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_ACTIVE;
1356 		spin_unlock_irqrestore(&ioctx->cmd.t_state_lock, flags);
1357 		break;
1358 	case SRPT_STATE_CMD_RSP_SENT:
1359 		/*
1360 		 * SRP_RSP sending failed or the SRP_RSP send completion has
1361 		 * not been received in time.
1362 		 */
1363 		srpt_unmap_sg_to_ib_sge(ioctx->ch, ioctx);
1364 		target_put_sess_cmd(ioctx->ch->sess, &ioctx->cmd);
1365 		break;
1366 	case SRPT_STATE_MGMT_RSP_SENT:
1367 		srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
1368 		target_put_sess_cmd(ioctx->ch->sess, &ioctx->cmd);
1369 		break;
1370 	default:
1371 		WARN(1, "Unexpected command state (%d)", state);
1372 		break;
1373 	}
1374 
1375 out:
1376 	return state;
1377 }
1378 
1379 /**
1380  * srpt_handle_send_err_comp() - Process an IB_WC_SEND error completion.
1381  */
1382 static void srpt_handle_send_err_comp(struct srpt_rdma_ch *ch, u64 wr_id)
1383 {
1384 	struct srpt_send_ioctx *ioctx;
1385 	enum srpt_command_state state;
1386 	struct se_cmd *cmd;
1387 	u32 index;
1388 
1389 	atomic_inc(&ch->sq_wr_avail);
1390 
1391 	index = idx_from_wr_id(wr_id);
1392 	ioctx = ch->ioctx_ring[index];
1393 	state = srpt_get_cmd_state(ioctx);
1394 	cmd = &ioctx->cmd;
1395 
1396 	WARN_ON(state != SRPT_STATE_CMD_RSP_SENT
1397 		&& state != SRPT_STATE_MGMT_RSP_SENT
1398 		&& state != SRPT_STATE_NEED_DATA
1399 		&& state != SRPT_STATE_DONE);
1400 
1401 	/* If SRP_RSP sending failed, undo the ch->req_lim change. */
1402 	if (state == SRPT_STATE_CMD_RSP_SENT
1403 	    || state == SRPT_STATE_MGMT_RSP_SENT)
1404 		atomic_dec(&ch->req_lim);
1405 
1406 	srpt_abort_cmd(ioctx);
1407 }
1408 
1409 /**
1410  * srpt_handle_send_comp() - Process an IB send completion notification.
1411  */
1412 static void srpt_handle_send_comp(struct srpt_rdma_ch *ch,
1413 				  struct srpt_send_ioctx *ioctx)
1414 {
1415 	enum srpt_command_state state;
1416 
1417 	atomic_inc(&ch->sq_wr_avail);
1418 
1419 	state = srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
1420 
1421 	if (WARN_ON(state != SRPT_STATE_CMD_RSP_SENT
1422 		    && state != SRPT_STATE_MGMT_RSP_SENT
1423 		    && state != SRPT_STATE_DONE))
1424 		pr_debug("state = %d\n", state);
1425 
1426 	if (state != SRPT_STATE_DONE) {
1427 		srpt_unmap_sg_to_ib_sge(ch, ioctx);
1428 		transport_generic_free_cmd(&ioctx->cmd, 0);
1429 	} else {
1430 		printk(KERN_ERR "IB completion has been received too late for"
1431 		       " wr_id = %u.\n", ioctx->ioctx.index);
1432 	}
1433 }
1434 
1435 /**
1436  * srpt_handle_rdma_comp() - Process an IB RDMA completion notification.
1437  *
1438  * XXX: what is now target_execute_cmd used to be asynchronous, and unmapping
1439  * the data that has been transferred via IB RDMA had to be postponed until the
1440  * check_stop_free() callback.  None of this is necessary anymore and needs to
1441  * be cleaned up.
1442  */
1443 static void srpt_handle_rdma_comp(struct srpt_rdma_ch *ch,
1444 				  struct srpt_send_ioctx *ioctx,
1445 				  enum srpt_opcode opcode)
1446 {
1447 	WARN_ON(ioctx->n_rdma <= 0);
1448 	atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
1449 
1450 	if (opcode == SRPT_RDMA_READ_LAST) {
1451 		if (srpt_test_and_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA,
1452 						SRPT_STATE_DATA_IN))
1453 			target_execute_cmd(&ioctx->cmd);
1454 		else
1455 			printk(KERN_ERR "%s[%d]: wrong state = %d\n", __func__,
1456 			       __LINE__, srpt_get_cmd_state(ioctx));
1457 	} else if (opcode == SRPT_RDMA_ABORT) {
1458 		ioctx->rdma_aborted = true;
1459 	} else {
1460 		WARN(true, "unexpected opcode %d\n", opcode);
1461 	}
1462 }
1463 
1464 /**
1465  * srpt_handle_rdma_err_comp() - Process an IB RDMA error completion.
1466  */
1467 static void srpt_handle_rdma_err_comp(struct srpt_rdma_ch *ch,
1468 				      struct srpt_send_ioctx *ioctx,
1469 				      enum srpt_opcode opcode)
1470 {
1471 	struct se_cmd *cmd;
1472 	enum srpt_command_state state;
1473 
1474 	cmd = &ioctx->cmd;
1475 	state = srpt_get_cmd_state(ioctx);
1476 	switch (opcode) {
1477 	case SRPT_RDMA_READ_LAST:
1478 		if (ioctx->n_rdma <= 0) {
1479 			printk(KERN_ERR "Received invalid RDMA read"
1480 			       " error completion with idx %d\n",
1481 			       ioctx->ioctx.index);
1482 			break;
1483 		}
1484 		atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
1485 		if (state == SRPT_STATE_NEED_DATA)
1486 			srpt_abort_cmd(ioctx);
1487 		else
1488 			printk(KERN_ERR "%s[%d]: wrong state = %d\n",
1489 			       __func__, __LINE__, state);
1490 		break;
1491 	case SRPT_RDMA_WRITE_LAST:
1492 		break;
1493 	default:
1494 		printk(KERN_ERR "%s[%d]: opcode = %u\n", __func__,
1495 		       __LINE__, opcode);
1496 		break;
1497 	}
1498 }
1499 
1500 /**
1501  * srpt_build_cmd_rsp() - Build an SRP_RSP response.
1502  * @ch: RDMA channel through which the request has been received.
1503  * @ioctx: I/O context associated with the SRP_CMD request. The response will
1504  *   be built in the buffer ioctx->buf points at and hence this function will
1505  *   overwrite the request data.
1506  * @tag: tag of the request for which this response is being generated.
1507  * @status: value for the STATUS field of the SRP_RSP information unit.
1508  *
1509  * Returns the size in bytes of the SRP_RSP response.
1510  *
1511  * An SRP_RSP response contains a SCSI status or service response. See also
1512  * section 6.9 in the SRP r16a document for the format of an SRP_RSP
1513  * response. See also SPC-2 for more information about sense data.
1514  */
1515 static int srpt_build_cmd_rsp(struct srpt_rdma_ch *ch,
1516 			      struct srpt_send_ioctx *ioctx, u64 tag,
1517 			      int status)
1518 {
1519 	struct srp_rsp *srp_rsp;
1520 	const u8 *sense_data;
1521 	int sense_data_len, max_sense_len;
1522 
1523 	/*
1524 	 * The lowest bit of all SAM-3 status codes is zero (see also
1525 	 * paragraph 5.3 in SAM-3).
1526 	 */
1527 	WARN_ON(status & 1);
1528 
1529 	srp_rsp = ioctx->ioctx.buf;
1530 	BUG_ON(!srp_rsp);
1531 
1532 	sense_data = ioctx->sense_data;
1533 	sense_data_len = ioctx->cmd.scsi_sense_length;
1534 	WARN_ON(sense_data_len > sizeof(ioctx->sense_data));
1535 
1536 	memset(srp_rsp, 0, sizeof *srp_rsp);
1537 	srp_rsp->opcode = SRP_RSP;
1538 	srp_rsp->req_lim_delta =
1539 		__constant_cpu_to_be32(1 + atomic_xchg(&ch->req_lim_delta, 0));
1540 	srp_rsp->tag = tag;
1541 	srp_rsp->status = status;
1542 
1543 	if (sense_data_len) {
1544 		BUILD_BUG_ON(MIN_MAX_RSP_SIZE <= sizeof(*srp_rsp));
1545 		max_sense_len = ch->max_ti_iu_len - sizeof(*srp_rsp);
1546 		if (sense_data_len > max_sense_len) {
1547 			printk(KERN_WARNING "truncated sense data from %d to %d"
1548 			       " bytes\n", sense_data_len, max_sense_len);
1549 			sense_data_len = max_sense_len;
1550 		}
1551 
1552 		srp_rsp->flags |= SRP_RSP_FLAG_SNSVALID;
1553 		srp_rsp->sense_data_len = cpu_to_be32(sense_data_len);
1554 		memcpy(srp_rsp + 1, sense_data, sense_data_len);
1555 	}
1556 
1557 	return sizeof(*srp_rsp) + sense_data_len;
1558 }
1559 
1560 /**
1561  * srpt_build_tskmgmt_rsp() - Build a task management response.
1562  * @ch:       RDMA channel through which the request has been received.
1563  * @ioctx:    I/O context in which the SRP_RSP response will be built.
1564  * @rsp_code: RSP_CODE that will be stored in the response.
1565  * @tag:      Tag of the request for which this response is being generated.
1566  *
1567  * Returns the size in bytes of the SRP_RSP response.
1568  *
1569  * An SRP_RSP response contains a SCSI status or service response. See also
1570  * section 6.9 in the SRP r16a document for the format of an SRP_RSP
1571  * response.
1572  */
1573 static int srpt_build_tskmgmt_rsp(struct srpt_rdma_ch *ch,
1574 				  struct srpt_send_ioctx *ioctx,
1575 				  u8 rsp_code, u64 tag)
1576 {
1577 	struct srp_rsp *srp_rsp;
1578 	int resp_data_len;
1579 	int resp_len;
1580 
1581 	resp_data_len = 4;
1582 	resp_len = sizeof(*srp_rsp) + resp_data_len;
1583 
1584 	srp_rsp = ioctx->ioctx.buf;
1585 	BUG_ON(!srp_rsp);
1586 	memset(srp_rsp, 0, sizeof *srp_rsp);
1587 
1588 	srp_rsp->opcode = SRP_RSP;
1589 	srp_rsp->req_lim_delta = __constant_cpu_to_be32(1
1590 				    + atomic_xchg(&ch->req_lim_delta, 0));
1591 	srp_rsp->tag = tag;
1592 
1593 	srp_rsp->flags |= SRP_RSP_FLAG_RSPVALID;
1594 	srp_rsp->resp_data_len = cpu_to_be32(resp_data_len);
1595 	srp_rsp->data[3] = rsp_code;
1596 
1597 	return resp_len;
1598 }
1599 
1600 #define NO_SUCH_LUN ((uint64_t)-1LL)
1601 
1602 /*
1603  * SCSI LUN addressing method. See also SAM-2 and the section about
1604  * eight byte LUNs.
1605  */
1606 enum scsi_lun_addr_method {
1607 	SCSI_LUN_ADDR_METHOD_PERIPHERAL   = 0,
1608 	SCSI_LUN_ADDR_METHOD_FLAT         = 1,
1609 	SCSI_LUN_ADDR_METHOD_LUN          = 2,
1610 	SCSI_LUN_ADDR_METHOD_EXTENDED_LUN = 3,
1611 };
1612 
1613 /*
1614  * srpt_unpack_lun() - Convert from network LUN to linear LUN.
1615  *
1616  * Convert an 2-byte, 4-byte, 6-byte or 8-byte LUN structure in network byte
1617  * order (big endian) to a linear LUN. Supports three LUN addressing methods:
1618  * peripheral, flat and logical unit. See also SAM-2, section 4.9.4 (page 40).
1619  */
1620 static uint64_t srpt_unpack_lun(const uint8_t *lun, int len)
1621 {
1622 	uint64_t res = NO_SUCH_LUN;
1623 	int addressing_method;
1624 
1625 	if (unlikely(len < 2)) {
1626 		printk(KERN_ERR "Illegal LUN length %d, expected 2 bytes or "
1627 		       "more", len);
1628 		goto out;
1629 	}
1630 
1631 	switch (len) {
1632 	case 8:
1633 		if ((*((__be64 *)lun) &
1634 		     __constant_cpu_to_be64(0x0000FFFFFFFFFFFFLL)) != 0)
1635 			goto out_err;
1636 		break;
1637 	case 4:
1638 		if (*((__be16 *)&lun[2]) != 0)
1639 			goto out_err;
1640 		break;
1641 	case 6:
1642 		if (*((__be32 *)&lun[2]) != 0)
1643 			goto out_err;
1644 		break;
1645 	case 2:
1646 		break;
1647 	default:
1648 		goto out_err;
1649 	}
1650 
1651 	addressing_method = (*lun) >> 6; /* highest two bits of byte 0 */
1652 	switch (addressing_method) {
1653 	case SCSI_LUN_ADDR_METHOD_PERIPHERAL:
1654 	case SCSI_LUN_ADDR_METHOD_FLAT:
1655 	case SCSI_LUN_ADDR_METHOD_LUN:
1656 		res = *(lun + 1) | (((*lun) & 0x3f) << 8);
1657 		break;
1658 
1659 	case SCSI_LUN_ADDR_METHOD_EXTENDED_LUN:
1660 	default:
1661 		printk(KERN_ERR "Unimplemented LUN addressing method %u",
1662 		       addressing_method);
1663 		break;
1664 	}
1665 
1666 out:
1667 	return res;
1668 
1669 out_err:
1670 	printk(KERN_ERR "Support for multi-level LUNs has not yet been"
1671 	       " implemented");
1672 	goto out;
1673 }
1674 
1675 static int srpt_check_stop_free(struct se_cmd *cmd)
1676 {
1677 	struct srpt_send_ioctx *ioctx = container_of(cmd,
1678 				struct srpt_send_ioctx, cmd);
1679 
1680 	return target_put_sess_cmd(ioctx->ch->sess, &ioctx->cmd);
1681 }
1682 
1683 /**
1684  * srpt_handle_cmd() - Process SRP_CMD.
