xref: /linux/drivers/infiniband/ulp/srpt/ib_srpt.c (revision 3e44c471a2dab210f7e9b1e5f7d4d54d52df59eb)
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_proto.h>
45 #include <scsi/scsi_tcq.h>
46 #include <target/configfs_macros.h>
47 #include <target/target_core_base.h>
48 #include <target/target_core_fabric_configfs.h>
49 #include <target/target_core_fabric.h>
50 #include <target/target_core_configfs.h>
51 #include "ib_srpt.h"
52 
53 /* Name of this kernel module. */
54 #define DRV_NAME		"ib_srpt"
55 #define DRV_VERSION		"2.0.0"
56 #define DRV_RELDATE		"2011-02-14"
57 
58 #define SRPT_ID_STRING	"Linux SRP target"
59 
60 #undef pr_fmt
61 #define pr_fmt(fmt) DRV_NAME " " fmt
62 
63 MODULE_AUTHOR("Vu Pham and Bart Van Assche");
64 MODULE_DESCRIPTION("InfiniBand SCSI RDMA Protocol target "
65 		   "v" DRV_VERSION " (" DRV_RELDATE ")");
66 MODULE_LICENSE("Dual BSD/GPL");
67 
68 /*
69  * Global Variables
70  */
71 
72 static u64 srpt_service_guid;
73 static DEFINE_SPINLOCK(srpt_dev_lock);	/* Protects srpt_dev_list. */
74 static LIST_HEAD(srpt_dev_list);	/* List of srpt_device structures. */
75 
76 static unsigned srp_max_req_size = DEFAULT_MAX_REQ_SIZE;
77 module_param(srp_max_req_size, int, 0444);
78 MODULE_PARM_DESC(srp_max_req_size,
79 		 "Maximum size of SRP request messages in bytes.");
80 
81 static int srpt_srq_size = DEFAULT_SRPT_SRQ_SIZE;
82 module_param(srpt_srq_size, int, 0444);
83 MODULE_PARM_DESC(srpt_srq_size,
84 		 "Shared receive queue (SRQ) size.");
85 
86 static int srpt_get_u64_x(char *buffer, struct kernel_param *kp)
87 {
88 	return sprintf(buffer, "0x%016llx", *(u64 *)kp->arg);
89 }
90 module_param_call(srpt_service_guid, NULL, srpt_get_u64_x, &srpt_service_guid,
91 		  0444);
92 MODULE_PARM_DESC(srpt_service_guid,
93 		 "Using this value for ioc_guid, id_ext, and cm_listen_id"
94 		 " instead of using the node_guid of the first HCA.");
95 
96 static struct ib_client srpt_client;
97 static const struct target_core_fabric_ops srpt_template;
98 static void srpt_release_channel(struct srpt_rdma_ch *ch);
99 static int srpt_queue_status(struct se_cmd *cmd);
100 
101 /**
102  * opposite_dma_dir() - Swap DMA_TO_DEVICE and DMA_FROM_DEVICE.
103  */
104 static inline
105 enum dma_data_direction opposite_dma_dir(enum dma_data_direction dir)
106 {
107 	switch (dir) {
108 	case DMA_TO_DEVICE:	return DMA_FROM_DEVICE;
109 	case DMA_FROM_DEVICE:	return DMA_TO_DEVICE;
110 	default:		return dir;
111 	}
112 }
113 
114 /**
115  * srpt_sdev_name() - Return the name associated with the HCA.
116  *
117  * Examples are ib0, ib1, ...
118  */
119 static inline const char *srpt_sdev_name(struct srpt_device *sdev)
120 {
121 	return sdev->device->name;
122 }
123 
124 static enum rdma_ch_state srpt_get_ch_state(struct srpt_rdma_ch *ch)
125 {
126 	unsigned long flags;
127 	enum rdma_ch_state state;
128 
129 	spin_lock_irqsave(&ch->spinlock, flags);
130 	state = ch->state;
131 	spin_unlock_irqrestore(&ch->spinlock, flags);
132 	return state;
133 }
134 
135 static enum rdma_ch_state
136 srpt_set_ch_state(struct srpt_rdma_ch *ch, enum rdma_ch_state new_state)
137 {
138 	unsigned long flags;
139 	enum rdma_ch_state prev;
140 
141 	spin_lock_irqsave(&ch->spinlock, flags);
142 	prev = ch->state;
143 	ch->state = new_state;
144 	spin_unlock_irqrestore(&ch->spinlock, flags);
145 	return prev;
146 }
147 
148 /**
149  * srpt_test_and_set_ch_state() - Test and set the channel state.
150  *
151  * Returns true if and only if the channel state has been set to the new state.
152  */
153 static bool
154 srpt_test_and_set_ch_state(struct srpt_rdma_ch *ch, enum rdma_ch_state old,
155 			   enum rdma_ch_state new)
156 {
157 	unsigned long flags;
158 	enum rdma_ch_state prev;
159 
160 	spin_lock_irqsave(&ch->spinlock, flags);
161 	prev = ch->state;
162 	if (prev == old)
163 		ch->state = new;
164 	spin_unlock_irqrestore(&ch->spinlock, flags);
165 	return prev == old;
166 }
167 
168 /**
169  * srpt_event_handler() - Asynchronous IB event callback function.
170  *
171  * Callback function called by the InfiniBand core when an asynchronous IB
172  * event occurs. This callback may occur in interrupt context. See also
173  * section 11.5.2, Set Asynchronous Event Handler in the InfiniBand
174  * Architecture Specification.
175  */
176 static void srpt_event_handler(struct ib_event_handler *handler,
177 			       struct ib_event *event)
178 {
179 	struct srpt_device *sdev;
180 	struct srpt_port *sport;
181 
182 	sdev = ib_get_client_data(event->device, &srpt_client);
183 	if (!sdev || sdev->device != event->device)
184 		return;
185 
186 	pr_debug("ASYNC event= %d on device= %s\n", event->event,
187 		 srpt_sdev_name(sdev));
188 
189 	switch (event->event) {
190 	case IB_EVENT_PORT_ERR:
191 		if (event->element.port_num <= sdev->device->phys_port_cnt) {
192 			sport = &sdev->port[event->element.port_num - 1];
193 			sport->lid = 0;
194 			sport->sm_lid = 0;
195 		}
196 		break;
197 	case IB_EVENT_PORT_ACTIVE:
198 	case IB_EVENT_LID_CHANGE:
199 	case IB_EVENT_PKEY_CHANGE:
200 	case IB_EVENT_SM_CHANGE:
201 	case IB_EVENT_CLIENT_REREGISTER:
202 	case IB_EVENT_GID_CHANGE:
203 		/* Refresh port data asynchronously. */
204 		if (event->element.port_num <= sdev->device->phys_port_cnt) {
205 			sport = &sdev->port[event->element.port_num - 1];
206 			if (!sport->lid && !sport->sm_lid)
207 				schedule_work(&sport->work);
208 		}
209 		break;
210 	default:
211 		pr_err("received unrecognized IB event %d\n",
212 		       event->event);
213 		break;
214 	}
215 }
216 
217 /**
218  * srpt_srq_event() - SRQ event callback function.
219  */
220 static void srpt_srq_event(struct ib_event *event, void *ctx)
221 {
222 	pr_info("SRQ event %d\n", event->event);
223 }
224 
225 /**
226  * srpt_qp_event() - QP event callback function.
227  */
228 static void srpt_qp_event(struct ib_event *event, struct srpt_rdma_ch *ch)
229 {
230 	pr_debug("QP event %d on cm_id=%p sess_name=%s state=%d\n",
231 		 event->event, ch->cm_id, ch->sess_name, srpt_get_ch_state(ch));
232 
233 	switch (event->event) {
234 	case IB_EVENT_COMM_EST:
235 		ib_cm_notify(ch->cm_id, event->event);
236 		break;
237 	case IB_EVENT_QP_LAST_WQE_REACHED:
238 		if (srpt_test_and_set_ch_state(ch, CH_DRAINING,
239 					       CH_RELEASING))
240 			srpt_release_channel(ch);
241 		else
242 			pr_debug("%s: state %d - ignored LAST_WQE.\n",
243 				 ch->sess_name, srpt_get_ch_state(ch));
244 		break;
245 	default:
246 		pr_err("received unrecognized IB QP event %d\n", event->event);
247 		break;
248 	}
249 }
250 
251 /**
252  * srpt_set_ioc() - Helper function for initializing an IOUnitInfo structure.
253  *
254  * @slot: one-based slot number.
255  * @value: four-bit value.
256  *
257  * Copies the lowest four bits of value in element slot of the array of four
258  * bit elements called c_list (controller list). The index slot is one-based.
259  */
260 static void srpt_set_ioc(u8 *c_list, u32 slot, u8 value)
261 {
262 	u16 id;
263 	u8 tmp;
264 
265 	id = (slot - 1) / 2;
266 	if (slot & 0x1) {
267 		tmp = c_list[id] & 0xf;
268 		c_list[id] = (value << 4) | tmp;
269 	} else {
270 		tmp = c_list[id] & 0xf0;
271 		c_list[id] = (value & 0xf) | tmp;
272 	}
273 }
274 
275 /**
276  * srpt_get_class_port_info() - Copy ClassPortInfo to a management datagram.
277  *
278  * See also section 16.3.3.1 ClassPortInfo in the InfiniBand Architecture
279  * Specification.
280  */
281 static void srpt_get_class_port_info(struct ib_dm_mad *mad)
282 {
283 	struct ib_class_port_info *cif;
284 
285 	cif = (struct ib_class_port_info *)mad->data;
286 	memset(cif, 0, sizeof *cif);
287 	cif->base_version = 1;
288 	cif->class_version = 1;
289 	cif->resp_time_value = 20;
290 
291 	mad->mad_hdr.status = 0;
292 }
293 
294 /**
295  * srpt_get_iou() - Write IOUnitInfo to a management datagram.
296  *
297  * See also section 16.3.3.3 IOUnitInfo in the InfiniBand Architecture
298  * Specification. See also section B.7, table B.6 in the SRP r16a document.
299  */
300 static void srpt_get_iou(struct ib_dm_mad *mad)
301 {
302 	struct ib_dm_iou_info *ioui;
303 	u8 slot;
304 	int i;
305 
306 	ioui = (struct ib_dm_iou_info *)mad->data;
307 	ioui->change_id = __constant_cpu_to_be16(1);
308 	ioui->max_controllers = 16;
309 
310 	/* set present for slot 1 and empty for the rest */
311 	srpt_set_ioc(ioui->controller_list, 1, 1);
312 	for (i = 1, slot = 2; i < 16; i++, slot++)
313 		srpt_set_ioc(ioui->controller_list, slot, 0);
314 
315 	mad->mad_hdr.status = 0;
316 }
317 
318 /**
319  * srpt_get_ioc() - Write IOControllerprofile to a management datagram.
320  *
321  * See also section 16.3.3.4 IOControllerProfile in the InfiniBand
322  * Architecture Specification. See also section B.7, table B.7 in the SRP
323  * r16a document.
324  */
325 static void srpt_get_ioc(struct srpt_port *sport, u32 slot,
326 			 struct ib_dm_mad *mad)
327 {
328 	struct srpt_device *sdev = sport->sdev;
329 	struct ib_dm_ioc_profile *iocp;
330 
331 	iocp = (struct ib_dm_ioc_profile *)mad->data;
332 
333 	if (!slot || slot > 16) {
334 		mad->mad_hdr.status
335 			= __constant_cpu_to_be16(DM_MAD_STATUS_INVALID_FIELD);
336 		return;
337 	}
338 
339 	if (slot > 2) {
340 		mad->mad_hdr.status
341 			= __constant_cpu_to_be16(DM_MAD_STATUS_NO_IOC);
342 		return;
343 	}
344 
345 	memset(iocp, 0, sizeof *iocp);
346 	strcpy(iocp->id_string, SRPT_ID_STRING);
347 	iocp->guid = cpu_to_be64(srpt_service_guid);
348 	iocp->vendor_id = cpu_to_be32(sdev->dev_attr.vendor_id);
349 	iocp->device_id = cpu_to_be32(sdev->dev_attr.vendor_part_id);
350 	iocp->device_version = cpu_to_be16(sdev->dev_attr.hw_ver);
351 	iocp->subsys_vendor_id = cpu_to_be32(sdev->dev_attr.vendor_id);
352 	iocp->subsys_device_id = 0x0;
353 	iocp->io_class = __constant_cpu_to_be16(SRP_REV16A_IB_IO_CLASS);
354 	iocp->io_subclass = __constant_cpu_to_be16(SRP_IO_SUBCLASS);
355 	iocp->protocol = __constant_cpu_to_be16(SRP_PROTOCOL);
356 	iocp->protocol_version = __constant_cpu_to_be16(SRP_PROTOCOL_VERSION);
357 	iocp->send_queue_depth = cpu_to_be16(sdev->srq_size);
358 	iocp->rdma_read_depth = 4;
359 	iocp->send_size = cpu_to_be32(srp_max_req_size);
360 	iocp->rdma_size = cpu_to_be32(min(sport->port_attrib.srp_max_rdma_size,
361 					  1U << 24));
362 	iocp->num_svc_entries = 1;
363 	iocp->op_cap_mask = SRP_SEND_TO_IOC | SRP_SEND_FROM_IOC |
364 		SRP_RDMA_READ_FROM_IOC | SRP_RDMA_WRITE_FROM_IOC;
365 
366 	mad->mad_hdr.status = 0;
367 }
368 
369 /**
370  * srpt_get_svc_entries() - Write ServiceEntries to a management datagram.
371  *
372  * See also section 16.3.3.5 ServiceEntries in the InfiniBand Architecture
373  * Specification. See also section B.7, table B.8 in the SRP r16a document.
374  */
375 static void srpt_get_svc_entries(u64 ioc_guid,
376 				 u16 slot, u8 hi, u8 lo, struct ib_dm_mad *mad)
377 {
378 	struct ib_dm_svc_entries *svc_entries;
379 
380 	WARN_ON(!ioc_guid);
381 
382 	if (!slot || slot > 16) {
383 		mad->mad_hdr.status
384 			= __constant_cpu_to_be16(DM_MAD_STATUS_INVALID_FIELD);
385 		return;
386 	}
387 
388 	if (slot > 2 || lo > hi || hi > 1) {
389 		mad->mad_hdr.status
390 			= __constant_cpu_to_be16(DM_MAD_STATUS_NO_IOC);
391 		return;
392 	}
393 
394 	svc_entries = (struct ib_dm_svc_entries *)mad->data;
395 	memset(svc_entries, 0, sizeof *svc_entries);
396 	svc_entries->service_entries[0].id = cpu_to_be64(ioc_guid);
397 	snprintf(svc_entries->service_entries[0].name,
398 		 sizeof(svc_entries->service_entries[0].name),
399 		 "%s%016llx",
400 		 SRP_SERVICE_NAME_PREFIX,
401 		 ioc_guid);
402 
403 	mad->mad_hdr.status = 0;
404 }
405 
406 /**
407  * srpt_mgmt_method_get() - Process a received management datagram.
408  * @sp:      source port through which the MAD has been received.
409  * @rq_mad:  received MAD.
410  * @rsp_mad: response MAD.
411  */
412 static void srpt_mgmt_method_get(struct srpt_port *sp, struct ib_mad *rq_mad,
413 				 struct ib_dm_mad *rsp_mad)
414 {
415 	u16 attr_id;
416 	u32 slot;
417 	u8 hi, lo;
418 
419 	attr_id = be16_to_cpu(rq_mad->mad_hdr.attr_id);
420 	switch (attr_id) {
421 	case DM_ATTR_CLASS_PORT_INFO:
422 		srpt_get_class_port_info(rsp_mad);
423 		break;
424 	case DM_ATTR_IOU_INFO:
425 		srpt_get_iou(rsp_mad);
426 		break;
427 	case DM_ATTR_IOC_PROFILE:
428 		slot = be32_to_cpu(rq_mad->mad_hdr.attr_mod);
429 		srpt_get_ioc(sp, slot, rsp_mad);
430 		break;
431 	case DM_ATTR_SVC_ENTRIES:
432 		slot = be32_to_cpu(rq_mad->mad_hdr.attr_mod);
433 		hi = (u8) ((slot >> 8) & 0xff);
434 		lo = (u8) (slot & 0xff);
435 		slot = (u16) ((slot >> 16) & 0xffff);
436 		srpt_get_svc_entries(srpt_service_guid,
437 				     slot, hi, lo, rsp_mad);
438 		break;
439 	default:
440 		rsp_mad->mad_hdr.status =
441 		    __constant_cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD_ATTR);
442 		break;
443 	}
444 }
445 
446 /**
447  * srpt_mad_send_handler() - Post MAD-send callback function.
448  */
449 static void srpt_mad_send_handler(struct ib_mad_agent *mad_agent,
450 				  struct ib_mad_send_wc *mad_wc)
451 {
452 	ib_destroy_ah(mad_wc->send_buf->ah);
453 	ib_free_send_mad(mad_wc->send_buf);
454 }
455 
456 /**
457  * srpt_mad_recv_handler() - MAD reception callback function.
458  */
459 static void srpt_mad_recv_handler(struct ib_mad_agent *mad_agent,
460 				  struct ib_mad_recv_wc *mad_wc)
461 {
462 	struct srpt_port *sport = (struct srpt_port *)mad_agent->context;
463 	struct ib_ah *ah;
464 	struct ib_mad_send_buf *rsp;
465 	struct ib_dm_mad *dm_mad;
466 
467 	if (!mad_wc || !mad_wc->recv_buf.mad)
468 		return;
469 
470 	ah = ib_create_ah_from_wc(mad_agent->qp->pd, mad_wc->wc,
471 				  mad_wc->recv_buf.grh, mad_agent->port_num);
472 	if (IS_ERR(ah))
473 		goto err;
474 
475 	BUILD_BUG_ON(offsetof(struct ib_dm_mad, data) != IB_MGMT_DEVICE_HDR);
476 
477 	rsp = ib_create_send_mad(mad_agent, mad_wc->wc->src_qp,
478 				 mad_wc->wc->pkey_index, 0,
479 				 IB_MGMT_DEVICE_HDR, IB_MGMT_DEVICE_DATA,
480 				 GFP_KERNEL,
481 				 IB_MGMT_BASE_VERSION);
482 	if (IS_ERR(rsp))
483 		goto err_rsp;
484 
485 	rsp->ah = ah;
486 
487 	dm_mad = rsp->mad;
488 	memcpy(dm_mad, mad_wc->recv_buf.mad, sizeof *dm_mad);
489 	dm_mad->mad_hdr.method = IB_MGMT_METHOD_GET_RESP;
490 	dm_mad->mad_hdr.status = 0;
491 
492 	switch (mad_wc->recv_buf.mad->mad_hdr.method) {
493 	case IB_MGMT_METHOD_GET:
494 		srpt_mgmt_method_get(sport, mad_wc->recv_buf.mad, dm_mad);
495 		break;
496 	case IB_MGMT_METHOD_SET:
497 		dm_mad->mad_hdr.status =
498 		    __constant_cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD_ATTR);
499 		break;
500 	default:
501 		dm_mad->mad_hdr.status =
502 		    __constant_cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD);
503 		break;
504 	}
505 
506 	if (!ib_post_send_mad(rsp, NULL)) {
507 		ib_free_recv_mad(mad_wc);
508 		/* will destroy_ah & free_send_mad in send completion */
509 		return;
510 	}
511 
512 	ib_free_send_mad(rsp);
513 
514 err_rsp:
515 	ib_destroy_ah(ah);
516 err:
517 	ib_free_recv_mad(mad_wc);
518 }
519 
520 /**
521  * srpt_refresh_port() - Configure a HCA port.
522  *
523  * Enable InfiniBand management datagram processing, update the cached sm_lid,
524  * lid and gid values, and register a callback function for processing MADs
525  * on the specified port.
526  *
527  * Note: It is safe to call this function more than once for the same port.
