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 <linux/inet.h>
45 #include <rdma/ib_cache.h>
46 #include <scsi/scsi_proto.h>
47 #include <scsi/scsi_tcq.h>
48 #include <target/target_core_base.h>
49 #include <target/target_core_fabric.h>
50 #include "ib_srpt.h"
51 
52 /* Name of this kernel module. */
53 #define DRV_NAME		"ib_srpt"
54 
55 #define SRPT_ID_STRING	"Linux SRP target"
56 
57 #undef pr_fmt
58 #define pr_fmt(fmt) DRV_NAME " " fmt
59 
60 MODULE_AUTHOR("Vu Pham and Bart Van Assche");
61 MODULE_DESCRIPTION("SCSI RDMA Protocol target driver");
62 MODULE_LICENSE("Dual BSD/GPL");
63 
64 /*
65  * Global Variables
66  */
67 
68 static u64 srpt_service_guid;
69 static DEFINE_SPINLOCK(srpt_dev_lock);	/* Protects srpt_dev_list. */
70 static LIST_HEAD(srpt_dev_list);	/* List of srpt_device structures. */
71 
72 static unsigned srp_max_req_size = DEFAULT_MAX_REQ_SIZE;
73 module_param(srp_max_req_size, int, 0444);
74 MODULE_PARM_DESC(srp_max_req_size,
75 		 "Maximum size of SRP request messages in bytes.");
76 
77 static int srpt_srq_size = DEFAULT_SRPT_SRQ_SIZE;
78 module_param(srpt_srq_size, int, 0444);
79 MODULE_PARM_DESC(srpt_srq_size,
80 		 "Shared receive queue (SRQ) size.");
81 
82 static int srpt_get_u64_x(char *buffer, const struct kernel_param *kp)
83 {
84 	return sprintf(buffer, "0x%016llx\n", *(u64 *)kp->arg);
85 }
86 module_param_call(srpt_service_guid, NULL, srpt_get_u64_x, &srpt_service_guid,
87 		  0444);
88 MODULE_PARM_DESC(srpt_service_guid,
89 		 "Using this value for ioc_guid, id_ext, and cm_listen_id instead of using the node_guid of the first HCA.");
90 
91 static struct ib_client srpt_client;
92 /* Protects both rdma_cm_port and rdma_cm_id. */
93 static DEFINE_MUTEX(rdma_cm_mutex);
94 /* Port number RDMA/CM will bind to. */
95 static u16 rdma_cm_port;
96 static struct rdma_cm_id *rdma_cm_id;
97 static void srpt_release_cmd(struct se_cmd *se_cmd);
98 static void srpt_free_ch(struct kref *kref);
99 static int srpt_queue_status(struct se_cmd *cmd);
100 static void srpt_recv_done(struct ib_cq *cq, struct ib_wc *wc);
101 static void srpt_send_done(struct ib_cq *cq, struct ib_wc *wc);
102 static void srpt_process_wait_list(struct srpt_rdma_ch *ch);
103 
104 /*
105  * The only allowed channel state changes are those that change the channel
106  * state into a state with a higher numerical value. Hence the new > prev test.
107  */
108 static bool srpt_set_ch_state(struct srpt_rdma_ch *ch, enum rdma_ch_state new)
109 {
110 	unsigned long flags;
111 	enum rdma_ch_state prev;
112 	bool changed = false;
113 
114 	spin_lock_irqsave(&ch->spinlock, flags);
115 	prev = ch->state;
116 	if (new > prev) {
117 		ch->state = new;
118 		changed = true;
119 	}
120 	spin_unlock_irqrestore(&ch->spinlock, flags);
121 
122 	return changed;
123 }
124 
125 /**
126  * srpt_event_handler - asynchronous IB event callback function
127  * @handler: IB event handler registered by ib_register_event_handler().
128  * @event: Description of the event that occurred.
129  *
130  * Callback function called by the InfiniBand core when an asynchronous IB
131  * event occurs. This callback may occur in interrupt context. See also
132  * section 11.5.2, Set Asynchronous Event Handler in the InfiniBand
133  * Architecture Specification.
134  */
135 static void srpt_event_handler(struct ib_event_handler *handler,
136 			       struct ib_event *event)
137 {
138 	struct srpt_device *sdev =
139 		container_of(handler, struct srpt_device, event_handler);
140 	struct srpt_port *sport;
141 	u8 port_num;
142 
143 	pr_debug("ASYNC event= %d on device= %s\n", event->event,
144 		 dev_name(&sdev->device->dev));
145 
146 	switch (event->event) {
147 	case IB_EVENT_PORT_ERR:
148 		port_num = event->element.port_num - 1;
149 		if (port_num < sdev->device->phys_port_cnt) {
150 			sport = &sdev->port[port_num];
151 			sport->lid = 0;
152 			sport->sm_lid = 0;
153 		} else {
154 			WARN(true, "event %d: port_num %d out of range 1..%d\n",
155 			     event->event, port_num + 1,
156 			     sdev->device->phys_port_cnt);
157 		}
158 		break;
159 	case IB_EVENT_PORT_ACTIVE:
160 	case IB_EVENT_LID_CHANGE:
161 	case IB_EVENT_PKEY_CHANGE:
162 	case IB_EVENT_SM_CHANGE:
163 	case IB_EVENT_CLIENT_REREGISTER:
164 	case IB_EVENT_GID_CHANGE:
165 		/* Refresh port data asynchronously. */
166 		port_num = event->element.port_num - 1;
167 		if (port_num < sdev->device->phys_port_cnt) {
168 			sport = &sdev->port[port_num];
169 			if (!sport->lid && !sport->sm_lid)
170 				schedule_work(&sport->work);
171 		} else {
172 			WARN(true, "event %d: port_num %d out of range 1..%d\n",
173 			     event->event, port_num + 1,
174 			     sdev->device->phys_port_cnt);
175 		}
176 		break;
177 	default:
178 		pr_err("received unrecognized IB event %d\n", event->event);
179 		break;
180 	}
181 }
182 
183 /**
184  * srpt_srq_event - SRQ event callback function
185  * @event: Description of the event that occurred.
186  * @ctx: Context pointer specified at SRQ creation time.
187  */
188 static void srpt_srq_event(struct ib_event *event, void *ctx)
189 {
190 	pr_debug("SRQ event %d\n", event->event);
191 }
192 
193 static const char *get_ch_state_name(enum rdma_ch_state s)
194 {
195 	switch (s) {
196 	case CH_CONNECTING:
197 		return "connecting";
198 	case CH_LIVE:
199 		return "live";
200 	case CH_DISCONNECTING:
201 		return "disconnecting";
202 	case CH_DRAINING:
203 		return "draining";
204 	case CH_DISCONNECTED:
205 		return "disconnected";
206 	}
207 	return "???";
208 }
209 
210 /**
211  * srpt_qp_event - QP event callback function
212  * @event: Description of the event that occurred.
213  * @ch: SRPT RDMA channel.
214  */
215 static void srpt_qp_event(struct ib_event *event, struct srpt_rdma_ch *ch)
216 {
217 	pr_debug("QP event %d on ch=%p sess_name=%s-%d state=%s\n",
218 		 event->event, ch, ch->sess_name, ch->qp->qp_num,
219 		 get_ch_state_name(ch->state));
220 
221 	switch (event->event) {
222 	case IB_EVENT_COMM_EST:
223 		if (ch->using_rdma_cm)
224 			rdma_notify(ch->rdma_cm.cm_id, event->event);
225 		else
226 			ib_cm_notify(ch->ib_cm.cm_id, event->event);
227 		break;
228 	case IB_EVENT_QP_LAST_WQE_REACHED:
229 		pr_debug("%s-%d, state %s: received Last WQE event.\n",
230 			 ch->sess_name, ch->qp->qp_num,
231 			 get_ch_state_name(ch->state));
232 		break;
233 	default:
234 		pr_err("received unrecognized IB QP event %d\n", event->event);
235 		break;
236 	}
237 }
238 
239 /**
240  * srpt_set_ioc - initialize a IOUnitInfo structure
241  * @c_list: controller list.
242  * @slot: one-based slot number.
243  * @value: four-bit value.
244  *
245  * Copies the lowest four bits of value in element slot of the array of four
246  * bit elements called c_list (controller list). The index slot is one-based.
247  */
248 static void srpt_set_ioc(u8 *c_list, u32 slot, u8 value)
249 {
250 	u16 id;
251 	u8 tmp;
252 
253 	id = (slot - 1) / 2;
254 	if (slot & 0x1) {
255 		tmp = c_list[id] & 0xf;
256 		c_list[id] = (value << 4) | tmp;
257 	} else {
258 		tmp = c_list[id] & 0xf0;
259 		c_list[id] = (value & 0xf) | tmp;
260 	}
261 }
262 
263 /**
264  * srpt_get_class_port_info - copy ClassPortInfo to a management datagram
265  * @mad: Datagram that will be sent as response to DM_ATTR_CLASS_PORT_INFO.
266  *
267  * See also section 16.3.3.1 ClassPortInfo in the InfiniBand Architecture
268  * Specification.
269  */
270 static void srpt_get_class_port_info(struct ib_dm_mad *mad)
271 {
272 	struct ib_class_port_info *cif;
273 
274 	cif = (struct ib_class_port_info *)mad->data;
275 	memset(cif, 0, sizeof(*cif));
276 	cif->base_version = 1;
277 	cif->class_version = 1;
278 
279 	ib_set_cpi_resp_time(cif, 20);
280 	mad->mad_hdr.status = 0;
281 }
282 
283 /**
284  * srpt_get_iou - write IOUnitInfo to a management datagram
285  * @mad: Datagram that will be sent as response to DM_ATTR_IOU_INFO.
286  *
287  * See also section 16.3.3.3 IOUnitInfo in the InfiniBand Architecture
288  * Specification. See also section B.7, table B.6 in the SRP r16a document.
289  */
290 static void srpt_get_iou(struct ib_dm_mad *mad)
291 {
292 	struct ib_dm_iou_info *ioui;
293 	u8 slot;
294 	int i;
295 
296 	ioui = (struct ib_dm_iou_info *)mad->data;
297 	ioui->change_id = cpu_to_be16(1);
298 	ioui->max_controllers = 16;
299 
300 	/* set present for slot 1 and empty for the rest */
301 	srpt_set_ioc(ioui->controller_list, 1, 1);
302 	for (i = 1, slot = 2; i < 16; i++, slot++)
303 		srpt_set_ioc(ioui->controller_list, slot, 0);
304 
305 	mad->mad_hdr.status = 0;
306 }
307 
308 /**
309  * srpt_get_ioc - write IOControllerprofile to a management datagram
310  * @sport: HCA port through which the MAD has been received.
311  * @slot: Slot number specified in DM_ATTR_IOC_PROFILE query.
312  * @mad: Datagram that will be sent as response to DM_ATTR_IOC_PROFILE.
313  *
314  * See also section 16.3.3.4 IOControllerProfile in the InfiniBand
315  * Architecture Specification. See also section B.7, table B.7 in the SRP
316  * r16a document.
317  */
318 static void srpt_get_ioc(struct srpt_port *sport, u32 slot,
319 			 struct ib_dm_mad *mad)
320 {
321 	struct srpt_device *sdev = sport->sdev;
322 	struct ib_dm_ioc_profile *iocp;
323 	int send_queue_depth;
324 
325 	iocp = (struct ib_dm_ioc_profile *)mad->data;
326 
327 	if (!slot || slot > 16) {
328 		mad->mad_hdr.status
329 			= cpu_to_be16(DM_MAD_STATUS_INVALID_FIELD);
330 		return;
331 	}
332 
333 	if (slot > 2) {
334 		mad->mad_hdr.status
335 			= cpu_to_be16(DM_MAD_STATUS_NO_IOC);
336 		return;
337 	}
338 
339 	if (sdev->use_srq)
340 		send_queue_depth = sdev->srq_size;
341 	else
342 		send_queue_depth = min(MAX_SRPT_RQ_SIZE,
343 				       sdev->device->attrs.max_qp_wr);
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->device->attrs.vendor_id);
349 	iocp->device_id = cpu_to_be32(sdev->device->attrs.vendor_part_id);
350 	iocp->device_version = cpu_to_be16(sdev->device->attrs.hw_ver);
351 	iocp->subsys_vendor_id = cpu_to_be32(sdev->device->attrs.vendor_id);
352 	iocp->subsys_device_id = 0x0;
353 	iocp->io_class = cpu_to_be16(SRP_REV16A_IB_IO_CLASS);
354 	iocp->io_subclass = cpu_to_be16(SRP_IO_SUBCLASS);
355 	iocp->protocol = cpu_to_be16(SRP_PROTOCOL);
356 	iocp->protocol_version = cpu_to_be16(SRP_PROTOCOL_VERSION);
357 	iocp->send_queue_depth = cpu_to_be16(send_queue_depth);
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  * @ioc_guid: I/O controller GUID to use in reply.
372  * @slot: I/O controller number.
373  * @hi: End of the range of service entries to be specified in the reply.
374  * @lo: Start of the range of service entries to be specified in the reply..
375  * @mad: Datagram that will be sent as response to DM_ATTR_SVC_ENTRIES.
376  *
377  * See also section 16.3.3.5 ServiceEntries in the InfiniBand Architecture
378  * Specification. See also section B.7, table B.8 in the SRP r16a document.
379  */
380 static void srpt_get_svc_entries(u64 ioc_guid,
381 				 u16 slot, u8 hi, u8 lo, struct ib_dm_mad *mad)
382 {
383 	struct ib_dm_svc_entries *svc_entries;
384 
385 	WARN_ON(!ioc_guid);
386 
387 	if (!slot || slot > 16) {
388 		mad->mad_hdr.status
389 			= cpu_to_be16(DM_MAD_STATUS_INVALID_FIELD);
390 		return;
391 	}
392 
393 	if (slot > 2 || lo > hi || hi > 1) {
394 		mad->mad_hdr.status
395 			= cpu_to_be16(DM_MAD_STATUS_NO_IOC);
396 		return;
397 	}
398 
399 	svc_entries = (struct ib_dm_svc_entries *)mad->data;
400 	memset(svc_entries, 0, sizeof(*svc_entries));
401 	svc_entries->service_entries[0].id = cpu_to_be64(ioc_guid);
402 	snprintf(svc_entries->service_entries[0].name,
403 		 sizeof(svc_entries->service_entries[0].name),
404 		 "%s%016llx",
405 		 SRP_SERVICE_NAME_PREFIX,
406 		 ioc_guid);
407 
408 	mad->mad_hdr.status = 0;
409 }
410 
411 /**
412  * srpt_mgmt_method_get - process a received management datagram
413  * @sp:      HCA port through which the MAD has been received.
414  * @rq_mad:  received MAD.
415  * @rsp_mad: response MAD.
416  */
417 static void srpt_mgmt_method_get(struct srpt_port *sp, struct ib_mad *rq_mad,
418 				 struct ib_dm_mad *rsp_mad)
419 {
420 	u16 attr_id;
421 	u32 slot;
422 	u8 hi, lo;
423 
424 	attr_id = be16_to_cpu(rq_mad->mad_hdr.attr_id);
425 	switch (attr_id) {
426 	case DM_ATTR_CLASS_PORT_INFO:
427 		srpt_get_class_port_info(rsp_mad);
428 		break;
429 	case DM_ATTR_IOU_INFO:
430 		srpt_get_iou(rsp_mad);
431 		break;
432 	case DM_ATTR_IOC_PROFILE:
433 		slot = be32_to_cpu(rq_mad->mad_hdr.attr_mod);
434 		srpt_get_ioc(sp, slot, rsp_mad);
435 		break;
436 	case DM_ATTR_SVC_ENTRIES:
437 		slot = be32_to_cpu(rq_mad->mad_hdr.attr_mod);
438 		hi = (u8) ((slot >> 8) & 0xff);
439 		lo = (u8) (slot & 0xff);
440 		slot = (u16) ((slot >> 16) & 0xffff);
441 		srpt_get_svc_entries(srpt_service_guid,
442 				     slot, hi, lo, rsp_mad);
443 		break;
444 	default:
445 		rsp_mad->mad_hdr.status =
446 		    cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD_ATTR);
447 		break;
448 	}
449 }
450 
451 /**
452  * srpt_mad_send_handler - MAD send completion callback
453  * @mad_agent: Return value of ib_register_mad_agent().
454  * @mad_wc: Work completion reporting that the MAD has been sent.
455  */
456 static void srpt_mad_send_handler(struct ib_mad_agent *mad_agent,
457 				  struct ib_mad_send_wc *mad_wc)
458 {
459 	rdma_destroy_ah(mad_wc->send_buf->ah, RDMA_DESTROY_AH_SLEEPABLE);
460 	ib_free_send_mad(mad_wc->send_buf);
461 }
462 
463 /**
464  * srpt_mad_recv_handler - MAD reception callback function
465  * @mad_agent: Return value of ib_register_mad_agent().
466  * @send_buf: Not used.
467  * @mad_wc: Work completion reporting that a MAD has been received.
468  */
469 static void srpt_mad_recv_handler(struct ib_mad_agent *mad_agent,
470 				  struct ib_mad_send_buf *send_buf,
471 				  struct ib_mad_recv_wc *mad_wc)
472 {
473 	struct srpt_port *sport = (struct srpt_port *)mad_agent->context;
474 	struct ib_ah *ah;
475 	struct ib_mad_send_buf *rsp;
476 	struct ib_dm_mad *dm_mad;
477 
478 	if (!mad_wc || !mad_wc->recv_buf.mad)
479 		return;
480 
481 	ah = ib_create_ah_from_wc(mad_agent->qp->pd, mad_wc->wc,
482 				  mad_wc->recv_buf.grh, mad_agent->port_num);
483 	if (IS_ERR(ah))
484 		goto err;
485 
486 	BUILD_BUG_ON(offsetof(struct ib_dm_mad, data) != IB_MGMT_DEVICE_HDR);
487 
488 	rsp = ib_create_send_mad(mad_agent, mad_wc->wc->src_qp,
489 				 mad_wc->wc->pkey_index, 0,
490 				 IB_MGMT_DEVICE_HDR, IB_MGMT_DEVICE_DATA,
491 				 GFP_KERNEL,
492 				 IB_MGMT_BASE_VERSION);
493 	if (IS_ERR(rsp))
494 		goto err_rsp;
495 
496 	rsp->ah = ah;
497 
498 	dm_mad = rsp->mad;
499 	memcpy(dm_mad, mad_wc->recv_buf.mad, sizeof(*dm_mad));
500 	dm_mad->mad_hdr.method = IB_MGMT_METHOD_GET_RESP;
501 	dm_mad->mad_hdr.status = 0;
502 
503 	switch (mad_wc->recv_buf.mad->mad_hdr.method) {
504 	case IB_MGMT_METHOD_GET:
505 		srpt_mgmt_method_get(sport, mad_wc->recv_buf.mad, dm_mad);
506 		break;
507 	case IB_MGMT_METHOD_SET:
508 		dm_mad->mad_hdr.status =
509 		    cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD_ATTR);
510 		break;
511 	default:
512 		dm_mad->mad_hdr.status =
513 		    cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD);
514 		break;
515 	}
516 
517 	if (!ib_post_send_mad(rsp, NULL)) {
518 		ib_free_recv_mad(mad_wc);
519 		/* will destroy_ah & free_send_mad in send completion */
520 		return;
521 	}
522 
523 	ib_free_send_mad(rsp);
524 
525 err_rsp:
526 	rdma_destroy_ah(ah, RDMA_DESTROY_AH_SLEEPABLE);
527 err:
528 	ib_free_recv_mad(mad_wc);
529 }
530 
531 static int srpt_format_guid(char *buf, unsigned int size, const __be64 *guid)
532 {
533 	const __be16 *g = (const __be16 *)guid;
534 
535 	return snprintf(buf, size, "%04x:%04x:%04x:%04x",
536 			be16_to_cpu(g[0]), be16_to_cpu(g[1]),
537 			be16_to_cpu(g[2]), be16_to_cpu(g[3]));
538 }
539 
540 /**
541  * srpt_refresh_port - configure a HCA port
542  * @sport: SRPT HCA port.
543  *
544  * Enable InfiniBand management datagram processing, update the cached sm_lid,
545  * lid and gid values, and register a callback function for processing MADs
546  * on the specified port.
547  *
548  * Note: It is safe to call this function more than once for the same port.