1685  */
1686 static int srpt_handle_cmd(struct srpt_rdma_ch *ch,
1687 			   struct srpt_recv_ioctx *recv_ioctx,
1688 			   struct srpt_send_ioctx *send_ioctx)
1689 {
1690 	struct se_cmd *cmd;
1691 	struct srp_cmd *srp_cmd;
1692 	uint64_t unpacked_lun;
1693 	u64 data_len;
1694 	enum dma_data_direction dir;
1695 	sense_reason_t ret;
1696 	int rc;
1697 
1698 	BUG_ON(!send_ioctx);
1699 
1700 	srp_cmd = recv_ioctx->ioctx.buf;
1701 	cmd = &send_ioctx->cmd;
1702 	send_ioctx->tag = srp_cmd->tag;
1703 
1704 	switch (srp_cmd->task_attr) {
1705 	case SRP_CMD_SIMPLE_Q:
1706 		cmd->sam_task_attr = MSG_SIMPLE_TAG;
1707 		break;
1708 	case SRP_CMD_ORDERED_Q:
1709 	default:
1710 		cmd->sam_task_attr = MSG_ORDERED_TAG;
1711 		break;
1712 	case SRP_CMD_HEAD_OF_Q:
1713 		cmd->sam_task_attr = MSG_HEAD_TAG;
1714 		break;
1715 	case SRP_CMD_ACA:
1716 		cmd->sam_task_attr = MSG_ACA_TAG;
1717 		break;
1718 	}
1719 
1720 	if (srpt_get_desc_tbl(send_ioctx, srp_cmd, &dir, &data_len)) {
1721 		printk(KERN_ERR "0x%llx: parsing SRP descriptor table failed.\n",
1722 		       srp_cmd->tag);
1723 		ret = TCM_INVALID_CDB_FIELD;
1724 		goto send_sense;
1725 	}
1726 
1727 	unpacked_lun = srpt_unpack_lun((uint8_t *)&srp_cmd->lun,
1728 				       sizeof(srp_cmd->lun));
1729 	rc = target_submit_cmd(cmd, ch->sess, srp_cmd->cdb,
1730 			&send_ioctx->sense_data[0], unpacked_lun, data_len,
1731 			MSG_SIMPLE_TAG, dir, TARGET_SCF_ACK_KREF);
1732 	if (rc != 0) {
1733 		ret = TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
1734 		goto send_sense;
1735 	}
1736 	return 0;
1737 
1738 send_sense:
1739 	transport_send_check_condition_and_sense(cmd, ret, 0);
1740 	return -1;
1741 }
1742 
1743 /**
1744  * srpt_rx_mgmt_fn_tag() - Process a task management function by tag.
1745  * @ch: RDMA channel of the task management request.
1746  * @fn: Task management function to perform.
1747  * @req_tag: Tag of the SRP task management request.
1748  * @mgmt_ioctx: I/O context of the task management request.
1749  *
1750  * Returns zero if the target core will process the task management
1751  * request asynchronously.
1752  *
1753  * Note: It is assumed that the initiator serializes tag-based task management
1754  * requests.
1755  */
1756 static int srpt_rx_mgmt_fn_tag(struct srpt_send_ioctx *ioctx, u64 tag)
1757 {
1758 	struct srpt_device *sdev;
1759 	struct srpt_rdma_ch *ch;
1760 	struct srpt_send_ioctx *target;
1761 	int ret, i;
1762 
1763 	ret = -EINVAL;
1764 	ch = ioctx->ch;
1765 	BUG_ON(!ch);
1766 	BUG_ON(!ch->sport);
1767 	sdev = ch->sport->sdev;
1768 	BUG_ON(!sdev);
1769 	spin_lock_irq(&sdev->spinlock);
1770 	for (i = 0; i < ch->rq_size; ++i) {
1771 		target = ch->ioctx_ring[i];
1772 		if (target->cmd.se_lun == ioctx->cmd.se_lun &&
1773 		    target->tag == tag &&
1774 		    srpt_get_cmd_state(target) != SRPT_STATE_DONE) {
1775 			ret = 0;
1776 			/* now let the target core abort &target->cmd; */
1777 			break;
1778 		}
1779 	}
1780 	spin_unlock_irq(&sdev->spinlock);
1781 	return ret;
1782 }
1783 
1784 static int srp_tmr_to_tcm(int fn)
1785 {
1786 	switch (fn) {
1787 	case SRP_TSK_ABORT_TASK:
1788 		return TMR_ABORT_TASK;
1789 	case SRP_TSK_ABORT_TASK_SET:
1790 		return TMR_ABORT_TASK_SET;
1791 	case SRP_TSK_CLEAR_TASK_SET:
1792 		return TMR_CLEAR_TASK_SET;
1793 	case SRP_TSK_LUN_RESET:
1794 		return TMR_LUN_RESET;
1795 	case SRP_TSK_CLEAR_ACA:
1796 		return TMR_CLEAR_ACA;
1797 	default:
1798 		return -1;
1799 	}
1800 }
1801 
1802 /**
1803  * srpt_handle_tsk_mgmt() - Process an SRP_TSK_MGMT information unit.
1804  *
1805  * Returns 0 if and only if the request will be processed by the target core.
1806  *
1807  * For more information about SRP_TSK_MGMT information units, see also section
1808  * 6.7 in the SRP r16a document.
1809  */
1810 static void srpt_handle_tsk_mgmt(struct srpt_rdma_ch *ch,
1811 				 struct srpt_recv_ioctx *recv_ioctx,
1812 				 struct srpt_send_ioctx *send_ioctx)
1813 {
1814 	struct srp_tsk_mgmt *srp_tsk;
1815 	struct se_cmd *cmd;
1816 	struct se_session *sess = ch->sess;
1817 	uint64_t unpacked_lun;
1818 	uint32_t tag = 0;
1819 	int tcm_tmr;
1820 	int rc;
1821 
1822 	BUG_ON(!send_ioctx);
1823 
1824 	srp_tsk = recv_ioctx->ioctx.buf;
1825 	cmd = &send_ioctx->cmd;
1826 
1827 	pr_debug("recv tsk_mgmt fn %d for task_tag %lld and cmd tag %lld"
1828 		 " cm_id %p sess %p\n", srp_tsk->tsk_mgmt_func,
1829 		 srp_tsk->task_tag, srp_tsk->tag, ch->cm_id, ch->sess);
1830 
1831 	srpt_set_cmd_state(send_ioctx, SRPT_STATE_MGMT);
1832 	send_ioctx->tag = srp_tsk->tag;
1833 	tcm_tmr = srp_tmr_to_tcm(srp_tsk->tsk_mgmt_func);
1834 	if (tcm_tmr < 0) {
1835 		send_ioctx->cmd.se_tmr_req->response =
1836 			TMR_TASK_MGMT_FUNCTION_NOT_SUPPORTED;
1837 		goto fail;
1838 	}
1839 	unpacked_lun = srpt_unpack_lun((uint8_t *)&srp_tsk->lun,
1840 				       sizeof(srp_tsk->lun));
1841 
1842 	if (srp_tsk->tsk_mgmt_func == SRP_TSK_ABORT_TASK) {
1843 		rc = srpt_rx_mgmt_fn_tag(send_ioctx, srp_tsk->task_tag);
1844 		if (rc < 0) {
1845 			send_ioctx->cmd.se_tmr_req->response =
1846 					TMR_TASK_DOES_NOT_EXIST;
1847 			goto fail;
1848 		}
1849 		tag = srp_tsk->task_tag;
1850 	}
1851 	rc = target_submit_tmr(&send_ioctx->cmd, sess, NULL, unpacked_lun,
1852 				srp_tsk, tcm_tmr, GFP_KERNEL, tag,
1853 				TARGET_SCF_ACK_KREF);
1854 	if (rc != 0) {
1855 		send_ioctx->cmd.se_tmr_req->response = TMR_FUNCTION_REJECTED;
1856 		goto fail;
1857 	}
1858 	return;
1859 fail:
1860 	transport_send_check_condition_and_sense(cmd, 0, 0); // XXX:
1861 }
1862 
1863 /**
1864  * srpt_handle_new_iu() - Process a newly received information unit.
1865  * @ch:    RDMA channel through which the information unit has been received.
1866  * @ioctx: SRPT I/O context associated with the information unit.
1867  */
1868 static void srpt_handle_new_iu(struct srpt_rdma_ch *ch,
1869 			       struct srpt_recv_ioctx *recv_ioctx,
1870 			       struct srpt_send_ioctx *send_ioctx)
1871 {
1872 	struct srp_cmd *srp_cmd;
1873 	enum rdma_ch_state ch_state;
1874 
1875 	BUG_ON(!ch);
1876 	BUG_ON(!recv_ioctx);
1877 
1878 	ib_dma_sync_single_for_cpu(ch->sport->sdev->device,
1879 				   recv_ioctx->ioctx.dma, srp_max_req_size,
1880 				   DMA_FROM_DEVICE);
1881 
1882 	ch_state = srpt_get_ch_state(ch);
1883 	if (unlikely(ch_state == CH_CONNECTING)) {
1884 		list_add_tail(&recv_ioctx->wait_list, &ch->cmd_wait_list);
1885 		goto out;
1886 	}
1887 
1888 	if (unlikely(ch_state != CH_LIVE))
1889 		goto out;
1890 
1891 	srp_cmd = recv_ioctx->ioctx.buf;
1892 	if (srp_cmd->opcode == SRP_CMD || srp_cmd->opcode == SRP_TSK_MGMT) {
1893 		if (!send_ioctx)
1894 			send_ioctx = srpt_get_send_ioctx(ch);
1895 		if (unlikely(!send_ioctx)) {
1896 			list_add_tail(&recv_ioctx->wait_list,
1897 				      &ch->cmd_wait_list);
1898 			goto out;
1899 		}
1900 	}
1901 
1902 	switch (srp_cmd->opcode) {
1903 	case SRP_CMD:
1904 		srpt_handle_cmd(ch, recv_ioctx, send_ioctx);
1905 		break;
1906 	case SRP_TSK_MGMT:
1907 		srpt_handle_tsk_mgmt(ch, recv_ioctx, send_ioctx);
1908 		break;
1909 	case SRP_I_LOGOUT:
1910 		printk(KERN_ERR "Not yet implemented: SRP_I_LOGOUT\n");
1911 		break;
1912 	case SRP_CRED_RSP:
1913 		pr_debug("received SRP_CRED_RSP\n");
1914 		break;
1915 	case SRP_AER_RSP:
1916 		pr_debug("received SRP_AER_RSP\n");
1917 		break;
1918 	case SRP_RSP:
1919 		printk(KERN_ERR "Received SRP_RSP\n");
1920 		break;
1921 	default:
1922 		printk(KERN_ERR "received IU with unknown opcode 0x%x\n",
1923 		       srp_cmd->opcode);
1924 		break;
1925 	}
1926 
1927 	srpt_post_recv(ch->sport->sdev, recv_ioctx);
1928 out:
1929 	return;
1930 }
1931 
1932 static void srpt_process_rcv_completion(struct ib_cq *cq,
1933 					struct srpt_rdma_ch *ch,
1934 					struct ib_wc *wc)
1935 {
1936 	struct srpt_device *sdev = ch->sport->sdev;
1937 	struct srpt_recv_ioctx *ioctx;
1938 	u32 index;
1939 
1940 	index = idx_from_wr_id(wc->wr_id);
1941 	if (wc->status == IB_WC_SUCCESS) {
1942 		int req_lim;
1943 
1944 		req_lim = atomic_dec_return(&ch->req_lim);
1945 		if (unlikely(req_lim < 0))
1946 			printk(KERN_ERR "req_lim = %d < 0\n", req_lim);
1947 		ioctx = sdev->ioctx_ring[index];
1948 		srpt_handle_new_iu(ch, ioctx, NULL);
1949 	} else {
1950 		printk(KERN_INFO "receiving failed for idx %u with status %d\n",
1951 		       index, wc->status);
1952 	}
1953 }
1954 
1955 /**
1956  * srpt_process_send_completion() - Process an IB send completion.
1957  *
1958  * Note: Although this has not yet been observed during tests, at least in
1959  * theory it is possible that the srpt_get_send_ioctx() call invoked by
1960  * srpt_handle_new_iu() fails. This is possible because the req_lim_delta
1961  * value in each response is set to one, and it is possible that this response
1962  * makes the initiator send a new request before the send completion for that
1963  * response has been processed. This could e.g. happen if the call to
1964  * srpt_put_send_iotcx() is delayed because of a higher priority interrupt or
1965  * if IB retransmission causes generation of the send completion to be
1966  * delayed. Incoming information units for which srpt_get_send_ioctx() fails
1967  * are queued on cmd_wait_list. The code below processes these delayed
1968  * requests one at a time.
