1 /*
2  * Copyright (c) 2006 - 2009 Mellanox Technology Inc.  All rights reserved.
3  * Copyright (C) 2008 - 2011 Bart Van Assche <bvanassche@acm.org>.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  *
33  */
34 
35 #include <linux/module.h>
36 #include <linux/init.h>
37 #include <linux/slab.h>
38 #include <linux/err.h>
39 #include <linux/ctype.h>
40 #include <linux/kthread.h>
41 #include <linux/string.h>
42 #include <linux/delay.h>
43 #include <linux/atomic.h>
44 #include <scsi/scsi_tcq.h>
45 #include <target/configfs_macros.h>
46 #include <target/target_core_base.h>
47 #include <target/target_core_fabric_configfs.h>
48 #include <target/target_core_fabric.h>
49 #include <target/target_core_configfs.h>
50 #include "ib_srpt.h"
51 
52 /* Name of this kernel module. */
53 #define DRV_NAME		"ib_srpt"
54 #define DRV_VERSION		"2.0.0"
55 #define DRV_RELDATE		"2011-02-14"
56 
57 #define SRPT_ID_STRING	"Linux SRP target"
58 
59 #undef pr_fmt
60 #define pr_fmt(fmt) DRV_NAME " " fmt
61 
62 MODULE_AUTHOR("Vu Pham and Bart Van Assche");
63 MODULE_DESCRIPTION("InfiniBand SCSI RDMA Protocol target "
64 		   "v" DRV_VERSION " (" DRV_RELDATE ")");
65 MODULE_LICENSE("Dual BSD/GPL");
66 
67 /*
68  * Global Variables
69  */
70 
71 static u64 srpt_service_guid;
72 static DEFINE_SPINLOCK(srpt_dev_lock);	/* Protects srpt_dev_list. */
73 static LIST_HEAD(srpt_dev_list);	/* List of srpt_device structures. */
74 
75 static unsigned srp_max_req_size = DEFAULT_MAX_REQ_SIZE;
76 module_param(srp_max_req_size, int, 0444);
77 MODULE_PARM_DESC(srp_max_req_size,
78 		 "Maximum size of SRP request messages in bytes.");
79 
80 static int srpt_srq_size = DEFAULT_SRPT_SRQ_SIZE;
81 module_param(srpt_srq_size, int, 0444);
82 MODULE_PARM_DESC(srpt_srq_size,
83 		 "Shared receive queue (SRQ) size.");
84 
85 static int srpt_get_u64_x(char *buffer, struct kernel_param *kp)
86 {
87 	return sprintf(buffer, "0x%016llx", *(u64 *)kp->arg);
88 }
89 module_param_call(srpt_service_guid, NULL, srpt_get_u64_x, &srpt_service_guid,
90 		  0444);
91 MODULE_PARM_DESC(srpt_service_guid,
92 		 "Using this value for ioc_guid, id_ext, and cm_listen_id"
93 		 " instead of using the node_guid of the first HCA.");
94 
95 static struct ib_client srpt_client;
96 static struct target_fabric_configfs *srpt_target;
97 static void srpt_release_channel(struct srpt_rdma_ch *ch);
98 static int srpt_queue_status(struct se_cmd *cmd);
99 
100 /**
101  * opposite_dma_dir() - Swap DMA_TO_DEVICE and DMA_FROM_DEVICE.
102  */
103 static inline
104 enum dma_data_direction opposite_dma_dir(enum dma_data_direction dir)
105 {
106 	switch (dir) {
107 	case DMA_TO_DEVICE:	return DMA_FROM_DEVICE;
108 	case DMA_FROM_DEVICE:	return DMA_TO_DEVICE;
109 	default:		return dir;
110 	}
111 }
112 
113 /**
114  * srpt_sdev_name() - Return the name associated with the HCA.
115  *
116  * Examples are ib0, ib1, ...
117  */
118 static inline const char *srpt_sdev_name(struct srpt_device *sdev)
119 {
120 	return sdev->device->name;
121 }
122 
123 static enum rdma_ch_state srpt_get_ch_state(struct srpt_rdma_ch *ch)
124 {
125 	unsigned long flags;
126 	enum rdma_ch_state state;
127 
128 	spin_lock_irqsave(&ch->spinlock, flags);
129 	state = ch->state;
130 	spin_unlock_irqrestore(&ch->spinlock, flags);
131 	return state;
132 }
133 
134 static enum rdma_ch_state
135 srpt_set_ch_state(struct srpt_rdma_ch *ch, enum rdma_ch_state new_state)
136 {
137 	unsigned long flags;
138 	enum rdma_ch_state prev;
139 
140 	spin_lock_irqsave(&ch->spinlock, flags);
141 	prev = ch->state;
142 	ch->state = new_state;
143 	spin_unlock_irqrestore(&ch->spinlock, flags);
144 	return prev;
145 }
146 
147 /**
148  * srpt_test_and_set_ch_state() - Test and set the channel state.
149  *
150  * Returns true if and only if the channel state has been set to the new state.
151  */
152 static bool
153 srpt_test_and_set_ch_state(struct srpt_rdma_ch *ch, enum rdma_ch_state old,
154 			   enum rdma_ch_state new)
155 {
156 	unsigned long flags;
157 	enum rdma_ch_state prev;
158 
159 	spin_lock_irqsave(&ch->spinlock, flags);
160 	prev = ch->state;
161 	if (prev == old)
162 		ch->state = new;
163 	spin_unlock_irqrestore(&ch->spinlock, flags);
164 	return prev == old;
165 }
166 
167 /**
168  * srpt_event_handler() - Asynchronous IB event callback function.
169  *
170  * Callback function called by the InfiniBand core when an asynchronous IB
171  * event occurs. This callback may occur in interrupt context. See also
172  * section 11.5.2, Set Asynchronous Event Handler in the InfiniBand
173  * Architecture Specification.
174  */
175 static void srpt_event_handler(struct ib_event_handler *handler,
176 			       struct ib_event *event)
177 {
178 	struct srpt_device *sdev;
179 	struct srpt_port *sport;
180 
181 	sdev = ib_get_client_data(event->device, &srpt_client);
182 	if (!sdev || sdev->device != event->device)
183 		return;
184 
185 	pr_debug("ASYNC event= %d on device= %s\n", event->event,
186 		 srpt_sdev_name(sdev));
187 
188 	switch (event->event) {
189 	case IB_EVENT_PORT_ERR:
190 		if (event->element.port_num <= sdev->device->phys_port_cnt) {
191 			sport = &sdev->port[event->element.port_num - 1];
192 			sport->lid = 0;
193 			sport->sm_lid = 0;
194 		}
195 		break;
196 	case IB_EVENT_PORT_ACTIVE:
197 	case IB_EVENT_LID_CHANGE:
198 	case IB_EVENT_PKEY_CHANGE:
199 	case IB_EVENT_SM_CHANGE:
200 	case IB_EVENT_CLIENT_REREGISTER:
201 		/* Refresh port data asynchronously. */
202 		if (event->element.port_num <= sdev->device->phys_port_cnt) {
203 			sport = &sdev->port[event->element.port_num - 1];
204 			if (!sport->lid && !sport->sm_lid)
205 				schedule_work(&sport->work);
206 		}
207 		break;
208 	default:
209 		printk(KERN_ERR "received unrecognized IB event %d\n",
210 		       event->event);
211 		break;
212 	}
213 }
214 
215 /**
216  * srpt_srq_event() - SRQ event callback function.
217  */
218 static void srpt_srq_event(struct ib_event *event, void *ctx)
219 {
220 	printk(KERN_INFO "SRQ event %d\n", event->event);
221 }
222 
223 /**
224  * srpt_qp_event() - QP event callback function.
225  */
226 static void srpt_qp_event(struct ib_event *event, struct srpt_rdma_ch *ch)
227 {
228 	pr_debug("QP event %d on cm_id=%p sess_name=%s state=%d\n",
229 		 event->event, ch->cm_id, ch->sess_name, srpt_get_ch_state(ch));
230 
231 	switch (event->event) {
232 	case IB_EVENT_COMM_EST:
233 		ib_cm_notify(ch->cm_id, event->event);
234 		break;
235 	case IB_EVENT_QP_LAST_WQE_REACHED:
236 		if (srpt_test_and_set_ch_state(ch, CH_DRAINING,
237 					       CH_RELEASING))
238 			srpt_release_channel(ch);
239 		else
240 			pr_debug("%s: state %d - ignored LAST_WQE.\n",
241 				 ch->sess_name, srpt_get_ch_state(ch));
242 		break;
243 	default:
244 		printk(KERN_ERR "received unrecognized IB QP event %d\n",
245 		       event->event);
246 		break;
247 	}
248 }
249 
250 /**
251  * srpt_set_ioc() - Helper function for initializing an IOUnitInfo structure.
252  *
253  * @slot: one-based slot number.
254  * @value: four-bit value.
255  *
256  * Copies the lowest four bits of value in element slot of the array of four
257  * bit elements called c_list (controller list). The index slot is one-based.
258  */
259 static void srpt_set_ioc(u8 *c_list, u32 slot, u8 value)
260 {
261 	u16 id;
262 	u8 tmp;
263 
264 	id = (slot - 1) / 2;
265 	if (slot & 0x1) {
266 		tmp = c_list[id] & 0xf;
267 		c_list[id] = (value << 4) | tmp;
268 	} else {
269 		tmp = c_list[id] & 0xf0;
270 		c_list[id] = (value & 0xf) | tmp;
271 	}
272 }
273 
274 /**
275  * srpt_get_class_port_info() - Copy ClassPortInfo to a management datagram.
276  *
277  * See also section 16.3.3.1 ClassPortInfo in the InfiniBand Architecture
278  * Specification.
279  */
280 static void srpt_get_class_port_info(struct ib_dm_mad *mad)
281 {
282 	struct ib_class_port_info *cif;
283 
284 	cif = (struct ib_class_port_info *)mad->data;
285 	memset(cif, 0, sizeof *cif);
286 	cif->base_version = 1;
287 	cif->class_version = 1;
288 	cif->resp_time_value = 20;
289 
290 	mad->mad_hdr.status = 0;
291 }
292 
293 /**
294  * srpt_get_iou() - Write IOUnitInfo to a management datagram.
295  *
296  * See also section 16.3.3.3 IOUnitInfo in the InfiniBand Architecture
297  * Specification. See also section B.7, table B.6 in the SRP r16a document.
298  */
299 static void srpt_get_iou(struct ib_dm_mad *mad)
300 {
301 	struct ib_dm_iou_info *ioui;
302 	u8 slot;
303 	int i;
304 
305 	ioui = (struct ib_dm_iou_info *)mad->data;
306 	ioui->change_id = __constant_cpu_to_be16(1);
307 	ioui->max_controllers = 16;
308 
309 	/* set present for slot 1 and empty for the rest */
310 	srpt_set_ioc(ioui->controller_list, 1, 1);
311 	for (i = 1, slot = 2; i < 16; i++, slot++)
312 		srpt_set_ioc(ioui->controller_list, slot, 0);
313 
314 	mad->mad_hdr.status = 0;
315 }
316 
317 /**
318  * srpt_get_ioc() - Write IOControllerprofile to a management datagram.
319  *
320  * See also section 16.3.3.4 IOControllerProfile in the InfiniBand
321  * Architecture Specification. See also section B.7, table B.7 in the SRP
322  * r16a document.
323  */
324 static void srpt_get_ioc(struct srpt_port *sport, u32 slot,
325 			 struct ib_dm_mad *mad)
326 {
327 	struct srpt_device *sdev = sport->sdev;
328 	struct ib_dm_ioc_profile *iocp;
329 
330 	iocp = (struct ib_dm_ioc_profile *)mad->data;
331 
332 	if (!slot || slot > 16) {
333 		mad->mad_hdr.status
334 			= __constant_cpu_to_be16(DM_MAD_STATUS_INVALID_FIELD);
335 		return;
336 	}
337 
338 	if (slot > 2) {
339 		mad->mad_hdr.status
340 			= __constant_cpu_to_be16(DM_MAD_STATUS_NO_IOC);
341 		return;
342 	}
343 
344 	memset(iocp, 0, sizeof *iocp);
345 	strcpy(iocp->id_string, SRPT_ID_STRING);
346 	iocp->guid = cpu_to_be64(srpt_service_guid);
347 	iocp->vendor_id = cpu_to_be32(sdev->dev_attr.vendor_id);
348 	iocp->device_id = cpu_to_be32(sdev->dev_attr.vendor_part_id);
349 	iocp->device_version = cpu_to_be16(sdev->dev_attr.hw_ver);
350 	iocp->subsys_vendor_id = cpu_to_be32(sdev->dev_attr.vendor_id);
351 	iocp->subsys_device_id = 0x0;
352 	iocp->io_class = __constant_cpu_to_be16(SRP_REV16A_IB_IO_CLASS);
353 	iocp->io_subclass = __constant_cpu_to_be16(SRP_IO_SUBCLASS);
354 	iocp->protocol = __constant_cpu_to_be16(SRP_PROTOCOL);
355 	iocp->protocol_version = __constant_cpu_to_be16(SRP_PROTOCOL_VERSION);
356 	iocp->send_queue_depth = cpu_to_be16(sdev->srq_size);
357 	iocp->rdma_read_depth = 4;
358 	iocp->send_size = cpu_to_be32(srp_max_req_size);
359 	iocp->rdma_size = cpu_to_be32(min(sport->port_attrib.srp_max_rdma_size,
360 					  1U << 24));
361 	iocp->num_svc_entries = 1;
362 	iocp->op_cap_mask = SRP_SEND_TO_IOC | SRP_SEND_FROM_IOC |
363 		SRP_RDMA_READ_FROM_IOC | SRP_RDMA_WRITE_FROM_IOC;
364 
365 	mad->mad_hdr.status = 0;
366 }
367 
368 /**
369  * srpt_get_svc_entries() - Write ServiceEntries to a management datagram.
370  *
371  * See also section 16.3.3.5 ServiceEntries in the InfiniBand Architecture
372  * Specification. See also section B.7, table B.8 in the SRP r16a document.
373  */
374 static void srpt_get_svc_entries(u64 ioc_guid,
375 				 u16 slot, u8 hi, u8 lo, struct ib_dm_mad *mad)
376 {
377 	struct ib_dm_svc_entries *svc_entries;
378 
379 	WARN_ON(!ioc_guid);
380 
381 	if (!slot || slot > 16) {
382 		mad->mad_hdr.status
383 			= __constant_cpu_to_be16(DM_MAD_STATUS_INVALID_FIELD);
384 		return;
385 	}
386 
387 	if (slot > 2 || lo > hi || hi > 1) {
388 		mad->mad_hdr.status
389 			= __constant_cpu_to_be16(DM_MAD_STATUS_NO_IOC);
390 		return;
391 	}
392 
393 	svc_entries = (struct ib_dm_svc_entries *)mad->data;
394 	memset(svc_entries, 0, sizeof *svc_entries);
395 	svc_entries->service_entries[0].id = cpu_to_be64(ioc_guid);
396 	snprintf(svc_entries->service_entries[0].name,
397 		 sizeof(svc_entries->service_entries[0].name),
398 		 "%s%016llx",
399 		 SRP_SERVICE_NAME_PREFIX,
400 		 ioc_guid);
401 
402 	mad->mad_hdr.status = 0;
403 }
404 
405 /**
406  * srpt_mgmt_method_get() - Process a received management datagram.
407  * @sp:      source port through which the MAD has been received.
408  * @rq_mad:  received MAD.
409  * @rsp_mad: response MAD.
410  */
411 static void srpt_mgmt_method_get(struct srpt_port *sp, struct ib_mad *rq_mad,
412 				 struct ib_dm_mad *rsp_mad)
413 {
414 	u16 attr_id;
415 	u32 slot;
416 	u8 hi, lo;
417 
418 	attr_id = be16_to_cpu(rq_mad->mad_hdr.attr_id);
419 	switch (attr_id) {
420 	case DM_ATTR_CLASS_PORT_INFO:
421 		srpt_get_class_port_info(rsp_mad);
422 		break;
423 	case DM_ATTR_IOU_INFO:
424 		srpt_get_iou(rsp_mad);
425 		break;
426 	case DM_ATTR_IOC_PROFILE:
427 		slot = be32_to_cpu(rq_mad->mad_hdr.attr_mod);
428 		srpt_get_ioc(sp, slot, rsp_mad);
429 		break;
430 	case DM_ATTR_SVC_ENTRIES:
431 		slot = be32_to_cpu(rq_mad->mad_hdr.attr_mod);
432 		hi = (u8) ((slot >> 8) & 0xff);
433 		lo = (u8) (slot & 0xff);
434 		slot = (u16) ((slot >> 16) & 0xffff);
435 		srpt_get_svc_entries(srpt_service_guid,
436 				     slot, hi, lo, rsp_mad);
437 		break;
438 	default:
439 		rsp_mad->mad_hdr.status =
440 		    __constant_cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD_ATTR);
441 		break;
442 	}
443 }
444 
445 /**
446  * srpt_mad_send_handler() - Post MAD-send callback function.
447  */
448 static void srpt_mad_send_handler(struct ib_mad_agent *mad_agent,
449 				  struct ib_mad_send_wc *mad_wc)
450 {
451 	ib_destroy_ah(mad_wc->send_buf->ah);
452 	ib_free_send_mad(mad_wc->send_buf);
453 }
454 
455 /**
456  * srpt_mad_recv_handler() - MAD reception callback function.
457  */
458 static void srpt_mad_recv_handler(struct ib_mad_agent *mad_agent,
459 				  struct ib_mad_recv_wc *mad_wc)
460 {
461 	struct srpt_port *sport = (struct srpt_port *)mad_agent->context;
462 	struct ib_ah *ah;
463 	struct ib_mad_send_buf *rsp;
464 	struct ib_dm_mad *dm_mad;
465 
466 	if (!mad_wc || !mad_wc->recv_buf.mad)
467 		return;
468 
469 	ah = ib_create_ah_from_wc(mad_agent->qp->pd, mad_wc->wc,
470 				  mad_wc->recv_buf.grh, mad_agent->port_num);
471 	if (IS_ERR(ah))
472 		goto err;
473 
474 	BUILD_BUG_ON(offsetof(struct ib_dm_mad, data) != IB_MGMT_DEVICE_HDR);
475 
476 	rsp = ib_create_send_mad(mad_agent, mad_wc->wc->src_qp,
477 				 mad_wc->wc->pkey_index, 0,
478 				 IB_MGMT_DEVICE_HDR, IB_MGMT_DEVICE_DATA,
479 				 GFP_KERNEL);
480 	if (IS_ERR(rsp))
481 		goto err_rsp;
482 
483 	rsp->ah = ah;
484 
485 	dm_mad = rsp->mad;
486 	memcpy(dm_mad, mad_wc->recv_buf.mad, sizeof *dm_mad);
487 	dm_mad->mad_hdr.method = IB_MGMT_METHOD_GET_RESP;
488 	dm_mad->mad_hdr.status = 0;
489 
490 	switch (mad_wc->recv_buf.mad->mad_hdr.method) {
491 	case IB_MGMT_METHOD_GET:
492 		srpt_mgmt_method_get(sport, mad_wc->recv_buf.mad, dm_mad);
493 		break;
494 	case IB_MGMT_METHOD_SET:
495 		dm_mad->mad_hdr.status =
496 		    __constant_cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD_ATTR);
497 		break;
498 	default:
499 		dm_mad->mad_hdr.status =
500 		    __constant_cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD);
501 		break;
502 	}
503 
504 	if (!ib_post_send_mad(rsp, NULL)) {
505 		ib_free_recv_mad(mad_wc);
506 		/* will destroy_ah & free_send_mad in send completion */
507 		return;
508 	}
509 
510 	ib_free_send_mad(rsp);
511 
512 err_rsp:
513 	ib_destroy_ah(ah);
514 err:
515 	ib_free_recv_mad(mad_wc);
516 }
517 
518 /**
519  * srpt_refresh_port() - Configure a HCA port.
520  *
521  * Enable InfiniBand management datagram processing, update the cached sm_lid,
522  * lid and gid values, and register a callback function for processing MADs
523  * on the specified port.
524  *
525  * Note: It is safe to call this function more than once for the same port.