549  */
550 static int srpt_refresh_port(struct srpt_port *sport)
551 {
552 	struct ib_mad_reg_req reg_req;
553 	struct ib_port_modify port_modify;
554 	struct ib_port_attr port_attr;
555 	int ret;
556 
557 	ret = ib_query_port(sport->sdev->device, sport->port, &port_attr);
558 	if (ret)
559 		return ret;
560 
561 	sport->sm_lid = port_attr.sm_lid;
562 	sport->lid = port_attr.lid;
563 
564 	ret = rdma_query_gid(sport->sdev->device, sport->port, 0, &sport->gid);
565 	if (ret)
566 		return ret;
567 
568 	sport->port_guid_id.wwn.priv = sport;
569 	srpt_format_guid(sport->port_guid_id.name,
570 			 sizeof(sport->port_guid_id.name),
571 			 &sport->gid.global.interface_id);
572 	sport->port_gid_id.wwn.priv = sport;
573 	snprintf(sport->port_gid_id.name, sizeof(sport->port_gid_id.name),
574 		 "0x%016llx%016llx",
575 		 be64_to_cpu(sport->gid.global.subnet_prefix),
576 		 be64_to_cpu(sport->gid.global.interface_id));
577 
578 	if (rdma_protocol_iwarp(sport->sdev->device, sport->port))
579 		return 0;
580 
581 	memset(&port_modify, 0, sizeof(port_modify));
582 	port_modify.set_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP;
583 	port_modify.clr_port_cap_mask = 0;
584 
585 	ret = ib_modify_port(sport->sdev->device, sport->port, 0, &port_modify);
586 	if (ret) {
587 		pr_warn("%s-%d: enabling device management failed (%d). Note: this is expected if SR-IOV is enabled.\n",
588 			dev_name(&sport->sdev->device->dev), sport->port, ret);
589 		return 0;
590 	}
591 
592 	if (!sport->mad_agent) {
593 		memset(&reg_req, 0, sizeof(reg_req));
594 		reg_req.mgmt_class = IB_MGMT_CLASS_DEVICE_MGMT;
595 		reg_req.mgmt_class_version = IB_MGMT_BASE_VERSION;
596 		set_bit(IB_MGMT_METHOD_GET, reg_req.method_mask);
597 		set_bit(IB_MGMT_METHOD_SET, reg_req.method_mask);
598 
599 		sport->mad_agent = ib_register_mad_agent(sport->sdev->device,
600 							 sport->port,
601 							 IB_QPT_GSI,
602 							 &reg_req, 0,
603 							 srpt_mad_send_handler,
604 							 srpt_mad_recv_handler,
605 							 sport, 0);
606 		if (IS_ERR(sport->mad_agent)) {
607 			pr_err("%s-%d: MAD agent registration failed (%ld). Note: this is expected if SR-IOV is enabled.\n",
608 			       dev_name(&sport->sdev->device->dev), sport->port,
609 			       PTR_ERR(sport->mad_agent));
610 			sport->mad_agent = NULL;
611 			memset(&port_modify, 0, sizeof(port_modify));
612 			port_modify.clr_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP;
613 			ib_modify_port(sport->sdev->device, sport->port, 0,
614 				       &port_modify);
615 
616 		}
617 	}
618 
619 	return 0;
620 }
621 
622 /**
623  * srpt_unregister_mad_agent - unregister MAD callback functions
624  * @sdev: SRPT HCA pointer.
625  *
626  * Note: It is safe to call this function more than once for the same device.
627  */
628 static void srpt_unregister_mad_agent(struct srpt_device *sdev)
629 {
630 	struct ib_port_modify port_modify = {
631 		.clr_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP,
632 	};
633 	struct srpt_port *sport;
634 	int i;
635 
636 	for (i = 1; i <= sdev->device->phys_port_cnt; i++) {
637 		sport = &sdev->port[i - 1];
638 		WARN_ON(sport->port != i);
639 		if (sport->mad_agent) {
640 			ib_modify_port(sdev->device, i, 0, &port_modify);
641 			ib_unregister_mad_agent(sport->mad_agent);
642 			sport->mad_agent = NULL;
643 		}
644 	}
645 }
646 
647 /**
648  * srpt_alloc_ioctx - allocate a SRPT I/O context structure
649  * @sdev: SRPT HCA pointer.
650  * @ioctx_size: I/O context size.
651  * @buf_cache: I/O buffer cache.
652  * @dir: DMA data direction.
653  */
654 static struct srpt_ioctx *srpt_alloc_ioctx(struct srpt_device *sdev,
655 					   int ioctx_size,
656 					   struct kmem_cache *buf_cache,
657 					   enum dma_data_direction dir)
658 {
659 	struct srpt_ioctx *ioctx;
660 
661 	ioctx = kzalloc(ioctx_size, GFP_KERNEL);
662 	if (!ioctx)
663 		goto err;
664 
665 	ioctx->buf = kmem_cache_alloc(buf_cache, GFP_KERNEL);
666 	if (!ioctx->buf)
667 		goto err_free_ioctx;
668 
669 	ioctx->dma = ib_dma_map_single(sdev->device, ioctx->buf,
670 				       kmem_cache_size(buf_cache), dir);
671 	if (ib_dma_mapping_error(sdev->device, ioctx->dma))
672 		goto err_free_buf;
673 
674 	return ioctx;
675 
676 err_free_buf:
677 	kmem_cache_free(buf_cache, ioctx->buf);
678 err_free_ioctx:
679 	kfree(ioctx);
680 err:
681 	return NULL;
682 }
683 
684 /**
685  * srpt_free_ioctx - free a SRPT I/O context structure
686  * @sdev: SRPT HCA pointer.
687  * @ioctx: I/O context pointer.
688  * @buf_cache: I/O buffer cache.
689  * @dir: DMA data direction.
690  */
691 static void srpt_free_ioctx(struct srpt_device *sdev, struct srpt_ioctx *ioctx,
692 			    struct kmem_cache *buf_cache,
693 			    enum dma_data_direction dir)
694 {
695 	if (!ioctx)
696 		return;
697 
698 	ib_dma_unmap_single(sdev->device, ioctx->dma,
699 			    kmem_cache_size(buf_cache), dir);
700 	kmem_cache_free(buf_cache, ioctx->buf);
701 	kfree(ioctx);
702 }
703 
704 /**
705  * srpt_alloc_ioctx_ring - allocate a ring of SRPT I/O context structures
706  * @sdev:       Device to allocate the I/O context ring for.
707  * @ring_size:  Number of elements in the I/O context ring.
708  * @ioctx_size: I/O context size.
709  * @buf_cache:  I/O buffer cache.
710  * @alignment_offset: Offset in each ring buffer at which the SRP information
711  *		unit starts.
712  * @dir:        DMA data direction.
713  */
714 static struct srpt_ioctx **srpt_alloc_ioctx_ring(struct srpt_device *sdev,
715 				int ring_size, int ioctx_size,
716 				struct kmem_cache *buf_cache,
717 				int alignment_offset,
718 				enum dma_data_direction dir)
719 {
720 	struct srpt_ioctx **ring;
721 	int i;
722 
723 	WARN_ON(ioctx_size != sizeof(struct srpt_recv_ioctx) &&
724 		ioctx_size != sizeof(struct srpt_send_ioctx));
725 
726 	ring = kvmalloc_array(ring_size, sizeof(ring[0]), GFP_KERNEL);
727 	if (!ring)
728 		goto out;
729 	for (i = 0; i < ring_size; ++i) {
730 		ring[i] = srpt_alloc_ioctx(sdev, ioctx_size, buf_cache, dir);
731 		if (!ring[i])
732 			goto err;
733 		ring[i]->index = i;
734 		ring[i]->offset = alignment_offset;
735 	}
736 	goto out;
737 
738 err:
739 	while (--i >= 0)
740 		srpt_free_ioctx(sdev, ring[i], buf_cache, dir);
741 	kvfree(ring);
742 	ring = NULL;
743 out:
744 	return ring;
745 }
746 
747 /**
748  * srpt_free_ioctx_ring - free the ring of SRPT I/O context structures
749  * @ioctx_ring: I/O context ring to be freed.
750  * @sdev: SRPT HCA pointer.
751  * @ring_size: Number of ring elements.
752  * @buf_cache: I/O buffer cache.
753  * @dir: DMA data direction.
754  */
755 static void srpt_free_ioctx_ring(struct srpt_ioctx **ioctx_ring,
756 				 struct srpt_device *sdev, int ring_size,
757 				 struct kmem_cache *buf_cache,
758 				 enum dma_data_direction dir)
759 {
760 	int i;
761 
762 	if (!ioctx_ring)
763 		return;
764 
765 	for (i = 0; i < ring_size; ++i)
766 		srpt_free_ioctx(sdev, ioctx_ring[i], buf_cache, dir);
767 	kvfree(ioctx_ring);
768 }
769 
770 /**
771  * srpt_set_cmd_state - set the state of a SCSI command
772  * @ioctx: Send I/O context.
773  * @new: New I/O context state.
774  *
775  * Does not modify the state of aborted commands. Returns the previous command
776  * state.
777  */
778 static enum srpt_command_state srpt_set_cmd_state(struct srpt_send_ioctx *ioctx,
779 						  enum srpt_command_state new)
780 {
781 	enum srpt_command_state previous;
782 
783 	previous = ioctx->state;
784 	if (previous != SRPT_STATE_DONE)
785 		ioctx->state = new;
786 
787 	return previous;
788 }
789 
790 /**
791  * srpt_test_and_set_cmd_state - test and set the state of a command
792  * @ioctx: Send I/O context.
793  * @old: Current I/O context state.
794  * @new: New I/O context state.
795  *
796  * Returns true if and only if the previous command state was equal to 'old'.
797  */
798 static bool srpt_test_and_set_cmd_state(struct srpt_send_ioctx *ioctx,
799 					enum srpt_command_state old,
800 					enum srpt_command_state new)
801 {
802 	enum srpt_command_state previous;
803 
804 	WARN_ON(!ioctx);
805 	WARN_ON(old == SRPT_STATE_DONE);
806 	WARN_ON(new == SRPT_STATE_NEW);
807 
808 	previous = ioctx->state;
809 	if (previous == old)
810 		ioctx->state = new;
811 
812 	return previous == old;
813 }
814 
815 /**
816  * srpt_post_recv - post an IB receive request
817  * @sdev: SRPT HCA pointer.
818  * @ch: SRPT RDMA channel.
819  * @ioctx: Receive I/O context pointer.
820  */
821 static int srpt_post_recv(struct srpt_device *sdev, struct srpt_rdma_ch *ch,
822 			  struct srpt_recv_ioctx *ioctx)
823 {
824 	struct ib_sge list;
825 	struct ib_recv_wr wr;
826 
827 	BUG_ON(!sdev);
828 	list.addr = ioctx->ioctx.dma + ioctx->ioctx.offset;
829 	list.length = srp_max_req_size;
830 	list.lkey = sdev->lkey;
831 
832 	ioctx->ioctx.cqe.done = srpt_recv_done;
833 	wr.wr_cqe = &ioctx->ioctx.cqe;
834 	wr.next = NULL;
835 	wr.sg_list = &list;
836 	wr.num_sge = 1;
837 
838 	if (sdev->use_srq)
839 		return ib_post_srq_recv(sdev->srq, &wr, NULL);
840 	else
841 		return ib_post_recv(ch->qp, &wr, NULL);
842 }
843 
844 /**
845  * srpt_zerolength_write - perform a zero-length RDMA write
846  * @ch: SRPT RDMA channel.
847  *
848  * A quote from the InfiniBand specification: C9-88: For an HCA responder
849  * using Reliable Connection service, for each zero-length RDMA READ or WRITE
850  * request, the R_Key shall not be validated, even if the request includes
851  * Immediate data.
852  */
853 static int srpt_zerolength_write(struct srpt_rdma_ch *ch)
854 {
855 	struct ib_rdma_wr wr = {
856 		.wr = {
857 			.next		= NULL,
858 			{ .wr_cqe	= &ch->zw_cqe, },
859 			.opcode		= IB_WR_RDMA_WRITE,
860 			.send_flags	= IB_SEND_SIGNALED,
861 		}
862 	};
863 
864 	pr_debug("%s-%d: queued zerolength write\n", ch->sess_name,
865 		 ch->qp->qp_num);
866 
867 	return ib_post_send(ch->qp, &wr.wr, NULL);
868 }
869 
870 static void srpt_zerolength_write_done(struct ib_cq *cq, struct ib_wc *wc)
871 {
872 	struct srpt_rdma_ch *ch = wc->qp->qp_context;
873 
874 	pr_debug("%s-%d wc->status %d\n", ch->sess_name, ch->qp->qp_num,
875 		 wc->status);
876 
877 	if (wc->status == IB_WC_SUCCESS) {
878 		srpt_process_wait_list(ch);
879 	} else {
880 		if (srpt_set_ch_state(ch, CH_DISCONNECTED))
881 			schedule_work(&ch->release_work);
882 		else
883 			pr_debug("%s-%d: already disconnected.\n",
884 				 ch->sess_name, ch->qp->qp_num);
885 	}
886 }
887 
888 static int srpt_alloc_rw_ctxs(struct srpt_send_ioctx *ioctx,
889 		struct srp_direct_buf *db, int nbufs, struct scatterlist **sg,
890 		unsigned *sg_cnt)
891 {
892 	enum dma_data_direction dir = target_reverse_dma_direction(&ioctx->cmd);
893 	struct srpt_rdma_ch *ch = ioctx->ch;
894 	struct scatterlist *prev = NULL;
895 	unsigned prev_nents;
896 	int ret, i;
897 
898 	if (nbufs == 1) {
899 		ioctx->rw_ctxs = &ioctx->s_rw_ctx;
900 	} else {
901 		ioctx->rw_ctxs = kmalloc_array(nbufs, sizeof(*ioctx->rw_ctxs),
902 			GFP_KERNEL);
903 		if (!ioctx->rw_ctxs)
904 			return -ENOMEM;
905 	}
906 
907 	for (i = ioctx->n_rw_ctx; i < nbufs; i++, db++) {
908 		struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i];
909 		u64 remote_addr = be64_to_cpu(db->va);
910 		u32 size = be32_to_cpu(db->len);
911 		u32 rkey = be32_to_cpu(db->key);
912 
913 		ret = target_alloc_sgl(&ctx->sg, &ctx->nents, size, false,
914 				i < nbufs - 1);
915 		if (ret)
916 			goto unwind;
917 
918 		ret = rdma_rw_ctx_init(&ctx->rw, ch->qp, ch->sport->port,
919 				ctx->sg, ctx->nents, 0, remote_addr, rkey, dir);
920 		if (ret < 0) {
921 			target_free_sgl(ctx->sg, ctx->nents);
922 			goto unwind;
923 		}
924 
925 		ioctx->n_rdma += ret;
926 		ioctx->n_rw_ctx++;
927 
928 		if (prev) {
929 			sg_unmark_end(&prev[prev_nents - 1]);
930 			sg_chain(prev, prev_nents + 1, ctx->sg);
931 		} else {
932 			*sg = ctx->sg;
933 		}
934 
935 		prev = ctx->sg;
936 		prev_nents = ctx->nents;
937 
938 		*sg_cnt += ctx->nents;
939 	}
940 
941 	return 0;
942 
943 unwind:
944 	while (--i >= 0) {
945 		struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i];
946 
947 		rdma_rw_ctx_destroy(&ctx->rw, ch->qp, ch->sport->port,
948 				ctx->sg, ctx->nents, dir);
949 		target_free_sgl(ctx->sg, ctx->nents);
950 	}
951 	if (ioctx->rw_ctxs != &ioctx->s_rw_ctx)
952 		kfree(ioctx->rw_ctxs);
953 	return ret;
954 }
955 
956 static void srpt_free_rw_ctxs(struct srpt_rdma_ch *ch,
957 				    struct srpt_send_ioctx *ioctx)
958 {
959 	enum dma_data_direction dir = target_reverse_dma_direction(&ioctx->cmd);
960 	int i;
961 
962 	for (i = 0; i < ioctx->n_rw_ctx; i++) {
963 		struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i];
964 
965 		rdma_rw_ctx_destroy(&ctx->rw, ch->qp, ch->sport->port,
966 				ctx->sg, ctx->nents, dir);
967 		target_free_sgl(ctx->sg, ctx->nents);
968 	}
969 
970 	if (ioctx->rw_ctxs != &ioctx->s_rw_ctx)
971 		kfree(ioctx->rw_ctxs);
972 }
973 
974 static inline void *srpt_get_desc_buf(struct srp_cmd *srp_cmd)
975 {
976 	/*
977 	 * The pointer computations below will only be compiled correctly
978 	 * if srp_cmd::add_data is declared as s8*, u8*, s8[] or u8[], so check
979 	 * whether srp_cmd::add_data has been declared as a byte pointer.
980 	 */
981 	BUILD_BUG_ON(!__same_type(srp_cmd->add_data[0], (s8)0) &&
982 		     !__same_type(srp_cmd->add_data[0], (u8)0));
983 
984 	/*
985 	 * According to the SRP spec, the lower two bits of the 'ADDITIONAL
986 	 * CDB LENGTH' field are reserved and the size in bytes of this field
987 	 * is four times the value specified in bits 3..7. Hence the "& ~3".
988 	 */
989 	return srp_cmd->add_data + (srp_cmd->add_cdb_len & ~3);
990 }
991 
992 /**
993  * srpt_get_desc_tbl - parse the data descriptors of a SRP_CMD request
994  * @recv_ioctx: I/O context associated with the received command @srp_cmd.
995  * @ioctx: I/O context that will be used for responding to the initiator.
996  * @srp_cmd: Pointer to the SRP_CMD request data.
997  * @dir: Pointer to the variable to which the transfer direction will be
998  *   written.
999  * @sg: [out] scatterlist for the parsed SRP_CMD.
1000  * @sg_cnt: [out] length of @sg.
1001  * @data_len: Pointer to the variable to which the total data length of all
1002  *   descriptors in the SRP_CMD request will be written.
1003  * @imm_data_offset: [in] Offset in SRP_CMD requests at which immediate data
1004  *   starts.
1005  *
1006  * This function initializes ioctx->nrbuf and ioctx->r_bufs.
1007  *
1008  * Returns -EINVAL when the SRP_CMD request contains inconsistent descriptors;
1009  * -ENOMEM when memory allocation fails and zero upon success.
1010  */
1011 static int srpt_get_desc_tbl(struct srpt_recv_ioctx *recv_ioctx,
1012 		struct srpt_send_ioctx *ioctx,
1013 		struct srp_cmd *srp_cmd, enum dma_data_direction *dir,
1014 		struct scatterlist **sg, unsigned int *sg_cnt, u64 *data_len,
1015 		u16 imm_data_offset)
1016 {
1017 	BUG_ON(!dir);
1018 	BUG_ON(!data_len);
1019 
1020 	/*
1021 	 * The lower four bits of the buffer format field contain the DATA-IN
1022 	 * buffer descriptor format, and the highest four bits contain the
1023 	 * DATA-OUT buffer descriptor format.