1969  */
1970 static void srpt_process_send_completion(struct ib_cq *cq,
1971 					 struct srpt_rdma_ch *ch,
1972 					 struct ib_wc *wc)
1973 {
1974 	struct srpt_send_ioctx *send_ioctx;
1975 	uint32_t index;
1976 	enum srpt_opcode opcode;
1977 
1978 	index = idx_from_wr_id(wc->wr_id);
1979 	opcode = opcode_from_wr_id(wc->wr_id);
1980 	send_ioctx = ch->ioctx_ring[index];
1981 	if (wc->status == IB_WC_SUCCESS) {
1982 		if (opcode == SRPT_SEND)
1983 			srpt_handle_send_comp(ch, send_ioctx);
1984 		else {
1985 			WARN_ON(opcode != SRPT_RDMA_ABORT &&
1986 				wc->opcode != IB_WC_RDMA_READ);
1987 			srpt_handle_rdma_comp(ch, send_ioctx, opcode);
1988 		}
1989 	} else {
1990 		if (opcode == SRPT_SEND) {
1991 			printk(KERN_INFO "sending response for idx %u failed"
1992 			       " with status %d\n", index, wc->status);
1993 			srpt_handle_send_err_comp(ch, wc->wr_id);
1994 		} else if (opcode != SRPT_RDMA_MID) {
1995 			printk(KERN_INFO "RDMA t %d for idx %u failed with"
1996 				" status %d", opcode, index, wc->status);
1997 			srpt_handle_rdma_err_comp(ch, send_ioctx, opcode);
1998 		}
1999 	}
2000 
2001 	while (unlikely(opcode == SRPT_SEND
2002 			&& !list_empty(&ch->cmd_wait_list)
2003 			&& srpt_get_ch_state(ch) == CH_LIVE
2004 			&& (send_ioctx = srpt_get_send_ioctx(ch)) != NULL)) {
2005 		struct srpt_recv_ioctx *recv_ioctx;
2006 
2007 		recv_ioctx = list_first_entry(&ch->cmd_wait_list,
2008 					      struct srpt_recv_ioctx,
2009 					      wait_list);
2010 		list_del(&recv_ioctx->wait_list);
2011 		srpt_handle_new_iu(ch, recv_ioctx, send_ioctx);
2012 	}
2013 }
2014 
2015 static void srpt_process_completion(struct ib_cq *cq, struct srpt_rdma_ch *ch)
2016 {
2017 	struct ib_wc *const wc = ch->wc;
2018 	int i, n;
2019 
2020 	WARN_ON(cq != ch->cq);
2021 
2022 	ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
2023 	while ((n = ib_poll_cq(cq, ARRAY_SIZE(ch->wc), wc)) > 0) {
2024 		for (i = 0; i < n; i++) {
2025 			if (opcode_from_wr_id(wc[i].wr_id) == SRPT_RECV)
2026 				srpt_process_rcv_completion(cq, ch, &wc[i]);
2027 			else
2028 				srpt_process_send_completion(cq, ch, &wc[i]);
2029 		}
2030 	}
2031 }
2032 
2033 /**
2034  * srpt_completion() - IB completion queue callback function.
2035  *
2036  * Notes:
2037  * - It is guaranteed that a completion handler will never be invoked
2038  *   concurrently on two different CPUs for the same completion queue. See also
2039  *   Documentation/infiniband/core_locking.txt and the implementation of
2040  *   handle_edge_irq() in kernel/irq/chip.c.
2041  * - When threaded IRQs are enabled, completion handlers are invoked in thread
2042  *   context instead of interrupt context.
2043  */
2044 static void srpt_completion(struct ib_cq *cq, void *ctx)
2045 {
2046 	struct srpt_rdma_ch *ch = ctx;
2047 
2048 	wake_up_interruptible(&ch->wait_queue);
2049 }
2050 
2051 static int srpt_compl_thread(void *arg)
2052 {
2053 	struct srpt_rdma_ch *ch;
2054 
2055 	/* Hibernation / freezing of the SRPT kernel thread is not supported. */
2056 	current->flags |= PF_NOFREEZE;
2057 
2058 	ch = arg;
2059 	BUG_ON(!ch);
2060 	printk(KERN_INFO "Session %s: kernel thread %s (PID %d) started\n",
2061 	       ch->sess_name, ch->thread->comm, current->pid);
2062 	while (!kthread_should_stop()) {
2063 		wait_event_interruptible(ch->wait_queue,
2064 			(srpt_process_completion(ch->cq, ch),
2065 			 kthread_should_stop()));
2066 	}
2067 	printk(KERN_INFO "Session %s: kernel thread %s (PID %d) stopped\n",
2068 	       ch->sess_name, ch->thread->comm, current->pid);
2069 	return 0;
2070 }
2071 
2072 /**
2073  * srpt_create_ch_ib() - Create receive and send completion queues.
2074  */
2075 static int srpt_create_ch_ib(struct srpt_rdma_ch *ch)
2076 {
2077 	struct ib_qp_init_attr *qp_init;
2078 	struct srpt_port *sport = ch->sport;
2079 	struct srpt_device *sdev = sport->sdev;
2080 	u32 srp_sq_size = sport->port_attrib.srp_sq_size;
2081 	int ret;
2082 
2083 	WARN_ON(ch->rq_size < 1);
2084 
2085 	ret = -ENOMEM;
2086 	qp_init = kzalloc(sizeof *qp_init, GFP_KERNEL);
2087 	if (!qp_init)
2088 		goto out;
2089 
2090 	ch->cq = ib_create_cq(sdev->device, srpt_completion, NULL, ch,
2091 			      ch->rq_size + srp_sq_size, 0);
2092 	if (IS_ERR(ch->cq)) {
2093 		ret = PTR_ERR(ch->cq);
2094 		printk(KERN_ERR "failed to create CQ cqe= %d ret= %d\n",
2095 		       ch->rq_size + srp_sq_size, ret);
2096 		goto out;
2097 	}
2098 
2099 	qp_init->qp_context = (void *)ch;
2100 	qp_init->event_handler
2101 		= (void(*)(struct ib_event *, void*))srpt_qp_event;
2102 	qp_init->send_cq = ch->cq;
2103 	qp_init->recv_cq = ch->cq;
2104 	qp_init->srq = sdev->srq;
2105 	qp_init->sq_sig_type = IB_SIGNAL_REQ_WR;
2106 	qp_init->qp_type = IB_QPT_RC;
2107 	qp_init->cap.max_send_wr = srp_sq_size;
2108 	qp_init->cap.max_send_sge = SRPT_DEF_SG_PER_WQE;
2109 
2110 	ch->qp = ib_create_qp(sdev->pd, qp_init);
2111 	if (IS_ERR(ch->qp)) {
2112 		ret = PTR_ERR(ch->qp);
2113 		printk(KERN_ERR "failed to create_qp ret= %d\n", ret);
2114 		goto err_destroy_cq;
2115 	}
2116 
2117 	atomic_set(&ch->sq_wr_avail, qp_init->cap.max_send_wr);
2118 
2119 	pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
2120 		 __func__, ch->cq->cqe, qp_init->cap.max_send_sge,
2121 		 qp_init->cap.max_send_wr, ch->cm_id);
2122 
2123 	ret = srpt_init_ch_qp(ch, ch->qp);
2124 	if (ret)
2125 		goto err_destroy_qp;
2126 
2127 	init_waitqueue_head(&ch->wait_queue);
2128 
2129 	pr_debug("creating thread for session %s\n", ch->sess_name);
2130 
2131 	ch->thread = kthread_run(srpt_compl_thread, ch, "ib_srpt_compl");
2132 	if (IS_ERR(ch->thread)) {
2133 		printk(KERN_ERR "failed to create kernel thread %ld\n",
2134 		       PTR_ERR(ch->thread));
2135 		ch->thread = NULL;
2136 		goto err_destroy_qp;
2137 	}
2138 
2139 out:
2140 	kfree(qp_init);
2141 	return ret;
2142 
2143 err_destroy_qp:
2144 	ib_destroy_qp(ch->qp);
2145 err_destroy_cq:
2146 	ib_destroy_cq(ch->cq);
2147 	goto out;
2148 }
2149 
2150 static void srpt_destroy_ch_ib(struct srpt_rdma_ch *ch)
2151 {
2152 	if (ch->thread)
2153 		kthread_stop(ch->thread);
2154 
2155 	ib_destroy_qp(ch->qp);
2156 	ib_destroy_cq(ch->cq);
2157 }
2158 
2159 /**
2160  * __srpt_close_ch() - Close an RDMA channel by setting the QP error state.
2161  *
2162  * Reset the QP and make sure all resources associated with the channel will
2163  * be deallocated at an appropriate time.
2164  *
2165  * Note: The caller must hold ch->sport->sdev->spinlock.
2166  */
2167 static void __srpt_close_ch(struct srpt_rdma_ch *ch)
2168 {
2169 	struct srpt_device *sdev;
2170 	enum rdma_ch_state prev_state;
2171 	unsigned long flags;
2172 
2173 	sdev = ch->sport->sdev;
2174 
2175 	spin_lock_irqsave(&ch->spinlock, flags);
2176 	prev_state = ch->state;
2177 	switch (prev_state) {
2178 	case CH_CONNECTING:
2179 	case CH_LIVE:
2180 		ch->state = CH_DISCONNECTING;
2181 		break;
2182 	default:
2183 		break;
2184 	}
2185 	spin_unlock_irqrestore(&ch->spinlock, flags);
2186 
2187 	switch (prev_state) {
2188 	case CH_CONNECTING:
2189 		ib_send_cm_rej(ch->cm_id, IB_CM_REJ_NO_RESOURCES, NULL, 0,
2190 			       NULL, 0);
2191 		/* fall through */
2192 	case CH_LIVE:
2193 		if (ib_send_cm_dreq(ch->cm_id, NULL, 0) < 0)
2194 			printk(KERN_ERR "sending CM DREQ failed.\n");
2195 		break;
2196 	case CH_DISCONNECTING:
2197 		break;
2198 	case CH_DRAINING:
2199 	case CH_RELEASING:
2200 		break;
2201 	}
2202 }
2203 
2204 /**
2205  * srpt_close_ch() - Close an RDMA channel.
2206  */
2207 static void srpt_close_ch(struct srpt_rdma_ch *ch)
2208 {
2209 	struct srpt_device *sdev;
2210 
2211 	sdev = ch->sport->sdev;
2212 	spin_lock_irq(&sdev->spinlock);
2213 	__srpt_close_ch(ch);
2214 	spin_unlock_irq(&sdev->spinlock);
2215 }
2216 
2217 /**
2218  * srpt_shutdown_session() - Whether or not a session may be shut down.
2219  */
2220 static int srpt_shutdown_session(struct se_session *se_sess)
2221 {
2222 	struct srpt_rdma_ch *ch = se_sess->fabric_sess_ptr;
2223 	unsigned long flags;
2224 
2225 	spin_lock_irqsave(&ch->spinlock, flags);
2226 	if (ch->in_shutdown) {
2227 		spin_unlock_irqrestore(&ch->spinlock, flags);
2228 		return true;
2229 	}
2230 
2231 	ch->in_shutdown = true;
2232 	target_sess_cmd_list_set_waiting(se_sess);
2233 	spin_unlock_irqrestore(&ch->spinlock, flags);
2234 
2235 	return true;
2236 }
2237 
2238 /**
2239  * srpt_drain_channel() - Drain a channel by resetting the IB queue pair.
2240  * @cm_id: Pointer to the CM ID of the channel to be drained.
2241  *
2242  * Note: Must be called from inside srpt_cm_handler to avoid a race between
2243  * accessing sdev->spinlock and the call to kfree(sdev) in srpt_remove_one()
2244  * (the caller of srpt_cm_handler holds the cm_id spinlock; srpt_remove_one()
2245  * waits until all target sessions for the associated IB device have been
2246  * unregistered and target session registration involves a call to
2247  * ib_destroy_cm_id(), which locks the cm_id spinlock and hence waits until
2248  * this function has finished).
2249  */
2250 static void srpt_drain_channel(struct ib_cm_id *cm_id)
2251 {
2252 	struct srpt_device *sdev;
2253 	struct srpt_rdma_ch *ch;
2254 	int ret;
2255 	bool do_reset = false;
2256 
2257 	WARN_ON_ONCE(irqs_disabled());
2258 
2259 	sdev = cm_id->context;
2260 	BUG_ON(!sdev);
2261 	spin_lock_irq(&sdev->spinlock);
2262 	list_for_each_entry(ch, &sdev->rch_list, list) {
2263 		if (ch->cm_id == cm_id) {
2264 			do_reset = srpt_test_and_set_ch_state(ch,
2265 					CH_CONNECTING, CH_DRAINING) ||
2266 				   srpt_test_and_set_ch_state(ch,
2267 					CH_LIVE, CH_DRAINING) ||
2268 				   srpt_test_and_set_ch_state(ch,
2269 					CH_DISCONNECTING, CH_DRAINING);
2270 			break;
2271 		}
2272 	}
2273 	spin_unlock_irq(&sdev->spinlock);
2274 
2275 	if (do_reset) {
2276 		if (ch->sess)
2277 			srpt_shutdown_session(ch->sess);
2278 
2279 		ret = srpt_ch_qp_err(ch);
2280 		if (ret < 0)
2281 			printk(KERN_ERR "Setting queue pair in error state"
2282 			       " failed: %d\n", ret);
2283 	}
2284 }
2285 
2286 /**
2287  * srpt_find_channel() - Look up an RDMA channel.
2288  * @cm_id: Pointer to the CM ID of the channel to be looked up.
2289  *
2290  * Return NULL if no matching RDMA channel has been found.
2291  */
2292 static struct srpt_rdma_ch *srpt_find_channel(struct srpt_device *sdev,
2293 					      struct ib_cm_id *cm_id)
2294 {
2295 	struct srpt_rdma_ch *ch;
2296 	bool found;
2297 
2298 	WARN_ON_ONCE(irqs_disabled());
2299 	BUG_ON(!sdev);
2300 
2301 	found = false;
2302 	spin_lock_irq(&sdev->spinlock);
2303 	list_for_each_entry(ch, &sdev->rch_list, list) {
2304 		if (ch->cm_id == cm_id) {
2305 			found = true;
2306 			break;
2307 		}
2308 	}
2309 	spin_unlock_irq(&sdev->spinlock);
2310 
2311 	return found ? ch : NULL;
2312 }
2313 
2314 /**
2315  * srpt_release_channel() - Release channel resources.
2316  *
2317  * Schedules the actual release because:
2318  * - Calling the ib_destroy_cm_id() call from inside an IB CM callback would
2319  *   trigger a deadlock.
2320  * - It is not safe to call TCM transport_* functions from interrupt context.
