xref: /linux/drivers/infiniband/ulp/srpt/ib_srpt.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
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 static DEFINE_MUTEX(srpt_mc_mutex);	/* Protects srpt_memory_caches. */
72 static DEFINE_XARRAY(srpt_memory_caches); /* See also srpt_memory_cache_entry */
73 
74 static unsigned srp_max_req_size = DEFAULT_MAX_REQ_SIZE;
75 module_param(srp_max_req_size, int, 0444);
76 MODULE_PARM_DESC(srp_max_req_size,
77 		 "Maximum size of SRP request messages in bytes.");
78 
79 static int srpt_srq_size = DEFAULT_SRPT_SRQ_SIZE;
80 module_param(srpt_srq_size, int, 0444);
81 MODULE_PARM_DESC(srpt_srq_size,
82 		 "Shared receive queue (SRQ) size.");
83 
84 static int srpt_set_u64_x(const char *buffer, const struct kernel_param *kp)
85 {
86 	return kstrtou64(buffer, 16, (u64 *)kp->arg);
87 }
88 static int srpt_get_u64_x(char *buffer, const struct kernel_param *kp)
89 {
90 	return sprintf(buffer, "0x%016llx\n", *(u64 *)kp->arg);
91 }
92 module_param_call(srpt_service_guid, srpt_set_u64_x, srpt_get_u64_x,
93 		  &srpt_service_guid, 0444);
94 MODULE_PARM_DESC(srpt_service_guid,
95 		 "Using this value for ioc_guid, id_ext, and cm_listen_id instead of using the node_guid of the first HCA.");
96 
97 static struct ib_client srpt_client;
98 /* Protects both rdma_cm_port and rdma_cm_id. */
99 static DEFINE_MUTEX(rdma_cm_mutex);
100 /* Port number RDMA/CM will bind to. */
101 static u16 rdma_cm_port;
102 static struct rdma_cm_id *rdma_cm_id;
103 static void srpt_release_cmd(struct se_cmd *se_cmd);
104 static void srpt_free_ch(struct kref *kref);
105 static int srpt_queue_status(struct se_cmd *cmd);
106 static void srpt_recv_done(struct ib_cq *cq, struct ib_wc *wc);
107 static void srpt_send_done(struct ib_cq *cq, struct ib_wc *wc);
108 static void srpt_process_wait_list(struct srpt_rdma_ch *ch);
109 
110 /* Type of the entries in srpt_memory_caches. */
111 struct srpt_memory_cache_entry {
112 	refcount_t ref;
113 	struct kmem_cache *c;
114 };
115 
116 static struct kmem_cache *srpt_cache_get(unsigned int object_size)
117 {
118 	struct srpt_memory_cache_entry *e;
119 	char name[32];
120 	void *res;
121 
122 	guard(mutex)(&srpt_mc_mutex);
123 	e = xa_load(&srpt_memory_caches, object_size);
124 	if (e) {
125 		refcount_inc(&e->ref);
126 		return e->c;
127 	}
128 	snprintf(name, sizeof(name), "srpt-%u", object_size);
129 	e = kmalloc(sizeof(*e), GFP_KERNEL);
130 	if (!e)
131 		return NULL;
132 	refcount_set(&e->ref, 1);
133 	e->c = kmem_cache_create(name, object_size, /*align=*/512, 0, NULL);
134 	if (!e->c)
135 		goto free_entry;
136 	res = xa_store(&srpt_memory_caches, object_size, e, GFP_KERNEL);
137 	if (xa_is_err(res))
138 		goto destroy_cache;
139 	return e->c;
140 
141 destroy_cache:
142 	kmem_cache_destroy(e->c);
143 
144 free_entry:
145 	kfree(e);
146 	return NULL;
147 }
148 
149 static void srpt_cache_put(struct kmem_cache *c)
150 {
151 	struct srpt_memory_cache_entry *e = NULL;
152 	unsigned long object_size;
153 
154 	guard(mutex)(&srpt_mc_mutex);
155 	xa_for_each(&srpt_memory_caches, object_size, e)
156 		if (e->c == c)
157 			break;
158 	if (WARN_ON_ONCE(!e))
159 		return;
160 	if (!refcount_dec_and_test(&e->ref))
161 		return;
162 	WARN_ON_ONCE(xa_erase(&srpt_memory_caches, object_size) != e);
163 	kmem_cache_destroy(e->c);
164 	kfree(e);
165 }
166 
167 /*
168  * The only allowed channel state changes are those that change the channel
169  * state into a state with a higher numerical value. Hence the new > prev test.
170  */
171 static bool srpt_set_ch_state(struct srpt_rdma_ch *ch, enum rdma_ch_state new)
172 {
173 	unsigned long flags;
174 	enum rdma_ch_state prev;
175 	bool changed = false;
176 
177 	spin_lock_irqsave(&ch->spinlock, flags);
178 	prev = ch->state;
179 	if (new > prev) {
180 		ch->state = new;
181 		changed = true;
182 	}
183 	spin_unlock_irqrestore(&ch->spinlock, flags);
184 
185 	return changed;
186 }
187 
188 /**
189  * srpt_event_handler - asynchronous IB event callback function
190  * @handler: IB event handler registered by ib_register_event_handler().
191  * @event: Description of the event that occurred.
192  *
193  * Callback function called by the InfiniBand core when an asynchronous IB
194  * event occurs. This callback may occur in interrupt context. See also
195  * section 11.5.2, Set Asynchronous Event Handler in the InfiniBand
196  * Architecture Specification.
197  */
198 static void srpt_event_handler(struct ib_event_handler *handler,
199 			       struct ib_event *event)
200 {
201 	struct srpt_device *sdev =
202 		container_of(handler, struct srpt_device, event_handler);
203 	struct srpt_port *sport;
204 	u8 port_num;
205 
206 	pr_debug("ASYNC event= %d on device= %s\n", event->event,
207 		 dev_name(&sdev->device->dev));
208 
209 	switch (event->event) {
210 	case IB_EVENT_PORT_ERR:
211 		port_num = event->element.port_num - 1;
212 		if (port_num < sdev->device->phys_port_cnt) {
213 			sport = &sdev->port[port_num];
214 			sport->lid = 0;
215 			sport->sm_lid = 0;
216 		} else {
217 			WARN(true, "event %d: port_num %d out of range 1..%d\n",
218 			     event->event, port_num + 1,
219 			     sdev->device->phys_port_cnt);
220 		}
221 		break;
222 	case IB_EVENT_PORT_ACTIVE:
223 	case IB_EVENT_LID_CHANGE:
224 	case IB_EVENT_PKEY_CHANGE:
225 	case IB_EVENT_SM_CHANGE:
226 	case IB_EVENT_CLIENT_REREGISTER:
227 	case IB_EVENT_GID_CHANGE:
228 		/* Refresh port data asynchronously. */
229 		port_num = event->element.port_num - 1;
230 		if (port_num < sdev->device->phys_port_cnt) {
231 			sport = &sdev->port[port_num];
232 			if (!sport->lid && !sport->sm_lid)
233 				schedule_work(&sport->work);
234 		} else {
235 			WARN(true, "event %d: port_num %d out of range 1..%d\n",
236 			     event->event, port_num + 1,
237 			     sdev->device->phys_port_cnt);
238 		}
239 		break;
240 	default:
241 		pr_err("received unrecognized IB event %d\n", event->event);
242 		break;
243 	}
244 }
245 
246 /**
247  * srpt_srq_event - SRQ event callback function
248  * @event: Description of the event that occurred.
249  * @ctx: Context pointer specified at SRQ creation time.
250  */
251 static void srpt_srq_event(struct ib_event *event, void *ctx)
252 {
253 	pr_debug("SRQ event %d\n", event->event);
254 }
255 
256 static const char *get_ch_state_name(enum rdma_ch_state s)
257 {
258 	switch (s) {
259 	case CH_CONNECTING:
260 		return "connecting";
261 	case CH_LIVE:
262 		return "live";
263 	case CH_DISCONNECTING:
264 		return "disconnecting";
265 	case CH_DRAINING:
266 		return "draining";
267 	case CH_DISCONNECTED:
268 		return "disconnected";
269 	}
270 	return "???";
271 }
272 
273 /**
274  * srpt_qp_event - QP event callback function
275  * @event: Description of the event that occurred.
276  * @ptr: SRPT RDMA channel.
277  */
278 static void srpt_qp_event(struct ib_event *event, void *ptr)
279 {
280 	struct srpt_rdma_ch *ch = ptr;
281 
282 	pr_debug("QP event %d on ch=%p sess_name=%s-%d state=%s\n",
283 		 event->event, ch, ch->sess_name, ch->qp->qp_num,
284 		 get_ch_state_name(ch->state));
285 
286 	switch (event->event) {
287 	case IB_EVENT_COMM_EST:
288 		if (ch->using_rdma_cm)
289 			rdma_notify(ch->rdma_cm.cm_id, event->event);
290 		else
291 			ib_cm_notify(ch->ib_cm.cm_id, event->event);
292 		break;
293 	case IB_EVENT_QP_LAST_WQE_REACHED:
294 		pr_debug("%s-%d, state %s: received Last WQE event.\n",
295 			 ch->sess_name, ch->qp->qp_num,
296 			 get_ch_state_name(ch->state));
297 		break;
298 	default:
299 		pr_err("received unrecognized IB QP event %d\n", event->event);
300 		break;
301 	}
302 }
303 
304 /**
305  * srpt_set_ioc - initialize a IOUnitInfo structure
306  * @c_list: controller list.
307  * @slot: one-based slot number.
308  * @value: four-bit value.
309  *
310  * Copies the lowest four bits of value in element slot of the array of four
311  * bit elements called c_list (controller list). The index slot is one-based.
312  */
313 static void srpt_set_ioc(u8 *c_list, u32 slot, u8 value)
314 {
315 	u16 id;
316 	u8 tmp;
317 
318 	id = (slot - 1) / 2;
319 	if (slot & 0x1) {
320 		tmp = c_list[id] & 0xf;
321 		c_list[id] = (value << 4) | tmp;
322 	} else {
323 		tmp = c_list[id] & 0xf0;
324 		c_list[id] = (value & 0xf) | tmp;
325 	}
326 }
327 
328 /**
329  * srpt_get_class_port_info - copy ClassPortInfo to a management datagram
330  * @mad: Datagram that will be sent as response to DM_ATTR_CLASS_PORT_INFO.
331  *
332  * See also section 16.3.3.1 ClassPortInfo in the InfiniBand Architecture
333  * Specification.
334  */
335 static void srpt_get_class_port_info(struct ib_dm_mad *mad)
336 {
337 	struct ib_class_port_info *cif;
338 
339 	cif = (struct ib_class_port_info *)mad->data;
340 	memset(cif, 0, sizeof(*cif));
341 	cif->base_version = 1;
342 	cif->class_version = 1;
343 
344 	ib_set_cpi_resp_time(cif, 20);
345 	mad->mad_hdr.status = 0;
346 }
347 
348 /**
349  * srpt_get_iou - write IOUnitInfo to a management datagram
350  * @mad: Datagram that will be sent as response to DM_ATTR_IOU_INFO.
351  *
352  * See also section 16.3.3.3 IOUnitInfo in the InfiniBand Architecture
353  * Specification. See also section B.7, table B.6 in the SRP r16a document.
354  */
355 static void srpt_get_iou(struct ib_dm_mad *mad)
356 {
357 	struct ib_dm_iou_info *ioui;
358 	u8 slot;
359 	int i;
360 
361 	ioui = (struct ib_dm_iou_info *)mad->data;
362 	ioui->change_id = cpu_to_be16(1);
363 	ioui->max_controllers = 16;
364 
365 	/* set present for slot 1 and empty for the rest */
366 	srpt_set_ioc(ioui->controller_list, 1, 1);
367 	for (i = 1, slot = 2; i < 16; i++, slot++)
368 		srpt_set_ioc(ioui->controller_list, slot, 0);
369 
370 	mad->mad_hdr.status = 0;
371 }
372 
373 /**
374  * srpt_get_ioc - write IOControllerprofile to a management datagram
375  * @sport: HCA port through which the MAD has been received.
376  * @slot: Slot number specified in DM_ATTR_IOC_PROFILE query.
377  * @mad: Datagram that will be sent as response to DM_ATTR_IOC_PROFILE.
378  *
379  * See also section 16.3.3.4 IOControllerProfile in the InfiniBand
380  * Architecture Specification. See also section B.7, table B.7 in the SRP
381  * r16a document.
382  */
383 static void srpt_get_ioc(struct srpt_port *sport, u32 slot,
384 			 struct ib_dm_mad *mad)
385 {
386 	struct srpt_device *sdev = sport->sdev;
387 	struct ib_dm_ioc_profile *iocp;
388 	int send_queue_depth;
389 
390 	iocp = (struct ib_dm_ioc_profile *)mad->data;
391 
392 	if (!slot || slot > 16) {
393 		mad->mad_hdr.status
394 			= cpu_to_be16(DM_MAD_STATUS_INVALID_FIELD);
395 		return;
396 	}
397 
398 	if (slot > 2) {
399 		mad->mad_hdr.status
400 			= cpu_to_be16(DM_MAD_STATUS_NO_IOC);
401 		return;
402 	}
403 
404 	if (sdev->use_srq)
405 		send_queue_depth = sdev->srq_size;
406 	else
407 		send_queue_depth = min(MAX_SRPT_RQ_SIZE,
408 				       sdev->device->attrs.max_qp_wr);
409 
410 	memset(iocp, 0, sizeof(*iocp));
411 	strcpy(iocp->id_string, SRPT_ID_STRING);
412 	iocp->guid = cpu_to_be64(srpt_service_guid);
413 	iocp->vendor_id = cpu_to_be32(sdev->device->attrs.vendor_id);
414 	iocp->device_id = cpu_to_be32(sdev->device->attrs.vendor_part_id);
415 	iocp->device_version = cpu_to_be16(sdev->device->attrs.hw_ver);
416 	iocp->subsys_vendor_id = cpu_to_be32(sdev->device->attrs.vendor_id);
417 	iocp->subsys_device_id = 0x0;
418 	iocp->io_class = cpu_to_be16(SRP_REV16A_IB_IO_CLASS);
419 	iocp->io_subclass = cpu_to_be16(SRP_IO_SUBCLASS);
420 	iocp->protocol = cpu_to_be16(SRP_PROTOCOL);
421 	iocp->protocol_version = cpu_to_be16(SRP_PROTOCOL_VERSION);
422 	iocp->send_queue_depth = cpu_to_be16(send_queue_depth);
423 	iocp->rdma_read_depth = 4;
424 	iocp->send_size = cpu_to_be32(srp_max_req_size);
425 	iocp->rdma_size = cpu_to_be32(min(sport->port_attrib.srp_max_rdma_size,
426 					  1U << 24));
427 	iocp->num_svc_entries = 1;
428 	iocp->op_cap_mask = SRP_SEND_TO_IOC | SRP_SEND_FROM_IOC |
429 		SRP_RDMA_READ_FROM_IOC | SRP_RDMA_WRITE_FROM_IOC;
430 
431 	mad->mad_hdr.status = 0;
432 }
433 
434 /**
435  * srpt_get_svc_entries - write ServiceEntries to a management datagram
436  * @ioc_guid: I/O controller GUID to use in reply.
437  * @slot: I/O controller number.
438  * @hi: End of the range of service entries to be specified in the reply.
439  * @lo: Start of the range of service entries to be specified in the reply..
440  * @mad: Datagram that will be sent as response to DM_ATTR_SVC_ENTRIES.
441  *
442  * See also section 16.3.3.5 ServiceEntries in the InfiniBand Architecture
443  * Specification. See also section B.7, table B.8 in the SRP r16a document.
444  */
445 static void srpt_get_svc_entries(u64 ioc_guid,
446 				 u16 slot, u8 hi, u8 lo, struct ib_dm_mad *mad)
447 {
448 	struct ib_dm_svc_entries *svc_entries;
449 
450 	WARN_ON(!ioc_guid);
451 
452 	if (!slot || slot > 16) {
453 		mad->mad_hdr.status
454 			= cpu_to_be16(DM_MAD_STATUS_INVALID_FIELD);
455 		return;
456 	}
457 
458 	if (slot > 2 || lo > hi || hi > 1) {
459 		mad->mad_hdr.status
460 			= cpu_to_be16(DM_MAD_STATUS_NO_IOC);
461 		return;
462 	}
463 
464 	svc_entries = (struct ib_dm_svc_entries *)mad->data;
465 	memset(svc_entries, 0, sizeof(*svc_entries));
466 	svc_entries->service_entries[0].id = cpu_to_be64(ioc_guid);
467 	snprintf(svc_entries->service_entries[0].name,
468 		 sizeof(svc_entries->service_entries[0].name),
469 		 "%s%016llx",
470 		 SRP_SERVICE_NAME_PREFIX,
471 		 ioc_guid);
472 
473 	mad->mad_hdr.status = 0;
474 }
475 
476 /**
477  * srpt_mgmt_method_get - process a received management datagram
478  * @sp:      HCA port through which the MAD has been received.
479  * @rq_mad:  received MAD.
480  * @rsp_mad: response MAD.
481  */
482 static void srpt_mgmt_method_get(struct srpt_port *sp, struct ib_mad *rq_mad,
483 				 struct ib_dm_mad *rsp_mad)
484 {
485 	u16 attr_id;
486 	u32 slot;
487 	u8 hi, lo;
488 
489 	attr_id = be16_to_cpu(rq_mad->mad_hdr.attr_id);
490 	switch (attr_id) {
491 	case DM_ATTR_CLASS_PORT_INFO:
492 		srpt_get_class_port_info(rsp_mad);
493 		break;
494 	case DM_ATTR_IOU_INFO:
495 		srpt_get_iou(rsp_mad);
496 		break;
497 	case DM_ATTR_IOC_PROFILE:
498 		slot = be32_to_cpu(rq_mad->mad_hdr.attr_mod);
499 		srpt_get_ioc(sp, slot, rsp_mad);
500 		break;
501 	case DM_ATTR_SVC_ENTRIES:
502 		slot = be32_to_cpu(rq_mad->mad_hdr.attr_mod);
503 		hi = (u8) ((slot >> 8) & 0xff);
504 		lo = (u8) (slot & 0xff);
505 		slot = (u16) ((slot >> 16) & 0xffff);
506 		srpt_get_svc_entries(srpt_service_guid,
507 				     slot, hi, lo, rsp_mad);
508 		break;
509 	default:
510 		rsp_mad->mad_hdr.status =
511 		    cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD_ATTR);
512 		break;
513 	}
514 }
515 
516 /**
517  * srpt_mad_send_handler - MAD send completion callback
518  * @mad_agent: Return value of ib_register_mad_agent().
519  * @mad_wc: Work completion reporting that the MAD has been sent.
520  */
521 static void srpt_mad_send_handler(struct ib_mad_agent *mad_agent,
522 				  struct ib_mad_send_wc *mad_wc)
523 {
524 	rdma_destroy_ah(mad_wc->send_buf->ah, RDMA_DESTROY_AH_SLEEPABLE);
525 	ib_free_send_mad(mad_wc->send_buf);
526 }
527 
528 /**
529  * srpt_mad_recv_handler - MAD reception callback function
530  * @mad_agent: Return value of ib_register_mad_agent().
531  * @send_buf: Not used.
532  * @mad_wc: Work completion reporting that a MAD has been received.
