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 	srpt_format_guid(sport->guid_name, ARRAY_SIZE(sport->guid_name),
569 			 &sport->gid.global.interface_id);
570 	snprintf(sport->gid_name, ARRAY_SIZE(sport->gid_name),
571 		 "0x%016llx%016llx",
572 		 be64_to_cpu(sport->gid.global.subnet_prefix),
573 		 be64_to_cpu(sport->gid.global.interface_id));
574 
575 	if (rdma_protocol_iwarp(sport->sdev->device, sport->port))
576 		return 0;
577 
578 	memset(&port_modify, 0, sizeof(port_modify));
579 	port_modify.set_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP;
580 	port_modify.clr_port_cap_mask = 0;
581 
582 	ret = ib_modify_port(sport->sdev->device, sport->port, 0, &port_modify);
583 	if (ret) {
584 		pr_warn("%s-%d: enabling device management failed (%d). Note: this is expected if SR-IOV is enabled.\n",
585 			dev_name(&sport->sdev->device->dev), sport->port, ret);
586 		return 0;
587 	}
588 
589 	if (!sport->mad_agent) {
590 		memset(&reg_req, 0, sizeof(reg_req));
591 		reg_req.mgmt_class = IB_MGMT_CLASS_DEVICE_MGMT;
592 		reg_req.mgmt_class_version = IB_MGMT_BASE_VERSION;
593 		set_bit(IB_MGMT_METHOD_GET, reg_req.method_mask);
594 		set_bit(IB_MGMT_METHOD_SET, reg_req.method_mask);
595 
596 		sport->mad_agent = ib_register_mad_agent(sport->sdev->device,
597 							 sport->port,
598 							 IB_QPT_GSI,
599 							 &reg_req, 0,
600 							 srpt_mad_send_handler,
601 							 srpt_mad_recv_handler,
602 							 sport, 0);
603 		if (IS_ERR(sport->mad_agent)) {
604 			pr_err("%s-%d: MAD agent registration failed (%ld). Note: this is expected if SR-IOV is enabled.\n",
605 			       dev_name(&sport->sdev->device->dev), sport->port,
606 			       PTR_ERR(sport->mad_agent));
607 			sport->mad_agent = NULL;
608 			memset(&port_modify, 0, sizeof(port_modify));
609 			port_modify.clr_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP;
610 			ib_modify_port(sport->sdev->device, sport->port, 0,
611 				       &port_modify);
612 
613 		}
614 	}
615 
616 	return 0;
617 }
618 
619 /**
620  * srpt_unregister_mad_agent - unregister MAD callback functions
621  * @sdev: SRPT HCA pointer.
622  * @port_cnt: number of ports with registered MAD
623  *
624  * Note: It is safe to call this function more than once for the same device.
625  */
626 static void srpt_unregister_mad_agent(struct srpt_device *sdev, int port_cnt)
627 {
628 	struct ib_port_modify port_modify = {
629 		.clr_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP,
630 	};
631 	struct srpt_port *sport;
632 	int i;
633 
634 	for (i = 1; i <= port_cnt; i++) {
635 		sport = &sdev->port[i - 1];
636 		WARN_ON(sport->port != i);
637 		if (sport->mad_agent) {
638 			ib_modify_port(sdev->device, i, 0, &port_modify);
639 			ib_unregister_mad_agent(sport->mad_agent);
640 			sport->mad_agent = NULL;
641 		}
642 	}
643 }
644 
645 /**
646  * srpt_alloc_ioctx - allocate a SRPT I/O context structure
647  * @sdev: SRPT HCA pointer.
648  * @ioctx_size: I/O context size.
649  * @buf_cache: I/O buffer cache.
650  * @dir: DMA data direction.
651  */
652 static struct srpt_ioctx *srpt_alloc_ioctx(struct srpt_device *sdev,
653 					   int ioctx_size,
654 					   struct kmem_cache *buf_cache,
655 					   enum dma_data_direction dir)
656 {
657 	struct srpt_ioctx *ioctx;
658 
659 	ioctx = kzalloc(ioctx_size, GFP_KERNEL);
660 	if (!ioctx)
661 		goto err;
662 
663 	ioctx->buf = kmem_cache_alloc(buf_cache, GFP_KERNEL);
664 	if (!ioctx->buf)
665 		goto err_free_ioctx;
666 
667 	ioctx->dma = ib_dma_map_single(sdev->device, ioctx->buf,
668 				       kmem_cache_size(buf_cache), dir);
669 	if (ib_dma_mapping_error(sdev->device, ioctx->dma))
670 		goto err_free_buf;
671 
672 	return ioctx;
673 
674 err_free_buf:
675 	kmem_cache_free(buf_cache, ioctx->buf);
676 err_free_ioctx:
677 	kfree(ioctx);
678 err:
679 	return NULL;
680 }
681 
682 /**
683  * srpt_free_ioctx - free a SRPT I/O context structure
684  * @sdev: SRPT HCA pointer.
685  * @ioctx: I/O context pointer.
686  * @buf_cache: I/O buffer cache.
687  * @dir: DMA data direction.
688  */
689 static void srpt_free_ioctx(struct srpt_device *sdev, struct srpt_ioctx *ioctx,
690 			    struct kmem_cache *buf_cache,
691 			    enum dma_data_direction dir)
692 {
693 	if (!ioctx)
694 		return;
695 
696 	ib_dma_unmap_single(sdev->device, ioctx->dma,
697 			    kmem_cache_size(buf_cache), dir);
698 	kmem_cache_free(buf_cache, ioctx->buf);
699 	kfree(ioctx);
700 }
701 
702 /**
703  * srpt_alloc_ioctx_ring - allocate a ring of SRPT I/O context structures
704  * @sdev:       Device to allocate the I/O context ring for.
705  * @ring_size:  Number of elements in the I/O context ring.
706  * @ioctx_size: I/O context size.
707  * @buf_cache:  I/O buffer cache.
708  * @alignment_offset: Offset in each ring buffer at which the SRP information
709  *		unit starts.
710  * @dir:        DMA data direction.
711  */
712 static struct srpt_ioctx **srpt_alloc_ioctx_ring(struct srpt_device *sdev,
713 				int ring_size, int ioctx_size,
714 				struct kmem_cache *buf_cache,
715 				int alignment_offset,
716 				enum dma_data_direction dir)
717 {
718 	struct srpt_ioctx **ring;
719 	int i;
720 
721 	WARN_ON(ioctx_size != sizeof(struct srpt_recv_ioctx) &&
722 		ioctx_size != sizeof(struct srpt_send_ioctx));
723 
724 	ring = kvmalloc_array(ring_size, sizeof(ring[0]), GFP_KERNEL);
725 	if (!ring)
726 		goto out;
727 	for (i = 0; i < ring_size; ++i) {
728 		ring[i] = srpt_alloc_ioctx(sdev, ioctx_size, buf_cache, dir);
729 		if (!ring[i])
730 			goto err;
731 		ring[i]->index = i;
732 		ring[i]->offset = alignment_offset;
733 	}
734 	goto out;
735 
736 err:
737 	while (--i >= 0)
738 		srpt_free_ioctx(sdev, ring[i], buf_cache, dir);
739 	kvfree(ring);
740 	ring = NULL;
741 out:
742 	return ring;
743 }
744 
745 /**
746  * srpt_free_ioctx_ring - free the ring of SRPT I/O context structures
747  * @ioctx_ring: I/O context ring to be freed.
748  * @sdev: SRPT HCA pointer.
749  * @ring_size: Number of ring elements.
750  * @buf_cache: I/O buffer cache.
751  * @dir: DMA data direction.
752  */
753 static void srpt_free_ioctx_ring(struct srpt_ioctx **ioctx_ring,
754 				 struct srpt_device *sdev, int ring_size,
755 				 struct kmem_cache *buf_cache,
756 				 enum dma_data_direction dir)
757 {
758 	int i;
759 
760 	if (!ioctx_ring)
761 		return;
762 
763 	for (i = 0; i < ring_size; ++i)
764 		srpt_free_ioctx(sdev, ioctx_ring[i], buf_cache, dir);
765 	kvfree(ioctx_ring);
766 }
767 
768 /**
769  * srpt_set_cmd_state - set the state of a SCSI command
770  * @ioctx: Send I/O context.
771  * @new: New I/O context state.
772  *
773  * Does not modify the state of aborted commands. Returns the previous command
774  * state.
775  */
776 static enum srpt_command_state srpt_set_cmd_state(struct srpt_send_ioctx *ioctx,
777 						  enum srpt_command_state new)
778 {
779 	enum srpt_command_state previous;
780 
781 	previous = ioctx->state;
782 	if (previous != SRPT_STATE_DONE)
783 		ioctx->state = new;
784 
785 	return previous;
786 }
787 
788 /**
789  * srpt_test_and_set_cmd_state - test and set the state of a command
790  * @ioctx: Send I/O context.
791  * @old: Current I/O context state.
792  * @new: New I/O context state.
793  *
794  * Returns true if and only if the previous command state was equal to 'old'.
795  */
796 static bool srpt_test_and_set_cmd_state(struct srpt_send_ioctx *ioctx,
797 					enum srpt_command_state old,
798 					enum srpt_command_state new)
799 {
800 	enum srpt_command_state previous;
801 
802 	WARN_ON(!ioctx);
803 	WARN_ON(old == SRPT_STATE_DONE);
804 	WARN_ON(new == SRPT_STATE_NEW);
805 
806 	previous = ioctx->state;
807 	if (previous == old)
808 		ioctx->state = new;
809 
810 	return previous == old;
811 }
812 
813 /**
814  * srpt_post_recv - post an IB receive request
815  * @sdev: SRPT HCA pointer.
816  * @ch: SRPT RDMA channel.
817  * @ioctx: Receive I/O context pointer.
818  */
819 static int srpt_post_recv(struct srpt_device *sdev, struct srpt_rdma_ch *ch,
820 			  struct srpt_recv_ioctx *ioctx)
821 {
822 	struct ib_sge list;
823 	struct ib_recv_wr wr;
824 
825 	BUG_ON(!sdev);
826 	list.addr = ioctx->ioctx.dma + ioctx->ioctx.offset;
827 	list.length = srp_max_req_size;
828 	list.lkey = sdev->lkey;
829 
830 	ioctx->ioctx.cqe.done = srpt_recv_done;
831 	wr.wr_cqe = &ioctx->ioctx.cqe;
832 	wr.next = NULL;
833 	wr.sg_list = &list;
834 	wr.num_sge = 1;
835 
836 	if (sdev->use_srq)
837 		return ib_post_srq_recv(sdev->srq, &wr, NULL);
838 	else
839 		return ib_post_recv(ch->qp, &wr, NULL);
840 }
841 
842 /**
843  * srpt_zerolength_write - perform a zero-length RDMA write
844  * @ch: SRPT RDMA channel.
845  *
846  * A quote from the InfiniBand specification: C9-88: For an HCA responder
847  * using Reliable Connection service, for each zero-length RDMA READ or WRITE
848  * request, the R_Key shall not be validated, even if the request includes
849  * Immediate data.
850  */
851 static int srpt_zerolength_write(struct srpt_rdma_ch *ch)
852 {
853 	struct ib_rdma_wr wr = {
854 		.wr = {
855 			.next		= NULL,
856 			{ .wr_cqe	= &ch->zw_cqe, },
857 			.opcode		= IB_WR_RDMA_WRITE,
858 			.send_flags	= IB_SEND_SIGNALED,
859 		}
860 	};
861 
862 	pr_debug("%s-%d: queued zerolength write\n", ch->sess_name,
863 		 ch->qp->qp_num);
864 
865 	return ib_post_send(ch->qp, &wr.wr, NULL);
866 }
867 
868 static void srpt_zerolength_write_done(struct ib_cq *cq, struct ib_wc *wc)
869 {
870 	struct srpt_rdma_ch *ch = wc->qp->qp_context;
871 
872 	pr_debug("%s-%d wc->status %d\n", ch->sess_name, ch->qp->qp_num,
873 		 wc->status);
874 
875 	if (wc->status == IB_WC_SUCCESS) {
876 		srpt_process_wait_list(ch);
877 	} else {
878 		if (srpt_set_ch_state(ch, CH_DISCONNECTED))
879 			schedule_work(&ch->release_work);
880 		else
881 			pr_debug("%s-%d: already disconnected.\n",
882 				 ch->sess_name, ch->qp->qp_num);
883 	}
884 }
885 
886 static int srpt_alloc_rw_ctxs(struct srpt_send_ioctx *ioctx,
887 		struct srp_direct_buf *db, int nbufs, struct scatterlist **sg,
888 		unsigned *sg_cnt)
889 {
890 	enum dma_data_direction dir = target_reverse_dma_direction(&ioctx->cmd);
891 	struct srpt_rdma_ch *ch = ioctx->ch;
892 	struct scatterlist *prev = NULL;
893 	unsigned prev_nents;
894 	int ret, i;
895 
896 	if (nbufs == 1) {
897 		ioctx->rw_ctxs = &ioctx->s_rw_ctx;
898 	} else {
899 		ioctx->rw_ctxs = kmalloc_array(nbufs, sizeof(*ioctx->rw_ctxs),
900 			GFP_KERNEL);
901 		if (!ioctx->rw_ctxs)
902 			return -ENOMEM;
903 	}
904 
905 	for (i = ioctx->n_rw_ctx; i < nbufs; i++, db++) {
906 		struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i];
907 		u64 remote_addr = be64_to_cpu(db->va);
908 		u32 size = be32_to_cpu(db->len);
909 		u32 rkey = be32_to_cpu(db->key);
910 
911 		ret = target_alloc_sgl(&ctx->sg, &ctx->nents, size, false,
912 				i < nbufs - 1);
913 		if (ret)
914 			goto unwind;
915 
916 		ret = rdma_rw_ctx_init(&ctx->rw, ch->qp, ch->sport->port,
917 				ctx->sg, ctx->nents, 0, remote_addr, rkey, dir);
918 		if (ret < 0) {
919 			target_free_sgl(ctx->sg, ctx->nents);
920 			goto unwind;
921 		}
922 
923 		ioctx->n_rdma += ret;
924 		ioctx->n_rw_ctx++;
925 
926 		if (prev) {
927 			sg_unmark_end(&prev[prev_nents - 1]);
928 			sg_chain(prev, prev_nents + 1, ctx->sg);
929 		} else {
930 			*sg = ctx->sg;
931 		}
932 
933 		prev = ctx->sg;
934 		prev_nents = ctx->nents;
935 
936 		*sg_cnt += ctx->nents;
937 	}
938 
939 	return 0;
940 
941 unwind:
942 	while (--i >= 0) {
943 		struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i];
944 
945 		rdma_rw_ctx_destroy(&ctx->rw, ch->qp, ch->sport->port,
946 				ctx->sg, ctx->nents, dir);
947 		target_free_sgl(ctx->sg, ctx->nents);
948 	}
949 	if (ioctx->rw_ctxs != &ioctx->s_rw_ctx)
950 		kfree(ioctx->rw_ctxs);
951 	return ret;
952 }
953 
954 static void srpt_free_rw_ctxs(struct srpt_rdma_ch *ch,
955 				    struct srpt_send_ioctx *ioctx)
956 {
957 	enum dma_data_direction dir = target_reverse_dma_direction(&ioctx->cmd);
958 	int i;
959 
960 	for (i = 0; i < ioctx->n_rw_ctx; i++) {
961 		struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i];
962 
963 		rdma_rw_ctx_destroy(&ctx->rw, ch->qp, ch->sport->port,
964 				ctx->sg, ctx->nents, dir);
965 		target_free_sgl(ctx->sg, ctx->nents);
966 	}
967 
968 	if (ioctx->rw_ctxs != &ioctx->s_rw_ctx)
969 		kfree(ioctx->rw_ctxs);
970 }
971 
972 static inline void *srpt_get_desc_buf(struct srp_cmd *srp_cmd)
973 {
974 	/*
975 	 * The pointer computations below will only be compiled correctly
976 	 * if srp_cmd::add_data is declared as s8*, u8*, s8[] or u8[], so check
977 	 * whether srp_cmd::add_data has been declared as a byte pointer.
978 	 */
979 	BUILD_BUG_ON(!__same_type(srp_cmd->add_data[0], (s8)0) &&
980 		     !__same_type(srp_cmd->add_data[0], (u8)0));
981 
982 	/*
983 	 * According to the SRP spec, the lower two bits of the 'ADDITIONAL
984 	 * CDB LENGTH' field are reserved and the size in bytes of this field
985 	 * is four times the value specified in bits 3..7. Hence the "& ~3".
986 	 */
987 	return srp_cmd->add_data + (srp_cmd->add_cdb_len & ~3);
988 }
989 
990 /**
991  * srpt_get_desc_tbl - parse the data descriptors of a SRP_CMD request
992  * @recv_ioctx: I/O context associated with the received command @srp_cmd.
993  * @ioctx: I/O context that will be used for responding to the initiator.
994  * @srp_cmd: Pointer to the SRP_CMD request data.
995  * @dir: Pointer to the variable to which the transfer direction will be
996  *   written.
997  * @sg: [out] scatterlist for the parsed SRP_CMD.
998  * @sg_cnt: [out] length of @sg.
999  * @data_len: Pointer to the variable to which the total data length of all
1000  *   descriptors in the SRP_CMD request will be written.
1001  * @imm_data_offset: [in] Offset in SRP_CMD requests at which immediate data
1002  *   starts.
1003  *
1004  * This function initializes ioctx->nrbuf and ioctx->r_bufs.
1005  *
1006  * Returns -EINVAL when the SRP_CMD request contains inconsistent descriptors;
1007  * -ENOMEM when memory allocation fails and zero upon success.