2321  */
2322 static void srpt_release_channel(struct srpt_rdma_ch *ch)
2323 {
2324 	schedule_work(&ch->release_work);
2325 }
2326 
2327 static void srpt_release_channel_work(struct work_struct *w)
2328 {
2329 	struct srpt_rdma_ch *ch;
2330 	struct srpt_device *sdev;
2331 	struct se_session *se_sess;
2332 
2333 	ch = container_of(w, struct srpt_rdma_ch, release_work);
2334 	pr_debug("ch = %p; ch->sess = %p; release_done = %p\n", ch, ch->sess,
2335 		 ch->release_done);
2336 
2337 	sdev = ch->sport->sdev;
2338 	BUG_ON(!sdev);
2339 
2340 	se_sess = ch->sess;
2341 	BUG_ON(!se_sess);
2342 
2343 	target_wait_for_sess_cmds(se_sess);
2344 
2345 	transport_deregister_session_configfs(se_sess);
2346 	transport_deregister_session(se_sess);
2347 	ch->sess = NULL;
2348 
2349 	ib_destroy_cm_id(ch->cm_id);
2350 
2351 	srpt_destroy_ch_ib(ch);
2352 
2353 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
2354 			     ch->sport->sdev, ch->rq_size,
2355 			     ch->rsp_size, DMA_TO_DEVICE);
2356 
2357 	spin_lock_irq(&sdev->spinlock);
2358 	list_del(&ch->list);
2359 	spin_unlock_irq(&sdev->spinlock);
2360 
2361 	if (ch->release_done)
2362 		complete(ch->release_done);
2363 
2364 	wake_up(&sdev->ch_releaseQ);
2365 
2366 	kfree(ch);
2367 }
2368 
2369 static struct srpt_node_acl *__srpt_lookup_acl(struct srpt_port *sport,
2370 					       u8 i_port_id[16])
2371 {
2372 	struct srpt_node_acl *nacl;
2373 
2374 	list_for_each_entry(nacl, &sport->port_acl_list, list)
2375 		if (memcmp(nacl->i_port_id, i_port_id,
2376 			   sizeof(nacl->i_port_id)) == 0)
2377 			return nacl;
2378 
2379 	return NULL;
2380 }
2381 
2382 static struct srpt_node_acl *srpt_lookup_acl(struct srpt_port *sport,
2383 					     u8 i_port_id[16])
2384 {
2385 	struct srpt_node_acl *nacl;
2386 
2387 	spin_lock_irq(&sport->port_acl_lock);
2388 	nacl = __srpt_lookup_acl(sport, i_port_id);
2389 	spin_unlock_irq(&sport->port_acl_lock);
2390 
2391 	return nacl;
2392 }
2393 
2394 /**
2395  * srpt_cm_req_recv() - Process the event IB_CM_REQ_RECEIVED.
2396  *
2397  * Ownership of the cm_id is transferred to the target session if this
2398  * functions returns zero. Otherwise the caller remains the owner of cm_id.
2399  */
2400 static int srpt_cm_req_recv(struct ib_cm_id *cm_id,
2401 			    struct ib_cm_req_event_param *param,
2402 			    void *private_data)
2403 {
2404 	struct srpt_device *sdev = cm_id->context;
2405 	struct srpt_port *sport = &sdev->port[param->port - 1];
2406 	struct srp_login_req *req;
2407 	struct srp_login_rsp *rsp;
2408 	struct srp_login_rej *rej;
2409 	struct ib_cm_rep_param *rep_param;
2410 	struct srpt_rdma_ch *ch, *tmp_ch;
2411 	struct srpt_node_acl *nacl;
2412 	u32 it_iu_len;
2413 	int i;
2414 	int ret = 0;
2415 
2416 	WARN_ON_ONCE(irqs_disabled());
2417 
2418 	if (WARN_ON(!sdev || !private_data))
2419 		return -EINVAL;
2420 
2421 	req = (struct srp_login_req *)private_data;
2422 
2423 	it_iu_len = be32_to_cpu(req->req_it_iu_len);
2424 
2425 	printk(KERN_INFO "Received SRP_LOGIN_REQ with i_port_id 0x%llx:0x%llx,"
2426 	       " t_port_id 0x%llx:0x%llx and it_iu_len %d on port %d"
2427 	       " (guid=0x%llx:0x%llx)\n",
2428 	       be64_to_cpu(*(__be64 *)&req->initiator_port_id[0]),
2429 	       be64_to_cpu(*(__be64 *)&req->initiator_port_id[8]),
2430 	       be64_to_cpu(*(__be64 *)&req->target_port_id[0]),
2431 	       be64_to_cpu(*(__be64 *)&req->target_port_id[8]),
2432 	       it_iu_len,
2433 	       param->port,
2434 	       be64_to_cpu(*(__be64 *)&sdev->port[param->port - 1].gid.raw[0]),
2435 	       be64_to_cpu(*(__be64 *)&sdev->port[param->port - 1].gid.raw[8]));
2436 
2437 	rsp = kzalloc(sizeof *rsp, GFP_KERNEL);
2438 	rej = kzalloc(sizeof *rej, GFP_KERNEL);
2439 	rep_param = kzalloc(sizeof *rep_param, GFP_KERNEL);
2440 
2441 	if (!rsp || !rej || !rep_param) {
2442 		ret = -ENOMEM;
2443 		goto out;
2444 	}
2445 
2446 	if (it_iu_len > srp_max_req_size || it_iu_len < 64) {
2447 		rej->reason = __constant_cpu_to_be32(
2448 				SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE);
2449 		ret = -EINVAL;
2450 		printk(KERN_ERR "rejected SRP_LOGIN_REQ because its"
2451 		       " length (%d bytes) is out of range (%d .. %d)\n",
2452 		       it_iu_len, 64, srp_max_req_size);
2453 		goto reject;
2454 	}
2455 
2456 	if (!sport->enabled) {
2457 		rej->reason = __constant_cpu_to_be32(
2458 			     SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2459 		ret = -EINVAL;
2460 		printk(KERN_ERR "rejected SRP_LOGIN_REQ because the target port"
2461 		       " has not yet been enabled\n");
2462 		goto reject;
2463 	}
2464 
2465 	if ((req->req_flags & SRP_MTCH_ACTION) == SRP_MULTICHAN_SINGLE) {
2466 		rsp->rsp_flags = SRP_LOGIN_RSP_MULTICHAN_NO_CHAN;
2467 
2468 		spin_lock_irq(&sdev->spinlock);
2469 
2470 		list_for_each_entry_safe(ch, tmp_ch, &sdev->rch_list, list) {
2471 			if (!memcmp(ch->i_port_id, req->initiator_port_id, 16)
2472 			    && !memcmp(ch->t_port_id, req->target_port_id, 16)
2473 			    && param->port == ch->sport->port
2474 			    && param->listen_id == ch->sport->sdev->cm_id
2475 			    && ch->cm_id) {
2476 				enum rdma_ch_state ch_state;
2477 
2478 				ch_state = srpt_get_ch_state(ch);
2479 				if (ch_state != CH_CONNECTING
2480 				    && ch_state != CH_LIVE)
2481 					continue;
2482 
2483 				/* found an existing channel */
2484 				pr_debug("Found existing channel %s"
2485 					 " cm_id= %p state= %d\n",
2486 					 ch->sess_name, ch->cm_id, ch_state);
2487 
2488 				__srpt_close_ch(ch);
2489 
2490 				rsp->rsp_flags =
2491 					SRP_LOGIN_RSP_MULTICHAN_TERMINATED;
2492 			}
2493 		}
2494 
2495 		spin_unlock_irq(&sdev->spinlock);
2496 
2497 	} else
2498 		rsp->rsp_flags = SRP_LOGIN_RSP_MULTICHAN_MAINTAINED;
2499 
2500 	if (*(__be64 *)req->target_port_id != cpu_to_be64(srpt_service_guid)
2501 	    || *(__be64 *)(req->target_port_id + 8) !=
2502 	       cpu_to_be64(srpt_service_guid)) {
2503 		rej->reason = __constant_cpu_to_be32(
2504 				SRP_LOGIN_REJ_UNABLE_ASSOCIATE_CHANNEL);
2505 		ret = -ENOMEM;
2506 		printk(KERN_ERR "rejected SRP_LOGIN_REQ because it"
2507 		       " has an invalid target port identifier.\n");
2508 		goto reject;
2509 	}
2510 
2511 	ch = kzalloc(sizeof *ch, GFP_KERNEL);
2512 	if (!ch) {
2513 		rej->reason = __constant_cpu_to_be32(
2514 					SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2515 		printk(KERN_ERR "rejected SRP_LOGIN_REQ because no memory.\n");
2516 		ret = -ENOMEM;
2517 		goto reject;
2518 	}
2519 
2520 	INIT_WORK(&ch->release_work, srpt_release_channel_work);
2521 	memcpy(ch->i_port_id, req->initiator_port_id, 16);
2522 	memcpy(ch->t_port_id, req->target_port_id, 16);
2523 	ch->sport = &sdev->port[param->port - 1];
2524 	ch->cm_id = cm_id;
2525 	/*
2526 	 * Avoid QUEUE_FULL conditions by limiting the number of buffers used
2527 	 * for the SRP protocol to the command queue size.
2528 	 */
2529 	ch->rq_size = SRPT_RQ_SIZE;
2530 	spin_lock_init(&ch->spinlock);
2531 	ch->state = CH_CONNECTING;
2532 	INIT_LIST_HEAD(&ch->cmd_wait_list);
2533 	ch->rsp_size = ch->sport->port_attrib.srp_max_rsp_size;
2534 
2535 	ch->ioctx_ring = (struct srpt_send_ioctx **)
2536 		srpt_alloc_ioctx_ring(ch->sport->sdev, ch->rq_size,
2537 				      sizeof(*ch->ioctx_ring[0]),
2538 				      ch->rsp_size, DMA_TO_DEVICE);
2539 	if (!ch->ioctx_ring)
2540 		goto free_ch;
2541 
2542 	INIT_LIST_HEAD(&ch->free_list);
2543 	for (i = 0; i < ch->rq_size; i++) {
2544 		ch->ioctx_ring[i]->ch = ch;
2545 		list_add_tail(&ch->ioctx_ring[i]->free_list, &ch->free_list);
2546 	}
2547 
2548 	ret = srpt_create_ch_ib(ch);
2549 	if (ret) {
2550 		rej->reason = __constant_cpu_to_be32(
2551 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2552 		printk(KERN_ERR "rejected SRP_LOGIN_REQ because creating"
2553 		       " a new RDMA channel failed.\n");
2554 		goto free_ring;
2555 	}
2556 
2557 	ret = srpt_ch_qp_rtr(ch, ch->qp);
2558 	if (ret) {
2559 		rej->reason = __constant_cpu_to_be32(
2560 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2561 		printk(KERN_ERR "rejected SRP_LOGIN_REQ because enabling"
2562 		       " RTR failed (error code = %d)\n", ret);
2563 		goto destroy_ib;
2564 	}
2565 	/*
2566 	 * Use the initator port identifier as the session name.
2567 	 */
2568 	snprintf(ch->sess_name, sizeof(ch->sess_name), "0x%016llx%016llx",
2569 			be64_to_cpu(*(__be64 *)ch->i_port_id),
2570 			be64_to_cpu(*(__be64 *)(ch->i_port_id + 8)));
2571 
2572 	pr_debug("registering session %s\n", ch->sess_name);
2573 
2574 	nacl = srpt_lookup_acl(sport, ch->i_port_id);
2575 	if (!nacl) {
2576 		printk(KERN_INFO "Rejected login because no ACL has been"
2577 		       " configured yet for initiator %s.\n", ch->sess_name);
2578 		rej->reason = __constant_cpu_to_be32(
2579 				SRP_LOGIN_REJ_CHANNEL_LIMIT_REACHED);
2580 		goto destroy_ib;
2581 	}
2582 
2583 	ch->sess = transport_init_session();
2584 	if (IS_ERR(ch->sess)) {
2585 		rej->reason = __constant_cpu_to_be32(
2586 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2587 		pr_debug("Failed to create session\n");
2588 		goto deregister_session;
2589 	}
2590 	ch->sess->se_node_acl = &nacl->nacl;
2591 	transport_register_session(&sport->port_tpg_1, &nacl->nacl, ch->sess, ch);
2592 
2593 	pr_debug("Establish connection sess=%p name=%s cm_id=%p\n", ch->sess,
2594 		 ch->sess_name, ch->cm_id);
2595 
2596 	/* create srp_login_response */
2597 	rsp->opcode = SRP_LOGIN_RSP;
2598 	rsp->tag = req->tag;
2599 	rsp->max_it_iu_len = req->req_it_iu_len;
2600 	rsp->max_ti_iu_len = req->req_it_iu_len;
2601 	ch->max_ti_iu_len = it_iu_len;
2602 	rsp->buf_fmt = __constant_cpu_to_be16(SRP_BUF_FORMAT_DIRECT
2603 					      | SRP_BUF_FORMAT_INDIRECT);
2604 	rsp->req_lim_delta = cpu_to_be32(ch->rq_size);
2605 	atomic_set(&ch->req_lim, ch->rq_size);
2606 	atomic_set(&ch->req_lim_delta, 0);
2607 
2608 	/* create cm reply */
2609 	rep_param->qp_num = ch->qp->qp_num;
2610 	rep_param->private_data = (void *)rsp;
2611 	rep_param->private_data_len = sizeof *rsp;
2612 	rep_param->rnr_retry_count = 7;
2613 	rep_param->flow_control = 1;
2614 	rep_param->failover_accepted = 0;
2615 	rep_param->srq = 1;
2616 	rep_param->responder_resources = 4;
2617 	rep_param->initiator_depth = 4;
2618 
2619 	ret = ib_send_cm_rep(cm_id, rep_param);
2620 	if (ret) {
2621 		printk(KERN_ERR "sending SRP_LOGIN_REQ response failed"
2622 		       " (error code = %d)\n", ret);
2623 		goto release_channel;
2624 	}
2625 
2626 	spin_lock_irq(&sdev->spinlock);
2627 	list_add_tail(&ch->list, &sdev->rch_list);
2628 	spin_unlock_irq(&sdev->spinlock);
2629 
2630 	goto out;
2631 
2632 release_channel:
2633 	srpt_set_ch_state(ch, CH_RELEASING);
2634 	transport_deregister_session_configfs(ch->sess);
2635 
2636 deregister_session:
2637 	transport_deregister_session(ch->sess);
2638 	ch->sess = NULL;
2639 
2640 destroy_ib:
2641 	srpt_destroy_ch_ib(ch);
2642 
2643 free_ring:
2644 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
2645 			     ch->sport->sdev, ch->rq_size,
2646 			     ch->rsp_size, DMA_TO_DEVICE);
2647 free_ch:
2648 	kfree(ch);
2649 
2650 reject:
2651 	rej->opcode = SRP_LOGIN_REJ;
2652 	rej->tag = req->tag;
2653 	rej->buf_fmt = __constant_cpu_to_be16(SRP_BUF_FORMAT_DIRECT
2654 					      | SRP_BUF_FORMAT_INDIRECT);
2655 
2656 	ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED, NULL, 0,
2657 			     (void *)rej, sizeof *rej);
2658 
2659 out:
2660 	kfree(rep_param);
2661 	kfree(rsp);
2662 	kfree(rej);
2663 
2664 	return ret;
2665 }
2666 
2667 static void srpt_cm_rej_recv(struct ib_cm_id *cm_id)
2668 {
2669 	printk(KERN_INFO "Received IB REJ for cm_id %p.\n", cm_id);
2670 	srpt_drain_channel(cm_id);
2671 }
2672 
2673 /**
2674  * srpt_cm_rtu_recv() - Process an IB_CM_RTU_RECEIVED or USER_ESTABLISHED event.
