xref: /linux/drivers/infiniband/ulp/srpt/ib_srpt.c (revision 5148fa52a12fa1b97c730b2fe321f2aad7ea041c)
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 	kref_init(&ioctx->kref);
1273 	spin_lock_init(&ioctx->spinlock);
1274 	ioctx->state = SRPT_STATE_NEW;
1275 	ioctx->n_rbuf = 0;
1276 	ioctx->rbufs = NULL;
1277 	ioctx->n_rdma = 0;
1278 	ioctx->n_rdma_ius = 0;
1279 	ioctx->rdma_ius = NULL;
1280 	ioctx->mapped_sg_count = 0;
1281 	init_completion(&ioctx->tx_done);
1282 	ioctx->queue_status_only = false;
1283 	/*
1284 	 * transport_init_se_cmd() does not initialize all fields, so do it
1285 	 * here.
1286 	 */
1287 	memset(&ioctx->cmd, 0, sizeof(ioctx->cmd));
1288 	memset(&ioctx->sense_data, 0, sizeof(ioctx->sense_data));
1289 
1290 	return ioctx;
1291 }
1292 
1293 /**
1294  * srpt_put_send_ioctx() - Free up resources.
1295  */
1296 static void srpt_put_send_ioctx(struct srpt_send_ioctx *ioctx)
1297 {
1298 	struct srpt_rdma_ch *ch;
1299 	unsigned long flags;
1300 
1301 	BUG_ON(!ioctx);
1302 	ch = ioctx->ch;
1303 	BUG_ON(!ch);
1304 
1305 	WARN_ON(srpt_get_cmd_state(ioctx) != SRPT_STATE_DONE);
1306 
1307 	srpt_unmap_sg_to_ib_sge(ioctx->ch, ioctx);
1308 	transport_generic_free_cmd(&ioctx->cmd, 0);
1309 
1310 	if (ioctx->n_rbuf > 1) {
1311 		kfree(ioctx->rbufs);
1312 		ioctx->rbufs = NULL;
1313 		ioctx->n_rbuf = 0;
1314 	}
1315 
1316 	spin_lock_irqsave(&ch->spinlock, flags);
1317 	list_add(&ioctx->free_list, &ch->free_list);
1318 	spin_unlock_irqrestore(&ch->spinlock, flags);
1319 }
1320 
1321 static void srpt_put_send_ioctx_kref(struct kref *kref)
1322 {
1323 	srpt_put_send_ioctx(container_of(kref, struct srpt_send_ioctx, kref));
1324 }
1325 
1326 /**
1327  * srpt_abort_cmd() - Abort a SCSI command.
1328  * @ioctx:   I/O context associated with the SCSI command.
1329  * @context: Preferred execution context.
1330  */
1331 static int srpt_abort_cmd(struct srpt_send_ioctx *ioctx)
1332 {
1333 	enum srpt_command_state state;
1334 	unsigned long flags;
1335 
1336 	BUG_ON(!ioctx);
1337 
1338 	/*
1339 	 * If the command is in a state where the target core is waiting for
1340 	 * the ib_srpt driver, change the state to the next state. Changing
1341 	 * the state of the command from SRPT_STATE_NEED_DATA to
1342 	 * SRPT_STATE_DATA_IN ensures that srpt_xmit_response() will call this
1343 	 * function a second time.
1344 	 */
1345 
1346 	spin_lock_irqsave(&ioctx->spinlock, flags);
1347 	state = ioctx->state;
1348 	switch (state) {
1349 	case SRPT_STATE_NEED_DATA:
1350 		ioctx->state = SRPT_STATE_DATA_IN;
1351 		break;
1352 	case SRPT_STATE_DATA_IN:
1353 	case SRPT_STATE_CMD_RSP_SENT:
1354 	case SRPT_STATE_MGMT_RSP_SENT:
1355 		ioctx->state = SRPT_STATE_DONE;
1356 		break;
1357 	default:
1358 		break;
1359 	}
1360 	spin_unlock_irqrestore(&ioctx->spinlock, flags);
1361 
1362 	if (state == SRPT_STATE_DONE)
1363 		goto out;
1364 
1365 	pr_debug("Aborting cmd with state %d and tag %lld\n", state,
1366 		 ioctx->tag);
1367 
1368 	switch (state) {
1369 	case SRPT_STATE_NEW:
1370 	case SRPT_STATE_DATA_IN:
1371 	case SRPT_STATE_MGMT:
1372 		/*
1373 		 * Do nothing - defer abort processing until
1374 		 * srpt_queue_response() is invoked.
1375 		 */
1376 		WARN_ON(!transport_check_aborted_status(&ioctx->cmd, false));
1377 		break;
1378 	case SRPT_STATE_NEED_DATA:
1379 		/* DMA_TO_DEVICE (write) - RDMA read error. */
1380 		spin_lock_irqsave(&ioctx->cmd.t_state_lock, flags);
1381 		ioctx->cmd.transport_state |= CMD_T_LUN_STOP;
1382 		spin_unlock_irqrestore(&ioctx->cmd.t_state_lock, flags);
1383 		transport_generic_handle_data(&ioctx->cmd);
1384 		break;
1385 	case SRPT_STATE_CMD_RSP_SENT:
1386 		/*
1387 		 * SRP_RSP sending failed or the SRP_RSP send completion has
1388 		 * not been received in time.
1389 		 */
1390 		srpt_unmap_sg_to_ib_sge(ioctx->ch, ioctx);
1391 		spin_lock_irqsave(&ioctx->cmd.t_state_lock, flags);
1392 		ioctx->cmd.transport_state |= CMD_T_LUN_STOP;
1393 		spin_unlock_irqrestore(&ioctx->cmd.t_state_lock, flags);
1394 		kref_put(&ioctx->kref, srpt_put_send_ioctx_kref);
1395 		break;
1396 	case SRPT_STATE_MGMT_RSP_SENT:
1397 		srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
1398 		kref_put(&ioctx->kref, srpt_put_send_ioctx_kref);
1399 		break;
1400 	default:
1401 		WARN_ON("ERROR: unexpected command state");
1402 		break;
1403 	}
1404 
1405 out:
1406 	return state;
1407 }
1408 
1409 /**
1410  * srpt_handle_send_err_comp() - Process an IB_WC_SEND error completion.
1411  */
1412 static void srpt_handle_send_err_comp(struct srpt_rdma_ch *ch, u64 wr_id)
1413 {
1414 	struct srpt_send_ioctx *ioctx;
1415 	enum srpt_command_state state;
1416 	struct se_cmd *cmd;
1417 	u32 index;
1418 
1419 	atomic_inc(&ch->sq_wr_avail);
1420 
1421 	index = idx_from_wr_id(wr_id);
1422 	ioctx = ch->ioctx_ring[index];
1423 	state = srpt_get_cmd_state(ioctx);
1424 	cmd = &ioctx->cmd;
1425 
1426 	WARN_ON(state != SRPT_STATE_CMD_RSP_SENT
1427 		&& state != SRPT_STATE_MGMT_RSP_SENT
1428 		&& state != SRPT_STATE_NEED_DATA
1429 		&& state != SRPT_STATE_DONE);
1430 
1431 	/* If SRP_RSP sending failed, undo the ch->req_lim change. */
1432 	if (state == SRPT_STATE_CMD_RSP_SENT
1433 	    || state == SRPT_STATE_MGMT_RSP_SENT)
1434 		atomic_dec(&ch->req_lim);
1435 
1436 	srpt_abort_cmd(ioctx);
1437 }
1438 
1439 /**
1440  * srpt_handle_send_comp() - Process an IB send completion notification.
1441  */
1442 static void srpt_handle_send_comp(struct srpt_rdma_ch *ch,
1443 				  struct srpt_send_ioctx *ioctx)
1444 {
1445 	enum srpt_command_state state;
1446 
1447 	atomic_inc(&ch->sq_wr_avail);
1448 
1449 	state = srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
1450 
1451 	if (WARN_ON(state != SRPT_STATE_CMD_RSP_SENT
1452 		    && state != SRPT_STATE_MGMT_RSP_SENT
1453 		    && state != SRPT_STATE_DONE))
1454 		pr_debug("state = %d\n", state);
1455 
1456 	if (state != SRPT_STATE_DONE)
1457 		kref_put(&ioctx->kref, srpt_put_send_ioctx_kref);
1458 	else
1459 		printk(KERN_ERR "IB completion has been received too late for"
1460 		       " wr_id = %u.\n", ioctx->ioctx.index);
1461 }
1462 
1463 /**
1464  * srpt_handle_rdma_comp() - Process an IB RDMA completion notification.
1465  *
1466  * Note: transport_generic_handle_data() is asynchronous so unmapping the
1467  * data that has been transferred via IB RDMA must be postponed until the
1468  * check_stop_free() callback.
1469  */
1470 static void srpt_handle_rdma_comp(struct srpt_rdma_ch *ch,
1471 				  struct srpt_send_ioctx *ioctx,
1472 				  enum srpt_opcode opcode)
1473 {
1474 	WARN_ON(ioctx->n_rdma <= 0);
1475 	atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
1476 
1477 	if (opcode == SRPT_RDMA_READ_LAST) {
1478 		if (srpt_test_and_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA,
1479 						SRPT_STATE_DATA_IN))
1480 			transport_generic_handle_data(&ioctx->cmd);
1481 		else
1482 			printk(KERN_ERR "%s[%d]: wrong state = %d\n", __func__,
1483 			       __LINE__, srpt_get_cmd_state(ioctx));
1484 	} else if (opcode == SRPT_RDMA_ABORT) {
1485 		ioctx->rdma_aborted = true;
1486 	} else {
1487 		WARN(true, "unexpected opcode %d\n", opcode);
1488 	}
1489 }
1490 
1491 /**
1492  * srpt_handle_rdma_err_comp() - Process an IB RDMA error completion.
1493  */
1494 static void srpt_handle_rdma_err_comp(struct srpt_rdma_ch *ch,
1495 				      struct srpt_send_ioctx *ioctx,
1496 				      enum srpt_opcode opcode)
1497 {
1498 	struct se_cmd *cmd;
1499 	enum srpt_command_state state;
1500 	unsigned long flags;
1501 
1502 	cmd = &ioctx->cmd;
1503 	state = srpt_get_cmd_state(ioctx);
1504 	switch (opcode) {
1505 	case SRPT_RDMA_READ_LAST:
1506 		if (ioctx->n_rdma <= 0) {
1507 			printk(KERN_ERR "Received invalid RDMA read"
1508 			       " error completion with idx %d\n",
1509 			       ioctx->ioctx.index);
1510 			break;
1511 		}
1512 		atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
1513 		if (state == SRPT_STATE_NEED_DATA)
1514 			srpt_abort_cmd(ioctx);
1515 		else
1516 			printk(KERN_ERR "%s[%d]: wrong state = %d\n",
1517 			       __func__, __LINE__, state);
1518 		break;
1519 	case SRPT_RDMA_WRITE_LAST:
1520 		spin_lock_irqsave(&ioctx->cmd.t_state_lock, flags);
1521 		ioctx->cmd.transport_state |= CMD_T_LUN_STOP;
1522 		spin_unlock_irqrestore(&ioctx->cmd.t_state_lock, flags);
1523 		break;
1524 	default:
1525 		printk(KERN_ERR "%s[%d]: opcode = %u\n", __func__,
1526 		       __LINE__, opcode);
1527 		break;
1528 	}
1529 }
1530 
1531 /**
1532  * srpt_build_cmd_rsp() - Build an SRP_RSP response.
1533  * @ch: RDMA channel through which the request has been received.
1534  * @ioctx: I/O context associated with the SRP_CMD request. The response will
1535  *   be built in the buffer ioctx->buf points at and hence this function will
1536  *   overwrite the request data.
1537  * @tag: tag of the request for which this response is being generated.
1538  * @status: value for the STATUS field of the SRP_RSP information unit.
1539  *
1540  * Returns the size in bytes of the SRP_RSP response.
1541  *
1542  * An SRP_RSP response contains a SCSI status or service response. See also
1543  * section 6.9 in the SRP r16a document for the format of an SRP_RSP
1544  * response. See also SPC-2 for more information about sense data.
1545  */
1546 static int srpt_build_cmd_rsp(struct srpt_rdma_ch *ch,
1547 			      struct srpt_send_ioctx *ioctx, u64 tag,
1548 			      int status)
1549 {
1550 	struct srp_rsp *srp_rsp;
1551 	const u8 *sense_data;
1552 	int sense_data_len, max_sense_len;
1553 
1554 	/*
1555 	 * The lowest bit of all SAM-3 status codes is zero (see also
1556 	 * paragraph 5.3 in SAM-3).
1557 	 */
1558 	WARN_ON(status & 1);
1559 
1560 	srp_rsp = ioctx->ioctx.buf;
1561 	BUG_ON(!srp_rsp);
1562 
1563 	sense_data = ioctx->sense_data;
1564 	sense_data_len = ioctx->cmd.scsi_sense_length;
1565 	WARN_ON(sense_data_len > sizeof(ioctx->sense_data));
1566 
1567 	memset(srp_rsp, 0, sizeof *srp_rsp);
1568 	srp_rsp->opcode = SRP_RSP;
1569 	srp_rsp->req_lim_delta =
1570 		__constant_cpu_to_be32(1 + atomic_xchg(&ch->req_lim_delta, 0));
1571 	srp_rsp->tag = tag;
1572 	srp_rsp->status = status;
1573 
1574 	if (sense_data_len) {
1575 		BUILD_BUG_ON(MIN_MAX_RSP_SIZE <= sizeof(*srp_rsp));
1576 		max_sense_len = ch->max_ti_iu_len - sizeof(*srp_rsp);
1577 		if (sense_data_len > max_sense_len) {
1578 			printk(KERN_WARNING "truncated sense data from %d to %d"
1579 			       " bytes\n", sense_data_len, max_sense_len);
1580 			sense_data_len = max_sense_len;
1581 		}
1582 
1583 		srp_rsp->flags |= SRP_RSP_FLAG_SNSVALID;
1584 		srp_rsp->sense_data_len = cpu_to_be32(sense_data_len);
1585 		memcpy(srp_rsp + 1, sense_data, sense_data_len);
1586 	}
1587 
1588 	return sizeof(*srp_rsp) + sense_data_len;
1589 }
1590 
1591 /**
1592  * srpt_build_tskmgmt_rsp() - Build a task management response.
1593  * @ch:       RDMA channel through which the request has been received.
1594  * @ioctx:    I/O context in which the SRP_RSP response will be built.
1595  * @rsp_code: RSP_CODE that will be stored in the response.
1596  * @tag:      Tag of the request for which this response is being generated.
1597  *
1598  * Returns the size in bytes of the SRP_RSP response.
1599  *
1600  * An SRP_RSP response contains a SCSI status or service response. See also
1601  * section 6.9 in the SRP r16a document for the format of an SRP_RSP
1602  * response.