1008  */
1009 static int srpt_get_desc_tbl(struct srpt_recv_ioctx *recv_ioctx,
1010 		struct srpt_send_ioctx *ioctx,
1011 		struct srp_cmd *srp_cmd, enum dma_data_direction *dir,
1012 		struct scatterlist **sg, unsigned int *sg_cnt, u64 *data_len,
1013 		u16 imm_data_offset)
1014 {
1015 	BUG_ON(!dir);
1016 	BUG_ON(!data_len);
1017 
1018 	/*
1019 	 * The lower four bits of the buffer format field contain the DATA-IN
1020 	 * buffer descriptor format, and the highest four bits contain the
1021 	 * DATA-OUT buffer descriptor format.
1022 	 */
1023 	if (srp_cmd->buf_fmt & 0xf)
1024 		/* DATA-IN: transfer data from target to initiator (read). */
1025 		*dir = DMA_FROM_DEVICE;
1026 	else if (srp_cmd->buf_fmt >> 4)
1027 		/* DATA-OUT: transfer data from initiator to target (write). */
1028 		*dir = DMA_TO_DEVICE;
1029 	else
1030 		*dir = DMA_NONE;
1031 
1032 	/* initialize data_direction early as srpt_alloc_rw_ctxs needs it */
1033 	ioctx->cmd.data_direction = *dir;
1034 
1035 	if (((srp_cmd->buf_fmt & 0xf) == SRP_DATA_DESC_DIRECT) ||
1036 	    ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_DIRECT)) {
1037 		struct srp_direct_buf *db = srpt_get_desc_buf(srp_cmd);
1038 
1039 		*data_len = be32_to_cpu(db->len);
1040 		return srpt_alloc_rw_ctxs(ioctx, db, 1, sg, sg_cnt);
1041 	} else if (((srp_cmd->buf_fmt & 0xf) == SRP_DATA_DESC_INDIRECT) ||
1042 		   ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_INDIRECT)) {
1043 		struct srp_indirect_buf *idb = srpt_get_desc_buf(srp_cmd);
1044 		int nbufs = be32_to_cpu(idb->table_desc.len) /
1045 				sizeof(struct srp_direct_buf);
1046 
1047 		if (nbufs >
1048 		    (srp_cmd->data_out_desc_cnt + srp_cmd->data_in_desc_cnt)) {
1049 			pr_err("received unsupported SRP_CMD request type (%u out + %u in != %u / %zu)\n",
1050 			       srp_cmd->data_out_desc_cnt,
1051 			       srp_cmd->data_in_desc_cnt,
1052 			       be32_to_cpu(idb->table_desc.len),
1053 			       sizeof(struct srp_direct_buf));
1054 			return -EINVAL;
1055 		}
1056 
1057 		*data_len = be32_to_cpu(idb->len);
1058 		return srpt_alloc_rw_ctxs(ioctx, idb->desc_list, nbufs,
1059 				sg, sg_cnt);
1060 	} else if ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_IMM) {
1061 		struct srp_imm_buf *imm_buf = srpt_get_desc_buf(srp_cmd);
1062 		void *data = (void *)srp_cmd + imm_data_offset;
1063 		uint32_t len = be32_to_cpu(imm_buf->len);
1064 		uint32_t req_size = imm_data_offset + len;
1065 
1066 		if (req_size > srp_max_req_size) {
1067 			pr_err("Immediate data (length %d + %d) exceeds request size %d\n",
1068 			       imm_data_offset, len, srp_max_req_size);
1069 			return -EINVAL;
1070 		}
1071 		if (recv_ioctx->byte_len < req_size) {
1072 			pr_err("Received too few data - %d < %d\n",
1073 			       recv_ioctx->byte_len, req_size);
1074 			return -EIO;
1075 		}
1076 		/*
1077 		 * The immediate data buffer descriptor must occur before the
1078 		 * immediate data itself.
1079 		 */
1080 		if ((void *)(imm_buf + 1) > (void *)data) {
1081 			pr_err("Received invalid write request\n");
1082 			return -EINVAL;
1083 		}
1084 		*data_len = len;
1085 		ioctx->recv_ioctx = recv_ioctx;
1086 		if ((uintptr_t)data & 511) {
1087 			pr_warn_once("Internal error - the receive buffers are not aligned properly.\n");
1088 			return -EINVAL;
1089 		}
1090 		sg_init_one(&ioctx->imm_sg, data, len);
1091 		*sg = &ioctx->imm_sg;
1092 		*sg_cnt = 1;
1093 		return 0;
1094 	} else {
1095 		*data_len = 0;
1096 		return 0;
1097 	}
1098 }
1099 
1100 /**
1101  * srpt_init_ch_qp - initialize queue pair attributes
1102  * @ch: SRPT RDMA channel.
1103  * @qp: Queue pair pointer.
1104  *
1105  * Initialized the attributes of queue pair 'qp' by allowing local write,
1106  * remote read and remote write. Also transitions 'qp' to state IB_QPS_INIT.
1107  */
1108 static int srpt_init_ch_qp(struct srpt_rdma_ch *ch, struct ib_qp *qp)
1109 {
1110 	struct ib_qp_attr *attr;
1111 	int ret;
1112 
1113 	WARN_ON_ONCE(ch->using_rdma_cm);
1114 
1115 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
1116 	if (!attr)
1117 		return -ENOMEM;
1118 
1119 	attr->qp_state = IB_QPS_INIT;
1120 	attr->qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1121 	attr->port_num = ch->sport->port;
1122 
1123 	ret = ib_find_cached_pkey(ch->sport->sdev->device, ch->sport->port,
1124 				  ch->pkey, &attr->pkey_index);
1125 	if (ret < 0)
1126 		pr_err("Translating pkey %#x failed (%d) - using index 0\n",
1127 		       ch->pkey, ret);
1128 
1129 	ret = ib_modify_qp(qp, attr,
1130 			   IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PORT |
1131 			   IB_QP_PKEY_INDEX);
1132 
1133 	kfree(attr);
1134 	return ret;
1135 }
1136 
1137 /**
1138  * srpt_ch_qp_rtr - change the state of a channel to 'ready to receive' (RTR)
1139  * @ch: channel of the queue pair.
1140  * @qp: queue pair to change the state of.
1141  *
1142  * Returns zero upon success and a negative value upon failure.
1143  *
1144  * Note: currently a struct ib_qp_attr takes 136 bytes on a 64-bit system.
1145  * If this structure ever becomes larger, it might be necessary to allocate
1146  * it dynamically instead of on the stack.
1147  */
1148 static int srpt_ch_qp_rtr(struct srpt_rdma_ch *ch, struct ib_qp *qp)
1149 {
1150 	struct ib_qp_attr qp_attr;
1151 	int attr_mask;
1152 	int ret;
1153 
1154 	WARN_ON_ONCE(ch->using_rdma_cm);
1155 
1156 	qp_attr.qp_state = IB_QPS_RTR;
1157 	ret = ib_cm_init_qp_attr(ch->ib_cm.cm_id, &qp_attr, &attr_mask);
1158 	if (ret)
1159 		goto out;
1160 
1161 	qp_attr.max_dest_rd_atomic = 4;
1162 
1163 	ret = ib_modify_qp(qp, &qp_attr, attr_mask);
1164 
1165 out:
1166 	return ret;
1167 }
1168 
1169 /**
1170  * srpt_ch_qp_rts - change the state of a channel to 'ready to send' (RTS)
1171  * @ch: channel of the queue pair.
1172  * @qp: queue pair to change the state of.
1173  *
1174  * Returns zero upon success and a negative value upon failure.
1175  *
1176  * Note: currently a struct ib_qp_attr takes 136 bytes on a 64-bit system.
1177  * If this structure ever becomes larger, it might be necessary to allocate
1178  * it dynamically instead of on the stack.
1179  */
1180 static int srpt_ch_qp_rts(struct srpt_rdma_ch *ch, struct ib_qp *qp)
1181 {
1182 	struct ib_qp_attr qp_attr;
1183 	int attr_mask;
1184 	int ret;
1185 
1186 	qp_attr.qp_state = IB_QPS_RTS;
1187 	ret = ib_cm_init_qp_attr(ch->ib_cm.cm_id, &qp_attr, &attr_mask);
1188 	if (ret)
1189 		goto out;
1190 
1191 	qp_attr.max_rd_atomic = 4;
1192 
1193 	ret = ib_modify_qp(qp, &qp_attr, attr_mask);
1194 
1195 out:
1196 	return ret;
1197 }
1198 
1199 /**
1200  * srpt_ch_qp_err - set the channel queue pair state to 'error'
1201  * @ch: SRPT RDMA channel.
1202  */
1203 static int srpt_ch_qp_err(struct srpt_rdma_ch *ch)
1204 {
1205 	struct ib_qp_attr qp_attr;
1206 
1207 	qp_attr.qp_state = IB_QPS_ERR;
1208 	return ib_modify_qp(ch->qp, &qp_attr, IB_QP_STATE);
1209 }
1210 
1211 /**
1212  * srpt_get_send_ioctx - obtain an I/O context for sending to the initiator
1213  * @ch: SRPT RDMA channel.
1214  */
1215 static struct srpt_send_ioctx *srpt_get_send_ioctx(struct srpt_rdma_ch *ch)
1216 {
1217 	struct srpt_send_ioctx *ioctx;
1218 	int tag, cpu;
1219 
1220 	BUG_ON(!ch);
1221 
1222 	tag = sbitmap_queue_get(&ch->sess->sess_tag_pool, &cpu);
1223 	if (tag < 0)
1224 		return NULL;
1225 
1226 	ioctx = ch->ioctx_ring[tag];
1227 	BUG_ON(ioctx->ch != ch);
1228 	ioctx->state = SRPT_STATE_NEW;
1229 	WARN_ON_ONCE(ioctx->recv_ioctx);
1230 	ioctx->n_rdma = 0;
1231 	ioctx->n_rw_ctx = 0;
1232 	ioctx->queue_status_only = false;
1233 	/*
1234 	 * transport_init_se_cmd() does not initialize all fields, so do it
1235 	 * here.
1236 	 */
1237 	memset(&ioctx->cmd, 0, sizeof(ioctx->cmd));
1238 	memset(&ioctx->sense_data, 0, sizeof(ioctx->sense_data));
1239 	ioctx->cmd.map_tag = tag;
1240 	ioctx->cmd.map_cpu = cpu;
1241 
1242 	return ioctx;
1243 }
1244 
1245 /**
1246  * srpt_abort_cmd - abort a SCSI command
1247  * @ioctx:   I/O context associated with the SCSI command.
1248  */
1249 static int srpt_abort_cmd(struct srpt_send_ioctx *ioctx)
1250 {
1251 	enum srpt_command_state state;
1252 
1253 	BUG_ON(!ioctx);
1254 
1255 	/*
1256 	 * If the command is in a state where the target core is waiting for
1257 	 * the ib_srpt driver, change the state to the next state.
1258 	 */
1259 
1260 	state = ioctx->state;
1261 	switch (state) {
1262 	case SRPT_STATE_NEED_DATA:
1263 		ioctx->state = SRPT_STATE_DATA_IN;
1264 		break;
1265 	case SRPT_STATE_CMD_RSP_SENT:
1266 	case SRPT_STATE_MGMT_RSP_SENT:
1267 		ioctx->state = SRPT_STATE_DONE;
1268 		break;
1269 	default:
1270 		WARN_ONCE(true, "%s: unexpected I/O context state %d\n",
1271 			  __func__, state);
1272 		break;
1273 	}
1274 
1275 	pr_debug("Aborting cmd with state %d -> %d and tag %lld\n", state,
1276 		 ioctx->state, ioctx->cmd.tag);
1277 
1278 	switch (state) {
1279 	case SRPT_STATE_NEW:
1280 	case SRPT_STATE_DATA_IN:
1281 	case SRPT_STATE_MGMT:
1282 	case SRPT_STATE_DONE:
1283 		/*
1284 		 * Do nothing - defer abort processing until
1285 		 * srpt_queue_response() is invoked.
1286 		 */
1287 		break;
1288 	case SRPT_STATE_NEED_DATA:
1289 		pr_debug("tag %#llx: RDMA read error\n", ioctx->cmd.tag);
1290 		transport_generic_request_failure(&ioctx->cmd,
1291 					TCM_CHECK_CONDITION_ABORT_CMD);
1292 		break;
1293 	case SRPT_STATE_CMD_RSP_SENT:
1294 		/*
1295 		 * SRP_RSP sending failed or the SRP_RSP send completion has
1296 		 * not been received in time.
1297 		 */
1298 		transport_generic_free_cmd(&ioctx->cmd, 0);
1299 		break;
1300 	case SRPT_STATE_MGMT_RSP_SENT:
1301 		transport_generic_free_cmd(&ioctx->cmd, 0);
1302 		break;
1303 	default:
1304 		WARN(1, "Unexpected command state (%d)", state);
1305 		break;
1306 	}
1307 
1308 	return state;
1309 }
1310 
1311 /**
1312  * srpt_rdma_read_done - RDMA read completion callback
1313  * @cq: Completion queue.
1314  * @wc: Work completion.
1315  *
1316  * XXX: what is now target_execute_cmd used to be asynchronous, and unmapping
1317  * the data that has been transferred via IB RDMA had to be postponed until the
1318  * check_stop_free() callback.  None of this is necessary anymore and needs to
1319  * be cleaned up.
1320  */
1321 static void srpt_rdma_read_done(struct ib_cq *cq, struct ib_wc *wc)
1322 {
1323 	struct srpt_rdma_ch *ch = wc->qp->qp_context;
1324 	struct srpt_send_ioctx *ioctx =
1325 		container_of(wc->wr_cqe, struct srpt_send_ioctx, rdma_cqe);
1326 
1327 	WARN_ON(ioctx->n_rdma <= 0);
1328 	atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
1329 	ioctx->n_rdma = 0;
1330 
1331 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
1332 		pr_info("RDMA_READ for ioctx 0x%p failed with status %d\n",
1333 			ioctx, wc->status);
1334 		srpt_abort_cmd(ioctx);
1335 		return;
1336 	}
1337 
1338 	if (srpt_test_and_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA,
1339 					SRPT_STATE_DATA_IN))
1340 		target_execute_cmd(&ioctx->cmd);
1341 	else
1342 		pr_err("%s[%d]: wrong state = %d\n", __func__,
1343 		       __LINE__, ioctx->state);
1344 }
1345 
1346 /**
1347  * srpt_build_cmd_rsp - build a SRP_RSP response
1348  * @ch: RDMA channel through which the request has been received.
1349  * @ioctx: I/O context associated with the SRP_CMD request. The response will
1350  *   be built in the buffer ioctx->buf points at and hence this function will
1351  *   overwrite the request data.
1352  * @tag: tag of the request for which this response is being generated.
1353  * @status: value for the STATUS field of the SRP_RSP information unit.
1354  *
1355  * Returns the size in bytes of the SRP_RSP response.
1356  *
1357  * An SRP_RSP response contains a SCSI status or service response. See also
1358  * section 6.9 in the SRP r16a document for the format of an SRP_RSP
1359  * response. See also SPC-2 for more information about sense data.
1360  */
1361 static int srpt_build_cmd_rsp(struct srpt_rdma_ch *ch,
1362 			      struct srpt_send_ioctx *ioctx, u64 tag,
1363 			      int status)
1364 {
1365 	struct se_cmd *cmd = &ioctx->cmd;
1366 	struct srp_rsp *srp_rsp;
1367 	const u8 *sense_data;
1368 	int sense_data_len, max_sense_len;
1369 	u32 resid = cmd->residual_count;
1370 
1371 	/*
1372 	 * The lowest bit of all SAM-3 status codes is zero (see also
1373 	 * paragraph 5.3 in SAM-3).
1374 	 */
1375 	WARN_ON(status & 1);
1376 
1377 	srp_rsp = ioctx->ioctx.buf;
1378 	BUG_ON(!srp_rsp);
1379 
1380 	sense_data = ioctx->sense_data;
1381 	sense_data_len = ioctx->cmd.scsi_sense_length;
1382 	WARN_ON(sense_data_len > sizeof(ioctx->sense_data));
1383 
1384 	memset(srp_rsp, 0, sizeof(*srp_rsp));
1385 	srp_rsp->opcode = SRP_RSP;
1386 	srp_rsp->req_lim_delta =
1387 		cpu_to_be32(1 + atomic_xchg(&ch->req_lim_delta, 0));
1388 	srp_rsp->tag = tag;
1389 	srp_rsp->status = status;
1390 
1391 	if (cmd->se_cmd_flags & SCF_UNDERFLOW_BIT) {
1392 		if (cmd->data_direction == DMA_TO_DEVICE) {
1393 			/* residual data from an underflow write */
1394 			srp_rsp->flags = SRP_RSP_FLAG_DOUNDER;
1395 			srp_rsp->data_out_res_cnt = cpu_to_be32(resid);
1396 		} else if (cmd->data_direction == DMA_FROM_DEVICE) {
1397 			/* residual data from an underflow read */
1398 			srp_rsp->flags = SRP_RSP_FLAG_DIUNDER;
1399 			srp_rsp->data_in_res_cnt = cpu_to_be32(resid);
1400 		}
1401 	} else if (cmd->se_cmd_flags & SCF_OVERFLOW_BIT) {
1402 		if (cmd->data_direction == DMA_TO_DEVICE) {
1403 			/* residual data from an overflow write */
1404 			srp_rsp->flags = SRP_RSP_FLAG_DOOVER;
1405 			srp_rsp->data_out_res_cnt = cpu_to_be32(resid);
1406 		} else if (cmd->data_direction == DMA_FROM_DEVICE) {
1407 			/* residual data from an overflow read */
1408 			srp_rsp->flags = SRP_RSP_FLAG_DIOVER;
1409 			srp_rsp->data_in_res_cnt = cpu_to_be32(resid);
1410 		}
1411 	}
1412 
1413 	if (sense_data_len) {
1414 		BUILD_BUG_ON(MIN_MAX_RSP_SIZE <= sizeof(*srp_rsp));
1415 		max_sense_len = ch->max_ti_iu_len - sizeof(*srp_rsp);
1416 		if (sense_data_len > max_sense_len) {
1417 			pr_warn("truncated sense data from %d to %d bytes\n",
1418 				sense_data_len, max_sense_len);
1419 			sense_data_len = max_sense_len;
1420 		}
1421 
1422 		srp_rsp->flags |= SRP_RSP_FLAG_SNSVALID;
1423 		srp_rsp->sense_data_len = cpu_to_be32(sense_data_len);
1424 		memcpy(srp_rsp->data, sense_data, sense_data_len);
1425 	}
1426 
1427 	return sizeof(*srp_rsp) + sense_data_len;
1428 }
1429 
1430 /**
1431  * srpt_build_tskmgmt_rsp - build a task management response
1432  * @ch:       RDMA channel through which the request has been received.
