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 const struct target_core_fabric_ops srpt_template;
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 	case IB_EVENT_GID_CHANGE:
202 		/* Refresh port data asynchronously. */
203 		if (event->element.port_num <= sdev->device->phys_port_cnt) {
204 			sport = &sdev->port[event->element.port_num - 1];
205 			if (!sport->lid && !sport->sm_lid)
206 				schedule_work(&sport->work);
207 		}
208 		break;
209 	default:
210 		pr_err("received unrecognized IB event %d\n",
211 		       event->event);
212 		break;
213 	}
214 }
215 
216 /**
217  * srpt_srq_event() - SRQ event callback function.
218  */
219 static void srpt_srq_event(struct ib_event *event, void *ctx)
220 {
221 	pr_info("SRQ event %d\n", event->event);
222 }
223 
224 /**
225  * srpt_qp_event() - QP event callback function.
226  */
227 static void srpt_qp_event(struct ib_event *event, struct srpt_rdma_ch *ch)
228 {
229 	pr_debug("QP event %d on cm_id=%p sess_name=%s state=%d\n",
230 		 event->event, ch->cm_id, ch->sess_name, srpt_get_ch_state(ch));
231 
232 	switch (event->event) {
233 	case IB_EVENT_COMM_EST:
234 		ib_cm_notify(ch->cm_id, event->event);
235 		break;
236 	case IB_EVENT_QP_LAST_WQE_REACHED:
237 		if (srpt_test_and_set_ch_state(ch, CH_DRAINING,
238 					       CH_RELEASING))
239 			srpt_release_channel(ch);
240 		else
241 			pr_debug("%s: state %d - ignored LAST_WQE.\n",
242 				 ch->sess_name, srpt_get_ch_state(ch));
243 		break;
244 	default:
245 		pr_err("received unrecognized IB QP event %d\n", 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, 0);
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 			pr_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 		pr_warn("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 			pr_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 ib_device *dev = ch->sport->sdev->device;
1082 	struct se_cmd *cmd;
1083 	struct scatterlist *sg, *sg_orig;
1084 	int sg_cnt;
1085 	enum dma_data_direction dir;
1086 	struct rdma_iu *riu;
1087 	struct srp_direct_buf *db;
1088 	dma_addr_t dma_addr;
1089 	struct ib_sge *sge;
1090 	u64 raddr;
1091 	u32 rsize;
1092 	u32 tsize;
1093 	u32 dma_len;
1094 	int count, nrdma;
1095 	int i, j, k;
1096 
1097 	BUG_ON(!ch);
1098 	BUG_ON(!ioctx);
1099 	cmd = &ioctx->cmd;
1100 	dir = cmd->data_direction;
1101 	BUG_ON(dir == DMA_NONE);
1102 
1103 	ioctx->sg = sg = sg_orig = cmd->t_data_sg;
1104 	ioctx->sg_cnt = sg_cnt = cmd->t_data_nents;
1105 
1106 	count = ib_dma_map_sg(ch->sport->sdev->device, sg, sg_cnt,
1107 			      opposite_dma_dir(dir));
1108 	if (unlikely(!count))
1109 		return -EAGAIN;
1110 
1111 	ioctx->mapped_sg_count = count;
1112 
1113 	if (ioctx->rdma_ius && ioctx->n_rdma_ius)
1114 		nrdma = ioctx->n_rdma_ius;
1115 	else {
1116 		nrdma = (count + SRPT_DEF_SG_PER_WQE - 1) / SRPT_DEF_SG_PER_WQE
1117 			+ ioctx->n_rbuf;
1118 
1119 		ioctx->rdma_ius = kzalloc(nrdma * sizeof *riu, GFP_KERNEL);
1120 		if (!ioctx->rdma_ius)
1121 			goto free_mem;
1122 
1123 		ioctx->n_rdma_ius = nrdma;
1124 	}
1125 
1126 	db = ioctx->rbufs;
1127 	tsize = cmd->data_length;
1128 	dma_len = ib_sg_dma_len(dev, &sg[0]);
1129 	riu = ioctx->rdma_ius;
1130 
1131 	/*
1132 	 * For each remote desc - calculate the #ib_sge.
1133 	 * If #ib_sge < SRPT_DEF_SG_PER_WQE per rdma operation then
1134 	 *      each remote desc rdma_iu is required a rdma wr;
1135 	 * else
1136 	 *      we need to allocate extra rdma_iu to carry extra #ib_sge in
1137 	 *      another rdma wr
1138 	 */
1139 	for (i = 0, j = 0;
1140 	     j < count && i < ioctx->n_rbuf && tsize > 0; ++i, ++riu, ++db) {
1141 		rsize = be32_to_cpu(db->len);
1142 		raddr = be64_to_cpu(db->va);
1143 		riu->raddr = raddr;
1144 		riu->rkey = be32_to_cpu(db->key);
1145 		riu->sge_cnt = 0;
1146 
1147 		/* calculate how many sge required for this remote_buf */
1148 		while (rsize > 0 && tsize > 0) {
1149 
1150 			if (rsize >= dma_len) {
1151 				tsize -= dma_len;
1152 				rsize -= dma_len;
1153 				raddr += dma_len;
1154 
1155 				if (tsize > 0) {
1156 					++j;
1157 					if (j < count) {
1158 						sg = sg_next(sg);
1159 						dma_len = ib_sg_dma_len(
1160 								dev, sg);
1161 					}
1162 				}
1163 			} else {
1164 				tsize -= rsize;
1165 				dma_len -= rsize;
1166 				rsize = 0;
1167 			}
1168 
1169 			++riu->sge_cnt;
1170 
1171 			if (rsize > 0 && riu->sge_cnt == SRPT_DEF_SG_PER_WQE) {
1172 				++ioctx->n_rdma;
1173 				riu->sge =
1174 				    kmalloc(riu->sge_cnt * sizeof *riu->sge,
1175 					    GFP_KERNEL);
1176 				if (!riu->sge)
1177 					goto free_mem;
1178 
1179 				++riu;
1180 				riu->sge_cnt = 0;
1181 				riu->raddr = raddr;
1182 				riu->rkey = be32_to_cpu(db->key);
1183 			}
1184 		}
1185 
1186 		++ioctx->n_rdma;
1187 		riu->sge = kmalloc(riu->sge_cnt * sizeof *riu->sge,
1188 				   GFP_KERNEL);
1189 		if (!riu->sge)
1190 			goto free_mem;
1191 	}
1192 
1193 	db = ioctx->rbufs;
1194 	tsize = cmd->data_length;
1195 	riu = ioctx->rdma_ius;
1196 	sg = sg_orig;
1197 	dma_len = ib_sg_dma_len(dev, &sg[0]);
1198 	dma_addr = ib_sg_dma_address(dev, &sg[0]);
1199 
1200 	/* this second loop is really mapped sg_addres to rdma_iu->ib_sge */
1201 	for (i = 0, j = 0;
1202 	     j < count && i < ioctx->n_rbuf && tsize > 0; ++i, ++riu, ++db) {
1203 		rsize = be32_to_cpu(db->len);
1204 		sge = riu->sge;
1205 		k = 0;
1206 
1207 		while (rsize > 0 && tsize > 0) {
1208 			sge->addr = dma_addr;
1209 			sge->lkey = ch->sport->sdev->mr->lkey;
1210 
1211 			if (rsize >= dma_len) {
1212 				sge->length =
1213 					(tsize < dma_len) ? tsize : dma_len;
1214 				tsize -= dma_len;
1215 				rsize -= dma_len;
1216 
1217 				if (tsize > 0) {
1218 					++j;
1219 					if (j < count) {
1220 						sg = sg_next(sg);
1221 						dma_len = ib_sg_dma_len(
1222 								dev, sg);
1223 						dma_addr = ib_sg_dma_address(
1224 								dev, sg);
1225 					}
1226 				}
1227 			} else {
1228 				sge->length = (tsize < rsize) ? tsize : rsize;
1229 				tsize -= rsize;
1230 				dma_len -= rsize;
1231 				dma_addr += rsize;
1232 				rsize = 0;
1233 			}
1234 
1235 			++k;
1236 			if (k == riu->sge_cnt && rsize > 0 && tsize > 0) {
1237 				++riu;
1238 				sge = riu->sge;
1239 				k = 0;
1240 			} else if (rsize > 0 && tsize > 0)
1241 				++sge;
1242 		}
1243 	}
1244 
1245 	return 0;
1246 
1247 free_mem:
1248 	srpt_unmap_sg_to_ib_sge(ch, ioctx);
1249 
1250 	return -ENOMEM;
1251 }
1252 
1253 /**
1254  * srpt_get_send_ioctx() - Obtain an I/O context for sending to the initiator.
1255  */
1256 static struct srpt_send_ioctx *srpt_get_send_ioctx(struct srpt_rdma_ch *ch)
1257 {
1258 	struct srpt_send_ioctx *ioctx;
1259 	unsigned long flags;
1260 
1261 	BUG_ON(!ch);
1262 
1263 	ioctx = NULL;
1264 	spin_lock_irqsave(&ch->spinlock, flags);
1265 	if (!list_empty(&ch->free_list)) {
1266 		ioctx = list_first_entry(&ch->free_list,
1267 					 struct srpt_send_ioctx, free_list);
1268 		list_del(&ioctx->free_list);
1269 	}
1270 	spin_unlock_irqrestore(&ch->spinlock, flags);
1271 
1272 	if (!ioctx)
1273 		return ioctx;
1274 
1275 	BUG_ON(ioctx->ch != ch);
1276 	spin_lock_init(&ioctx->spinlock);
1277 	ioctx->state = SRPT_STATE_NEW;
1278 	ioctx->n_rbuf = 0;
1279 	ioctx->rbufs = NULL;
1280 	ioctx->n_rdma = 0;
1281 	ioctx->n_rdma_ius = 0;
1282 	ioctx->rdma_ius = NULL;
1283 	ioctx->mapped_sg_count = 0;
1284 	init_completion(&ioctx->tx_done);
1285 	ioctx->queue_status_only = false;
1286 	/*
1287 	 * transport_init_se_cmd() does not initialize all fields, so do it
1288 	 * here.
1289 	 */
1290 	memset(&ioctx->cmd, 0, sizeof(ioctx->cmd));
1291 	memset(&ioctx->sense_data, 0, sizeof(ioctx->sense_data));
1292 
1293 	return ioctx;
1294 }
1295 
1296 /**
1297  * srpt_abort_cmd() - Abort a SCSI command.
1298  * @ioctx:   I/O context associated with the SCSI command.
1299  * @context: Preferred execution context.
1300  */
1301 static int srpt_abort_cmd(struct srpt_send_ioctx *ioctx)
1302 {
1303 	enum srpt_command_state state;
1304 	unsigned long flags;
1305 
1306 	BUG_ON(!ioctx);
1307 
1308 	/*
1309 	 * If the command is in a state where the target core is waiting for
1310 	 * the ib_srpt driver, change the state to the next state. Changing
1311 	 * the state of the command from SRPT_STATE_NEED_DATA to
1312 	 * SRPT_STATE_DATA_IN ensures that srpt_xmit_response() will call this
1313 	 * function a second time.
1314 	 */
1315 
1316 	spin_lock_irqsave(&ioctx->spinlock, flags);
1317 	state = ioctx->state;
1318 	switch (state) {
1319 	case SRPT_STATE_NEED_DATA:
1320 		ioctx->state = SRPT_STATE_DATA_IN;
1321 		break;
1322 	case SRPT_STATE_DATA_IN:
1323 	case SRPT_STATE_CMD_RSP_SENT:
1324 	case SRPT_STATE_MGMT_RSP_SENT:
1325 		ioctx->state = SRPT_STATE_DONE;
1326 		break;
1327 	default:
1328 		break;
1329 	}
1330 	spin_unlock_irqrestore(&ioctx->spinlock, flags);
1331 
1332 	if (state == SRPT_STATE_DONE) {
1333 		struct srpt_rdma_ch *ch = ioctx->ch;
1334 
1335 		BUG_ON(ch->sess == NULL);
1336 
1337 		target_put_sess_cmd(ch->sess, &ioctx->cmd);
1338 		goto out;
1339 	}
1340 
1341 	pr_debug("Aborting cmd with state %d and tag %lld\n", state,
1342 		 ioctx->tag);
1343 
1344 	switch (state) {
1345 	case SRPT_STATE_NEW:
1346 	case SRPT_STATE_DATA_IN:
1347 	case SRPT_STATE_MGMT:
1348 		/*
1349 		 * Do nothing - defer abort processing until
1350 		 * srpt_queue_response() is invoked.
1351 		 */
1352 		WARN_ON(!transport_check_aborted_status(&ioctx->cmd, false));
1353 		break;
1354 	case SRPT_STATE_NEED_DATA:
1355 		/* DMA_TO_DEVICE (write) - RDMA read error. */
1356 
1357 		/* XXX(hch): this is a horrible layering violation.. */
1358 		spin_lock_irqsave(&ioctx->cmd.t_state_lock, flags);
1359 		ioctx->cmd.transport_state &= ~CMD_T_ACTIVE;
1360 		spin_unlock_irqrestore(&ioctx->cmd.t_state_lock, flags);
1361 		break;
1362 	case SRPT_STATE_CMD_RSP_SENT:
1363 		/*
1364 		 * SRP_RSP sending failed or the SRP_RSP send completion has
1365 		 * not been received in time.
1366 		 */
1367 		srpt_unmap_sg_to_ib_sge(ioctx->ch, ioctx);
1368 		target_put_sess_cmd(ioctx->ch->sess, &ioctx->cmd);
1369 		break;
1370 	case SRPT_STATE_MGMT_RSP_SENT:
1371 		srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
1372 		target_put_sess_cmd(ioctx->ch->sess, &ioctx->cmd);
1373 		break;
1374 	default:
1375 		WARN(1, "Unexpected command state (%d)", state);
1376 		break;
1377 	}
1378 
1379 out:
1380 	return state;
1381 }
1382 
1383 /**
1384  * srpt_handle_send_err_comp() - Process an IB_WC_SEND error completion.
1385  */
1386 static void srpt_handle_send_err_comp(struct srpt_rdma_ch *ch, u64 wr_id)
1387 {
1388 	struct srpt_send_ioctx *ioctx;
1389 	enum srpt_command_state state;
1390 	struct se_cmd *cmd;
1391 	u32 index;
1392 
1393 	atomic_inc(&ch->sq_wr_avail);
1394 
1395 	index = idx_from_wr_id(wr_id);
1396 	ioctx = ch->ioctx_ring[index];
1397 	state = srpt_get_cmd_state(ioctx);
1398 	cmd = &ioctx->cmd;
1399 
1400 	WARN_ON(state != SRPT_STATE_CMD_RSP_SENT
1401 		&& state != SRPT_STATE_MGMT_RSP_SENT
1402 		&& state != SRPT_STATE_NEED_DATA
1403 		&& state != SRPT_STATE_DONE);
1404 
1405 	/* If SRP_RSP sending failed, undo the ch->req_lim change. */
1406 	if (state == SRPT_STATE_CMD_RSP_SENT
1407 	    || state == SRPT_STATE_MGMT_RSP_SENT)
1408 		atomic_dec(&ch->req_lim);
1409 
1410 	srpt_abort_cmd(ioctx);
1411 }
1412 
1413 /**
1414  * srpt_handle_send_comp() - Process an IB send completion notification.
