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) KBUILD_MODNAME ": " 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 bool prefer_fr;
70 static bool register_always;
71 static int topspin_workarounds = 1;
72 
73 module_param(srp_sg_tablesize, uint, 0444);
74 MODULE_PARM_DESC(srp_sg_tablesize, "Deprecated name for cmd_sg_entries");
75 
76 module_param(cmd_sg_entries, uint, 0444);
77 MODULE_PARM_DESC(cmd_sg_entries,
78 		 "Default number of gather/scatter entries in the SRP command (default is 12, max 255)");
79 
80 module_param(indirect_sg_entries, uint, 0444);
81 MODULE_PARM_DESC(indirect_sg_entries,
82 		 "Default max number of gather/scatter entries (default is 12, max is " __stringify(SCSI_MAX_SG_CHAIN_SEGMENTS) ")");
83 
84 module_param(allow_ext_sg, bool, 0444);
85 MODULE_PARM_DESC(allow_ext_sg,
86 		  "Default behavior when there are more than cmd_sg_entries S/G entries after mapping; fails the request when false (default false)");
87 
88 module_param(topspin_workarounds, int, 0444);
89 MODULE_PARM_DESC(topspin_workarounds,
90 		 "Enable workarounds for Topspin/Cisco SRP target bugs if != 0");
91 
92 module_param(prefer_fr, bool, 0444);
93 MODULE_PARM_DESC(prefer_fr,
94 "Whether to use fast registration if both FMR and fast registration are supported");
95 
96 module_param(register_always, bool, 0444);
97 MODULE_PARM_DESC(register_always,
98 		 "Use memory registration even for contiguous memory regions");
99 
100 static struct kernel_param_ops srp_tmo_ops;
101 
102 static int srp_reconnect_delay = 10;
103 module_param_cb(reconnect_delay, &srp_tmo_ops, &srp_reconnect_delay,
104 		S_IRUGO | S_IWUSR);
105 MODULE_PARM_DESC(reconnect_delay, "Time between successive reconnect attempts");
106 
107 static int srp_fast_io_fail_tmo = 15;
108 module_param_cb(fast_io_fail_tmo, &srp_tmo_ops, &srp_fast_io_fail_tmo,
109 		S_IRUGO | S_IWUSR);
110 MODULE_PARM_DESC(fast_io_fail_tmo,
111 		 "Number of seconds between the observation of a transport"
112 		 " layer error and failing all I/O. \"off\" means that this"
113 		 " functionality is disabled.");
114 
115 static int srp_dev_loss_tmo = 600;
116 module_param_cb(dev_loss_tmo, &srp_tmo_ops, &srp_dev_loss_tmo,
117 		S_IRUGO | S_IWUSR);
118 MODULE_PARM_DESC(dev_loss_tmo,
119 		 "Maximum number of seconds that the SRP transport should"
120 		 " insulate transport layer errors. After this time has been"
121 		 " exceeded the SCSI host is removed. Should be"
122 		 " between 1 and " __stringify(SCSI_DEVICE_BLOCK_MAX_TIMEOUT)
123 		 " if fast_io_fail_tmo has not been set. \"off\" means that"
124 		 " this functionality is disabled.");
125 
126 static void srp_add_one(struct ib_device *device);
127 static void srp_remove_one(struct ib_device *device);
128 static void srp_recv_completion(struct ib_cq *cq, void *target_ptr);
129 static void srp_send_completion(struct ib_cq *cq, void *target_ptr);
130 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event);
131 
132 static struct scsi_transport_template *ib_srp_transport_template;
133 static struct workqueue_struct *srp_remove_wq;
134 
135 static struct ib_client srp_client = {
136 	.name   = "srp",
137 	.add    = srp_add_one,
138 	.remove = srp_remove_one
139 };
140 
141 static struct ib_sa_client srp_sa_client;
142 
143 static int srp_tmo_get(char *buffer, const struct kernel_param *kp)
144 {
145 	int tmo = *(int *)kp->arg;
146 
147 	if (tmo >= 0)
148 		return sprintf(buffer, "%d", tmo);
149 	else
150 		return sprintf(buffer, "off");
151 }
152 
153 static int srp_tmo_set(const char *val, const struct kernel_param *kp)
154 {
155 	int tmo, res;
156 
157 	if (strncmp(val, "off", 3) != 0) {
158 		res = kstrtoint(val, 0, &tmo);
159 		if (res)
160 			goto out;
161 	} else {
162 		tmo = -1;
163 	}
164 	if (kp->arg == &srp_reconnect_delay)
165 		res = srp_tmo_valid(tmo, srp_fast_io_fail_tmo,
166 				    srp_dev_loss_tmo);
167 	else if (kp->arg == &srp_fast_io_fail_tmo)
168 		res = srp_tmo_valid(srp_reconnect_delay, tmo, srp_dev_loss_tmo);
169 	else
170 		res = srp_tmo_valid(srp_reconnect_delay, srp_fast_io_fail_tmo,
171 				    tmo);
172 	if (res)
173 		goto out;
174 	*(int *)kp->arg = tmo;
175 
176 out:
177 	return res;
178 }
179 
180 static struct kernel_param_ops srp_tmo_ops = {
181 	.get = srp_tmo_get,
182 	.set = srp_tmo_set,
183 };
184 
185 static inline struct srp_target_port *host_to_target(struct Scsi_Host *host)
186 {
187 	return (struct srp_target_port *) host->hostdata;
188 }
189 
190 static const char *srp_target_info(struct Scsi_Host *host)
191 {
192 	return host_to_target(host)->target_name;
193 }
194 
195 static int srp_target_is_topspin(struct srp_target_port *target)
196 {
197 	static const u8 topspin_oui[3] = { 0x00, 0x05, 0xad };
198 	static const u8 cisco_oui[3]   = { 0x00, 0x1b, 0x0d };
199 
200 	return topspin_workarounds &&
201 		(!memcmp(&target->ioc_guid, topspin_oui, sizeof topspin_oui) ||
202 		 !memcmp(&target->ioc_guid, cisco_oui, sizeof cisco_oui));
203 }
204 
205 static struct srp_iu *srp_alloc_iu(struct srp_host *host, size_t size,
206 				   gfp_t gfp_mask,
207 				   enum dma_data_direction direction)
208 {
209 	struct srp_iu *iu;
210 
211 	iu = kmalloc(sizeof *iu, gfp_mask);
212 	if (!iu)
213 		goto out;
214 
215 	iu->buf = kzalloc(size, gfp_mask);
216 	if (!iu->buf)
217 		goto out_free_iu;
218 
219 	iu->dma = ib_dma_map_single(host->srp_dev->dev, iu->buf, size,
220 				    direction);
221 	if (ib_dma_mapping_error(host->srp_dev->dev, iu->dma))
222 		goto out_free_buf;
223 
224 	iu->size      = size;
225 	iu->direction = direction;
226 
227 	return iu;
228 
229 out_free_buf:
230 	kfree(iu->buf);
231 out_free_iu:
232 	kfree(iu);
233 out:
234 	return NULL;
235 }
236 
237 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
238 {
239 	if (!iu)
240 		return;
241 
242 	ib_dma_unmap_single(host->srp_dev->dev, iu->dma, iu->size,
243 			    iu->direction);
244 	kfree(iu->buf);
245 	kfree(iu);
246 }
247 
248 static void srp_qp_event(struct ib_event *event, void *context)
249 {
250 	pr_debug("QP event %d\n", event->event);
251 }
252 
253 static int srp_init_qp(struct srp_target_port *target,
254 		       struct ib_qp *qp)
255 {
256 	struct ib_qp_attr *attr;
257 	int ret;
258 
259 	attr = kmalloc(sizeof *attr, GFP_KERNEL);
260 	if (!attr)
261 		return -ENOMEM;
262 
263 	ret = ib_find_pkey(target->srp_host->srp_dev->dev,
264 			   target->srp_host->port,
265 			   be16_to_cpu(target->path.pkey),
266 			   &attr->pkey_index);
267 	if (ret)
268 		goto out;
269 
270 	attr->qp_state        = IB_QPS_INIT;
271 	attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
272 				    IB_ACCESS_REMOTE_WRITE);
273 	attr->port_num        = target->srp_host->port;
274 
275 	ret = ib_modify_qp(qp, attr,
276 			   IB_QP_STATE		|
277 			   IB_QP_PKEY_INDEX	|
278 			   IB_QP_ACCESS_FLAGS	|
279 			   IB_QP_PORT);
280 
281 out:
282 	kfree(attr);
283 	return ret;
284 }
285 
286 static int srp_new_cm_id(struct srp_target_port *target)
287 {
288 	struct ib_cm_id *new_cm_id;
289 
290 	new_cm_id = ib_create_cm_id(target->srp_host->srp_dev->dev,
291 				    srp_cm_handler, target);
292 	if (IS_ERR(new_cm_id))
293 		return PTR_ERR(new_cm_id);
294 
295 	if (target->cm_id)
296 		ib_destroy_cm_id(target->cm_id);
297 	target->cm_id = new_cm_id;
298 
299 	return 0;
300 }
301 
302 static struct ib_fmr_pool *srp_alloc_fmr_pool(struct srp_target_port *target)
303 {
304 	struct srp_device *dev = target->srp_host->srp_dev;
305 	struct ib_fmr_pool_param fmr_param;
306 
307 	memset(&fmr_param, 0, sizeof(fmr_param));
308 	fmr_param.pool_size	    = target->scsi_host->can_queue;
309 	fmr_param.dirty_watermark   = fmr_param.pool_size / 4;
310 	fmr_param.cache		    = 1;
311 	fmr_param.max_pages_per_fmr = dev->max_pages_per_mr;
312 	fmr_param.page_shift	    = ilog2(dev->mr_page_size);
313 	fmr_param.access	    = (IB_ACCESS_LOCAL_WRITE |
314 				       IB_ACCESS_REMOTE_WRITE |
315 				       IB_ACCESS_REMOTE_READ);
316 
317 	return ib_create_fmr_pool(dev->pd, &fmr_param);
318 }
319 
320 /**
321  * srp_destroy_fr_pool() - free the resources owned by a pool
322  * @pool: Fast registration pool to be destroyed.
323  */
324 static void srp_destroy_fr_pool(struct srp_fr_pool *pool)
325 {
326 	int i;
327 	struct srp_fr_desc *d;
328 
329 	if (!pool)
330 		return;
331 
332 	for (i = 0, d = &pool->desc[0]; i < pool->size; i++, d++) {
333 		if (d->frpl)
334 			ib_free_fast_reg_page_list(d->frpl);
335 		if (d->mr)
336 			ib_dereg_mr(d->mr);
337 	}
338 	kfree(pool);
339 }
340 
341 /**
342  * srp_create_fr_pool() - allocate and initialize a pool for fast registration
343  * @device:            IB device to allocate fast registration descriptors for.
344  * @pd:                Protection domain associated with the FR descriptors.
345  * @pool_size:         Number of descriptors to allocate.
346  * @max_page_list_len: Maximum fast registration work request page list length.
347  */
348 static struct srp_fr_pool *srp_create_fr_pool(struct ib_device *device,
349 					      struct ib_pd *pd, int pool_size,
350 					      int max_page_list_len)
351 {
352 	struct srp_fr_pool *pool;
353 	struct srp_fr_desc *d;
354 	struct ib_mr *mr;
355 	struct ib_fast_reg_page_list *frpl;
356 	int i, ret = -EINVAL;
357 
358 	if (pool_size <= 0)
359 		goto err;
360 	ret = -ENOMEM;
361 	pool = kzalloc(sizeof(struct srp_fr_pool) +
362 		       pool_size * sizeof(struct srp_fr_desc), GFP_KERNEL);
363 	if (!pool)
364 		goto err;
365 	pool->size = pool_size;
366 	pool->max_page_list_len = max_page_list_len;
367 	spin_lock_init(&pool->lock);
368 	INIT_LIST_HEAD(&pool->free_list);
369 
370 	for (i = 0, d = &pool->desc[0]; i < pool->size; i++, d++) {
371 		mr = ib_alloc_fast_reg_mr(pd, max_page_list_len);
372 		if (IS_ERR(mr)) {
373 			ret = PTR_ERR(mr);
374 			goto destroy_pool;
375 		}
376 		d->mr = mr;
377 		frpl = ib_alloc_fast_reg_page_list(device, max_page_list_len);
378 		if (IS_ERR(frpl)) {
379 			ret = PTR_ERR(frpl);
380 			goto destroy_pool;
381 		}
382 		d->frpl = frpl;
383 		list_add_tail(&d->entry, &pool->free_list);
384 	}
385 
386 out:
387 	return pool;
388 
389 destroy_pool:
390 	srp_destroy_fr_pool(pool);
391 
392 err:
393 	pool = ERR_PTR(ret);
394 	goto out;
395 }
396 
397 /**
398  * srp_fr_pool_get() - obtain a descriptor suitable for fast registration
399  * @pool: Pool to obtain descriptor from.
400  */
401 static struct srp_fr_desc *srp_fr_pool_get(struct srp_fr_pool *pool)
402 {
403 	struct srp_fr_desc *d = NULL;
404 	unsigned long flags;
405 
406 	spin_lock_irqsave(&pool->lock, flags);
407 	if (!list_empty(&pool->free_list)) {
408 		d = list_first_entry(&pool->free_list, typeof(*d), entry);
409 		list_del(&d->entry);
410 	}
411 	spin_unlock_irqrestore(&pool->lock, flags);
412 
413 	return d;
414 }
415 
416 /**
417  * srp_fr_pool_put() - put an FR descriptor back in the free list
418  * @pool: Pool the descriptor was allocated from.
419  * @desc: Pointer to an array of fast registration descriptor pointers.
420  * @n:    Number of descriptors to put back.
421  *
422  * Note: The caller must already have queued an invalidation request for
423  * desc->mr->rkey before calling this function.
424  */
425 static void srp_fr_pool_put(struct srp_fr_pool *pool, struct srp_fr_desc **desc,
426 			    int n)
427 {
428 	unsigned long flags;
429 	int i;
430 
431 	spin_lock_irqsave(&pool->lock, flags);
432 	for (i = 0; i < n; i++)
433 		list_add(&desc[i]->entry, &pool->free_list);
434 	spin_unlock_irqrestore(&pool->lock, flags);
435 }
436 
437 static struct srp_fr_pool *srp_alloc_fr_pool(struct srp_target_port *target)
438 {
439 	struct srp_device *dev = target->srp_host->srp_dev;
440 
441 	return srp_create_fr_pool(dev->dev, dev->pd,
442 				  target->scsi_host->can_queue,
443 				  dev->max_pages_per_mr);
444 }
445 
446 static int srp_create_target_ib(struct srp_target_port *target)
447 {
448 	struct srp_device *dev = target->srp_host->srp_dev;
449 	struct ib_qp_init_attr *init_attr;
450 	struct ib_cq *recv_cq, *send_cq;
451 	struct ib_qp *qp;
452 	struct ib_fmr_pool *fmr_pool = NULL;
453 	struct srp_fr_pool *fr_pool = NULL;
454 	const int m = 1 + dev->use_fast_reg;
455 	int ret;
456 
457 	init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
458 	if (!init_attr)
459 		return -ENOMEM;
460 
461 	recv_cq = ib_create_cq(dev->dev, srp_recv_completion, NULL, target,
462 			       target->queue_size, target->comp_vector);
463 	if (IS_ERR(recv_cq)) {
464 		ret = PTR_ERR(recv_cq);
465 		goto err;
466 	}
467 
468 	send_cq = ib_create_cq(dev->dev, srp_send_completion, NULL, target,
469 			       m * target->queue_size, target->comp_vector);
470 	if (IS_ERR(send_cq)) {
471 		ret = PTR_ERR(send_cq);
472 		goto err_recv_cq;
473 	}
474 
475 	ib_req_notify_cq(recv_cq, IB_CQ_NEXT_COMP);
476 
477 	init_attr->event_handler       = srp_qp_event;
478 	init_attr->cap.max_send_wr     = m * target->queue_size;
479 	init_attr->cap.max_recv_wr     = target->queue_size;
480 	init_attr->cap.max_recv_sge    = 1;
481 	init_attr->cap.max_send_sge    = 1;
482 	init_attr->sq_sig_type         = IB_SIGNAL_REQ_WR;
483 	init_attr->qp_type             = IB_QPT_RC;
484 	init_attr->send_cq             = send_cq;
485 	init_attr->recv_cq             = recv_cq;
486 
487 	qp = ib_create_qp(dev->pd, init_attr);
488 	if (IS_ERR(qp)) {
489 		ret = PTR_ERR(qp);
490 		goto err_send_cq;
491 	}
492 
493 	ret = srp_init_qp(target, qp);
494 	if (ret)
495 		goto err_qp;
496 
497 	if (dev->use_fast_reg && dev->has_fr) {
498 		fr_pool = srp_alloc_fr_pool(target);
499 		if (IS_ERR(fr_pool)) {
500 			ret = PTR_ERR(fr_pool);
501 			shost_printk(KERN_WARNING, target->scsi_host, PFX
502 				     "FR pool allocation failed (%d)\n", ret);
503 			goto err_qp;
504 		}
505 		if (target->fr_pool)
506 			srp_destroy_fr_pool(target->fr_pool);
507 		target->fr_pool = fr_pool;
508 	} else if (!dev->use_fast_reg && dev->has_fmr) {
509 		fmr_pool = srp_alloc_fmr_pool(target);
510 		if (IS_ERR(fmr_pool)) {
511 			ret = PTR_ERR(fmr_pool);
512 			shost_printk(KERN_WARNING, target->scsi_host, PFX
513 				     "FMR pool allocation failed (%d)\n", ret);
514 			goto err_qp;
515 		}
516 		if (target->fmr_pool)
517 			ib_destroy_fmr_pool(target->fmr_pool);
518 		target->fmr_pool = fmr_pool;
519 	}
520 
521 	if (target->qp)
522 		ib_destroy_qp(target->qp);
523 	if (target->recv_cq)
524 		ib_destroy_cq(target->recv_cq);
525 	if (target->send_cq)
526 		ib_destroy_cq(target->send_cq);
527 
528 	target->qp = qp;
529 	target->recv_cq = recv_cq;
530 	target->send_cq = send_cq;
531 
532 	kfree(init_attr);
533 	return 0;
534 
535 err_qp:
536 	ib_destroy_qp(qp);
537 
538 err_send_cq:
539 	ib_destroy_cq(send_cq);
540 
541 err_recv_cq:
542 	ib_destroy_cq(recv_cq);
543 
544 err:
545 	kfree(init_attr);
546 	return ret;
547 }
548 
549 /*
550  * Note: this function may be called without srp_alloc_iu_bufs() having been
551  * invoked. Hence the target->[rt]x_ring checks.