528  */
529 static int srpt_refresh_port(struct srpt_port *sport)
530 {
531 	struct ib_mad_reg_req reg_req;
532 	struct ib_port_modify port_modify;
533 	struct ib_port_attr port_attr;
534 	int ret;
535 
536 	memset(&port_modify, 0, sizeof port_modify);
537 	port_modify.set_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP;
538 	port_modify.clr_port_cap_mask = 0;
539 
540 	ret = ib_modify_port(sport->sdev->device, sport->port, 0, &port_modify);
541 	if (ret)
542 		goto err_mod_port;
543 
544 	ret = ib_query_port(sport->sdev->device, sport->port, &port_attr);
545 	if (ret)
546 		goto err_query_port;
547 
548 	sport->sm_lid = port_attr.sm_lid;
549 	sport->lid = port_attr.lid;
550 
551 	ret = ib_query_gid(sport->sdev->device, sport->port, 0, &sport->gid);
552 	if (ret)
553 		goto err_query_port;
554 
555 	if (!sport->mad_agent) {
556 		memset(&reg_req, 0, sizeof reg_req);
557 		reg_req.mgmt_class = IB_MGMT_CLASS_DEVICE_MGMT;
558 		reg_req.mgmt_class_version = IB_MGMT_BASE_VERSION;
559 		set_bit(IB_MGMT_METHOD_GET, reg_req.method_mask);
560 		set_bit(IB_MGMT_METHOD_SET, reg_req.method_mask);
561 
562 		sport->mad_agent = ib_register_mad_agent(sport->sdev->device,
563 							 sport->port,
564 							 IB_QPT_GSI,
565 							 &reg_req, 0,
566 							 srpt_mad_send_handler,
567 							 srpt_mad_recv_handler,
568 							 sport, 0);
569 		if (IS_ERR(sport->mad_agent)) {
570 			ret = PTR_ERR(sport->mad_agent);
571 			sport->mad_agent = NULL;
572 			goto err_query_port;
573 		}
574 	}
575 
576 	return 0;
577 
578 err_query_port:
579 
580 	port_modify.set_port_cap_mask = 0;
581 	port_modify.clr_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP;
582 	ib_modify_port(sport->sdev->device, sport->port, 0, &port_modify);
583 
584 err_mod_port:
585 
586 	return ret;
587 }
588 
589 /**
590  * srpt_unregister_mad_agent() - Unregister MAD callback functions.
591  *
592  * Note: It is safe to call this function more than once for the same device.
593  */
594 static void srpt_unregister_mad_agent(struct srpt_device *sdev)
595 {
596 	struct ib_port_modify port_modify = {
597 		.clr_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP,
598 	};
599 	struct srpt_port *sport;
600 	int i;
601 
602 	for (i = 1; i <= sdev->device->phys_port_cnt; i++) {
603 		sport = &sdev->port[i - 1];
604 		WARN_ON(sport->port != i);
605 		if (ib_modify_port(sdev->device, i, 0, &port_modify) < 0)
606 			pr_err("disabling MAD processing failed.\n");
607 		if (sport->mad_agent) {
608 			ib_unregister_mad_agent(sport->mad_agent);
609 			sport->mad_agent = NULL;
610 		}
611 	}
612 }
613 
614 /**
615  * srpt_alloc_ioctx() - Allocate an SRPT I/O context structure.
616  */
617 static struct srpt_ioctx *srpt_alloc_ioctx(struct srpt_device *sdev,
618 					   int ioctx_size, int dma_size,
619 					   enum dma_data_direction dir)
620 {
621 	struct srpt_ioctx *ioctx;
622 
623 	ioctx = kmalloc(ioctx_size, GFP_KERNEL);
624 	if (!ioctx)
625 		goto err;
626 
627 	ioctx->buf = kmalloc(dma_size, GFP_KERNEL);
628 	if (!ioctx->buf)
629 		goto err_free_ioctx;
630 
631 	ioctx->dma = ib_dma_map_single(sdev->device, ioctx->buf, dma_size, dir);
632 	if (ib_dma_mapping_error(sdev->device, ioctx->dma))
633 		goto err_free_buf;
634 
635 	return ioctx;
636 
637 err_free_buf:
638 	kfree(ioctx->buf);
639 err_free_ioctx:
640 	kfree(ioctx);
641 err:
642 	return NULL;
643 }
644 
645 /**
646  * srpt_free_ioctx() - Free an SRPT I/O context structure.
647  */
648 static void srpt_free_ioctx(struct srpt_device *sdev, struct srpt_ioctx *ioctx,
649 			    int dma_size, enum dma_data_direction dir)
650 {
651 	if (!ioctx)
652 		return;
653 
654 	ib_dma_unmap_single(sdev->device, ioctx->dma, dma_size, dir);
655 	kfree(ioctx->buf);
656 	kfree(ioctx);
657 }
658 
659 /**
660  * srpt_alloc_ioctx_ring() - Allocate a ring of SRPT I/O context structures.
661  * @sdev:       Device to allocate the I/O context ring for.
662  * @ring_size:  Number of elements in the I/O context ring.
663  * @ioctx_size: I/O context size.
664  * @dma_size:   DMA buffer size.
665  * @dir:        DMA data direction.
666  */
667 static struct srpt_ioctx **srpt_alloc_ioctx_ring(struct srpt_device *sdev,
668 				int ring_size, int ioctx_size,
669 				int dma_size, enum dma_data_direction dir)
670 {
671 	struct srpt_ioctx **ring;
672 	int i;
673 
674 	WARN_ON(ioctx_size != sizeof(struct srpt_recv_ioctx)
675 		&& ioctx_size != sizeof(struct srpt_send_ioctx));
676 
677 	ring = kmalloc(ring_size * sizeof(ring[0]), GFP_KERNEL);
678 	if (!ring)
679 		goto out;
680 	for (i = 0; i < ring_size; ++i) {
681 		ring[i] = srpt_alloc_ioctx(sdev, ioctx_size, dma_size, dir);
682 		if (!ring[i])
683 			goto err;
684 		ring[i]->index = i;
685 	}
686 	goto out;
687 
688 err:
689 	while (--i >= 0)
690 		srpt_free_ioctx(sdev, ring[i], dma_size, dir);
691 	kfree(ring);
692 	ring = NULL;
693 out:
694 	return ring;
695 }
696 
697 /**
698  * srpt_free_ioctx_ring() - Free the ring of SRPT I/O context structures.
699  */
700 static void srpt_free_ioctx_ring(struct srpt_ioctx **ioctx_ring,
701 				 struct srpt_device *sdev, int ring_size,
702 				 int dma_size, enum dma_data_direction dir)
703 {
704 	int i;
705 
706 	for (i = 0; i < ring_size; ++i)
707 		srpt_free_ioctx(sdev, ioctx_ring[i], dma_size, dir);
708 	kfree(ioctx_ring);
709 }
710 
711 /**
712  * srpt_get_cmd_state() - Get the state of a SCSI command.
713  */
714 static enum srpt_command_state srpt_get_cmd_state(struct srpt_send_ioctx *ioctx)
715 {
716 	enum srpt_command_state state;
717 	unsigned long flags;
718 
719 	BUG_ON(!ioctx);
720 
721 	spin_lock_irqsave(&ioctx->spinlock, flags);
722 	state = ioctx->state;
723 	spin_unlock_irqrestore(&ioctx->spinlock, flags);
724 	return state;
725 }
726 
727 /**
728  * srpt_set_cmd_state() - Set the state of a SCSI command.
729  *
730  * Does not modify the state of aborted commands. Returns the previous command
731  * state.
732  */
733 static enum srpt_command_state srpt_set_cmd_state(struct srpt_send_ioctx *ioctx,
734 						  enum srpt_command_state new)
735 {
736 	enum srpt_command_state previous;
737 	unsigned long flags;
738 
739 	BUG_ON(!ioctx);
740 
741 	spin_lock_irqsave(&ioctx->spinlock, flags);
742 	previous = ioctx->state;
743 	if (previous != SRPT_STATE_DONE)
744 		ioctx->state = new;
745 	spin_unlock_irqrestore(&ioctx->spinlock, flags);
746 
747 	return previous;
748 }
749 
750 /**
751  * srpt_test_and_set_cmd_state() - Test and set the state of a command.
752  *
753  * Returns true if and only if the previous command state was equal to 'old'.
754  */
755 static bool srpt_test_and_set_cmd_state(struct srpt_send_ioctx *ioctx,
756 					enum srpt_command_state old,
757 					enum srpt_command_state new)
758 {
759 	enum srpt_command_state previous;
760 	unsigned long flags;
761 
762 	WARN_ON(!ioctx);
763 	WARN_ON(old == SRPT_STATE_DONE);
764 	WARN_ON(new == SRPT_STATE_NEW);
765 
766 	spin_lock_irqsave(&ioctx->spinlock, flags);
767 	previous = ioctx->state;
768 	if (previous == old)
769 		ioctx->state = new;
770 	spin_unlock_irqrestore(&ioctx->spinlock, flags);
771 	return previous == old;
772 }
773 
774 /**
775  * srpt_post_recv() - Post an IB receive request.
776  */
777 static int srpt_post_recv(struct srpt_device *sdev,
778 			  struct srpt_recv_ioctx *ioctx)
779 {
780 	struct ib_sge list;
781 	struct ib_recv_wr wr, *bad_wr;
782 
783 	BUG_ON(!sdev);
784 	wr.wr_id = encode_wr_id(SRPT_RECV, ioctx->ioctx.index);
785 
786 	list.addr = ioctx->ioctx.dma;
787 	list.length = srp_max_req_size;
788 	list.lkey = sdev->mr->lkey;
789 
790 	wr.next = NULL;
791 	wr.sg_list = &list;
792 	wr.num_sge = 1;
793 
794 	return ib_post_srq_recv(sdev->srq, &wr, &bad_wr);
795 }
796 
797 /**
798  * srpt_post_send() - Post an IB send request.
799  *
800  * Returns zero upon success and a non-zero value upon failure.
801  */
802 static int srpt_post_send(struct srpt_rdma_ch *ch,
803 			  struct srpt_send_ioctx *ioctx, int len)
804 {
805 	struct ib_sge list;
806 	struct ib_send_wr wr, *bad_wr;
807 	struct srpt_device *sdev = ch->sport->sdev;
808 	int ret;
809 
810 	atomic_inc(&ch->req_lim);
811 
812 	ret = -ENOMEM;
813 	if (unlikely(atomic_dec_return(&ch->sq_wr_avail) < 0)) {
814 		pr_warn("IB send queue full (needed 1)\n");
815 		goto out;
816 	}
817 
818 	ib_dma_sync_single_for_device(sdev->device, ioctx->ioctx.dma, len,
819 				      DMA_TO_DEVICE);
820 
821 	list.addr = ioctx->ioctx.dma;
822 	list.length = len;
823 	list.lkey = sdev->mr->lkey;
824 
825 	wr.next = NULL;
826 	wr.wr_id = encode_wr_id(SRPT_SEND, ioctx->ioctx.index);
827 	wr.sg_list = &list;
828 	wr.num_sge = 1;
829 	wr.opcode = IB_WR_SEND;
830 	wr.send_flags = IB_SEND_SIGNALED;
831 
832 	ret = ib_post_send(ch->qp, &wr, &bad_wr);
833 
834 out:
835 	if (ret < 0) {
836 		atomic_inc(&ch->sq_wr_avail);
837 		atomic_dec(&ch->req_lim);
838 	}
839 	return ret;
840 }
841 
842 /**
843  * srpt_get_desc_tbl() - Parse the data descriptors of an SRP_CMD request.
844  * @ioctx: Pointer to the I/O context associated with the request.
845  * @srp_cmd: Pointer to the SRP_CMD request data.
846  * @dir: Pointer to the variable to which the transfer direction will be
847  *   written.
848  * @data_len: Pointer to the variable to which the total data length of all
849  *   descriptors in the SRP_CMD request will be written.
850  *
851  * This function initializes ioctx->nrbuf and ioctx->r_bufs.
852  *
853  * Returns -EINVAL when the SRP_CMD request contains inconsistent descriptors;
854  * -ENOMEM when memory allocation fails and zero upon success.
855  */
856 static int srpt_get_desc_tbl(struct srpt_send_ioctx *ioctx,
857 			     struct srp_cmd *srp_cmd,
858 			     enum dma_data_direction *dir, u64 *data_len)
859 {
860 	struct srp_indirect_buf *idb;
861 	struct srp_direct_buf *db;
862 	unsigned add_cdb_offset;
863 	int ret;
864 
865 	/*
866 	 * The pointer computations below will only be compiled correctly
867 	 * if srp_cmd::add_data is declared as s8*, u8*, s8[] or u8[], so check
868 	 * whether srp_cmd::add_data has been declared as a byte pointer.
869 	 */
870 	BUILD_BUG_ON(!__same_type(srp_cmd->add_data[0], (s8)0)
871 		     && !__same_type(srp_cmd->add_data[0], (u8)0));
872 
873 	BUG_ON(!dir);
874 	BUG_ON(!data_len);
875 
876 	ret = 0;
877 	*data_len = 0;
878 
879 	/*
880 	 * The lower four bits of the buffer format field contain the DATA-IN
881 	 * buffer descriptor format, and the highest four bits contain the
882 	 * DATA-OUT buffer descriptor format.
883 	 */
884 	*dir = DMA_NONE;
885 	if (srp_cmd->buf_fmt & 0xf)
886 		/* DATA-IN: transfer data from target to initiator (read). */
887 		*dir = DMA_FROM_DEVICE;
888 	else if (srp_cmd->buf_fmt >> 4)
889 		/* DATA-OUT: transfer data from initiator to target (write). */
890 		*dir = DMA_TO_DEVICE;
891 
892 	/*
893 	 * According to the SRP spec, the lower two bits of the 'ADDITIONAL
894 	 * CDB LENGTH' field are reserved and the size in bytes of this field
895 	 * is four times the value specified in bits 3..7. Hence the "& ~3".
896 	 */
897 	add_cdb_offset = srp_cmd->add_cdb_len & ~3;
898 	if (((srp_cmd->buf_fmt & 0xf) == SRP_DATA_DESC_DIRECT) ||
899 	    ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_DIRECT)) {
900 		ioctx->n_rbuf = 1;
901 		ioctx->rbufs = &ioctx->single_rbuf;
902 
903 		db = (struct srp_direct_buf *)(srp_cmd->add_data
904 					       + add_cdb_offset);
905 		memcpy(ioctx->rbufs, db, sizeof *db);
906 		*data_len = be32_to_cpu(db->len);
907 	} else if (((srp_cmd->buf_fmt & 0xf) == SRP_DATA_DESC_INDIRECT) ||
908 		   ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_INDIRECT)) {
909 		idb = (struct srp_indirect_buf *)(srp_cmd->add_data
910 						  + add_cdb_offset);
911 
912 		ioctx->n_rbuf = be32_to_cpu(idb->table_desc.len) / sizeof *db;
913 
914 		if (ioctx->n_rbuf >
915 		    (srp_cmd->data_out_desc_cnt + srp_cmd->data_in_desc_cnt)) {
916 			pr_err("received unsupported SRP_CMD request"
917 			       " type (%u out + %u in != %u / %zu)\n",
918 			       srp_cmd->data_out_desc_cnt,
919 			       srp_cmd->data_in_desc_cnt,
920 			       be32_to_cpu(idb->table_desc.len),
921 			       sizeof(*db));
922 			ioctx->n_rbuf = 0;
923 			ret = -EINVAL;
924 			goto out;
925 		}
926 
927 		if (ioctx->n_rbuf == 1)
928 			ioctx->rbufs = &ioctx->single_rbuf;
929 		else {
930 			ioctx->rbufs =
931 				kmalloc(ioctx->n_rbuf * sizeof *db, GFP_ATOMIC);
932 			if (!ioctx->rbufs) {
933 				ioctx->n_rbuf = 0;
934 				ret = -ENOMEM;
935 				goto out;
936 			}
937 		}
938 
939 		db = idb->desc_list;
940 		memcpy(ioctx->rbufs, db, ioctx->n_rbuf * sizeof *db);
941 		*data_len = be32_to_cpu(idb->len);
942 	}
943 out:
944 	return ret;
945 }
946 
947 /**
948  * srpt_init_ch_qp() - Initialize queue pair attributes.
949  *
950  * Initialized the attributes of queue pair 'qp' by allowing local write,
951  * remote read and remote write. Also transitions 'qp' to state IB_QPS_INIT.
952  */
953 static int srpt_init_ch_qp(struct srpt_rdma_ch *ch, struct ib_qp *qp)
954 {
955 	struct ib_qp_attr *attr;
956 	int ret;
957 
958 	attr = kzalloc(sizeof *attr, GFP_KERNEL);
959 	if (!attr)
960 		return -ENOMEM;
961 
962 	attr->qp_state = IB_QPS_INIT;
963 	attr->qp_access_flags = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_READ |
964 	    IB_ACCESS_REMOTE_WRITE;
965 	attr->port_num = ch->sport->port;
966 	attr->pkey_index = 0;
967 
968 	ret = ib_modify_qp(qp, attr,
969 			   IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PORT |
970 			   IB_QP_PKEY_INDEX);
971 
972 	kfree(attr);
973 	return ret;
974 }
975 
976 /**
977  * srpt_ch_qp_rtr() - Change the state of a channel to 'ready to receive' (RTR).
978  * @ch: channel of the queue pair.
979  * @qp: queue pair to change the state of.
980  *
981  * Returns zero upon success and a negative value upon failure.
982  *
983  * Note: currently a struct ib_qp_attr takes 136 bytes on a 64-bit system.
984  * If this structure ever becomes larger, it might be necessary to allocate
985  * it dynamically instead of on the stack.
986  */
987 static int srpt_ch_qp_rtr(struct srpt_rdma_ch *ch, struct ib_qp *qp)
988 {
989 	struct ib_qp_attr qp_attr;
990 	int attr_mask;
991 	int ret;
992 
993 	qp_attr.qp_state = IB_QPS_RTR;
994 	ret = ib_cm_init_qp_attr(ch->cm_id, &qp_attr, &attr_mask);
995 	if (ret)
996 		goto out;
997 
998 	qp_attr.max_dest_rd_atomic = 4;
999 
1000 	ret = ib_modify_qp(qp, &qp_attr, attr_mask);
1001 
1002 out:
1003 	return ret;
1004 }
1005 
1006 /**
1007  * srpt_ch_qp_rts() - Change the state of a channel to 'ready to send' (RTS).
1008  * @ch: channel of the queue pair.
1009  * @qp: queue pair to change the state of.
1010  *
1011  * Returns zero upon success and a negative value upon failure.
1012  *
1013  * Note: currently a struct ib_qp_attr takes 136 bytes on a 64-bit system.
1014  * If this structure ever becomes larger, it might be necessary to allocate
1015  * it dynamically instead of on the stack.
1016  */
1017 static int srpt_ch_qp_rts(struct srpt_rdma_ch *ch, struct ib_qp *qp)
1018 {
1019 	struct ib_qp_attr qp_attr;
1020 	int attr_mask;
1021 	int ret;
1022 
1023 	qp_attr.qp_state = IB_QPS_RTS;
1024 	ret = ib_cm_init_qp_attr(ch->cm_id, &qp_attr, &attr_mask);
1025 	if (ret)
1026 		goto out;
1027 
1028 	qp_attr.max_rd_atomic = 4;
1029 
1030 	ret = ib_modify_qp(qp, &qp_attr, attr_mask);
1031 
1032 out:
1033 	return ret;
1034 }
1035 
1036 /**
1037  * srpt_ch_qp_err() - Set the channel queue pair state to 'error'.
1038  */
1039 static int srpt_ch_qp_err(struct srpt_rdma_ch *ch)
1040 {
1041 	struct ib_qp_attr qp_attr;
1042 
1043 	qp_attr.qp_state = IB_QPS_ERR;
1044 	return ib_modify_qp(ch->qp, &qp_attr, IB_QP_STATE);
1045 }
1046 
1047 /**
1048  * srpt_unmap_sg_to_ib_sge() - Unmap an IB SGE list.
1049  */
1050 static void srpt_unmap_sg_to_ib_sge(struct srpt_rdma_ch *ch,
1051 				    struct srpt_send_ioctx *ioctx)
1052 {
1053 	struct scatterlist *sg;
1054 	enum dma_data_direction dir;
1055 
1056 	BUG_ON(!ch);
1057 	BUG_ON(!ioctx);
1058 	BUG_ON(ioctx->n_rdma && !ioctx->rdma_ius);
1059 
1060 	while (ioctx->n_rdma)
1061 		kfree(ioctx->rdma_ius[--ioctx->n_rdma].sge);
1062 
1063 	kfree(ioctx->rdma_ius);
1064 	ioctx->rdma_ius = NULL;
1065 
1066 	if (ioctx->mapped_sg_count) {
1067 		sg = ioctx->sg;
1068 		WARN_ON(!sg);
1069 		dir = ioctx->cmd.data_direction;
1070 		BUG_ON(dir == DMA_NONE);
1071 		ib_dma_unmap_sg(ch->sport->sdev->device, sg, ioctx->sg_cnt,
1072 				opposite_dma_dir(dir));
1073 		ioctx->mapped_sg_count = 0;
1074 	}
1075 }
1076 
1077 /**
1078  * srpt_map_sg_to_ib_sge() - Map an SG list to an IB SGE list.
