmnd) {
793 		scmnd = req->scmnd;
794 		req->scmnd = NULL;
795 	} else if (req->scmnd == scmnd) {
796 		req->scmnd = NULL;
797 	} else {
798 		scmnd = NULL;
799 	}
800 	spin_unlock_irqrestore(&target->lock, flags);
801 
802 	return scmnd;
803 }
804 
805 /**
806  * srp_free_req() - Unmap data and add request to the free request list.
807  */
808 static void srp_free_req(struct srp_target_port *target,
809 			 struct srp_request *req, struct scsi_cmnd *scmnd,
810 			 s32 req_lim_delta)
811 {
812 	unsigned long flags;
813 
814 	srp_unmap_data(scmnd, target, req);
815 
816 	spin_lock_irqsave(&target->lock, flags);
817 	target->req_lim += req_lim_delta;
818 	list_add_tail(&req->list, &target->free_reqs);
819 	spin_unlock_irqrestore(&target->lock, flags);
820 }
821 
822 static void srp_finish_req(struct srp_target_port *target,
823 			   struct srp_request *req, int result)
824 {
825 	struct scsi_cmnd *scmnd = srp_claim_req(target, req, NULL);
826 
827 	if (scmnd) {
828 		srp_free_req(target, req, scmnd, 0);
829 		scmnd->result = result;
830 		scmnd->scsi_done(scmnd);
831 	}
832 }
833 
834 static void srp_terminate_io(struct srp_rport *rport)
835 {
836 	struct srp_target_port *target = rport->lld_data;
837 	int i;
838 
839 	for (i = 0; i < target->req_ring_size; ++i) {
840 		struct srp_request *req = &target->req_ring[i];
841 		srp_finish_req(target, req, DID_TRANSPORT_FAILFAST << 16);
842 	}
843 }
844 
845 /*
846  * It is up to the caller to ensure that srp_rport_reconnect() calls are
847  * serialized and that no concurrent srp_queuecommand(), srp_abort(),
848  * srp_reset_device() or srp_reset_host() calls will occur while this function
849  * is in progress. One way to realize that is not to call this function
850  * directly but to call srp_reconnect_rport() instead since that last function
851  * serializes calls of this function via rport->mutex and also blocks
852  * srp_queuecommand() calls before invoking this function.
853  */
854 static int srp_rport_reconnect(struct srp_rport *rport)
855 {
856 	struct srp_target_port *target = rport->lld_data;
857 	int i, ret;
858 
859 	srp_disconnect_target(target);
860 	/*
861 	 * Now get a new local CM ID so that we avoid confusing the target in
862 	 * case things are really fouled up. Doing so also ensures that all CM
863 	 * callbacks will have finished before a new QP is allocated.
864 	 */
865 	ret = srp_new_cm_id(target);
866 	/*
867 	 * Whether or not creating a new CM ID succeeded, create a new
868 	 * QP. This guarantees that all completion callback function
869 	 * invocations have finished before request resetting starts.
870 	 */
871 	if (ret == 0)
872 		ret = srp_create_target_ib(target);
873 	else
874 		srp_create_target_ib(target);
875 
876 	for (i = 0; i < target->req_ring_size; ++i) {
877 		struct srp_request *req = &target->req_ring[i];
878 		srp_finish_req(target, req, DID_RESET << 16);
879 	}
880 
881 	INIT_LIST_HEAD(&target->free_tx);
882 	for (i = 0; i < target->queue_size; ++i)
883 		list_add(&target->tx_ring[i]->list, &target->free_tx);
884 
885 	if (ret == 0)
886 		ret = srp_connect_target(target);
887 
888 	if (ret == 0)
889 		shost_printk(KERN_INFO, target->scsi_host,
890 			     PFX "reconnect succeeded\n");
891 
892 	return ret;
893 }
894 
895 static void srp_map_desc(struct srp_map_state *state, dma_addr_t dma_addr,
896 			 unsigned int dma_len, u32 rkey)
897 {
898 	struct srp_direct_buf *desc = state->desc;
899 
900 	desc->va = cpu_to_be64(dma_addr);
901 	desc->key = cpu_to_be32(rkey);
902 	desc->len = cpu_to_be32(dma_len);
903 
904 	state->total_len += dma_len;
905 	state->desc++;
906 	state->ndesc++;
907 }
908 
909 static int srp_map_finish_fmr(struct srp_map_state *state,
910 			      struct srp_target_port *target)
911 {
912 	struct srp_device *dev = target->srp_host->srp_dev;
913 	struct ib_pool_fmr *fmr;
914 	u64 io_addr = 0;
915 
916 	if (!state->npages)
917 		return 0;
918 
919 	if (state->npages == 1) {
920 		srp_map_desc(state, state->base_dma_addr, state->fmr_len,
921 			     target->rkey);
922 		state->npages = state->fmr_len = 0;
923 		return 0;
924 	}
925 
926 	fmr = ib_fmr_pool_map_phys(dev->fmr_pool, state->pages,
927 				   state->npages, io_addr);
928 	if (IS_ERR(fmr))
929 		return PTR_ERR(fmr);
930 
931 	*state->next_fmr++ = fmr;
932 	state->nfmr++;
933 
934 	srp_map_desc(state, 0, state->fmr_len, fmr->fmr->rkey);
935 	state->npages = state->fmr_len = 0;
936 	return 0;
937 }
938 
939 static void srp_map_update_start(struct srp_map_state *state,
940 				 struct scatterlist *sg, int sg_index,
941 				 dma_addr_t dma_addr)
942 {
943 	state->unmapped_sg = sg;
944 	state->unmapped_index = sg_index;
945 	state->unmapped_addr = dma_addr;
946 }
947 
948 static int srp_map_sg_entry(struct srp_map_state *state,
949 			    struct srp_target_port *target,
950 			    struct scatterlist *sg, int sg_index,
951 			    int use_fmr)
952 {
953 	struct srp_device *dev = target->srp_host->srp_dev;
954 	struct ib_device *ibdev = dev->dev;
955 	dma_addr_t dma_addr = ib_sg_dma_address(ibdev, sg);
956 	unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
957 	unsigned int len;
958 	int ret;
959 
960 	if (!dma_len)
961 		return 0;
962 
963 	if (use_fmr == SRP_MAP_NO_FMR) {
964 		/* Once we're in direct map mode for a request, we don't
965 		 * go back to FMR mode, so no need to update anything
966 		 * other than the descriptor.
967 		 */
968 		srp_map_desc(state, dma_addr, dma_len, target->rkey);
969 		return 0;
970 	}
971 
972 	/* If we start at an offset into the FMR page, don't merge into
973 	 * the current FMR. Finish it out, and use the kernel's MR for this
974 	 * sg entry. This is to avoid potential bugs on some SRP targets
975 	 * that were never quite defined, but went away when the initiator
976 	 * avoided using FMR on such page fragments.
977 	 */
978 	if (dma_addr & ~dev->fmr_page_mask || dma_len > dev->fmr_max_size) {
979 		ret = srp_map_finish_fmr(state, target);
980 		if (ret)
981 			return ret;
982 
983 		srp_map_desc(state, dma_addr, dma_len, target->rkey);
984 		srp_map_update_start(state, NULL, 0, 0);
985 		return 0;
986 	}
987 
988 	/* If this is the first sg to go into the FMR, save our position.
989 	 * We need to know the first unmapped entry, its index, and the
990 	 * first unmapped address within that entry to be able to restart
991 	 * mapping after an error.
992 	 */
993 	if (!state->unmapped_sg)
994 		srp_map_update_start(state, sg, sg_index, dma_addr);
995 
996 	while (dma_len) {
997 		if (state->npages == SRP_FMR_SIZE) {
998 			ret = srp_map_finish_fmr(state, target);
999 			if (ret)
1000 				return ret;
1001 
1002 			srp_map_update_start(state, sg, sg_index, dma_addr);
1003 		}
1004 
1005 		len = min_t(unsigned int, dma_len, dev->fmr_page_size);
1006 
1007 		if (!state->npages)
1008 			state->base_dma_addr = dma_addr;
1009 		state->pages[state->npages++] = dma_addr;
1010 		state->fmr_len += len;
1011 		dma_addr += len;
1012 		dma_len -= len;
1013 	}
1014 
1015 	/* If the last entry of the FMR wasn't a full page, then we need to
1016 	 * close it out and start a new one -- we can only merge at page
1017 	 * boundries.
1018 	 */
1019 	ret = 0;
1020 	if (len != dev->fmr_page_size) {
1021 		ret = srp_map_finish_fmr(state, target);
1022 		if (!ret)
1023 			srp_map_update_start(state, NULL, 0, 0);
1024 	}
1025 	return ret;
1026 }
1027 
1028 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_target_port *target,
1029 			struct srp_request *req)
1030 {
1031 	struct scatterlist *scat, *sg;
1032 	struct srp_cmd *cmd = req->cmd->buf;
1033 	int i, len, nents, count, use_fmr;
1034 	struct srp_device *dev;
1035 	struct ib_device *ibdev;
1036 	struct srp_map_state state;
1037 	struct srp_indirect_buf *indirect_hdr;
1038 	u32 table_len;
1039 	u8 fmt;
1040 
1041 	if (!scsi_sglist(scmnd) || scmnd->sc_data_direction == DMA_NONE)
1042 		return sizeof (struct srp_cmd);
1043 
1044 	if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
1045 	    scmnd->sc_data_direction != DMA_TO_DEVICE) {
1046 		shost_printk(KERN_WARNING, target->scsi_host,
1047 			     PFX "Unhandled data direction %d\n",
1048 			     scmnd->sc_data_direction);
1049 		return -EINVAL;
1050 	}
1051 
1052 	nents = scsi_sg_count(scmnd);
1053 	scat  = scsi_sglist(scmnd);
1054 
1055 	dev = target->srp_host->srp_dev;
1056 	ibdev = dev->dev;
1057 
1058 	count = ib_dma_map_sg(ibdev, scat, nents, scmnd->sc_data_direction);
1059 	if (unlikely(count == 0))
1060 		return -EIO;
1061 
1062 	fmt = SRP_DATA_DESC_DIRECT;
1063 	len = sizeof (struct srp_cmd) +	sizeof (struct srp_direct_buf);
1064 
1065 	if (count == 1) {
1066 		/*
1067 		 * The midlayer only generated a single gather/scatter
1068 		 * entry, or DMA mapping coalesced everything to a
1069 		 * single entry.  So a direct descriptor along with
1070 		 * the DMA MR suffices.
1071 		 */
1072 		struct srp_direct_buf *buf = (void *) cmd->add_data;
1073 
1074 		buf->va  = cpu_to_be64(ib_sg_dma_address(ibdev, scat));
1075 		buf->key = cpu_to_be32(target->rkey);
1076 		buf->len = cpu_to_be32(ib_sg_dma_len(ibdev, scat));
1077 
1078 		req->nfmr = 0;
1079 		goto map_complete;
1080 	}
1081 
1082 	/* We have more than one scatter/gather entry, so build our indirect
1083 	 * descriptor table, trying to merge as many entries with FMR as we
1084 	 * can.