552  */
553 static void srp_free_target_ib(struct srp_target_port *target)
554 {
555 	struct srp_device *dev = target->srp_host->srp_dev;
556 	int i;
557 
558 	if (dev->use_fast_reg) {
559 		if (target->fr_pool)
560 			srp_destroy_fr_pool(target->fr_pool);
561 	} else {
562 		if (target->fmr_pool)
563 			ib_destroy_fmr_pool(target->fmr_pool);
564 	}
565 	ib_destroy_qp(target->qp);
566 	ib_destroy_cq(target->send_cq);
567 	ib_destroy_cq(target->recv_cq);
568 
569 	target->qp = NULL;
570 	target->send_cq = target->recv_cq = NULL;
571 
572 	if (target->rx_ring) {
573 		for (i = 0; i < target->queue_size; ++i)
574 			srp_free_iu(target->srp_host, target->rx_ring[i]);
575 		kfree(target->rx_ring);
576 		target->rx_ring = NULL;
577 	}
578 	if (target->tx_ring) {
579 		for (i = 0; i < target->queue_size; ++i)
580 			srp_free_iu(target->srp_host, target->tx_ring[i]);
581 		kfree(target->tx_ring);
582 		target->tx_ring = NULL;
583 	}
584 }
585 
586 static void srp_path_rec_completion(int status,
587 				    struct ib_sa_path_rec *pathrec,
588 				    void *target_ptr)
589 {
590 	struct srp_target_port *target = target_ptr;
591 
592 	target->status = status;
593 	if (status)
594 		shost_printk(KERN_ERR, target->scsi_host,
595 			     PFX "Got failed path rec status %d\n", status);
596 	else
597 		target->path = *pathrec;
598 	complete(&target->done);
599 }
600 
601 static int srp_lookup_path(struct srp_target_port *target)
602 {
603 	int ret;
604 
605 	target->path.numb_path = 1;
606 
607 	init_completion(&target->done);
608 
609 	target->path_query_id = ib_sa_path_rec_get(&srp_sa_client,
610 						   target->srp_host->srp_dev->dev,
611 						   target->srp_host->port,
612 						   &target->path,
613 						   IB_SA_PATH_REC_SERVICE_ID	|
614 						   IB_SA_PATH_REC_DGID		|
615 						   IB_SA_PATH_REC_SGID		|
616 						   IB_SA_PATH_REC_NUMB_PATH	|
617 						   IB_SA_PATH_REC_PKEY,
618 						   SRP_PATH_REC_TIMEOUT_MS,
619 						   GFP_KERNEL,
620 						   srp_path_rec_completion,
621 						   target, &target->path_query);
622 	if (target->path_query_id < 0)
623 		return target->path_query_id;
624 
625 	ret = wait_for_completion_interruptible(&target->done);
626 	if (ret < 0)
627 		return ret;
628 
629 	if (target->status < 0)
630 		shost_printk(KERN_WARNING, target->scsi_host,
631 			     PFX "Path record query failed\n");
632 
633 	return target->status;
634 }
635 
636 static int srp_send_req(struct srp_target_port *target)
637 {
638 	struct {
639 		struct ib_cm_req_param param;
640 		struct srp_login_req   priv;
641 	} *req = NULL;
642 	int status;
643 
644 	req = kzalloc(sizeof *req, GFP_KERNEL);
645 	if (!req)
646 		return -ENOMEM;
647 
648 	req->param.primary_path 	      = &target->path;
649 	req->param.alternate_path 	      = NULL;
650 	req->param.service_id 		      = target->service_id;
651 	req->param.qp_num 		      = target->qp->qp_num;
652 	req->param.qp_type 		      = target->qp->qp_type;
653 	req->param.private_data 	      = &req->priv;
654 	req->param.private_data_len 	      = sizeof req->priv;
655 	req->param.flow_control 	      = 1;
656 
657 	get_random_bytes(&req->param.starting_psn, 4);
658 	req->param.starting_psn 	     &= 0xffffff;
659 
660 	/*
661 	 * Pick some arbitrary defaults here; we could make these
662 	 * module parameters if anyone cared about setting them.
663 	 */
664 	req->param.responder_resources	      = 4;
665 	req->param.remote_cm_response_timeout = 20;
666 	req->param.local_cm_response_timeout  = 20;
667 	req->param.retry_count                = target->tl_retry_count;
668 	req->param.rnr_retry_count 	      = 7;
669 	req->param.max_cm_retries 	      = 15;
670 
671 	req->priv.opcode     	= SRP_LOGIN_REQ;
672 	req->priv.tag        	= 0;
673 	req->priv.req_it_iu_len = cpu_to_be32(target->max_iu_len);
674 	req->priv.req_buf_fmt 	= cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
675 					      SRP_BUF_FORMAT_INDIRECT);
676 	/*
677 	 * In the published SRP specification (draft rev. 16a), the
678 	 * port identifier format is 8 bytes of ID extension followed
679 	 * by 8 bytes of GUID.  Older drafts put the two halves in the
680 	 * opposite order, so that the GUID comes first.
681 	 *
682 	 * Targets conforming to these obsolete drafts can be
683 	 * recognized by the I/O Class they report.
684 	 */
685 	if (target->io_class == SRP_REV10_IB_IO_CLASS) {
686 		memcpy(req->priv.initiator_port_id,
687 		       &target->path.sgid.global.interface_id, 8);
688 		memcpy(req->priv.initiator_port_id + 8,
689 		       &target->initiator_ext, 8);
690 		memcpy(req->priv.target_port_id,     &target->ioc_guid, 8);
691 		memcpy(req->priv.target_port_id + 8, &target->id_ext, 8);
692 	} else {
693 		memcpy(req->priv.initiator_port_id,
694 		       &target->initiator_ext, 8);
695 		memcpy(req->priv.initiator_port_id + 8,
696 		       &target->path.sgid.global.interface_id, 8);
697 		memcpy(req->priv.target_port_id,     &target->id_ext, 8);
698 		memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
699 	}
700 
701 	/*
702 	 * Topspin/Cisco SRP targets will reject our login unless we
703 	 * zero out the first 8 bytes of our initiator port ID and set
704 	 * the second 8 bytes to the local node GUID.
705 	 */
706 	if (srp_target_is_topspin(target)) {
707 		shost_printk(KERN_DEBUG, target->scsi_host,
708 			     PFX "Topspin/Cisco initiator port ID workaround "
709 			     "activated for target GUID %016llx\n",
710 			     (unsigned long long) be64_to_cpu(target->ioc_guid));
711 		memset(req->priv.initiator_port_id, 0, 8);
712 		memcpy(req->priv.initiator_port_id + 8,
713 		       &target->srp_host->srp_dev->dev->node_guid, 8);
714 	}
715 
716 	status = ib_send_cm_req(target->cm_id, &req->param);
717 
718 	kfree(req);
719 
720 	return status;
721 }
722 
723 static bool srp_queue_remove_work(struct srp_target_port *target)
724 {
725 	bool changed = false;
726 
727 	spin_lock_irq(&target->lock);
728 	if (target->state != SRP_TARGET_REMOVED) {
729 		target->state = SRP_TARGET_REMOVED;
730 		changed = true;
731 	}
732 	spin_unlock_irq(&target->lock);
733 
734 	if (changed)
735 		queue_work(srp_remove_wq, &target->remove_work);
736 
737 	return changed;
738 }
739 
740 static bool srp_change_conn_state(struct srp_target_port *target,
741 				  bool connected)
742 {
743 	bool changed = false;
744 
745 	spin_lock_irq(&target->lock);
746 	if (target->connected != connected) {
747 		target->connected = connected;
748 		changed = true;
749 	}
750 	spin_unlock_irq(&target->lock);
751 
752 	return changed;
753 }
754 
755 static void srp_disconnect_target(struct srp_target_port *target)
756 {
757 	if (srp_change_conn_state(target, false)) {
758 		/* XXX should send SRP_I_LOGOUT request */
759 
760 		if (ib_send_cm_dreq(target->cm_id, NULL, 0)) {
761 			shost_printk(KERN_DEBUG, target->scsi_host,
762 				     PFX "Sending CM DREQ failed\n");
763 		}
764 	}
765 }
766 
767 static void srp_free_req_data(struct srp_target_port *target)
768 {
769 	struct srp_device *dev = target->srp_host->srp_dev;
770 	struct ib_device *ibdev = dev->dev;
771 	struct srp_request *req;
772 	int i;
773 
774 	if (!target->req_ring)
775 		return;
776 
777 	for (i = 0; i < target->req_ring_size; ++i) {
778 		req = &target->req_ring[i];
779 		if (dev->use_fast_reg)
780 			kfree(req->fr_list);
781 		else
782 			kfree(req->fmr_list);
783 		kfree(req->map_page);
784 		if (req->indirect_dma_addr) {
785 			ib_dma_unmap_single(ibdev, req->indirect_dma_addr,
786 					    target->indirect_size,
787 					    DMA_TO_DEVICE);
788 		}
789 		kfree(req->indirect_desc);
790 	}
791 
792 	kfree(target->req_ring);
793 	target->req_ring = NULL;
794 }
795 
796 static int srp_alloc_req_data(struct srp_target_port *target)
797 {
798 	struct srp_device *srp_dev = target->srp_host->srp_dev;
799 	struct ib_device *ibdev = srp_dev->dev;
800 	struct srp_request *req;
801 	void *mr_list;
802 	dma_addr_t dma_addr;
803 	int i, ret = -ENOMEM;
804 
805 	INIT_LIST_HEAD(&target->free_reqs);
806 
807 	target->req_ring = kzalloc(target->req_ring_size *
808 				   sizeof(*target->req_ring), GFP_KERNEL);
809 	if (!target->req_ring)
810 		goto out;
811 
812 	for (i = 0; i < target->req_ring_size; ++i) {
813 		req = &target->req_ring[i];
814 		mr_list = kmalloc(target->cmd_sg_cnt * sizeof(void *),
815 				  GFP_KERNEL);
816 		if (!mr_list)
817 			goto out;
818 		if (srp_dev->use_fast_reg)
819 			req->fr_list = mr_list;
820 		else
821 			req->fmr_list = mr_list;
822 		req->map_page = kmalloc(srp_dev->max_pages_per_mr *
823 					sizeof(void *), GFP_KERNEL);
824 		if (!req->map_page)
825 			goto out;
826 		req->indirect_desc = kmalloc(target->indirect_size, GFP_KERNEL);
827 		if (!req->indirect_desc)
828 			goto out;
829 
830 		dma_addr = ib_dma_map_single(ibdev, req->indirect_desc,
831 					     target->indirect_size,
832 					     DMA_TO_DEVICE);
833 		if (ib_dma_mapping_error(ibdev, dma_addr))
834 			goto out;
835 
836 		req->indirect_dma_addr = dma_addr;
837 		req->index = i;
838 		list_add_tail(&req->list, &target->free_reqs);
839 	}
840 	ret = 0;
841 
842 out:
843 	return ret;
844 }
845 
846 /**
847  * srp_del_scsi_host_attr() - Remove attributes defined in the host template.
848  * @shost: SCSI host whose attributes to remove from sysfs.
849  *
850  * Note: Any attributes defined in the host template and that did not exist
851  * before invocation of this function will be ignored.
852  */
853 static void srp_del_scsi_host_attr(struct Scsi_Host *shost)
854 {
855 	struct device_attribute **attr;
856 
857 	for (attr = shost->hostt->shost_attrs; attr && *attr; ++attr)
858 		device_remove_file(&shost->shost_dev, *attr);
859 }
860 
861 static void srp_remove_target(struct srp_target_port *target)
862 {
863 	WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
864 
865 	srp_del_scsi_host_attr(target->scsi_host);
866 	srp_rport_get(target->rport);
867 	srp_remove_host(target->scsi_host);
868 	scsi_remove_host(target->scsi_host);
869 	srp_stop_rport_timers(target->rport);
870 	srp_disconnect_target(target);
871 	ib_destroy_cm_id(target->cm_id);
872 	srp_free_target_ib(target);
873 	cancel_work_sync(&target->tl_err_work);
874 	srp_rport_put(target->rport);
875 	srp_free_req_data(target);
876 
877 	spin_lock(&target->srp_host->target_lock);
878 	list_del(&target->list);
879 	spin_unlock(&target->srp_host->target_lock);
880 
881 	scsi_host_put(target->scsi_host);
882 }
883 
884 static void srp_remove_work(struct work_struct *work)
885 {
886 	struct srp_target_port *target =
887 		container_of(work, struct srp_target_port, remove_work);
888 
889 	WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
890 
891 	srp_remove_target(target);
892 }
893 
894 static void srp_rport_delete(struct srp_rport *rport)
895 {
896 	struct srp_target_port *target = rport->lld_data;
897 
898 	srp_queue_remove_work(target);
899 }
900 
901 static int srp_connect_target(struct srp_target_port *target)
902 {
903 	int retries = 3;
904 	int ret;
905 
906 	WARN_ON_ONCE(target->connected);
907 
908 	target->qp_in_error = false;
909 
910 	ret = srp_lookup_path(target);
911 	if (ret)
912 		return ret;
913 
914 	while (1) {
915 		init_completion(&target->done);
916 		ret = srp_send_req(target);
917 		if (ret)
918 			return ret;
919 		ret = wait_for_completion_interruptible(&target->done);
920 		if (ret < 0)
921 			return ret;
922 
923 		/*
924 		 * The CM event handling code will set status to
925 		 * SRP_PORT_REDIRECT if we get a port redirect REJ
926 		 * back, or SRP_DLID_REDIRECT if we get a lid/qp
927 		 * redirect REJ back.
928 		 */
929 		switch (target->status) {
930 		case 0:
931 			srp_change_conn_state(target, true);
932 			return 0;
933 
934 		case SRP_PORT_REDIRECT:
935 			ret = srp_lookup_path(target);
936 			if (ret)
937 				return ret;
938 			break;
939 
940 		case SRP_DLID_REDIRECT:
941 			break;
942 
943 		case SRP_STALE_CONN:
944 			/* Our current CM id was stale, and is now in timewait.