1433  * @ioctx:    I/O context in which the SRP_RSP response will be built.
1434  * @rsp_code: RSP_CODE that will be stored in the response.
1435  * @tag:      Tag of the request for which this response is being generated.
1436  *
1437  * Returns the size in bytes of the SRP_RSP response.
1438  *
1439  * An SRP_RSP response contains a SCSI status or service response. See also
1440  * section 6.9 in the SRP r16a document for the format of an SRP_RSP
1441  * response.
1442  */
1443 static int srpt_build_tskmgmt_rsp(struct srpt_rdma_ch *ch,
1444 				  struct srpt_send_ioctx *ioctx,
1445 				  u8 rsp_code, u64 tag)
1446 {
1447 	struct srp_rsp *srp_rsp;
1448 	int resp_data_len;
1449 	int resp_len;
1450 
1451 	resp_data_len = 4;
1452 	resp_len = sizeof(*srp_rsp) + resp_data_len;
1453 
1454 	srp_rsp = ioctx->ioctx.buf;
1455 	BUG_ON(!srp_rsp);
1456 	memset(srp_rsp, 0, sizeof(*srp_rsp));
1457 
1458 	srp_rsp->opcode = SRP_RSP;
1459 	srp_rsp->req_lim_delta =
1460 		cpu_to_be32(1 + atomic_xchg(&ch->req_lim_delta, 0));
1461 	srp_rsp->tag = tag;
1462 
1463 	srp_rsp->flags |= SRP_RSP_FLAG_RSPVALID;
1464 	srp_rsp->resp_data_len = cpu_to_be32(resp_data_len);
1465 	srp_rsp->data[3] = rsp_code;
1466 
1467 	return resp_len;
1468 }
1469 
1470 static int srpt_check_stop_free(struct se_cmd *cmd)
1471 {
1472 	struct srpt_send_ioctx *ioctx = container_of(cmd,
1473 				struct srpt_send_ioctx, cmd);
1474 
1475 	return target_put_sess_cmd(&ioctx->cmd);
1476 }
1477 
1478 /**
1479  * srpt_handle_cmd - process a SRP_CMD information unit
1480  * @ch: SRPT RDMA channel.
1481  * @recv_ioctx: Receive I/O context.
1482  * @send_ioctx: Send I/O context.
1483  */
1484 static void srpt_handle_cmd(struct srpt_rdma_ch *ch,
1485 			    struct srpt_recv_ioctx *recv_ioctx,
1486 			    struct srpt_send_ioctx *send_ioctx)
1487 {
1488 	struct se_cmd *cmd;
1489 	struct srp_cmd *srp_cmd;
1490 	struct scatterlist *sg = NULL;
1491 	unsigned sg_cnt = 0;
1492 	u64 data_len;
1493 	enum dma_data_direction dir;
1494 	int rc;
1495 
1496 	BUG_ON(!send_ioctx);
1497 
1498 	srp_cmd = recv_ioctx->ioctx.buf + recv_ioctx->ioctx.offset;
1499 	cmd = &send_ioctx->cmd;
1500 	cmd->tag = srp_cmd->tag;
1501 
1502 	switch (srp_cmd->task_attr) {
1503 	case SRP_CMD_SIMPLE_Q:
1504 		cmd->sam_task_attr = TCM_SIMPLE_TAG;
1505 		break;
1506 	case SRP_CMD_ORDERED_Q:
1507 	default:
1508 		cmd->sam_task_attr = TCM_ORDERED_TAG;
1509 		break;
1510 	case SRP_CMD_HEAD_OF_Q:
1511 		cmd->sam_task_attr = TCM_HEAD_TAG;
1512 		break;
1513 	case SRP_CMD_ACA:
1514 		cmd->sam_task_attr = TCM_ACA_TAG;
1515 		break;
1516 	}
1517 
1518 	rc = srpt_get_desc_tbl(recv_ioctx, send_ioctx, srp_cmd, &dir,
1519 			       &sg, &sg_cnt, &data_len, ch->imm_data_offset);
1520 	if (rc) {
1521 		if (rc != -EAGAIN) {
1522 			pr_err("0x%llx: parsing SRP descriptor table failed.\n",
1523 			       srp_cmd->tag);
1524 		}
1525 		goto busy;
1526 	}
1527 
1528 	rc = target_init_cmd(cmd, ch->sess, &send_ioctx->sense_data[0],
1529 			     scsilun_to_int(&srp_cmd->lun), data_len,
1530 			     TCM_SIMPLE_TAG, dir, TARGET_SCF_ACK_KREF);
1531 	if (rc != 0) {
1532 		pr_debug("target_submit_cmd() returned %d for tag %#llx\n", rc,
1533 			 srp_cmd->tag);
1534 		goto busy;
1535 	}
1536 
1537 	if (target_submit_prep(cmd, srp_cmd->cdb, sg, sg_cnt, NULL, 0, NULL, 0,
1538 			       GFP_KERNEL))
1539 		return;
1540 
1541 	target_submit(cmd);
1542 	return;
1543 
1544 busy:
1545 	target_send_busy(cmd);
1546 }
1547 
1548 static int srp_tmr_to_tcm(int fn)
1549 {
1550 	switch (fn) {
1551 	case SRP_TSK_ABORT_TASK:
1552 		return TMR_ABORT_TASK;
1553 	case SRP_TSK_ABORT_TASK_SET:
1554 		return TMR_ABORT_TASK_SET;
1555 	case SRP_TSK_CLEAR_TASK_SET:
1556 		return TMR_CLEAR_TASK_SET;
1557 	case SRP_TSK_LUN_RESET:
1558 		return TMR_LUN_RESET;
1559 	case SRP_TSK_CLEAR_ACA:
1560 		return TMR_CLEAR_ACA;
1561 	default:
1562 		return -1;
1563 	}
1564 }
1565 
1566 /**
1567  * srpt_handle_tsk_mgmt - process a SRP_TSK_MGMT information unit
1568  * @ch: SRPT RDMA channel.
1569  * @recv_ioctx: Receive I/O context.
1570  * @send_ioctx: Send I/O context.
1571  *
1572  * Returns 0 if and only if the request will be processed by the target core.
1573  *
1574  * For more information about SRP_TSK_MGMT information units, see also section
1575  * 6.7 in the SRP r16a document.
1576  */
1577 static void srpt_handle_tsk_mgmt(struct srpt_rdma_ch *ch,
1578 				 struct srpt_recv_ioctx *recv_ioctx,
1579 				 struct srpt_send_ioctx *send_ioctx)
1580 {
1581 	struct srp_tsk_mgmt *srp_tsk;
1582 	struct se_cmd *cmd;
1583 	struct se_session *sess = ch->sess;
1584 	int tcm_tmr;
1585 	int rc;
1586 
1587 	BUG_ON(!send_ioctx);
1588 
1589 	srp_tsk = recv_ioctx->ioctx.buf + recv_ioctx->ioctx.offset;
1590 	cmd = &send_ioctx->cmd;
1591 
1592 	pr_debug("recv tsk_mgmt fn %d for task_tag %lld and cmd tag %lld ch %p sess %p\n",
1593 		 srp_tsk->tsk_mgmt_func, srp_tsk->task_tag, srp_tsk->tag, ch,
1594 		 ch->sess);
1595 
1596 	srpt_set_cmd_state(send_ioctx, SRPT_STATE_MGMT);
1597 	send_ioctx->cmd.tag = srp_tsk->tag;
1598 	tcm_tmr = srp_tmr_to_tcm(srp_tsk->tsk_mgmt_func);
1599 	rc = target_submit_tmr(&send_ioctx->cmd, sess, NULL,
1600 			       scsilun_to_int(&srp_tsk->lun), srp_tsk, tcm_tmr,
1601 			       GFP_KERNEL, srp_tsk->task_tag,
1602 			       TARGET_SCF_ACK_KREF);
1603 	if (rc != 0) {
1604 		send_ioctx->cmd.se_tmr_req->response = TMR_FUNCTION_REJECTED;
1605 		cmd->se_tfo->queue_tm_rsp(cmd);
1606 	}
1607 	return;
1608 }
1609 
1610 /**
1611  * srpt_handle_new_iu - process a newly received information unit
1612  * @ch:    RDMA channel through which the information unit has been received.
1613  * @recv_ioctx: Receive I/O context associated with the information unit.
1614  */
1615 static bool
1616 srpt_handle_new_iu(struct srpt_rdma_ch *ch, struct srpt_recv_ioctx *recv_ioctx)
1617 {
1618 	struct srpt_send_ioctx *send_ioctx = NULL;
1619 	struct srp_cmd *srp_cmd;
1620 	bool res = false;
1621 	u8 opcode;
1622 
1623 	BUG_ON(!ch);
1624 	BUG_ON(!recv_ioctx);
1625 
1626 	if (unlikely(ch->state == CH_CONNECTING))
1627 		goto push;
1628 
1629 	ib_dma_sync_single_for_cpu(ch->sport->sdev->device,
1630 				   recv_ioctx->ioctx.dma,
1631 				   recv_ioctx->ioctx.offset + srp_max_req_size,
1632 				   DMA_FROM_DEVICE);
1633 
1634 	srp_cmd = recv_ioctx->ioctx.buf + recv_ioctx->ioctx.offset;
1635 	opcode = srp_cmd->opcode;
1636 	if (opcode == SRP_CMD || opcode == SRP_TSK_MGMT) {
1637 		send_ioctx = srpt_get_send_ioctx(ch);
1638 		if (unlikely(!send_ioctx))
1639 			goto push;
1640 	}
1641 
1642 	if (!list_empty(&recv_ioctx->wait_list)) {
1643 		WARN_ON_ONCE(!ch->processing_wait_list);
1644 		list_del_init(&recv_ioctx->wait_list);
1645 	}
1646 
1647 	switch (opcode) {
1648 	case SRP_CMD:
1649 		srpt_handle_cmd(ch, recv_ioctx, send_ioctx);
1650 		break;
1651 	case SRP_TSK_MGMT:
1652 		srpt_handle_tsk_mgmt(ch, recv_ioctx, send_ioctx);
1653 		break;
1654 	case SRP_I_LOGOUT:
1655 		pr_err("Not yet implemented: SRP_I_LOGOUT\n");
1656 		break;
1657 	case SRP_CRED_RSP:
1658 		pr_debug("received SRP_CRED_RSP\n");
1659 		break;
1660 	case SRP_AER_RSP:
1661 		pr_debug("received SRP_AER_RSP\n");
1662 		break;
1663 	case SRP_RSP:
1664 		pr_err("Received SRP_RSP\n");
1665 		break;
1666 	default:
1667 		pr_err("received IU with unknown opcode 0x%x\n", opcode);
1668 		break;
1669 	}
1670 
1671 	if (!send_ioctx || !send_ioctx->recv_ioctx)
1672 		srpt_post_recv(ch->sport->sdev, ch, recv_ioctx);
1673 	res = true;
1674 
1675 out:
1676 	return res;
1677 
1678 push:
1679 	if (list_empty(&recv_ioctx->wait_list)) {
1680 		WARN_ON_ONCE(ch->processing_wait_list);
1681 		list_add_tail(&recv_ioctx->wait_list, &ch->cmd_wait_list);
1682 	}
1683 	goto out;
1684 }
1685 
1686 static void srpt_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1687 {
1688 	struct srpt_rdma_ch *ch = wc->qp->qp_context;
1689 	struct srpt_recv_ioctx *ioctx =
1690 		container_of(wc->wr_cqe, struct srpt_recv_ioctx, ioctx.cqe);
1691 
1692 	if (wc->status == IB_WC_SUCCESS) {
1693 		int req_lim;
1694 
1695 		req_lim = atomic_dec_return(&ch->req_lim);
1696 		if (unlikely(req_lim < 0))
1697 			pr_err("req_lim = %d < 0\n", req_lim);
1698 		ioctx->byte_len = wc->byte_len;
1699 		srpt_handle_new_iu(ch, ioctx);
1700 	} else {
1701 		pr_info_ratelimited("receiving failed for ioctx %p with status %d\n",
1702 				    ioctx, wc->status);
1703 	}
1704 }
1705 
1706 /*
1707  * This function must be called from the context in which RDMA completions are
1708  * processed because it accesses the wait list without protection against
1709  * access from other threads.
1710  */
1711 static void srpt_process_wait_list(struct srpt_rdma_ch *ch)
1712 {
1713 	struct srpt_recv_ioctx *recv_ioctx, *tmp;
1714 
1715 	WARN_ON_ONCE(ch->state == CH_CONNECTING);
1716 
1717 	if (list_empty(&ch->cmd_wait_list))
1718 		return;
1719 
1720 	WARN_ON_ONCE(ch->processing_wait_list);
1721 	ch->processing_wait_list = true;
1722 	list_for_each_entry_safe(recv_ioctx, tmp, &ch->cmd_wait_list,
1723 				 wait_list) {
1724 		if (!srpt_handle_new_iu(ch, recv_ioctx))
1725 			break;
1726 	}
1727 	ch->processing_wait_list = false;
1728 }
1729 
1730 /**
1731  * srpt_send_done - send completion callback
1732  * @cq: Completion queue.
1733  * @wc: Work completion.
1734  *
1735  * Note: Although this has not yet been observed during tests, at least in
1736  * theory it is possible that the srpt_get_send_ioctx() call invoked by
1737  * srpt_handle_new_iu() fails. This is possible because the req_lim_delta
1738  * value in each response is set to one, and it is possible that this response
1739  * makes the initiator send a new request before the send completion for that
1740  * response has been processed. This could e.g. happen if the call to
1741  * srpt_put_send_iotcx() is delayed because of a higher priority interrupt or
1742  * if IB retransmission causes generation of the send completion to be
1743  * delayed. Incoming information units for which srpt_get_send_ioctx() fails
1744  * are queued on cmd_wait_list. The code below processes these delayed
1745  * requests one at a time.
1746  */
1747 static void srpt_send_done(struct ib_cq *cq, struct ib_wc *wc)
1748 {
1749 	struct srpt_rdma_ch *ch = wc->qp->qp_context;
1750 	struct srpt_send_ioctx *ioctx =
1751 		container_of(wc->wr_cqe, struct srpt_send_ioctx, ioctx.cqe);
1752 	enum srpt_command_state state;
1753 
1754 	state = srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
1755 
1756 	WARN_ON(state != SRPT_STATE_CMD_RSP_SENT &&
1757 		state != SRPT_STATE_MGMT_RSP_SENT);
1758 
1759 	atomic_add(1 + ioctx->n_rdma, &ch->sq_wr_avail);
1760 
1761 	if (wc->status != IB_WC_SUCCESS)
1762 		pr_info("sending response for ioctx 0x%p failed with status %d\n",
1763 			ioctx, wc->status);
1764 
1765 	if (state != SRPT_STATE_DONE) {
1766 		transport_generic_free_cmd(&ioctx->cmd, 0);
1767 	} else {
1768 		pr_err("IB completion has been received too late for wr_id = %u.\n",
1769 		       ioctx->ioctx.index);
1770 	}
1771 
1772 	srpt_process_wait_list(ch);
1773 }
1774 
1775 /**
1776  * srpt_create_ch_ib - create receive and send completion queues
1777  * @ch: SRPT RDMA channel.
1778  */
1779 static int srpt_create_ch_ib(struct srpt_rdma_ch *ch)
1780 {
1781 	struct ib_qp_init_attr *qp_init;
1782 	struct srpt_port *sport = ch->sport;
1783 	struct srpt_device *sdev = sport->sdev;
1784 	const struct ib_device_attr *attrs = &sdev->device->attrs;
1785 	int sq_size = sport->port_attrib.srp_sq_size;
1786 	int i, ret;
1787 
1788 	WARN_ON(ch->rq_size < 1);
1789 
1790 	ret = -ENOMEM;
1791 	qp_init = kzalloc(sizeof(*qp_init), GFP_KERNEL);
1792 	if (!qp_init)
1793 		goto out;
1794 
1795 retry:
1796 	ch->cq = ib_cq_pool_get(sdev->device, ch->rq_size + sq_size, -1,
1797 				 IB_POLL_WORKQUEUE);
1798 	if (IS_ERR(ch->cq)) {
1799 		ret = PTR_ERR(ch->cq);
1800 		pr_err("failed to create CQ cqe= %d ret= %d\n",
1801 		       ch->rq_size + sq_size, ret);
1802 		goto out;
1803 	}
1804 	ch->cq_size = ch->rq_size + sq_size;
1805 
1806 	qp_init->qp_context = (void *)ch;
1807 	qp_init->event_handler
1808 		= (void(*)(struct ib_event *, void*))srpt_qp_event;
1809 	qp_init->send_cq = ch->cq;
1810 	qp_init->recv_cq = ch->cq;
1811 	qp_init->sq_sig_type = IB_SIGNAL_REQ_WR;
1812 	qp_init->qp_type = IB_QPT_RC;
1813 	/*
1814 	 * We divide up our send queue size into half SEND WRs to send the
1815 	 * completions, and half R/W contexts to actually do the RDMA
1816 	 * READ/WRITE transfers.  Note that we need to allocate CQ slots for
1817 	 * both both, as RDMA contexts will also post completions for the
1818 	 * RDMA READ case.