1415  */
1416 static void srpt_handle_send_comp(struct srpt_rdma_ch *ch,
1417 				  struct srpt_send_ioctx *ioctx)
1418 {
1419 	enum srpt_command_state state;
1420 
1421 	atomic_inc(&ch->sq_wr_avail);
1422 
1423 	state = srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
1424 
1425 	if (WARN_ON(state != SRPT_STATE_CMD_RSP_SENT
1426 		    && state != SRPT_STATE_MGMT_RSP_SENT
1427 		    && state != SRPT_STATE_DONE))
1428 		pr_debug("state = %d\n", state);
1429 
1430 	if (state != SRPT_STATE_DONE) {
1431 		srpt_unmap_sg_to_ib_sge(ch, ioctx);
1432 		transport_generic_free_cmd(&ioctx->cmd, 0);
1433 	} else {
1434 		pr_err("IB completion has been received too late for"
1435 		       " wr_id = %u.\n", ioctx->ioctx.index);
1436 	}
1437 }
1438 
1439 /**
1440  * srpt_handle_rdma_comp() - Process an IB RDMA completion notification.
1441  *
1442  * XXX: what is now target_execute_cmd used to be asynchronous, and unmapping
1443  * the data that has been transferred via IB RDMA had to be postponed until the
1444  * check_stop_free() callback.  None of this is necessary anymore and needs to
1445  * be cleaned up.
1446  */
1447 static void srpt_handle_rdma_comp(struct srpt_rdma_ch *ch,
1448 				  struct srpt_send_ioctx *ioctx,
1449 				  enum srpt_opcode opcode)
1450 {
1451 	WARN_ON(ioctx->n_rdma <= 0);
1452 	atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
1453 
1454 	if (opcode == SRPT_RDMA_READ_LAST) {
1455 		if (srpt_test_and_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA,
1456 						SRPT_STATE_DATA_IN))
1457 			target_execute_cmd(&ioctx->cmd);
1458 		else
1459 			pr_err("%s[%d]: wrong state = %d\n", __func__,
1460 			       __LINE__, srpt_get_cmd_state(ioctx));
1461 	} else if (opcode == SRPT_RDMA_ABORT) {
1462 		ioctx->rdma_aborted = true;
1463 	} else {
1464 		WARN(true, "unexpected opcode %d\n", opcode);
1465 	}
1466 }
1467 
1468 /**
1469  * srpt_handle_rdma_err_comp() - Process an IB RDMA error completion.
1470  */
1471 static void srpt_handle_rdma_err_comp(struct srpt_rdma_ch *ch,
1472 				      struct srpt_send_ioctx *ioctx,
1473 				      enum srpt_opcode opcode)
1474 {
1475 	struct se_cmd *cmd;
1476 	enum srpt_command_state state;
1477 
1478 	cmd = &ioctx->cmd;
1479 	state = srpt_get_cmd_state(ioctx);
1480 	switch (opcode) {
1481 	case SRPT_RDMA_READ_LAST:
1482 		if (ioctx->n_rdma <= 0) {
1483 			pr_err("Received invalid RDMA read"
1484 			       " error completion with idx %d\n",
1485 			       ioctx->ioctx.index);
1486 			break;
1487 		}
1488 		atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
1489 		if (state == SRPT_STATE_NEED_DATA)
1490 			srpt_abort_cmd(ioctx);
1491 		else
1492 			pr_err("%s[%d]: wrong state = %d\n",
1493 			       __func__, __LINE__, state);
1494 		break;
1495 	case SRPT_RDMA_WRITE_LAST:
1496 		break;
1497 	default:
1498 		pr_err("%s[%d]: opcode = %u\n", __func__, __LINE__, opcode);
1499 		break;
1500 	}
1501 }
1502 
1503 /**
1504  * srpt_build_cmd_rsp() - Build an SRP_RSP response.
1505  * @ch: RDMA channel through which the request has been received.
1506  * @ioctx: I/O context associated with the SRP_CMD request. The response will
1507  *   be built in the buffer ioctx->buf points at and hence this function will
1508  *   overwrite the request data.
1509  * @tag: tag of the request for which this response is being generated.
1510  * @status: value for the STATUS field of the SRP_RSP information unit.
1511  *
1512  * Returns the size in bytes of the SRP_RSP response.
1513  *
1514  * An SRP_RSP response contains a SCSI status or service response. See also
1515  * section 6.9 in the SRP r16a document for the format of an SRP_RSP
1516  * response. See also SPC-2 for more information about sense data.
1517  */
1518 static int srpt_build_cmd_rsp(struct srpt_rdma_ch *ch,
1519 			      struct srpt_send_ioctx *ioctx, u64 tag,
1520 			      int status)
1521 {
1522 	struct srp_rsp *srp_rsp;
1523 	const u8 *sense_data;
1524 	int sense_data_len, max_sense_len;
1525 
1526 	/*
1527 	 * The lowest bit of all SAM-3 status codes is zero (see also
1528 	 * paragraph 5.3 in SAM-3).
1529 	 */
1530 	WARN_ON(status & 1);
1531 
1532 	srp_rsp = ioctx->ioctx.buf;
1533 	BUG_ON(!srp_rsp);
1534 
1535 	sense_data = ioctx->sense_data;
1536 	sense_data_len = ioctx->cmd.scsi_sense_length;
1537 	WARN_ON(sense_data_len > sizeof(ioctx->sense_data));
1538 
1539 	memset(srp_rsp, 0, sizeof *srp_rsp);
1540 	srp_rsp->opcode = SRP_RSP;
1541 	srp_rsp->req_lim_delta =
1542 		__constant_cpu_to_be32(1 + atomic_xchg(&ch->req_lim_delta, 0));
1543 	srp_rsp->tag = tag;
1544 	srp_rsp->status = status;
1545 
1546 	if (sense_data_len) {
1547 		BUILD_BUG_ON(MIN_MAX_RSP_SIZE <= sizeof(*srp_rsp));
1548 		max_sense_len = ch->max_ti_iu_len - sizeof(*srp_rsp);
1549 		if (sense_data_len > max_sense_len) {
1550 			pr_warn("truncated sense data from %d to %d"
1551 				" bytes\n", sense_data_len, max_sense_len);
1552 			sense_data_len = max_sense_len;
1553 		}
1554 
1555 		srp_rsp->flags |= SRP_RSP_FLAG_SNSVALID;
1556 		srp_rsp->sense_data_len = cpu_to_be32(sense_data_len);
1557 		memcpy(srp_rsp + 1, sense_data, sense_data_len);
1558 	}
1559 
1560 	return sizeof(*srp_rsp) + sense_data_len;
1561 }
1562 
1563 /**
1564  * srpt_build_tskmgmt_rsp() - Build a task management response.
1565  * @ch:       RDMA channel through which the request has been received.
1566  * @ioctx:    I/O context in which the SRP_RSP response will be built.
1567  * @rsp_code: RSP_CODE that will be stored in the response.
1568  * @tag:      Tag of the request for which this response is being generated.
1569  *
1570  * Returns the size in bytes of the SRP_RSP response.
1571  *
1572  * An SRP_RSP response contains a SCSI status or service response. See also
1573  * section 6.9 in the SRP r16a document for the format of an SRP_RSP
1574  * response.
1575  */
1576 static int srpt_build_tskmgmt_rsp(struct srpt_rdma_ch *ch,
1577 				  struct srpt_send_ioctx *ioctx,
1578 				  u8 rsp_code, u64 tag)
1579 {
1580 	struct srp_rsp *srp_rsp;
1581 	int resp_data_len;
1582 	int resp_len;
1583 
1584 	resp_data_len = 4;
1585 	resp_len = sizeof(*srp_rsp) + resp_data_len;
1586 
1587 	srp_rsp = ioctx->ioctx.buf;
1588 	BUG_ON(!srp_rsp);
1589 	memset(srp_rsp, 0, sizeof *srp_rsp);
1590 
1591 	srp_rsp->opcode = SRP_RSP;
1592 	srp_rsp->req_lim_delta = __constant_cpu_to_be32(1
1593 				    + atomic_xchg(&ch->req_lim_delta, 0));
1594 	srp_rsp->tag = tag;
1595 
1596 	srp_rsp->flags |= SRP_RSP_FLAG_RSPVALID;
1597 	srp_rsp->resp_data_len = cpu_to_be32(resp_data_len);
1598 	srp_rsp->data[3] = rsp_code;
1599 
1600 	return resp_len;
1601 }
1602 
1603 #define NO_SUCH_LUN ((uint64_t)-1LL)
1604 
1605 /*
1606  * SCSI LUN addressing method. See also SAM-2 and the section about
1607  * eight byte LUNs.
1608  */
1609 enum scsi_lun_addr_method {
1610 	SCSI_LUN_ADDR_METHOD_PERIPHERAL   = 0,
1611 	SCSI_LUN_ADDR_METHOD_FLAT         = 1,
1612 	SCSI_LUN_ADDR_METHOD_LUN          = 2,
1613 	SCSI_LUN_ADDR_METHOD_EXTENDED_LUN = 3,
1614 };
1615 
1616 /*
1617  * srpt_unpack_lun() - Convert from network LUN to linear LUN.
1618  *
1619  * Convert an 2-byte, 4-byte, 6-byte or 8-byte LUN structure in network byte
1620  * order (big endian) to a linear LUN. Supports three LUN addressing methods:
1621  * peripheral, flat and logical unit. See also SAM-2, section 4.9.4 (page 40).
1622  */
1623 static uint64_t srpt_unpack_lun(const uint8_t *lun, int len)
1624 {
1625 	uint64_t res = NO_SUCH_LUN;
1626 	int addressing_method;
1627 
1628 	if (unlikely(len < 2)) {
1629 		pr_err("Illegal LUN length %d, expected 2 bytes or more\n",
1630 		       len);
1631 		goto out;
1632 	}
1633 
1634 	switch (len) {
1635 	case 8:
1636 		if ((*((__be64 *)lun) &
1637 		     __constant_cpu_to_be64(0x0000FFFFFFFFFFFFLL)) != 0)
1638 			goto out_err;
1639 		break;
1640 	case 4:
1641 		if (*((__be16 *)&lun[2]) != 0)
1642 			goto out_err;
1643 		break;
1644 	case 6:
1645 		if (*((__be32 *)&lun[2]) != 0)
1646 			goto out_err;
1647 		break;
1648 	case 2:
1649 		break;
1650 	default:
1651 		goto out_err;
1652 	}
1653 
1654 	addressing_method = (*lun) >> 6; /* highest two bits of byte 0 */
1655 	switch (addressing_method) {
1656 	case SCSI_LUN_ADDR_METHOD_PERIPHERAL:
1657 	case SCSI_LUN_ADDR_METHOD_FLAT:
1658 	case SCSI_LUN_ADDR_METHOD_LUN:
1659 		res = *(lun + 1) | (((*lun) & 0x3f) << 8);
1660 		break;
1661 
1662 	case SCSI_LUN_ADDR_METHOD_EXTENDED_LUN:
1663 	default:
1664 		pr_err("Unimplemented LUN addressing method %u\n",
1665 		       addressing_method);
1666 		break;
1667 	}
1668 
1669 out:
1670 	return res;
1671 
1672 out_err:
1673 	pr_err("Support for multi-level LUNs has not yet been implemented\n");
1674 	goto out;
1675 }
1676 
1677 static int srpt_check_stop_free(struct se_cmd *cmd)
1678 {
1679 	struct srpt_send_ioctx *ioctx = container_of(cmd,
1680 				struct srpt_send_ioctx, cmd);
1681 
1682 	return target_put_sess_cmd(ioctx->ch->sess, &ioctx->cmd);
1683 }
1684 
1685 /**
1686  * srpt_handle_cmd() - Process SRP_CMD.
1687  */
1688 static int srpt_handle_cmd(struct srpt_rdma_ch *ch,
1689 			   struct srpt_recv_ioctx *recv_ioctx,
1690 			   struct srpt_send_ioctx *send_ioctx)
1691 {
1692 	struct se_cmd *cmd;
1693 	struct srp_cmd *srp_cmd;
1694 	uint64_t unpacked_lun;
1695 	u64 data_len;
1696 	enum dma_data_direction dir;
1697 	sense_reason_t ret;
1698 	int rc;
1699 
1700 	BUG_ON(!send_ioctx);
1701 
1702 	srp_cmd = recv_ioctx->ioctx.buf;
1703 	cmd = &send_ioctx->cmd;
1704 	send_ioctx->tag = srp_cmd->tag;
1705 
1706 	switch (srp_cmd->task_attr) {
1707 	case SRP_CMD_SIMPLE_Q:
1708 		cmd->sam_task_attr = TCM_SIMPLE_TAG;
1709 		break;
1710 	case SRP_CMD_ORDERED_Q:
1711 	default:
1712 		cmd->sam_task_attr = TCM_ORDERED_TAG;
1713 		break;
1714 	case SRP_CMD_HEAD_OF_Q:
1715 		cmd->sam_task_attr = TCM_HEAD_TAG;
1716 		break;
1717 	case SRP_CMD_ACA:
1718 		cmd->sam_task_attr = TCM_ACA_TAG;
1719 		break;
1720 	}
1721 
1722 	if (srpt_get_desc_tbl(send_ioctx, srp_cmd, &dir, &data_len)) {
1723 		pr_err("0x%llx: parsing SRP descriptor table failed.\n",
1724 		       srp_cmd->tag);
1725 		ret = TCM_INVALID_CDB_FIELD;
1726 		goto send_sense;
1727 	}
1728 
1729 	unpacked_lun = srpt_unpack_lun((uint8_t *)&srp_cmd->lun,
1730 				       sizeof(srp_cmd->lun));
1731 	rc = target_submit_cmd(cmd, ch->sess, srp_cmd->cdb,
1732 			&send_ioctx->sense_data[0], unpacked_lun, data_len,
1733 			TCM_SIMPLE_TAG, dir, TARGET_SCF_ACK_KREF);
1734 	if (rc != 0) {
1735 		ret = TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
1736 		goto send_sense;
1737 	}
1738 	return 0;
1739 
1740 send_sense:
1741 	transport_send_check_condition_and_sense(cmd, ret, 0);
1742 	return -1;
1743 }
1744 
1745 /**
1746  * srpt_rx_mgmt_fn_tag() - Process a task management function by tag.
1747  * @ch: RDMA channel of the task management request.
1748  * @fn: Task management function to perform.
1749  * @req_tag: Tag of the SRP task management request.
1750  * @mgmt_ioctx: I/O context of the task management request.
1751  *
1752  * Returns zero if the target core will process the task management
1753  * request asynchronously.
1754  *
1755  * Note: It is assumed that the initiator serializes tag-based task management
1756  * requests.