945 			 * Try to reconnect with a new one.
946 			 */
947 			if (!retries-- || srp_new_cm_id(target)) {
948 				shost_printk(KERN_ERR, target->scsi_host, PFX
949 					     "giving up on stale connection\n");
950 				target->status = -ECONNRESET;
951 				return target->status;
952 			}
953 
954 			shost_printk(KERN_ERR, target->scsi_host, PFX
955 				     "retrying stale connection\n");
956 			break;
957 
958 		default:
959 			return target->status;
960 		}
961 	}
962 }
963 
964 static int srp_inv_rkey(struct srp_target_port *target, u32 rkey)
965 {
966 	struct ib_send_wr *bad_wr;
967 	struct ib_send_wr wr = {
968 		.opcode		    = IB_WR_LOCAL_INV,
969 		.wr_id		    = LOCAL_INV_WR_ID_MASK,
970 		.next		    = NULL,
971 		.num_sge	    = 0,
972 		.send_flags	    = 0,
973 		.ex.invalidate_rkey = rkey,
974 	};
975 
976 	return ib_post_send(target->qp, &wr, &bad_wr);
977 }
978 
979 static void srp_unmap_data(struct scsi_cmnd *scmnd,
980 			   struct srp_target_port *target,
981 			   struct srp_request *req)
982 {
983 	struct srp_device *dev = target->srp_host->srp_dev;
984 	struct ib_device *ibdev = dev->dev;
985 	int i, res;
986 
987 	if (!scsi_sglist(scmnd) ||
988 	    (scmnd->sc_data_direction != DMA_TO_DEVICE &&
989 	     scmnd->sc_data_direction != DMA_FROM_DEVICE))
990 		return;
991 
992 	if (dev->use_fast_reg) {
993 		struct srp_fr_desc **pfr;
994 
995 		for (i = req->nmdesc, pfr = req->fr_list; i > 0; i--, pfr++) {
996 			res = srp_inv_rkey(target, (*pfr)->mr->rkey);
997 			if (res < 0) {
998 				shost_printk(KERN_ERR, target->scsi_host, PFX
999 				  "Queueing INV WR for rkey %#x failed (%d)\n",
1000 				  (*pfr)->mr->rkey, res);
1001 				queue_work(system_long_wq,
1002 					   &target->tl_err_work);
1003 			}
1004 		}
1005 		if (req->nmdesc)
1006 			srp_fr_pool_put(target->fr_pool, req->fr_list,
1007 					req->nmdesc);
1008 	} else {
1009 		struct ib_pool_fmr **pfmr;
1010 
1011 		for (i = req->nmdesc, pfmr = req->fmr_list; i > 0; i--, pfmr++)
1012 			ib_fmr_pool_unmap(*pfmr);
1013 	}
1014 
1015 	ib_dma_unmap_sg(ibdev, scsi_sglist(scmnd), scsi_sg_count(scmnd),
1016 			scmnd->sc_data_direction);
1017 }
1018 
1019 /**
1020  * srp_claim_req - Take ownership of the scmnd associated with a request.
1021  * @target: SRP target port.
1022  * @req: SRP request.
1023  * @sdev: If not NULL, only take ownership for this SCSI device.
1024  * @scmnd: If NULL, take ownership of @req->scmnd. If not NULL, only take
1025  *         ownership of @req->scmnd if it equals @scmnd.
1026  *
1027  * Return value:
1028  * Either NULL or a pointer to the SCSI command the caller became owner of.
1029  */
1030 static struct scsi_cmnd *srp_claim_req(struct srp_target_port *target,
1031 				       struct srp_request *req,
1032 				       struct scsi_device *sdev,
1033 				       struct scsi_cmnd *scmnd)
1034 {
1035 	unsigned long flags;
1036 
1037 	spin_lock_irqsave(&target->lock, flags);
1038 	if (req->scmnd &&
1039 	    (!sdev || req->scmnd->device == sdev) &&
1040 	    (!scmnd || req->scmnd == scmnd)) {
1041 		scmnd = req->scmnd;
1042 		req->scmnd = NULL;
1043 	} else {
1044 		scmnd = NULL;
1045 	}
1046 	spin_unlock_irqrestore(&target->lock, flags);
1047 
1048 	return scmnd;
1049 }
1050 
1051 /**
1052  * srp_free_req() - Unmap data and add request to the free request list.
1053  * @target: SRP target port.
1054  * @req:    Request to be freed.
1055  * @scmnd:  SCSI command associated with @req.
1056  * @req_lim_delta: Amount to be added to @target->req_lim.
1057  */
1058 static void srp_free_req(struct srp_target_port *target,
1059 			 struct srp_request *req, struct scsi_cmnd *scmnd,
1060 			 s32 req_lim_delta)
1061 {
1062 	unsigned long flags;
1063 
1064 	srp_unmap_data(scmnd, target, req);
1065 
1066 	spin_lock_irqsave(&target->lock, flags);
1067 	target->req_lim += req_lim_delta;
1068 	list_add_tail(&req->list, &target->free_reqs);
1069 	spin_unlock_irqrestore(&target->lock, flags);
1070 }
1071 
1072 static void srp_finish_req(struct srp_target_port *target,
1073 			   struct srp_request *req, struct scsi_device *sdev,
1074 			   int result)
1075 {
1076 	struct scsi_cmnd *scmnd = srp_claim_req(target, req, sdev, NULL);
1077 
1078 	if (scmnd) {
1079 		srp_free_req(target, req, scmnd, 0);
1080 		scmnd->result = result;
1081 		scmnd->scsi_done(scmnd);
1082 	}
1083 }
1084 
1085 static void srp_terminate_io(struct srp_rport *rport)
1086 {
1087 	struct srp_target_port *target = rport->lld_data;
1088 	struct Scsi_Host *shost = target->scsi_host;
1089 	struct scsi_device *sdev;
1090 	int i;
1091 
1092 	/*
1093 	 * Invoking srp_terminate_io() while srp_queuecommand() is running
1094 	 * is not safe. Hence the warning statement below.
1095 	 */
1096 	shost_for_each_device(sdev, shost)
1097 		WARN_ON_ONCE(sdev->request_queue->request_fn_active);
1098 
1099 	for (i = 0; i < target->req_ring_size; ++i) {
1100 		struct srp_request *req = &target->req_ring[i];
1101 		srp_finish_req(target, req, NULL, DID_TRANSPORT_FAILFAST << 16);
1102 	}
1103 }
1104 
1105 /*
1106  * It is up to the caller to ensure that srp_rport_reconnect() calls are
1107  * serialized and that no concurrent srp_queuecommand(), srp_abort(),
1108  * srp_reset_device() or srp_reset_host() calls will occur while this function
1109  * is in progress. One way to realize that is not to call this function
1110  * directly but to call srp_reconnect_rport() instead since that last function
1111  * serializes calls of this function via rport->mutex and also blocks
1112  * srp_queuecommand() calls before invoking this function.
1113  */
1114 static int srp_rport_reconnect(struct srp_rport *rport)
1115 {
1116 	struct srp_target_port *target = rport->lld_data;
1117 	int i, ret;
1118 
1119 	srp_disconnect_target(target);
1120 	/*
1121 	 * Now get a new local CM ID so that we avoid confusing the target in
1122 	 * case things are really fouled up. Doing so also ensures that all CM
1123 	 * callbacks will have finished before a new QP is allocated.
1124 	 */
1125 	ret = srp_new_cm_id(target);
1126 
1127 	for (i = 0; i < target->req_ring_size; ++i) {
1128 		struct srp_request *req = &target->req_ring[i];
1129 		srp_finish_req(target, req, NULL, DID_RESET << 16);
1130 	}
1131 
1132 	/*
1133 	 * Whether or not creating a new CM ID succeeded, create a new
1134 	 * QP. This guarantees that all callback functions for the old QP have
1135 	 * finished before any send requests are posted on the new QP.
1136 	 */
1137 	ret += srp_create_target_ib(target);
1138 
1139 	INIT_LIST_HEAD(&target->free_tx);
1140 	for (i = 0; i < target->queue_size; ++i)
1141 		list_add(&target->tx_ring[i]->list, &target->free_tx);
1142 
1143 	if (ret == 0)
1144 		ret = srp_connect_target(target);
1145 
1146 	if (ret == 0)
1147 		shost_printk(KERN_INFO, target->scsi_host,
1148 			     PFX "reconnect succeeded\n");
1149 
1150 	return ret;
1151 }
1152 
1153 static void srp_map_desc(struct srp_map_state *state, dma_addr_t dma_addr,
1154 			 unsigned int dma_len, u32 rkey)
1155 {
1156 	struct srp_direct_buf *desc = state->desc;
1157 
1158 	desc->va = cpu_to_be64(dma_addr);
1159 	desc->key = cpu_to_be32(rkey);
1160 	desc->len = cpu_to_be32(dma_len);
1161 
1162 	state->total_len += dma_len;
1163 	state->desc++;
1164 	state->ndesc++;
1165 }
1166 
1167 static int srp_map_finish_fmr(struct srp_map_state *state,
1168 			      struct srp_target_port *target)
1169 {
1170 	struct ib_pool_fmr *fmr;
1171 	u64 io_addr = 0;
1172 
1173 	fmr = ib_fmr_pool_map_phys(target->fmr_pool, state->pages,
1174 				   state->npages, io_addr);
1175 	if (IS_ERR(fmr))
1176 		return PTR_ERR(fmr);
1177 
1178 	*state->next_fmr++ = fmr;
1179 	state->nmdesc++;
1180 
1181 	srp_map_desc(state, 0, state->dma_len, fmr->fmr->rkey);
1182 
1183 	return 0;
1184 }
1185 
1186 static int srp_map_finish_fr(struct srp_map_state *state,
1187 			     struct srp_target_port *target)
1188 {
1189 	struct srp_device *dev = target->srp_host->srp_dev;
1190 	struct ib_send_wr *bad_wr;
1191 	struct ib_send_wr wr;
1192 	struct srp_fr_desc *desc;
1193 	u32 rkey;
1194 
1195 	desc = srp_fr_pool_get(target->fr_pool);
1196 	if (!desc)
1197 		return -ENOMEM;
1198 
1199 	rkey = ib_inc_rkey(desc->mr->rkey);
1200 	ib_update_fast_reg_key(desc->mr, rkey);
1201 
1202 	memcpy(desc->frpl->page_list, state->pages,
1203 	       sizeof(state->pages[0]) * state->npages);
1204 
1205 	memset(&wr, 0, sizeof(wr));
1206 	wr.opcode = IB_WR_FAST_REG_MR;
1207 	wr.wr_id = FAST_REG_WR_ID_MASK;
1208 	wr.wr.fast_reg.iova_start = state->base_dma_addr;
1209 	wr.wr.fast_reg.page_list = desc->frpl;
1210 	wr.wr.fast_reg.page_list_len = state->npages;
1211 	wr.wr.fast_reg.page_shift = ilog2(dev->mr_page_size);
1212 	wr.wr.fast_reg.length = state->dma_len;
1213 	wr.wr.fast_reg.access_flags = (IB_ACCESS_LOCAL_WRITE |
1214 				       IB_ACCESS_REMOTE_READ |
1215 				       IB_ACCESS_REMOTE_WRITE);
1216 	wr.wr.fast_reg.rkey = desc->mr->lkey;
1217 
1218 	*state->next_fr++ = desc;
1219 	state->nmdesc++;
1220 
1221 	srp_map_desc(state, state->base_dma_addr, state->dma_len,
1222 		     desc->mr->rkey);
1223 
1224 	return ib_post_send(target->qp, &wr, &bad_wr);
1225 }
1226 
1227 static int srp_finish_mapping(struct srp_map_state *state,
1228 			      struct srp_target_port *target)
1229 {
1230 	int ret = 0;
1231 
1232 	if (state->npages == 0)
1233 		return 0;
1234 
1235 	if (state->npages == 1 && !register_always)
1236 		srp_map_desc(state, state->base_dma_addr, state->dma_len,
1237 			     target->rkey);
1238 	else
1239 		ret = target->srp_host->srp_dev->use_fast_reg ?
1240 			srp_map_finish_fr(state, target) :
1241 			srp_map_finish_fmr(state, target);
1242 
1243 	if (ret == 0) {
1244 		state->npages = 0;
1245 		state->dma_len = 0;
1246 	}
1247 
1248 	return ret;
1249 }
1250 
1251 static void srp_map_update_start(struct srp_map_state *state,
1252 				 struct scatterlist *sg, int sg_index,
1253 				 dma_addr_t dma_addr)
1254 {
1255 	state->unmapped_sg = sg;
1256 	state->unmapped_index = sg_index;
1257 	state->unmapped_addr = dma_addr;
1258 }
1259 
1260 static int srp_map_sg_entry(struct srp_map_state *state,
1261 			    struct srp_target_port *target,
1262 			    struct scatterlist *sg, int sg_index,
1263 			    bool use_mr)
1264 {
1265 	struct srp_device *dev = target->srp_host->srp_dev;
1266 	struct ib_device *ibdev = dev->dev;
1267 	dma_addr_t dma_addr = ib_sg_dma_address(ibdev, sg);
1268 	unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
1269 	unsigned int len;
1270 	int ret;
1271 
1272 	if (!dma_len)
1273 		return 0;
1274 
1275 	if (!use_mr) {
1276 		/*
1277 		 * Once we're in direct map mode for a request, we don't
1278 		 * go back to FMR or FR mode, so no need to update anything
1279 		 * other than the descriptor.
1280 		 */
1281 		srp_map_desc(state, dma_addr, dma_len, target->rkey);
1282 		return 0;
1283 	}
1284 
1285 	/*
1286 	 * Since not all RDMA HW drivers support non-zero page offsets for
1287 	 * FMR, if we start at an offset into a page, don't merge into the
1288 	 * current FMR mapping. Finish it out, and use the kernel's MR for
1289 	 * this sg entry.
1290 	 */
1291 	if ((!dev->use_fast_reg && dma_addr & ~dev->mr_page_mask) ||
1292 	    dma_len > dev->mr_max_size) {
1293 		ret = srp_finish_mapping(state, target);
1294 		if (ret)
1295 			return ret;
1296 
1297 		srp_map_desc(state, dma_addr, dma_len, target->rkey);
1298 		srp_map_update_start(state, NULL, 0, 0);
1299 		return 0;
1300 	}
1301 
1302 	/*
1303 	 * If this is the first sg that will be mapped via FMR or via FR, save
1304 	 * our position. We need to know the first unmapped entry, its index,
1305 	 * and the first unmapped address within that entry to be able to
1306 	 * restart mapping after an error.
1307 	 */
1308 	if (!state->unmapped_sg)
1309 		srp_map_update_start(state, sg, sg_index, dma_addr);
1310 
1311 	while (dma_len) {
1312 		unsigned offset = dma_addr & ~dev->mr_page_mask;
1313 		if (state->npages == dev->max_pages_per_mr || offset != 0) {
1314 			ret = srp_finish_mapping(state, target);
1315 			if (ret)
1316 				return ret;
1317 
1318 			srp_map_update_start(state, sg, sg_index, dma_addr);
1319 		}
1320 
1321 		len = min_t(unsigned int, dma_len, dev->mr_page_size - offset);
1322 
1323 		if (!state->npages)
1324 			state->base_dma_addr = dma_addr;
1325 		state->pages[state->npages++] = dma_addr & dev->mr_page_mask;
1326 		state->dma_len += len;
1327 		dma_addr += len;
1328 		dma_len -= len;
1329 	}
1330 
1331 	/*
1332 	 * If the last entry of the MR wasn't a full page, then we need to
1333 	 * close it out and start a new one -- we can only merge at page
1334 	 * boundries.
1335 	 */
1336 	ret = 0;
1337 	if (len != dev->mr_page_size) {
1338 		ret = srp_finish_mapping(state, target);
1339 		if (!ret)
1340 			srp_map_update_start(state, NULL, 0, 0);
1341 	}
1342 	return ret;
1343 }
1344 
1345 static int srp_map_sg(struct srp_map_state *state,
1346 		      struct srp_target_port *target, struct srp_request *req,
1347 		      struct scatterlist *scat, int count)
1348 {
1349 	struct srp_device *dev = target->srp_host->srp_dev;
1350 	struct ib_device *ibdev = dev->dev;
1351 	struct scatterlist *sg;
1352 	int i;
1353 	bool use_mr;
1354 
1355 	state->desc	= req->indirect_desc;
1356 	state->pages	= req->map_page;
1357 	if (dev->use_fast_reg) {
1358 		state->next_fr = req->fr_list;
1359 		use_mr = !!target->fr_pool;
1360 	} else {
1361 		state->next_fmr = req->fmr_list;
1362 		use_mr = !!target->fmr_pool;
1363 	}
1364 
1365 	for_each_sg(scat, sg, count, i) {
1366 		if (srp_map_sg_entry(state, target, sg, i, use_mr)) {
1367 			/*
1368 			 * Memory registration failed, so backtrack to the
1369 			 * first unmapped entry and continue on without using
1370 			 * memory registration.