2675  *
2676  * An IB_CM_RTU_RECEIVED message indicates that the connection is established
2677  * and that the recipient may begin transmitting (RTU = ready to use).
2678  */
2679 static void srpt_cm_rtu_recv(struct ib_cm_id *cm_id)
2680 {
2681 	struct srpt_rdma_ch *ch;
2682 	int ret;
2683 
2684 	ch = srpt_find_channel(cm_id->context, cm_id);
2685 	BUG_ON(!ch);
2686 
2687 	if (srpt_test_and_set_ch_state(ch, CH_CONNECTING, CH_LIVE)) {
2688 		struct srpt_recv_ioctx *ioctx, *ioctx_tmp;
2689 
2690 		ret = srpt_ch_qp_rts(ch, ch->qp);
2691 
2692 		list_for_each_entry_safe(ioctx, ioctx_tmp, &ch->cmd_wait_list,
2693 					 wait_list) {
2694 			list_del(&ioctx->wait_list);
2695 			srpt_handle_new_iu(ch, ioctx, NULL);
2696 		}
2697 		if (ret)
2698 			srpt_close_ch(ch);
2699 	}
2700 }
2701 
2702 static void srpt_cm_timewait_exit(struct ib_cm_id *cm_id)
2703 {
2704 	printk(KERN_INFO "Received IB TimeWait exit for cm_id %p.\n", cm_id);
2705 	srpt_drain_channel(cm_id);
2706 }
2707 
2708 static void srpt_cm_rep_error(struct ib_cm_id *cm_id)
2709 {
2710 	printk(KERN_INFO "Received IB REP error for cm_id %p.\n", cm_id);
2711 	srpt_drain_channel(cm_id);
2712 }
2713 
2714 /**
2715  * srpt_cm_dreq_recv() - Process reception of a DREQ message.
2716  */
2717 static void srpt_cm_dreq_recv(struct ib_cm_id *cm_id)
2718 {
2719 	struct srpt_rdma_ch *ch;
2720 	unsigned long flags;
2721 	bool send_drep = false;
2722 
2723 	ch = srpt_find_channel(cm_id->context, cm_id);
2724 	BUG_ON(!ch);
2725 
2726 	pr_debug("cm_id= %p ch->state= %d\n", cm_id, srpt_get_ch_state(ch));
2727 
2728 	spin_lock_irqsave(&ch->spinlock, flags);
2729 	switch (ch->state) {
2730 	case CH_CONNECTING:
2731 	case CH_LIVE:
2732 		send_drep = true;
2733 		ch->state = CH_DISCONNECTING;
2734 		break;
2735 	case CH_DISCONNECTING:
2736 	case CH_DRAINING:
2737 	case CH_RELEASING:
2738 		WARN(true, "unexpected channel state %d\n", ch->state);
2739 		break;
2740 	}
2741 	spin_unlock_irqrestore(&ch->spinlock, flags);
2742 
2743 	if (send_drep) {
2744 		if (ib_send_cm_drep(ch->cm_id, NULL, 0) < 0)
2745 			printk(KERN_ERR "Sending IB DREP failed.\n");
2746 		printk(KERN_INFO "Received DREQ and sent DREP for session %s.\n",
2747 		       ch->sess_name);
2748 	}
2749 }
2750 
2751 /**
2752  * srpt_cm_drep_recv() - Process reception of a DREP message.
2753  */
2754 static void srpt_cm_drep_recv(struct ib_cm_id *cm_id)
2755 {
2756 	printk(KERN_INFO "Received InfiniBand DREP message for cm_id %p.\n",
2757 	       cm_id);
2758 	srpt_drain_channel(cm_id);
2759 }
2760 
2761 /**
2762  * srpt_cm_handler() - IB connection manager callback function.
2763  *
2764  * A non-zero return value will cause the caller destroy the CM ID.
2765  *
2766  * Note: srpt_cm_handler() must only return a non-zero value when transferring
2767  * ownership of the cm_id to a channel by srpt_cm_req_recv() failed. Returning
2768  * a non-zero value in any other case will trigger a race with the
2769  * ib_destroy_cm_id() call in srpt_release_channel().
2770  */
2771 static int srpt_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
2772 {
2773 	int ret;
2774 
2775 	ret = 0;
2776 	switch (event->event) {
2777 	case IB_CM_REQ_RECEIVED:
2778 		ret = srpt_cm_req_recv(cm_id, &event->param.req_rcvd,
2779 				       event->private_data);
2780 		break;
2781 	case IB_CM_REJ_RECEIVED:
2782 		srpt_cm_rej_recv(cm_id);
2783 		break;
2784 	case IB_CM_RTU_RECEIVED:
2785 	case IB_CM_USER_ESTABLISHED:
2786 		srpt_cm_rtu_recv(cm_id);
2787 		break;
2788 	case IB_CM_DREQ_RECEIVED:
2789 		srpt_cm_dreq_recv(cm_id);
2790 		break;
2791 	case IB_CM_DREP_RECEIVED:
2792 		srpt_cm_drep_recv(cm_id);
2793 		break;
2794 	case IB_CM_TIMEWAIT_EXIT:
2795 		srpt_cm_timewait_exit(cm_id);
2796 		break;
2797 	case IB_CM_REP_ERROR:
2798 		srpt_cm_rep_error(cm_id);
2799 		break;
2800 	case IB_CM_DREQ_ERROR:
2801 		printk(KERN_INFO "Received IB DREQ ERROR event.\n");
2802 		break;
2803 	case IB_CM_MRA_RECEIVED:
2804 		printk(KERN_INFO "Received IB MRA event\n");
2805 		break;
2806 	default:
2807 		printk(KERN_ERR "received unrecognized IB CM event %d\n",
2808 		       event->event);
2809 		break;
2810 	}
2811 
2812 	return ret;
2813 }
2814 
2815 /**
2816  * srpt_perform_rdmas() - Perform IB RDMA.
2817  *
2818  * Returns zero upon success or a negative number upon failure.
2819  */
2820 static int srpt_perform_rdmas(struct srpt_rdma_ch *ch,
2821 			      struct srpt_send_ioctx *ioctx)
2822 {
2823 	struct ib_send_wr wr;
2824 	struct ib_send_wr *bad_wr;
2825 	struct rdma_iu *riu;
2826 	int i;
2827 	int ret;
2828 	int sq_wr_avail;
2829 	enum dma_data_direction dir;
2830 	const int n_rdma = ioctx->n_rdma;
2831 
2832 	dir = ioctx->cmd.data_direction;
2833 	if (dir == DMA_TO_DEVICE) {
2834 		/* write */
2835 		ret = -ENOMEM;
2836 		sq_wr_avail = atomic_sub_return(n_rdma, &ch->sq_wr_avail);
2837 		if (sq_wr_avail < 0) {
2838 			printk(KERN_WARNING "IB send queue full (needed %d)\n",
2839 			       n_rdma);
2840 			goto out;
2841 		}
2842 	}
2843 
2844 	ioctx->rdma_aborted = false;
2845 	ret = 0;
2846 	riu = ioctx->rdma_ius;
2847 	memset(&wr, 0, sizeof wr);
2848 
2849 	for (i = 0; i < n_rdma; ++i, ++riu) {
2850 		if (dir == DMA_FROM_DEVICE) {
2851 			wr.opcode = IB_WR_RDMA_WRITE;
2852 			wr.wr_id = encode_wr_id(i == n_rdma - 1 ?
2853 						SRPT_RDMA_WRITE_LAST :
2854 						SRPT_RDMA_MID,
2855 						ioctx->ioctx.index);
2856 		} else {
2857 			wr.opcode = IB_WR_RDMA_READ;
2858 			wr.wr_id = encode_wr_id(i == n_rdma - 1 ?
2859 						SRPT_RDMA_READ_LAST :
2860 						SRPT_RDMA_MID,
2861 						ioctx->ioctx.index);
2862 		}
2863 		wr.next = NULL;
2864 		wr.wr.rdma.remote_addr = riu->raddr;
2865 		wr.wr.rdma.rkey = riu->rkey;
2866 		wr.num_sge = riu->sge_cnt;
2867 		wr.sg_list = riu->sge;
2868 
2869 		/* only get completion event for the last rdma write */
2870 		if (i == (n_rdma - 1) && dir == DMA_TO_DEVICE)
2871 			wr.send_flags = IB_SEND_SIGNALED;
2872 
2873 		ret = ib_post_send(ch->qp, &wr, &bad_wr);
2874 		if (ret)
2875 			break;
2876 	}
2877 
2878 	if (ret)
2879 		printk(KERN_ERR "%s[%d]: ib_post_send() returned %d for %d/%d",
2880 				 __func__, __LINE__, ret, i, n_rdma);
2881 	if (ret && i > 0) {
2882 		wr.num_sge = 0;
2883 		wr.wr_id = encode_wr_id(SRPT_RDMA_ABORT, ioctx->ioctx.index);
2884 		wr.send_flags = IB_SEND_SIGNALED;
2885 		while (ch->state == CH_LIVE &&
2886 			ib_post_send(ch->qp, &wr, &bad_wr) != 0) {
2887 			printk(KERN_INFO "Trying to abort failed RDMA transfer [%d]",
2888 				ioctx->ioctx.index);
2889 			msleep(1000);
2890 		}
2891 		while (ch->state != CH_RELEASING && !ioctx->rdma_aborted) {
2892 			printk(KERN_INFO "Waiting until RDMA abort finished [%d]",
2893 				ioctx->ioctx.index);
2894 			msleep(1000);
2895 		}
2896 	}
2897 out:
2898 	if (unlikely(dir == DMA_TO_DEVICE && ret < 0))
2899 		atomic_add(n_rdma, &ch->sq_wr_avail);
2900 	return ret;
2901 }
2902 
2903 /**
2904  * srpt_xfer_data() - Start data transfer from initiator to target.
2905  */
2906 static int srpt_xfer_data(struct srpt_rdma_ch *ch,
2907 			  struct srpt_send_ioctx *ioctx)
2908 {
2909 	int ret;
2910 
2911 	ret = srpt_map_sg_to_ib_sge(ch, ioctx);
2912 	if (ret) {
2913 		printk(KERN_ERR "%s[%d] ret=%d\n", __func__, __LINE__, ret);
2914 		goto out;
2915 	}
2916 
2917 	ret = srpt_perform_rdmas(ch, ioctx);
2918 	if (ret) {
2919 		if (ret == -EAGAIN || ret == -ENOMEM)
2920 			printk(KERN_INFO "%s[%d] queue full -- ret=%d\n",
2921 				   __func__, __LINE__, ret);
2922 		else
2923 			printk(KERN_ERR "%s[%d] fatal error -- ret=%d\n",
2924 			       __func__, __LINE__, ret);
2925 		goto out_unmap;
2926 	}
2927 
2928 out:
2929 	return ret;
2930 out_unmap:
2931 	srpt_unmap_sg_to_ib_sge(ch, ioctx);
2932 	goto out;
2933 }
2934 
2935 static int srpt_write_pending_status(struct se_cmd *se_cmd)
2936 {
2937 	struct srpt_send_ioctx *ioctx;
2938 
2939 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
2940 	return srpt_get_cmd_state(ioctx) == SRPT_STATE_NEED_DATA;
2941 }
2942 
2943 /*
2944  * srpt_write_pending() - Start data transfer from initiator to target (write).
2945  */
2946 static int srpt_write_pending(struct se_cmd *se_cmd)
2947 {
2948 	struct srpt_rdma_ch *ch;
2949 	struct srpt_send_ioctx *ioctx;
2950 	enum srpt_command_state new_state;
2951 	enum rdma_ch_state ch_state;
2952 	int ret;
2953 
2954 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
2955 
2956 	new_state = srpt_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA);
2957 	WARN_ON(new_state == SRPT_STATE_DONE);
2958 
2959 	ch = ioctx->ch;
2960 	BUG_ON(!ch);
2961 
2962 	ch_state = srpt_get_ch_state(ch);
2963 	switch (ch_state) {
2964 	case CH_CONNECTING:
2965 		WARN(true, "unexpected channel state %d\n", ch_state);
2966 		ret = -EINVAL;
2967 		goto out;
2968 	case CH_LIVE:
2969 		break;
2970 	case CH_DISCONNECTING:
2971 	case CH_DRAINING:
2972 	case CH_RELEASING:
2973 		pr_debug("cmd with tag %lld: channel disconnecting\n",
2974 			 ioctx->tag);
2975 		srpt_set_cmd_state(ioctx, SRPT_STATE_DATA_IN);
2976 		ret = -EINVAL;
2977 		goto out;
2978 	}
2979 	ret = srpt_xfer_data(ch, ioctx);
2980 
2981 out:
2982 	return ret;
2983 }
2984 
2985 static u8 tcm_to_srp_tsk_mgmt_status(const int tcm_mgmt_status)
2986 {
2987 	switch (tcm_mgmt_status) {
2988 	case TMR_FUNCTION_COMPLETE:
2989 		return SRP_TSK_MGMT_SUCCESS;
2990 	case TMR_FUNCTION_REJECTED:
2991 		return SRP_TSK_MGMT_FUNC_NOT_SUPP;
2992 	}
2993 	return SRP_TSK_MGMT_FAILED;
2994 }
2995 
2996 /**
2997  * srpt_queue_response() - Transmits the response to a SCSI command.
2998  *
2999  * Callback function called by the TCM core. Must not block since it can be
3000  * invoked on the context of the IB completion handler.