533  */
534 static void srpt_mad_recv_handler(struct ib_mad_agent *mad_agent,
535 				  struct ib_mad_send_buf *send_buf,
536 				  struct ib_mad_recv_wc *mad_wc)
537 {
538 	struct srpt_port *sport = (struct srpt_port *)mad_agent->context;
539 	struct ib_ah *ah;
540 	struct ib_mad_send_buf *rsp;
541 	struct ib_dm_mad *dm_mad;
542 
543 	if (!mad_wc || !mad_wc->recv_buf.mad)
544 		return;
545 
546 	ah = ib_create_ah_from_wc(mad_agent->qp->pd, mad_wc->wc,
547 				  mad_wc->recv_buf.grh, mad_agent->port_num);
548 	if (IS_ERR(ah))
549 		goto err;
550 
551 	BUILD_BUG_ON(offsetof(struct ib_dm_mad, data) != IB_MGMT_DEVICE_HDR);
552 
553 	rsp = ib_create_send_mad(mad_agent, mad_wc->wc->src_qp,
554 				 mad_wc->wc->pkey_index, 0,
555 				 IB_MGMT_DEVICE_HDR, IB_MGMT_DEVICE_DATA,
556 				 GFP_KERNEL,
557 				 IB_MGMT_BASE_VERSION);
558 	if (IS_ERR(rsp))
559 		goto err_rsp;
560 
561 	rsp->ah = ah;
562 
563 	dm_mad = rsp->mad;
564 	memcpy(dm_mad, mad_wc->recv_buf.mad, sizeof(*dm_mad));
565 	dm_mad->mad_hdr.method = IB_MGMT_METHOD_GET_RESP;
566 	dm_mad->mad_hdr.status = 0;
567 
568 	switch (mad_wc->recv_buf.mad->mad_hdr.method) {
569 	case IB_MGMT_METHOD_GET:
570 		srpt_mgmt_method_get(sport, mad_wc->recv_buf.mad, dm_mad);
571 		break;
572 	case IB_MGMT_METHOD_SET:
573 		dm_mad->mad_hdr.status =
574 		    cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD_ATTR);
575 		break;
576 	default:
577 		dm_mad->mad_hdr.status =
578 		    cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD);
579 		break;
580 	}
581 
582 	if (!ib_post_send_mad(rsp, NULL)) {
583 		ib_free_recv_mad(mad_wc);
584 		/* will destroy_ah & free_send_mad in send completion */
585 		return;
586 	}
587 
588 	ib_free_send_mad(rsp);
589 
590 err_rsp:
591 	rdma_destroy_ah(ah, RDMA_DESTROY_AH_SLEEPABLE);
592 err:
593 	ib_free_recv_mad(mad_wc);
594 }
595 
596 static int srpt_format_guid(char *buf, unsigned int size, const __be64 *guid)
597 {
598 	const __be16 *g = (const __be16 *)guid;
599 
600 	return snprintf(buf, size, "%04x:%04x:%04x:%04x",
601 			be16_to_cpu(g[0]), be16_to_cpu(g[1]),
602 			be16_to_cpu(g[2]), be16_to_cpu(g[3]));
603 }
604 
605 /**
606  * srpt_refresh_port - configure a HCA port
607  * @sport: SRPT HCA port.
608  *
609  * Enable InfiniBand management datagram processing, update the cached sm_lid,
610  * lid and gid values, and register a callback function for processing MADs
611  * on the specified port.
612  *
613  * Note: It is safe to call this function more than once for the same port.
614  */
615 static int srpt_refresh_port(struct srpt_port *sport)
616 {
617 	struct ib_mad_agent *mad_agent;
618 	struct ib_mad_reg_req reg_req;
619 	struct ib_port_modify port_modify;
620 	struct ib_port_attr port_attr;
621 	int ret;
622 
623 	ret = ib_query_port(sport->sdev->device, sport->port, &port_attr);
624 	if (ret)
625 		return ret;
626 
627 	sport->sm_lid = port_attr.sm_lid;
628 	sport->lid = port_attr.lid;
629 
630 	ret = rdma_query_gid(sport->sdev->device, sport->port, 0, &sport->gid);
631 	if (ret)
632 		return ret;
633 
634 	srpt_format_guid(sport->guid_name, ARRAY_SIZE(sport->guid_name),
635 			 &sport->gid.global.interface_id);
636 	snprintf(sport->gid_name, ARRAY_SIZE(sport->gid_name),
637 		 "0x%016llx%016llx",
638 		 be64_to_cpu(sport->gid.global.subnet_prefix),
639 		 be64_to_cpu(sport->gid.global.interface_id));
640 
641 	if (rdma_protocol_iwarp(sport->sdev->device, sport->port))
642 		return 0;
643 
644 	memset(&port_modify, 0, sizeof(port_modify));
645 	port_modify.set_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP;
646 	port_modify.clr_port_cap_mask = 0;
647 
648 	ret = ib_modify_port(sport->sdev->device, sport->port, 0, &port_modify);
649 	if (ret) {
650 		pr_warn("%s-%d: enabling device management failed (%d). Note: this is expected if SR-IOV is enabled.\n",
651 			dev_name(&sport->sdev->device->dev), sport->port, ret);
652 		return 0;
653 	}
654 
655 	if (!sport->mad_agent) {
656 		memset(&reg_req, 0, sizeof(reg_req));
657 		reg_req.mgmt_class = IB_MGMT_CLASS_DEVICE_MGMT;
658 		reg_req.mgmt_class_version = IB_MGMT_BASE_VERSION;
659 		set_bit(IB_MGMT_METHOD_GET, reg_req.method_mask);
660 		set_bit(IB_MGMT_METHOD_SET, reg_req.method_mask);
661 
662 		mad_agent = ib_register_mad_agent(sport->sdev->device,
663 						  sport->port,
664 						  IB_QPT_GSI,
665 						  &reg_req, 0,
666 						  srpt_mad_send_handler,
667 						  srpt_mad_recv_handler,
668 						  sport, 0);
669 		if (IS_ERR(mad_agent)) {
670 			pr_err("%s-%d: MAD agent registration failed (%ld). Note: this is expected if SR-IOV is enabled.\n",
671 			       dev_name(&sport->sdev->device->dev), sport->port,
672 			       PTR_ERR(mad_agent));
673 			sport->mad_agent = NULL;
674 			memset(&port_modify, 0, sizeof(port_modify));
675 			port_modify.clr_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP;
676 			ib_modify_port(sport->sdev->device, sport->port, 0,
677 				       &port_modify);
678 			return 0;
679 		}
680 
681 		sport->mad_agent = mad_agent;
682 	}
683 
684 	return 0;
685 }
686 
687 /**
688  * srpt_unregister_mad_agent - unregister MAD callback functions
689  * @sdev: SRPT HCA pointer.
690  * @port_cnt: number of ports with registered MAD
691  *
692  * Note: It is safe to call this function more than once for the same device.
693  */
694 static void srpt_unregister_mad_agent(struct srpt_device *sdev, int port_cnt)
695 {
696 	struct ib_port_modify port_modify = {
697 		.clr_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP,
698 	};
699 	struct srpt_port *sport;
700 	int i;
701 
702 	for (i = 1; i <= port_cnt; i++) {
703 		sport = &sdev->port[i - 1];
704 		WARN_ON(sport->port != i);
705 		if (sport->mad_agent) {
706 			ib_modify_port(sdev->device, i, 0, &port_modify);
707 			ib_unregister_mad_agent(sport->mad_agent);
708 			sport->mad_agent = NULL;
709 		}
710 	}
711 }
712 
713 /**
714  * srpt_alloc_ioctx - allocate a SRPT I/O context structure
715  * @sdev: SRPT HCA pointer.
716  * @ioctx_size: I/O context size.
717  * @buf_cache: I/O buffer cache.
718  * @dir: DMA data direction.
719  */
720 static struct srpt_ioctx *srpt_alloc_ioctx(struct srpt_device *sdev,
721 					   int ioctx_size,
722 					   struct kmem_cache *buf_cache,
723 					   enum dma_data_direction dir)
724 {
725 	struct srpt_ioctx *ioctx;
726 
727 	ioctx = kzalloc(ioctx_size, GFP_KERNEL);
728 	if (!ioctx)
729 		goto err;
730 
731 	ioctx->buf = kmem_cache_alloc(buf_cache, GFP_KERNEL);
732 	if (!ioctx->buf)
733 		goto err_free_ioctx;
734 
735 	ioctx->dma = ib_dma_map_single(sdev->device, ioctx->buf,
736 				       kmem_cache_size(buf_cache), dir);
737 	if (ib_dma_mapping_error(sdev->device, ioctx->dma))
738 		goto err_free_buf;
739 
740 	return ioctx;
741 
742 err_free_buf:
743 	kmem_cache_free(buf_cache, ioctx->buf);
744 err_free_ioctx:
745 	kfree(ioctx);
746 err:
747 	return NULL;
748 }
749 
750 /**
751  * srpt_free_ioctx - free a SRPT I/O context structure
752  * @sdev: SRPT HCA pointer.
753  * @ioctx: I/O context pointer.
754  * @buf_cache: I/O buffer cache.
755  * @dir: DMA data direction.
756  */
757 static void srpt_free_ioctx(struct srpt_device *sdev, struct srpt_ioctx *ioctx,
758 			    struct kmem_cache *buf_cache,
759 			    enum dma_data_direction dir)
760 {
761 	if (!ioctx)
762 		return;
763 
764 	ib_dma_unmap_single(sdev->device, ioctx->dma,
765 			    kmem_cache_size(buf_cache), dir);
766 	kmem_cache_free(buf_cache, ioctx->buf);
767 	kfree(ioctx);
768 }
769 
770 /**
771  * srpt_alloc_ioctx_ring - allocate a ring of SRPT I/O context structures
772  * @sdev:       Device to allocate the I/O context ring for.
773  * @ring_size:  Number of elements in the I/O context ring.
774  * @ioctx_size: I/O context size.
775  * @buf_cache:  I/O buffer cache.
776  * @alignment_offset: Offset in each ring buffer at which the SRP information
777  *		unit starts.
778  * @dir:        DMA data direction.
779  */
780 static struct srpt_ioctx **srpt_alloc_ioctx_ring(struct srpt_device *sdev,
781 				int ring_size, int ioctx_size,
782 				struct kmem_cache *buf_cache,
783 				int alignment_offset,
784 				enum dma_data_direction dir)
785 {
786 	struct srpt_ioctx **ring;
787 	int i;
788 
789 	WARN_ON(ioctx_size != sizeof(struct srpt_recv_ioctx) &&
790 		ioctx_size != sizeof(struct srpt_send_ioctx));
791 
792 	ring = kvmalloc_array(ring_size, sizeof(ring[0]), GFP_KERNEL);
793 	if (!ring)
794 		goto out;
795 	for (i = 0; i < ring_size; ++i) {
796 		ring[i] = srpt_alloc_ioctx(sdev, ioctx_size, buf_cache, dir);
797 		if (!ring[i])
798 			goto err;
799 		ring[i]->index = i;
800 		ring[i]->offset = alignment_offset;
801 	}
802 	goto out;
803 
804 err:
805 	while (--i >= 0)
806 		srpt_free_ioctx(sdev, ring[i], buf_cache, dir);
807 	kvfree(ring);
808 	ring = NULL;
809 out:
810 	return ring;
811 }
812 
813 /**
814  * srpt_free_ioctx_ring - free the ring of SRPT I/O context structures
815  * @ioctx_ring: I/O context ring to be freed.
816  * @sdev: SRPT HCA pointer.
817  * @ring_size: Number of ring elements.
818  * @buf_cache: I/O buffer cache.
819  * @dir: DMA data direction.
820  */
821 static void srpt_free_ioctx_ring(struct srpt_ioctx **ioctx_ring,
822 				 struct srpt_device *sdev, int ring_size,
823 				 struct kmem_cache *buf_cache,
824 				 enum dma_data_direction dir)
825 {
826 	int i;
827 
828 	if (!ioctx_ring)
829 		return;
830 
831 	for (i = 0; i < ring_size; ++i)
832 		srpt_free_ioctx(sdev, ioctx_ring[i], buf_cache, dir);
833 	kvfree(ioctx_ring);
834 }
835 
836 /**
837  * srpt_set_cmd_state - set the state of a SCSI command
838  * @ioctx: Send I/O context.
839  * @new: New I/O context state.
840  *
841  * Does not modify the state of aborted commands. Returns the previous command
842  * state.
843  */
844 static enum srpt_command_state srpt_set_cmd_state(struct srpt_send_ioctx *ioctx,
845 						  enum srpt_command_state new)
846 {
847 	enum srpt_command_state previous;
848 
849 	previous = ioctx->state;
850 	if (previous != SRPT_STATE_DONE)
851 		ioctx->state = new;
852 
853 	return previous;
854 }
855 
856 /**
857  * srpt_test_and_set_cmd_state - test and set the state of a command
858  * @ioctx: Send I/O context.
859  * @old: Current I/O context state.
860  * @new: New I/O context state.
861  *
862  * Returns true if and only if the previous command state was equal to 'old'.
863  */
864 static bool srpt_test_and_set_cmd_state(struct srpt_send_ioctx *ioctx,
865 					enum srpt_command_state old,
866 					enum srpt_command_state new)
867 {
868 	enum srpt_command_state previous;
869 
870 	WARN_ON(!ioctx);
871 	WARN_ON(old == SRPT_STATE_DONE);
872 	WARN_ON(new == SRPT_STATE_NEW);
873 
874 	previous = ioctx->state;
875 	if (previous == old)
876 		ioctx->state = new;
877 
878 	return previous == old;
879 }
880 
881 /**
882  * srpt_post_recv - post an IB receive request
883  * @sdev: SRPT HCA pointer.
884  * @ch: SRPT RDMA channel.
885  * @ioctx: Receive I/O context pointer.
886  */
887 static int srpt_post_recv(struct srpt_device *sdev, struct srpt_rdma_ch *ch,
888 			  struct srpt_recv_ioctx *ioctx)
889 {
890 	struct ib_sge list;
891 	struct ib_recv_wr wr;
892 
893 	BUG_ON(!sdev);
894 	list.addr = ioctx->ioctx.dma + ioctx->ioctx.offset;
895 	list.length = srp_max_req_size;
896 	list.lkey = sdev->lkey;
897 
898 	ioctx->ioctx.cqe.done = srpt_recv_done;
899 	wr.wr_cqe = &ioctx->ioctx.cqe;
900 	wr.next = NULL;
901 	wr.sg_list = &list;
902 	wr.num_sge = 1;
903 
904 	if (sdev->use_srq)
905 		return ib_post_srq_recv(sdev->srq, &wr, NULL);
906 	else
907 		return ib_post_recv(ch->qp, &wr, NULL);
908 }
909 
910 /**
911  * srpt_zerolength_write - perform a zero-length RDMA write
912  * @ch: SRPT RDMA channel.
913  *
914  * A quote from the InfiniBand specification: C9-88: For an HCA responder
915  * using Reliable Connection service, for each zero-length RDMA READ or WRITE
916  * request, the R_Key shall not be validated, even if the request includes
917  * Immediate data.
918  */
919 static int srpt_zerolength_write(struct srpt_rdma_ch *ch)
920 {
921 	struct ib_rdma_wr wr = {
922 		.wr = {
923 			.next		= NULL,
924 			{ .wr_cqe	= &ch->zw_cqe, },
925 			.opcode		= IB_WR_RDMA_WRITE,
926 			.send_flags	= IB_SEND_SIGNALED,
927 		}
928 	};
929 
930 	pr_debug("%s-%d: queued zerolength write\n", ch->sess_name,
931 		 ch->qp->qp_num);
932 
933 	return ib_post_send(ch->qp, &wr.wr, NULL);
934 }
935 
936 static void srpt_zerolength_write_done(struct ib_cq *cq, struct ib_wc *wc)
937 {
938 	struct srpt_rdma_ch *ch = wc->qp->qp_context;
939 
940 	pr_debug("%s-%d wc->status %d\n", ch->sess_name, ch->qp->qp_num,
941 		 wc->status);
942 
943 	if (wc->status == IB_WC_SUCCESS) {
944 		srpt_process_wait_list(ch);
945 	} else {
946 		if (srpt_set_ch_state(ch, CH_DISCONNECTED))
947 			schedule_work(&ch->release_work);
948 		else
949 			pr_debug("%s-%d: already disconnected.\n",
950 				 ch->sess_name, ch->qp->qp_num);
951 	}
952 }
953 
954 static int srpt_alloc_rw_ctxs(struct srpt_send_ioctx *ioctx,
955 		struct srp_direct_buf *db, int nbufs, struct scatterlist **sg,
956 		unsigned *sg_cnt)
957 {
958 	enum dma_data_direction dir = target_reverse_dma_direction(&ioctx->cmd);
959 	struct srpt_rdma_ch *ch = ioctx->ch;
960 	struct scatterlist *prev = NULL;
961 	unsigned prev_nents;
962 	int ret, i;
963 
964 	if (nbufs == 1) {
965 		ioctx->rw_ctxs = &ioctx->s_rw_ctx;
966 	} else {
967 		ioctx->rw_ctxs = kmalloc_array(nbufs, sizeof(*ioctx->rw_ctxs),
968 			GFP_KERNEL);
969 		if (!ioctx->rw_ctxs)
970 			return -ENOMEM;
971 	}
972 
973 	for (i = ioctx->n_rw_ctx; i < nbufs; i++, db++) {
974 		struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i];
975 		u64 remote_addr = be64_to_cpu(db->va);
976 		u32 size = be32_to_cpu(db->len);
977 		u32 rkey = be32_to_cpu(db->key);
978 
979 		ret = target_alloc_sgl(&ctx->sg, &ctx->nents, size, false,
980 				i < nbufs - 1);
981 		if (ret)
982 			goto unwind;
983 
984 		ret = rdma_rw_ctx_init(&ctx->rw, ch->qp, ch->sport->port,
985 				ctx->sg, ctx->nents, 0, remote_addr, rkey, dir);
986 		if (ret < 0) {
987 			target_free_sgl(ctx->sg, ctx->nents);
988 			goto unwind;
989 		}
990 
991 		ioctx->n_rdma += ret;
992 		ioctx->n_rw_ctx++;
993 
994 		if (prev) {
995 			sg_unmark_end(&prev[prev_nents - 1]);
996 			sg_chain(prev, prev_nents + 1, ctx->sg);
997 		} else {
998 			*sg = ctx->sg;
999 		}
1000 
1001 		prev = ctx->sg;
1002 		prev_nents = ctx->nents;
1003 
1004 		*sg_cnt += ctx->nents;
1005 	}
1006 
1007 	return 0;
1008 
1009 unwind:
1010 	while (--i >= 0) {
1011 		struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i];
1012 
1013 		rdma_rw_ctx_destroy(&ctx->rw, ch->qp, ch->sport->port,
1014 				ctx->sg, ctx->nents, dir);
1015 		target_free_sgl(ctx->sg, ctx->nents);
1016 	}
1017 	if (ioctx->rw_ctxs != &ioctx->s_rw_ctx)
1018 		kfree(ioctx->rw_ctxs);
1019 	return ret;
1020 }
1021 
1022 static void srpt_free_rw_ctxs(struct srpt_rdma_ch *ch,
1023 				    struct srpt_send_ioctx *ioctx)
1024 {
1025 	enum dma_data_direction dir = target_reverse_dma_direction(&ioctx->cmd);
1026 	int i;
1027 
1028 	for (i = 0; i < ioctx->n_rw_ctx; i++) {
1029 		struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i];
1030 
1031 		rdma_rw_ctx_destroy(&ctx->rw, ch->qp, ch->sport->port,
1032 				ctx->sg, ctx->nents, dir);
1033 		target_free_sgl(ctx->sg, ctx->nents);
1034 	}
1035 
1036 	if (ioctx->rw_ctxs != &ioctx->s_rw_ctx)
1037 		kfree(ioctx->rw_ctxs);
1038 }
1039 
1040 static inline void *srpt_get_desc_buf(struct srp_cmd *srp_cmd)
1041 {
1042 	/*
1043 	 * The pointer computations below will only be compiled correctly
1044 	 * if srp_cmd::add_data is declared as s8*, u8*, s8[] or u8[], so check
1045 	 * whether srp_cmd::add_data has been declared as a byte pointer.
1046 	 */
1047 	BUILD_BUG_ON(!__same_type(srp_cmd->add_data[0], (s8)0) &&
1048 		     !__same_type(srp_cmd->add_data[0], (u8)0));
1049 
1050 	/*
1051 	 * According to the SRP spec, the lower two bits of the 'ADDITIONAL
1052 	 * CDB LENGTH' field are reserved and the size in bytes of this field
1053 	 * is four times the value specified in bits 3..7. Hence the "& ~3".
1054 	 */
1055 	return srp_cmd->add_data + (srp_cmd->add_cdb_len & ~3);
1056 }
1057 
1058 /**
1059  * srpt_get_desc_tbl - parse the data descriptors of a SRP_CMD request
1060  * @recv_ioctx: I/O context associated with the received command @srp_cmd.
1061  * @ioctx: I/O context that will be used for responding to the initiator.
1062  * @srp_cmd: Pointer to the SRP_CMD request data.
1063  * @dir: Pointer to the variable to which the transfer direction will be
1064  *   written.
1065  * @sg: [out] scatterlist for the parsed SRP_CMD.
1066  * @sg_cnt: [out] length of @sg.
1067  * @data_len: Pointer to the variable to which the total data length of all
1068  *   descriptors in the SRP_CMD request will be written.
1069  * @imm_data_offset: [in] Offset in SRP_CMD requests at which immediate data
1070  *   starts.
1071  *
1072  * This function initializes ioctx->nrbuf and ioctx->r_bufs.
1073  *
1074  * Returns -EINVAL when the SRP_CMD request contains inconsistent descriptors;
1075  * -ENOMEM when memory allocation fails and zero upon success.