1371 			 */
1372 			dma_addr_t dma_addr;
1373 			unsigned int dma_len;
1374 
1375 backtrack:
1376 			sg = state->unmapped_sg;
1377 			i = state->unmapped_index;
1378 
1379 			dma_addr = ib_sg_dma_address(ibdev, sg);
1380 			dma_len = ib_sg_dma_len(ibdev, sg);
1381 			dma_len -= (state->unmapped_addr - dma_addr);
1382 			dma_addr = state->unmapped_addr;
1383 			use_mr = false;
1384 			srp_map_desc(state, dma_addr, dma_len, target->rkey);
1385 		}
1386 	}
1387 
1388 	if (use_mr && srp_finish_mapping(state, target))
1389 		goto backtrack;
1390 
1391 	req->nmdesc = state->nmdesc;
1392 
1393 	return 0;
1394 }
1395 
1396 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_target_port *target,
1397 			struct srp_request *req)
1398 {
1399 	struct scatterlist *scat;
1400 	struct srp_cmd *cmd = req->cmd->buf;
1401 	int len, nents, count;
1402 	struct srp_device *dev;
1403 	struct ib_device *ibdev;
1404 	struct srp_map_state state;
1405 	struct srp_indirect_buf *indirect_hdr;
1406 	u32 table_len;
1407 	u8 fmt;
1408 
1409 	if (!scsi_sglist(scmnd) || scmnd->sc_data_direction == DMA_NONE)
1410 		return sizeof (struct srp_cmd);
1411 
1412 	if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
1413 	    scmnd->sc_data_direction != DMA_TO_DEVICE) {
1414 		shost_printk(KERN_WARNING, target->scsi_host,
1415 			     PFX "Unhandled data direction %d\n",
1416 			     scmnd->sc_data_direction);
1417 		return -EINVAL;
1418 	}
1419 
1420 	nents = scsi_sg_count(scmnd);
1421 	scat  = scsi_sglist(scmnd);
1422 
1423 	dev = target->srp_host->srp_dev;
1424 	ibdev = dev->dev;
1425 
1426 	count = ib_dma_map_sg(ibdev, scat, nents, scmnd->sc_data_direction);
1427 	if (unlikely(count == 0))
1428 		return -EIO;
1429 
1430 	fmt = SRP_DATA_DESC_DIRECT;
1431 	len = sizeof (struct srp_cmd) +	sizeof (struct srp_direct_buf);
1432 
1433 	if (count == 1 && !register_always) {
1434 		/*
1435 		 * The midlayer only generated a single gather/scatter
1436 		 * entry, or DMA mapping coalesced everything to a
1437 		 * single entry.  So a direct descriptor along with
1438 		 * the DMA MR suffices.
1439 		 */
1440 		struct srp_direct_buf *buf = (void *) cmd->add_data;
1441 
1442 		buf->va  = cpu_to_be64(ib_sg_dma_address(ibdev, scat));
1443 		buf->key = cpu_to_be32(target->rkey);
1444 		buf->len = cpu_to_be32(ib_sg_dma_len(ibdev, scat));
1445 
1446 		req->nmdesc = 0;
1447 		goto map_complete;
1448 	}
1449 
1450 	/*
1451 	 * We have more than one scatter/gather entry, so build our indirect
1452 	 * descriptor table, trying to merge as many entries as we can.
1453 	 */
1454 	indirect_hdr = (void *) cmd->add_data;
1455 
1456 	ib_dma_sync_single_for_cpu(ibdev, req->indirect_dma_addr,
1457 				   target->indirect_size, DMA_TO_DEVICE);
1458 
1459 	memset(&state, 0, sizeof(state));
1460 	srp_map_sg(&state, target, req, scat, count);
1461 
1462 	/* We've mapped the request, now pull as much of the indirect
1463 	 * descriptor table as we can into the command buffer. If this
1464 	 * target is not using an external indirect table, we are
1465 	 * guaranteed to fit into the command, as the SCSI layer won't
1466 	 * give us more S/G entries than we allow.
1467 	 */
1468 	if (state.ndesc == 1) {
1469 		/*
1470 		 * Memory registration collapsed the sg-list into one entry,
1471 		 * so use a direct descriptor.
1472 		 */
1473 		struct srp_direct_buf *buf = (void *) cmd->add_data;
1474 
1475 		*buf = req->indirect_desc[0];
1476 		goto map_complete;
1477 	}
1478 
1479 	if (unlikely(target->cmd_sg_cnt < state.ndesc &&
1480 						!target->allow_ext_sg)) {
1481 		shost_printk(KERN_ERR, target->scsi_host,
1482 			     "Could not fit S/G list into SRP_CMD\n");
1483 		return -EIO;
1484 	}
1485 
1486 	count = min(state.ndesc, target->cmd_sg_cnt);
1487 	table_len = state.ndesc * sizeof (struct srp_direct_buf);
1488 
1489 	fmt = SRP_DATA_DESC_INDIRECT;
1490 	len = sizeof(struct srp_cmd) + sizeof (struct srp_indirect_buf);
1491 	len += count * sizeof (struct srp_direct_buf);
1492 
1493 	memcpy(indirect_hdr->desc_list, req->indirect_desc,
1494 	       count * sizeof (struct srp_direct_buf));
1495 
1496 	indirect_hdr->table_desc.va = cpu_to_be64(req->indirect_dma_addr);
1497 	indirect_hdr->table_desc.key = cpu_to_be32(target->rkey);
1498 	indirect_hdr->table_desc.len = cpu_to_be32(table_len);
1499 	indirect_hdr->len = cpu_to_be32(state.total_len);
1500 
1501 	if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1502 		cmd->data_out_desc_cnt = count;
1503 	else
1504 		cmd->data_in_desc_cnt = count;
1505 
1506 	ib_dma_sync_single_for_device(ibdev, req->indirect_dma_addr, table_len,
1507 				      DMA_TO_DEVICE);
1508 
1509 map_complete:
1510 	if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1511 		cmd->buf_fmt = fmt << 4;
1512 	else
1513 		cmd->buf_fmt = fmt;
1514 
1515 	return len;
1516 }
1517 
1518 /*
1519  * Return an IU and possible credit to the free pool
1520  */
1521 static void srp_put_tx_iu(struct srp_target_port *target, struct srp_iu *iu,
1522 			  enum srp_iu_type iu_type)
1523 {
1524 	unsigned long flags;
1525 
1526 	spin_lock_irqsave(&target->lock, flags);
1527 	list_add(&iu->list, &target->free_tx);
1528 	if (iu_type != SRP_IU_RSP)
1529 		++target->req_lim;
1530 	spin_unlock_irqrestore(&target->lock, flags);
1531 }
1532 
1533 /*
1534  * Must be called with target->lock held to protect req_lim and free_tx.
1535  * If IU is not sent, it must be returned using srp_put_tx_iu().
1536  *
1537  * Note:
1538  * An upper limit for the number of allocated information units for each
1539  * request type is:
1540  * - SRP_IU_CMD: SRP_CMD_SQ_SIZE, since the SCSI mid-layer never queues
1541  *   more than Scsi_Host.can_queue requests.
1542  * - SRP_IU_TSK_MGMT: SRP_TSK_MGMT_SQ_SIZE.
1543  * - SRP_IU_RSP: 1, since a conforming SRP target never sends more than
1544  *   one unanswered SRP request to an initiator.
1545  */
1546 static struct srp_iu *__srp_get_tx_iu(struct srp_target_port *target,
1547 				      enum srp_iu_type iu_type)
1548 {
1549 	s32 rsv = (iu_type == SRP_IU_TSK_MGMT) ? 0 : SRP_TSK_MGMT_SQ_SIZE;
1550 	struct srp_iu *iu;
1551 
1552 	srp_send_completion(target->send_cq, target);
1553 
1554 	if (list_empty(&target->free_tx))
1555 		return NULL;
1556 
1557 	/* Initiator responses to target requests do not consume credits */
1558 	if (iu_type != SRP_IU_RSP) {
1559 		if (target->req_lim <= rsv) {
1560 			++target->zero_req_lim;
1561 			return NULL;
1562 		}
1563 
1564 		--target->req_lim;
1565 	}
1566 
1567 	iu = list_first_entry(&target->free_tx, struct srp_iu, list);
1568 	list_del(&iu->list);
1569 	return iu;
1570 }
1571 
1572 static int srp_post_send(struct srp_target_port *target,
1573 			 struct srp_iu *iu, int len)
1574 {
1575 	struct ib_sge list;
1576 	struct ib_send_wr wr, *bad_wr;
1577 
1578 	list.addr   = iu->dma;
1579 	list.length = len;
1580 	list.lkey   = target->lkey;
1581 
1582 	wr.next       = NULL;
1583 	wr.wr_id      = (uintptr_t) iu;
1584 	wr.sg_list    = &list;
1585 	wr.num_sge    = 1;
1586 	wr.opcode     = IB_WR_SEND;
1587 	wr.send_flags = IB_SEND_SIGNALED;
1588 
1589 	return ib_post_send(target->qp, &wr, &bad_wr);
1590 }
1591 
1592 static int srp_post_recv(struct srp_target_port *target, struct srp_iu *iu)
1593 {
1594 	struct ib_recv_wr wr, *bad_wr;
1595 	struct ib_sge list;
1596 
1597 	list.addr   = iu->dma;
1598 	list.length = iu->size;
1599 	list.lkey   = target->lkey;
1600 
1601 	wr.next     = NULL;
1602 	wr.wr_id    = (uintptr_t) iu;
1603 	wr.sg_list  = &list;
1604 	wr.num_sge  = 1;
1605 
1606 	return ib_post_recv(target->qp, &wr, &bad_wr);
1607 }
1608 
1609 static void srp_process_rsp(struct srp_target_port *target, struct srp_rsp *rsp)
1610 {
1611 	struct srp_request *req;
1612 	struct scsi_cmnd *scmnd;
1613 	unsigned long flags;
1614 
1615 	if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
1616 		spin_lock_irqsave(&target->lock, flags);
1617 		target->req_lim += be32_to_cpu(rsp->req_lim_delta);
1618 		spin_unlock_irqrestore(&target->lock, flags);
1619 
1620 		target->tsk_mgmt_status = -1;
1621 		if (be32_to_cpu(rsp->resp_data_len) >= 4)
1622 			target->tsk_mgmt_status = rsp->data[3];
1623 		complete(&target->tsk_mgmt_done);
1624 	} else {
1625 		req = &target->req_ring[rsp->tag];
1626 		scmnd = srp_claim_req(target, req, NULL, NULL);
1627 		if (!scmnd) {
1628 			shost_printk(KERN_ERR, target->scsi_host,
1629 				     "Null scmnd for RSP w/tag %016llx\n",
1630 				     (unsigned long long) rsp->tag);
1631 
1632 			spin_lock_irqsave(&target->lock, flags);
1633 			target->req_lim += be32_to_cpu(rsp->req_lim_delta);
1634 			spin_unlock_irqrestore(&target->lock, flags);
1635 
1636 			return;
1637 		}
1638 		scmnd->result = rsp->status;
1639 
1640 		if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
1641 			memcpy(scmnd->sense_buffer, rsp->data +
1642 			       be32_to_cpu(rsp->resp_data_len),
1643 			       min_t(int, be32_to_cpu(rsp->sense_data_len),
1644 				     SCSI_SENSE_BUFFERSIZE));
1645 		}
1646 
1647 		if (unlikely(rsp->flags & SRP_RSP_FLAG_DIUNDER))
1648 			scsi_set_resid(scmnd, be32_to_cpu(rsp->data_in_res_cnt));
1649 		else if (unlikely(rsp->flags & SRP_RSP_FLAG_DIOVER))
1650 			scsi_set_resid(scmnd, -be32_to_cpu(rsp->data_in_res_cnt));
1651 		else if (unlikely(rsp->flags & SRP_RSP_FLAG_DOUNDER))
1652 			scsi_set_resid(scmnd, be32_to_cpu(rsp->data_out_res_cnt));
1653 		else if (unlikely(rsp->flags & SRP_RSP_FLAG_DOOVER))
1654 			scsi_set_resid(scmnd, -be32_to_cpu(rsp->data_out_res_cnt));
1655 
1656 		srp_free_req(target, req, scmnd,
1657 			     be32_to_cpu(rsp->req_lim_delta));
1658 
1659 		scmnd->host_scribble = NULL;
1660 		scmnd->scsi_done(scmnd);
1661 	}
1662 }
1663 
1664 static int srp_response_common(struct srp_target_port *target, s32 req_delta,
1665 			       void *rsp, int len)
1666 {
1667 	struct ib_device *dev = target->srp_host->srp_dev->dev;
1668 	unsigned long flags;
1669 	struct srp_iu *iu;
1670 	int err;
1671 
1672 	spin_lock_irqsave(&target->lock, flags);
1673 	target->req_lim += req_delta;
1674 	iu = __srp_get_tx_iu(target, SRP_IU_RSP);
1675 	spin_unlock_irqrestore(&target->lock, flags);
1676 
1677 	if (!iu) {
1678 		shost_printk(KERN_ERR, target->scsi_host, PFX
1679 			     "no IU available to send response\n");
1680 		return 1;
1681 	}
1682 
1683 	ib_dma_sync_single_for_cpu(dev, iu->dma, len, DMA_TO_DEVICE);
1684 	memcpy(iu->buf, rsp, len);
1685 	ib_dma_sync_single_for_device(dev, iu->dma, len, DMA_TO_DEVICE);
1686 
1687 	err = srp_post_send(target, iu, len);
1688 	if (err) {
1689 		shost_printk(KERN_ERR, target->scsi_host, PFX
1690 			     "unable to post response: %d\n", err);
1691 		srp_put_tx_iu(target, iu, SRP_IU_RSP);
1692 	}
1693 
1694 	return err;
1695 }
1696 
1697 static void srp_process_cred_req(struct srp_target_port *target,
1698 				 struct srp_cred_req *req)
1699 {
1700 	struct srp_cred_rsp rsp = {
1701 		.opcode = SRP_CRED_RSP,
1702 		.tag = req->tag,
1703 	};
1704 	s32 delta = be32_to_cpu(req->req_lim_delta);
1705 
1706 	if (srp_response_common(target, delta, &rsp, sizeof rsp))
1707 		shost_printk(KERN_ERR, target->scsi_host, PFX
1708 			     "problems processing SRP_CRED_REQ\n");
1709 }
1710 
1711 static void srp_process_aer_req(struct srp_target_port *target,
1712 				struct srp_aer_req *req)
1713 {
1714 	struct srp_aer_rsp rsp = {
1715 		.opcode = SRP_AER_RSP,
1716 		.tag = req->tag,
1717 	};
1718 	s32 delta = be32_to_cpu(req->req_lim_delta);
1719 
1720 	shost_printk(KERN_ERR, target->scsi_host, PFX
1721 		     "ignoring AER for LUN %llu\n", be64_to_cpu(req->lun));
1722 
1723 	if (srp_response_common(target, delta, &rsp, sizeof rsp))
1724 		shost_printk(KERN_ERR, target->scsi_host, PFX
1725 			     "problems processing SRP_AER_REQ\n");
1726 }
1727 
1728 static void srp_handle_recv(struct srp_target_port *target, struct ib_wc *wc)
1729 {
1730 	struct ib_device *dev = target->srp_host->srp_dev->dev;
1731 	struct srp_iu *iu = (struct srp_iu *) (uintptr_t) wc->wr_id;
1732 	int res;
1733 	u8 opcode;
1734 
1735 	ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_ti_iu_len,
1736 				   DMA_FROM_DEVICE);
1737 
1738 	opcode = *(u8 *) iu->buf;
1739 
1740 	if (0) {
1741 		shost_printk(KERN_ERR, target->scsi_host,
1742 			     PFX "recv completion, opcode 0x%02x\n", opcode);
1743 		print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 8, 1,
1744 			       iu->buf, wc->byte_len, true);
1745 	}
1746 
1747 	switch (opcode) {
1748 	case SRP_RSP:
1749 		srp_process_rsp(target, iu->buf);
1750 		break;
1751 
1752 	case SRP_CRED_REQ:
1753 		srp_process_cred_req(target, iu->buf);
1754 		break;
1755 
1756 	case SRP_AER_REQ:
1757 		srp_process_aer_req(target, iu->buf);
1758 		break;
1759 
1760 	case SRP_T_LOGOUT:
1761 		/* XXX Handle target logout */
1762 		shost_printk(KERN_WARNING, target->scsi_host,
1763 			     PFX "Got target logout request\n");
1764 		break;
1765 
1766 	default:
1767 		shost_printk(KERN_WARNING, target->scsi_host,
1768 			     PFX "Unhandled SRP opcode 0x%02x\n", opcode);
1769 		break;
1770 	}
1771 
1772 	ib_dma_sync_single_for_device(dev, iu->dma, target->max_ti_iu_len,
1773 				      DMA_FROM_DEVICE);
1774 
1775 	res = srp_post_recv(target, iu);
1776 	if (res != 0)
1777 		shost_printk(KERN_ERR, target->scsi_host,
1778 			     PFX "Recv failed with error code %d\n", res);
1779 }
1780 
1781 /**
1782  * srp_tl_err_work() - handle a transport layer error
1783  * @work: Work structure embedded in an SRP target port.