1024 	 */
1025 	if (srp_cmd->buf_fmt & 0xf)
1026 		/* DATA-IN: transfer data from target to initiator (read). */
1027 		*dir = DMA_FROM_DEVICE;
1028 	else if (srp_cmd->buf_fmt >> 4)
1029 		/* DATA-OUT: transfer data from initiator to target (write). */
1030 		*dir = DMA_TO_DEVICE;
1031 	else
1032 		*dir = DMA_NONE;
1033 
1034 	/* initialize data_direction early as srpt_alloc_rw_ctxs needs it */
1035 	ioctx->cmd.data_direction = *dir;
1036 
1037 	if (((srp_cmd->buf_fmt & 0xf) == SRP_DATA_DESC_DIRECT) ||
1038 	    ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_DIRECT)) {
1039 		struct srp_direct_buf *db = srpt_get_desc_buf(srp_cmd);
1040 
1041 		*data_len = be32_to_cpu(db->len);
1042 		return srpt_alloc_rw_ctxs(ioctx, db, 1, sg, sg_cnt);
1043 	} else if (((srp_cmd->buf_fmt & 0xf) == SRP_DATA_DESC_INDIRECT) ||
1044 		   ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_INDIRECT)) {
1045 		struct srp_indirect_buf *idb = srpt_get_desc_buf(srp_cmd);
1046 		int nbufs = be32_to_cpu(idb->table_desc.len) /
1047 				sizeof(struct srp_direct_buf);
1048 
1049 		if (nbufs >
1050 		    (srp_cmd->data_out_desc_cnt + srp_cmd->data_in_desc_cnt)) {
1051 			pr_err("received unsupported SRP_CMD request type (%u out + %u in != %u / %zu)\n",
1052 			       srp_cmd->data_out_desc_cnt,
1053 			       srp_cmd->data_in_desc_cnt,
1054 			       be32_to_cpu(idb->table_desc.len),
1055 			       sizeof(struct srp_direct_buf));
1056 			return -EINVAL;
1057 		}
1058 
1059 		*data_len = be32_to_cpu(idb->len);
1060 		return srpt_alloc_rw_ctxs(ioctx, idb->desc_list, nbufs,
1061 				sg, sg_cnt);
1062 	} else if ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_IMM) {
1063 		struct srp_imm_buf *imm_buf = srpt_get_desc_buf(srp_cmd);
1064 		void *data = (void *)srp_cmd + imm_data_offset;
1065 		uint32_t len = be32_to_cpu(imm_buf->len);
1066 		uint32_t req_size = imm_data_offset + len;
1067 
1068 		if (req_size > srp_max_req_size) {
1069 			pr_err("Immediate data (length %d + %d) exceeds request size %d\n",
1070 			       imm_data_offset, len, srp_max_req_size);
1071 			return -EINVAL;
1072 		}
1073 		if (recv_ioctx->byte_len < req_size) {
1074 			pr_err("Received too few data - %d < %d\n",
1075 			       recv_ioctx->byte_len, req_size);
1076 			return -EIO;
1077 		}
1078 		/*
1079 		 * The immediate data buffer descriptor must occur before the
1080 		 * immediate data itself.
1081 		 */
1082 		if ((void *)(imm_buf + 1) > (void *)data) {
1083 			pr_err("Received invalid write request\n");
1084 			return -EINVAL;
1085 		}
1086 		*data_len = len;
1087 		ioctx->recv_ioctx = recv_ioctx;
1088 		if ((uintptr_t)data & 511) {
1089 			pr_warn_once("Internal error - the receive buffers are not aligned properly.\n");
1090 			return -EINVAL;
1091 		}
1092 		sg_init_one(&ioctx->imm_sg, data, len);
1093 		*sg = &ioctx->imm_sg;
1094 		*sg_cnt = 1;
1095 		return 0;
1096 	} else {
1097 		*data_len = 0;
1098 		return 0;
1099 	}
1100 }
1101 
1102 /**
1103  * srpt_init_ch_qp - initialize queue pair attributes
1104  * @ch: SRPT RDMA channel.
1105  * @qp: Queue pair pointer.
1106  *
1107  * Initialized the attributes of queue pair 'qp' by allowing local write,
1108  * remote read and remote write. Also transitions 'qp' to state IB_QPS_INIT.
1109  */
1110 static int srpt_init_ch_qp(struct srpt_rdma_ch *ch, struct ib_qp *qp)
1111 {
1112 	struct ib_qp_attr *attr;
1113 	int ret;
1114 
1115 	WARN_ON_ONCE(ch->using_rdma_cm);
1116 
1117 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
1118 	if (!attr)
1119 		return -ENOMEM;
1120 
1121 	attr->qp_state = IB_QPS_INIT;
1122 	attr->qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1123 	attr->port_num = ch->sport->port;
1124 
1125 	ret = ib_find_cached_pkey(ch->sport->sdev->device, ch->sport->port,
1126 				  ch->pkey, &attr->pkey_index);
1127 	if (ret < 0)
1128 		pr_err("Translating pkey %#x failed (%d) - using index 0\n",
1129 		       ch->pkey, ret);
1130 
1131 	ret = ib_modify_qp(qp, attr,
1132 			   IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PORT |
1133 			   IB_QP_PKEY_INDEX);
1134 
1135 	kfree(attr);
1136 	return ret;
1137 }
1138 
1139 /**
1140  * srpt_ch_qp_rtr - change the state of a channel to 'ready to receive' (RTR)
1141  * @ch: channel of the queue pair.
1142  * @qp: queue pair to change the state of.
1143  *
1144  * Returns zero upon success and a negative value upon failure.
1145  *
1146  * Note: currently a struct ib_qp_attr takes 136 bytes on a 64-bit system.
1147  * If this structure ever becomes larger, it might be necessary to allocate
1148  * it dynamically instead of on the stack.
1149  */
1150 static int srpt_ch_qp_rtr(struct srpt_rdma_ch *ch, struct ib_qp *qp)
1151 {
1152 	struct ib_qp_attr qp_attr;
1153 	int attr_mask;
1154 	int ret;
1155 
1156 	WARN_ON_ONCE(ch->using_rdma_cm);
1157 
1158 	qp_attr.qp_state = IB_QPS_RTR;
1159 	ret = ib_cm_init_qp_attr(ch->ib_cm.cm_id, &qp_attr, &attr_mask);
1160 	if (ret)
1161 		goto out;
1162 
1163 	qp_attr.max_dest_rd_atomic = 4;
1164 
1165 	ret = ib_modify_qp(qp, &qp_attr, attr_mask);
1166 
1167 out:
1168 	return ret;
1169 }
1170 
1171 /**
1172  * srpt_ch_qp_rts - change the state of a channel to 'ready to send' (RTS)
1173  * @ch: channel of the queue pair.
1174  * @qp: queue pair to change the state of.
1175  *
1176  * Returns zero upon success and a negative value upon failure.
1177  *
1178  * Note: currently a struct ib_qp_attr takes 136 bytes on a 64-bit system.
1179  * If this structure ever becomes larger, it might be necessary to allocate
1180  * it dynamically instead of on the stack.
1181  */
1182 static int srpt_ch_qp_rts(struct srpt_rdma_ch *ch, struct ib_qp *qp)
1183 {
1184 	struct ib_qp_attr qp_attr;
1185 	int attr_mask;
1186 	int ret;
1187 
1188 	qp_attr.qp_state = IB_QPS_RTS;
1189 	ret = ib_cm_init_qp_attr(ch->ib_cm.cm_id, &qp_attr, &attr_mask);
1190 	if (ret)
1191 		goto out;
1192 
1193 	qp_attr.max_rd_atomic = 4;
1194 
1195 	ret = ib_modify_qp(qp, &qp_attr, attr_mask);
1196 
1197 out:
1198 	return ret;
1199 }
1200 
1201 /**
1202  * srpt_ch_qp_err - set the channel queue pair state to 'error'
1203  * @ch: SRPT RDMA channel.
1204  */
1205 static int srpt_ch_qp_err(struct srpt_rdma_ch *ch)
1206 {
1207 	struct ib_qp_attr qp_attr;
1208 
1209 	qp_attr.qp_state = IB_QPS_ERR;
1210 	return ib_modify_qp(ch->qp, &qp_attr, IB_QP_STATE);
1211 }
1212 
1213 /**
1214  * srpt_get_send_ioctx - obtain an I/O context for sending to the initiator
1215  * @ch: SRPT RDMA channel.
1216  */
1217 static struct srpt_send_ioctx *srpt_get_send_ioctx(struct srpt_rdma_ch *ch)
1218 {
1219 	struct srpt_send_ioctx *ioctx;
1220 	int tag, cpu;
1221 
1222 	BUG_ON(!ch);
1223 
1224 	tag = sbitmap_queue_get(&ch->sess->sess_tag_pool, &cpu);
1225 	if (tag < 0)
1226 		return NULL;
1227 
1228 	ioctx = ch->ioctx_ring[tag];
1229 	BUG_ON(ioctx->ch != ch);
1230 	ioctx->state = SRPT_STATE_NEW;
1231 	WARN_ON_ONCE(ioctx->recv_ioctx);
1232 	ioctx->n_rdma = 0;
1233 	ioctx->n_rw_ctx = 0;
1234 	ioctx->queue_status_only = false;
1235 	/*
1236 	 * transport_init_se_cmd() does not initialize all fields, so do it
1237 	 * here.
1238 	 */
1239 	memset(&ioctx->cmd, 0, sizeof(ioctx->cmd));
1240 	memset(&ioctx->sense_data, 0, sizeof(ioctx->sense_data));
1241 	ioctx->cmd.map_tag = tag;
1242 	ioctx->cmd.map_cpu = cpu;
1243 
1244 	return ioctx;
1245 }
1246 
1247 /**
1248  * srpt_abort_cmd - abort a SCSI command
1249  * @ioctx:   I/O context associated with the SCSI command.
1250  */
1251 static int srpt_abort_cmd(struct srpt_send_ioctx *ioctx)
1252 {
1253 	enum srpt_command_state state;
1254 
1255 	BUG_ON(!ioctx);
1256 
1257 	/*
1258 	 * If the command is in a state where the target core is waiting for
1259 	 * the ib_srpt driver, change the state to the next state.
1260 	 */
1261 
1262 	state = ioctx->state;
1263 	switch (state) {
1264 	case SRPT_STATE_NEED_DATA:
1265 		ioctx->state = SRPT_STATE_DATA_IN;
1266 		break;
1267 	case SRPT_STATE_CMD_RSP_SENT:
1268 	case SRPT_STATE_MGMT_RSP_SENT:
1269 		ioctx->state = SRPT_STATE_DONE;
1270 		break;
1271 	default:
1272 		WARN_ONCE(true, "%s: unexpected I/O context state %d\n",
1273 			  __func__, state);
1274 		break;
1275 	}
1276 
1277 	pr_debug("Aborting cmd with state %d -> %d and tag %lld\n", state,
1278 		 ioctx->state, ioctx->cmd.tag);
1279 
1280 	switch (state) {
1281 	case SRPT_STATE_NEW:
1282 	case SRPT_STATE_DATA_IN:
1283 	case SRPT_STATE_MGMT:
1284 	case SRPT_STATE_DONE:
1285 		/*
1286 		 * Do nothing - defer abort processing until
1287 		 * srpt_queue_response() is invoked.
1288 		 */
1289 		break;
1290 	case SRPT_STATE_NEED_DATA:
1291 		pr_debug("tag %#llx: RDMA read error\n", ioctx->cmd.tag);
1292 		transport_generic_request_failure(&ioctx->cmd,
1293 					TCM_CHECK_CONDITION_ABORT_CMD);
1294 		break;
1295 	case SRPT_STATE_CMD_RSP_SENT:
1296 		/*
1297 		 * SRP_RSP sending failed or the SRP_RSP send completion has
1298 		 * not been received in time.
1299 		 */
1300 		transport_generic_free_cmd(&ioctx->cmd, 0);
1301 		break;
1302 	case SRPT_STATE_MGMT_RSP_SENT:
1303 		transport_generic_free_cmd(&ioctx->cmd, 0);
1304 		break;
1305 	default:
1306 		WARN(1, "Unexpected command state (%d)", state);
1307 		break;
1308 	}
1309 
1310 	return state;
1311 }
1312 
1313 /**
1314  * srpt_rdma_read_done - RDMA read completion callback
1315  * @cq: Completion queue.
1316  * @wc: Work completion.
1317  *
1318  * XXX: what is now target_execute_cmd used to be asynchronous, and unmapping
1319  * the data that has been transferred via IB RDMA had to be postponed until the
1320  * check_stop_free() callback.  None of this is necessary anymore and needs to
1321  * be cleaned up.
1322  */
1323 static void srpt_rdma_read_done(struct ib_cq *cq, struct ib_wc *wc)
1324 {
1325 	struct srpt_rdma_ch *ch = wc->qp->qp_context;
1326 	struct srpt_send_ioctx *ioctx =
1327 		container_of(wc->wr_cqe, struct srpt_send_ioctx, rdma_cqe);
1328 
1329 	WARN_ON(ioctx->n_rdma <= 0);
1330 	atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
1331 	ioctx->n_rdma = 0;
1332 
1333 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
1334 		pr_info("RDMA_READ for ioctx 0x%p failed with status %d\n",
1335 			ioctx, wc->status);
1336 		srpt_abort_cmd(ioctx);
1337 		return;
1338 	}
1339 
1340 	if (srpt_test_and_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA,
1341 					SRPT_STATE_DATA_IN))
1342 		target_execute_cmd(&ioctx->cmd);
1343 	else
1344 		pr_err("%s[%d]: wrong state = %d\n", __func__,
1345 		       __LINE__, ioctx->state);
1346 }
1347 
1348 /**
1349  * srpt_build_cmd_rsp - build a SRP_RSP response
1350  * @ch: RDMA channel through which the request has been received.
1351  * @ioctx: I/O context associated with the SRP_CMD request. The response will
1352  *   be built in the buffer ioctx->buf points at and hence this function will
1353  *   overwrite the request data.
1354  * @tag: tag of the request for which this response is being generated.
1355  * @status: value for the STATUS field of the SRP_RSP information unit.
1356  *
1357  * Returns the size in bytes of the SRP_RSP response.
1358  *
1359  * An SRP_RSP response contains a SCSI status or service response. See also
1360  * section 6.9 in the SRP r16a document for the format of an SRP_RSP
1361  * response. See also SPC-2 for more information about sense data.
1362  */
1363 static int srpt_build_cmd_rsp(struct srpt_rdma_ch *ch,
1364 			      struct srpt_send_ioctx *ioctx, u64 tag,
1365 			      int status)
1366 {
1367 	struct se_cmd *cmd = &ioctx->cmd;
1368 	struct srp_rsp *srp_rsp;
1369 	const u8 *sense_data;
1370 	int sense_data_len, max_sense_len;
1371 	u32 resid = cmd->residual_count;
1372 
1373 	/*
1374 	 * The lowest bit of all SAM-3 status codes is zero (see also
1375 	 * paragraph 5.3 in SAM-3).
1376 	 */
1377 	WARN_ON(status & 1);
1378 
1379 	srp_rsp = ioctx->ioctx.buf;
1380 	BUG_ON(!srp_rsp);
1381 
1382 	sense_data = ioctx->sense_data;
1383 	sense_data_len = ioctx->cmd.scsi_sense_length;
1384 	WARN_ON(sense_data_len > sizeof(ioctx->sense_data));
1385 
1386 	memset(srp_rsp, 0, sizeof(*srp_rsp));
1387 	srp_rsp->opcode = SRP_RSP;
1388 	srp_rsp->req_lim_delta =
1389 		cpu_to_be32(1 + atomic_xchg(&ch->req_lim_delta, 0));
1390 	srp_rsp->tag = tag;
1391 	srp_rsp->status = status;
1392 
1393 	if (cmd->se_cmd_flags & SCF_UNDERFLOW_BIT) {
1394 		if (cmd->data_direction == DMA_TO_DEVICE) {
1395 			/* residual data from an underflow write */
1396 			srp_rsp->flags = SRP_RSP_FLAG_DOUNDER;
1397 			srp_rsp->data_out_res_cnt = cpu_to_be32(resid);
1398 		} else if (cmd->data_direction == DMA_FROM_DEVICE) {
1399 			/* residual data from an underflow read */
1400 			srp_rsp->flags = SRP_RSP_FLAG_DIUNDER;
1401 			srp_rsp->data_in_res_cnt = cpu_to_be32(resid);
1402 		}
1403 	} else if (cmd->se_cmd_flags & SCF_OVERFLOW_BIT) {
1404 		if (cmd->data_direction == DMA_TO_DEVICE) {
1405 			/* residual data from an overflow write */
1406 			srp_rsp->flags = SRP_RSP_FLAG_DOOVER;
1407 			srp_rsp->data_out_res_cnt = cpu_to_be32(resid);
1408 		} else if (cmd->data_direction == DMA_FROM_DEVICE) {
1409 			/* residual data from an overflow read */
1410 			srp_rsp->flags = SRP_RSP_FLAG_DIOVER;
1411 			srp_rsp->data_in_res_cnt = cpu_to_be32(resid);
1412 		}
1413 	}
1414 
1415 	if (sense_data_len) {
1416 		BUILD_BUG_ON(MIN_MAX_RSP_SIZE <= sizeof(*srp_rsp));
1417 		max_sense_len = ch->max_ti_iu_len - sizeof(*srp_rsp);
1418 		if (sense_data_len > max_sense_len) {
1419 			pr_warn("truncated sense data from %d to %d bytes\n",
1420 				sense_data_len, max_sense_len);
1421 			sense_data_len = max_sense_len;
1422 		}
1423 
1424 		srp_rsp->flags |= SRP_RSP_FLAG_SNSVALID;
1425 		srp_rsp->sense_data_len = cpu_to_be32(sense_data_len);
1426 		memcpy(srp_rsp + 1, sense_data, sense_data_len);
1427 	}
1428 
1429 	return sizeof(*srp_rsp) + sense_data_len;
1430 }
1431 
1432 /**
1433  * srpt_build_tskmgmt_rsp - build a task management response
1434  * @ch:       RDMA channel through which the request has been received.
1435  * @ioctx:    I/O context in which the SRP_RSP response will be built.
1436  * @rsp_code: RSP_CODE that will be stored in the response.
1437  * @tag:      Tag of the request for which this response is being generated.
1438  *
1439  * Returns the size in bytes of the SRP_RSP response.
1440  *
1441  * An SRP_RSP response contains a SCSI status or service response. See also
1442  * section 6.9 in the SRP r16a document for the format of an SRP_RSP
1443  * response.
1444  */
1445 static int srpt_build_tskmgmt_rsp(struct srpt_rdma_ch *ch,
1446 				  struct srpt_send_ioctx *ioctx,
1447 				  u8 rsp_code, u64 tag)
1448 {
1449 	struct srp_rsp *srp_rsp;
1450 	int resp_data_len;
1451 	int resp_len;
1452 
1453 	resp_data_len = 4;
1454 	resp_len = sizeof(*srp_rsp) + resp_data_len;
1455 
1456 	srp_rsp = ioctx->ioctx.buf;
1457 	BUG_ON(!srp_rsp);
1458 	memset(srp_rsp, 0, sizeof(*srp_rsp));
1459 
1460 	srp_rsp->opcode = SRP_RSP;
1461 	srp_rsp->req_lim_delta =
1462 		cpu_to_be32(1 + atomic_xchg(&ch->req_lim_delta, 0));
1463 	srp_rsp->tag = tag;
1464 
1465 	srp_rsp->flags |= SRP_RSP_FLAG_RSPVALID;
1466 	srp_rsp->resp_data_len = cpu_to_be32(resp_data_len);
1467 	srp_rsp->data[3] = rsp_code;
1468 
1469 	return resp_len;
1470 }
1471 
1472 static int srpt_check_stop_free(struct se_cmd *cmd)
1473 {
1474 	struct srpt_send_ioctx *ioctx = container_of(cmd,
1475 				struct srpt_send_ioctx, cmd);
1476 
1477 	return target_put_sess_cmd(&ioctx->cmd);
1478 }
1479 
1480 /**
1481  * srpt_handle_cmd - process a SRP_CMD information unit
1482  * @ch: SRPT RDMA channel.
1483  * @recv_ioctx: Receive I/O context.
1484  * @send_ioctx: Send I/O context.