1076  */
1077 static int srpt_get_desc_tbl(struct srpt_recv_ioctx *recv_ioctx,
1078 		struct srpt_send_ioctx *ioctx,
1079 		struct srp_cmd *srp_cmd, enum dma_data_direction *dir,
1080 		struct scatterlist **sg, unsigned int *sg_cnt, u64 *data_len,
1081 		u16 imm_data_offset)
1082 {
1083 	BUG_ON(!dir);
1084 	BUG_ON(!data_len);
1085 
1086 	/*
1087 	 * The lower four bits of the buffer format field contain the DATA-IN
1088 	 * buffer descriptor format, and the highest four bits contain the
1089 	 * DATA-OUT buffer descriptor format.
1090 	 */
1091 	if (srp_cmd->buf_fmt & 0xf)
1092 		/* DATA-IN: transfer data from target to initiator (read). */
1093 		*dir = DMA_FROM_DEVICE;
1094 	else if (srp_cmd->buf_fmt >> 4)
1095 		/* DATA-OUT: transfer data from initiator to target (write). */
1096 		*dir = DMA_TO_DEVICE;
1097 	else
1098 		*dir = DMA_NONE;
1099 
1100 	/* initialize data_direction early as srpt_alloc_rw_ctxs needs it */
1101 	ioctx->cmd.data_direction = *dir;
1102 
1103 	if (((srp_cmd->buf_fmt & 0xf) == SRP_DATA_DESC_DIRECT) ||
1104 	    ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_DIRECT)) {
1105 		struct srp_direct_buf *db = srpt_get_desc_buf(srp_cmd);
1106 
1107 		*data_len = be32_to_cpu(db->len);
1108 		return srpt_alloc_rw_ctxs(ioctx, db, 1, sg, sg_cnt);
1109 	} else if (((srp_cmd->buf_fmt & 0xf) == SRP_DATA_DESC_INDIRECT) ||
1110 		   ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_INDIRECT)) {
1111 		struct srp_indirect_buf *idb = srpt_get_desc_buf(srp_cmd);
1112 		int nbufs = be32_to_cpu(idb->table_desc.len) /
1113 				sizeof(struct srp_direct_buf);
1114 
1115 		if (nbufs >
1116 		    (srp_cmd->data_out_desc_cnt + srp_cmd->data_in_desc_cnt)) {
1117 			pr_err("received unsupported SRP_CMD request type (%u out + %u in != %u / %zu)\n",
1118 			       srp_cmd->data_out_desc_cnt,
1119 			       srp_cmd->data_in_desc_cnt,
1120 			       be32_to_cpu(idb->table_desc.len),
1121 			       sizeof(struct srp_direct_buf));
1122 			return -EINVAL;
1123 		}
1124 
1125 		*data_len = be32_to_cpu(idb->len);
1126 		return srpt_alloc_rw_ctxs(ioctx, idb->desc_list, nbufs,
1127 				sg, sg_cnt);
1128 	} else if ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_IMM) {
1129 		struct srp_imm_buf *imm_buf = srpt_get_desc_buf(srp_cmd);
1130 		void *data = (void *)srp_cmd + imm_data_offset;
1131 		uint32_t len = be32_to_cpu(imm_buf->len);
1132 		uint32_t req_size = imm_data_offset + len;
1133 
1134 		if (req_size > srp_max_req_size) {
1135 			pr_err("Immediate data (length %d + %d) exceeds request size %d\n",
1136 			       imm_data_offset, len, srp_max_req_size);
1137 			return -EINVAL;
1138 		}
1139 		if (recv_ioctx->byte_len < req_size) {
1140 			pr_err("Received too few data - %d < %d\n",
1141 			       recv_ioctx->byte_len, req_size);
1142 			return -EIO;
1143 		}
1144 		/*
1145 		 * The immediate data buffer descriptor must occur before the
1146 		 * immediate data itself.
1147 		 */
1148 		if ((void *)(imm_buf + 1) > (void *)data) {
1149 			pr_err("Received invalid write request\n");
1150 			return -EINVAL;
1151 		}
1152 		*data_len = len;
1153 		ioctx->recv_ioctx = recv_ioctx;
1154 		if ((uintptr_t)data & 511) {
1155 			pr_warn_once("Internal error - the receive buffers are not aligned properly.\n");
1156 			return -EINVAL;
1157 		}
1158 		sg_init_one(&ioctx->imm_sg, data, len);
1159 		*sg = &ioctx->imm_sg;
1160 		*sg_cnt = 1;
1161 		return 0;
1162 	} else {
1163 		*data_len = 0;
1164 		return 0;
1165 	}
1166 }
1167 
1168 /**
1169  * srpt_init_ch_qp - initialize queue pair attributes
1170  * @ch: SRPT RDMA channel.
1171  * @qp: Queue pair pointer.
1172  *
1173  * Initialized the attributes of queue pair 'qp' by allowing local write,
1174  * remote read and remote write. Also transitions 'qp' to state IB_QPS_INIT.
1175  */
1176 static int srpt_init_ch_qp(struct srpt_rdma_ch *ch, struct ib_qp *qp)
1177 {
1178 	struct ib_qp_attr *attr;
1179 	int ret;
1180 
1181 	WARN_ON_ONCE(ch->using_rdma_cm);
1182 
1183 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
1184 	if (!attr)
1185 		return -ENOMEM;
1186 
1187 	attr->qp_state = IB_QPS_INIT;
1188 	attr->qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1189 	attr->port_num = ch->sport->port;
1190 
1191 	ret = ib_find_cached_pkey(ch->sport->sdev->device, ch->sport->port,
1192 				  ch->pkey, &attr->pkey_index);
1193 	if (ret < 0)
1194 		pr_err("Translating pkey %#x failed (%d) - using index 0\n",
1195 		       ch->pkey, ret);
1196 
1197 	ret = ib_modify_qp(qp, attr,
1198 			   IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PORT |
1199 			   IB_QP_PKEY_INDEX);
1200 
1201 	kfree(attr);
1202 	return ret;
1203 }
1204 
1205 /**
1206  * srpt_ch_qp_rtr - change the state of a channel to 'ready to receive' (RTR)
1207  * @ch: channel of the queue pair.
1208  * @qp: queue pair to change the state of.
1209  *
1210  * Returns zero upon success and a negative value upon failure.
1211  *
1212  * Note: currently a struct ib_qp_attr takes 136 bytes on a 64-bit system.
1213  * If this structure ever becomes larger, it might be necessary to allocate
1214  * it dynamically instead of on the stack.
1215  */
1216 static int srpt_ch_qp_rtr(struct srpt_rdma_ch *ch, struct ib_qp *qp)
1217 {
1218 	struct ib_qp_attr qp_attr;
1219 	int attr_mask;
1220 	int ret;
1221 
1222 	WARN_ON_ONCE(ch->using_rdma_cm);
1223 
1224 	qp_attr.qp_state = IB_QPS_RTR;
1225 	ret = ib_cm_init_qp_attr(ch->ib_cm.cm_id, &qp_attr, &attr_mask);
1226 	if (ret)
1227 		goto out;
1228 
1229 	qp_attr.max_dest_rd_atomic = 4;
1230 
1231 	ret = ib_modify_qp(qp, &qp_attr, attr_mask);
1232 
1233 out:
1234 	return ret;
1235 }
1236 
1237 /**
1238  * srpt_ch_qp_rts - change the state of a channel to 'ready to send' (RTS)
1239  * @ch: channel of the queue pair.
1240  * @qp: queue pair to change the state of.
1241  *
1242  * Returns zero upon success and a negative value upon failure.
1243  *
1244  * Note: currently a struct ib_qp_attr takes 136 bytes on a 64-bit system.
1245  * If this structure ever becomes larger, it might be necessary to allocate
1246  * it dynamically instead of on the stack.
1247  */
1248 static int srpt_ch_qp_rts(struct srpt_rdma_ch *ch, struct ib_qp *qp)
1249 {
1250 	struct ib_qp_attr qp_attr;
1251 	int attr_mask;
1252 	int ret;
1253 
1254 	qp_attr.qp_state = IB_QPS_RTS;
1255 	ret = ib_cm_init_qp_attr(ch->ib_cm.cm_id, &qp_attr, &attr_mask);
1256 	if (ret)
1257 		goto out;
1258 
1259 	qp_attr.max_rd_atomic = 4;
1260 
1261 	ret = ib_modify_qp(qp, &qp_attr, attr_mask);
1262 
1263 out:
1264 	return ret;
1265 }
1266 
1267 /**
1268  * srpt_ch_qp_err - set the channel queue pair state to 'error'
1269  * @ch: SRPT RDMA channel.
1270  */
1271 static int srpt_ch_qp_err(struct srpt_rdma_ch *ch)
1272 {
1273 	struct ib_qp_attr qp_attr;
1274 
1275 	qp_attr.qp_state = IB_QPS_ERR;
1276 	return ib_modify_qp(ch->qp, &qp_attr, IB_QP_STATE);
1277 }
1278 
1279 /**
1280  * srpt_get_send_ioctx - obtain an I/O context for sending to the initiator
1281  * @ch: SRPT RDMA channel.
1282  */
1283 static struct srpt_send_ioctx *srpt_get_send_ioctx(struct srpt_rdma_ch *ch)
1284 {
1285 	struct srpt_send_ioctx *ioctx;
1286 	int tag, cpu;
1287 
1288 	BUG_ON(!ch);
1289 
1290 	tag = sbitmap_queue_get(&ch->sess->sess_tag_pool, &cpu);
1291 	if (tag < 0)
1292 		return NULL;
1293 
1294 	ioctx = ch->ioctx_ring[tag];
1295 	BUG_ON(ioctx->ch != ch);
1296 	ioctx->state = SRPT_STATE_NEW;
1297 	WARN_ON_ONCE(ioctx->recv_ioctx);
1298 	ioctx->n_rdma = 0;
1299 	ioctx->n_rw_ctx = 0;
1300 	ioctx->queue_status_only = false;
1301 	/*
1302 	 * transport_init_se_cmd() does not initialize all fields, so do it
1303 	 * here.
1304 	 */
1305 	memset(&ioctx->cmd, 0, sizeof(ioctx->cmd));
1306 	memset(&ioctx->sense_data, 0, sizeof(ioctx->sense_data));
1307 	ioctx->cmd.map_tag = tag;
1308 	ioctx->cmd.map_cpu = cpu;
1309 
1310 	return ioctx;
1311 }
1312 
1313 /**
1314  * srpt_abort_cmd - abort a SCSI command
1315  * @ioctx:   I/O context associated with the SCSI command.
1316  */
1317 static int srpt_abort_cmd(struct srpt_send_ioctx *ioctx)
1318 {
1319 	enum srpt_command_state state;
1320 
1321 	BUG_ON(!ioctx);
1322 
1323 	/*
1324 	 * If the command is in a state where the target core is waiting for
1325 	 * the ib_srpt driver, change the state to the next state.
1326 	 */
1327 
1328 	state = ioctx->state;
1329 	switch (state) {
1330 	case SRPT_STATE_NEED_DATA:
1331 		ioctx->state = SRPT_STATE_DATA_IN;
1332 		break;
1333 	case SRPT_STATE_CMD_RSP_SENT:
1334 	case SRPT_STATE_MGMT_RSP_SENT:
1335 		ioctx->state = SRPT_STATE_DONE;
1336 		break;
1337 	default:
1338 		WARN_ONCE(true, "%s: unexpected I/O context state %d\n",
1339 			  __func__, state);
1340 		break;
1341 	}
1342 
1343 	pr_debug("Aborting cmd with state %d -> %d and tag %lld\n", state,
1344 		 ioctx->state, ioctx->cmd.tag);
1345 
1346 	switch (state) {
1347 	case SRPT_STATE_NEW:
1348 	case SRPT_STATE_DATA_IN:
1349 	case SRPT_STATE_MGMT:
1350 	case SRPT_STATE_DONE:
1351 		/*
1352 		 * Do nothing - defer abort processing until
1353 		 * srpt_queue_response() is invoked.
1354 		 */
1355 		break;
1356 	case SRPT_STATE_NEED_DATA:
1357 		pr_debug("tag %#llx: RDMA read error\n", ioctx->cmd.tag);
1358 		transport_generic_request_failure(&ioctx->cmd,
1359 					TCM_CHECK_CONDITION_ABORT_CMD);
1360 		break;
1361 	case SRPT_STATE_CMD_RSP_SENT:
1362 		/*
1363 		 * SRP_RSP sending failed or the SRP_RSP send completion has
1364 		 * not been received in time.
1365 		 */
1366 		transport_generic_free_cmd(&ioctx->cmd, 0);
1367 		break;
1368 	case SRPT_STATE_MGMT_RSP_SENT:
1369 		transport_generic_free_cmd(&ioctx->cmd, 0);
1370 		break;
1371 	default:
1372 		WARN(1, "Unexpected command state (%d)", state);
1373 		break;
1374 	}
1375 
1376 	return state;
1377 }
1378 
1379 /**
1380  * srpt_rdma_read_done - RDMA read completion callback
1381  * @cq: Completion queue.
1382  * @wc: Work completion.
1383  *
1384  * XXX: what is now target_execute_cmd used to be asynchronous, and unmapping
1385  * the data that has been transferred via IB RDMA had to be postponed until the
1386  * check_stop_free() callback.  None of this is necessary anymore and needs to
1387  * be cleaned up.
1388  */
1389 static void srpt_rdma_read_done(struct ib_cq *cq, struct ib_wc *wc)
1390 {
1391 	struct srpt_rdma_ch *ch = wc->qp->qp_context;
1392 	struct srpt_send_ioctx *ioctx =
1393 		container_of(wc->wr_cqe, struct srpt_send_ioctx, rdma_cqe);
1394 
1395 	WARN_ON(ioctx->n_rdma <= 0);
1396 	atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
1397 	ioctx->n_rdma = 0;
1398 
1399 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
1400 		pr_info("RDMA_READ for ioctx 0x%p failed with status %d\n",
1401 			ioctx, wc->status);
1402 		srpt_abort_cmd(ioctx);
1403 		return;
1404 	}
1405 
1406 	if (srpt_test_and_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA,
1407 					SRPT_STATE_DATA_IN))
1408 		target_execute_cmd(&ioctx->cmd);
1409 	else
1410 		pr_err("%s[%d]: wrong state = %d\n", __func__,
1411 		       __LINE__, ioctx->state);
1412 }
1413 
1414 /**
1415  * srpt_build_cmd_rsp - build a SRP_RSP response
1416  * @ch: RDMA channel through which the request has been received.
1417  * @ioctx: I/O context associated with the SRP_CMD request. The response will
1418  *   be built in the buffer ioctx->buf points at and hence this function will
1419  *   overwrite the request data.
1420  * @tag: tag of the request for which this response is being generated.
1421  * @status: value for the STATUS field of the SRP_RSP information unit.
1422  *
1423  * Returns the size in bytes of the SRP_RSP response.
1424  *
1425  * An SRP_RSP response contains a SCSI status or service response. See also
1426  * section 6.9 in the SRP r16a document for the format of an SRP_RSP
1427  * response. See also SPC-2 for more information about sense data.
1428  */
1429 static int srpt_build_cmd_rsp(struct srpt_rdma_ch *ch,
1430 			      struct srpt_send_ioctx *ioctx, u64 tag,
1431 			      int status)
1432 {
1433 	struct se_cmd *cmd = &ioctx->cmd;
1434 	struct srp_rsp *srp_rsp;
1435 	const u8 *sense_data;
1436 	int sense_data_len, max_sense_len;
1437 	u32 resid = cmd->residual_count;
1438 
1439 	/*
1440 	 * The lowest bit of all SAM-3 status codes is zero (see also
1441 	 * paragraph 5.3 in SAM-3).
1442 	 */
1443 	WARN_ON(status & 1);
1444 
1445 	srp_rsp = ioctx->ioctx.buf;
1446 	BUG_ON(!srp_rsp);
1447 
1448 	sense_data = ioctx->sense_data;
1449 	sense_data_len = ioctx->cmd.scsi_sense_length;
1450 	WARN_ON(sense_data_len > sizeof(ioctx->sense_data));
1451 
1452 	memset(srp_rsp, 0, sizeof(*srp_rsp));
1453 	srp_rsp->opcode = SRP_RSP;
1454 	srp_rsp->req_lim_delta =
1455 		cpu_to_be32(1 + atomic_xchg(&ch->req_lim_delta, 0));
1456 	srp_rsp->tag = tag;
1457 	srp_rsp->status = status;
1458 
1459 	if (cmd->se_cmd_flags & SCF_UNDERFLOW_BIT) {
1460 		if (cmd->data_direction == DMA_TO_DEVICE) {
1461 			/* residual data from an underflow write */
1462 			srp_rsp->flags = SRP_RSP_FLAG_DOUNDER;
1463 			srp_rsp->data_out_res_cnt = cpu_to_be32(resid);
1464 		} else if (cmd->data_direction == DMA_FROM_DEVICE) {
1465 			/* residual data from an underflow read */
1466 			srp_rsp->flags = SRP_RSP_FLAG_DIUNDER;
1467 			srp_rsp->data_in_res_cnt = cpu_to_be32(resid);
1468 		}
1469 	} else if (cmd->se_cmd_flags & SCF_OVERFLOW_BIT) {
1470 		if (cmd->data_direction == DMA_TO_DEVICE) {
1471 			/* residual data from an overflow write */
1472 			srp_rsp->flags = SRP_RSP_FLAG_DOOVER;
1473 			srp_rsp->data_out_res_cnt = cpu_to_be32(resid);
1474 		} else if (cmd->data_direction == DMA_FROM_DEVICE) {
1475 			/* residual data from an overflow read */
1476 			srp_rsp->flags = SRP_RSP_FLAG_DIOVER;
1477 			srp_rsp->data_in_res_cnt = cpu_to_be32(resid);
1478 		}
1479 	}
1480 
1481 	if (sense_data_len) {
1482 		BUILD_BUG_ON(MIN_MAX_RSP_SIZE <= sizeof(*srp_rsp));
1483 		max_sense_len = ch->max_ti_iu_len - sizeof(*srp_rsp);
1484 		if (sense_data_len > max_sense_len) {
1485 			pr_warn("truncated sense data from %d to %d bytes\n",
1486 				sense_data_len, max_sense_len);
1487 			sense_data_len = max_sense_len;
1488 		}
1489 
1490 		srp_rsp->flags |= SRP_RSP_FLAG_SNSVALID;
1491 		srp_rsp->sense_data_len = cpu_to_be32(sense_data_len);
1492 		memcpy(srp_rsp->data, sense_data, sense_data_len);
1493 	}
1494 
1495 	return sizeof(*srp_rsp) + sense_data_len;
1496 }
1497 
1498 /**
1499  * srpt_build_tskmgmt_rsp - build a task management response
1500  * @ch:       RDMA channel through which the request has been received.
1501  * @ioctx:    I/O context in which the SRP_RSP response will be built.
1502  * @rsp_code: RSP_CODE that will be stored in the response.
1503  * @tag:      Tag of the request for which this response is being generated.
1504  *
1505  * Returns the size in bytes of the SRP_RSP response.
1506  *
1507  * An SRP_RSP response contains a SCSI status or service response. See also
1508  * section 6.9 in the SRP r16a document for the format of an SRP_RSP
1509  * response.
1510  */
1511 static int srpt_build_tskmgmt_rsp(struct srpt_rdma_ch *ch,
1512 				  struct srpt_send_ioctx *ioctx,
1513 				  u8 rsp_code, u64 tag)
1514 {
1515 	struct srp_rsp *srp_rsp;
1516 	int resp_data_len;
1517 	int resp_len;
1518 
1519 	resp_data_len = 4;
1520 	resp_len = sizeof(*srp_rsp) + resp_data_len;
1521 
1522 	srp_rsp = ioctx->ioctx.buf;
1523 	BUG_ON(!srp_rsp);
1524 	memset(srp_rsp, 0, sizeof(*srp_rsp));
1525 
1526 	srp_rsp->opcode = SRP_RSP;
1527 	srp_rsp->req_lim_delta =
1528 		cpu_to_be32(1 + atomic_xchg(&ch->req_lim_delta, 0));
1529 	srp_rsp->tag = tag;
1530 
1531 	srp_rsp->flags |= SRP_RSP_FLAG_RSPVALID;
1532 	srp_rsp->resp_data_len = cpu_to_be32(resp_data_len);
1533 	srp_rsp->data[3] = rsp_code;
1534 
1535 	return resp_len;
1536 }
1537 
1538 static int srpt_check_stop_free(struct se_cmd *cmd)
1539 {
1540 	struct srpt_send_ioctx *ioctx = container_of(cmd,
1541 				struct srpt_send_ioctx, cmd);
1542 
1543 	return target_put_sess_cmd(&ioctx->cmd);
1544 }
1545 
1546 /**
1547  * srpt_handle_cmd - process a SRP_CMD information unit
1548  * @ch: SRPT RDMA channel.