526  */
527 static int srpt_refresh_port(struct srpt_port *sport)
528 {
529 	struct ib_mad_reg_req reg_req;
530 	struct ib_port_modify port_modify;
531 	struct ib_port_attr port_attr;
532 	int ret;
533 
534 	memset(&port_modify, 0, sizeof port_modify);
535 	port_modify.set_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP;
536 	port_modify.clr_port_cap_mask = 0;
537 
538 	ret = ib_modify_port(sport->sdev->device, sport->port, 0, &port_modify);
539 	if (ret)
540 		goto err_mod_port;
541 
542 	ret = ib_query_port(sport->sdev->device, sport->port, &port_attr);
543 	if (ret)
544 		goto err_query_port;
545 
546 	sport->sm_lid = port_attr.sm_lid;
547 	sport->lid = port_attr.lid;
548 
549 	ret = ib_query_gid(sport->sdev->device, sport->port, 0, &sport->gid);
550 	if (ret)
551 		goto err_query_port;
552 
553 	if (!sport->mad_agent) {
554 		memset(&reg_req, 0, sizeof reg_req);
555 		reg_req.mgmt_class = IB_MGMT_CLASS_DEVICE_MGMT;
556 		reg_req.mgmt_class_version = IB_MGMT_BASE_VERSION;
557 		set_bit(IB_MGMT_METHOD_GET, reg_req.method_mask);
558 		set_bit(IB_MGMT_METHOD_SET, reg_req.method_mask);
559 
560 		sport->mad_agent = ib_register_mad_agent(sport->sdev->device,
561 							 sport->port,
562 							 IB_QPT_GSI,
563 							 &reg_req, 0,
564 							 srpt_mad_send_handler,
565 							 srpt_mad_recv_handler,
566 							 sport);
567 		if (IS_ERR(sport->mad_agent)) {
568 			ret = PTR_ERR(sport->mad_agent);
569 			sport->mad_agent = NULL;
570 			goto err_query_port;
571 		}
572 	}
573 
574 	return 0;
575 
576 err_query_port:
577 
578 	port_modify.set_port_cap_mask = 0;
579 	port_modify.clr_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP;
580 	ib_modify_port(sport->sdev->device, sport->port, 0, &port_modify);
581 
582 err_mod_port:
583 
584 	return ret;
585 }
586 
587 /**
588  * srpt_unregister_mad_agent() - Unregister MAD callback functions.
589  *
590  * Note: It is safe to call this function more than once for the same device.
591  */
592 static void srpt_unregister_mad_agent(struct srpt_device *sdev)
593 {
594 	struct ib_port_modify port_modify = {
595 		.clr_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP,
596 	};
597 	struct srpt_port *sport;
598 	int i;
599 
600 	for (i = 1; i <= sdev->device->phys_port_cnt; i++) {
601 		sport = &sdev->port[i - 1];
602 		WARN_ON(sport->port != i);
603 		if (ib_modify_port(sdev->device, i, 0, &port_modify) < 0)
604 			printk(KERN_ERR "disabling MAD processing failed.\n");
605 		if (sport->mad_agent) {
606 			ib_unregister_mad_agent(sport->mad_agent);
607 			sport->mad_agent = NULL;
608 		}
609 	}
610 }
611 
612 /**
613  * srpt_alloc_ioctx() - Allocate an SRPT I/O context structure.
614  */
615 static struct srpt_ioctx *srpt_alloc_ioctx(struct srpt_device *sdev,
616 					   int ioctx_size, int dma_size,
617 					   enum dma_data_direction dir)
618 {
619 	struct srpt_ioctx *ioctx;
620 
621 	ioctx = kmalloc(ioctx_size, GFP_KERNEL);
622 	if (!ioctx)
623 		goto err;
624 
625 	ioctx->buf = kmalloc(dma_size, GFP_KERNEL);
626 	if (!ioctx->buf)
627 		goto err_free_ioctx;
628 
629 	ioctx->dma = ib_dma_map_single(sdev->device, ioctx->buf, dma_size, dir);
630 	if (ib_dma_mapping_error(sdev->device, ioctx->dma))
631 		goto err_free_buf;
632 
633 	return ioctx;
634 
635 err_free_buf:
636 	kfree(ioctx->buf);
637 err_free_ioctx:
638 	kfree(ioctx);
639 err:
640 	return NULL;
641 }
642 
643 /**
644  * srpt_free_ioctx() - Free an SRPT I/O context structure.
645  */
646 static void srpt_free_ioctx(struct srpt_device *sdev, struct srpt_ioctx *ioctx,
647 			    int dma_size, enum dma_data_direction dir)
648 {
649 	if (!ioctx)
650 		return;
651 
652 	ib_dma_unmap_single(sdev->device, ioctx->dma, dma_size, dir);
653 	kfree(ioctx->buf);
654 	kfree(ioctx);
655 }
656 
657 /**
658  * srpt_alloc_ioctx_ring() - Allocate a ring of SRPT I/O context structures.
659  * @sdev:       Device to allocate the I/O context ring for.
660  * @ring_size:  Number of elements in the I/O context ring.
661  * @ioctx_size: I/O context size.
662  * @dma_size:   DMA buffer size.
663  * @dir:        DMA data direction.
664  */
665 static struct srpt_ioctx **srpt_alloc_ioctx_ring(struct srpt_device *sdev,
666 				int ring_size, int ioctx_size,
667 				int dma_size, enum dma_data_direction dir)
668 {
669 	struct srpt_ioctx **ring;
670 	int i;
671 
672 	WARN_ON(ioctx_size != sizeof(struct srpt_recv_ioctx)
673 		&& ioctx_size != sizeof(struct srpt_send_ioctx));
674 
675 	ring = kmalloc(ring_size * sizeof(ring[0]), GFP_KERNEL);
676 	if (!ring)
677 		goto out;
678 	for (i = 0; i < ring_size; ++i) {
679 		ring[i] = srpt_alloc_ioctx(sdev, ioctx_size, dma_size, dir);
680 		if (!ring[i])
681 			goto err;
682 		ring[i]->index = i;
683 	}
684 	goto out;
685 
686 err:
687 	while (--i >= 0)
688 		srpt_free_ioctx(sdev, ring[i], dma_size, dir);
689 	kfree(ring);
690 	ring = NULL;
691 out:
692 	return ring;
693 }
694 
695 /**
696  * srpt_free_ioctx_ring() - Free the ring of SRPT I/O context structures.
697  */
698 static void srpt_free_ioctx_ring(struct srpt_ioctx **ioctx_ring,
699 				 struct srpt_device *sdev, int ring_size,
700 				 int dma_size, enum dma_data_direction dir)
701 {
702 	int i;
703 
704 	for (i = 0; i < ring_size; ++i)
705 		srpt_free_ioctx(sdev, ioctx_ring[i], dma_size, dir);
706 	kfree(ioctx_ring);
707 }
708 
709 /**
710  * srpt_get_cmd_state() - Get the state of a SCSI command.
711  */
712 static enum srpt_command_state srpt_get_cmd_state(struct srpt_send_ioctx *ioctx)
713 {
714 	enum srpt_command_state state;
715 	unsigned long flags;
716 
717 	BUG_ON(!ioctx);
718 
719 	spin_lock_irqsave(&ioctx->spinlock, flags);
720 	state = ioctx->state;
721 	spin_unlock_irqrestore(&ioctx->spinlock, flags);
722 	return state;
723 }
724 
725 /**
726  * srpt_set_cmd_state() - Set the state of a SCSI command.
727  *
728  * Does not modify the state of aborted commands. Returns the previous command
729  * state.
730  */
731 static enum srpt_command_state srpt_set_cmd_state(struct srpt_send_ioctx *ioctx,
732 						  enum srpt_command_state new)
733 {
734 	enum srpt_command_state previous;
735 	unsigned long flags;
736 
737 	BUG_ON(!ioctx);
738 
739 	spin_lock_irqsave(&ioctx->spinlock, flags);
740 	previous = ioctx->state;
741 	if (previous != SRPT_STATE_DONE)
742 		ioctx->state = new;
743 	spin_unlock_irqrestore(&ioctx->spinlock, flags);
744 
745 	return previous;
746 }
747 
748 /**
749  * srpt_test_and_set_cmd_state() - Test and set the state of a command.
750  *
751  * Returns true if and only if the previous command state was equal to 'old'.
752  */
753 static bool srpt_test_and_set_cmd_state(struct srpt_send_ioctx *ioctx,
754 					enum srpt_command_state old,
755 					enum srpt_command_state new)
756 {
757 	enum srpt_command_state previous;
758 	unsigned long flags;
759 
760 	WARN_ON(!ioctx);
761 	WARN_ON(old == SRPT_STATE_DONE);
762 	WARN_ON(new == SRPT_STATE_NEW);
763 
764 	spin_lock_irqsave(&ioctx->spinlock, flags);
765 	previous = ioctx->state;
766 	if (previous == old)
767 		ioctx->state = new;
768 	spin_unlock_irqrestore(&ioctx->spinlock, flags);
769 	return previous == old;
770 }
771 
772 /**
773  * srpt_post_recv() - Post an IB receive request.
774  */
775 static int srpt_post_recv(struct srpt_device *sdev,
776 			  struct srpt_recv_ioctx *ioctx)
777 {
778 	struct ib_sge list;
779 	struct ib_recv_wr wr, *bad_wr;
780 
781 	BUG_ON(!sdev);
782 	wr.wr_id = encode_wr_id(SRPT_RECV, ioctx->ioctx.index);
783 
784 	list.addr = ioctx->ioctx.dma;
785 	list.length = srp_max_req_size;
786 	list.lkey = sdev->mr->lkey;
787 
788 	wr.next = NULL;
789 	wr.sg_list = &list;
790 	wr.num_sge = 1;
791 
792 	return ib_post_srq_recv(sdev->srq, &wr, &bad_wr);
793 }
794 
795 /**
796  * srpt_post_send() - Post an IB send request.
797  *
798  * Returns zero upon success and a non-zero value upon failure.
799  */
800 static int srpt_post_send(struct srpt_rdma_ch *ch,
801 			  struct srpt_send_ioctx *ioctx, int len)
802 {
803 	struct ib_sge list;
804 	struct ib_send_wr wr, *bad_wr;
805 	struct srpt_device *sdev = ch->sport->sdev;
806 	int ret;
807 
808 	atomic_inc(&ch->req_lim);
809 
810 	ret = -ENOMEM;
811 	if (unlikely(atomic_dec_return(&ch->sq_wr_avail) < 0)) {
812 		printk(KERN_WARNING "IB send queue full (needed 1)\n");
813 		goto out;
814 	}
815 
816 	ib_dma_sync_single_for_device(sdev->device, ioctx->ioctx.dma, len,
817 				      DMA_TO_DEVICE);
818 
819 	list.addr = ioctx->ioctx.dma;
820 	list.length = len;
821 	list.lkey = sdev->mr->lkey;
822 
823 	wr.next = NULL;
824 	wr.wr_id = encode_wr_id(SRPT_SEND, ioctx->ioctx.index);
825 	wr.sg_list = &list;
826 	wr.num_sge = 1;
827 	wr.opcode = IB_WR_SEND;
828 	wr.send_flags = IB_SEND_SIGNALED;
829 
830 	ret = ib_post_send(ch->qp, &wr, &bad_wr);
831 
832 out:
833 	if (ret < 0) {
834 		atomic_inc(&ch->sq_wr_avail);
835 		atomic_dec(&ch->req_lim);
836 	}
837 	return ret;
838 }
839 
840 /**
841  * srpt_get_desc_tbl() - Parse the data descriptors of an SRP_CMD request.
842  * @ioctx: Pointer to the I/O context associated with the request.
843  * @srp_cmd: Pointer to the SRP_CMD request data.
844  * @dir: Pointer to the variable to which the transfer direction will be
845  *   written.
846  * @data_len: Pointer to the variable to which the total data length of all
847  *   descriptors in the SRP_CMD request will be written.
848  *
849  * This function initializes ioctx->nrbuf and ioctx->r_bufs.
850  *
851  * Returns -EINVAL when the SRP_CMD request contains inconsistent descriptors;
852  * -ENOMEM when memory allocation fails and zero upon success.
853  */
854 static int srpt_get_desc_tbl(struct srpt_send_ioctx *ioctx,
855 			     struct srp_cmd *srp_cmd,
856 			     enum dma_data_direction *dir, u64 *data_len)
857 {
858 	struct srp_indirect_buf *idb;
859 	struct srp_direct_buf *db;
860 	unsigned add_cdb_offset;
861 	int ret;
862 
863 	/*
864 	 * The pointer computations below will only be compiled correctly
865 	 * if srp_cmd::add_data is declared as s8*, u8*, s8[] or u8[], so check
866 	 * whether srp_cmd::add_data has been declared as a byte pointer.
867 	 */
868 	BUILD_BUG_ON(!__same_type(srp_cmd->add_data[0], (s8)0)
869 		     && !__same_type(srp_cmd->add_data[0], (u8)0));
870 
871 	BUG_ON(!dir);
872 	BUG_ON(!data_len);
873 
874 	ret = 0;
875 	*data_len = 0;
876 
877 	/*
878 	 * The lower four bits of the buffer format field contain the DATA-IN
879 	 * buffer descriptor format, and the highest four bits contain the
880 	 * DATA-OUT buffer descriptor format.
881 	 */
882 	*dir = DMA_NONE;
883 	if (srp_cmd->buf_fmt & 0xf)
884 		/* DATA-IN: transfer data from target to initiator (read). */
885 		*dir = DMA_FROM_DEVICE;
886 	else if (srp_cmd->buf_fmt >> 4)
887 		/* DATA-OUT: transfer data from initiator to target (write). */
888 		*dir = DMA_TO_DEVICE;
889 
890 	/*
891 	 * According to the SRP spec, the lower two bits of the 'ADDITIONAL
892 	 * CDB LENGTH' field are reserved and the size in bytes of this field
893 	 * is four times the value specified in bits 3..7. Hence the "& ~3".
894 	 */
895 	add_cdb_offset = srp_cmd->add_cdb_len & ~3;
896 	if (((srp_cmd->buf_fmt & 0xf) == SRP_DATA_DESC_DIRECT) ||
897 	    ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_DIRECT)) {
898 		ioctx->n_rbuf = 1;
899 		ioctx->rbufs = &ioctx->single_rbuf;
900 
901 		db = (struct srp_direct_buf *)(srp_cmd->add_data
902 					       + add_cdb_offset);
903 		memcpy(ioctx->rbufs, db, sizeof *db);
904 		*data_len = be32_to_cpu(db->len);
905 	} else if (((srp_cmd->buf_fmt & 0xf) == SRP_DATA_DESC_INDIRECT) ||
906 		   ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_INDIRECT)) {
907 		idb = (struct srp_indirect_buf *)(srp_cmd->add_data
908 						  + add_cdb_offset);
909 
910 		ioctx->n_rbuf = be32_to_cpu(idb->table_desc.len) / sizeof *db;
911 
912 		if (ioctx->n_rbuf >
913 		    (srp_cmd->data_out_desc_cnt + srp_cmd->data_in_desc_cnt)) {
914 			printk(KERN_ERR "received unsupported SRP_CMD request"
915 			       " type (%u out + %u in != %u / %zu)\n",
916 			       srp_cmd->data_out_desc_cnt,
917 			       srp_cmd->data_in_desc_cnt,
918 			       be32_to_cpu(idb->table_desc.len),
919 			       sizeof(*db));
920 			ioctx->n_rbuf = 0;
921 			ret = -EINVAL;
922 			goto out;
923 		}
924 
925 		if (ioctx->n_rbuf == 1)
926 			ioctx->rbufs = &ioctx->single_rbuf;
927 		else {
928 			ioctx->rbufs =
929 				kmalloc(ioctx->n_rbuf * sizeof *db, GFP_ATOMIC);
930 			if (!ioctx->rbufs) {
931 				ioctx->n_rbuf = 0;
932 				ret = -ENOMEM;
933 				goto out;
934 			}
935 		}
936 
937 		db = idb->desc_list;
938 		memcpy(ioctx->rbufs, db, ioctx->n_rbuf * sizeof *db);
939 		*data_len = be32_to_cpu(idb->len);
940 	}
941 out:
942 	return ret;
943 }
944 
945 /**
946  * srpt_init_ch_qp() - Initialize queue pair attributes.
947  *
948  * Initialized the attributes of queue pair 'qp' by allowing local write,
949  * remote read and remote write. Also transitions 'qp' to state IB_QPS_INIT.
950  */
951 static int srpt_init_ch_qp(struct srpt_rdma_ch *ch, struct ib_qp *qp)
952 {
953 	struct ib_qp_attr *attr;
954 	int ret;
955 
956 	attr = kzalloc(sizeof *attr, GFP_KERNEL);
957 	if (!attr)
958 		return -ENOMEM;
959 
960 	attr->qp_state = IB_QPS_INIT;
961 	attr->qp_access_flags = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_READ |
962 	    IB_ACCESS_REMOTE_WRITE;
963 	attr->port_num = ch->sport->port;
964 	attr->pkey_index = 0;
965 
966 	ret = ib_modify_qp(qp, attr,
967 			   IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PORT |
968 			   IB_QP_PKEY_INDEX);
969 
970 	kfree(attr);
971 	return ret;
972 }
973 
974 /**
975  * srpt_ch_qp_rtr() - Change the state of a channel to 'ready to receive' (RTR).
976  * @ch: channel of the queue pair.
977  * @qp: queue pair to change the state of.
978  *
979  * Returns zero upon success and a negative value upon failure.
980  *
981  * Note: currently a struct ib_qp_attr takes 136 bytes on a 64-bit system.
982  * If this structure ever becomes larger, it might be necessary to allocate
983  * it dynamically instead of on the stack.
984  */
985 static int srpt_ch_qp_rtr(struct srpt_rdma_ch *ch, struct ib_qp *qp)
986 {
987 	struct ib_qp_attr qp_attr;
988 	int attr_mask;
989 	int ret;
990 
991 	qp_attr.qp_state = IB_QPS_RTR;
992 	ret = ib_cm_init_qp_attr(ch->cm_id, &qp_attr, &attr_mask);
993 	if (ret)
994 		goto out;
995 
996 	qp_attr.max_dest_rd_atomic = 4;
997 
998 	ret = ib_modify_qp(qp, &qp_attr, attr_mask);
999 
1000 out:
1001 	return ret;
1002 }
1003 
1004 /**
1005  * srpt_ch_qp_rts() - Change the state of a channel to 'ready to send' (RTS).
1006  * @ch: channel of the queue pair.
1007  * @qp: queue pair to change the state of.
1008  *
1009  * Returns zero upon success and a negative value upon failure.
1010  *
1011  * Note: currently a struct ib_qp_attr takes 136 bytes on a 64-bit system.
1012  * If this structure ever becomes larger, it might be necessary to allocate
1013  * it dynamically instead of on the stack.
1014  */
1015 static int srpt_ch_qp_rts(struct srpt_rdma_ch *ch, struct ib_qp *qp)
1016 {
1017 	struct ib_qp_attr qp_attr;
1018 	int attr_mask;
1019 	int ret;
1020 
1021 	qp_attr.qp_state = IB_QPS_RTS;
1022 	ret = ib_cm_init_qp_attr(ch->cm_id, &qp_attr, &attr_mask);
1023 	if (ret)
1024 		goto out;
1025 
1026 	qp_attr.max_rd_atomic = 4;
1027 
1028 	ret = ib_modify_qp(qp, &qp_attr, attr_mask);
1029 
1030 out:
1031 	return ret;
1032 }
1033 
1034 /**
1035  * srpt_ch_qp_err() - Set the channel queue pair state to 'error'.
1036  */
1037 static int srpt_ch_qp_err(struct srpt_rdma_ch *ch)
1038 {
1039 	struct ib_qp_attr qp_attr;
1040 
1041 	qp_attr.qp_state = IB_QPS_ERR;
1042 	return ib_modify_qp(ch->qp, &qp_attr, IB_QP_STATE);
1043 }
1044 
1045 /**
1046  * srpt_unmap_sg_to_ib_sge() - Unmap an IB SGE list.
1047  */
1048 static void srpt_unmap_sg_to_ib_sge(struct srpt_rdma_ch *ch,
1049 				    struct srpt_send_ioctx *ioctx)
1050 {
1051 	struct scatterlist *sg;
1052 	enum dma_data_direction dir;
1053 
1054 	BUG_ON(!ch);
1055 	BUG_ON(!ioctx);
1056 	BUG_ON(ioctx->n_rdma && !ioctx->rdma_ius);
1057 
1058 	while (ioctx->n_rdma)
1059 		kfree(ioctx->rdma_ius[--ioctx->n_rdma].sge);
1060 
1061 	kfree(ioctx->rdma_ius);
1062 	ioctx->rdma_ius = NULL;
1063 
1064 	if (ioctx->mapped_sg_count) {
1065 		sg = ioctx->sg;
1066 		WARN_ON(!sg);
1067 		dir = ioctx->cmd.data_direction;
1068 		BUG_ON(dir == DMA_NONE);
1069 		ib_dma_unmap_sg(ch->sport->sdev->device, sg, ioctx->sg_cnt,
1070 				opposite_dma_dir(dir));
1071 		ioctx->mapped_sg_count = 0;
1072 	}
1073 }
1074 
1075 /**
1076  * srpt_map_sg_to_ib_sge() - Map an SG list to an IB SGE list.