1485  */
1486 static void srpt_handle_cmd(struct srpt_rdma_ch *ch,
1487 			    struct srpt_recv_ioctx *recv_ioctx,
1488 			    struct srpt_send_ioctx *send_ioctx)
1489 {
1490 	struct se_cmd *cmd;
1491 	struct srp_cmd *srp_cmd;
1492 	struct scatterlist *sg = NULL;
1493 	unsigned sg_cnt = 0;
1494 	u64 data_len;
1495 	enum dma_data_direction dir;
1496 	int rc;
1497 
1498 	BUG_ON(!send_ioctx);
1499 
1500 	srp_cmd = recv_ioctx->ioctx.buf + recv_ioctx->ioctx.offset;
1501 	cmd = &send_ioctx->cmd;
1502 	cmd->tag = srp_cmd->tag;
1503 
1504 	switch (srp_cmd->task_attr) {
1505 	case SRP_CMD_SIMPLE_Q:
1506 		cmd->sam_task_attr = TCM_SIMPLE_TAG;
1507 		break;
1508 	case SRP_CMD_ORDERED_Q:
1509 	default:
1510 		cmd->sam_task_attr = TCM_ORDERED_TAG;
1511 		break;
1512 	case SRP_CMD_HEAD_OF_Q:
1513 		cmd->sam_task_attr = TCM_HEAD_TAG;
1514 		break;
1515 	case SRP_CMD_ACA:
1516 		cmd->sam_task_attr = TCM_ACA_TAG;
1517 		break;
1518 	}
1519 
1520 	rc = srpt_get_desc_tbl(recv_ioctx, send_ioctx, srp_cmd, &dir,
1521 			       &sg, &sg_cnt, &data_len, ch->imm_data_offset);
1522 	if (rc) {
1523 		if (rc != -EAGAIN) {
1524 			pr_err("0x%llx: parsing SRP descriptor table failed.\n",
1525 			       srp_cmd->tag);
1526 		}
1527 		goto busy;
1528 	}
1529 
1530 	rc = target_submit_cmd_map_sgls(cmd, ch->sess, srp_cmd->cdb,
1531 			       &send_ioctx->sense_data[0],
1532 			       scsilun_to_int(&srp_cmd->lun), data_len,
1533 			       TCM_SIMPLE_TAG, dir, TARGET_SCF_ACK_KREF,
1534 			       sg, sg_cnt, NULL, 0, NULL, 0);
1535 	if (rc != 0) {
1536 		pr_debug("target_submit_cmd() returned %d for tag %#llx\n", rc,
1537 			 srp_cmd->tag);
1538 		goto busy;
1539 	}
1540 	return;
1541 
1542 busy:
1543 	target_send_busy(cmd);
1544 }
1545 
1546 static int srp_tmr_to_tcm(int fn)
1547 {
1548 	switch (fn) {
1549 	case SRP_TSK_ABORT_TASK:
1550 		return TMR_ABORT_TASK;
1551 	case SRP_TSK_ABORT_TASK_SET:
1552 		return TMR_ABORT_TASK_SET;
1553 	case SRP_TSK_CLEAR_TASK_SET:
1554 		return TMR_CLEAR_TASK_SET;
1555 	case SRP_TSK_LUN_RESET:
1556 		return TMR_LUN_RESET;
1557 	case SRP_TSK_CLEAR_ACA:
1558 		return TMR_CLEAR_ACA;
1559 	default:
1560 		return -1;
1561 	}
1562 }
1563 
1564 /**
1565  * srpt_handle_tsk_mgmt - process a SRP_TSK_MGMT information unit
1566  * @ch: SRPT RDMA channel.
1567  * @recv_ioctx: Receive I/O context.
1568  * @send_ioctx: Send I/O context.
1569  *
1570  * Returns 0 if and only if the request will be processed by the target core.
1571  *
1572  * For more information about SRP_TSK_MGMT information units, see also section
1573  * 6.7 in the SRP r16a document.
1574  */
1575 static void srpt_handle_tsk_mgmt(struct srpt_rdma_ch *ch,
1576 				 struct srpt_recv_ioctx *recv_ioctx,
1577 				 struct srpt_send_ioctx *send_ioctx)
1578 {
1579 	struct srp_tsk_mgmt *srp_tsk;
1580 	struct se_cmd *cmd;
1581 	struct se_session *sess = ch->sess;
1582 	int tcm_tmr;
1583 	int rc;
1584 
1585 	BUG_ON(!send_ioctx);
1586 
1587 	srp_tsk = recv_ioctx->ioctx.buf + recv_ioctx->ioctx.offset;
1588 	cmd = &send_ioctx->cmd;
1589 
1590 	pr_debug("recv tsk_mgmt fn %d for task_tag %lld and cmd tag %lld ch %p sess %p\n",
1591 		 srp_tsk->tsk_mgmt_func, srp_tsk->task_tag, srp_tsk->tag, ch,
1592 		 ch->sess);
1593 
1594 	srpt_set_cmd_state(send_ioctx, SRPT_STATE_MGMT);
1595 	send_ioctx->cmd.tag = srp_tsk->tag;
1596 	tcm_tmr = srp_tmr_to_tcm(srp_tsk->tsk_mgmt_func);
1597 	rc = target_submit_tmr(&send_ioctx->cmd, sess, NULL,
1598 			       scsilun_to_int(&srp_tsk->lun), srp_tsk, tcm_tmr,
1599 			       GFP_KERNEL, srp_tsk->task_tag,
1600 			       TARGET_SCF_ACK_KREF);
1601 	if (rc != 0) {
1602 		send_ioctx->cmd.se_tmr_req->response = TMR_FUNCTION_REJECTED;
1603 		cmd->se_tfo->queue_tm_rsp(cmd);
1604 	}
1605 	return;
1606 }
1607 
1608 /**
1609  * srpt_handle_new_iu - process a newly received information unit
1610  * @ch:    RDMA channel through which the information unit has been received.
1611  * @recv_ioctx: Receive I/O context associated with the information unit.
1612  */
1613 static bool
1614 srpt_handle_new_iu(struct srpt_rdma_ch *ch, struct srpt_recv_ioctx *recv_ioctx)
1615 {
1616 	struct srpt_send_ioctx *send_ioctx = NULL;
1617 	struct srp_cmd *srp_cmd;
1618 	bool res = false;
1619 	u8 opcode;
1620 
1621 	BUG_ON(!ch);
1622 	BUG_ON(!recv_ioctx);
1623 
1624 	if (unlikely(ch->state == CH_CONNECTING))
1625 		goto push;
1626 
1627 	ib_dma_sync_single_for_cpu(ch->sport->sdev->device,
1628 				   recv_ioctx->ioctx.dma,
1629 				   recv_ioctx->ioctx.offset + srp_max_req_size,
1630 				   DMA_FROM_DEVICE);
1631 
1632 	srp_cmd = recv_ioctx->ioctx.buf + recv_ioctx->ioctx.offset;
1633 	opcode = srp_cmd->opcode;
1634 	if (opcode == SRP_CMD || opcode == SRP_TSK_MGMT) {
1635 		send_ioctx = srpt_get_send_ioctx(ch);
1636 		if (unlikely(!send_ioctx))
1637 			goto push;
1638 	}
1639 
1640 	if (!list_empty(&recv_ioctx->wait_list)) {
1641 		WARN_ON_ONCE(!ch->processing_wait_list);
1642 		list_del_init(&recv_ioctx->wait_list);
1643 	}
1644 
1645 	switch (opcode) {
1646 	case SRP_CMD:
1647 		srpt_handle_cmd(ch, recv_ioctx, send_ioctx);
1648 		break;
1649 	case SRP_TSK_MGMT:
1650 		srpt_handle_tsk_mgmt(ch, recv_ioctx, send_ioctx);
1651 		break;
1652 	case SRP_I_LOGOUT:
1653 		pr_err("Not yet implemented: SRP_I_LOGOUT\n");
1654 		break;
1655 	case SRP_CRED_RSP:
1656 		pr_debug("received SRP_CRED_RSP\n");
1657 		break;
1658 	case SRP_AER_RSP:
1659 		pr_debug("received SRP_AER_RSP\n");
1660 		break;
1661 	case SRP_RSP:
1662 		pr_err("Received SRP_RSP\n");
1663 		break;
1664 	default:
1665 		pr_err("received IU with unknown opcode 0x%x\n", opcode);
1666 		break;
1667 	}
1668 
1669 	if (!send_ioctx || !send_ioctx->recv_ioctx)
1670 		srpt_post_recv(ch->sport->sdev, ch, recv_ioctx);
1671 	res = true;
1672 
1673 out:
1674 	return res;
1675 
1676 push:
1677 	if (list_empty(&recv_ioctx->wait_list)) {
1678 		WARN_ON_ONCE(ch->processing_wait_list);
1679 		list_add_tail(&recv_ioctx->wait_list, &ch->cmd_wait_list);
1680 	}
1681 	goto out;
1682 }
1683 
1684 static void srpt_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1685 {
1686 	struct srpt_rdma_ch *ch = wc->qp->qp_context;
1687 	struct srpt_recv_ioctx *ioctx =
1688 		container_of(wc->wr_cqe, struct srpt_recv_ioctx, ioctx.cqe);
1689 
1690 	if (wc->status == IB_WC_SUCCESS) {
1691 		int req_lim;
1692 
1693 		req_lim = atomic_dec_return(&ch->req_lim);
1694 		if (unlikely(req_lim < 0))
1695 			pr_err("req_lim = %d < 0\n", req_lim);
1696 		ioctx->byte_len = wc->byte_len;
1697 		srpt_handle_new_iu(ch, ioctx);
1698 	} else {
1699 		pr_info_ratelimited("receiving failed for ioctx %p with status %d\n",
1700 				    ioctx, wc->status);
1701 	}
1702 }
1703 
1704 /*
1705  * This function must be called from the context in which RDMA completions are
1706  * processed because it accesses the wait list without protection against
1707  * access from other threads.
1708  */
1709 static void srpt_process_wait_list(struct srpt_rdma_ch *ch)
1710 {
1711 	struct srpt_recv_ioctx *recv_ioctx, *tmp;
1712 
1713 	WARN_ON_ONCE(ch->state == CH_CONNECTING);
1714 
1715 	if (list_empty(&ch->cmd_wait_list))
1716 		return;
1717 
1718 	WARN_ON_ONCE(ch->processing_wait_list);
1719 	ch->processing_wait_list = true;
1720 	list_for_each_entry_safe(recv_ioctx, tmp, &ch->cmd_wait_list,
1721 				 wait_list) {
1722 		if (!srpt_handle_new_iu(ch, recv_ioctx))
1723 			break;
1724 	}
1725 	ch->processing_wait_list = false;
1726 }
1727 
1728 /**
1729  * srpt_send_done - send completion callback
1730  * @cq: Completion queue.
1731  * @wc: Work completion.
1732  *
1733  * Note: Although this has not yet been observed during tests, at least in
1734  * theory it is possible that the srpt_get_send_ioctx() call invoked by
1735  * srpt_handle_new_iu() fails. This is possible because the req_lim_delta
1736  * value in each response is set to one, and it is possible that this response
1737  * makes the initiator send a new request before the send completion for that
1738  * response has been processed. This could e.g. happen if the call to
1739  * srpt_put_send_iotcx() is delayed because of a higher priority interrupt or
1740  * if IB retransmission causes generation of the send completion to be
1741  * delayed. Incoming information units for which srpt_get_send_ioctx() fails
1742  * are queued on cmd_wait_list. The code below processes these delayed
1743  * requests one at a time.
1744  */
1745 static void srpt_send_done(struct ib_cq *cq, struct ib_wc *wc)
1746 {
1747 	struct srpt_rdma_ch *ch = wc->qp->qp_context;
1748 	struct srpt_send_ioctx *ioctx =
1749 		container_of(wc->wr_cqe, struct srpt_send_ioctx, ioctx.cqe);
1750 	enum srpt_command_state state;
1751 
1752 	state = srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
1753 
1754 	WARN_ON(state != SRPT_STATE_CMD_RSP_SENT &&
1755 		state != SRPT_STATE_MGMT_RSP_SENT);
1756 
1757 	atomic_add(1 + ioctx->n_rdma, &ch->sq_wr_avail);
1758 
1759 	if (wc->status != IB_WC_SUCCESS)
1760 		pr_info("sending response for ioctx 0x%p failed with status %d\n",
1761 			ioctx, wc->status);
1762 
1763 	if (state != SRPT_STATE_DONE) {
1764 		transport_generic_free_cmd(&ioctx->cmd, 0);
1765 	} else {
1766 		pr_err("IB completion has been received too late for wr_id = %u.\n",
1767 		       ioctx->ioctx.index);
1768 	}
1769 
1770 	srpt_process_wait_list(ch);
1771 }
1772 
1773 /**
1774  * srpt_create_ch_ib - create receive and send completion queues
1775  * @ch: SRPT RDMA channel.
1776  */
1777 static int srpt_create_ch_ib(struct srpt_rdma_ch *ch)
1778 {
1779 	struct ib_qp_init_attr *qp_init;
1780 	struct srpt_port *sport = ch->sport;
1781 	struct srpt_device *sdev = sport->sdev;
1782 	const struct ib_device_attr *attrs = &sdev->device->attrs;
1783 	int sq_size = sport->port_attrib.srp_sq_size;
1784 	int i, ret;
1785 
1786 	WARN_ON(ch->rq_size < 1);
1787 
1788 	ret = -ENOMEM;
1789 	qp_init = kzalloc(sizeof(*qp_init), GFP_KERNEL);
1790 	if (!qp_init)
1791 		goto out;
1792 
1793 retry:
1794 	ch->cq = ib_cq_pool_get(sdev->device, ch->rq_size + sq_size, -1,
1795 				 IB_POLL_WORKQUEUE);
1796 	if (IS_ERR(ch->cq)) {
1797 		ret = PTR_ERR(ch->cq);
1798 		pr_err("failed to create CQ cqe= %d ret= %d\n",
1799 		       ch->rq_size + sq_size, ret);
1800 		goto out;
1801 	}
1802 	ch->cq_size = ch->rq_size + sq_size;
1803 
1804 	qp_init->qp_context = (void *)ch;
1805 	qp_init->event_handler
1806 		= (void(*)(struct ib_event *, void*))srpt_qp_event;
1807 	qp_init->send_cq = ch->cq;
1808 	qp_init->recv_cq = ch->cq;
1809 	qp_init->sq_sig_type = IB_SIGNAL_REQ_WR;
1810 	qp_init->qp_type = IB_QPT_RC;
1811 	/*
1812 	 * We divide up our send queue size into half SEND WRs to send the
1813 	 * completions, and half R/W contexts to actually do the RDMA
1814 	 * READ/WRITE transfers.  Note that we need to allocate CQ slots for
1815 	 * both both, as RDMA contexts will also post completions for the
1816 	 * RDMA READ case.
1817 	 */
1818 	qp_init->cap.max_send_wr = min(sq_size / 2, attrs->max_qp_wr);
1819 	qp_init->cap.max_rdma_ctxs = sq_size / 2;
1820 	qp_init->cap.max_send_sge = attrs->max_send_sge;
1821 	qp_init->cap.max_recv_sge = 1;
1822 	qp_init->port_num = ch->sport->port;
1823 	if (sdev->use_srq)
1824 		qp_init->srq = sdev->srq;
1825 	else
1826 		qp_init->cap.max_recv_wr = ch->rq_size;
1827 
1828 	if (ch->using_rdma_cm) {
1829 		ret = rdma_create_qp(ch->rdma_cm.cm_id, sdev->pd, qp_init);
1830 		ch->qp = ch->rdma_cm.cm_id->qp;
1831 	} else {
1832 		ch->qp = ib_create_qp(sdev->pd, qp_init);
1833 		if (!IS_ERR(ch->qp)) {
1834 			ret = srpt_init_ch_qp(ch, ch->qp);
1835 			if (ret)
1836 				ib_destroy_qp(ch->qp);
1837 		} else {
1838 			ret = PTR_ERR(ch->qp);
1839 		}
1840 	}
1841 	if (ret) {
1842 		bool retry = sq_size > MIN_SRPT_SQ_SIZE;
1843 
1844 		if (retry) {
1845 			pr_debug("failed to create queue pair with sq_size = %d (%d) - retrying\n",
1846 				 sq_size, ret);
1847 			ib_cq_pool_put(ch->cq, ch->cq_size);
1848 			sq_size = max(sq_size / 2, MIN_SRPT_SQ_SIZE);
1849 			goto retry;
1850 		} else {
1851 			pr_err("failed to create queue pair with sq_size = %d (%d)\n",
1852 			       sq_size, ret);
1853 			goto err_destroy_cq;
1854 		}
1855 	}
1856 
1857 	atomic_set(&ch->sq_wr_avail, qp_init->cap.max_send_wr);
1858 
1859 	pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d ch= %p\n",
1860 		 __func__, ch->cq->cqe, qp_init->cap.max_send_sge,
1861 		 qp_init->cap.max_send_wr, ch);
1862 
1863 	if (!sdev->use_srq)
1864 		for (i = 0; i < ch->rq_size; i++)
1865 			srpt_post_recv(sdev, ch, ch->ioctx_recv_ring[i]);
1866 
1867 out:
1868 	kfree(qp_init);
1869 	return ret;
1870 
1871 err_destroy_cq:
1872 	ch->qp = NULL;
1873 	ib_cq_pool_put(ch->cq, ch->cq_size);
1874 	goto out;
1875 }
1876 
1877 static void srpt_destroy_ch_ib(struct srpt_rdma_ch *ch)
1878 {
1879 	ib_destroy_qp(ch->qp);
1880 	ib_cq_pool_put(ch->cq, ch->cq_size);
1881 }
1882 
1883 /**
1884  * srpt_close_ch - close a RDMA channel
1885  * @ch: SRPT RDMA channel.
1886  *
1887  * Make sure all resources associated with the channel will be deallocated at
1888  * an appropriate time.
1889  *
1890  * Returns true if and only if the channel state has been modified into
1891  * CH_DRAINING.
1892  */
1893 static bool srpt_close_ch(struct srpt_rdma_ch *ch)
1894 {
1895 	int ret;
1896 
1897 	if (!srpt_set_ch_state(ch, CH_DRAINING)) {
1898 		pr_debug("%s: already closed\n", ch->sess_name);
1899 		return false;
1900 	}
1901 
1902 	kref_get(&ch->kref);
1903 
1904 	ret = srpt_ch_qp_err(ch);
1905 	if (ret < 0)
1906 		pr_err("%s-%d: changing queue pair into error state failed: %d\n",
1907 		       ch->sess_name, ch->qp->qp_num, ret);
1908 
1909 	ret = srpt_zerolength_write(ch);
1910 	if (ret < 0) {
1911 		pr_err("%s-%d: queuing zero-length write failed: %d\n",
1912 		       ch->sess_name, ch->qp->qp_num, ret);
1913 		if (srpt_set_ch_state(ch, CH_DISCONNECTED))
1914 			schedule_work(&ch->release_work);
1915 		else
1916 			WARN_ON_ONCE(true);
1917 	}
1918 
1919 	kref_put(&ch->kref, srpt_free_ch);
1920 
1921 	return true;
1922 }
1923 
1924 /*
1925  * Change the channel state into CH_DISCONNECTING. If a channel has not yet
1926  * reached the connected state, close it. If a channel is in the connected
1927  * state, send a DREQ. If a DREQ has been received, send a DREP. Note: it is
1928  * the responsibility of the caller to ensure that this function is not
1929  * invoked concurrently with the code that accepts a connection. This means
1930  * that this function must either be invoked from inside a CM callback
1931  * function or that it must be invoked with the srpt_port.mutex held.
1932  */
1933 static int srpt_disconnect_ch(struct srpt_rdma_ch *ch)
1934 {
1935 	int ret;
1936 
1937 	if (!srpt_set_ch_state(ch, CH_DISCONNECTING))
1938 		return -ENOTCONN;
1939 
1940 	if (ch->using_rdma_cm) {
1941 		ret = rdma_disconnect(ch->rdma_cm.cm_id);
1942 	} else {
1943 		ret = ib_send_cm_dreq(ch->ib_cm.cm_id, NULL, 0);
1944 		if (ret < 0)
1945 			ret = ib_send_cm_drep(ch->ib_cm.cm_id, NULL, 0);
1946 	}
1947 
1948 	if (ret < 0 && srpt_close_ch(ch))
1949 		ret = 0;
1950 
1951 	return ret;
1952 }
1953 
1954 /* Send DREQ and wait for DREP. */
1955 static void srpt_disconnect_ch_sync(struct srpt_rdma_ch *ch)
1956 {
1957 	DECLARE_COMPLETION_ONSTACK(closed);
1958 	struct srpt_port *sport = ch->sport;
1959 
1960 	pr_debug("ch %s-%d state %d\n", ch->sess_name, ch->qp->qp_num,
1961 		 ch->state);
1962 
1963 	ch->closed = &closed;
1964 
1965 	mutex_lock(&sport->mutex);
1966 	srpt_disconnect_ch(ch);
1967 	mutex_unlock(&sport->mutex);
1968 
1969 	while (wait_for_completion_timeout(&closed, 5 * HZ) == 0)
1970 		pr_info("%s(%s-%d state %d): still waiting ...\n", __func__,
1971 			ch->sess_name, ch->qp->qp_num, ch->state);
1972 
1973 }
1974 
1975 static void __srpt_close_all_ch(struct srpt_port *sport)
1976 {
1977 	struct srpt_nexus *nexus;
1978 	struct srpt_rdma_ch *ch;
1979 
1980 	lockdep_assert_held(&sport->mutex);
1981 
1982 	list_for_each_entry(nexus, &sport->nexus_list, entry) {
1983 		list_for_each_entry(ch, &nexus->ch_list, list) {
1984 			if (srpt_disconnect_ch(ch) >= 0)
1985 				pr_info("Closing channel %s-%d because target %s_%d has been disabled\n",
1986 					ch->sess_name, ch->qp->qp_num,
1987 					dev_name(&sport->sdev->device->dev),
1988 					sport->port);
1989 			srpt_close_ch(ch);
1990 		}
1991 	}
1992 }
1993 
1994 /*
1995  * Look up (i_port_id, t_port_id) in sport->nexus_list. Create an entry if
1996  * it does not yet exist.