3001  */
3002 static void srpt_queue_response(struct se_cmd *cmd)
3003 {
3004 	struct srpt_rdma_ch *ch;
3005 	struct srpt_send_ioctx *ioctx;
3006 	enum srpt_command_state state;
3007 	unsigned long flags;
3008 	int ret;
3009 	enum dma_data_direction dir;
3010 	int resp_len;
3011 	u8 srp_tm_status;
3012 
3013 	ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
3014 	ch = ioctx->ch;
3015 	BUG_ON(!ch);
3016 
3017 	spin_lock_irqsave(&ioctx->spinlock, flags);
3018 	state = ioctx->state;
3019 	switch (state) {
3020 	case SRPT_STATE_NEW:
3021 	case SRPT_STATE_DATA_IN:
3022 		ioctx->state = SRPT_STATE_CMD_RSP_SENT;
3023 		break;
3024 	case SRPT_STATE_MGMT:
3025 		ioctx->state = SRPT_STATE_MGMT_RSP_SENT;
3026 		break;
3027 	default:
3028 		WARN(true, "ch %p; cmd %d: unexpected command state %d\n",
3029 			ch, ioctx->ioctx.index, ioctx->state);
3030 		break;
3031 	}
3032 	spin_unlock_irqrestore(&ioctx->spinlock, flags);
3033 
3034 	if (unlikely(transport_check_aborted_status(&ioctx->cmd, false)
3035 		     || WARN_ON_ONCE(state == SRPT_STATE_CMD_RSP_SENT))) {
3036 		atomic_inc(&ch->req_lim_delta);
3037 		srpt_abort_cmd(ioctx);
3038 		return;
3039 	}
3040 
3041 	dir = ioctx->cmd.data_direction;
3042 
3043 	/* For read commands, transfer the data to the initiator. */
3044 	if (dir == DMA_FROM_DEVICE && ioctx->cmd.data_length &&
3045 	    !ioctx->queue_status_only) {
3046 		ret = srpt_xfer_data(ch, ioctx);
3047 		if (ret) {
3048 			printk(KERN_ERR "xfer_data failed for tag %llu\n",
3049 			       ioctx->tag);
3050 			return;
3051 		}
3052 	}
3053 
3054 	if (state != SRPT_STATE_MGMT)
3055 		resp_len = srpt_build_cmd_rsp(ch, ioctx, ioctx->tag,
3056 					      cmd->scsi_status);
3057 	else {
3058 		srp_tm_status
3059 			= tcm_to_srp_tsk_mgmt_status(cmd->se_tmr_req->response);
3060 		resp_len = srpt_build_tskmgmt_rsp(ch, ioctx, srp_tm_status,
3061 						 ioctx->tag);
3062 	}
3063 	ret = srpt_post_send(ch, ioctx, resp_len);
3064 	if (ret) {
3065 		printk(KERN_ERR "sending cmd response failed for tag %llu\n",
3066 		       ioctx->tag);
3067 		srpt_unmap_sg_to_ib_sge(ch, ioctx);
3068 		srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
3069 		target_put_sess_cmd(ioctx->ch->sess, &ioctx->cmd);
3070 	}
3071 }
3072 
3073 static int srpt_queue_data_in(struct se_cmd *cmd)
3074 {
3075 	srpt_queue_response(cmd);
3076 	return 0;
3077 }
3078 
3079 static void srpt_queue_tm_rsp(struct se_cmd *cmd)
3080 {
3081 	srpt_queue_response(cmd);
3082 }
3083 
3084 static int srpt_queue_status(struct se_cmd *cmd)
3085 {
3086 	struct srpt_send_ioctx *ioctx;
3087 
3088 	ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
3089 	BUG_ON(ioctx->sense_data != cmd->sense_buffer);
3090 	if (cmd->se_cmd_flags &
3091 	    (SCF_TRANSPORT_TASK_SENSE | SCF_EMULATED_TASK_SENSE))
3092 		WARN_ON(cmd->scsi_status != SAM_STAT_CHECK_CONDITION);
3093 	ioctx->queue_status_only = true;
3094 	srpt_queue_response(cmd);
3095 	return 0;
3096 }
3097 
3098 static void srpt_refresh_port_work(struct work_struct *work)
3099 {
3100 	struct srpt_port *sport = container_of(work, struct srpt_port, work);
3101 
3102 	srpt_refresh_port(sport);
3103 }
3104 
3105 static int srpt_ch_list_empty(struct srpt_device *sdev)
3106 {
3107 	int res;
3108 
3109 	spin_lock_irq(&sdev->spinlock);
3110 	res = list_empty(&sdev->rch_list);
3111 	spin_unlock_irq(&sdev->spinlock);
3112 
3113 	return res;
3114 }
3115 
3116 /**
3117  * srpt_release_sdev() - Free the channel resources associated with a target.
3118  */
3119 static int srpt_release_sdev(struct srpt_device *sdev)
3120 {
3121 	struct srpt_rdma_ch *ch, *tmp_ch;
3122 	int res;
3123 
3124 	WARN_ON_ONCE(irqs_disabled());
3125 
3126 	BUG_ON(!sdev);
3127 
3128 	spin_lock_irq(&sdev->spinlock);
3129 	list_for_each_entry_safe(ch, tmp_ch, &sdev->rch_list, list)
3130 		__srpt_close_ch(ch);
3131 	spin_unlock_irq(&sdev->spinlock);
3132 
3133 	res = wait_event_interruptible(sdev->ch_releaseQ,
3134 				       srpt_ch_list_empty(sdev));
3135 	if (res)
3136 		printk(KERN_ERR "%s: interrupted.\n", __func__);
3137 
3138 	return 0;
3139 }
3140 
3141 static struct srpt_port *__srpt_lookup_port(const char *name)
3142 {
3143 	struct ib_device *dev;
3144 	struct srpt_device *sdev;
3145 	struct srpt_port *sport;
3146 	int i;
3147 
3148 	list_for_each_entry(sdev, &srpt_dev_list, list) {
3149 		dev = sdev->device;
3150 		if (!dev)
3151 			continue;
3152 
3153 		for (i = 0; i < dev->phys_port_cnt; i++) {
3154 			sport = &sdev->port[i];
3155 
3156 			if (!strcmp(sport->port_guid, name))
3157 				return sport;
3158 		}
3159 	}
3160 
3161 	return NULL;
3162 }
3163 
3164 static struct srpt_port *srpt_lookup_port(const char *name)
3165 {
3166 	struct srpt_port *sport;
3167 
3168 	spin_lock(&srpt_dev_lock);
3169 	sport = __srpt_lookup_port(name);
3170 	spin_unlock(&srpt_dev_lock);
3171 
3172 	return sport;
3173 }
3174 
3175 /**
3176  * srpt_add_one() - Infiniband device addition callback function.
3177  */
3178 static void srpt_add_one(struct ib_device *device)
3179 {
3180 	struct srpt_device *sdev;
3181 	struct srpt_port *sport;
3182 	struct ib_srq_init_attr srq_attr;
3183 	int i;
3184 
3185 	pr_debug("device = %p, device->dma_ops = %p\n", device,
3186 		 device->dma_ops);
3187 
3188 	sdev = kzalloc(sizeof *sdev, GFP_KERNEL);
3189 	if (!sdev)
3190 		goto err;
3191 
3192 	sdev->device = device;
3193 	INIT_LIST_HEAD(&sdev->rch_list);
3194 	init_waitqueue_head(&sdev->ch_releaseQ);
3195 	spin_lock_init(&sdev->spinlock);
3196 
3197 	if (ib_query_device(device, &sdev->dev_attr))
3198 		goto free_dev;
3199 
3200 	sdev->pd = ib_alloc_pd(device);
3201 	if (IS_ERR(sdev->pd))
3202 		goto free_dev;
3203 
3204 	sdev->mr = ib_get_dma_mr(sdev->pd, IB_ACCESS_LOCAL_WRITE);
3205 	if (IS_ERR(sdev->mr))
3206 		goto err_pd;
3207 
3208 	sdev->srq_size = min(srpt_srq_size, sdev->dev_attr.max_srq_wr);
3209 
3210 	srq_attr.event_handler = srpt_srq_event;
3211 	srq_attr.srq_context = (void *)sdev;
3212 	srq_attr.attr.max_wr = sdev->srq_size;
3213 	srq_attr.attr.max_sge = 1;
3214 	srq_attr.attr.srq_limit = 0;
3215 	srq_attr.srq_type = IB_SRQT_BASIC;
3216 
3217 	sdev->srq = ib_create_srq(sdev->pd, &srq_attr);
3218 	if (IS_ERR(sdev->srq))
3219 		goto err_mr;
3220 
3221 	pr_debug("%s: create SRQ #wr= %d max_allow=%d dev= %s\n",
3222 		 __func__, sdev->srq_size, sdev->dev_attr.max_srq_wr,
3223 		 device->name);
3224 
3225 	if (!srpt_service_guid)
3226 		srpt_service_guid = be64_to_cpu(device->node_guid);
3227 
3228 	sdev->cm_id = ib_create_cm_id(device, srpt_cm_handler, sdev);
3229 	if (IS_ERR(sdev->cm_id))
3230 		goto err_srq;
3231 
3232 	/* print out target login information */
3233 	pr_debug("Target login info: id_ext=%016llx,ioc_guid=%016llx,"
3234 		 "pkey=ffff,service_id=%016llx\n", srpt_service_guid,
3235 		 srpt_service_guid, srpt_service_guid);
3236 
3237 	/*
3238 	 * We do not have a consistent service_id (ie. also id_ext of target_id)
3239 	 * to identify this target. We currently use the guid of the first HCA
3240 	 * in the system as service_id; therefore, the target_id will change
3241 	 * if this HCA is gone bad and replaced by different HCA
3242 	 */
3243 	if (ib_cm_listen(sdev->cm_id, cpu_to_be64(srpt_service_guid), 0, NULL))
3244 		goto err_cm;
3245 
3246 	INIT_IB_EVENT_HANDLER(&sdev->event_handler, sdev->device,
3247 			      srpt_event_handler);
3248 	if (ib_register_event_handler(&sdev->event_handler))
3249 		goto err_cm;
3250 
3251 	sdev->ioctx_ring = (struct srpt_recv_ioctx **)
3252 		srpt_alloc_ioctx_ring(sdev, sdev->srq_size,
3253 				      sizeof(*sdev->ioctx_ring[0]),
3254 				      srp_max_req_size, DMA_FROM_DEVICE);
3255 	if (!sdev->ioctx_ring)
3256 		goto err_event;
3257 
3258 	for (i = 0; i < sdev->srq_size; ++i)
3259 		srpt_post_recv(sdev, sdev->ioctx_ring[i]);
3260 
3261 	WARN_ON(sdev->device->phys_port_cnt > ARRAY_SIZE(sdev->port));
3262 
3263 	for (i = 1; i <= sdev->device->phys_port_cnt; i++) {
3264 		sport = &sdev->port[i - 1];
3265 		sport->sdev = sdev;
3266 		sport->port = i;
3267 		sport->port_attrib.srp_max_rdma_size = DEFAULT_MAX_RDMA_SIZE;
3268 		sport->port_attrib.srp_max_rsp_size = DEFAULT_MAX_RSP_SIZE;
3269 		sport->port_attrib.srp_sq_size = DEF_SRPT_SQ_SIZE;
3270 		INIT_WORK(&sport->work, srpt_refresh_port_work);
3271 		INIT_LIST_HEAD(&sport->port_acl_list);
3272 		spin_lock_init(&sport->port_acl_lock);
3273 
3274 		if (srpt_refresh_port(sport)) {
3275 			printk(KERN_ERR "MAD registration failed for %s-%d.\n",
3276 			       srpt_sdev_name(sdev), i);
3277 			goto err_ring;
3278 		}
3279 		snprintf(sport->port_guid, sizeof(sport->port_guid),
3280 			"0x%016llx%016llx",
3281 			be64_to_cpu(sport->gid.global.subnet_prefix),
3282 			be64_to_cpu(sport->gid.global.interface_id));
3283 	}
3284 
3285 	spin_lock(&srpt_dev_lock);
3286 	list_add_tail(&sdev->list, &srpt_dev_list);
3287 	spin_unlock(&srpt_dev_lock);
3288 
3289 out:
3290 	ib_set_client_data(device, &srpt_client, sdev);
3291 	pr_debug("added %s.\n", device->name);
3292 	return;
3293 
3294 err_ring:
3295 	srpt_free_ioctx_ring((struct srpt_ioctx **)sdev->ioctx_ring, sdev,
3296 			     sdev->srq_size, srp_max_req_size,
3297 			     DMA_FROM_DEVICE);
3298 err_event:
3299 	ib_unregister_event_handler(&sdev->event_handler);
3300 err_cm:
3301 	ib_destroy_cm_id(sdev->cm_id);
3302 err_srq:
3303 	ib_destroy_srq(sdev->srq);
3304 err_mr:
3305 	ib_dereg_mr(sdev->mr);
3306 err_pd:
3307 	ib_dealloc_pd(sdev->pd);
3308 free_dev:
3309 	kfree(sdev);
3310 err:
3311 	sdev = NULL;
3312 	printk(KERN_INFO "%s(%s) failed.\n", __func__, device->name);
3313 	goto out;
3314 }
3315 
3316 /**
3317  * srpt_remove_one() - InfiniBand device removal callback function.
3318  */
3319 static void srpt_remove_one(struct ib_device *device)
3320 {
3321 	struct srpt_device *sdev;
3322 	int i;
3323 
3324 	sdev = ib_get_client_data(device, &srpt_client);
3325 	if (!sdev) {
3326 		printk(KERN_INFO "%s(%s): nothing to do.\n", __func__,
3327 		       device->name);
3328 		return;
3329 	}
3330 
3331 	srpt_unregister_mad_agent(sdev);
3332 
3333 	ib_unregister_event_handler(&sdev->event_handler);
3334 
3335 	/* Cancel any work queued by the just unregistered IB event handler. */
3336 	for (i = 0; i < sdev->device->phys_port_cnt; i++)
3337 		cancel_work_sync(&sdev->port[i].work);
3338 
3339 	ib_destroy_cm_id(sdev->cm_id);
3340 
3341 	/*
3342 	 * Unregistering a target must happen after destroying sdev->cm_id
3343 	 * such that no new SRP_LOGIN_REQ information units can arrive while
3344 	 * destroying the target.