1077  */
1078 static int srpt_map_sg_to_ib_sge(struct srpt_rdma_ch *ch,
1079 				 struct srpt_send_ioctx *ioctx)
1080 {
1081 	struct se_cmd *cmd;
1082 	struct scatterlist *sg, *sg_orig;
1083 	int sg_cnt;
1084 	enum dma_data_direction dir;
1085 	struct rdma_iu *riu;
1086 	struct srp_direct_buf *db;
1087 	dma_addr_t dma_addr;
1088 	struct ib_sge *sge;
1089 	u64 raddr;
1090 	u32 rsize;
1091 	u32 tsize;
1092 	u32 dma_len;
1093 	int count, nrdma;
1094 	int i, j, k;
1095 
1096 	BUG_ON(!ch);
1097 	BUG_ON(!ioctx);
1098 	cmd = &ioctx->cmd;
1099 	dir = cmd->data_direction;
1100 	BUG_ON(dir == DMA_NONE);
1101 
1102 	ioctx->sg = sg = sg_orig = cmd->t_data_sg;
1103 	ioctx->sg_cnt = sg_cnt = cmd->t_data_nents;
1104 
1105 	count = ib_dma_map_sg(ch->sport->sdev->device, sg, sg_cnt,
1106 			      opposite_dma_dir(dir));
1107 	if (unlikely(!count))
1108 		return -EAGAIN;
1109 
1110 	ioctx->mapped_sg_count = count;
1111 
1112 	if (ioctx->rdma_ius && ioctx->n_rdma_ius)
1113 		nrdma = ioctx->n_rdma_ius;
1114 	else {
1115 		nrdma = (count + SRPT_DEF_SG_PER_WQE - 1) / SRPT_DEF_SG_PER_WQE
1116 			+ ioctx->n_rbuf;
1117 
1118 		ioctx->rdma_ius = kzalloc(nrdma * sizeof *riu, GFP_KERNEL);
1119 		if (!ioctx->rdma_ius)
1120 			goto free_mem;
1121 
1122 		ioctx->n_rdma_ius = nrdma;
1123 	}
1124 
1125 	db = ioctx->rbufs;
1126 	tsize = cmd->data_length;
1127 	dma_len = sg_dma_len(&sg[0]);
1128 	riu = ioctx->rdma_ius;
1129 
1130 	/*
1131 	 * For each remote desc - calculate the #ib_sge.
1132 	 * If #ib_sge < SRPT_DEF_SG_PER_WQE per rdma operation then
1133 	 *      each remote desc rdma_iu is required a rdma wr;
1134 	 * else
1135 	 *      we need to allocate extra rdma_iu to carry extra #ib_sge in
1136 	 *      another rdma wr
1137 	 */
1138 	for (i = 0, j = 0;
1139 	     j < count && i < ioctx->n_rbuf && tsize > 0; ++i, ++riu, ++db) {
1140 		rsize = be32_to_cpu(db->len);
1141 		raddr = be64_to_cpu(db->va);
1142 		riu->raddr = raddr;
1143 		riu->rkey = be32_to_cpu(db->key);
1144 		riu->sge_cnt = 0;
1145 
1146 		/* calculate how many sge required for this remote_buf */
1147 		while (rsize > 0 && tsize > 0) {
1148 
1149 			if (rsize >= dma_len) {
1150 				tsize -= dma_len;
1151 				rsize -= dma_len;
1152 				raddr += dma_len;
1153 
1154 				if (tsize > 0) {
1155 					++j;
1156 					if (j < count) {
1157 						sg = sg_next(sg);
1158 						dma_len = sg_dma_len(sg);
1159 					}
1160 				}
1161 			} else {
1162 				tsize -= rsize;
1163 				dma_len -= rsize;
1164 				rsize = 0;
1165 			}
1166 
1167 			++riu->sge_cnt;
1168 
1169 			if (rsize > 0 && riu->sge_cnt == SRPT_DEF_SG_PER_WQE) {
1170 				++ioctx->n_rdma;
1171 				riu->sge =
1172 				    kmalloc(riu->sge_cnt * sizeof *riu->sge,
1173 					    GFP_KERNEL);
1174 				if (!riu->sge)
1175 					goto free_mem;
1176 
1177 				++riu;
1178 				riu->sge_cnt = 0;
1179 				riu->raddr = raddr;
1180 				riu->rkey = be32_to_cpu(db->key);
1181 			}
1182 		}
1183 
1184 		++ioctx->n_rdma;
1185 		riu->sge = kmalloc(riu->sge_cnt * sizeof *riu->sge,
1186 				   GFP_KERNEL);
1187 		if (!riu->sge)
1188 			goto free_mem;
1189 	}
1190 
1191 	db = ioctx->rbufs;
1192 	tsize = cmd->data_length;
1193 	riu = ioctx->rdma_ius;
1194 	sg = sg_orig;
1195 	dma_len = sg_dma_len(&sg[0]);
1196 	dma_addr = sg_dma_address(&sg[0]);
1197 
1198 	/* this second loop is really mapped sg_addres to rdma_iu->ib_sge */
1199 	for (i = 0, j = 0;
1200 	     j < count && i < ioctx->n_rbuf && tsize > 0; ++i, ++riu, ++db) {
1201 		rsize = be32_to_cpu(db->len);
1202 		sge = riu->sge;
1203 		k = 0;
1204 
1205 		while (rsize > 0 && tsize > 0) {
1206 			sge->addr = dma_addr;
1207 			sge->lkey = ch->sport->sdev->mr->lkey;
1208 
1209 			if (rsize >= dma_len) {
1210 				sge->length =
1211 					(tsize < dma_len) ? tsize : dma_len;
1212 				tsize -= dma_len;
1213 				rsize -= dma_len;
1214 
1215 				if (tsize > 0) {
1216 					++j;
1217 					if (j < count) {
1218 						sg = sg_next(sg);
1219 						dma_len = sg_dma_len(sg);
1220 						dma_addr = sg_dma_address(sg);
1221 					}
1222 				}
1223 			} else {
1224 				sge->length = (tsize < rsize) ? tsize : rsize;
1225 				tsize -= rsize;
1226 				dma_len -= rsize;
1227 				dma_addr += rsize;
1228 				rsize = 0;
1229 			}
1230 
1231 			++k;
1232 			if (k == riu->sge_cnt && rsize > 0 && tsize > 0) {
1233 				++riu;
1234 				sge = riu->sge;
1235 				k = 0;
1236 			} else if (rsize > 0 && tsize > 0)
1237 				++sge;
1238 		}
1239 	}
1240 
1241 	return 0;
1242 
1243 free_mem:
1244 	srpt_unmap_sg_to_ib_sge(ch, ioctx);
1245 
1246 	return -ENOMEM;
1247 }
1248 
1249 /**
1250  * srpt_get_send_ioctx() - Obtain an I/O context for sending to the initiator.
1251  */
1252 static struct srpt_send_ioctx *srpt_get_send_ioctx(struct srpt_rdma_ch *ch)
1253 {
1254 	struct srpt_send_ioctx *ioctx;
1255 	unsigned long flags;
1256 
1257 	BUG_ON(!ch);
1258 
1259 	ioctx = NULL;
1260 	spin_lock_irqsave(&ch->spinlock, flags);
1261 	if (!list_empty(&ch->free_list)) {
1262 		ioctx = list_first_entry(&ch->free_list,
1263 					 struct srpt_send_ioctx, free_list);
1264 		list_del(&ioctx->free_list);
1265 	}
1266 	spin_unlock_irqrestore(&ch->spinlock, flags);
1267 
1268 	if (!ioctx)
1269 		return ioctx;
1270 
1271 	BUG_ON(ioctx->ch != ch);
1272 	spin_lock_init(&ioctx->spinlock);
1273 	ioctx->state = SRPT_STATE_NEW;
1274 	ioctx->n_rbuf = 0;
1275 	ioctx->rbufs = NULL;
1276 	ioctx->n_rdma = 0;
1277 	ioctx->n_rdma_ius = 0;
1278 	ioctx->rdma_ius = NULL;
1279 	ioctx->mapped_sg_count = 0;
1280 	init_completion(&ioctx->tx_done);
1281 	ioctx->queue_status_only = false;
1282 	/*
1283 	 * transport_init_se_cmd() does not initialize all fields, so do it
1284 	 * here.
1285 	 */
1286 	memset(&ioctx->cmd, 0, sizeof(ioctx->cmd));
1287 	memset(&ioctx->sense_data, 0, sizeof(ioctx->sense_data));
1288 
1289 	return ioctx;
1290 }
1291 
1292 /**
1293  * srpt_abort_cmd() - Abort a SCSI command.
1294  * @ioctx:   I/O context associated with the SCSI command.
1295  * @context: Preferred execution context.
1296  */
1297 static int srpt_abort_cmd(struct srpt_send_ioctx *ioctx)
1298 {
1299 	enum srpt_command_state state;
1300 	unsigned long flags;
1301 
1302 	BUG_ON(!ioctx);
1303 
1304 	/*
1305 	 * If the command is in a state where the target core is waiting for
1306 	 * the ib_srpt driver, change the state to the next state. Changing
1307 	 * the state of the command from SRPT_STATE_NEED_DATA to
1308 	 * SRPT_STATE_DATA_IN ensures that srpt_xmit_response() will call this
1309 	 * function a second time.
1310 	 */
1311 
1312 	spin_lock_irqsave(&ioctx->spinlock, flags);
1313 	state = ioctx->state;
1314 	switch (state) {
1315 	case SRPT_STATE_NEED_DATA:
1316 		ioctx->state = SRPT_STATE_DATA_IN;
1317 		break;
1318 	case SRPT_STATE_DATA_IN:
1319 	case SRPT_STATE_CMD_RSP_SENT:
1320 	case SRPT_STATE_MGMT_RSP_SENT:
1321 		ioctx->state = SRPT_STATE_DONE;
1322 		break;
1323 	default:
1324 		break;
1325 	}
1326 	spin_unlock_irqrestore(&ioctx->spinlock, flags);
1327 
1328 	if (state == SRPT_STATE_DONE) {
1329 		struct srpt_rdma_ch *ch = ioctx->ch;
1330 
1331 		BUG_ON(ch->sess == NULL);
1332 
1333 		target_put_sess_cmd(ch->sess, &ioctx->cmd);
1334 		goto out;
1335 	}
1336 
1337 	pr_debug("Aborting cmd with state %d and tag %lld\n", state,
1338 		 ioctx->tag);
1339 
1340 	switch (state) {
1341 	case SRPT_STATE_NEW:
1342 	case SRPT_STATE_DATA_IN:
1343 	case SRPT_STATE_MGMT:
1344 		/*
1345 		 * Do nothing - defer abort processing until
1346 		 * srpt_queue_response() is invoked.
1347 		 */
1348 		WARN_ON(!transport_check_aborted_status(&ioctx->cmd, false));
1349 		break;
1350 	case SRPT_STATE_NEED_DATA:
1351 		/* DMA_TO_DEVICE (write) - RDMA read error. */
1352 
1353 		/* XXX(hch): this is a horrible layering violation.. */
1354 		spin_lock_irqsave(&ioctx->cmd.t_state_lock, flags);
1355 		ioctx->cmd.transport_state |= CMD_T_LUN_STOP;
1356 		ioctx->cmd.transport_state &= ~CMD_T_ACTIVE;
1357 		spin_unlock_irqrestore(&ioctx->cmd.t_state_lock, flags);
1358 
1359 		complete(&ioctx->cmd.transport_lun_stop_comp);
1360 		break;
1361 	case SRPT_STATE_CMD_RSP_SENT:
1362 		/*
1363 		 * SRP_RSP sending failed or the SRP_RSP send completion has
1364 		 * not been received in time.
1365 		 */
1366 		srpt_unmap_sg_to_ib_sge(ioctx->ch, ioctx);
1367 		spin_lock_irqsave(&ioctx->cmd.t_state_lock, flags);
1368 		ioctx->cmd.transport_state |= CMD_T_LUN_STOP;
1369 		spin_unlock_irqrestore(&ioctx->cmd.t_state_lock, flags);
1370 		target_put_sess_cmd(ioctx->ch->sess, &ioctx->cmd);
1371 		break;
1372 	case SRPT_STATE_MGMT_RSP_SENT:
1373 		srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
1374 		target_put_sess_cmd(ioctx->ch->sess, &ioctx->cmd);
1375 		break;
1376 	default:
1377 		WARN(1, "Unexpected command state (%d)", state);
1378 		break;
1379 	}
1380 
1381 out:
1382 	return state;
1383 }
1384 
1385 /**
1386  * srpt_handle_send_err_comp() - Process an IB_WC_SEND error completion.
1387  */
1388 static void srpt_handle_send_err_comp(struct srpt_rdma_ch *ch, u64 wr_id)
1389 {
1390 	struct srpt_send_ioctx *ioctx;
1391 	enum srpt_command_state state;
1392 	struct se_cmd *cmd;
1393 	u32 index;
1394 
1395 	atomic_inc(&ch->sq_wr_avail);
1396 
1397 	index = idx_from_wr_id(wr_id);
1398 	ioctx = ch->ioctx_ring[index];
1399 	state = srpt_get_cmd_state(ioctx);
1400 	cmd = &ioctx->cmd;
1401 
1402 	WARN_ON(state != SRPT_STATE_CMD_RSP_SENT
1403 		&& state != SRPT_STATE_MGMT_RSP_SENT
1404 		&& state != SRPT_STATE_NEED_DATA
1405 		&& state != SRPT_STATE_DONE);
1406 
1407 	/* If SRP_RSP sending failed, undo the ch->req_lim change. */
1408 	if (state == SRPT_STATE_CMD_RSP_SENT
1409 	    || state == SRPT_STATE_MGMT_RSP_SENT)
1410 		atomic_dec(&ch->req_lim);
1411 
1412 	srpt_abort_cmd(ioctx);
1413 }
1414 
1415 /**
1416  * srpt_handle_send_comp() - Process an IB send completion notification.
1417  */
1418 static void srpt_handle_send_comp(struct srpt_rdma_ch *ch,
1419 				  struct srpt_send_ioctx *ioctx)
1420 {
1421 	enum srpt_command_state state;
1422 
1423 	atomic_inc(&ch->sq_wr_avail);
1424 
1425 	state = srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
1426 
1427 	if (WARN_ON(state != SRPT_STATE_CMD_RSP_SENT
1428 		    && state != SRPT_STATE_MGMT_RSP_SENT
1429 		    && state != SRPT_STATE_DONE))
1430 		pr_debug("state = %d\n", state);
1431 
1432 	if (state != SRPT_STATE_DONE) {
1433 		srpt_unmap_sg_to_ib_sge(ch, ioctx);
1434 		transport_generic_free_cmd(&ioctx->cmd, 0);
1435 	} else {
1436 		printk(KERN_ERR "IB completion has been received too late for"
1437 		       " wr_id = %u.\n", ioctx->ioctx.index);
1438 	}
1439 }
1440 
1441 /**
1442  * srpt_handle_rdma_comp() - Process an IB RDMA completion notification.
1443  *
1444  * XXX: what is now target_execute_cmd used to be asynchronous, and unmapping
1445  * the data that has been transferred via IB RDMA had to be postponed until the
1446  * check_stop_free() callback.  None of this is necessary anymore and needs to
1447  * be cleaned up.
1448  */
1449 static void srpt_handle_rdma_comp(struct srpt_rdma_ch *ch,
1450 				  struct srpt_send_ioctx *ioctx,
1451 				  enum srpt_opcode opcode)
1452 {
1453 	WARN_ON(ioctx->n_rdma <= 0);
1454 	atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
1455 
1456 	if (opcode == SRPT_RDMA_READ_LAST) {
1457 		if (srpt_test_and_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA,
1458 						SRPT_STATE_DATA_IN))
1459 			target_execute_cmd(&ioctx->cmd);
1460 		else
1461 			printk(KERN_ERR "%s[%d]: wrong state = %d\n", __func__,
1462 			       __LINE__, srpt_get_cmd_state(ioctx));
1463 	} else if (opcode == SRPT_RDMA_ABORT) {
1464 		ioctx->rdma_aborted = true;
1465 	} else {
1466 		WARN(true, "unexpected opcode %d\n", opcode);
1467 	}
1468 }
1469 
1470 /**
1471  * srpt_handle_rdma_err_comp() - Process an IB RDMA error completion.
1472  */
1473 static void srpt_handle_rdma_err_comp(struct srpt_rdma_ch *ch,
1474 				      struct srpt_send_ioctx *ioctx,
1475 				      enum srpt_opcode opcode)
1476 {
1477 	struct se_cmd *cmd;
1478 	enum srpt_command_state state;
1479 	unsigned long flags;
1480 
1481 	cmd = &ioctx->cmd;
1482 	state = srpt_get_cmd_state(ioctx);
1483 	switch (opcode) {
1484 	case SRPT_RDMA_READ_LAST:
1485 		if (ioctx->n_rdma <= 0) {
1486 			printk(KERN_ERR "Received invalid RDMA read"
1487 			       " error completion with idx %d\n",
1488 			       ioctx->ioctx.index);
1489 			break;
1490 		}
1491 		atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
1492 		if (state == SRPT_STATE_NEED_DATA)
1493 			srpt_abort_cmd(ioctx);
1494 		else
1495 			printk(KERN_ERR "%s[%d]: wrong state = %d\n",
1496 			       __func__, __LINE__, state);
1497 		break;
1498 	case SRPT_RDMA_WRITE_LAST:
1499 		spin_lock_irqsave(&ioctx->cmd.t_state_lock, flags);
1500 		ioctx->cmd.transport_state |= CMD_T_LUN_STOP;
1501 		spin_unlock_irqrestore(&ioctx->cmd.t_state_lock, flags);
1502 		break;
1503 	default:
1504 		printk(KERN_ERR "%s[%d]: opcode = %u\n", __func__,
1505 		       __LINE__, opcode);
1506 		break;
1507 	}
1508 }
1509 
1510 /**
1511  * srpt_build_cmd_rsp() - Build an SRP_RSP response.
1512  * @ch: RDMA channel through which the request has been received.
1513  * @ioctx: I/O context associated with the SRP_CMD request. The response will
1514  *   be built in the buffer ioctx->buf points at and hence this function will
1515  *   overwrite the request data.
1516  * @tag: tag of the request for which this response is being generated.
1517  * @status: value for the STATUS field of the SRP_RSP information unit.
1518  *
1519  * Returns the size in bytes of the SRP_RSP response.
1520  *
1521  * An SRP_RSP response contains a SCSI status or service response. See also
1522  * section 6.9 in the SRP r16a document for the format of an SRP_RSP
1523  * response. See also SPC-2 for more information about sense data.
1524  */
1525 static int srpt_build_cmd_rsp(struct srpt_rdma_ch *ch,
1526 			      struct srpt_send_ioctx *ioctx, u64 tag,
1527 			      int status)
1528 {
1529 	struct srp_rsp *srp_rsp;
1530 	const u8 *sense_data;
1531 	int sense_data_len, max_sense_len;
1532 
1533 	/*
1534 	 * The lowest bit of all SAM-3 status codes is zero (see also
1535 	 * paragraph 5.3 in SAM-3).
1536 	 */
1537 	WARN_ON(status & 1);
1538 
1539 	srp_rsp = ioctx->ioctx.buf;
1540 	BUG_ON(!srp_rsp);
1541 
1542 	sense_data = ioctx->sense_data;
1543 	sense_data_len = ioctx->cmd.scsi_sense_length;
1544 	WARN_ON(sense_data_len > sizeof(ioctx->sense_data));
1545 
1546 	memset(srp_rsp, 0, sizeof *srp_rsp);
1547 	srp_rsp->opcode = SRP_RSP;
1548 	srp_rsp->req_lim_delta =
1549 		__constant_cpu_to_be32(1 + atomic_xchg(&ch->req_lim_delta, 0));
1550 	srp_rsp->tag = tag;
1551 	srp_rsp->status = status;
1552 
1553 	if (sense_data_len) {
1554 		BUILD_BUG_ON(MIN_MAX_RSP_SIZE <= sizeof(*srp_rsp));
1555 		max_sense_len = ch->max_ti_iu_len - sizeof(*srp_rsp);
1556 		if (sense_data_len > max_sense_len) {
1557 			printk(KERN_WARNING "truncated sense data from %d to %d"
1558 			       " bytes\n", sense_data_len, max_sense_len);
1559 			sense_data_len = max_sense_len;
1560 		}
1561 
1562 		srp_rsp->flags |= SRP_RSP_FLAG_SNSVALID;
1563 		srp_rsp->sense_data_len = cpu_to_be32(sense_data_len);
1564 		memcpy(srp_rsp + 1, sense_data, sense_data_len);
1565 	}
1566 
1567 	return sizeof(*srp_rsp) + sense_data_len;
1568 }
1569 
1570 /**
1571  * srpt_build_tskmgmt_rsp() - Build a task management response.