1784  *
1785  * Note: This function may get invoked before the rport has been created,
1786  * hence the target->rport test.
1787  */
1788 static void srp_tl_err_work(struct work_struct *work)
1789 {
1790 	struct srp_target_port *target;
1791 
1792 	target = container_of(work, struct srp_target_port, tl_err_work);
1793 	if (target->rport)
1794 		srp_start_tl_fail_timers(target->rport);
1795 }
1796 
1797 static void srp_handle_qp_err(u64 wr_id, enum ib_wc_status wc_status,
1798 			      bool send_err, struct srp_target_port *target)
1799 {
1800 	if (target->connected && !target->qp_in_error) {
1801 		if (wr_id & LOCAL_INV_WR_ID_MASK) {
1802 			shost_printk(KERN_ERR, target->scsi_host, PFX
1803 				     "LOCAL_INV failed with status %d\n",
1804 				     wc_status);
1805 		} else if (wr_id & FAST_REG_WR_ID_MASK) {
1806 			shost_printk(KERN_ERR, target->scsi_host, PFX
1807 				     "FAST_REG_MR failed status %d\n",
1808 				     wc_status);
1809 		} else {
1810 			shost_printk(KERN_ERR, target->scsi_host,
1811 				     PFX "failed %s status %d for iu %p\n",
1812 				     send_err ? "send" : "receive",
1813 				     wc_status, (void *)(uintptr_t)wr_id);
1814 		}
1815 		queue_work(system_long_wq, &target->tl_err_work);
1816 	}
1817 	target->qp_in_error = true;
1818 }
1819 
1820 static void srp_recv_completion(struct ib_cq *cq, void *target_ptr)
1821 {
1822 	struct srp_target_port *target = target_ptr;
1823 	struct ib_wc wc;
1824 
1825 	ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1826 	while (ib_poll_cq(cq, 1, &wc) > 0) {
1827 		if (likely(wc.status == IB_WC_SUCCESS)) {
1828 			srp_handle_recv(target, &wc);
1829 		} else {
1830 			srp_handle_qp_err(wc.wr_id, wc.status, false, target);
1831 		}
1832 	}
1833 }
1834 
1835 static void srp_send_completion(struct ib_cq *cq, void *target_ptr)
1836 {
1837 	struct srp_target_port *target = target_ptr;
1838 	struct ib_wc wc;
1839 	struct srp_iu *iu;
1840 
1841 	while (ib_poll_cq(cq, 1, &wc) > 0) {
1842 		if (likely(wc.status == IB_WC_SUCCESS)) {
1843 			iu = (struct srp_iu *) (uintptr_t) wc.wr_id;
1844 			list_add(&iu->list, &target->free_tx);
1845 		} else {
1846 			srp_handle_qp_err(wc.wr_id, wc.status, true, target);
1847 		}
1848 	}
1849 }
1850 
1851 static int srp_queuecommand(struct Scsi_Host *shost, struct scsi_cmnd *scmnd)
1852 {
1853 	struct srp_target_port *target = host_to_target(shost);
1854 	struct srp_rport *rport = target->rport;
1855 	struct srp_request *req;
1856 	struct srp_iu *iu;
1857 	struct srp_cmd *cmd;
1858 	struct ib_device *dev;
1859 	unsigned long flags;
1860 	int len, ret;
1861 	const bool in_scsi_eh = !in_interrupt() && current == shost->ehandler;
1862 
1863 	/*
1864 	 * The SCSI EH thread is the only context from which srp_queuecommand()
1865 	 * can get invoked for blocked devices (SDEV_BLOCK /
1866 	 * SDEV_CREATED_BLOCK). Avoid racing with srp_reconnect_rport() by
1867 	 * locking the rport mutex if invoked from inside the SCSI EH.
1868 	 */
1869 	if (in_scsi_eh)
1870 		mutex_lock(&rport->mutex);
1871 
1872 	scmnd->result = srp_chkready(target->rport);
1873 	if (unlikely(scmnd->result))
1874 		goto err;
1875 
1876 	spin_lock_irqsave(&target->lock, flags);
1877 	iu = __srp_get_tx_iu(target, SRP_IU_CMD);
1878 	if (!iu)
1879 		goto err_unlock;
1880 
1881 	req = list_first_entry(&target->free_reqs, struct srp_request, list);
1882 	list_del(&req->list);
1883 	spin_unlock_irqrestore(&target->lock, flags);
1884 
1885 	dev = target->srp_host->srp_dev->dev;
1886 	ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_iu_len,
1887 				   DMA_TO_DEVICE);
1888 
1889 	scmnd->host_scribble = (void *) req;
1890 
1891 	cmd = iu->buf;
1892 	memset(cmd, 0, sizeof *cmd);
1893 
1894 	cmd->opcode = SRP_CMD;
1895 	cmd->lun    = cpu_to_be64((u64) scmnd->device->lun << 48);
1896 	cmd->tag    = req->index;
1897 	memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
1898 
1899 	req->scmnd    = scmnd;
1900 	req->cmd      = iu;
1901 
1902 	len = srp_map_data(scmnd, target, req);
1903 	if (len < 0) {
1904 		shost_printk(KERN_ERR, target->scsi_host,
1905 			     PFX "Failed to map data (%d)\n", len);
1906 		/*
1907 		 * If we ran out of memory descriptors (-ENOMEM) because an
1908 		 * application is queuing many requests with more than
1909 		 * max_pages_per_mr sg-list elements, tell the SCSI mid-layer
1910 		 * to reduce queue depth temporarily.
1911 		 */
1912 		scmnd->result = len == -ENOMEM ?
1913 			DID_OK << 16 | QUEUE_FULL << 1 : DID_ERROR << 16;
1914 		goto err_iu;
1915 	}
1916 
1917 	ib_dma_sync_single_for_device(dev, iu->dma, target->max_iu_len,
1918 				      DMA_TO_DEVICE);
1919 
1920 	if (srp_post_send(target, iu, len)) {
1921 		shost_printk(KERN_ERR, target->scsi_host, PFX "Send failed\n");
1922 		goto err_unmap;
1923 	}
1924 
1925 	ret = 0;
1926 
1927 unlock_rport:
1928 	if (in_scsi_eh)
1929 		mutex_unlock(&rport->mutex);
1930 
1931 	return ret;
1932 
1933 err_unmap:
1934 	srp_unmap_data(scmnd, target, req);
1935 
1936 err_iu:
1937 	srp_put_tx_iu(target, iu, SRP_IU_CMD);
1938 
1939 	/*
1940 	 * Avoid that the loops that iterate over the request ring can
1941 	 * encounter a dangling SCSI command pointer.
1942 	 */
1943 	req->scmnd = NULL;
1944 
1945 	spin_lock_irqsave(&target->lock, flags);
1946 	list_add(&req->list, &target->free_reqs);
1947 
1948 err_unlock:
1949 	spin_unlock_irqrestore(&target->lock, flags);
1950 
1951 err:
1952 	if (scmnd->result) {
1953 		scmnd->scsi_done(scmnd);
1954 		ret = 0;
1955 	} else {
1956 		ret = SCSI_MLQUEUE_HOST_BUSY;
1957 	}
1958 
1959 	goto unlock_rport;
1960 }
1961 
1962 /*
1963  * Note: the resources allocated in this function are freed in
1964  * srp_free_target_ib().
1965  */
1966 static int srp_alloc_iu_bufs(struct srp_target_port *target)
1967 {
1968 	int i;
1969 
1970 	target->rx_ring = kzalloc(target->queue_size * sizeof(*target->rx_ring),
1971 				  GFP_KERNEL);
1972 	if (!target->rx_ring)
1973 		goto err_no_ring;
1974 	target->tx_ring = kzalloc(target->queue_size * sizeof(*target->tx_ring),
1975 				  GFP_KERNEL);
1976 	if (!target->tx_ring)
1977 		goto err_no_ring;
1978 
1979 	for (i = 0; i < target->queue_size; ++i) {
1980 		target->rx_ring[i] = srp_alloc_iu(target->srp_host,
1981 						  target->max_ti_iu_len,
1982 						  GFP_KERNEL, DMA_FROM_DEVICE);
1983 		if (!target->rx_ring[i])
1984 			goto err;
1985 	}
1986 
1987 	for (i = 0; i < target->queue_size; ++i) {
1988 		target->tx_ring[i] = srp_alloc_iu(target->srp_host,
1989 						  target->max_iu_len,
1990 						  GFP_KERNEL, DMA_TO_DEVICE);
1991 		if (!target->tx_ring[i])
1992 			goto err;
1993 
1994 		list_add(&target->tx_ring[i]->list, &target->free_tx);
1995 	}
1996 
1997 	return 0;
1998 
1999 err:
2000 	for (i = 0; i < target->queue_size; ++i) {
2001 		srp_free_iu(target->srp_host, target->rx_ring[i]);
2002 		srp_free_iu(target->srp_host, target->tx_ring[i]);
2003 	}
2004 
2005 
2006 err_no_ring:
2007 	kfree(target->tx_ring);
2008 	target->tx_ring = NULL;
2009 	kfree(target->rx_ring);
2010 	target->rx_ring = NULL;
2011 
2012 	return -ENOMEM;
2013 }
2014 
2015 static uint32_t srp_compute_rq_tmo(struct ib_qp_attr *qp_attr, int attr_mask)
2016 {
2017 	uint64_t T_tr_ns, max_compl_time_ms;
2018 	uint32_t rq_tmo_jiffies;
2019 
2020 	/*
2021 	 * According to section 11.2.4.2 in the IBTA spec (Modify Queue Pair,
2022 	 * table 91), both the QP timeout and the retry count have to be set
2023 	 * for RC QP's during the RTR to RTS transition.
2024 	 */
2025 	WARN_ON_ONCE((attr_mask & (IB_QP_TIMEOUT | IB_QP_RETRY_CNT)) !=
2026 		     (IB_QP_TIMEOUT | IB_QP_RETRY_CNT));
2027 
2028 	/*
2029 	 * Set target->rq_tmo_jiffies to one second more than the largest time
2030 	 * it can take before an error completion is generated. See also
2031 	 * C9-140..142 in the IBTA spec for more information about how to
2032 	 * convert the QP Local ACK Timeout value to nanoseconds.
2033 	 */
2034 	T_tr_ns = 4096 * (1ULL << qp_attr->timeout);
2035 	max_compl_time_ms = qp_attr->retry_cnt * 4 * T_tr_ns;
2036 	do_div(max_compl_time_ms, NSEC_PER_MSEC);
2037 	rq_tmo_jiffies = msecs_to_jiffies(max_compl_time_ms + 1000);
2038 
2039 	return rq_tmo_jiffies;
2040 }
2041 
2042 static void srp_cm_rep_handler(struct ib_cm_id *cm_id,
2043 			       struct srp_login_rsp *lrsp,
2044 			       struct srp_target_port *target)
2045 {
2046 	struct ib_qp_attr *qp_attr = NULL;
2047 	int attr_mask = 0;
2048 	int ret;
2049 	int i;
2050 
2051 	if (lrsp->opcode == SRP_LOGIN_RSP) {
2052 		target->max_ti_iu_len = be32_to_cpu(lrsp->max_ti_iu_len);
2053 		target->req_lim       = be32_to_cpu(lrsp->req_lim_delta);
2054 
2055 		/*
2056 		 * Reserve credits for task management so we don't
2057 		 * bounce requests back to the SCSI mid-layer.
2058 		 */
2059 		target->scsi_host->can_queue
2060 			= min(target->req_lim - SRP_TSK_MGMT_SQ_SIZE,
2061 			      target->scsi_host->can_queue);
2062 		target->scsi_host->cmd_per_lun
2063 			= min_t(int, target->scsi_host->can_queue,
2064 				target->scsi_host->cmd_per_lun);
2065 	} else {
2066 		shost_printk(KERN_WARNING, target->scsi_host,
2067 			     PFX "Unhandled RSP opcode %#x\n", lrsp->opcode);
2068 		ret = -ECONNRESET;
2069 		goto error;
2070 	}
2071 
2072 	if (!target->rx_ring) {
2073 		ret = srp_alloc_iu_bufs(target);
2074 		if (ret)
2075 			goto error;
2076 	}
2077 
2078 	ret = -ENOMEM;
2079 	qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
2080 	if (!qp_attr)
2081 		goto error;
2082 
2083 	qp_attr->qp_state = IB_QPS_RTR;
2084 	ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
2085 	if (ret)
2086 		goto error_free;
2087 
2088 	ret = ib_modify_qp(target->qp, qp_attr, attr_mask);
2089 	if (ret)
2090 		goto error_free;
2091 
2092 	for (i = 0; i < target->queue_size; i++) {
2093 		struct srp_iu *iu = target->rx_ring[i];
2094 		ret = srp_post_recv(target, iu);
2095 		if (ret)
2096 			goto error_free;
2097 	}
2098 
2099 	qp_attr->qp_state = IB_QPS_RTS;
2100 	ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
2101 	if (ret)
2102 		goto error_free;
2103 
2104 	target->rq_tmo_jiffies = srp_compute_rq_tmo(qp_attr, attr_mask);
2105 
2106 	ret = ib_modify_qp(target->qp, qp_attr, attr_mask);
2107 	if (ret)
2108 		goto error_free;
2109 
2110 	ret = ib_send_cm_rtu(cm_id, NULL, 0);
2111 
2112 error_free:
2113 	kfree(qp_attr);
2114 
2115 error:
2116 	target->status = ret;
2117 }
2118 
2119 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
2120 			       struct ib_cm_event *event,
2121 			       struct srp_target_port *target)
2122 {
2123 	struct Scsi_Host *shost = target->scsi_host;
2124 	struct ib_class_port_info *cpi;
2125 	int opcode;
2126 
2127 	switch (event->param.rej_rcvd.reason) {
2128 	case IB_CM_REJ_PORT_CM_REDIRECT:
2129 		cpi = event->param.rej_rcvd.ari;
2130 		target->path.dlid = cpi->redirect_lid;
2131 		target->path.pkey = cpi->redirect_pkey;
2132 		cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
2133 		memcpy(target->path.dgid.raw, cpi->redirect_gid, 16);
2134 
2135 		target->status = target->path.dlid ?
2136 			SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
2137 		break;
2138 
2139 	case IB_CM_REJ_PORT_REDIRECT:
2140 		if (srp_target_is_topspin(target)) {
2141 			/*
2142 			 * Topspin/Cisco SRP gateways incorrectly send
2143 			 * reject reason code 25 when they mean 24
2144 			 * (port redirect).
