1 /*
2  * Copyright (c) 2006 - 2009 Mellanox Technology Inc.  All rights reserved.
3  * Copyright (C) 2008 - 2011 Bart Van Assche <bvanassche@acm.org>.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  *
33  */
34 
35 #include <linux/module.h>
36 #include <linux/init.h>
37 #include <linux/slab.h>
38 #include <linux/err.h>
39 #include <linux/ctype.h>
40 #include <linux/kthread.h>
41 #include <linux/string.h>
42 #include <linux/delay.h>
43 #include <linux/atomic.h>
44 #include <linux/inet.h>
45 #include <rdma/ib_cache.h>
46 #include <scsi/scsi_proto.h>
47 #include <scsi/scsi_tcq.h>
48 #include <target/target_core_base.h>
49 #include <target/target_core_fabric.h>
50 #include "ib_srpt.h"
51 
52 /* Name of this kernel module. */
53 #define DRV_NAME		"ib_srpt"
54 
55 #define SRPT_ID_STRING	"Linux SRP target"
56 
57 #undef pr_fmt
58 #define pr_fmt(fmt) DRV_NAME " " fmt
59 
60 MODULE_AUTHOR("Vu Pham and Bart Van Assche");
61 MODULE_DESCRIPTION("SCSI RDMA Protocol target driver");
62 MODULE_LICENSE("Dual BSD/GPL");
63 
64 /*
65  * Global Variables
66  */
67 
68 static u64 srpt_service_guid;
69 static DEFINE_SPINLOCK(srpt_dev_lock);	/* Protects srpt_dev_list. */
70 static LIST_HEAD(srpt_dev_list);	/* List of srpt_device structures. */
71 
72 static unsigned srp_max_req_size = DEFAULT_MAX_REQ_SIZE;
73 module_param(srp_max_req_size, int, 0444);
74 MODULE_PARM_DESC(srp_max_req_size,
75 		 "Maximum size of SRP request messages in bytes.");
76 
77 static int srpt_srq_size = DEFAULT_SRPT_SRQ_SIZE;
78 module_param(srpt_srq_size, int, 0444);
79 MODULE_PARM_DESC(srpt_srq_size,
80 		 "Shared receive queue (SRQ) size.");
81 
82 static int srpt_get_u64_x(char *buffer, const struct kernel_param *kp)
83 {
84 	return sprintf(buffer, "0x%016llx\n", *(u64 *)kp->arg);
85 }
86 module_param_call(srpt_service_guid, NULL, srpt_get_u64_x, &srpt_service_guid,
87 		  0444);
88 MODULE_PARM_DESC(srpt_service_guid,
89 		 "Using this value for ioc_guid, id_ext, and cm_listen_id instead of using the node_guid of the first HCA.");
90 
91 static struct ib_client srpt_client;
92 /* Protects both rdma_cm_port and rdma_cm_id. */
93 static DEFINE_MUTEX(rdma_cm_mutex);
94 /* Port number RDMA/CM will bind to. */
95 static u16 rdma_cm_port;
96 static struct rdma_cm_id *rdma_cm_id;
97 static void srpt_release_cmd(struct se_cmd *se_cmd);
98 static void srpt_free_ch(struct kref *kref);
99 static int srpt_queue_status(struct se_cmd *cmd);
100 static void srpt_recv_done(struct ib_cq *cq, struct ib_wc *wc);
101 static void srpt_send_done(struct ib_cq *cq, struct ib_wc *wc);
102 static void srpt_process_wait_list(struct srpt_rdma_ch *ch);
103 
104 /*
105  * The only allowed channel state changes are those that change the channel
106  * state into a state with a higher numerical value. Hence the new > prev test.
107  */
108 static bool srpt_set_ch_state(struct srpt_rdma_ch *ch, enum rdma_ch_state new)
109 {
110 	unsigned long flags;
111 	enum rdma_ch_state prev;
112 	bool changed = false;
113 
114 	spin_lock_irqsave(&ch->spinlock, flags);
115 	prev = ch->state;
116 	if (new > prev) {
117 		ch->state = new;
118 		changed = true;
119 	}
120 	spin_unlock_irqrestore(&ch->spinlock, flags);
121 
122 	return changed;
123 }
124 
125 /**
126  * srpt_event_handler - asynchronous IB event callback function
127  * @handler: IB event handler registered by ib_register_event_handler().
128  * @event: Description of the event that occurred.
129  *
130  * Callback function called by the InfiniBand core when an asynchronous IB
131  * event occurs. This callback may occur in interrupt context. See also
132  * section 11.5.2, Set Asynchronous Event Handler in the InfiniBand
133  * Architecture Specification.