1549  * @recv_ioctx: Receive I/O context.
1550  * @send_ioctx: Send I/O context.
1551  */
1552 static void srpt_handle_cmd(struct srpt_rdma_ch *ch,
1553 			    struct srpt_recv_ioctx *recv_ioctx,
1554 			    struct srpt_send_ioctx *send_ioctx)
1555 {
1556 	struct se_cmd *cmd;
1557 	struct srp_cmd *srp_cmd;
1558 	struct scatterlist *sg = NULL;
1559 	unsigned sg_cnt = 0;
1560 	u64 data_len;
1561 	enum dma_data_direction dir;
1562 	int rc;
1563 
1564 	BUG_ON(!send_ioctx);
1565 
1566 	srp_cmd = recv_ioctx->ioctx.buf + recv_ioctx->ioctx.offset;
1567 	cmd = &send_ioctx->cmd;
1568 	cmd->tag = srp_cmd->tag;
1569 
1570 	switch (srp_cmd->task_attr) {
1571 	case SRP_CMD_SIMPLE_Q:
1572 		cmd->sam_task_attr = TCM_SIMPLE_TAG;
1573 		break;
1574 	case SRP_CMD_ORDERED_Q:
1575 	default:
1576 		cmd->sam_task_attr = TCM_ORDERED_TAG;
1577 		break;
1578 	case SRP_CMD_HEAD_OF_Q:
1579 		cmd->sam_task_attr = TCM_HEAD_TAG;
1580 		break;
1581 	case SRP_CMD_ACA:
1582 		cmd->sam_task_attr = TCM_ACA_TAG;
1583 		break;
1584 	}
1585 
1586 	rc = srpt_get_desc_tbl(recv_ioctx, send_ioctx, srp_cmd, &dir,
1587 			       &sg, &sg_cnt, &data_len, ch->imm_data_offset);
1588 	if (rc) {
1589 		if (rc != -EAGAIN) {
1590 			pr_err("0x%llx: parsing SRP descriptor table failed.\n",
1591 			       srp_cmd->tag);
1592 		}
1593 		goto busy;
1594 	}
1595 
1596 	rc = target_init_cmd(cmd, ch->sess, &send_ioctx->sense_data[0],
1597 			     scsilun_to_int(&srp_cmd->lun), data_len,
1598 			     TCM_SIMPLE_TAG, dir, TARGET_SCF_ACK_KREF);
1599 	if (rc != 0) {
1600 		pr_debug("target_submit_cmd() returned %d for tag %#llx\n", rc,
1601 			 srp_cmd->tag);
1602 		goto busy;
1603 	}
1604 
1605 	if (target_submit_prep(cmd, srp_cmd->cdb, sg, sg_cnt, NULL, 0, NULL, 0,
1606 			       GFP_KERNEL))
1607 		return;
1608 
1609 	target_submit(cmd);
1610 	return;
1611 
1612 busy:
1613 	target_send_busy(cmd);
1614 }
1615 
1616 static int srp_tmr_to_tcm(int fn)
1617 {
1618 	switch (fn) {
1619 	case SRP_TSK_ABORT_TASK:
1620 		return TMR_ABORT_TASK;
1621 	case SRP_TSK_ABORT_TASK_SET:
1622 		return TMR_ABORT_TASK_SET;
1623 	case SRP_TSK_CLEAR_TASK_SET:
1624 		return TMR_CLEAR_TASK_SET;
1625 	case SRP_TSK_LUN_RESET:
1626 		return TMR_LUN_RESET;
1627 	case SRP_TSK_CLEAR_ACA:
1628 		return TMR_CLEAR_ACA;
1629 	default:
1630 		return -1;
1631 	}
1632 }
1633 
1634 /**
1635  * srpt_handle_tsk_mgmt - process a SRP_TSK_MGMT information unit
1636  * @ch: SRPT RDMA channel.
1637  * @recv_ioctx: Receive I/O context.
1638  * @send_ioctx: Send I/O context.
1639  *
1640  * Returns 0 if and only if the request will be processed by the target core.
1641  *
1642  * For more information about SRP_TSK_MGMT information units, see also section
1643  * 6.7 in the SRP r16a document.
1644  */
1645 static void srpt_handle_tsk_mgmt(struct srpt_rdma_ch *ch,
1646 				 struct srpt_recv_ioctx *recv_ioctx,
1647 				 struct srpt_send_ioctx *send_ioctx)
1648 {
1649 	struct srp_tsk_mgmt *srp_tsk;
1650 	struct se_cmd *cmd;
1651 	struct se_session *sess = ch->sess;
1652 	int tcm_tmr;
1653 	int rc;
1654 
1655 	BUG_ON(!send_ioctx);
1656 
1657 	srp_tsk = recv_ioctx->ioctx.buf + recv_ioctx->ioctx.offset;
1658 	cmd = &send_ioctx->cmd;
1659 
1660 	pr_debug("recv tsk_mgmt fn %d for task_tag %lld and cmd tag %lld ch %p sess %p\n",
1661 		 srp_tsk->tsk_mgmt_func, srp_tsk->task_tag, srp_tsk->tag, ch,
1662 		 ch->sess);
1663 
1664 	srpt_set_cmd_state(send_ioctx, SRPT_STATE_MGMT);
1665 	send_ioctx->cmd.tag = srp_tsk->tag;
1666 	tcm_tmr = srp_tmr_to_tcm(srp_tsk->tsk_mgmt_func);
1667 	rc = target_submit_tmr(&send_ioctx->cmd, sess, NULL,
1668 			       scsilun_to_int(&srp_tsk->lun), srp_tsk, tcm_tmr,
1669 			       GFP_KERNEL, srp_tsk->task_tag,
1670 			       TARGET_SCF_ACK_KREF);
1671 	if (rc != 0) {
1672 		send_ioctx->cmd.se_tmr_req->response = TMR_FUNCTION_REJECTED;
1673 		cmd->se_tfo->queue_tm_rsp(cmd);
1674 	}
1675 	return;
1676 }
1677 
1678 /**
1679  * srpt_handle_new_iu - process a newly received information unit
1680  * @ch:    RDMA channel through which the information unit has been received.
1681  * @recv_ioctx: Receive I/O context associated with the information unit.
1682  */
1683 static bool
1684 srpt_handle_new_iu(struct srpt_rdma_ch *ch, struct srpt_recv_ioctx *recv_ioctx)
1685 {
1686 	struct srpt_send_ioctx *send_ioctx = NULL;
1687 	struct srp_cmd *srp_cmd;
1688 	bool res = false;
1689 	u8 opcode;
1690 
1691 	BUG_ON(!ch);
1692 	BUG_ON(!recv_ioctx);
1693 
1694 	if (unlikely(ch->state == CH_CONNECTING))
1695 		goto push;
1696 
1697 	ib_dma_sync_single_for_cpu(ch->sport->sdev->device,
1698 				   recv_ioctx->ioctx.dma,
1699 				   recv_ioctx->ioctx.offset + srp_max_req_size,
1700 				   DMA_FROM_DEVICE);
1701 
1702 	srp_cmd = recv_ioctx->ioctx.buf + recv_ioctx->ioctx.offset;
1703 	opcode = srp_cmd->opcode;
1704 	if (opcode == SRP_CMD || opcode == SRP_TSK_MGMT) {
1705 		send_ioctx = srpt_get_send_ioctx(ch);
1706 		if (unlikely(!send_ioctx))
1707 			goto push;
1708 	}
1709 
1710 	if (!list_empty(&recv_ioctx->wait_list)) {
1711 		WARN_ON_ONCE(!ch->processing_wait_list);
1712 		list_del_init(&recv_ioctx->wait_list);
1713 	}
1714 
1715 	switch (opcode) {
1716 	case SRP_CMD:
1717 		srpt_handle_cmd(ch, recv_ioctx, send_ioctx);
1718 		break;
1719 	case SRP_TSK_MGMT:
1720 		srpt_handle_tsk_mgmt(ch, recv_ioctx, send_ioctx);
1721 		break;
1722 	case SRP_I_LOGOUT:
1723 		pr_err("Not yet implemented: SRP_I_LOGOUT\n");
1724 		break;
1725 	case SRP_CRED_RSP:
1726 		pr_debug("received SRP_CRED_RSP\n");
1727 		break;
1728 	case SRP_AER_RSP:
1729 		pr_debug("received SRP_AER_RSP\n");
1730 		break;
1731 	case SRP_RSP:
1732 		pr_err("Received SRP_RSP\n");
1733 		break;
1734 	default:
1735 		pr_err("received IU with unknown opcode 0x%x\n", opcode);
1736 		break;
1737 	}
1738 
1739 	if (!send_ioctx || !send_ioctx->recv_ioctx)
1740 		srpt_post_recv(ch->sport->sdev, ch, recv_ioctx);
1741 	res = true;
1742 
1743 out:
1744 	return res;
1745 
1746 push:
1747 	if (list_empty(&recv_ioctx->wait_list)) {
1748 		WARN_ON_ONCE(ch->processing_wait_list);
1749 		list_add_tail(&recv_ioctx->wait_list, &ch->cmd_wait_list);
1750 	}
1751 	goto out;
1752 }
1753 
1754 static void srpt_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1755 {
1756 	struct srpt_rdma_ch *ch = wc->qp->qp_context;
1757 	struct srpt_recv_ioctx *ioctx =
1758 		container_of(wc->wr_cqe, struct srpt_recv_ioctx, ioctx.cqe);
1759 
1760 	if (wc->status == IB_WC_SUCCESS) {
1761 		int req_lim;
1762 
1763 		req_lim = atomic_dec_return(&ch->req_lim);
1764 		if (unlikely(req_lim < 0))
1765 			pr_err("req_lim = %d < 0\n", req_lim);
1766 		ioctx->byte_len = wc->byte_len;
1767 		srpt_handle_new_iu(ch, ioctx);
1768 	} else {
1769 		pr_info_ratelimited("receiving failed for ioctx %p with status %d\n",
1770 				    ioctx, wc->status);
1771 	}
1772 }
1773 
1774 /*
1775  * This function must be called from the context in which RDMA completions are
1776  * processed because it accesses the wait list without protection against
1777  * access from other threads.
1778  */
1779 static void srpt_process_wait_list(struct srpt_rdma_ch *ch)
1780 {
1781 	struct srpt_recv_ioctx *recv_ioctx, *tmp;
1782 
1783 	WARN_ON_ONCE(ch->state == CH_CONNECTING);
1784 
1785 	if (list_empty(&ch->cmd_wait_list))
1786 		return;
1787 
1788 	WARN_ON_ONCE(ch->processing_wait_list);
1789 	ch->processing_wait_list = true;
1790 	list_for_each_entry_safe(recv_ioctx, tmp, &ch->cmd_wait_list,
1791 				 wait_list) {
1792 		if (!srpt_handle_new_iu(ch, recv_ioctx))
1793 			break;
1794 	}
1795 	ch->processing_wait_list = false;
1796 }
1797 
1798 /**
1799  * srpt_send_done - send completion callback
1800  * @cq: Completion queue.
1801  * @wc: Work completion.
1802  *
1803  * Note: Although this has not yet been observed during tests, at least in
1804  * theory it is possible that the srpt_get_send_ioctx() call invoked by
1805  * srpt_handle_new_iu() fails. This is possible because the req_lim_delta
1806  * value in each response is set to one, and it is possible that this response
1807  * makes the initiator send a new request before the send completion for that
1808  * response has been processed. This could e.g. happen if the call to
1809  * srpt_put_send_iotcx() is delayed because of a higher priority interrupt or
1810  * if IB retransmission causes generation of the send completion to be
1811  * delayed. Incoming information units for which srpt_get_send_ioctx() fails
1812  * are queued on cmd_wait_list. The code below processes these delayed
1813  * requests one at a time.
1814  */
1815 static void srpt_send_done(struct ib_cq *cq, struct ib_wc *wc)
1816 {
1817 	struct srpt_rdma_ch *ch = wc->qp->qp_context;
1818 	struct srpt_send_ioctx *ioctx =
1819 		container_of(wc->wr_cqe, struct srpt_send_ioctx, ioctx.cqe);
1820 	enum srpt_command_state state;
1821 
1822 	state = srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
1823 
1824 	WARN_ON(state != SRPT_STATE_CMD_RSP_SENT &&
1825 		state != SRPT_STATE_MGMT_RSP_SENT);
1826 
1827 	atomic_add(1 + ioctx->n_rdma, &ch->sq_wr_avail);
1828 
1829 	if (wc->status != IB_WC_SUCCESS)
1830 		pr_info("sending response for ioctx 0x%p failed with status %d\n",
1831 			ioctx, wc->status);
1832 
1833 	if (state != SRPT_STATE_DONE) {
1834 		transport_generic_free_cmd(&ioctx->cmd, 0);
1835 	} else {
1836 		pr_err("IB completion has been received too late for wr_id = %u.\n",
1837 		       ioctx->ioctx.index);
1838 	}
1839 
1840 	srpt_process_wait_list(ch);
1841 }
1842 
1843 /**
1844  * srpt_create_ch_ib - create receive and send completion queues
1845  * @ch: SRPT RDMA channel.
1846  */
1847 static int srpt_create_ch_ib(struct srpt_rdma_ch *ch)
1848 {
1849 	struct ib_qp_init_attr *qp_init;
1850 	struct srpt_port *sport = ch->sport;
1851 	struct srpt_device *sdev = sport->sdev;
1852 	const struct ib_device_attr *attrs = &sdev->device->attrs;
1853 	int sq_size = sport->port_attrib.srp_sq_size;
1854 	int i, ret;
1855 
1856 	WARN_ON(ch->rq_size < 1);
1857 
1858 	ret = -ENOMEM;
1859 	qp_init = kzalloc(sizeof(*qp_init), GFP_KERNEL);
1860 	if (!qp_init)
1861 		goto out;
1862 
1863 retry:
1864 	ch->cq = ib_cq_pool_get(sdev->device, ch->rq_size + sq_size, -1,
1865 				 IB_POLL_WORKQUEUE);
1866 	if (IS_ERR(ch->cq)) {
1867 		ret = PTR_ERR(ch->cq);
1868 		pr_err("failed to create CQ cqe= %d ret= %d\n",
1869 		       ch->rq_size + sq_size, ret);
1870 		goto out;
1871 	}
1872 	ch->cq_size = ch->rq_size + sq_size;
1873 
1874 	qp_init->qp_context = (void *)ch;
1875 	qp_init->event_handler = srpt_qp_event;
1876 	qp_init->send_cq = ch->cq;
1877 	qp_init->recv_cq = ch->cq;
1878 	qp_init->sq_sig_type = IB_SIGNAL_REQ_WR;
1879 	qp_init->qp_type = IB_QPT_RC;
1880 	/*
1881 	 * We divide up our send queue size into half SEND WRs to send the
1882 	 * completions, and half R/W contexts to actually do the RDMA
1883 	 * READ/WRITE transfers.  Note that we need to allocate CQ slots for
1884 	 * both both, as RDMA contexts will also post completions for the
1885 	 * RDMA READ case.
1886 	 */
1887 	qp_init->cap.max_send_wr = min(sq_size / 2, attrs->max_qp_wr);
1888 	qp_init->cap.max_rdma_ctxs = sq_size / 2;
1889 	qp_init->cap.max_send_sge = attrs->max_send_sge;
1890 	qp_init->cap.max_recv_sge = 1;
1891 	qp_init->port_num = ch->sport->port;
1892 	if (sdev->use_srq)
1893 		qp_init->srq = sdev->srq;
1894 	else
1895 		qp_init->cap.max_recv_wr = ch->rq_size;
1896 
1897 	if (ch->using_rdma_cm) {
1898 		ret = rdma_create_qp(ch->rdma_cm.cm_id, sdev->pd, qp_init);
1899 		ch->qp = ch->rdma_cm.cm_id->qp;
1900 	} else {
1901 		ch->qp = ib_create_qp(sdev->pd, qp_init);
1902 		if (!IS_ERR(ch->qp)) {
1903 			ret = srpt_init_ch_qp(ch, ch->qp);
1904 			if (ret)
1905 				ib_destroy_qp(ch->qp);
1906 		} else {
1907 			ret = PTR_ERR(ch->qp);
1908 		}
1909 	}
1910 	if (ret) {
1911 		bool retry = sq_size > MIN_SRPT_SQ_SIZE;
1912 
1913 		if (retry) {
1914 			pr_debug("failed to create queue pair with sq_size = %d (%d) - retrying\n",
1915 				 sq_size, ret);
1916 			ib_cq_pool_put(ch->cq, ch->cq_size);
1917 			sq_size = max(sq_size / 2, MIN_SRPT_SQ_SIZE);
1918 			goto retry;
1919 		} else {
1920 			pr_err("failed to create queue pair with sq_size = %d (%d)\n",
1921 			       sq_size, ret);
1922 			goto err_destroy_cq;
1923 		}
1924 	}
1925 
1926 	atomic_set(&ch->sq_wr_avail, qp_init->cap.max_send_wr);
1927 
1928 	pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d ch= %p\n",
1929 		 __func__, ch->cq->cqe, qp_init->cap.max_send_sge,
1930 		 qp_init->cap.max_send_wr, ch);
1931 
1932 	if (!sdev->use_srq)
1933 		for (i = 0; i < ch->rq_size; i++)
1934 			srpt_post_recv(sdev, ch, ch->ioctx_recv_ring[i]);
1935 
1936 out:
1937 	kfree(qp_init);
1938 	return ret;
1939 
1940 err_destroy_cq:
1941 	ch->qp = NULL;
1942 	ib_cq_pool_put(ch->cq, ch->cq_size);
1943 	goto out;
1944 }
1945 
1946 static void srpt_destroy_ch_ib(struct srpt_rdma_ch *ch)
1947 {
1948 	ib_destroy_qp(ch->qp);
1949 	ib_cq_pool_put(ch->cq, ch->cq_size);
1950 }
1951 
1952 /**
1953  * srpt_close_ch - close a RDMA channel
1954  * @ch: SRPT RDMA channel.
1955  *
1956  * Make sure all resources associated with the channel will be deallocated at
1957  * an appropriate time.
1958  *
1959  * Returns true if and only if the channel state has been modified into
1960  * CH_DRAINING.
1961  */
1962 static bool srpt_close_ch(struct srpt_rdma_ch *ch)
1963 {
1964 	int ret;
1965 
1966 	if (!srpt_set_ch_state(ch, CH_DRAINING)) {
1967 		pr_debug("%s: already closed\n", ch->sess_name);
1968 		return false;
1969 	}
1970 
1971 	kref_get(&ch->kref);
1972 
1973 	ret = srpt_ch_qp_err(ch);
1974 	if (ret < 0)
1975 		pr_err("%s-%d: changing queue pair into error state failed: %d\n",
1976 		       ch->sess_name, ch->qp->qp_num, ret);
1977 
1978 	ret = srpt_zerolength_write(ch);
1979 	if (ret < 0) {
1980 		pr_err("%s-%d: queuing zero-length write failed: %d\n",
1981 		       ch->sess_name, ch->qp->qp_num, ret);
1982 		if (srpt_set_ch_state(ch, CH_DISCONNECTED))
1983 			schedule_work(&ch->release_work);
1984 		else
1985 			WARN_ON_ONCE(true);
1986 	}
1987 
1988 	kref_put(&ch->kref, srpt_free_ch);
1989 
1990 	return true;
1991 }
1992 
1993 /*
1994  * Change the channel state into CH_DISCONNECTING. If a channel has not yet
1995  * reached the connected state, close it. If a channel is in the connected
1996  * state, send a DREQ. If a DREQ has been received, send a DREP. Note: it is
1997  * the responsibility of the caller to ensure that this function is not
1998  * invoked concurrently with the code that accepts a connection. This means
1999  * that this function must either be invoked from inside a CM callback
2000  * function or that it must be invoked with the srpt_port.mutex held.