1085 	 */
1086 	indirect_hdr = (void *) cmd->add_data;
1087 
1088 	ib_dma_sync_single_for_cpu(ibdev, req->indirect_dma_addr,
1089 				   target->indirect_size, DMA_TO_DEVICE);
1090 
1091 	memset(&state, 0, sizeof(state));
1092 	state.desc	= req->indirect_desc;
1093 	state.pages	= req->map_page;
1094 	state.next_fmr	= req->fmr_list;
1095 
1096 	use_fmr = dev->fmr_pool ? SRP_MAP_ALLOW_FMR : SRP_MAP_NO_FMR;
1097 
1098 	for_each_sg(scat, sg, count, i) {
1099 		if (srp_map_sg_entry(&state, target, sg, i, use_fmr)) {
1100 			/* FMR mapping failed, so backtrack to the first
1101 			 * unmapped entry and continue on without using FMR.
1102 			 */
1103 			dma_addr_t dma_addr;
1104 			unsigned int dma_len;
1105 
1106 backtrack:
1107 			sg = state.unmapped_sg;
1108 			i = state.unmapped_index;
1109 
1110 			dma_addr = ib_sg_dma_address(ibdev, sg);
1111 			dma_len = ib_sg_dma_len(ibdev, sg);
1112 			dma_len -= (state.unmapped_addr - dma_addr);
1113 			dma_addr = state.unmapped_addr;
1114 			use_fmr = SRP_MAP_NO_FMR;
1115 			srp_map_desc(&state, dma_addr, dma_len, target->rkey);
1116 		}
1117 	}
1118 
1119 	if (use_fmr == SRP_MAP_ALLOW_FMR && srp_map_finish_fmr(&state, target))
1120 		goto backtrack;
1121 
1122 	/* We've mapped the request, now pull as much of the indirect
1123 	 * descriptor table as we can into the command buffer. If this
1124 	 * target is not using an external indirect table, we are
1125 	 * guaranteed to fit into the command, as the SCSI layer won't
1126 	 * give us more S/G entries than we allow.
1127 	 */
1128 	req->nfmr = state.nfmr;
1129 	if (state.ndesc == 1) {
1130 		/* FMR mapping was able to collapse this to one entry,
1131 		 * so use a direct descriptor.
1132 		 */
1133 		struct srp_direct_buf *buf = (void *) cmd->add_data;
1134 
1135 		*buf = req->indirect_desc[0];
1136 		goto map_complete;
1137 	}
1138 
1139 	if (unlikely(target->cmd_sg_cnt < state.ndesc &&
1140 						!target->allow_ext_sg)) {
1141 		shost_printk(KERN_ERR, target->scsi_host,
1142 			     "Could not fit S/G list into SRP_CMD\n");
1143 		return -EIO;
1144 	}
1145 
1146 	count = min(state.ndesc, target->cmd_sg_cnt);
1147 	table_len = state.ndesc * sizeof (struct srp_direct_buf);
1148 
1149 	fmt = SRP_DATA_DESC_INDIRECT;
1150 	len = sizeof(struct srp_cmd) + sizeof (struct srp_indirect_buf);
1151 	len += count * sizeof (struct srp_direct_buf);
1152 
1153 	memcpy(indirect_hdr->desc_list, req->indirect_desc,
1154 	       count * sizeof (struct srp_direct_buf));
1155 
1156 	indirect_hdr->table_desc.va = cpu_to_be64(req->indirect_dma_addr);
1157 	indirect_hdr->table_desc.key = cpu_to_be32(target->rkey);
1158 	indirect_hdr->table_desc.len = cpu_to_be32(table_len);
1159 	indirect_hdr->len = cpu_to_be32(state.total_len);
1160 
1161 	if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1162 		cmd->data_out_desc_cnt = count;
1163 	else
1164 		cmd->data_in_desc_cnt = count;
1165 
1166 	ib_dma_sync_single_for_device(ibdev, req->indirect_dma_addr, table_len,
1167 				      DMA_TO_DEVICE);
1168 
1169 map_complete:
1170 	if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1171 		cmd->buf_fmt = fmt << 4;
1172 	else
1173 		cmd->buf_fmt = fmt;
1174 
1175 	return len;
1176 }
1177 
1178 /*
1179  * Return an IU and possible credit to the free pool
1180  */
1181 static void srp_put_tx_iu(struct srp_target_port *target, struct srp_iu *iu,
1182 			  enum srp_iu_type iu_type)
1183 {
1184 	unsigned long flags;
1185 
1186 	spin_lock_irqsave(&target->lock, flags);
1187 	list_add(&iu->list, &target->free_tx);
1188 	if (iu_type != SRP_IU_RSP)
1189 		++target->req_lim;
1190 	spin_unlock_irqrestore(&target->lock, flags);
1191 }
1192 
1193 /*
1194  * Must be called with target->lock held to protect req_lim and free_tx.
1195  * If IU is not sent, it must be returned using srp_put_tx_iu().
1196  *
1197  * Note:
1198  * An upper limit for the number of allocated information units for each
1199  * request type is:
1200  * - SRP_IU_CMD: SRP_CMD_SQ_SIZE, since the SCSI mid-layer never queues
1201  *   more than Scsi_Host.can_queue requests.
1202  * - SRP_IU_TSK_MGMT: SRP_TSK_MGMT_SQ_SIZE.
1203  * - SRP_IU_RSP: 1, since a conforming SRP target never sends more than
1204  *   one unanswered SRP request to an initiator.
1205  */
1206 static struct srp_iu *__srp_get_tx_iu(struct srp_target_port *target,
1207 				      enum srp_iu_type iu_type)
1208 {
1209 	s32 rsv = (iu_type == SRP_IU_TSK_MGMT) ? 0 : SRP_TSK_MGMT_SQ_SIZE;
1210 	struct srp_iu *iu;
1211 
1212 	srp_send_completion(target->send_cq, target);
1213 
1214 	if (list_empty(&target->free_tx))
1215 		return NULL;
1216 
1217 	/* Initiator responses to target requests do not consume credits */
1218 	if (iu_type != SRP_IU_RSP) {
1219 		if (target->req_lim <= rsv) {
1220 			++target->zero_req_lim;
1221 			return NULL;
1222 		}
1223 
1224 		--target->req_lim;
1225 	}
1226 
1227 	iu = list_first_entry(&target->free_tx, struct srp_iu, list);
1228 	list_del(&iu->list);
1229 	return iu;
1230 }
1231 
1232 static int srp_post_send(struct srp_target_port *target,
1233 			 struct srp_iu *iu, int len)
1234 {
1235 	struct ib_sge list;
1236 	struct ib_send_wr wr, *bad_wr;
1237 
1238 	list.addr   = iu->dma;
1239 	list.length = len;
1240 	list.lkey   = target->lkey;
1241 
1242 	wr.next       = NULL;
1243 	wr.wr_id      = (uintptr_t) iu;
1244 	wr.sg_list    = &list;
1245 	wr.num_sge    = 1;
1246 	wr.opcode     = IB_WR_SEND;
1247 	wr.send_flags = IB_SEND_SIGNALED;
1248 
1249 	return ib_post_send(target->qp, &wr, &bad_wr);
1250 }
1251 
1252 static int srp_post_recv(struct srp_target_port *target, struct srp_iu *iu)
1253 {
1254 	struct ib_recv_wr wr, *bad_wr;
1255 	struct ib_sge list;
1256 
1257 	list.addr   = iu->dma;
1258 	list.length = iu->size;
1259 	list.lkey   = target->lkey;
1260 
1261 	wr.next     = NULL;
1262 	wr.wr_id    = (uintptr_t) iu;
1263 	wr.sg_list  = &list;
1264 	wr.num_sge  = 1;
1265 
1266 	return ib_post_recv(target->qp, &wr, &bad_wr);
1267 }
1268 
1269 static void srp_process_rsp(struct srp_target_port *target, struct srp_rsp *rsp)
1270 {
1271 	struct srp_request *req;
1272 	struct scsi_cmnd *scmnd;
1273 	unsigned long flags;
1274 
1275 	if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
1276 		spin_lock_irqsave(&target->lock, flags);
1277 		target->req_lim += be32_to_cpu(rsp->req_lim_delta);
1278 		spin_unlock_irqrestore(&target->lock, flags);
1279 
1280 		target->tsk_mgmt_status = -1;
1281 		if (be32_to_cpu(rsp->resp_data_len) >= 4)
1282 			target->tsk_mgmt_status = rsp->data[3];
1283 		complete(&target->tsk_mgmt_done);
1284 	} else {
1285 		req = &target->req_ring[rsp->tag];
1286 		scmnd = srp_claim_req(target, req, NULL);
1287 		if (!scmnd) {
1288 			shost_printk(KERN_ERR, target->scsi_host,
1289 				     "Null scmnd for RSP w/tag %016llx\n",
1290 				     (unsigned long long) rsp->tag);
1291 
1292 			spin_lock_irqsave(&target->lock, flags);
1293 			target->req_lim += be32_to_cpu(rsp->req_lim_delta);
1294 			spin_unlock_irqrestore(&target->lock, flags);
1295 
1296 			return;
1297 		}
1298 		scmnd->result = rsp->status;
1299 
1300 		if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
1301 			memcpy(scmnd->sense_buffer, rsp->data +
1302 			       be32_to_cpu(rsp->resp_data_len),
1303 			       min_t(int, be32_to_cpu(rsp->sense_data_len),
1304 				     SCSI_SENSE_BUFFERSIZE));
1305 		}
1306 
1307 		if (rsp->flags & (SRP_RSP_FLAG_DOOVER | SRP_RSP_FLAG_DOUNDER))
1308 			scsi_set_resid(scmnd, be32_to_cpu(rsp->data_out_res_cnt));
1309 		else if (rsp->flags & (SRP_RSP_FLAG_DIOVER | SRP_RSP_FLAG_DIUNDER))
1310 			scsi_set_resid(scmnd, be32_to_cpu(rsp->data_in_res_cnt));
1311 
1312 		srp_free_req(target, req, scmnd,
1313 			     be32_to_cpu(rsp->req_lim_delta));
1314 
1315 		scmnd->host_scribble = NULL;
1316 		scmnd->scsi_done(scmnd);
1317 	}
1318 }
1319 
1320 static int srp_response_common(struct srp_target_port *target, s32 req_delta,
1321 			       void *rsp, int len)
1322 {
1323 	struct ib_device *dev = target->srp_host->srp_dev->dev;
1324 	unsigned long flags;
1325 	struct srp_iu *iu;
1326 	int err;
1327 
1328 	spin_lock_irqsave(&target->lock, flags);