134  */
135 static void srpt_event_handler(struct ib_event_handler *handler,
136 			       struct ib_event *event)
137 {
138 	struct srpt_device *sdev =
139 		container_of(handler, struct srpt_device, event_handler);
140 	struct srpt_port *sport;
141 	u8 port_num;
142 
143 	pr_debug("ASYNC event= %d on device= %s\n", event->event,
144 		 dev_name(&sdev->device->dev));
145 
146 	switch (event->event) {
147 	case IB_EVENT_PORT_ERR:
148 		port_num = event->element.port_num - 1;
149 		if (port_num < sdev->device->phys_port_cnt) {
150 			sport = &sdev->port[port_num];
151 			sport->lid = 0;
152 			sport->sm_lid = 0;
153 		} else {
154 			WARN(true, "event %d: port_num %d out of range 1..%d\n",
155 			     event->event, port_num + 1,
156 			     sdev->device->phys_port_cnt);
157 		}
158 		break;
159 	case IB_EVENT_PORT_ACTIVE:
160 	case IB_EVENT_LID_CHANGE:
161 	case IB_EVENT_PKEY_CHANGE:
162 	case IB_EVENT_SM_CHANGE:
163 	case IB_EVENT_CLIENT_REREGISTER:
164 	case IB_EVENT_GID_CHANGE:
165 		/* Refresh port data asynchronously. */
166 		port_num = event->element.port_num - 1;
167 		if (port_num < sdev->device->phys_port_cnt) {
168 			sport = &sdev->port[port_num];
169 			if (!sport->lid && !sport->sm_lid)
170 				schedule_work(&sport->work);
171 		} else {
172 			WARN(true, "event %d: port_num %d out of range 1..%d\n",
173 			     event->event, port_num + 1,
174 			     sdev->device->phys_port_cnt);
175 		}
176 		break;
177 	default:
178 		pr_err("received unrecognized IB event %d\n", event->event);
179 		break;
180 	}
181 }
182 
183 /**
184  * srpt_srq_event - SRQ event callback function
185  * @event: Description of the event that occurred.
186  * @ctx: Context pointer specified at SRQ creation time.
187  */
188 static void srpt_srq_event(struct ib_event *event, void *ctx)
189 {
190 	pr_debug("SRQ event %d\n", event->event);
191 }
192 
193 static const char *get_ch_state_name(enum rdma_ch_state s)
194 {
195 	switch (s) {
196 	case CH_CONNECTING:
197 		return "connecting";
198 	case CH_LIVE:
199 		return "live";
200 	case CH_DISCONNECTING:
201 		return "disconnecting";
202 	case CH_DRAINING:
203 		return "draining";
204 	case CH_DISCONNECTED:
205 		return "disconnected";
206 	}
207 	return "???";
208 }
209 
210 /**
211  * srpt_qp_event - QP event callback function
212  * @event: Description of the event that occurred.
213  * @ch: SRPT RDMA channel.
214  */
215 static void srpt_qp_event(struct ib_event *event, struct srpt_rdma_ch *ch)
216 {
217 	pr_debug("QP event %d on ch=%p sess_name=%s-%d state=%s\n",
218 		 event->event, ch, ch->sess_name, ch->qp->qp_num,
219 		 get_ch_state_name(ch->state));
220 
221 	switch (event->event) {
222 	case IB_EVENT_COMM_EST:
223 		if (ch->using_rdma_cm)
224 			rdma_notify(ch->rdma_cm.cm_id, event->event);
225 		else
226 			ib_cm_notify(ch->ib_cm.cm_id, event->event);
227 		break;
228 	case IB_EVENT_QP_LAST_WQE_REACHED:
229 		pr_debug("%s-%d, state %s: received Last WQE event.\n",
230 			 ch->sess_name, ch->qp->qp_num,
231 			 get_ch_state_name(ch->state));
232 		break;
233 	default:
234 		pr_err("received unrecognized IB QP event %d\n", event->event);
235 		break;
236 	}
237 }
238 
239 /**
240  * srpt_set_ioc - initialize a IOUnitInfo structure
241  * @c_list: controller list.
242  * @slot: one-based slot number.
243  * @value: four-bit value.
244  *
245  * Copies the lowest four bits of value in element slot of the array of four
246  * bit elements called c_list (controller list). The index slot is one-based.