1603  */
1604 static int srpt_build_tskmgmt_rsp(struct srpt_rdma_ch *ch,
1605 				  struct srpt_send_ioctx *ioctx,
1606 				  u8 rsp_code, u64 tag)
1607 {
1608 	struct srp_rsp *srp_rsp;
1609 	int resp_data_len;
1610 	int resp_len;
1611 
1612 	resp_data_len = (rsp_code == SRP_TSK_MGMT_SUCCESS) ? 0 : 4;
1613 	resp_len = sizeof(*srp_rsp) + resp_data_len;
1614 
1615 	srp_rsp = ioctx->ioctx.buf;
1616 	BUG_ON(!srp_rsp);
1617 	memset(srp_rsp, 0, sizeof *srp_rsp);
1618 
1619 	srp_rsp->opcode = SRP_RSP;
1620 	srp_rsp->req_lim_delta = __constant_cpu_to_be32(1
1621 				    + atomic_xchg(&ch->req_lim_delta, 0));
1622 	srp_rsp->tag = tag;
1623 
1624 	if (rsp_code != SRP_TSK_MGMT_SUCCESS) {
1625 		srp_rsp->flags |= SRP_RSP_FLAG_RSPVALID;
1626 		srp_rsp->resp_data_len = cpu_to_be32(resp_data_len);
1627 		srp_rsp->data[3] = rsp_code;
1628 	}
1629 
1630 	return resp_len;
1631 }
1632 
1633 #define NO_SUCH_LUN ((uint64_t)-1LL)
1634 
1635 /*
1636  * SCSI LUN addressing method. See also SAM-2 and the section about
1637  * eight byte LUNs.
1638  */
1639 enum scsi_lun_addr_method {
1640 	SCSI_LUN_ADDR_METHOD_PERIPHERAL   = 0,
1641 	SCSI_LUN_ADDR_METHOD_FLAT         = 1,
1642 	SCSI_LUN_ADDR_METHOD_LUN          = 2,
1643 	SCSI_LUN_ADDR_METHOD_EXTENDED_LUN = 3,
1644 };
1645 
1646 /*
1647  * srpt_unpack_lun() - Convert from network LUN to linear LUN.
1648  *
1649  * Convert an 2-byte, 4-byte, 6-byte or 8-byte LUN structure in network byte
1650  * order (big endian) to a linear LUN. Supports three LUN addressing methods:
1651  * peripheral, flat and logical unit. See also SAM-2, section 4.9.4 (page 40).
1652  */
1653 static uint64_t srpt_unpack_lun(const uint8_t *lun, int len)
1654 {
1655 	uint64_t res = NO_SUCH_LUN;
1656 	int addressing_method;
1657 
1658 	if (unlikely(len < 2)) {
1659 		printk(KERN_ERR "Illegal LUN length %d, expected 2 bytes or "
1660 		       "more", len);
1661 		goto out;
1662 	}
1663 
1664 	switch (len) {
1665 	case 8:
1666 		if ((*((__be64 *)lun) &
1667 		     __constant_cpu_to_be64(0x0000FFFFFFFFFFFFLL)) != 0)
1668 			goto out_err;
1669 		break;
1670 	case 4:
1671 		if (*((__be16 *)&lun[2]) != 0)
1672 			goto out_err;
1673 		break;
1674 	case 6:
1675 		if (*((__be32 *)&lun[2]) != 0)
1676 			goto out_err;
1677 		break;
1678 	case 2:
1679 		break;
1680 	default:
1681 		goto out_err;
1682 	}
1683 
1684 	addressing_method = (*lun) >> 6; /* highest two bits of byte 0 */
1685 	switch (addressing_method) {
1686 	case SCSI_LUN_ADDR_METHOD_PERIPHERAL:
1687 	case SCSI_LUN_ADDR_METHOD_FLAT:
1688 	case SCSI_LUN_ADDR_METHOD_LUN:
1689 		res = *(lun + 1) | (((*lun) & 0x3f) << 8);
1690 		break;
1691 
1692 	case SCSI_LUN_ADDR_METHOD_EXTENDED_LUN:
1693 	default:
1694 		printk(KERN_ERR "Unimplemented LUN addressing method %u",
1695 		       addressing_method);
1696 		break;
1697 	}
1698 
1699 out:
1700 	return res;
1701 
1702 out_err:
1703 	printk(KERN_ERR "Support for multi-level LUNs has not yet been"
1704 	       " implemented");
1705 	goto out;
1706 }
1707 
1708 static int srpt_check_stop_free(struct se_cmd *cmd)
1709 {
1710 	struct srpt_send_ioctx *ioctx;
1711 
1712 	ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
1713 	return kref_put(&ioctx->kref, srpt_put_send_ioctx_kref);
1714 }
1715 
1716 /**
1717  * srpt_handle_cmd() - Process SRP_CMD.
1718  */
1719 static int srpt_handle_cmd(struct srpt_rdma_ch *ch,
1720 			   struct srpt_recv_ioctx *recv_ioctx,
1721 			   struct srpt_send_ioctx *send_ioctx)
1722 {
1723 	struct se_cmd *cmd;
1724 	struct srp_cmd *srp_cmd;
1725 	uint64_t unpacked_lun;
1726 	u64 data_len;
1727 	enum dma_data_direction dir;
1728 	int ret;
1729 
1730 	BUG_ON(!send_ioctx);
1731 
1732 	srp_cmd = recv_ioctx->ioctx.buf;
1733 	kref_get(&send_ioctx->kref);
1734 	cmd = &send_ioctx->cmd;
1735 	send_ioctx->tag = srp_cmd->tag;
1736 
1737 	switch (srp_cmd->task_attr) {
1738 	case SRP_CMD_SIMPLE_Q:
1739 		cmd->sam_task_attr = MSG_SIMPLE_TAG;
1740 		break;
1741 	case SRP_CMD_ORDERED_Q:
1742 	default:
1743 		cmd->sam_task_attr = MSG_ORDERED_TAG;
1744 		break;
1745 	case SRP_CMD_HEAD_OF_Q:
1746 		cmd->sam_task_attr = MSG_HEAD_TAG;
1747 		break;
1748 	case SRP_CMD_ACA:
1749 		cmd->sam_task_attr = MSG_ACA_TAG;
1750 		break;
1751 	}
1752 
1753 	ret = srpt_get_desc_tbl(send_ioctx, srp_cmd, &dir, &data_len);
1754 	if (ret) {
1755 		printk(KERN_ERR "0x%llx: parsing SRP descriptor table failed.\n",
1756 		       srp_cmd->tag);
1757 		cmd->se_cmd_flags |= SCF_SCSI_CDB_EXCEPTION;
1758 		cmd->scsi_sense_reason = TCM_INVALID_CDB_FIELD;
1759 		kref_put(&send_ioctx->kref, srpt_put_send_ioctx_kref);
1760 		goto send_sense;
1761 	}
1762 
1763 	cmd->data_length = data_len;
1764 	cmd->data_direction = dir;
1765 	unpacked_lun = srpt_unpack_lun((uint8_t *)&srp_cmd->lun,
1766 				       sizeof(srp_cmd->lun));
1767 	if (transport_lookup_cmd_lun(cmd, unpacked_lun) < 0) {
1768 		kref_put(&send_ioctx->kref, srpt_put_send_ioctx_kref);
1769 		goto send_sense;
1770 	}
1771 	ret = target_setup_cmd_from_cdb(cmd, srp_cmd->cdb);
1772 	if (ret < 0) {
1773 		kref_put(&send_ioctx->kref, srpt_put_send_ioctx_kref);
1774 		if (cmd->se_cmd_flags & SCF_SCSI_RESERVATION_CONFLICT) {
1775 			srpt_queue_status(cmd);
1776 			return 0;
1777 		} else
1778 			goto send_sense;
1779 	}
1780 
1781 	transport_handle_cdb_direct(cmd);
1782 	return 0;
1783 
1784 send_sense:
1785 	transport_send_check_condition_and_sense(cmd, cmd->scsi_sense_reason,
1786 						 0);
1787 	return -1;
1788 }
1789 
1790 /**
1791  * srpt_rx_mgmt_fn_tag() - Process a task management function by tag.
1792  * @ch: RDMA channel of the task management request.
1793  * @fn: Task management function to perform.
1794  * @req_tag: Tag of the SRP task management request.
1795  * @mgmt_ioctx: I/O context of the task management request.
1796  *
1797  * Returns zero if the target core will process the task management
1798  * request asynchronously.
1799  *
1800  * Note: It is assumed that the initiator serializes tag-based task management
1801  * requests.
1802  */
1803 static int srpt_rx_mgmt_fn_tag(struct srpt_send_ioctx *ioctx, u64 tag)
1804 {
1805 	struct srpt_device *sdev;
1806 	struct srpt_rdma_ch *ch;
1807 	struct srpt_send_ioctx *target;
1808 	int ret, i;
1809 
1810 	ret = -EINVAL;
1811 	ch = ioctx->ch;
1812 	BUG_ON(!ch);
1813 	BUG_ON(!ch->sport);
1814 	sdev = ch->sport->sdev;
1815 	BUG_ON(!sdev);
1816 	spin_lock_irq(&sdev->spinlock);
1817 	for (i = 0; i < ch->rq_size; ++i) {
1818 		target = ch->ioctx_ring[i];
1819 		if (target->cmd.se_lun == ioctx->cmd.se_lun &&
1820 		    target->tag == tag &&
1821 		    srpt_get_cmd_state(target) != SRPT_STATE_DONE) {
1822 			ret = 0;
1823 			/* now let the target core abort &target->cmd; */
1824 			break;
1825 		}
1826 	}
1827 	spin_unlock_irq(&sdev->spinlock);
1828 	return ret;
1829 }
1830 
1831 static int srp_tmr_to_tcm(int fn)
1832 {
1833 	switch (fn) {
1834 	case SRP_TSK_ABORT_TASK:
1835 		return TMR_ABORT_TASK;
1836 	case SRP_TSK_ABORT_TASK_SET:
1837 		return TMR_ABORT_TASK_SET;
1838 	case SRP_TSK_CLEAR_TASK_SET:
1839 		return TMR_CLEAR_TASK_SET;
1840 	case SRP_TSK_LUN_RESET:
1841 		return TMR_LUN_RESET;
1842 	case SRP_TSK_CLEAR_ACA:
1843 		return TMR_CLEAR_ACA;
1844 	default:
1845 		return -1;
1846 	}
1847 }
1848 
1849 /**
1850  * srpt_handle_tsk_mgmt() - Process an SRP_TSK_MGMT information unit.
1851  *
1852  * Returns 0 if and only if the request will be processed by the target core.
1853  *
1854  * For more information about SRP_TSK_MGMT information units, see also section
1855  * 6.7 in the SRP r16a document.
1856  */
1857 static void srpt_handle_tsk_mgmt(struct srpt_rdma_ch *ch,
1858 				 struct srpt_recv_ioctx *recv_ioctx,
1859 				 struct srpt_send_ioctx *send_ioctx)
1860 {
1861 	struct srp_tsk_mgmt *srp_tsk;
1862 	struct se_cmd *cmd;
1863 	uint64_t unpacked_lun;
1864 	int tcm_tmr;
1865 	int res;
1866 
1867 	BUG_ON(!send_ioctx);
1868 
1869 	srp_tsk = recv_ioctx->ioctx.buf;
1870 	cmd = &send_ioctx->cmd;
1871 
1872 	pr_debug("recv tsk_mgmt fn %d for task_tag %lld and cmd tag %lld"
1873 		 " cm_id %p sess %p\n", srp_tsk->tsk_mgmt_func,
1874 		 srp_tsk->task_tag, srp_tsk->tag, ch->cm_id, ch->sess);
1875 
1876 	srpt_set_cmd_state(send_ioctx, SRPT_STATE_MGMT);
1877 	send_ioctx->tag = srp_tsk->tag;
1878 	tcm_tmr = srp_tmr_to_tcm(srp_tsk->tsk_mgmt_func);
1879 	if (tcm_tmr < 0) {
1880 		send_ioctx->cmd.se_cmd_flags |= SCF_SCSI_CDB_EXCEPTION;
1881 		send_ioctx->cmd.se_tmr_req->response =
1882 			TMR_TASK_MGMT_FUNCTION_NOT_SUPPORTED;
1883 		goto process_tmr;
1884 	}
1885 	res = core_tmr_alloc_req(cmd, NULL, tcm_tmr, GFP_KERNEL);
1886 	if (res < 0) {
1887 		send_ioctx->cmd.se_cmd_flags |= SCF_SCSI_CDB_EXCEPTION;
1888 		send_ioctx->cmd.se_tmr_req->response = TMR_FUNCTION_REJECTED;
1889 		goto process_tmr;
1890 	}
1891 
1892 	unpacked_lun = srpt_unpack_lun((uint8_t *)&srp_tsk->lun,
1893 				       sizeof(srp_tsk->lun));
1894 	res = transport_lookup_tmr_lun(&send_ioctx->cmd, unpacked_lun);
1895 	if (res) {
1896 		pr_debug("rejecting TMR for LUN %lld\n", unpacked_lun);
1897 		send_ioctx->cmd.se_cmd_flags |= SCF_SCSI_CDB_EXCEPTION;
1898 		send_ioctx->cmd.se_tmr_req->response = TMR_LUN_DOES_NOT_EXIST;
1899 		goto process_tmr;
1900 	}
1901 
1902 	if (srp_tsk->tsk_mgmt_func == SRP_TSK_ABORT_TASK)
1903 		srpt_rx_mgmt_fn_tag(send_ioctx, srp_tsk->task_tag);
1904 
1905 process_tmr:
1906 	kref_get(&send_ioctx->kref);
1907 	if (!(send_ioctx->cmd.se_cmd_flags & SCF_SCSI_CDB_EXCEPTION))
1908 		transport_generic_handle_tmr(&send_ioctx->cmd);
1909 	else
1910 		transport_send_check_condition_and_sense(cmd,
1911 						cmd->scsi_sense_reason, 0);
1912 
1913 }
1914 
1915 /**
1916  * srpt_handle_new_iu() - Process a newly received information unit.
1917  * @ch:    RDMA channel through which the information unit has been received.
1918  * @ioctx: SRPT I/O context associated with the information unit.