1079  */
1080 static int srpt_map_sg_to_ib_sge(struct srpt_rdma_ch *ch,
1081 				 struct srpt_send_ioctx *ioctx)
1082 {
1083 	struct ib_device *dev = ch->sport->sdev->device;
1084 	struct se_cmd *cmd;
1085 	struct scatterlist *sg, *sg_orig;
1086 	int sg_cnt;
1087 	enum dma_data_direction dir;
1088 	struct rdma_iu *riu;
1089 	struct srp_direct_buf *db;
1090 	dma_addr_t dma_addr;
1091 	struct ib_sge *sge;
1092 	u64 raddr;
1093 	u32 rsize;
1094 	u32 tsize;
1095 	u32 dma_len;
1096 	int count, nrdma;
1097 	int i, j, k;
1098 
1099 	BUG_ON(!ch);
1100 	BUG_ON(!ioctx);
1101 	cmd = &ioctx->cmd;
1102 	dir = cmd->data_direction;
1103 	BUG_ON(dir == DMA_NONE);
1104 
1105 	ioctx->sg = sg = sg_orig = cmd->t_data_sg;
1106 	ioctx->sg_cnt = sg_cnt = cmd->t_data_nents;
1107 
1108 	count = ib_dma_map_sg(ch->sport->sdev->device, sg, sg_cnt,
1109 			      opposite_dma_dir(dir));
1110 	if (unlikely(!count))
1111 		return -EAGAIN;
1112 
1113 	ioctx->mapped_sg_count = count;
1114 
1115 	if (ioctx->rdma_ius && ioctx->n_rdma_ius)
1116 		nrdma = ioctx->n_rdma_ius;
1117 	else {
1118 		nrdma = (count + SRPT_DEF_SG_PER_WQE - 1) / SRPT_DEF_SG_PER_WQE
1119 			+ ioctx->n_rbuf;
1120 
1121 		ioctx->rdma_ius = kzalloc(nrdma * sizeof *riu, GFP_KERNEL);
1122 		if (!ioctx->rdma_ius)
1123 			goto free_mem;
1124 
1125 		ioctx->n_rdma_ius = nrdma;
1126 	}
1127 
1128 	db = ioctx->rbufs;
1129 	tsize = cmd->data_length;
1130 	dma_len = ib_sg_dma_len(dev, &sg[0]);
1131 	riu = ioctx->rdma_ius;
1132 
1133 	/*
1134 	 * For each remote desc - calculate the #ib_sge.
1135 	 * If #ib_sge < SRPT_DEF_SG_PER_WQE per rdma operation then
1136 	 *      each remote desc rdma_iu is required a rdma wr;
1137 	 * else
1138 	 *      we need to allocate extra rdma_iu to carry extra #ib_sge in
1139 	 *      another rdma wr
1140 	 */
1141 	for (i = 0, j = 0;
1142 	     j < count && i < ioctx->n_rbuf && tsize > 0; ++i, ++riu, ++db) {
1143 		rsize = be32_to_cpu(db->len);
1144 		raddr = be64_to_cpu(db->va);
1145 		riu->raddr = raddr;
1146 		riu->rkey = be32_to_cpu(db->key);
1147 		riu->sge_cnt = 0;
1148 
1149 		/* calculate how many sge required for this remote_buf */
1150 		while (rsize > 0 && tsize > 0) {
1151 
1152 			if (rsize >= dma_len) {
1153 				tsize -= dma_len;
1154 				rsize -= dma_len;
1155 				raddr += dma_len;
1156 
1157 				if (tsize > 0) {
1158 					++j;
1159 					if (j < count) {
1160 						sg = sg_next(sg);
1161 						dma_len = ib_sg_dma_len(
1162 								dev, sg);
1163 					}
1164 				}
1165 			} else {
1166 				tsize -= rsize;
1167 				dma_len -= rsize;
1168 				rsize = 0;
1169 			}
1170 
1171 			++riu->sge_cnt;
1172 
1173 			if (rsize > 0 && riu->sge_cnt == SRPT_DEF_SG_PER_WQE) {
1174 				++ioctx->n_rdma;
1175 				riu->sge =
1176 				    kmalloc(riu->sge_cnt * sizeof *riu->sge,
1177 					    GFP_KERNEL);
1178 				if (!riu->sge)
1179 					goto free_mem;
1180 
1181 				++riu;
1182 				riu->sge_cnt = 0;
1183 				riu->raddr = raddr;
1184 				riu->rkey = be32_to_cpu(db->key);
1185 			}
1186 		}
1187 
1188 		++ioctx->n_rdma;
1189 		riu->sge = kmalloc(riu->sge_cnt * sizeof *riu->sge,
1190 				   GFP_KERNEL);
1191 		if (!riu->sge)
1192 			goto free_mem;
1193 	}
1194 
1195 	db = ioctx->rbufs;
1196 	tsize = cmd->data_length;
1197 	riu = ioctx->rdma_ius;
1198 	sg = sg_orig;
1199 	dma_len = ib_sg_dma_len(dev, &sg[0]);
1200 	dma_addr = ib_sg_dma_address(dev, &sg[0]);
1201 
1202 	/* this second loop is really mapped sg_addres to rdma_iu->ib_sge */
1203 	for (i = 0, j = 0;
1204 	     j < count && i < ioctx->n_rbuf && tsize > 0; ++i, ++riu, ++db) {
1205 		rsize = be32_to_cpu(db->len);
1206 		sge = riu->sge;
1207 		k = 0;
1208 
1209 		while (rsize > 0 && tsize > 0) {
1210 			sge->addr = dma_addr;
1211 			sge->lkey = ch->sport->sdev->mr->lkey;
1212 
1213 			if (rsize >= dma_len) {
1214 				sge->length =
1215 					(tsize < dma_len) ? tsize : dma_len;
1216 				tsize -= dma_len;
1217 				rsize -= dma_len;
1218 
1219 				if (tsize > 0) {
1220 					++j;
1221 					if (j < count) {
1222 						sg = sg_next(sg);
1223 						dma_len = ib_sg_dma_len(
1224 								dev, sg);
1225 						dma_addr = ib_sg_dma_address(
1226 								dev, sg);
1227 					}
1228 				}
1229 			} else {
1230 				sge->length = (tsize < rsize) ? tsize : rsize;
1231 				tsize -= rsize;
1232 				dma_len -= rsize;
1233 				dma_addr += rsize;
1234 				rsize = 0;
1235 			}
1236 
1237 			++k;
1238 			if (k == riu->sge_cnt && rsize > 0 && tsize > 0) {
1239 				++riu;
1240 				sge = riu->sge;
1241 				k = 0;
1242 			} else if (rsize > 0 && tsize > 0)
1243 				++sge;
1244 		}
1245 	}
1246 
1247 	return 0;
1248 
1249 free_mem:
1250 	srpt_unmap_sg_to_ib_sge(ch, ioctx);
1251 
1252 	return -ENOMEM;
1253 }
1254 
1255 /**
1256  * srpt_get_send_ioctx() - Obtain an I/O context for sending to the initiator.
1257  */
1258 static struct srpt_send_ioctx *srpt_get_send_ioctx(struct srpt_rdma_ch *ch)
1259 {
1260 	struct srpt_send_ioctx *ioctx;
1261 	unsigned long flags;
1262 
1263 	BUG_ON(!ch);
1264 
1265 	ioctx = NULL;
1266 	spin_lock_irqsave(&ch->spinlock, flags);
1267 	if (!list_empty(&ch->free_list)) {
1268 		ioctx = list_first_entry(&ch->free_list,
1269 					 struct srpt_send_ioctx, free_list);
1270 		list_del(&ioctx->free_list);
1271 	}
1272 	spin_unlock_irqrestore(&ch->spinlock, flags);
1273 
1274 	if (!ioctx)
1275 		return ioctx;
1276 
1277 	BUG_ON(ioctx->ch != ch);
1278 	spin_lock_init(&ioctx->spinlock);
1279 	ioctx->state = SRPT_STATE_NEW;
1280 	ioctx->n_rbuf = 0;
1281 	ioctx->rbufs = NULL;
1282 	ioctx->n_rdma = 0;
1283 	ioctx->n_rdma_ius = 0;
1284 	ioctx->rdma_ius = NULL;
1285 	ioctx->mapped_sg_count = 0;
1286 	init_completion(&ioctx->tx_done);
1287 	ioctx->queue_status_only = false;
1288 	/*
1289 	 * transport_init_se_cmd() does not initialize all fields, so do it
1290 	 * here.
1291 	 */
1292 	memset(&ioctx->cmd, 0, sizeof(ioctx->cmd));
1293 	memset(&ioctx->sense_data, 0, sizeof(ioctx->sense_data));
1294 
1295 	return ioctx;
1296 }
1297 
1298 /**
1299  * srpt_abort_cmd() - Abort a SCSI command.
1300  * @ioctx:   I/O context associated with the SCSI command.
1301  * @context: Preferred execution context.
1302  */
1303 static int srpt_abort_cmd(struct srpt_send_ioctx *ioctx)
1304 {
1305 	enum srpt_command_state state;
1306 	unsigned long flags;
1307 
1308 	BUG_ON(!ioctx);
1309 
1310 	/*
1311 	 * If the command is in a state where the target core is waiting for
1312 	 * the ib_srpt driver, change the state to the next state. Changing
1313 	 * the state of the command from SRPT_STATE_NEED_DATA to
1314 	 * SRPT_STATE_DATA_IN ensures that srpt_xmit_response() will call this
1315 	 * function a second time.
1316 	 */
1317 
1318 	spin_lock_irqsave(&ioctx->spinlock, flags);
1319 	state = ioctx->state;
1320 	switch (state) {
1321 	case SRPT_STATE_NEED_DATA:
1322 		ioctx->state = SRPT_STATE_DATA_IN;
1323 		break;
1324 	case SRPT_STATE_DATA_IN:
1325 	case SRPT_STATE_CMD_RSP_SENT:
1326 	case SRPT_STATE_MGMT_RSP_SENT:
1327 		ioctx->state = SRPT_STATE_DONE;
1328 		break;
1329 	default:
1330 		break;
1331 	}
1332 	spin_unlock_irqrestore(&ioctx->spinlock, flags);
1333 
1334 	if (state == SRPT_STATE_DONE) {
1335 		struct srpt_rdma_ch *ch = ioctx->ch;
1336 
1337 		BUG_ON(ch->sess == NULL);
1338 
1339 		target_put_sess_cmd(ch->sess, &ioctx->cmd);
1340 		goto out;
1341 	}
1342 
1343 	pr_debug("Aborting cmd with state %d and tag %lld\n", state,
1344 		 ioctx->tag);
1345 
1346 	switch (state) {
1347 	case SRPT_STATE_NEW:
1348 	case SRPT_STATE_DATA_IN:
1349 	case SRPT_STATE_MGMT:
1350 		/*
1351 		 * Do nothing - defer abort processing until
1352 		 * srpt_queue_response() is invoked.
1353 		 */
1354 		WARN_ON(!transport_check_aborted_status(&ioctx->cmd, false));
1355 		break;
1356 	case SRPT_STATE_NEED_DATA:
1357 		/* DMA_TO_DEVICE (write) - RDMA read error. */
1358 
1359 		/* XXX(hch): this is a horrible layering violation.. */
1360 		spin_lock_irqsave(&ioctx->cmd.t_state_lock, flags);
1361 		ioctx->cmd.transport_state &= ~CMD_T_ACTIVE;
1362 		spin_unlock_irqrestore(&ioctx->cmd.t_state_lock, flags);
1363 		break;
1364 	case SRPT_STATE_CMD_RSP_SENT:
1365 		/*
1366 		 * SRP_RSP sending failed or the SRP_RSP send completion has
1367 		 * not been received in time.
1368 		 */
1369 		srpt_unmap_sg_to_ib_sge(ioctx->ch, ioctx);
1370 		target_put_sess_cmd(ioctx->ch->sess, &ioctx->cmd);
1371 		break;
1372 	case SRPT_STATE_MGMT_RSP_SENT:
1373 		srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
1374 		target_put_sess_cmd(ioctx->ch->sess, &ioctx->cmd);
1375 		break;
1376 	default:
1377 		WARN(1, "Unexpected command state (%d)", state);
1378 		break;
1379 	}
1380 
1381 out:
1382 	return state;
1383 }
1384 
1385 /**
1386  * srpt_handle_send_err_comp() - Process an IB_WC_SEND error completion.
1387  */
1388 static void srpt_handle_send_err_comp(struct srpt_rdma_ch *ch, u64 wr_id)
1389 {
1390 	struct srpt_send_ioctx *ioctx;
1391 	enum srpt_command_state state;
1392 	struct se_cmd *cmd;
1393 	u32 index;
1394 
1395 	atomic_inc(&ch->sq_wr_avail);
1396 
1397 	index = idx_from_wr_id(wr_id);
1398 	ioctx = ch->ioctx_ring[index];
1399 	state = srpt_get_cmd_state(ioctx);
1400 	cmd = &ioctx->cmd;
1401 
1402 	WARN_ON(state != SRPT_STATE_CMD_RSP_SENT
1403 		&& state != SRPT_STATE_MGMT_RSP_SENT
1404 		&& state != SRPT_STATE_NEED_DATA
1405 		&& state != SRPT_STATE_DONE);
1406 
1407 	/* If SRP_RSP sending failed, undo the ch->req_lim change. */
1408 	if (state == SRPT_STATE_CMD_RSP_SENT
1409 	    || state == SRPT_STATE_MGMT_RSP_SENT)
1410 		atomic_dec(&ch->req_lim);
1411 
1412 	srpt_abort_cmd(ioctx);
1413 }
1414 
1415 /**
1416  * srpt_handle_send_comp() - Process an IB send completion notification.
1417  */
1418 static void srpt_handle_send_comp(struct srpt_rdma_ch *ch,
1419 				  struct srpt_send_ioctx *ioctx)
1420 {
1421 	enum srpt_command_state state;
1422 
1423 	atomic_inc(&ch->sq_wr_avail);
1424 
1425 	state = srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
1426 
1427 	if (WARN_ON(state != SRPT_STATE_CMD_RSP_SENT
1428 		    && state != SRPT_STATE_MGMT_RSP_SENT
1429 		    && state != SRPT_STATE_DONE))
1430 		pr_debug("state = %d\n", state);
1431 
1432 	if (state != SRPT_STATE_DONE) {
1433 		srpt_unmap_sg_to_ib_sge(ch, ioctx);
1434 		transport_generic_free_cmd(&ioctx->cmd, 0);
1435 	} else {
1436 		pr_err("IB completion has been received too late for"
1437 		       " wr_id = %u.\n", ioctx->ioctx.index);
1438 	}
1439 }
1440 
1441 /**
1442  * srpt_handle_rdma_comp() - Process an IB RDMA completion notification.
1443  *
1444  * XXX: what is now target_execute_cmd used to be asynchronous, and unmapping
1445  * the data that has been transferred via IB RDMA had to be postponed until the
1446  * check_stop_free() callback.  None of this is necessary anymore and needs to
1447  * be cleaned up.
1448  */
1449 static void srpt_handle_rdma_comp(struct srpt_rdma_ch *ch,
1450 				  struct srpt_send_ioctx *ioctx,
1451 				  enum srpt_opcode opcode)
1452 {
1453 	WARN_ON(ioctx->n_rdma <= 0);
1454 	atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
1455 
1456 	if (opcode == SRPT_RDMA_READ_LAST) {
1457 		if (srpt_test_and_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA,
1458 						SRPT_STATE_DATA_IN))
1459 			target_execute_cmd(&ioctx->cmd);
1460 		else
1461 			pr_err("%s[%d]: wrong state = %d\n", __func__,
1462 			       __LINE__, srpt_get_cmd_state(ioctx));
1463 	} else if (opcode == SRPT_RDMA_ABORT) {
1464 		ioctx->rdma_aborted = true;
1465 	} else {
1466 		WARN(true, "unexpected opcode %d\n", opcode);
1467 	}
1468 }
1469 
1470 /**
1471  * srpt_handle_rdma_err_comp() - Process an IB RDMA error completion.
1472  */
1473 static void srpt_handle_rdma_err_comp(struct srpt_rdma_ch *ch,
1474 				      struct srpt_send_ioctx *ioctx,
1475 				      enum srpt_opcode opcode)
1476 {
1477 	struct se_cmd *cmd;
1478 	enum srpt_command_state state;
1479 
1480 	cmd = &ioctx->cmd;
1481 	state = srpt_get_cmd_state(ioctx);
1482 	switch (opcode) {
1483 	case SRPT_RDMA_READ_LAST:
1484 		if (ioctx->n_rdma <= 0) {
1485 			pr_err("Received invalid RDMA read"
1486 			       " error completion with idx %d\n",
1487 			       ioctx->ioctx.index);
1488 			break;
1489 		}
1490 		atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
1491 		if (state == SRPT_STATE_NEED_DATA)
1492 			srpt_abort_cmd(ioctx);
1493 		else
1494 			pr_err("%s[%d]: wrong state = %d\n",
1495 			       __func__, __LINE__, state);
1496 		break;
1497 	case SRPT_RDMA_WRITE_LAST:
1498 		break;
1499 	default:
1500 		pr_err("%s[%d]: opcode = %u\n", __func__, __LINE__, opcode);
1501 		break;
1502 	}
1503 }
1504 
1505 /**
1506  * srpt_build_cmd_rsp() - Build an SRP_RSP response.
1507  * @ch: RDMA channel through which the request has been received.
1508  * @ioctx: I/O context associated with the SRP_CMD request. The response will
1509  *   be built in the buffer ioctx->buf points at and hence this function will
1510  *   overwrite the request data.
1511  * @tag: tag of the request for which this response is being generated.
1512  * @status: value for the STATUS field of the SRP_RSP information unit.
1513  *
1514  * Returns the size in bytes of the SRP_RSP response.
1515  *
1516  * An SRP_RSP response contains a SCSI status or service response. See also
1517  * section 6.9 in the SRP r16a document for the format of an SRP_RSP
1518  * response. See also SPC-2 for more information about sense data.
1519  */
1520 static int srpt_build_cmd_rsp(struct srpt_rdma_ch *ch,
1521 			      struct srpt_send_ioctx *ioctx, u64 tag,
1522 			      int status)
1523 {
1524 	struct srp_rsp *srp_rsp;
1525 	const u8 *sense_data;
1526 	int sense_data_len, max_sense_len;
1527 
1528 	/*
1529 	 * The lowest bit of all SAM-3 status codes is zero (see also
1530 	 * paragraph 5.3 in SAM-3).
1531 	 */
1532 	WARN_ON(status & 1);
1533 
1534 	srp_rsp = ioctx->ioctx.buf;
1535 	BUG_ON(!srp_rsp);
1536 
1537 	sense_data = ioctx->sense_data;
1538 	sense_data_len = ioctx->cmd.scsi_sense_length;
1539 	WARN_ON(sense_data_len > sizeof(ioctx->sense_data));
1540 
1541 	memset(srp_rsp, 0, sizeof *srp_rsp);
1542 	srp_rsp->opcode = SRP_RSP;
1543 	srp_rsp->req_lim_delta =
1544 		__constant_cpu_to_be32(1 + atomic_xchg(&ch->req_lim_delta, 0));
1545 	srp_rsp->tag = tag;
1546 	srp_rsp->status = status;
1547 
1548 	if (sense_data_len) {
1549 		BUILD_BUG_ON(MIN_MAX_RSP_SIZE <= sizeof(*srp_rsp));
1550 		max_sense_len = ch->max_ti_iu_len - sizeof(*srp_rsp);
1551 		if (sense_data_len > max_sense_len) {
1552 			pr_warn("truncated sense data from %d to %d"
1553 				" bytes\n", sense_data_len, max_sense_len);
1554 			sense_data_len = max_sense_len;
1555 		}
1556 
1557 		srp_rsp->flags |= SRP_RSP_FLAG_SNSVALID;
1558 		srp_rsp->sense_data_len = cpu_to_be32(sense_data_len);
1559 		memcpy(srp_rsp + 1, sense_data, sense_data_len);
1560 	}
1561 
1562 	return sizeof(*srp_rsp) + sense_data_len;
1563 }
1564 
1565 /**
1566  * srpt_build_tskmgmt_rsp() - Build a task management response.