1757  */
1758 static int srpt_rx_mgmt_fn_tag(struct srpt_send_ioctx *ioctx, u64 tag)
1759 {
1760 	struct srpt_device *sdev;
1761 	struct srpt_rdma_ch *ch;
1762 	struct srpt_send_ioctx *target;
1763 	int ret, i;
1764 
1765 	ret = -EINVAL;
1766 	ch = ioctx->ch;
1767 	BUG_ON(!ch);
1768 	BUG_ON(!ch->sport);
1769 	sdev = ch->sport->sdev;
1770 	BUG_ON(!sdev);
1771 	spin_lock_irq(&sdev->spinlock);
1772 	for (i = 0; i < ch->rq_size; ++i) {
1773 		target = ch->ioctx_ring[i];
1774 		if (target->cmd.se_lun == ioctx->cmd.se_lun &&
1775 		    target->tag == tag &&
1776 		    srpt_get_cmd_state(target) != SRPT_STATE_DONE) {
1777 			ret = 0;
1778 			/* now let the target core abort &target->cmd; */
1779 			break;
1780 		}
1781 	}
1782 	spin_unlock_irq(&sdev->spinlock);
1783 	return ret;
1784 }
1785 
1786 static int srp_tmr_to_tcm(int fn)
1787 {
1788 	switch (fn) {
1789 	case SRP_TSK_ABORT_TASK:
1790 		return TMR_ABORT_TASK;
1791 	case SRP_TSK_ABORT_TASK_SET:
1792 		return TMR_ABORT_TASK_SET;
1793 	case SRP_TSK_CLEAR_TASK_SET:
1794 		return TMR_CLEAR_TASK_SET;
1795 	case SRP_TSK_LUN_RESET:
1796 		return TMR_LUN_RESET;
1797 	case SRP_TSK_CLEAR_ACA:
1798 		return TMR_CLEAR_ACA;
1799 	default:
1800 		return -1;
1801 	}
1802 }
1803 
1804 /**
1805  * srpt_handle_tsk_mgmt() - Process an SRP_TSK_MGMT information unit.
1806  *
1807  * Returns 0 if and only if the request will be processed by the target core.
1808  *
1809  * For more information about SRP_TSK_MGMT information units, see also section
1810  * 6.7 in the SRP r16a document.
1811  */
1812 static void srpt_handle_tsk_mgmt(struct srpt_rdma_ch *ch,
1813 				 struct srpt_recv_ioctx *recv_ioctx,
1814 				 struct srpt_send_ioctx *send_ioctx)
1815 {
1816 	struct srp_tsk_mgmt *srp_tsk;
1817 	struct se_cmd *cmd;
1818 	struct se_session *sess = ch->sess;
1819 	uint64_t unpacked_lun;
1820 	uint32_t tag = 0;
1821 	int tcm_tmr;
1822 	int rc;
1823 
1824 	BUG_ON(!send_ioctx);
1825 
1826 	srp_tsk = recv_ioctx->ioctx.buf;
1827 	cmd = &send_ioctx->cmd;
1828 
1829 	pr_debug("recv tsk_mgmt fn %d for task_tag %lld and cmd tag %lld"
1830 		 " cm_id %p sess %p\n", srp_tsk->tsk_mgmt_func,
1831 		 srp_tsk->task_tag, srp_tsk->tag, ch->cm_id, ch->sess);
1832 
1833 	srpt_set_cmd_state(send_ioctx, SRPT_STATE_MGMT);
1834 	send_ioctx->tag = srp_tsk->tag;
1835 	tcm_tmr = srp_tmr_to_tcm(srp_tsk->tsk_mgmt_func);
1836 	if (tcm_tmr < 0) {
1837 		send_ioctx->cmd.se_tmr_req->response =
1838 			TMR_TASK_MGMT_FUNCTION_NOT_SUPPORTED;
1839 		goto fail;
1840 	}
1841 	unpacked_lun = srpt_unpack_lun((uint8_t *)&srp_tsk->lun,
1842 				       sizeof(srp_tsk->lun));
1843 
1844 	if (srp_tsk->tsk_mgmt_func == SRP_TSK_ABORT_TASK) {
1845 		rc = srpt_rx_mgmt_fn_tag(send_ioctx, srp_tsk->task_tag);
1846 		if (rc < 0) {
1847 			send_ioctx->cmd.se_tmr_req->response =
1848 					TMR_TASK_DOES_NOT_EXIST;
1849 			goto fail;
1850 		}
1851 		tag = srp_tsk->task_tag;
1852 	}
1853 	rc = target_submit_tmr(&send_ioctx->cmd, sess, NULL, unpacked_lun,
1854 				srp_tsk, tcm_tmr, GFP_KERNEL, tag,
1855 				TARGET_SCF_ACK_KREF);
1856 	if (rc != 0) {
1857 		send_ioctx->cmd.se_tmr_req->response = TMR_FUNCTION_REJECTED;
1858 		goto fail;
1859 	}
1860 	return;
1861 fail:
1862 	transport_send_check_condition_and_sense(cmd, 0, 0); // XXX:
1863 }
1864 
1865 /**
1866  * srpt_handle_new_iu() - Process a newly received information unit.
1867  * @ch:    RDMA channel through which the information unit has been received.
1868  * @ioctx: SRPT I/O context associated with the information unit.
1869  */
1870 static void srpt_handle_new_iu(struct srpt_rdma_ch *ch,
1871 			       struct srpt_recv_ioctx *recv_ioctx,
1872 			       struct srpt_send_ioctx *send_ioctx)
1873 {
1874 	struct srp_cmd *srp_cmd;
1875 	enum rdma_ch_state ch_state;
1876 
1877 	BUG_ON(!ch);
1878 	BUG_ON(!recv_ioctx);
1879 
1880 	ib_dma_sync_single_for_cpu(ch->sport->sdev->device,
1881 				   recv_ioctx->ioctx.dma, srp_max_req_size,
1882 				   DMA_FROM_DEVICE);
1883 
1884 	ch_state = srpt_get_ch_state(ch);
1885 	if (unlikely(ch_state == CH_CONNECTING)) {
1886 		list_add_tail(&recv_ioctx->wait_list, &ch->cmd_wait_list);
1887 		goto out;
1888 	}
1889 
1890 	if (unlikely(ch_state != CH_LIVE))
1891 		goto out;
1892 
1893 	srp_cmd = recv_ioctx->ioctx.buf;
1894 	if (srp_cmd->opcode == SRP_CMD || srp_cmd->opcode == SRP_TSK_MGMT) {
1895 		if (!send_ioctx)
1896 			send_ioctx = srpt_get_send_ioctx(ch);
1897 		if (unlikely(!send_ioctx)) {
1898 			list_add_tail(&recv_ioctx->wait_list,
1899 				      &ch->cmd_wait_list);
1900 			goto out;
1901 		}
1902 	}
1903 
1904 	switch (srp_cmd->opcode) {
1905 	case SRP_CMD:
1906 		srpt_handle_cmd(ch, recv_ioctx, send_ioctx);
1907 		break;
1908 	case SRP_TSK_MGMT:
1909 		srpt_handle_tsk_mgmt(ch, recv_ioctx, send_ioctx);
1910 		break;
1911 	case SRP_I_LOGOUT:
1912 		pr_err("Not yet implemented: SRP_I_LOGOUT\n");
1913 		break;
1914 	case SRP_CRED_RSP:
1915 		pr_debug("received SRP_CRED_RSP\n");
1916 		break;
1917 	case SRP_AER_RSP:
1918 		pr_debug("received SRP_AER_RSP\n");
1919 		break;
1920 	case SRP_RSP:
1921 		pr_err("Received SRP_RSP\n");
1922 		break;
1923 	default:
1924 		pr_err("received IU with unknown opcode 0x%x\n",
1925 		       srp_cmd->opcode);
1926 		break;
1927 	}
1928 
1929 	srpt_post_recv(ch->sport->sdev, recv_ioctx);
1930 out:
1931 	return;
1932 }
1933 
1934 static void srpt_process_rcv_completion(struct ib_cq *cq,
1935 					struct srpt_rdma_ch *ch,
1936 					struct ib_wc *wc)
1937 {
1938 	struct srpt_device *sdev = ch->sport->sdev;
1939 	struct srpt_recv_ioctx *ioctx;
1940 	u32 index;
1941 
1942 	index = idx_from_wr_id(wc->wr_id);
1943 	if (wc->status == IB_WC_SUCCESS) {
1944 		int req_lim;
1945 
1946 		req_lim = atomic_dec_return(&ch->req_lim);
1947 		if (unlikely(req_lim < 0))
1948 			pr_err("req_lim = %d < 0\n", req_lim);
1949 		ioctx = sdev->ioctx_ring[index];
1950 		srpt_handle_new_iu(ch, ioctx, NULL);
1951 	} else {
1952 		pr_info("receiving failed for idx %u with status %d\n",
1953 			index, wc->status);
1954 	}
1955 }
1956 
1957 /**
1958  * srpt_process_send_completion() - Process an IB send completion.
1959  *
1960  * Note: Although this has not yet been observed during tests, at least in
1961  * theory it is possible that the srpt_get_send_ioctx() call invoked by
1962  * srpt_handle_new_iu() fails. This is possible because the req_lim_delta
1963  * value in each response is set to one, and it is possible that this response
1964  * makes the initiator send a new request before the send completion for that
1965  * response has been processed. This could e.g. happen if the call to
1966  * srpt_put_send_iotcx() is delayed because of a higher priority interrupt or
1967  * if IB retransmission causes generation of the send completion to be
1968  * delayed. Incoming information units for which srpt_get_send_ioctx() fails
1969  * are queued on cmd_wait_list. The code below processes these delayed
1970  * requests one at a time.
1971  */
1972 static void srpt_process_send_completion(struct ib_cq *cq,
1973 					 struct srpt_rdma_ch *ch,
1974 					 struct ib_wc *wc)
1975 {
1976 	struct srpt_send_ioctx *send_ioctx;
1977 	uint32_t index;
1978 	enum srpt_opcode opcode;
1979 
1980 	index = idx_from_wr_id(wc->wr_id);
1981 	opcode = opcode_from_wr_id(wc->wr_id);
1982 	send_ioctx = ch->ioctx_ring[index];
1983 	if (wc->status == IB_WC_SUCCESS) {
1984 		if (opcode == SRPT_SEND)
1985 			srpt_handle_send_comp(ch, send_ioctx);
1986 		else {
1987 			WARN_ON(opcode != SRPT_RDMA_ABORT &&
1988 				wc->opcode != IB_WC_RDMA_READ);
1989 			srpt_handle_rdma_comp(ch, send_ioctx, opcode);
1990 		}
1991 	} else {
1992 		if (opcode == SRPT_SEND) {
1993 			pr_info("sending response for idx %u failed"
1994 				" with status %d\n", index, wc->status);
1995 			srpt_handle_send_err_comp(ch, wc->wr_id);
1996 		} else if (opcode != SRPT_RDMA_MID) {
1997 			pr_info("RDMA t %d for idx %u failed with"
1998 				" status %d\n", opcode, index, wc->status);
1999 			srpt_handle_rdma_err_comp(ch, send_ioctx, opcode);
2000 		}
2001 	}
2002 
2003 	while (unlikely(opcode == SRPT_SEND
2004 			&& !list_empty(&ch->cmd_wait_list)
2005 			&& srpt_get_ch_state(ch) == CH_LIVE
2006 			&& (send_ioctx = srpt_get_send_ioctx(ch)) != NULL)) {
2007 		struct srpt_recv_ioctx *recv_ioctx;
2008 
2009 		recv_ioctx = list_first_entry(&ch->cmd_wait_list,
2010 					      struct srpt_recv_ioctx,
2011 					      wait_list);
2012 		list_del(&recv_ioctx->wait_list);
2013 		srpt_handle_new_iu(ch, recv_ioctx, send_ioctx);
2014 	}
2015 }
2016 
2017 static void srpt_process_completion(struct ib_cq *cq, struct srpt_rdma_ch *ch)
2018 {
2019 	struct ib_wc *const wc = ch->wc;
2020 	int i, n;
2021 
2022 	WARN_ON(cq != ch->cq);
2023 
2024 	ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
2025 	while ((n = ib_poll_cq(cq, ARRAY_SIZE(ch->wc), wc)) > 0) {
2026 		for (i = 0; i < n; i++) {
2027 			if (opcode_from_wr_id(wc[i].wr_id) == SRPT_RECV)
2028 				srpt_process_rcv_completion(cq, ch, &wc[i]);
2029 			else
2030 				srpt_process_send_completion(cq, ch, &wc[i]);
2031 		}
2032 	}
2033 }
2034 
2035 /**
2036  * srpt_completion() - IB completion queue callback function.
2037  *
2038  * Notes:
2039  * - It is guaranteed that a completion handler will never be invoked
2040  *   concurrently on two different CPUs for the same completion queue. See also
2041  *   Documentation/infiniband/core_locking.txt and the implementation of
2042  *   handle_edge_irq() in kernel/irq/chip.c.
2043  * - When threaded IRQs are enabled, completion handlers are invoked in thread
2044  *   context instead of interrupt context.
2045  */
2046 static void srpt_completion(struct ib_cq *cq, void *ctx)
2047 {
2048 	struct srpt_rdma_ch *ch = ctx;
2049 
2050 	wake_up_interruptible(&ch->wait_queue);
2051 }
2052 
2053 static int srpt_compl_thread(void *arg)
2054 {
2055 	struct srpt_rdma_ch *ch;
2056 
2057 	/* Hibernation / freezing of the SRPT kernel thread is not supported. */
2058 	current->flags |= PF_NOFREEZE;
2059 
2060 	ch = arg;
2061 	BUG_ON(!ch);
2062 	pr_info("Session %s: kernel thread %s (PID %d) started\n",
2063 		ch->sess_name, ch->thread->comm, current->pid);
2064 	while (!kthread_should_stop()) {
2065 		wait_event_interruptible(ch->wait_queue,
2066 			(srpt_process_completion(ch->cq, ch),
2067 			 kthread_should_stop()));
2068 	}
2069 	pr_info("Session %s: kernel thread %s (PID %d) stopped\n",
2070 		ch->sess_name, ch->thread->comm, current->pid);
2071 	return 0;
2072 }
2073 
2074 /**
2075  * srpt_create_ch_ib() - Create receive and send completion queues.