1919  */
1920 static void srpt_handle_new_iu(struct srpt_rdma_ch *ch,
1921 			       struct srpt_recv_ioctx *recv_ioctx,
1922 			       struct srpt_send_ioctx *send_ioctx)
1923 {
1924 	struct srp_cmd *srp_cmd;
1925 	enum rdma_ch_state ch_state;
1926 
1927 	BUG_ON(!ch);
1928 	BUG_ON(!recv_ioctx);
1929 
1930 	ib_dma_sync_single_for_cpu(ch->sport->sdev->device,
1931 				   recv_ioctx->ioctx.dma, srp_max_req_size,
1932 				   DMA_FROM_DEVICE);
1933 
1934 	ch_state = srpt_get_ch_state(ch);
1935 	if (unlikely(ch_state == CH_CONNECTING)) {
1936 		list_add_tail(&recv_ioctx->wait_list, &ch->cmd_wait_list);
1937 		goto out;
1938 	}
1939 
1940 	if (unlikely(ch_state != CH_LIVE))
1941 		goto out;
1942 
1943 	srp_cmd = recv_ioctx->ioctx.buf;
1944 	if (srp_cmd->opcode == SRP_CMD || srp_cmd->opcode == SRP_TSK_MGMT) {
1945 		if (!send_ioctx)
1946 			send_ioctx = srpt_get_send_ioctx(ch);
1947 		if (unlikely(!send_ioctx)) {
1948 			list_add_tail(&recv_ioctx->wait_list,
1949 				      &ch->cmd_wait_list);
1950 			goto out;
1951 		}
1952 	}
1953 
1954 	transport_init_se_cmd(&send_ioctx->cmd, &srpt_target->tf_ops, ch->sess,
1955 			      0, DMA_NONE, MSG_SIMPLE_TAG,
1956 			      send_ioctx->sense_data);
1957 
1958 	switch (srp_cmd->opcode) {
1959 	case SRP_CMD:
1960 		srpt_handle_cmd(ch, recv_ioctx, send_ioctx);
1961 		break;
1962 	case SRP_TSK_MGMT:
1963 		srpt_handle_tsk_mgmt(ch, recv_ioctx, send_ioctx);
1964 		break;
1965 	case SRP_I_LOGOUT:
1966 		printk(KERN_ERR "Not yet implemented: SRP_I_LOGOUT\n");
1967 		break;
1968 	case SRP_CRED_RSP:
1969 		pr_debug("received SRP_CRED_RSP\n");
1970 		break;
1971 	case SRP_AER_RSP:
1972 		pr_debug("received SRP_AER_RSP\n");
1973 		break;
1974 	case SRP_RSP:
1975 		printk(KERN_ERR "Received SRP_RSP\n");
1976 		break;
1977 	default:
1978 		printk(KERN_ERR "received IU with unknown opcode 0x%x\n",
1979 		       srp_cmd->opcode);
1980 		break;
1981 	}
1982 
1983 	srpt_post_recv(ch->sport->sdev, recv_ioctx);
1984 out:
1985 	return;
1986 }
1987 
1988 static void srpt_process_rcv_completion(struct ib_cq *cq,
1989 					struct srpt_rdma_ch *ch,
1990 					struct ib_wc *wc)
1991 {
1992 	struct srpt_device *sdev = ch->sport->sdev;
1993 	struct srpt_recv_ioctx *ioctx;
1994 	u32 index;
1995 
1996 	index = idx_from_wr_id(wc->wr_id);
1997 	if (wc->status == IB_WC_SUCCESS) {
1998 		int req_lim;
1999 
2000 		req_lim = atomic_dec_return(&ch->req_lim);
2001 		if (unlikely(req_lim < 0))
2002 			printk(KERN_ERR "req_lim = %d < 0\n", req_lim);
2003 		ioctx = sdev->ioctx_ring[index];
2004 		srpt_handle_new_iu(ch, ioctx, NULL);
2005 	} else {
2006 		printk(KERN_INFO "receiving failed for idx %u with status %d\n",
2007 		       index, wc->status);
2008 	}
2009 }
2010 
2011 /**
2012  * srpt_process_send_completion() - Process an IB send completion.
2013  *
2014  * Note: Although this has not yet been observed during tests, at least in
2015  * theory it is possible that the srpt_get_send_ioctx() call invoked by
2016  * srpt_handle_new_iu() fails. This is possible because the req_lim_delta
2017  * value in each response is set to one, and it is possible that this response
2018  * makes the initiator send a new request before the send completion for that
2019  * response has been processed. This could e.g. happen if the call to
2020  * srpt_put_send_iotcx() is delayed because of a higher priority interrupt or
2021  * if IB retransmission causes generation of the send completion to be
2022  * delayed. Incoming information units for which srpt_get_send_ioctx() fails
2023  * are queued on cmd_wait_list. The code below processes these delayed
2024  * requests one at a time.
2025  */
2026 static void srpt_process_send_completion(struct ib_cq *cq,
2027 					 struct srpt_rdma_ch *ch,
2028 					 struct ib_wc *wc)
2029 {
2030 	struct srpt_send_ioctx *send_ioctx;
2031 	uint32_t index;
2032 	enum srpt_opcode opcode;
2033 
2034 	index = idx_from_wr_id(wc->wr_id);
2035 	opcode = opcode_from_wr_id(wc->wr_id);
2036 	send_ioctx = ch->ioctx_ring[index];
2037 	if (wc->status == IB_WC_SUCCESS) {
2038 		if (opcode == SRPT_SEND)
2039 			srpt_handle_send_comp(ch, send_ioctx);
2040 		else {
2041 			WARN_ON(opcode != SRPT_RDMA_ABORT &&
2042 				wc->opcode != IB_WC_RDMA_READ);
2043 			srpt_handle_rdma_comp(ch, send_ioctx, opcode);
2044 		}
2045 	} else {
2046 		if (opcode == SRPT_SEND) {
2047 			printk(KERN_INFO "sending response for idx %u failed"
2048 			       " with status %d\n", index, wc->status);
2049 			srpt_handle_send_err_comp(ch, wc->wr_id);
2050 		} else if (opcode != SRPT_RDMA_MID) {
2051 			printk(KERN_INFO "RDMA t %d for idx %u failed with"
2052 				" status %d", opcode, index, wc->status);
2053 			srpt_handle_rdma_err_comp(ch, send_ioctx, opcode);
2054 		}
2055 	}
2056 
2057 	while (unlikely(opcode == SRPT_SEND
2058 			&& !list_empty(&ch->cmd_wait_list)
2059 			&& srpt_get_ch_state(ch) == CH_LIVE
2060 			&& (send_ioctx = srpt_get_send_ioctx(ch)) != NULL)) {
2061 		struct srpt_recv_ioctx *recv_ioctx;
2062 
2063 		recv_ioctx = list_first_entry(&ch->cmd_wait_list,
2064 					      struct srpt_recv_ioctx,
2065 					      wait_list);
2066 		list_del(&recv_ioctx->wait_list);
2067 		srpt_handle_new_iu(ch, recv_ioctx, send_ioctx);
2068 	}
2069 }
2070 
2071 static void srpt_process_completion(struct ib_cq *cq, struct srpt_rdma_ch *ch)
2072 {
2073 	struct ib_wc *const wc = ch->wc;
2074 	int i, n;
2075 
2076 	WARN_ON(cq != ch->cq);
2077 
2078 	ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
2079 	while ((n = ib_poll_cq(cq, ARRAY_SIZE(ch->wc), wc)) > 0) {
2080 		for (i = 0; i < n; i++) {
2081 			if (opcode_from_wr_id(wc[i].wr_id) == SRPT_RECV)
2082 				srpt_process_rcv_completion(cq, ch, &wc[i]);
2083 			else
2084 				srpt_process_send_completion(cq, ch, &wc[i]);
2085 		}
2086 	}
2087 }
2088 
2089 /**
2090  * srpt_completion() - IB completion queue callback function.
2091  *
2092  * Notes:
2093  * - It is guaranteed that a completion handler will never be invoked
2094  *   concurrently on two different CPUs for the same completion queue. See also
2095  *   Documentation/infiniband/core_locking.txt and the implementation of
2096  *   handle_edge_irq() in kernel/irq/chip.c.
2097  * - When threaded IRQs are enabled, completion handlers are invoked in thread
2098  *   context instead of interrupt context.
2099  */
2100 static void srpt_completion(struct ib_cq *cq, void *ctx)
2101 {
2102 	struct srpt_rdma_ch *ch = ctx;
2103 
2104 	wake_up_interruptible(&ch->wait_queue);
2105 }
2106 
2107 static int srpt_compl_thread(void *arg)
2108 {
2109 	struct srpt_rdma_ch *ch;
2110 
2111 	/* Hibernation / freezing of the SRPT kernel thread is not supported. */
2112 	current->flags |= PF_NOFREEZE;
2113 
2114 	ch = arg;
2115 	BUG_ON(!ch);
2116 	printk(KERN_INFO "Session %s: kernel thread %s (PID %d) started\n",
2117 	       ch->sess_name, ch->thread->comm, current->pid);
2118 	while (!kthread_should_stop()) {
2119 		wait_event_interruptible(ch->wait_queue,
2120 			(srpt_process_completion(ch->cq, ch),
2121 			 kthread_should_stop()));
2122 	}
2123 	printk(KERN_INFO "Session %s: kernel thread %s (PID %d) stopped\n",
2124 	       ch->sess_name, ch->thread->comm, current->pid);
2125 	return 0;
2126 }
2127 
2128 /**
2129  * srpt_create_ch_ib() - Create receive and send completion queues.
2130  */
2131 static int srpt_create_ch_ib(struct srpt_rdma_ch *ch)
2132 {
2133 	struct ib_qp_init_attr *qp_init;
2134 	struct srpt_port *sport = ch->sport;
2135 	struct srpt_device *sdev = sport->sdev;
2136 	u32 srp_sq_size = sport->port_attrib.srp_sq_size;
2137 	int ret;
2138 
2139 	WARN_ON(ch->rq_size < 1);
2140 
2141 	ret = -ENOMEM;
2142 	qp_init = kzalloc(sizeof *qp_init, GFP_KERNEL);
2143 	if (!qp_init)
2144 		goto out;
2145 
2146 	ch->cq = ib_create_cq(sdev->device, srpt_completion, NULL, ch,
2147 			      ch->rq_size + srp_sq_size, 0);
2148 	if (IS_ERR(ch->cq)) {
2149 		ret = PTR_ERR(ch->cq);
2150 		printk(KERN_ERR "failed to create CQ cqe= %d ret= %d\n",
2151 		       ch->rq_size + srp_sq_size, ret);
2152 		goto out;
2153 	}
2154 
2155 	qp_init->qp_context = (void *)ch;
2156 	qp_init->event_handler
2157 		= (void(*)(struct ib_event *, void*))srpt_qp_event;
2158 	qp_init->send_cq = ch->cq;
2159 	qp_init->recv_cq = ch->cq;
2160 	qp_init->srq = sdev->srq;
2161 	qp_init->sq_sig_type = IB_SIGNAL_REQ_WR;
2162 	qp_init->qp_type = IB_QPT_RC;
2163 	qp_init->cap.max_send_wr = srp_sq_size;
2164 	qp_init->cap.max_send_sge = SRPT_DEF_SG_PER_WQE;
2165 
2166 	ch->qp = ib_create_qp(sdev->pd, qp_init);
2167 	if (IS_ERR(ch->qp)) {
2168 		ret = PTR_ERR(ch->qp);
2169 		printk(KERN_ERR "failed to create_qp ret= %d\n", ret);
2170 		goto err_destroy_cq;
2171 	}
2172 
2173 	atomic_set(&ch->sq_wr_avail, qp_init->cap.max_send_wr);
2174 
2175 	pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
2176 		 __func__, ch->cq->cqe, qp_init->cap.max_send_sge,
2177 		 qp_init->cap.max_send_wr, ch->cm_id);
2178 
2179 	ret = srpt_init_ch_qp(ch, ch->qp);
2180 	if (ret)
2181 		goto err_destroy_qp;
2182 
2183 	init_waitqueue_head(&ch->wait_queue);
2184 
2185 	pr_debug("creating thread for session %s\n", ch->sess_name);
2186 
2187 	ch->thread = kthread_run(srpt_compl_thread, ch, "ib_srpt_compl");
2188 	if (IS_ERR(ch->thread)) {
2189 		printk(KERN_ERR "failed to create kernel thread %ld\n",
2190 		       PTR_ERR(ch->thread));
2191 		ch->thread = NULL;
2192 		goto err_destroy_qp;
2193 	}
2194 
2195 out:
2196 	kfree(qp_init);
2197 	return ret;
2198 
2199 err_destroy_qp:
2200 	ib_destroy_qp(ch->qp);
2201 err_destroy_cq:
2202 	ib_destroy_cq(ch->cq);
2203 	goto out;
2204 }
2205 
2206 static void srpt_destroy_ch_ib(struct srpt_rdma_ch *ch)
2207 {
2208 	if (ch->thread)
2209 		kthread_stop(ch->thread);
2210 
2211 	ib_destroy_qp(ch->qp);
2212 	ib_destroy_cq(ch->cq);
2213 }
2214 
2215 /**
2216  * __srpt_close_ch() - Close an RDMA channel by setting the QP error state.
2217  *
2218  * Reset the QP and make sure all resources associated with the channel will
2219  * be deallocated at an appropriate time.
2220  *
2221  * Note: The caller must hold ch->sport->sdev->spinlock.
2222  */
2223 static void __srpt_close_ch(struct srpt_rdma_ch *ch)
2224 {
2225 	struct srpt_device *sdev;
2226 	enum rdma_ch_state prev_state;
2227 	unsigned long flags;
2228 
2229 	sdev = ch->sport->sdev;
2230 
2231 	spin_lock_irqsave(&ch->spinlock, flags);
2232 	prev_state = ch->state;
2233 	switch (prev_state) {
2234 	case CH_CONNECTING:
2235 	case CH_LIVE:
2236 		ch->state = CH_DISCONNECTING;
2237 		break;
2238 	default:
2239 		break;
2240 	}
2241 	spin_unlock_irqrestore(&ch->spinlock, flags);
2242 
2243 	switch (prev_state) {
2244 	case CH_CONNECTING:
2245 		ib_send_cm_rej(ch->cm_id, IB_CM_REJ_NO_RESOURCES, NULL, 0,
2246 			       NULL, 0);
2247 		/* fall through */
2248 	case CH_LIVE:
2249 		if (ib_send_cm_dreq(ch->cm_id, NULL, 0) < 0)
2250 			printk(KERN_ERR "sending CM DREQ failed.\n");
2251 		break;
2252 	case CH_DISCONNECTING:
2253 		break;
2254 	case CH_DRAINING:
2255 	case CH_RELEASING:
2256 		break;
2257 	}
2258 }
2259 
2260 /**
2261  * srpt_close_ch() - Close an RDMA channel.
2262  */
2263 static void srpt_close_ch(struct srpt_rdma_ch *ch)
2264 {
2265 	struct srpt_device *sdev;
2266 
2267 	sdev = ch->sport->sdev;
2268 	spin_lock_irq(&sdev->spinlock);
2269 	__srpt_close_ch(ch);
2270 	spin_unlock_irq(&sdev->spinlock);
2271 }
2272 
2273 /**
2274  * srpt_drain_channel() - Drain a channel by resetting the IB queue pair.
2275  * @cm_id: Pointer to the CM ID of the channel to be drained.