3345 	 */
3346 	spin_lock(&srpt_dev_lock);
3347 	list_del(&sdev->list);
3348 	spin_unlock(&srpt_dev_lock);
3349 	srpt_release_sdev(sdev);
3350 
3351 	ib_destroy_srq(sdev->srq);
3352 	ib_dereg_mr(sdev->mr);
3353 	ib_dealloc_pd(sdev->pd);
3354 
3355 	srpt_free_ioctx_ring((struct srpt_ioctx **)sdev->ioctx_ring, sdev,
3356 			     sdev->srq_size, srp_max_req_size, DMA_FROM_DEVICE);
3357 	sdev->ioctx_ring = NULL;
3358 	kfree(sdev);
3359 }
3360 
3361 static struct ib_client srpt_client = {
3362 	.name = DRV_NAME,
3363 	.add = srpt_add_one,
3364 	.remove = srpt_remove_one
3365 };
3366 
3367 static int srpt_check_true(struct se_portal_group *se_tpg)
3368 {
3369 	return 1;
3370 }
3371 
3372 static int srpt_check_false(struct se_portal_group *se_tpg)
3373 {
3374 	return 0;
3375 }
3376 
3377 static char *srpt_get_fabric_name(void)
3378 {
3379 	return "srpt";
3380 }
3381 
3382 static u8 srpt_get_fabric_proto_ident(struct se_portal_group *se_tpg)
3383 {
3384 	return SCSI_TRANSPORTID_PROTOCOLID_SRP;
3385 }
3386 
3387 static char *srpt_get_fabric_wwn(struct se_portal_group *tpg)
3388 {
3389 	struct srpt_port *sport = container_of(tpg, struct srpt_port, port_tpg_1);
3390 
3391 	return sport->port_guid;
3392 }
3393 
3394 static u16 srpt_get_tag(struct se_portal_group *tpg)
3395 {
3396 	return 1;
3397 }
3398 
3399 static u32 srpt_get_default_depth(struct se_portal_group *se_tpg)
3400 {
3401 	return 1;
3402 }
3403 
3404 static u32 srpt_get_pr_transport_id(struct se_portal_group *se_tpg,
3405 				    struct se_node_acl *se_nacl,
3406 				    struct t10_pr_registration *pr_reg,
3407 				    int *format_code, unsigned char *buf)
3408 {
3409 	struct srpt_node_acl *nacl;
3410 	struct spc_rdma_transport_id *tr_id;
3411 
3412 	nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3413 	tr_id = (void *)buf;
3414 	tr_id->protocol_identifier = SCSI_TRANSPORTID_PROTOCOLID_SRP;
3415 	memcpy(tr_id->i_port_id, nacl->i_port_id, sizeof(tr_id->i_port_id));
3416 	return sizeof(*tr_id);
3417 }
3418 
3419 static u32 srpt_get_pr_transport_id_len(struct se_portal_group *se_tpg,
3420 					struct se_node_acl *se_nacl,
3421 					struct t10_pr_registration *pr_reg,
3422 					int *format_code)
3423 {
3424 	*format_code = 0;
3425 	return sizeof(struct spc_rdma_transport_id);
3426 }
3427 
3428 static char *srpt_parse_pr_out_transport_id(struct se_portal_group *se_tpg,
3429 					    const char *buf, u32 *out_tid_len,
3430 					    char **port_nexus_ptr)
3431 {
3432 	struct spc_rdma_transport_id *tr_id;
3433 
3434 	*port_nexus_ptr = NULL;
3435 	*out_tid_len = sizeof(struct spc_rdma_transport_id);
3436 	tr_id = (void *)buf;
3437 	return (char *)tr_id->i_port_id;
3438 }
3439 
3440 static struct se_node_acl *srpt_alloc_fabric_acl(struct se_portal_group *se_tpg)
3441 {
3442 	struct srpt_node_acl *nacl;
3443 
3444 	nacl = kzalloc(sizeof(struct srpt_node_acl), GFP_KERNEL);
3445 	if (!nacl) {
3446 		printk(KERN_ERR "Unable to allocate struct srpt_node_acl\n");
3447 		return NULL;
3448 	}
3449 
3450 	return &nacl->nacl;
3451 }
3452 
3453 static void srpt_release_fabric_acl(struct se_portal_group *se_tpg,
3454 				    struct se_node_acl *se_nacl)
3455 {
3456 	struct srpt_node_acl *nacl;
3457 
3458 	nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3459 	kfree(nacl);
3460 }
3461 
3462 static u32 srpt_tpg_get_inst_index(struct se_portal_group *se_tpg)
3463 {
3464 	return 1;
3465 }
3466 
3467 static void srpt_release_cmd(struct se_cmd *se_cmd)
3468 {
3469 	struct srpt_send_ioctx *ioctx = container_of(se_cmd,
3470 				struct srpt_send_ioctx, cmd);
3471 	struct srpt_rdma_ch *ch = ioctx->ch;
3472 	unsigned long flags;
3473 
3474 	WARN_ON(ioctx->state != SRPT_STATE_DONE);
3475 	WARN_ON(ioctx->mapped_sg_count != 0);
3476 
3477 	if (ioctx->n_rbuf > 1) {
3478 		kfree(ioctx->rbufs);
3479 		ioctx->rbufs = NULL;
3480 		ioctx->n_rbuf = 0;
3481 	}
3482 
3483 	spin_lock_irqsave(&ch->spinlock, flags);
3484 	list_add(&ioctx->free_list, &ch->free_list);
3485 	spin_unlock_irqrestore(&ch->spinlock, flags);
3486 }
3487 
3488 /**
3489  * srpt_close_session() - Forcibly close a session.
3490  *
3491  * Callback function invoked by the TCM core to clean up sessions associated
3492  * with a node ACL when the user invokes
3493  * rmdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3494  */
3495 static void srpt_close_session(struct se_session *se_sess)
3496 {
3497 	DECLARE_COMPLETION_ONSTACK(release_done);
3498 	struct srpt_rdma_ch *ch;
3499 	struct srpt_device *sdev;
3500 	int res;
3501 
3502 	ch = se_sess->fabric_sess_ptr;
3503 	WARN_ON(ch->sess != se_sess);
3504 
3505 	pr_debug("ch %p state %d\n", ch, srpt_get_ch_state(ch));
3506 
3507 	sdev = ch->sport->sdev;
3508 	spin_lock_irq(&sdev->spinlock);
3509 	BUG_ON(ch->release_done);
3510 	ch->release_done = &release_done;
3511 	__srpt_close_ch(ch);
3512 	spin_unlock_irq(&sdev->spinlock);
3513 
3514 	res = wait_for_completion_timeout(&release_done, 60 * HZ);
3515 	WARN_ON(res <= 0);
3516 }
3517 
3518 /**
3519  * srpt_sess_get_index() - Return the value of scsiAttIntrPortIndex (SCSI-MIB).
3520  *
3521  * A quote from RFC 4455 (SCSI-MIB) about this MIB object:
3522  * This object represents an arbitrary integer used to uniquely identify a
3523  * particular attached remote initiator port to a particular SCSI target port
3524  * within a particular SCSI target device within a particular SCSI instance.
3525  */
3526 static u32 srpt_sess_get_index(struct se_session *se_sess)
3527 {
3528 	return 0;
3529 }
3530 
3531 static void srpt_set_default_node_attrs(struct se_node_acl *nacl)
3532 {
3533 }
3534 
3535 static u32 srpt_get_task_tag(struct se_cmd *se_cmd)
3536 {
3537 	struct srpt_send_ioctx *ioctx;
3538 
3539 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
3540 	return ioctx->tag;
3541 }
3542 
3543 /* Note: only used from inside debug printk's by the TCM core. */
3544 static int srpt_get_tcm_cmd_state(struct se_cmd *se_cmd)
3545 {
3546 	struct srpt_send_ioctx *ioctx;
3547 
3548 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
3549 	return srpt_get_cmd_state(ioctx);
3550 }
3551 
3552 /**
3553  * srpt_parse_i_port_id() - Parse an initiator port ID.
3554  * @name: ASCII representation of a 128-bit initiator port ID.
3555  * @i_port_id: Binary 128-bit port ID.
3556  */
3557 static int srpt_parse_i_port_id(u8 i_port_id[16], const char *name)
3558 {
3559 	const char *p;
3560 	unsigned len, count, leading_zero_bytes;
3561 	int ret, rc;
3562 
3563 	p = name;
3564 	if (strnicmp(p, "0x", 2) == 0)
3565 		p += 2;
3566 	ret = -EINVAL;
3567 	len = strlen(p);
3568 	if (len % 2)
3569 		goto out;
3570 	count = min(len / 2, 16U);
3571 	leading_zero_bytes = 16 - count;
3572 	memset(i_port_id, 0, leading_zero_bytes);
3573 	rc = hex2bin(i_port_id + leading_zero_bytes, p, count);
3574 	if (rc < 0)
3575 		pr_debug("hex2bin failed for srpt_parse_i_port_id: %d\n", rc);
3576 	ret = 0;
3577 out:
3578 	return ret;
3579 }
3580 
3581 /*
3582  * configfs callback function invoked for
3583  * mkdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3584  */
3585 static struct se_node_acl *srpt_make_nodeacl(struct se_portal_group *tpg,
3586 					     struct config_group *group,
3587 					     const char *name)
3588 {
3589 	struct srpt_port *sport = container_of(tpg, struct srpt_port, port_tpg_1);
3590 	struct se_node_acl *se_nacl, *se_nacl_new;
3591 	struct srpt_node_acl *nacl;
3592 	int ret = 0;
3593 	u32 nexus_depth = 1;
3594 	u8 i_port_id[16];
3595 
3596 	if (srpt_parse_i_port_id(i_port_id, name) < 0) {
3597 		printk(KERN_ERR "invalid initiator port ID %s\n", name);
3598 		ret = -EINVAL;
3599 		goto err;
3600 	}
3601 
3602 	se_nacl_new = srpt_alloc_fabric_acl(tpg);
3603 	if (!se_nacl_new) {
3604 		ret = -ENOMEM;
3605 		goto err;
3606 	}
3607 	/*
3608 	 * nacl_new may be released by core_tpg_add_initiator_node_acl()
3609 	 * when converting a node ACL from demo mode to explict
3610 	 */
3611 	se_nacl = core_tpg_add_initiator_node_acl(tpg, se_nacl_new, name,
3612 						  nexus_depth);
3613 	if (IS_ERR(se_nacl)) {
3614 		ret = PTR_ERR(se_nacl);
3615 		goto err;
3616 	}
3617 	/* Locate our struct srpt_node_acl and set sdev and i_port_id. */
3618 	nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3619 	memcpy(&nacl->i_port_id[0], &i_port_id[0], 16);
3620 	nacl->sport = sport;
3621 
3622 	spin_lock_irq(&sport->port_acl_lock);
3623 	list_add_tail(&nacl->list, &sport->port_acl_list);
3624 	spin_unlock_irq(&sport->port_acl_lock);
3625 
3626 	return se_nacl;
3627 err:
3628 	return ERR_PTR(ret);
3629 }
3630 
3631 /*
3632  * configfs callback function invoked for
3633  * rmdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3634  */
3635 static void srpt_drop_nodeacl(struct se_node_acl *se_nacl)
3636 {
3637 	struct srpt_node_acl *nacl;
3638 	struct srpt_device *sdev;
3639 	struct srpt_port *sport;
3640 
3641 	nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3642 	sport = nacl->sport;
3643 	sdev = sport->sdev;
3644 	spin_lock_irq(&sport->port_acl_lock);
3645 	list_del(&nacl->list);
3646 	spin_unlock_irq(&sport->port_acl_lock);
3647 	core_tpg_del_initiator_node_acl(&sport->port_tpg_1, se_nacl, 1);
3648 	srpt_release_fabric_acl(NULL, se_nacl);
3649 }
3650 
3651 static ssize_t srpt_tpg_attrib_show_srp_max_rdma_size(
3652 	struct se_portal_group *se_tpg,
3653 	char *page)
3654 {
3655 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3656 
3657 	return sprintf(page, "%u\n", sport->port_attrib.srp_max_rdma_size);
3658 }
3659 
3660 static ssize_t srpt_tpg_attrib_store_srp_max_rdma_size(
3661 	struct se_portal_group *se_tpg,
3662 	const char *page,
3663 	size_t count)
3664 {
3665 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3666 	unsigned long val;
3667 	int ret;
3668 
3669 	ret = strict_strtoul(page, 0, &val);
3670 	if (ret < 0) {
3671 		pr_err("strict_strtoul() failed with ret: %d\n", ret);
3672 		return -EINVAL;
3673 	}
3674 	if (val > MAX_SRPT_RDMA_SIZE) {
3675 		pr_err("val: %lu exceeds MAX_SRPT_RDMA_SIZE: %d\n", val,
3676 			MAX_SRPT_RDMA_SIZE);
3677 		return -EINVAL;
3678 	}
3679 	if (val < DEFAULT_MAX_RDMA_SIZE) {
3680 		pr_err("val: %lu smaller than DEFAULT_MAX_RDMA_SIZE: %d\n",
3681 			val, DEFAULT_MAX_RDMA_SIZE);
3682 		return -EINVAL;
3683 	}
3684 	sport->port_attrib.srp_max_rdma_size = val;
3685 
3686 	return count;
3687 }
3688 
3689 TF_TPG_ATTRIB_ATTR(srpt, srp_max_rdma_size, S_IRUGO | S_IWUSR);
3690 
3691 static ssize_t srpt_tpg_attrib_show_srp_max_rsp_size(
3692 	struct se_portal_group *se_tpg,
3693 	char *page)
3694 {
3695 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3696 
3697 	return sprintf(page, "%u\n", sport->port_attrib.srp_max_rsp_size);
3698 }
3699 