1567  * @ch:       RDMA channel through which the request has been received.
1568  * @ioctx:    I/O context in which the SRP_RSP response will be built.
1569  * @rsp_code: RSP_CODE that will be stored in the response.
1570  * @tag:      Tag of the request for which this response is being generated.
1571  *
1572  * Returns the size in bytes of the SRP_RSP response.
1573  *
1574  * An SRP_RSP response contains a SCSI status or service response. See also
1575  * section 6.9 in the SRP r16a document for the format of an SRP_RSP
1576  * response.
1577  */
1578 static int srpt_build_tskmgmt_rsp(struct srpt_rdma_ch *ch,
1579 				  struct srpt_send_ioctx *ioctx,
1580 				  u8 rsp_code, u64 tag)
1581 {
1582 	struct srp_rsp *srp_rsp;
1583 	int resp_data_len;
1584 	int resp_len;
1585 
1586 	resp_data_len = 4;
1587 	resp_len = sizeof(*srp_rsp) + resp_data_len;
1588 
1589 	srp_rsp = ioctx->ioctx.buf;
1590 	BUG_ON(!srp_rsp);
1591 	memset(srp_rsp, 0, sizeof *srp_rsp);
1592 
1593 	srp_rsp->opcode = SRP_RSP;
1594 	srp_rsp->req_lim_delta = __constant_cpu_to_be32(1
1595 				    + atomic_xchg(&ch->req_lim_delta, 0));
1596 	srp_rsp->tag = tag;
1597 
1598 	srp_rsp->flags |= SRP_RSP_FLAG_RSPVALID;
1599 	srp_rsp->resp_data_len = cpu_to_be32(resp_data_len);
1600 	srp_rsp->data[3] = rsp_code;
1601 
1602 	return resp_len;
1603 }
1604 
1605 #define NO_SUCH_LUN ((uint64_t)-1LL)
1606 
1607 /*
1608  * SCSI LUN addressing method. See also SAM-2 and the section about
1609  * eight byte LUNs.
1610  */
1611 enum scsi_lun_addr_method {
1612 	SCSI_LUN_ADDR_METHOD_PERIPHERAL   = 0,
1613 	SCSI_LUN_ADDR_METHOD_FLAT         = 1,
1614 	SCSI_LUN_ADDR_METHOD_LUN          = 2,
1615 	SCSI_LUN_ADDR_METHOD_EXTENDED_LUN = 3,
1616 };
1617 
1618 /*
1619  * srpt_unpack_lun() - Convert from network LUN to linear LUN.
1620  *
1621  * Convert an 2-byte, 4-byte, 6-byte or 8-byte LUN structure in network byte
1622  * order (big endian) to a linear LUN. Supports three LUN addressing methods:
1623  * peripheral, flat and logical unit. See also SAM-2, section 4.9.4 (page 40).
1624  */
1625 static uint64_t srpt_unpack_lun(const uint8_t *lun, int len)
1626 {
1627 	uint64_t res = NO_SUCH_LUN;
1628 	int addressing_method;
1629 
1630 	if (unlikely(len < 2)) {
1631 		pr_err("Illegal LUN length %d, expected 2 bytes or more\n",
1632 		       len);
1633 		goto out;
1634 	}
1635 
1636 	switch (len) {
1637 	case 8:
1638 		if ((*((__be64 *)lun) &
1639 		     __constant_cpu_to_be64(0x0000FFFFFFFFFFFFLL)) != 0)
1640 			goto out_err;
1641 		break;
1642 	case 4:
1643 		if (*((__be16 *)&lun[2]) != 0)
1644 			goto out_err;
1645 		break;
1646 	case 6:
1647 		if (*((__be32 *)&lun[2]) != 0)
1648 			goto out_err;
1649 		break;
1650 	case 2:
1651 		break;
1652 	default:
1653 		goto out_err;
1654 	}
1655 
1656 	addressing_method = (*lun) >> 6; /* highest two bits of byte 0 */
1657 	switch (addressing_method) {
1658 	case SCSI_LUN_ADDR_METHOD_PERIPHERAL:
1659 	case SCSI_LUN_ADDR_METHOD_FLAT:
1660 	case SCSI_LUN_ADDR_METHOD_LUN:
1661 		res = *(lun + 1) | (((*lun) & 0x3f) << 8);
1662 		break;
1663 
1664 	case SCSI_LUN_ADDR_METHOD_EXTENDED_LUN:
1665 	default:
1666 		pr_err("Unimplemented LUN addressing method %u\n",
1667 		       addressing_method);
1668 		break;
1669 	}
1670 
1671 out:
1672 	return res;
1673 
1674 out_err:
1675 	pr_err("Support for multi-level LUNs has not yet been implemented\n");
1676 	goto out;
1677 }
1678 
1679 static int srpt_check_stop_free(struct se_cmd *cmd)
1680 {
1681 	struct srpt_send_ioctx *ioctx = container_of(cmd,
1682 				struct srpt_send_ioctx, cmd);
1683 
1684 	return target_put_sess_cmd(ioctx->ch->sess, &ioctx->cmd);
1685 }
1686 
1687 /**
1688  * srpt_handle_cmd() - Process SRP_CMD.
1689  */
1690 static int srpt_handle_cmd(struct srpt_rdma_ch *ch,
1691 			   struct srpt_recv_ioctx *recv_ioctx,
1692 			   struct srpt_send_ioctx *send_ioctx)
1693 {
1694 	struct se_cmd *cmd;
1695 	struct srp_cmd *srp_cmd;
1696 	uint64_t unpacked_lun;
1697 	u64 data_len;
1698 	enum dma_data_direction dir;
1699 	sense_reason_t ret;
1700 	int rc;
1701 
1702 	BUG_ON(!send_ioctx);
1703 
1704 	srp_cmd = recv_ioctx->ioctx.buf;
1705 	cmd = &send_ioctx->cmd;
1706 	send_ioctx->tag = srp_cmd->tag;
1707 
1708 	switch (srp_cmd->task_attr) {
1709 	case SRP_CMD_SIMPLE_Q:
1710 		cmd->sam_task_attr = TCM_SIMPLE_TAG;
1711 		break;
1712 	case SRP_CMD_ORDERED_Q:
1713 	default:
1714 		cmd->sam_task_attr = TCM_ORDERED_TAG;
1715 		break;
1716 	case SRP_CMD_HEAD_OF_Q:
1717 		cmd->sam_task_attr = TCM_HEAD_TAG;
1718 		break;
1719 	case SRP_CMD_ACA:
1720 		cmd->sam_task_attr = TCM_ACA_TAG;
1721 		break;
1722 	}
1723 
1724 	if (srpt_get_desc_tbl(send_ioctx, srp_cmd, &dir, &data_len)) {
1725 		pr_err("0x%llx: parsing SRP descriptor table failed.\n",
1726 		       srp_cmd->tag);
1727 		ret = TCM_INVALID_CDB_FIELD;
1728 		goto send_sense;
1729 	}
1730 
1731 	unpacked_lun = srpt_unpack_lun((uint8_t *)&srp_cmd->lun,
1732 				       sizeof(srp_cmd->lun));
1733 	rc = target_submit_cmd(cmd, ch->sess, srp_cmd->cdb,
1734 			&send_ioctx->sense_data[0], unpacked_lun, data_len,
1735 			TCM_SIMPLE_TAG, dir, TARGET_SCF_ACK_KREF);
1736 	if (rc != 0) {
1737 		ret = TCM_LOGICAL_UNIT_COMMUNICATION_FAILURE;
1738 		goto send_sense;
1739 	}
1740 	return 0;
1741 
1742 send_sense:
1743 	transport_send_check_condition_and_sense(cmd, ret, 0);
1744 	return -1;
1745 }
1746 
1747 /**
1748  * srpt_rx_mgmt_fn_tag() - Process a task management function by tag.
1749  * @ch: RDMA channel of the task management request.
1750  * @fn: Task management function to perform.
1751  * @req_tag: Tag of the SRP task management request.
1752  * @mgmt_ioctx: I/O context of the task management request.
1753  *
1754  * Returns zero if the target core will process the task management
1755  * request asynchronously.
1756  *
1757  * Note: It is assumed that the initiator serializes tag-based task management
1758  * requests.
1759  */
1760 static int srpt_rx_mgmt_fn_tag(struct srpt_send_ioctx *ioctx, u64 tag)
1761 {
1762 	struct srpt_device *sdev;
1763 	struct srpt_rdma_ch *ch;
1764 	struct srpt_send_ioctx *target;
1765 	int ret, i;
1766 
1767 	ret = -EINVAL;
1768 	ch = ioctx->ch;
1769 	BUG_ON(!ch);
1770 	BUG_ON(!ch->sport);
1771 	sdev = ch->sport->sdev;
1772 	BUG_ON(!sdev);
1773 	spin_lock_irq(&sdev->spinlock);
1774 	for (i = 0; i < ch->rq_size; ++i) {
1775 		target = ch->ioctx_ring[i];
1776 		if (target->cmd.se_lun == ioctx->cmd.se_lun &&
1777 		    target->tag == tag &&
1778 		    srpt_get_cmd_state(target) != SRPT_STATE_DONE) {
1779 			ret = 0;
1780 			/* now let the target core abort &target->cmd; */
1781 			break;
1782 		}
1783 	}
1784 	spin_unlock_irq(&sdev->spinlock);
1785 	return ret;
1786 }
1787 
1788 static int srp_tmr_to_tcm(int fn)
1789 {
1790 	switch (fn) {
1791 	case SRP_TSK_ABORT_TASK:
1792 		return TMR_ABORT_TASK;
1793 	case SRP_TSK_ABORT_TASK_SET:
1794 		return TMR_ABORT_TASK_SET;
1795 	case SRP_TSK_CLEAR_TASK_SET:
1796 		return TMR_CLEAR_TASK_SET;
1797 	case SRP_TSK_LUN_RESET:
1798 		return TMR_LUN_RESET;
1799 	case SRP_TSK_CLEAR_ACA:
1800 		return TMR_CLEAR_ACA;
1801 	default:
1802 		return -1;
1803 	}
1804 }
1805 
1806 /**
1807  * srpt_handle_tsk_mgmt() - Process an SRP_TSK_MGMT information unit.
1808  *
1809  * Returns 0 if and only if the request will be processed by the target core.
1810  *
1811  * For more information about SRP_TSK_MGMT information units, see also section
1812  * 6.7 in the SRP r16a document.
1813  */
1814 static void srpt_handle_tsk_mgmt(struct srpt_rdma_ch *ch,
1815 				 struct srpt_recv_ioctx *recv_ioctx,
1816 				 struct srpt_send_ioctx *send_ioctx)
1817 {
1818 	struct srp_tsk_mgmt *srp_tsk;
1819 	struct se_cmd *cmd;
1820 	struct se_session *sess = ch->sess;
1821 	uint64_t unpacked_lun;
1822 	uint32_t tag = 0;
1823 	int tcm_tmr;
1824 	int rc;
1825 
1826 	BUG_ON(!send_ioctx);
1827 
1828 	srp_tsk = recv_ioctx->ioctx.buf;
1829 	cmd = &send_ioctx->cmd;
1830 
1831 	pr_debug("recv tsk_mgmt fn %d for task_tag %lld and cmd tag %lld"
1832 		 " cm_id %p sess %p\n", srp_tsk->tsk_mgmt_func,
1833 		 srp_tsk->task_tag, srp_tsk->tag, ch->cm_id, ch->sess);
1834 
1835 	srpt_set_cmd_state(send_ioctx, SRPT_STATE_MGMT);
1836 	send_ioctx->tag = srp_tsk->tag;
1837 	tcm_tmr = srp_tmr_to_tcm(srp_tsk->tsk_mgmt_func);
1838 	if (tcm_tmr < 0) {
1839 		send_ioctx->cmd.se_tmr_req->response =
1840 			TMR_TASK_MGMT_FUNCTION_NOT_SUPPORTED;
1841 		goto fail;
1842 	}
1843 	unpacked_lun = srpt_unpack_lun((uint8_t *)&srp_tsk->lun,
1844 				       sizeof(srp_tsk->lun));
1845 
1846 	if (srp_tsk->tsk_mgmt_func == SRP_TSK_ABORT_TASK) {
1847 		rc = srpt_rx_mgmt_fn_tag(send_ioctx, srp_tsk->task_tag);
1848 		if (rc < 0) {
1849 			send_ioctx->cmd.se_tmr_req->response =
1850 					TMR_TASK_DOES_NOT_EXIST;
1851 			goto fail;
1852 		}
1853 		tag = srp_tsk->task_tag;
1854 	}
1855 	rc = target_submit_tmr(&send_ioctx->cmd, sess, NULL, unpacked_lun,
1856 				srp_tsk, tcm_tmr, GFP_KERNEL, tag,
1857 				TARGET_SCF_ACK_KREF);
1858 	if (rc != 0) {
1859 		send_ioctx->cmd.se_tmr_req->response = TMR_FUNCTION_REJECTED;
1860 		goto fail;
1861 	}
1862 	return;
1863 fail:
1864 	transport_send_check_condition_and_sense(cmd, 0, 0); // XXX:
1865 }
1866 
1867 /**
1868  * srpt_handle_new_iu() - Process a newly received information unit.
1869  * @ch:    RDMA channel through which the information unit has been received.
1870  * @ioctx: SRPT I/O context associated with the information unit.
1871  */
1872 static void srpt_handle_new_iu(struct srpt_rdma_ch *ch,
1873 			       struct srpt_recv_ioctx *recv_ioctx,
1874 			       struct srpt_send_ioctx *send_ioctx)
1875 {
1876 	struct srp_cmd *srp_cmd;
1877 	enum rdma_ch_state ch_state;
1878 
1879 	BUG_ON(!ch);
1880 	BUG_ON(!recv_ioctx);
1881 
1882 	ib_dma_sync_single_for_cpu(ch->sport->sdev->device,
1883 				   recv_ioctx->ioctx.dma, srp_max_req_size,
1884 				   DMA_FROM_DEVICE);
1885 
1886 	ch_state = srpt_get_ch_state(ch);
1887 	if (unlikely(ch_state == CH_CONNECTING)) {
1888 		list_add_tail(&recv_ioctx->wait_list, &ch->cmd_wait_list);
1889 		goto out;
1890 	}
1891 
1892 	if (unlikely(ch_state != CH_LIVE))
1893 		goto out;
1894 
1895 	srp_cmd = recv_ioctx->ioctx.buf;
1896 	if (srp_cmd->opcode == SRP_CMD || srp_cmd->opcode == SRP_TSK_MGMT) {
1897 		if (!send_ioctx)
1898 			send_ioctx = srpt_get_send_ioctx(ch);
1899 		if (unlikely(!send_ioctx)) {
1900 			list_add_tail(&recv_ioctx->wait_list,
1901 				      &ch->cmd_wait_list);
1902 			goto out;
1903 		}
1904 	}
1905 
1906 	switch (srp_cmd->opcode) {
1907 	case SRP_CMD:
1908 		srpt_handle_cmd(ch, recv_ioctx, send_ioctx);
1909 		break;
1910 	case SRP_TSK_MGMT:
1911 		srpt_handle_tsk_mgmt(ch, recv_ioctx, send_ioctx);
1912 		break;
1913 	case SRP_I_LOGOUT:
1914 		pr_err("Not yet implemented: SRP_I_LOGOUT\n");
1915 		break;
1916 	case SRP_CRED_RSP:
1917 		pr_debug("received SRP_CRED_RSP\n");
1918 		break;
1919 	case SRP_AER_RSP:
1920 		pr_debug("received SRP_AER_RSP\n");
1921 		break;
1922 	case SRP_RSP:
1923 		pr_err("Received SRP_RSP\n");
1924 		break;
1925 	default:
1926 		pr_err("received IU with unknown opcode 0x%x\n",
1927 		       srp_cmd->opcode);
1928 		break;
1929 	}
1930 
1931 	srpt_post_recv(ch->sport->sdev, recv_ioctx);
1932 out:
1933 	return;
1934 }
1935 
1936 static void srpt_process_rcv_completion(struct ib_cq *cq,
1937 					struct srpt_rdma_ch *ch,
1938 					struct ib_wc *wc)
1939 {
1940 	struct srpt_device *sdev = ch->sport->sdev;
1941 	struct srpt_recv_ioctx *ioctx;
1942 	u32 index;
1943 
1944 	index = idx_from_wr_id(wc->wr_id);
1945 	if (wc->status == IB_WC_SUCCESS) {
1946 		int req_lim;
1947 
1948 		req_lim = atomic_dec_return(&ch->req_lim);
1949 		if (unlikely(req_lim < 0))
1950 			pr_err("req_lim = %d < 0\n", req_lim);
1951 		ioctx = sdev->ioctx_ring[index];
1952 		srpt_handle_new_iu(ch, ioctx, NULL);
1953 	} else {
1954 		pr_info("receiving failed for idx %u with status %d\n",
1955 			index, wc->status);
1956 	}
1957 }
1958 
1959 /**
1960  * srpt_process_send_completion() - Process an IB send completion.
1961  *
1962  * Note: Although this has not yet been observed during tests, at least in
1963  * theory it is possible that the srpt_get_send_ioctx() call invoked by
1964  * srpt_handle_new_iu() fails. This is possible because the req_lim_delta
1965  * value in each response is set to one, and it is possible that this response
1966  * makes the initiator send a new request before the send completion for that
1967  * response has been processed. This could e.g. happen if the call to
1968  * srpt_put_send_iotcx() is delayed because of a higher priority interrupt or
1969  * if IB retransmission causes generation of the send completion to be
1970  * delayed. Incoming information units for which srpt_get_send_ioctx() fails
1971  * are queued on cmd_wait_list. The code below processes these delayed
1972  * requests one at a time.
1973  */
1974 static void srpt_process_send_completion(struct ib_cq *cq,
1975 					 struct srpt_rdma_ch *ch,
1976 					 struct ib_wc *wc)
1977 {
1978 	struct srpt_send_ioctx *send_ioctx;
1979 	uint32_t index;
1980 	enum srpt_opcode opcode;
1981 
1982 	index = idx_from_wr_id(wc->wr_id);
1983 	opcode = opcode_from_wr_id(wc->wr_id);
1984 	send_ioctx = ch->ioctx_ring[index];
1985 	if (wc->status == IB_WC_SUCCESS) {
1986 		if (opcode == SRPT_SEND)
1987 			srpt_handle_send_comp(ch, send_ioctx);
1988 		else {
1989 			WARN_ON(opcode != SRPT_RDMA_ABORT &&
1990 				wc->opcode != IB_WC_RDMA_READ);
1991 			srpt_handle_rdma_comp(ch, send_ioctx, opcode);
1992 		}
1993 	} else {
1994 		if (opcode == SRPT_SEND) {
1995 			pr_info("sending response for idx %u failed"
1996 				" with status %d\n", index, wc->status);
1997 			srpt_handle_send_err_comp(ch, wc->wr_id);
1998 		} else if (opcode != SRPT_RDMA_MID) {
1999 			pr_info("RDMA t %d for idx %u failed with"
2000 				" status %d\n", opcode, index, wc->status);
2001 			srpt_handle_rdma_err_comp(ch, send_ioctx, opcode);
2002 		}
2003 	}
2004 
2005 	while (unlikely(opcode == SRPT_SEND
2006 			&& !list_empty(&ch->cmd_wait_list)
2007 			&& srpt_get_ch_state(ch) == CH_LIVE
2008 			&& (send_ioctx = srpt_get_send_ioctx(ch)) != NULL)) {
2009 		struct srpt_recv_ioctx *recv_ioctx;
2010 
2011 		recv_ioctx = list_first_entry(&ch->cmd_wait_list,
2012 					      struct srpt_recv_ioctx,
2013 					      wait_list);
2014 		list_del(&recv_ioctx->wait_list);
2015 		srpt_handle_new_iu(ch, recv_ioctx, send_ioctx);
2016 	}
2017 }
2018 
2019 static void srpt_process_completion(struct ib_cq *cq, struct srpt_rdma_ch *ch)
2020 {
2021 	struct ib_wc *const wc = ch->wc;
2022 	int i, n;
2023 
2024 	WARN_ON(cq != ch->cq);
2025 
2026 	ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
2027 	while ((n = ib_poll_cq(cq, ARRAY_SIZE(ch->wc), wc)) > 0) {
2028 		for (i = 0; i < n; i++) {
2029 			if (opcode_from_wr_id(wc[i].wr_id) == SRPT_RECV)
2030 				srpt_process_rcv_completion(cq, ch, &wc[i]);
2031 			else
2032 				srpt_process_send_completion(cq, ch, &wc[i]);
2033 		}
2034 	}
2035 }
2036 
2037 /**
2038  * srpt_completion() - IB completion queue callback function.