1819 	 */
1820 	qp_init->cap.max_send_wr = min(sq_size / 2, attrs->max_qp_wr);
1821 	qp_init->cap.max_rdma_ctxs = sq_size / 2;
1822 	qp_init->cap.max_send_sge = attrs->max_send_sge;
1823 	qp_init->cap.max_recv_sge = 1;
1824 	qp_init->port_num = ch->sport->port;
1825 	if (sdev->use_srq)
1826 		qp_init->srq = sdev->srq;
1827 	else
1828 		qp_init->cap.max_recv_wr = ch->rq_size;
1829 
1830 	if (ch->using_rdma_cm) {
1831 		ret = rdma_create_qp(ch->rdma_cm.cm_id, sdev->pd, qp_init);
1832 		ch->qp = ch->rdma_cm.cm_id->qp;
1833 	} else {
1834 		ch->qp = ib_create_qp(sdev->pd, qp_init);
1835 		if (!IS_ERR(ch->qp)) {
1836 			ret = srpt_init_ch_qp(ch, ch->qp);
1837 			if (ret)
1838 				ib_destroy_qp(ch->qp);
1839 		} else {
1840 			ret = PTR_ERR(ch->qp);
1841 		}
1842 	}
1843 	if (ret) {
1844 		bool retry = sq_size > MIN_SRPT_SQ_SIZE;
1845 
1846 		if (retry) {
1847 			pr_debug("failed to create queue pair with sq_size = %d (%d) - retrying\n",
1848 				 sq_size, ret);
1849 			ib_cq_pool_put(ch->cq, ch->cq_size);
1850 			sq_size = max(sq_size / 2, MIN_SRPT_SQ_SIZE);
1851 			goto retry;
1852 		} else {
1853 			pr_err("failed to create queue pair with sq_size = %d (%d)\n",
1854 			       sq_size, ret);
1855 			goto err_destroy_cq;
1856 		}
1857 	}
1858 
1859 	atomic_set(&ch->sq_wr_avail, qp_init->cap.max_send_wr);
1860 
1861 	pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d ch= %p\n",
1862 		 __func__, ch->cq->cqe, qp_init->cap.max_send_sge,
1863 		 qp_init->cap.max_send_wr, ch);
1864 
1865 	if (!sdev->use_srq)
1866 		for (i = 0; i < ch->rq_size; i++)
1867 			srpt_post_recv(sdev, ch, ch->ioctx_recv_ring[i]);
1868 
1869 out:
1870 	kfree(qp_init);
1871 	return ret;
1872 
1873 err_destroy_cq:
1874 	ch->qp = NULL;
1875 	ib_cq_pool_put(ch->cq, ch->cq_size);
1876 	goto out;
1877 }
1878 
1879 static void srpt_destroy_ch_ib(struct srpt_rdma_ch *ch)
1880 {
1881 	ib_destroy_qp(ch->qp);
1882 	ib_cq_pool_put(ch->cq, ch->cq_size);
1883 }
1884 
1885 /**
1886  * srpt_close_ch - close a RDMA channel
1887  * @ch: SRPT RDMA channel.
1888  *
1889  * Make sure all resources associated with the channel will be deallocated at
1890  * an appropriate time.
1891  *
1892  * Returns true if and only if the channel state has been modified into
1893  * CH_DRAINING.
1894  */
1895 static bool srpt_close_ch(struct srpt_rdma_ch *ch)
1896 {
1897 	int ret;
1898 
1899 	if (!srpt_set_ch_state(ch, CH_DRAINING)) {
1900 		pr_debug("%s: already closed\n", ch->sess_name);
1901 		return false;
1902 	}
1903 
1904 	kref_get(&ch->kref);
1905 
1906 	ret = srpt_ch_qp_err(ch);
1907 	if (ret < 0)
1908 		pr_err("%s-%d: changing queue pair into error state failed: %d\n",
1909 		       ch->sess_name, ch->qp->qp_num, ret);
1910 
1911 	ret = srpt_zerolength_write(ch);
1912 	if (ret < 0) {
1913 		pr_err("%s-%d: queuing zero-length write failed: %d\n",
1914 		       ch->sess_name, ch->qp->qp_num, ret);
1915 		if (srpt_set_ch_state(ch, CH_DISCONNECTED))
1916 			schedule_work(&ch->release_work);
1917 		else
1918 			WARN_ON_ONCE(true);
1919 	}
1920 
1921 	kref_put(&ch->kref, srpt_free_ch);
1922 
1923 	return true;
1924 }
1925 
1926 /*
1927  * Change the channel state into CH_DISCONNECTING. If a channel has not yet
1928  * reached the connected state, close it. If a channel is in the connected
1929  * state, send a DREQ. If a DREQ has been received, send a DREP. Note: it is
1930  * the responsibility of the caller to ensure that this function is not
1931  * invoked concurrently with the code that accepts a connection. This means
1932  * that this function must either be invoked from inside a CM callback
1933  * function or that it must be invoked with the srpt_port.mutex held.
1934  */
1935 static int srpt_disconnect_ch(struct srpt_rdma_ch *ch)
1936 {
1937 	int ret;
1938 
1939 	if (!srpt_set_ch_state(ch, CH_DISCONNECTING))
1940 		return -ENOTCONN;
1941 
1942 	if (ch->using_rdma_cm) {
1943 		ret = rdma_disconnect(ch->rdma_cm.cm_id);
1944 	} else {
1945 		ret = ib_send_cm_dreq(ch->ib_cm.cm_id, NULL, 0);
1946 		if (ret < 0)
1947 			ret = ib_send_cm_drep(ch->ib_cm.cm_id, NULL, 0);
1948 	}
1949 
1950 	if (ret < 0 && srpt_close_ch(ch))
1951 		ret = 0;
1952 
1953 	return ret;
1954 }
1955 
1956 /* Send DREQ and wait for DREP. */
1957 static void srpt_disconnect_ch_sync(struct srpt_rdma_ch *ch)
1958 {
1959 	DECLARE_COMPLETION_ONSTACK(closed);
1960 	struct srpt_port *sport = ch->sport;
1961 
1962 	pr_debug("ch %s-%d state %d\n", ch->sess_name, ch->qp->qp_num,
1963 		 ch->state);
1964 
1965 	ch->closed = &closed;
1966 
1967 	mutex_lock(&sport->mutex);
1968 	srpt_disconnect_ch(ch);
1969 	mutex_unlock(&sport->mutex);
1970 
1971 	while (wait_for_completion_timeout(&closed, 5 * HZ) == 0)
1972 		pr_info("%s(%s-%d state %d): still waiting ...\n", __func__,
1973 			ch->sess_name, ch->qp->qp_num, ch->state);
1974 
1975 }
1976 
1977 static void __srpt_close_all_ch(struct srpt_port *sport)
1978 {
1979 	struct srpt_nexus *nexus;
1980 	struct srpt_rdma_ch *ch;
1981 
1982 	lockdep_assert_held(&sport->mutex);
1983 
1984 	list_for_each_entry(nexus, &sport->nexus_list, entry) {
1985 		list_for_each_entry(ch, &nexus->ch_list, list) {
1986 			if (srpt_disconnect_ch(ch) >= 0)
1987 				pr_info("Closing channel %s-%d because target %s_%d has been disabled\n",
1988 					ch->sess_name, ch->qp->qp_num,
1989 					dev_name(&sport->sdev->device->dev),
1990 					sport->port);
1991 			srpt_close_ch(ch);
1992 		}
1993 	}
1994 }
1995 
1996 /*
1997  * Look up (i_port_id, t_port_id) in sport->nexus_list. Create an entry if
1998  * it does not yet exist.
1999  */
2000 static struct srpt_nexus *srpt_get_nexus(struct srpt_port *sport,
2001 					 const u8 i_port_id[16],
2002 					 const u8 t_port_id[16])
2003 {
2004 	struct srpt_nexus *nexus = NULL, *tmp_nexus = NULL, *n;
2005 
2006 	for (;;) {
2007 		mutex_lock(&sport->mutex);
2008 		list_for_each_entry(n, &sport->nexus_list, entry) {
2009 			if (memcmp(n->i_port_id, i_port_id, 16) == 0 &&
2010 			    memcmp(n->t_port_id, t_port_id, 16) == 0) {
2011 				nexus = n;
2012 				break;
2013 			}
2014 		}
2015 		if (!nexus && tmp_nexus) {
2016 			list_add_tail_rcu(&tmp_nexus->entry,
2017 					  &sport->nexus_list);
2018 			swap(nexus, tmp_nexus);
2019 		}
2020 		mutex_unlock(&sport->mutex);
2021 
2022 		if (nexus)
2023 			break;
2024 		tmp_nexus = kzalloc(sizeof(*nexus), GFP_KERNEL);
2025 		if (!tmp_nexus) {
2026 			nexus = ERR_PTR(-ENOMEM);
2027 			break;
2028 		}
2029 		INIT_LIST_HEAD(&tmp_nexus->ch_list);
2030 		memcpy(tmp_nexus->i_port_id, i_port_id, 16);
2031 		memcpy(tmp_nexus->t_port_id, t_port_id, 16);
2032 	}
2033 
2034 	kfree(tmp_nexus);
2035 
2036 	return nexus;
2037 }
2038 
2039 static void srpt_set_enabled(struct srpt_port *sport, bool enabled)
2040 	__must_hold(&sport->mutex)
2041 {
2042 	lockdep_assert_held(&sport->mutex);
2043 
2044 	if (sport->enabled == enabled)
2045 		return;
2046 	sport->enabled = enabled;
2047 	if (!enabled)
2048 		__srpt_close_all_ch(sport);
2049 }
2050 
2051 static void srpt_drop_sport_ref(struct srpt_port *sport)
2052 {
2053 	if (atomic_dec_return(&sport->refcount) == 0 && sport->freed_channels)
2054 		complete(sport->freed_channels);
2055 }
2056 
2057 static void srpt_free_ch(struct kref *kref)
2058 {
2059 	struct srpt_rdma_ch *ch = container_of(kref, struct srpt_rdma_ch, kref);
2060 
2061 	srpt_drop_sport_ref(ch->sport);
2062 	kfree_rcu(ch, rcu);
2063 }
2064 
2065 /*
2066  * Shut down the SCSI target session, tell the connection manager to
2067  * disconnect the associated RDMA channel, transition the QP to the error
2068  * state and remove the channel from the channel list. This function is
2069  * typically called from inside srpt_zerolength_write_done(). Concurrent
2070  * srpt_zerolength_write() calls from inside srpt_close_ch() are possible
2071  * as long as the channel is on sport->nexus_list.
2072  */
2073 static void srpt_release_channel_work(struct work_struct *w)
2074 {
2075 	struct srpt_rdma_ch *ch;
2076 	struct srpt_device *sdev;
2077 	struct srpt_port *sport;
2078 	struct se_session *se_sess;
2079 
2080 	ch = container_of(w, struct srpt_rdma_ch, release_work);
2081 	pr_debug("%s-%d\n", ch->sess_name, ch->qp->qp_num);
2082 
2083 	sdev = ch->sport->sdev;
2084 	BUG_ON(!sdev);
2085 
2086 	se_sess = ch->sess;
2087 	BUG_ON(!se_sess);
2088 
2089 	target_stop_session(se_sess);
2090 	target_wait_for_sess_cmds(se_sess);
2091 
2092 	target_remove_session(se_sess);
2093 	ch->sess = NULL;
2094 
2095 	if (ch->using_rdma_cm)
2096 		rdma_destroy_id(ch->rdma_cm.cm_id);
2097 	else
2098 		ib_destroy_cm_id(ch->ib_cm.cm_id);
2099 
2100 	sport = ch->sport;
2101 	mutex_lock(&sport->mutex);
2102 	list_del_rcu(&ch->list);
2103 	mutex_unlock(&sport->mutex);
2104 
2105 	if (ch->closed)
2106 		complete(ch->closed);
2107 
2108 	srpt_destroy_ch_ib(ch);
2109 
2110 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
2111 			     ch->sport->sdev, ch->rq_size,
2112 			     ch->rsp_buf_cache, DMA_TO_DEVICE);
2113 
2114 	kmem_cache_destroy(ch->rsp_buf_cache);
2115 
2116 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_recv_ring,
2117 			     sdev, ch->rq_size,
2118 			     ch->req_buf_cache, DMA_FROM_DEVICE);
2119 
2120 	kmem_cache_destroy(ch->req_buf_cache);
2121 
2122 	kref_put(&ch->kref, srpt_free_ch);
2123 }
2124 
2125 /**
2126  * srpt_cm_req_recv - process the event IB_CM_REQ_RECEIVED
2127  * @sdev: HCA through which the login request was received.
2128  * @ib_cm_id: IB/CM connection identifier in case of IB/CM.
2129  * @rdma_cm_id: RDMA/CM connection identifier in case of RDMA/CM.
2130  * @port_num: Port through which the REQ message was received.
2131  * @pkey: P_Key of the incoming connection.
2132  * @req: SRP login request.
2133  * @src_addr: GID (IB/CM) or IP address (RDMA/CM) of the port that submitted
2134  * the login request.
2135  *
2136  * Ownership of the cm_id is transferred to the target session if this
2137  * function returns zero. Otherwise the caller remains the owner of cm_id.
2138  */
2139 static int srpt_cm_req_recv(struct srpt_device *const sdev,
2140 			    struct ib_cm_id *ib_cm_id,
2141 			    struct rdma_cm_id *rdma_cm_id,
2142 			    u8 port_num, __be16 pkey,
2143 			    const struct srp_login_req *req,
2144 			    const char *src_addr)
2145 {
2146 	struct srpt_port *sport = &sdev->port[port_num - 1];
2147 	struct srpt_nexus *nexus;
2148 	struct srp_login_rsp *rsp = NULL;
2149 	struct srp_login_rej *rej = NULL;
2150 	union {
2151 		struct rdma_conn_param rdma_cm;
2152 		struct ib_cm_rep_param ib_cm;
2153 	} *rep_param = NULL;
2154 	struct srpt_rdma_ch *ch = NULL;
2155 	char i_port_id[36];
2156 	u32 it_iu_len;
2157 	int i, tag_num, tag_size, ret;
2158 	struct srpt_tpg *stpg;
2159 
2160 	WARN_ON_ONCE(irqs_disabled());
2161 
2162 	it_iu_len = be32_to_cpu(req->req_it_iu_len);
2163 
2164 	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",
2165 		req->initiator_port_id, req->target_port_id, it_iu_len,
2166 		port_num, &sport->gid, be16_to_cpu(pkey));
2167 
2168 	nexus = srpt_get_nexus(sport, req->initiator_port_id,
2169 			       req->target_port_id);
2170 	if (IS_ERR(nexus)) {
2171 		ret = PTR_ERR(nexus);
2172 		goto out;
2173 	}
2174 
2175 	ret = -ENOMEM;
2176 	rsp = kzalloc(sizeof(*rsp), GFP_KERNEL);
2177 	rej = kzalloc(sizeof(*rej), GFP_KERNEL);
2178 	rep_param = kzalloc(sizeof(*rep_param), GFP_KERNEL);
2179 	if (!rsp || !rej || !rep_param)
2180 		goto out;
2181 
2182 	ret = -EINVAL;
2183 	if (it_iu_len > srp_max_req_size || it_iu_len < 64) {
2184 		rej->reason = cpu_to_be32(
2185 				SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE);
2186 		pr_err("rejected SRP_LOGIN_REQ because its length (%d bytes) is out of range (%d .. %d)\n",
2187 		       it_iu_len, 64, srp_max_req_size);
2188 		goto reject;
2189 	}
2190 
2191 	if (!sport->enabled) {
2192 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2193 		pr_info("rejected SRP_LOGIN_REQ because target port %s_%d has not yet been enabled\n",
2194 			dev_name(&sport->sdev->device->dev), port_num);
2195 		goto reject;
2196 	}
2197 
2198 	if (*(__be64 *)req->target_port_id != cpu_to_be64(srpt_service_guid)
2199 	    || *(__be64 *)(req->target_port_id + 8) !=
2200 	       cpu_to_be64(srpt_service_guid)) {
2201 		rej->reason = cpu_to_be32(
2202 				SRP_LOGIN_REJ_UNABLE_ASSOCIATE_CHANNEL);
2203 		pr_err("rejected SRP_LOGIN_REQ because it has an invalid target port identifier.\n");
2204 		goto reject;
2205 	}
2206 
2207 	ret = -ENOMEM;
2208 	ch = kzalloc(sizeof(*ch), GFP_KERNEL);
2209 	if (!ch) {
2210 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2211 		pr_err("rejected SRP_LOGIN_REQ because out of memory.\n");
2212 		goto reject;
2213 	}
2214 
2215 	kref_init(&ch->kref);
2216 	ch->pkey = be16_to_cpu(pkey);
2217 	ch->nexus = nexus;
2218 	ch->zw_cqe.done = srpt_zerolength_write_done;
2219 	INIT_WORK(&ch->release_work, srpt_release_channel_work);
2220 	ch->sport = sport;
2221 	if (rdma_cm_id) {
2222 		ch->using_rdma_cm = true;
2223 		ch->rdma_cm.cm_id = rdma_cm_id;
2224 		rdma_cm_id->context = ch;
2225 	} else {
2226 		ch->ib_cm.cm_id = ib_cm_id;
2227 		ib_cm_id->context = ch;
2228 	}
2229 	/*
2230 	 * ch->rq_size should be at least as large as the initiator queue
2231 	 * depth to avoid that the initiator driver has to report QUEUE_FULL
2232 	 * to the SCSI mid-layer.