1997  */
1998 static struct srpt_nexus *srpt_get_nexus(struct srpt_port *sport,
1999 					 const u8 i_port_id[16],
2000 					 const u8 t_port_id[16])
2001 {
2002 	struct srpt_nexus *nexus = NULL, *tmp_nexus = NULL, *n;
2003 
2004 	for (;;) {
2005 		mutex_lock(&sport->mutex);
2006 		list_for_each_entry(n, &sport->nexus_list, entry) {
2007 			if (memcmp(n->i_port_id, i_port_id, 16) == 0 &&
2008 			    memcmp(n->t_port_id, t_port_id, 16) == 0) {
2009 				nexus = n;
2010 				break;
2011 			}
2012 		}
2013 		if (!nexus && tmp_nexus) {
2014 			list_add_tail_rcu(&tmp_nexus->entry,
2015 					  &sport->nexus_list);
2016 			swap(nexus, tmp_nexus);
2017 		}
2018 		mutex_unlock(&sport->mutex);
2019 
2020 		if (nexus)
2021 			break;
2022 		tmp_nexus = kzalloc(sizeof(*nexus), GFP_KERNEL);
2023 		if (!tmp_nexus) {
2024 			nexus = ERR_PTR(-ENOMEM);
2025 			break;
2026 		}
2027 		INIT_LIST_HEAD(&tmp_nexus->ch_list);
2028 		memcpy(tmp_nexus->i_port_id, i_port_id, 16);
2029 		memcpy(tmp_nexus->t_port_id, t_port_id, 16);
2030 	}
2031 
2032 	kfree(tmp_nexus);
2033 
2034 	return nexus;
2035 }
2036 
2037 static void srpt_set_enabled(struct srpt_port *sport, bool enabled)
2038 	__must_hold(&sport->mutex)
2039 {
2040 	lockdep_assert_held(&sport->mutex);
2041 
2042 	if (sport->enabled == enabled)
2043 		return;
2044 	sport->enabled = enabled;
2045 	if (!enabled)
2046 		__srpt_close_all_ch(sport);
2047 }
2048 
2049 static void srpt_drop_sport_ref(struct srpt_port *sport)
2050 {
2051 	if (atomic_dec_return(&sport->refcount) == 0 && sport->freed_channels)
2052 		complete(sport->freed_channels);
2053 }
2054 
2055 static void srpt_free_ch(struct kref *kref)
2056 {
2057 	struct srpt_rdma_ch *ch = container_of(kref, struct srpt_rdma_ch, kref);
2058 
2059 	srpt_drop_sport_ref(ch->sport);
2060 	kfree_rcu(ch, rcu);
2061 }
2062 
2063 /*
2064  * Shut down the SCSI target session, tell the connection manager to
2065  * disconnect the associated RDMA channel, transition the QP to the error
2066  * state and remove the channel from the channel list. This function is
2067  * typically called from inside srpt_zerolength_write_done(). Concurrent
2068  * srpt_zerolength_write() calls from inside srpt_close_ch() are possible
2069  * as long as the channel is on sport->nexus_list.
2070  */
2071 static void srpt_release_channel_work(struct work_struct *w)
2072 {
2073 	struct srpt_rdma_ch *ch;
2074 	struct srpt_device *sdev;
2075 	struct srpt_port *sport;
2076 	struct se_session *se_sess;
2077 
2078 	ch = container_of(w, struct srpt_rdma_ch, release_work);
2079 	pr_debug("%s-%d\n", ch->sess_name, ch->qp->qp_num);
2080 
2081 	sdev = ch->sport->sdev;
2082 	BUG_ON(!sdev);
2083 
2084 	se_sess = ch->sess;
2085 	BUG_ON(!se_sess);
2086 
2087 	target_sess_cmd_list_set_waiting(se_sess);
2088 	target_wait_for_sess_cmds(se_sess);
2089 
2090 	target_remove_session(se_sess);
2091 	ch->sess = NULL;
2092 
2093 	if (ch->using_rdma_cm)
2094 		rdma_destroy_id(ch->rdma_cm.cm_id);
2095 	else
2096 		ib_destroy_cm_id(ch->ib_cm.cm_id);
2097 
2098 	sport = ch->sport;
2099 	mutex_lock(&sport->mutex);
2100 	list_del_rcu(&ch->list);
2101 	mutex_unlock(&sport->mutex);
2102 
2103 	if (ch->closed)
2104 		complete(ch->closed);
2105 
2106 	srpt_destroy_ch_ib(ch);
2107 
2108 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
2109 			     ch->sport->sdev, ch->rq_size,
2110 			     ch->rsp_buf_cache, DMA_TO_DEVICE);
2111 
2112 	kmem_cache_destroy(ch->rsp_buf_cache);
2113 
2114 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_recv_ring,
2115 			     sdev, ch->rq_size,
2116 			     ch->req_buf_cache, DMA_FROM_DEVICE);
2117 
2118 	kmem_cache_destroy(ch->req_buf_cache);
2119 
2120 	kref_put(&ch->kref, srpt_free_ch);
2121 }
2122 
2123 /**
2124  * srpt_cm_req_recv - process the event IB_CM_REQ_RECEIVED
2125  * @sdev: HCA through which the login request was received.
2126  * @ib_cm_id: IB/CM connection identifier in case of IB/CM.
2127  * @rdma_cm_id: RDMA/CM connection identifier in case of RDMA/CM.
2128  * @port_num: Port through which the REQ message was received.
2129  * @pkey: P_Key of the incoming connection.
2130  * @req: SRP login request.
2131  * @src_addr: GID (IB/CM) or IP address (RDMA/CM) of the port that submitted
2132  * the login request.
2133  *
2134  * Ownership of the cm_id is transferred to the target session if this
2135  * function returns zero. Otherwise the caller remains the owner of cm_id.
2136  */
2137 static int srpt_cm_req_recv(struct srpt_device *const sdev,
2138 			    struct ib_cm_id *ib_cm_id,
2139 			    struct rdma_cm_id *rdma_cm_id,
2140 			    u8 port_num, __be16 pkey,
2141 			    const struct srp_login_req *req,
2142 			    const char *src_addr)
2143 {
2144 	struct srpt_port *sport = &sdev->port[port_num - 1];
2145 	struct srpt_nexus *nexus;
2146 	struct srp_login_rsp *rsp = NULL;
2147 	struct srp_login_rej *rej = NULL;
2148 	union {
2149 		struct rdma_conn_param rdma_cm;
2150 		struct ib_cm_rep_param ib_cm;
2151 	} *rep_param = NULL;
2152 	struct srpt_rdma_ch *ch = NULL;
2153 	char i_port_id[36];
2154 	u32 it_iu_len;
2155 	int i, tag_num, tag_size, ret;
2156 	struct srpt_tpg *stpg;
2157 
2158 	WARN_ON_ONCE(irqs_disabled());
2159 
2160 	it_iu_len = be32_to_cpu(req->req_it_iu_len);
2161 
2162 	pr_info("Received SRP_LOGIN_REQ with i_port_id %pI6, t_port_id %pI6 and it_iu_len %d on port %d (guid=%pI6); pkey %#04x\n",
2163 		req->initiator_port_id, req->target_port_id, it_iu_len,
2164 		port_num, &sport->gid, be16_to_cpu(pkey));
2165 
2166 	nexus = srpt_get_nexus(sport, req->initiator_port_id,
2167 			       req->target_port_id);
2168 	if (IS_ERR(nexus)) {
2169 		ret = PTR_ERR(nexus);
2170 		goto out;
2171 	}
2172 
2173 	ret = -ENOMEM;
2174 	rsp = kzalloc(sizeof(*rsp), GFP_KERNEL);
2175 	rej = kzalloc(sizeof(*rej), GFP_KERNEL);
2176 	rep_param = kzalloc(sizeof(*rep_param), GFP_KERNEL);
2177 	if (!rsp || !rej || !rep_param)
2178 		goto out;
2179 
2180 	ret = -EINVAL;
2181 	if (it_iu_len > srp_max_req_size || it_iu_len < 64) {
2182 		rej->reason = cpu_to_be32(
2183 				SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE);
2184 		pr_err("rejected SRP_LOGIN_REQ because its length (%d bytes) is out of range (%d .. %d)\n",
2185 		       it_iu_len, 64, srp_max_req_size);
2186 		goto reject;
2187 	}
2188 
2189 	if (!sport->enabled) {
2190 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2191 		pr_info("rejected SRP_LOGIN_REQ because target port %s_%d has not yet been enabled\n",
2192 			dev_name(&sport->sdev->device->dev), port_num);
2193 		goto reject;
2194 	}
2195 
2196 	if (*(__be64 *)req->target_port_id != cpu_to_be64(srpt_service_guid)
2197 	    || *(__be64 *)(req->target_port_id + 8) !=
2198 	       cpu_to_be64(srpt_service_guid)) {
2199 		rej->reason = cpu_to_be32(
2200 				SRP_LOGIN_REJ_UNABLE_ASSOCIATE_CHANNEL);
2201 		pr_err("rejected SRP_LOGIN_REQ because it has an invalid target port identifier.\n");
2202 		goto reject;
2203 	}
2204 
2205 	ret = -ENOMEM;
2206 	ch = kzalloc(sizeof(*ch), GFP_KERNEL);
2207 	if (!ch) {
2208 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2209 		pr_err("rejected SRP_LOGIN_REQ because out of memory.\n");
2210 		goto reject;
2211 	}
2212 
2213 	kref_init(&ch->kref);
2214 	ch->pkey = be16_to_cpu(pkey);
2215 	ch->nexus = nexus;
2216 	ch->zw_cqe.done = srpt_zerolength_write_done;
2217 	INIT_WORK(&ch->release_work, srpt_release_channel_work);
2218 	ch->sport = sport;
2219 	if (ib_cm_id) {
2220 		ch->ib_cm.cm_id = ib_cm_id;
2221 		ib_cm_id->context = ch;
2222 	} else {
2223 		ch->using_rdma_cm = true;
2224 		ch->rdma_cm.cm_id = rdma_cm_id;
2225 		rdma_cm_id->context = ch;
2226 	}
2227 	/*
2228 	 * ch->rq_size should be at least as large as the initiator queue
2229 	 * depth to avoid that the initiator driver has to report QUEUE_FULL
2230 	 * to the SCSI mid-layer.
2231 	 */
2232 	ch->rq_size = min(MAX_SRPT_RQ_SIZE, sdev->device->attrs.max_qp_wr);
2233 	spin_lock_init(&ch->spinlock);
2234 	ch->state = CH_CONNECTING;
2235 	INIT_LIST_HEAD(&ch->cmd_wait_list);
2236 	ch->max_rsp_size = ch->sport->port_attrib.srp_max_rsp_size;
2237 
2238 	ch->rsp_buf_cache = kmem_cache_create("srpt-rsp-buf", ch->max_rsp_size,
2239 					      512, 0, NULL);
2240 	if (!ch->rsp_buf_cache)
2241 		goto free_ch;
2242 
2243 	ch->ioctx_ring = (struct srpt_send_ioctx **)
2244 		srpt_alloc_ioctx_ring(ch->sport->sdev, ch->rq_size,
2245 				      sizeof(*ch->ioctx_ring[0]),
2246 				      ch->rsp_buf_cache, 0, DMA_TO_DEVICE);
2247 	if (!ch->ioctx_ring) {
2248 		pr_err("rejected SRP_LOGIN_REQ because creating a new QP SQ ring failed.\n");
2249 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2250 		goto free_rsp_cache;
2251 	}
2252 
2253 	for (i = 0; i < ch->rq_size; i++)
2254 		ch->ioctx_ring[i]->ch = ch;
2255 	if (!sdev->use_srq) {
2256 		u16 imm_data_offset = req->req_flags & SRP_IMMED_REQUESTED ?
2257 			be16_to_cpu(req->imm_data_offset) : 0;
2258 		u16 alignment_offset;
2259 		u32 req_sz;
2260 
2261 		if (req->req_flags & SRP_IMMED_REQUESTED)
2262 			pr_debug("imm_data_offset = %d\n",
2263 				 be16_to_cpu(req->imm_data_offset));
2264 		if (imm_data_offset >= sizeof(struct srp_cmd)) {
2265 			ch->imm_data_offset = imm_data_offset;
2266 			rsp->rsp_flags |= SRP_LOGIN_RSP_IMMED_SUPP;
2267 		} else {
2268 			ch->imm_data_offset = 0;
2269 		}
2270 		alignment_offset = round_up(imm_data_offset, 512) -
2271 			imm_data_offset;
2272 		req_sz = alignment_offset + imm_data_offset + srp_max_req_size;
2273 		ch->req_buf_cache = kmem_cache_create("srpt-req-buf", req_sz,
2274 						      512, 0, NULL);
2275 		if (!ch->req_buf_cache)
2276 			goto free_rsp_ring;
2277 
2278 		ch->ioctx_recv_ring = (struct srpt_recv_ioctx **)
2279 			srpt_alloc_ioctx_ring(ch->sport->sdev, ch->rq_size,
2280 					      sizeof(*ch->ioctx_recv_ring[0]),
2281 					      ch->req_buf_cache,
2282 					      alignment_offset,
2283 					      DMA_FROM_DEVICE);
2284 		if (!ch->ioctx_recv_ring) {
2285 			pr_err("rejected SRP_LOGIN_REQ because creating a new QP RQ ring failed.\n");
2286 			rej->reason =
2287 			    cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2288 			goto free_recv_cache;
2289 		}
2290 		for (i = 0; i < ch->rq_size; i++)
2291 			INIT_LIST_HEAD(&ch->ioctx_recv_ring[i]->wait_list);
2292 	}
2293 
2294 	ret = srpt_create_ch_ib(ch);
2295 	if (ret) {
2296 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2297 		pr_err("rejected SRP_LOGIN_REQ because creating a new RDMA channel failed.\n");
2298 		goto free_recv_ring;
2299 	}
2300 
2301 	strlcpy(ch->sess_name, src_addr, sizeof(ch->sess_name));
2302 	snprintf(i_port_id, sizeof(i_port_id), "0x%016llx%016llx",
2303 			be64_to_cpu(*(__be64 *)nexus->i_port_id),
2304 			be64_to_cpu(*(__be64 *)(nexus->i_port_id + 8)));
2305 
2306 	pr_debug("registering src addr %s or i_port_id %s\n", ch->sess_name,
2307 		 i_port_id);
2308 
2309 	tag_num = ch->rq_size;
2310 	tag_size = 1; /* ib_srpt does not use se_sess->sess_cmd_map */
2311 
2312 	mutex_lock(&sport->port_guid_id.mutex);
2313 	list_for_each_entry(stpg, &sport->port_guid_id.tpg_list, entry) {
2314 		if (!IS_ERR_OR_NULL(ch->sess))
2315 			break;
2316 		ch->sess = target_setup_session(&stpg->tpg, tag_num,
2317 						tag_size, TARGET_PROT_NORMAL,
2318 						ch->sess_name, ch, NULL);
2319 	}
2320 	mutex_unlock(&sport->port_guid_id.mutex);
2321 
2322 	mutex_lock(&sport->port_gid_id.mutex);
2323 	list_for_each_entry(stpg, &sport->port_gid_id.tpg_list, entry) {
2324 		if (!IS_ERR_OR_NULL(ch->sess))
2325 			break;
2326 		ch->sess = target_setup_session(&stpg->tpg, tag_num,
2327 					tag_size, TARGET_PROT_NORMAL, i_port_id,
2328 					ch, NULL);
2329 		if (!IS_ERR_OR_NULL(ch->sess))
2330 			break;
2331 		/* Retry without leading "0x" */
2332 		ch->sess = target_setup_session(&stpg->tpg, tag_num,
2333 						tag_size, TARGET_PROT_NORMAL,
2334 						i_port_id + 2, ch, NULL);
2335 	}
2336 	mutex_unlock(&sport->port_gid_id.mutex);
2337 
2338 	if (IS_ERR_OR_NULL(ch->sess)) {
2339 		WARN_ON_ONCE(ch->sess == NULL);
2340 		ret = PTR_ERR(ch->sess);
2341 		ch->sess = NULL;
2342 		pr_info("Rejected login for initiator %s: ret = %d.\n",
2343 			ch->sess_name, ret);
2344 		rej->reason = cpu_to_be32(ret == -ENOMEM ?
2345 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES :
2346 				SRP_LOGIN_REJ_CHANNEL_LIMIT_REACHED);
2347 		goto destroy_ib;
2348 	}
2349 
2350 	/*
2351 	 * Once a session has been created destruction of srpt_rdma_ch objects
2352 	 * will decrement sport->refcount. Hence increment sport->refcount now.
2353 	 */
2354 	atomic_inc(&sport->refcount);
2355 
2356 	mutex_lock(&sport->mutex);
2357 
2358 	if ((req->req_flags & SRP_MTCH_ACTION) == SRP_MULTICHAN_SINGLE) {
2359 		struct srpt_rdma_ch *ch2;
2360 
2361 		list_for_each_entry(ch2, &nexus->ch_list, list) {
2362 			if (srpt_disconnect_ch(ch2) < 0)
2363 				continue;
2364 			pr_info("Relogin - closed existing channel %s\n",
2365 				ch2->sess_name);
2366 			rsp->rsp_flags |= SRP_LOGIN_RSP_MULTICHAN_TERMINATED;
2367 		}
2368 	} else {
2369 		rsp->rsp_flags |= SRP_LOGIN_RSP_MULTICHAN_MAINTAINED;
2370 	}
2371 
2372 	list_add_tail_rcu(&ch->list, &nexus->ch_list);
2373 
2374 	if (!sport->enabled) {
2375 		rej->reason = cpu_to_be32(
2376 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2377 		pr_info("rejected SRP_LOGIN_REQ because target %s_%d is not enabled\n",
2378 			dev_name(&sdev->device->dev), port_num);
2379 		mutex_unlock(&sport->mutex);
2380 		goto reject;
2381 	}
2382 
2383 	mutex_unlock(&sport->mutex);
2384 
2385 	ret = ch->using_rdma_cm ? 0 : srpt_ch_qp_rtr(ch, ch->qp);
2386 	if (ret) {
2387 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2388 		pr_err("rejected SRP_LOGIN_REQ because enabling RTR failed (error code = %d)\n",
2389 		       ret);
2390 		goto reject;
2391 	}
2392 
2393 	pr_debug("Establish connection sess=%p name=%s ch=%p\n", ch->sess,
2394 		 ch->sess_name, ch);
2395 
2396 	/* create srp_login_response */
2397 	rsp->opcode = SRP_LOGIN_RSP;
2398 	rsp->tag = req->tag;
2399 	rsp->max_it_iu_len = cpu_to_be32(srp_max_req_size);
2400 	rsp->max_ti_iu_len = req->req_it_iu_len;
2401 	ch->max_ti_iu_len = it_iu_len;
2402 	rsp->buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
2403 				   SRP_BUF_FORMAT_INDIRECT);
2404 	rsp->req_lim_delta = cpu_to_be32(ch->rq_size);
2405 	atomic_set(&ch->req_lim, ch->rq_size);
2406 	atomic_set(&ch->req_lim_delta, 0);
2407 
2408 	/* create cm reply */
2409 	if (ch->using_rdma_cm) {
2410 		rep_param->rdma_cm.private_data = (void *)rsp;
2411 		rep_param->rdma_cm.private_data_len = sizeof(*rsp);
2412 		rep_param->rdma_cm.rnr_retry_count = 7;
2413 		rep_param->rdma_cm.flow_control = 1;
2414 		rep_param->rdma_cm.responder_resources = 4;
2415 		rep_param->rdma_cm.initiator_depth = 4;
2416 	} else {
2417 		rep_param->ib_cm.qp_num = ch->qp->qp_num;
2418 		rep_param->ib_cm.private_data = (void *)rsp;
2419 		rep_param->ib_cm.private_data_len = sizeof(*rsp);
2420 		rep_param->ib_cm.rnr_retry_count = 7;
2421 		rep_param->ib_cm.flow_control = 1;
2422 		rep_param->ib_cm.failover_accepted = 0;
2423 		rep_param->ib_cm.srq = 1;
2424 		rep_param->ib_cm.responder_resources = 4;
2425 		rep_param->ib_cm.initiator_depth = 4;
2426 	}
2427 
2428 	/*
2429 	 * Hold the sport mutex while accepting a connection to avoid that
2430 	 * srpt_disconnect_ch() is invoked concurrently with this code.