2145 			 */
2146 			memcpy(target->path.dgid.raw,
2147 			       event->param.rej_rcvd.ari, 16);
2148 
2149 			shost_printk(KERN_DEBUG, shost,
2150 				     PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
2151 				     (unsigned long long) be64_to_cpu(target->path.dgid.global.subnet_prefix),
2152 				     (unsigned long long) be64_to_cpu(target->path.dgid.global.interface_id));
2153 
2154 			target->status = SRP_PORT_REDIRECT;
2155 		} else {
2156 			shost_printk(KERN_WARNING, shost,
2157 				     "  REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
2158 			target->status = -ECONNRESET;
2159 		}
2160 		break;
2161 
2162 	case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
2163 		shost_printk(KERN_WARNING, shost,
2164 			    "  REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
2165 		target->status = -ECONNRESET;
2166 		break;
2167 
2168 	case IB_CM_REJ_CONSUMER_DEFINED:
2169 		opcode = *(u8 *) event->private_data;
2170 		if (opcode == SRP_LOGIN_REJ) {
2171 			struct srp_login_rej *rej = event->private_data;
2172 			u32 reason = be32_to_cpu(rej->reason);
2173 
2174 			if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
2175 				shost_printk(KERN_WARNING, shost,
2176 					     PFX "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
2177 			else
2178 				shost_printk(KERN_WARNING, shost, PFX
2179 					     "SRP LOGIN from %pI6 to %pI6 REJECTED, reason 0x%08x\n",
2180 					     target->path.sgid.raw,
2181 					     target->orig_dgid, reason);
2182 		} else
2183 			shost_printk(KERN_WARNING, shost,
2184 				     "  REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
2185 				     " opcode 0x%02x\n", opcode);
2186 		target->status = -ECONNRESET;
2187 		break;
2188 
2189 	case IB_CM_REJ_STALE_CONN:
2190 		shost_printk(KERN_WARNING, shost, "  REJ reason: stale connection\n");
2191 		target->status = SRP_STALE_CONN;
2192 		break;
2193 
2194 	default:
2195 		shost_printk(KERN_WARNING, shost, "  REJ reason 0x%x\n",
2196 			     event->param.rej_rcvd.reason);
2197 		target->status = -ECONNRESET;
2198 	}
2199 }
2200 
2201 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
2202 {
2203 	struct srp_target_port *target = cm_id->context;
2204 	int comp = 0;
2205 
2206 	switch (event->event) {
2207 	case IB_CM_REQ_ERROR:
2208 		shost_printk(KERN_DEBUG, target->scsi_host,
2209 			     PFX "Sending CM REQ failed\n");
2210 		comp = 1;
2211 		target->status = -ECONNRESET;
2212 		break;
2213 
2214 	case IB_CM_REP_RECEIVED:
2215 		comp = 1;
2216 		srp_cm_rep_handler(cm_id, event->private_data, target);
2217 		break;
2218 
2219 	case IB_CM_REJ_RECEIVED:
2220 		shost_printk(KERN_DEBUG, target->scsi_host, PFX "REJ received\n");
2221 		comp = 1;
2222 
2223 		srp_cm_rej_handler(cm_id, event, target);
2224 		break;
2225 
2226 	case IB_CM_DREQ_RECEIVED:
2227 		shost_printk(KERN_WARNING, target->scsi_host,
2228 			     PFX "DREQ received - connection closed\n");
2229 		srp_change_conn_state(target, false);
2230 		if (ib_send_cm_drep(cm_id, NULL, 0))
2231 			shost_printk(KERN_ERR, target->scsi_host,
2232 				     PFX "Sending CM DREP failed\n");
2233 		queue_work(system_long_wq, &target->tl_err_work);
2234 		break;
2235 
2236 	case IB_CM_TIMEWAIT_EXIT:
2237 		shost_printk(KERN_ERR, target->scsi_host,
2238 			     PFX "connection closed\n");
2239 		comp = 1;
2240 
2241 		target->status = 0;
2242 		break;
2243 
2244 	case IB_CM_MRA_RECEIVED:
2245 	case IB_CM_DREQ_ERROR:
2246 	case IB_CM_DREP_RECEIVED:
2247 		break;
2248 
2249 	default:
2250 		shost_printk(KERN_WARNING, target->scsi_host,
2251 			     PFX "Unhandled CM event %d\n", event->event);
2252 		break;
2253 	}
2254 
2255 	if (comp)
2256 		complete(&target->done);
2257 
2258 	return 0;
2259 }
2260 
2261 /**
2262  * srp_change_queue_type - changing device queue tag type
2263  * @sdev: scsi device struct
2264  * @tag_type: requested tag type
2265  *
2266  * Returns queue tag type.
2267  */
2268 static int
2269 srp_change_queue_type(struct scsi_device *sdev, int tag_type)
2270 {
2271 	if (sdev->tagged_supported) {
2272 		scsi_set_tag_type(sdev, tag_type);
2273 		if (tag_type)
2274 			scsi_activate_tcq(sdev, sdev->queue_depth);
2275 		else
2276 			scsi_deactivate_tcq(sdev, sdev->queue_depth);
2277 	} else
2278 		tag_type = 0;
2279 
2280 	return tag_type;
2281 }
2282 
2283 /**
2284  * srp_change_queue_depth - setting device queue depth
2285  * @sdev: scsi device struct
2286  * @qdepth: requested queue depth
2287  * @reason: SCSI_QDEPTH_DEFAULT/SCSI_QDEPTH_QFULL/SCSI_QDEPTH_RAMP_UP
2288  * (see include/scsi/scsi_host.h for definition)
2289  *
2290  * Returns queue depth.
2291  */
2292 static int
2293 srp_change_queue_depth(struct scsi_device *sdev, int qdepth, int reason)
2294 {
2295 	struct Scsi_Host *shost = sdev->host;
2296 	int max_depth;
2297 	if (reason == SCSI_QDEPTH_DEFAULT || reason == SCSI_QDEPTH_RAMP_UP) {
2298 		max_depth = shost->can_queue;
2299 		if (!sdev->tagged_supported)
2300 			max_depth = 1;
2301 		if (qdepth > max_depth)
2302 			qdepth = max_depth;
2303 		scsi_adjust_queue_depth(sdev, scsi_get_tag_type(sdev), qdepth);
2304 	} else if (reason == SCSI_QDEPTH_QFULL)
2305 		scsi_track_queue_full(sdev, qdepth);
2306 	else
2307 		return -EOPNOTSUPP;
2308 
2309 	return sdev->queue_depth;
2310 }
2311 
2312 static int srp_send_tsk_mgmt(struct srp_target_port *target,
2313 			     u64 req_tag, unsigned int lun, u8 func)
2314 {
2315 	struct srp_rport *rport = target->rport;
2316 	struct ib_device *dev = target->srp_host->srp_dev->dev;
2317 	struct srp_iu *iu;
2318 	struct srp_tsk_mgmt *tsk_mgmt;
2319 
2320 	if (!target->connected || target->qp_in_error)
2321 		return -1;
2322 
2323 	init_completion(&target->tsk_mgmt_done);
2324 
2325 	/*
2326 	 * Lock the rport mutex to avoid that srp_create_target_ib() is
2327 	 * invoked while a task management function is being sent.
2328 	 */
2329 	mutex_lock(&rport->mutex);
2330 	spin_lock_irq(&target->lock);
2331 	iu = __srp_get_tx_iu(target, SRP_IU_TSK_MGMT);
2332 	spin_unlock_irq(&target->lock);
2333 
2334 	if (!iu) {
2335 		mutex_unlock(&rport->mutex);
2336 
2337 		return -1;
2338 	}
2339 
2340 	ib_dma_sync_single_for_cpu(dev, iu->dma, sizeof *tsk_mgmt,
2341 				   DMA_TO_DEVICE);
2342 	tsk_mgmt = iu->buf;
2343 	memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
2344 
2345 	tsk_mgmt->opcode 	= SRP_TSK_MGMT;
2346 	tsk_mgmt->lun		= cpu_to_be64((u64) lun << 48);
2347 	tsk_mgmt->tag		= req_tag | SRP_TAG_TSK_MGMT;
2348 	tsk_mgmt->tsk_mgmt_func = func;
2349 	tsk_mgmt->task_tag	= req_tag;
2350 
2351 	ib_dma_sync_single_for_device(dev, iu->dma, sizeof *tsk_mgmt,
2352 				      DMA_TO_DEVICE);
2353 	if (srp_post_send(target, iu, sizeof *tsk_mgmt)) {
2354 		srp_put_tx_iu(target, iu, SRP_IU_TSK_MGMT);
2355 		mutex_unlock(&rport->mutex);
2356 
2357 		return -1;
2358 	}
2359 	mutex_unlock(&rport->mutex);
2360 
2361 	if (!wait_for_completion_timeout(&target->tsk_mgmt_done,
2362 					 msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
2363 		return -1;
2364 
2365 	return 0;
2366 }
2367 
2368 static int srp_abort(struct scsi_cmnd *scmnd)
2369 {
2370 	struct srp_target_port *target = host_to_target(scmnd->device->host);
2371 	struct srp_request *req = (struct srp_request *) scmnd->host_scribble;
2372 	int ret;
2373 
2374 	shost_printk(KERN_ERR, target->scsi_host, "SRP abort called\n");
2375 
2376 	if (!req || !srp_claim_req(target, req, NULL, scmnd))
2377 		return SUCCESS;
2378 	if (srp_send_tsk_mgmt(target, req->index, scmnd->device->lun,
2379 			      SRP_TSK_ABORT_TASK) == 0)
2380 		ret = SUCCESS;
2381 	else if (target->rport->state == SRP_RPORT_LOST)
2382 		ret = FAST_IO_FAIL;
2383 	else
2384 		ret = FAILED;
2385 	srp_free_req(target, req, scmnd, 0);
2386 	scmnd->result = DID_ABORT << 16;
2387 	scmnd->scsi_done(scmnd);
2388 
2389 	return ret;
2390 }
2391 
2392 static int srp_reset_device(struct scsi_cmnd *scmnd)
2393 {
2394 	struct srp_target_port *target = host_to_target(scmnd->device->host);
2395 	int i;
2396 
2397 	shost_printk(KERN_ERR, target->scsi_host, "SRP reset_device called\n");
2398 
2399 	if (srp_send_tsk_mgmt(target, SRP_TAG_NO_REQ, scmnd->device->lun,
2400 			      SRP_TSK_LUN_RESET))
2401 		return FAILED;
2402 	if (target->tsk_mgmt_status)
2403 		return FAILED;
2404 
2405 	for (i = 0; i < target->req_ring_size; ++i) {
2406 		struct srp_request *req = &target->req_ring[i];
2407 		srp_finish_req(target, req, scmnd->device, DID_RESET << 16);
2408 	}
2409 
2410 	return SUCCESS;
2411 }
2412 
2413 static int srp_reset_host(struct scsi_cmnd *scmnd)
2414 {
2415 	struct srp_target_port *target = host_to_target(scmnd->device->host);
2416 
2417 	shost_printk(KERN_ERR, target->scsi_host, PFX "SRP reset_host called\n");
2418 
2419 	return srp_reconnect_rport(target->rport) == 0 ? SUCCESS : FAILED;
2420 }
2421 
2422 static int srp_slave_configure(struct scsi_device *sdev)
2423 {
2424 	struct Scsi_Host *shost = sdev->host;
2425 	struct srp_target_port *target = host_to_target(shost);
2426 	struct request_queue *q = sdev->request_queue;
2427 	unsigned long timeout;
2428 
2429 	if (sdev->type == TYPE_DISK) {
2430 		timeout = max_t(unsigned, 30 * HZ, target->rq_tmo_jiffies);
2431 		blk_queue_rq_timeout(q, timeout);
2432 	}
2433 
2434 	return 0;
2435 }
2436 
2437 static ssize_t show_id_ext(struct device *dev, struct device_attribute *attr,
2438 			   char *buf)
2439 {
2440 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2441 
2442 	return sprintf(buf, "0x%016llx\n",
2443 		       (unsigned long long) be64_to_cpu(target->id_ext));
2444 }
2445 
2446 static ssize_t show_ioc_guid(struct device *dev, struct device_attribute *attr,
2447 			     char *buf)
2448 {
2449 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2450 
2451 	return sprintf(buf, "0x%016llx\n",
2452 		       (unsigned long long) be64_to_cpu(target->ioc_guid));
2453 }
2454 
2455 static ssize_t show_service_id(struct device *dev,
2456 			       struct device_attribute *attr, char *buf)
2457 {
2458 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2459 
2460 	return sprintf(buf, "0x%016llx\n",
2461 		       (unsigned long long) be64_to_cpu(target->service_id));
2462 }
2463 
2464 static ssize_t show_pkey(struct device *dev, struct device_attribute *attr,
2465 			 char *buf)
2466 {
2467 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2468 
2469 	return sprintf(buf, "0x%04x\n", be16_to_cpu(target->path.pkey));
2470 }
2471 
2472 static ssize_t show_sgid(struct device *dev, struct device_attribute *attr,
2473 			 char *buf)
2474 {
2475 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2476 
2477 	return sprintf(buf, "%pI6\n", target->path.sgid.raw);
2478 }
2479 
2480 static ssize_t show_dgid(struct device *dev, struct device_attribute *attr,
2481 			 char *buf)
2482 {
2483 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2484 
2485 	return sprintf(buf, "%pI6\n", target->path.dgid.raw);
2486 }
2487 
2488 static ssize_t show_orig_dgid(struct device *dev,
2489 			      struct device_attribute *attr, char *buf)
2490 {
2491 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2492 
2493 	return sprintf(buf, "%pI6\n", target->orig_dgid);
2494 }
2495 
2496 static ssize_t show_req_lim(struct device *dev,
2497 			    struct device_attribute *attr, char *buf)
2498 {
2499 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2500 
2501 	return sprintf(buf, "%d\n", target->req_lim);
2502 }
2503 
2504 static ssize_t show_zero_req_lim(struct device *dev,
2505 				 struct device_attribute *attr, char *buf)
2506 {
2507 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2508 
2509 	return sprintf(buf, "%d\n", target->zero_req_lim);
2510 }
2511 
2512 static ssize_t show_local_ib_port(struct device *dev,
2513 				  struct device_attribute *attr, char *buf)
2514 {
2515 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2516 
2517 	return sprintf(buf, "%d\n", target->srp_host->port);
2518 }
2519 
2520 static ssize_t show_local_ib_device(struct device *dev,
2521 				    struct device_attribute *attr, char *buf)
2522 {
2523 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2524 
2525 	return sprintf(buf, "%s\n", target->srp_host->srp_dev->dev->name);
2526 }
2527 
2528 static ssize_t show_comp_vector(struct device *dev,
2529 				struct device_attribute *attr, char *buf)
2530 {
2531 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2532 
2533 	return sprintf(buf, "%d\n", target->comp_vector);
2534 }
2535 
2536 static ssize_t show_tl_retry_count(struct device *dev,
2537 				   struct device_attribute *attr, char *buf)
2538 {
2539 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2540 
2541 	return sprintf(buf, "%d\n", target->tl_retry_count);
2542 }
2543 
2544 static ssize_t show_cmd_sg_entries(struct device *dev,
2545 				   struct device_attribute *attr, char *buf)
2546 {
2547 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2548 
2549 	return sprintf(buf, "%u\n", target->cmd_sg_cnt);
2550 }
2551 
2552 static ssize_t show_allow_ext_sg(struct device *dev,
2553 				 struct device_attribute *attr, char *buf)
2554 {
2555 	struct srp_target_port *target = host_to_target(class_to_shost(dev));
2556 
2557 	return sprintf(buf, "%s\n", target->allow_ext_sg ? "true" : "false");
2558 }
2559 
2560 static DEVICE_ATTR(id_ext,	    S_IRUGO, show_id_ext,	   NULL);
2561 static DEVICE_ATTR(ioc_guid,	    S_IRUGO, show_ioc_guid,	   NULL);
2562 static DEVICE_ATTR(service_id,	    S_IRUGO, show_service_id,	   NULL);
2563 static DEVICE_ATTR(pkey,	    S_IRUGO, show_pkey,		   NULL);
2564 static DEVICE_ATTR(sgid,	    S_IRUGO, show_sgid,		   NULL);
2565 static DEVICE_ATTR(dgid,	    S_IRUGO, show_dgid,		   NULL);
2566 static DEVICE_ATTR(orig_dgid,	    S_IRUGO, show_orig_dgid,	   NULL);
2567 static DEVICE_ATTR(req_lim,         S_IRUGO, show_req_lim,         NULL);
2568 static DEVICE_ATTR(zero_req_lim,    S_IRUGO, show_zero_req_lim,	   NULL);
2569 static DEVICE_ATTR(local_ib_port,   S_IRUGO, show_local_ib_port,   NULL);
2570 static DEVICE_ATTR(local_ib_device, S_IRUGO, show_local_ib_device, NULL);
2571 static DEVICE_ATTR(comp_vector,     S_IRUGO, show_comp_vector,     NULL);
2572 static DEVICE_ATTR(tl_retry_count,  S_IRUGO, show_tl_retry_count,  NULL);
2573 static DEVICE_ATTR(cmd_sg_entries,  S_IRUGO, show_cmd_sg_entries,  NULL);
2574 static DEVICE_ATTR(allow_ext_sg,    S_IRUGO, show_allow_ext_sg,    NULL);
2575 
2576 static struct device_attribute *srp_host_attrs[] = {
2577 	&dev_attr_id_ext,
2578 	&dev_attr_ioc_guid,
2579 	&dev_attr_service_id,
2580 	&dev_attr_pkey,
2581 	&dev_attr_sgid,
2582 	&dev_attr_dgid,
2583 	&dev_attr_orig_dgid,
2584 	&dev_attr_req_lim,
2585 	&dev_attr_zero_req_lim,
2586 	&dev_attr_local_ib_port,
2587 	&dev_attr_local_ib_device,
2588 	&dev_attr_comp_vector,
2589 	&dev_attr_tl_retry_count,
2590 	&dev_attr_cmd_sg_entries,
2591 	&dev_attr_allow_ext_sg,
2592 	NULL
2593 };
2594 
2595 static struct scsi_host_template srp_template = {
2596 	.module				= THIS_MODULE,
2597 	.name				= "InfiniBand SRP initiator",
2598 	.proc_name			= DRV_NAME,
2599 	.slave_configure		= srp_slave_configure,
2600 	.info				= srp_target_info,
2601 	.queuecommand			= srp_queuecommand,
2602 	.change_queue_depth             = srp_change_queue_depth,
2603 	.change_queue_type              = srp_change_queue_type,
2604 	.eh_abort_handler		= srp_abort,
2605 	.eh_device_reset_handler	= srp_reset_device,
2606 	.eh_host_reset_handler		= srp_reset_host,
2607 	.skip_settle_delay		= true,
2608 	.sg_tablesize			= SRP_DEF_SG_TABLESIZE,
2609 	.can_queue			= SRP_DEFAULT_CMD_SQ_SIZE,
2610 	.this_id			= -1,
2611 	.cmd_per_lun			= SRP_DEFAULT_CMD_SQ_SIZE,
2612 	.use_clustering			= ENABLE_CLUSTERING,
2613 	.shost_attrs			= srp_host_attrs
2614 };
2615 
2616 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
2617 {
2618 	struct srp_rport_identifiers ids;
2619 	struct srp_rport *rport;
2620 
2621 	sprintf(target->target_name, "SRP.T10:%016llX",
2622 		 (unsigned long long) be64_to_cpu(target->id_ext));
2623 
2624 	if (scsi_add_host(target->scsi_host, host->srp_dev->dev->dma_device))
2625 		return -ENODEV;
2626 
2627 	memcpy(ids.port_id, &target->id_ext, 8);
2628 	memcpy(ids.port_id + 8, &target->ioc_guid, 8);
2629 	ids.roles = SRP_RPORT_ROLE_TARGET;
2630 	rport = srp_rport_add(target->scsi_host, &ids);
2631 	if (IS_ERR(rport)) {
2632 		scsi_remove_host(target->scsi_host);
2633 		return PTR_ERR(rport);
2634 	}
2635 
2636 	rport->lld_data = target;
2637 	target->rport = rport;
2638 
2639 	spin_lock(&host->target_lock);
2640 	list_add_tail(&target->list, &host->target_list);
2641 	spin_unlock(&host->target_lock);
2642 
2643 	target->state = SRP_TARGET_LIVE;
2644 
2645 	scsi_scan_target(&target->scsi_host->shost_gendev,
2646 			 0, target->scsi_id, SCAN_WILD_CARD, 0);
2647 
2648 	return 0;
2649 }
2650 
2651 static void srp_release_dev(struct device *dev)
2652 {
2653 	struct srp_host *host =
2654 		container_of(dev, struct srp_host, dev);
2655 
2656 	complete(&host->released);
2657 }
2658 
2659 static struct class srp_class = {
2660 	.name    = "infiniband_srp",
2661 	.dev_release = srp_release_dev
2662 };
2663 
2664 /**
2665  * srp_conn_unique() - check whether the connection to a target is unique
2666  * @host:   SRP host.