2001  */
2002 static int srpt_disconnect_ch(struct srpt_rdma_ch *ch)
2003 {
2004 	int ret;
2005 
2006 	if (!srpt_set_ch_state(ch, CH_DISCONNECTING))
2007 		return -ENOTCONN;
2008 
2009 	if (ch->using_rdma_cm) {
2010 		ret = rdma_disconnect(ch->rdma_cm.cm_id);
2011 	} else {
2012 		ret = ib_send_cm_dreq(ch->ib_cm.cm_id, NULL, 0);
2013 		if (ret < 0)
2014 			ret = ib_send_cm_drep(ch->ib_cm.cm_id, NULL, 0);
2015 	}
2016 
2017 	if (ret < 0 && srpt_close_ch(ch))
2018 		ret = 0;
2019 
2020 	return ret;
2021 }
2022 
2023 /* Send DREQ and wait for DREP. */
2024 static void srpt_disconnect_ch_sync(struct srpt_rdma_ch *ch)
2025 {
2026 	DECLARE_COMPLETION_ONSTACK(closed);
2027 	struct srpt_port *sport = ch->sport;
2028 
2029 	pr_debug("ch %s-%d state %d\n", ch->sess_name, ch->qp->qp_num,
2030 		 ch->state);
2031 
2032 	ch->closed = &closed;
2033 
2034 	mutex_lock(&sport->mutex);
2035 	srpt_disconnect_ch(ch);
2036 	mutex_unlock(&sport->mutex);
2037 
2038 	while (wait_for_completion_timeout(&closed, 5 * HZ) == 0)
2039 		pr_info("%s(%s-%d state %d): still waiting ...\n", __func__,
2040 			ch->sess_name, ch->qp->qp_num, ch->state);
2041 
2042 }
2043 
2044 static void __srpt_close_all_ch(struct srpt_port *sport)
2045 {
2046 	struct srpt_nexus *nexus;
2047 	struct srpt_rdma_ch *ch;
2048 
2049 	lockdep_assert_held(&sport->mutex);
2050 
2051 	list_for_each_entry(nexus, &sport->nexus_list, entry) {
2052 		list_for_each_entry(ch, &nexus->ch_list, list) {
2053 			if (srpt_disconnect_ch(ch) >= 0)
2054 				pr_info("Closing channel %s-%d because target %s_%d has been disabled\n",
2055 					ch->sess_name, ch->qp->qp_num,
2056 					dev_name(&sport->sdev->device->dev),
2057 					sport->port);
2058 			srpt_close_ch(ch);
2059 		}
2060 	}
2061 }
2062 
2063 /*
2064  * Look up (i_port_id, t_port_id) in sport->nexus_list. Create an entry if
2065  * it does not yet exist.
2066  */
2067 static struct srpt_nexus *srpt_get_nexus(struct srpt_port *sport,
2068 					 const u8 i_port_id[16],
2069 					 const u8 t_port_id[16])
2070 {
2071 	struct srpt_nexus *nexus = NULL, *tmp_nexus = NULL, *n;
2072 
2073 	for (;;) {
2074 		mutex_lock(&sport->mutex);
2075 		list_for_each_entry(n, &sport->nexus_list, entry) {
2076 			if (memcmp(n->i_port_id, i_port_id, 16) == 0 &&
2077 			    memcmp(n->t_port_id, t_port_id, 16) == 0) {
2078 				nexus = n;
2079 				break;
2080 			}
2081 		}
2082 		if (!nexus && tmp_nexus) {
2083 			list_add_tail_rcu(&tmp_nexus->entry,
2084 					  &sport->nexus_list);
2085 			swap(nexus, tmp_nexus);
2086 		}
2087 		mutex_unlock(&sport->mutex);
2088 
2089 		if (nexus)
2090 			break;
2091 		tmp_nexus = kzalloc(sizeof(*nexus), GFP_KERNEL);
2092 		if (!tmp_nexus) {
2093 			nexus = ERR_PTR(-ENOMEM);
2094 			break;
2095 		}
2096 		INIT_LIST_HEAD(&tmp_nexus->ch_list);
2097 		memcpy(tmp_nexus->i_port_id, i_port_id, 16);
2098 		memcpy(tmp_nexus->t_port_id, t_port_id, 16);
2099 	}
2100 
2101 	kfree(tmp_nexus);
2102 
2103 	return nexus;
2104 }
2105 
2106 static void srpt_set_enabled(struct srpt_port *sport, bool enabled)
2107 	__must_hold(&sport->mutex)
2108 {
2109 	lockdep_assert_held(&sport->mutex);
2110 
2111 	if (sport->enabled == enabled)
2112 		return;
2113 	sport->enabled = enabled;
2114 	if (!enabled)
2115 		__srpt_close_all_ch(sport);
2116 }
2117 
2118 static void srpt_drop_sport_ref(struct srpt_port *sport)
2119 {
2120 	if (atomic_dec_return(&sport->refcount) == 0 && sport->freed_channels)
2121 		complete(sport->freed_channels);
2122 }
2123 
2124 static void srpt_free_ch(struct kref *kref)
2125 {
2126 	struct srpt_rdma_ch *ch = container_of(kref, struct srpt_rdma_ch, kref);
2127 
2128 	srpt_drop_sport_ref(ch->sport);
2129 	kfree_rcu(ch, rcu);
2130 }
2131 
2132 /*
2133  * Shut down the SCSI target session, tell the connection manager to
2134  * disconnect the associated RDMA channel, transition the QP to the error
2135  * state and remove the channel from the channel list. This function is
2136  * typically called from inside srpt_zerolength_write_done(). Concurrent
2137  * srpt_zerolength_write() calls from inside srpt_close_ch() are possible
2138  * as long as the channel is on sport->nexus_list.
2139  */
2140 static void srpt_release_channel_work(struct work_struct *w)
2141 {
2142 	struct srpt_rdma_ch *ch;
2143 	struct srpt_device *sdev;
2144 	struct srpt_port *sport;
2145 	struct se_session *se_sess;
2146 
2147 	ch = container_of(w, struct srpt_rdma_ch, release_work);
2148 	pr_debug("%s-%d\n", ch->sess_name, ch->qp->qp_num);
2149 
2150 	sdev = ch->sport->sdev;
2151 	BUG_ON(!sdev);
2152 
2153 	se_sess = ch->sess;
2154 	BUG_ON(!se_sess);
2155 
2156 	target_stop_session(se_sess);
2157 	target_wait_for_sess_cmds(se_sess);
2158 
2159 	target_remove_session(se_sess);
2160 	ch->sess = NULL;
2161 
2162 	if (ch->using_rdma_cm)
2163 		rdma_destroy_id(ch->rdma_cm.cm_id);
2164 	else
2165 		ib_destroy_cm_id(ch->ib_cm.cm_id);
2166 
2167 	sport = ch->sport;
2168 	mutex_lock(&sport->mutex);
2169 	list_del_rcu(&ch->list);
2170 	mutex_unlock(&sport->mutex);
2171 
2172 	if (ch->closed)
2173 		complete(ch->closed);
2174 
2175 	srpt_destroy_ch_ib(ch);
2176 
2177 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
2178 			     ch->sport->sdev, ch->rq_size,
2179 			     ch->rsp_buf_cache, DMA_TO_DEVICE);
2180 
2181 	srpt_cache_put(ch->rsp_buf_cache);
2182 
2183 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_recv_ring,
2184 			     sdev, ch->rq_size,
2185 			     ch->req_buf_cache, DMA_FROM_DEVICE);
2186 
2187 	srpt_cache_put(ch->req_buf_cache);
2188 
2189 	kref_put(&ch->kref, srpt_free_ch);
2190 }
2191 
2192 /**
2193  * srpt_cm_req_recv - process the event IB_CM_REQ_RECEIVED
2194  * @sdev: HCA through which the login request was received.
2195  * @ib_cm_id: IB/CM connection identifier in case of IB/CM.
2196  * @rdma_cm_id: RDMA/CM connection identifier in case of RDMA/CM.
2197  * @port_num: Port through which the REQ message was received.
2198  * @pkey: P_Key of the incoming connection.
2199  * @req: SRP login request.
2200  * @src_addr: GID (IB/CM) or IP address (RDMA/CM) of the port that submitted
2201  * the login request.
2202  *
2203  * Ownership of the cm_id is transferred to the target session if this
2204  * function returns zero. Otherwise the caller remains the owner of cm_id.
2205  */
2206 static int srpt_cm_req_recv(struct srpt_device *const sdev,
2207 			    struct ib_cm_id *ib_cm_id,
2208 			    struct rdma_cm_id *rdma_cm_id,
2209 			    u8 port_num, __be16 pkey,
2210 			    const struct srp_login_req *req,
2211 			    const char *src_addr)
2212 {
2213 	struct srpt_port *sport = &sdev->port[port_num - 1];
2214 	struct srpt_nexus *nexus;
2215 	struct srp_login_rsp *rsp = NULL;
2216 	struct srp_login_rej *rej = NULL;
2217 	union {
2218 		struct rdma_conn_param rdma_cm;
2219 		struct ib_cm_rep_param ib_cm;
2220 	} *rep_param = NULL;
2221 	struct srpt_rdma_ch *ch = NULL;
2222 	char i_port_id[36];
2223 	u32 it_iu_len;
2224 	int i, tag_num, tag_size, ret;
2225 	struct srpt_tpg *stpg;
2226 
2227 	WARN_ON_ONCE(irqs_disabled());
2228 
2229 	it_iu_len = be32_to_cpu(req->req_it_iu_len);
2230 
2231 	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",
2232 		req->initiator_port_id, req->target_port_id, it_iu_len,
2233 		port_num, &sport->gid, be16_to_cpu(pkey));
2234 
2235 	nexus = srpt_get_nexus(sport, req->initiator_port_id,
2236 			       req->target_port_id);
2237 	if (IS_ERR(nexus)) {
2238 		ret = PTR_ERR(nexus);
2239 		goto out;
2240 	}
2241 
2242 	ret = -ENOMEM;
2243 	rsp = kzalloc(sizeof(*rsp), GFP_KERNEL);
2244 	rej = kzalloc(sizeof(*rej), GFP_KERNEL);
2245 	rep_param = kzalloc(sizeof(*rep_param), GFP_KERNEL);
2246 	if (!rsp || !rej || !rep_param)
2247 		goto out;
2248 
2249 	ret = -EINVAL;
2250 	if (it_iu_len > srp_max_req_size || it_iu_len < 64) {
2251 		rej->reason = cpu_to_be32(
2252 				SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE);
2253 		pr_err("rejected SRP_LOGIN_REQ because its length (%d bytes) is out of range (%d .. %d)\n",
2254 		       it_iu_len, 64, srp_max_req_size);
2255 		goto reject;
2256 	}
2257 
2258 	if (!sport->enabled) {
2259 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2260 		pr_info("rejected SRP_LOGIN_REQ because target port %s_%d has not yet been enabled\n",
2261 			dev_name(&sport->sdev->device->dev), port_num);
2262 		goto reject;
2263 	}
2264 
2265 	if (*(__be64 *)req->target_port_id != cpu_to_be64(srpt_service_guid)
2266 	    || *(__be64 *)(req->target_port_id + 8) !=
2267 	       cpu_to_be64(srpt_service_guid)) {
2268 		rej->reason = cpu_to_be32(
2269 				SRP_LOGIN_REJ_UNABLE_ASSOCIATE_CHANNEL);
2270 		pr_err("rejected SRP_LOGIN_REQ because it has an invalid target port identifier.\n");
2271 		goto reject;
2272 	}
2273 
2274 	ret = -ENOMEM;
2275 	ch = kzalloc(sizeof(*ch), GFP_KERNEL);
2276 	if (!ch) {
2277 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2278 		pr_err("rejected SRP_LOGIN_REQ because out of memory.\n");
2279 		goto reject;
2280 	}
2281 
2282 	kref_init(&ch->kref);
2283 	ch->pkey = be16_to_cpu(pkey);
2284 	ch->nexus = nexus;
2285 	ch->zw_cqe.done = srpt_zerolength_write_done;
2286 	INIT_WORK(&ch->release_work, srpt_release_channel_work);
2287 	ch->sport = sport;
2288 	if (rdma_cm_id) {
2289 		ch->using_rdma_cm = true;
2290 		ch->rdma_cm.cm_id = rdma_cm_id;
2291 		rdma_cm_id->context = ch;
2292 	} else {
2293 		ch->ib_cm.cm_id = ib_cm_id;
2294 		ib_cm_id->context = ch;
2295 	}
2296 	/*
2297 	 * ch->rq_size should be at least as large as the initiator queue
2298 	 * depth to avoid that the initiator driver has to report QUEUE_FULL
2299 	 * to the SCSI mid-layer.
2300 	 */
2301 	ch->rq_size = min(MAX_SRPT_RQ_SIZE, sdev->device->attrs.max_qp_wr);
2302 	spin_lock_init(&ch->spinlock);
2303 	ch->state = CH_CONNECTING;
2304 	INIT_LIST_HEAD(&ch->cmd_wait_list);
2305 	ch->max_rsp_size = ch->sport->port_attrib.srp_max_rsp_size;
2306 
2307 	ch->rsp_buf_cache = srpt_cache_get(ch->max_rsp_size);
2308 	if (!ch->rsp_buf_cache)
2309 		goto free_ch;
2310 
2311 	ch->ioctx_ring = (struct srpt_send_ioctx **)
2312 		srpt_alloc_ioctx_ring(ch->sport->sdev, ch->rq_size,
2313 				      sizeof(*ch->ioctx_ring[0]),
2314 				      ch->rsp_buf_cache, 0, DMA_TO_DEVICE);
2315 	if (!ch->ioctx_ring) {
2316 		pr_err("rejected SRP_LOGIN_REQ because creating a new QP SQ ring failed.\n");
2317 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2318 		goto free_rsp_cache;
2319 	}
2320 
2321 	for (i = 0; i < ch->rq_size; i++)
2322 		ch->ioctx_ring[i]->ch = ch;
2323 	if (!sdev->use_srq) {
2324 		u16 imm_data_offset = req->req_flags & SRP_IMMED_REQUESTED ?
2325 			be16_to_cpu(req->imm_data_offset) : 0;
2326 		u16 alignment_offset;
2327 		u32 req_sz;
2328 
2329 		if (req->req_flags & SRP_IMMED_REQUESTED)
2330 			pr_debug("imm_data_offset = %d\n",
2331 				 be16_to_cpu(req->imm_data_offset));
2332 		if (imm_data_offset >= sizeof(struct srp_cmd)) {
2333 			ch->imm_data_offset = imm_data_offset;
2334 			rsp->rsp_flags |= SRP_LOGIN_RSP_IMMED_SUPP;
2335 		} else {
2336 			ch->imm_data_offset = 0;
2337 		}
2338 		alignment_offset = round_up(imm_data_offset, 512) -
2339 			imm_data_offset;
2340 		req_sz = alignment_offset + imm_data_offset + srp_max_req_size;
2341 		ch->req_buf_cache = srpt_cache_get(req_sz);
2342 		if (!ch->req_buf_cache)
2343 			goto free_rsp_ring;
2344 
2345 		ch->ioctx_recv_ring = (struct srpt_recv_ioctx **)
2346 			srpt_alloc_ioctx_ring(ch->sport->sdev, ch->rq_size,
2347 					      sizeof(*ch->ioctx_recv_ring[0]),
2348 					      ch->req_buf_cache,
2349 					      alignment_offset,
2350 					      DMA_FROM_DEVICE);
2351 		if (!ch->ioctx_recv_ring) {
2352 			pr_err("rejected SRP_LOGIN_REQ because creating a new QP RQ ring failed.\n");
2353 			rej->reason =
2354 			    cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2355 			goto free_recv_cache;
2356 		}
2357 		for (i = 0; i < ch->rq_size; i++)
2358 			INIT_LIST_HEAD(&ch->ioctx_recv_ring[i]->wait_list);
2359 	}
2360 
2361 	ret = srpt_create_ch_ib(ch);
2362 	if (ret) {
2363 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2364 		pr_err("rejected SRP_LOGIN_REQ because creating a new RDMA channel failed.\n");
2365 		goto free_recv_ring;
2366 	}
2367 
2368 	strscpy(ch->sess_name, src_addr, sizeof(ch->sess_name));
2369 	snprintf(i_port_id, sizeof(i_port_id), "0x%016llx%016llx",
2370 			be64_to_cpu(*(__be64 *)nexus->i_port_id),
2371 			be64_to_cpu(*(__be64 *)(nexus->i_port_id + 8)));
2372 
2373 	pr_debug("registering src addr %s or i_port_id %s\n", ch->sess_name,
2374 		 i_port_id);
2375 
2376 	tag_num = ch->rq_size;
2377 	tag_size = 1; /* ib_srpt does not use se_sess->sess_cmd_map */
2378 
2379 	if (sport->guid_id) {
2380 		mutex_lock(&sport->guid_id->mutex);
2381 		list_for_each_entry(stpg, &sport->guid_id->tpg_list, entry) {
2382 			if (!IS_ERR_OR_NULL(ch->sess))
2383 				break;
2384 			ch->sess = target_setup_session(&stpg->tpg, tag_num,
2385 						tag_size, TARGET_PROT_NORMAL,
2386 						ch->sess_name, ch, NULL);
2387 		}
2388 		mutex_unlock(&sport->guid_id->mutex);
2389 	}
2390 
2391 	if (sport->gid_id) {
2392 		mutex_lock(&sport->gid_id->mutex);
2393 		list_for_each_entry(stpg, &sport->gid_id->tpg_list, entry) {
2394 			if (!IS_ERR_OR_NULL(ch->sess))
2395 				break;
2396 			ch->sess = target_setup_session(&stpg->tpg, tag_num,
2397 					tag_size, TARGET_PROT_NORMAL, i_port_id,
2398 					ch, NULL);
2399 			if (!IS_ERR_OR_NULL(ch->sess))
2400 				break;
2401 			/* Retry without leading "0x" */
2402 			ch->sess = target_setup_session(&stpg->tpg, tag_num,
2403 						tag_size, TARGET_PROT_NORMAL,
2404 						i_port_id + 2, ch, NULL);
2405 		}
2406 		mutex_unlock(&sport->gid_id->mutex);
2407 	}
2408 
2409 	if (IS_ERR_OR_NULL(ch->sess)) {
2410 		WARN_ON_ONCE(ch->sess == NULL);
2411 		ret = PTR_ERR(ch->sess);
2412 		ch->sess = NULL;
2413 		pr_info("Rejected login for initiator %s: ret = %d.\n",
2414 			ch->sess_name, ret);
2415 		rej->reason = cpu_to_be32(ret == -ENOMEM ?
2416 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES :
2417 				SRP_LOGIN_REJ_CHANNEL_LIMIT_REACHED);
2418 		goto destroy_ib;
2419 	}
2420 
2421 	/*
2422 	 * Once a session has been created destruction of srpt_rdma_ch objects
2423 	 * will decrement sport->refcount. Hence increment sport->refcount now.
2424 	 */
2425 	atomic_inc(&sport->refcount);
2426 
2427 	mutex_lock(&sport->mutex);
2428 
2429 	if ((req->req_flags & SRP_MTCH_ACTION) == SRP_MULTICHAN_SINGLE) {
2430 		struct srpt_rdma_ch *ch2;
2431 
2432 		list_for_each_entry(ch2, &nexus->ch_list, list) {
2433 			if (srpt_disconnect_ch(ch2) < 0)
2434 				continue;
2435 			pr_info("Relogin - closed existing channel %s\n",
2436 				ch2->sess_name);
2437 			rsp->rsp_flags |= SRP_LOGIN_RSP_MULTICHAN_TERMINATED;
2438 		}
2439 	} else {
2440 		rsp->rsp_flags |= SRP_LOGIN_RSP_MULTICHAN_MAINTAINED;
2441 	}
2442 
2443 	list_add_tail_rcu(&ch->list, &nexus->ch_list);
2444 
2445 	if (!sport->enabled) {
2446 		rej->reason = cpu_to_be32(
2447 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2448 		pr_info("rejected SRP_LOGIN_REQ because target %s_%d is not enabled\n",
2449 			dev_name(&sdev->device->dev), port_num);
2450 		mutex_unlock(&sport->mutex);
2451 		ret = -EINVAL;
2452 		goto reject;
2453 	}
2454 
2455 	mutex_unlock(&sport->mutex);
2456 
2457 	ret = ch->using_rdma_cm ? 0 : srpt_ch_qp_rtr(ch, ch->qp);
2458 	if (ret) {
2459 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2460 		pr_err("rejected SRP_LOGIN_REQ because enabling RTR failed (error code = %d)\n",
2461 		       ret);
2462 		goto reject;
2463 	}
2464 
2465 	pr_debug("Establish connection sess=%p name=%s ch=%p\n", ch->sess,
2466 		 ch->sess_name, ch);
2467 
2468 	/* create srp_login_response */
2469 	rsp->opcode = SRP_LOGIN_RSP;
2470 	rsp->tag = req->tag;
2471 	rsp->max_it_iu_len = cpu_to_be32(srp_max_req_size);
2472 	rsp->max_ti_iu_len = req->req_it_iu_len;
2473 	ch->max_ti_iu_len = it_iu_len;
2474 	rsp->buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
2475 				   SRP_BUF_FORMAT_INDIRECT);
2476 	rsp->req_lim_delta = cpu_to_be32(ch->rq_size);
2477 	atomic_set(&ch->req_lim, ch->rq_size);
2478 	atomic_set(&ch->req_lim_delta, 0);
2479 
2480 	/* create cm reply */
2481 	if (ch->using_rdma_cm) {
2482 		rep_param->rdma_cm.private_data = (void *)rsp;
2483 		rep_param->rdma_cm.private_data_len = sizeof(*rsp);
2484 		rep_param->rdma_cm.rnr_retry_count = 7;
2485 		rep_param->rdma_cm.flow_control = 1;
2486 		rep_param->rdma_cm.responder_resources = 4;
2487 		rep_param->rdma_cm.initiator_depth = 4;
2488 	} else {
2489 		rep_param->ib_cm.qp_num = ch->qp->qp_num;
2490 		rep_param->ib_cm.private_data = (void *)rsp;
2491 		rep_param->ib_cm.private_data_len = sizeof(*rsp);
2492 		rep_param->ib_cm.rnr_retry_count = 7;
2493 		rep_param->ib_cm.flow_control = 1;
2494 		rep_param->ib_cm.failover_accepted = 0;
2495 		rep_param->ib_cm.srq = 1;
2496 		rep_param->ib_cm.responder_resources = 4;
2497 		rep_param->ib_cm.initiator_depth = 4;
2498 	}
2499 
2500 	/*
2501 	 * Hold the sport mutex while accepting a connection to avoid that
2502 	 * srpt_disconnect_ch() is invoked concurrently with this code.