1329 	target->req_lim += req_delta;
1330 	iu = __srp_get_tx_iu(target, SRP_IU_RSP);
1331 	spin_unlock_irqrestore(&target->lock, flags);
1332 
1333 	if (!iu) {
1334 		shost_printk(KERN_ERR, target->scsi_host, PFX
1335 			     "no IU available to send response\n");
1336 		return 1;
1337 	}
1338 
1339 	ib_dma_sync_single_for_cpu(dev, iu->dma, len, DMA_TO_DEVICE);
1340 	memcpy(iu->buf, rsp, len);
1341 	ib_dma_sync_single_for_device(dev, iu->dma, len, DMA_TO_DEVICE);
1342 
1343 	err = srp_post_send(target, iu, len);
1344 	if (err) {
1345 		shost_printk(KERN_ERR, target->scsi_host, PFX
1346 			     "unable to post response: %d\n", err);
1347 		srp_put_tx_iu(target, iu, SRP_IU_RSP);
1348 	}
1349 
1350 	return err;
1351 }
1352 
1353 static void srp_process_cred_req(struct srp_target_port *target,
1354 				 struct srp_cred_req *req)
1355 {
1356 	struct srp_cred_rsp rsp = {
1357 		.opcode = SRP_CRED_RSP,
1358 		.tag = req->tag,
1359 	};
1360 	s32 delta = be32_to_cpu(req->req_lim_delta);
1361 
1362 	if (srp_response_common(target, delta, &rsp, sizeof rsp))
1363 		shost_printk(KERN_ERR, target->scsi_host, PFX
1364 			     "problems processing SRP_CRED_REQ\n");
1365 }
1366 
1367 static void srp_process_aer_req(struct srp_target_port *target,
1368 				struct srp_aer_req *req)
1369 {
1370 	struct srp_aer_rsp rsp = {
1371 		.opcode = SRP_AER_RSP,
1372 		.tag = req->tag,
1373 	};
1374 	s32 delta = be32_to_cpu(req->req_lim_delta);
1375 
1376 	shost_printk(KERN_ERR, target->scsi_host, PFX
1377 		     "ignoring AER for LUN %llu\n", be64_to_cpu(req->lun));
1378 
1379 	if (srp_response_common(target, delta, &rsp, sizeof rsp))
1380 		shost_printk(KERN_ERR, target->scsi_host, PFX
1381 			     "problems processing SRP_AER_REQ\n");
1382 }
1383 
1384 static void srp_handle_recv(struct srp_target_port *target, struct ib_wc *wc)
1385 {
1386 	struct ib_device *dev = target->srp_host->srp_dev->dev;
1387 	struct srp_iu *iu = (struct srp_iu *) (uintptr_t) wc->wr_id;
1388 	int res;
1389 	u8 opcode;
1390 
1391 	ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_ti_iu_len,
1392 				   DMA_FROM_DEVICE);
1393 
1394 	opcode = *(u8 *) iu->buf;
1395 
1396 	if (0) {
1397 		shost_printk(KERN_ERR, target->scsi_host,
1398 			     PFX "recv completion, opcode 0x%02x\n", opcode);
1399 		print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 8, 1,
1400 			       iu->buf, wc->byte_len, true);
1401 	}
1402 
1403 	switch (opcode) {
1404 	case SRP_RSP:
1405 		srp_process_rsp(target, iu->buf);
1406 		break;
1407 
1408 	case SRP_CRED_REQ:
1409 		srp_process_cred_req(target, iu->buf);
1410 		break;
1411 
1412 	case SRP_AER_REQ:
1413 		srp_process_aer_req(target, iu->buf);
1414 		break;
1415 
1416 	case SRP_T_LOGOUT:
1417 		/* XXX Handle target logout */
1418 		shost_printk(KERN_WARNING, target->scsi_host,
1419 			     PFX "Got target logout request\n");
1420 		break;
1421 
1422 	default:
1423 		shost_printk(KERN_WARNING, target->scsi_host,
1424 			     PFX "Unhandled SRP opcode 0x%02x\n", opcode);
1425 		break;
1426 	}
1427 
1428 	ib_dma_sync_single_for_device(dev, iu->dma, target->max_ti_iu_len,
1429 				      DMA_FROM_DEVICE);
1430 
1431 	res = srp_post_recv(target, iu);
1432 	if (res != 0)
1433 		shost_printk(KERN_ERR, target->scsi_host,
1434 			     PFX "Recv failed with error code %d\n", res);
1435 }
1436 
1437 /**
1438  * srp_tl_err_work() - handle a transport layer error
1439  *
1440  * Note: This function may get invoked before the rport has been created,
1441  * hence the target->rport test.
1442  */
1443 static void srp_tl_err_work(struct work_struct *work)
1444 {
1445 	struct srp_target_port *target;
1446 
1447 	target = container_of(work, struct srp_target_port, tl_err_work);
1448 	if (target->rport)
1449 		srp_start_tl_fail_timers(target->rport);
1450 }
1451 
1452 static void srp_handle_qp_err(enum ib_wc_status wc_status, bool send_err,
1453 			      struct srp_target_port *target)
1454 {
1455 	if (target->connected && !target->qp_in_error) {
1456 		shost_printk(KERN_ERR, target->scsi_host,
1457 			     PFX "failed %s status %d\n",
1458 			     send_err ? "send" : "receive",
1459 			     wc_status);
1460 		queue_work(system_long_wq, &target->tl_err_work);
1461 	}
1462 	target->qp_in_error = true;
1463 }
1464 
1465 static void srp_recv_completion(struct ib_cq *cq, void *target_ptr)
1466 {
1467 	struct srp_target_port *target = target_ptr;
1468 	struct ib_wc wc;
1469 
1470 	ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1471 	while (ib_poll_cq(cq, 1, &wc) > 0) {
1472 		if (likely(wc.status == IB_WC_SUCCESS)) {
1473 			srp_handle_recv(target, &wc);
1474 		} else {
1475 			srp_handle_qp_err(wc.status, false, target);
1476 		}
1477 	}
1478 }
1479 
1480 static void srp_send_completion(struct ib_cq *cq, void *target_ptr)
1481 {
1482 	struct srp_target_port *target = target_ptr;
1483 	struct ib_wc wc;
1484 	struct srp_iu *iu;
1485 
1486 	while (ib_poll_cq(cq, 1, &wc) > 0) {
1487 		if (likely(wc.status == IB_WC_SUCCESS)) {
1488 			iu = (struct srp_iu *) (uintptr_t) wc.wr_id;
1489 			list_add(&iu->list, &target->free_tx);
1490 		} else {
1491 			srp_handle_qp_err(wc.status, true, target);
1492 		}
1493 	}
1494 }
1495 
1496 static int srp_queuecommand(struct Scsi_Host *shost, struct scsi_cmnd *scmnd)
1497 {
1498 	struct srp_target_port *target = host_to_target(shost);
1499 	struct srp_rport *rport = target->rport;
1500 	struct srp_request *req;
1501 	struct srp_iu *iu;
1502 	struct srp_cmd *cmd;
1503 	struct ib_device *dev;
1504 	unsigned long flags;
1505 	int len, result;
1506 	const bool in_scsi_eh = !in_interrupt() && current == shost->ehandler;
1507 
1508 	/*
1509 	 * The SCSI EH thread is the only context from which srp_queuecommand()
1510 	 * can get invoked for blocked devices (SDEV_BLOCK /
1511 	 * SDEV_CREATED_BLOCK). Avoid racing with srp_reconnect_rport() by
1512 	 * locking the rport mutex if invoked from inside the SCSI EH.
1513 	 */
1514 	if (in_scsi_eh)
1515 		mutex_lock(&rport->mutex);
1516 
1517 	result = srp_chkready(target->rport);
1518 	if (unlikely(result)) {
1519 		scmnd->result = result;
1520 		scmnd->scsi_done(scmnd);
1521 		goto unlock_rport;
1522 	}
1523 
1524 	spin_lock_irqsave(&target->lock, flags);
1525 	iu = __srp_get_tx_iu(target, SRP_IU_CMD);
1526 	if (!iu)
1527 		goto err_unlock;
1528 
1529 	req = list_first_entry(&target->free_reqs, struct srp_request, list);
1530 	list_del(&req->list);
1531 	spin_unlock_irqrestore(&target->lock, flags);
1532 
1533 	dev = target->srp_host->srp_dev->dev;
1534 	ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_iu_len,
1535 				   DMA_TO_DEVICE);
1536 
1537 	scmnd->result        = 0;
1538 	scmnd->host_scribble = (void *) req;
1539 
1540 	cmd = iu->buf;
1541 	memset(cmd, 0, sizeof *cmd);
1542 
1543 	cmd->opcode = SRP_CMD;
1544 	cmd->lun    = cpu_to_be64((u64) scmnd->device->lun << 48);
1545 	cmd->tag    = req->index;
1546 	memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
1547 
1548 	req->scmnd    = scmnd;
1549 	req->cmd      = iu;
1550 
1551 	len = srp_map_data(scmnd, target, req);
1552 	if (len < 0) {
1553 		shost_printk(KERN_ERR, target->scsi_host,
1554 			     PFX "Failed to map data\n");
1555 		goto err_iu;
1556 	}
1557 
1558 	ib_dma_sync_single_for_device(dev, iu->dma, target->max_iu_len,
1559 				      DMA_TO_DEVICE);
1560 
1561 	if (srp_post_send(target, iu, len)) {
1562 		shost_printk(KERN_ERR, target->scsi_host, PFX "Send failed\n");
1563 		goto err_unmap;
1564 	}
1565 
1566 unlock_rport:
1567 	if (in_scsi_eh)
1568 		mutex_unlock(&rport->mutex);
1569 
1570 	return 0;
1571 
1572 err_unmap:
1573 	srp_unmap_data(scmnd, target, req);
1574 
1575 err_iu:
1576 	srp_put_tx_iu(target, iu, SRP_IU_CMD);
1577 
1578 	spin_lock_irqsave(&target->lock, flags);
1579 	list_add(&req->list, &target->free_reqs);
1580 
1581 err_unlock:
1582 	spin_unlock_irqrestore(&target->lock, flags);
1583 
1584 	if (in_scsi_eh)
1585 		mutex_unlock(&rport->mutex);
1586 
1587 	return SCSI_MLQUEUE_HOST_BUSY;
1588 }
1589 
1590 /*
1591  * Note: the resources allocated in this function are freed in
1592  * srp_free_target_ib().