2276  *
2277  * Note: Must be called from inside srpt_cm_handler to avoid a race between
2278  * accessing sdev->spinlock and the call to kfree(sdev) in srpt_remove_one()
2279  * (the caller of srpt_cm_handler holds the cm_id spinlock; srpt_remove_one()
2280  * waits until all target sessions for the associated IB device have been
2281  * unregistered and target session registration involves a call to
2282  * ib_destroy_cm_id(), which locks the cm_id spinlock and hence waits until
2283  * this function has finished).
2284  */
2285 static void srpt_drain_channel(struct ib_cm_id *cm_id)
2286 {
2287 	struct srpt_device *sdev;
2288 	struct srpt_rdma_ch *ch;
2289 	int ret;
2290 	bool do_reset = false;
2291 
2292 	WARN_ON_ONCE(irqs_disabled());
2293 
2294 	sdev = cm_id->context;
2295 	BUG_ON(!sdev);
2296 	spin_lock_irq(&sdev->spinlock);
2297 	list_for_each_entry(ch, &sdev->rch_list, list) {
2298 		if (ch->cm_id == cm_id) {
2299 			do_reset = srpt_test_and_set_ch_state(ch,
2300 					CH_CONNECTING, CH_DRAINING) ||
2301 				   srpt_test_and_set_ch_state(ch,
2302 					CH_LIVE, CH_DRAINING) ||
2303 				   srpt_test_and_set_ch_state(ch,
2304 					CH_DISCONNECTING, CH_DRAINING);
2305 			break;
2306 		}
2307 	}
2308 	spin_unlock_irq(&sdev->spinlock);
2309 
2310 	if (do_reset) {
2311 		ret = srpt_ch_qp_err(ch);
2312 		if (ret < 0)
2313 			printk(KERN_ERR "Setting queue pair in error state"
2314 			       " failed: %d\n", ret);
2315 	}
2316 }
2317 
2318 /**
2319  * srpt_find_channel() - Look up an RDMA channel.
2320  * @cm_id: Pointer to the CM ID of the channel to be looked up.
2321  *
2322  * Return NULL if no matching RDMA channel has been found.
2323  */
2324 static struct srpt_rdma_ch *srpt_find_channel(struct srpt_device *sdev,
2325 					      struct ib_cm_id *cm_id)
2326 {
2327 	struct srpt_rdma_ch *ch;
2328 	bool found;
2329 
2330 	WARN_ON_ONCE(irqs_disabled());
2331 	BUG_ON(!sdev);
2332 
2333 	found = false;
2334 	spin_lock_irq(&sdev->spinlock);
2335 	list_for_each_entry(ch, &sdev->rch_list, list) {
2336 		if (ch->cm_id == cm_id) {
2337 			found = true;
2338 			break;
2339 		}
2340 	}
2341 	spin_unlock_irq(&sdev->spinlock);
2342 
2343 	return found ? ch : NULL;
2344 }
2345 
2346 /**
2347  * srpt_release_channel() - Release channel resources.
2348  *
2349  * Schedules the actual release because:
2350  * - Calling the ib_destroy_cm_id() call from inside an IB CM callback would
2351  *   trigger a deadlock.
2352  * - It is not safe to call TCM transport_* functions from interrupt context.
2353  */
2354 static void srpt_release_channel(struct srpt_rdma_ch *ch)
2355 {
2356 	schedule_work(&ch->release_work);
2357 }
2358 
2359 static void srpt_release_channel_work(struct work_struct *w)
2360 {
2361 	struct srpt_rdma_ch *ch;
2362 	struct srpt_device *sdev;
2363 
2364 	ch = container_of(w, struct srpt_rdma_ch, release_work);
2365 	pr_debug("ch = %p; ch->sess = %p; release_done = %p\n", ch, ch->sess,
2366 		 ch->release_done);
2367 
2368 	sdev = ch->sport->sdev;
2369 	BUG_ON(!sdev);
2370 
2371 	transport_deregister_session_configfs(ch->sess);
2372 	transport_deregister_session(ch->sess);
2373 	ch->sess = NULL;
2374 
2375 	srpt_destroy_ch_ib(ch);
2376 
2377 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
2378 			     ch->sport->sdev, ch->rq_size,
2379 			     ch->rsp_size, DMA_TO_DEVICE);
2380 
2381 	spin_lock_irq(&sdev->spinlock);
2382 	list_del(&ch->list);
2383 	spin_unlock_irq(&sdev->spinlock);
2384 
2385 	ib_destroy_cm_id(ch->cm_id);
2386 
2387 	if (ch->release_done)
2388 		complete(ch->release_done);
2389 
2390 	wake_up(&sdev->ch_releaseQ);
2391 
2392 	kfree(ch);
2393 }
2394 
2395 static struct srpt_node_acl *__srpt_lookup_acl(struct srpt_port *sport,
2396 					       u8 i_port_id[16])
2397 {
2398 	struct srpt_node_acl *nacl;
2399 
2400 	list_for_each_entry(nacl, &sport->port_acl_list, list)
2401 		if (memcmp(nacl->i_port_id, i_port_id,
2402 			   sizeof(nacl->i_port_id)) == 0)
2403 			return nacl;
2404 
2405 	return NULL;
2406 }
2407 
2408 static struct srpt_node_acl *srpt_lookup_acl(struct srpt_port *sport,
2409 					     u8 i_port_id[16])
2410 {
2411 	struct srpt_node_acl *nacl;
2412 
2413 	spin_lock_irq(&sport->port_acl_lock);
2414 	nacl = __srpt_lookup_acl(sport, i_port_id);
2415 	spin_unlock_irq(&sport->port_acl_lock);
2416 
2417 	return nacl;
2418 }
2419 
2420 /**
2421  * srpt_cm_req_recv() - Process the event IB_CM_REQ_RECEIVED.
2422  *
2423  * Ownership of the cm_id is transferred to the target session if this
2424  * functions returns zero. Otherwise the caller remains the owner of cm_id.
2425  */
2426 static int srpt_cm_req_recv(struct ib_cm_id *cm_id,
2427 			    struct ib_cm_req_event_param *param,
2428 			    void *private_data)
2429 {
2430 	struct srpt_device *sdev = cm_id->context;
2431 	struct srpt_port *sport = &sdev->port[param->port - 1];
2432 	struct srp_login_req *req;
2433 	struct srp_login_rsp *rsp;
2434 	struct srp_login_rej *rej;
2435 	struct ib_cm_rep_param *rep_param;
2436 	struct srpt_rdma_ch *ch, *tmp_ch;
2437 	struct srpt_node_acl *nacl;
2438 	u32 it_iu_len;
2439 	int i;
2440 	int ret = 0;
2441 
2442 	WARN_ON_ONCE(irqs_disabled());
2443 
2444 	if (WARN_ON(!sdev || !private_data))
2445 		return -EINVAL;
2446 
2447 	req = (struct srp_login_req *)private_data;
2448 
2449 	it_iu_len = be32_to_cpu(req->req_it_iu_len);
2450 
2451 	printk(KERN_INFO "Received SRP_LOGIN_REQ with i_port_id 0x%llx:0x%llx,"
2452 	       " t_port_id 0x%llx:0x%llx and it_iu_len %d on port %d"
2453 	       " (guid=0x%llx:0x%llx)\n",
2454 	       be64_to_cpu(*(__be64 *)&req->initiator_port_id[0]),
2455 	       be64_to_cpu(*(__be64 *)&req->initiator_port_id[8]),
2456 	       be64_to_cpu(*(__be64 *)&req->target_port_id[0]),
2457 	       be64_to_cpu(*(__be64 *)&req->target_port_id[8]),
2458 	       it_iu_len,
2459 	       param->port,
2460 	       be64_to_cpu(*(__be64 *)&sdev->port[param->port - 1].gid.raw[0]),
2461 	       be64_to_cpu(*(__be64 *)&sdev->port[param->port - 1].gid.raw[8]));
2462 
2463 	rsp = kzalloc(sizeof *rsp, GFP_KERNEL);
2464 	rej = kzalloc(sizeof *rej, GFP_KERNEL);
2465 	rep_param = kzalloc(sizeof *rep_param, GFP_KERNEL);
2466 
2467 	if (!rsp || !rej || !rep_param) {
2468 		ret = -ENOMEM;
2469 		goto out;
2470 	}
2471 
2472 	if (it_iu_len > srp_max_req_size || it_iu_len < 64) {
2473 		rej->reason = __constant_cpu_to_be32(
2474 				SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE);
2475 		ret = -EINVAL;
2476 		printk(KERN_ERR "rejected SRP_LOGIN_REQ because its"
2477 		       " length (%d bytes) is out of range (%d .. %d)\n",
2478 		       it_iu_len, 64, srp_max_req_size);
2479 		goto reject;
2480 	}
2481 
2482 	if (!sport->enabled) {
2483 		rej->reason = __constant_cpu_to_be32(
2484 			     SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2485 		ret = -EINVAL;
2486 		printk(KERN_ERR "rejected SRP_LOGIN_REQ because the target port"
2487 		       " has not yet been enabled\n");
2488 		goto reject;
2489 	}
2490 
2491 	if ((req->req_flags & SRP_MTCH_ACTION) == SRP_MULTICHAN_SINGLE) {
2492 		rsp->rsp_flags = SRP_LOGIN_RSP_MULTICHAN_NO_CHAN;
2493 
2494 		spin_lock_irq(&sdev->spinlock);
2495 
2496 		list_for_each_entry_safe(ch, tmp_ch, &sdev->rch_list, list) {
2497 			if (!memcmp(ch->i_port_id, req->initiator_port_id, 16)
2498 			    && !memcmp(ch->t_port_id, req->target_port_id, 16)
2499 			    && param->port == ch->sport->port
2500 			    && param->listen_id == ch->sport->sdev->cm_id
2501 			    && ch->cm_id) {
2502 				enum rdma_ch_state ch_state;
2503 
2504 				ch_state = srpt_get_ch_state(ch);
2505 				if (ch_state != CH_CONNECTING
2506 				    && ch_state != CH_LIVE)
2507 					continue;
2508 
2509 				/* found an existing channel */
2510 				pr_debug("Found existing channel %s"
2511 					 " cm_id= %p state= %d\n",
2512 					 ch->sess_name, ch->cm_id, ch_state);
2513 
2514 				__srpt_close_ch(ch);
2515 
2516 				rsp->rsp_flags =
2517 					SRP_LOGIN_RSP_MULTICHAN_TERMINATED;
2518 			}
2519 		}
2520 
2521 		spin_unlock_irq(&sdev->spinlock);
2522 
2523 	} else
2524 		rsp->rsp_flags = SRP_LOGIN_RSP_MULTICHAN_MAINTAINED;
2525 
2526 	if (*(__be64 *)req->target_port_id != cpu_to_be64(srpt_service_guid)
2527 	    || *(__be64 *)(req->target_port_id + 8) !=
2528 	       cpu_to_be64(srpt_service_guid)) {
2529 		rej->reason = __constant_cpu_to_be32(
2530 				SRP_LOGIN_REJ_UNABLE_ASSOCIATE_CHANNEL);
2531 		ret = -ENOMEM;
2532 		printk(KERN_ERR "rejected SRP_LOGIN_REQ because it"
2533 		       " has an invalid target port identifier.\n");
2534 		goto reject;
2535 	}
2536 
2537 	ch = kzalloc(sizeof *ch, GFP_KERNEL);
2538 	if (!ch) {
2539 		rej->reason = __constant_cpu_to_be32(
2540 					SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2541 		printk(KERN_ERR "rejected SRP_LOGIN_REQ because no memory.\n");
2542 		ret = -ENOMEM;
2543 		goto reject;
2544 	}
2545 
2546 	INIT_WORK(&ch->release_work, srpt_release_channel_work);
2547 	memcpy(ch->i_port_id, req->initiator_port_id, 16);
2548 	memcpy(ch->t_port_id, req->target_port_id, 16);
2549 	ch->sport = &sdev->port[param->port - 1];
2550 	ch->cm_id = cm_id;
2551 	/*
2552 	 * Avoid QUEUE_FULL conditions by limiting the number of buffers used
2553 	 * for the SRP protocol to the command queue size.
2554 	 */
2555 	ch->rq_size = SRPT_RQ_SIZE;
2556 	spin_lock_init(&ch->spinlock);
2557 	ch->state = CH_CONNECTING;
2558 	INIT_LIST_HEAD(&ch->cmd_wait_list);
2559 	ch->rsp_size = ch->sport->port_attrib.srp_max_rsp_size;
2560 
2561 	ch->ioctx_ring = (struct srpt_send_ioctx **)
2562 		srpt_alloc_ioctx_ring(ch->sport->sdev, ch->rq_size,
2563 				      sizeof(*ch->ioctx_ring[0]),
2564 				      ch->rsp_size, DMA_TO_DEVICE);
2565 	if (!ch->ioctx_ring)
2566 		goto free_ch;
2567 
2568 	INIT_LIST_HEAD(&ch->free_list);
2569 	for (i = 0; i < ch->rq_size; i++) {
2570 		ch->ioctx_ring[i]->ch = ch;
2571 		list_add_tail(&ch->ioctx_ring[i]->free_list, &ch->free_list);
2572 	}
2573 
2574 	ret = srpt_create_ch_ib(ch);
2575 	if (ret) {
2576 		rej->reason = __constant_cpu_to_be32(
2577 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2578 		printk(KERN_ERR "rejected SRP_LOGIN_REQ because creating"
2579 		       " a new RDMA channel failed.\n");
2580 		goto free_ring;
2581 	}
2582 
2583 	ret = srpt_ch_qp_rtr(ch, ch->qp);
2584 	if (ret) {
2585 		rej->reason = __constant_cpu_to_be32(
2586 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2587 		printk(KERN_ERR "rejected SRP_LOGIN_REQ because enabling"
2588 		       " RTR failed (error code = %d)\n", ret);
2589 		goto destroy_ib;
2590 	}
2591 	/*
2592 	 * Use the initator port identifier as the session name.