2233 	 */
2234 	ch->rq_size = min(MAX_SRPT_RQ_SIZE, sdev->device->attrs.max_qp_wr);
2235 	spin_lock_init(&ch->spinlock);
2236 	ch->state = CH_CONNECTING;
2237 	INIT_LIST_HEAD(&ch->cmd_wait_list);
2238 	ch->max_rsp_size = ch->sport->port_attrib.srp_max_rsp_size;
2239 
2240 	ch->rsp_buf_cache = kmem_cache_create("srpt-rsp-buf", ch->max_rsp_size,
2241 					      512, 0, NULL);
2242 	if (!ch->rsp_buf_cache)
2243 		goto free_ch;
2244 
2245 	ch->ioctx_ring = (struct srpt_send_ioctx **)
2246 		srpt_alloc_ioctx_ring(ch->sport->sdev, ch->rq_size,
2247 				      sizeof(*ch->ioctx_ring[0]),
2248 				      ch->rsp_buf_cache, 0, DMA_TO_DEVICE);
2249 	if (!ch->ioctx_ring) {
2250 		pr_err("rejected SRP_LOGIN_REQ because creating a new QP SQ ring failed.\n");
2251 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2252 		goto free_rsp_cache;
2253 	}
2254 
2255 	for (i = 0; i < ch->rq_size; i++)
2256 		ch->ioctx_ring[i]->ch = ch;
2257 	if (!sdev->use_srq) {
2258 		u16 imm_data_offset = req->req_flags & SRP_IMMED_REQUESTED ?
2259 			be16_to_cpu(req->imm_data_offset) : 0;
2260 		u16 alignment_offset;
2261 		u32 req_sz;
2262 
2263 		if (req->req_flags & SRP_IMMED_REQUESTED)
2264 			pr_debug("imm_data_offset = %d\n",
2265 				 be16_to_cpu(req->imm_data_offset));
2266 		if (imm_data_offset >= sizeof(struct srp_cmd)) {
2267 			ch->imm_data_offset = imm_data_offset;
2268 			rsp->rsp_flags |= SRP_LOGIN_RSP_IMMED_SUPP;
2269 		} else {
2270 			ch->imm_data_offset = 0;
2271 		}
2272 		alignment_offset = round_up(imm_data_offset, 512) -
2273 			imm_data_offset;
2274 		req_sz = alignment_offset + imm_data_offset + srp_max_req_size;
2275 		ch->req_buf_cache = kmem_cache_create("srpt-req-buf", req_sz,
2276 						      512, 0, NULL);
2277 		if (!ch->req_buf_cache)
2278 			goto free_rsp_ring;
2279 
2280 		ch->ioctx_recv_ring = (struct srpt_recv_ioctx **)
2281 			srpt_alloc_ioctx_ring(ch->sport->sdev, ch->rq_size,
2282 					      sizeof(*ch->ioctx_recv_ring[0]),
2283 					      ch->req_buf_cache,
2284 					      alignment_offset,
2285 					      DMA_FROM_DEVICE);
2286 		if (!ch->ioctx_recv_ring) {
2287 			pr_err("rejected SRP_LOGIN_REQ because creating a new QP RQ ring failed.\n");
2288 			rej->reason =
2289 			    cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2290 			goto free_recv_cache;
2291 		}
2292 		for (i = 0; i < ch->rq_size; i++)
2293 			INIT_LIST_HEAD(&ch->ioctx_recv_ring[i]->wait_list);
2294 	}
2295 
2296 	ret = srpt_create_ch_ib(ch);
2297 	if (ret) {
2298 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2299 		pr_err("rejected SRP_LOGIN_REQ because creating a new RDMA channel failed.\n");
2300 		goto free_recv_ring;
2301 	}
2302 
2303 	strscpy(ch->sess_name, src_addr, sizeof(ch->sess_name));
2304 	snprintf(i_port_id, sizeof(i_port_id), "0x%016llx%016llx",
2305 			be64_to_cpu(*(__be64 *)nexus->i_port_id),
2306 			be64_to_cpu(*(__be64 *)(nexus->i_port_id + 8)));
2307 
2308 	pr_debug("registering src addr %s or i_port_id %s\n", ch->sess_name,
2309 		 i_port_id);
2310 
2311 	tag_num = ch->rq_size;
2312 	tag_size = 1; /* ib_srpt does not use se_sess->sess_cmd_map */
2313 
2314 	if (sport->guid_id) {
2315 		mutex_lock(&sport->guid_id->mutex);
2316 		list_for_each_entry(stpg, &sport->guid_id->tpg_list, entry) {
2317 			if (!IS_ERR_OR_NULL(ch->sess))
2318 				break;
2319 			ch->sess = target_setup_session(&stpg->tpg, tag_num,
2320 						tag_size, TARGET_PROT_NORMAL,
2321 						ch->sess_name, ch, NULL);
2322 		}
2323 		mutex_unlock(&sport->guid_id->mutex);
2324 	}
2325 
2326 	if (sport->gid_id) {
2327 		mutex_lock(&sport->gid_id->mutex);
2328 		list_for_each_entry(stpg, &sport->gid_id->tpg_list, entry) {
2329 			if (!IS_ERR_OR_NULL(ch->sess))
2330 				break;
2331 			ch->sess = target_setup_session(&stpg->tpg, tag_num,
2332 					tag_size, TARGET_PROT_NORMAL, i_port_id,
2333 					ch, NULL);
2334 			if (!IS_ERR_OR_NULL(ch->sess))
2335 				break;
2336 			/* Retry without leading "0x" */
2337 			ch->sess = target_setup_session(&stpg->tpg, tag_num,
2338 						tag_size, TARGET_PROT_NORMAL,
2339 						i_port_id + 2, ch, NULL);
2340 		}
2341 		mutex_unlock(&sport->gid_id->mutex);
2342 	}
2343 
2344 	if (IS_ERR_OR_NULL(ch->sess)) {
2345 		WARN_ON_ONCE(ch->sess == NULL);
2346 		ret = PTR_ERR(ch->sess);
2347 		ch->sess = NULL;
2348 		pr_info("Rejected login for initiator %s: ret = %d.\n",
2349 			ch->sess_name, ret);
2350 		rej->reason = cpu_to_be32(ret == -ENOMEM ?
2351 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES :
2352 				SRP_LOGIN_REJ_CHANNEL_LIMIT_REACHED);
2353 		goto destroy_ib;
2354 	}
2355 
2356 	/*
2357 	 * Once a session has been created destruction of srpt_rdma_ch objects
2358 	 * will decrement sport->refcount. Hence increment sport->refcount now.
2359 	 */
2360 	atomic_inc(&sport->refcount);
2361 
2362 	mutex_lock(&sport->mutex);
2363 
2364 	if ((req->req_flags & SRP_MTCH_ACTION) == SRP_MULTICHAN_SINGLE) {
2365 		struct srpt_rdma_ch *ch2;
2366 
2367 		list_for_each_entry(ch2, &nexus->ch_list, list) {
2368 			if (srpt_disconnect_ch(ch2) < 0)
2369 				continue;
2370 			pr_info("Relogin - closed existing channel %s\n",
2371 				ch2->sess_name);
2372 			rsp->rsp_flags |= SRP_LOGIN_RSP_MULTICHAN_TERMINATED;
2373 		}
2374 	} else {
2375 		rsp->rsp_flags |= SRP_LOGIN_RSP_MULTICHAN_MAINTAINED;
2376 	}
2377 
2378 	list_add_tail_rcu(&ch->list, &nexus->ch_list);
2379 
2380 	if (!sport->enabled) {
2381 		rej->reason = cpu_to_be32(
2382 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2383 		pr_info("rejected SRP_LOGIN_REQ because target %s_%d is not enabled\n",
2384 			dev_name(&sdev->device->dev), port_num);
2385 		mutex_unlock(&sport->mutex);
2386 		ret = -EINVAL;
2387 		goto reject;
2388 	}
2389 
2390 	mutex_unlock(&sport->mutex);
2391 
2392 	ret = ch->using_rdma_cm ? 0 : srpt_ch_qp_rtr(ch, ch->qp);
2393 	if (ret) {
2394 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2395 		pr_err("rejected SRP_LOGIN_REQ because enabling RTR failed (error code = %d)\n",
2396 		       ret);
2397 		goto reject;
2398 	}
2399 
2400 	pr_debug("Establish connection sess=%p name=%s ch=%p\n", ch->sess,
2401 		 ch->sess_name, ch);
2402 
2403 	/* create srp_login_response */
2404 	rsp->opcode = SRP_LOGIN_RSP;
2405 	rsp->tag = req->tag;
2406 	rsp->max_it_iu_len = cpu_to_be32(srp_max_req_size);
2407 	rsp->max_ti_iu_len = req->req_it_iu_len;
2408 	ch->max_ti_iu_len = it_iu_len;
2409 	rsp->buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
2410 				   SRP_BUF_FORMAT_INDIRECT);
2411 	rsp->req_lim_delta = cpu_to_be32(ch->rq_size);
2412 	atomic_set(&ch->req_lim, ch->rq_size);
2413 	atomic_set(&ch->req_lim_delta, 0);
2414 
2415 	/* create cm reply */
2416 	if (ch->using_rdma_cm) {
2417 		rep_param->rdma_cm.private_data = (void *)rsp;
2418 		rep_param->rdma_cm.private_data_len = sizeof(*rsp);
2419 		rep_param->rdma_cm.rnr_retry_count = 7;
2420 		rep_param->rdma_cm.flow_control = 1;
2421 		rep_param->rdma_cm.responder_resources = 4;
2422 		rep_param->rdma_cm.initiator_depth = 4;
2423 	} else {
2424 		rep_param->ib_cm.qp_num = ch->qp->qp_num;
2425 		rep_param->ib_cm.private_data = (void *)rsp;
2426 		rep_param->ib_cm.private_data_len = sizeof(*rsp);
2427 		rep_param->ib_cm.rnr_retry_count = 7;
2428 		rep_param->ib_cm.flow_control = 1;
2429 		rep_param->ib_cm.failover_accepted = 0;
2430 		rep_param->ib_cm.srq = 1;
2431 		rep_param->ib_cm.responder_resources = 4;
2432 		rep_param->ib_cm.initiator_depth = 4;
2433 	}
2434 
2435 	/*
2436 	 * Hold the sport mutex while accepting a connection to avoid that
2437 	 * srpt_disconnect_ch() is invoked concurrently with this code.
2438 	 */
2439 	mutex_lock(&sport->mutex);
2440 	if (sport->enabled && ch->state == CH_CONNECTING) {
2441 		if (ch->using_rdma_cm)
2442 			ret = rdma_accept(rdma_cm_id, &rep_param->rdma_cm);
2443 		else
2444 			ret = ib_send_cm_rep(ib_cm_id, &rep_param->ib_cm);
2445 	} else {
2446 		ret = -EINVAL;
2447 	}
2448 	mutex_unlock(&sport->mutex);
2449 
2450 	switch (ret) {
2451 	case 0:
2452 		break;
2453 	case -EINVAL:
2454 		goto reject;
2455 	default:
2456 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2457 		pr_err("sending SRP_LOGIN_REQ response failed (error code = %d)\n",
2458 		       ret);
2459 		goto reject;
2460 	}
2461 
2462 	goto out;
2463 
2464 destroy_ib:
2465 	srpt_destroy_ch_ib(ch);
2466 
2467 free_recv_ring:
2468 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_recv_ring,
2469 			     ch->sport->sdev, ch->rq_size,
2470 			     ch->req_buf_cache, DMA_FROM_DEVICE);
2471 
2472 free_recv_cache:
2473 	kmem_cache_destroy(ch->req_buf_cache);
2474 
2475 free_rsp_ring:
2476 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
2477 			     ch->sport->sdev, ch->rq_size,
2478 			     ch->rsp_buf_cache, DMA_TO_DEVICE);
2479 
2480 free_rsp_cache:
2481 	kmem_cache_destroy(ch->rsp_buf_cache);
2482 
2483 free_ch:
2484 	if (rdma_cm_id)
2485 		rdma_cm_id->context = NULL;
2486 	else
2487 		ib_cm_id->context = NULL;
2488 	kfree(ch);
2489 	ch = NULL;
2490 
2491 	WARN_ON_ONCE(ret == 0);
2492 
2493 reject:
2494 	pr_info("Rejecting login with reason %#x\n", be32_to_cpu(rej->reason));
2495 	rej->opcode = SRP_LOGIN_REJ;
2496 	rej->tag = req->tag;
2497 	rej->buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
2498 				   SRP_BUF_FORMAT_INDIRECT);
2499 
2500 	if (rdma_cm_id)
2501 		rdma_reject(rdma_cm_id, rej, sizeof(*rej),
2502 			    IB_CM_REJ_CONSUMER_DEFINED);
2503 	else
2504 		ib_send_cm_rej(ib_cm_id, IB_CM_REJ_CONSUMER_DEFINED, NULL, 0,
2505 			       rej, sizeof(*rej));
2506 
2507 	if (ch && ch->sess) {
2508 		srpt_close_ch(ch);
2509 		/*
2510 		 * Tell the caller not to free cm_id since
2511 		 * srpt_release_channel_work() will do that.
2512 		 */
2513 		ret = 0;
2514 	}
2515 
2516 out:
2517 	kfree(rep_param);
2518 	kfree(rsp);
2519 	kfree(rej);
2520 
2521 	return ret;
2522 }
2523 
2524 static int srpt_ib_cm_req_recv(struct ib_cm_id *cm_id,
2525 			       const struct ib_cm_req_event_param *param,
2526 			       void *private_data)
2527 {
2528 	char sguid[40];
2529 
2530 	srpt_format_guid(sguid, sizeof(sguid),
2531 			 &param->primary_path->dgid.global.interface_id);
2532 
2533 	return srpt_cm_req_recv(cm_id->context, cm_id, NULL, param->port,
2534 				param->primary_path->pkey,
2535 				private_data, sguid);
2536 }
2537 
2538 static int srpt_rdma_cm_req_recv(struct rdma_cm_id *cm_id,
2539 				 struct rdma_cm_event *event)
2540 {
2541 	struct srpt_device *sdev;
2542 	struct srp_login_req req;
2543 	const struct srp_login_req_rdma *req_rdma;
2544 	struct sa_path_rec *path_rec = cm_id->route.path_rec;
2545 	char src_addr[40];
2546 
2547 	sdev = ib_get_client_data(cm_id->device, &srpt_client);
2548 	if (!sdev)
2549 		return -ECONNREFUSED;
2550 
2551 	if (event->param.conn.private_data_len < sizeof(*req_rdma))
2552 		return -EINVAL;
2553 
2554 	/* Transform srp_login_req_rdma into srp_login_req. */
2555 	req_rdma = event->param.conn.private_data;
2556 	memset(&req, 0, sizeof(req));
2557 	req.opcode		= req_rdma->opcode;
2558 	req.tag			= req_rdma->tag;
2559 	req.req_it_iu_len	= req_rdma->req_it_iu_len;
2560 	req.req_buf_fmt		= req_rdma->req_buf_fmt;
2561 	req.req_flags		= req_rdma->req_flags;
2562 	memcpy(req.initiator_port_id, req_rdma->initiator_port_id, 16);
2563 	memcpy(req.target_port_id, req_rdma->target_port_id, 16);
2564 	req.imm_data_offset	= req_rdma->imm_data_offset;
2565 
2566 	snprintf(src_addr, sizeof(src_addr), "%pIS",
2567 		 &cm_id->route.addr.src_addr);
2568 
2569 	return srpt_cm_req_recv(sdev, NULL, cm_id, cm_id->port_num,
2570 				path_rec ? path_rec->pkey : 0, &req, src_addr);
2571 }
2572 
2573 static void srpt_cm_rej_recv(struct srpt_rdma_ch *ch,
2574 			     enum ib_cm_rej_reason reason,
2575 			     const u8 *private_data,
2576 			     u8 private_data_len)
2577 {
2578 	char *priv = NULL;
2579 	int i;
2580 
2581 	if (private_data_len && (priv = kmalloc(private_data_len * 3 + 1,
2582 						GFP_KERNEL))) {
2583 		for (i = 0; i < private_data_len; i++)
2584 			sprintf(priv + 3 * i, " %02x", private_data[i]);
2585 	}
2586 	pr_info("Received CM REJ for ch %s-%d; reason %d%s%s.\n",
2587 		ch->sess_name, ch->qp->qp_num, reason, private_data_len ?
2588 		"; private data" : "", priv ? priv : " (?)");
2589 	kfree(priv);
2590 }
2591 
2592 /**
2593  * srpt_cm_rtu_recv - process an IB_CM_RTU_RECEIVED or USER_ESTABLISHED event
2594  * @ch: SRPT RDMA channel.
2595  *
2596  * An RTU (ready to use) message indicates that the connection has been
2597  * established and that the recipient may begin transmitting.
2598  */
2599 static void srpt_cm_rtu_recv(struct srpt_rdma_ch *ch)
2600 {
2601 	int ret;
2602 
2603 	ret = ch->using_rdma_cm ? 0 : srpt_ch_qp_rts(ch, ch->qp);
2604 	if (ret < 0) {
2605 		pr_err("%s-%d: QP transition to RTS failed\n", ch->sess_name,
2606 		       ch->qp->qp_num);
2607 		srpt_close_ch(ch);
2608 		return;
2609 	}
2610 
2611 	/*
2612 	 * Note: calling srpt_close_ch() if the transition to the LIVE state
2613 	 * fails is not necessary since that means that that function has
2614 	 * already been invoked from another thread.
2615 	 */
2616 	if (!srpt_set_ch_state(ch, CH_LIVE)) {
2617 		pr_err("%s-%d: channel transition to LIVE state failed\n",
2618 		       ch->sess_name, ch->qp->qp_num);
2619 		return;
2620 	}
2621 
2622 	/* Trigger wait list processing. */
2623 	ret = srpt_zerolength_write(ch);
2624 	WARN_ONCE(ret < 0, "%d\n", ret);
2625 }
2626 
2627 /**
2628  * srpt_cm_handler - IB connection manager callback function
2629  * @cm_id: IB/CM connection identifier.
2630  * @event: IB/CM event.
2631  *
2632  * A non-zero return value will cause the caller destroy the CM ID.
2633  *
2634  * Note: srpt_cm_handler() must only return a non-zero value when transferring
2635  * ownership of the cm_id to a channel by srpt_cm_req_recv() failed. Returning
2636  * a non-zero value in any other case will trigger a race with the
2637  * ib_destroy_cm_id() call in srpt_release_channel().