1593  */
1594 static int srp_alloc_iu_bufs(struct srp_target_port *target)
1595 {
1596 	int i;
1597 
1598 	target->rx_ring = kzalloc(target->queue_size * sizeof(*target->rx_ring),
1599 				  GFP_KERNEL);
1600 	if (!target->rx_ring)
1601 		goto err_no_ring;
1602 	target->tx_ring = kzalloc(target->queue_size * sizeof(*target->tx_ring),
1603 				  GFP_KERNEL);
1604 	if (!target->tx_ring)
1605 		goto err_no_ring;
1606 
1607 	for (i = 0; i < target->queue_size; ++i) {
1608 		target->rx_ring[i] = srp_alloc_iu(target->srp_host,
1609 						  target->max_ti_iu_len,
1610 						  GFP_KERNEL, DMA_FROM_DEVICE);
1611 		if (!target->rx_ring[i])
1612 			goto err;
1613 	}
1614 
1615 	for (i = 0; i < target->queue_size; ++i) {
1616 		target->tx_ring[i] = srp_alloc_iu(target->srp_host,
1617 						  target->max_iu_len,
1618 						  GFP_KERNEL, DMA_TO_DEVICE);
1619 		if (!target->tx_ring[i])
1620 			goto err;
1621 
1622 		list_add(&target->tx_ring[i]->list, &target->free_tx);
1623 	}
1624 
1625 	return 0;
1626 
1627 err:
1628 	for (i = 0; i < target->queue_size; ++i) {
1629 		srp_free_iu(target->srp_host, target->rx_ring[i]);
1630 		srp_free_iu(target->srp_host, target->tx_ring[i]);
1631 	}
1632 
1633 
1634 err_no_ring:
1635 	kfree(target->tx_ring);
1636 	target->tx_ring = NULL;
1637 	kfree(target->rx_ring);
1638 	target->rx_ring = NULL;
1639 
1640 	return -ENOMEM;
1641 }
1642 
1643 static uint32_t srp_compute_rq_tmo(struct ib_qp_attr *qp_attr, int attr_mask)
1644 {
1645 	uint64_t T_tr_ns, max_compl_time_ms;
1646 	uint32_t rq_tmo_jiffies;
1647 
1648 	/*
1649 	 * According to section 11.2.4.2 in the IBTA spec (Modify Queue Pair,
1650 	 * table 91), both the QP timeout and the retry count have to be set
1651 	 * for RC QP's during the RTR to RTS transition.
1652 	 */
1653 	WARN_ON_ONCE((attr_mask & (IB_QP_TIMEOUT | IB_QP_RETRY_CNT)) !=
1654 		     (IB_QP_TIMEOUT | IB_QP_RETRY_CNT));
1655 
1656 	/*
1657 	 * Set target->rq_tmo_jiffies to one second more than the largest time
1658 	 * it can take before an error completion is generated. See also
1659 	 * C9-140..142 in the IBTA spec for more information about how to
1660 	 * convert the QP Local ACK Timeout value to nanoseconds.
1661 	 */
1662 	T_tr_ns = 4096 * (1ULL << qp_attr->timeout);
1663 	max_compl_time_ms = qp_attr->retry_cnt * 4 * T_tr_ns;
1664 	do_div(max_compl_time_ms, NSEC_PER_MSEC);
1665 	rq_tmo_jiffies = msecs_to_jiffies(max_compl_time_ms + 1000);
1666 
1667 	return rq_tmo_jiffies;
1668 }
1669 
1670 static void srp_cm_rep_handler(struct ib_cm_id *cm_id,
1671 			       struct srp_login_rsp *lrsp,
1672 			       struct srp_target_port *target)
1673 {
1674 	struct ib_qp_attr *qp_attr = NULL;
1675 	int attr_mask = 0;
1676 	int ret;
1677 	int i;
1678 
1679 	if (lrsp->opcode == SRP_LOGIN_RSP) {
1680 		target->max_ti_iu_len = be32_to_cpu(lrsp->max_ti_iu_len);
1681 		target->req_lim       = be32_to_cpu(lrsp->req_lim_delta);
1682 
1683 		/*
1684 		 * Reserve credits for task management so we don't
1685 		 * bounce requests back to the SCSI mid-layer.
1686 		 */
1687 		target->scsi_host->can_queue
1688 			= min(target->req_lim - SRP_TSK_MGMT_SQ_SIZE,
1689 			      target->scsi_host->can_queue);
1690 		target->scsi_host->cmd_per_lun
1691 			= min_t(int, target->scsi_host->can_queue,
1692 				target->scsi_host->cmd_per_lun);
1693 	} else {
1694 		shost_printk(KERN_WARNING, target->scsi_host,
1695 			     PFX "Unhandled RSP opcode %#x\n", lrsp->opcode);
1696 		ret = -ECONNRESET;
1697 		goto error;
1698 	}
1699 
1700 	if (!target->rx_ring) {
1701 		ret = srp_alloc_iu_bufs(target);
1702 		if (ret)
1703 			goto error;
1704 	}
1705 
1706 	ret = -ENOMEM;
1707 	qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
1708 	if (!qp_attr)
1709 		goto error;
1710 
1711 	qp_attr->qp_state = IB_QPS_RTR;
1712 	ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1713 	if (ret)
1714 		goto error_free;
1715 
1716 	ret = ib_modify_qp(target->qp, qp_attr, attr_mask);
1717 	if (ret)
1718 		goto error_free;
1719 
1720 	for (i = 0; i < target->queue_size; i++) {
1721 		struct srp_iu *iu = target->rx_ring[i];
1722 		ret = srp_post_recv(target, iu);
1723 		if (ret)
1724 			goto error_free;
1725 	}
1726 
1727 	qp_attr->qp_state = IB_QPS_RTS;
1728 	ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1729 	if (ret)
1730 		goto error_free;
1731 
1732 	target->rq_tmo_jiffies = srp_compute_rq_tmo(qp_attr, attr_mask);
1733 
1734 	ret = ib_modify_qp(target->qp, qp_attr, attr_mask);
1735 	if (ret)
1736 		goto error_free;
1737 
1738 	ret = ib_send_cm_rtu(cm_id, NULL, 0);
1739 
1740 error_free:
1741 	kfree(qp_attr);
1742 
1743 error:
1744 	target->status = ret;
1745 }
1746 
1747 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
1748 			       struct ib_cm_event *event,
1749 			       struct srp_target_port *target)
1750 {
1751 	struct Scsi_Host *shost = target->scsi_host;
1752 	struct ib_class_port_info *cpi;
1753 	int opcode;
1754 
1755 	switch (event->param.rej_rcvd.reason) {
1756 	case IB_CM_REJ_PORT_CM_REDIRECT:
1757 		cpi = event->param.rej_rcvd.ari;
1758 		target->path.dlid = cpi->redirect_lid;
1759 		target->path.pkey = cpi->redirect_pkey;
1760 		cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
1761 		memcpy(target->path.dgid.raw, cpi->redirect_gid, 16);
1762 
1763 		target->status = target->path.dlid ?
1764 			SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
1765 		break;
1766 
1767 	case IB_CM_REJ_PORT_REDIRECT:
1768 		if (srp_target_is_topspin(target)) {
1769 			/*
1770 			 * Topspin/Cisco SRP gateways incorrectly send
1771 			 * reject reason code 25 when they mean 24
1772 			 * (port redirect).
1773 			 */
1774 			memcpy(target->path.dgid.raw,
1775 			       event->param.rej_rcvd.ari, 16);
1776 
1777 			shost_printk(KERN_DEBUG, shost,
1778 				     PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
1779 				     (unsigned long long) be64_to_cpu(target->path.dgid.global.subnet_prefix),
1780 				     (unsigned long long) be64_to_cpu(target->path.dgid.global.interface_id));
1781 
1782 			target->status = SRP_PORT_REDIRECT;
1783 		} else {
1784 			shost_printk(KERN_WARNING, shost,
1785 				     "  REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
1786 			target->status = -ECONNRESET;
1787 		}
1788 		break;
1789 
1790 	case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
1791 		shost_printk(KERN_WARNING, shost,
1792 			    "  REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
1793 		target->status = -ECONNRESET;
1794 		break;
1795 
1796 	case IB_CM_REJ_CONSUMER_DEFINED:
1797 		opcode = *(u8 *) event->private_data;
1798 		if (opcode == SRP_LOGIN_REJ) {
1799 			struct srp_login_rej *rej = event->private_data;
1800 			u32 reason = be32_to_cpu(rej->reason);
1801 
1802 			if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
1803 				shost_printk(KERN_WARNING, shost,
1804 					     PFX "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
1805 			else
1806 				shost_printk(KERN_WARNING, shost,
1807 					    PFX "SRP LOGIN REJECTED, reason 0x%08x\n", reason);
1808 		} else
1809 			shost_printk(KERN_WARNING, shost,
1810 				     "  REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
1811 				     " opcode 0x%02x\n", opcode);
1812 		target->status = -ECONNRESET;
1813 		break;
1814 
1815 	case IB_CM_REJ_STALE_CONN:
1816 		shost_printk(KERN_WARNING, shost, "  REJ reason: stale connection\n");
1817 		target->status = SRP_STALE_CONN;
1818 		break;
1819 
1820 	default:
1821 		shost_printk(KERN_WARNING, shost, "  REJ reason 0x%x\n",
1822 			     event->param.rej_rcvd.reason);
1823 		target->status = -ECONNRESET;
1824 	}
1825 }
1826 
1827 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
1828 {
1829 	struct srp_target_port *target = cm_id->context;
1830 	int comp = 0;
1831 
1832 	switch (event->event) {
1833 	case IB_CM_REQ_ERROR:
1834 		shost_printk(KERN_DEBUG, target->scsi_host,
1835 			     PFX "Sending CM REQ failed\n");
1836 		comp = 1;
1837 		target->status = -ECONNRESET;
1838 		break;
1839 
1840 	case IB_CM_REP_RECEIVED:
1841 		comp = 1;
1842 		srp_cm_rep_handler(cm_id, event->private_data, target);
1843 		break;
1844 
1845 	case IB_CM_REJ_RECEIVED:
1846 		shost_printk(KERN_DEBUG, target->scsi_host, PFX "REJ received\n");
1847 		comp = 1;
1848 
1849 		srp_cm_rej_handler(cm_id, event, target);
1850 		break;
1851 
1852 	case IB_CM_DREQ_RECEIVED:
1853 		shost_printk(KERN_WARNING, target->scsi_host,
1854 			     PFX "DREQ received - connection closed\n");
1855 		srp_change_conn_state(target, false);
1856 		if (ib_send_cm_drep(cm_id, NULL, 0))
1857 			shost_printk(KERN_ERR, target->scsi_host,
1858 				     PFX "Sending CM DREP failed\n");
1859 		queue_work(system_long_wq, &target->tl_err_work);
1860 		break;
1861 
1862 	case IB_CM_TIMEWAIT_EXIT:
1863 		shost_printk(KERN_ERR, target->scsi_host,
1864 			     PFX "connection closed\n");
1865 
1866 		target->status = 0;
1867 		break;
1868 
1869 	case IB_CM_MRA_RECEIVED:
1870 	case IB_CM_DREQ_ERROR:
1871 	case IB_CM_DREP_RECEIVED:
1872 		break;
1873 
1874 	default:
1875 		shost_printk(KERN_WARNING, target->scsi_host,
1876 			     PFX "Unhandled CM event %d\n", event->event);
1877 		break;
1878 	}
1879 
1880 	if (comp)
1881 		complete(&target->done);
1882 
1883 	return 0;
1884 }
1885 
1886 /**
1887  * srp_change_queue_type - changing device queue tag type
1888  * @sdev: scsi device struct
1889  * @tag_type: requested tag type
1890  *
1891  * Returns queue tag type.