2593 	 */
2594 	snprintf(ch->sess_name, sizeof(ch->sess_name), "0x%016llx%016llx",
2595 			be64_to_cpu(*(__be64 *)ch->i_port_id),
2596 			be64_to_cpu(*(__be64 *)(ch->i_port_id + 8)));
2597 
2598 	pr_debug("registering session %s\n", ch->sess_name);
2599 
2600 	nacl = srpt_lookup_acl(sport, ch->i_port_id);
2601 	if (!nacl) {
2602 		printk(KERN_INFO "Rejected login because no ACL has been"
2603 		       " configured yet for initiator %s.\n", ch->sess_name);
2604 		rej->reason = __constant_cpu_to_be32(
2605 				SRP_LOGIN_REJ_CHANNEL_LIMIT_REACHED);
2606 		goto destroy_ib;
2607 	}
2608 
2609 	ch->sess = transport_init_session();
2610 	if (IS_ERR(ch->sess)) {
2611 		rej->reason = __constant_cpu_to_be32(
2612 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2613 		pr_debug("Failed to create session\n");
2614 		goto deregister_session;
2615 	}
2616 	ch->sess->se_node_acl = &nacl->nacl;
2617 	transport_register_session(&sport->port_tpg_1, &nacl->nacl, ch->sess, ch);
2618 
2619 	pr_debug("Establish connection sess=%p name=%s cm_id=%p\n", ch->sess,
2620 		 ch->sess_name, ch->cm_id);
2621 
2622 	/* create srp_login_response */
2623 	rsp->opcode = SRP_LOGIN_RSP;
2624 	rsp->tag = req->tag;
2625 	rsp->max_it_iu_len = req->req_it_iu_len;
2626 	rsp->max_ti_iu_len = req->req_it_iu_len;
2627 	ch->max_ti_iu_len = it_iu_len;
2628 	rsp->buf_fmt = __constant_cpu_to_be16(SRP_BUF_FORMAT_DIRECT
2629 					      | SRP_BUF_FORMAT_INDIRECT);
2630 	rsp->req_lim_delta = cpu_to_be32(ch->rq_size);
2631 	atomic_set(&ch->req_lim, ch->rq_size);
2632 	atomic_set(&ch->req_lim_delta, 0);
2633 
2634 	/* create cm reply */
2635 	rep_param->qp_num = ch->qp->qp_num;
2636 	rep_param->private_data = (void *)rsp;
2637 	rep_param->private_data_len = sizeof *rsp;
2638 	rep_param->rnr_retry_count = 7;
2639 	rep_param->flow_control = 1;
2640 	rep_param->failover_accepted = 0;
2641 	rep_param->srq = 1;
2642 	rep_param->responder_resources = 4;
2643 	rep_param->initiator_depth = 4;
2644 
2645 	ret = ib_send_cm_rep(cm_id, rep_param);
2646 	if (ret) {
2647 		printk(KERN_ERR "sending SRP_LOGIN_REQ response failed"
2648 		       " (error code = %d)\n", ret);
2649 		goto release_channel;
2650 	}
2651 
2652 	spin_lock_irq(&sdev->spinlock);
2653 	list_add_tail(&ch->list, &sdev->rch_list);
2654 	spin_unlock_irq(&sdev->spinlock);
2655 
2656 	goto out;
2657 
2658 release_channel:
2659 	srpt_set_ch_state(ch, CH_RELEASING);
2660 	transport_deregister_session_configfs(ch->sess);
2661 
2662 deregister_session:
2663 	transport_deregister_session(ch->sess);
2664 	ch->sess = NULL;
2665 
2666 destroy_ib:
2667 	srpt_destroy_ch_ib(ch);
2668 
2669 free_ring:
2670 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
2671 			     ch->sport->sdev, ch->rq_size,
2672 			     ch->rsp_size, DMA_TO_DEVICE);
2673 free_ch:
2674 	kfree(ch);
2675 
2676 reject:
2677 	rej->opcode = SRP_LOGIN_REJ;
2678 	rej->tag = req->tag;
2679 	rej->buf_fmt = __constant_cpu_to_be16(SRP_BUF_FORMAT_DIRECT
2680 					      | SRP_BUF_FORMAT_INDIRECT);
2681 
2682 	ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED, NULL, 0,
2683 			     (void *)rej, sizeof *rej);
2684 
2685 out:
2686 	kfree(rep_param);
2687 	kfree(rsp);
2688 	kfree(rej);
2689 
2690 	return ret;
2691 }
2692 
2693 static void srpt_cm_rej_recv(struct ib_cm_id *cm_id)
2694 {
2695 	printk(KERN_INFO "Received IB REJ for cm_id %p.\n", cm_id);
2696 	srpt_drain_channel(cm_id);
2697 }
2698 
2699 /**
2700  * srpt_cm_rtu_recv() - Process an IB_CM_RTU_RECEIVED or USER_ESTABLISHED event.
2701  *
2702  * An IB_CM_RTU_RECEIVED message indicates that the connection is established
2703  * and that the recipient may begin transmitting (RTU = ready to use).
2704  */
2705 static void srpt_cm_rtu_recv(struct ib_cm_id *cm_id)
2706 {
2707 	struct srpt_rdma_ch *ch;
2708 	int ret;
2709 
2710 	ch = srpt_find_channel(cm_id->context, cm_id);
2711 	BUG_ON(!ch);
2712 
2713 	if (srpt_test_and_set_ch_state(ch, CH_CONNECTING, CH_LIVE)) {
2714 		struct srpt_recv_ioctx *ioctx, *ioctx_tmp;
2715 
2716 		ret = srpt_ch_qp_rts(ch, ch->qp);
2717 
2718 		list_for_each_entry_safe(ioctx, ioctx_tmp, &ch->cmd_wait_list,
2719 					 wait_list) {
2720 			list_del(&ioctx->wait_list);
2721 			srpt_handle_new_iu(ch, ioctx, NULL);
2722 		}
2723 		if (ret)
2724 			srpt_close_ch(ch);
2725 	}
2726 }
2727 
2728 static void srpt_cm_timewait_exit(struct ib_cm_id *cm_id)
2729 {
2730 	printk(KERN_INFO "Received IB TimeWait exit for cm_id %p.\n", cm_id);
2731 	srpt_drain_channel(cm_id);
2732 }
2733 
2734 static void srpt_cm_rep_error(struct ib_cm_id *cm_id)
2735 {
2736 	printk(KERN_INFO "Received IB REP error for cm_id %p.\n", cm_id);
2737 	srpt_drain_channel(cm_id);
2738 }
2739 
2740 /**
2741  * srpt_cm_dreq_recv() - Process reception of a DREQ message.
2742  */
2743 static void srpt_cm_dreq_recv(struct ib_cm_id *cm_id)
2744 {
2745 	struct srpt_rdma_ch *ch;
2746 	unsigned long flags;
2747 	bool send_drep = false;
2748 
2749 	ch = srpt_find_channel(cm_id->context, cm_id);
2750 	BUG_ON(!ch);
2751 
2752 	pr_debug("cm_id= %p ch->state= %d\n", cm_id, srpt_get_ch_state(ch));
2753 
2754 	spin_lock_irqsave(&ch->spinlock, flags);
2755 	switch (ch->state) {
2756 	case CH_CONNECTING:
2757 	case CH_LIVE:
2758 		send_drep = true;
2759 		ch->state = CH_DISCONNECTING;
2760 		break;
2761 	case CH_DISCONNECTING:
2762 	case CH_DRAINING:
2763 	case CH_RELEASING:
2764 		WARN(true, "unexpected channel state %d\n", ch->state);
2765 		break;
2766 	}
2767 	spin_unlock_irqrestore(&ch->spinlock, flags);
2768 
2769 	if (send_drep) {
2770 		if (ib_send_cm_drep(ch->cm_id, NULL, 0) < 0)
2771 			printk(KERN_ERR "Sending IB DREP failed.\n");
2772 		printk(KERN_INFO "Received DREQ and sent DREP for session %s.\n",
2773 		       ch->sess_name);
2774 	}
2775 }
2776 
2777 /**
2778  * srpt_cm_drep_recv() - Process reception of a DREP message.
2779  */
2780 static void srpt_cm_drep_recv(struct ib_cm_id *cm_id)
2781 {
2782 	printk(KERN_INFO "Received InfiniBand DREP message for cm_id %p.\n",
2783 	       cm_id);
2784 	srpt_drain_channel(cm_id);
2785 }
2786 
2787 /**
2788  * srpt_cm_handler() - IB connection manager callback function.
2789  *
2790  * A non-zero return value will cause the caller destroy the CM ID.
2791  *
2792  * Note: srpt_cm_handler() must only return a non-zero value when transferring
2793  * ownership of the cm_id to a channel by srpt_cm_req_recv() failed. Returning
2794  * a non-zero value in any other case will trigger a race with the
2795  * ib_destroy_cm_id() call in srpt_release_channel().
2796  */
2797 static int srpt_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
2798 {
2799 	int ret;
2800 
2801 	ret = 0;
2802 	switch (event->event) {
2803 	case IB_CM_REQ_RECEIVED:
2804 		ret = srpt_cm_req_recv(cm_id, &event->param.req_rcvd,
2805 				       event->private_data);
2806 		break;
2807 	case IB_CM_REJ_RECEIVED:
2808 		srpt_cm_rej_recv(cm_id);
2809 		break;
2810 	case IB_CM_RTU_RECEIVED:
2811 	case IB_CM_USER_ESTABLISHED:
2812 		srpt_cm_rtu_recv(cm_id);
2813 		break;
2814 	case IB_CM_DREQ_RECEIVED:
2815 		srpt_cm_dreq_recv(cm_id);
2816 		break;
2817 	case IB_CM_DREP_RECEIVED:
2818 		srpt_cm_drep_recv(cm_id);
2819 		break;
2820 	case IB_CM_TIMEWAIT_EXIT:
2821 		srpt_cm_timewait_exit(cm_id);
2822 		break;
2823 	case IB_CM_REP_ERROR:
2824 		srpt_cm_rep_error(cm_id);
2825 		break;
2826 	case IB_CM_DREQ_ERROR:
2827 		printk(KERN_INFO "Received IB DREQ ERROR event.\n");
2828 		break;
2829 	case IB_CM_MRA_RECEIVED:
2830 		printk(KERN_INFO "Received IB MRA event\n");
2831 		break;
2832 	default:
2833 		printk(KERN_ERR "received unrecognized IB CM event %d\n",
2834 		       event->event);
2835 		break;
2836 	}
2837 
2838 	return ret;
2839 }
2840 
2841 /**
2842  * srpt_perform_rdmas() - Perform IB RDMA.
2843  *
2844  * Returns zero upon success or a negative number upon failure.
2845  */
2846 static int srpt_perform_rdmas(struct srpt_rdma_ch *ch,
2847 			      struct srpt_send_ioctx *ioctx)
2848 {
2849 	struct ib_send_wr wr;
2850 	struct ib_send_wr *bad_wr;
2851 	struct rdma_iu *riu;
2852 	int i;
2853 	int ret;
2854 	int sq_wr_avail;
2855 	enum dma_data_direction dir;
2856 	const int n_rdma = ioctx->n_rdma;
2857 
2858 	dir = ioctx->cmd.data_direction;
2859 	if (dir == DMA_TO_DEVICE) {
2860 		/* write */
2861 		ret = -ENOMEM;
2862 		sq_wr_avail = atomic_sub_return(n_rdma, &ch->sq_wr_avail);
2863 		if (sq_wr_avail < 0) {
2864 			printk(KERN_WARNING "IB send queue full (needed %d)\n",
2865 			       n_rdma);
2866 			goto out;
2867 		}
2868 	}
2869 
2870 	ioctx->rdma_aborted = false;
2871 	ret = 0;
2872 	riu = ioctx->rdma_ius;
2873 	memset(&wr, 0, sizeof wr);
2874 
2875 	for (i = 0; i < n_rdma; ++i, ++riu) {
2876 		if (dir == DMA_FROM_DEVICE) {
2877 			wr.opcode = IB_WR_RDMA_WRITE;
2878 			wr.wr_id = encode_wr_id(i == n_rdma - 1 ?
2879 						SRPT_RDMA_WRITE_LAST :
2880 						SRPT_RDMA_MID,
2881 						ioctx->ioctx.index);
2882 		} else {
2883 			wr.opcode = IB_WR_RDMA_READ;
2884 			wr.wr_id = encode_wr_id(i == n_rdma - 1 ?
2885 						SRPT_RDMA_READ_LAST :
2886 						SRPT_RDMA_MID,
2887 						ioctx->ioctx.index);
2888 		}
2889 		wr.next = NULL;
2890 		wr.wr.rdma.remote_addr = riu->raddr;
2891 		wr.wr.rdma.rkey = riu->rkey;
2892 		wr.num_sge = riu->sge_cnt;
2893 		wr.sg_list = riu->sge;
2894 
2895 		/* only get completion event for the last rdma write */
2896 		if (i == (n_rdma - 1) && dir == DMA_TO_DEVICE)
2897 			wr.send_flags = IB_SEND_SIGNALED;
2898 
2899 		ret = ib_post_send(ch->qp, &wr, &bad_wr);
2900 		if (ret)
2901 			break;
2902 	}
2903 
2904 	if (ret)
2905 		printk(KERN_ERR "%s[%d]: ib_post_send() returned %d for %d/%d",
2906 				 __func__, __LINE__, ret, i, n_rdma);
2907 	if (ret && i > 0) {
2908 		wr.num_sge = 0;
2909 		wr.wr_id = encode_wr_id(SRPT_RDMA_ABORT, ioctx->ioctx.index);
2910 		wr.send_flags = IB_SEND_SIGNALED;
2911 		while (ch->state == CH_LIVE &&
2912 			ib_post_send(ch->qp, &wr, &bad_wr) != 0) {
2913 			printk(KERN_INFO "Trying to abort failed RDMA transfer [%d]",
2914 				ioctx->ioctx.index);
2915 			msleep(1000);
2916 		}
2917 		while (ch->state != CH_RELEASING && !ioctx->rdma_aborted) {
2918 			printk(KERN_INFO "Waiting until RDMA abort finished [%d]",
2919 				ioctx->ioctx.index);
2920 			msleep(1000);
2921 		}
2922 	}
2923 out:
2924 	if (unlikely(dir == DMA_TO_DEVICE && ret < 0))
2925 		atomic_add(n_rdma, &ch->sq_wr_avail);
2926 	return ret;
2927 }
2928 
2929 /**
2930  * srpt_xfer_data() - Start data transfer from initiator to target.
2931  */
2932 static int srpt_xfer_data(struct srpt_rdma_ch *ch,
2933 			  struct srpt_send_ioctx *ioctx)
2934 {
2935 	int ret;
2936 
2937 	ret = srpt_map_sg_to_ib_sge(ch, ioctx);
2938 	if (ret) {
2939 		printk(KERN_ERR "%s[%d] ret=%d\n", __func__, __LINE__, ret);
2940 		goto out;
2941 	}
2942 
2943 	ret = srpt_perform_rdmas(ch, ioctx);
2944 	if (ret) {
2945 		if (ret == -EAGAIN || ret == -ENOMEM)
2946 			printk(KERN_INFO "%s[%d] queue full -- ret=%d\n",
2947 				   __func__, __LINE__, ret);
2948 		else
2949 			printk(KERN_ERR "%s[%d] fatal error -- ret=%d\n",
2950 			       __func__, __LINE__, ret);
2951 		goto out_unmap;
2952 	}
2953 
2954 out:
2955 	return ret;
2956 out_unmap:
2957 	srpt_unmap_sg_to_ib_sge(ch, ioctx);
2958 	goto out;
2959 }
2960 
2961 static int srpt_write_pending_status(struct se_cmd *se_cmd)
2962 {
2963 	struct srpt_send_ioctx *ioctx;
2964 
2965 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
2966 	return srpt_get_cmd_state(ioctx) == SRPT_STATE_NEED_DATA;
2967 }
2968 
2969 /*
2970  * srpt_write_pending() - Start data transfer from initiator to target (write).