2076  */
2077 static int srpt_create_ch_ib(struct srpt_rdma_ch *ch)
2078 {
2079 	struct ib_qp_init_attr *qp_init;
2080 	struct srpt_port *sport = ch->sport;
2081 	struct srpt_device *sdev = sport->sdev;
2082 	u32 srp_sq_size = sport->port_attrib.srp_sq_size;
2083 	int ret;
2084 
2085 	WARN_ON(ch->rq_size < 1);
2086 
2087 	ret = -ENOMEM;
2088 	qp_init = kzalloc(sizeof *qp_init, GFP_KERNEL);
2089 	if (!qp_init)
2090 		goto out;
2091 
2092 retry:
2093 	ch->cq = ib_create_cq(sdev->device, srpt_completion, NULL, ch,
2094 			      ch->rq_size + srp_sq_size, 0);
2095 	if (IS_ERR(ch->cq)) {
2096 		ret = PTR_ERR(ch->cq);
2097 		pr_err("failed to create CQ cqe= %d ret= %d\n",
2098 		       ch->rq_size + srp_sq_size, ret);
2099 		goto out;
2100 	}
2101 
2102 	qp_init->qp_context = (void *)ch;
2103 	qp_init->event_handler
2104 		= (void(*)(struct ib_event *, void*))srpt_qp_event;
2105 	qp_init->send_cq = ch->cq;
2106 	qp_init->recv_cq = ch->cq;
2107 	qp_init->srq = sdev->srq;
2108 	qp_init->sq_sig_type = IB_SIGNAL_REQ_WR;
2109 	qp_init->qp_type = IB_QPT_RC;
2110 	qp_init->cap.max_send_wr = srp_sq_size;
2111 	qp_init->cap.max_send_sge = SRPT_DEF_SG_PER_WQE;
2112 
2113 	ch->qp = ib_create_qp(sdev->pd, qp_init);
2114 	if (IS_ERR(ch->qp)) {
2115 		ret = PTR_ERR(ch->qp);
2116 		if (ret == -ENOMEM) {
2117 			srp_sq_size /= 2;
2118 			if (srp_sq_size >= MIN_SRPT_SQ_SIZE) {
2119 				ib_destroy_cq(ch->cq);
2120 				goto retry;
2121 			}
2122 		}
2123 		pr_err("failed to create_qp ret= %d\n", ret);
2124 		goto err_destroy_cq;
2125 	}
2126 
2127 	atomic_set(&ch->sq_wr_avail, qp_init->cap.max_send_wr);
2128 
2129 	pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
2130 		 __func__, ch->cq->cqe, qp_init->cap.max_send_sge,
2131 		 qp_init->cap.max_send_wr, ch->cm_id);
2132 
2133 	ret = srpt_init_ch_qp(ch, ch->qp);
2134 	if (ret)
2135 		goto err_destroy_qp;
2136 
2137 	init_waitqueue_head(&ch->wait_queue);
2138 
2139 	pr_debug("creating thread for session %s\n", ch->sess_name);
2140 
2141 	ch->thread = kthread_run(srpt_compl_thread, ch, "ib_srpt_compl");
2142 	if (IS_ERR(ch->thread)) {
2143 		pr_err("failed to create kernel thread %ld\n",
2144 		       PTR_ERR(ch->thread));
2145 		ch->thread = NULL;
2146 		goto err_destroy_qp;
2147 	}
2148 
2149 out:
2150 	kfree(qp_init);
2151 	return ret;
2152 
2153 err_destroy_qp:
2154 	ib_destroy_qp(ch->qp);
2155 err_destroy_cq:
2156 	ib_destroy_cq(ch->cq);
2157 	goto out;
2158 }
2159 
2160 static void srpt_destroy_ch_ib(struct srpt_rdma_ch *ch)
2161 {
2162 	if (ch->thread)
2163 		kthread_stop(ch->thread);
2164 
2165 	ib_destroy_qp(ch->qp);
2166 	ib_destroy_cq(ch->cq);
2167 }
2168 
2169 /**
2170  * __srpt_close_ch() - Close an RDMA channel by setting the QP error state.
2171  *
2172  * Reset the QP and make sure all resources associated with the channel will
2173  * be deallocated at an appropriate time.
2174  *
2175  * Note: The caller must hold ch->sport->sdev->spinlock.
2176  */
2177 static void __srpt_close_ch(struct srpt_rdma_ch *ch)
2178 {
2179 	struct srpt_device *sdev;
2180 	enum rdma_ch_state prev_state;
2181 	unsigned long flags;
2182 
2183 	sdev = ch->sport->sdev;
2184 
2185 	spin_lock_irqsave(&ch->spinlock, flags);
2186 	prev_state = ch->state;
2187 	switch (prev_state) {
2188 	case CH_CONNECTING:
2189 	case CH_LIVE:
2190 		ch->state = CH_DISCONNECTING;
2191 		break;
2192 	default:
2193 		break;
2194 	}
2195 	spin_unlock_irqrestore(&ch->spinlock, flags);
2196 
2197 	switch (prev_state) {
2198 	case CH_CONNECTING:
2199 		ib_send_cm_rej(ch->cm_id, IB_CM_REJ_NO_RESOURCES, NULL, 0,
2200 			       NULL, 0);
2201 		/* fall through */
2202 	case CH_LIVE:
2203 		if (ib_send_cm_dreq(ch->cm_id, NULL, 0) < 0)
2204 			pr_err("sending CM DREQ failed.\n");
2205 		break;
2206 	case CH_DISCONNECTING:
2207 		break;
2208 	case CH_DRAINING:
2209 	case CH_RELEASING:
2210 		break;
2211 	}
2212 }
2213 
2214 /**
2215  * srpt_close_ch() - Close an RDMA channel.
2216  */
2217 static void srpt_close_ch(struct srpt_rdma_ch *ch)
2218 {
2219 	struct srpt_device *sdev;
2220 
2221 	sdev = ch->sport->sdev;
2222 	spin_lock_irq(&sdev->spinlock);
2223 	__srpt_close_ch(ch);
2224 	spin_unlock_irq(&sdev->spinlock);
2225 }
2226 
2227 /**
2228  * srpt_shutdown_session() - Whether or not a session may be shut down.
2229  */
2230 static int srpt_shutdown_session(struct se_session *se_sess)
2231 {
2232 	struct srpt_rdma_ch *ch = se_sess->fabric_sess_ptr;
2233 	unsigned long flags;
2234 
2235 	spin_lock_irqsave(&ch->spinlock, flags);
2236 	if (ch->in_shutdown) {
2237 		spin_unlock_irqrestore(&ch->spinlock, flags);
2238 		return true;
2239 	}
2240 
2241 	ch->in_shutdown = true;
2242 	target_sess_cmd_list_set_waiting(se_sess);
2243 	spin_unlock_irqrestore(&ch->spinlock, flags);
2244 
2245 	return true;
2246 }
2247 
2248 /**
2249  * srpt_drain_channel() - Drain a channel by resetting the IB queue pair.
2250  * @cm_id: Pointer to the CM ID of the channel to be drained.
2251  *
2252  * Note: Must be called from inside srpt_cm_handler to avoid a race between
2253  * accessing sdev->spinlock and the call to kfree(sdev) in srpt_remove_one()
2254  * (the caller of srpt_cm_handler holds the cm_id spinlock; srpt_remove_one()
2255  * waits until all target sessions for the associated IB device have been
2256  * unregistered and target session registration involves a call to
2257  * ib_destroy_cm_id(), which locks the cm_id spinlock and hence waits until
2258  * this function has finished).
2259  */
2260 static void srpt_drain_channel(struct ib_cm_id *cm_id)
2261 {
2262 	struct srpt_device *sdev;
2263 	struct srpt_rdma_ch *ch;
2264 	int ret;
2265 	bool do_reset = false;
2266 
2267 	WARN_ON_ONCE(irqs_disabled());
2268 
2269 	sdev = cm_id->context;
2270 	BUG_ON(!sdev);
2271 	spin_lock_irq(&sdev->spinlock);
2272 	list_for_each_entry(ch, &sdev->rch_list, list) {
2273 		if (ch->cm_id == cm_id) {
2274 			do_reset = srpt_test_and_set_ch_state(ch,
2275 					CH_CONNECTING, CH_DRAINING) ||
2276 				   srpt_test_and_set_ch_state(ch,
2277 					CH_LIVE, CH_DRAINING) ||
2278 				   srpt_test_and_set_ch_state(ch,
2279 					CH_DISCONNECTING, CH_DRAINING);
2280 			break;
2281 		}
2282 	}
2283 	spin_unlock_irq(&sdev->spinlock);
2284 
2285 	if (do_reset) {
2286 		if (ch->sess)
2287 			srpt_shutdown_session(ch->sess);
2288 
2289 		ret = srpt_ch_qp_err(ch);
2290 		if (ret < 0)
2291 			pr_err("Setting queue pair in error state"
2292 			       " failed: %d\n", ret);
2293 	}
2294 }
2295 
2296 /**
2297  * srpt_find_channel() - Look up an RDMA channel.
2298  * @cm_id: Pointer to the CM ID of the channel to be looked up.
2299  *
2300  * Return NULL if no matching RDMA channel has been found.
2301  */
2302 static struct srpt_rdma_ch *srpt_find_channel(struct srpt_device *sdev,
2303 					      struct ib_cm_id *cm_id)
2304 {
2305 	struct srpt_rdma_ch *ch;
2306 	bool found;
2307 
2308 	WARN_ON_ONCE(irqs_disabled());
2309 	BUG_ON(!sdev);
2310 
2311 	found = false;
2312 	spin_lock_irq(&sdev->spinlock);
2313 	list_for_each_entry(ch, &sdev->rch_list, list) {
2314 		if (ch->cm_id == cm_id) {
2315 			found = true;
2316 			break;
2317 		}
2318 	}
2319 	spin_unlock_irq(&sdev->spinlock);
2320 
2321 	return found ? ch : NULL;
2322 }
2323 
2324 /**
2325  * srpt_release_channel() - Release channel resources.
2326  *
2327  * Schedules the actual release because:
2328  * - Calling the ib_destroy_cm_id() call from inside an IB CM callback would
2329  *   trigger a deadlock.
2330  * - It is not safe to call TCM transport_* functions from interrupt context.
2331  */
2332 static void srpt_release_channel(struct srpt_rdma_ch *ch)
2333 {
2334 	schedule_work(&ch->release_work);
2335 }
2336 
2337 static void srpt_release_channel_work(struct work_struct *w)
2338 {
2339 	struct srpt_rdma_ch *ch;
2340 	struct srpt_device *sdev;
2341 	struct se_session *se_sess;
2342 
2343 	ch = container_of(w, struct srpt_rdma_ch, release_work);
2344 	pr_debug("ch = %p; ch->sess = %p; release_done = %p\n", ch, ch->sess,
2345 		 ch->release_done);
2346 
2347 	sdev = ch->sport->sdev;
2348 	BUG_ON(!sdev);
2349 
2350 	se_sess = ch->sess;
2351 	BUG_ON(!se_sess);
2352 
2353 	target_wait_for_sess_cmds(se_sess);
2354 
2355 	transport_deregister_session_configfs(se_sess);
2356 	transport_deregister_session(se_sess);
2357 	ch->sess = NULL;
2358 
2359 	ib_destroy_cm_id(ch->cm_id);
2360 
2361 	srpt_destroy_ch_ib(ch);
2362 
2363 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
2364 			     ch->sport->sdev, ch->rq_size,
2365 			     ch->rsp_size, DMA_TO_DEVICE);
2366 
2367 	spin_lock_irq(&sdev->spinlock);
2368 	list_del(&ch->list);
2369 	spin_unlock_irq(&sdev->spinlock);
2370 
2371 	if (ch->release_done)
2372 		complete(ch->release_done);
2373 
2374 	wake_up(&sdev->ch_releaseQ);
2375 
2376 	kfree(ch);
2377 }
2378 
2379 static struct srpt_node_acl *__srpt_lookup_acl(struct srpt_port *sport,
2380 					       u8 i_port_id[16])
2381 {
2382 	struct srpt_node_acl *nacl;
2383 
2384 	list_for_each_entry(nacl, &sport->port_acl_list, list)
2385 		if (memcmp(nacl->i_port_id, i_port_id,
2386 			   sizeof(nacl->i_port_id)) == 0)
2387 			return nacl;
2388 
2389 	return NULL;
2390 }
2391 
2392 static struct srpt_node_acl *srpt_lookup_acl(struct srpt_port *sport,
2393 					     u8 i_port_id[16])
2394 {
2395 	struct srpt_node_acl *nacl;
2396 
2397 	spin_lock_irq(&sport->port_acl_lock);
2398 	nacl = __srpt_lookup_acl(sport, i_port_id);
2399 	spin_unlock_irq(&sport->port_acl_lock);
2400 
2401 	return nacl;
2402 }
2403 
2404 /**
2405  * srpt_cm_req_recv() - Process the event IB_CM_REQ_RECEIVED.
2406  *
2407  * Ownership of the cm_id is transferred to the target session if this
2408  * functions returns zero. Otherwise the caller remains the owner of cm_id.
2409  */
2410 static int srpt_cm_req_recv(struct ib_cm_id *cm_id,
2411 			    struct ib_cm_req_event_param *param,
2412 			    void *private_data)
2413 {
2414 	struct srpt_device *sdev = cm_id->context;
2415 	struct srpt_port *sport = &sdev->port[param->port - 1];
2416 	struct srp_login_req *req;
2417 	struct srp_login_rsp *rsp;
2418 	struct srp_login_rej *rej;
2419 	struct ib_cm_rep_param *rep_param;
2420 	struct srpt_rdma_ch *ch, *tmp_ch;
2421 	struct srpt_node_acl *nacl;
2422 	u32 it_iu_len;
2423 	int i;
2424 	int ret = 0;
2425 
2426 	WARN_ON_ONCE(irqs_disabled());
2427 
2428 	if (WARN_ON(!sdev || !private_data))
2429 		return -EINVAL;
2430 
2431 	req = (struct srp_login_req *)private_data;
2432 
2433 	it_iu_len = be32_to_cpu(req->req_it_iu_len);
2434 
2435 	pr_info("Received SRP_LOGIN_REQ with i_port_id 0x%llx:0x%llx,"
2436 		" t_port_id 0x%llx:0x%llx and it_iu_len %d on port %d"
2437 		" (guid=0x%llx:0x%llx)\n",
2438 		be64_to_cpu(*(__be64 *)&req->initiator_port_id[0]),
2439 		be64_to_cpu(*(__be64 *)&req->initiator_port_id[8]),
2440 		be64_to_cpu(*(__be64 *)&req->target_port_id[0]),
2441 		be64_to_cpu(*(__be64 *)&req->target_port_id[8]),
2442 		it_iu_len,
2443 		param->port,
2444 		be64_to_cpu(*(__be64 *)&sdev->port[param->port - 1].gid.raw[0]),
2445 		be64_to_cpu(*(__be64 *)&sdev->port[param->port - 1].gid.raw[8]));
2446 
2447 	rsp = kzalloc(sizeof *rsp, GFP_KERNEL);
2448 	rej = kzalloc(sizeof *rej, GFP_KERNEL);
2449 	rep_param = kzalloc(sizeof *rep_param, GFP_KERNEL);
2450 
2451 	if (!rsp || !rej || !rep_param) {
2452 		ret = -ENOMEM;
2453 		goto out;
2454 	}
2455 
2456 	if (it_iu_len > srp_max_req_size || it_iu_len < 64) {
2457 		rej->reason = __constant_cpu_to_be32(
2458 				SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE);
2459 		ret = -EINVAL;
2460 		pr_err("rejected SRP_LOGIN_REQ because its"
2461 		       " length (%d bytes) is out of range (%d .. %d)\n",
2462 		       it_iu_len, 64, srp_max_req_size);
2463 		goto reject;
2464 	}
2465 
2466 	if (!sport->enabled) {
2467 		rej->reason = __constant_cpu_to_be32(
2468 			     SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2469 		ret = -EINVAL;
2470 		pr_err("rejected SRP_LOGIN_REQ because the target port"
2471 		       " has not yet been enabled\n");
2472 		goto reject;
2473 	}
2474 
2475 	if ((req->req_flags & SRP_MTCH_ACTION) == SRP_MULTICHAN_SINGLE) {
2476 		rsp->rsp_flags = SRP_LOGIN_RSP_MULTICHAN_NO_CHAN;
2477 
2478 		spin_lock_irq(&sdev->spinlock);
2479 
2480 		list_for_each_entry_safe(ch, tmp_ch, &sdev->rch_list, list) {
2481 			if (!memcmp(ch->i_port_id, req->initiator_port_id, 16)
2482 			    && !memcmp(ch->t_port_id, req->target_port_id, 16)
2483 			    && param->port == ch->sport->port
2484 			    && param->listen_id == ch->sport->sdev->cm_id
2485 			    && ch->cm_id) {
2486 				enum rdma_ch_state ch_state;
2487 
2488 				ch_state = srpt_get_ch_state(ch);
2489 				if (ch_state != CH_CONNECTING
2490 				    && ch_state != CH_LIVE)
2491 					continue;
2492 
2493 				/* found an existing channel */
2494 				pr_debug("Found existing channel %s"
2495 					 " cm_id= %p state= %d\n",
2496 					 ch->sess_name, ch->cm_id, ch_state);
2497 
2498 				__srpt_close_ch(ch);
2499 
2500 				rsp->rsp_flags =
2501 					SRP_LOGIN_RSP_MULTICHAN_TERMINATED;
2502 			}
2503 		}
2504 
2505 		spin_unlock_irq(&sdev->spinlock);
2506 
2507 	} else
2508 		rsp->rsp_flags = SRP_LOGIN_RSP_MULTICHAN_MAINTAINED;
2509 
2510 	if (*(__be64 *)req->target_port_id != cpu_to_be64(srpt_service_guid)
2511 	    || *(__be64 *)(req->target_port_id + 8) !=
2512 	       cpu_to_be64(srpt_service_guid)) {
2513 		rej->reason = __constant_cpu_to_be32(
2514 				SRP_LOGIN_REJ_UNABLE_ASSOCIATE_CHANNEL);
2515 		ret = -ENOMEM;
2516 		pr_err("rejected SRP_LOGIN_REQ because it"
2517 		       " has an invalid target port identifier.\n");
2518 		goto reject;
2519 	}
2520 
2521 	ch = kzalloc(sizeof *ch, GFP_KERNEL);
2522 	if (!ch) {
2523 		rej->reason = __constant_cpu_to_be32(
2524 					SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2525 		pr_err("rejected SRP_LOGIN_REQ because no memory.\n");
2526 		ret = -ENOMEM;
2527 		goto reject;
2528 	}
2529 
2530 	INIT_WORK(&ch->release_work, srpt_release_channel_work);
2531 	memcpy(ch->i_port_id, req->initiator_port_id, 16);
2532 	memcpy(ch->t_port_id, req->target_port_id, 16);
2533 	ch->sport = &sdev->port[param->port - 1];
2534 	ch->cm_id = cm_id;
2535 	/*
2536 	 * Avoid QUEUE_FULL conditions by limiting the number of buffers used
2537 	 * for the SRP protocol to the command queue size.