1572  * @ch:       RDMA channel through which the request has been received.
1573  * @ioctx:    I/O context in which the SRP_RSP response will be built.
1574  * @rsp_code: RSP_CODE that will be stored in the response.
1575  * @tag:      Tag of the request for which this response is being generated.
1576  *
1577  * Returns the size in bytes of the SRP_RSP response.
1578  *
1579  * An SRP_RSP response contains a SCSI status or service response. See also
1580  * section 6.9 in the SRP r16a document for the format of an SRP_RSP
1581  * response.
1582  */
1583 static int srpt_build_tskmgmt_rsp(struct srpt_rdma_ch *ch,
1584 				  struct srpt_send_ioctx *ioctx,
1585 				  u8 rsp_code, u64 tag)
1586 {
1587 	struct srp_rsp *srp_rsp;
1588 	int resp_data_len;
1589 	int resp_len;
1590 
1591 	resp_data_len = 4;
1592 	resp_len = sizeof(*srp_rsp) + resp_data_len;
1593 
1594 	srp_rsp = ioctx->ioctx.buf;
1595 	BUG_ON(!srp_rsp);
1596 	memset(srp_rsp, 0, sizeof *srp_rsp);
1597 
1598 	srp_rsp->opcode = SRP_RSP;
1599 	srp_rsp->req_lim_delta = __constant_cpu_to_be32(1
1600 				    + atomic_xchg(&ch->req_lim_delta, 0));
1601 	srp_rsp->tag = tag;
1602 
1603 	srp_rsp->flags |= SRP_RSP_FLAG_RSPVALID;
1604 	srp_rsp->resp_data_len = cpu_to_be32(resp_data_len);
1605 	srp_rsp->data[3] = rsp_code;
1606 
1607 	return resp_len;
1608 }
1609 
1610 #define NO_SUCH_LUN ((uint64_t)-1LL)
1611 
1612 /*
1613  * SCSI LUN addressing method. See also SAM-2 and the section about
1614  * eight byte LUNs.
1615  */
1616 enum scsi_lun_addr_method {
1617 	SCSI_LUN_ADDR_METHOD_PERIPHERAL   = 0,
1618 	SCSI_LUN_ADDR_METHOD_FLAT         = 1,
1619 	SCSI_LUN_ADDR_METHOD_LUN          = 2,
1620 	SCSI_LUN_ADDR_METHOD_EXTENDED_LUN = 3,
1621 };
1622 
1623 /*
1624  * srpt_unpack_lun() - Convert from network LUN to linear LUN.
1625  *
1626  * Convert an 2-byte, 4-byte, 6-byte or 8-byte LUN structure in network byte
1627  * order (big endian) to a linear LUN. Supports three LUN addressing methods:
1628  * peripheral, flat and logical unit. See also SAM-2, section 4.9.4 (page 40).
1629  */
1630 static uint64_t srpt_unpack_lun(const uint8_t *lun, int len)
1631 {
1632 	uint64_t res = NO_SUCH_LUN;
1633 	int addressing_method;
1634 
1635 	if (unlikely(len < 2)) {
1636 		printk(KERN_ERR "Illegal LUN length %d, expected 2 bytes or "
1637 		       "more", len);
1638 		goto out;
1639 	}
1640 
1641 	switch (len) {
1642 	case 8:
1643 		if ((*((__be64 *)lun) &
1644 		     __constant_cpu_to_be64(0x0000FFFFFFFFFFFFLL)) != 0)
1645 			goto out_err;
1646 		break;
1647 	case 4:
1648 		if (*((__be16 *)&lun[2]) != 0)
1649 			goto out_err;
1650 		break;
1651 	case 6:
1652 		if (*((__be32 *)&lun[2]) != 0)
1653 			goto out_err;
1654 		break;
1655 	case 2:
1656 		break;
1657 	default:
1658 		goto out_err;
1659 	}
1660 
1661 	addressing_method = (*lun) >> 6; /* highest two bits of byte 0 */
1662 	switch (addressing_method) {
1663 	case SCSI_LUN_ADDR_METHOD_PERIPHERAL:
1664 	case SCSI_LUN_ADDR_METHOD_FLAT:
1665 	case SCSI_LUN_ADDR_METHOD_LUN:
1666 		res = *(lun + 1) | (((*lun) & 0x3f) << 8);
1667 		break;
1668 
1669 	case SCSI_LUN_ADDR_METHOD_EXTENDED_LUN:
1670 	default:
1671 		printk(KERN_ERR "Unimplemented LUN addressing method %u",
1672 		       addressing_method);
1673 		break;
1674 	}
1675 
1676 out:
1677 	return res;
1678 
1679 out_err:
1680 	printk(KERN_ERR "Support for multi-level LUNs has not yet been"
1681 	       " implemented");
1682 	goto out;
1683 }
1684 
1685 static int srpt_check_stop_free(struct se_cmd *cmd)
1686 {
1687 	struct srpt_send_ioctx *ioctx = container_of(cmd,
1688 				struct srpt_send_ioctx, cmd);
1689 
1690 	return target_put_sess_cmd(ioctx->ch->sess, &ioctx->cmd);
1691 }
1692 
1693 /**
1694  * srpt_handle_cmd() - Process SRP_CMD.
1695  */
1696 static int srpt_handle_cmd(struct srpt_rdma_ch *ch,
1697 			   struct srpt_recv_ioctx *recv_ioctx,
1698 			   struct srpt_send_ioctx *send_ioctx)
1699 {
1700 	struct se_cmd *cmd;
1701 	struct srp_cmd *srp_cmd;
1702 	uint64_t unpacked_lun;
1703 	u64 data_len;
1704 	enum dma_data_direction dir;
1705 	sense_reason_t ret;
1706 	int rc;
1707 
1708 	BUG_ON(!send_ioctx);
1709 
1710 	srp_cmd = recv_ioctx->ioctx.buf;
1711 	cmd = &send_ioctx->cmd;
1712 	send_ioctx->tag = srp_cmd->tag;
1713 
1714 	switch (srp_cmd->task_attr) {
1715 	case SRP_CMD_SIMPLE_Q:
1716 		cmd->sam_task_attr = MSG_SIMPLE_TAG;
1717 		break;
1718 	case SRP_CMD_ORDERED_Q:
1719 	default:
1720 		cmd->sam_task_attr = MSG_ORDERED_TAG;
1721 		break;
1722 	case SRP_CMD_HEAD_OF_Q:
1723 		cmd->sam_task_attr = MSG_HEAD_TAG;
1724 		break;
1725 	case SRP_CMD_ACA:
1726 		cmd->sam_task_attr = MSG_ACA_TAG;
1727 		break;
1728 	}
1729 
1730 	if (srpt_get_desc_tbl(send_ioctx, srp_cmd, &dir, &data_len)) {
1731 		printk(KERN_ERR "0x%llx: parsing SRP descriptor table failed.\n",
1732 		       srp_cmd->tag);
1733 		ret = TCM_INVALID_CDB_FIELD;
1734 		goto send_sense;
1735 	}
1736 
1737 	unpacked_lun = srpt_unpack_lun((uint8_t *)&srp_cmd->lun,
1738 				       sizeof(srp_cmd->lun));
1739 	rc = target_submit_cmd(cmd, ch->sess, srp_cmd->cdb,
1740 			&send_ioctx->sense_data[0], unpacked_lun, data_len,
1741 			MSG_SIMPLE_TAG, dir, TARGET_SCF_ACK_KREF);
1742 	if (rc != 0) {
1743 		ret = TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
1744 		goto send_sense;
1745 	}
1746 	return 0;
1747 
1748 send_sense:
1749 	transport_send_check_condition_and_sense(cmd, ret, 0);
1750 	return -1;
1751 }
1752 
1753 /**
1754  * srpt_rx_mgmt_fn_tag() - Process a task management function by tag.
1755  * @ch: RDMA channel of the task management request.
1756  * @fn: Task management function to perform.
1757  * @req_tag: Tag of the SRP task management request.
1758  * @mgmt_ioctx: I/O context of the task management request.
1759  *
1760  * Returns zero if the target core will process the task management
1761  * request asynchronously.
1762  *
1763  * Note: It is assumed that the initiator serializes tag-based task management
1764  * requests.
1765  */
1766 static int srpt_rx_mgmt_fn_tag(struct srpt_send_ioctx *ioctx, u64 tag)
1767 {
1768 	struct srpt_device *sdev;
1769 	struct srpt_rdma_ch *ch;
1770 	struct srpt_send_ioctx *target;
1771 	int ret, i;
1772 
1773 	ret = -EINVAL;
1774 	ch = ioctx->ch;
1775 	BUG_ON(!ch);
1776 	BUG_ON(!ch->sport);
1777 	sdev = ch->sport->sdev;
1778 	BUG_ON(!sdev);
1779 	spin_lock_irq(&sdev->spinlock);
1780 	for (i = 0; i < ch->rq_size; ++i) {
1781 		target = ch->ioctx_ring[i];
1782 		if (target->cmd.se_lun == ioctx->cmd.se_lun &&
1783 		    target->tag == tag &&
1784 		    srpt_get_cmd_state(target) != SRPT_STATE_DONE) {
1785 			ret = 0;
1786 			/* now let the target core abort &target->cmd; */
1787 			break;
1788 		}
1789 	}
1790 	spin_unlock_irq(&sdev->spinlock);
1791 	return ret;
1792 }
1793 
1794 static int srp_tmr_to_tcm(int fn)
1795 {
1796 	switch (fn) {
1797 	case SRP_TSK_ABORT_TASK:
1798 		return TMR_ABORT_TASK;
1799 	case SRP_TSK_ABORT_TASK_SET:
1800 		return TMR_ABORT_TASK_SET;
1801 	case SRP_TSK_CLEAR_TASK_SET:
1802 		return TMR_CLEAR_TASK_SET;
1803 	case SRP_TSK_LUN_RESET:
1804 		return TMR_LUN_RESET;
1805 	case SRP_TSK_CLEAR_ACA:
1806 		return TMR_CLEAR_ACA;
1807 	default:
1808 		return -1;
1809 	}
1810 }
1811 
1812 /**
1813  * srpt_handle_tsk_mgmt() - Process an SRP_TSK_MGMT information unit.
1814  *
1815  * Returns 0 if and only if the request will be processed by the target core.
1816  *
1817  * For more information about SRP_TSK_MGMT information units, see also section
1818  * 6.7 in the SRP r16a document.
1819  */
1820 static void srpt_handle_tsk_mgmt(struct srpt_rdma_ch *ch,
1821 				 struct srpt_recv_ioctx *recv_ioctx,
1822 				 struct srpt_send_ioctx *send_ioctx)
1823 {
1824 	struct srp_tsk_mgmt *srp_tsk;
1825 	struct se_cmd *cmd;
1826 	struct se_session *sess = ch->sess;
1827 	uint64_t unpacked_lun;
1828 	uint32_t tag = 0;
1829 	int tcm_tmr;
1830 	int rc;
1831 
1832 	BUG_ON(!send_ioctx);
1833 
1834 	srp_tsk = recv_ioctx->ioctx.buf;
1835 	cmd = &send_ioctx->cmd;
1836 
1837 	pr_debug("recv tsk_mgmt fn %d for task_tag %lld and cmd tag %lld"
1838 		 " cm_id %p sess %p\n", srp_tsk->tsk_mgmt_func,
1839 		 srp_tsk->task_tag, srp_tsk->tag, ch->cm_id, ch->sess);
1840 
1841 	srpt_set_cmd_state(send_ioctx, SRPT_STATE_MGMT);
1842 	send_ioctx->tag = srp_tsk->tag;
1843 	tcm_tmr = srp_tmr_to_tcm(srp_tsk->tsk_mgmt_func);
1844 	if (tcm_tmr < 0) {
1845 		send_ioctx->cmd.se_tmr_req->response =
1846 			TMR_TASK_MGMT_FUNCTION_NOT_SUPPORTED;
1847 		goto fail;
1848 	}
1849 	unpacked_lun = srpt_unpack_lun((uint8_t *)&srp_tsk->lun,
1850 				       sizeof(srp_tsk->lun));
1851 
1852 	if (srp_tsk->tsk_mgmt_func == SRP_TSK_ABORT_TASK) {
1853 		rc = srpt_rx_mgmt_fn_tag(send_ioctx, srp_tsk->task_tag);
1854 		if (rc < 0) {
1855 			send_ioctx->cmd.se_tmr_req->response =
1856 					TMR_TASK_DOES_NOT_EXIST;
1857 			goto fail;
1858 		}
1859 		tag = srp_tsk->task_tag;
1860 	}
1861 	rc = target_submit_tmr(&send_ioctx->cmd, sess, NULL, unpacked_lun,
1862 				srp_tsk, tcm_tmr, GFP_KERNEL, tag,
1863 				TARGET_SCF_ACK_KREF);
1864 	if (rc != 0) {
1865 		send_ioctx->cmd.se_tmr_req->response = TMR_FUNCTION_REJECTED;
1866 		goto fail;
1867 	}
1868 	return;
1869 fail:
1870 	transport_send_check_condition_and_sense(cmd, 0, 0); // XXX:
1871 }
1872 
1873 /**
1874  * srpt_handle_new_iu() - Process a newly received information unit.
1875  * @ch:    RDMA channel through which the information unit has been received.
1876  * @ioctx: SRPT I/O context associated with the information unit.
1877  */
1878 static void srpt_handle_new_iu(struct srpt_rdma_ch *ch,
1879 			       struct srpt_recv_ioctx *recv_ioctx,
1880 			       struct srpt_send_ioctx *send_ioctx)
1881 {
1882 	struct srp_cmd *srp_cmd;
1883 	enum rdma_ch_state ch_state;
1884 
1885 	BUG_ON(!ch);
1886 	BUG_ON(!recv_ioctx);
1887 
1888 	ib_dma_sync_single_for_cpu(ch->sport->sdev->device,
1889 				   recv_ioctx->ioctx.dma, srp_max_req_size,
1890 				   DMA_FROM_DEVICE);
1891 
1892 	ch_state = srpt_get_ch_state(ch);
1893 	if (unlikely(ch_state == CH_CONNECTING)) {
1894 		list_add_tail(&recv_ioctx->wait_list, &ch->cmd_wait_list);
1895 		goto out;
1896 	}
1897 
1898 	if (unlikely(ch_state != CH_LIVE))
1899 		goto out;
1900 
1901 	srp_cmd = recv_ioctx->ioctx.buf;
1902 	if (srp_cmd->opcode == SRP_CMD || srp_cmd->opcode == SRP_TSK_MGMT) {
1903 		if (!send_ioctx)
1904 			send_ioctx = srpt_get_send_ioctx(ch);
1905 		if (unlikely(!send_ioctx)) {
1906 			list_add_tail(&recv_ioctx->wait_list,
1907 				      &ch->cmd_wait_list);
1908 			goto out;
1909 		}
1910 	}
1911 
1912 	switch (srp_cmd->opcode) {
1913 	case SRP_CMD:
1914 		srpt_handle_cmd(ch, recv_ioctx, send_ioctx);
1915 		break;
1916 	case SRP_TSK_MGMT:
1917 		srpt_handle_tsk_mgmt(ch, recv_ioctx, send_ioctx);
1918 		break;
1919 	case SRP_I_LOGOUT:
1920 		printk(KERN_ERR "Not yet implemented: SRP_I_LOGOUT\n");
1921 		break;
1922 	case SRP_CRED_RSP:
1923 		pr_debug("received SRP_CRED_RSP\n");
1924 		break;
1925 	case SRP_AER_RSP:
1926 		pr_debug("received SRP_AER_RSP\n");
1927 		break;
1928 	case SRP_RSP:
1929 		printk(KERN_ERR "Received SRP_RSP\n");
1930 		break;
1931 	default:
1932 		printk(KERN_ERR "received IU with unknown opcode 0x%x\n",
1933 		       srp_cmd->opcode);
1934 		break;
1935 	}
1936 
1937 	srpt_post_recv(ch->sport->sdev, recv_ioctx);
1938 out:
1939 	return;
1940 }
1941 
1942 static void srpt_process_rcv_completion(struct ib_cq *cq,
1943 					struct srpt_rdma_ch *ch,
1944 					struct ib_wc *wc)
1945 {
1946 	struct srpt_device *sdev = ch->sport->sdev;
1947 	struct srpt_recv_ioctx *ioctx;
1948 	u32 index;
1949 
1950 	index = idx_from_wr_id(wc->wr_id);
1951 	if (wc->status == IB_WC_SUCCESS) {
1952 		int req_lim;
1953 
1954 		req_lim = atomic_dec_return(&ch->req_lim);
1955 		if (unlikely(req_lim < 0))
1956 			printk(KERN_ERR "req_lim = %d < 0\n", req_lim);
1957 		ioctx = sdev->ioctx_ring[index];
1958 		srpt_handle_new_iu(ch, ioctx, NULL);
1959 	} else {
1960 		printk(KERN_INFO "receiving failed for idx %u with status %d\n",
1961 		       index, wc->status);
1962 	}
1963 }
1964 
1965 /**
1966  * srpt_process_send_completion() - Process an IB send completion.
1967  *
1968  * Note: Although this has not yet been observed during tests, at least in
1969  * theory it is possible that the srpt_get_send_ioctx() call invoked by
1970  * srpt_handle_new_iu() fails. This is possible because the req_lim_delta
1971  * value in each response is set to one, and it is possible that this response
1972  * makes the initiator send a new request before the send completion for that
1973  * response has been processed. This could e.g. happen if the call to
1974  * srpt_put_send_iotcx() is delayed because of a higher priority interrupt or
1975  * if IB retransmission causes generation of the send completion to be
1976  * delayed. Incoming information units for which srpt_get_send_ioctx() fails
1977  * are queued on cmd_wait_list. The code below processes these delayed
1978  * requests one at a time.
1979  */
1980 static void srpt_process_send_completion(struct ib_cq *cq,
1981 					 struct srpt_rdma_ch *ch,
1982 					 struct ib_wc *wc)
1983 {
1984 	struct srpt_send_ioctx *send_ioctx;
1985 	uint32_t index;
1986 	enum srpt_opcode opcode;
1987 
1988 	index = idx_from_wr_id(wc->wr_id);
1989 	opcode = opcode_from_wr_id(wc->wr_id);
1990 	send_ioctx = ch->ioctx_ring[index];
1991 	if (wc->status == IB_WC_SUCCESS) {
1992 		if (opcode == SRPT_SEND)
1993 			srpt_handle_send_comp(ch, send_ioctx);
1994 		else {
1995 			WARN_ON(opcode != SRPT_RDMA_ABORT &&
1996 				wc->opcode != IB_WC_RDMA_READ);
1997 			srpt_handle_rdma_comp(ch, send_ioctx, opcode);
1998 		}
1999 	} else {
2000 		if (opcode == SRPT_SEND) {
2001 			printk(KERN_INFO "sending response for idx %u failed"
2002 			       " with status %d\n", index, wc->status);
2003 			srpt_handle_send_err_comp(ch, wc->wr_id);
2004 		} else if (opcode != SRPT_RDMA_MID) {
2005 			printk(KERN_INFO "RDMA t %d for idx %u failed with"
2006 				" status %d", opcode, index, wc->status);
2007 			srpt_handle_rdma_err_comp(ch, send_ioctx, opcode);
2008 		}
2009 	}
2010 
2011 	while (unlikely(opcode == SRPT_SEND
2012 			&& !list_empty(&ch->cmd_wait_list)
2013 			&& srpt_get_ch_state(ch) == CH_LIVE
2014 			&& (send_ioctx = srpt_get_send_ioctx(ch)) != NULL)) {
2015 		struct srpt_recv_ioctx *recv_ioctx;
2016 
2017 		recv_ioctx = list_first_entry(&ch->cmd_wait_list,
2018 					      struct srpt_recv_ioctx,
2019 					      wait_list);
2020 		list_del(&recv_ioctx->wait_list);
2021 		srpt_handle_new_iu(ch, recv_ioctx, send_ioctx);
2022 	}
2023 }
2024 
2025 static void srpt_process_completion(struct ib_cq *cq, struct srpt_rdma_ch *ch)
2026 {
2027 	struct ib_wc *const wc = ch->wc;
2028 	int i, n;
2029 
2030 	WARN_ON(cq != ch->cq);
2031 
2032 	ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
2033 	while ((n = ib_poll_cq(cq, ARRAY_SIZE(ch->wc), wc)) > 0) {
2034 		for (i = 0; i < n; i++) {
2035 			if (opcode_from_wr_id(wc[i].wr_id) == SRPT_RECV)
2036 				srpt_process_rcv_completion(cq, ch, &wc[i]);
2037 			else
2038 				srpt_process_send_completion(cq, ch, &wc[i]);
2039 		}
2040 	}
2041 }
2042 
2043 /**
2044  * srpt_completion() - IB completion queue callback function.