2971  */
2972 static int srpt_write_pending(struct se_cmd *se_cmd)
2973 {
2974 	struct srpt_rdma_ch *ch;
2975 	struct srpt_send_ioctx *ioctx;
2976 	enum srpt_command_state new_state;
2977 	enum rdma_ch_state ch_state;
2978 	int ret;
2979 
2980 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
2981 
2982 	new_state = srpt_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA);
2983 	WARN_ON(new_state == SRPT_STATE_DONE);
2984 
2985 	ch = ioctx->ch;
2986 	BUG_ON(!ch);
2987 
2988 	ch_state = srpt_get_ch_state(ch);
2989 	switch (ch_state) {
2990 	case CH_CONNECTING:
2991 		WARN(true, "unexpected channel state %d\n", ch_state);
2992 		ret = -EINVAL;
2993 		goto out;
2994 	case CH_LIVE:
2995 		break;
2996 	case CH_DISCONNECTING:
2997 	case CH_DRAINING:
2998 	case CH_RELEASING:
2999 		pr_debug("cmd with tag %lld: channel disconnecting\n",
3000 			 ioctx->tag);
3001 		srpt_set_cmd_state(ioctx, SRPT_STATE_DATA_IN);
3002 		ret = -EINVAL;
3003 		goto out;
3004 	}
3005 	ret = srpt_xfer_data(ch, ioctx);
3006 
3007 out:
3008 	return ret;
3009 }
3010 
3011 static u8 tcm_to_srp_tsk_mgmt_status(const int tcm_mgmt_status)
3012 {
3013 	switch (tcm_mgmt_status) {
3014 	case TMR_FUNCTION_COMPLETE:
3015 		return SRP_TSK_MGMT_SUCCESS;
3016 	case TMR_FUNCTION_REJECTED:
3017 		return SRP_TSK_MGMT_FUNC_NOT_SUPP;
3018 	}
3019 	return SRP_TSK_MGMT_FAILED;
3020 }
3021 
3022 /**
3023  * srpt_queue_response() - Transmits the response to a SCSI command.
3024  *
3025  * Callback function called by the TCM core. Must not block since it can be
3026  * invoked on the context of the IB completion handler.
3027  */
3028 static int srpt_queue_response(struct se_cmd *cmd)
3029 {
3030 	struct srpt_rdma_ch *ch;
3031 	struct srpt_send_ioctx *ioctx;
3032 	enum srpt_command_state state;
3033 	unsigned long flags;
3034 	int ret;
3035 	enum dma_data_direction dir;
3036 	int resp_len;
3037 	u8 srp_tm_status;
3038 
3039 	ret = 0;
3040 
3041 	ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
3042 	ch = ioctx->ch;
3043 	BUG_ON(!ch);
3044 
3045 	spin_lock_irqsave(&ioctx->spinlock, flags);
3046 	state = ioctx->state;
3047 	switch (state) {
3048 	case SRPT_STATE_NEW:
3049 	case SRPT_STATE_DATA_IN:
3050 		ioctx->state = SRPT_STATE_CMD_RSP_SENT;
3051 		break;
3052 	case SRPT_STATE_MGMT:
3053 		ioctx->state = SRPT_STATE_MGMT_RSP_SENT;
3054 		break;
3055 	default:
3056 		WARN(true, "ch %p; cmd %d: unexpected command state %d\n",
3057 			ch, ioctx->ioctx.index, ioctx->state);
3058 		break;
3059 	}
3060 	spin_unlock_irqrestore(&ioctx->spinlock, flags);
3061 
3062 	if (unlikely(transport_check_aborted_status(&ioctx->cmd, false)
3063 		     || WARN_ON_ONCE(state == SRPT_STATE_CMD_RSP_SENT))) {
3064 		atomic_inc(&ch->req_lim_delta);
3065 		srpt_abort_cmd(ioctx);
3066 		goto out;
3067 	}
3068 
3069 	dir = ioctx->cmd.data_direction;
3070 
3071 	/* For read commands, transfer the data to the initiator. */
3072 	if (dir == DMA_FROM_DEVICE && ioctx->cmd.data_length &&
3073 	    !ioctx->queue_status_only) {
3074 		ret = srpt_xfer_data(ch, ioctx);
3075 		if (ret) {
3076 			printk(KERN_ERR "xfer_data failed for tag %llu\n",
3077 			       ioctx->tag);
3078 			goto out;
3079 		}
3080 	}
3081 
3082 	if (state != SRPT_STATE_MGMT)
3083 		resp_len = srpt_build_cmd_rsp(ch, ioctx, ioctx->tag,
3084 					      cmd->scsi_status);
3085 	else {
3086 		srp_tm_status
3087 			= tcm_to_srp_tsk_mgmt_status(cmd->se_tmr_req->response);
3088 		resp_len = srpt_build_tskmgmt_rsp(ch, ioctx, srp_tm_status,
3089 						 ioctx->tag);
3090 	}
3091 	ret = srpt_post_send(ch, ioctx, resp_len);
3092 	if (ret) {
3093 		printk(KERN_ERR "sending cmd response failed for tag %llu\n",
3094 		       ioctx->tag);
3095 		srpt_unmap_sg_to_ib_sge(ch, ioctx);
3096 		srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
3097 		kref_put(&ioctx->kref, srpt_put_send_ioctx_kref);
3098 	}
3099 
3100 out:
3101 	return ret;
3102 }
3103 
3104 static int srpt_queue_status(struct se_cmd *cmd)
3105 {
3106 	struct srpt_send_ioctx *ioctx;
3107 
3108 	ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
3109 	BUG_ON(ioctx->sense_data != cmd->sense_buffer);
3110 	if (cmd->se_cmd_flags &
3111 	    (SCF_TRANSPORT_TASK_SENSE | SCF_EMULATED_TASK_SENSE))
3112 		WARN_ON(cmd->scsi_status != SAM_STAT_CHECK_CONDITION);
3113 	ioctx->queue_status_only = true;
3114 	return srpt_queue_response(cmd);
3115 }
3116 
3117 static void srpt_refresh_port_work(struct work_struct *work)
3118 {
3119 	struct srpt_port *sport = container_of(work, struct srpt_port, work);
3120 
3121 	srpt_refresh_port(sport);
3122 }
3123 
3124 static int srpt_ch_list_empty(struct srpt_device *sdev)
3125 {
3126 	int res;
3127 
3128 	spin_lock_irq(&sdev->spinlock);
3129 	res = list_empty(&sdev->rch_list);
3130 	spin_unlock_irq(&sdev->spinlock);
3131 
3132 	return res;
3133 }
3134 
3135 /**
3136  * srpt_release_sdev() - Free the channel resources associated with a target.
3137  */
3138 static int srpt_release_sdev(struct srpt_device *sdev)
3139 {
3140 	struct srpt_rdma_ch *ch, *tmp_ch;
3141 	int res;
3142 
3143 	WARN_ON_ONCE(irqs_disabled());
3144 
3145 	BUG_ON(!sdev);
3146 
3147 	spin_lock_irq(&sdev->spinlock);
3148 	list_for_each_entry_safe(ch, tmp_ch, &sdev->rch_list, list)
3149 		__srpt_close_ch(ch);
3150 	spin_unlock_irq(&sdev->spinlock);
3151 
3152 	res = wait_event_interruptible(sdev->ch_releaseQ,
3153 				       srpt_ch_list_empty(sdev));
3154 	if (res)
3155 		printk(KERN_ERR "%s: interrupted.\n", __func__);
3156 
3157 	return 0;
3158 }
3159 
3160 static struct srpt_port *__srpt_lookup_port(const char *name)
3161 {
3162 	struct ib_device *dev;
3163 	struct srpt_device *sdev;
3164 	struct srpt_port *sport;
3165 	int i;
3166 
3167 	list_for_each_entry(sdev, &srpt_dev_list, list) {
3168 		dev = sdev->device;
3169 		if (!dev)
3170 			continue;
3171 
3172 		for (i = 0; i < dev->phys_port_cnt; i++) {
3173 			sport = &sdev->port[i];
3174 
3175 			if (!strcmp(sport->port_guid, name))
3176 				return sport;
3177 		}
3178 	}
3179 
3180 	return NULL;
3181 }
3182 
3183 static struct srpt_port *srpt_lookup_port(const char *name)
3184 {
3185 	struct srpt_port *sport;
3186 
3187 	spin_lock(&srpt_dev_lock);
3188 	sport = __srpt_lookup_port(name);
3189 	spin_unlock(&srpt_dev_lock);
3190 
3191 	return sport;
3192 }
3193 
3194 /**
3195  * srpt_add_one() - Infiniband device addition callback function.
3196  */
3197 static void srpt_add_one(struct ib_device *device)
3198 {
3199 	struct srpt_device *sdev;
3200 	struct srpt_port *sport;
3201 	struct ib_srq_init_attr srq_attr;
3202 	int i;
3203 
3204 	pr_debug("device = %p, device->dma_ops = %p\n", device,
3205 		 device->dma_ops);
3206 
3207 	sdev = kzalloc(sizeof *sdev, GFP_KERNEL);
3208 	if (!sdev)
3209 		goto err;
3210 
3211 	sdev->device = device;
3212 	INIT_LIST_HEAD(&sdev->rch_list);
3213 	init_waitqueue_head(&sdev->ch_releaseQ);
3214 	spin_lock_init(&sdev->spinlock);
3215 
3216 	if (ib_query_device(device, &sdev->dev_attr))
3217 		goto free_dev;
3218 
3219 	sdev->pd = ib_alloc_pd(device);
3220 	if (IS_ERR(sdev->pd))
3221 		goto free_dev;
3222 
3223 	sdev->mr = ib_get_dma_mr(sdev->pd, IB_ACCESS_LOCAL_WRITE);
3224 	if (IS_ERR(sdev->mr))
3225 		goto err_pd;
3226 
3227 	sdev->srq_size = min(srpt_srq_size, sdev->dev_attr.max_srq_wr);
3228 
3229 	srq_attr.event_handler = srpt_srq_event;
3230 	srq_attr.srq_context = (void *)sdev;
3231 	srq_attr.attr.max_wr = sdev->srq_size;
3232 	srq_attr.attr.max_sge = 1;
3233 	srq_attr.attr.srq_limit = 0;
3234 	srq_attr.srq_type = IB_SRQT_BASIC;
3235 
3236 	sdev->srq = ib_create_srq(sdev->pd, &srq_attr);
3237 	if (IS_ERR(sdev->srq))
3238 		goto err_mr;
3239 
3240 	pr_debug("%s: create SRQ #wr= %d max_allow=%d dev= %s\n",
3241 		 __func__, sdev->srq_size, sdev->dev_attr.max_srq_wr,
3242 		 device->name);
3243 
3244 	if (!srpt_service_guid)
3245 		srpt_service_guid = be64_to_cpu(device->node_guid);
3246 
3247 	sdev->cm_id = ib_create_cm_id(device, srpt_cm_handler, sdev);
3248 	if (IS_ERR(sdev->cm_id))
3249 		goto err_srq;
3250 
3251 	/* print out target login information */
3252 	pr_debug("Target login info: id_ext=%016llx,ioc_guid=%016llx,"
3253 		 "pkey=ffff,service_id=%016llx\n", srpt_service_guid,
3254 		 srpt_service_guid, srpt_service_guid);
3255 
3256 	/*
3257 	 * We do not have a consistent service_id (ie. also id_ext of target_id)
3258 	 * to identify this target. We currently use the guid of the first HCA
3259 	 * in the system as service_id; therefore, the target_id will change
3260 	 * if this HCA is gone bad and replaced by different HCA
3261 	 */
3262 	if (ib_cm_listen(sdev->cm_id, cpu_to_be64(srpt_service_guid), 0, NULL))
3263 		goto err_cm;
3264 
3265 	INIT_IB_EVENT_HANDLER(&sdev->event_handler, sdev->device,
3266 			      srpt_event_handler);
3267 	if (ib_register_event_handler(&sdev->event_handler))
3268 		goto err_cm;
3269 
3270 	sdev->ioctx_ring = (struct srpt_recv_ioctx **)
3271 		srpt_alloc_ioctx_ring(sdev, sdev->srq_size,
3272 				      sizeof(*sdev->ioctx_ring[0]),
3273 				      srp_max_req_size, DMA_FROM_DEVICE);
3274 	if (!sdev->ioctx_ring)
3275 		goto err_event;
3276 
3277 	for (i = 0; i < sdev->srq_size; ++i)
3278 		srpt_post_recv(sdev, sdev->ioctx_ring[i]);
3279 
3280 	WARN_ON(sdev->device->phys_port_cnt > ARRAY_SIZE(sdev->port));
3281 
3282 	for (i = 1; i <= sdev->device->phys_port_cnt; i++) {
3283 		sport = &sdev->port[i - 1];
3284 		sport->sdev = sdev;
3285 		sport->port = i;
3286 		sport->port_attrib.srp_max_rdma_size = DEFAULT_MAX_RDMA_SIZE;
3287 		sport->port_attrib.srp_max_rsp_size = DEFAULT_MAX_RSP_SIZE;
3288 		sport->port_attrib.srp_sq_size = DEF_SRPT_SQ_SIZE;
3289 		INIT_WORK(&sport->work, srpt_refresh_port_work);
3290 		INIT_LIST_HEAD(&sport->port_acl_list);
3291 		spin_lock_init(&sport->port_acl_lock);
3292 
3293 		if (srpt_refresh_port(sport)) {
3294 			printk(KERN_ERR "MAD registration failed for %s-%d.\n",
3295 			       srpt_sdev_name(sdev), i);
3296 			goto err_ring;
3297 		}
3298 		snprintf(sport->port_guid, sizeof(sport->port_guid),
3299 			"0x%016llx%016llx",
3300 			be64_to_cpu(sport->gid.global.subnet_prefix),
3301 			be64_to_cpu(sport->gid.global.interface_id));
3302 	}
3303 
3304 	spin_lock(&srpt_dev_lock);
3305 	list_add_tail(&sdev->list, &srpt_dev_list);
3306 	spin_unlock(&srpt_dev_lock);
3307 
3308 out:
3309 	ib_set_client_data(device, &srpt_client, sdev);
3310 	pr_debug("added %s.\n", device->name);
3311 	return;
3312 
3313 err_ring:
3314 	srpt_free_ioctx_ring((struct srpt_ioctx **)sdev->ioctx_ring, sdev,
3315 			     sdev->srq_size, srp_max_req_size,
3316 			     DMA_FROM_DEVICE);
3317 err_event:
3318 	ib_unregister_event_handler(&sdev->event_handler);
3319 err_cm:
3320 	ib_destroy_cm_id(sdev->cm_id);
3321 err_srq:
3322 	ib_destroy_srq(sdev->srq);
3323 err_mr:
3324 	ib_dereg_mr(sdev->mr);
3325 err_pd:
3326 	ib_dealloc_pd(sdev->pd);
3327 free_dev:
3328 	kfree(sdev);
3329 err:
3330 	sdev = NULL;
3331 	printk(KERN_INFO "%s(%s) failed.\n", __func__, device->name);
3332 	goto out;
3333 }
3334 
3335 /**
3336  * srpt_remove_one() - InfiniBand device removal callback function.