2667  * @target: SRP target port.
2668  */
2669 static bool srp_conn_unique(struct srp_host *host,
2670 			    struct srp_target_port *target)
2671 {
2672 	struct srp_target_port *t;
2673 	bool ret = false;
2674 
2675 	if (target->state == SRP_TARGET_REMOVED)
2676 		goto out;
2677 
2678 	ret = true;
2679 
2680 	spin_lock(&host->target_lock);
2681 	list_for_each_entry(t, &host->target_list, list) {
2682 		if (t != target &&
2683 		    target->id_ext == t->id_ext &&
2684 		    target->ioc_guid == t->ioc_guid &&
2685 		    target->initiator_ext == t->initiator_ext) {
2686 			ret = false;
2687 			break;
2688 		}
2689 	}
2690 	spin_unlock(&host->target_lock);
2691 
2692 out:
2693 	return ret;
2694 }
2695 
2696 /*
2697  * Target ports are added by writing
2698  *
2699  *     id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
2700  *     pkey=<P_Key>,service_id=<service ID>
2701  *
2702  * to the add_target sysfs attribute.
2703  */
2704 enum {
2705 	SRP_OPT_ERR		= 0,
2706 	SRP_OPT_ID_EXT		= 1 << 0,
2707 	SRP_OPT_IOC_GUID	= 1 << 1,
2708 	SRP_OPT_DGID		= 1 << 2,
2709 	SRP_OPT_PKEY		= 1 << 3,
2710 	SRP_OPT_SERVICE_ID	= 1 << 4,
2711 	SRP_OPT_MAX_SECT	= 1 << 5,
2712 	SRP_OPT_MAX_CMD_PER_LUN	= 1 << 6,
2713 	SRP_OPT_IO_CLASS	= 1 << 7,
2714 	SRP_OPT_INITIATOR_EXT	= 1 << 8,
2715 	SRP_OPT_CMD_SG_ENTRIES	= 1 << 9,
2716 	SRP_OPT_ALLOW_EXT_SG	= 1 << 10,
2717 	SRP_OPT_SG_TABLESIZE	= 1 << 11,
2718 	SRP_OPT_COMP_VECTOR	= 1 << 12,
2719 	SRP_OPT_TL_RETRY_COUNT	= 1 << 13,
2720 	SRP_OPT_QUEUE_SIZE	= 1 << 14,
2721 	SRP_OPT_ALL		= (SRP_OPT_ID_EXT	|
2722 				   SRP_OPT_IOC_GUID	|
2723 				   SRP_OPT_DGID		|
2724 				   SRP_OPT_PKEY		|
2725 				   SRP_OPT_SERVICE_ID),
2726 };
2727 
2728 static const match_table_t srp_opt_tokens = {
2729 	{ SRP_OPT_ID_EXT,		"id_ext=%s" 		},
2730 	{ SRP_OPT_IOC_GUID,		"ioc_guid=%s" 		},
2731 	{ SRP_OPT_DGID,			"dgid=%s" 		},
2732 	{ SRP_OPT_PKEY,			"pkey=%x" 		},
2733 	{ SRP_OPT_SERVICE_ID,		"service_id=%s"		},
2734 	{ SRP_OPT_MAX_SECT,		"max_sect=%d" 		},
2735 	{ SRP_OPT_MAX_CMD_PER_LUN,	"max_cmd_per_lun=%d" 	},
2736 	{ SRP_OPT_IO_CLASS,		"io_class=%x"		},
2737 	{ SRP_OPT_INITIATOR_EXT,	"initiator_ext=%s"	},
2738 	{ SRP_OPT_CMD_SG_ENTRIES,	"cmd_sg_entries=%u"	},
2739 	{ SRP_OPT_ALLOW_EXT_SG,		"allow_ext_sg=%u"	},
2740 	{ SRP_OPT_SG_TABLESIZE,		"sg_tablesize=%u"	},
2741 	{ SRP_OPT_COMP_VECTOR,		"comp_vector=%u"	},
2742 	{ SRP_OPT_TL_RETRY_COUNT,	"tl_retry_count=%u"	},
2743 	{ SRP_OPT_QUEUE_SIZE,		"queue_size=%d"		},
2744 	{ SRP_OPT_ERR,			NULL 			}
2745 };
2746 
2747 static int srp_parse_options(const char *buf, struct srp_target_port *target)
2748 {
2749 	char *options, *sep_opt;
2750 	char *p;
2751 	char dgid[3];
2752 	substring_t args[MAX_OPT_ARGS];
2753 	int opt_mask = 0;
2754 	int token;
2755 	int ret = -EINVAL;
2756 	int i;
2757 
2758 	options = kstrdup(buf, GFP_KERNEL);
2759 	if (!options)
2760 		return -ENOMEM;
2761 
2762 	sep_opt = options;
2763 	while ((p = strsep(&sep_opt, ",")) != NULL) {
2764 		if (!*p)
2765 			continue;
2766 
2767 		token = match_token(p, srp_opt_tokens, args);
2768 		opt_mask |= token;
2769 
2770 		switch (token) {
2771 		case SRP_OPT_ID_EXT:
2772 			p = match_strdup(args);
2773 			if (!p) {
2774 				ret = -ENOMEM;
2775 				goto out;
2776 			}
2777 			target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2778 			kfree(p);
2779 			break;
2780 
2781 		case SRP_OPT_IOC_GUID:
2782 			p = match_strdup(args);
2783 			if (!p) {
2784 				ret = -ENOMEM;
2785 				goto out;
2786 			}
2787 			target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
2788 			kfree(p);
2789 			break;
2790 
2791 		case SRP_OPT_DGID:
2792 			p = match_strdup(args);
2793 			if (!p) {
2794 				ret = -ENOMEM;
2795 				goto out;
2796 			}
2797 			if (strlen(p) != 32) {
2798 				pr_warn("bad dest GID parameter '%s'\n", p);
2799 				kfree(p);
2800 				goto out;
2801 			}
2802 
2803 			for (i = 0; i < 16; ++i) {
2804 				strlcpy(dgid, p + i * 2, 3);
2805 				target->path.dgid.raw[i] = simple_strtoul(dgid, NULL, 16);
2806 			}
2807 			kfree(p);
2808 			memcpy(target->orig_dgid, target->path.dgid.raw, 16);
2809 			break;
2810 
2811 		case SRP_OPT_PKEY:
2812 			if (match_hex(args, &token)) {
2813 				pr_warn("bad P_Key parameter '%s'\n", p);
2814 				goto out;
2815 			}
2816 			target->path.pkey = cpu_to_be16(token);
2817 			break;
2818 
2819 		case SRP_OPT_SERVICE_ID:
2820 			p = match_strdup(args);
2821 			if (!p) {
2822 				ret = -ENOMEM;
2823 				goto out;
2824 			}
2825 			target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
2826 			target->path.service_id = target->service_id;
2827 			kfree(p);
2828 			break;
2829 
2830 		case SRP_OPT_MAX_SECT:
2831 			if (match_int(args, &token)) {
2832 				pr_warn("bad max sect parameter '%s'\n", p);
2833 				goto out;
2834 			}
2835 			target->scsi_host->max_sectors = token;
2836 			break;
2837 
2838 		case SRP_OPT_QUEUE_SIZE:
2839 			if (match_int(args, &token) || token < 1) {
2840 				pr_warn("bad queue_size parameter '%s'\n", p);
2841 				goto out;
2842 			}
2843 			target->scsi_host->can_queue = token;
2844 			target->queue_size = token + SRP_RSP_SQ_SIZE +
2845 					     SRP_TSK_MGMT_SQ_SIZE;
2846 			if (!(opt_mask & SRP_OPT_MAX_CMD_PER_LUN))
2847 				target->scsi_host->cmd_per_lun = token;
2848 			break;
2849 
2850 		case SRP_OPT_MAX_CMD_PER_LUN:
2851 			if (match_int(args, &token) || token < 1) {
2852 				pr_warn("bad max cmd_per_lun parameter '%s'\n",
2853 					p);
2854 				goto out;
2855 			}
2856 			target->scsi_host->cmd_per_lun = token;
2857 			break;
2858 
2859 		case SRP_OPT_IO_CLASS:
2860 			if (match_hex(args, &token)) {
2861 				pr_warn("bad IO class parameter '%s'\n", p);
2862 				goto out;
2863 			}
2864 			if (token != SRP_REV10_IB_IO_CLASS &&
2865 			    token != SRP_REV16A_IB_IO_CLASS) {
2866 				pr_warn("unknown IO class parameter value %x specified (use %x or %x).\n",
2867 					token, SRP_REV10_IB_IO_CLASS,
2868 					SRP_REV16A_IB_IO_CLASS);
2869 				goto out;
2870 			}
2871 			target->io_class = token;
2872 			break;
2873 
2874 		case SRP_OPT_INITIATOR_EXT:
2875 			p = match_strdup(args);
2876 			if (!p) {
2877 				ret = -ENOMEM;
2878 				goto out;
2879 			}
2880 			target->initiator_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2881 			kfree(p);
2882 			break;
2883 
2884 		case SRP_OPT_CMD_SG_ENTRIES:
2885 			if (match_int(args, &token) || token < 1 || token > 255) {
2886 				pr_warn("bad max cmd_sg_entries parameter '%s'\n",
2887 					p);
2888 				goto out;
2889 			}
2890 			target->cmd_sg_cnt = token;
2891 			break;
2892 
2893 		case SRP_OPT_ALLOW_EXT_SG:
2894 			if (match_int(args, &token)) {
2895 				pr_warn("bad allow_ext_sg parameter '%s'\n", p);
2896 				goto out;
2897 			}
2898 			target->allow_ext_sg = !!token;
2899 			break;
2900 
2901 		case SRP_OPT_SG_TABLESIZE:
2902 			if (match_int(args, &token) || token < 1 ||
2903 					token > SCSI_MAX_SG_CHAIN_SEGMENTS) {
2904 				pr_warn("bad max sg_tablesize parameter '%s'\n",
2905 					p);
2906 				goto out;
2907 			}
2908 			target->sg_tablesize = token;
2909 			break;
2910 
2911 		case SRP_OPT_COMP_VECTOR:
2912 			if (match_int(args, &token) || token < 0) {
2913 				pr_warn("bad comp_vector parameter '%s'\n", p);
2914 				goto out;
2915 			}
2916 			target->comp_vector = token;
2917 			break;
2918 
2919 		case SRP_OPT_TL_RETRY_COUNT:
2920 			if (match_int(args, &token) || token < 2 || token > 7) {
2921 				pr_warn("bad tl_retry_count parameter '%s' (must be a number between 2 and 7)\n",
2922 					p);
2923 				goto out;
2924 			}
2925 			target->tl_retry_count = token;
2926 			break;
2927 
2928 		default:
2929 			pr_warn("unknown parameter or missing value '%s' in target creation request\n",
2930 				p);
2931 			goto out;
2932 		}
2933 	}
2934 
2935 	if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
2936 		ret = 0;
2937 	else
2938 		for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
2939 			if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
2940 			    !(srp_opt_tokens[i].token & opt_mask))
2941 				pr_warn("target creation request is missing parameter '%s'\n",
2942 					srp_opt_tokens[i].pattern);
2943 
2944 	if (target->scsi_host->cmd_per_lun > target->scsi_host->can_queue
2945 	    && (opt_mask & SRP_OPT_MAX_CMD_PER_LUN))
2946 		pr_warn("cmd_per_lun = %d > queue_size = %d\n",
2947 			target->scsi_host->cmd_per_lun,
2948 			target->scsi_host->can_queue);
2949 
2950 out:
2951 	kfree(options);
2952 	return ret;
2953 }
2954 
2955 static ssize_t srp_create_target(struct device *dev,
2956 				 struct device_attribute *attr,
2957 				 const char *buf, size_t count)
2958 {
2959 	struct srp_host *host =
2960 		container_of(dev, struct srp_host, dev);
2961 	struct Scsi_Host *target_host;
2962 	struct srp_target_port *target;
2963 	struct srp_device *srp_dev = host->srp_dev;
2964 	struct ib_device *ibdev = srp_dev->dev;
2965 	int ret;
2966 
2967 	target_host = scsi_host_alloc(&srp_template,
2968 				      sizeof (struct srp_target_port));
2969 	if (!target_host)
2970 		return -ENOMEM;
2971 
2972 	target_host->transportt  = ib_srp_transport_template;
2973 	target_host->max_channel = 0;
2974 	target_host->max_id      = 1;
2975 	target_host->max_lun     = SRP_MAX_LUN;
2976 	target_host->max_cmd_len = sizeof ((struct srp_cmd *) (void *) 0L)->cdb;
2977 
2978 	target = host_to_target(target_host);
2979 
2980 	target->io_class	= SRP_REV16A_IB_IO_CLASS;
2981 	target->scsi_host	= target_host;
2982 	target->srp_host	= host;
2983 	target->lkey		= host->srp_dev->mr->lkey;
2984 	target->rkey		= host->srp_dev->mr->rkey;
2985 	target->cmd_sg_cnt	= cmd_sg_entries;
2986 	target->sg_tablesize	= indirect_sg_entries ? : cmd_sg_entries;
2987 	target->allow_ext_sg	= allow_ext_sg;
2988 	target->tl_retry_count	= 7;
2989 	target->queue_size	= SRP_DEFAULT_QUEUE_SIZE;
2990 
2991 	mutex_lock(&host->add_target_mutex);
2992 
2993 	ret = srp_parse_options(buf, target);
2994 	if (ret)
2995 		goto err;
2996 
2997 	target->req_ring_size = target->queue_size - SRP_TSK_MGMT_SQ_SIZE;
2998 
2999 	if (!srp_conn_unique(target->srp_host, target)) {
3000 		shost_printk(KERN_INFO, target->scsi_host,
3001 			     PFX "Already connected to target port with id_ext=%016llx;ioc_guid=%016llx;initiator_ext=%016llx\n",
3002 			     be64_to_cpu(target->id_ext),
3003 			     be64_to_cpu(target->ioc_guid),
3004 			     be64_to_cpu(target->initiator_ext));
3005 		ret = -EEXIST;
3006 		goto err;
3007 	}
3008 
3009 	if (!srp_dev->has_fmr && !srp_dev->has_fr && !target->allow_ext_sg &&
3010 	    target->cmd_sg_cnt < target->sg_tablesize) {
3011 		pr_warn("No MR pool and no external indirect descriptors, limiting sg_tablesize to cmd_sg_cnt\n");
3012 		target->sg_tablesize = target->cmd_sg_cnt;