3700 static ssize_t srpt_tpg_attrib_store_srp_max_rsp_size(
3701 	struct se_portal_group *se_tpg,
3702 	const char *page,
3703 	size_t count)
3704 {
3705 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3706 	unsigned long val;
3707 	int ret;
3708 
3709 	ret = strict_strtoul(page, 0, &val);
3710 	if (ret < 0) {
3711 		pr_err("strict_strtoul() failed with ret: %d\n", ret);
3712 		return -EINVAL;
3713 	}
3714 	if (val > MAX_SRPT_RSP_SIZE) {
3715 		pr_err("val: %lu exceeds MAX_SRPT_RSP_SIZE: %d\n", val,
3716 			MAX_SRPT_RSP_SIZE);
3717 		return -EINVAL;
3718 	}
3719 	if (val < MIN_MAX_RSP_SIZE) {
3720 		pr_err("val: %lu smaller than MIN_MAX_RSP_SIZE: %d\n", val,
3721 			MIN_MAX_RSP_SIZE);
3722 		return -EINVAL;
3723 	}
3724 	sport->port_attrib.srp_max_rsp_size = val;
3725 
3726 	return count;
3727 }
3728 
3729 TF_TPG_ATTRIB_ATTR(srpt, srp_max_rsp_size, S_IRUGO | S_IWUSR);
3730 
3731 static ssize_t srpt_tpg_attrib_show_srp_sq_size(
3732 	struct se_portal_group *se_tpg,
3733 	char *page)
3734 {
3735 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3736 
3737 	return sprintf(page, "%u\n", sport->port_attrib.srp_sq_size);
3738 }
3739 
3740 static ssize_t srpt_tpg_attrib_store_srp_sq_size(
3741 	struct se_portal_group *se_tpg,
3742 	const char *page,
3743 	size_t count)
3744 {
3745 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3746 	unsigned long val;
3747 	int ret;
3748 
3749 	ret = strict_strtoul(page, 0, &val);
3750 	if (ret < 0) {
3751 		pr_err("strict_strtoul() failed with ret: %d\n", ret);
3752 		return -EINVAL;
3753 	}
3754 	if (val > MAX_SRPT_SRQ_SIZE) {
3755 		pr_err("val: %lu exceeds MAX_SRPT_SRQ_SIZE: %d\n", val,
3756 			MAX_SRPT_SRQ_SIZE);
3757 		return -EINVAL;
3758 	}
3759 	if (val < MIN_SRPT_SRQ_SIZE) {
3760 		pr_err("val: %lu smaller than MIN_SRPT_SRQ_SIZE: %d\n", val,
3761 			MIN_SRPT_SRQ_SIZE);
3762 		return -EINVAL;
3763 	}
3764 	sport->port_attrib.srp_sq_size = val;
3765 
3766 	return count;
3767 }
3768 
3769 TF_TPG_ATTRIB_ATTR(srpt, srp_sq_size, S_IRUGO | S_IWUSR);
3770 
3771 static struct configfs_attribute *srpt_tpg_attrib_attrs[] = {
3772 	&srpt_tpg_attrib_srp_max_rdma_size.attr,
3773 	&srpt_tpg_attrib_srp_max_rsp_size.attr,
3774 	&srpt_tpg_attrib_srp_sq_size.attr,
3775 	NULL,
3776 };
3777 
3778 static ssize_t srpt_tpg_show_enable(
3779 	struct se_portal_group *se_tpg,
3780 	char *page)
3781 {
3782 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3783 
3784 	return snprintf(page, PAGE_SIZE, "%d\n", (sport->enabled) ? 1: 0);
3785 }
3786 
3787 static ssize_t srpt_tpg_store_enable(
3788 	struct se_portal_group *se_tpg,
3789 	const char *page,
3790 	size_t count)
3791 {
3792 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3793 	unsigned long tmp;
3794         int ret;
3795 
3796 	ret = strict_strtoul(page, 0, &tmp);
3797 	if (ret < 0) {
3798 		printk(KERN_ERR "Unable to extract srpt_tpg_store_enable\n");
3799 		return -EINVAL;
3800 	}
3801 
3802 	if ((tmp != 0) && (tmp != 1)) {
3803 		printk(KERN_ERR "Illegal value for srpt_tpg_store_enable: %lu\n", tmp);
3804 		return -EINVAL;
3805 	}
3806 	if (tmp == 1)
3807 		sport->enabled = true;
3808 	else
3809 		sport->enabled = false;
3810 
3811 	return count;
3812 }
3813 
3814 TF_TPG_BASE_ATTR(srpt, enable, S_IRUGO | S_IWUSR);
3815 
3816 static struct configfs_attribute *srpt_tpg_attrs[] = {
3817 	&srpt_tpg_enable.attr,
3818 	NULL,
3819 };
3820 
3821 /**
3822  * configfs callback invoked for
3823  * mkdir /sys/kernel/config/target/$driver/$port/$tpg
3824  */
3825 static struct se_portal_group *srpt_make_tpg(struct se_wwn *wwn,
3826 					     struct config_group *group,
3827 					     const char *name)
3828 {
3829 	struct srpt_port *sport = container_of(wwn, struct srpt_port, port_wwn);
3830 	int res;
3831 
3832 	/* Initialize sport->port_wwn and sport->port_tpg_1 */
3833 	res = core_tpg_register(&srpt_target->tf_ops, &sport->port_wwn,
3834 			&sport->port_tpg_1, sport, TRANSPORT_TPG_TYPE_NORMAL);
3835 	if (res)
3836 		return ERR_PTR(res);
3837 
3838 	return &sport->port_tpg_1;
3839 }
3840 
3841 /**
3842  * configfs callback invoked for
3843  * rmdir /sys/kernel/config/target/$driver/$port/$tpg
3844  */
3845 static void srpt_drop_tpg(struct se_portal_group *tpg)
3846 {
3847 	struct srpt_port *sport = container_of(tpg,
3848 				struct srpt_port, port_tpg_1);
3849 
3850 	sport->enabled = false;
3851 	core_tpg_deregister(&sport->port_tpg_1);
3852 }
3853 
3854 /**
3855  * configfs callback invoked for
3856  * mkdir /sys/kernel/config/target/$driver/$port
3857  */
3858 static struct se_wwn *srpt_make_tport(struct target_fabric_configfs *tf,
3859 				      struct config_group *group,
3860 				      const char *name)
3861 {
3862 	struct srpt_port *sport;
3863 	int ret;
3864 
3865 	sport = srpt_lookup_port(name);
3866 	pr_debug("make_tport(%s)\n", name);
3867 	ret = -EINVAL;
3868 	if (!sport)
3869 		goto err;
3870 
3871 	return &sport->port_wwn;
3872 
3873 err:
3874 	return ERR_PTR(ret);
3875 }
3876 
3877 /**
3878  * configfs callback invoked for
3879  * rmdir /sys/kernel/config/target/$driver/$port
3880  */
3881 static void srpt_drop_tport(struct se_wwn *wwn)
3882 {
3883 	struct srpt_port *sport = container_of(wwn, struct srpt_port, port_wwn);
3884 
3885 	pr_debug("drop_tport(%s\n", config_item_name(&sport->port_wwn.wwn_group.cg_item));
3886 }
3887 
3888 static ssize_t srpt_wwn_show_attr_version(struct target_fabric_configfs *tf,
3889 					      char *buf)
3890 {
3891 	return scnprintf(buf, PAGE_SIZE, "%s\n", DRV_VERSION);
3892 }
3893 
3894 TF_WWN_ATTR_RO(srpt, version);
3895 
3896 static struct configfs_attribute *srpt_wwn_attrs[] = {
3897 	&srpt_wwn_version.attr,
3898 	NULL,
3899 };
3900 
3901 static struct target_core_fabric_ops srpt_template = {
3902 	.get_fabric_name		= srpt_get_fabric_name,
3903 	.get_fabric_proto_ident		= srpt_get_fabric_proto_ident,
3904 	.tpg_get_wwn			= srpt_get_fabric_wwn,
3905 	.tpg_get_tag			= srpt_get_tag,
3906 	.tpg_get_default_depth		= srpt_get_default_depth,
3907 	.tpg_get_pr_transport_id	= srpt_get_pr_transport_id,
3908 	.tpg_get_pr_transport_id_len	= srpt_get_pr_transport_id_len,
3909 	.tpg_parse_pr_out_transport_id	= srpt_parse_pr_out_transport_id,
3910 	.tpg_check_demo_mode		= srpt_check_false,
3911 	.tpg_check_demo_mode_cache	= srpt_check_true,
3912 	.tpg_check_demo_mode_write_protect = srpt_check_true,
3913 	.tpg_check_prod_mode_write_protect = srpt_check_false,
3914 	.tpg_alloc_fabric_acl		= srpt_alloc_fabric_acl,
3915 	.tpg_release_fabric_acl		= srpt_release_fabric_acl,
3916 	.tpg_get_inst_index		= srpt_tpg_get_inst_index,
3917 	.release_cmd			= srpt_release_cmd,
3918 	.check_stop_free		= srpt_check_stop_free,
3919 	.shutdown_session		= srpt_shutdown_session,
3920 	.close_session			= srpt_close_session,
3921 	.sess_get_index			= srpt_sess_get_index,
3922 	.sess_get_initiator_sid		= NULL,
3923 	.write_pending			= srpt_write_pending,
3924 	.write_pending_status		= srpt_write_pending_status,
3925 	.set_default_node_attributes	= srpt_set_default_node_attrs,
3926 	.get_task_tag			= srpt_get_task_tag,
3927 	.get_cmd_state			= srpt_get_tcm_cmd_state,
3928 	.queue_data_in			= srpt_queue_data_in,
3929 	.queue_status			= srpt_queue_status,
3930 	.queue_tm_rsp			= srpt_queue_tm_rsp,
3931 	/*
3932 	 * Setup function pointers for generic logic in
3933 	 * target_core_fabric_configfs.c
3934 	 */
3935 	.fabric_make_wwn		= srpt_make_tport,
3936 	.fabric_drop_wwn		= srpt_drop_tport,
3937 	.fabric_make_tpg		= srpt_make_tpg,
3938 	.fabric_drop_tpg		= srpt_drop_tpg,
3939 	.fabric_post_link		= NULL,
3940 	.fabric_pre_unlink		= NULL,
3941 	.fabric_make_np			= NULL,
3942 	.fabric_drop_np			= NULL,
3943 	.fabric_make_nodeacl		= srpt_make_nodeacl,
3944 	.fabric_drop_nodeacl		= srpt_drop_nodeacl,
3945 };
3946 
3947 /**
3948  * srpt_init_module() - Kernel module initialization.
3949  *
3950  * Note: Since ib_register_client() registers callback functions, and since at
3951  * least one of these callback functions (srpt_add_one()) calls target core
3952  * functions, this driver must be registered with the target core before
3953  * ib_register_client() is called.
3954  */
3955 static int __init srpt_init_module(void)
3956 {
3957 	int ret;
3958 
3959 	ret = -EINVAL;
3960 	if (srp_max_req_size < MIN_MAX_REQ_SIZE) {
3961 		printk(KERN_ERR "invalid value %d for kernel module parameter"
3962 		       " srp_max_req_size -- must be at least %d.\n",
3963 		       srp_max_req_size, MIN_MAX_REQ_SIZE);
3964 		goto out;
3965 	}
3966 
3967 	if (srpt_srq_size < MIN_SRPT_SRQ_SIZE
3968 	    || srpt_srq_size > MAX_SRPT_SRQ_SIZE) {
3969 		printk(KERN_ERR "invalid value %d for kernel module parameter"
3970 		       " srpt_srq_size -- must be in the range [%d..%d].\n",
3971 		       srpt_srq_size, MIN_SRPT_SRQ_SIZE, MAX_SRPT_SRQ_SIZE);
3972 		goto out;
3973 	}
3974 
3975 	srpt_target = target_fabric_configfs_init(THIS_MODULE, "srpt");
3976 	if (IS_ERR(srpt_target)) {
3977 		printk(KERN_ERR "couldn't register\n");
3978 		ret = PTR_ERR(srpt_target);
3979 		goto out;
3980 	}
3981 
3982 	srpt_target->tf_ops = srpt_template;
3983 
3984 	/*
3985 	 * Set up default attribute lists.
3986 	 */
3987 	srpt_target->tf_cit_tmpl.tfc_wwn_cit.ct_attrs = srpt_wwn_attrs;
3988 	srpt_target->tf_cit_tmpl.tfc_tpg_base_cit.ct_attrs = srpt_tpg_attrs;
3989 	srpt_target->tf_cit_tmpl.tfc_tpg_attrib_cit.ct_attrs = srpt_tpg_attrib_attrs;
3990 	srpt_target->tf_cit_tmpl.tfc_tpg_param_cit.ct_attrs = NULL;
3991 	srpt_target->tf_cit_tmpl.tfc_tpg_np_base_cit.ct_attrs = NULL;
3992 	srpt_target->tf_cit_tmpl.tfc_tpg_nacl_base_cit.ct_attrs = NULL;
3993 	srpt_target->tf_cit_tmpl.tfc_tpg_nacl_attrib_cit.ct_attrs = NULL;
3994 	srpt_target->tf_cit_tmpl.tfc_tpg_nacl_auth_cit.ct_attrs = NULL;
3995 	srpt_target->tf_cit_tmpl.tfc_tpg_nacl_param_cit.ct_attrs = NULL;
3996 
3997 	ret = target_fabric_configfs_register(srpt_target);
3998 	if (ret < 0) {
3999 		printk(KERN_ERR "couldn't register\n");
4000 		goto out_free_target;
4001 	}
4002 
4003 	ret = ib_register_client(&srpt_client);
4004 	if (ret) {
4005 		printk(KERN_ERR "couldn't register IB client\n");
4006 		goto out_unregister_target;
4007 	}
4008 
4009 	return 0;
4010 
4011 out_unregister_target:
4012 	target_fabric_configfs_deregister(srpt_target);
4013 	srpt_target = NULL;
4014 out_free_target:
4015 	if (srpt_target)
4016 		target_fabric_configfs_free(srpt_target);
4017 out:
4018 	return ret;
4019 }
4020 
4021 static void __exit srpt_cleanup_module(void)
4022 {
4023 	ib_unregister_client(&srpt_client);
4024 	target_fabric_configfs_deregister(srpt_target);
4025 	srpt_target = NULL;
4026 }
4027 
4028 module_init(srpt_init_module);
4029 module_exit(srpt_cleanup_module);
4030