2503 	 */
2504 	mutex_lock(&sport->mutex);
2505 	if (sport->enabled && ch->state == CH_CONNECTING) {
2506 		if (ch->using_rdma_cm)
2507 			ret = rdma_accept(rdma_cm_id, &rep_param->rdma_cm);
2508 		else
2509 			ret = ib_send_cm_rep(ib_cm_id, &rep_param->ib_cm);
2510 	} else {
2511 		ret = -EINVAL;
2512 	}
2513 	mutex_unlock(&sport->mutex);
2514 
2515 	switch (ret) {
2516 	case 0:
2517 		break;
2518 	case -EINVAL:
2519 		goto reject;
2520 	default:
2521 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2522 		pr_err("sending SRP_LOGIN_REQ response failed (error code = %d)\n",
2523 		       ret);
2524 		goto reject;
2525 	}
2526 
2527 	goto out;
2528 
2529 destroy_ib:
2530 	srpt_destroy_ch_ib(ch);
2531 
2532 free_recv_ring:
2533 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_recv_ring,
2534 			     ch->sport->sdev, ch->rq_size,
2535 			     ch->req_buf_cache, DMA_FROM_DEVICE);
2536 
2537 free_recv_cache:
2538 	srpt_cache_put(ch->req_buf_cache);
2539 
2540 free_rsp_ring:
2541 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
2542 			     ch->sport->sdev, ch->rq_size,
2543 			     ch->rsp_buf_cache, DMA_TO_DEVICE);
2544 
2545 free_rsp_cache:
2546 	srpt_cache_put(ch->rsp_buf_cache);
2547 
2548 free_ch:
2549 	if (rdma_cm_id)
2550 		rdma_cm_id->context = NULL;
2551 	else
2552 		ib_cm_id->context = NULL;
2553 	kfree(ch);
2554 	ch = NULL;
2555 
2556 	WARN_ON_ONCE(ret == 0);
2557 
2558 reject:
2559 	pr_info("Rejecting login with reason %#x\n", be32_to_cpu(rej->reason));
2560 	rej->opcode = SRP_LOGIN_REJ;
2561 	rej->tag = req->tag;
2562 	rej->buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
2563 				   SRP_BUF_FORMAT_INDIRECT);
2564 
2565 	if (rdma_cm_id)
2566 		rdma_reject(rdma_cm_id, rej, sizeof(*rej),
2567 			    IB_CM_REJ_CONSUMER_DEFINED);
2568 	else
2569 		ib_send_cm_rej(ib_cm_id, IB_CM_REJ_CONSUMER_DEFINED, NULL, 0,
2570 			       rej, sizeof(*rej));
2571 
2572 	if (ch && ch->sess) {
2573 		srpt_close_ch(ch);
2574 		/*
2575 		 * Tell the caller not to free cm_id since
2576 		 * srpt_release_channel_work() will do that.
2577 		 */
2578 		ret = 0;
2579 	}
2580 
2581 out:
2582 	kfree(rep_param);
2583 	kfree(rsp);
2584 	kfree(rej);
2585 
2586 	return ret;
2587 }
2588 
2589 static int srpt_ib_cm_req_recv(struct ib_cm_id *cm_id,
2590 			       const struct ib_cm_req_event_param *param,
2591 			       void *private_data)
2592 {
2593 	char sguid[40];
2594 
2595 	srpt_format_guid(sguid, sizeof(sguid),
2596 			 &param->primary_path->dgid.global.interface_id);
2597 
2598 	return srpt_cm_req_recv(cm_id->context, cm_id, NULL, param->port,
2599 				param->primary_path->pkey,
2600 				private_data, sguid);
2601 }
2602 
2603 static int srpt_rdma_cm_req_recv(struct rdma_cm_id *cm_id,
2604 				 struct rdma_cm_event *event)
2605 {
2606 	struct srpt_device *sdev;
2607 	struct srp_login_req req;
2608 	const struct srp_login_req_rdma *req_rdma;
2609 	struct sa_path_rec *path_rec = cm_id->route.path_rec;
2610 	char src_addr[40];
2611 
2612 	sdev = ib_get_client_data(cm_id->device, &srpt_client);
2613 	if (!sdev)
2614 		return -ECONNREFUSED;
2615 
2616 	if (event->param.conn.private_data_len < sizeof(*req_rdma))
2617 		return -EINVAL;
2618 
2619 	/* Transform srp_login_req_rdma into srp_login_req. */
2620 	req_rdma = event->param.conn.private_data;
2621 	memset(&req, 0, sizeof(req));
2622 	req.opcode		= req_rdma->opcode;
2623 	req.tag			= req_rdma->tag;
2624 	req.req_it_iu_len	= req_rdma->req_it_iu_len;
2625 	req.req_buf_fmt		= req_rdma->req_buf_fmt;
2626 	req.req_flags		= req_rdma->req_flags;
2627 	memcpy(req.initiator_port_id, req_rdma->initiator_port_id, 16);
2628 	memcpy(req.target_port_id, req_rdma->target_port_id, 16);
2629 	req.imm_data_offset	= req_rdma->imm_data_offset;
2630 
2631 	snprintf(src_addr, sizeof(src_addr), "%pIS",
2632 		 &cm_id->route.addr.src_addr);
2633 
2634 	return srpt_cm_req_recv(sdev, NULL, cm_id, cm_id->port_num,
2635 				path_rec ? path_rec->pkey : 0, &req, src_addr);
2636 }
2637 
2638 static void srpt_cm_rej_recv(struct srpt_rdma_ch *ch,
2639 			     enum ib_cm_rej_reason reason,
2640 			     const u8 *private_data,
2641 			     u8 private_data_len)
2642 {
2643 	char *priv = NULL;
2644 	int i;
2645 
2646 	if (private_data_len && (priv = kmalloc(private_data_len * 3 + 1,
2647 						GFP_KERNEL))) {
2648 		for (i = 0; i < private_data_len; i++)
2649 			sprintf(priv + 3 * i, " %02x", private_data[i]);
2650 	}
2651 	pr_info("Received CM REJ for ch %s-%d; reason %d%s%s.\n",
2652 		ch->sess_name, ch->qp->qp_num, reason, private_data_len ?
2653 		"; private data" : "", priv ? priv : " (?)");
2654 	kfree(priv);
2655 }
2656 
2657 /**
2658  * srpt_cm_rtu_recv - process an IB_CM_RTU_RECEIVED or USER_ESTABLISHED event
2659  * @ch: SRPT RDMA channel.
2660  *
2661  * An RTU (ready to use) message indicates that the connection has been
2662  * established and that the recipient may begin transmitting.
2663  */
2664 static void srpt_cm_rtu_recv(struct srpt_rdma_ch *ch)
2665 {
2666 	int ret;
2667 
2668 	ret = ch->using_rdma_cm ? 0 : srpt_ch_qp_rts(ch, ch->qp);
2669 	if (ret < 0) {
2670 		pr_err("%s-%d: QP transition to RTS failed\n", ch->sess_name,
2671 		       ch->qp->qp_num);
2672 		srpt_close_ch(ch);
2673 		return;
2674 	}
2675 
2676 	/*
2677 	 * Note: calling srpt_close_ch() if the transition to the LIVE state
2678 	 * fails is not necessary since that means that that function has
2679 	 * already been invoked from another thread.
2680 	 */
2681 	if (!srpt_set_ch_state(ch, CH_LIVE)) {
2682 		pr_err("%s-%d: channel transition to LIVE state failed\n",
2683 		       ch->sess_name, ch->qp->qp_num);
2684 		return;
2685 	}
2686 
2687 	/* Trigger wait list processing. */
2688 	ret = srpt_zerolength_write(ch);
2689 	WARN_ONCE(ret < 0, "%d\n", ret);
2690 }
2691 
2692 /**
2693  * srpt_cm_handler - IB connection manager callback function
2694  * @cm_id: IB/CM connection identifier.
2695  * @event: IB/CM event.
2696  *
2697  * A non-zero return value will cause the caller destroy the CM ID.
2698  *
2699  * Note: srpt_cm_handler() must only return a non-zero value when transferring
2700  * ownership of the cm_id to a channel by srpt_cm_req_recv() failed. Returning
2701  * a non-zero value in any other case will trigger a race with the
2702  * ib_destroy_cm_id() call in srpt_release_channel().
2703  */
2704 static int srpt_cm_handler(struct ib_cm_id *cm_id,
2705 			   const struct ib_cm_event *event)
2706 {
2707 	struct srpt_rdma_ch *ch = cm_id->context;
2708 	int ret;
2709 
2710 	ret = 0;
2711 	switch (event->event) {
2712 	case IB_CM_REQ_RECEIVED:
2713 		ret = srpt_ib_cm_req_recv(cm_id, &event->param.req_rcvd,
2714 					  event->private_data);
2715 		break;
2716 	case IB_CM_REJ_RECEIVED:
2717 		srpt_cm_rej_recv(ch, event->param.rej_rcvd.reason,
2718 				 event->private_data,
2719 				 IB_CM_REJ_PRIVATE_DATA_SIZE);
2720 		break;
2721 	case IB_CM_RTU_RECEIVED:
2722 	case IB_CM_USER_ESTABLISHED:
2723 		srpt_cm_rtu_recv(ch);
2724 		break;
2725 	case IB_CM_DREQ_RECEIVED:
2726 		srpt_disconnect_ch(ch);
2727 		break;
2728 	case IB_CM_DREP_RECEIVED:
2729 		pr_info("Received CM DREP message for ch %s-%d.\n",
2730 			ch->sess_name, ch->qp->qp_num);
2731 		srpt_close_ch(ch);
2732 		break;
2733 	case IB_CM_TIMEWAIT_EXIT:
2734 		pr_info("Received CM TimeWait exit for ch %s-%d.\n",
2735 			ch->sess_name, ch->qp->qp_num);
2736 		srpt_close_ch(ch);
2737 		break;
2738 	case IB_CM_REP_ERROR:
2739 		pr_info("Received CM REP error for ch %s-%d.\n", ch->sess_name,
2740 			ch->qp->qp_num);
2741 		break;
2742 	case IB_CM_DREQ_ERROR:
2743 		pr_info("Received CM DREQ ERROR event.\n");
2744 		break;
2745 	case IB_CM_MRA_RECEIVED:
2746 		pr_info("Received CM MRA event\n");
2747 		break;
2748 	default:
2749 		pr_err("received unrecognized CM event %d\n", event->event);
2750 		break;
2751 	}
2752 
2753 	return ret;
2754 }
2755 
2756 static int srpt_rdma_cm_handler(struct rdma_cm_id *cm_id,
2757 				struct rdma_cm_event *event)
2758 {
2759 	struct srpt_rdma_ch *ch = cm_id->context;
2760 	int ret = 0;
2761 
2762 	switch (event->event) {
2763 	case RDMA_CM_EVENT_CONNECT_REQUEST:
2764 		ret = srpt_rdma_cm_req_recv(cm_id, event);
2765 		break;
2766 	case RDMA_CM_EVENT_REJECTED:
2767 		srpt_cm_rej_recv(ch, event->status,
2768 				 event->param.conn.private_data,
2769 				 event->param.conn.private_data_len);
2770 		break;
2771 	case RDMA_CM_EVENT_ESTABLISHED:
2772 		srpt_cm_rtu_recv(ch);
2773 		break;
2774 	case RDMA_CM_EVENT_DISCONNECTED:
2775 		if (ch->state < CH_DISCONNECTING)
2776 			srpt_disconnect_ch(ch);
2777 		else
2778 			srpt_close_ch(ch);
2779 		break;
2780 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
2781 		srpt_close_ch(ch);
2782 		break;
2783 	case RDMA_CM_EVENT_UNREACHABLE:
2784 		pr_info("Received CM REP error for ch %s-%d.\n", ch->sess_name,
2785 			ch->qp->qp_num);
2786 		break;
2787 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
2788 	case RDMA_CM_EVENT_ADDR_CHANGE:
2789 		break;
2790 	default:
2791 		pr_err("received unrecognized RDMA CM event %d\n",
2792 		       event->event);
2793 		break;
2794 	}
2795 
2796 	return ret;
2797 }
2798 
2799 /*
2800  * srpt_write_pending - Start data transfer from initiator to target (write).
2801  */
2802 static int srpt_write_pending(struct se_cmd *se_cmd)
2803 {
2804 	struct srpt_send_ioctx *ioctx =
2805 		container_of(se_cmd, struct srpt_send_ioctx, cmd);
2806 	struct srpt_rdma_ch *ch = ioctx->ch;
2807 	struct ib_send_wr *first_wr = NULL;
2808 	struct ib_cqe *cqe = &ioctx->rdma_cqe;
2809 	enum srpt_command_state new_state;
2810 	int ret, i;
2811 
2812 	if (ioctx->recv_ioctx) {
2813 		srpt_set_cmd_state(ioctx, SRPT_STATE_DATA_IN);
2814 		target_execute_cmd(&ioctx->cmd);
2815 		return 0;
2816 	}
2817 
2818 	new_state = srpt_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA);
2819 	WARN_ON(new_state == SRPT_STATE_DONE);
2820 
2821 	if (atomic_sub_return(ioctx->n_rdma, &ch->sq_wr_avail) < 0) {
2822 		pr_warn("%s: IB send queue full (needed %d)\n",
2823 				__func__, ioctx->n_rdma);
2824 		ret = -ENOMEM;
2825 		goto out_undo;
2826 	}
2827 
2828 	cqe->done = srpt_rdma_read_done;
2829 	for (i = ioctx->n_rw_ctx - 1; i >= 0; i--) {
2830 		struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i];
2831 
2832 		first_wr = rdma_rw_ctx_wrs(&ctx->rw, ch->qp, ch->sport->port,
2833 				cqe, first_wr);
2834 		cqe = NULL;
2835 	}
2836 
2837 	ret = ib_post_send(ch->qp, first_wr, NULL);
2838 	if (ret) {
2839 		pr_err("%s: ib_post_send() returned %d for %d (avail: %d)\n",
2840 			 __func__, ret, ioctx->n_rdma,
2841 			 atomic_read(&ch->sq_wr_avail));
2842 		goto out_undo;
2843 	}
2844 
2845 	return 0;
2846 out_undo:
2847 	atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
2848 	return ret;
2849 }
2850 
2851 static u8 tcm_to_srp_tsk_mgmt_status(const int tcm_mgmt_status)
2852 {
2853 	switch (tcm_mgmt_status) {
2854 	case TMR_FUNCTION_COMPLETE:
2855 		return SRP_TSK_MGMT_SUCCESS;
2856 	case TMR_FUNCTION_REJECTED:
2857 		return SRP_TSK_MGMT_FUNC_NOT_SUPP;
2858 	}
2859 	return SRP_TSK_MGMT_FAILED;
2860 }
2861 
2862 /**
2863  * srpt_queue_response - transmit the response to a SCSI command
2864  * @cmd: SCSI target command.
2865  *
2866  * Callback function called by the TCM core. Must not block since it can be
2867  * invoked on the context of the IB completion handler.
2868  */
2869 static void srpt_queue_response(struct se_cmd *cmd)
2870 {
2871 	struct srpt_send_ioctx *ioctx =
2872 		container_of(cmd, struct srpt_send_ioctx, cmd);
2873 	struct srpt_rdma_ch *ch = ioctx->ch;
2874 	struct srpt_device *sdev = ch->sport->sdev;
2875 	struct ib_send_wr send_wr, *first_wr = &send_wr;
2876 	struct ib_sge sge;
2877 	enum srpt_command_state state;
2878 	int resp_len, ret, i;
2879 	u8 srp_tm_status;
2880 
2881 	state = ioctx->state;
2882 	switch (state) {
2883 	case SRPT_STATE_NEW:
2884 	case SRPT_STATE_DATA_IN:
2885 		ioctx->state = SRPT_STATE_CMD_RSP_SENT;
2886 		break;
2887 	case SRPT_STATE_MGMT:
2888 		ioctx->state = SRPT_STATE_MGMT_RSP_SENT;
2889 		break;
2890 	default:
2891 		WARN(true, "ch %p; cmd %d: unexpected command state %d\n",
2892 			ch, ioctx->ioctx.index, ioctx->state);
2893 		break;
2894 	}
2895 
2896 	if (WARN_ON_ONCE(state == SRPT_STATE_CMD_RSP_SENT))
2897 		return;
2898 
2899 	/* For read commands, transfer the data to the initiator. */
2900 	if (ioctx->cmd.data_direction == DMA_FROM_DEVICE &&
2901 	    ioctx->cmd.data_length &&
2902 	    !ioctx->queue_status_only) {
2903 		for (i = ioctx->n_rw_ctx - 1; i >= 0; i--) {
2904 			struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i];
2905 
2906 			first_wr = rdma_rw_ctx_wrs(&ctx->rw, ch->qp,
2907 					ch->sport->port, NULL, first_wr);
2908 		}
2909 	}
2910 
2911 	if (state != SRPT_STATE_MGMT)
2912 		resp_len = srpt_build_cmd_rsp(ch, ioctx, ioctx->cmd.tag,
2913 					      cmd->scsi_status);
2914 	else {
2915 		srp_tm_status
2916 			= tcm_to_srp_tsk_mgmt_status(cmd->se_tmr_req->response);
2917 		resp_len = srpt_build_tskmgmt_rsp(ch, ioctx, srp_tm_status,
2918 						 ioctx->cmd.tag);
2919 	}
2920 
2921 	atomic_inc(&ch->req_lim);
2922 
2923 	if (unlikely(atomic_sub_return(1 + ioctx->n_rdma,
2924 			&ch->sq_wr_avail) < 0)) {
2925 		pr_warn("%s: IB send queue full (needed %d)\n",
2926 				__func__, ioctx->n_rdma);
2927 		goto out;
2928 	}
2929 
2930 	ib_dma_sync_single_for_device(sdev->device, ioctx->ioctx.dma, resp_len,
2931 				      DMA_TO_DEVICE);
2932 
2933 	sge.addr = ioctx->ioctx.dma;
2934 	sge.length = resp_len;
2935 	sge.lkey = sdev->lkey;
2936 
2937 	ioctx->ioctx.cqe.done = srpt_send_done;
2938 	send_wr.next = NULL;
2939 	send_wr.wr_cqe = &ioctx->ioctx.cqe;
2940 	send_wr.sg_list = &sge;
2941 	send_wr.num_sge = 1;
2942 	send_wr.opcode = IB_WR_SEND;
2943 	send_wr.send_flags = IB_SEND_SIGNALED;
2944 
2945 	ret = ib_post_send(ch->qp, first_wr, NULL);
2946 	if (ret < 0) {
2947 		pr_err("%s: sending cmd response failed for tag %llu (%d)\n",
2948 			__func__, ioctx->cmd.tag, ret);
2949 		goto out;
2950 	}
2951 
2952 	return;
2953 
2954 out:
2955 	atomic_add(1 + ioctx->n_rdma, &ch->sq_wr_avail);
2956 	atomic_dec(&ch->req_lim);
2957 	srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
2958 	target_put_sess_cmd(&ioctx->cmd);
2959 }
2960 
2961 static int srpt_queue_data_in(struct se_cmd *cmd)
2962 {
2963 	srpt_queue_response(cmd);
2964 	return 0;
2965 }
2966 
2967 static void srpt_queue_tm_rsp(struct se_cmd *cmd)
2968 {
2969 	srpt_queue_response(cmd);
2970 }
2971 
2972 /*
2973  * This function is called for aborted commands if no response is sent to the
2974  * initiator. Make sure that the credits freed by aborting a command are
2975  * returned to the initiator the next time a response is sent by incrementing
2976  * ch->req_lim_delta.