1892  */
1893 static int
1894 srp_change_queue_type(struct scsi_device *sdev, int tag_type)
1895 {
1896 	if (sdev->tagged_supported) {
1897 		scsi_set_tag_type(sdev, tag_type);
1898 		if (tag_type)
1899 			scsi_activate_tcq(sdev, sdev->queue_depth);
1900 		else
1901 			scsi_deactivate_tcq(sdev, sdev->queue_depth);
1902 	} else
1903 		tag_type = 0;
1904 
1905 	return tag_type;
1906 }
1907 
1908 /**
1909  * srp_change_queue_depth - setting device queue depth
1910  * @sdev: scsi device struct
1911  * @qdepth: requested queue depth
1912  * @reason: SCSI_QDEPTH_DEFAULT/SCSI_QDEPTH_QFULL/SCSI_QDEPTH_RAMP_UP
1913  * (see include/scsi/scsi_host.h for definition)
1914  *
1915  * Returns queue depth.
1916  */
1917 static int
1918 srp_change_queue_depth(struct scsi_device *sdev, int qdepth, int reason)
1919 {
1920 	struct Scsi_Host *shost = sdev->host;
1921 	int max_depth;
1922 	if (reason == SCSI_QDEPTH_DEFAULT || reason == SCSI_QDEPTH_RAMP_UP) {
1923 		max_depth = shost->can_queue;
1924 		if (!sdev->tagged_supported)
1925 			max_depth = 1;
1926 		if (qdepth > max_depth)
1927 			qdepth = max_depth;
1928 		scsi_adjust_queue_depth(sdev, scsi_get_tag_type(sdev), qdepth);
1929 	} else if (reason == SCSI_QDEPTH_QFULL)
1930 		scsi_track_queue_full(sdev, qdepth);
1931 	else
1932 		return -EOPNOTSUPP;
1933 
1934 	return sdev->queue_depth;
1935 }
1936 
1937 static int srp_send_tsk_mgmt(struct srp_target_port *target,
1938 			     u64 req_tag, unsigned int lun, u8 func)
1939 {
1940 	struct srp_rport *rport = target->rport;
1941 	struct ib_device *dev = target->srp_host->srp_dev->dev;
1942 	struct srp_iu *iu;
1943 	struct srp_tsk_mgmt *tsk_mgmt;
1944 
1945 	if (!target->connected || target->qp_in_error)
1946 		return -1;
1947 
1948 	init_completion(&target->tsk_mgmt_done);
1949 
1950 	/*
1951 	 * Lock the rport mutex to avoid that srp_create_target_ib() is
1952 	 * invoked while a task management function is being sent.
1953 	 */
1954 	mutex_lock(&rport->mutex);
1955 	spin_lock_irq(&target->lock);
1956 	iu = __srp_get_tx_iu(target, SRP_IU_TSK_MGMT);
1957 	spin_unlock_irq(&target->lock);
1958 
1959 	if (!iu) {
1960 		mutex_unlock(&rport->mutex);
1961 
1962 		return -1;
1963 	}
1964 
1965 	ib_dma_sync_single_for_cpu(dev, iu->dma, sizeof *tsk_mgmt,
1966 				   DMA_TO_DEVICE);
1967 	tsk_mgmt = iu->buf;
1968 	memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
1969 
1970 	tsk_mgmt->opcode 	= SRP_TSK_MGMT;
1971 	tsk_mgmt->lun		= cpu_to_be64((u64) lun << 48);
1972 	tsk_mgmt->tag		= req_tag | SRP_TAG_TSK_MGMT;
1973 	tsk_mgmt->tsk_mgmt_func = func;
1974 	tsk_mgmt->task_tag	= req_tag;
1975 
1976 	ib_dma_sync_single_for_device(dev, iu->dma, sizeof *tsk_mgmt,
1977 				      DMA_TO_DEVICE);
1978 	if (srp_post_send(target, iu, sizeof *tsk_mgmt)) {
1979 		srp_put_tx_iu(target, iu, SRP_IU_TSK_MGMT);
1980 		mutex_unlock(&rport->mutex);
1981 
1982 		return -1;
1983 	}
1984 	mutex_unlock(&rport->mutex);
1985 
1986 	if (!wait_for_completion_timeout(&target->tsk_mgmt_done,
1987 					 msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
1988 		return -1;
1989 
1990 	return 0;
1991 }
1992 
1993 static int srp_abort(struct scsi_cmnd *scmnd)
1994 {
1995 	struct srp_target_port *target = host_to_target(scmnd->device->host);
1996 	struct srp_request *req = (struct srp_request *) scmnd->host_scribble;
1997 	int ret;
1998 
1999 	shost_printk(KERN_ERR, target->scsi_host, "SRP abort called\n");
2000 
2001 	if (!req || !srp_claim_req(target, req, scmnd))
2002 		return SUCCESS;
2003 	if (srp_send_tsk_mgmt(target, req->index, scmnd->device->lun,
2004 			      SRP_TSK_ABORT_TASK) == 0)
2005 		ret = SUCCESS;
2006 	else if (target->rport->state == SRP_RPORT_LOST)
2007 		ret = FAST_IO_FAIL;
2008 	else
2009 		ret = FAILED;
2010 	srp_free_req(target, req, scmnd, 0);
2011 	scmnd->result = DID_ABORT << 16;
2012 	scmnd->scsi_done(scmnd);
2013 
2014 	return ret;
2015 }
2016 
2017 static int srp_reset_device(struct scsi_cmnd *scmnd)
2018 {
2019 	struct srp_target_port *target = host_to_target(scmnd->device->host);
2020 	int i;
2021 
2022 	shost_printk(KERN_ERR, target->scsi_host, "SRP reset_device called\n");
2023 
2024 	if (srp_send_tsk_mgmt(target, SRP_TAG_NO_REQ, scmnd->device->lun,
2025 			      SRP_TSK_LUN_RESET))
2026 		return FAILED;
2027 	if (target->tsk_mgmt_status)
2028 		return FAILED;
2029 
2030 	for (i = 0; i < target->req_ring_size; ++i) {
2031 		struct srp_request *req = &target->req_ring[i];
2032 		if (req->scmnd && req->scmnd->device == scmnd->device)
2033 			srp_finish_req(target, req, DID_RESET << 16);
2034 	}
2035 
2036 	return SUCCESS;
2037 }
2038 
2039 static int srp_reset_host(struct scsi_cmnd *scmnd)
2040 {
2041 	struct srp_target_port *target = host_to_target(scmnd->device->host);
2042 
2043 	shost_printk(KERN_ERR, target->scsi_host, PFX "SRP reset_host called\n");
2044 
2045 	return srp_reconnect_rport(target->rport) == 0 ? SUCCESS : FAILED;
2046 }
2047 
2048 static int srp_slave_configure(struct scsi_device *sdev)
2049 {
2050 	struct Scsi_Host *shost = sdev->host;
2051 	struct srp_target_port *target = host_to_target(shost);
2052 	struct request_queue *q = sdev->request_queue;
2053 	unsigned long timeout;
2054 
2055 	if (sdev->type == TYPE_DISK) {
2056 		timeout = max_t(unsigned, 30 * HZ, target->rq_tmo_jiffies);
2057 		blk_queue_rq_timeout(q, timeout);
2058 	}
2059 
2060 	return 0;
2061 }
2062 
2063 static ssize_t show_id_ext(struct device *dev, struct device_attribute *attr,
2064 			   char *buf)
2065 {
2066 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2067 
2068 	return sprintf(buf, "0x%016llx\n",
2069 		       (unsigned long long) be64_to_cpu(target->id_ext));
2070 }
2071 
2072 static ssize_t show_ioc_guid(struct device *dev, struct device_attribute *attr,
2073 			     char *buf)
2074 {
2075 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2076 
2077 	return sprintf(buf, "0x%016llx\n",
2078 		       (unsigned long long) be64_to_cpu(target->ioc_guid));
2079 }
2080 
2081 static ssize_t show_service_id(struct device *dev,
2082 			       struct device_attribute *attr, char *buf)
2083 {
2084 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2085 
2086 	return sprintf(buf, "0x%016llx\n",
2087 		       (unsigned long long) be64_to_cpu(target->service_id));
2088 }
2089 
2090 static ssize_t show_pkey(struct device *dev, struct device_attribute *attr,
2091 			 char *buf)
2092 {
2093 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2094 
2095 	return sprintf(buf, "0x%04x\n", be16_to_cpu(target->path.pkey));
2096 }
2097 
2098 static ssize_t show_sgid(struct device *dev, struct device_attribute *attr,
2099 			 char *buf)
2100 {
2101 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2102 
2103 	return sprintf(buf, "%pI6\n", target->path.sgid.raw);
2104 }
2105 
2106 static ssize_t show_dgid(struct device *dev, struct device_attribute *attr,
2107 			 char *buf)
2108 {
2109 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2110 
2111 	return sprintf(buf, "%pI6\n", target->path.dgid.raw);
2112 }
2113 
2114 static ssize_t show_orig_dgid(struct device *dev,
2115 			      struct device_attribute *attr, char *buf)
2116 {
2117 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2118 
2119 	return sprintf(buf, "%pI6\n", target->orig_dgid);
2120 }
2121 
2122 static ssize_t show_req_lim(struct device *dev,
2123 			    struct device_attribute *attr, char *buf)
2124 {
2125 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2126 
2127 	return sprintf(buf, "%d\n", target->req_lim);
2128 }
2129 
2130 static ssize_t show_zero_req_lim(struct device *dev,
2131 				 struct device_attribute *attr, char *buf)
2132 {
2133 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2134 
2135 	return sprintf(buf, "%d\n", target->zero_req_lim);
2136 }
2137 
2138 static ssize_t show_local_ib_port(struct device *dev,
2139 				  struct device_attribute *attr, char *buf)
2140 {
2141 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2142 
2143 	return sprintf(buf, "%d\n", target->srp_host->port);
2144 }
2145 
2146 static ssize_t show_local_ib_device(struct device *dev,
2147 				    struct device_attribute *attr, char *buf)
2148 {
2149 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2150 
2151 	return sprintf(buf, "%s\n", target->srp_host->srp_dev->dev->name);
2152 }
2153 
2154 static ssize_t show_comp_vector(struct device *dev,
2155 				struct device_attribute *attr, char *buf)
2156 {
2157 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2158 
2159 	return sprintf(buf, "%d\n", target->comp_vector);
2160 }
2161 
2162 static ssize_t show_tl_retry_count(struct device *dev,
2163 				   struct device_attribute *attr, char *buf)
2164 {
2165 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2166 
2167 	return sprintf(buf, "%d\n", target->tl_retry_count);
2168 }
2169 
2170 static ssize_t show_cmd_sg_entries(struct device *dev,
2171 				   struct device_attribute *attr, char *buf)
2172 {
2173 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2174 
2175 	return sprintf(buf, "%u\n", target->cmd_sg_cnt);
2176 }
2177 
2178 static ssize_t show_allow_ext_sg(struct device *dev,
2179 				 struct device_attribute *attr, char *buf)
2180 {
2181 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2182 
2183 	return sprintf(buf, "%s\n", target->allow_ext_sg ? "true" : "false");
2184 }
2185 
2186 static DEVICE_ATTR(id_ext,	    S_IRUGO, show_id_ext,	   NULL);