2045  *
2046  * Notes:
2047  * - It is guaranteed that a completion handler will never be invoked
2048  *   concurrently on two different CPUs for the same completion queue. See also
2049  *   Documentation/infiniband/core_locking.txt and the implementation of
2050  *   handle_edge_irq() in kernel/irq/chip.c.
2051  * - When threaded IRQs are enabled, completion handlers are invoked in thread
2052  *   context instead of interrupt context.
2053  */
2054 static void srpt_completion(struct ib_cq *cq, void *ctx)
2055 {
2056 	struct srpt_rdma_ch *ch = ctx;
2057 
2058 	wake_up_interruptible(&ch->wait_queue);
2059 }
2060 
2061 static int srpt_compl_thread(void *arg)
2062 {
2063 	struct srpt_rdma_ch *ch;
2064 
2065 	/* Hibernation / freezing of the SRPT kernel thread is not supported. */
2066 	current->flags |= PF_NOFREEZE;
2067 
2068 	ch = arg;
2069 	BUG_ON(!ch);
2070 	printk(KERN_INFO "Session %s: kernel thread %s (PID %d) started\n",
2071 	       ch->sess_name, ch->thread->comm, current->pid);
2072 	while (!kthread_should_stop()) {
2073 		wait_event_interruptible(ch->wait_queue,
2074 			(srpt_process_completion(ch->cq, ch),
2075 			 kthread_should_stop()));
2076 	}
2077 	printk(KERN_INFO "Session %s: kernel thread %s (PID %d) stopped\n",
2078 	       ch->sess_name, ch->thread->comm, current->pid);
2079 	return 0;
2080 }
2081 
2082 /**
2083  * srpt_create_ch_ib() - Create receive and send completion queues.
2084  */
2085 static int srpt_create_ch_ib(struct srpt_rdma_ch *ch)
2086 {
2087 	struct ib_qp_init_attr *qp_init;
2088 	struct srpt_port *sport = ch->sport;
2089 	struct srpt_device *sdev = sport->sdev;
2090 	u32 srp_sq_size = sport->port_attrib.srp_sq_size;
2091 	int ret;
2092 
2093 	WARN_ON(ch->rq_size < 1);
2094 
2095 	ret = -ENOMEM;
2096 	qp_init = kzalloc(sizeof *qp_init, GFP_KERNEL);
2097 	if (!qp_init)
2098 		goto out;
2099 
2100 	ch->cq = ib_create_cq(sdev->device, srpt_completion, NULL, ch,
2101 			      ch->rq_size + srp_sq_size, 0);
2102 	if (IS_ERR(ch->cq)) {
2103 		ret = PTR_ERR(ch->cq);
2104 		printk(KERN_ERR "failed to create CQ cqe= %d ret= %d\n",
2105 		       ch->rq_size + srp_sq_size, ret);
2106 		goto out;
2107 	}
2108 
2109 	qp_init->qp_context = (void *)ch;
2110 	qp_init->event_handler
2111 		= (void(*)(struct ib_event *, void*))srpt_qp_event;
2112 	qp_init->send_cq = ch->cq;
2113 	qp_init->recv_cq = ch->cq;
2114 	qp_init->srq = sdev->srq;
2115 	qp_init->sq_sig_type = IB_SIGNAL_REQ_WR;
2116 	qp_init->qp_type = IB_QPT_RC;
2117 	qp_init->cap.max_send_wr = srp_sq_size;
2118 	qp_init->cap.max_send_sge = SRPT_DEF_SG_PER_WQE;
2119 
2120 	ch->qp = ib_create_qp(sdev->pd, qp_init);
2121 	if (IS_ERR(ch->qp)) {
2122 		ret = PTR_ERR(ch->qp);
2123 		printk(KERN_ERR "failed to create_qp ret= %d\n", ret);
2124 		goto err_destroy_cq;
2125 	}
2126 
2127 	atomic_set(&ch->sq_wr_avail, qp_init->cap.max_send_wr);
2128 
2129 	pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
2130 		 __func__, ch->cq->cqe, qp_init->cap.max_send_sge,
2131 		 qp_init->cap.max_send_wr, ch->cm_id);
2132 
2133 	ret = srpt_init_ch_qp(ch, ch->qp);
2134 	if (ret)
2135 		goto err_destroy_qp;
2136 
2137 	init_waitqueue_head(&ch->wait_queue);
2138 
2139 	pr_debug("creating thread for session %s\n", ch->sess_name);
2140 
2141 	ch->thread = kthread_run(srpt_compl_thread, ch, "ib_srpt_compl");
2142 	if (IS_ERR(ch->thread)) {
2143 		printk(KERN_ERR "failed to create kernel thread %ld\n",
2144 		       PTR_ERR(ch->thread));
2145 		ch->thread = NULL;
2146 		goto err_destroy_qp;
2147 	}
2148 
2149 out:
2150 	kfree(qp_init);
2151 	return ret;
2152 
2153 err_destroy_qp:
2154 	ib_destroy_qp(ch->qp);
2155 err_destroy_cq:
2156 	ib_destroy_cq(ch->cq);
2157 	goto out;
2158 }
2159 
2160 static void srpt_destroy_ch_ib(struct srpt_rdma_ch *ch)
2161 {
2162 	if (ch->thread)
2163 		kthread_stop(ch->thread);
2164 
2165 	ib_destroy_qp(ch->qp);
2166 	ib_destroy_cq(ch->cq);
2167 }
2168 
2169 /**
2170  * __srpt_close_ch() - Close an RDMA channel by setting the QP error state.
2171  *
2172  * Reset the QP and make sure all resources associated with the channel will
2173  * be deallocated at an appropriate time.
2174  *
2175  * Note: The caller must hold ch->sport->sdev->spinlock.
2176  */
2177 static void __srpt_close_ch(struct srpt_rdma_ch *ch)
2178 {
2179 	struct srpt_device *sdev;
2180 	enum rdma_ch_state prev_state;
2181 	unsigned long flags;
2182 
2183 	sdev = ch->sport->sdev;
2184 
2185 	spin_lock_irqsave(&ch->spinlock, flags);
2186 	prev_state = ch->state;
2187 	switch (prev_state) {
2188 	case CH_CONNECTING:
2189 	case CH_LIVE:
2190 		ch->state = CH_DISCONNECTING;
2191 		break;
2192 	default:
2193 		break;
2194 	}
2195 	spin_unlock_irqrestore(&ch->spinlock, flags);
2196 
2197 	switch (prev_state) {
2198 	case CH_CONNECTING:
2199 		ib_send_cm_rej(ch->cm_id, IB_CM_REJ_NO_RESOURCES, NULL, 0,
2200 			       NULL, 0);
2201 		/* fall through */
2202 	case CH_LIVE:
2203 		if (ib_send_cm_dreq(ch->cm_id, NULL, 0) < 0)
2204 			printk(KERN_ERR "sending CM DREQ failed.\n");
2205 		break;
2206 	case CH_DISCONNECTING:
2207 		break;
2208 	case CH_DRAINING:
2209 	case CH_RELEASING:
2210 		break;
2211 	}
2212 }
2213 
2214 /**
2215  * srpt_close_ch() - Close an RDMA channel.
2216  */
2217 static void srpt_close_ch(struct srpt_rdma_ch *ch)
2218 {
2219 	struct srpt_device *sdev;
2220 
2221 	sdev = ch->sport->sdev;
2222 	spin_lock_irq(&sdev->spinlock);
2223 	__srpt_close_ch(ch);
2224 	spin_unlock_irq(&sdev->spinlock);
2225 }
2226 
2227 /**
2228  * srpt_shutdown_session() - Whether or not a session may be shut down.
2229  */
2230 static int srpt_shutdown_session(struct se_session *se_sess)
2231 {
2232 	struct srpt_rdma_ch *ch = se_sess->fabric_sess_ptr;
2233 	unsigned long flags;
2234 
2235 	spin_lock_irqsave(&ch->spinlock, flags);
2236 	if (ch->in_shutdown) {
2237 		spin_unlock_irqrestore(&ch->spinlock, flags);
2238 		return true;
2239 	}
2240 
2241 	ch->in_shutdown = true;
2242 	target_sess_cmd_list_set_waiting(se_sess);
2243 	spin_unlock_irqrestore(&ch->spinlock, flags);
2244 
2245 	return true;
2246 }
2247 
2248 /**
2249  * srpt_drain_channel() - Drain a channel by resetting the IB queue pair.
2250  * @cm_id: Pointer to the CM ID of the channel to be drained.
2251  *
2252  * Note: Must be called from inside srpt_cm_handler to avoid a race between
2253  * accessing sdev->spinlock and the call to kfree(sdev) in srpt_remove_one()
2254  * (the caller of srpt_cm_handler holds the cm_id spinlock; srpt_remove_one()
2255  * waits until all target sessions for the associated IB device have been
2256  * unregistered and target session registration involves a call to
2257  * ib_destroy_cm_id(), which locks the cm_id spinlock and hence waits until
2258  * this function has finished).
2259  */
2260 static void srpt_drain_channel(struct ib_cm_id *cm_id)
2261 {
2262 	struct srpt_device *sdev;
2263 	struct srpt_rdma_ch *ch;
2264 	int ret;
2265 	bool do_reset = false;
2266 
2267 	WARN_ON_ONCE(irqs_disabled());
2268 
2269 	sdev = cm_id->context;
2270 	BUG_ON(!sdev);
2271 	spin_lock_irq(&sdev->spinlock);
2272 	list_for_each_entry(ch, &sdev->rch_list, list) {
2273 		if (ch->cm_id == cm_id) {
2274 			do_reset = srpt_test_and_set_ch_state(ch,
2275 					CH_CONNECTING, CH_DRAINING) ||
2276 				   srpt_test_and_set_ch_state(ch,
2277 					CH_LIVE, CH_DRAINING) ||
2278 				   srpt_test_and_set_ch_state(ch,
2279 					CH_DISCONNECTING, CH_DRAINING);
2280 			break;
2281 		}
2282 	}
2283 	spin_unlock_irq(&sdev->spinlock);
2284 
2285 	if (do_reset) {
2286 		if (ch->sess)
2287 			srpt_shutdown_session(ch->sess);
2288 
2289 		ret = srpt_ch_qp_err(ch);
2290 		if (ret < 0)
2291 			printk(KERN_ERR "Setting queue pair in error state"
2292 			       " failed: %d\n", ret);
2293 	}
2294 }
2295 
2296 /**
2297  * srpt_find_channel() - Look up an RDMA channel.
2298  * @cm_id: Pointer to the CM ID of the channel to be looked up.
2299  *
2300  * Return NULL if no matching RDMA channel has been found.
2301  */
2302 static struct srpt_rdma_ch *srpt_find_channel(struct srpt_device *sdev,
2303 					      struct ib_cm_id *cm_id)
2304 {
2305 	struct srpt_rdma_ch *ch;
2306 	bool found;
2307 
2308 	WARN_ON_ONCE(irqs_disabled());
2309 	BUG_ON(!sdev);
2310 
2311 	found = false;
2312 	spin_lock_irq(&sdev->spinlock);
2313 	list_for_each_entry(ch, &sdev->rch_list, list) {
2314 		if (ch->cm_id == cm_id) {
2315 			found = true;
2316 			break;
2317 		}
2318 	}
2319 	spin_unlock_irq(&sdev->spinlock);
2320 
2321 	return found ? ch : NULL;
2322 }
2323 
2324 /**
2325  * srpt_release_channel() - Release channel resources.
2326  *
2327  * Schedules the actual release because:
2328  * - Calling the ib_destroy_cm_id() call from inside an IB CM callback would
2329  *   trigger a deadlock.
2330  * - It is not safe to call TCM transport_* functions from interrupt context.
2331  */
2332 static void srpt_release_channel(struct srpt_rdma_ch *ch)
2333 {
2334 	schedule_work(&ch->release_work);
2335 }
2336 
2337 static void srpt_release_channel_work(struct work_struct *w)
2338 {
2339 	struct srpt_rdma_ch *ch;
2340 	struct srpt_device *sdev;
2341 	struct se_session *se_sess;
2342 
2343 	ch = container_of(w, struct srpt_rdma_ch, release_work);
2344 	pr_debug("ch = %p; ch->sess = %p; release_done = %p\n", ch, ch->sess,
2345 		 ch->release_done);
2346 
2347 	sdev = ch->sport->sdev;
2348 	BUG_ON(!sdev);
2349 
2350 	se_sess = ch->sess;
2351 	BUG_ON(!se_sess);
2352 
2353 	target_wait_for_sess_cmds(se_sess);
2354 
2355 	transport_deregister_session_configfs(se_sess);
2356 	transport_deregister_session(se_sess);
2357 	ch->sess = NULL;
2358 
2359 	ib_destroy_cm_id(ch->cm_id);
2360 
2361 	srpt_destroy_ch_ib(ch);
2362 
2363 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
2364 			     ch->sport->sdev, ch->rq_size,
2365 			     ch->rsp_size, DMA_TO_DEVICE);
2366 
2367 	spin_lock_irq(&sdev->spinlock);
2368 	list_del(&ch->list);
2369 	spin_unlock_irq(&sdev->spinlock);
2370 
2371 	if (ch->release_done)
2372 		complete(ch->release_done);
2373 
2374 	wake_up(&sdev->ch_releaseQ);
2375 
2376 	kfree(ch);
2377 }
2378 
2379 static struct srpt_node_acl *__srpt_lookup_acl(struct srpt_port *sport,
2380 					       u8 i_port_id[16])
2381 {
2382 	struct srpt_node_acl *nacl;
2383 
2384 	list_for_each_entry(nacl, &sport->port_acl_list, list)
2385 		if (memcmp(nacl->i_port_id, i_port_id,
2386 			   sizeof(nacl->i_port_id)) == 0)
2387 			return nacl;
2388 
2389 	return NULL;
2390 }
2391 
2392 static struct srpt_node_acl *srpt_lookup_acl(struct srpt_port *sport,
2393 					     u8 i_port_id[16])
2394 {
2395 	struct srpt_node_acl *nacl;
2396 
2397 	spin_lock_irq(&sport->port_acl_lock);
2398 	nacl = __srpt_lookup_acl(sport, i_port_id);
2399 	spin_unlock_irq(&sport->port_acl_lock);
2400 
2401 	return nacl;
2402 }
2403 
2404 /**
2405  * srpt_cm_req_recv() - Process the event IB_CM_REQ_RECEIVED.
2406  *
2407  * Ownership of the cm_id is transferred to the target session if this
2408  * functions returns zero. Otherwise the caller remains the owner of cm_id.
2409  */
2410 static int srpt_cm_req_recv(struct ib_cm_id *cm_id,
2411 			    struct ib_cm_req_event_param *param,
2412 			    void *private_data)
2413 {
2414 	struct srpt_device *sdev = cm_id->context;
2415 	struct srpt_port *sport = &sdev->port[param->port - 1];
2416 	struct srp_login_req *req;
2417 	struct srp_login_rsp *rsp;
2418 	struct srp_login_rej *rej;
2419 	struct ib_cm_rep_param *rep_param;
2420 	struct srpt_rdma_ch *ch, *tmp_ch;
2421 	struct srpt_node_acl *nacl;
2422 	u32 it_iu_len;
2423 	int i;
2424 	int ret = 0;
2425 
2426 	WARN_ON_ONCE(irqs_disabled());
2427 
2428 	if (WARN_ON(!sdev || !private_data))
2429 		return -EINVAL;
2430 
2431 	req = (struct srp_login_req *)private_data;
2432 
2433 	it_iu_len = be32_to_cpu(req->req_it_iu_len);
2434 
2435 	printk(KERN_INFO "Received SRP_LOGIN_REQ with i_port_id 0x%llx:0x%llx,"
2436 	       " t_port_id 0x%llx:0x%llx and it_iu_len %d on port %d"
2437 	       " (guid=0x%llx:0x%llx)\n",
2438 	       be64_to_cpu(*(__be64 *)&req->initiator_port_id[0]),
2439 	       be64_to_cpu(*(__be64 *)&req->initiator_port_id[8]),
2440 	       be64_to_cpu(*(__be64 *)&req->target_port_id[0]),
2441 	       be64_to_cpu(*(__be64 *)&req->target_port_id[8]),
2442 	       it_iu_len,
2443 	       param->port,
2444 	       be64_to_cpu(*(__be64 *)&sdev->port[param->port - 1].gid.raw[0]),
2445 	       be64_to_cpu(*(__be64 *)&sdev->port[param->port - 1].gid.raw[8]));
2446 
2447 	rsp = kzalloc(sizeof *rsp, GFP_KERNEL);
2448 	rej = kzalloc(sizeof *rej, GFP_KERNEL);
2449 	rep_param = kzalloc(sizeof *rep_param, GFP_KERNEL);
2450 
2451 	if (!rsp || !rej || !rep_param) {
2452 		ret = -ENOMEM;
2453 		goto out;
2454 	}
2455 
2456 	if (it_iu_len > srp_max_req_size || it_iu_len < 64) {
2457 		rej->reason = __constant_cpu_to_be32(
2458 				SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE);
2459 		ret = -EINVAL;
2460 		printk(KERN_ERR "rejected SRP_LOGIN_REQ because its"
2461 		       " length (%d bytes) is out of range (%d .. %d)\n",
2462 		       it_iu_len, 64, srp_max_req_size);
2463 		goto reject;
2464 	}
2465 
2466 	if (!sport->enabled) {
2467 		rej->reason = __constant_cpu_to_be32(
2468 			     SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2469 		ret = -EINVAL;
2470 		printk(KERN_ERR "rejected SRP_LOGIN_REQ because the target port"
2471 		       " has not yet been enabled\n");
2472 		goto reject;
2473 	}
2474 
2475 	if ((req->req_flags & SRP_MTCH_ACTION) == SRP_MULTICHAN_SINGLE) {
2476 		rsp->rsp_flags = SRP_LOGIN_RSP_MULTICHAN_NO_CHAN;
2477 
2478 		spin_lock_irq(&sdev->spinlock);
2479 
2480 		list_for_each_entry_safe(ch, tmp_ch, &sdev->rch_list, list) {
2481 			if (!memcmp(ch->i_port_id, req->initiator_port_id, 16)
2482 			    && !memcmp(ch->t_port_id, req->target_port_id, 16)
2483 			    && param->port == ch->sport->port
2484 			    && param->listen_id == ch->sport->sdev->cm_id
2485 			    && ch->cm_id) {
2486 				enum rdma_ch_state ch_state;
2487 
2488 				ch_state = srpt_get_ch_state(ch);
2489 				if (ch_state != CH_CONNECTING
2490 				    && ch_state != CH_LIVE)
2491 					continue;
2492 
2493 				/* found an existing channel */
2494 				pr_debug("Found existing channel %s"
2495 					 " cm_id= %p state= %d\n",
2496 					 ch->sess_name, ch->cm_id, ch_state);
2497 
2498 				__srpt_close_ch(ch);
2499 
2500 				rsp->rsp_flags =
2501 					SRP_LOGIN_RSP_MULTICHAN_TERMINATED;
2502 			}
2503 		}
2504 
2505 		spin_unlock_irq(&sdev->spinlock);
2506 
2507 	} else
2508 		rsp->rsp_flags = SRP_LOGIN_RSP_MULTICHAN_MAINTAINED;
2509 
2510 	if (*(__be64 *)req->target_port_id != cpu_to_be64(srpt_service_guid)
2511 	    || *(__be64 *)(req->target_port_id + 8) !=
2512 	       cpu_to_be64(srpt_service_guid)) {
2513 		rej->reason = __constant_cpu_to_be32(
2514 				SRP_LOGIN_REJ_UNABLE_ASSOCIATE_CHANNEL);
2515 		ret = -ENOMEM;
2516 		printk(KERN_ERR "rejected SRP_LOGIN_REQ because it"
2517 		       " has an invalid target port identifier.\n");
2518 		goto reject;
2519 	}
2520 
2521 	ch = kzalloc(sizeof *ch, GFP_KERNEL);
2522 	if (!ch) {
2523 		rej->reason = __constant_cpu_to_be32(
2524 					SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2525 		printk(KERN_ERR "rejected SRP_LOGIN_REQ because no memory.\n");
2526 		ret = -ENOMEM;
2527 		goto reject;
2528 	}
2529 
2530 	INIT_WORK(&ch->release_work, srpt_release_channel_work);
2531 	memcpy(ch->i_port_id, req->initiator_port_id, 16);
2532 	memcpy(ch->t_port_id, req->target_port_id, 16);
2533 	ch->sport = &sdev->port[param->port - 1];
2534 	ch->cm_id = cm_id;
2535 	/*
2536 	 * Avoid QUEUE_FULL conditions by limiting the number of buffers used
2537 	 * for the SRP protocol to the command queue size.