2977  */
2978 static void srpt_aborted_task(struct se_cmd *cmd)
2979 {
2980 	struct srpt_send_ioctx *ioctx = container_of(cmd,
2981 				struct srpt_send_ioctx, cmd);
2982 	struct srpt_rdma_ch *ch = ioctx->ch;
2983 
2984 	atomic_inc(&ch->req_lim_delta);
2985 }
2986 
2987 static int srpt_queue_status(struct se_cmd *cmd)
2988 {
2989 	struct srpt_send_ioctx *ioctx;
2990 
2991 	ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
2992 	BUG_ON(ioctx->sense_data != cmd->sense_buffer);
2993 	if (cmd->se_cmd_flags &
2994 	    (SCF_TRANSPORT_TASK_SENSE | SCF_EMULATED_TASK_SENSE))
2995 		WARN_ON(cmd->scsi_status != SAM_STAT_CHECK_CONDITION);
2996 	ioctx->queue_status_only = true;
2997 	srpt_queue_response(cmd);
2998 	return 0;
2999 }
3000 
3001 static void srpt_refresh_port_work(struct work_struct *work)
3002 {
3003 	struct srpt_port *sport = container_of(work, struct srpt_port, work);
3004 
3005 	srpt_refresh_port(sport);
3006 }
3007 
3008 /**
3009  * srpt_release_sport - disable login and wait for associated channels
3010  * @sport: SRPT HCA port.
3011  */
3012 static int srpt_release_sport(struct srpt_port *sport)
3013 {
3014 	DECLARE_COMPLETION_ONSTACK(c);
3015 	struct srpt_nexus *nexus, *next_n;
3016 	struct srpt_rdma_ch *ch;
3017 
3018 	WARN_ON_ONCE(irqs_disabled());
3019 
3020 	sport->freed_channels = &c;
3021 
3022 	mutex_lock(&sport->mutex);
3023 	srpt_set_enabled(sport, false);
3024 	mutex_unlock(&sport->mutex);
3025 
3026 	while (atomic_read(&sport->refcount) > 0 &&
3027 	       wait_for_completion_timeout(&c, 5 * HZ) <= 0) {
3028 		pr_info("%s_%d: waiting for unregistration of %d sessions ...\n",
3029 			dev_name(&sport->sdev->device->dev), sport->port,
3030 			atomic_read(&sport->refcount));
3031 		rcu_read_lock();
3032 		list_for_each_entry(nexus, &sport->nexus_list, entry) {
3033 			list_for_each_entry(ch, &nexus->ch_list, list) {
3034 				pr_info("%s-%d: state %s\n",
3035 					ch->sess_name, ch->qp->qp_num,
3036 					get_ch_state_name(ch->state));
3037 			}
3038 		}
3039 		rcu_read_unlock();
3040 	}
3041 
3042 	mutex_lock(&sport->mutex);
3043 	list_for_each_entry_safe(nexus, next_n, &sport->nexus_list, entry) {
3044 		list_del(&nexus->entry);
3045 		kfree_rcu(nexus, rcu);
3046 	}
3047 	mutex_unlock(&sport->mutex);
3048 
3049 	return 0;
3050 }
3051 
3052 struct port_and_port_id {
3053 	struct srpt_port *sport;
3054 	struct srpt_port_id **port_id;
3055 };
3056 
3057 static struct port_and_port_id __srpt_lookup_port(const char *name)
3058 {
3059 	struct ib_device *dev;
3060 	struct srpt_device *sdev;
3061 	struct srpt_port *sport;
3062 	int i;
3063 
3064 	list_for_each_entry(sdev, &srpt_dev_list, list) {
3065 		dev = sdev->device;
3066 		if (!dev)
3067 			continue;
3068 
3069 		for (i = 0; i < dev->phys_port_cnt; i++) {
3070 			sport = &sdev->port[i];
3071 
3072 			if (strcmp(sport->guid_name, name) == 0) {
3073 				kref_get(&sdev->refcnt);
3074 				return (struct port_and_port_id){
3075 					sport, &sport->guid_id};
3076 			}
3077 			if (strcmp(sport->gid_name, name) == 0) {
3078 				kref_get(&sdev->refcnt);
3079 				return (struct port_and_port_id){
3080 					sport, &sport->gid_id};
3081 			}
3082 		}
3083 	}
3084 
3085 	return (struct port_and_port_id){};
3086 }
3087 
3088 /**
3089  * srpt_lookup_port() - Look up an RDMA port by name
3090  * @name: ASCII port name
3091  *
3092  * Increments the RDMA port reference count if an RDMA port pointer is returned.
3093  * The caller must drop that reference count by calling srpt_port_put_ref().
3094  */
3095 static struct port_and_port_id srpt_lookup_port(const char *name)
3096 {
3097 	struct port_and_port_id papi;
3098 
3099 	spin_lock(&srpt_dev_lock);
3100 	papi = __srpt_lookup_port(name);
3101 	spin_unlock(&srpt_dev_lock);
3102 
3103 	return papi;
3104 }
3105 
3106 static void srpt_free_srq(struct srpt_device *sdev)
3107 {
3108 	if (!sdev->srq)
3109 		return;
3110 
3111 	ib_destroy_srq(sdev->srq);
3112 	srpt_free_ioctx_ring((struct srpt_ioctx **)sdev->ioctx_ring, sdev,
3113 			     sdev->srq_size, sdev->req_buf_cache,
3114 			     DMA_FROM_DEVICE);
3115 	srpt_cache_put(sdev->req_buf_cache);
3116 	sdev->srq = NULL;
3117 }
3118 
3119 static int srpt_alloc_srq(struct srpt_device *sdev)
3120 {
3121 	struct ib_srq_init_attr srq_attr = {
3122 		.event_handler = srpt_srq_event,
3123 		.srq_context = (void *)sdev,
3124 		.attr.max_wr = sdev->srq_size,
3125 		.attr.max_sge = 1,
3126 		.srq_type = IB_SRQT_BASIC,
3127 	};
3128 	struct ib_device *device = sdev->device;
3129 	struct ib_srq *srq;
3130 	int i;
3131 
3132 	WARN_ON_ONCE(sdev->srq);
3133 	srq = ib_create_srq(sdev->pd, &srq_attr);
3134 	if (IS_ERR(srq)) {
3135 		pr_debug("ib_create_srq() failed: %ld\n", PTR_ERR(srq));
3136 		return PTR_ERR(srq);
3137 	}
3138 
3139 	pr_debug("create SRQ #wr= %d max_allow=%d dev= %s\n", sdev->srq_size,
3140 		 sdev->device->attrs.max_srq_wr, dev_name(&device->dev));
3141 
3142 	sdev->req_buf_cache = srpt_cache_get(srp_max_req_size);
3143 	if (!sdev->req_buf_cache)
3144 		goto free_srq;
3145 
3146 	sdev->ioctx_ring = (struct srpt_recv_ioctx **)
3147 		srpt_alloc_ioctx_ring(sdev, sdev->srq_size,
3148 				      sizeof(*sdev->ioctx_ring[0]),
3149 				      sdev->req_buf_cache, 0, DMA_FROM_DEVICE);
3150 	if (!sdev->ioctx_ring)
3151 		goto free_cache;
3152 
3153 	sdev->use_srq = true;
3154 	sdev->srq = srq;
3155 
3156 	for (i = 0; i < sdev->srq_size; ++i) {
3157 		INIT_LIST_HEAD(&sdev->ioctx_ring[i]->wait_list);
3158 		srpt_post_recv(sdev, NULL, sdev->ioctx_ring[i]);
3159 	}
3160 
3161 	return 0;
3162 
3163 free_cache:
3164 	srpt_cache_put(sdev->req_buf_cache);
3165 
3166 free_srq:
3167 	ib_destroy_srq(srq);
3168 	return -ENOMEM;
3169 }
3170 
3171 static int srpt_use_srq(struct srpt_device *sdev, bool use_srq)
3172 {
3173 	struct ib_device *device = sdev->device;
3174 	int ret = 0;
3175 
3176 	if (!use_srq) {
3177 		srpt_free_srq(sdev);
3178 		sdev->use_srq = false;
3179 	} else if (use_srq && !sdev->srq) {
3180 		ret = srpt_alloc_srq(sdev);
3181 	}
3182 	pr_debug("%s(%s): use_srq = %d; ret = %d\n", __func__,
3183 		 dev_name(&device->dev), sdev->use_srq, ret);
3184 	return ret;
3185 }
3186 
3187 static void srpt_free_sdev(struct kref *refcnt)
3188 {
3189 	struct srpt_device *sdev = container_of(refcnt, typeof(*sdev), refcnt);
3190 
3191 	kfree(sdev);
3192 }
3193 
3194 static void srpt_sdev_put(struct srpt_device *sdev)
3195 {
3196 	kref_put(&sdev->refcnt, srpt_free_sdev);
3197 }
3198 
3199 /**
3200  * srpt_add_one - InfiniBand device addition callback function
3201  * @device: Describes a HCA.
3202  */
3203 static int srpt_add_one(struct ib_device *device)
3204 {
3205 	struct srpt_device *sdev;
3206 	struct srpt_port *sport;
3207 	int ret;
3208 	u32 i;
3209 
3210 	pr_debug("device = %p\n", device);
3211 
3212 	sdev = kzalloc(struct_size(sdev, port, device->phys_port_cnt),
3213 		       GFP_KERNEL);
3214 	if (!sdev)
3215 		return -ENOMEM;
3216 
3217 	kref_init(&sdev->refcnt);
3218 	sdev->device = device;
3219 	mutex_init(&sdev->sdev_mutex);
3220 
3221 	sdev->pd = ib_alloc_pd(device, 0);
3222 	if (IS_ERR(sdev->pd)) {
3223 		ret = PTR_ERR(sdev->pd);
3224 		goto free_dev;
3225 	}
3226 
3227 	sdev->lkey = sdev->pd->local_dma_lkey;
3228 
3229 	sdev->srq_size = min(srpt_srq_size, sdev->device->attrs.max_srq_wr);
3230 
3231 	srpt_use_srq(sdev, sdev->port[0].port_attrib.use_srq);
3232 
3233 	if (!srpt_service_guid)
3234 		srpt_service_guid = be64_to_cpu(device->node_guid);
3235 
3236 	if (rdma_port_get_link_layer(device, 1) == IB_LINK_LAYER_INFINIBAND)
3237 		sdev->cm_id = ib_create_cm_id(device, srpt_cm_handler, sdev);
3238 	if (IS_ERR(sdev->cm_id)) {
3239 		pr_info("ib_create_cm_id() failed: %ld\n",
3240 			PTR_ERR(sdev->cm_id));
3241 		ret = PTR_ERR(sdev->cm_id);
3242 		sdev->cm_id = NULL;
3243 		if (!rdma_cm_id)
3244 			goto err_ring;
3245 	}
3246 
3247 	/* print out target login information */
3248 	pr_debug("Target login info: id_ext=%016llx,ioc_guid=%016llx,pkey=ffff,service_id=%016llx\n",
3249 		 srpt_service_guid, srpt_service_guid, srpt_service_guid);
3250 
3251 	/*
3252 	 * We do not have a consistent service_id (ie. also id_ext of target_id)
3253 	 * to identify this target. We currently use the guid of the first HCA
3254 	 * in the system as service_id; therefore, the target_id will change
3255 	 * if this HCA is gone bad and replaced by different HCA
3256 	 */
3257 	ret = sdev->cm_id ?
3258 		ib_cm_listen(sdev->cm_id, cpu_to_be64(srpt_service_guid)) :
3259 		0;
3260 	if (ret < 0) {
3261 		pr_err("ib_cm_listen() failed: %d (cm_id state = %d)\n", ret,
3262 		       sdev->cm_id->state);
3263 		goto err_cm;
3264 	}
3265 
3266 	INIT_IB_EVENT_HANDLER(&sdev->event_handler, sdev->device,
3267 			      srpt_event_handler);
3268 
3269 	for (i = 1; i <= sdev->device->phys_port_cnt; i++) {
3270 		sport = &sdev->port[i - 1];
3271 		INIT_LIST_HEAD(&sport->nexus_list);
3272 		mutex_init(&sport->mutex);
3273 		sport->sdev = sdev;
3274 		sport->port = i;
3275 		sport->port_attrib.srp_max_rdma_size = DEFAULT_MAX_RDMA_SIZE;
3276 		sport->port_attrib.srp_max_rsp_size = DEFAULT_MAX_RSP_SIZE;
3277 		sport->port_attrib.srp_sq_size = DEF_SRPT_SQ_SIZE;
3278 		sport->port_attrib.use_srq = false;
3279 		INIT_WORK(&sport->work, srpt_refresh_port_work);
3280 
3281 		ret = srpt_refresh_port(sport);
3282 		if (ret) {
3283 			pr_err("MAD registration failed for %s-%d.\n",
3284 			       dev_name(&sdev->device->dev), i);
3285 			i--;
3286 			goto err_port;
3287 		}
3288 	}
3289 
3290 	ib_register_event_handler(&sdev->event_handler);
3291 	spin_lock(&srpt_dev_lock);
3292 	list_add_tail(&sdev->list, &srpt_dev_list);
3293 	spin_unlock(&srpt_dev_lock);
3294 
3295 	ib_set_client_data(device, &srpt_client, sdev);
3296 	pr_debug("added %s.\n", dev_name(&device->dev));
3297 	return 0;
3298 
3299 err_port:
3300 	srpt_unregister_mad_agent(sdev, i);
3301 err_cm:
3302 	if (sdev->cm_id)
3303 		ib_destroy_cm_id(sdev->cm_id);
3304 err_ring:
3305 	srpt_free_srq(sdev);
3306 	ib_dealloc_pd(sdev->pd);
3307 free_dev:
3308 	srpt_sdev_put(sdev);
3309 	pr_info("%s(%s) failed.\n", __func__, dev_name(&device->dev));
3310 	return ret;
3311 }
3312 
3313 /**
3314  * srpt_remove_one - InfiniBand device removal callback function
3315  * @device: Describes a HCA.
3316  * @client_data: The value passed as the third argument to ib_set_client_data().
3317  */
3318 static void srpt_remove_one(struct ib_device *device, void *client_data)
3319 {
3320 	struct srpt_device *sdev = client_data;
3321 	int i;
3322 
3323 	srpt_unregister_mad_agent(sdev, sdev->device->phys_port_cnt);
3324 
3325 	ib_unregister_event_handler(&sdev->event_handler);
3326 
3327 	/* Cancel any work queued by the just unregistered IB event handler. */
3328 	for (i = 0; i < sdev->device->phys_port_cnt; i++)
3329 		cancel_work_sync(&sdev->port[i].work);
3330 
3331 	if (sdev->cm_id)
3332 		ib_destroy_cm_id(sdev->cm_id);
3333 
3334 	ib_set_client_data(device, &srpt_client, NULL);
3335 
3336 	/*
3337 	 * Unregistering a target must happen after destroying sdev->cm_id
3338 	 * such that no new SRP_LOGIN_REQ information units can arrive while
3339 	 * destroying the target.
3340 	 */
3341 	spin_lock(&srpt_dev_lock);
3342 	list_del(&sdev->list);
3343 	spin_unlock(&srpt_dev_lock);
3344 
3345 	for (i = 0; i < sdev->device->phys_port_cnt; i++)
3346 		srpt_release_sport(&sdev->port[i]);
3347 
3348 	srpt_free_srq(sdev);
3349 
3350 	ib_dealloc_pd(sdev->pd);
3351 
3352 	srpt_sdev_put(sdev);
3353 }
3354 
3355 static struct ib_client srpt_client = {
3356 	.name = DRV_NAME,
3357 	.add = srpt_add_one,
3358 	.remove = srpt_remove_one
3359 };
3360 
3361 static int srpt_check_true(struct se_portal_group *se_tpg)
3362 {
3363 	return 1;
3364 }
3365 
3366 static struct srpt_port *srpt_tpg_to_sport(struct se_portal_group *tpg)
3367 {
3368 	return tpg->se_tpg_wwn->priv;
3369 }
3370 
3371 static struct srpt_port_id *srpt_wwn_to_sport_id(struct se_wwn *wwn)
3372 {
3373 	struct srpt_port *sport = wwn->priv;
3374 
3375 	if (sport->guid_id && &sport->guid_id->wwn == wwn)
3376 		return sport->guid_id;
3377 	if (sport->gid_id && &sport->gid_id->wwn == wwn)
3378 		return sport->gid_id;
3379 	WARN_ON_ONCE(true);
3380 	return NULL;
3381 }
3382 
3383 static char *srpt_get_fabric_wwn(struct se_portal_group *tpg)
3384 {
3385 	struct srpt_tpg *stpg = container_of(tpg, typeof(*stpg), tpg);
3386 
3387 	return stpg->sport_id->name;
3388 }
3389 
3390 static u16 srpt_get_tag(struct se_portal_group *tpg)
3391 {
3392 	return 1;
3393 }
3394 
3395 static void srpt_release_cmd(struct se_cmd *se_cmd)
3396 {
3397 	struct srpt_send_ioctx *ioctx = container_of(se_cmd,
3398 				struct srpt_send_ioctx, cmd);
3399 	struct srpt_rdma_ch *ch = ioctx->ch;
3400 	struct srpt_recv_ioctx *recv_ioctx = ioctx->recv_ioctx;
3401 
3402 	WARN_ON_ONCE(ioctx->state != SRPT_STATE_DONE &&
3403 		     !(ioctx->cmd.transport_state & CMD_T_ABORTED));
3404 
3405 	if (recv_ioctx) {
3406 		WARN_ON_ONCE(!list_empty(&recv_ioctx->wait_list));
3407 		ioctx->recv_ioctx = NULL;
3408 		srpt_post_recv(ch->sport->sdev, ch, recv_ioctx);
3409 	}
3410 
3411 	if (ioctx->n_rw_ctx) {
3412 		srpt_free_rw_ctxs(ch, ioctx);
3413 		ioctx->n_rw_ctx = 0;
3414 	}
3415 
3416 	target_free_tag(se_cmd->se_sess, se_cmd);
3417 }
3418 
3419 /**
3420  * srpt_close_session - forcibly close a session
3421  * @se_sess: SCSI target session.
3422  *
3423  * Callback function invoked by the TCM core to clean up sessions associated
3424  * with a node ACL when the user invokes
3425  * rmdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3426  */
3427 static void srpt_close_session(struct se_session *se_sess)
3428 {
3429 	struct srpt_rdma_ch *ch = se_sess->fabric_sess_ptr;
3430 
3431 	srpt_disconnect_ch_sync(ch);
3432 }
3433 
3434 /* Note: only used from inside debug printk's by the TCM core. */
3435 static int srpt_get_tcm_cmd_state(struct se_cmd *se_cmd)
3436 {
3437 	struct srpt_send_ioctx *ioctx;
3438 
3439 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
3440 	return ioctx->state;
3441 }
3442 
3443 static int srpt_parse_guid(u64 *guid, const char *name)
3444 {
3445 	u16 w[4];
3446 	int ret = -EINVAL;
3447 
3448 	if (sscanf(name, "%hx:%hx:%hx:%hx", &w[0], &w[1], &w[2], &w[3]) != 4)
3449 		goto out;
3450 	*guid = get_unaligned_be64(w);
3451 	ret = 0;
3452 out:
3453 	return ret;
3454 }
3455 
3456 /**
3457  * srpt_parse_i_port_id - parse an initiator port ID
3458  * @name: ASCII representation of a 128-bit initiator port ID.
3459  * @i_port_id: Binary 128-bit port ID.