2039  *
2040  * Notes:
2041  * - It is guaranteed that a completion handler will never be invoked
2042  *   concurrently on two different CPUs for the same completion queue. See also
2043  *   Documentation/infiniband/core_locking.txt and the implementation of
2044  *   handle_edge_irq() in kernel/irq/chip.c.
2045  * - When threaded IRQs are enabled, completion handlers are invoked in thread
2046  *   context instead of interrupt context.
2047  */
2048 static void srpt_completion(struct ib_cq *cq, void *ctx)
2049 {
2050 	struct srpt_rdma_ch *ch = ctx;
2051 
2052 	wake_up_interruptible(&ch->wait_queue);
2053 }
2054 
2055 static int srpt_compl_thread(void *arg)
2056 {
2057 	struct srpt_rdma_ch *ch;
2058 
2059 	/* Hibernation / freezing of the SRPT kernel thread is not supported. */
2060 	current->flags |= PF_NOFREEZE;
2061 
2062 	ch = arg;
2063 	BUG_ON(!ch);
2064 	pr_info("Session %s: kernel thread %s (PID %d) started\n",
2065 		ch->sess_name, ch->thread->comm, current->pid);
2066 	while (!kthread_should_stop()) {
2067 		wait_event_interruptible(ch->wait_queue,
2068 			(srpt_process_completion(ch->cq, ch),
2069 			 kthread_should_stop()));
2070 	}
2071 	pr_info("Session %s: kernel thread %s (PID %d) stopped\n",
2072 		ch->sess_name, ch->thread->comm, current->pid);
2073 	return 0;
2074 }
2075 
2076 /**
2077  * srpt_create_ch_ib() - Create receive and send completion queues.
2078  */
2079 static int srpt_create_ch_ib(struct srpt_rdma_ch *ch)
2080 {
2081 	struct ib_qp_init_attr *qp_init;
2082 	struct srpt_port *sport = ch->sport;
2083 	struct srpt_device *sdev = sport->sdev;
2084 	u32 srp_sq_size = sport->port_attrib.srp_sq_size;
2085 	struct ib_cq_init_attr cq_attr = {};
2086 	int ret;
2087 
2088 	WARN_ON(ch->rq_size < 1);
2089 
2090 	ret = -ENOMEM;
2091 	qp_init = kzalloc(sizeof *qp_init, GFP_KERNEL);
2092 	if (!qp_init)
2093 		goto out;
2094 
2095 retry:
2096 	cq_attr.cqe = ch->rq_size + srp_sq_size;
2097 	ch->cq = ib_create_cq(sdev->device, srpt_completion, NULL, ch,
2098 			      &cq_attr);
2099 	if (IS_ERR(ch->cq)) {
2100 		ret = PTR_ERR(ch->cq);
2101 		pr_err("failed to create CQ cqe= %d ret= %d\n",
2102 		       ch->rq_size + srp_sq_size, ret);
2103 		goto out;
2104 	}
2105 
2106 	qp_init->qp_context = (void *)ch;
2107 	qp_init->event_handler
2108 		= (void(*)(struct ib_event *, void*))srpt_qp_event;
2109 	qp_init->send_cq = ch->cq;
2110 	qp_init->recv_cq = ch->cq;
2111 	qp_init->srq = sdev->srq;
2112 	qp_init->sq_sig_type = IB_SIGNAL_REQ_WR;
2113 	qp_init->qp_type = IB_QPT_RC;
2114 	qp_init->cap.max_send_wr = srp_sq_size;
2115 	qp_init->cap.max_send_sge = SRPT_DEF_SG_PER_WQE;
2116 
2117 	ch->qp = ib_create_qp(sdev->pd, qp_init);
2118 	if (IS_ERR(ch->qp)) {
2119 		ret = PTR_ERR(ch->qp);
2120 		if (ret == -ENOMEM) {
2121 			srp_sq_size /= 2;
2122 			if (srp_sq_size >= MIN_SRPT_SQ_SIZE) {
2123 				ib_destroy_cq(ch->cq);
2124 				goto retry;
2125 			}
2126 		}
2127 		pr_err("failed to create_qp ret= %d\n", ret);
2128 		goto err_destroy_cq;
2129 	}
2130 
2131 	atomic_set(&ch->sq_wr_avail, qp_init->cap.max_send_wr);
2132 
2133 	pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
2134 		 __func__, ch->cq->cqe, qp_init->cap.max_send_sge,
2135 		 qp_init->cap.max_send_wr, ch->cm_id);
2136 
2137 	ret = srpt_init_ch_qp(ch, ch->qp);
2138 	if (ret)
2139 		goto err_destroy_qp;
2140 
2141 	init_waitqueue_head(&ch->wait_queue);
2142 
2143 	pr_debug("creating thread for session %s\n", ch->sess_name);
2144 
2145 	ch->thread = kthread_run(srpt_compl_thread, ch, "ib_srpt_compl");
2146 	if (IS_ERR(ch->thread)) {
2147 		pr_err("failed to create kernel thread %ld\n",
2148 		       PTR_ERR(ch->thread));
2149 		ch->thread = NULL;
2150 		goto err_destroy_qp;
2151 	}
2152 
2153 out:
2154 	kfree(qp_init);
2155 	return ret;
2156 
2157 err_destroy_qp:
2158 	ib_destroy_qp(ch->qp);
2159 err_destroy_cq:
2160 	ib_destroy_cq(ch->cq);
2161 	goto out;
2162 }
2163 
2164 static void srpt_destroy_ch_ib(struct srpt_rdma_ch *ch)
2165 {
2166 	if (ch->thread)
2167 		kthread_stop(ch->thread);
2168 
2169 	ib_destroy_qp(ch->qp);
2170 	ib_destroy_cq(ch->cq);
2171 }
2172 
2173 /**
2174  * __srpt_close_ch() - Close an RDMA channel by setting the QP error state.
2175  *
2176  * Reset the QP and make sure all resources associated with the channel will
2177  * be deallocated at an appropriate time.
2178  *
2179  * Note: The caller must hold ch->sport->sdev->spinlock.
2180  */
2181 static void __srpt_close_ch(struct srpt_rdma_ch *ch)
2182 {
2183 	struct srpt_device *sdev;
2184 	enum rdma_ch_state prev_state;
2185 	unsigned long flags;
2186 
2187 	sdev = ch->sport->sdev;
2188 
2189 	spin_lock_irqsave(&ch->spinlock, flags);
2190 	prev_state = ch->state;
2191 	switch (prev_state) {
2192 	case CH_CONNECTING:
2193 	case CH_LIVE:
2194 		ch->state = CH_DISCONNECTING;
2195 		break;
2196 	default:
2197 		break;
2198 	}
2199 	spin_unlock_irqrestore(&ch->spinlock, flags);
2200 
2201 	switch (prev_state) {
2202 	case CH_CONNECTING:
2203 		ib_send_cm_rej(ch->cm_id, IB_CM_REJ_NO_RESOURCES, NULL, 0,
2204 			       NULL, 0);
2205 		/* fall through */
2206 	case CH_LIVE:
2207 		if (ib_send_cm_dreq(ch->cm_id, NULL, 0) < 0)
2208 			pr_err("sending CM DREQ failed.\n");
2209 		break;
2210 	case CH_DISCONNECTING:
2211 		break;
2212 	case CH_DRAINING:
2213 	case CH_RELEASING:
2214 		break;
2215 	}
2216 }
2217 
2218 /**
2219  * srpt_close_ch() - Close an RDMA channel.
2220  */
2221 static void srpt_close_ch(struct srpt_rdma_ch *ch)
2222 {
2223 	struct srpt_device *sdev;
2224 
2225 	sdev = ch->sport->sdev;
2226 	spin_lock_irq(&sdev->spinlock);
2227 	__srpt_close_ch(ch);
2228 	spin_unlock_irq(&sdev->spinlock);
2229 }
2230 
2231 /**
2232  * srpt_shutdown_session() - Whether or not a session may be shut down.
2233  */
2234 static int srpt_shutdown_session(struct se_session *se_sess)
2235 {
2236 	struct srpt_rdma_ch *ch = se_sess->fabric_sess_ptr;
2237 	unsigned long flags;
2238 
2239 	spin_lock_irqsave(&ch->spinlock, flags);
2240 	if (ch->in_shutdown) {
2241 		spin_unlock_irqrestore(&ch->spinlock, flags);
2242 		return true;
2243 	}
2244 
2245 	ch->in_shutdown = true;
2246 	target_sess_cmd_list_set_waiting(se_sess);
2247 	spin_unlock_irqrestore(&ch->spinlock, flags);
2248 
2249 	return true;
2250 }
2251 
2252 /**
2253  * srpt_drain_channel() - Drain a channel by resetting the IB queue pair.
2254  * @cm_id: Pointer to the CM ID of the channel to be drained.
2255  *
2256  * Note: Must be called from inside srpt_cm_handler to avoid a race between
2257  * accessing sdev->spinlock and the call to kfree(sdev) in srpt_remove_one()
2258  * (the caller of srpt_cm_handler holds the cm_id spinlock; srpt_remove_one()
2259  * waits until all target sessions for the associated IB device have been
2260  * unregistered and target session registration involves a call to
2261  * ib_destroy_cm_id(), which locks the cm_id spinlock and hence waits until
2262  * this function has finished).
2263  */
2264 static void srpt_drain_channel(struct ib_cm_id *cm_id)
2265 {
2266 	struct srpt_device *sdev;
2267 	struct srpt_rdma_ch *ch;
2268 	int ret;
2269 	bool do_reset = false;
2270 
2271 	WARN_ON_ONCE(irqs_disabled());
2272 
2273 	sdev = cm_id->context;
2274 	BUG_ON(!sdev);
2275 	spin_lock_irq(&sdev->spinlock);
2276 	list_for_each_entry(ch, &sdev->rch_list, list) {
2277 		if (ch->cm_id == cm_id) {
2278 			do_reset = srpt_test_and_set_ch_state(ch,
2279 					CH_CONNECTING, CH_DRAINING) ||
2280 				   srpt_test_and_set_ch_state(ch,
2281 					CH_LIVE, CH_DRAINING) ||
2282 				   srpt_test_and_set_ch_state(ch,
2283 					CH_DISCONNECTING, CH_DRAINING);
2284 			break;
2285 		}
2286 	}
2287 	spin_unlock_irq(&sdev->spinlock);
2288 
2289 	if (do_reset) {
2290 		if (ch->sess)
2291 			srpt_shutdown_session(ch->sess);
2292 
2293 		ret = srpt_ch_qp_err(ch);
2294 		if (ret < 0)
2295 			pr_err("Setting queue pair in error state"
2296 			       " failed: %d\n", ret);
2297 	}
2298 }
2299 
2300 /**
2301  * srpt_find_channel() - Look up an RDMA channel.
2302  * @cm_id: Pointer to the CM ID of the channel to be looked up.
2303  *
2304  * Return NULL if no matching RDMA channel has been found.
2305  */
2306 static struct srpt_rdma_ch *srpt_find_channel(struct srpt_device *sdev,
2307 					      struct ib_cm_id *cm_id)
2308 {
2309 	struct srpt_rdma_ch *ch;
2310 	bool found;
2311 
2312 	WARN_ON_ONCE(irqs_disabled());
2313 	BUG_ON(!sdev);
2314 
2315 	found = false;
2316 	spin_lock_irq(&sdev->spinlock);
2317 	list_for_each_entry(ch, &sdev->rch_list, list) {
2318 		if (ch->cm_id == cm_id) {
2319 			found = true;
2320 			break;
2321 		}
2322 	}
2323 	spin_unlock_irq(&sdev->spinlock);
2324 
2325 	return found ? ch : NULL;
2326 }
2327 
2328 /**
2329  * srpt_release_channel() - Release channel resources.
2330  *
2331  * Schedules the actual release because:
2332  * - Calling the ib_destroy_cm_id() call from inside an IB CM callback would
2333  *   trigger a deadlock.
2334  * - It is not safe to call TCM transport_* functions from interrupt context.
2335  */
2336 static void srpt_release_channel(struct srpt_rdma_ch *ch)
2337 {
2338 	schedule_work(&ch->release_work);
2339 }
2340 
2341 static void srpt_release_channel_work(struct work_struct *w)
2342 {
2343 	struct srpt_rdma_ch *ch;
2344 	struct srpt_device *sdev;
2345 	struct se_session *se_sess;
2346 
2347 	ch = container_of(w, struct srpt_rdma_ch, release_work);
2348 	pr_debug("ch = %p; ch->sess = %p; release_done = %p\n", ch, ch->sess,
2349 		 ch->release_done);
2350 
2351 	sdev = ch->sport->sdev;
2352 	BUG_ON(!sdev);
2353 
2354 	se_sess = ch->sess;
2355 	BUG_ON(!se_sess);
2356 
2357 	target_wait_for_sess_cmds(se_sess);
2358 
2359 	transport_deregister_session_configfs(se_sess);
2360 	transport_deregister_session(se_sess);
2361 	ch->sess = NULL;
2362 
2363 	ib_destroy_cm_id(ch->cm_id);
2364 
2365 	srpt_destroy_ch_ib(ch);
2366 
2367 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
2368 			     ch->sport->sdev, ch->rq_size,
2369 			     ch->rsp_size, DMA_TO_DEVICE);
2370 
2371 	spin_lock_irq(&sdev->spinlock);
2372 	list_del(&ch->list);
2373 	spin_unlock_irq(&sdev->spinlock);
2374 
2375 	if (ch->release_done)
2376 		complete(ch->release_done);
2377 
2378 	wake_up(&sdev->ch_releaseQ);
2379 
2380 	kfree(ch);
2381 }
2382 
2383 static struct srpt_node_acl *__srpt_lookup_acl(struct srpt_port *sport,
2384 					       u8 i_port_id[16])
2385 {
2386 	struct srpt_node_acl *nacl;
2387 
2388 	list_for_each_entry(nacl, &sport->port_acl_list, list)
2389 		if (memcmp(nacl->i_port_id, i_port_id,
2390 			   sizeof(nacl->i_port_id)) == 0)
2391 			return nacl;
2392 
2393 	return NULL;
2394 }
2395 
2396 static struct srpt_node_acl *srpt_lookup_acl(struct srpt_port *sport,
2397 					     u8 i_port_id[16])
2398 {
2399 	struct srpt_node_acl *nacl;
2400 
2401 	spin_lock_irq(&sport->port_acl_lock);
2402 	nacl = __srpt_lookup_acl(sport, i_port_id);
2403 	spin_unlock_irq(&sport->port_acl_lock);
2404 
2405 	return nacl;
2406 }
2407 
2408 /**
2409  * srpt_cm_req_recv() - Process the event IB_CM_REQ_RECEIVED.
2410  *
2411  * Ownership of the cm_id is transferred to the target session if this
2412  * functions returns zero. Otherwise the caller remains the owner of cm_id.
2413  */
2414 static int srpt_cm_req_recv(struct ib_cm_id *cm_id,
2415 			    struct ib_cm_req_event_param *param,
2416 			    void *private_data)
2417 {
2418 	struct srpt_device *sdev = cm_id->context;
2419 	struct srpt_port *sport = &sdev->port[param->port - 1];
2420 	struct srp_login_req *req;
2421 	struct srp_login_rsp *rsp;
2422 	struct srp_login_rej *rej;
2423 	struct ib_cm_rep_param *rep_param;
2424 	struct srpt_rdma_ch *ch, *tmp_ch;
2425 	struct srpt_node_acl *nacl;
2426 	u32 it_iu_len;
2427 	int i;
2428 	int ret = 0;
2429 
2430 	WARN_ON_ONCE(irqs_disabled());
2431 
2432 	if (WARN_ON(!sdev || !private_data))
2433 		return -EINVAL;
2434 
2435 	req = (struct srp_login_req *)private_data;
2436 
2437 	it_iu_len = be32_to_cpu(req->req_it_iu_len);
2438 
2439 	pr_info("Received SRP_LOGIN_REQ with i_port_id 0x%llx:0x%llx,"
2440 		" t_port_id 0x%llx:0x%llx and it_iu_len %d on port %d"
2441 		" (guid=0x%llx:0x%llx)\n",
2442 		be64_to_cpu(*(__be64 *)&req->initiator_port_id[0]),
2443 		be64_to_cpu(*(__be64 *)&req->initiator_port_id[8]),
2444 		be64_to_cpu(*(__be64 *)&req->target_port_id[0]),
2445 		be64_to_cpu(*(__be64 *)&req->target_port_id[8]),
2446 		it_iu_len,
2447 		param->port,
2448 		be64_to_cpu(*(__be64 *)&sdev->port[param->port - 1].gid.raw[0]),
2449 		be64_to_cpu(*(__be64 *)&sdev->port[param->port - 1].gid.raw[8]));
2450 
2451 	rsp = kzalloc(sizeof *rsp, GFP_KERNEL);
2452 	rej = kzalloc(sizeof *rej, GFP_KERNEL);
2453 	rep_param = kzalloc(sizeof *rep_param, GFP_KERNEL);
2454 
2455 	if (!rsp || !rej || !rep_param) {
2456 		ret = -ENOMEM;
2457 		goto out;
2458 	}
2459 
2460 	if (it_iu_len > srp_max_req_size || it_iu_len < 64) {
2461 		rej->reason = __constant_cpu_to_be32(
2462 				SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE);
2463 		ret = -EINVAL;
2464 		pr_err("rejected SRP_LOGIN_REQ because its"
2465 		       " length (%d bytes) is out of range (%d .. %d)\n",
2466 		       it_iu_len, 64, srp_max_req_size);
2467 		goto reject;
2468 	}
2469 
2470 	if (!sport->enabled) {
2471 		rej->reason = __constant_cpu_to_be32(
2472 			     SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2473 		ret = -EINVAL;
2474 		pr_err("rejected SRP_LOGIN_REQ because the target port"
2475 		       " has not yet been enabled\n");
2476 		goto reject;
2477 	}
2478 
2479 	if ((req->req_flags & SRP_MTCH_ACTION) == SRP_MULTICHAN_SINGLE) {
2480 		rsp->rsp_flags = SRP_LOGIN_RSP_MULTICHAN_NO_CHAN;
2481 
2482 		spin_lock_irq(&sdev->spinlock);
2483 
2484 		list_for_each_entry_safe(ch, tmp_ch, &sdev->rch_list, list) {
2485 			if (!memcmp(ch->i_port_id, req->initiator_port_id, 16)
2486 			    && !memcmp(ch->t_port_id, req->target_port_id, 16)
2487 			    && param->port == ch->sport->port
2488 			    && param->listen_id == ch->sport->sdev->cm_id
2489 			    && ch->cm_id) {
2490 				enum rdma_ch_state ch_state;
2491 
2492 				ch_state = srpt_get_ch_state(ch);
2493 				if (ch_state != CH_CONNECTING
2494 				    && ch_state != CH_LIVE)
2495 					continue;
2496 
2497 				/* found an existing channel */
2498 				pr_debug("Found existing channel %s"
2499 					 " cm_id= %p state= %d\n",
2500 					 ch->sess_name, ch->cm_id, ch_state);
2501 
2502 				__srpt_close_ch(ch);
2503 
2504 				rsp->rsp_flags =
2505 					SRP_LOGIN_RSP_MULTICHAN_TERMINATED;
2506 			}
2507 		}
2508 
2509 		spin_unlock_irq(&sdev->spinlock);
2510 
2511 	} else
2512 		rsp->rsp_flags = SRP_LOGIN_RSP_MULTICHAN_MAINTAINED;
2513 
2514 	if (*(__be64 *)req->target_port_id != cpu_to_be64(srpt_service_guid)
2515 	    || *(__be64 *)(req->target_port_id + 8) !=
2516 	       cpu_to_be64(srpt_service_guid)) {
2517 		rej->reason = __constant_cpu_to_be32(
2518 				SRP_LOGIN_REJ_UNABLE_ASSOCIATE_CHANNEL);
2519 		ret = -ENOMEM;
2520 		pr_err("rejected SRP_LOGIN_REQ because it"
2521 		       " has an invalid target port identifier.\n");
2522 		goto reject;
2523 	}
2524 
2525 	ch = kzalloc(sizeof *ch, GFP_KERNEL);
2526 	if (!ch) {
2527 		rej->reason = __constant_cpu_to_be32(
2528 					SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2529 		pr_err("rejected SRP_LOGIN_REQ because no memory.\n");
2530 		ret = -ENOMEM;
2531 		goto reject;
2532 	}
2533 
2534 	INIT_WORK(&ch->release_work, srpt_release_channel_work);
2535 	memcpy(ch->i_port_id, req->initiator_port_id, 16);
2536 	memcpy(ch->t_port_id, req->target_port_id, 16);
2537 	ch->sport = &sdev->port[param->port - 1];
2538 	ch->cm_id = cm_id;
2539 	/*
2540 	 * Avoid QUEUE_FULL conditions by limiting the number of buffers used
2541 	 * for the SRP protocol to the command queue size.