2187 static DEVICE_ATTR(ioc_guid,	    S_IRUGO, show_ioc_guid,	   NULL);
2188 static DEVICE_ATTR(service_id,	    S_IRUGO, show_service_id,	   NULL);
2189 static DEVICE_ATTR(pkey,	    S_IRUGO, show_pkey,		   NULL);
2190 static DEVICE_ATTR(sgid,	    S_IRUGO, show_sgid,		   NULL);
2191 static DEVICE_ATTR(dgid,	    S_IRUGO, show_dgid,		   NULL);
2192 static DEVICE_ATTR(orig_dgid,	    S_IRUGO, show_orig_dgid,	   NULL);
2193 static DEVICE_ATTR(req_lim,         S_IRUGO, show_req_lim,         NULL);
2194 static DEVICE_ATTR(zero_req_lim,    S_IRUGO, show_zero_req_lim,	   NULL);
2195 static DEVICE_ATTR(local_ib_port,   S_IRUGO, show_local_ib_port,   NULL);
2196 static DEVICE_ATTR(local_ib_device, S_IRUGO, show_local_ib_device, NULL);
2197 static DEVICE_ATTR(comp_vector,     S_IRUGO, show_comp_vector,     NULL);
2198 static DEVICE_ATTR(tl_retry_count,  S_IRUGO, show_tl_retry_count,  NULL);
2199 static DEVICE_ATTR(cmd_sg_entries,  S_IRUGO, show_cmd_sg_entries,  NULL);
2200 static DEVICE_ATTR(allow_ext_sg,    S_IRUGO, show_allow_ext_sg,    NULL);
2201 
2202 static struct device_attribute *srp_host_attrs[] = {
2203 	&dev_attr_id_ext,
2204 	&dev_attr_ioc_guid,
2205 	&dev_attr_service_id,
2206 	&dev_attr_pkey,
2207 	&dev_attr_sgid,
2208 	&dev_attr_dgid,
2209 	&dev_attr_orig_dgid,
2210 	&dev_attr_req_lim,
2211 	&dev_attr_zero_req_lim,
2212 	&dev_attr_local_ib_port,
2213 	&dev_attr_local_ib_device,
2214 	&dev_attr_comp_vector,
2215 	&dev_attr_tl_retry_count,
2216 	&dev_attr_cmd_sg_entries,
2217 	&dev_attr_allow_ext_sg,
2218 	NULL
2219 };
2220 
2221 static struct scsi_host_template srp_template = {
2222 	.module				= THIS_MODULE,
2223 	.name				= "InfiniBand SRP initiator",
2224 	.proc_name			= DRV_NAME,
2225 	.slave_configure		= srp_slave_configure,
2226 	.info				= srp_target_info,
2227 	.queuecommand			= srp_queuecommand,
2228 	.change_queue_depth             = srp_change_queue_depth,
2229 	.change_queue_type              = srp_change_queue_type,
2230 	.eh_abort_handler		= srp_abort,
2231 	.eh_device_reset_handler	= srp_reset_device,
2232 	.eh_host_reset_handler		= srp_reset_host,
2233 	.skip_settle_delay		= true,
2234 	.sg_tablesize			= SRP_DEF_SG_TABLESIZE,
2235 	.can_queue			= SRP_DEFAULT_CMD_SQ_SIZE,
2236 	.this_id			= -1,
2237 	.cmd_per_lun			= SRP_DEFAULT_CMD_SQ_SIZE,
2238 	.use_clustering			= ENABLE_CLUSTERING,
2239 	.shost_attrs			= srp_host_attrs
2240 };
2241 
2242 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
2243 {
2244 	struct srp_rport_identifiers ids;
2245 	struct srp_rport *rport;
2246 
2247 	sprintf(target->target_name, "SRP.T10:%016llX",
2248 		 (unsigned long long) be64_to_cpu(target->id_ext));
2249 
2250 	if (scsi_add_host(target->scsi_host, host->srp_dev->dev->dma_device))
2251 		return -ENODEV;
2252 
2253 	memcpy(ids.port_id, &target->id_ext, 8);
2254 	memcpy(ids.port_id + 8, &target->ioc_guid, 8);
2255 	ids.roles = SRP_RPORT_ROLE_TARGET;
2256 	rport = srp_rport_add(target->scsi_host, &ids);
2257 	if (IS_ERR(rport)) {
2258 		scsi_remove_host(target->scsi_host);
2259 		return PTR_ERR(rport);
2260 	}
2261 
2262 	rport->lld_data = target;
2263 	target->rport = rport;
2264 
2265 	spin_lock(&host->target_lock);
2266 	list_add_tail(&target->list, &host->target_list);
2267 	spin_unlock(&host->target_lock);
2268 
2269 	target->state = SRP_TARGET_LIVE;
2270 
2271 	scsi_scan_target(&target->scsi_host->shost_gendev,
2272 			 0, target->scsi_id, SCAN_WILD_CARD, 0);
2273 
2274 	return 0;
2275 }
2276 
2277 static void srp_release_dev(struct device *dev)
2278 {
2279 	struct srp_host *host =
2280 		container_of(dev, struct srp_host, dev);
2281 
2282 	complete(&host->released);
2283 }
2284 
2285 static struct class srp_class = {
2286 	.name    = "infiniband_srp",
2287 	.dev_release = srp_release_dev
2288 };
2289 
2290 /**
2291  * srp_conn_unique() - check whether the connection to a target is unique
2292  */
2293 static bool srp_conn_unique(struct srp_host *host,
2294 			    struct srp_target_port *target)
2295 {
2296 	struct srp_target_port *t;
2297 	bool ret = false;
2298 
2299 	if (target->state == SRP_TARGET_REMOVED)
2300 		goto out;
2301 
2302 	ret = true;
2303 
2304 	spin_lock(&host->target_lock);
2305 	list_for_each_entry(t, &host->target_list, list) {
2306 		if (t != target &&
2307 		    target->id_ext == t->id_ext &&
2308 		    target->ioc_guid == t->ioc_guid &&
2309 		    target->initiator_ext == t->initiator_ext) {
2310 			ret = false;
2311 			break;
2312 		}
2313 	}
2314 	spin_unlock(&host->target_lock);
2315 
2316 out:
2317 	return ret;
2318 }
2319 
2320 /*
2321  * Target ports are added by writing
2322  *
2323  *     id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
2324  *     pkey=<P_Key>,service_id=<service ID>
2325  *
2326  * to the add_target sysfs attribute.
2327  */
2328 enum {
2329 	SRP_OPT_ERR		= 0,
2330 	SRP_OPT_ID_EXT		= 1 << 0,
2331 	SRP_OPT_IOC_GUID	= 1 << 1,
2332 	SRP_OPT_DGID		= 1 << 2,
2333 	SRP_OPT_PKEY		= 1 << 3,
2334 	SRP_OPT_SERVICE_ID	= 1 << 4,
2335 	SRP_OPT_MAX_SECT	= 1 << 5,
2336 	SRP_OPT_MAX_CMD_PER_LUN	= 1 << 6,
2337 	SRP_OPT_IO_CLASS	= 1 << 7,
2338 	SRP_OPT_INITIATOR_EXT	= 1 << 8,
2339 	SRP_OPT_CMD_SG_ENTRIES	= 1 << 9,
2340 	SRP_OPT_ALLOW_EXT_SG	= 1 << 10,
2341 	SRP_OPT_SG_TABLESIZE	= 1 << 11,
2342 	SRP_OPT_COMP_VECTOR	= 1 << 12,
2343 	SRP_OPT_TL_RETRY_COUNT	= 1 << 13,
2344 	SRP_OPT_QUEUE_SIZE	= 1 << 14,
2345 	SRP_OPT_ALL		= (SRP_OPT_ID_EXT	|
2346 				   SRP_OPT_IOC_GUID	|
2347 				   SRP_OPT_DGID		|
2348 				   SRP_OPT_PKEY		|
2349 				   SRP_OPT_SERVICE_ID),
2350 };
2351 
2352 static const match_table_t srp_opt_tokens = {
2353 	{ SRP_OPT_ID_EXT,		"id_ext=%s" 		},
2354 	{ SRP_OPT_IOC_GUID,		"ioc_guid=%s" 		},
2355 	{ SRP_OPT_DGID,			"dgid=%s" 		},
2356 	{ SRP_OPT_PKEY,			"pkey=%x" 		},
2357 	{ SRP_OPT_SERVICE_ID,		"service_id=%s"		},
2358 	{ SRP_OPT_MAX_SECT,		"max_sect=%d" 		},
2359 	{ SRP_OPT_MAX_CMD_PER_LUN,	"max_cmd_per_lun=%d" 	},
2360 	{ SRP_OPT_IO_CLASS,		"io_class=%x"		},
2361 	{ SRP_OPT_INITIATOR_EXT,	"initiator_ext=%s"	},
2362 	{ SRP_OPT_CMD_SG_ENTRIES,	"cmd_sg_entries=%u"	},
2363 	{ SRP_OPT_ALLOW_EXT_SG,		"allow_ext_sg=%u"	},
2364 	{ SRP_OPT_SG_TABLESIZE,		"sg_tablesize=%u"	},
2365 	{ SRP_OPT_COMP_VECTOR,		"comp_vector=%u"	},
2366 	{ SRP_OPT_TL_RETRY_COUNT,	"tl_retry_count=%u"	},
2367 	{ SRP_OPT_QUEUE_SIZE,		"queue_size=%d"		},
2368 	{ SRP_OPT_ERR,			NULL 			}
2369 };
2370 
2371 static int srp_parse_options(const char *buf, struct srp_target_port *target)
2372 {
2373 	char *options, *sep_opt;
2374 	char *p;
2375 	char dgid[3];
2376 	substring_t args[MAX_OPT_ARGS];
2377 	int opt_mask = 0;
2378 	int token;
2379 	int ret = -EINVAL;
2380 	int i;
2381 
2382 	options = kstrdup(buf, GFP_KERNEL);
2383 	if (!options)
2384 		return -ENOMEM;
2385 
2386 	sep_opt = options;
2387 	while ((p = strsep(&sep_opt, ",")) != NULL) {
2388 		if (!*p)
2389 			continue;
2390 
2391 		token = match_token(p, srp_opt_tokens, args);
2392 		opt_mask |= token;
2393 
2394 		switch (token) {
2395 		case SRP_OPT_ID_EXT:
2396 			p = match_strdup(args);
2397 			if (!p) {
2398 				ret = -ENOMEM;
2399 				goto out;
2400 			}
2401 			target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2402 			kfree(p);
2403 			break;
2404 
2405 		case SRP_OPT_IOC_GUID:
2406 			p = match_strdup(args);
2407 			if (!p) {
2408 				ret = -ENOMEM;
2409 				goto out;
2410 			}
2411 			target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
2412 			kfree(p);
2413 			break;
2414 
2415 		case SRP_OPT_DGID:
2416 			p = match_strdup(args);
2417 			if (!p) {
2418 				ret = -ENOMEM;
2419 				goto out;
2420 			}
2421 			if (strlen(p) != 32) {
2422 				pr_warn("bad dest GID parameter '%s'\n", p);
2423 				kfree(p);
2424 				goto out;
2425 			}
2426 
2427 			for (i = 0; i < 16; ++i) {
2428 				strlcpy(dgid, p + i * 2, 3);
2429 				target->path.dgid.raw[i] = simple_strtoul(dgid, NULL, 16);
2430 			}
2431 			kfree(p);
2432 			memcpy(target->orig_dgid, target->path.dgid.raw, 16);
2433 			break;
2434 
2435 		case SRP_OPT_PKEY:
2436 			if (match_hex(args, &token)) {
2437 				pr_warn("bad P_Key parameter '%s'\n", p);
2438 				goto out;
2439 			}
2440 			target->path.pkey = cpu_to_be16(token);
2441 			break;
2442 
2443 		case SRP_OPT_SERVICE_ID:
2444 			p = match_strdup(args);
2445 			if (!p) {
2446 				ret = -ENOMEM;
2447 				goto out;
2448 			}
2449 			target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
2450 			target->path.service_id = target->service_id;
2451 			kfree(p);
2452 			break;
2453 
2454 		case SRP_OPT_MAX_SECT:
2455 			if (match_int(args, &token)) {
2456 				pr_warn("bad max sect parameter '%s'\n", p);
2457 				goto out;
2458 			}
2459 			target->scsi_host->max_sectors = token;
2460 			break;
2461 
2462 		case SRP_OPT_QUEUE_SIZE:
2463 			if (match_int(args, &token) || token < 1) {
2464 				pr_warn("bad queue_size parameter '%s'\n", p);
2465 				goto out;
2466 			}
2467 			target->scsi_host->can_queue = token;
2468 			target->queue_size = token + SRP_RSP_SQ_SIZE +
2469 					     SRP_TSK_MGMT_SQ_SIZE;
2470 			if (!(opt_mask & SRP_OPT_MAX_CMD_PER_LUN))
2471 				target->scsi_host->cmd_per_lun = token;
2472 			break;
2473 
2474 		case SRP_OPT_MAX_CMD_PER_LUN:
2475 			if (match_int(args, &token) || token < 1) {
2476 				pr_warn("bad max cmd_per_lun parameter '%s'\n",
2477 					p);
2478 				goto out;
2479 			}
2480 			target->scsi_host->cmd_per_lun = token;
2481 			break;
2482 
2483 		case SRP_OPT_IO_CLASS:
2484 			if (match_hex(args, &token)) {
2485 				pr_warn("bad IO class parameter '%s'\n", p);