2431 	 */
2432 	mutex_lock(&sport->mutex);
2433 	if (sport->enabled && ch->state == CH_CONNECTING) {
2434 		if (ch->using_rdma_cm)
2435 			ret = rdma_accept(rdma_cm_id, &rep_param->rdma_cm);
2436 		else
2437 			ret = ib_send_cm_rep(ib_cm_id, &rep_param->ib_cm);
2438 	} else {
2439 		ret = -EINVAL;
2440 	}
2441 	mutex_unlock(&sport->mutex);
2442 
2443 	switch (ret) {
2444 	case 0:
2445 		break;
2446 	case -EINVAL:
2447 		goto reject;
2448 	default:
2449 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2450 		pr_err("sending SRP_LOGIN_REQ response failed (error code = %d)\n",
2451 		       ret);
2452 		goto reject;
2453 	}
2454 
2455 	goto out;
2456 
2457 destroy_ib:
2458 	srpt_destroy_ch_ib(ch);
2459 
2460 free_recv_ring:
2461 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_recv_ring,
2462 			     ch->sport->sdev, ch->rq_size,
2463 			     ch->req_buf_cache, DMA_FROM_DEVICE);
2464 
2465 free_recv_cache:
2466 	kmem_cache_destroy(ch->req_buf_cache);
2467 
2468 free_rsp_ring:
2469 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
2470 			     ch->sport->sdev, ch->rq_size,
2471 			     ch->rsp_buf_cache, DMA_TO_DEVICE);
2472 
2473 free_rsp_cache:
2474 	kmem_cache_destroy(ch->rsp_buf_cache);
2475 
2476 free_ch:
2477 	if (rdma_cm_id)
2478 		rdma_cm_id->context = NULL;
2479 	else
2480 		ib_cm_id->context = NULL;
2481 	kfree(ch);
2482 	ch = NULL;
2483 
2484 	WARN_ON_ONCE(ret == 0);
2485 
2486 reject:
2487 	pr_info("Rejecting login with reason %#x\n", be32_to_cpu(rej->reason));
2488 	rej->opcode = SRP_LOGIN_REJ;
2489 	rej->tag = req->tag;
2490 	rej->buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
2491 				   SRP_BUF_FORMAT_INDIRECT);
2492 
2493 	if (rdma_cm_id)
2494 		rdma_reject(rdma_cm_id, rej, sizeof(*rej),
2495 			    IB_CM_REJ_CONSUMER_DEFINED);
2496 	else
2497 		ib_send_cm_rej(ib_cm_id, IB_CM_REJ_CONSUMER_DEFINED, NULL, 0,
2498 			       rej, sizeof(*rej));
2499 
2500 	if (ch && ch->sess) {
2501 		srpt_close_ch(ch);
2502 		/*
2503 		 * Tell the caller not to free cm_id since
2504 		 * srpt_release_channel_work() will do that.
2505 		 */
2506 		ret = 0;
2507 	}
2508 
2509 out:
2510 	kfree(rep_param);
2511 	kfree(rsp);
2512 	kfree(rej);
2513 
2514 	return ret;
2515 }
2516 
2517 static int srpt_ib_cm_req_recv(struct ib_cm_id *cm_id,
2518 			       const struct ib_cm_req_event_param *param,
2519 			       void *private_data)
2520 {
2521 	char sguid[40];
2522 
2523 	srpt_format_guid(sguid, sizeof(sguid),
2524 			 &param->primary_path->dgid.global.interface_id);
2525 
2526 	return srpt_cm_req_recv(cm_id->context, cm_id, NULL, param->port,
2527 				param->primary_path->pkey,
2528 				private_data, sguid);
2529 }
2530 
2531 static int srpt_rdma_cm_req_recv(struct rdma_cm_id *cm_id,
2532 				 struct rdma_cm_event *event)
2533 {
2534 	struct srpt_device *sdev;
2535 	struct srp_login_req req;
2536 	const struct srp_login_req_rdma *req_rdma;
2537 	struct sa_path_rec *path_rec = cm_id->route.path_rec;
2538 	char src_addr[40];
2539 
2540 	sdev = ib_get_client_data(cm_id->device, &srpt_client);
2541 	if (!sdev)
2542 		return -ECONNREFUSED;
2543 
2544 	if (event->param.conn.private_data_len < sizeof(*req_rdma))
2545 		return -EINVAL;
2546 
2547 	/* Transform srp_login_req_rdma into srp_login_req. */
2548 	req_rdma = event->param.conn.private_data;
2549 	memset(&req, 0, sizeof(req));
2550 	req.opcode		= req_rdma->opcode;
2551 	req.tag			= req_rdma->tag;
2552 	req.req_it_iu_len	= req_rdma->req_it_iu_len;
2553 	req.req_buf_fmt		= req_rdma->req_buf_fmt;
2554 	req.req_flags		= req_rdma->req_flags;
2555 	memcpy(req.initiator_port_id, req_rdma->initiator_port_id, 16);
2556 	memcpy(req.target_port_id, req_rdma->target_port_id, 16);
2557 	req.imm_data_offset	= req_rdma->imm_data_offset;
2558 
2559 	snprintf(src_addr, sizeof(src_addr), "%pIS",
2560 		 &cm_id->route.addr.src_addr);
2561 
2562 	return srpt_cm_req_recv(sdev, NULL, cm_id, cm_id->port_num,
2563 				path_rec ? path_rec->pkey : 0, &req, src_addr);
2564 }
2565 
2566 static void srpt_cm_rej_recv(struct srpt_rdma_ch *ch,
2567 			     enum ib_cm_rej_reason reason,
2568 			     const u8 *private_data,
2569 			     u8 private_data_len)
2570 {
2571 	char *priv = NULL;
2572 	int i;
2573 
2574 	if (private_data_len && (priv = kmalloc(private_data_len * 3 + 1,
2575 						GFP_KERNEL))) {
2576 		for (i = 0; i < private_data_len; i++)
2577 			sprintf(priv + 3 * i, " %02x", private_data[i]);
2578 	}
2579 	pr_info("Received CM REJ for ch %s-%d; reason %d%s%s.\n",
2580 		ch->sess_name, ch->qp->qp_num, reason, private_data_len ?
2581 		"; private data" : "", priv ? priv : " (?)");
2582 	kfree(priv);
2583 }
2584 
2585 /**
2586  * srpt_cm_rtu_recv - process an IB_CM_RTU_RECEIVED or USER_ESTABLISHED event
2587  * @ch: SRPT RDMA channel.
2588  *
2589  * An RTU (ready to use) message indicates that the connection has been
2590  * established and that the recipient may begin transmitting.
2591  */
2592 static void srpt_cm_rtu_recv(struct srpt_rdma_ch *ch)
2593 {
2594 	int ret;
2595 
2596 	ret = ch->using_rdma_cm ? 0 : srpt_ch_qp_rts(ch, ch->qp);
2597 	if (ret < 0) {
2598 		pr_err("%s-%d: QP transition to RTS failed\n", ch->sess_name,
2599 		       ch->qp->qp_num);
2600 		srpt_close_ch(ch);
2601 		return;
2602 	}
2603 
2604 	/*
2605 	 * Note: calling srpt_close_ch() if the transition to the LIVE state
2606 	 * fails is not necessary since that means that that function has
2607 	 * already been invoked from another thread.
2608 	 */
2609 	if (!srpt_set_ch_state(ch, CH_LIVE)) {
2610 		pr_err("%s-%d: channel transition to LIVE state failed\n",
2611 		       ch->sess_name, ch->qp->qp_num);
2612 		return;
2613 	}
2614 
2615 	/* Trigger wait list processing. */
2616 	ret = srpt_zerolength_write(ch);
2617 	WARN_ONCE(ret < 0, "%d\n", ret);
2618 }
2619 
2620 /**
2621  * srpt_cm_handler - IB connection manager callback function
2622  * @cm_id: IB/CM connection identifier.
2623  * @event: IB/CM event.
2624  *
2625  * A non-zero return value will cause the caller destroy the CM ID.
2626  *
2627  * Note: srpt_cm_handler() must only return a non-zero value when transferring
2628  * ownership of the cm_id to a channel by srpt_cm_req_recv() failed. Returning
2629  * a non-zero value in any other case will trigger a race with the
2630  * ib_destroy_cm_id() call in srpt_release_channel().
2631  */
2632 static int srpt_cm_handler(struct ib_cm_id *cm_id,
2633 			   const struct ib_cm_event *event)
2634 {
2635 	struct srpt_rdma_ch *ch = cm_id->context;
2636 	int ret;
2637 
2638 	ret = 0;
2639 	switch (event->event) {
2640 	case IB_CM_REQ_RECEIVED:
2641 		ret = srpt_ib_cm_req_recv(cm_id, &event->param.req_rcvd,
2642 					  event->private_data);
2643 		break;
2644 	case IB_CM_REJ_RECEIVED:
2645 		srpt_cm_rej_recv(ch, event->param.rej_rcvd.reason,
2646 				 event->private_data,
2647 				 IB_CM_REJ_PRIVATE_DATA_SIZE);
2648 		break;
2649 	case IB_CM_RTU_RECEIVED:
2650 	case IB_CM_USER_ESTABLISHED:
2651 		srpt_cm_rtu_recv(ch);
2652 		break;
2653 	case IB_CM_DREQ_RECEIVED:
2654 		srpt_disconnect_ch(ch);
2655 		break;
2656 	case IB_CM_DREP_RECEIVED:
2657 		pr_info("Received CM DREP message for ch %s-%d.\n",
2658 			ch->sess_name, ch->qp->qp_num);
2659 		srpt_close_ch(ch);
2660 		break;
2661 	case IB_CM_TIMEWAIT_EXIT:
2662 		pr_info("Received CM TimeWait exit for ch %s-%d.\n",
2663 			ch->sess_name, ch->qp->qp_num);
2664 		srpt_close_ch(ch);
2665 		break;
2666 	case IB_CM_REP_ERROR:
2667 		pr_info("Received CM REP error for ch %s-%d.\n", ch->sess_name,
2668 			ch->qp->qp_num);
2669 		break;
2670 	case IB_CM_DREQ_ERROR:
2671 		pr_info("Received CM DREQ ERROR event.\n");
2672 		break;
2673 	case IB_CM_MRA_RECEIVED:
2674 		pr_info("Received CM MRA event\n");
2675 		break;
2676 	default:
2677 		pr_err("received unrecognized CM event %d\n", event->event);
2678 		break;
2679 	}
2680 
2681 	return ret;
2682 }
2683 
2684 static int srpt_rdma_cm_handler(struct rdma_cm_id *cm_id,
2685 				struct rdma_cm_event *event)
2686 {
2687 	struct srpt_rdma_ch *ch = cm_id->context;
2688 	int ret = 0;
2689 
2690 	switch (event->event) {
2691 	case RDMA_CM_EVENT_CONNECT_REQUEST:
2692 		ret = srpt_rdma_cm_req_recv(cm_id, event);
2693 		break;
2694 	case RDMA_CM_EVENT_REJECTED:
2695 		srpt_cm_rej_recv(ch, event->status,
2696 				 event->param.conn.private_data,
2697 				 event->param.conn.private_data_len);
2698 		break;
2699 	case RDMA_CM_EVENT_ESTABLISHED:
2700 		srpt_cm_rtu_recv(ch);
2701 		break;
2702 	case RDMA_CM_EVENT_DISCONNECTED:
2703 		if (ch->state < CH_DISCONNECTING)
2704 			srpt_disconnect_ch(ch);
2705 		else
2706 			srpt_close_ch(ch);
2707 		break;
2708 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
2709 		srpt_close_ch(ch);
2710 		break;
2711 	case RDMA_CM_EVENT_UNREACHABLE:
2712 		pr_info("Received CM REP error for ch %s-%d.\n", ch->sess_name,
2713 			ch->qp->qp_num);
2714 		break;
2715 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
2716 	case RDMA_CM_EVENT_ADDR_CHANGE:
2717 		break;
2718 	default:
2719 		pr_err("received unrecognized RDMA CM event %d\n",
2720 		       event->event);
2721 		break;
2722 	}
2723 
2724 	return ret;
2725 }
2726 
2727 /*
2728  * srpt_write_pending - Start data transfer from initiator to target (write).
2729  */
2730 static int srpt_write_pending(struct se_cmd *se_cmd)
2731 {
2732 	struct srpt_send_ioctx *ioctx =
2733 		container_of(se_cmd, struct srpt_send_ioctx, cmd);
2734 	struct srpt_rdma_ch *ch = ioctx->ch;
2735 	struct ib_send_wr *first_wr = NULL;
2736 	struct ib_cqe *cqe = &ioctx->rdma_cqe;
2737 	enum srpt_command_state new_state;
2738 	int ret, i;
2739 
2740 	if (ioctx->recv_ioctx) {
2741 		srpt_set_cmd_state(ioctx, SRPT_STATE_DATA_IN);
2742 		target_execute_cmd(&ioctx->cmd);
2743 		return 0;
2744 	}
2745 
2746 	new_state = srpt_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA);
2747 	WARN_ON(new_state == SRPT_STATE_DONE);
2748 
2749 	if (atomic_sub_return(ioctx->n_rdma, &ch->sq_wr_avail) < 0) {
2750 		pr_warn("%s: IB send queue full (needed %d)\n",
2751 				__func__, ioctx->n_rdma);
2752 		ret = -ENOMEM;
2753 		goto out_undo;
2754 	}
2755 
2756 	cqe->done = srpt_rdma_read_done;
2757 	for (i = ioctx->n_rw_ctx - 1; i >= 0; i--) {
2758 		struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i];
2759 
2760 		first_wr = rdma_rw_ctx_wrs(&ctx->rw, ch->qp, ch->sport->port,
2761 				cqe, first_wr);
2762 		cqe = NULL;
2763 	}
2764 
2765 	ret = ib_post_send(ch->qp, first_wr, NULL);
2766 	if (ret) {
2767 		pr_err("%s: ib_post_send() returned %d for %d (avail: %d)\n",
2768 			 __func__, ret, ioctx->n_rdma,
2769 			 atomic_read(&ch->sq_wr_avail));
2770 		goto out_undo;
2771 	}
2772 
2773 	return 0;
2774 out_undo:
2775 	atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
2776 	return ret;
2777 }
2778 
2779 static u8 tcm_to_srp_tsk_mgmt_status(const int tcm_mgmt_status)
2780 {
2781 	switch (tcm_mgmt_status) {
2782 	case TMR_FUNCTION_COMPLETE:
2783 		return SRP_TSK_MGMT_SUCCESS;
2784 	case TMR_FUNCTION_REJECTED:
2785 		return SRP_TSK_MGMT_FUNC_NOT_SUPP;
2786 	}
2787 	return SRP_TSK_MGMT_FAILED;
2788 }
2789 
2790 /**
2791  * srpt_queue_response - transmit the response to a SCSI command
2792  * @cmd: SCSI target command.
2793  *
2794  * Callback function called by the TCM core. Must not block since it can be
2795  * invoked on the context of the IB completion handler.
2796  */
2797 static void srpt_queue_response(struct se_cmd *cmd)
2798 {
2799 	struct srpt_send_ioctx *ioctx =
2800 		container_of(cmd, struct srpt_send_ioctx, cmd);
2801 	struct srpt_rdma_ch *ch = ioctx->ch;
2802 	struct srpt_device *sdev = ch->sport->sdev;
2803 	struct ib_send_wr send_wr, *first_wr = &send_wr;
2804 	struct ib_sge sge;
2805 	enum srpt_command_state state;
2806 	int resp_len, ret, i;
2807 	u8 srp_tm_status;
2808 
2809 	state = ioctx->state;
2810 	switch (state) {
2811 	case SRPT_STATE_NEW:
2812 	case SRPT_STATE_DATA_IN:
2813 		ioctx->state = SRPT_STATE_CMD_RSP_SENT;
2814 		break;
2815 	case SRPT_STATE_MGMT:
2816 		ioctx->state = SRPT_STATE_MGMT_RSP_SENT;
2817 		break;
2818 	default:
2819 		WARN(true, "ch %p; cmd %d: unexpected command state %d\n",
2820 			ch, ioctx->ioctx.index, ioctx->state);
2821 		break;
2822 	}
2823 
2824 	if (WARN_ON_ONCE(state == SRPT_STATE_CMD_RSP_SENT))
2825 		return;
2826 
2827 	/* For read commands, transfer the data to the initiator. */
2828 	if (ioctx->cmd.data_direction == DMA_FROM_DEVICE &&
2829 	    ioctx->cmd.data_length &&
2830 	    !ioctx->queue_status_only) {
2831 		for (i = ioctx->n_rw_ctx - 1; i >= 0; i--) {
2832 			struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i];
2833 
2834 			first_wr = rdma_rw_ctx_wrs(&ctx->rw, ch->qp,
2835 					ch->sport->port, NULL, first_wr);
2836 		}
2837 	}
2838 
2839 	if (state != SRPT_STATE_MGMT)
2840 		resp_len = srpt_build_cmd_rsp(ch, ioctx, ioctx->cmd.tag,
2841 					      cmd->scsi_status);
2842 	else {
2843 		srp_tm_status
2844 			= tcm_to_srp_tsk_mgmt_status(cmd->se_tmr_req->response);
2845 		resp_len = srpt_build_tskmgmt_rsp(ch, ioctx, srp_tm_status,
2846 						 ioctx->cmd.tag);
2847 	}
2848 
2849 	atomic_inc(&ch->req_lim);
2850 
2851 	if (unlikely(atomic_sub_return(1 + ioctx->n_rdma,
2852 			&ch->sq_wr_avail) < 0)) {
2853 		pr_warn("%s: IB send queue full (needed %d)\n",
2854 				__func__, ioctx->n_rdma);
2855 		ret = -ENOMEM;
2856 		goto out;
2857 	}
2858 
2859 	ib_dma_sync_single_for_device(sdev->device, ioctx->ioctx.dma, resp_len,
2860 				      DMA_TO_DEVICE);
2861 
2862 	sge.addr = ioctx->ioctx.dma;
2863 	sge.length = resp_len;
2864 	sge.lkey = sdev->lkey;
2865 
2866 	ioctx->ioctx.cqe.done = srpt_send_done;
2867 	send_wr.next = NULL;
2868 	send_wr.wr_cqe = &ioctx->ioctx.cqe;
2869 	send_wr.sg_list = &sge;
2870 	send_wr.num_sge = 1;
2871 	send_wr.opcode = IB_WR_SEND;
2872 	send_wr.send_flags = IB_SEND_SIGNALED;
2873 
2874 	ret = ib_post_send(ch->qp, first_wr, NULL);
2875 	if (ret < 0) {
2876 		pr_err("%s: sending cmd response failed for tag %llu (%d)\n",
2877 			__func__, ioctx->cmd.tag, ret);
2878 		goto out;
2879 	}
2880 
2881 	return;
2882 
2883 out:
2884 	atomic_add(1 + ioctx->n_rdma, &ch->sq_wr_avail);
2885 	atomic_dec(&ch->req_lim);
2886 	srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
2887 	target_put_sess_cmd(&ioctx->cmd);
2888 }
2889 
2890 static int srpt_queue_data_in(struct se_cmd *cmd)
2891 {
2892 	srpt_queue_response(cmd);
2893 	return 0;
2894 }
2895 
2896 static void srpt_queue_tm_rsp(struct se_cmd *cmd)
2897 {
2898 	srpt_queue_response(cmd);
2899 }
2900 
2901 /*
2902  * This function is called for aborted commands if no response is sent to the
2903  * initiator. Make sure that the credits freed by aborting a command are
2904  * returned to the initiator the next time a response is sent by incrementing
2905  * ch->req_lim_delta.