2538 	 */
2539 	ch->rq_size = SRPT_RQ_SIZE;
2540 	spin_lock_init(&ch->spinlock);
2541 	ch->state = CH_CONNECTING;
2542 	INIT_LIST_HEAD(&ch->cmd_wait_list);
2543 	ch->rsp_size = ch->sport->port_attrib.srp_max_rsp_size;
2544 
2545 	ch->ioctx_ring = (struct srpt_send_ioctx **)
2546 		srpt_alloc_ioctx_ring(ch->sport->sdev, ch->rq_size,
2547 				      sizeof(*ch->ioctx_ring[0]),
2548 				      ch->rsp_size, DMA_TO_DEVICE);
2549 	if (!ch->ioctx_ring)
2550 		goto free_ch;
2551 
2552 	INIT_LIST_HEAD(&ch->free_list);
2553 	for (i = 0; i < ch->rq_size; i++) {
2554 		ch->ioctx_ring[i]->ch = ch;
2555 		list_add_tail(&ch->ioctx_ring[i]->free_list, &ch->free_list);
2556 	}
2557 
2558 	ret = srpt_create_ch_ib(ch);
2559 	if (ret) {
2560 		rej->reason = __constant_cpu_to_be32(
2561 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2562 		pr_err("rejected SRP_LOGIN_REQ because creating"
2563 		       " a new RDMA channel failed.\n");
2564 		goto free_ring;
2565 	}
2566 
2567 	ret = srpt_ch_qp_rtr(ch, ch->qp);
2568 	if (ret) {
2569 		rej->reason = __constant_cpu_to_be32(
2570 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2571 		pr_err("rejected SRP_LOGIN_REQ because enabling"
2572 		       " RTR failed (error code = %d)\n", ret);
2573 		goto destroy_ib;
2574 	}
2575 	/*
2576 	 * Use the initator port identifier as the session name.
2577 	 */
2578 	snprintf(ch->sess_name, sizeof(ch->sess_name), "0x%016llx%016llx",
2579 			be64_to_cpu(*(__be64 *)ch->i_port_id),
2580 			be64_to_cpu(*(__be64 *)(ch->i_port_id + 8)));
2581 
2582 	pr_debug("registering session %s\n", ch->sess_name);
2583 
2584 	nacl = srpt_lookup_acl(sport, ch->i_port_id);
2585 	if (!nacl) {
2586 		pr_info("Rejected login because no ACL has been"
2587 			" configured yet for initiator %s.\n", ch->sess_name);
2588 		rej->reason = __constant_cpu_to_be32(
2589 				SRP_LOGIN_REJ_CHANNEL_LIMIT_REACHED);
2590 		goto destroy_ib;
2591 	}
2592 
2593 	ch->sess = transport_init_session(TARGET_PROT_NORMAL);
2594 	if (IS_ERR(ch->sess)) {
2595 		rej->reason = __constant_cpu_to_be32(
2596 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2597 		pr_debug("Failed to create session\n");
2598 		goto deregister_session;
2599 	}
2600 	ch->sess->se_node_acl = &nacl->nacl;
2601 	transport_register_session(&sport->port_tpg_1, &nacl->nacl, ch->sess, ch);
2602 
2603 	pr_debug("Establish connection sess=%p name=%s cm_id=%p\n", ch->sess,
2604 		 ch->sess_name, ch->cm_id);
2605 
2606 	/* create srp_login_response */
2607 	rsp->opcode = SRP_LOGIN_RSP;
2608 	rsp->tag = req->tag;
2609 	rsp->max_it_iu_len = req->req_it_iu_len;
2610 	rsp->max_ti_iu_len = req->req_it_iu_len;
2611 	ch->max_ti_iu_len = it_iu_len;
2612 	rsp->buf_fmt = __constant_cpu_to_be16(SRP_BUF_FORMAT_DIRECT
2613 					      | SRP_BUF_FORMAT_INDIRECT);
2614 	rsp->req_lim_delta = cpu_to_be32(ch->rq_size);
2615 	atomic_set(&ch->req_lim, ch->rq_size);
2616 	atomic_set(&ch->req_lim_delta, 0);
2617 
2618 	/* create cm reply */
2619 	rep_param->qp_num = ch->qp->qp_num;
2620 	rep_param->private_data = (void *)rsp;
2621 	rep_param->private_data_len = sizeof *rsp;
2622 	rep_param->rnr_retry_count = 7;
2623 	rep_param->flow_control = 1;
2624 	rep_param->failover_accepted = 0;
2625 	rep_param->srq = 1;
2626 	rep_param->responder_resources = 4;
2627 	rep_param->initiator_depth = 4;
2628 
2629 	ret = ib_send_cm_rep(cm_id, rep_param);
2630 	if (ret) {
2631 		pr_err("sending SRP_LOGIN_REQ response failed"
2632 		       " (error code = %d)\n", ret);
2633 		goto release_channel;
2634 	}
2635 
2636 	spin_lock_irq(&sdev->spinlock);
2637 	list_add_tail(&ch->list, &sdev->rch_list);
2638 	spin_unlock_irq(&sdev->spinlock);
2639 
2640 	goto out;
2641 
2642 release_channel:
2643 	srpt_set_ch_state(ch, CH_RELEASING);
2644 	transport_deregister_session_configfs(ch->sess);
2645 
2646 deregister_session:
2647 	transport_deregister_session(ch->sess);
2648 	ch->sess = NULL;
2649 
2650 destroy_ib:
2651 	srpt_destroy_ch_ib(ch);
2652 
2653 free_ring:
2654 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
2655 			     ch->sport->sdev, ch->rq_size,
2656 			     ch->rsp_size, DMA_TO_DEVICE);
2657 free_ch:
2658 	kfree(ch);
2659 
2660 reject:
2661 	rej->opcode = SRP_LOGIN_REJ;
2662 	rej->tag = req->tag;
2663 	rej->buf_fmt = __constant_cpu_to_be16(SRP_BUF_FORMAT_DIRECT
2664 					      | SRP_BUF_FORMAT_INDIRECT);
2665 
2666 	ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED, NULL, 0,
2667 			     (void *)rej, sizeof *rej);
2668 
2669 out:
2670 	kfree(rep_param);
2671 	kfree(rsp);
2672 	kfree(rej);
2673 
2674 	return ret;
2675 }
2676 
2677 static void srpt_cm_rej_recv(struct ib_cm_id *cm_id)
2678 {
2679 	pr_info("Received IB REJ for cm_id %p.\n", cm_id);
2680 	srpt_drain_channel(cm_id);
2681 }
2682 
2683 /**
2684  * srpt_cm_rtu_recv() - Process an IB_CM_RTU_RECEIVED or USER_ESTABLISHED event.
2685  *
2686  * An IB_CM_RTU_RECEIVED message indicates that the connection is established
2687  * and that the recipient may begin transmitting (RTU = ready to use).
2688  */
2689 static void srpt_cm_rtu_recv(struct ib_cm_id *cm_id)
2690 {
2691 	struct srpt_rdma_ch *ch;
2692 	int ret;
2693 
2694 	ch = srpt_find_channel(cm_id->context, cm_id);
2695 	BUG_ON(!ch);
2696 
2697 	if (srpt_test_and_set_ch_state(ch, CH_CONNECTING, CH_LIVE)) {
2698 		struct srpt_recv_ioctx *ioctx, *ioctx_tmp;
2699 
2700 		ret = srpt_ch_qp_rts(ch, ch->qp);
2701 
2702 		list_for_each_entry_safe(ioctx, ioctx_tmp, &ch->cmd_wait_list,
2703 					 wait_list) {
2704 			list_del(&ioctx->wait_list);
2705 			srpt_handle_new_iu(ch, ioctx, NULL);
2706 		}
2707 		if (ret)
2708 			srpt_close_ch(ch);
2709 	}
2710 }
2711 
2712 static void srpt_cm_timewait_exit(struct ib_cm_id *cm_id)
2713 {
2714 	pr_info("Received IB TimeWait exit for cm_id %p.\n", cm_id);
2715 	srpt_drain_channel(cm_id);
2716 }
2717 
2718 static void srpt_cm_rep_error(struct ib_cm_id *cm_id)
2719 {
2720 	pr_info("Received IB REP error for cm_id %p.\n", cm_id);
2721 	srpt_drain_channel(cm_id);
2722 }
2723 
2724 /**
2725  * srpt_cm_dreq_recv() - Process reception of a DREQ message.
2726  */
2727 static void srpt_cm_dreq_recv(struct ib_cm_id *cm_id)
2728 {
2729 	struct srpt_rdma_ch *ch;
2730 	unsigned long flags;
2731 	bool send_drep = false;
2732 
2733 	ch = srpt_find_channel(cm_id->context, cm_id);
2734 	BUG_ON(!ch);
2735 
2736 	pr_debug("cm_id= %p ch->state= %d\n", cm_id, srpt_get_ch_state(ch));
2737 
2738 	spin_lock_irqsave(&ch->spinlock, flags);
2739 	switch (ch->state) {
2740 	case CH_CONNECTING:
2741 	case CH_LIVE:
2742 		send_drep = true;
2743 		ch->state = CH_DISCONNECTING;
2744 		break;
2745 	case CH_DISCONNECTING:
2746 	case CH_DRAINING:
2747 	case CH_RELEASING:
2748 		WARN(true, "unexpected channel state %d\n", ch->state);
2749 		break;
2750 	}
2751 	spin_unlock_irqrestore(&ch->spinlock, flags);
2752 
2753 	if (send_drep) {
2754 		if (ib_send_cm_drep(ch->cm_id, NULL, 0) < 0)
2755 			pr_err("Sending IB DREP failed.\n");
2756 		pr_info("Received DREQ and sent DREP for session %s.\n",
2757 			ch->sess_name);
2758 	}
2759 }
2760 
2761 /**
2762  * srpt_cm_drep_recv() - Process reception of a DREP message.
2763  */
2764 static void srpt_cm_drep_recv(struct ib_cm_id *cm_id)
2765 {
2766 	pr_info("Received InfiniBand DREP message for cm_id %p.\n", cm_id);
2767 	srpt_drain_channel(cm_id);
2768 }
2769 
2770 /**
2771  * srpt_cm_handler() - IB connection manager callback function.
2772  *
2773  * A non-zero return value will cause the caller destroy the CM ID.
2774  *
2775  * Note: srpt_cm_handler() must only return a non-zero value when transferring
2776  * ownership of the cm_id to a channel by srpt_cm_req_recv() failed. Returning
2777  * a non-zero value in any other case will trigger a race with the
2778  * ib_destroy_cm_id() call in srpt_release_channel().
2779  */
2780 static int srpt_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
2781 {
2782 	int ret;
2783 
2784 	ret = 0;
2785 	switch (event->event) {
2786 	case IB_CM_REQ_RECEIVED:
2787 		ret = srpt_cm_req_recv(cm_id, &event->param.req_rcvd,
2788 				       event->private_data);
2789 		break;
2790 	case IB_CM_REJ_RECEIVED:
2791 		srpt_cm_rej_recv(cm_id);
2792 		break;
2793 	case IB_CM_RTU_RECEIVED:
2794 	case IB_CM_USER_ESTABLISHED:
2795 		srpt_cm_rtu_recv(cm_id);
2796 		break;
2797 	case IB_CM_DREQ_RECEIVED:
2798 		srpt_cm_dreq_recv(cm_id);
2799 		break;
2800 	case IB_CM_DREP_RECEIVED:
2801 		srpt_cm_drep_recv(cm_id);
2802 		break;
2803 	case IB_CM_TIMEWAIT_EXIT:
2804 		srpt_cm_timewait_exit(cm_id);
2805 		break;
2806 	case IB_CM_REP_ERROR:
2807 		srpt_cm_rep_error(cm_id);
2808 		break;
2809 	case IB_CM_DREQ_ERROR:
2810 		pr_info("Received IB DREQ ERROR event.\n");
2811 		break;
2812 	case IB_CM_MRA_RECEIVED:
2813 		pr_info("Received IB MRA event\n");
2814 		break;
2815 	default:
2816 		pr_err("received unrecognized IB CM event %d\n", event->event);
2817 		break;
2818 	}
2819 
2820 	return ret;
2821 }
2822 
2823 /**
2824  * srpt_perform_rdmas() - Perform IB RDMA.
2825  *
2826  * Returns zero upon success or a negative number upon failure.
2827  */
2828 static int srpt_perform_rdmas(struct srpt_rdma_ch *ch,
2829 			      struct srpt_send_ioctx *ioctx)
2830 {
2831 	struct ib_send_wr wr;
2832 	struct ib_send_wr *bad_wr;
2833 	struct rdma_iu *riu;
2834 	int i;
2835 	int ret;
2836 	int sq_wr_avail;
2837 	enum dma_data_direction dir;
2838 	const int n_rdma = ioctx->n_rdma;
2839 
2840 	dir = ioctx->cmd.data_direction;
2841 	if (dir == DMA_TO_DEVICE) {
2842 		/* write */
2843 		ret = -ENOMEM;
2844 		sq_wr_avail = atomic_sub_return(n_rdma, &ch->sq_wr_avail);
2845 		if (sq_wr_avail < 0) {
2846 			pr_warn("IB send queue full (needed %d)\n",
2847 				n_rdma);
2848 			goto out;
2849 		}
2850 	}
2851 
2852 	ioctx->rdma_aborted = false;
2853 	ret = 0;
2854 	riu = ioctx->rdma_ius;
2855 	memset(&wr, 0, sizeof wr);
2856 
2857 	for (i = 0; i < n_rdma; ++i, ++riu) {
2858 		if (dir == DMA_FROM_DEVICE) {
2859 			wr.opcode = IB_WR_RDMA_WRITE;
2860 			wr.wr_id = encode_wr_id(i == n_rdma - 1 ?