3337  */
3338 static void srpt_remove_one(struct ib_device *device)
3339 {
3340 	struct srpt_device *sdev;
3341 	int i;
3342 
3343 	sdev = ib_get_client_data(device, &srpt_client);
3344 	if (!sdev) {
3345 		printk(KERN_INFO "%s(%s): nothing to do.\n", __func__,
3346 		       device->name);
3347 		return;
3348 	}
3349 
3350 	srpt_unregister_mad_agent(sdev);
3351 
3352 	ib_unregister_event_handler(&sdev->event_handler);
3353 
3354 	/* Cancel any work queued by the just unregistered IB event handler. */
3355 	for (i = 0; i < sdev->device->phys_port_cnt; i++)
3356 		cancel_work_sync(&sdev->port[i].work);
3357 
3358 	ib_destroy_cm_id(sdev->cm_id);
3359 
3360 	/*
3361 	 * Unregistering a target must happen after destroying sdev->cm_id
3362 	 * such that no new SRP_LOGIN_REQ information units can arrive while
3363 	 * destroying the target.
3364 	 */
3365 	spin_lock(&srpt_dev_lock);
3366 	list_del(&sdev->list);
3367 	spin_unlock(&srpt_dev_lock);
3368 	srpt_release_sdev(sdev);
3369 
3370 	ib_destroy_srq(sdev->srq);
3371 	ib_dereg_mr(sdev->mr);
3372 	ib_dealloc_pd(sdev->pd);
3373 
3374 	srpt_free_ioctx_ring((struct srpt_ioctx **)sdev->ioctx_ring, sdev,
3375 			     sdev->srq_size, srp_max_req_size, DMA_FROM_DEVICE);
3376 	sdev->ioctx_ring = NULL;
3377 	kfree(sdev);
3378 }
3379 
3380 static struct ib_client srpt_client = {
3381 	.name = DRV_NAME,
3382 	.add = srpt_add_one,
3383 	.remove = srpt_remove_one
3384 };
3385 
3386 static int srpt_check_true(struct se_portal_group *se_tpg)
3387 {
3388 	return 1;
3389 }
3390 
3391 static int srpt_check_false(struct se_portal_group *se_tpg)
3392 {
3393 	return 0;
3394 }
3395 
3396 static char *srpt_get_fabric_name(void)
3397 {
3398 	return "srpt";
3399 }
3400 
3401 static u8 srpt_get_fabric_proto_ident(struct se_portal_group *se_tpg)
3402 {
3403 	return SCSI_TRANSPORTID_PROTOCOLID_SRP;
3404 }
3405 
3406 static char *srpt_get_fabric_wwn(struct se_portal_group *tpg)
3407 {
3408 	struct srpt_port *sport = container_of(tpg, struct srpt_port, port_tpg_1);
3409 
3410 	return sport->port_guid;
3411 }
3412 
3413 static u16 srpt_get_tag(struct se_portal_group *tpg)
3414 {
3415 	return 1;
3416 }
3417 
3418 static u32 srpt_get_default_depth(struct se_portal_group *se_tpg)
3419 {
3420 	return 1;
3421 }
3422 
3423 static u32 srpt_get_pr_transport_id(struct se_portal_group *se_tpg,
3424 				    struct se_node_acl *se_nacl,
3425 				    struct t10_pr_registration *pr_reg,
3426 				    int *format_code, unsigned char *buf)
3427 {
3428 	struct srpt_node_acl *nacl;
3429 	struct spc_rdma_transport_id *tr_id;
3430 
3431 	nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3432 	tr_id = (void *)buf;
3433 	tr_id->protocol_identifier = SCSI_TRANSPORTID_PROTOCOLID_SRP;
3434 	memcpy(tr_id->i_port_id, nacl->i_port_id, sizeof(tr_id->i_port_id));
3435 	return sizeof(*tr_id);
3436 }
3437 
3438 static u32 srpt_get_pr_transport_id_len(struct se_portal_group *se_tpg,
3439 					struct se_node_acl *se_nacl,
3440 					struct t10_pr_registration *pr_reg,
3441 					int *format_code)
3442 {
3443 	*format_code = 0;
3444 	return sizeof(struct spc_rdma_transport_id);
3445 }
3446 
3447 static char *srpt_parse_pr_out_transport_id(struct se_portal_group *se_tpg,
3448 					    const char *buf, u32 *out_tid_len,
3449 					    char **port_nexus_ptr)
3450 {
3451 	struct spc_rdma_transport_id *tr_id;
3452 
3453 	*port_nexus_ptr = NULL;
3454 	*out_tid_len = sizeof(struct spc_rdma_transport_id);
3455 	tr_id = (void *)buf;
3456 	return (char *)tr_id->i_port_id;
3457 }
3458 
3459 static struct se_node_acl *srpt_alloc_fabric_acl(struct se_portal_group *se_tpg)
3460 {
3461 	struct srpt_node_acl *nacl;
3462 
3463 	nacl = kzalloc(sizeof(struct srpt_node_acl), GFP_KERNEL);
3464 	if (!nacl) {
3465 		printk(KERN_ERR "Unable to allocate struct srpt_node_acl\n");
3466 		return NULL;
3467 	}
3468 
3469 	return &nacl->nacl;
3470 }
3471 
3472 static void srpt_release_fabric_acl(struct se_portal_group *se_tpg,
3473 				    struct se_node_acl *se_nacl)
3474 {
3475 	struct srpt_node_acl *nacl;
3476 
3477 	nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3478 	kfree(nacl);
3479 }
3480 
3481 static u32 srpt_tpg_get_inst_index(struct se_portal_group *se_tpg)
3482 {
3483 	return 1;
3484 }
3485 
3486 static void srpt_release_cmd(struct se_cmd *se_cmd)
3487 {
3488 }
3489 
3490 /**
3491  * srpt_shutdown_session() - Whether or not a session may be shut down.
3492  */
3493 static int srpt_shutdown_session(struct se_session *se_sess)
3494 {
3495 	return true;
3496 }
3497 
3498 /**
3499  * srpt_close_session() - Forcibly close a session.
3500  *
3501  * Callback function invoked by the TCM core to clean up sessions associated
3502  * with a node ACL when the user invokes
3503  * rmdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3504  */
3505 static void srpt_close_session(struct se_session *se_sess)
3506 {
3507 	DECLARE_COMPLETION_ONSTACK(release_done);
3508 	struct srpt_rdma_ch *ch;
3509 	struct srpt_device *sdev;
3510 	int res;
3511 
3512 	ch = se_sess->fabric_sess_ptr;
3513 	WARN_ON(ch->sess != se_sess);
3514 
3515 	pr_debug("ch %p state %d\n", ch, srpt_get_ch_state(ch));
3516 
3517 	sdev = ch->sport->sdev;
3518 	spin_lock_irq(&sdev->spinlock);
3519 	BUG_ON(ch->release_done);
3520 	ch->release_done = &release_done;
3521 	__srpt_close_ch(ch);
3522 	spin_unlock_irq(&sdev->spinlock);
3523 
3524 	res = wait_for_completion_timeout(&release_done, 60 * HZ);
3525 	WARN_ON(res <= 0);
3526 }
3527 
3528 /**
3529  * srpt_sess_get_index() - Return the value of scsiAttIntrPortIndex (SCSI-MIB).
3530  *
3531  * A quote from RFC 4455 (SCSI-MIB) about this MIB object:
3532  * This object represents an arbitrary integer used to uniquely identify a
3533  * particular attached remote initiator port to a particular SCSI target port
3534  * within a particular SCSI target device within a particular SCSI instance.
3535  */
3536 static u32 srpt_sess_get_index(struct se_session *se_sess)
3537 {
3538 	return 0;
3539 }
3540 
3541 static void srpt_set_default_node_attrs(struct se_node_acl *nacl)
3542 {
3543 }
3544 
3545 static u32 srpt_get_task_tag(struct se_cmd *se_cmd)
3546 {
3547 	struct srpt_send_ioctx *ioctx;
3548 
3549 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
3550 	return ioctx->tag;
3551 }
3552 
3553 /* Note: only used from inside debug printk's by the TCM core. */
3554 static int srpt_get_tcm_cmd_state(struct se_cmd *se_cmd)
3555 {
3556 	struct srpt_send_ioctx *ioctx;
3557 
3558 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
3559 	return srpt_get_cmd_state(ioctx);
3560 }
3561 
3562 static u16 srpt_set_fabric_sense_len(struct se_cmd *cmd, u32 sense_length)
3563 {
3564 	return 0;
3565 }
3566 
3567 static u16 srpt_get_fabric_sense_len(void)
3568 {
3569 	return 0;
3570 }
3571 
3572 /**
3573  * srpt_parse_i_port_id() - Parse an initiator port ID.
3574  * @name: ASCII representation of a 128-bit initiator port ID.
3575  * @i_port_id: Binary 128-bit port ID.
3576  */
3577 static int srpt_parse_i_port_id(u8 i_port_id[16], const char *name)
3578 {
3579 	const char *p;
3580 	unsigned len, count, leading_zero_bytes;
3581 	int ret, rc;
3582 
3583 	p = name;
3584 	if (strnicmp(p, "0x", 2) == 0)
3585 		p += 2;
3586 	ret = -EINVAL;
3587 	len = strlen(p);
3588 	if (len % 2)
3589 		goto out;
3590 	count = min(len / 2, 16U);
3591 	leading_zero_bytes = 16 - count;
3592 	memset(i_port_id, 0, leading_zero_bytes);
3593 	rc = hex2bin(i_port_id + leading_zero_bytes, p, count);
3594 	if (rc < 0)
3595 		pr_debug("hex2bin failed for srpt_parse_i_port_id: %d\n", rc);
3596 	ret = 0;
3597 out:
3598 	return ret;
3599 }
3600 
3601 /*
3602  * configfs callback function invoked for
3603  * mkdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3604  */
3605 static struct se_node_acl *srpt_make_nodeacl(struct se_portal_group *tpg,
3606 					     struct config_group *group,
3607 					     const char *name)
3608 {
3609 	struct srpt_port *sport = container_of(tpg, struct srpt_port, port_tpg_1);
3610 	struct se_node_acl *se_nacl, *se_nacl_new;
3611 	struct srpt_node_acl *nacl;
3612 	int ret = 0;
3613 	u32 nexus_depth = 1;
3614 	u8 i_port_id[16];
3615 
3616 	if (srpt_parse_i_port_id(i_port_id, name) < 0) {
3617 		printk(KERN_ERR "invalid initiator port ID %s\n", name);
3618 		ret = -EINVAL;
3619 		goto err;
3620 	}
3621 
3622 	se_nacl_new = srpt_alloc_fabric_acl(tpg);
3623 	if (!se_nacl_new) {
3624 		ret = -ENOMEM;
3625 		goto err;
3626 	}
3627 	/*
3628 	 * nacl_new may be released by core_tpg_add_initiator_node_acl()
3629 	 * when converting a node ACL from demo mode to explict
3630 	 */
3631 	se_nacl = core_tpg_add_initiator_node_acl(tpg, se_nacl_new, name,
3632 						  nexus_depth);
3633 	if (IS_ERR(se_nacl)) {
3634 		ret = PTR_ERR(se_nacl);
3635 		goto err;
3636 	}
3637 	/* Locate our struct srpt_node_acl and set sdev and i_port_id. */
3638 	nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3639 	memcpy(&nacl->i_port_id[0], &i_port_id[0], 16);
3640 	nacl->sport = sport;
3641 
3642 	spin_lock_irq(&sport->port_acl_lock);
3643 	list_add_tail(&nacl->list, &sport->port_acl_list);
3644 	spin_unlock_irq(&sport->port_acl_lock);
3645 
3646 	return se_nacl;
3647 err:
3648 	return ERR_PTR(ret);
3649 }
3650 
3651 /*
3652  * configfs callback function invoked for
3653  * rmdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3654  */
3655 static void srpt_drop_nodeacl(struct se_node_acl *se_nacl)
3656 {
3657 	struct srpt_node_acl *nacl;
3658 	struct srpt_device *sdev;
3659 	struct srpt_port *sport;
3660 
3661 	nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3662 	sport = nacl->sport;
3663 	sdev = sport->sdev;
3664 	spin_lock_irq(&sport->port_acl_lock);
3665 	list_del(&nacl->list);
3666 	spin_unlock_irq(&sport->port_acl_lock);
3667 	core_tpg_del_initiator_node_acl(&sport->port_tpg_1, se_nacl, 1);
3668 	srpt_release_fabric_acl(NULL, se_nacl);
3669 }
3670 
3671 static ssize_t srpt_tpg_attrib_show_srp_max_rdma_size(
3672 	struct se_portal_group *se_tpg,
3673 	char *page)
3674 {
3675 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3676 
3677 	return sprintf(page, "%u\n", sport->port_attrib.srp_max_rdma_size);
3678 }
3679 
3680 static ssize_t srpt_tpg_attrib_store_srp_max_rdma_size(
3681 	struct se_portal_group *se_tpg,
3682 	const char *page,
3683 	size_t count)
3684 {
3685 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3686 	unsigned long val;
3687 	int ret;
3688 
3689 	ret = strict_strtoul(page, 0, &val);
3690 	if (ret < 0) {
3691 		pr_err("strict_strtoul() failed with ret: %d\n", ret);
3692 		return -EINVAL;
3693 	}
3694 	if (val > MAX_SRPT_RDMA_SIZE) {
3695 		pr_err("val: %lu exceeds MAX_SRPT_RDMA_SIZE: %d\n", val,
3696 			MAX_SRPT_RDMA_SIZE);
3697 		return -EINVAL;
3698 	}
3699 	if (val < DEFAULT_MAX_RDMA_SIZE) {
3700 		pr_err("val: %lu smaller than DEFAULT_MAX_RDMA_SIZE: %d\n",
3701 			val, DEFAULT_MAX_RDMA_SIZE);
3702 		return -EINVAL;
3703 	}
3704 	sport->port_attrib.srp_max_rdma_size = val;
3705 
3706 	return count;
3707 }
3708 
3709 TF_TPG_ATTRIB_ATTR(srpt, srp_max_rdma_size, S_IRUGO | S_IWUSR);