2906  */
2907 static void srpt_aborted_task(struct se_cmd *cmd)
2908 {
2909 	struct srpt_send_ioctx *ioctx = container_of(cmd,
2910 				struct srpt_send_ioctx, cmd);
2911 	struct srpt_rdma_ch *ch = ioctx->ch;
2912 
2913 	atomic_inc(&ch->req_lim_delta);
2914 }
2915 
2916 static int srpt_queue_status(struct se_cmd *cmd)
2917 {
2918 	struct srpt_send_ioctx *ioctx;
2919 
2920 	ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
2921 	BUG_ON(ioctx->sense_data != cmd->sense_buffer);
2922 	if (cmd->se_cmd_flags &
2923 	    (SCF_TRANSPORT_TASK_SENSE | SCF_EMULATED_TASK_SENSE))
2924 		WARN_ON(cmd->scsi_status != SAM_STAT_CHECK_CONDITION);
2925 	ioctx->queue_status_only = true;
2926 	srpt_queue_response(cmd);
2927 	return 0;
2928 }
2929 
2930 static void srpt_refresh_port_work(struct work_struct *work)
2931 {
2932 	struct srpt_port *sport = container_of(work, struct srpt_port, work);
2933 
2934 	srpt_refresh_port(sport);
2935 }
2936 
2937 /**
2938  * srpt_release_sport - disable login and wait for associated channels
2939  * @sport: SRPT HCA port.
2940  */
2941 static int srpt_release_sport(struct srpt_port *sport)
2942 {
2943 	DECLARE_COMPLETION_ONSTACK(c);
2944 	struct srpt_nexus *nexus, *next_n;
2945 	struct srpt_rdma_ch *ch;
2946 
2947 	WARN_ON_ONCE(irqs_disabled());
2948 
2949 	sport->freed_channels = &c;
2950 
2951 	mutex_lock(&sport->mutex);
2952 	srpt_set_enabled(sport, false);
2953 	mutex_unlock(&sport->mutex);
2954 
2955 	while (atomic_read(&sport->refcount) > 0 &&
2956 	       wait_for_completion_timeout(&c, 5 * HZ) <= 0) {
2957 		pr_info("%s_%d: waiting for unregistration of %d sessions ...\n",
2958 			dev_name(&sport->sdev->device->dev), sport->port,
2959 			atomic_read(&sport->refcount));
2960 		rcu_read_lock();
2961 		list_for_each_entry(nexus, &sport->nexus_list, entry) {
2962 			list_for_each_entry(ch, &nexus->ch_list, list) {
2963 				pr_info("%s-%d: state %s\n",
2964 					ch->sess_name, ch->qp->qp_num,
2965 					get_ch_state_name(ch->state));
2966 			}
2967 		}
2968 		rcu_read_unlock();
2969 	}
2970 
2971 	mutex_lock(&sport->mutex);
2972 	list_for_each_entry_safe(nexus, next_n, &sport->nexus_list, entry) {
2973 		list_del(&nexus->entry);
2974 		kfree_rcu(nexus, rcu);
2975 	}
2976 	mutex_unlock(&sport->mutex);
2977 
2978 	return 0;
2979 }
2980 
2981 static struct se_wwn *__srpt_lookup_wwn(const char *name)
2982 {
2983 	struct ib_device *dev;
2984 	struct srpt_device *sdev;
2985 	struct srpt_port *sport;
2986 	int i;
2987 
2988 	list_for_each_entry(sdev, &srpt_dev_list, list) {
2989 		dev = sdev->device;
2990 		if (!dev)
2991 			continue;
2992 
2993 		for (i = 0; i < dev->phys_port_cnt; i++) {
2994 			sport = &sdev->port[i];
2995 
2996 			if (strcmp(sport->port_guid_id.name, name) == 0)
2997 				return &sport->port_guid_id.wwn;
2998 			if (strcmp(sport->port_gid_id.name, name) == 0)
2999 				return &sport->port_gid_id.wwn;
3000 		}
3001 	}
3002 
3003 	return NULL;
3004 }
3005 
3006 static struct se_wwn *srpt_lookup_wwn(const char *name)
3007 {
3008 	struct se_wwn *wwn;
3009 
3010 	spin_lock(&srpt_dev_lock);
3011 	wwn = __srpt_lookup_wwn(name);
3012 	spin_unlock(&srpt_dev_lock);
3013 
3014 	return wwn;
3015 }
3016 
3017 static void srpt_free_srq(struct srpt_device *sdev)
3018 {
3019 	if (!sdev->srq)
3020 		return;
3021 
3022 	ib_destroy_srq(sdev->srq);
3023 	srpt_free_ioctx_ring((struct srpt_ioctx **)sdev->ioctx_ring, sdev,
3024 			     sdev->srq_size, sdev->req_buf_cache,
3025 			     DMA_FROM_DEVICE);
3026 	kmem_cache_destroy(sdev->req_buf_cache);
3027 	sdev->srq = NULL;
3028 }
3029 
3030 static int srpt_alloc_srq(struct srpt_device *sdev)
3031 {
3032 	struct ib_srq_init_attr srq_attr = {
3033 		.event_handler = srpt_srq_event,
3034 		.srq_context = (void *)sdev,
3035 		.attr.max_wr = sdev->srq_size,
3036 		.attr.max_sge = 1,
3037 		.srq_type = IB_SRQT_BASIC,
3038 	};
3039 	struct ib_device *device = sdev->device;
3040 	struct ib_srq *srq;
3041 	int i;
3042 
3043 	WARN_ON_ONCE(sdev->srq);
3044 	srq = ib_create_srq(sdev->pd, &srq_attr);
3045 	if (IS_ERR(srq)) {
3046 		pr_debug("ib_create_srq() failed: %ld\n", PTR_ERR(srq));
3047 		return PTR_ERR(srq);
3048 	}
3049 
3050 	pr_debug("create SRQ #wr= %d max_allow=%d dev= %s\n", sdev->srq_size,
3051 		 sdev->device->attrs.max_srq_wr, dev_name(&device->dev));
3052 
3053 	sdev->req_buf_cache = kmem_cache_create("srpt-srq-req-buf",
3054 						srp_max_req_size, 0, 0, NULL);
3055 	if (!sdev->req_buf_cache)
3056 		goto free_srq;
3057 
3058 	sdev->ioctx_ring = (struct srpt_recv_ioctx **)
3059 		srpt_alloc_ioctx_ring(sdev, sdev->srq_size,
3060 				      sizeof(*sdev->ioctx_ring[0]),
3061 				      sdev->req_buf_cache, 0, DMA_FROM_DEVICE);
3062 	if (!sdev->ioctx_ring)
3063 		goto free_cache;
3064 
3065 	sdev->use_srq = true;
3066 	sdev->srq = srq;
3067 
3068 	for (i = 0; i < sdev->srq_size; ++i) {
3069 		INIT_LIST_HEAD(&sdev->ioctx_ring[i]->wait_list);
3070 		srpt_post_recv(sdev, NULL, sdev->ioctx_ring[i]);
3071 	}
3072 
3073 	return 0;
3074 
3075 free_cache:
3076 	kmem_cache_destroy(sdev->req_buf_cache);
3077 
3078 free_srq:
3079 	ib_destroy_srq(srq);
3080 	return -ENOMEM;
3081 }
3082 
3083 static int srpt_use_srq(struct srpt_device *sdev, bool use_srq)
3084 {
3085 	struct ib_device *device = sdev->device;
3086 	int ret = 0;
3087 
3088 	if (!use_srq) {
3089 		srpt_free_srq(sdev);
3090 		sdev->use_srq = false;
3091 	} else if (use_srq && !sdev->srq) {
3092 		ret = srpt_alloc_srq(sdev);
3093 	}
3094 	pr_debug("%s(%s): use_srq = %d; ret = %d\n", __func__,
3095 		 dev_name(&device->dev), sdev->use_srq, ret);
3096 	return ret;
3097 }
3098 
3099 /**
3100  * srpt_add_one - InfiniBand device addition callback function
3101  * @device: Describes a HCA.
3102  */
3103 static int srpt_add_one(struct ib_device *device)
3104 {
3105 	struct srpt_device *sdev;
3106 	struct srpt_port *sport;
3107 	int i, ret;
3108 
3109 	pr_debug("device = %p\n", device);
3110 
3111 	sdev = kzalloc(struct_size(sdev, port, device->phys_port_cnt),
3112 		       GFP_KERNEL);
3113 	if (!sdev)
3114 		return -ENOMEM;
3115 
3116 	sdev->device = device;
3117 	mutex_init(&sdev->sdev_mutex);
3118 
3119 	sdev->pd = ib_alloc_pd(device, 0);
3120 	if (IS_ERR(sdev->pd)) {
3121 		ret = PTR_ERR(sdev->pd);
3122 		goto free_dev;
3123 	}
3124 
3125 	sdev->lkey = sdev->pd->local_dma_lkey;
3126 
3127 	sdev->srq_size = min(srpt_srq_size, sdev->device->attrs.max_srq_wr);
3128 
3129 	srpt_use_srq(sdev, sdev->port[0].port_attrib.use_srq);
3130 
3131 	if (!srpt_service_guid)
3132 		srpt_service_guid = be64_to_cpu(device->node_guid);
3133 
3134 	if (rdma_port_get_link_layer(device, 1) == IB_LINK_LAYER_INFINIBAND)
3135 		sdev->cm_id = ib_create_cm_id(device, srpt_cm_handler, sdev);
3136 	if (IS_ERR(sdev->cm_id)) {
3137 		pr_info("ib_create_cm_id() failed: %ld\n",
3138 			PTR_ERR(sdev->cm_id));
3139 		ret = PTR_ERR(sdev->cm_id);
3140 		sdev->cm_id = NULL;
3141 		if (!rdma_cm_id)
3142 			goto err_ring;
3143 	}
3144 
3145 	/* print out target login information */
3146 	pr_debug("Target login info: id_ext=%016llx,ioc_guid=%016llx,pkey=ffff,service_id=%016llx\n",
3147 		 srpt_service_guid, srpt_service_guid, srpt_service_guid);
3148 
3149 	/*
3150 	 * We do not have a consistent service_id (ie. also id_ext of target_id)
3151 	 * to identify this target. We currently use the guid of the first HCA
3152 	 * in the system as service_id; therefore, the target_id will change
3153 	 * if this HCA is gone bad and replaced by different HCA
3154 	 */
3155 	ret = sdev->cm_id ?
3156 		ib_cm_listen(sdev->cm_id, cpu_to_be64(srpt_service_guid), 0) :
3157 		0;
3158 	if (ret < 0) {
3159 		pr_err("ib_cm_listen() failed: %d (cm_id state = %d)\n", ret,
3160 		       sdev->cm_id->state);
3161 		goto err_cm;
3162 	}
3163 
3164 	INIT_IB_EVENT_HANDLER(&sdev->event_handler, sdev->device,
3165 			      srpt_event_handler);
3166 	ib_register_event_handler(&sdev->event_handler);
3167 
3168 	for (i = 1; i <= sdev->device->phys_port_cnt; i++) {
3169 		sport = &sdev->port[i - 1];
3170 		INIT_LIST_HEAD(&sport->nexus_list);
3171 		mutex_init(&sport->mutex);
3172 		sport->sdev = sdev;
3173 		sport->port = i;
3174 		sport->port_attrib.srp_max_rdma_size = DEFAULT_MAX_RDMA_SIZE;
3175 		sport->port_attrib.srp_max_rsp_size = DEFAULT_MAX_RSP_SIZE;
3176 		sport->port_attrib.srp_sq_size = DEF_SRPT_SQ_SIZE;
3177 		sport->port_attrib.use_srq = false;
3178 		INIT_WORK(&sport->work, srpt_refresh_port_work);
3179 		mutex_init(&sport->port_guid_id.mutex);
3180 		INIT_LIST_HEAD(&sport->port_guid_id.tpg_list);
3181 		mutex_init(&sport->port_gid_id.mutex);
3182 		INIT_LIST_HEAD(&sport->port_gid_id.tpg_list);
3183 
3184 		ret = srpt_refresh_port(sport);
3185 		if (ret) {
3186 			pr_err("MAD registration failed for %s-%d.\n",
3187 			       dev_name(&sdev->device->dev), i);
3188 			goto err_event;
3189 		}
3190 	}
3191 
3192 	spin_lock(&srpt_dev_lock);
3193 	list_add_tail(&sdev->list, &srpt_dev_list);
3194 	spin_unlock(&srpt_dev_lock);
3195 
3196 	ib_set_client_data(device, &srpt_client, sdev);
3197 	pr_debug("added %s.\n", dev_name(&device->dev));
3198 	return 0;
3199 
3200 err_event:
3201 	ib_unregister_event_handler(&sdev->event_handler);
3202 err_cm:
3203 	if (sdev->cm_id)
3204 		ib_destroy_cm_id(sdev->cm_id);
3205 err_ring:
3206 	srpt_free_srq(sdev);
3207 	ib_dealloc_pd(sdev->pd);
3208 free_dev:
3209 	kfree(sdev);
3210 	pr_info("%s(%s) failed.\n", __func__, dev_name(&device->dev));
3211 	return ret;
3212 }
3213 
3214 /**
3215  * srpt_remove_one - InfiniBand device removal callback function
3216  * @device: Describes a HCA.
3217  * @client_data: The value passed as the third argument to ib_set_client_data().
3218  */
3219 static void srpt_remove_one(struct ib_device *device, void *client_data)
3220 {
3221 	struct srpt_device *sdev = client_data;
3222 	int i;
3223 
3224 	srpt_unregister_mad_agent(sdev);
3225 
3226 	ib_unregister_event_handler(&sdev->event_handler);
3227 
3228 	/* Cancel any work queued by the just unregistered IB event handler. */
3229 	for (i = 0; i < sdev->device->phys_port_cnt; i++)
3230 		cancel_work_sync(&sdev->port[i].work);
3231 
3232 	if (sdev->cm_id)
3233 		ib_destroy_cm_id(sdev->cm_id);
3234 
3235 	ib_set_client_data(device, &srpt_client, NULL);
3236 
3237 	/*
3238 	 * Unregistering a target must happen after destroying sdev->cm_id
3239 	 * such that no new SRP_LOGIN_REQ information units can arrive while
3240 	 * destroying the target.
3241 	 */
3242 	spin_lock(&srpt_dev_lock);
3243 	list_del(&sdev->list);
3244 	spin_unlock(&srpt_dev_lock);
3245 
3246 	for (i = 0; i < sdev->device->phys_port_cnt; i++)
3247 		srpt_release_sport(&sdev->port[i]);
3248 
3249 	srpt_free_srq(sdev);
3250 
3251 	ib_dealloc_pd(sdev->pd);
3252 
3253 	kfree(sdev);
3254 }
3255 
3256 static struct ib_client srpt_client = {
3257 	.name = DRV_NAME,
3258 	.add = srpt_add_one,
3259 	.remove = srpt_remove_one
3260 };
3261 
3262 static int srpt_check_true(struct se_portal_group *se_tpg)
3263 {
3264 	return 1;
3265 }
3266 
3267 static int srpt_check_false(struct se_portal_group *se_tpg)
3268 {
3269 	return 0;
3270 }
3271 
3272 static struct srpt_port *srpt_tpg_to_sport(struct se_portal_group *tpg)
3273 {
3274 	return tpg->se_tpg_wwn->priv;
3275 }
3276 
3277 static struct srpt_port_id *srpt_wwn_to_sport_id(struct se_wwn *wwn)
3278 {
3279 	struct srpt_port *sport = wwn->priv;
3280 
3281 	if (wwn == &sport->port_guid_id.wwn)
3282 		return &sport->port_guid_id;
3283 	if (wwn == &sport->port_gid_id.wwn)
3284 		return &sport->port_gid_id;
3285 	WARN_ON_ONCE(true);
3286 	return NULL;
3287 }
3288 
3289 static char *srpt_get_fabric_wwn(struct se_portal_group *tpg)
3290 {
3291 	struct srpt_tpg *stpg = container_of(tpg, typeof(*stpg), tpg);
3292 
3293 	return stpg->sport_id->name;
3294 }
3295 
3296 static u16 srpt_get_tag(struct se_portal_group *tpg)
3297 {
3298 	return 1;
3299 }
3300 
3301 static u32 srpt_tpg_get_inst_index(struct se_portal_group *se_tpg)
3302 {
3303 	return 1;
3304 }
3305 
3306 static void srpt_release_cmd(struct se_cmd *se_cmd)
3307 {
3308 	struct srpt_send_ioctx *ioctx = container_of(se_cmd,
3309 				struct srpt_send_ioctx, cmd);
3310 	struct srpt_rdma_ch *ch = ioctx->ch;
3311 	struct srpt_recv_ioctx *recv_ioctx = ioctx->recv_ioctx;
3312 
3313 	WARN_ON_ONCE(ioctx->state != SRPT_STATE_DONE &&
3314 		     !(ioctx->cmd.transport_state & CMD_T_ABORTED));
3315 
3316 	if (recv_ioctx) {
3317 		WARN_ON_ONCE(!list_empty(&recv_ioctx->wait_list));
3318 		ioctx->recv_ioctx = NULL;
3319 		srpt_post_recv(ch->sport->sdev, ch, recv_ioctx);
3320 	}
3321 
3322 	if (ioctx->n_rw_ctx) {
3323 		srpt_free_rw_ctxs(ch, ioctx);
3324 		ioctx->n_rw_ctx = 0;
3325 	}
3326 
3327 	target_free_tag(se_cmd->se_sess, se_cmd);
3328 }
3329 
3330 /**
3331  * srpt_close_session - forcibly close a session
3332  * @se_sess: SCSI target session.
3333  *
3334  * Callback function invoked by the TCM core to clean up sessions associated
3335  * with a node ACL when the user invokes
3336  * rmdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3337  */
3338 static void srpt_close_session(struct se_session *se_sess)
3339 {
3340 	struct srpt_rdma_ch *ch = se_sess->fabric_sess_ptr;
3341 
3342 	srpt_disconnect_ch_sync(ch);
3343 }
3344 
3345 /**
3346  * srpt_sess_get_index - return the value of scsiAttIntrPortIndex (SCSI-MIB)
3347  * @se_sess: SCSI target session.
3348  *
3349  * A quote from RFC 4455 (SCSI-MIB) about this MIB object:
3350  * This object represents an arbitrary integer used to uniquely identify a
3351  * particular attached remote initiator port to a particular SCSI target port
3352  * within a particular SCSI target device within a particular SCSI instance.
3353  */
3354 static u32 srpt_sess_get_index(struct se_session *se_sess)
3355 {
3356 	return 0;
3357 }
3358 
3359 static void srpt_set_default_node_attrs(struct se_node_acl *nacl)
3360 {
3361 }
3362 
3363 /* Note: only used from inside debug printk's by the TCM core. */
3364 static int srpt_get_tcm_cmd_state(struct se_cmd *se_cmd)
3365 {
3366 	struct srpt_send_ioctx *ioctx;
3367 
3368 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
3369 	return ioctx->state;
3370 }
3371 
3372 static int srpt_parse_guid(u64 *guid, const char *name)
3373 {
3374 	u16 w[4];
3375 	int ret = -EINVAL;
3376 
3377 	if (sscanf(name, "%hx:%hx:%hx:%hx", &w[0], &w[1], &w[2], &w[3]) != 4)
3378 		goto out;
3379 	*guid = get_unaligned_be64(w);
3380 	ret = 0;
3381 out:
3382 	return ret;
3383 }
3384 
3385 /**
3386  * srpt_parse_i_port_id - parse an initiator port ID
3387  * @name: ASCII representation of a 128-bit initiator port ID.
3388  * @i_port_id: Binary 128-bit port ID.