2486 				goto out;
2487 			}
2488 			if (token != SRP_REV10_IB_IO_CLASS &&
2489 			    token != SRP_REV16A_IB_IO_CLASS) {
2490 				pr_warn("unknown IO class parameter value %x specified (use %x or %x).\n",
2491 					token, SRP_REV10_IB_IO_CLASS,
2492 					SRP_REV16A_IB_IO_CLASS);
2493 				goto out;
2494 			}
2495 			target->io_class = token;
2496 			break;
2497 
2498 		case SRP_OPT_INITIATOR_EXT:
2499 			p = match_strdup(args);
2500 			if (!p) {
2501 				ret = -ENOMEM;
2502 				goto out;
2503 			}
2504 			target->initiator_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2505 			kfree(p);
2506 			break;
2507 
2508 		case SRP_OPT_CMD_SG_ENTRIES:
2509 			if (match_int(args, &token) || token < 1 || token > 255) {
2510 				pr_warn("bad max cmd_sg_entries parameter '%s'\n",
2511 					p);
2512 				goto out;
2513 			}
2514 			target->cmd_sg_cnt = token;
2515 			break;
2516 
2517 		case SRP_OPT_ALLOW_EXT_SG:
2518 			if (match_int(args, &token)) {
2519 				pr_warn("bad allow_ext_sg parameter '%s'\n", p);
2520 				goto out;
2521 			}
2522 			target->allow_ext_sg = !!token;
2523 			break;
2524 
2525 		case SRP_OPT_SG_TABLESIZE:
2526 			if (match_int(args, &token) || token < 1 ||
2527 					token > SCSI_MAX_SG_CHAIN_SEGMENTS) {
2528 				pr_warn("bad max sg_tablesize parameter '%s'\n",
2529 					p);
2530 				goto out;
2531 			}
2532 			target->sg_tablesize = token;
2533 			break;
2534 
2535 		case SRP_OPT_COMP_VECTOR:
2536 			if (match_int(args, &token) || token < 0) {
2537 				pr_warn("bad comp_vector parameter '%s'\n", p);
2538 				goto out;
2539 			}
2540 			target->comp_vector = token;
2541 			break;
2542 
2543 		case SRP_OPT_TL_RETRY_COUNT:
2544 			if (match_int(args, &token) || token < 2 || token > 7) {
2545 				pr_warn("bad tl_retry_count parameter '%s' (must be a number between 2 and 7)\n",
2546 					p);
2547 				goto out;
2548 			}
2549 			target->tl_retry_count = token;
2550 			break;
2551 
2552 		default:
2553 			pr_warn("unknown parameter or missing value '%s' in target creation request\n",
2554 				p);
2555 			goto out;
2556 		}
2557 	}
2558 
2559 	if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
2560 		ret = 0;
2561 	else
2562 		for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
2563 			if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
2564 			    !(srp_opt_tokens[i].token & opt_mask))
2565 				pr_warn("target creation request is missing parameter '%s'\n",
2566 					srp_opt_tokens[i].pattern);
2567 
2568 	if (target->scsi_host->cmd_per_lun > target->scsi_host->can_queue
2569 	    && (opt_mask & SRP_OPT_MAX_CMD_PER_LUN))
2570 		pr_warn("cmd_per_lun = %d > queue_size = %d\n",
2571 			target->scsi_host->cmd_per_lun,
2572 			target->scsi_host->can_queue);
2573 
2574 out:
2575 	kfree(options);
2576 	return ret;
2577 }
2578 
2579 static ssize_t srp_create_target(struct device *dev,
2580 				 struct device_attribute *attr,
2581 				 const char *buf, size_t count)
2582 {
2583 	struct srp_host *host =
2584 		container_of(dev, struct srp_host, dev);
2585 	struct Scsi_Host *target_host;
2586 	struct srp_target_port *target;
2587 	struct ib_device *ibdev = host->srp_dev->dev;
2588 	int ret;
2589 
2590 	target_host = scsi_host_alloc(&srp_template,
2591 				      sizeof (struct srp_target_port));
2592 	if (!target_host)
2593 		return -ENOMEM;
2594 
2595 	target_host->transportt  = ib_srp_transport_template;
2596 	target_host->max_channel = 0;
2597 	target_host->max_id      = 1;
2598 	target_host->max_lun     = SRP_MAX_LUN;
2599 	target_host->max_cmd_len = sizeof ((struct srp_cmd *) (void *) 0L)->cdb;
2600 
2601 	target = host_to_target(target_host);
2602 
2603 	target->io_class	= SRP_REV16A_IB_IO_CLASS;
2604 	target->scsi_host	= target_host;
2605 	target->srp_host	= host;
2606 	target->lkey		= host->srp_dev->mr->lkey;
2607 	target->rkey		= host->srp_dev->mr->rkey;
2608 	target->cmd_sg_cnt	= cmd_sg_entries;
2609 	target->sg_tablesize	= indirect_sg_entries ? : cmd_sg_entries;
2610 	target->allow_ext_sg	= allow_ext_sg;
2611 	target->tl_retry_count	= 7;
2612 	target->queue_size	= SRP_DEFAULT_QUEUE_SIZE;
2613 
2614 	ret = srp_parse_options(buf, target);
2615 	if (ret)
2616 		goto err;
2617 
2618 	target->req_ring_size = target->queue_size - SRP_TSK_MGMT_SQ_SIZE;
2619 
2620 	if (!srp_conn_unique(target->srp_host, target)) {
2621 		shost_printk(KERN_INFO, target->scsi_host,
2622 			     PFX "Already connected to target port with id_ext=%016llx;ioc_guid=%016llx;initiator_ext=%016llx\n",
2623 			     be64_to_cpu(target->id_ext),
2624 			     be64_to_cpu(target->ioc_guid),
2625 			     be64_to_cpu(target->initiator_ext));
2626 		ret = -EEXIST;
2627 		goto err;
2628 	}
2629 
2630 	if (!host->srp_dev->fmr_pool && !target->allow_ext_sg &&
2631 				target->cmd_sg_cnt < target->sg_tablesize) {
2632 		pr_warn("No FMR pool and no external indirect descriptors, limiting sg_tablesize to cmd_sg_cnt\n");
2633 		target->sg_tablesize = target->cmd_sg_cnt;
2634 	}
2635 
2636 	target_host->sg_tablesize = target->sg_tablesize;
2637 	target->indirect_size = target->sg_tablesize *
2638 				sizeof (struct srp_direct_buf);
2639 	target->max_iu_len = sizeof (struct srp_cmd) +
2640 			     sizeof (struct srp_indirect_buf) +
2641 			     target->cmd_sg_cnt * sizeof (struct srp_direct_buf);
2642 
2643 	INIT_WORK(&target->tl_err_work, srp_tl_err_work);
2644 	INIT_WORK(&target->remove_work, srp_remove_work);
2645 	spin_lock_init(&target->lock);
2646 	INIT_LIST_HEAD(&target->free_tx);
2647 	ret = srp_alloc_req_data(target);
2648 	if (ret)
2649 		goto err_free_mem;
2650 
2651 	ib_query_gid(ibdev, host->port, 0, &target->path.sgid);
2652 
2653 	shost_printk(KERN_DEBUG, target->scsi_host, PFX
2654 		     "new target: id_ext %016llx ioc_guid %016llx pkey %04x "
2655 		     "service_id %016llx dgid %pI6\n",
2656 	       (unsigned long long) be64_to_cpu(target->id_ext),
2657 	       (unsigned long long) be64_to_cpu(target->ioc_guid),
2658 	       be16_to_cpu(target->path.pkey),
2659 	       (unsigned long long) be64_to_cpu(target->service_id),
2660 	       target->path.dgid.raw);
2661 
2662 	ret = srp_create_target_ib(target);
2663 	if (ret)
2664 		goto err_free_mem;
2665 
2666 	ret = srp_new_cm_id(target);
2667 	if (ret)
2668 		goto err_free_ib;
2669 
2670 	ret = srp_connect_target(target);
2671 	if (ret) {
2672 		shost_printk(KERN_ERR, target->scsi_host,
2673 			     PFX "Connection failed\n");
2674 		goto err_cm_id;
2675 	}
2676 
2677 	ret = srp_add_target(host, target);
2678 	if (ret)
2679 		goto err_disconnect;
2680 
2681 	return count;
2682 
2683 err_disconnect:
2684 	srp_disconnect_target(target);
2685 
2686 err_cm_id:
2687 	ib_destroy_cm_id(target->cm_id);
2688 
2689 err_free_ib:
2690 	srp_free_target_ib(target);
2691 
2692 err_free_mem:
2693 	srp_free_req_data(target);
2694 
2695 err:
2696 	scsi_host_put(target_host);
2697 
2698 	return ret;
2699 }
2700 
2701 static DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
2702 
2703 static ssize_t show_ibdev(struct device *dev, struct device_attribute *attr,
2704 			  char *buf)
2705 {
2706 	struct srp_host *host = container_of(dev, struct srp_host, dev);
2707 
2708 	return sprintf(buf, "%s\n", host->srp_dev->dev->name);
2709 }
2710 
2711 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
2712 
2713 static ssize_t show_port(struct device *dev, struct device_attribute *attr,
2714 			 char *buf)
2715 {
2716 	struct srp_host *host = container_of(dev, struct srp_host, dev);
2717 
2718 	return sprintf(buf, "%d\n", host->port);
2719 }
2720 
2721 static DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
2722 
2723 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
2724 {
2725 	struct srp_host *host;
2726 
2727 	host = kzalloc(sizeof *host, GFP_KERNEL);
2728 	if (!host)
2729 		return NULL;
2730 
2731 	INIT_LIST_HEAD(&host->target_list);
2732 	spin_lock_init(&host->target_lock);
2733 	init_completion(&host->released);
2734 	host->srp_dev = device;
2735 	host->port = port;
2736 
2737 	host->dev.class = &srp_class;
2738 	host->dev.parent = device->dev->dma_device;
2739 	dev_set_name(&host->dev, "srp-%s-%d", device->dev->name, port);
2740 
2741 	if (device_register(&host->dev))
2742 		goto free_host;
2743 	if (device_create_file(&host->dev, &dev_attr_add_target))
2744 		goto err_class;
2745 	if (device_create_file(&host->dev, &dev_attr_ibdev))
2746 		goto err_class;
2747 	if (device_create_file(&host->dev, &dev_attr_port))
2748 		goto err_class;
2749 
2750 	return host;
2751 
2752 err_class:
2753 	device_unregister(&host->dev);
2754 
2755 free_host:
2756 	kfree(host);
2757 
2758 	return NULL;
2759 }
2760 
2761 static void srp_add_one(struct ib_device *device)
2762 {
2763 	struct srp_device *srp_dev;
2764 	struct ib_device_attr *dev_attr;
2765 	struct ib_fmr_pool_param fmr_param;
2766 	struct srp_host *host;
2767 	int max_pages_per_fmr, fmr_page_shift, s, e, p;
2768 
2769 	dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
2770 	if (!dev_attr)
2771 		return;
2772 
2773 	if (ib_query_device(device, dev_attr)) {
2774 		pr_warn("Query device failed for %s\n", device->name);
2775 		goto free_attr;
2776 	}
2777 
2778 	srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
2779 	if (!srp_dev)
2780 		goto free_attr;
2781 
2782 	/*
2783 	 * Use the smallest page size supported by the HCA, down to a
2784 	 * minimum of 4096 bytes. We're unlikely to build large sglists
2785 	 * out of smaller entries.