2638  */
2639 static int srpt_cm_handler(struct ib_cm_id *cm_id,
2640 			   const struct ib_cm_event *event)
2641 {
2642 	struct srpt_rdma_ch *ch = cm_id->context;
2643 	int ret;
2644 
2645 	ret = 0;
2646 	switch (event->event) {
2647 	case IB_CM_REQ_RECEIVED:
2648 		ret = srpt_ib_cm_req_recv(cm_id, &event->param.req_rcvd,
2649 					  event->private_data);
2650 		break;
2651 	case IB_CM_REJ_RECEIVED:
2652 		srpt_cm_rej_recv(ch, event->param.rej_rcvd.reason,
2653 				 event->private_data,
2654 				 IB_CM_REJ_PRIVATE_DATA_SIZE);
2655 		break;
2656 	case IB_CM_RTU_RECEIVED:
2657 	case IB_CM_USER_ESTABLISHED:
2658 		srpt_cm_rtu_recv(ch);
2659 		break;
2660 	case IB_CM_DREQ_RECEIVED:
2661 		srpt_disconnect_ch(ch);
2662 		break;
2663 	case IB_CM_DREP_RECEIVED:
2664 		pr_info("Received CM DREP message for ch %s-%d.\n",
2665 			ch->sess_name, ch->qp->qp_num);
2666 		srpt_close_ch(ch);
2667 		break;
2668 	case IB_CM_TIMEWAIT_EXIT:
2669 		pr_info("Received CM TimeWait exit for ch %s-%d.\n",
2670 			ch->sess_name, ch->qp->qp_num);
2671 		srpt_close_ch(ch);
2672 		break;
2673 	case IB_CM_REP_ERROR:
2674 		pr_info("Received CM REP error for ch %s-%d.\n", ch->sess_name,
2675 			ch->qp->qp_num);
2676 		break;
2677 	case IB_CM_DREQ_ERROR:
2678 		pr_info("Received CM DREQ ERROR event.\n");
2679 		break;
2680 	case IB_CM_MRA_RECEIVED:
2681 		pr_info("Received CM MRA event\n");
2682 		break;
2683 	default:
2684 		pr_err("received unrecognized CM event %d\n", event->event);
2685 		break;
2686 	}
2687 
2688 	return ret;
2689 }
2690 
2691 static int srpt_rdma_cm_handler(struct rdma_cm_id *cm_id,
2692 				struct rdma_cm_event *event)
2693 {
2694 	struct srpt_rdma_ch *ch = cm_id->context;
2695 	int ret = 0;
2696 
2697 	switch (event->event) {
2698 	case RDMA_CM_EVENT_CONNECT_REQUEST:
2699 		ret = srpt_rdma_cm_req_recv(cm_id, event);
2700 		break;
2701 	case RDMA_CM_EVENT_REJECTED:
2702 		srpt_cm_rej_recv(ch, event->status,
2703 				 event->param.conn.private_data,
2704 				 event->param.conn.private_data_len);
2705 		break;
2706 	case RDMA_CM_EVENT_ESTABLISHED:
2707 		srpt_cm_rtu_recv(ch);
2708 		break;
2709 	case RDMA_CM_EVENT_DISCONNECTED:
2710 		if (ch->state < CH_DISCONNECTING)
2711 			srpt_disconnect_ch(ch);
2712 		else
2713 			srpt_close_ch(ch);
2714 		break;
2715 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
2716 		srpt_close_ch(ch);
2717 		break;
2718 	case RDMA_CM_EVENT_UNREACHABLE:
2719 		pr_info("Received CM REP error for ch %s-%d.\n", ch->sess_name,
2720 			ch->qp->qp_num);
2721 		break;
2722 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
2723 	case RDMA_CM_EVENT_ADDR_CHANGE:
2724 		break;
2725 	default:
2726 		pr_err("received unrecognized RDMA CM event %d\n",
2727 		       event->event);
2728 		break;
2729 	}
2730 
2731 	return ret;
2732 }
2733 
2734 /*
2735  * srpt_write_pending - Start data transfer from initiator to target (write).
2736  */
2737 static int srpt_write_pending(struct se_cmd *se_cmd)
2738 {
2739 	struct srpt_send_ioctx *ioctx =
2740 		container_of(se_cmd, struct srpt_send_ioctx, cmd);
2741 	struct srpt_rdma_ch *ch = ioctx->ch;
2742 	struct ib_send_wr *first_wr = NULL;
2743 	struct ib_cqe *cqe = &ioctx->rdma_cqe;
2744 	enum srpt_command_state new_state;
2745 	int ret, i;
2746 
2747 	if (ioctx->recv_ioctx) {
2748 		srpt_set_cmd_state(ioctx, SRPT_STATE_DATA_IN);
2749 		target_execute_cmd(&ioctx->cmd);
2750 		return 0;
2751 	}
2752 
2753 	new_state = srpt_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA);
2754 	WARN_ON(new_state == SRPT_STATE_DONE);
2755 
2756 	if (atomic_sub_return(ioctx->n_rdma, &ch->sq_wr_avail) < 0) {
2757 		pr_warn("%s: IB send queue full (needed %d)\n",
2758 				__func__, ioctx->n_rdma);
2759 		ret = -ENOMEM;
2760 		goto out_undo;
2761 	}
2762 
2763 	cqe->done = srpt_rdma_read_done;
2764 	for (i = ioctx->n_rw_ctx - 1; i >= 0; i--) {
2765 		struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i];
2766 
2767 		first_wr = rdma_rw_ctx_wrs(&ctx->rw, ch->qp, ch->sport->port,
2768 				cqe, first_wr);
2769 		cqe = NULL;
2770 	}
2771 
2772 	ret = ib_post_send(ch->qp, first_wr, NULL);
2773 	if (ret) {
2774 		pr_err("%s: ib_post_send() returned %d for %d (avail: %d)\n",
2775 			 __func__, ret, ioctx->n_rdma,
2776 			 atomic_read(&ch->sq_wr_avail));
2777 		goto out_undo;
2778 	}
2779 
2780 	return 0;
2781 out_undo:
2782 	atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
2783 	return ret;
2784 }
2785 
2786 static u8 tcm_to_srp_tsk_mgmt_status(const int tcm_mgmt_status)
2787 {
2788 	switch (tcm_mgmt_status) {
2789 	case TMR_FUNCTION_COMPLETE:
2790 		return SRP_TSK_MGMT_SUCCESS;
2791 	case TMR_FUNCTION_REJECTED:
2792 		return SRP_TSK_MGMT_FUNC_NOT_SUPP;
2793 	}
2794 	return SRP_TSK_MGMT_FAILED;
2795 }
2796 
2797 /**
2798  * srpt_queue_response - transmit the response to a SCSI command
2799  * @cmd: SCSI target command.
2800  *
2801  * Callback function called by the TCM core. Must not block since it can be
2802  * invoked on the context of the IB completion handler.
2803  */
2804 static void srpt_queue_response(struct se_cmd *cmd)
2805 {
2806 	struct srpt_send_ioctx *ioctx =
2807 		container_of(cmd, struct srpt_send_ioctx, cmd);
2808 	struct srpt_rdma_ch *ch = ioctx->ch;
2809 	struct srpt_device *sdev = ch->sport->sdev;
2810 	struct ib_send_wr send_wr, *first_wr = &send_wr;
2811 	struct ib_sge sge;
2812 	enum srpt_command_state state;
2813 	int resp_len, ret, i;
2814 	u8 srp_tm_status;
2815 
2816 	state = ioctx->state;
2817 	switch (state) {
2818 	case SRPT_STATE_NEW:
2819 	case SRPT_STATE_DATA_IN:
2820 		ioctx->state = SRPT_STATE_CMD_RSP_SENT;
2821 		break;
2822 	case SRPT_STATE_MGMT:
2823 		ioctx->state = SRPT_STATE_MGMT_RSP_SENT;
2824 		break;
2825 	default:
2826 		WARN(true, "ch %p; cmd %d: unexpected command state %d\n",
2827 			ch, ioctx->ioctx.index, ioctx->state);
2828 		break;
2829 	}
2830 
2831 	if (WARN_ON_ONCE(state == SRPT_STATE_CMD_RSP_SENT))
2832 		return;
2833 
2834 	/* For read commands, transfer the data to the initiator. */
2835 	if (ioctx->cmd.data_direction == DMA_FROM_DEVICE &&
2836 	    ioctx->cmd.data_length &&
2837 	    !ioctx->queue_status_only) {
2838 		for (i = ioctx->n_rw_ctx - 1; i >= 0; i--) {
2839 			struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i];
2840 
2841 			first_wr = rdma_rw_ctx_wrs(&ctx->rw, ch->qp,
2842 					ch->sport->port, NULL, first_wr);
2843 		}
2844 	}
2845 
2846 	if (state != SRPT_STATE_MGMT)
2847 		resp_len = srpt_build_cmd_rsp(ch, ioctx, ioctx->cmd.tag,
2848 					      cmd->scsi_status);
2849 	else {
2850 		srp_tm_status
2851 			= tcm_to_srp_tsk_mgmt_status(cmd->se_tmr_req->response);
2852 		resp_len = srpt_build_tskmgmt_rsp(ch, ioctx, srp_tm_status,
2853 						 ioctx->cmd.tag);
2854 	}
2855 
2856 	atomic_inc(&ch->req_lim);
2857 
2858 	if (unlikely(atomic_sub_return(1 + ioctx->n_rdma,
2859 			&ch->sq_wr_avail) < 0)) {
2860 		pr_warn("%s: IB send queue full (needed %d)\n",
2861 				__func__, ioctx->n_rdma);
2862 		goto out;
2863 	}
2864 
2865 	ib_dma_sync_single_for_device(sdev->device, ioctx->ioctx.dma, resp_len,
2866 				      DMA_TO_DEVICE);
2867 
2868 	sge.addr = ioctx->ioctx.dma;
2869 	sge.length = resp_len;
2870 	sge.lkey = sdev->lkey;
2871 
2872 	ioctx->ioctx.cqe.done = srpt_send_done;
2873 	send_wr.next = NULL;
2874 	send_wr.wr_cqe = &ioctx->ioctx.cqe;
2875 	send_wr.sg_list = &sge;
2876 	send_wr.num_sge = 1;
2877 	send_wr.opcode = IB_WR_SEND;
2878 	send_wr.send_flags = IB_SEND_SIGNALED;
2879 
2880 	ret = ib_post_send(ch->qp, first_wr, NULL);
2881 	if (ret < 0) {
2882 		pr_err("%s: sending cmd response failed for tag %llu (%d)\n",
2883 			__func__, ioctx->cmd.tag, ret);
2884 		goto out;
2885 	}
2886 
2887 	return;
2888 
2889 out:
2890 	atomic_add(1 + ioctx->n_rdma, &ch->sq_wr_avail);
2891 	atomic_dec(&ch->req_lim);
2892 	srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
2893 	target_put_sess_cmd(&ioctx->cmd);
2894 }
2895 
2896 static int srpt_queue_data_in(struct se_cmd *cmd)
2897 {
2898 	srpt_queue_response(cmd);
2899 	return 0;
2900 }
2901 
2902 static void srpt_queue_tm_rsp(struct se_cmd *cmd)
2903 {
2904 	srpt_queue_response(cmd);
2905 }
2906 
2907 /*
2908  * This function is called for aborted commands if no response is sent to the
2909  * initiator. Make sure that the credits freed by aborting a command are
2910  * returned to the initiator the next time a response is sent by incrementing
2911  * ch->req_lim_delta.
2912  */
2913 static void srpt_aborted_task(struct se_cmd *cmd)
2914 {
2915 	struct srpt_send_ioctx *ioctx = container_of(cmd,
2916 				struct srpt_send_ioctx, cmd);
2917 	struct srpt_rdma_ch *ch = ioctx->ch;
2918 
2919 	atomic_inc(&ch->req_lim_delta);
2920 }
2921 
2922 static int srpt_queue_status(struct se_cmd *cmd)
2923 {
2924 	struct srpt_send_ioctx *ioctx;
2925 
2926 	ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
2927 	BUG_ON(ioctx->sense_data != cmd->sense_buffer);
2928 	if (cmd->se_cmd_flags &
2929 	    (SCF_TRANSPORT_TASK_SENSE | SCF_EMULATED_TASK_SENSE))
2930 		WARN_ON(cmd->scsi_status != SAM_STAT_CHECK_CONDITION);
2931 	ioctx->queue_status_only = true;
2932 	srpt_queue_response(cmd);
2933 	return 0;
2934 }
2935 
2936 static void srpt_refresh_port_work(struct work_struct *work)
2937 {
2938 	struct srpt_port *sport = container_of(work, struct srpt_port, work);
2939 
2940 	srpt_refresh_port(sport);
2941 }
2942 
2943 /**
2944  * srpt_release_sport - disable login and wait for associated channels
2945  * @sport: SRPT HCA port.
2946  */
2947 static int srpt_release_sport(struct srpt_port *sport)
2948 {
2949 	DECLARE_COMPLETION_ONSTACK(c);
2950 	struct srpt_nexus *nexus, *next_n;
2951 	struct srpt_rdma_ch *ch;
2952 
2953 	WARN_ON_ONCE(irqs_disabled());
2954 
2955 	sport->freed_channels = &c;
2956 
2957 	mutex_lock(&sport->mutex);
2958 	srpt_set_enabled(sport, false);
2959 	mutex_unlock(&sport->mutex);
2960 
2961 	while (atomic_read(&sport->refcount) > 0 &&
2962 	       wait_for_completion_timeout(&c, 5 * HZ) <= 0) {
2963 		pr_info("%s_%d: waiting for unregistration of %d sessions ...\n",
2964 			dev_name(&sport->sdev->device->dev), sport->port,
2965 			atomic_read(&sport->refcount));
2966 		rcu_read_lock();
2967 		list_for_each_entry(nexus, &sport->nexus_list, entry) {
2968 			list_for_each_entry(ch, &nexus->ch_list, list) {
2969 				pr_info("%s-%d: state %s\n",
2970 					ch->sess_name, ch->qp->qp_num,
2971 					get_ch_state_name(ch->state));
2972 			}
2973 		}
2974 		rcu_read_unlock();
2975 	}
2976 
2977 	mutex_lock(&sport->mutex);
2978 	list_for_each_entry_safe(nexus, next_n, &sport->nexus_list, entry) {
2979 		list_del(&nexus->entry);
2980 		kfree_rcu(nexus, rcu);
2981 	}
2982 	mutex_unlock(&sport->mutex);
2983 
2984 	return 0;
2985 }
2986 
2987 struct port_and_port_id {
2988 	struct srpt_port *sport;
2989 	struct srpt_port_id **port_id;
2990 };
2991 
2992 static struct port_and_port_id __srpt_lookup_port(const char *name)
2993 {
2994 	struct ib_device *dev;
2995 	struct srpt_device *sdev;
2996 	struct srpt_port *sport;
2997 	int i;
2998 
2999 	list_for_each_entry(sdev, &srpt_dev_list, list) {
3000 		dev = sdev->device;
3001 		if (!dev)
3002 			continue;
3003 
3004 		for (i = 0; i < dev->phys_port_cnt; i++) {
3005 			sport = &sdev->port[i];
3006 
3007 			if (strcmp(sport->guid_name, name) == 0) {
3008 				kref_get(&sdev->refcnt);
3009 				return (struct port_and_port_id){
3010 					sport, &sport->guid_id};
3011 			}
3012 			if (strcmp(sport->gid_name, name) == 0) {
3013 				kref_get(&sdev->refcnt);
3014 				return (struct port_and_port_id){
3015 					sport, &sport->gid_id};
3016 			}
3017 		}
3018 	}
3019 
3020 	return (struct port_and_port_id){};
3021 }
3022 
3023 /**
3024  * srpt_lookup_port() - Look up an RDMA port by name
3025  * @name: ASCII port name
3026  *
3027  * Increments the RDMA port reference count if an RDMA port pointer is returned.
3028  * The caller must drop that reference count by calling srpt_port_put_ref().