2538 	 */
2539 	ch->rq_size = SRPT_RQ_SIZE;
2540 	spin_lock_init(&ch->spinlock);
2541 	ch->state = CH_CONNECTING;
2542 	INIT_LIST_HEAD(&ch->cmd_wait_list);
2543 	ch->rsp_size = ch->sport->port_attrib.srp_max_rsp_size;
2544 
2545 	ch->ioctx_ring = (struct srpt_send_ioctx **)
2546 		srpt_alloc_ioctx_ring(ch->sport->sdev, ch->rq_size,
2547 				      sizeof(*ch->ioctx_ring[0]),
2548 				      ch->rsp_size, DMA_TO_DEVICE);
2549 	if (!ch->ioctx_ring)
2550 		goto free_ch;
2551 
2552 	INIT_LIST_HEAD(&ch->free_list);
2553 	for (i = 0; i < ch->rq_size; i++) {
2554 		ch->ioctx_ring[i]->ch = ch;
2555 		list_add_tail(&ch->ioctx_ring[i]->free_list, &ch->free_list);
2556 	}
2557 
2558 	ret = srpt_create_ch_ib(ch);
2559 	if (ret) {
2560 		rej->reason = __constant_cpu_to_be32(
2561 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2562 		printk(KERN_ERR "rejected SRP_LOGIN_REQ because creating"
2563 		       " a new RDMA channel failed.\n");
2564 		goto free_ring;
2565 	}
2566 
2567 	ret = srpt_ch_qp_rtr(ch, ch->qp);
2568 	if (ret) {
2569 		rej->reason = __constant_cpu_to_be32(
2570 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2571 		printk(KERN_ERR "rejected SRP_LOGIN_REQ because enabling"
2572 		       " RTR failed (error code = %d)\n", ret);
2573 		goto destroy_ib;
2574 	}
2575 	/*
2576 	 * Use the initator port identifier as the session name.
2577 	 */
2578 	snprintf(ch->sess_name, sizeof(ch->sess_name), "0x%016llx%016llx",
2579 			be64_to_cpu(*(__be64 *)ch->i_port_id),
2580 			be64_to_cpu(*(__be64 *)(ch->i_port_id + 8)));
2581 
2582 	pr_debug("registering session %s\n", ch->sess_name);
2583 
2584 	nacl = srpt_lookup_acl(sport, ch->i_port_id);
2585 	if (!nacl) {
2586 		printk(KERN_INFO "Rejected login because no ACL has been"
2587 		       " configured yet for initiator %s.\n", ch->sess_name);
2588 		rej->reason = __constant_cpu_to_be32(
2589 				SRP_LOGIN_REJ_CHANNEL_LIMIT_REACHED);
2590 		goto destroy_ib;
2591 	}
2592 
2593 	ch->sess = transport_init_session();
2594 	if (IS_ERR(ch->sess)) {
2595 		rej->reason = __constant_cpu_to_be32(
2596 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2597 		pr_debug("Failed to create session\n");
2598 		goto deregister_session;
2599 	}
2600 	ch->sess->se_node_acl = &nacl->nacl;
2601 	transport_register_session(&sport->port_tpg_1, &nacl->nacl, ch->sess, ch);
2602 
2603 	pr_debug("Establish connection sess=%p name=%s cm_id=%p\n", ch->sess,
2604 		 ch->sess_name, ch->cm_id);
2605 
2606 	/* create srp_login_response */
2607 	rsp->opcode = SRP_LOGIN_RSP;
2608 	rsp->tag = req->tag;
2609 	rsp->max_it_iu_len = req->req_it_iu_len;
2610 	rsp->max_ti_iu_len = req->req_it_iu_len;
2611 	ch->max_ti_iu_len = it_iu_len;
2612 	rsp->buf_fmt = __constant_cpu_to_be16(SRP_BUF_FORMAT_DIRECT
2613 					      | SRP_BUF_FORMAT_INDIRECT);
2614 	rsp->req_lim_delta = cpu_to_be32(ch->rq_size);
2615 	atomic_set(&ch->req_lim, ch->rq_size);
2616 	atomic_set(&ch->req_lim_delta, 0);
2617 
2618 	/* create cm reply */
2619 	rep_param->qp_num = ch->qp->qp_num;
2620 	rep_param->private_data = (void *)rsp;
2621 	rep_param->private_data_len = sizeof *rsp;
2622 	rep_param->rnr_retry_count = 7;
2623 	rep_param->flow_control = 1;
2624 	rep_param->failover_accepted = 0;
2625 	rep_param->srq = 1;
2626 	rep_param->responder_resources = 4;
2627 	rep_param->initiator_depth = 4;
2628 
2629 	ret = ib_send_cm_rep(cm_id, rep_param);
2630 	if (ret) {
2631 		printk(KERN_ERR "sending SRP_LOGIN_REQ response failed"
2632 		       " (error code = %d)\n", ret);
2633 		goto release_channel;
2634 	}
2635 
2636 	spin_lock_irq(&sdev->spinlock);
2637 	list_add_tail(&ch->list, &sdev->rch_list);
2638 	spin_unlock_irq(&sdev->spinlock);
2639 
2640 	goto out;
2641 
2642 release_channel:
2643 	srpt_set_ch_state(ch, CH_RELEASING);
2644 	transport_deregister_session_configfs(ch->sess);
2645 
2646 deregister_session:
2647 	transport_deregister_session(ch->sess);
2648 	ch->sess = NULL;
2649 
2650 destroy_ib:
2651 	srpt_destroy_ch_ib(ch);
2652 
2653 free_ring:
2654 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
2655 			     ch->sport->sdev, ch->rq_size,
2656 			     ch->rsp_size, DMA_TO_DEVICE);
2657 free_ch:
2658 	kfree(ch);
2659 
2660 reject:
2661 	rej->opcode = SRP_LOGIN_REJ;
2662 	rej->tag = req->tag;
2663 	rej->buf_fmt = __constant_cpu_to_be16(SRP_BUF_FORMAT_DIRECT
2664 					      | SRP_BUF_FORMAT_INDIRECT);
2665 
2666 	ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED, NULL, 0,
2667 			     (void *)rej, sizeof *rej);
2668 
2669 out:
2670 	kfree(rep_param);
2671 	kfree(rsp);
2672 	kfree(rej);
2673 
2674 	return ret;
2675 }
2676 
2677 static void srpt_cm_rej_recv(struct ib_cm_id *cm_id)
2678 {
2679 	printk(KERN_INFO "Received IB REJ for cm_id %p.\n", cm_id);
2680 	srpt_drain_channel(cm_id);
2681 }
2682 
2683 /**
2684  * srpt_cm_rtu_recv() - Process an IB_CM_RTU_RECEIVED or USER_ESTABLISHED event.
2685  *
2686  * An IB_CM_RTU_RECEIVED message indicates that the connection is established
2687  * and that the recipient may begin transmitting (RTU = ready to use).
2688  */
2689 static void srpt_cm_rtu_recv(struct ib_cm_id *cm_id)
2690 {
2691 	struct srpt_rdma_ch *ch;
2692 	int ret;
2693 
2694 	ch = srpt_find_channel(cm_id->context, cm_id);
2695 	BUG_ON(!ch);
2696 
2697 	if (srpt_test_and_set_ch_state(ch, CH_CONNECTING, CH_LIVE)) {
2698 		struct srpt_recv_ioctx *ioctx, *ioctx_tmp;
2699 
2700 		ret = srpt_ch_qp_rts(ch, ch->qp);
2701 
2702 		list_for_each_entry_safe(ioctx, ioctx_tmp, &ch->cmd_wait_list,
2703 					 wait_list) {
2704 			list_del(&ioctx->wait_list);
2705 			srpt_handle_new_iu(ch, ioctx, NULL);
2706 		}
2707 		if (ret)
2708 			srpt_close_ch(ch);
2709 	}
2710 }
2711 
2712 static void srpt_cm_timewait_exit(struct ib_cm_id *cm_id)
2713 {
2714 	printk(KERN_INFO "Received IB TimeWait exit for cm_id %p.\n", cm_id);
2715 	srpt_drain_channel(cm_id);
2716 }
2717 
2718 static void srpt_cm_rep_error(struct ib_cm_id *cm_id)
2719 {
2720 	printk(KERN_INFO "Received IB REP error for cm_id %p.\n", cm_id);
2721 	srpt_drain_channel(cm_id);
2722 }
2723 
2724 /**
2725  * srpt_cm_dreq_recv() - Process reception of a DREQ message.
2726  */
2727 static void srpt_cm_dreq_recv(struct ib_cm_id *cm_id)
2728 {
2729 	struct srpt_rdma_ch *ch;
2730 	unsigned long flags;
2731 	bool send_drep = false;
2732 
2733 	ch = srpt_find_channel(cm_id->context, cm_id);
2734 	BUG_ON(!ch);
2735 
2736 	pr_debug("cm_id= %p ch->state= %d\n", cm_id, srpt_get_ch_state(ch));
2737 
2738 	spin_lock_irqsave(&ch->spinlock, flags);
2739 	switch (ch->state) {
2740 	case CH_CONNECTING:
2741 	case CH_LIVE:
2742 		send_drep = true;
2743 		ch->state = CH_DISCONNECTING;
2744 		break;
2745 	case CH_DISCONNECTING:
2746 	case CH_DRAINING:
2747 	case CH_RELEASING:
2748 		WARN(true, "unexpected channel state %d\n", ch->state);
2749 		break;
2750 	}
2751 	spin_unlock_irqrestore(&ch->spinlock, flags);
2752 
2753 	if (send_drep) {
2754 		if (ib_send_cm_drep(ch->cm_id, NULL, 0) < 0)
2755 			printk(KERN_ERR "Sending IB DREP failed.\n");
2756 		printk(KERN_INFO "Received DREQ and sent DREP for session %s.\n",
2757 		       ch->sess_name);
2758 	}
2759 }
2760 
2761 /**
2762  * srpt_cm_drep_recv() - Process reception of a DREP message.
2763  */
2764 static void srpt_cm_drep_recv(struct ib_cm_id *cm_id)
2765 {
2766 	printk(KERN_INFO "Received InfiniBand DREP message for cm_id %p.\n",
2767 	       cm_id);
2768 	srpt_drain_channel(cm_id);
2769 }
2770 
2771 /**
2772  * srpt_cm_handler() - IB connection manager callback function.
2773  *
2774  * A non-zero return value will cause the caller destroy the CM ID.
2775  *
2776  * Note: srpt_cm_handler() must only return a non-zero value when transferring
2777  * ownership of the cm_id to a channel by srpt_cm_req_recv() failed. Returning
2778  * a non-zero value in any other case will trigger a race with the
2779  * ib_destroy_cm_id() call in srpt_release_channel().
2780  */
2781 static int srpt_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
2782 {
2783 	int ret;
2784 
2785 	ret = 0;
2786 	switch (event->event) {
2787 	case IB_CM_REQ_RECEIVED:
2788 		ret = srpt_cm_req_recv(cm_id, &event->param.req_rcvd,
2789 				       event->private_data);
2790 		break;
2791 	case IB_CM_REJ_RECEIVED:
2792 		srpt_cm_rej_recv(cm_id);
2793 		break;
2794 	case IB_CM_RTU_RECEIVED:
2795 	case IB_CM_USER_ESTABLISHED:
2796 		srpt_cm_rtu_recv(cm_id);
2797 		break;
2798 	case IB_CM_DREQ_RECEIVED:
2799 		srpt_cm_dreq_recv(cm_id);
2800 		break;
2801 	case IB_CM_DREP_RECEIVED:
2802 		srpt_cm_drep_recv(cm_id);
2803 		break;
2804 	case IB_CM_TIMEWAIT_EXIT:
2805 		srpt_cm_timewait_exit(cm_id);
2806 		break;
2807 	case IB_CM_REP_ERROR:
2808 		srpt_cm_rep_error(cm_id);
2809 		break;
2810 	case IB_CM_DREQ_ERROR:
2811 		printk(KERN_INFO "Received IB DREQ ERROR event.\n");
2812 		break;
2813 	case IB_CM_MRA_RECEIVED:
2814 		printk(KERN_INFO "Received IB MRA event\n");
2815 		break;
2816 	default:
2817 		printk(KERN_ERR "received unrecognized IB CM event %d\n",
2818 		       event->event);
2819 		break;
2820 	}
2821 
2822 	return ret;
2823 }
2824 
2825 /**
2826  * srpt_perform_rdmas() - Perform IB RDMA.
2827  *
2828  * Returns zero upon success or a negative number upon failure.
2829  */
2830 static int srpt_perform_rdmas(struct srpt_rdma_ch *ch,
2831 			      struct srpt_send_ioctx *ioctx)
2832 {
2833 	struct ib_send_wr wr;
2834 	struct ib_send_wr *bad_wr;
2835 	struct rdma_iu *riu;
2836 	int i;
2837 	int ret;
2838 	int sq_wr_avail;
2839 	enum dma_data_direction dir;
2840 	const int n_rdma = ioctx->n_rdma;
2841 
2842 	dir = ioctx->cmd.data_direction;
2843 	if (dir == DMA_TO_DEVICE) {
2844 		/* write */
2845 		ret = -ENOMEM;
2846 		sq_wr_avail = atomic_sub_return(n_rdma, &ch->sq_wr_avail);
2847 		if (sq_wr_avail < 0) {
2848 			printk(KERN_WARNING "IB send queue full (needed %d)\n",
2849 			       n_rdma);
2850 			goto out;
2851 		}
2852 	}
2853 
2854 	ioctx->rdma_aborted = false;
2855 	ret = 0;
2856 	riu = ioctx->rdma_ius;
2857 	memset(&wr, 0, sizeof wr);
2858 
2859 	for (i = 0; i < n_rdma; ++i, ++riu) {
2860 		if (dir == DMA_FROM_DEVICE) {
2861 			wr.opcode = IB_WR_RDMA_WRITE;
2862 			wr.wr_id = encode_wr_id(i == n_rdma - 1 ?
2863 						SRPT_RDMA_WRITE_LAST :
2864 						SRPT_RDMA_MID,
2865 						ioctx->ioctx.index);
2866 		} else {
2867 			wr.opcode = IB_WR_RDMA_READ;
2868 			wr.wr_id = encode_wr_id(i == n_rdma - 1 ?
2869 						SRPT_RDMA_READ_LAST :
2870 						SRPT_RDMA_MID,
2871 						ioctx->ioctx.index);
2872 		}
2873 		wr.next = NULL;
2874 		wr.wr.rdma.remote_addr = riu->raddr;
2875 		wr.wr.rdma.rkey = riu->rkey;
2876 		wr.num_sge = riu->sge_cnt;
2877 		wr.sg_list = riu->sge;
2878 
2879 		/* only get completion event for the last rdma write */
2880 		if (i == (n_rdma - 1) && dir == DMA_TO_DEVICE)
2881 			wr.send_flags = IB_SEND_SIGNALED;
2882 
2883 		ret = ib_post_send(ch->qp, &wr, &bad_wr);
2884 		if (ret)
2885 			break;
2886 	}
2887 
2888 	if (ret)
2889 		printk(KERN_ERR "%s[%d]: ib_post_send() returned %d for %d/%d",
2890 				 __func__, __LINE__, ret, i, n_rdma);
2891 	if (ret && i > 0) {
2892 		wr.num_sge = 0;
2893 		wr.wr_id = encode_wr_id(SRPT_RDMA_ABORT, ioctx->ioctx.index);
2894 		wr.send_flags = IB_SEND_SIGNALED;
2895 		while (ch->state == CH_LIVE &&
2896 			ib_post_send(ch->qp, &wr, &bad_wr) != 0) {
2897 			printk(KERN_INFO "Trying to abort failed RDMA transfer [%d]",
2898 				ioctx->ioctx.index);
2899 			msleep(1000);
2900 		}
2901 		while (ch->state != CH_RELEASING && !ioctx->rdma_aborted) {
2902 			printk(KERN_INFO "Waiting until RDMA abort finished [%d]",
2903 				ioctx->ioctx.index);
2904 			msleep(1000);
2905 		}
2906 	}
2907 out:
2908 	if (unlikely(dir == DMA_TO_DEVICE && ret < 0))
2909 		atomic_add(n_rdma, &ch->sq_wr_avail);
2910 	return ret;
2911 }
2912 
2913 /**
2914  * srpt_xfer_data() - Start data transfer from initiator to target.
2915  */
2916 static int srpt_xfer_data(struct srpt_rdma_ch *ch,
2917 			  struct srpt_send_ioctx *ioctx)
2918 {
2919 	int ret;
2920 
2921 	ret = srpt_map_sg_to_ib_sge(ch, ioctx);
2922 	if (ret) {
2923 		printk(KERN_ERR "%s[%d] ret=%d\n", __func__, __LINE__, ret);
2924 		goto out;
2925 	}
2926 
2927 	ret = srpt_perform_rdmas(ch, ioctx);
2928 	if (ret) {
2929 		if (ret == -EAGAIN || ret == -ENOMEM)
2930 			printk(KERN_INFO "%s[%d] queue full -- ret=%d\n",
2931 				   __func__, __LINE__, ret);
2932 		else
2933 			printk(KERN_ERR "%s[%d] fatal error -- ret=%d\n",
2934 			       __func__, __LINE__, ret);
2935 		goto out_unmap;
2936 	}
2937 
2938 out:
2939 	return ret;
2940 out_unmap:
2941 	srpt_unmap_sg_to_ib_sge(ch, ioctx);
2942 	goto out;
2943 }
2944 
2945 static int srpt_write_pending_status(struct se_cmd *se_cmd)
2946 {
2947 	struct srpt_send_ioctx *ioctx;
2948 
2949 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
2950 	return srpt_get_cmd_state(ioctx) == SRPT_STATE_NEED_DATA;
2951 }
2952 
2953 /*
2954  * srpt_write_pending() - Start data transfer from initiator to target (write).
2955  */
2956 static int srpt_write_pending(struct se_cmd *se_cmd)
2957 {
2958 	struct srpt_rdma_ch *ch;
2959 	struct srpt_send_ioctx *ioctx;
2960 	enum srpt_command_state new_state;
2961 	enum rdma_ch_state ch_state;
2962 	int ret;
2963 
2964 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
2965 
2966 	new_state = srpt_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA);
2967 	WARN_ON(new_state == SRPT_STATE_DONE);
2968 
2969 	ch = ioctx->ch;
2970 	BUG_ON(!ch);
2971 
2972 	ch_state = srpt_get_ch_state(ch);
2973 	switch (ch_state) {
2974 	case CH_CONNECTING:
2975 		WARN(true, "unexpected channel state %d\n", ch_state);
2976 		ret = -EINVAL;
2977 		goto out;
2978 	case CH_LIVE:
2979 		break;
2980 	case CH_DISCONNECTING:
2981 	case CH_DRAINING:
2982 	case CH_RELEASING:
2983 		pr_debug("cmd with tag %lld: channel disconnecting\n",
2984 			 ioctx->tag);
2985 		srpt_set_cmd_state(ioctx, SRPT_STATE_DATA_IN);
2986 		ret = -EINVAL;
2987 		goto out;
2988 	}
2989 	ret = srpt_xfer_data(ch, ioctx);
2990 
2991 out:
2992 	return ret;
2993 }
2994 
2995 static u8 tcm_to_srp_tsk_mgmt_status(const int tcm_mgmt_status)
2996 {
2997 	switch (tcm_mgmt_status) {
2998 	case TMR_FUNCTION_COMPLETE:
2999 		return SRP_TSK_MGMT_SUCCESS;
3000 	case TMR_FUNCTION_REJECTED:
3001 		return SRP_TSK_MGMT_FUNC_NOT_SUPP;
3002 	}
3003 	return SRP_TSK_MGMT_FAILED;
3004 }
3005 
3006 /**
3007  * srpt_queue_response() - Transmits the response to a SCSI command.
3008  *
3009  * Callback function called by the TCM core. Must not block since it can be
3010  * invoked on the context of the IB completion handler.