2542 	 */
2543 	ch->rq_size = SRPT_RQ_SIZE;
2544 	spin_lock_init(&ch->spinlock);
2545 	ch->state = CH_CONNECTING;
2546 	INIT_LIST_HEAD(&ch->cmd_wait_list);
2547 	ch->rsp_size = ch->sport->port_attrib.srp_max_rsp_size;
2548 
2549 	ch->ioctx_ring = (struct srpt_send_ioctx **)
2550 		srpt_alloc_ioctx_ring(ch->sport->sdev, ch->rq_size,
2551 				      sizeof(*ch->ioctx_ring[0]),
2552 				      ch->rsp_size, DMA_TO_DEVICE);
2553 	if (!ch->ioctx_ring)
2554 		goto free_ch;
2555 
2556 	INIT_LIST_HEAD(&ch->free_list);
2557 	for (i = 0; i < ch->rq_size; i++) {
2558 		ch->ioctx_ring[i]->ch = ch;
2559 		list_add_tail(&ch->ioctx_ring[i]->free_list, &ch->free_list);
2560 	}
2561 
2562 	ret = srpt_create_ch_ib(ch);
2563 	if (ret) {
2564 		rej->reason = __constant_cpu_to_be32(
2565 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2566 		pr_err("rejected SRP_LOGIN_REQ because creating"
2567 		       " a new RDMA channel failed.\n");
2568 		goto free_ring;
2569 	}
2570 
2571 	ret = srpt_ch_qp_rtr(ch, ch->qp);
2572 	if (ret) {
2573 		rej->reason = __constant_cpu_to_be32(
2574 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2575 		pr_err("rejected SRP_LOGIN_REQ because enabling"
2576 		       " RTR failed (error code = %d)\n", ret);
2577 		goto destroy_ib;
2578 	}
2579 	/*
2580 	 * Use the initator port identifier as the session name.
2581 	 */
2582 	snprintf(ch->sess_name, sizeof(ch->sess_name), "0x%016llx%016llx",
2583 			be64_to_cpu(*(__be64 *)ch->i_port_id),
2584 			be64_to_cpu(*(__be64 *)(ch->i_port_id + 8)));
2585 
2586 	pr_debug("registering session %s\n", ch->sess_name);
2587 
2588 	nacl = srpt_lookup_acl(sport, ch->i_port_id);
2589 	if (!nacl) {
2590 		pr_info("Rejected login because no ACL has been"
2591 			" configured yet for initiator %s.\n", ch->sess_name);
2592 		rej->reason = __constant_cpu_to_be32(
2593 				SRP_LOGIN_REJ_CHANNEL_LIMIT_REACHED);
2594 		goto destroy_ib;
2595 	}
2596 
2597 	ch->sess = transport_init_session(TARGET_PROT_NORMAL);
2598 	if (IS_ERR(ch->sess)) {
2599 		rej->reason = __constant_cpu_to_be32(
2600 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2601 		pr_debug("Failed to create session\n");
2602 		goto deregister_session;
2603 	}
2604 	ch->sess->se_node_acl = &nacl->nacl;
2605 	transport_register_session(&sport->port_tpg_1, &nacl->nacl, ch->sess, ch);
2606 
2607 	pr_debug("Establish connection sess=%p name=%s cm_id=%p\n", ch->sess,
2608 		 ch->sess_name, ch->cm_id);
2609 
2610 	/* create srp_login_response */
2611 	rsp->opcode = SRP_LOGIN_RSP;
2612 	rsp->tag = req->tag;
2613 	rsp->max_it_iu_len = req->req_it_iu_len;
2614 	rsp->max_ti_iu_len = req->req_it_iu_len;
2615 	ch->max_ti_iu_len = it_iu_len;
2616 	rsp->buf_fmt = __constant_cpu_to_be16(SRP_BUF_FORMAT_DIRECT
2617 					      | SRP_BUF_FORMAT_INDIRECT);
2618 	rsp->req_lim_delta = cpu_to_be32(ch->rq_size);
2619 	atomic_set(&ch->req_lim, ch->rq_size);
2620 	atomic_set(&ch->req_lim_delta, 0);
2621 
2622 	/* create cm reply */
2623 	rep_param->qp_num = ch->qp->qp_num;
2624 	rep_param->private_data = (void *)rsp;
2625 	rep_param->private_data_len = sizeof *rsp;
2626 	rep_param->rnr_retry_count = 7;
2627 	rep_param->flow_control = 1;
2628 	rep_param->failover_accepted = 0;
2629 	rep_param->srq = 1;
2630 	rep_param->responder_resources = 4;
2631 	rep_param->initiator_depth = 4;
2632 
2633 	ret = ib_send_cm_rep(cm_id, rep_param);
2634 	if (ret) {
2635 		pr_err("sending SRP_LOGIN_REQ response failed"
2636 		       " (error code = %d)\n", ret);
2637 		goto release_channel;
2638 	}
2639 
2640 	spin_lock_irq(&sdev->spinlock);
2641 	list_add_tail(&ch->list, &sdev->rch_list);
2642 	spin_unlock_irq(&sdev->spinlock);
2643 
2644 	goto out;
2645 
2646 release_channel:
2647 	srpt_set_ch_state(ch, CH_RELEASING);
2648 	transport_deregister_session_configfs(ch->sess);
2649 
2650 deregister_session:
2651 	transport_deregister_session(ch->sess);
2652 	ch->sess = NULL;
2653 
2654 destroy_ib:
2655 	srpt_destroy_ch_ib(ch);
2656 
2657 free_ring:
2658 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
2659 			     ch->sport->sdev, ch->rq_size,
2660 			     ch->rsp_size, DMA_TO_DEVICE);
2661 free_ch:
2662 	kfree(ch);
2663 
2664 reject:
2665 	rej->opcode = SRP_LOGIN_REJ;
2666 	rej->tag = req->tag;
2667 	rej->buf_fmt = __constant_cpu_to_be16(SRP_BUF_FORMAT_DIRECT
2668 					      | SRP_BUF_FORMAT_INDIRECT);
2669 
2670 	ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED, NULL, 0,
2671 			     (void *)rej, sizeof *rej);
2672 
2673 out:
2674 	kfree(rep_param);
2675 	kfree(rsp);
2676 	kfree(rej);
2677 
2678 	return ret;
2679 }
2680 
2681 static void srpt_cm_rej_recv(struct ib_cm_id *cm_id)
2682 {
2683 	pr_info("Received IB REJ for cm_id %p.\n", cm_id);
2684 	srpt_drain_channel(cm_id);
2685 }
2686 
2687 /**
2688  * srpt_cm_rtu_recv() - Process an IB_CM_RTU_RECEIVED or USER_ESTABLISHED event.
2689  *
2690  * An IB_CM_RTU_RECEIVED message indicates that the connection is established
2691  * and that the recipient may begin transmitting (RTU = ready to use).
2692  */
2693 static void srpt_cm_rtu_recv(struct ib_cm_id *cm_id)
2694 {
2695 	struct srpt_rdma_ch *ch;
2696 	int ret;
2697 
2698 	ch = srpt_find_channel(cm_id->context, cm_id);
2699 	BUG_ON(!ch);
2700 
2701 	if (srpt_test_and_set_ch_state(ch, CH_CONNECTING, CH_LIVE)) {
2702 		struct srpt_recv_ioctx *ioctx, *ioctx_tmp;
2703 
2704 		ret = srpt_ch_qp_rts(ch, ch->qp);
2705 
2706 		list_for_each_entry_safe(ioctx, ioctx_tmp, &ch->cmd_wait_list,
2707 					 wait_list) {
2708 			list_del(&ioctx->wait_list);
2709 			srpt_handle_new_iu(ch, ioctx, NULL);
2710 		}
2711 		if (ret)
2712 			srpt_close_ch(ch);
2713 	}
2714 }
2715 
2716 static void srpt_cm_timewait_exit(struct ib_cm_id *cm_id)
2717 {
2718 	pr_info("Received IB TimeWait exit for cm_id %p.\n", cm_id);
2719 	srpt_drain_channel(cm_id);
2720 }
2721 
2722 static void srpt_cm_rep_error(struct ib_cm_id *cm_id)
2723 {
2724 	pr_info("Received IB REP error for cm_id %p.\n", cm_id);
2725 	srpt_drain_channel(cm_id);
2726 }
2727 
2728 /**
2729  * srpt_cm_dreq_recv() - Process reception of a DREQ message.
2730  */
2731 static void srpt_cm_dreq_recv(struct ib_cm_id *cm_id)
2732 {
2733 	struct srpt_rdma_ch *ch;
2734 	unsigned long flags;
2735 	bool send_drep = false;
2736 
2737 	ch = srpt_find_channel(cm_id->context, cm_id);
2738 	BUG_ON(!ch);
2739 
2740 	pr_debug("cm_id= %p ch->state= %d\n", cm_id, srpt_get_ch_state(ch));
2741 
2742 	spin_lock_irqsave(&ch->spinlock, flags);
2743 	switch (ch->state) {
2744 	case CH_CONNECTING:
2745 	case CH_LIVE:
2746 		send_drep = true;
2747 		ch->state = CH_DISCONNECTING;
2748 		break;
2749 	case CH_DISCONNECTING:
2750 	case CH_DRAINING:
2751 	case CH_RELEASING:
2752 		WARN(true, "unexpected channel state %d\n", ch->state);
2753 		break;
2754 	}
2755 	spin_unlock_irqrestore(&ch->spinlock, flags);
2756 
2757 	if (send_drep) {
2758 		if (ib_send_cm_drep(ch->cm_id, NULL, 0) < 0)
2759 			pr_err("Sending IB DREP failed.\n");
2760 		pr_info("Received DREQ and sent DREP for session %s.\n",
2761 			ch->sess_name);
2762 	}
2763 }
2764 
2765 /**
2766  * srpt_cm_drep_recv() - Process reception of a DREP message.
2767  */
2768 static void srpt_cm_drep_recv(struct ib_cm_id *cm_id)
2769 {
2770 	pr_info("Received InfiniBand DREP message for cm_id %p.\n", cm_id);
2771 	srpt_drain_channel(cm_id);
2772 }
2773 
2774 /**
2775  * srpt_cm_handler() - IB connection manager callback function.
2776  *
2777  * A non-zero return value will cause the caller destroy the CM ID.
2778  *
2779  * Note: srpt_cm_handler() must only return a non-zero value when transferring
2780  * ownership of the cm_id to a channel by srpt_cm_req_recv() failed. Returning
2781  * a non-zero value in any other case will trigger a race with the
2782  * ib_destroy_cm_id() call in srpt_release_channel().
2783  */
2784 static int srpt_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
2785 {
2786 	int ret;
2787 
2788 	ret = 0;
2789 	switch (event->event) {
2790 	case IB_CM_REQ_RECEIVED:
2791 		ret = srpt_cm_req_recv(cm_id, &event->param.req_rcvd,
2792 				       event->private_data);
2793 		break;
2794 	case IB_CM_REJ_RECEIVED:
2795 		srpt_cm_rej_recv(cm_id);
2796 		break;
2797 	case IB_CM_RTU_RECEIVED:
2798 	case IB_CM_USER_ESTABLISHED:
2799 		srpt_cm_rtu_recv(cm_id);
2800 		break;
2801 	case IB_CM_DREQ_RECEIVED:
2802 		srpt_cm_dreq_recv(cm_id);
2803 		break;
2804 	case IB_CM_DREP_RECEIVED:
2805 		srpt_cm_drep_recv(cm_id);
2806 		break;
2807 	case IB_CM_TIMEWAIT_EXIT:
2808 		srpt_cm_timewait_exit(cm_id);
2809 		break;
2810 	case IB_CM_REP_ERROR:
2811 		srpt_cm_rep_error(cm_id);
2812 		break;
2813 	case IB_CM_DREQ_ERROR:
2814 		pr_info("Received IB DREQ ERROR event.\n");
2815 		break;
2816 	case IB_CM_MRA_RECEIVED:
2817 		pr_info("Received IB MRA event\n");
2818 		break;
2819 	default:
2820 		pr_err("received unrecognized IB CM event %d\n", event->event);
2821 		break;
2822 	}
2823 
2824 	return ret;
2825 }
2826 
2827 /**
2828  * srpt_perform_rdmas() - Perform IB RDMA.
2829  *
2830  * Returns zero upon success or a negative number upon failure.
2831  */
2832 static int srpt_perform_rdmas(struct srpt_rdma_ch *ch,
2833 			      struct srpt_send_ioctx *ioctx)
2834 {
2835 	struct ib_send_wr wr;
2836 	struct ib_send_wr *bad_wr;
2837 	struct rdma_iu *riu;
2838 	int i;
2839 	int ret;
2840 	int sq_wr_avail;
2841 	enum dma_data_direction dir;
2842 	const int n_rdma = ioctx->n_rdma;
2843 
2844 	dir = ioctx->cmd.data_direction;
2845 	if (dir == DMA_TO_DEVICE) {
2846 		/* write */
2847 		ret = -ENOMEM;
2848 		sq_wr_avail = atomic_sub_return(n_rdma, &ch->sq_wr_avail);
2849 		if (sq_wr_avail < 0) {
2850 			pr_warn("IB send queue full (needed %d)\n",
2851 				n_rdma);
2852 			goto out;
2853 		}
2854 	}
2855 
2856 	ioctx->rdma_aborted = false;
2857 	ret = 0;
2858 	riu = ioctx->rdma_ius;
2859 	memset(&wr, 0, sizeof wr);
2860 
2861 	for (i = 0; i < n_rdma; ++i, ++riu) {
2862 		if (dir == DMA_FROM_DEVICE) {
2863 			wr.opcode = IB_WR_RDMA_WRITE;
2864 			wr.wr_id = encode_wr_id(i == n_rdma - 1 ?
2865 						SRPT_RDMA_WRITE_LAST :
2866 						SRPT_RDMA_MID,
2867 						ioctx->ioctx.index);
2868 		} else {
2869 			wr.opcode = IB_WR_RDMA_READ;
2870 			wr.wr_id = encode_wr_id(i == n_rdma - 1 ?
2871 						SRPT_RDMA_READ_LAST :
2872 						SRPT_RDMA_MID,
2873 						ioctx->ioctx.index);
2874 		}
2875 		wr.next = NULL;
2876 		wr.wr.rdma.remote_addr = riu->raddr;
2877 		wr.wr.rdma.rkey = riu->rkey;
2878 		wr.num_sge = riu->sge_cnt;
2879 		wr.sg_list = riu->sge;
2880 
2881 		/* only get completion event for the last rdma write */
2882 		if (i == (n_rdma - 1) && dir == DMA_TO_DEVICE)
2883 			wr.send_flags = IB_SEND_SIGNALED;
2884 
2885 		ret = ib_post_send(ch->qp, &wr, &bad_wr);
2886 		if (ret)
2887 			break;
2888 	}
2889 
2890 	if (ret)
2891 		pr_err("%s[%d]: ib_post_send() returned %d for %d/%d\n",
2892 				 __func__, __LINE__, ret, i, n_rdma);
2893 	if (ret && i > 0) {
2894 		wr.num_sge = 0;
2895 		wr.wr_id = encode_wr_id(SRPT_RDMA_ABORT, ioctx->ioctx.index);
2896 		wr.send_flags = IB_SEND_SIGNALED;
2897 		while (ch->state == CH_LIVE &&
2898 			ib_post_send(ch->qp, &wr, &bad_wr) != 0) {
2899 			pr_info("Trying to abort failed RDMA transfer [%d]\n",
2900 				ioctx->ioctx.index);
2901 			msleep(1000);
2902 		}
2903 		while (ch->state != CH_RELEASING && !ioctx->rdma_aborted) {
2904 			pr_info("Waiting until RDMA abort finished [%d]\n",
2905 				ioctx->ioctx.index);
2906 			msleep(1000);
2907 		}
2908 	}
2909 out:
2910 	if (unlikely(dir == DMA_TO_DEVICE && ret < 0))
2911 		atomic_add(n_rdma, &ch->sq_wr_avail);
2912 	return ret;
2913 }
2914 
2915 /**
2916  * srpt_xfer_data() - Start data transfer from initiator to target.
2917  */
2918 static int srpt_xfer_data(struct srpt_rdma_ch *ch,
2919 			  struct srpt_send_ioctx *ioctx)
2920 {
2921 	int ret;
2922 
2923 	ret = srpt_map_sg_to_ib_sge(ch, ioctx);
2924 	if (ret) {
2925 		pr_err("%s[%d] ret=%d\n", __func__, __LINE__, ret);
2926 		goto out;
2927 	}
2928 
2929 	ret = srpt_perform_rdmas(ch, ioctx);
2930 	if (ret) {
2931 		if (ret == -EAGAIN || ret == -ENOMEM)
2932 			pr_info("%s[%d] queue full -- ret=%d\n",
2933 				__func__, __LINE__, ret);
2934 		else
2935 			pr_err("%s[%d] fatal error -- ret=%d\n",
2936 			       __func__, __LINE__, ret);
2937 		goto out_unmap;
2938 	}
2939 
2940 out:
2941 	return ret;
2942 out_unmap:
2943 	srpt_unmap_sg_to_ib_sge(ch, ioctx);
2944 	goto out;
2945 }
2946 
2947 static int srpt_write_pending_status(struct se_cmd *se_cmd)
2948 {
2949 	struct srpt_send_ioctx *ioctx;
2950 
2951 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
2952 	return srpt_get_cmd_state(ioctx) == SRPT_STATE_NEED_DATA;
2953 }
2954 
2955 /*
2956  * srpt_write_pending() - Start data transfer from initiator to target (write).
2957  */
2958 static int srpt_write_pending(struct se_cmd *se_cmd)
2959 {
2960 	struct srpt_rdma_ch *ch;
2961 	struct srpt_send_ioctx *ioctx;
2962 	enum srpt_command_state new_state;
2963 	enum rdma_ch_state ch_state;
2964 	int ret;
2965 
2966 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
2967 
2968 	new_state = srpt_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA);
2969 	WARN_ON(new_state == SRPT_STATE_DONE);
2970 
2971 	ch = ioctx->ch;
2972 	BUG_ON(!ch);
2973 
2974 	ch_state = srpt_get_ch_state(ch);
2975 	switch (ch_state) {
2976 	case CH_CONNECTING:
2977 		WARN(true, "unexpected channel state %d\n", ch_state);
2978 		ret = -EINVAL;
2979 		goto out;
2980 	case CH_LIVE:
2981 		break;
2982 	case CH_DISCONNECTING:
2983 	case CH_DRAINING:
2984 	case CH_RELEASING:
2985 		pr_debug("cmd with tag %lld: channel disconnecting\n",
2986 			 ioctx->tag);
2987 		srpt_set_cmd_state(ioctx, SRPT_STATE_DATA_IN);
2988 		ret = -EINVAL;
2989 		goto out;
2990 	}
2991 	ret = srpt_xfer_data(ch, ioctx);
2992 
2993 out:
2994 	return ret;
2995 }
2996 
2997 static u8 tcm_to_srp_tsk_mgmt_status(const int tcm_mgmt_status)
2998 {
2999 	switch (tcm_mgmt_status) {
3000 	case TMR_FUNCTION_COMPLETE:
3001 		return SRP_TSK_MGMT_SUCCESS;
3002 	case TMR_FUNCTION_REJECTED:
3003 		return SRP_TSK_MGMT_FUNC_NOT_SUPP;
3004 	}
3005 	return SRP_TSK_MGMT_FAILED;
3006 }
3007 
3008 /**
3009  * srpt_queue_response() - Transmits the response to a SCSI command.
3010  *
3011  * Callback function called by the TCM core. Must not block since it can be
3012  * invoked on the context of the IB completion handler.