2786 	 */
2787 	fmr_page_shift		= max(12, ffs(dev_attr->page_size_cap) - 1);
2788 	srp_dev->fmr_page_size	= 1 << fmr_page_shift;
2789 	srp_dev->fmr_page_mask	= ~((u64) srp_dev->fmr_page_size - 1);
2790 	srp_dev->fmr_max_size	= srp_dev->fmr_page_size * SRP_FMR_SIZE;
2791 
2792 	INIT_LIST_HEAD(&srp_dev->dev_list);
2793 
2794 	srp_dev->dev = device;
2795 	srp_dev->pd  = ib_alloc_pd(device);
2796 	if (IS_ERR(srp_dev->pd))
2797 		goto free_dev;
2798 
2799 	srp_dev->mr = ib_get_dma_mr(srp_dev->pd,
2800 				    IB_ACCESS_LOCAL_WRITE |
2801 				    IB_ACCESS_REMOTE_READ |
2802 				    IB_ACCESS_REMOTE_WRITE);
2803 	if (IS_ERR(srp_dev->mr))
2804 		goto err_pd;
2805 
2806 	for (max_pages_per_fmr = SRP_FMR_SIZE;
2807 			max_pages_per_fmr >= SRP_FMR_MIN_SIZE;
2808 			max_pages_per_fmr /= 2, srp_dev->fmr_max_size /= 2) {
2809 		memset(&fmr_param, 0, sizeof fmr_param);
2810 		fmr_param.pool_size	    = SRP_FMR_POOL_SIZE;
2811 		fmr_param.dirty_watermark   = SRP_FMR_DIRTY_SIZE;
2812 		fmr_param.cache		    = 1;
2813 		fmr_param.max_pages_per_fmr = max_pages_per_fmr;
2814 		fmr_param.page_shift	    = fmr_page_shift;
2815 		fmr_param.access	    = (IB_ACCESS_LOCAL_WRITE |
2816 					       IB_ACCESS_REMOTE_WRITE |
2817 					       IB_ACCESS_REMOTE_READ);
2818 
2819 		srp_dev->fmr_pool = ib_create_fmr_pool(srp_dev->pd, &fmr_param);
2820 		if (!IS_ERR(srp_dev->fmr_pool))
2821 			break;
2822 	}
2823 
2824 	if (IS_ERR(srp_dev->fmr_pool))
2825 		srp_dev->fmr_pool = NULL;
2826 
2827 	if (device->node_type == RDMA_NODE_IB_SWITCH) {
2828 		s = 0;
2829 		e = 0;
2830 	} else {
2831 		s = 1;
2832 		e = device->phys_port_cnt;
2833 	}
2834 
2835 	for (p = s; p <= e; ++p) {
2836 		host = srp_add_port(srp_dev, p);
2837 		if (host)
2838 			list_add_tail(&host->list, &srp_dev->dev_list);
2839 	}
2840 
2841 	ib_set_client_data(device, &srp_client, srp_dev);
2842 
2843 	goto free_attr;
2844 
2845 err_pd:
2846 	ib_dealloc_pd(srp_dev->pd);
2847 
2848 free_dev:
2849 	kfree(srp_dev);
2850 
2851 free_attr:
2852 	kfree(dev_attr);
2853 }
2854 
2855 static void srp_remove_one(struct ib_device *device)
2856 {
2857 	struct srp_device *srp_dev;
2858 	struct srp_host *host, *tmp_host;
2859 	struct srp_target_port *target;
2860 
2861 	srp_dev = ib_get_client_data(device, &srp_client);
2862 	if (!srp_dev)
2863 		return;
2864 
2865 	list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
2866 		device_unregister(&host->dev);
2867 		/*
2868 		 * Wait for the sysfs entry to go away, so that no new
2869 		 * target ports can be created.
2870 		 */
2871 		wait_for_completion(&host->released);
2872 
2873 		/*
2874 		 * Remove all target ports.
2875 		 */
2876 		spin_lock(&host->target_lock);
2877 		list_for_each_entry(target, &host->target_list, list)
2878 			srp_queue_remove_work(target);
2879 		spin_unlock(&host->target_lock);
2880 
2881 		/*
2882 		 * Wait for target port removal tasks.
2883 		 */
2884 		flush_workqueue(system_long_wq);
2885 
2886 		kfree(host);
2887 	}
2888 
2889 	if (srp_dev->fmr_pool)
2890 		ib_destroy_fmr_pool(srp_dev->fmr_pool);
2891 	ib_dereg_mr(srp_dev->mr);
2892 	ib_dealloc_pd(srp_dev->pd);
2893 
2894 	kfree(srp_dev);
2895 }
2896 
2897 static struct srp_function_template ib_srp_transport_functions = {
2898 	.has_rport_state	 = true,
2899 	.reset_timer_if_blocked	 = true,
2900 	.reconnect_delay	 = &srp_reconnect_delay,
2901 	.fast_io_fail_tmo	 = &srp_fast_io_fail_tmo,
2902 	.dev_loss_tmo		 = &srp_dev_loss_tmo,
2903 	.reconnect		 = srp_rport_reconnect,
2904 	.rport_delete		 = srp_rport_delete,
2905 	.terminate_rport_io	 = srp_terminate_io,
2906 };
2907 
2908 static int __init srp_init_module(void)
2909 {
2910 	int ret;
2911 
2912 	BUILD_BUG_ON(FIELD_SIZEOF(struct ib_wc, wr_id) < sizeof(void *));
2913 
2914 	if (srp_sg_tablesize) {
2915 		pr_warn("srp_sg_tablesize is deprecated, please use cmd_sg_entries\n");
2916 		if (!cmd_sg_entries)
2917 			cmd_sg_entries = srp_sg_tablesize;
2918 	}
2919 
2920 	if (!cmd_sg_entries)
2921 		cmd_sg_entries = SRP_DEF_SG_TABLESIZE;
2922 
2923 	if (cmd_sg_entries > 255) {
2924 		pr_warn("Clamping cmd_sg_entries to 255\n");
2925 		cmd_sg_entries = 255;
2926 	}
2927 
2928 	if (!indirect_sg_entries)
2929 		indirect_sg_entries = cmd_sg_entries;
2930 	else if (indirect_sg_entries < cmd_sg_entries) {
2931 		pr_warn("Bumping up indirect_sg_entries to match cmd_sg_entries (%u)\n",
2932 			cmd_sg_entries);
2933 		indirect_sg_entries = cmd_sg_entries;
2934 	}
2935 
2936 	ib_srp_transport_template =
2937 		srp_attach_transport(&ib_srp_transport_functions);
2938 	if (!ib_srp_transport_template)
2939 		return -ENOMEM;
2940 
2941 	ret = class_register(&srp_class);
2942 	if (ret) {
2943 		pr_err("couldn't register class infiniband_srp\n");
2944 		srp_release_transport(ib_srp_transport_template);
2945 		return ret;
2946 	}
2947 
2948 	ib_sa_register_client(&srp_sa_client);
2949 
2950 	ret = ib_register_client(&srp_client);
2951 	if (ret) {
2952 		pr_err("couldn't register IB client\n");
2953 		srp_release_transport(ib_srp_transport_template);
2954 		ib_sa_unregister_client(&srp_sa_client);
2955 		class_unregister(&srp_class);
2956 		return ret;
2957 	}
2958 
2959 	return 0;
2960 }
2961 
2962 static void __exit srp_cleanup_module(void)
2963 {
2964 	ib_unregister_client(&srp_client);
2965 	ib_sa_unregister_client(&srp_sa_client);
2966 	class_unregister(&srp_class);
2967 	srp_release_transport(ib_srp_transport_template);
2968 }
2969 
2970 module_init(srp_init_module);
2971 module_exit(srp_cleanup_module);
2972