3029  */
3030 static struct port_and_port_id srpt_lookup_port(const char *name)
3031 {
3032 	struct port_and_port_id papi;
3033 
3034 	spin_lock(&srpt_dev_lock);
3035 	papi = __srpt_lookup_port(name);
3036 	spin_unlock(&srpt_dev_lock);
3037 
3038 	return papi;
3039 }
3040 
3041 static void srpt_free_srq(struct srpt_device *sdev)
3042 {
3043 	if (!sdev->srq)
3044 		return;
3045 
3046 	ib_destroy_srq(sdev->srq);
3047 	srpt_free_ioctx_ring((struct srpt_ioctx **)sdev->ioctx_ring, sdev,
3048 			     sdev->srq_size, sdev->req_buf_cache,
3049 			     DMA_FROM_DEVICE);
3050 	kmem_cache_destroy(sdev->req_buf_cache);
3051 	sdev->srq = NULL;
3052 }
3053 
3054 static int srpt_alloc_srq(struct srpt_device *sdev)
3055 {
3056 	struct ib_srq_init_attr srq_attr = {
3057 		.event_handler = srpt_srq_event,
3058 		.srq_context = (void *)sdev,
3059 		.attr.max_wr = sdev->srq_size,
3060 		.attr.max_sge = 1,
3061 		.srq_type = IB_SRQT_BASIC,
3062 	};
3063 	struct ib_device *device = sdev->device;
3064 	struct ib_srq *srq;
3065 	int i;
3066 
3067 	WARN_ON_ONCE(sdev->srq);
3068 	srq = ib_create_srq(sdev->pd, &srq_attr);
3069 	if (IS_ERR(srq)) {
3070 		pr_debug("ib_create_srq() failed: %ld\n", PTR_ERR(srq));
3071 		return PTR_ERR(srq);
3072 	}
3073 
3074 	pr_debug("create SRQ #wr= %d max_allow=%d dev= %s\n", sdev->srq_size,
3075 		 sdev->device->attrs.max_srq_wr, dev_name(&device->dev));
3076 
3077 	sdev->req_buf_cache = kmem_cache_create("srpt-srq-req-buf",
3078 						srp_max_req_size, 0, 0, NULL);
3079 	if (!sdev->req_buf_cache)
3080 		goto free_srq;
3081 
3082 	sdev->ioctx_ring = (struct srpt_recv_ioctx **)
3083 		srpt_alloc_ioctx_ring(sdev, sdev->srq_size,
3084 				      sizeof(*sdev->ioctx_ring[0]),
3085 				      sdev->req_buf_cache, 0, DMA_FROM_DEVICE);
3086 	if (!sdev->ioctx_ring)
3087 		goto free_cache;
3088 
3089 	sdev->use_srq = true;
3090 	sdev->srq = srq;
3091 
3092 	for (i = 0; i < sdev->srq_size; ++i) {
3093 		INIT_LIST_HEAD(&sdev->ioctx_ring[i]->wait_list);
3094 		srpt_post_recv(sdev, NULL, sdev->ioctx_ring[i]);
3095 	}
3096 
3097 	return 0;
3098 
3099 free_cache:
3100 	kmem_cache_destroy(sdev->req_buf_cache);
3101 
3102 free_srq:
3103 	ib_destroy_srq(srq);
3104 	return -ENOMEM;
3105 }
3106 
3107 static int srpt_use_srq(struct srpt_device *sdev, bool use_srq)
3108 {
3109 	struct ib_device *device = sdev->device;
3110 	int ret = 0;
3111 
3112 	if (!use_srq) {
3113 		srpt_free_srq(sdev);
3114 		sdev->use_srq = false;
3115 	} else if (use_srq && !sdev->srq) {
3116 		ret = srpt_alloc_srq(sdev);
3117 	}
3118 	pr_debug("%s(%s): use_srq = %d; ret = %d\n", __func__,
3119 		 dev_name(&device->dev), sdev->use_srq, ret);
3120 	return ret;
3121 }
3122 
3123 static void srpt_free_sdev(struct kref *refcnt)
3124 {
3125 	struct srpt_device *sdev = container_of(refcnt, typeof(*sdev), refcnt);
3126 
3127 	kfree(sdev);
3128 }
3129 
3130 static void srpt_sdev_put(struct srpt_device *sdev)
3131 {
3132 	kref_put(&sdev->refcnt, srpt_free_sdev);
3133 }
3134 
3135 /**
3136  * srpt_add_one - InfiniBand device addition callback function
3137  * @device: Describes a HCA.
3138  */
3139 static int srpt_add_one(struct ib_device *device)
3140 {
3141 	struct srpt_device *sdev;
3142 	struct srpt_port *sport;
3143 	int ret;
3144 	u32 i;
3145 
3146 	pr_debug("device = %p\n", device);
3147 
3148 	sdev = kzalloc(struct_size(sdev, port, device->phys_port_cnt),
3149 		       GFP_KERNEL);
3150 	if (!sdev)
3151 		return -ENOMEM;
3152 
3153 	kref_init(&sdev->refcnt);
3154 	sdev->device = device;
3155 	mutex_init(&sdev->sdev_mutex);
3156 
3157 	sdev->pd = ib_alloc_pd(device, 0);
3158 	if (IS_ERR(sdev->pd)) {
3159 		ret = PTR_ERR(sdev->pd);
3160 		goto free_dev;
3161 	}
3162 
3163 	sdev->lkey = sdev->pd->local_dma_lkey;
3164 
3165 	sdev->srq_size = min(srpt_srq_size, sdev->device->attrs.max_srq_wr);
3166 
3167 	srpt_use_srq(sdev, sdev->port[0].port_attrib.use_srq);
3168 
3169 	if (!srpt_service_guid)
3170 		srpt_service_guid = be64_to_cpu(device->node_guid);
3171 
3172 	if (rdma_port_get_link_layer(device, 1) == IB_LINK_LAYER_INFINIBAND)
3173 		sdev->cm_id = ib_create_cm_id(device, srpt_cm_handler, sdev);
3174 	if (IS_ERR(sdev->cm_id)) {
3175 		pr_info("ib_create_cm_id() failed: %ld\n",
3176 			PTR_ERR(sdev->cm_id));
3177 		ret = PTR_ERR(sdev->cm_id);
3178 		sdev->cm_id = NULL;
3179 		if (!rdma_cm_id)
3180 			goto err_ring;
3181 	}
3182 
3183 	/* print out target login information */
3184 	pr_debug("Target login info: id_ext=%016llx,ioc_guid=%016llx,pkey=ffff,service_id=%016llx\n",
3185 		 srpt_service_guid, srpt_service_guid, srpt_service_guid);
3186 
3187 	/*
3188 	 * We do not have a consistent service_id (ie. also id_ext of target_id)
3189 	 * to identify this target. We currently use the guid of the first HCA
3190 	 * in the system as service_id; therefore, the target_id will change
3191 	 * if this HCA is gone bad and replaced by different HCA
3192 	 */
3193 	ret = sdev->cm_id ?
3194 		ib_cm_listen(sdev->cm_id, cpu_to_be64(srpt_service_guid)) :
3195 		0;
3196 	if (ret < 0) {
3197 		pr_err("ib_cm_listen() failed: %d (cm_id state = %d)\n", ret,
3198 		       sdev->cm_id->state);
3199 		goto err_cm;
3200 	}
3201 
3202 	INIT_IB_EVENT_HANDLER(&sdev->event_handler, sdev->device,
3203 			      srpt_event_handler);
3204 	ib_register_event_handler(&sdev->event_handler);
3205 
3206 	for (i = 1; i <= sdev->device->phys_port_cnt; i++) {
3207 		sport = &sdev->port[i - 1];
3208 		INIT_LIST_HEAD(&sport->nexus_list);
3209 		mutex_init(&sport->mutex);
3210 		sport->sdev = sdev;
3211 		sport->port = i;
3212 		sport->port_attrib.srp_max_rdma_size = DEFAULT_MAX_RDMA_SIZE;
3213 		sport->port_attrib.srp_max_rsp_size = DEFAULT_MAX_RSP_SIZE;
3214 		sport->port_attrib.srp_sq_size = DEF_SRPT_SQ_SIZE;
3215 		sport->port_attrib.use_srq = false;
3216 		INIT_WORK(&sport->work, srpt_refresh_port_work);
3217 
3218 		ret = srpt_refresh_port(sport);
3219 		if (ret) {
3220 			pr_err("MAD registration failed for %s-%d.\n",
3221 			       dev_name(&sdev->device->dev), i);
3222 			i--;
3223 			goto err_port;
3224 		}
3225 	}
3226 
3227 	spin_lock(&srpt_dev_lock);
3228 	list_add_tail(&sdev->list, &srpt_dev_list);
3229 	spin_unlock(&srpt_dev_lock);
3230 
3231 	ib_set_client_data(device, &srpt_client, sdev);
3232 	pr_debug("added %s.\n", dev_name(&device->dev));
3233 	return 0;
3234 
3235 err_port:
3236 	srpt_unregister_mad_agent(sdev, i);
3237 	ib_unregister_event_handler(&sdev->event_handler);
3238 err_cm:
3239 	if (sdev->cm_id)
3240 		ib_destroy_cm_id(sdev->cm_id);
3241 err_ring:
3242 	srpt_free_srq(sdev);
3243 	ib_dealloc_pd(sdev->pd);
3244 free_dev:
3245 	srpt_sdev_put(sdev);
3246 	pr_info("%s(%s) failed.\n", __func__, dev_name(&device->dev));
3247 	return ret;
3248 }
3249 
3250 /**
3251  * srpt_remove_one - InfiniBand device removal callback function
3252  * @device: Describes a HCA.
3253  * @client_data: The value passed as the third argument to ib_set_client_data().
3254  */
3255 static void srpt_remove_one(struct ib_device *device, void *client_data)
3256 {
3257 	struct srpt_device *sdev = client_data;
3258 	int i;
3259 
3260 	srpt_unregister_mad_agent(sdev, sdev->device->phys_port_cnt);
3261 
3262 	ib_unregister_event_handler(&sdev->event_handler);
3263 
3264 	/* Cancel any work queued by the just unregistered IB event handler. */
3265 	for (i = 0; i < sdev->device->phys_port_cnt; i++)
3266 		cancel_work_sync(&sdev->port[i].work);
3267 
3268 	if (sdev->cm_id)
3269 		ib_destroy_cm_id(sdev->cm_id);
3270 
3271 	ib_set_client_data(device, &srpt_client, NULL);
3272 
3273 	/*
3274 	 * Unregistering a target must happen after destroying sdev->cm_id
3275 	 * such that no new SRP_LOGIN_REQ information units can arrive while
3276 	 * destroying the target.
3277 	 */
3278 	spin_lock(&srpt_dev_lock);
3279 	list_del(&sdev->list);
3280 	spin_unlock(&srpt_dev_lock);
3281 
3282 	for (i = 0; i < sdev->device->phys_port_cnt; i++)
3283 		srpt_release_sport(&sdev->port[i]);
3284 
3285 	srpt_free_srq(sdev);
3286 
3287 	ib_dealloc_pd(sdev->pd);
3288 
3289 	srpt_sdev_put(sdev);
3290 }
3291 
3292 static struct ib_client srpt_client = {
3293 	.name = DRV_NAME,
3294 	.add = srpt_add_one,
3295 	.remove = srpt_remove_one
3296 };
3297 
3298 static int srpt_check_true(struct se_portal_group *se_tpg)
3299 {
3300 	return 1;
3301 }
3302 
3303 static int srpt_check_false(struct se_portal_group *se_tpg)
3304 {
3305 	return 0;
3306 }
3307 
3308 static struct srpt_port *srpt_tpg_to_sport(struct se_portal_group *tpg)
3309 {
3310 	return tpg->se_tpg_wwn->priv;
3311 }
3312 
3313 static struct srpt_port_id *srpt_wwn_to_sport_id(struct se_wwn *wwn)
3314 {
3315 	struct srpt_port *sport = wwn->priv;
3316 
3317 	if (sport->guid_id && &sport->guid_id->wwn == wwn)
3318 		return sport->guid_id;
3319 	if (sport->gid_id && &sport->gid_id->wwn == wwn)
3320 		return sport->gid_id;
3321 	WARN_ON_ONCE(true);
3322 	return NULL;
3323 }
3324 
3325 static char *srpt_get_fabric_wwn(struct se_portal_group *tpg)
3326 {
3327 	struct srpt_tpg *stpg = container_of(tpg, typeof(*stpg), tpg);
3328 
3329 	return stpg->sport_id->name;
3330 }
3331 
3332 static u16 srpt_get_tag(struct se_portal_group *tpg)
3333 {
3334 	return 1;
3335 }
3336 
3337 static u32 srpt_tpg_get_inst_index(struct se_portal_group *se_tpg)
3338 {
3339 	return 1;
3340 }
3341 
3342 static void srpt_release_cmd(struct se_cmd *se_cmd)
3343 {
3344 	struct srpt_send_ioctx *ioctx = container_of(se_cmd,
3345 				struct srpt_send_ioctx, cmd);
3346 	struct srpt_rdma_ch *ch = ioctx->ch;
3347 	struct srpt_recv_ioctx *recv_ioctx = ioctx->recv_ioctx;
3348 
3349 	WARN_ON_ONCE(ioctx->state != SRPT_STATE_DONE &&
3350 		     !(ioctx->cmd.transport_state & CMD_T_ABORTED));
3351 
3352 	if (recv_ioctx) {
3353 		WARN_ON_ONCE(!list_empty(&recv_ioctx->wait_list));
3354 		ioctx->recv_ioctx = NULL;
3355 		srpt_post_recv(ch->sport->sdev, ch, recv_ioctx);
3356 	}
3357 
3358 	if (ioctx->n_rw_ctx) {
3359 		srpt_free_rw_ctxs(ch, ioctx);
3360 		ioctx->n_rw_ctx = 0;
3361 	}
3362 
3363 	target_free_tag(se_cmd->se_sess, se_cmd);
3364 }
3365 
3366 /**
3367  * srpt_close_session - forcibly close a session
3368  * @se_sess: SCSI target session.
3369  *
3370  * Callback function invoked by the TCM core to clean up sessions associated
3371  * with a node ACL when the user invokes
3372  * rmdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3373  */
3374 static void srpt_close_session(struct se_session *se_sess)
3375 {
3376 	struct srpt_rdma_ch *ch = se_sess->fabric_sess_ptr;
3377 
3378 	srpt_disconnect_ch_sync(ch);
3379 }
3380 
3381 /**
3382  * srpt_sess_get_index - return the value of scsiAttIntrPortIndex (SCSI-MIB)
3383  * @se_sess: SCSI target session.
3384  *
3385  * A quote from RFC 4455 (SCSI-MIB) about this MIB object:
3386  * This object represents an arbitrary integer used to uniquely identify a
3387  * particular attached remote initiator port to a particular SCSI target port
3388  * within a particular SCSI target device within a particular SCSI instance.
3389  */
3390 static u32 srpt_sess_get_index(struct se_session *se_sess)
3391 {
3392 	return 0;
3393 }
3394 
3395 static void srpt_set_default_node_attrs(struct se_node_acl *nacl)
3396 {
3397 }
3398 
3399 /* Note: only used from inside debug printk's by the TCM core. */
3400 static int srpt_get_tcm_cmd_state(struct se_cmd *se_cmd)
3401 {
3402 	struct srpt_send_ioctx *ioctx;
3403 
3404 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
3405 	return ioctx->state;
3406 }
3407 
3408 static int srpt_parse_guid(u64 *guid, const char *name)
3409 {
3410 	u16 w[4];
3411 	int ret = -EINVAL;
3412 
3413 	if (sscanf(name, "%hx:%hx:%hx:%hx", &w[0], &w[1], &w[2], &w[3]) != 4)
3414 		goto out;
3415 	*guid = get_unaligned_be64(w);
3416 	ret = 0;
3417 out:
3418 	return ret;
3419 }
3420 
3421 /**
3422  * srpt_parse_i_port_id - parse an initiator port ID
3423  * @name: ASCII representation of a 128-bit initiator port ID.
3424  * @i_port_id: Binary 128-bit port ID.