3013  */
3014 static void srpt_queue_response(struct se_cmd *cmd)
3015 {
3016 	struct srpt_rdma_ch *ch;
3017 	struct srpt_send_ioctx *ioctx;
3018 	enum srpt_command_state state;
3019 	unsigned long flags;
3020 	int ret;
3021 	enum dma_data_direction dir;
3022 	int resp_len;
3023 	u8 srp_tm_status;
3024 
3025 	ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
3026 	ch = ioctx->ch;
3027 	BUG_ON(!ch);
3028 
3029 	spin_lock_irqsave(&ioctx->spinlock, flags);
3030 	state = ioctx->state;
3031 	switch (state) {
3032 	case SRPT_STATE_NEW:
3033 	case SRPT_STATE_DATA_IN:
3034 		ioctx->state = SRPT_STATE_CMD_RSP_SENT;
3035 		break;
3036 	case SRPT_STATE_MGMT:
3037 		ioctx->state = SRPT_STATE_MGMT_RSP_SENT;
3038 		break;
3039 	default:
3040 		WARN(true, "ch %p; cmd %d: unexpected command state %d\n",
3041 			ch, ioctx->ioctx.index, ioctx->state);
3042 		break;
3043 	}
3044 	spin_unlock_irqrestore(&ioctx->spinlock, flags);
3045 
3046 	if (unlikely(transport_check_aborted_status(&ioctx->cmd, false)
3047 		     || WARN_ON_ONCE(state == SRPT_STATE_CMD_RSP_SENT))) {
3048 		atomic_inc(&ch->req_lim_delta);
3049 		srpt_abort_cmd(ioctx);
3050 		return;
3051 	}
3052 
3053 	dir = ioctx->cmd.data_direction;
3054 
3055 	/* For read commands, transfer the data to the initiator. */
3056 	if (dir == DMA_FROM_DEVICE && ioctx->cmd.data_length &&
3057 	    !ioctx->queue_status_only) {
3058 		ret = srpt_xfer_data(ch, ioctx);
3059 		if (ret) {
3060 			pr_err("xfer_data failed for tag %llu\n",
3061 			       ioctx->tag);
3062 			return;
3063 		}
3064 	}
3065 
3066 	if (state != SRPT_STATE_MGMT)
3067 		resp_len = srpt_build_cmd_rsp(ch, ioctx, ioctx->tag,
3068 					      cmd->scsi_status);
3069 	else {
3070 		srp_tm_status
3071 			= tcm_to_srp_tsk_mgmt_status(cmd->se_tmr_req->response);
3072 		resp_len = srpt_build_tskmgmt_rsp(ch, ioctx, srp_tm_status,
3073 						 ioctx->tag);
3074 	}
3075 	ret = srpt_post_send(ch, ioctx, resp_len);
3076 	if (ret) {
3077 		pr_err("sending cmd response failed for tag %llu\n",
3078 		       ioctx->tag);
3079 		srpt_unmap_sg_to_ib_sge(ch, ioctx);
3080 		srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
3081 		target_put_sess_cmd(ioctx->ch->sess, &ioctx->cmd);
3082 	}
3083 }
3084 
3085 static int srpt_queue_data_in(struct se_cmd *cmd)
3086 {
3087 	srpt_queue_response(cmd);
3088 	return 0;
3089 }
3090 
3091 static void srpt_queue_tm_rsp(struct se_cmd *cmd)
3092 {
3093 	srpt_queue_response(cmd);
3094 }
3095 
3096 static void srpt_aborted_task(struct se_cmd *cmd)
3097 {
3098 	struct srpt_send_ioctx *ioctx = container_of(cmd,
3099 				struct srpt_send_ioctx, cmd);
3100 
3101 	srpt_unmap_sg_to_ib_sge(ioctx->ch, ioctx);
3102 }
3103 
3104 static int srpt_queue_status(struct se_cmd *cmd)
3105 {
3106 	struct srpt_send_ioctx *ioctx;
3107 
3108 	ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
3109 	BUG_ON(ioctx->sense_data != cmd->sense_buffer);
3110 	if (cmd->se_cmd_flags &
3111 	    (SCF_TRANSPORT_TASK_SENSE | SCF_EMULATED_TASK_SENSE))
3112 		WARN_ON(cmd->scsi_status != SAM_STAT_CHECK_CONDITION);
3113 	ioctx->queue_status_only = true;
3114 	srpt_queue_response(cmd);
3115 	return 0;
3116 }
3117 
3118 static void srpt_refresh_port_work(struct work_struct *work)
3119 {
3120 	struct srpt_port *sport = container_of(work, struct srpt_port, work);
3121 
3122 	srpt_refresh_port(sport);
3123 }
3124 
3125 static int srpt_ch_list_empty(struct srpt_device *sdev)
3126 {
3127 	int res;
3128 
3129 	spin_lock_irq(&sdev->spinlock);
3130 	res = list_empty(&sdev->rch_list);
3131 	spin_unlock_irq(&sdev->spinlock);
3132 
3133 	return res;
3134 }
3135 
3136 /**
3137  * srpt_release_sdev() - Free the channel resources associated with a target.
3138  */
3139 static int srpt_release_sdev(struct srpt_device *sdev)
3140 {
3141 	struct srpt_rdma_ch *ch, *tmp_ch;
3142 	int res;
3143 
3144 	WARN_ON_ONCE(irqs_disabled());
3145 
3146 	BUG_ON(!sdev);
3147 
3148 	spin_lock_irq(&sdev->spinlock);
3149 	list_for_each_entry_safe(ch, tmp_ch, &sdev->rch_list, list)
3150 		__srpt_close_ch(ch);
3151 	spin_unlock_irq(&sdev->spinlock);
3152 
3153 	res = wait_event_interruptible(sdev->ch_releaseQ,
3154 				       srpt_ch_list_empty(sdev));
3155 	if (res)
3156 		pr_err("%s: interrupted.\n", __func__);
3157 
3158 	return 0;
3159 }
3160 
3161 static struct srpt_port *__srpt_lookup_port(const char *name)
3162 {
3163 	struct ib_device *dev;
3164 	struct srpt_device *sdev;
3165 	struct srpt_port *sport;
3166 	int i;
3167 
3168 	list_for_each_entry(sdev, &srpt_dev_list, list) {
3169 		dev = sdev->device;
3170 		if (!dev)
3171 			continue;
3172 
3173 		for (i = 0; i < dev->phys_port_cnt; i++) {
3174 			sport = &sdev->port[i];
3175 
3176 			if (!strcmp(sport->port_guid, name))
3177 				return sport;
3178 		}
3179 	}
3180 
3181 	return NULL;
3182 }
3183 
3184 static struct srpt_port *srpt_lookup_port(const char *name)
3185 {
3186 	struct srpt_port *sport;
3187 
3188 	spin_lock(&srpt_dev_lock);
3189 	sport = __srpt_lookup_port(name);
3190 	spin_unlock(&srpt_dev_lock);
3191 
3192 	return sport;
3193 }
3194 
3195 /**
3196  * srpt_add_one() - Infiniband device addition callback function.
3197  */
3198 static void srpt_add_one(struct ib_device *device)
3199 {
3200 	struct srpt_device *sdev;
3201 	struct srpt_port *sport;
3202 	struct ib_srq_init_attr srq_attr;
3203 	int i;
3204 
3205 	pr_debug("device = %p, device->dma_ops = %p\n", device,
3206 		 device->dma_ops);
3207 
3208 	sdev = kzalloc(sizeof *sdev, GFP_KERNEL);
3209 	if (!sdev)
3210 		goto err;
3211 
3212 	sdev->device = device;
3213 	INIT_LIST_HEAD(&sdev->rch_list);
3214 	init_waitqueue_head(&sdev->ch_releaseQ);
3215 	spin_lock_init(&sdev->spinlock);
3216 
3217 	if (ib_query_device(device, &sdev->dev_attr))
3218 		goto free_dev;
3219 
3220 	sdev->pd = ib_alloc_pd(device);
3221 	if (IS_ERR(sdev->pd))
3222 		goto free_dev;
3223 
3224 	sdev->mr = ib_get_dma_mr(sdev->pd, IB_ACCESS_LOCAL_WRITE);
3225 	if (IS_ERR(sdev->mr))
3226 		goto err_pd;
3227 
3228 	sdev->srq_size = min(srpt_srq_size, sdev->dev_attr.max_srq_wr);
3229 
3230 	srq_attr.event_handler = srpt_srq_event;
3231 	srq_attr.srq_context = (void *)sdev;
3232 	srq_attr.attr.max_wr = sdev->srq_size;
3233 	srq_attr.attr.max_sge = 1;
3234 	srq_attr.attr.srq_limit = 0;
3235 	srq_attr.srq_type = IB_SRQT_BASIC;
3236 
3237 	sdev->srq = ib_create_srq(sdev->pd, &srq_attr);
3238 	if (IS_ERR(sdev->srq))
3239 		goto err_mr;
3240 
3241 	pr_debug("%s: create SRQ #wr= %d max_allow=%d dev= %s\n",
3242 		 __func__, sdev->srq_size, sdev->dev_attr.max_srq_wr,
3243 		 device->name);
3244 
3245 	if (!srpt_service_guid)
3246 		srpt_service_guid = be64_to_cpu(device->node_guid);
3247 
3248 	sdev->cm_id = ib_create_cm_id(device, srpt_cm_handler, sdev);
3249 	if (IS_ERR(sdev->cm_id))
3250 		goto err_srq;
3251 
3252 	/* print out target login information */
3253 	pr_debug("Target login info: id_ext=%016llx,ioc_guid=%016llx,"
3254 		 "pkey=ffff,service_id=%016llx\n", srpt_service_guid,
3255 		 srpt_service_guid, srpt_service_guid);
3256 
3257 	/*
3258 	 * We do not have a consistent service_id (ie. also id_ext of target_id)
3259 	 * to identify this target. We currently use the guid of the first HCA
3260 	 * in the system as service_id; therefore, the target_id will change
3261 	 * if this HCA is gone bad and replaced by different HCA
3262 	 */
3263 	if (ib_cm_listen(sdev->cm_id, cpu_to_be64(srpt_service_guid), 0, NULL))
3264 		goto err_cm;
3265 
3266 	INIT_IB_EVENT_HANDLER(&sdev->event_handler, sdev->device,
3267 			      srpt_event_handler);
3268 	if (ib_register_event_handler(&sdev->event_handler))
3269 		goto err_cm;
3270 
3271 	sdev->ioctx_ring = (struct srpt_recv_ioctx **)
3272 		srpt_alloc_ioctx_ring(sdev, sdev->srq_size,
3273 				      sizeof(*sdev->ioctx_ring[0]),
3274 				      srp_max_req_size, DMA_FROM_DEVICE);
3275 	if (!sdev->ioctx_ring)
3276 		goto err_event;
3277 
3278 	for (i = 0; i < sdev->srq_size; ++i)
3279 		srpt_post_recv(sdev, sdev->ioctx_ring[i]);
3280 
3281 	WARN_ON(sdev->device->phys_port_cnt > ARRAY_SIZE(sdev->port));
3282 
3283 	for (i = 1; i <= sdev->device->phys_port_cnt; i++) {
3284 		sport = &sdev->port[i - 1];
3285 		sport->sdev = sdev;
3286 		sport->port = i;
3287 		sport->port_attrib.srp_max_rdma_size = DEFAULT_MAX_RDMA_SIZE;
3288 		sport->port_attrib.srp_max_rsp_size = DEFAULT_MAX_RSP_SIZE;
3289 		sport->port_attrib.srp_sq_size = DEF_SRPT_SQ_SIZE;
3290 		INIT_WORK(&sport->work, srpt_refresh_port_work);
3291 		INIT_LIST_HEAD(&sport->port_acl_list);
3292 		spin_lock_init(&sport->port_acl_lock);
3293 
3294 		if (srpt_refresh_port(sport)) {
3295 			pr_err("MAD registration failed for %s-%d.\n",
3296 			       srpt_sdev_name(sdev), i);
3297 			goto err_ring;
3298 		}
3299 		snprintf(sport->port_guid, sizeof(sport->port_guid),
3300 			"0x%016llx%016llx",
3301 			be64_to_cpu(sport->gid.global.subnet_prefix),
3302 			be64_to_cpu(sport->gid.global.interface_id));
3303 	}
3304 
3305 	spin_lock(&srpt_dev_lock);
3306 	list_add_tail(&sdev->list, &srpt_dev_list);
3307 	spin_unlock(&srpt_dev_lock);
3308 
3309 out:
3310 	ib_set_client_data(device, &srpt_client, sdev);
3311 	pr_debug("added %s.\n", device->name);
3312 	return;
3313 
3314 err_ring:
3315 	srpt_free_ioctx_ring((struct srpt_ioctx **)sdev->ioctx_ring, sdev,
3316 			     sdev->srq_size, srp_max_req_size,
3317 			     DMA_FROM_DEVICE);
3318 err_event:
3319 	ib_unregister_event_handler(&sdev->event_handler);
3320 err_cm:
3321 	ib_destroy_cm_id(sdev->cm_id);
3322 err_srq:
3323 	ib_destroy_srq(sdev->srq);
3324 err_mr:
3325 	ib_dereg_mr(sdev->mr);
3326 err_pd:
3327 	ib_dealloc_pd(sdev->pd);
3328 free_dev:
3329 	kfree(sdev);
3330 err:
3331 	sdev = NULL;
3332 	pr_info("%s(%s) failed.\n", __func__, device->name);
3333 	goto out;
3334 }
3335 
3336 /**
3337  * srpt_remove_one() - InfiniBand device removal callback function.
3338  */
3339 static void srpt_remove_one(struct ib_device *device)
3340 {
3341 	struct srpt_device *sdev;
3342 	int i;
3343 
3344 	sdev = ib_get_client_data(device, &srpt_client);
3345 	if (!sdev) {
3346 		pr_info("%s(%s): nothing to do.\n", __func__, device->name);
3347 		return;
3348 	}
3349 
3350 	srpt_unregister_mad_agent(sdev);
3351 
3352 	ib_unregister_event_handler(&sdev->event_handler);
3353 
3354 	/* Cancel any work queued by the just unregistered IB event handler. */
3355 	for (i = 0; i < sdev->device->phys_port_cnt; i++)
3356 		cancel_work_sync(&sdev->port[i].work);
3357 
3358 	ib_destroy_cm_id(sdev->cm_id);
3359 
3360 	/*
3361 	 * Unregistering a target must happen after destroying sdev->cm_id
3362 	 * such that no new SRP_LOGIN_REQ information units can arrive while
3363 	 * destroying the target.
3364 	 */
3365 	spin_lock(&srpt_dev_lock);
3366 	list_del(&sdev->list);
3367 	spin_unlock(&srpt_dev_lock);
3368 	srpt_release_sdev(sdev);
3369 
3370 	ib_destroy_srq(sdev->srq);
3371 	ib_dereg_mr(sdev->mr);
3372 	ib_dealloc_pd(sdev->pd);
3373 
3374 	srpt_free_ioctx_ring((struct srpt_ioctx **)sdev->ioctx_ring, sdev,
3375 			     sdev->srq_size, srp_max_req_size, DMA_FROM_DEVICE);
3376 	sdev->ioctx_ring = NULL;
3377 	kfree(sdev);
3378 }
3379 
3380 static struct ib_client srpt_client = {
3381 	.name = DRV_NAME,
3382 	.add = srpt_add_one,
3383 	.remove = srpt_remove_one
3384 };
3385 
3386 static int srpt_check_true(struct se_portal_group *se_tpg)
3387 {
3388 	return 1;
3389 }
3390 
3391 static int srpt_check_false(struct se_portal_group *se_tpg)
3392 {
3393 	return 0;
3394 }
3395 
3396 static char *srpt_get_fabric_name(void)
3397 {
3398 	return "srpt";
3399 }
3400 
3401 static u8 srpt_get_fabric_proto_ident(struct se_portal_group *se_tpg)
3402 {
3403 	return SCSI_TRANSPORTID_PROTOCOLID_SRP;
3404 }
3405 
3406 static char *srpt_get_fabric_wwn(struct se_portal_group *tpg)
3407 {
3408 	struct srpt_port *sport = container_of(tpg, struct srpt_port, port_tpg_1);
3409 
3410 	return sport->port_guid;
3411 }
3412 
3413 static u16 srpt_get_tag(struct se_portal_group *tpg)
3414 {
3415 	return 1;
3416 }
3417 
3418 static u32 srpt_get_default_depth(struct se_portal_group *se_tpg)
3419 {
3420 	return 1;
3421 }
3422 
3423 static u32 srpt_get_pr_transport_id(struct se_portal_group *se_tpg,
3424 				    struct se_node_acl *se_nacl,
3425 				    struct t10_pr_registration *pr_reg,
3426 				    int *format_code, unsigned char *buf)
3427 {
3428 	struct srpt_node_acl *nacl;
3429 	struct spc_rdma_transport_id *tr_id;
3430 
3431 	nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3432 	tr_id = (void *)buf;
3433 	tr_id->protocol_identifier = SCSI_TRANSPORTID_PROTOCOLID_SRP;
3434 	memcpy(tr_id->i_port_id, nacl->i_port_id, sizeof(tr_id->i_port_id));
3435 	return sizeof(*tr_id);
3436 }
3437 
3438 static u32 srpt_get_pr_transport_id_len(struct se_portal_group *se_tpg,
3439 					struct se_node_acl *se_nacl,
3440 					struct t10_pr_registration *pr_reg,
3441 					int *format_code)
3442 {
3443 	*format_code = 0;
3444 	return sizeof(struct spc_rdma_transport_id);
3445 }
3446 
3447 static char *srpt_parse_pr_out_transport_id(struct se_portal_group *se_tpg,
3448 					    const char *buf, u32 *out_tid_len,
3449 					    char **port_nexus_ptr)
3450 {
3451 	struct spc_rdma_transport_id *tr_id;
3452 
3453 	*port_nexus_ptr = NULL;
3454 	*out_tid_len = sizeof(struct spc_rdma_transport_id);
3455 	tr_id = (void *)buf;
3456 	return (char *)tr_id->i_port_id;
3457 }
3458 
3459 static struct se_node_acl *srpt_alloc_fabric_acl(struct se_portal_group *se_tpg)
3460 {
3461 	struct srpt_node_acl *nacl;
3462 
3463 	nacl = kzalloc(sizeof(struct srpt_node_acl), GFP_KERNEL);
3464 	if (!nacl) {
3465 		pr_err("Unable to allocate struct srpt_node_acl\n");
3466 		return NULL;
3467 	}
3468 
3469 	return &nacl->nacl;
3470 }
3471 
3472 static void srpt_release_fabric_acl(struct se_portal_group *se_tpg,
3473 				    struct se_node_acl *se_nacl)
3474 {
3475 	struct srpt_node_acl *nacl;
3476 
3477 	nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3478 	kfree(nacl);
3479 }
3480 
3481 static u32 srpt_tpg_get_inst_index(struct se_portal_group *se_tpg)
3482 {
3483 	return 1;
3484 }
3485 
3486 static void srpt_release_cmd(struct se_cmd *se_cmd)
3487 {
3488 	struct srpt_send_ioctx *ioctx = container_of(se_cmd,
3489 				struct srpt_send_ioctx, cmd);
3490 	struct srpt_rdma_ch *ch = ioctx->ch;
3491 	unsigned long flags;
3492 
3493 	WARN_ON(ioctx->state != SRPT_STATE_DONE);
3494 	WARN_ON(ioctx->mapped_sg_count != 0);
3495 
3496 	if (ioctx->n_rbuf > 1) {
3497 		kfree(ioctx->rbufs);
3498 		ioctx->rbufs = NULL;
3499 		ioctx->n_rbuf = 0;
3500 	}
3501 
3502 	spin_lock_irqsave(&ch->spinlock, flags);
3503 	list_add(&ioctx->free_list, &ch->free_list);
3504 	spin_unlock_irqrestore(&ch->spinlock, flags);
3505 }
3506 
3507 /**
3508  * srpt_close_session() - Forcibly close a session.