3011  */
3012 static void srpt_queue_response(struct se_cmd *cmd)
3013 {
3014 	struct srpt_rdma_ch *ch;
3015 	struct srpt_send_ioctx *ioctx;
3016 	enum srpt_command_state state;
3017 	unsigned long flags;
3018 	int ret;
3019 	enum dma_data_direction dir;
3020 	int resp_len;
3021 	u8 srp_tm_status;
3022 
3023 	ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
3024 	ch = ioctx->ch;
3025 	BUG_ON(!ch);
3026 
3027 	spin_lock_irqsave(&ioctx->spinlock, flags);
3028 	state = ioctx->state;
3029 	switch (state) {
3030 	case SRPT_STATE_NEW:
3031 	case SRPT_STATE_DATA_IN:
3032 		ioctx->state = SRPT_STATE_CMD_RSP_SENT;
3033 		break;
3034 	case SRPT_STATE_MGMT:
3035 		ioctx->state = SRPT_STATE_MGMT_RSP_SENT;
3036 		break;
3037 	default:
3038 		WARN(true, "ch %p; cmd %d: unexpected command state %d\n",
3039 			ch, ioctx->ioctx.index, ioctx->state);
3040 		break;
3041 	}
3042 	spin_unlock_irqrestore(&ioctx->spinlock, flags);
3043 
3044 	if (unlikely(transport_check_aborted_status(&ioctx->cmd, false)
3045 		     || WARN_ON_ONCE(state == SRPT_STATE_CMD_RSP_SENT))) {
3046 		atomic_inc(&ch->req_lim_delta);
3047 		srpt_abort_cmd(ioctx);
3048 		return;
3049 	}
3050 
3051 	dir = ioctx->cmd.data_direction;
3052 
3053 	/* For read commands, transfer the data to the initiator. */
3054 	if (dir == DMA_FROM_DEVICE && ioctx->cmd.data_length &&
3055 	    !ioctx->queue_status_only) {
3056 		ret = srpt_xfer_data(ch, ioctx);
3057 		if (ret) {
3058 			printk(KERN_ERR "xfer_data failed for tag %llu\n",
3059 			       ioctx->tag);
3060 			return;
3061 		}
3062 	}
3063 
3064 	if (state != SRPT_STATE_MGMT)
3065 		resp_len = srpt_build_cmd_rsp(ch, ioctx, ioctx->tag,
3066 					      cmd->scsi_status);
3067 	else {
3068 		srp_tm_status
3069 			= tcm_to_srp_tsk_mgmt_status(cmd->se_tmr_req->response);
3070 		resp_len = srpt_build_tskmgmt_rsp(ch, ioctx, srp_tm_status,
3071 						 ioctx->tag);
3072 	}
3073 	ret = srpt_post_send(ch, ioctx, resp_len);
3074 	if (ret) {
3075 		printk(KERN_ERR "sending cmd response failed for tag %llu\n",
3076 		       ioctx->tag);
3077 		srpt_unmap_sg_to_ib_sge(ch, ioctx);
3078 		srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
3079 		target_put_sess_cmd(ioctx->ch->sess, &ioctx->cmd);
3080 	}
3081 }
3082 
3083 static int srpt_queue_data_in(struct se_cmd *cmd)
3084 {
3085 	srpt_queue_response(cmd);
3086 	return 0;
3087 }
3088 
3089 static void srpt_queue_tm_rsp(struct se_cmd *cmd)
3090 {
3091 	srpt_queue_response(cmd);
3092 }
3093 
3094 static int srpt_queue_status(struct se_cmd *cmd)
3095 {
3096 	struct srpt_send_ioctx *ioctx;
3097 
3098 	ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
3099 	BUG_ON(ioctx->sense_data != cmd->sense_buffer);
3100 	if (cmd->se_cmd_flags &
3101 	    (SCF_TRANSPORT_TASK_SENSE | SCF_EMULATED_TASK_SENSE))
3102 		WARN_ON(cmd->scsi_status != SAM_STAT_CHECK_CONDITION);
3103 	ioctx->queue_status_only = true;
3104 	srpt_queue_response(cmd);
3105 	return 0;
3106 }
3107 
3108 static void srpt_refresh_port_work(struct work_struct *work)
3109 {
3110 	struct srpt_port *sport = container_of(work, struct srpt_port, work);
3111 
3112 	srpt_refresh_port(sport);
3113 }
3114 
3115 static int srpt_ch_list_empty(struct srpt_device *sdev)
3116 {
3117 	int res;
3118 
3119 	spin_lock_irq(&sdev->spinlock);
3120 	res = list_empty(&sdev->rch_list);
3121 	spin_unlock_irq(&sdev->spinlock);
3122 
3123 	return res;
3124 }
3125 
3126 /**
3127  * srpt_release_sdev() - Free the channel resources associated with a target.
3128  */
3129 static int srpt_release_sdev(struct srpt_device *sdev)
3130 {
3131 	struct srpt_rdma_ch *ch, *tmp_ch;
3132 	int res;
3133 
3134 	WARN_ON_ONCE(irqs_disabled());
3135 
3136 	BUG_ON(!sdev);
3137 
3138 	spin_lock_irq(&sdev->spinlock);
3139 	list_for_each_entry_safe(ch, tmp_ch, &sdev->rch_list, list)
3140 		__srpt_close_ch(ch);
3141 	spin_unlock_irq(&sdev->spinlock);
3142 
3143 	res = wait_event_interruptible(sdev->ch_releaseQ,
3144 				       srpt_ch_list_empty(sdev));
3145 	if (res)
3146 		printk(KERN_ERR "%s: interrupted.\n", __func__);
3147 
3148 	return 0;
3149 }
3150 
3151 static struct srpt_port *__srpt_lookup_port(const char *name)
3152 {
3153 	struct ib_device *dev;
3154 	struct srpt_device *sdev;
3155 	struct srpt_port *sport;
3156 	int i;
3157 
3158 	list_for_each_entry(sdev, &srpt_dev_list, list) {
3159 		dev = sdev->device;
3160 		if (!dev)
3161 			continue;
3162 
3163 		for (i = 0; i < dev->phys_port_cnt; i++) {
3164 			sport = &sdev->port[i];
3165 
3166 			if (!strcmp(sport->port_guid, name))
3167 				return sport;
3168 		}
3169 	}
3170 
3171 	return NULL;
3172 }
3173 
3174 static struct srpt_port *srpt_lookup_port(const char *name)
3175 {
3176 	struct srpt_port *sport;
3177 
3178 	spin_lock(&srpt_dev_lock);
3179 	sport = __srpt_lookup_port(name);
3180 	spin_unlock(&srpt_dev_lock);
3181 
3182 	return sport;
3183 }
3184 
3185 /**
3186  * srpt_add_one() - Infiniband device addition callback function.
3187  */
3188 static void srpt_add_one(struct ib_device *device)
3189 {
3190 	struct srpt_device *sdev;
3191 	struct srpt_port *sport;
3192 	struct ib_srq_init_attr srq_attr;
3193 	int i;
3194 
3195 	pr_debug("device = %p, device->dma_ops = %p\n", device,
3196 		 device->dma_ops);
3197 
3198 	sdev = kzalloc(sizeof *sdev, GFP_KERNEL);
3199 	if (!sdev)
3200 		goto err;
3201 
3202 	sdev->device = device;
3203 	INIT_LIST_HEAD(&sdev->rch_list);
3204 	init_waitqueue_head(&sdev->ch_releaseQ);
3205 	spin_lock_init(&sdev->spinlock);
3206 
3207 	if (ib_query_device(device, &sdev->dev_attr))
3208 		goto free_dev;
3209 
3210 	sdev->pd = ib_alloc_pd(device);
3211 	if (IS_ERR(sdev->pd))
3212 		goto free_dev;
3213 
3214 	sdev->mr = ib_get_dma_mr(sdev->pd, IB_ACCESS_LOCAL_WRITE);
3215 	if (IS_ERR(sdev->mr))
3216 		goto err_pd;
3217 
3218 	sdev->srq_size = min(srpt_srq_size, sdev->dev_attr.max_srq_wr);
3219 
3220 	srq_attr.event_handler = srpt_srq_event;
3221 	srq_attr.srq_context = (void *)sdev;
3222 	srq_attr.attr.max_wr = sdev->srq_size;
3223 	srq_attr.attr.max_sge = 1;
3224 	srq_attr.attr.srq_limit = 0;
3225 	srq_attr.srq_type = IB_SRQT_BASIC;
3226 
3227 	sdev->srq = ib_create_srq(sdev->pd, &srq_attr);
3228 	if (IS_ERR(sdev->srq))
3229 		goto err_mr;
3230 
3231 	pr_debug("%s: create SRQ #wr= %d max_allow=%d dev= %s\n",
3232 		 __func__, sdev->srq_size, sdev->dev_attr.max_srq_wr,
3233 		 device->name);
3234 
3235 	if (!srpt_service_guid)
3236 		srpt_service_guid = be64_to_cpu(device->node_guid);
3237 
3238 	sdev->cm_id = ib_create_cm_id(device, srpt_cm_handler, sdev);
3239 	if (IS_ERR(sdev->cm_id))
3240 		goto err_srq;
3241 
3242 	/* print out target login information */
3243 	pr_debug("Target login info: id_ext=%016llx,ioc_guid=%016llx,"
3244 		 "pkey=ffff,service_id=%016llx\n", srpt_service_guid,
3245 		 srpt_service_guid, srpt_service_guid);
3246 
3247 	/*
3248 	 * We do not have a consistent service_id (ie. also id_ext of target_id)
3249 	 * to identify this target. We currently use the guid of the first HCA
3250 	 * in the system as service_id; therefore, the target_id will change
3251 	 * if this HCA is gone bad and replaced by different HCA
3252 	 */
3253 	if (ib_cm_listen(sdev->cm_id, cpu_to_be64(srpt_service_guid), 0, NULL))
3254 		goto err_cm;
3255 
3256 	INIT_IB_EVENT_HANDLER(&sdev->event_handler, sdev->device,
3257 			      srpt_event_handler);
3258 	if (ib_register_event_handler(&sdev->event_handler))
3259 		goto err_cm;
3260 
3261 	sdev->ioctx_ring = (struct srpt_recv_ioctx **)
3262 		srpt_alloc_ioctx_ring(sdev, sdev->srq_size,
3263 				      sizeof(*sdev->ioctx_ring[0]),
3264 				      srp_max_req_size, DMA_FROM_DEVICE);
3265 	if (!sdev->ioctx_ring)
3266 		goto err_event;
3267 
3268 	for (i = 0; i < sdev->srq_size; ++i)
3269 		srpt_post_recv(sdev, sdev->ioctx_ring[i]);
3270 
3271 	WARN_ON(sdev->device->phys_port_cnt > ARRAY_SIZE(sdev->port));
3272 
3273 	for (i = 1; i <= sdev->device->phys_port_cnt; i++) {
3274 		sport = &sdev->port[i - 1];
3275 		sport->sdev = sdev;
3276 		sport->port = i;
3277 		sport->port_attrib.srp_max_rdma_size = DEFAULT_MAX_RDMA_SIZE;
3278 		sport->port_attrib.srp_max_rsp_size = DEFAULT_MAX_RSP_SIZE;
3279 		sport->port_attrib.srp_sq_size = DEF_SRPT_SQ_SIZE;
3280 		INIT_WORK(&sport->work, srpt_refresh_port_work);
3281 		INIT_LIST_HEAD(&sport->port_acl_list);
3282 		spin_lock_init(&sport->port_acl_lock);
3283 
3284 		if (srpt_refresh_port(sport)) {
3285 			printk(KERN_ERR "MAD registration failed for %s-%d.\n",
3286 			       srpt_sdev_name(sdev), i);
3287 			goto err_ring;
3288 		}
3289 		snprintf(sport->port_guid, sizeof(sport->port_guid),
3290 			"0x%016llx%016llx",
3291 			be64_to_cpu(sport->gid.global.subnet_prefix),
3292 			be64_to_cpu(sport->gid.global.interface_id));
3293 	}
3294 
3295 	spin_lock(&srpt_dev_lock);
3296 	list_add_tail(&sdev->list, &srpt_dev_list);
3297 	spin_unlock(&srpt_dev_lock);
3298 
3299 out:
3300 	ib_set_client_data(device, &srpt_client, sdev);
3301 	pr_debug("added %s.\n", device->name);
3302 	return;
3303 
3304 err_ring:
3305 	srpt_free_ioctx_ring((struct srpt_ioctx **)sdev->ioctx_ring, sdev,
3306 			     sdev->srq_size, srp_max_req_size,
3307 			     DMA_FROM_DEVICE);
3308 err_event:
3309 	ib_unregister_event_handler(&sdev->event_handler);
3310 err_cm:
3311 	ib_destroy_cm_id(sdev->cm_id);
3312 err_srq:
3313 	ib_destroy_srq(sdev->srq);
3314 err_mr:
3315 	ib_dereg_mr(sdev->mr);
3316 err_pd:
3317 	ib_dealloc_pd(sdev->pd);
3318 free_dev:
3319 	kfree(sdev);
3320 err:
3321 	sdev = NULL;
3322 	printk(KERN_INFO "%s(%s) failed.\n", __func__, device->name);
3323 	goto out;
3324 }
3325 
3326 /**
3327  * srpt_remove_one() - InfiniBand device removal callback function.
3328  */
3329 static void srpt_remove_one(struct ib_device *device)
3330 {
3331 	struct srpt_device *sdev;
3332 	int i;
3333 
3334 	sdev = ib_get_client_data(device, &srpt_client);
3335 	if (!sdev) {
3336 		printk(KERN_INFO "%s(%s): nothing to do.\n", __func__,
3337 		       device->name);
3338 		return;
3339 	}
3340 
3341 	srpt_unregister_mad_agent(sdev);
3342 
3343 	ib_unregister_event_handler(&sdev->event_handler);
3344 
3345 	/* Cancel any work queued by the just unregistered IB event handler. */
3346 	for (i = 0; i < sdev->device->phys_port_cnt; i++)
3347 		cancel_work_sync(&sdev->port[i].work);
3348 
3349 	ib_destroy_cm_id(sdev->cm_id);
3350 
3351 	/*
3352 	 * Unregistering a target must happen after destroying sdev->cm_id
3353 	 * such that no new SRP_LOGIN_REQ information units can arrive while
3354 	 * destroying the target.
3355 	 */
3356 	spin_lock(&srpt_dev_lock);
3357 	list_del(&sdev->list);
3358 	spin_unlock(&srpt_dev_lock);
3359 	srpt_release_sdev(sdev);
3360 
3361 	ib_destroy_srq(sdev->srq);
3362 	ib_dereg_mr(sdev->mr);
3363 	ib_dealloc_pd(sdev->pd);
3364 
3365 	srpt_free_ioctx_ring((struct srpt_ioctx **)sdev->ioctx_ring, sdev,
3366 			     sdev->srq_size, srp_max_req_size, DMA_FROM_DEVICE);
3367 	sdev->ioctx_ring = NULL;
3368 	kfree(sdev);
3369 }
3370 
3371 static struct ib_client srpt_client = {
3372 	.name = DRV_NAME,
3373 	.add = srpt_add_one,
3374 	.remove = srpt_remove_one
3375 };
3376 
3377 static int srpt_check_true(struct se_portal_group *se_tpg)
3378 {
3379 	return 1;
3380 }
3381 
3382 static int srpt_check_false(struct se_portal_group *se_tpg)
3383 {
3384 	return 0;
3385 }
3386 
3387 static char *srpt_get_fabric_name(void)
3388 {
3389 	return "srpt";
3390 }
3391 
3392 static u8 srpt_get_fabric_proto_ident(struct se_portal_group *se_tpg)
3393 {
3394 	return SCSI_TRANSPORTID_PROTOCOLID_SRP;
3395 }
3396 
3397 static char *srpt_get_fabric_wwn(struct se_portal_group *tpg)
3398 {
3399 	struct srpt_port *sport = container_of(tpg, struct srpt_port, port_tpg_1);
3400 
3401 	return sport->port_guid;
3402 }
3403 
3404 static u16 srpt_get_tag(struct se_portal_group *tpg)
3405 {
3406 	return 1;
3407 }
3408 
3409 static u32 srpt_get_default_depth(struct se_portal_group *se_tpg)
3410 {
3411 	return 1;
3412 }
3413 
3414 static u32 srpt_get_pr_transport_id(struct se_portal_group *se_tpg,
3415 				    struct se_node_acl *se_nacl,
3416 				    struct t10_pr_registration *pr_reg,
3417 				    int *format_code, unsigned char *buf)
3418 {
3419 	struct srpt_node_acl *nacl;
3420 	struct spc_rdma_transport_id *tr_id;
3421 
3422 	nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3423 	tr_id = (void *)buf;
3424 	tr_id->protocol_identifier = SCSI_TRANSPORTID_PROTOCOLID_SRP;
3425 	memcpy(tr_id->i_port_id, nacl->i_port_id, sizeof(tr_id->i_port_id));
3426 	return sizeof(*tr_id);
3427 }
3428 
3429 static u32 srpt_get_pr_transport_id_len(struct se_portal_group *se_tpg,
3430 					struct se_node_acl *se_nacl,
3431 					struct t10_pr_registration *pr_reg,
3432 					int *format_code)
3433 {
3434 	*format_code = 0;
3435 	return sizeof(struct spc_rdma_transport_id);
3436 }
3437 
3438 static char *srpt_parse_pr_out_transport_id(struct se_portal_group *se_tpg,
3439 					    const char *buf, u32 *out_tid_len,
3440 					    char **port_nexus_ptr)
3441 {
3442 	struct spc_rdma_transport_id *tr_id;
3443 
3444 	*port_nexus_ptr = NULL;
3445 	*out_tid_len = sizeof(struct spc_rdma_transport_id);
3446 	tr_id = (void *)buf;
3447 	return (char *)tr_id->i_port_id;
3448 }
3449 
3450 static struct se_node_acl *srpt_alloc_fabric_acl(struct se_portal_group *se_tpg)
3451 {
3452 	struct srpt_node_acl *nacl;
3453 
3454 	nacl = kzalloc(sizeof(struct srpt_node_acl), GFP_KERNEL);
3455 	if (!nacl) {
3456 		printk(KERN_ERR "Unable to allocate struct srpt_node_acl\n");
3457 		return NULL;
3458 	}
3459 
3460 	return &nacl->nacl;
3461 }
3462 
3463 static void srpt_release_fabric_acl(struct se_portal_group *se_tpg,
3464 				    struct se_node_acl *se_nacl)
3465 {
3466 	struct srpt_node_acl *nacl;
3467 
3468 	nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3469 	kfree(nacl);
3470 }
3471 
3472 static u32 srpt_tpg_get_inst_index(struct se_portal_group *se_tpg)
3473 {
3474 	return 1;
3475 }
3476 
3477 static void srpt_release_cmd(struct se_cmd *se_cmd)
3478 {
3479 	struct srpt_send_ioctx *ioctx = container_of(se_cmd,
3480 				struct srpt_send_ioctx, cmd);
3481 	struct srpt_rdma_ch *ch = ioctx->ch;
3482 	unsigned long flags;
3483 
3484 	WARN_ON(ioctx->state != SRPT_STATE_DONE);
3485 	WARN_ON(ioctx->mapped_sg_count != 0);
3486 
3487 	if (ioctx->n_rbuf > 1) {
3488 		kfree(ioctx->rbufs);
3489 		ioctx->rbufs = NULL;
3490 		ioctx->n_rbuf = 0;
3491 	}
3492 
3493 	spin_lock_irqsave(&ch->spinlock, flags);
3494 	list_add(&ioctx->free_list, &ch->free_list);
3495 	spin_unlock_irqrestore(&ch->spinlock, flags);
3496 }
3497 
3498 /**
3499  * srpt_close_session() - Forcibly close a session.
3500  *
3501  * Callback function invoked by the TCM core to clean up sessions associated
3502  * with a node ACL when the user invokes
3503  * rmdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3504  */
3505 static void srpt_close_session(struct se_session *se_sess)
3506 {
3507 	DECLARE_COMPLETION_ONSTACK(release_done);
3508 	struct srpt_rdma_ch *ch;
3509 	struct srpt_device *sdev;
3510 	int res;
3511 
3512 	ch = se_sess->fabric_sess_ptr;
3513 	WARN_ON(ch->sess != se_sess);
3514 
3515 	pr_debug("ch %p state %d\n", ch, srpt_get_ch_state(ch));
3516 
3517 	sdev = ch->sport->sdev;
3518 	spin_lock_irq(&sdev->spinlock);
3519 	BUG_ON(ch->release_done);
3520 	ch->release_done = &release_done;
3521 	__srpt_close_ch(ch);
3522 	spin_unlock_irq(&sdev->spinlock);
3523 
3524 	res = wait_for_completion_timeout(&release_done, 60 * HZ);
3525 	WARN_ON(res <= 0);
3526 }
3527 
3528 /**
3529  * srpt_sess_get_index() - Return the value of scsiAttIntrPortIndex (SCSI-MIB).
3530  *
3531  * A quote from RFC 4455 (SCSI-MIB) about this MIB object:
3532  * This object represents an arbitrary integer used to uniquely identify a
3533  * particular attached remote initiator port to a particular SCSI target port
3534  * within a particular SCSI target device within a particular SCSI instance.