3389  */
3390 static int srpt_parse_i_port_id(u8 i_port_id[16], const char *name)
3391 {
3392 	const char *p;
3393 	unsigned len, count, leading_zero_bytes;
3394 	int ret;
3395 
3396 	p = name;
3397 	if (strncasecmp(p, "0x", 2) == 0)
3398 		p += 2;
3399 	ret = -EINVAL;
3400 	len = strlen(p);
3401 	if (len % 2)
3402 		goto out;
3403 	count = min(len / 2, 16U);
3404 	leading_zero_bytes = 16 - count;
3405 	memset(i_port_id, 0, leading_zero_bytes);
3406 	ret = hex2bin(i_port_id + leading_zero_bytes, p, count);
3407 
3408 out:
3409 	return ret;
3410 }
3411 
3412 /*
3413  * configfs callback function invoked for mkdir
3414  * /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3415  *
3416  * i_port_id must be an initiator port GUID, GID or IP address. See also the
3417  * target_alloc_session() calls in this driver. Examples of valid initiator
3418  * port IDs:
3419  * 0x0000000000000000505400fffe4a0b7b
3420  * 0000000000000000505400fffe4a0b7b
3421  * 5054:00ff:fe4a:0b7b
3422  * 192.168.122.76
3423  */
3424 static int srpt_init_nodeacl(struct se_node_acl *se_nacl, const char *name)
3425 {
3426 	struct sockaddr_storage sa;
3427 	u64 guid;
3428 	u8 i_port_id[16];
3429 	int ret;
3430 
3431 	ret = srpt_parse_guid(&guid, name);
3432 	if (ret < 0)
3433 		ret = srpt_parse_i_port_id(i_port_id, name);
3434 	if (ret < 0)
3435 		ret = inet_pton_with_scope(&init_net, AF_UNSPEC, name, NULL,
3436 					   &sa);
3437 	if (ret < 0)
3438 		pr_err("invalid initiator port ID %s\n", name);
3439 	return ret;
3440 }
3441 
3442 static ssize_t srpt_tpg_attrib_srp_max_rdma_size_show(struct config_item *item,
3443 		char *page)
3444 {
3445 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3446 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3447 
3448 	return sprintf(page, "%u\n", sport->port_attrib.srp_max_rdma_size);
3449 }
3450 
3451 static ssize_t srpt_tpg_attrib_srp_max_rdma_size_store(struct config_item *item,
3452 		const char *page, size_t count)
3453 {
3454 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3455 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3456 	unsigned long val;
3457 	int ret;
3458 
3459 	ret = kstrtoul(page, 0, &val);
3460 	if (ret < 0) {
3461 		pr_err("kstrtoul() failed with ret: %d\n", ret);
3462 		return -EINVAL;
3463 	}
3464 	if (val > MAX_SRPT_RDMA_SIZE) {
3465 		pr_err("val: %lu exceeds MAX_SRPT_RDMA_SIZE: %d\n", val,
3466 			MAX_SRPT_RDMA_SIZE);
3467 		return -EINVAL;
3468 	}
3469 	if (val < DEFAULT_MAX_RDMA_SIZE) {
3470 		pr_err("val: %lu smaller than DEFAULT_MAX_RDMA_SIZE: %d\n",
3471 			val, DEFAULT_MAX_RDMA_SIZE);
3472 		return -EINVAL;
3473 	}
3474 	sport->port_attrib.srp_max_rdma_size = val;
3475 
3476 	return count;
3477 }
3478 
3479 static ssize_t srpt_tpg_attrib_srp_max_rsp_size_show(struct config_item *item,
3480 		char *page)
3481 {
3482 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3483 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3484 
3485 	return sprintf(page, "%u\n", sport->port_attrib.srp_max_rsp_size);
3486 }
3487 
3488 static ssize_t srpt_tpg_attrib_srp_max_rsp_size_store(struct config_item *item,
3489 		const char *page, size_t count)
3490 {
3491 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3492 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3493 	unsigned long val;
3494 	int ret;
3495 
3496 	ret = kstrtoul(page, 0, &val);
3497 	if (ret < 0) {
3498 		pr_err("kstrtoul() failed with ret: %d\n", ret);
3499 		return -EINVAL;
3500 	}
3501 	if (val > MAX_SRPT_RSP_SIZE) {
3502 		pr_err("val: %lu exceeds MAX_SRPT_RSP_SIZE: %d\n", val,
3503 			MAX_SRPT_RSP_SIZE);
3504 		return -EINVAL;
3505 	}
3506 	if (val < MIN_MAX_RSP_SIZE) {
3507 		pr_err("val: %lu smaller than MIN_MAX_RSP_SIZE: %d\n", val,
3508 			MIN_MAX_RSP_SIZE);
3509 		return -EINVAL;
3510 	}
3511 	sport->port_attrib.srp_max_rsp_size = val;
3512 
3513 	return count;
3514 }
3515 
3516 static ssize_t srpt_tpg_attrib_srp_sq_size_show(struct config_item *item,
3517 		char *page)
3518 {
3519 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3520 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3521 
3522 	return sprintf(page, "%u\n", sport->port_attrib.srp_sq_size);
3523 }
3524 
3525 static ssize_t srpt_tpg_attrib_srp_sq_size_store(struct config_item *item,
3526 		const char *page, size_t count)
3527 {
3528 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3529 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3530 	unsigned long val;
3531 	int ret;
3532 
3533 	ret = kstrtoul(page, 0, &val);
3534 	if (ret < 0) {
3535 		pr_err("kstrtoul() failed with ret: %d\n", ret);
3536 		return -EINVAL;
3537 	}
3538 	if (val > MAX_SRPT_SRQ_SIZE) {
3539 		pr_err("val: %lu exceeds MAX_SRPT_SRQ_SIZE: %d\n", val,
3540 			MAX_SRPT_SRQ_SIZE);
3541 		return -EINVAL;
3542 	}
3543 	if (val < MIN_SRPT_SRQ_SIZE) {
3544 		pr_err("val: %lu smaller than MIN_SRPT_SRQ_SIZE: %d\n", val,
3545 			MIN_SRPT_SRQ_SIZE);
3546 		return -EINVAL;
3547 	}
3548 	sport->port_attrib.srp_sq_size = val;
3549 
3550 	return count;
3551 }
3552 
3553 static ssize_t srpt_tpg_attrib_use_srq_show(struct config_item *item,
3554 					    char *page)
3555 {
3556 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3557 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3558 
3559 	return sprintf(page, "%d\n", sport->port_attrib.use_srq);
3560 }
3561 
3562 static ssize_t srpt_tpg_attrib_use_srq_store(struct config_item *item,
3563 					     const char *page, size_t count)
3564 {
3565 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3566 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3567 	struct srpt_device *sdev = sport->sdev;
3568 	unsigned long val;
3569 	bool enabled;
3570 	int ret;
3571 
3572 	ret = kstrtoul(page, 0, &val);
3573 	if (ret < 0)
3574 		return ret;
3575 	if (val != !!val)
3576 		return -EINVAL;
3577 
3578 	ret = mutex_lock_interruptible(&sdev->sdev_mutex);
3579 	if (ret < 0)
3580 		return ret;
3581 	ret = mutex_lock_interruptible(&sport->mutex);
3582 	if (ret < 0)
3583 		goto unlock_sdev;
3584 	enabled = sport->enabled;
3585 	/* Log out all initiator systems before changing 'use_srq'. */
3586 	srpt_set_enabled(sport, false);
3587 	sport->port_attrib.use_srq = val;
3588 	srpt_use_srq(sdev, sport->port_attrib.use_srq);
3589 	srpt_set_enabled(sport, enabled);
3590 	ret = count;
3591 	mutex_unlock(&sport->mutex);
3592 unlock_sdev:
3593 	mutex_unlock(&sdev->sdev_mutex);
3594 
3595 	return ret;
3596 }
3597 
3598 CONFIGFS_ATTR(srpt_tpg_attrib_,  srp_max_rdma_size);
3599 CONFIGFS_ATTR(srpt_tpg_attrib_,  srp_max_rsp_size);
3600 CONFIGFS_ATTR(srpt_tpg_attrib_,  srp_sq_size);
3601 CONFIGFS_ATTR(srpt_tpg_attrib_,  use_srq);
3602 
3603 static struct configfs_attribute *srpt_tpg_attrib_attrs[] = {
3604 	&srpt_tpg_attrib_attr_srp_max_rdma_size,
3605 	&srpt_tpg_attrib_attr_srp_max_rsp_size,
3606 	&srpt_tpg_attrib_attr_srp_sq_size,
3607 	&srpt_tpg_attrib_attr_use_srq,
3608 	NULL,
3609 };
3610 
3611 static struct rdma_cm_id *srpt_create_rdma_id(struct sockaddr *listen_addr)
3612 {
3613 	struct rdma_cm_id *rdma_cm_id;
3614 	int ret;
3615 
3616 	rdma_cm_id = rdma_create_id(&init_net, srpt_rdma_cm_handler,
3617 				    NULL, RDMA_PS_TCP, IB_QPT_RC);
3618 	if (IS_ERR(rdma_cm_id)) {
3619 		pr_err("RDMA/CM ID creation failed: %ld\n",
3620 		       PTR_ERR(rdma_cm_id));
3621 		goto out;
3622 	}
3623 
3624 	ret = rdma_bind_addr(rdma_cm_id, listen_addr);
3625 	if (ret) {
3626 		char addr_str[64];
3627 
3628 		snprintf(addr_str, sizeof(addr_str), "%pISp", listen_addr);
3629 		pr_err("Binding RDMA/CM ID to address %s failed: %d\n",
3630 		       addr_str, ret);
3631 		rdma_destroy_id(rdma_cm_id);
3632 		rdma_cm_id = ERR_PTR(ret);
3633 		goto out;
3634 	}
3635 
3636 	ret = rdma_listen(rdma_cm_id, 128);
3637 	if (ret) {
3638 		pr_err("rdma_listen() failed: %d\n", ret);
3639 		rdma_destroy_id(rdma_cm_id);
3640 		rdma_cm_id = ERR_PTR(ret);
3641 	}
3642 
3643 out:
3644 	return rdma_cm_id;
3645 }
3646 
3647 static ssize_t srpt_rdma_cm_port_show(struct config_item *item, char *page)
3648 {
3649 	return sprintf(page, "%d\n", rdma_cm_port);
3650 }
3651 
3652 static ssize_t srpt_rdma_cm_port_store(struct config_item *item,
3653 				       const char *page, size_t count)
3654 {
3655 	struct sockaddr_in  addr4 = { .sin_family  = AF_INET  };
3656 	struct sockaddr_in6 addr6 = { .sin6_family = AF_INET6 };
3657 	struct rdma_cm_id *new_id = NULL;
3658 	u16 val;
3659 	int ret;
3660 
3661 	ret = kstrtou16(page, 0, &val);
3662 	if (ret < 0)
3663 		return ret;
3664 	ret = count;
3665 	if (rdma_cm_port == val)
3666 		goto out;
3667 
3668 	if (val) {
3669 		addr6.sin6_port = cpu_to_be16(val);
3670 		new_id = srpt_create_rdma_id((struct sockaddr *)&addr6);
3671 		if (IS_ERR(new_id)) {
3672 			addr4.sin_port = cpu_to_be16(val);
3673 			new_id = srpt_create_rdma_id((struct sockaddr *)&addr4);
3674 			if (IS_ERR(new_id)) {
3675 				ret = PTR_ERR(new_id);
3676 				goto out;
3677 			}
3678 		}
3679 	}
3680 
3681 	mutex_lock(&rdma_cm_mutex);
3682 	rdma_cm_port = val;
3683 	swap(rdma_cm_id, new_id);
3684 	mutex_unlock(&rdma_cm_mutex);
3685 
3686 	if (new_id)
3687 		rdma_destroy_id(new_id);
3688 	ret = count;
3689 out:
3690 	return ret;
3691 }
3692 
3693 CONFIGFS_ATTR(srpt_, rdma_cm_port);
3694 
3695 static struct configfs_attribute *srpt_da_attrs[] = {
3696 	&srpt_attr_rdma_cm_port,
3697 	NULL,
3698 };
3699 
3700 static ssize_t srpt_tpg_enable_show(struct config_item *item, char *page)
3701 {
3702 	struct se_portal_group *se_tpg = to_tpg(item);
3703 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3704 
3705 	return snprintf(page, PAGE_SIZE, "%d\n", sport->enabled);
3706 }
3707 
3708 static ssize_t srpt_tpg_enable_store(struct config_item *item,
3709 		const char *page, size_t count)
3710 {
3711 	struct se_portal_group *se_tpg = to_tpg(item);
3712 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3713 	unsigned long tmp;
3714 	int ret;
3715 
3716 	ret = kstrtoul(page, 0, &tmp);
3717 	if (ret < 0) {
3718 		pr_err("Unable to extract srpt_tpg_store_enable\n");
3719 		return -EINVAL;
3720 	}
3721 
3722 	if ((tmp != 0) && (tmp != 1)) {
3723 		pr_err("Illegal value for srpt_tpg_store_enable: %lu\n", tmp);
3724 		return -EINVAL;
3725 	}
3726 
3727 	mutex_lock(&sport->mutex);
3728 	srpt_set_enabled(sport, tmp);
3729 	mutex_unlock(&sport->mutex);
3730 
3731 	return count;
3732 }
3733 
3734 CONFIGFS_ATTR(srpt_tpg_, enable);
3735 
3736 static struct configfs_attribute *srpt_tpg_attrs[] = {
3737 	&srpt_tpg_attr_enable,
3738 	NULL,
3739 };
3740 
3741 /**
3742  * srpt_make_tpg - configfs callback invoked for mkdir /sys/kernel/config/target/$driver/$port/$tpg
3743  * @wwn: Corresponds to $driver/$port.
3744  * @name: $tpg.
3745  */
3746 static struct se_portal_group *srpt_make_tpg(struct se_wwn *wwn,
3747 					     const char *name)
3748 {
3749 	struct srpt_port_id *sport_id = srpt_wwn_to_sport_id(wwn);
3750 	struct srpt_tpg *stpg;
3751 	int res = -ENOMEM;
3752 
3753 	stpg = kzalloc(sizeof(*stpg), GFP_KERNEL);
3754 	if (!stpg)
3755 		return ERR_PTR(res);
3756 	stpg->sport_id = sport_id;
3757 	res = core_tpg_register(wwn, &stpg->tpg, SCSI_PROTOCOL_SRP);
3758 	if (res) {
3759 		kfree(stpg);
3760 		return ERR_PTR(res);
3761 	}
3762 
3763 	mutex_lock(&sport_id->mutex);
3764 	list_add_tail(&stpg->entry, &sport_id->tpg_list);
3765 	mutex_unlock(&sport_id->mutex);
3766 
3767 	return &stpg->tpg;
3768 }
3769 
3770 /**
3771  * srpt_drop_tpg - configfs callback invoked for rmdir /sys/kernel/config/target/$driver/$port/$tpg
3772  * @tpg: Target portal group to deregister.
3773  */
3774 static void srpt_drop_tpg(struct se_portal_group *tpg)
3775 {
3776 	struct srpt_tpg *stpg = container_of(tpg, typeof(*stpg), tpg);
3777 	struct srpt_port_id *sport_id = stpg->sport_id;
3778 	struct srpt_port *sport = srpt_tpg_to_sport(tpg);
3779 
3780 	mutex_lock(&sport_id->mutex);
3781 	list_del(&stpg->entry);
3782 	mutex_unlock(&sport_id->mutex);
3783 
3784 	sport->enabled = false;
3785 	core_tpg_deregister(tpg);
3786 	kfree(stpg);
3787 }
3788 
3789 /**
3790  * srpt_make_tport - configfs callback invoked for mkdir /sys/kernel/config/target/$driver/$port
3791  * @tf: Not used.
3792  * @group: Not used.
3793  * @name: $port.
3794  */
3795 static struct se_wwn *srpt_make_tport(struct target_fabric_configfs *tf,
3796 				      struct config_group *group,
3797 				      const char *name)
3798 {
3799 	return srpt_lookup_wwn(name) ? : ERR_PTR(-EINVAL);
3800 }
3801 
3802 /**
3803  * srpt_drop_tport - configfs callback invoked for rmdir /sys/kernel/config/target/$driver/$port
3804  * @wwn: $port.
3805  */
3806 static void srpt_drop_tport(struct se_wwn *wwn)
3807 {
3808 }
3809 
3810 static ssize_t srpt_wwn_version_show(struct config_item *item, char *buf)
3811 {
3812 	return scnprintf(buf, PAGE_SIZE, "\n");
3813 }
3814 
3815 CONFIGFS_ATTR_RO(srpt_wwn_, version);
3816 
3817 static struct configfs_attribute *srpt_wwn_attrs[] = {
3818 	&srpt_wwn_attr_version,
3819 	NULL,
3820 };
3821 
3822 static const struct target_core_fabric_ops srpt_template = {
3823 	.module				= THIS_MODULE,
3824 	.fabric_name			= "srpt",
3825 	.tpg_get_wwn			= srpt_get_fabric_wwn,
3826 	.tpg_get_tag			= srpt_get_tag,
3827 	.tpg_check_demo_mode		= srpt_check_false,
3828 	.tpg_check_demo_mode_cache	= srpt_check_true,
3829 	.tpg_check_demo_mode_write_protect = srpt_check_true,
3830 	.tpg_check_prod_mode_write_protect = srpt_check_false,
3831 	.tpg_get_inst_index		= srpt_tpg_get_inst_index,
3832 	.release_cmd			= srpt_release_cmd,
3833 	.check_stop_free		= srpt_check_stop_free,
3834 	.close_session			= srpt_close_session,
3835 	.sess_get_index			= srpt_sess_get_index,
3836 	.sess_get_initiator_sid		= NULL,
3837 	.write_pending			= srpt_write_pending,
3838 	.set_default_node_attributes	= srpt_set_default_node_attrs,
3839 	.get_cmd_state			= srpt_get_tcm_cmd_state,
3840 	.queue_data_in			= srpt_queue_data_in,
3841 	.queue_status			= srpt_queue_status,
3842 	.queue_tm_rsp			= srpt_queue_tm_rsp,
3843 	.aborted_task			= srpt_aborted_task,
3844 	/*
3845 	 * Setup function pointers for generic logic in
3846 	 * target_core_fabric_configfs.c
3847 	 */
3848 	.fabric_make_wwn		= srpt_make_tport,
3849 	.fabric_drop_wwn		= srpt_drop_tport,
3850 	.fabric_make_tpg		= srpt_make_tpg,
3851 	.fabric_drop_tpg		= srpt_drop_tpg,
3852 	.fabric_init_nodeacl		= srpt_init_nodeacl,
3853 
3854 	.tfc_discovery_attrs		= srpt_da_attrs,
3855 	.tfc_wwn_attrs			= srpt_wwn_attrs,
3856 	.tfc_tpg_base_attrs		= srpt_tpg_attrs,
3857 	.tfc_tpg_attrib_attrs		= srpt_tpg_attrib_attrs,
3858 };
3859 
3860 /**
3861  * srpt_init_module - kernel module initialization
3862  *
3863  * Note: Since ib_register_client() registers callback functions, and since at
3864  * least one of these callback functions (srpt_add_one()) calls target core
3865  * functions, this driver must be registered with the target core before
3866  * ib_register_client() is called.
3867  */
3868 static int __init srpt_init_module(void)
3869 {
3870 	int ret;
3871 
3872 	ret = -EINVAL;
3873 	if (srp_max_req_size < MIN_MAX_REQ_SIZE) {
3874 		pr_err("invalid value %d for kernel module parameter srp_max_req_size -- must be at least %d.\n",
3875 		       srp_max_req_size, MIN_MAX_REQ_SIZE);
3876 		goto out;
3877 	}
3878 
3879 	if (srpt_srq_size < MIN_SRPT_SRQ_SIZE
3880 	    || srpt_srq_size > MAX_SRPT_SRQ_SIZE) {
3881 		pr_err("invalid value %d for kernel module parameter srpt_srq_size -- must be in the range [%d..%d].\n",
3882 		       srpt_srq_size, MIN_SRPT_SRQ_SIZE, MAX_SRPT_SRQ_SIZE);
3883 		goto out;
3884 	}
3885 
3886 	ret = target_register_template(&srpt_template);
3887 	if (ret)
3888 		goto out;
3889 
3890 	ret = ib_register_client(&srpt_client);
3891 	if (ret) {
3892 		pr_err("couldn't register IB client\n");
3893 		goto out_unregister_target;
3894 	}
3895 
3896 	return 0;
3897 
3898 out_unregister_target:
3899 	target_unregister_template(&srpt_template);
3900 out:
3901 	return ret;
3902 }
3903 
3904 static void __exit srpt_cleanup_module(void)
3905 {
3906 	if (rdma_cm_id)
3907 		rdma_destroy_id(rdma_cm_id);
3908 	ib_unregister_client(&srpt_client);
3909 	target_unregister_template(&srpt_template);
3910 }
3911 
3912 module_init(srpt_init_module);
3913 module_exit(srpt_cleanup_module);
3914