3509  *
3510  * Callback function invoked by the TCM core to clean up sessions associated
3511  * with a node ACL when the user invokes
3512  * rmdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3513  */
3514 static void srpt_close_session(struct se_session *se_sess)
3515 {
3516 	DECLARE_COMPLETION_ONSTACK(release_done);
3517 	struct srpt_rdma_ch *ch;
3518 	struct srpt_device *sdev;
3519 	unsigned long res;
3520 
3521 	ch = se_sess->fabric_sess_ptr;
3522 	WARN_ON(ch->sess != se_sess);
3523 
3524 	pr_debug("ch %p state %d\n", ch, srpt_get_ch_state(ch));
3525 
3526 	sdev = ch->sport->sdev;
3527 	spin_lock_irq(&sdev->spinlock);
3528 	BUG_ON(ch->release_done);
3529 	ch->release_done = &release_done;
3530 	__srpt_close_ch(ch);
3531 	spin_unlock_irq(&sdev->spinlock);
3532 
3533 	res = wait_for_completion_timeout(&release_done, 60 * HZ);
3534 	WARN_ON(res == 0);
3535 }
3536 
3537 /**
3538  * srpt_sess_get_index() - Return the value of scsiAttIntrPortIndex (SCSI-MIB).
3539  *
3540  * A quote from RFC 4455 (SCSI-MIB) about this MIB object:
3541  * This object represents an arbitrary integer used to uniquely identify a
3542  * particular attached remote initiator port to a particular SCSI target port
3543  * within a particular SCSI target device within a particular SCSI instance.
3544  */
3545 static u32 srpt_sess_get_index(struct se_session *se_sess)
3546 {
3547 	return 0;
3548 }
3549 
3550 static void srpt_set_default_node_attrs(struct se_node_acl *nacl)
3551 {
3552 }
3553 
3554 static u32 srpt_get_task_tag(struct se_cmd *se_cmd)
3555 {
3556 	struct srpt_send_ioctx *ioctx;
3557 
3558 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
3559 	return ioctx->tag;
3560 }
3561 
3562 /* Note: only used from inside debug printk's by the TCM core. */
3563 static int srpt_get_tcm_cmd_state(struct se_cmd *se_cmd)
3564 {
3565 	struct srpt_send_ioctx *ioctx;
3566 
3567 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
3568 	return srpt_get_cmd_state(ioctx);
3569 }
3570 
3571 /**
3572  * srpt_parse_i_port_id() - Parse an initiator port ID.
3573  * @name: ASCII representation of a 128-bit initiator port ID.
3574  * @i_port_id: Binary 128-bit port ID.
3575  */
3576 static int srpt_parse_i_port_id(u8 i_port_id[16], const char *name)
3577 {
3578 	const char *p;
3579 	unsigned len, count, leading_zero_bytes;
3580 	int ret, rc;
3581 
3582 	p = name;
3583 	if (strncasecmp(p, "0x", 2) == 0)
3584 		p += 2;
3585 	ret = -EINVAL;
3586 	len = strlen(p);
3587 	if (len % 2)
3588 		goto out;
3589 	count = min(len / 2, 16U);
3590 	leading_zero_bytes = 16 - count;
3591 	memset(i_port_id, 0, leading_zero_bytes);
3592 	rc = hex2bin(i_port_id + leading_zero_bytes, p, count);
3593 	if (rc < 0)
3594 		pr_debug("hex2bin failed for srpt_parse_i_port_id: %d\n", rc);
3595 	ret = 0;
3596 out:
3597 	return ret;
3598 }
3599 
3600 /*
3601  * configfs callback function invoked for
3602  * mkdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3603  */
3604 static struct se_node_acl *srpt_make_nodeacl(struct se_portal_group *tpg,
3605 					     struct config_group *group,
3606 					     const char *name)
3607 {
3608 	struct srpt_port *sport = container_of(tpg, struct srpt_port, port_tpg_1);
3609 	struct se_node_acl *se_nacl, *se_nacl_new;
3610 	struct srpt_node_acl *nacl;
3611 	int ret = 0;
3612 	u32 nexus_depth = 1;
3613 	u8 i_port_id[16];
3614 
3615 	if (srpt_parse_i_port_id(i_port_id, name) < 0) {
3616 		pr_err("invalid initiator port ID %s\n", name);
3617 		ret = -EINVAL;
3618 		goto err;
3619 	}
3620 
3621 	se_nacl_new = srpt_alloc_fabric_acl(tpg);
3622 	if (!se_nacl_new) {
3623 		ret = -ENOMEM;
3624 		goto err;
3625 	}
3626 	/*
3627 	 * nacl_new may be released by core_tpg_add_initiator_node_acl()
3628 	 * when converting a node ACL from demo mode to explict
3629 	 */
3630 	se_nacl = core_tpg_add_initiator_node_acl(tpg, se_nacl_new, name,
3631 						  nexus_depth);
3632 	if (IS_ERR(se_nacl)) {
3633 		ret = PTR_ERR(se_nacl);
3634 		goto err;
3635 	}
3636 	/* Locate our struct srpt_node_acl and set sdev and i_port_id. */
3637 	nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3638 	memcpy(&nacl->i_port_id[0], &i_port_id[0], 16);
3639 	nacl->sport = sport;
3640 
3641 	spin_lock_irq(&sport->port_acl_lock);
3642 	list_add_tail(&nacl->list, &sport->port_acl_list);
3643 	spin_unlock_irq(&sport->port_acl_lock);
3644 
3645 	return se_nacl;
3646 err:
3647 	return ERR_PTR(ret);
3648 }
3649 
3650 /*
3651  * configfs callback function invoked for
3652  * rmdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3653  */
3654 static void srpt_drop_nodeacl(struct se_node_acl *se_nacl)
3655 {
3656 	struct srpt_node_acl *nacl;
3657 	struct srpt_device *sdev;
3658 	struct srpt_port *sport;
3659 
3660 	nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3661 	sport = nacl->sport;
3662 	sdev = sport->sdev;
3663 	spin_lock_irq(&sport->port_acl_lock);
3664 	list_del(&nacl->list);
3665 	spin_unlock_irq(&sport->port_acl_lock);
3666 	core_tpg_del_initiator_node_acl(&sport->port_tpg_1, se_nacl, 1);
3667 	srpt_release_fabric_acl(NULL, se_nacl);
3668 }
3669 
3670 static ssize_t srpt_tpg_attrib_show_srp_max_rdma_size(
3671 	struct se_portal_group *se_tpg,
3672 	char *page)
3673 {
3674 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3675 
3676 	return sprintf(page, "%u\n", sport->port_attrib.srp_max_rdma_size);
3677 }
3678 
3679 static ssize_t srpt_tpg_attrib_store_srp_max_rdma_size(
3680 	struct se_portal_group *se_tpg,
3681 	const char *page,
3682 	size_t count)
3683 {
3684 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3685 	unsigned long val;
3686 	int ret;
3687 
3688 	ret = kstrtoul(page, 0, &val);
3689 	if (ret < 0) {
3690 		pr_err("kstrtoul() failed with ret: %d\n", ret);
3691 		return -EINVAL;
3692 	}
3693 	if (val > MAX_SRPT_RDMA_SIZE) {
3694 		pr_err("val: %lu exceeds MAX_SRPT_RDMA_SIZE: %d\n", val,
3695 			MAX_SRPT_RDMA_SIZE);
3696 		return -EINVAL;
3697 	}
3698 	if (val < DEFAULT_MAX_RDMA_SIZE) {
3699 		pr_err("val: %lu smaller than DEFAULT_MAX_RDMA_SIZE: %d\n",
3700 			val, DEFAULT_MAX_RDMA_SIZE);
3701 		return -EINVAL;
3702 	}
3703 	sport->port_attrib.srp_max_rdma_size = val;
3704 
3705 	return count;
3706 }
3707 
3708 TF_TPG_ATTRIB_ATTR(srpt, srp_max_rdma_size, S_IRUGO | S_IWUSR);
3709 
3710 static ssize_t srpt_tpg_attrib_show_srp_max_rsp_size(
3711 	struct se_portal_group *se_tpg,
3712 	char *page)
3713 {
3714 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3715 
3716 	return sprintf(page, "%u\n", sport->port_attrib.srp_max_rsp_size);
3717 }
3718 
3719 static ssize_t srpt_tpg_attrib_store_srp_max_rsp_size(
3720 	struct se_portal_group *se_tpg,
3721 	const char *page,
3722 	size_t count)
3723 {
3724 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3725 	unsigned long val;
3726 	int ret;
3727 
3728 	ret = kstrtoul(page, 0, &val);
3729 	if (ret < 0) {
3730 		pr_err("kstrtoul() failed with ret: %d\n", ret);
3731 		return -EINVAL;
3732 	}
3733 	if (val > MAX_SRPT_RSP_SIZE) {
3734 		pr_err("val: %lu exceeds MAX_SRPT_RSP_SIZE: %d\n", val,
3735 			MAX_SRPT_RSP_SIZE);
3736 		return -EINVAL;
3737 	}
3738 	if (val < MIN_MAX_RSP_SIZE) {
3739 		pr_err("val: %lu smaller than MIN_MAX_RSP_SIZE: %d\n", val,
3740 			MIN_MAX_RSP_SIZE);
3741 		return -EINVAL;
3742 	}
3743 	sport->port_attrib.srp_max_rsp_size = val;
3744 
3745 	return count;
3746 }
3747 
3748 TF_TPG_ATTRIB_ATTR(srpt, srp_max_rsp_size, S_IRUGO | S_IWUSR);
3749 
3750 static ssize_t srpt_tpg_attrib_show_srp_sq_size(
3751 	struct se_portal_group *se_tpg,
3752 	char *page)
3753 {
3754 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3755 
3756 	return sprintf(page, "%u\n", sport->port_attrib.srp_sq_size);
3757 }
3758 
3759 static ssize_t srpt_tpg_attrib_store_srp_sq_size(
3760 	struct se_portal_group *se_tpg,
3761 	const char *page,
3762 	size_t count)
3763 {
3764 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3765 	unsigned long val;
3766 	int ret;
3767 
3768 	ret = kstrtoul(page, 0, &val);
3769 	if (ret < 0) {
3770 		pr_err("kstrtoul() failed with ret: %d\n", ret);
3771 		return -EINVAL;
3772 	}
3773 	if (val > MAX_SRPT_SRQ_SIZE) {
3774 		pr_err("val: %lu exceeds MAX_SRPT_SRQ_SIZE: %d\n", val,
3775 			MAX_SRPT_SRQ_SIZE);
3776 		return -EINVAL;
3777 	}
3778 	if (val < MIN_SRPT_SRQ_SIZE) {
3779 		pr_err("val: %lu smaller than MIN_SRPT_SRQ_SIZE: %d\n", val,
3780 			MIN_SRPT_SRQ_SIZE);
3781 		return -EINVAL;
3782 	}
3783 	sport->port_attrib.srp_sq_size = val;
3784 
3785 	return count;
3786 }
3787 
3788 TF_TPG_ATTRIB_ATTR(srpt, srp_sq_size, S_IRUGO | S_IWUSR);
3789 
3790 static struct configfs_attribute *srpt_tpg_attrib_attrs[] = {
3791 	&srpt_tpg_attrib_srp_max_rdma_size.attr,
3792 	&srpt_tpg_attrib_srp_max_rsp_size.attr,
3793 	&srpt_tpg_attrib_srp_sq_size.attr,
3794 	NULL,
3795 };
3796 
3797 static ssize_t srpt_tpg_show_enable(
3798 	struct se_portal_group *se_tpg,
3799 	char *page)
3800 {
3801 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3802 
3803 	return snprintf(page, PAGE_SIZE, "%d\n", (sport->enabled) ? 1: 0);
3804 }
3805 
3806 static ssize_t srpt_tpg_store_enable(
3807 	struct se_portal_group *se_tpg,
3808 	const char *page,
3809 	size_t count)
3810 {
3811 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3812 	unsigned long tmp;
3813         int ret;
3814 
3815 	ret = kstrtoul(page, 0, &tmp);
3816 	if (ret < 0) {
3817 		pr_err("Unable to extract srpt_tpg_store_enable\n");
3818 		return -EINVAL;
3819 	}
3820 
3821 	if ((tmp != 0) && (tmp != 1)) {
3822 		pr_err("Illegal value for srpt_tpg_store_enable: %lu\n", tmp);
3823 		return -EINVAL;
3824 	}
3825 	if (tmp == 1)
3826 		sport->enabled = true;
3827 	else
3828 		sport->enabled = false;
3829 
3830 	return count;
3831 }
3832 
3833 TF_TPG_BASE_ATTR(srpt, enable, S_IRUGO | S_IWUSR);
3834 
3835 static struct configfs_attribute *srpt_tpg_attrs[] = {
3836 	&srpt_tpg_enable.attr,
3837 	NULL,
3838 };
3839 
3840 /**
3841  * configfs callback invoked for
3842  * mkdir /sys/kernel/config/target/$driver/$port/$tpg
3843  */
3844 static struct se_portal_group *srpt_make_tpg(struct se_wwn *wwn,
3845 					     struct config_group *group,
3846 					     const char *name)
3847 {
3848 	struct srpt_port *sport = container_of(wwn, struct srpt_port, port_wwn);
3849 	int res;
3850 
3851 	/* Initialize sport->port_wwn and sport->port_tpg_1 */
3852 	res = core_tpg_register(&srpt_template, &sport->port_wwn,
3853 			&sport->port_tpg_1, sport, TRANSPORT_TPG_TYPE_NORMAL);
3854 	if (res)
3855 		return ERR_PTR(res);
3856 
3857 	return &sport->port_tpg_1;
3858 }
3859 
3860 /**
3861  * configfs callback invoked for
3862  * rmdir /sys/kernel/config/target/$driver/$port/$tpg
3863  */
3864 static void srpt_drop_tpg(struct se_portal_group *tpg)
3865 {
3866 	struct srpt_port *sport = container_of(tpg,
3867 				struct srpt_port, port_tpg_1);
3868 
3869 	sport->enabled = false;
3870 	core_tpg_deregister(&sport->port_tpg_1);
3871 }
3872 
3873 /**
3874  * configfs callback invoked for
3875  * mkdir /sys/kernel/config/target/$driver/$port
3876  */
3877 static struct se_wwn *srpt_make_tport(struct target_fabric_configfs *tf,
3878 				      struct config_group *group,
3879 				      const char *name)
3880 {
3881 	struct srpt_port *sport;
3882 	int ret;
3883 
3884 	sport = srpt_lookup_port(name);
3885 	pr_debug("make_tport(%s)\n", name);
3886 	ret = -EINVAL;
3887 	if (!sport)
3888 		goto err;
3889 
3890 	return &sport->port_wwn;
3891 
3892 err:
3893 	return ERR_PTR(ret);
3894 }
3895 
3896 /**
3897  * configfs callback invoked for
3898  * rmdir /sys/kernel/config/target/$driver/$port
3899  */
3900 static void srpt_drop_tport(struct se_wwn *wwn)
3901 {
3902 	struct srpt_port *sport = container_of(wwn, struct srpt_port, port_wwn);
3903 
3904 	pr_debug("drop_tport(%s\n", config_item_name(&sport->port_wwn.wwn_group.cg_item));
3905 }
3906 
3907 static ssize_t srpt_wwn_show_attr_version(struct target_fabric_configfs *tf,
3908 					      char *buf)
3909 {
3910 	return scnprintf(buf, PAGE_SIZE, "%s\n", DRV_VERSION);
3911 }
3912 
3913 TF_WWN_ATTR_RO(srpt, version);
3914 
3915 static struct configfs_attribute *srpt_wwn_attrs[] = {
3916 	&srpt_wwn_version.attr,
3917 	NULL,
3918 };
3919 
3920 static const struct target_core_fabric_ops srpt_template = {
3921 	.module				= THIS_MODULE,
3922 	.name				= "srpt",
3923 	.get_fabric_name		= srpt_get_fabric_name,
3924 	.get_fabric_proto_ident		= srpt_get_fabric_proto_ident,
3925 	.tpg_get_wwn			= srpt_get_fabric_wwn,
3926 	.tpg_get_tag			= srpt_get_tag,
3927 	.tpg_get_default_depth		= srpt_get_default_depth,
3928 	.tpg_get_pr_transport_id	= srpt_get_pr_transport_id,
3929 	.tpg_get_pr_transport_id_len	= srpt_get_pr_transport_id_len,
3930 	.tpg_parse_pr_out_transport_id	= srpt_parse_pr_out_transport_id,
3931 	.tpg_check_demo_mode		= srpt_check_false,
3932 	.tpg_check_demo_mode_cache	= srpt_check_true,
3933 	.tpg_check_demo_mode_write_protect = srpt_check_true,
3934 	.tpg_check_prod_mode_write_protect = srpt_check_false,
3935 	.tpg_alloc_fabric_acl		= srpt_alloc_fabric_acl,
3936 	.tpg_release_fabric_acl		= srpt_release_fabric_acl,
3937 	.tpg_get_inst_index		= srpt_tpg_get_inst_index,
3938 	.release_cmd			= srpt_release_cmd,
3939 	.check_stop_free		= srpt_check_stop_free,
3940 	.shutdown_session		= srpt_shutdown_session,
3941 	.close_session			= srpt_close_session,
3942 	.sess_get_index			= srpt_sess_get_index,
3943 	.sess_get_initiator_sid		= NULL,
3944 	.write_pending			= srpt_write_pending,
3945 	.write_pending_status		= srpt_write_pending_status,
3946 	.set_default_node_attributes	= srpt_set_default_node_attrs,
3947 	.get_task_tag			= srpt_get_task_tag,
3948 	.get_cmd_state			= srpt_get_tcm_cmd_state,
3949 	.queue_data_in			= srpt_queue_data_in,
3950 	.queue_status			= srpt_queue_status,
3951 	.queue_tm_rsp			= srpt_queue_tm_rsp,
3952 	.aborted_task			= srpt_aborted_task,
3953 	/*
3954 	 * Setup function pointers for generic logic in
3955 	 * target_core_fabric_configfs.c
3956 	 */
3957 	.fabric_make_wwn		= srpt_make_tport,
3958 	.fabric_drop_wwn		= srpt_drop_tport,
3959 	.fabric_make_tpg		= srpt_make_tpg,
3960 	.fabric_drop_tpg		= srpt_drop_tpg,
3961 	.fabric_post_link		= NULL,
3962 	.fabric_pre_unlink		= NULL,
3963 	.fabric_make_np			= NULL,
3964 	.fabric_drop_np			= NULL,
3965 	.fabric_make_nodeacl		= srpt_make_nodeacl,
3966 	.fabric_drop_nodeacl		= srpt_drop_nodeacl,
3967 
3968 	.tfc_wwn_attrs			= srpt_wwn_attrs,
3969 	.tfc_tpg_base_attrs		= srpt_tpg_attrs,
3970 	.tfc_tpg_attrib_attrs		= srpt_tpg_attrib_attrs,
3971 };
3972 
3973 /**
3974  * srpt_init_module() - Kernel module initialization.
3975  *
3976  * Note: Since ib_register_client() registers callback functions, and since at
3977  * least one of these callback functions (srpt_add_one()) calls target core
3978  * functions, this driver must be registered with the target core before
3979  * ib_register_client() is called.
3980  */
3981 static int __init srpt_init_module(void)
3982 {
3983 	int ret;
3984 
3985 	ret = -EINVAL;
3986 	if (srp_max_req_size < MIN_MAX_REQ_SIZE) {
3987 		pr_err("invalid value %d for kernel module parameter"
3988 		       " srp_max_req_size -- must be at least %d.\n",
3989 		       srp_max_req_size, MIN_MAX_REQ_SIZE);
3990 		goto out;
3991 	}
3992 
3993 	if (srpt_srq_size < MIN_SRPT_SRQ_SIZE
3994 	    || srpt_srq_size > MAX_SRPT_SRQ_SIZE) {
3995 		pr_err("invalid value %d for kernel module parameter"
3996 		       " srpt_srq_size -- must be in the range [%d..%d].\n",
3997 		       srpt_srq_size, MIN_SRPT_SRQ_SIZE, MAX_SRPT_SRQ_SIZE);
3998 		goto out;
3999 	}
4000 
4001 	ret = target_register_template(&srpt_template);
4002 	if (ret)
4003 		goto out;
4004 
4005 	ret = ib_register_client(&srpt_client);
4006 	if (ret) {
4007 		pr_err("couldn't register IB client\n");
4008 		goto out_unregister_target;
4009 	}
4010 
4011 	return 0;
4012 
4013 out_unregister_target:
4014 	target_unregister_template(&srpt_template);
4015 out:
4016 	return ret;
4017 }
4018 
4019 static void __exit srpt_cleanup_module(void)
4020 {
4021 	ib_unregister_client(&srpt_client);
4022 	target_unregister_template(&srpt_template);
4023 }
4024 
4025 module_init(srpt_init_module);
4026 module_exit(srpt_cleanup_module);
4027