3535  */
3536 static u32 srpt_sess_get_index(struct se_session *se_sess)
3537 {
3538 	return 0;
3539 }
3540 
3541 static void srpt_set_default_node_attrs(struct se_node_acl *nacl)
3542 {
3543 }
3544 
3545 static u32 srpt_get_task_tag(struct se_cmd *se_cmd)
3546 {
3547 	struct srpt_send_ioctx *ioctx;
3548 
3549 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
3550 	return ioctx->tag;
3551 }
3552 
3553 /* Note: only used from inside debug printk's by the TCM core. */
3554 static int srpt_get_tcm_cmd_state(struct se_cmd *se_cmd)
3555 {
3556 	struct srpt_send_ioctx *ioctx;
3557 
3558 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
3559 	return srpt_get_cmd_state(ioctx);
3560 }
3561 
3562 /**
3563  * srpt_parse_i_port_id() - Parse an initiator port ID.
3564  * @name: ASCII representation of a 128-bit initiator port ID.
3565  * @i_port_id: Binary 128-bit port ID.
3566  */
3567 static int srpt_parse_i_port_id(u8 i_port_id[16], const char *name)
3568 {
3569 	const char *p;
3570 	unsigned len, count, leading_zero_bytes;
3571 	int ret, rc;
3572 
3573 	p = name;
3574 	if (strnicmp(p, "0x", 2) == 0)
3575 		p += 2;
3576 	ret = -EINVAL;
3577 	len = strlen(p);
3578 	if (len % 2)
3579 		goto out;
3580 	count = min(len / 2, 16U);
3581 	leading_zero_bytes = 16 - count;
3582 	memset(i_port_id, 0, leading_zero_bytes);
3583 	rc = hex2bin(i_port_id + leading_zero_bytes, p, count);
3584 	if (rc < 0)
3585 		pr_debug("hex2bin failed for srpt_parse_i_port_id: %d\n", rc);
3586 	ret = 0;
3587 out:
3588 	return ret;
3589 }
3590 
3591 /*
3592  * configfs callback function invoked for
3593  * mkdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3594  */
3595 static struct se_node_acl *srpt_make_nodeacl(struct se_portal_group *tpg,
3596 					     struct config_group *group,
3597 					     const char *name)
3598 {
3599 	struct srpt_port *sport = container_of(tpg, struct srpt_port, port_tpg_1);
3600 	struct se_node_acl *se_nacl, *se_nacl_new;
3601 	struct srpt_node_acl *nacl;
3602 	int ret = 0;
3603 	u32 nexus_depth = 1;
3604 	u8 i_port_id[16];
3605 
3606 	if (srpt_parse_i_port_id(i_port_id, name) < 0) {
3607 		printk(KERN_ERR "invalid initiator port ID %s\n", name);
3608 		ret = -EINVAL;
3609 		goto err;
3610 	}
3611 
3612 	se_nacl_new = srpt_alloc_fabric_acl(tpg);
3613 	if (!se_nacl_new) {
3614 		ret = -ENOMEM;
3615 		goto err;
3616 	}
3617 	/*
3618 	 * nacl_new may be released by core_tpg_add_initiator_node_acl()
3619 	 * when converting a node ACL from demo mode to explict
3620 	 */
3621 	se_nacl = core_tpg_add_initiator_node_acl(tpg, se_nacl_new, name,
3622 						  nexus_depth);
3623 	if (IS_ERR(se_nacl)) {
3624 		ret = PTR_ERR(se_nacl);
3625 		goto err;
3626 	}
3627 	/* Locate our struct srpt_node_acl and set sdev and i_port_id. */
3628 	nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3629 	memcpy(&nacl->i_port_id[0], &i_port_id[0], 16);
3630 	nacl->sport = sport;
3631 
3632 	spin_lock_irq(&sport->port_acl_lock);
3633 	list_add_tail(&nacl->list, &sport->port_acl_list);
3634 	spin_unlock_irq(&sport->port_acl_lock);
3635 
3636 	return se_nacl;
3637 err:
3638 	return ERR_PTR(ret);
3639 }
3640 
3641 /*
3642  * configfs callback function invoked for
3643  * rmdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3644  */
3645 static void srpt_drop_nodeacl(struct se_node_acl *se_nacl)
3646 {
3647 	struct srpt_node_acl *nacl;
3648 	struct srpt_device *sdev;
3649 	struct srpt_port *sport;
3650 
3651 	nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3652 	sport = nacl->sport;
3653 	sdev = sport->sdev;
3654 	spin_lock_irq(&sport->port_acl_lock);
3655 	list_del(&nacl->list);
3656 	spin_unlock_irq(&sport->port_acl_lock);
3657 	core_tpg_del_initiator_node_acl(&sport->port_tpg_1, se_nacl, 1);
3658 	srpt_release_fabric_acl(NULL, se_nacl);
3659 }
3660 
3661 static ssize_t srpt_tpg_attrib_show_srp_max_rdma_size(
3662 	struct se_portal_group *se_tpg,
3663 	char *page)
3664 {
3665 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3666 
3667 	return sprintf(page, "%u\n", sport->port_attrib.srp_max_rdma_size);
3668 }
3669 
3670 static ssize_t srpt_tpg_attrib_store_srp_max_rdma_size(
3671 	struct se_portal_group *se_tpg,
3672 	const char *page,
3673 	size_t count)
3674 {
3675 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3676 	unsigned long val;
3677 	int ret;
3678 
3679 	ret = strict_strtoul(page, 0, &val);
3680 	if (ret < 0) {
3681 		pr_err("strict_strtoul() failed with ret: %d\n", ret);
3682 		return -EINVAL;
3683 	}
3684 	if (val > MAX_SRPT_RDMA_SIZE) {
3685 		pr_err("val: %lu exceeds MAX_SRPT_RDMA_SIZE: %d\n", val,
3686 			MAX_SRPT_RDMA_SIZE);
3687 		return -EINVAL;
3688 	}
3689 	if (val < DEFAULT_MAX_RDMA_SIZE) {
3690 		pr_err("val: %lu smaller than DEFAULT_MAX_RDMA_SIZE: %d\n",
3691 			val, DEFAULT_MAX_RDMA_SIZE);
3692 		return -EINVAL;
3693 	}
3694 	sport->port_attrib.srp_max_rdma_size = val;
3695 
3696 	return count;
3697 }
3698 
3699 TF_TPG_ATTRIB_ATTR(srpt, srp_max_rdma_size, S_IRUGO | S_IWUSR);
3700 
3701 static ssize_t srpt_tpg_attrib_show_srp_max_rsp_size(
3702 	struct se_portal_group *se_tpg,
3703 	char *page)
3704 {
3705 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3706 
3707 	return sprintf(page, "%u\n", sport->port_attrib.srp_max_rsp_size);
3708 }
3709 
3710 static ssize_t srpt_tpg_attrib_store_srp_max_rsp_size(
3711 	struct se_portal_group *se_tpg,
3712 	const char *page,
3713 	size_t count)
3714 {
3715 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3716 	unsigned long val;
3717 	int ret;
3718 
3719 	ret = strict_strtoul(page, 0, &val);
3720 	if (ret < 0) {
3721 		pr_err("strict_strtoul() failed with ret: %d\n", ret);
3722 		return -EINVAL;
3723 	}
3724 	if (val > MAX_SRPT_RSP_SIZE) {
3725 		pr_err("val: %lu exceeds MAX_SRPT_RSP_SIZE: %d\n", val,
3726 			MAX_SRPT_RSP_SIZE);
3727 		return -EINVAL;
3728 	}
3729 	if (val < MIN_MAX_RSP_SIZE) {
3730 		pr_err("val: %lu smaller than MIN_MAX_RSP_SIZE: %d\n", val,
3731 			MIN_MAX_RSP_SIZE);
3732 		return -EINVAL;
3733 	}
3734 	sport->port_attrib.srp_max_rsp_size = val;
3735 
3736 	return count;
3737 }
3738 
3739 TF_TPG_ATTRIB_ATTR(srpt, srp_max_rsp_size, S_IRUGO | S_IWUSR);
3740 
3741 static ssize_t srpt_tpg_attrib_show_srp_sq_size(
3742 	struct se_portal_group *se_tpg,
3743 	char *page)
3744 {
3745 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3746 
3747 	return sprintf(page, "%u\n", sport->port_attrib.srp_sq_size);
3748 }
3749 
3750 static ssize_t srpt_tpg_attrib_store_srp_sq_size(
3751 	struct se_portal_group *se_tpg,
3752 	const char *page,
3753 	size_t count)
3754 {
3755 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3756 	unsigned long val;
3757 	int ret;
3758 
3759 	ret = strict_strtoul(page, 0, &val);
3760 	if (ret < 0) {
3761 		pr_err("strict_strtoul() failed with ret: %d\n", ret);
3762 		return -EINVAL;
3763 	}
3764 	if (val > MAX_SRPT_SRQ_SIZE) {
3765 		pr_err("val: %lu exceeds MAX_SRPT_SRQ_SIZE: %d\n", val,
3766 			MAX_SRPT_SRQ_SIZE);
3767 		return -EINVAL;
3768 	}
3769 	if (val < MIN_SRPT_SRQ_SIZE) {
3770 		pr_err("val: %lu smaller than MIN_SRPT_SRQ_SIZE: %d\n", val,
3771 			MIN_SRPT_SRQ_SIZE);
3772 		return -EINVAL;
3773 	}
3774 	sport->port_attrib.srp_sq_size = val;
3775 
3776 	return count;
3777 }
3778 
3779 TF_TPG_ATTRIB_ATTR(srpt, srp_sq_size, S_IRUGO | S_IWUSR);
3780 
3781 static struct configfs_attribute *srpt_tpg_attrib_attrs[] = {
3782 	&srpt_tpg_attrib_srp_max_rdma_size.attr,
3783 	&srpt_tpg_attrib_srp_max_rsp_size.attr,
3784 	&srpt_tpg_attrib_srp_sq_size.attr,
3785 	NULL,
3786 };
3787 
3788 static ssize_t srpt_tpg_show_enable(
3789 	struct se_portal_group *se_tpg,
3790 	char *page)
3791 {
3792 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3793 
3794 	return snprintf(page, PAGE_SIZE, "%d\n", (sport->enabled) ? 1: 0);
3795 }
3796 
3797 static ssize_t srpt_tpg_store_enable(
3798 	struct se_portal_group *se_tpg,
3799 	const char *page,
3800 	size_t count)
3801 {
3802 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3803 	unsigned long tmp;
3804         int ret;
3805 
3806 	ret = strict_strtoul(page, 0, &tmp);
3807 	if (ret < 0) {
3808 		printk(KERN_ERR "Unable to extract srpt_tpg_store_enable\n");
3809 		return -EINVAL;
3810 	}
3811 
3812 	if ((tmp != 0) && (tmp != 1)) {
3813 		printk(KERN_ERR "Illegal value for srpt_tpg_store_enable: %lu\n", tmp);
3814 		return -EINVAL;
3815 	}
3816 	if (tmp == 1)
3817 		sport->enabled = true;
3818 	else
3819 		sport->enabled = false;
3820 
3821 	return count;
3822 }
3823 
3824 TF_TPG_BASE_ATTR(srpt, enable, S_IRUGO | S_IWUSR);
3825 
3826 static struct configfs_attribute *srpt_tpg_attrs[] = {
3827 	&srpt_tpg_enable.attr,
3828 	NULL,
3829 };
3830 
3831 /**
3832  * configfs callback invoked for
3833  * mkdir /sys/kernel/config/target/$driver/$port/$tpg
3834  */
3835 static struct se_portal_group *srpt_make_tpg(struct se_wwn *wwn,
3836 					     struct config_group *group,
3837 					     const char *name)
3838 {
3839 	struct srpt_port *sport = container_of(wwn, struct srpt_port, port_wwn);
3840 	int res;
3841 
3842 	/* Initialize sport->port_wwn and sport->port_tpg_1 */
3843 	res = core_tpg_register(&srpt_target->tf_ops, &sport->port_wwn,
3844 			&sport->port_tpg_1, sport, TRANSPORT_TPG_TYPE_NORMAL);
3845 	if (res)
3846 		return ERR_PTR(res);
3847 
3848 	return &sport->port_tpg_1;
3849 }
3850 
3851 /**
3852  * configfs callback invoked for
3853  * rmdir /sys/kernel/config/target/$driver/$port/$tpg
3854  */
3855 static void srpt_drop_tpg(struct se_portal_group *tpg)
3856 {
3857 	struct srpt_port *sport = container_of(tpg,
3858 				struct srpt_port, port_tpg_1);
3859 
3860 	sport->enabled = false;
3861 	core_tpg_deregister(&sport->port_tpg_1);
3862 }
3863 
3864 /**
3865  * configfs callback invoked for
3866  * mkdir /sys/kernel/config/target/$driver/$port
3867  */
3868 static struct se_wwn *srpt_make_tport(struct target_fabric_configfs *tf,
3869 				      struct config_group *group,
3870 				      const char *name)
3871 {
3872 	struct srpt_port *sport;
3873 	int ret;
3874 
3875 	sport = srpt_lookup_port(name);
3876 	pr_debug("make_tport(%s)\n", name);
3877 	ret = -EINVAL;
3878 	if (!sport)
3879 		goto err;
3880 
3881 	return &sport->port_wwn;
3882 
3883 err:
3884 	return ERR_PTR(ret);
3885 }
3886 
3887 /**
3888  * configfs callback invoked for
3889  * rmdir /sys/kernel/config/target/$driver/$port
3890  */
3891 static void srpt_drop_tport(struct se_wwn *wwn)
3892 {
3893 	struct srpt_port *sport = container_of(wwn, struct srpt_port, port_wwn);
3894 
3895 	pr_debug("drop_tport(%s\n", config_item_name(&sport->port_wwn.wwn_group.cg_item));
3896 }
3897 
3898 static ssize_t srpt_wwn_show_attr_version(struct target_fabric_configfs *tf,
3899 					      char *buf)
3900 {
3901 	return scnprintf(buf, PAGE_SIZE, "%s\n", DRV_VERSION);
3902 }
3903 
3904 TF_WWN_ATTR_RO(srpt, version);
3905 
3906 static struct configfs_attribute *srpt_wwn_attrs[] = {
3907 	&srpt_wwn_version.attr,
3908 	NULL,
3909 };
3910 
3911 static struct target_core_fabric_ops srpt_template = {
3912 	.get_fabric_name		= srpt_get_fabric_name,
3913 	.get_fabric_proto_ident		= srpt_get_fabric_proto_ident,
3914 	.tpg_get_wwn			= srpt_get_fabric_wwn,
3915 	.tpg_get_tag			= srpt_get_tag,
3916 	.tpg_get_default_depth		= srpt_get_default_depth,
3917 	.tpg_get_pr_transport_id	= srpt_get_pr_transport_id,
3918 	.tpg_get_pr_transport_id_len	= srpt_get_pr_transport_id_len,
3919 	.tpg_parse_pr_out_transport_id	= srpt_parse_pr_out_transport_id,
3920 	.tpg_check_demo_mode		= srpt_check_false,
3921 	.tpg_check_demo_mode_cache	= srpt_check_true,
3922 	.tpg_check_demo_mode_write_protect = srpt_check_true,
3923 	.tpg_check_prod_mode_write_protect = srpt_check_false,
3924 	.tpg_alloc_fabric_acl		= srpt_alloc_fabric_acl,
3925 	.tpg_release_fabric_acl		= srpt_release_fabric_acl,
3926 	.tpg_get_inst_index		= srpt_tpg_get_inst_index,
3927 	.release_cmd			= srpt_release_cmd,
3928 	.check_stop_free		= srpt_check_stop_free,
3929 	.shutdown_session		= srpt_shutdown_session,
3930 	.close_session			= srpt_close_session,
3931 	.sess_get_index			= srpt_sess_get_index,
3932 	.sess_get_initiator_sid		= NULL,
3933 	.write_pending			= srpt_write_pending,
3934 	.write_pending_status		= srpt_write_pending_status,
3935 	.set_default_node_attributes	= srpt_set_default_node_attrs,
3936 	.get_task_tag			= srpt_get_task_tag,
3937 	.get_cmd_state			= srpt_get_tcm_cmd_state,
3938 	.queue_data_in			= srpt_queue_data_in,
3939 	.queue_status			= srpt_queue_status,
3940 	.queue_tm_rsp			= srpt_queue_tm_rsp,
3941 	/*
3942 	 * Setup function pointers for generic logic in
3943 	 * target_core_fabric_configfs.c
3944 	 */
3945 	.fabric_make_wwn		= srpt_make_tport,
3946 	.fabric_drop_wwn		= srpt_drop_tport,
3947 	.fabric_make_tpg		= srpt_make_tpg,
3948 	.fabric_drop_tpg		= srpt_drop_tpg,
3949 	.fabric_post_link		= NULL,
3950 	.fabric_pre_unlink		= NULL,
3951 	.fabric_make_np			= NULL,
3952 	.fabric_drop_np			= NULL,
3953 	.fabric_make_nodeacl		= srpt_make_nodeacl,
3954 	.fabric_drop_nodeacl		= srpt_drop_nodeacl,
3955 };
3956 
3957 /**
3958  * srpt_init_module() - Kernel module initialization.
3959  *
3960  * Note: Since ib_register_client() registers callback functions, and since at
3961  * least one of these callback functions (srpt_add_one()) calls target core
3962  * functions, this driver must be registered with the target core before
3963  * ib_register_client() is called.
3964  */
3965 static int __init srpt_init_module(void)
3966 {
3967 	int ret;
3968 
3969 	ret = -EINVAL;
3970 	if (srp_max_req_size < MIN_MAX_REQ_SIZE) {
3971 		printk(KERN_ERR "invalid value %d for kernel module parameter"
3972 		       " srp_max_req_size -- must be at least %d.\n",
3973 		       srp_max_req_size, MIN_MAX_REQ_SIZE);
3974 		goto out;
3975 	}
3976 
3977 	if (srpt_srq_size < MIN_SRPT_SRQ_SIZE
3978 	    || srpt_srq_size > MAX_SRPT_SRQ_SIZE) {
3979 		printk(KERN_ERR "invalid value %d for kernel module parameter"
3980 		       " srpt_srq_size -- must be in the range [%d..%d].\n",
3981 		       srpt_srq_size, MIN_SRPT_SRQ_SIZE, MAX_SRPT_SRQ_SIZE);
3982 		goto out;
3983 	}
3984 
3985 	srpt_target = target_fabric_configfs_init(THIS_MODULE, "srpt");
3986 	if (IS_ERR(srpt_target)) {
3987 		printk(KERN_ERR "couldn't register\n");
3988 		ret = PTR_ERR(srpt_target);
3989 		goto out;
3990 	}
3991 
3992 	srpt_target->tf_ops = srpt_template;
3993 
3994 	/*
3995 	 * Set up default attribute lists.
3996 	 */
3997 	srpt_target->tf_cit_tmpl.tfc_wwn_cit.ct_attrs = srpt_wwn_attrs;
3998 	srpt_target->tf_cit_tmpl.tfc_tpg_base_cit.ct_attrs = srpt_tpg_attrs;
3999 	srpt_target->tf_cit_tmpl.tfc_tpg_attrib_cit.ct_attrs = srpt_tpg_attrib_attrs;
4000 	srpt_target->tf_cit_tmpl.tfc_tpg_param_cit.ct_attrs = NULL;
4001 	srpt_target->tf_cit_tmpl.tfc_tpg_np_base_cit.ct_attrs = NULL;
4002 	srpt_target->tf_cit_tmpl.tfc_tpg_nacl_base_cit.ct_attrs = NULL;
4003 	srpt_target->tf_cit_tmpl.tfc_tpg_nacl_attrib_cit.ct_attrs = NULL;
4004 	srpt_target->tf_cit_tmpl.tfc_tpg_nacl_auth_cit.ct_attrs = NULL;
4005 	srpt_target->tf_cit_tmpl.tfc_tpg_nacl_param_cit.ct_attrs = NULL;
4006 
4007 	ret = target_fabric_configfs_register(srpt_target);
4008 	if (ret < 0) {
4009 		printk(KERN_ERR "couldn't register\n");
4010 		goto out_free_target;
4011 	}
4012 
4013 	ret = ib_register_client(&srpt_client);
4014 	if (ret) {
4015 		printk(KERN_ERR "couldn't register IB client\n");
4016 		goto out_unregister_target;
4017 	}
4018 
4019 	return 0;
4020 
4021 out_unregister_target:
4022 	target_fabric_configfs_deregister(srpt_target);
4023 	srpt_target = NULL;
4024 out_free_target:
4025 	if (srpt_target)
4026 		target_fabric_configfs_free(srpt_target);
4027 out:
4028 	return ret;
4029 }
4030 
4031 static void __exit srpt_cleanup_module(void)
4032 {
4033 	ib_unregister_client(&srpt_client);
4034 	target_fabric_configfs_deregister(srpt_target);
4035 	srpt_target = NULL;
4036 }
4037 
4038 module_init(srpt_init_module);
4039 module_exit(srpt_cleanup_module);
4040