3013 	}
3014 
3015 	target_host->sg_tablesize = target->sg_tablesize;
3016 	target->indirect_size = target->sg_tablesize *
3017 				sizeof (struct srp_direct_buf);
3018 	target->max_iu_len = sizeof (struct srp_cmd) +
3019 			     sizeof (struct srp_indirect_buf) +
3020 			     target->cmd_sg_cnt * sizeof (struct srp_direct_buf);
3021 
3022 	INIT_WORK(&target->tl_err_work, srp_tl_err_work);
3023 	INIT_WORK(&target->remove_work, srp_remove_work);
3024 	spin_lock_init(&target->lock);
3025 	INIT_LIST_HEAD(&target->free_tx);
3026 	ret = srp_alloc_req_data(target);
3027 	if (ret)
3028 		goto err_free_mem;
3029 
3030 	ret = ib_query_gid(ibdev, host->port, 0, &target->path.sgid);
3031 	if (ret)
3032 		goto err_free_mem;
3033 
3034 	ret = srp_create_target_ib(target);
3035 	if (ret)
3036 		goto err_free_mem;
3037 
3038 	ret = srp_new_cm_id(target);
3039 	if (ret)
3040 		goto err_free_ib;
3041 
3042 	ret = srp_connect_target(target);
3043 	if (ret) {
3044 		shost_printk(KERN_ERR, target->scsi_host,
3045 			     PFX "Connection failed\n");
3046 		goto err_cm_id;
3047 	}
3048 
3049 	ret = srp_add_target(host, target);
3050 	if (ret)
3051 		goto err_disconnect;
3052 
3053 	shost_printk(KERN_DEBUG, target->scsi_host, PFX
3054 		     "new target: id_ext %016llx ioc_guid %016llx pkey %04x service_id %016llx sgid %pI6 dgid %pI6\n",
3055 		     be64_to_cpu(target->id_ext),
3056 		     be64_to_cpu(target->ioc_guid),
3057 		     be16_to_cpu(target->path.pkey),
3058 		     be64_to_cpu(target->service_id),
3059 		     target->path.sgid.raw, target->path.dgid.raw);
3060 
3061 	ret = count;
3062 
3063 out:
3064 	mutex_unlock(&host->add_target_mutex);
3065 	return ret;
3066 
3067 err_disconnect:
3068 	srp_disconnect_target(target);
3069 
3070 err_cm_id:
3071 	ib_destroy_cm_id(target->cm_id);
3072 
3073 err_free_ib:
3074 	srp_free_target_ib(target);
3075 
3076 err_free_mem:
3077 	srp_free_req_data(target);
3078 
3079 err:
3080 	scsi_host_put(target_host);
3081 	goto out;
3082 }
3083 
3084 static DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
3085 
3086 static ssize_t show_ibdev(struct device *dev, struct device_attribute *attr,
3087 			  char *buf)
3088 {
3089 	struct srp_host *host = container_of(dev, struct srp_host, dev);
3090 
3091 	return sprintf(buf, "%s\n", host->srp_dev->dev->name);
3092 }
3093 
3094 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
3095 
3096 static ssize_t show_port(struct device *dev, struct device_attribute *attr,
3097 			 char *buf)
3098 {
3099 	struct srp_host *host = container_of(dev, struct srp_host, dev);
3100 
3101 	return sprintf(buf, "%d\n", host->port);
3102 }
3103 
3104 static DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
3105 
3106 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
3107 {
3108 	struct srp_host *host;
3109 
3110 	host = kzalloc(sizeof *host, GFP_KERNEL);
3111 	if (!host)
3112 		return NULL;
3113 
3114 	INIT_LIST_HEAD(&host->target_list);
3115 	spin_lock_init(&host->target_lock);
3116 	init_completion(&host->released);
3117 	mutex_init(&host->add_target_mutex);
3118 	host->srp_dev = device;
3119 	host->port = port;
3120 
3121 	host->dev.class = &srp_class;
3122 	host->dev.parent = device->dev->dma_device;
3123 	dev_set_name(&host->dev, "srp-%s-%d", device->dev->name, port);
3124 
3125 	if (device_register(&host->dev))
3126 		goto free_host;
3127 	if (device_create_file(&host->dev, &dev_attr_add_target))
3128 		goto err_class;
3129 	if (device_create_file(&host->dev, &dev_attr_ibdev))
3130 		goto err_class;
3131 	if (device_create_file(&host->dev, &dev_attr_port))
3132 		goto err_class;
3133 
3134 	return host;
3135 
3136 err_class:
3137 	device_unregister(&host->dev);
3138 
3139 free_host:
3140 	kfree(host);
3141 
3142 	return NULL;
3143 }
3144 
3145 static void srp_add_one(struct ib_device *device)
3146 {
3147 	struct srp_device *srp_dev;
3148 	struct ib_device_attr *dev_attr;
3149 	struct srp_host *host;
3150 	int mr_page_shift, s, e, p;
3151 	u64 max_pages_per_mr;
3152 
3153 	dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
3154 	if (!dev_attr)
3155 		return;
3156 
3157 	if (ib_query_device(device, dev_attr)) {
3158 		pr_warn("Query device failed for %s\n", device->name);
3159 		goto free_attr;
3160 	}
3161 
3162 	srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
3163 	if (!srp_dev)
3164 		goto free_attr;
3165 
3166 	srp_dev->has_fmr = (device->alloc_fmr && device->dealloc_fmr &&
3167 			    device->map_phys_fmr && device->unmap_fmr);
3168 	srp_dev->has_fr = (dev_attr->device_cap_flags &
3169 			   IB_DEVICE_MEM_MGT_EXTENSIONS);
3170 	if (!srp_dev->has_fmr && !srp_dev->has_fr)
3171 		dev_warn(&device->dev, "neither FMR nor FR is supported\n");
3172 
3173 	srp_dev->use_fast_reg = (srp_dev->has_fr &&
3174 				 (!srp_dev->has_fmr || prefer_fr));
3175 
3176 	/*
3177 	 * Use the smallest page size supported by the HCA, down to a
3178 	 * minimum of 4096 bytes. We're unlikely to build large sglists
3179 	 * out of smaller entries.
3180 	 */
3181 	mr_page_shift		= max(12, ffs(dev_attr->page_size_cap) - 1);
3182 	srp_dev->mr_page_size	= 1 << mr_page_shift;
3183 	srp_dev->mr_page_mask	= ~((u64) srp_dev->mr_page_size - 1);
3184 	max_pages_per_mr	= dev_attr->max_mr_size;
3185 	do_div(max_pages_per_mr, srp_dev->mr_page_size);
3186 	srp_dev->max_pages_per_mr = min_t(u64, SRP_MAX_PAGES_PER_MR,
3187 					  max_pages_per_mr);
3188 	if (srp_dev->use_fast_reg) {
3189 		srp_dev->max_pages_per_mr =
3190 			min_t(u32, srp_dev->max_pages_per_mr,
3191 			      dev_attr->max_fast_reg_page_list_len);
3192 	}
3193 	srp_dev->mr_max_size	= srp_dev->mr_page_size *
3194 				   srp_dev->max_pages_per_mr;
3195 	pr_debug("%s: mr_page_shift = %d, dev_attr->max_mr_size = %#llx, dev_attr->max_fast_reg_page_list_len = %u, max_pages_per_mr = %d, mr_max_size = %#x\n",
3196 		 device->name, mr_page_shift, dev_attr->max_mr_size,
3197 		 dev_attr->max_fast_reg_page_list_len,
3198 		 srp_dev->max_pages_per_mr, srp_dev->mr_max_size);
3199 
3200 	INIT_LIST_HEAD(&srp_dev->dev_list);
3201 
3202 	srp_dev->dev = device;
3203 	srp_dev->pd  = ib_alloc_pd(device);
3204 	if (IS_ERR(srp_dev->pd))
3205 		goto free_dev;
3206 
3207 	srp_dev->mr = ib_get_dma_mr(srp_dev->pd,
3208 				    IB_ACCESS_LOCAL_WRITE |
3209 				    IB_ACCESS_REMOTE_READ |
3210 				    IB_ACCESS_REMOTE_WRITE);
3211 	if (IS_ERR(srp_dev->mr))
3212 		goto err_pd;
3213 
3214 	if (device->node_type == RDMA_NODE_IB_SWITCH) {
3215 		s = 0;
3216 		e = 0;
3217 	} else {
3218 		s = 1;
3219 		e = device->phys_port_cnt;
3220 	}
3221 
3222 	for (p = s; p <= e; ++p) {
3223 		host = srp_add_port(srp_dev, p);
3224 		if (host)
3225 			list_add_tail(&host->list, &srp_dev->dev_list);
3226 	}
3227 
3228 	ib_set_client_data(device, &srp_client, srp_dev);
3229 
3230 	goto free_attr;
3231 
3232 err_pd:
3233 	ib_dealloc_pd(srp_dev->pd);
3234 
3235 free_dev:
3236 	kfree(srp_dev);
3237 
3238 free_attr:
3239 	kfree(dev_attr);
3240 }
3241 
3242 static void srp_remove_one(struct ib_device *device)
3243 {
3244 	struct srp_device *srp_dev;
3245 	struct srp_host *host, *tmp_host;
3246 	struct srp_target_port *target;
3247 
3248 	srp_dev = ib_get_client_data(device, &srp_client);
3249 	if (!srp_dev)
3250 		return;
3251 
3252 	list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
3253 		device_unregister(&host->dev);
3254 		/*
3255 		 * Wait for the sysfs entry to go away, so that no new
3256 		 * target ports can be created.
3257 		 */
3258 		wait_for_completion(&host->released);
3259 
3260 		/*
3261 		 * Remove all target ports.
3262 		 */
3263 		spin_lock(&host->target_lock);
3264 		list_for_each_entry(target, &host->target_list, list)
3265 			srp_queue_remove_work(target);
3266 		spin_unlock(&host->target_lock);
3267 
3268 		/*
3269 		 * Wait for tl_err and target port removal tasks.
3270 		 */
3271 		flush_workqueue(system_long_wq);
3272 		flush_workqueue(srp_remove_wq);
3273 
3274 		kfree(host);
3275 	}
3276 
3277 	ib_dereg_mr(srp_dev->mr);
3278 	ib_dealloc_pd(srp_dev->pd);
3279 
3280 	kfree(srp_dev);
3281 }
3282 
3283 static struct srp_function_template ib_srp_transport_functions = {
3284 	.has_rport_state	 = true,
3285 	.reset_timer_if_blocked	 = true,
3286 	.reconnect_delay	 = &srp_reconnect_delay,
3287 	.fast_io_fail_tmo	 = &srp_fast_io_fail_tmo,
3288 	.dev_loss_tmo		 = &srp_dev_loss_tmo,
3289 	.reconnect		 = srp_rport_reconnect,
3290 	.rport_delete		 = srp_rport_delete,
3291 	.terminate_rport_io	 = srp_terminate_io,
3292 };
3293 
3294 static int __init srp_init_module(void)
3295 {
3296 	int ret;
3297 
3298 	BUILD_BUG_ON(FIELD_SIZEOF(struct ib_wc, wr_id) < sizeof(void *));
3299 
3300 	if (srp_sg_tablesize) {
3301 		pr_warn("srp_sg_tablesize is deprecated, please use cmd_sg_entries\n");
3302 		if (!cmd_sg_entries)
3303 			cmd_sg_entries = srp_sg_tablesize;
3304 	}
3305 
3306 	if (!cmd_sg_entries)
3307 		cmd_sg_entries = SRP_DEF_SG_TABLESIZE;
3308 
3309 	if (cmd_sg_entries > 255) {
3310 		pr_warn("Clamping cmd_sg_entries to 255\n");
3311 		cmd_sg_entries = 255;
3312 	}
3313 
3314 	if (!indirect_sg_entries)
3315 		indirect_sg_entries = cmd_sg_entries;
3316 	else if (indirect_sg_entries < cmd_sg_entries) {
3317 		pr_warn("Bumping up indirect_sg_entries to match cmd_sg_entries (%u)\n",
3318 			cmd_sg_entries);
3319 		indirect_sg_entries = cmd_sg_entries;
3320 	}
3321 
3322 	srp_remove_wq = create_workqueue("srp_remove");
3323 	if (!srp_remove_wq) {
3324 		ret = -ENOMEM;
3325 		goto out;
3326 	}
3327 
3328 	ret = -ENOMEM;
3329 	ib_srp_transport_template =
3330 		srp_attach_transport(&ib_srp_transport_functions);
3331 	if (!ib_srp_transport_template)
3332 		goto destroy_wq;
3333 
3334 	ret = class_register(&srp_class);
3335 	if (ret) {
3336 		pr_err("couldn't register class infiniband_srp\n");
3337 		goto release_tr;
3338 	}
3339 
3340 	ib_sa_register_client(&srp_sa_client);
3341 
3342 	ret = ib_register_client(&srp_client);
3343 	if (ret) {
3344 		pr_err("couldn't register IB client\n");
3345 		goto unreg_sa;
3346 	}
3347 
3348 out:
3349 	return ret;
3350 
3351 unreg_sa:
3352 	ib_sa_unregister_client(&srp_sa_client);
3353 	class_unregister(&srp_class);
3354 
3355 release_tr:
3356 	srp_release_transport(ib_srp_transport_template);
3357 
3358 destroy_wq:
3359 	destroy_workqueue(srp_remove_wq);
3360 	goto out;
3361 }
3362 
3363 static void __exit srp_cleanup_module(void)
3364 {
3365 	ib_unregister_client(&srp_client);
3366 	ib_sa_unregister_client(&srp_sa_client);
3367 	class_unregister(&srp_class);
3368 	srp_release_transport(ib_srp_transport_template);
3369 	destroy_workqueue(srp_remove_wq);
3370 }
3371 
3372 module_init(srp_init_module);
3373 module_exit(srp_cleanup_module);
3374