3425  */
3426 static int srpt_parse_i_port_id(u8 i_port_id[16], const char *name)
3427 {
3428 	const char *p;
3429 	unsigned len, count, leading_zero_bytes;
3430 	int ret;
3431 
3432 	p = name;
3433 	if (strncasecmp(p, "0x", 2) == 0)
3434 		p += 2;
3435 	ret = -EINVAL;
3436 	len = strlen(p);
3437 	if (len % 2)
3438 		goto out;
3439 	count = min(len / 2, 16U);
3440 	leading_zero_bytes = 16 - count;
3441 	memset(i_port_id, 0, leading_zero_bytes);
3442 	ret = hex2bin(i_port_id + leading_zero_bytes, p, count);
3443 
3444 out:
3445 	return ret;
3446 }
3447 
3448 /*
3449  * configfs callback function invoked for mkdir
3450  * /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3451  *
3452  * i_port_id must be an initiator port GUID, GID or IP address. See also the
3453  * target_alloc_session() calls in this driver. Examples of valid initiator
3454  * port IDs:
3455  * 0x0000000000000000505400fffe4a0b7b
3456  * 0000000000000000505400fffe4a0b7b
3457  * 5054:00ff:fe4a:0b7b
3458  * 192.168.122.76
3459  */
3460 static int srpt_init_nodeacl(struct se_node_acl *se_nacl, const char *name)
3461 {
3462 	struct sockaddr_storage sa;
3463 	u64 guid;
3464 	u8 i_port_id[16];
3465 	int ret;
3466 
3467 	ret = srpt_parse_guid(&guid, name);
3468 	if (ret < 0)
3469 		ret = srpt_parse_i_port_id(i_port_id, name);
3470 	if (ret < 0)
3471 		ret = inet_pton_with_scope(&init_net, AF_UNSPEC, name, NULL,
3472 					   &sa);
3473 	if (ret < 0)
3474 		pr_err("invalid initiator port ID %s\n", name);
3475 	return ret;
3476 }
3477 
3478 static ssize_t srpt_tpg_attrib_srp_max_rdma_size_show(struct config_item *item,
3479 		char *page)
3480 {
3481 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3482 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3483 
3484 	return sysfs_emit(page, "%u\n", sport->port_attrib.srp_max_rdma_size);
3485 }
3486 
3487 static ssize_t srpt_tpg_attrib_srp_max_rdma_size_store(struct config_item *item,
3488 		const char *page, size_t count)
3489 {
3490 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3491 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3492 	unsigned long val;
3493 	int ret;
3494 
3495 	ret = kstrtoul(page, 0, &val);
3496 	if (ret < 0) {
3497 		pr_err("kstrtoul() failed with ret: %d\n", ret);
3498 		return -EINVAL;
3499 	}
3500 	if (val > MAX_SRPT_RDMA_SIZE) {
3501 		pr_err("val: %lu exceeds MAX_SRPT_RDMA_SIZE: %d\n", val,
3502 			MAX_SRPT_RDMA_SIZE);
3503 		return -EINVAL;
3504 	}
3505 	if (val < DEFAULT_MAX_RDMA_SIZE) {
3506 		pr_err("val: %lu smaller than DEFAULT_MAX_RDMA_SIZE: %d\n",
3507 			val, DEFAULT_MAX_RDMA_SIZE);
3508 		return -EINVAL;
3509 	}
3510 	sport->port_attrib.srp_max_rdma_size = val;
3511 
3512 	return count;
3513 }
3514 
3515 static ssize_t srpt_tpg_attrib_srp_max_rsp_size_show(struct config_item *item,
3516 		char *page)
3517 {
3518 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3519 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3520 
3521 	return sysfs_emit(page, "%u\n", sport->port_attrib.srp_max_rsp_size);
3522 }
3523 
3524 static ssize_t srpt_tpg_attrib_srp_max_rsp_size_store(struct config_item *item,
3525 		const char *page, size_t count)
3526 {
3527 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3528 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3529 	unsigned long val;
3530 	int ret;
3531 
3532 	ret = kstrtoul(page, 0, &val);
3533 	if (ret < 0) {
3534 		pr_err("kstrtoul() failed with ret: %d\n", ret);
3535 		return -EINVAL;
3536 	}
3537 	if (val > MAX_SRPT_RSP_SIZE) {
3538 		pr_err("val: %lu exceeds MAX_SRPT_RSP_SIZE: %d\n", val,
3539 			MAX_SRPT_RSP_SIZE);
3540 		return -EINVAL;
3541 	}
3542 	if (val < MIN_MAX_RSP_SIZE) {
3543 		pr_err("val: %lu smaller than MIN_MAX_RSP_SIZE: %d\n", val,
3544 			MIN_MAX_RSP_SIZE);
3545 		return -EINVAL;
3546 	}
3547 	sport->port_attrib.srp_max_rsp_size = val;
3548 
3549 	return count;
3550 }
3551 
3552 static ssize_t srpt_tpg_attrib_srp_sq_size_show(struct config_item *item,
3553 		char *page)
3554 {
3555 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3556 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3557 
3558 	return sysfs_emit(page, "%u\n", sport->port_attrib.srp_sq_size);
3559 }
3560 
3561 static ssize_t srpt_tpg_attrib_srp_sq_size_store(struct config_item *item,
3562 		const char *page, size_t count)
3563 {
3564 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3565 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3566 	unsigned long val;
3567 	int ret;
3568 
3569 	ret = kstrtoul(page, 0, &val);
3570 	if (ret < 0) {
3571 		pr_err("kstrtoul() failed with ret: %d\n", ret);
3572 		return -EINVAL;
3573 	}
3574 	if (val > MAX_SRPT_SRQ_SIZE) {
3575 		pr_err("val: %lu exceeds MAX_SRPT_SRQ_SIZE: %d\n", val,
3576 			MAX_SRPT_SRQ_SIZE);
3577 		return -EINVAL;
3578 	}
3579 	if (val < MIN_SRPT_SRQ_SIZE) {
3580 		pr_err("val: %lu smaller than MIN_SRPT_SRQ_SIZE: %d\n", val,
3581 			MIN_SRPT_SRQ_SIZE);
3582 		return -EINVAL;
3583 	}
3584 	sport->port_attrib.srp_sq_size = val;
3585 
3586 	return count;
3587 }
3588 
3589 static ssize_t srpt_tpg_attrib_use_srq_show(struct config_item *item,
3590 					    char *page)
3591 {
3592 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3593 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3594 
3595 	return sysfs_emit(page, "%d\n", sport->port_attrib.use_srq);
3596 }
3597 
3598 static ssize_t srpt_tpg_attrib_use_srq_store(struct config_item *item,
3599 					     const char *page, size_t count)
3600 {
3601 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3602 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3603 	struct srpt_device *sdev = sport->sdev;
3604 	unsigned long val;
3605 	bool enabled;
3606 	int ret;
3607 
3608 	ret = kstrtoul(page, 0, &val);
3609 	if (ret < 0)
3610 		return ret;
3611 	if (val != !!val)
3612 		return -EINVAL;
3613 
3614 	ret = mutex_lock_interruptible(&sdev->sdev_mutex);
3615 	if (ret < 0)
3616 		return ret;
3617 	ret = mutex_lock_interruptible(&sport->mutex);
3618 	if (ret < 0)
3619 		goto unlock_sdev;
3620 	enabled = sport->enabled;
3621 	/* Log out all initiator systems before changing 'use_srq'. */
3622 	srpt_set_enabled(sport, false);
3623 	sport->port_attrib.use_srq = val;
3624 	srpt_use_srq(sdev, sport->port_attrib.use_srq);
3625 	srpt_set_enabled(sport, enabled);
3626 	ret = count;
3627 	mutex_unlock(&sport->mutex);
3628 unlock_sdev:
3629 	mutex_unlock(&sdev->sdev_mutex);
3630 
3631 	return ret;
3632 }
3633 
3634 CONFIGFS_ATTR(srpt_tpg_attrib_,  srp_max_rdma_size);
3635 CONFIGFS_ATTR(srpt_tpg_attrib_,  srp_max_rsp_size);
3636 CONFIGFS_ATTR(srpt_tpg_attrib_,  srp_sq_size);
3637 CONFIGFS_ATTR(srpt_tpg_attrib_,  use_srq);
3638 
3639 static struct configfs_attribute *srpt_tpg_attrib_attrs[] = {
3640 	&srpt_tpg_attrib_attr_srp_max_rdma_size,
3641 	&srpt_tpg_attrib_attr_srp_max_rsp_size,
3642 	&srpt_tpg_attrib_attr_srp_sq_size,
3643 	&srpt_tpg_attrib_attr_use_srq,
3644 	NULL,
3645 };
3646 
3647 static struct rdma_cm_id *srpt_create_rdma_id(struct sockaddr *listen_addr)
3648 {
3649 	struct rdma_cm_id *rdma_cm_id;
3650 	int ret;
3651 
3652 	rdma_cm_id = rdma_create_id(&init_net, srpt_rdma_cm_handler,
3653 				    NULL, RDMA_PS_TCP, IB_QPT_RC);
3654 	if (IS_ERR(rdma_cm_id)) {
3655 		pr_err("RDMA/CM ID creation failed: %ld\n",
3656 		       PTR_ERR(rdma_cm_id));
3657 		goto out;
3658 	}
3659 
3660 	ret = rdma_bind_addr(rdma_cm_id, listen_addr);
3661 	if (ret) {
3662 		char addr_str[64];
3663 
3664 		snprintf(addr_str, sizeof(addr_str), "%pISp", listen_addr);
3665 		pr_err("Binding RDMA/CM ID to address %s failed: %d\n",
3666 		       addr_str, ret);
3667 		rdma_destroy_id(rdma_cm_id);
3668 		rdma_cm_id = ERR_PTR(ret);
3669 		goto out;
3670 	}
3671 
3672 	ret = rdma_listen(rdma_cm_id, 128);
3673 	if (ret) {
3674 		pr_err("rdma_listen() failed: %d\n", ret);
3675 		rdma_destroy_id(rdma_cm_id);
3676 		rdma_cm_id = ERR_PTR(ret);
3677 	}
3678 
3679 out:
3680 	return rdma_cm_id;
3681 }
3682 
3683 static ssize_t srpt_rdma_cm_port_show(struct config_item *item, char *page)
3684 {
3685 	return sysfs_emit(page, "%d\n", rdma_cm_port);
3686 }
3687 
3688 static ssize_t srpt_rdma_cm_port_store(struct config_item *item,
3689 				       const char *page, size_t count)
3690 {
3691 	struct sockaddr_in  addr4 = { .sin_family  = AF_INET  };
3692 	struct sockaddr_in6 addr6 = { .sin6_family = AF_INET6 };
3693 	struct rdma_cm_id *new_id = NULL;
3694 	u16 val;
3695 	int ret;
3696 
3697 	ret = kstrtou16(page, 0, &val);
3698 	if (ret < 0)
3699 		return ret;
3700 	ret = count;
3701 	if (rdma_cm_port == val)
3702 		goto out;
3703 
3704 	if (val) {
3705 		addr6.sin6_port = cpu_to_be16(val);
3706 		new_id = srpt_create_rdma_id((struct sockaddr *)&addr6);
3707 		if (IS_ERR(new_id)) {
3708 			addr4.sin_port = cpu_to_be16(val);
3709 			new_id = srpt_create_rdma_id((struct sockaddr *)&addr4);
3710 			if (IS_ERR(new_id)) {
3711 				ret = PTR_ERR(new_id);
3712 				goto out;
3713 			}
3714 		}
3715 	}
3716 
3717 	mutex_lock(&rdma_cm_mutex);
3718 	rdma_cm_port = val;
3719 	swap(rdma_cm_id, new_id);
3720 	mutex_unlock(&rdma_cm_mutex);
3721 
3722 	if (new_id)
3723 		rdma_destroy_id(new_id);
3724 	ret = count;
3725 out:
3726 	return ret;
3727 }
3728 
3729 CONFIGFS_ATTR(srpt_, rdma_cm_port);
3730 
3731 static struct configfs_attribute *srpt_da_attrs[] = {
3732 	&srpt_attr_rdma_cm_port,
3733 	NULL,
3734 };
3735 
3736 static int srpt_enable_tpg(struct se_portal_group *se_tpg, bool enable)
3737 {
3738 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3739 
3740 	mutex_lock(&sport->mutex);
3741 	srpt_set_enabled(sport, enable);
3742 	mutex_unlock(&sport->mutex);
3743 
3744 	return 0;
3745 }
3746 
3747 /**
3748  * srpt_make_tpg - configfs callback invoked for mkdir /sys/kernel/config/target/$driver/$port/$tpg
3749  * @wwn: Corresponds to $driver/$port.
3750  * @name: $tpg.
3751  */
3752 static struct se_portal_group *srpt_make_tpg(struct se_wwn *wwn,
3753 					     const char *name)
3754 {
3755 	struct srpt_port_id *sport_id = srpt_wwn_to_sport_id(wwn);
3756 	struct srpt_tpg *stpg;
3757 	int res = -ENOMEM;
3758 
3759 	stpg = kzalloc(sizeof(*stpg), GFP_KERNEL);
3760 	if (!stpg)
3761 		return ERR_PTR(res);
3762 	stpg->sport_id = sport_id;
3763 	res = core_tpg_register(wwn, &stpg->tpg, SCSI_PROTOCOL_SRP);
3764 	if (res) {
3765 		kfree(stpg);
3766 		return ERR_PTR(res);
3767 	}
3768 
3769 	mutex_lock(&sport_id->mutex);
3770 	list_add_tail(&stpg->entry, &sport_id->tpg_list);
3771 	mutex_unlock(&sport_id->mutex);
3772 
3773 	return &stpg->tpg;
3774 }
3775 
3776 /**
3777  * srpt_drop_tpg - configfs callback invoked for rmdir /sys/kernel/config/target/$driver/$port/$tpg
3778  * @tpg: Target portal group to deregister.
3779  */
3780 static void srpt_drop_tpg(struct se_portal_group *tpg)
3781 {
3782 	struct srpt_tpg *stpg = container_of(tpg, typeof(*stpg), tpg);
3783 	struct srpt_port_id *sport_id = stpg->sport_id;
3784 	struct srpt_port *sport = srpt_tpg_to_sport(tpg);
3785 
3786 	mutex_lock(&sport_id->mutex);
3787 	list_del(&stpg->entry);
3788 	mutex_unlock(&sport_id->mutex);
3789 
3790 	sport->enabled = false;
3791 	core_tpg_deregister(tpg);
3792 	kfree(stpg);
3793 }
3794 
3795 /**
3796  * srpt_make_tport - configfs callback invoked for mkdir /sys/kernel/config/target/$driver/$port
3797  * @tf: Not used.
3798  * @group: Not used.
3799  * @name: $port.
3800  */
3801 static struct se_wwn *srpt_make_tport(struct target_fabric_configfs *tf,
3802 				      struct config_group *group,
3803 				      const char *name)
3804 {
3805 	struct port_and_port_id papi = srpt_lookup_port(name);
3806 	struct srpt_port *sport = papi.sport;
3807 	struct srpt_port_id *port_id;
3808 
3809 	if (!papi.port_id)
3810 		return ERR_PTR(-EINVAL);
3811 	if (*papi.port_id) {
3812 		/* Attempt to create a directory that already exists. */
3813 		WARN_ON_ONCE(true);
3814 		return &(*papi.port_id)->wwn;
3815 	}
3816 	port_id = kzalloc(sizeof(*port_id), GFP_KERNEL);
3817 	if (!port_id) {
3818 		srpt_sdev_put(sport->sdev);
3819 		return ERR_PTR(-ENOMEM);
3820 	}
3821 	mutex_init(&port_id->mutex);
3822 	INIT_LIST_HEAD(&port_id->tpg_list);
3823 	port_id->wwn.priv = sport;
3824 	memcpy(port_id->name, port_id == sport->guid_id ? sport->guid_name :
3825 	       sport->gid_name, ARRAY_SIZE(port_id->name));
3826 
3827 	*papi.port_id = port_id;
3828 
3829 	return &port_id->wwn;
3830 }
3831 
3832 /**
3833  * srpt_drop_tport - configfs callback invoked for rmdir /sys/kernel/config/target/$driver/$port
3834  * @wwn: $port.
3835  */
3836 static void srpt_drop_tport(struct se_wwn *wwn)
3837 {
3838 	struct srpt_port_id *port_id = container_of(wwn, typeof(*port_id), wwn);
3839 	struct srpt_port *sport = wwn->priv;
3840 
3841 	if (sport->guid_id == port_id)
3842 		sport->guid_id = NULL;
3843 	else if (sport->gid_id == port_id)
3844 		sport->gid_id = NULL;
3845 	else
3846 		WARN_ON_ONCE(true);
3847 
3848 	srpt_sdev_put(sport->sdev);
3849 	kfree(port_id);
3850 }
3851 
3852 static ssize_t srpt_wwn_version_show(struct config_item *item, char *buf)
3853 {
3854 	return sysfs_emit(buf, "\n");
3855 }
3856 
3857 CONFIGFS_ATTR_RO(srpt_wwn_, version);
3858 
3859 static struct configfs_attribute *srpt_wwn_attrs[] = {
3860 	&srpt_wwn_attr_version,
3861 	NULL,
3862 };
3863 
3864 static const struct target_core_fabric_ops srpt_template = {
3865 	.module				= THIS_MODULE,
3866 	.fabric_name			= "srpt",
3867 	.tpg_get_wwn			= srpt_get_fabric_wwn,
3868 	.tpg_get_tag			= srpt_get_tag,
3869 	.tpg_check_demo_mode		= srpt_check_false,
3870 	.tpg_check_demo_mode_cache	= srpt_check_true,
3871 	.tpg_check_demo_mode_write_protect = srpt_check_true,
3872 	.tpg_check_prod_mode_write_protect = srpt_check_false,
3873 	.tpg_get_inst_index		= srpt_tpg_get_inst_index,
3874 	.release_cmd			= srpt_release_cmd,
3875 	.check_stop_free		= srpt_check_stop_free,
3876 	.close_session			= srpt_close_session,
3877 	.sess_get_index			= srpt_sess_get_index,
3878 	.sess_get_initiator_sid		= NULL,
3879 	.write_pending			= srpt_write_pending,
3880 	.set_default_node_attributes	= srpt_set_default_node_attrs,
3881 	.get_cmd_state			= srpt_get_tcm_cmd_state,
3882 	.queue_data_in			= srpt_queue_data_in,
3883 	.queue_status			= srpt_queue_status,
3884 	.queue_tm_rsp			= srpt_queue_tm_rsp,
3885 	.aborted_task			= srpt_aborted_task,
3886 	/*
3887 	 * Setup function pointers for generic logic in
3888 	 * target_core_fabric_configfs.c
3889 	 */
3890 	.fabric_make_wwn		= srpt_make_tport,
3891 	.fabric_drop_wwn		= srpt_drop_tport,
3892 	.fabric_make_tpg		= srpt_make_tpg,
3893 	.fabric_enable_tpg		= srpt_enable_tpg,
3894 	.fabric_drop_tpg		= srpt_drop_tpg,
3895 	.fabric_init_nodeacl		= srpt_init_nodeacl,
3896 
3897 	.tfc_discovery_attrs		= srpt_da_attrs,
3898 	.tfc_wwn_attrs			= srpt_wwn_attrs,
3899 	.tfc_tpg_attrib_attrs		= srpt_tpg_attrib_attrs,
3900 };
3901 
3902 /**
3903  * srpt_init_module - kernel module initialization
3904  *
3905  * Note: Since ib_register_client() registers callback functions, and since at
3906  * least one of these callback functions (srpt_add_one()) calls target core
3907  * functions, this driver must be registered with the target core before
3908  * ib_register_client() is called.
3909  */
3910 static int __init srpt_init_module(void)
3911 {
3912 	int ret;
3913 
3914 	ret = -EINVAL;
3915 	if (srp_max_req_size < MIN_MAX_REQ_SIZE) {
3916 		pr_err("invalid value %d for kernel module parameter srp_max_req_size -- must be at least %d.\n",
3917 		       srp_max_req_size, MIN_MAX_REQ_SIZE);
3918 		goto out;
3919 	}
3920 
3921 	if (srpt_srq_size < MIN_SRPT_SRQ_SIZE
3922 	    || srpt_srq_size > MAX_SRPT_SRQ_SIZE) {
3923 		pr_err("invalid value %d for kernel module parameter srpt_srq_size -- must be in the range [%d..%d].\n",
3924 		       srpt_srq_size, MIN_SRPT_SRQ_SIZE, MAX_SRPT_SRQ_SIZE);
3925 		goto out;
3926 	}
3927 
3928 	ret = target_register_template(&srpt_template);
3929 	if (ret)
3930 		goto out;
3931 
3932 	ret = ib_register_client(&srpt_client);
3933 	if (ret) {
3934 		pr_err("couldn't register IB client\n");
3935 		goto out_unregister_target;
3936 	}
3937 
3938 	return 0;
3939 
3940 out_unregister_target:
3941 	target_unregister_template(&srpt_template);
3942 out:
3943 	return ret;
3944 }
3945 
3946 static void __exit srpt_cleanup_module(void)
3947 {
3948 	if (rdma_cm_id)
3949 		rdma_destroy_id(rdma_cm_id);
3950 	ib_unregister_client(&srpt_client);
3951 	target_unregister_template(&srpt_template);
3952 }
3953 
3954 module_init(srpt_init_module);
3955 module_exit(srpt_cleanup_module);
3956