3710 
3711 static ssize_t srpt_tpg_attrib_show_srp_max_rsp_size(
3712 	struct se_portal_group *se_tpg,
3713 	char *page)
3714 {
3715 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3716 
3717 	return sprintf(page, "%u\n", sport->port_attrib.srp_max_rsp_size);
3718 }
3719 
3720 static ssize_t srpt_tpg_attrib_store_srp_max_rsp_size(
3721 	struct se_portal_group *se_tpg,
3722 	const char *page,
3723 	size_t count)
3724 {
3725 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3726 	unsigned long val;
3727 	int ret;
3728 
3729 	ret = strict_strtoul(page, 0, &val);
3730 	if (ret < 0) {
3731 		pr_err("strict_strtoul() failed with ret: %d\n", ret);
3732 		return -EINVAL;
3733 	}
3734 	if (val > MAX_SRPT_RSP_SIZE) {
3735 		pr_err("val: %lu exceeds MAX_SRPT_RSP_SIZE: %d\n", val,
3736 			MAX_SRPT_RSP_SIZE);
3737 		return -EINVAL;
3738 	}
3739 	if (val < MIN_MAX_RSP_SIZE) {
3740 		pr_err("val: %lu smaller than MIN_MAX_RSP_SIZE: %d\n", val,
3741 			MIN_MAX_RSP_SIZE);
3742 		return -EINVAL;
3743 	}
3744 	sport->port_attrib.srp_max_rsp_size = val;
3745 
3746 	return count;
3747 }
3748 
3749 TF_TPG_ATTRIB_ATTR(srpt, srp_max_rsp_size, S_IRUGO | S_IWUSR);
3750 
3751 static ssize_t srpt_tpg_attrib_show_srp_sq_size(
3752 	struct se_portal_group *se_tpg,
3753 	char *page)
3754 {
3755 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3756 
3757 	return sprintf(page, "%u\n", sport->port_attrib.srp_sq_size);
3758 }
3759 
3760 static ssize_t srpt_tpg_attrib_store_srp_sq_size(
3761 	struct se_portal_group *se_tpg,
3762 	const char *page,
3763 	size_t count)
3764 {
3765 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3766 	unsigned long val;
3767 	int ret;
3768 
3769 	ret = strict_strtoul(page, 0, &val);
3770 	if (ret < 0) {
3771 		pr_err("strict_strtoul() failed with ret: %d\n", ret);
3772 		return -EINVAL;
3773 	}
3774 	if (val > MAX_SRPT_SRQ_SIZE) {
3775 		pr_err("val: %lu exceeds MAX_SRPT_SRQ_SIZE: %d\n", val,
3776 			MAX_SRPT_SRQ_SIZE);
3777 		return -EINVAL;
3778 	}
3779 	if (val < MIN_SRPT_SRQ_SIZE) {
3780 		pr_err("val: %lu smaller than MIN_SRPT_SRQ_SIZE: %d\n", val,
3781 			MIN_SRPT_SRQ_SIZE);
3782 		return -EINVAL;
3783 	}
3784 	sport->port_attrib.srp_sq_size = val;
3785 
3786 	return count;
3787 }
3788 
3789 TF_TPG_ATTRIB_ATTR(srpt, srp_sq_size, S_IRUGO | S_IWUSR);
3790 
3791 static struct configfs_attribute *srpt_tpg_attrib_attrs[] = {
3792 	&srpt_tpg_attrib_srp_max_rdma_size.attr,
3793 	&srpt_tpg_attrib_srp_max_rsp_size.attr,
3794 	&srpt_tpg_attrib_srp_sq_size.attr,
3795 	NULL,
3796 };
3797 
3798 static ssize_t srpt_tpg_show_enable(
3799 	struct se_portal_group *se_tpg,
3800 	char *page)
3801 {
3802 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3803 
3804 	return snprintf(page, PAGE_SIZE, "%d\n", (sport->enabled) ? 1: 0);
3805 }
3806 
3807 static ssize_t srpt_tpg_store_enable(
3808 	struct se_portal_group *se_tpg,
3809 	const char *page,
3810 	size_t count)
3811 {
3812 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3813 	unsigned long tmp;
3814         int ret;
3815 
3816 	ret = strict_strtoul(page, 0, &tmp);
3817 	if (ret < 0) {
3818 		printk(KERN_ERR "Unable to extract srpt_tpg_store_enable\n");
3819 		return -EINVAL;
3820 	}
3821 
3822 	if ((tmp != 0) && (tmp != 1)) {
3823 		printk(KERN_ERR "Illegal value for srpt_tpg_store_enable: %lu\n", tmp);
3824 		return -EINVAL;
3825 	}
3826 	if (tmp == 1)
3827 		sport->enabled = true;
3828 	else
3829 		sport->enabled = false;
3830 
3831 	return count;
3832 }
3833 
3834 TF_TPG_BASE_ATTR(srpt, enable, S_IRUGO | S_IWUSR);
3835 
3836 static struct configfs_attribute *srpt_tpg_attrs[] = {
3837 	&srpt_tpg_enable.attr,
3838 	NULL,
3839 };
3840 
3841 /**
3842  * configfs callback invoked for
3843  * mkdir /sys/kernel/config/target/$driver/$port/$tpg
3844  */
3845 static struct se_portal_group *srpt_make_tpg(struct se_wwn *wwn,
3846 					     struct config_group *group,
3847 					     const char *name)
3848 {
3849 	struct srpt_port *sport = container_of(wwn, struct srpt_port, port_wwn);
3850 	int res;
3851 
3852 	/* Initialize sport->port_wwn and sport->port_tpg_1 */
3853 	res = core_tpg_register(&srpt_target->tf_ops, &sport->port_wwn,
3854 			&sport->port_tpg_1, sport, TRANSPORT_TPG_TYPE_NORMAL);
3855 	if (res)
3856 		return ERR_PTR(res);
3857 
3858 	return &sport->port_tpg_1;
3859 }
3860 
3861 /**
3862  * configfs callback invoked for
3863  * rmdir /sys/kernel/config/target/$driver/$port/$tpg
3864  */
3865 static void srpt_drop_tpg(struct se_portal_group *tpg)
3866 {
3867 	struct srpt_port *sport = container_of(tpg,
3868 				struct srpt_port, port_tpg_1);
3869 
3870 	sport->enabled = false;
3871 	core_tpg_deregister(&sport->port_tpg_1);
3872 }
3873 
3874 /**
3875  * configfs callback invoked for
3876  * mkdir /sys/kernel/config/target/$driver/$port
3877  */
3878 static struct se_wwn *srpt_make_tport(struct target_fabric_configfs *tf,
3879 				      struct config_group *group,
3880 				      const char *name)
3881 {
3882 	struct srpt_port *sport;
3883 	int ret;
3884 
3885 	sport = srpt_lookup_port(name);
3886 	pr_debug("make_tport(%s)\n", name);
3887 	ret = -EINVAL;
3888 	if (!sport)
3889 		goto err;
3890 
3891 	return &sport->port_wwn;
3892 
3893 err:
3894 	return ERR_PTR(ret);
3895 }
3896 
3897 /**
3898  * configfs callback invoked for
3899  * rmdir /sys/kernel/config/target/$driver/$port
3900  */
3901 static void srpt_drop_tport(struct se_wwn *wwn)
3902 {
3903 	struct srpt_port *sport = container_of(wwn, struct srpt_port, port_wwn);
3904 
3905 	pr_debug("drop_tport(%s\n", config_item_name(&sport->port_wwn.wwn_group.cg_item));
3906 }
3907 
3908 static ssize_t srpt_wwn_show_attr_version(struct target_fabric_configfs *tf,
3909 					      char *buf)
3910 {
3911 	return scnprintf(buf, PAGE_SIZE, "%s\n", DRV_VERSION);
3912 }
3913 
3914 TF_WWN_ATTR_RO(srpt, version);
3915 
3916 static struct configfs_attribute *srpt_wwn_attrs[] = {
3917 	&srpt_wwn_version.attr,
3918 	NULL,
3919 };
3920 
3921 static struct target_core_fabric_ops srpt_template = {
3922 	.get_fabric_name		= srpt_get_fabric_name,
3923 	.get_fabric_proto_ident		= srpt_get_fabric_proto_ident,
3924 	.tpg_get_wwn			= srpt_get_fabric_wwn,
3925 	.tpg_get_tag			= srpt_get_tag,
3926 	.tpg_get_default_depth		= srpt_get_default_depth,
3927 	.tpg_get_pr_transport_id	= srpt_get_pr_transport_id,
3928 	.tpg_get_pr_transport_id_len	= srpt_get_pr_transport_id_len,
3929 	.tpg_parse_pr_out_transport_id	= srpt_parse_pr_out_transport_id,
3930 	.tpg_check_demo_mode		= srpt_check_false,
3931 	.tpg_check_demo_mode_cache	= srpt_check_true,
3932 	.tpg_check_demo_mode_write_protect = srpt_check_true,
3933 	.tpg_check_prod_mode_write_protect = srpt_check_false,
3934 	.tpg_alloc_fabric_acl		= srpt_alloc_fabric_acl,
3935 	.tpg_release_fabric_acl		= srpt_release_fabric_acl,
3936 	.tpg_get_inst_index		= srpt_tpg_get_inst_index,
3937 	.release_cmd			= srpt_release_cmd,
3938 	.check_stop_free		= srpt_check_stop_free,
3939 	.shutdown_session		= srpt_shutdown_session,
3940 	.close_session			= srpt_close_session,
3941 	.sess_get_index			= srpt_sess_get_index,
3942 	.sess_get_initiator_sid		= NULL,
3943 	.write_pending			= srpt_write_pending,
3944 	.write_pending_status		= srpt_write_pending_status,
3945 	.set_default_node_attributes	= srpt_set_default_node_attrs,
3946 	.get_task_tag			= srpt_get_task_tag,
3947 	.get_cmd_state			= srpt_get_tcm_cmd_state,
3948 	.queue_data_in			= srpt_queue_response,
3949 	.queue_status			= srpt_queue_status,
3950 	.queue_tm_rsp			= srpt_queue_response,
3951 	.get_fabric_sense_len		= srpt_get_fabric_sense_len,
3952 	.set_fabric_sense_len		= srpt_set_fabric_sense_len,
3953 	/*
3954 	 * Setup function pointers for generic logic in
3955 	 * target_core_fabric_configfs.c
3956 	 */
3957 	.fabric_make_wwn		= srpt_make_tport,
3958 	.fabric_drop_wwn		= srpt_drop_tport,
3959 	.fabric_make_tpg		= srpt_make_tpg,
3960 	.fabric_drop_tpg		= srpt_drop_tpg,
3961 	.fabric_post_link		= NULL,
3962 	.fabric_pre_unlink		= NULL,
3963 	.fabric_make_np			= NULL,
3964 	.fabric_drop_np			= NULL,
3965 	.fabric_make_nodeacl		= srpt_make_nodeacl,
3966 	.fabric_drop_nodeacl		= srpt_drop_nodeacl,
3967 };
3968 
3969 /**
3970  * srpt_init_module() - Kernel module initialization.
3971  *
3972  * Note: Since ib_register_client() registers callback functions, and since at
3973  * least one of these callback functions (srpt_add_one()) calls target core
3974  * functions, this driver must be registered with the target core before
3975  * ib_register_client() is called.
3976  */
3977 static int __init srpt_init_module(void)
3978 {
3979 	int ret;
3980 
3981 	ret = -EINVAL;
3982 	if (srp_max_req_size < MIN_MAX_REQ_SIZE) {
3983 		printk(KERN_ERR "invalid value %d for kernel module parameter"
3984 		       " srp_max_req_size -- must be at least %d.\n",
3985 		       srp_max_req_size, MIN_MAX_REQ_SIZE);
3986 		goto out;
3987 	}
3988 
3989 	if (srpt_srq_size < MIN_SRPT_SRQ_SIZE
3990 	    || srpt_srq_size > MAX_SRPT_SRQ_SIZE) {
3991 		printk(KERN_ERR "invalid value %d for kernel module parameter"
3992 		       " srpt_srq_size -- must be in the range [%d..%d].\n",
3993 		       srpt_srq_size, MIN_SRPT_SRQ_SIZE, MAX_SRPT_SRQ_SIZE);
3994 		goto out;
3995 	}
3996 
3997 	srpt_target = target_fabric_configfs_init(THIS_MODULE, "srpt");
3998 	if (IS_ERR(srpt_target)) {
3999 		printk(KERN_ERR "couldn't register\n");
4000 		ret = PTR_ERR(srpt_target);
4001 		goto out;
4002 	}
4003 
4004 	srpt_target->tf_ops = srpt_template;
4005 
4006 	/*
4007 	 * Set up default attribute lists.
4008 	 */
4009 	srpt_target->tf_cit_tmpl.tfc_wwn_cit.ct_attrs = srpt_wwn_attrs;
4010 	srpt_target->tf_cit_tmpl.tfc_tpg_base_cit.ct_attrs = srpt_tpg_attrs;
4011 	srpt_target->tf_cit_tmpl.tfc_tpg_attrib_cit.ct_attrs = srpt_tpg_attrib_attrs;
4012 	srpt_target->tf_cit_tmpl.tfc_tpg_param_cit.ct_attrs = NULL;
4013 	srpt_target->tf_cit_tmpl.tfc_tpg_np_base_cit.ct_attrs = NULL;
4014 	srpt_target->tf_cit_tmpl.tfc_tpg_nacl_base_cit.ct_attrs = NULL;
4015 	srpt_target->tf_cit_tmpl.tfc_tpg_nacl_attrib_cit.ct_attrs = NULL;
4016 	srpt_target->tf_cit_tmpl.tfc_tpg_nacl_auth_cit.ct_attrs = NULL;
4017 	srpt_target->tf_cit_tmpl.tfc_tpg_nacl_param_cit.ct_attrs = NULL;
4018 
4019 	ret = target_fabric_configfs_register(srpt_target);
4020 	if (ret < 0) {
4021 		printk(KERN_ERR "couldn't register\n");
4022 		goto out_free_target;
4023 	}
4024 
4025 	ret = ib_register_client(&srpt_client);
4026 	if (ret) {
4027 		printk(KERN_ERR "couldn't register IB client\n");
4028 		goto out_unregister_target;
4029 	}
4030 
4031 	return 0;
4032 
4033 out_unregister_target:
4034 	target_fabric_configfs_deregister(srpt_target);
4035 	srpt_target = NULL;
4036 out_free_target:
4037 	if (srpt_target)
4038 		target_fabric_configfs_free(srpt_target);
4039 out:
4040 	return ret;
4041 }
4042 
4043 static void __exit srpt_cleanup_module(void)
4044 {
4045 	ib_unregister_client(&srpt_client);
4046 	target_fabric_configfs_deregister(srpt_target);
4047 	srpt_target = NULL;
4048 }
4049 
4050 module_init(srpt_init_module);
4051 module_exit(srpt_cleanup_module);
4052