2861 						SRPT_RDMA_WRITE_LAST :
2862 						SRPT_RDMA_MID,
2863 						ioctx->ioctx.index);
2864 		} else {
2865 			wr.opcode = IB_WR_RDMA_READ;
2866 			wr.wr_id = encode_wr_id(i == n_rdma - 1 ?
2867 						SRPT_RDMA_READ_LAST :
2868 						SRPT_RDMA_MID,
2869 						ioctx->ioctx.index);
2870 		}
2871 		wr.next = NULL;
2872 		wr.wr.rdma.remote_addr = riu->raddr;
2873 		wr.wr.rdma.rkey = riu->rkey;
2874 		wr.num_sge = riu->sge_cnt;
2875 		wr.sg_list = riu->sge;
2876 
2877 		/* only get completion event for the last rdma write */
2878 		if (i == (n_rdma - 1) && dir == DMA_TO_DEVICE)
2879 			wr.send_flags = IB_SEND_SIGNALED;
2880 
2881 		ret = ib_post_send(ch->qp, &wr, &bad_wr);
2882 		if (ret)
2883 			break;
2884 	}
2885 
2886 	if (ret)
2887 		pr_err("%s[%d]: ib_post_send() returned %d for %d/%d\n",
2888 				 __func__, __LINE__, ret, i, n_rdma);
2889 	if (ret && i > 0) {
2890 		wr.num_sge = 0;
2891 		wr.wr_id = encode_wr_id(SRPT_RDMA_ABORT, ioctx->ioctx.index);
2892 		wr.send_flags = IB_SEND_SIGNALED;
2893 		while (ch->state == CH_LIVE &&
2894 			ib_post_send(ch->qp, &wr, &bad_wr) != 0) {
2895 			pr_info("Trying to abort failed RDMA transfer [%d]\n",
2896 				ioctx->ioctx.index);
2897 			msleep(1000);
2898 		}
2899 		while (ch->state != CH_RELEASING && !ioctx->rdma_aborted) {
2900 			pr_info("Waiting until RDMA abort finished [%d]\n",
2901 				ioctx->ioctx.index);
2902 			msleep(1000);
2903 		}
2904 	}
2905 out:
2906 	if (unlikely(dir == DMA_TO_DEVICE && ret < 0))
2907 		atomic_add(n_rdma, &ch->sq_wr_avail);
2908 	return ret;
2909 }
2910 
2911 /**
2912  * srpt_xfer_data() - Start data transfer from initiator to target.
2913  */
2914 static int srpt_xfer_data(struct srpt_rdma_ch *ch,
2915 			  struct srpt_send_ioctx *ioctx)
2916 {
2917 	int ret;
2918 
2919 	ret = srpt_map_sg_to_ib_sge(ch, ioctx);
2920 	if (ret) {
2921 		pr_err("%s[%d] ret=%d\n", __func__, __LINE__, ret);
2922 		goto out;
2923 	}
2924 
2925 	ret = srpt_perform_rdmas(ch, ioctx);
2926 	if (ret) {
2927 		if (ret == -EAGAIN || ret == -ENOMEM)
2928 			pr_info("%s[%d] queue full -- ret=%d\n",
2929 				__func__, __LINE__, ret);
2930 		else
2931 			pr_err("%s[%d] fatal error -- ret=%d\n",
2932 			       __func__, __LINE__, ret);
2933 		goto out_unmap;
2934 	}
2935 
2936 out:
2937 	return ret;
2938 out_unmap:
2939 	srpt_unmap_sg_to_ib_sge(ch, ioctx);
2940 	goto out;
2941 }
2942 
2943 static int srpt_write_pending_status(struct se_cmd *se_cmd)
2944 {
2945 	struct srpt_send_ioctx *ioctx;
2946 
2947 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
2948 	return srpt_get_cmd_state(ioctx) == SRPT_STATE_NEED_DATA;
2949 }
2950 
2951 /*
2952  * srpt_write_pending() - Start data transfer from initiator to target (write).
2953  */
2954 static int srpt_write_pending(struct se_cmd *se_cmd)
2955 {
2956 	struct srpt_rdma_ch *ch;
2957 	struct srpt_send_ioctx *ioctx;
2958 	enum srpt_command_state new_state;
2959 	enum rdma_ch_state ch_state;
2960 	int ret;
2961 
2962 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
2963 
2964 	new_state = srpt_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA);
2965 	WARN_ON(new_state == SRPT_STATE_DONE);
2966 
2967 	ch = ioctx->ch;
2968 	BUG_ON(!ch);
2969 
2970 	ch_state = srpt_get_ch_state(ch);
2971 	switch (ch_state) {
2972 	case CH_CONNECTING:
2973 		WARN(true, "unexpected channel state %d\n", ch_state);
2974 		ret = -EINVAL;
2975 		goto out;
2976 	case CH_LIVE:
2977 		break;
2978 	case CH_DISCONNECTING:
2979 	case CH_DRAINING:
2980 	case CH_RELEASING:
2981 		pr_debug("cmd with tag %lld: channel disconnecting\n",
2982 			 ioctx->tag);
2983 		srpt_set_cmd_state(ioctx, SRPT_STATE_DATA_IN);
2984 		ret = -EINVAL;
2985 		goto out;
2986 	}
2987 	ret = srpt_xfer_data(ch, ioctx);
2988 
2989 out:
2990 	return ret;
2991 }
2992 
2993 static u8 tcm_to_srp_tsk_mgmt_status(const int tcm_mgmt_status)
2994 {
2995 	switch (tcm_mgmt_status) {
2996 	case TMR_FUNCTION_COMPLETE:
2997 		return SRP_TSK_MGMT_SUCCESS;
2998 	case TMR_FUNCTION_REJECTED:
2999 		return SRP_TSK_MGMT_FUNC_NOT_SUPP;
3000 	}
3001 	return SRP_TSK_MGMT_FAILED;
3002 }
3003 
3004 /**
3005  * srpt_queue_response() - Transmits the response to a SCSI command.
3006  *
3007  * Callback function called by the TCM core. Must not block since it can be
3008  * invoked on the context of the IB completion handler.
3009  */
3010 static void srpt_queue_response(struct se_cmd *cmd)
3011 {
3012 	struct srpt_rdma_ch *ch;
3013 	struct srpt_send_ioctx *ioctx;
3014 	enum srpt_command_state state;
3015 	unsigned long flags;
3016 	int ret;
3017 	enum dma_data_direction dir;
3018 	int resp_len;
3019 	u8 srp_tm_status;
3020 
3021 	ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
3022 	ch = ioctx->ch;
3023 	BUG_ON(!ch);
3024 
3025 	spin_lock_irqsave(&ioctx->spinlock, flags);
3026 	state = ioctx->state;
3027 	switch (state) {
3028 	case SRPT_STATE_NEW:
3029 	case SRPT_STATE_DATA_IN:
3030 		ioctx->state = SRPT_STATE_CMD_RSP_SENT;
3031 		break;
3032 	case SRPT_STATE_MGMT:
3033 		ioctx->state = SRPT_STATE_MGMT_RSP_SENT;
3034 		break;
3035 	default:
3036 		WARN(true, "ch %p; cmd %d: unexpected command state %d\n",
3037 			ch, ioctx->ioctx.index, ioctx->state);
3038 		break;
3039 	}
3040 	spin_unlock_irqrestore(&ioctx->spinlock, flags);
3041 
3042 	if (unlikely(transport_check_aborted_status(&ioctx->cmd, false)
3043 		     || WARN_ON_ONCE(state == SRPT_STATE_CMD_RSP_SENT))) {
3044 		atomic_inc(&ch->req_lim_delta);
3045 		srpt_abort_cmd(ioctx);
3046 		return;
3047 	}
3048 
3049 	dir = ioctx->cmd.data_direction;
3050 
3051 	/* For read commands, transfer the data to the initiator. */
3052 	if (dir == DMA_FROM_DEVICE && ioctx->cmd.data_length &&
3053 	    !ioctx->queue_status_only) {
3054 		ret = srpt_xfer_data(ch, ioctx);
3055 		if (ret) {
3056 			pr_err("xfer_data failed for tag %llu\n",
3057 			       ioctx->tag);
3058 			return;
3059 		}
3060 	}
3061 
3062 	if (state != SRPT_STATE_MGMT)
3063 		resp_len = srpt_build_cmd_rsp(ch, ioctx, ioctx->tag,
3064 					      cmd->scsi_status);
3065 	else {
3066 		srp_tm_status
3067 			= tcm_to_srp_tsk_mgmt_status(cmd->se_tmr_req->response);
3068 		resp_len = srpt_build_tskmgmt_rsp(ch, ioctx, srp_tm_status,
3069 						 ioctx->tag);
3070 	}
3071 	ret = srpt_post_send(ch, ioctx, resp_len);
3072 	if (ret) {
3073 		pr_err("sending cmd response failed for tag %llu\n",
3074 		       ioctx->tag);
3075 		srpt_unmap_sg_to_ib_sge(ch, ioctx);
3076 		srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
3077 		target_put_sess_cmd(ioctx->ch->sess, &ioctx->cmd);
3078 	}
3079 }
3080 
3081 static int srpt_queue_data_in(struct se_cmd *cmd)
3082 {
3083 	srpt_queue_response(cmd);
3084 	return 0;
3085 }
3086 
3087 static void srpt_queue_tm_rsp(struct se_cmd *cmd)
3088 {
3089 	srpt_queue_response(cmd);
3090 }
3091 
3092 static void srpt_aborted_task(struct se_cmd *cmd)
3093 {
3094 	struct srpt_send_ioctx *ioctx = container_of(cmd,
3095 				struct srpt_send_ioctx, cmd);
3096 
3097 	srpt_unmap_sg_to_ib_sge(ioctx->ch, ioctx);
3098 }
3099 
3100 static int srpt_queue_status(struct se_cmd *cmd)
3101 {
3102 	struct srpt_send_ioctx *ioctx;
3103 
3104 	ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
3105 	BUG_ON(ioctx->sense_data != cmd->sense_buffer);
3106 	if (cmd->se_cmd_flags &
3107 	    (SCF_TRANSPORT_TASK_SENSE | SCF_EMULATED_TASK_SENSE))
3108 		WARN_ON(cmd->scsi_status != SAM_STAT_CHECK_CONDITION);
3109 	ioctx->queue_status_only = true;
3110 	srpt_queue_response(cmd);
3111 	return 0;
3112 }
3113 
3114 static void srpt_refresh_port_work(struct work_struct *work)
3115 {
3116 	struct srpt_port *sport = container_of(work, struct srpt_port, work);
3117 
3118 	srpt_refresh_port(sport);
3119 }
3120 
3121 static int srpt_ch_list_empty(struct srpt_device *sdev)
3122 {
3123 	int res;
3124 
3125 	spin_lock_irq(&sdev->spinlock);
3126 	res = list_empty(&sdev->rch_list);
3127 	spin_unlock_irq(&sdev->spinlock);
3128 
3129 	return res;
3130 }
3131 
3132 /**
3133  * srpt_release_sdev() - Free the channel resources associated with a target.
3134  */
3135 static int srpt_release_sdev(struct srpt_device *sdev)
3136 {
3137 	struct srpt_rdma_ch *ch, *tmp_ch;
3138 	int res;
3139 
3140 	WARN_ON_ONCE(irqs_disabled());
3141 
3142 	BUG_ON(!sdev);
3143 
3144 	spin_lock_irq(&sdev->spinlock);
3145 	list_for_each_entry_safe(ch, tmp_ch, &sdev->rch_list, list)
3146 		__srpt_close_ch(ch);
3147 	spin_unlock_irq(&sdev->spinlock);
3148 
3149 	res = wait_event_interruptible(sdev->ch_releaseQ,
3150 				       srpt_ch_list_empty(sdev));
3151 	if (res)
3152 		pr_err("%s: interrupted.\n", __func__);
3153 
3154 	return 0;
3155 }
3156 
3157 static struct srpt_port *__srpt_lookup_port(const char *name)
3158 {
3159 	struct ib_device *dev;
3160 	struct srpt_device *sdev;
3161 	struct srpt_port *sport;
3162 	int i;
3163 
3164 	list_for_each_entry(sdev, &srpt_dev_list, list) {
3165 		dev = sdev->device;
3166 		if (!dev)
3167 			continue;
3168 
3169 		for (i = 0; i < dev->phys_port_cnt; i++) {
3170 			sport = &sdev->port[i];
3171 
3172 			if (!strcmp(sport->port_guid, name))
3173 				return sport;
3174 		}
3175 	}
3176 
3177 	return NULL;
3178 }
3179 
3180 static struct srpt_port *srpt_lookup_port(const char *name)
3181 {
3182 	struct srpt_port *sport;
3183 
3184 	spin_lock(&srpt_dev_lock);
3185 	sport = __srpt_lookup_port(name);
3186 	spin_unlock(&srpt_dev_lock);
3187 
3188 	return sport;
3189 }
3190 
3191 /**
3192  * srpt_add_one() - Infiniband device addition callback function.