3460  */
3461 static int srpt_parse_i_port_id(u8 i_port_id[16], const char *name)
3462 {
3463 	const char *p;
3464 	unsigned len, count, leading_zero_bytes;
3465 	int ret;
3466 
3467 	p = name;
3468 	if (strncasecmp(p, "0x", 2) == 0)
3469 		p += 2;
3470 	ret = -EINVAL;
3471 	len = strlen(p);
3472 	if (len % 2)
3473 		goto out;
3474 	count = min(len / 2, 16U);
3475 	leading_zero_bytes = 16 - count;
3476 	memset(i_port_id, 0, leading_zero_bytes);
3477 	ret = hex2bin(i_port_id + leading_zero_bytes, p, count);
3478 
3479 out:
3480 	return ret;
3481 }
3482 
3483 /*
3484  * configfs callback function invoked for mkdir
3485  * /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3486  *
3487  * i_port_id must be an initiator port GUID, GID or IP address. See also the
3488  * target_alloc_session() calls in this driver. Examples of valid initiator
3489  * port IDs:
3490  * 0x0000000000000000505400fffe4a0b7b
3491  * 0000000000000000505400fffe4a0b7b
3492  * 5054:00ff:fe4a:0b7b
3493  * 192.168.122.76
3494  */
3495 static int srpt_init_nodeacl(struct se_node_acl *se_nacl, const char *name)
3496 {
3497 	struct sockaddr_storage sa;
3498 	u64 guid;
3499 	u8 i_port_id[16];
3500 	int ret;
3501 
3502 	ret = srpt_parse_guid(&guid, name);
3503 	if (ret < 0)
3504 		ret = srpt_parse_i_port_id(i_port_id, name);
3505 	if (ret < 0)
3506 		ret = inet_pton_with_scope(&init_net, AF_UNSPEC, name, NULL,
3507 					   &sa);
3508 	if (ret < 0)
3509 		pr_err("invalid initiator port ID %s\n", name);
3510 	return ret;
3511 }
3512 
3513 static ssize_t srpt_tpg_attrib_srp_max_rdma_size_show(struct config_item *item,
3514 		char *page)
3515 {
3516 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3517 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3518 
3519 	return sysfs_emit(page, "%u\n", sport->port_attrib.srp_max_rdma_size);
3520 }
3521 
3522 static ssize_t srpt_tpg_attrib_srp_max_rdma_size_store(struct config_item *item,
3523 		const char *page, size_t count)
3524 {
3525 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3526 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3527 	unsigned long val;
3528 	int ret;
3529 
3530 	ret = kstrtoul(page, 0, &val);
3531 	if (ret < 0) {
3532 		pr_err("kstrtoul() failed with ret: %d\n", ret);
3533 		return -EINVAL;
3534 	}
3535 	if (val > MAX_SRPT_RDMA_SIZE) {
3536 		pr_err("val: %lu exceeds MAX_SRPT_RDMA_SIZE: %d\n", val,
3537 			MAX_SRPT_RDMA_SIZE);
3538 		return -EINVAL;
3539 	}
3540 	if (val < DEFAULT_MAX_RDMA_SIZE) {
3541 		pr_err("val: %lu smaller than DEFAULT_MAX_RDMA_SIZE: %d\n",
3542 			val, DEFAULT_MAX_RDMA_SIZE);
3543 		return -EINVAL;
3544 	}
3545 	sport->port_attrib.srp_max_rdma_size = val;
3546 
3547 	return count;
3548 }
3549 
3550 static ssize_t srpt_tpg_attrib_srp_max_rsp_size_show(struct config_item *item,
3551 		char *page)
3552 {
3553 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3554 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3555 
3556 	return sysfs_emit(page, "%u\n", sport->port_attrib.srp_max_rsp_size);
3557 }
3558 
3559 static ssize_t srpt_tpg_attrib_srp_max_rsp_size_store(struct config_item *item,
3560 		const char *page, size_t count)
3561 {
3562 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3563 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3564 	unsigned long val;
3565 	int ret;
3566 
3567 	ret = kstrtoul(page, 0, &val);
3568 	if (ret < 0) {
3569 		pr_err("kstrtoul() failed with ret: %d\n", ret);
3570 		return -EINVAL;
3571 	}
3572 	if (val > MAX_SRPT_RSP_SIZE) {
3573 		pr_err("val: %lu exceeds MAX_SRPT_RSP_SIZE: %d\n", val,
3574 			MAX_SRPT_RSP_SIZE);
3575 		return -EINVAL;
3576 	}
3577 	if (val < MIN_MAX_RSP_SIZE) {
3578 		pr_err("val: %lu smaller than MIN_MAX_RSP_SIZE: %d\n", val,
3579 			MIN_MAX_RSP_SIZE);
3580 		return -EINVAL;
3581 	}
3582 	sport->port_attrib.srp_max_rsp_size = val;
3583 
3584 	return count;
3585 }
3586 
3587 static ssize_t srpt_tpg_attrib_srp_sq_size_show(struct config_item *item,
3588 		char *page)
3589 {
3590 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3591 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3592 
3593 	return sysfs_emit(page, "%u\n", sport->port_attrib.srp_sq_size);
3594 }
3595 
3596 static ssize_t srpt_tpg_attrib_srp_sq_size_store(struct config_item *item,
3597 		const char *page, size_t count)
3598 {
3599 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3600 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3601 	unsigned long val;
3602 	int ret;
3603 
3604 	ret = kstrtoul(page, 0, &val);
3605 	if (ret < 0) {
3606 		pr_err("kstrtoul() failed with ret: %d\n", ret);
3607 		return -EINVAL;
3608 	}
3609 	if (val > MAX_SRPT_SRQ_SIZE) {
3610 		pr_err("val: %lu exceeds MAX_SRPT_SRQ_SIZE: %d\n", val,
3611 			MAX_SRPT_SRQ_SIZE);
3612 		return -EINVAL;
3613 	}
3614 	if (val < MIN_SRPT_SRQ_SIZE) {
3615 		pr_err("val: %lu smaller than MIN_SRPT_SRQ_SIZE: %d\n", val,
3616 			MIN_SRPT_SRQ_SIZE);
3617 		return -EINVAL;
3618 	}
3619 	sport->port_attrib.srp_sq_size = val;
3620 
3621 	return count;
3622 }
3623 
3624 static ssize_t srpt_tpg_attrib_use_srq_show(struct config_item *item,
3625 					    char *page)
3626 {
3627 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3628 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3629 
3630 	return sysfs_emit(page, "%d\n", sport->port_attrib.use_srq);
3631 }
3632 
3633 static ssize_t srpt_tpg_attrib_use_srq_store(struct config_item *item,
3634 					     const char *page, size_t count)
3635 {
3636 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3637 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3638 	struct srpt_device *sdev = sport->sdev;
3639 	unsigned long val;
3640 	bool enabled;
3641 	int ret;
3642 
3643 	ret = kstrtoul(page, 0, &val);
3644 	if (ret < 0)
3645 		return ret;
3646 	if (val != !!val)
3647 		return -EINVAL;
3648 
3649 	ret = mutex_lock_interruptible(&sdev->sdev_mutex);
3650 	if (ret < 0)
3651 		return ret;
3652 	ret = mutex_lock_interruptible(&sport->mutex);
3653 	if (ret < 0)
3654 		goto unlock_sdev;
3655 	enabled = sport->enabled;
3656 	/* Log out all initiator systems before changing 'use_srq'. */
3657 	srpt_set_enabled(sport, false);
3658 	sport->port_attrib.use_srq = val;
3659 	srpt_use_srq(sdev, sport->port_attrib.use_srq);
3660 	srpt_set_enabled(sport, enabled);
3661 	ret = count;
3662 	mutex_unlock(&sport->mutex);
3663 unlock_sdev:
3664 	mutex_unlock(&sdev->sdev_mutex);
3665 
3666 	return ret;
3667 }
3668 
3669 CONFIGFS_ATTR(srpt_tpg_attrib_,  srp_max_rdma_size);
3670 CONFIGFS_ATTR(srpt_tpg_attrib_,  srp_max_rsp_size);
3671 CONFIGFS_ATTR(srpt_tpg_attrib_,  srp_sq_size);
3672 CONFIGFS_ATTR(srpt_tpg_attrib_,  use_srq);
3673 
3674 static struct configfs_attribute *srpt_tpg_attrib_attrs[] = {
3675 	&srpt_tpg_attrib_attr_srp_max_rdma_size,
3676 	&srpt_tpg_attrib_attr_srp_max_rsp_size,
3677 	&srpt_tpg_attrib_attr_srp_sq_size,
3678 	&srpt_tpg_attrib_attr_use_srq,
3679 	NULL,
3680 };
3681 
3682 static struct rdma_cm_id *srpt_create_rdma_id(struct sockaddr *listen_addr)
3683 {
3684 	struct rdma_cm_id *rdma_cm_id;
3685 	int ret;
3686 
3687 	rdma_cm_id = rdma_create_id(&init_net, srpt_rdma_cm_handler,
3688 				    NULL, RDMA_PS_TCP, IB_QPT_RC);
3689 	if (IS_ERR(rdma_cm_id)) {
3690 		pr_err("RDMA/CM ID creation failed: %ld\n",
3691 		       PTR_ERR(rdma_cm_id));
3692 		goto out;
3693 	}
3694 
3695 	ret = rdma_bind_addr(rdma_cm_id, listen_addr);
3696 	if (ret) {
3697 		char addr_str[64];
3698 
3699 		snprintf(addr_str, sizeof(addr_str), "%pISp", listen_addr);
3700 		pr_err("Binding RDMA/CM ID to address %s failed: %d\n",
3701 		       addr_str, ret);
3702 		rdma_destroy_id(rdma_cm_id);
3703 		rdma_cm_id = ERR_PTR(ret);
3704 		goto out;
3705 	}
3706 
3707 	ret = rdma_listen(rdma_cm_id, 128);
3708 	if (ret) {
3709 		pr_err("rdma_listen() failed: %d\n", ret);
3710 		rdma_destroy_id(rdma_cm_id);
3711 		rdma_cm_id = ERR_PTR(ret);
3712 	}
3713 
3714 out:
3715 	return rdma_cm_id;
3716 }
3717 
3718 static ssize_t srpt_rdma_cm_port_show(struct config_item *item, char *page)
3719 {
3720 	return sysfs_emit(page, "%d\n", rdma_cm_port);
3721 }
3722 
3723 static ssize_t srpt_rdma_cm_port_store(struct config_item *item,
3724 				       const char *page, size_t count)
3725 {
3726 	struct sockaddr_in  addr4 = { .sin_family  = AF_INET  };
3727 	struct sockaddr_in6 addr6 = { .sin6_family = AF_INET6 };
3728 	struct rdma_cm_id *new_id = NULL;
3729 	u16 val;
3730 	int ret;
3731 
3732 	ret = kstrtou16(page, 0, &val);
3733 	if (ret < 0)
3734 		return ret;
3735 	ret = count;
3736 	if (rdma_cm_port == val)
3737 		goto out;
3738 
3739 	if (val) {
3740 		addr6.sin6_port = cpu_to_be16(val);
3741 		new_id = srpt_create_rdma_id((struct sockaddr *)&addr6);
3742 		if (IS_ERR(new_id)) {
3743 			addr4.sin_port = cpu_to_be16(val);
3744 			new_id = srpt_create_rdma_id((struct sockaddr *)&addr4);
3745 			if (IS_ERR(new_id)) {
3746 				ret = PTR_ERR(new_id);
3747 				goto out;
3748 			}
3749 		}
3750 	}
3751 
3752 	mutex_lock(&rdma_cm_mutex);
3753 	rdma_cm_port = val;
3754 	swap(rdma_cm_id, new_id);
3755 	mutex_unlock(&rdma_cm_mutex);
3756 
3757 	if (new_id)
3758 		rdma_destroy_id(new_id);
3759 	ret = count;
3760 out:
3761 	return ret;
3762 }
3763 
3764 CONFIGFS_ATTR(srpt_, rdma_cm_port);
3765 
3766 static struct configfs_attribute *srpt_da_attrs[] = {
3767 	&srpt_attr_rdma_cm_port,
3768 	NULL,
3769 };
3770 
3771 static int srpt_enable_tpg(struct se_portal_group *se_tpg, bool enable)
3772 {
3773 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3774 
3775 	mutex_lock(&sport->mutex);
3776 	srpt_set_enabled(sport, enable);
3777 	mutex_unlock(&sport->mutex);
3778 
3779 	return 0;
3780 }
3781 
3782 /**
3783  * srpt_make_tpg - configfs callback invoked for mkdir /sys/kernel/config/target/$driver/$port/$tpg
3784  * @wwn: Corresponds to $driver/$port.
3785  * @name: $tpg.
3786  */
3787 static struct se_portal_group *srpt_make_tpg(struct se_wwn *wwn,
3788 					     const char *name)
3789 {
3790 	struct srpt_port_id *sport_id = srpt_wwn_to_sport_id(wwn);
3791 	struct srpt_tpg *stpg;
3792 	int res = -ENOMEM;
3793 
3794 	stpg = kzalloc(sizeof(*stpg), GFP_KERNEL);
3795 	if (!stpg)
3796 		return ERR_PTR(res);
3797 	stpg->sport_id = sport_id;
3798 	res = core_tpg_register(wwn, &stpg->tpg, SCSI_PROTOCOL_SRP);
3799 	if (res) {
3800 		kfree(stpg);
3801 		return ERR_PTR(res);
3802 	}
3803 
3804 	mutex_lock(&sport_id->mutex);
3805 	list_add_tail(&stpg->entry, &sport_id->tpg_list);
3806 	mutex_unlock(&sport_id->mutex);
3807 
3808 	return &stpg->tpg;
3809 }
3810 
3811 /**
3812  * srpt_drop_tpg - configfs callback invoked for rmdir /sys/kernel/config/target/$driver/$port/$tpg
3813  * @tpg: Target portal group to deregister.
3814  */
3815 static void srpt_drop_tpg(struct se_portal_group *tpg)
3816 {
3817 	struct srpt_tpg *stpg = container_of(tpg, typeof(*stpg), tpg);
3818 	struct srpt_port_id *sport_id = stpg->sport_id;
3819 	struct srpt_port *sport = srpt_tpg_to_sport(tpg);
3820 
3821 	mutex_lock(&sport_id->mutex);
3822 	list_del(&stpg->entry);
3823 	mutex_unlock(&sport_id->mutex);
3824 
3825 	sport->enabled = false;
3826 	core_tpg_deregister(tpg);
3827 	kfree(stpg);
3828 }
3829 
3830 /**
3831  * srpt_make_tport - configfs callback invoked for mkdir /sys/kernel/config/target/$driver/$port
3832  * @tf: Not used.
3833  * @group: Not used.
3834  * @name: $port.
3835  */
3836 static struct se_wwn *srpt_make_tport(struct target_fabric_configfs *tf,
3837 				      struct config_group *group,
3838 				      const char *name)
3839 {
3840 	struct port_and_port_id papi = srpt_lookup_port(name);
3841 	struct srpt_port *sport = papi.sport;
3842 	struct srpt_port_id *port_id;
3843 
3844 	if (!papi.port_id)
3845 		return ERR_PTR(-EINVAL);
3846 	if (*papi.port_id) {
3847 		/* Attempt to create a directory that already exists. */
3848 		WARN_ON_ONCE(true);
3849 		return &(*papi.port_id)->wwn;
3850 	}
3851 	port_id = kzalloc(sizeof(*port_id), GFP_KERNEL);
3852 	if (!port_id) {
3853 		srpt_sdev_put(sport->sdev);
3854 		return ERR_PTR(-ENOMEM);
3855 	}
3856 	mutex_init(&port_id->mutex);
3857 	INIT_LIST_HEAD(&port_id->tpg_list);
3858 	port_id->wwn.priv = sport;
3859 	memcpy(port_id->name, port_id == sport->guid_id ? sport->guid_name :
3860 	       sport->gid_name, ARRAY_SIZE(port_id->name));
3861 
3862 	*papi.port_id = port_id;
3863 
3864 	return &port_id->wwn;
3865 }
3866 
3867 /**
3868  * srpt_drop_tport - configfs callback invoked for rmdir /sys/kernel/config/target/$driver/$port
3869  * @wwn: $port.
3870  */
3871 static void srpt_drop_tport(struct se_wwn *wwn)
3872 {
3873 	struct srpt_port_id *port_id = container_of(wwn, typeof(*port_id), wwn);
3874 	struct srpt_port *sport = wwn->priv;
3875 
3876 	if (sport->guid_id == port_id)
3877 		sport->guid_id = NULL;
3878 	else if (sport->gid_id == port_id)
3879 		sport->gid_id = NULL;
3880 	else
3881 		WARN_ON_ONCE(true);
3882 
3883 	srpt_sdev_put(sport->sdev);
3884 	kfree(port_id);
3885 }
3886 
3887 static ssize_t srpt_wwn_version_show(struct config_item *item, char *buf)
3888 {
3889 	return sysfs_emit(buf, "\n");
3890 }
3891 
3892 CONFIGFS_ATTR_RO(srpt_wwn_, version);
3893 
3894 static struct configfs_attribute *srpt_wwn_attrs[] = {
3895 	&srpt_wwn_attr_version,
3896 	NULL,
3897 };
3898 
3899 static const struct target_core_fabric_ops srpt_template = {
3900 	.module				= THIS_MODULE,
3901 	.fabric_name			= "srpt",
3902 	.tpg_get_wwn			= srpt_get_fabric_wwn,
3903 	.tpg_get_tag			= srpt_get_tag,
3904 	.tpg_check_demo_mode_cache	= srpt_check_true,
3905 	.tpg_check_demo_mode_write_protect = srpt_check_true,
3906 	.release_cmd			= srpt_release_cmd,
3907 	.check_stop_free		= srpt_check_stop_free,
3908 	.close_session			= srpt_close_session,
3909 	.sess_get_initiator_sid		= NULL,
3910 	.write_pending			= srpt_write_pending,
3911 	.get_cmd_state			= srpt_get_tcm_cmd_state,
3912 	.queue_data_in			= srpt_queue_data_in,
3913 	.queue_status			= srpt_queue_status,
3914 	.queue_tm_rsp			= srpt_queue_tm_rsp,
3915 	.aborted_task			= srpt_aborted_task,
3916 	/*
3917 	 * Setup function pointers for generic logic in
3918 	 * target_core_fabric_configfs.c
3919 	 */
3920 	.fabric_make_wwn		= srpt_make_tport,
3921 	.fabric_drop_wwn		= srpt_drop_tport,
3922 	.fabric_make_tpg		= srpt_make_tpg,
3923 	.fabric_enable_tpg		= srpt_enable_tpg,
3924 	.fabric_drop_tpg		= srpt_drop_tpg,
3925 	.fabric_init_nodeacl		= srpt_init_nodeacl,
3926 
3927 	.tfc_discovery_attrs		= srpt_da_attrs,
3928 	.tfc_wwn_attrs			= srpt_wwn_attrs,
3929 	.tfc_tpg_attrib_attrs		= srpt_tpg_attrib_attrs,
3930 
3931 	.default_submit_type		= TARGET_DIRECT_SUBMIT,
3932 	.direct_submit_supp		= 1,
3933 };
3934 
3935 /**
3936  * srpt_init_module - kernel module initialization
3937  *
3938  * Note: Since ib_register_client() registers callback functions, and since at
3939  * least one of these callback functions (srpt_add_one()) calls target core
3940  * functions, this driver must be registered with the target core before
3941  * ib_register_client() is called.
3942  */
3943 static int __init srpt_init_module(void)
3944 {
3945 	int ret;
3946 
3947 	ret = -EINVAL;
3948 	if (srp_max_req_size < MIN_MAX_REQ_SIZE) {
3949 		pr_err("invalid value %d for kernel module parameter srp_max_req_size -- must be at least %d.\n",
3950 		       srp_max_req_size, MIN_MAX_REQ_SIZE);
3951 		goto out;
3952 	}
3953 
3954 	if (srpt_srq_size < MIN_SRPT_SRQ_SIZE
3955 	    || srpt_srq_size > MAX_SRPT_SRQ_SIZE) {
3956 		pr_err("invalid value %d for kernel module parameter srpt_srq_size -- must be in the range [%d..%d].\n",
3957 		       srpt_srq_size, MIN_SRPT_SRQ_SIZE, MAX_SRPT_SRQ_SIZE);
3958 		goto out;
3959 	}
3960 
3961 	ret = target_register_template(&srpt_template);
3962 	if (ret)
3963 		goto out;
3964 
3965 	ret = ib_register_client(&srpt_client);
3966 	if (ret) {
3967 		pr_err("couldn't register IB client\n");
3968 		goto out_unregister_target;
3969 	}
3970 
3971 	return 0;
3972 
3973 out_unregister_target:
3974 	target_unregister_template(&srpt_template);
3975 out:
3976 	return ret;
3977 }
3978 
3979 static void __exit srpt_cleanup_module(void)
3980 {
3981 	if (rdma_cm_id)
3982 		rdma_destroy_id(rdma_cm_id);
3983 	ib_unregister_client(&srpt_client);
3984 	target_unregister_template(&srpt_template);
3985 }
3986 
3987 module_init(srpt_init_module);
3988 module_exit(srpt_cleanup_module);
3989