3193  */
3194 static void srpt_add_one(struct ib_device *device)
3195 {
3196 	struct srpt_device *sdev;
3197 	struct srpt_port *sport;
3198 	struct ib_srq_init_attr srq_attr;
3199 	int i;
3200 
3201 	pr_debug("device = %p, device->dma_ops = %p\n", device,
3202 		 device->dma_ops);
3203 
3204 	sdev = kzalloc(sizeof *sdev, GFP_KERNEL);
3205 	if (!sdev)
3206 		goto err;
3207 
3208 	sdev->device = device;
3209 	INIT_LIST_HEAD(&sdev->rch_list);
3210 	init_waitqueue_head(&sdev->ch_releaseQ);
3211 	spin_lock_init(&sdev->spinlock);
3212 
3213 	if (ib_query_device(device, &sdev->dev_attr))
3214 		goto free_dev;
3215 
3216 	sdev->pd = ib_alloc_pd(device);
3217 	if (IS_ERR(sdev->pd))
3218 		goto free_dev;
3219 
3220 	sdev->mr = ib_get_dma_mr(sdev->pd, IB_ACCESS_LOCAL_WRITE);
3221 	if (IS_ERR(sdev->mr))
3222 		goto err_pd;
3223 
3224 	sdev->srq_size = min(srpt_srq_size, sdev->dev_attr.max_srq_wr);
3225 
3226 	srq_attr.event_handler = srpt_srq_event;
3227 	srq_attr.srq_context = (void *)sdev;
3228 	srq_attr.attr.max_wr = sdev->srq_size;
3229 	srq_attr.attr.max_sge = 1;
3230 	srq_attr.attr.srq_limit = 0;
3231 	srq_attr.srq_type = IB_SRQT_BASIC;
3232 
3233 	sdev->srq = ib_create_srq(sdev->pd, &srq_attr);
3234 	if (IS_ERR(sdev->srq))
3235 		goto err_mr;
3236 
3237 	pr_debug("%s: create SRQ #wr= %d max_allow=%d dev= %s\n",
3238 		 __func__, sdev->srq_size, sdev->dev_attr.max_srq_wr,
3239 		 device->name);
3240 
3241 	if (!srpt_service_guid)
3242 		srpt_service_guid = be64_to_cpu(device->node_guid);
3243 
3244 	sdev->cm_id = ib_create_cm_id(device, srpt_cm_handler, sdev);
3245 	if (IS_ERR(sdev->cm_id))
3246 		goto err_srq;
3247 
3248 	/* print out target login information */
3249 	pr_debug("Target login info: id_ext=%016llx,ioc_guid=%016llx,"
3250 		 "pkey=ffff,service_id=%016llx\n", srpt_service_guid,
3251 		 srpt_service_guid, srpt_service_guid);
3252 
3253 	/*
3254 	 * We do not have a consistent service_id (ie. also id_ext of target_id)
3255 	 * to identify this target. We currently use the guid of the first HCA
3256 	 * in the system as service_id; therefore, the target_id will change
3257 	 * if this HCA is gone bad and replaced by different HCA
3258 	 */
3259 	if (ib_cm_listen(sdev->cm_id, cpu_to_be64(srpt_service_guid), 0, NULL))
3260 		goto err_cm;
3261 
3262 	INIT_IB_EVENT_HANDLER(&sdev->event_handler, sdev->device,
3263 			      srpt_event_handler);
3264 	if (ib_register_event_handler(&sdev->event_handler))
3265 		goto err_cm;
3266 
3267 	sdev->ioctx_ring = (struct srpt_recv_ioctx **)
3268 		srpt_alloc_ioctx_ring(sdev, sdev->srq_size,
3269 				      sizeof(*sdev->ioctx_ring[0]),
3270 				      srp_max_req_size, DMA_FROM_DEVICE);
3271 	if (!sdev->ioctx_ring)
3272 		goto err_event;
3273 
3274 	for (i = 0; i < sdev->srq_size; ++i)
3275 		srpt_post_recv(sdev, sdev->ioctx_ring[i]);
3276 
3277 	WARN_ON(sdev->device->phys_port_cnt > ARRAY_SIZE(sdev->port));
3278 
3279 	for (i = 1; i <= sdev->device->phys_port_cnt; i++) {
3280 		sport = &sdev->port[i - 1];
3281 		sport->sdev = sdev;
3282 		sport->port = i;
3283 		sport->port_attrib.srp_max_rdma_size = DEFAULT_MAX_RDMA_SIZE;
3284 		sport->port_attrib.srp_max_rsp_size = DEFAULT_MAX_RSP_SIZE;
3285 		sport->port_attrib.srp_sq_size = DEF_SRPT_SQ_SIZE;
3286 		INIT_WORK(&sport->work, srpt_refresh_port_work);
3287 		INIT_LIST_HEAD(&sport->port_acl_list);
3288 		spin_lock_init(&sport->port_acl_lock);
3289 
3290 		if (srpt_refresh_port(sport)) {
3291 			pr_err("MAD registration failed for %s-%d.\n",
3292 			       srpt_sdev_name(sdev), i);
3293 			goto err_ring;
3294 		}
3295 		snprintf(sport->port_guid, sizeof(sport->port_guid),
3296 			"0x%016llx%016llx",
3297 			be64_to_cpu(sport->gid.global.subnet_prefix),
3298 			be64_to_cpu(sport->gid.global.interface_id));
3299 	}
3300 
3301 	spin_lock(&srpt_dev_lock);
3302 	list_add_tail(&sdev->list, &srpt_dev_list);
3303 	spin_unlock(&srpt_dev_lock);
3304 
3305 out:
3306 	ib_set_client_data(device, &srpt_client, sdev);
3307 	pr_debug("added %s.\n", device->name);
3308 	return;
3309 
3310 err_ring:
3311 	srpt_free_ioctx_ring((struct srpt_ioctx **)sdev->ioctx_ring, sdev,
3312 			     sdev->srq_size, srp_max_req_size,
3313 			     DMA_FROM_DEVICE);
3314 err_event:
3315 	ib_unregister_event_handler(&sdev->event_handler);
3316 err_cm:
3317 	ib_destroy_cm_id(sdev->cm_id);
3318 err_srq:
3319 	ib_destroy_srq(sdev->srq);
3320 err_mr:
3321 	ib_dereg_mr(sdev->mr);
3322 err_pd:
3323 	ib_dealloc_pd(sdev->pd);
3324 free_dev:
3325 	kfree(sdev);
3326 err:
3327 	sdev = NULL;
3328 	pr_info("%s(%s) failed.\n", __func__, device->name);
3329 	goto out;
3330 }
3331 
3332 /**
3333  * srpt_remove_one() - InfiniBand device removal callback function.
3334  */
3335 static void srpt_remove_one(struct ib_device *device)
3336 {
3337 	struct srpt_device *sdev;
3338 	int i;
3339 
3340 	sdev = ib_get_client_data(device, &srpt_client);
3341 	if (!sdev) {
3342 		pr_info("%s(%s): nothing to do.\n", __func__, device->name);
3343 		return;
3344 	}
3345 
3346 	srpt_unregister_mad_agent(sdev);
3347 
3348 	ib_unregister_event_handler(&sdev->event_handler);
3349 
3350 	/* Cancel any work queued by the just unregistered IB event handler. */
3351 	for (i = 0; i < sdev->device->phys_port_cnt; i++)
3352 		cancel_work_sync(&sdev->port[i].work);
3353 
3354 	ib_destroy_cm_id(sdev->cm_id);
3355 
3356 	/*
3357 	 * Unregistering a target must happen after destroying sdev->cm_id
3358 	 * such that no new SRP_LOGIN_REQ information units can arrive while
3359 	 * destroying the target.
3360 	 */
3361 	spin_lock(&srpt_dev_lock);
3362 	list_del(&sdev->list);
3363 	spin_unlock(&srpt_dev_lock);
3364 	srpt_release_sdev(sdev);
3365 
3366 	ib_destroy_srq(sdev->srq);
3367 	ib_dereg_mr(sdev->mr);
3368 	ib_dealloc_pd(sdev->pd);
3369 
3370 	srpt_free_ioctx_ring((struct srpt_ioctx **)sdev->ioctx_ring, sdev,
3371 			     sdev->srq_size, srp_max_req_size, DMA_FROM_DEVICE);
3372 	sdev->ioctx_ring = NULL;
3373 	kfree(sdev);
3374 }
3375 
3376 static struct ib_client srpt_client = {
3377 	.name = DRV_NAME,
3378 	.add = srpt_add_one,
3379 	.remove = srpt_remove_one
3380 };
3381 
3382 static int srpt_check_true(struct se_portal_group *se_tpg)
3383 {
3384 	return 1;
3385 }
3386 
3387 static int srpt_check_false(struct se_portal_group *se_tpg)
3388 {
3389 	return 0;
3390 }
3391 
3392 static char *srpt_get_fabric_name(void)
3393 {
3394 	return "srpt";
3395 }
3396 
3397 static u8 srpt_get_fabric_proto_ident(struct se_portal_group *se_tpg)
3398 {
3399 	return SCSI_TRANSPORTID_PROTOCOLID_SRP;
3400 }
3401 
3402 static char *srpt_get_fabric_wwn(struct se_portal_group *tpg)
3403 {
3404 	struct srpt_port *sport = container_of(tpg, struct srpt_port, port_tpg_1);
3405 
3406 	return sport->port_guid;
3407 }
3408 
3409 static u16 srpt_get_tag(struct se_portal_group *tpg)
3410 {
3411 	return 1;
3412 }
3413 
3414 static u32 srpt_get_default_depth(struct se_portal_group *se_tpg)
3415 {
3416 	return 1;
3417 }
3418 
3419 static u32 srpt_get_pr_transport_id(struct se_portal_group *se_tpg,
3420 				    struct se_node_acl *se_nacl,
3421 				    struct t10_pr_registration *pr_reg,
3422 				    int *format_code, unsigned char *buf)
3423 {
3424 	struct srpt_node_acl *nacl;
3425 	struct spc_rdma_transport_id *tr_id;
3426 
3427 	nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3428 	tr_id = (void *)buf;
3429 	tr_id->protocol_identifier = SCSI_TRANSPORTID_PROTOCOLID_SRP;
3430 	memcpy(tr_id->i_port_id, nacl->i_port_id, sizeof(tr_id->i_port_id));
3431 	return sizeof(*tr_id);
3432 }
3433 
3434 static u32 srpt_get_pr_transport_id_len(struct se_portal_group *se_tpg,
3435 					struct se_node_acl *se_nacl,
3436 					struct t10_pr_registration *pr_reg,
3437 					int *format_code)
3438 {
3439 	*format_code = 0;
3440 	return sizeof(struct spc_rdma_transport_id);
3441 }
3442 
3443 static char *srpt_parse_pr_out_transport_id(struct se_portal_group *se_tpg,
3444 					    const char *buf, u32 *out_tid_len,
3445 					    char **port_nexus_ptr)
3446 {
3447 	struct spc_rdma_transport_id *tr_id;
3448 
3449 	*port_nexus_ptr = NULL;
3450 	*out_tid_len = sizeof(struct spc_rdma_transport_id);
3451 	tr_id = (void *)buf;
3452 	return (char *)tr_id->i_port_id;
3453 }
3454 
3455 static struct se_node_acl *srpt_alloc_fabric_acl(struct se_portal_group *se_tpg)
3456 {
3457 	struct srpt_node_acl *nacl;
3458 
3459 	nacl = kzalloc(sizeof(struct srpt_node_acl), GFP_KERNEL);
3460 	if (!nacl) {
3461 		pr_err("Unable to allocate struct srpt_node_acl\n");
3462 		return NULL;
3463 	}
3464 
3465 	return &nacl->nacl;
3466 }
3467 
3468 static void srpt_release_fabric_acl(struct se_portal_group *se_tpg,
3469 				    struct se_node_acl *se_nacl)
3470 {
3471 	struct srpt_node_acl *nacl;
3472 
3473 	nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3474 	kfree(nacl);
3475 }
3476 
3477 static u32 srpt_tpg_get_inst_index(struct se_portal_group *se_tpg)
3478 {
3479 	return 1;
3480 }
3481 
3482 static void srpt_release_cmd(struct se_cmd *se_cmd)
3483 {
3484 	struct srpt_send_ioctx *ioctx = container_of(se_cmd,
3485 				struct srpt_send_ioctx, cmd);
3486 	struct srpt_rdma_ch *ch = ioctx->ch;
3487 	unsigned long flags;
3488 
3489 	WARN_ON(ioctx->state != SRPT_STATE_DONE);
3490 	WARN_ON(ioctx->mapped_sg_count != 0);
3491 
3492 	if (ioctx->n_rbuf > 1) {
3493 		kfree(ioctx->rbufs);
3494 		ioctx->rbufs = NULL;
3495 		ioctx->n_rbuf = 0;
3496 	}
3497 
3498 	spin_lock_irqsave(&ch->spinlock, flags);
3499 	list_add(&ioctx->free_list, &ch->free_list);
3500 	spin_unlock_irqrestore(&ch->spinlock, flags);
3501 }
3502 
3503 /**
3504  * srpt_close_session() - Forcibly close a session.
3505  *
3506  * Callback function invoked by the TCM core to clean up sessions associated
3507  * with a node ACL when the user invokes
3508  * rmdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3509  */
3510 static void srpt_close_session(struct se_session *se_sess)
3511 {
3512 	DECLARE_COMPLETION_ONSTACK(release_done);
3513 	struct srpt_rdma_ch *ch;
3514 	struct srpt_device *sdev;
3515 	unsigned long res;
3516 
3517 	ch = se_sess->fabric_sess_ptr;
3518 	WARN_ON(ch->sess != se_sess);
3519 
3520 	pr_debug("ch %p state %d\n", ch, srpt_get_ch_state(ch));
3521 
3522 	sdev = ch->sport->sdev;
3523 	spin_lock_irq(&sdev->spinlock);
3524 	BUG_ON(ch->release_done);
3525 	ch->release_done = &release_done;
3526 	__srpt_close_ch(ch);
3527 	spin_unlock_irq(&sdev->spinlock);
3528 
3529 	res = wait_for_completion_timeout(&release_done, 60 * HZ);
3530 	WARN_ON(res == 0);
3531 }
3532 
3533 /**
3534  * srpt_sess_get_index() - Return the value of scsiAttIntrPortIndex (SCSI-MIB).
3535  *
3536  * A quote from RFC 4455 (SCSI-MIB) about this MIB object:
3537  * This object represents an arbitrary integer used to uniquely identify a
3538  * particular attached remote initiator port to a particular SCSI target port
3539  * within a particular SCSI target device within a particular SCSI instance.
3540  */
3541 static u32 srpt_sess_get_index(struct se_session *se_sess)
3542 {
3543 	return 0;
3544 }
3545 
3546 static void srpt_set_default_node_attrs(struct se_node_acl *nacl)
3547 {
3548 }
3549 
3550 static u32 srpt_get_task_tag(struct se_cmd *se_cmd)
3551 {
3552 	struct srpt_send_ioctx *ioctx;
3553 
3554 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
3555 	return ioctx->tag;
3556 }
3557 
3558 /* Note: only used from inside debug printk's by the TCM core. */
3559 static int srpt_get_tcm_cmd_state(struct se_cmd *se_cmd)
3560 {
3561 	struct srpt_send_ioctx *ioctx;
3562 
3563 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
3564 	return srpt_get_cmd_state(ioctx);
3565 }
3566 
3567 /**
3568  * srpt_parse_i_port_id() - Parse an initiator port ID.
3569  * @name: ASCII representation of a 128-bit initiator port ID.
3570  * @i_port_id: Binary 128-bit port ID.
3571  */
3572 static int srpt_parse_i_port_id(u8 i_port_id[16], const char *name)
3573 {
3574 	const char *p;
3575 	unsigned len, count, leading_zero_bytes;
3576 	int ret, rc;
3577 
3578 	p = name;
3579 	if (strncasecmp(p, "0x", 2) == 0)
3580 		p += 2;
3581 	ret = -EINVAL;
3582 	len = strlen(p);
3583 	if (len % 2)
3584 		goto out;
3585 	count = min(len / 2, 16U);
3586 	leading_zero_bytes = 16 - count;
3587 	memset(i_port_id, 0, leading_zero_bytes);
3588 	rc = hex2bin(i_port_id + leading_zero_bytes, p, count);
3589 	if (rc < 0)
3590 		pr_debug("hex2bin failed for srpt_parse_i_port_id: %d\n", rc);
3591 	ret = 0;
3592 out:
3593 	return ret;
3594 }
3595 
3596 /*
3597  * configfs callback function invoked for
3598  * mkdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3599  */
3600 static struct se_node_acl *srpt_make_nodeacl(struct se_portal_group *tpg,
3601 					     struct config_group *group,
3602 					     const char *name)
3603 {
3604 	struct srpt_port *sport = container_of(tpg, struct srpt_port, port_tpg_1);
3605 	struct se_node_acl *se_nacl, *se_nacl_new;
3606 	struct srpt_node_acl *nacl;
3607 	int ret = 0;
3608 	u32 nexus_depth = 1;
3609 	u8 i_port_id[16];
3610 
3611 	if (srpt_parse_i_port_id(i_port_id, name) < 0) {
3612 		pr_err("invalid initiator port ID %s\n", name);
3613 		ret = -EINVAL;
3614 		goto err;
3615 	}
3616 
3617 	se_nacl_new = srpt_alloc_fabric_acl(tpg);
3618 	if (!se_nacl_new) {
3619 		ret = -ENOMEM;
3620 		goto err;
3621 	}
3622 	/*
3623 	 * nacl_new may be released by core_tpg_add_initiator_node_acl()
3624 	 * when converting a node ACL from demo mode to explict
3625 	 */
3626 	se_nacl = core_tpg_add_initiator_node_acl(tpg, se_nacl_new, name,
3627 						  nexus_depth);
3628 	if (IS_ERR(se_nacl)) {
3629 		ret = PTR_ERR(se_nacl);
3630 		goto err;
3631 	}
3632 	/* Locate our struct srpt_node_acl and set sdev and i_port_id. */
3633 	nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3634 	memcpy(&nacl->i_port_id[0], &i_port_id[0], 16);
3635 	nacl->sport = sport;
3636 
3637 	spin_lock_irq(&sport->port_acl_lock);
3638 	list_add_tail(&nacl->list, &sport->port_acl_list);
3639 	spin_unlock_irq(&sport->port_acl_lock);
3640 
3641 	return se_nacl;
3642 err:
3643 	return ERR_PTR(ret);
3644 }
3645 
3646 /*
3647  * configfs callback function invoked for
3648  * rmdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3649  */
3650 static void srpt_drop_nodeacl(struct se_node_acl *se_nacl)
3651 {
3652 	struct srpt_node_acl *nacl;
3653 	struct srpt_device *sdev;
3654 	struct srpt_port *sport;
3655 
3656 	nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3657 	sport = nacl->sport;
3658 	sdev = sport->sdev;
3659 	spin_lock_irq(&sport->port_acl_lock);
3660 	list_del(&nacl->list);
3661 	spin_unlock_irq(&sport->port_acl_lock);
3662 	core_tpg_del_initiator_node_acl(&sport->port_tpg_1, se_nacl, 1);
3663 	srpt_release_fabric_acl(NULL, se_nacl);
3664 }
3665 
3666 static ssize_t srpt_tpg_attrib_show_srp_max_rdma_size(
3667 	struct se_portal_group *se_tpg,
3668 	char *page)
3669 {
3670 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3671 
3672 	return sprintf(page, "%u\n", sport->port_attrib.srp_max_rdma_size);
3673 }
3674 
3675 static ssize_t srpt_tpg_attrib_store_srp_max_rdma_size(
3676 	struct se_portal_group *se_tpg,
3677 	const char *page,
3678 	size_t count)
3679 {
3680 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3681 	unsigned long val;
3682 	int ret;
3683 
3684 	ret = kstrtoul(page, 0, &val);
3685 	if (ret < 0) {
3686 		pr_err("kstrtoul() failed with ret: %d\n", ret);
3687 		return -EINVAL;
3688 	}
3689 	if (val > MAX_SRPT_RDMA_SIZE) {
3690 		pr_err("val: %lu exceeds MAX_SRPT_RDMA_SIZE: %d\n", val,
3691 			MAX_SRPT_RDMA_SIZE);
3692 		return -EINVAL;
3693 	}
3694 	if (val < DEFAULT_MAX_RDMA_SIZE) {
3695 		pr_err("val: %lu smaller than DEFAULT_MAX_RDMA_SIZE: %d\n",
3696 			val, DEFAULT_MAX_RDMA_SIZE);
3697 		return -EINVAL;
3698 	}
3699 	sport->port_attrib.srp_max_rdma_size = val;
3700 
3701 	return count;
3702 }
3703 
3704 TF_TPG_ATTRIB_ATTR(srpt, srp_max_rdma_size, S_IRUGO | S_IWUSR);
3705 
3706 static ssize_t srpt_tpg_attrib_show_srp_max_rsp_size(
3707 	struct se_portal_group *se_tpg,
3708 	char *page)
3709 {
3710 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3711 
3712 	return sprintf(page, "%u\n", sport->port_attrib.srp_max_rsp_size);
3713 }
3714 
3715 static ssize_t srpt_tpg_attrib_store_srp_max_rsp_size(
3716 	struct se_portal_group *se_tpg,
3717 	const char *page,
3718 	size_t count)
3719 {
3720 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3721 	unsigned long val;
3722 	int ret;
3723 
3724 	ret = kstrtoul(page, 0, &val);
3725 	if (ret < 0) {
3726 		pr_err("kstrtoul() failed with ret: %d\n", ret);
3727 		return -EINVAL;
3728 	}
3729 	if (val > MAX_SRPT_RSP_SIZE) {
3730 		pr_err("val: %lu exceeds MAX_SRPT_RSP_SIZE: %d\n", val,
3731 			MAX_SRPT_RSP_SIZE);
3732 		return -EINVAL;
3733 	}
3734 	if (val < MIN_MAX_RSP_SIZE) {
3735 		pr_err("val: %lu smaller than MIN_MAX_RSP_SIZE: %d\n", val,
3736 			MIN_MAX_RSP_SIZE);
3737 		return -EINVAL;
3738 	}
3739 	sport->port_attrib.srp_max_rsp_size = val;
3740 
3741 	return count;
3742 }
3743 
3744 TF_TPG_ATTRIB_ATTR(srpt, srp_max_rsp_size, S_IRUGO | S_IWUSR);
3745 
3746 static ssize_t srpt_tpg_attrib_show_srp_sq_size(
3747 	struct se_portal_group *se_tpg,
3748 	char *page)
3749 {
3750 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3751 
3752 	return sprintf(page, "%u\n", sport->port_attrib.srp_sq_size);
3753 }
3754 
3755 static ssize_t srpt_tpg_attrib_store_srp_sq_size(
3756 	struct se_portal_group *se_tpg,
3757 	const char *page,
3758 	size_t count)
3759 {
3760 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3761 	unsigned long val;
3762 	int ret;
3763 
3764 	ret = kstrtoul(page, 0, &val);
3765 	if (ret < 0) {
3766 		pr_err("kstrtoul() failed with ret: %d\n", ret);
3767 		return -EINVAL;
3768 	}
3769 	if (val > MAX_SRPT_SRQ_SIZE) {
3770 		pr_err("val: %lu exceeds MAX_SRPT_SRQ_SIZE: %d\n", val,
3771 			MAX_SRPT_SRQ_SIZE);
3772 		return -EINVAL;
3773 	}
3774 	if (val < MIN_SRPT_SRQ_SIZE) {
3775 		pr_err("val: %lu smaller than MIN_SRPT_SRQ_SIZE: %d\n", val,
3776 			MIN_SRPT_SRQ_SIZE);
3777 		return -EINVAL;
3778 	}
3779 	sport->port_attrib.srp_sq_size = val;
3780 
3781 	return count;
3782 }
3783 
3784 TF_TPG_ATTRIB_ATTR(srpt, srp_sq_size, S_IRUGO | S_IWUSR);
3785 
3786 static struct configfs_attribute *srpt_tpg_attrib_attrs[] = {
3787 	&srpt_tpg_attrib_srp_max_rdma_size.attr,
3788 	&srpt_tpg_attrib_srp_max_rsp_size.attr,
3789 	&srpt_tpg_attrib_srp_sq_size.attr,
3790 	NULL,
3791 };
3792 
3793 static ssize_t srpt_tpg_show_enable(
3794 	struct se_portal_group *se_tpg,
3795 	char *page)
3796 {
3797 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3798 
3799 	return snprintf(page, PAGE_SIZE, "%d\n", (sport->enabled) ? 1: 0);
3800 }
3801 
3802 static ssize_t srpt_tpg_store_enable(
3803 	struct se_portal_group *se_tpg,
3804 	const char *page,
3805 	size_t count)
3806 {
3807 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3808 	unsigned long tmp;
3809         int ret;
3810 
3811 	ret = kstrtoul(page, 0, &tmp);
3812 	if (ret < 0) {
3813 		pr_err("Unable to extract srpt_tpg_store_enable\n");
3814 		return -EINVAL;
3815 	}
3816 
3817 	if ((tmp != 0) && (tmp != 1)) {
3818 		pr_err("Illegal value for srpt_tpg_store_enable: %lu\n", tmp);
3819 		return -EINVAL;
3820 	}
3821 	if (tmp == 1)
3822 		sport->enabled = true;
3823 	else
3824 		sport->enabled = false;
3825 
3826 	return count;
3827 }
3828 
3829 TF_TPG_BASE_ATTR(srpt, enable, S_IRUGO | S_IWUSR);
3830 
3831 static struct configfs_attribute *srpt_tpg_attrs[] = {
3832 	&srpt_tpg_enable.attr,
3833 	NULL,
3834 };
3835 
3836 /**
3837  * configfs callback invoked for
3838  * mkdir /sys/kernel/config/target/$driver/$port/$tpg
3839  */
3840 static struct se_portal_group *srpt_make_tpg(struct se_wwn *wwn,
3841 					     struct config_group *group,
3842 					     const char *name)
3843 {
3844 	struct srpt_port *sport = container_of(wwn, struct srpt_port, port_wwn);
3845 	int res;
3846 
3847 	/* Initialize sport->port_wwn and sport->port_tpg_1 */
3848 	res = core_tpg_register(&srpt_template, &sport->port_wwn,
3849 			&sport->port_tpg_1, sport, TRANSPORT_TPG_TYPE_NORMAL);
3850 	if (res)
3851 		return ERR_PTR(res);
3852 
3853 	return &sport->port_tpg_1;
3854 }
3855 
3856 /**
3857  * configfs callback invoked for
3858  * rmdir /sys/kernel/config/target/$driver/$port/$tpg
3859  */
3860 static void srpt_drop_tpg(struct se_portal_group *tpg)
3861 {
3862 	struct srpt_port *sport = container_of(tpg,
3863 				struct srpt_port, port_tpg_1);
3864 
3865 	sport->enabled = false;
3866 	core_tpg_deregister(&sport->port_tpg_1);
3867 }
3868 
3869 /**
3870  * configfs callback invoked for
3871  * mkdir /sys/kernel/config/target/$driver/$port
3872  */
3873 static struct se_wwn *srpt_make_tport(struct target_fabric_configfs *tf,
3874 				      struct config_group *group,
3875 				      const char *name)
3876 {
3877 	struct srpt_port *sport;
3878 	int ret;
3879 
3880 	sport = srpt_lookup_port(name);
3881 	pr_debug("make_tport(%s)\n", name);
3882 	ret = -EINVAL;
3883 	if (!sport)
3884 		goto err;
3885 
3886 	return &sport->port_wwn;
3887 
3888 err:
3889 	return ERR_PTR(ret);
3890 }
3891 
3892 /**
3893  * configfs callback invoked for
3894  * rmdir /sys/kernel/config/target/$driver/$port
3895  */
3896 static void srpt_drop_tport(struct se_wwn *wwn)
3897 {
3898 	struct srpt_port *sport = container_of(wwn, struct srpt_port, port_wwn);
3899 
3900 	pr_debug("drop_tport(%s\n", config_item_name(&sport->port_wwn.wwn_group.cg_item));
3901 }
3902 
3903 static ssize_t srpt_wwn_show_attr_version(struct target_fabric_configfs *tf,
3904 					      char *buf)
3905 {
3906 	return scnprintf(buf, PAGE_SIZE, "%s\n", DRV_VERSION);
3907 }
3908 
3909 TF_WWN_ATTR_RO(srpt, version);
3910 
3911 static struct configfs_attribute *srpt_wwn_attrs[] = {
3912 	&srpt_wwn_version.attr,
3913 	NULL,
3914 };
3915 
3916 static const struct target_core_fabric_ops srpt_template = {
3917 	.module				= THIS_MODULE,
3918 	.name				= "srpt",
3919 	.get_fabric_name		= srpt_get_fabric_name,
3920 	.get_fabric_proto_ident		= srpt_get_fabric_proto_ident,
3921 	.tpg_get_wwn			= srpt_get_fabric_wwn,
3922 	.tpg_get_tag			= srpt_get_tag,
3923 	.tpg_get_default_depth		= srpt_get_default_depth,
3924 	.tpg_get_pr_transport_id	= srpt_get_pr_transport_id,
3925 	.tpg_get_pr_transport_id_len	= srpt_get_pr_transport_id_len,
3926 	.tpg_parse_pr_out_transport_id	= srpt_parse_pr_out_transport_id,
3927 	.tpg_check_demo_mode		= srpt_check_false,
3928 	.tpg_check_demo_mode_cache	= srpt_check_true,
3929 	.tpg_check_demo_mode_write_protect = srpt_check_true,
3930 	.tpg_check_prod_mode_write_protect = srpt_check_false,
3931 	.tpg_alloc_fabric_acl		= srpt_alloc_fabric_acl,
3932 	.tpg_release_fabric_acl		= srpt_release_fabric_acl,
3933 	.tpg_get_inst_index		= srpt_tpg_get_inst_index,
3934 	.release_cmd			= srpt_release_cmd,
3935 	.check_stop_free		= srpt_check_stop_free,
3936 	.shutdown_session		= srpt_shutdown_session,
3937 	.close_session			= srpt_close_session,
3938 	.sess_get_index			= srpt_sess_get_index,
3939 	.sess_get_initiator_sid		= NULL,
3940 	.write_pending			= srpt_write_pending,
3941 	.write_pending_status		= srpt_write_pending_status,
3942 	.set_default_node_attributes	= srpt_set_default_node_attrs,
3943 	.get_task_tag			= srpt_get_task_tag,
3944 	.get_cmd_state			= srpt_get_tcm_cmd_state,
3945 	.queue_data_in			= srpt_queue_data_in,
3946 	.queue_status			= srpt_queue_status,
3947 	.queue_tm_rsp			= srpt_queue_tm_rsp,
3948 	.aborted_task			= srpt_aborted_task,
3949 	/*
3950 	 * Setup function pointers for generic logic in
3951 	 * target_core_fabric_configfs.c
3952 	 */
3953 	.fabric_make_wwn		= srpt_make_tport,
3954 	.fabric_drop_wwn		= srpt_drop_tport,
3955 	.fabric_make_tpg		= srpt_make_tpg,
3956 	.fabric_drop_tpg		= srpt_drop_tpg,
3957 	.fabric_post_link		= NULL,
3958 	.fabric_pre_unlink		= NULL,
3959 	.fabric_make_np			= NULL,
3960 	.fabric_drop_np			= NULL,
3961 	.fabric_make_nodeacl		= srpt_make_nodeacl,
3962 	.fabric_drop_nodeacl		= srpt_drop_nodeacl,
3963 
3964 	.tfc_wwn_attrs			= srpt_wwn_attrs,
3965 	.tfc_tpg_base_attrs		= srpt_tpg_attrs,
3966 	.tfc_tpg_attrib_attrs		= srpt_tpg_attrib_attrs,
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 		pr_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 		pr_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 	ret = target_register_template(&srpt_template);
3998 	if (ret)
3999 		goto out;
4000 
4001 	ret = ib_register_client(&srpt_client);
4002 	if (ret) {
4003 		pr_err("couldn't register IB client\n");
4004 		goto out_unregister_target;
4005 	}
4006 
4007 	return 0;
4008 
4009 out_unregister_target:
4010 	target_unregister_template(&srpt_template);
4011 out:
4012 	return ret;
4013 }
4014 
4015 static void __exit srpt_cleanup_module(void)
4016 {
4017 	ib_unregister_client(&srpt_client);
4018 	target_unregister_template(&srpt_template);
4019 }
4020 
4021 module_init(srpt_init_module);
4022 module_exit(srpt_cleanup_module);
4023