247  */
248 static void srpt_set_ioc(u8 *c_list, u32 slot, u8 value)
249 {
250 	u16 id;
251 	u8 tmp;
252 
253 	id = (slot - 1) / 2;
254 	if (slot & 0x1) {
255 		tmp = c_list[id] & 0xf;
256 		c_list[id] = (value << 4) | tmp;
257 	} else {
258 		tmp = c_list[id] & 0xf0;
259 		c_list[id] = (value & 0xf) | tmp;
260 	}
261 }
262 
263 /**
264  * srpt_get_class_port_info - copy ClassPortInfo to a management datagram
265  * @mad: Datagram that will be sent as response to DM_ATTR_CLASS_PORT_INFO.
266  *
267  * See also section 16.3.3.1 ClassPortInfo in the InfiniBand Architecture
268  * Specification.
269  */
270 static void srpt_get_class_port_info(struct ib_dm_mad *mad)
271 {
272 	struct ib_class_port_info *cif;
273 
274 	cif = (struct ib_class_port_info *)mad->data;
275 	memset(cif, 0, sizeof(*cif));
276 	cif->base_version = 1;
277 	cif->class_version = 1;
278 
279 	ib_set_cpi_resp_time(cif, 20);
280 	mad->mad_hdr.status = 0;
281 }
282 
283 /**
284  * srpt_get_iou - write IOUnitInfo to a management datagram
285  * @mad: Datagram that will be sent as response to DM_ATTR_IOU_INFO.
286  *
287  * See also section 16.3.3.3 IOUnitInfo in the InfiniBand Architecture
288  * Specification. See also section B.7, table B.6 in the SRP r16a document.
289  */
290 static void srpt_get_iou(struct ib_dm_mad *mad)
291 {
292 	struct ib_dm_iou_info *ioui;
293 	u8 slot;
294 	int i;
295 
296 	ioui = (struct ib_dm_iou_info *)mad->data;
297 	ioui->change_id = cpu_to_be16(1);
298 	ioui->max_controllers = 16;
299 
300 	/* set present for slot 1 and empty for the rest */
301 	srpt_set_ioc(ioui->controller_list, 1, 1);
302 	for (i = 1, slot = 2; i < 16; i++, slot++)
303 		srpt_set_ioc(ioui->controller_list, slot, 0);
304 
305 	mad->mad_hdr.status = 0;
306 }
307 
308 /**
309  * srpt_get_ioc - write IOControllerprofile to a management datagram
310  * @sport: HCA port through which the MAD has been received.
311  * @slot: Slot number specified in DM_ATTR_IOC_PROFILE query.
312  * @mad: Datagram that will be sent as response to DM_ATTR_IOC_PROFILE.
313  *
314  * See also section 16.3.3.4 IOControllerProfile in the InfiniBand
315  * Architecture Specification. See also section B.7, table B.7 in the SRP
316  * r16a document.
317  */
318 static void srpt_get_ioc(struct srpt_port *sport, u32 slot,
319 			 struct ib_dm_mad *mad)
320 {
321 	struct srpt_device *sdev = sport->sdev;
322 	struct ib_dm_ioc_profile *iocp;
323 	int send_queue_depth;
324 
325 	iocp = (struct ib_dm_ioc_profile *)mad->data;
326 
327 	if (!slot || slot > 16) {
328 		mad->mad_hdr.status
329 			= cpu_to_be16(DM_MAD_STATUS_INVALID_FIELD);
330 		return;
331 	}
332 
333 	if (slot > 2) {
334 		mad->mad_hdr.status
335 			= cpu_to_be16(DM_MAD_STATUS_NO_IOC);
336 		return;
337 	}
338 
339 	if (sdev->use_srq)
340 		send_queue_depth = sdev->srq_size;
341 	else
342 		send_queue_depth = min(MAX_SRPT_RQ_SIZE,
343 				       sdev->device->attrs.max_qp_wr);
344 
345 	memset(iocp, 0, sizeof(*iocp));
346 	strcpy(iocp->id_string, SRPT_ID_STRING);
347 	iocp->guid = cpu_to_be64(srpt_service_guid);
348 	iocp->vendor_id = cpu_to_be32(sdev->device->attrs.vendor_id);
349 	iocp->device_id = cpu_to_be32(sdev->device->attrs.vendor_part_id);
350 	iocp->device_version = cpu_to_be16(sdev->device->attrs.hw_ver);
351 	iocp->subsys_vendor_id = cpu_to_be32(sdev->device->attrs.vendor_id);
352 	iocp->subsys_device_id = 0x0;
353 	iocp->io_class = cpu_to_be16(SRP_REV16A_IB_IO_CLASS);
354 	iocp->io_subclass = cpu_to_be16(SRP_IO_SUBCLASS);
355 	iocp->protocol = cpu_to_be16(SRP_PROTOCOL);
356 	iocp->protocol_version = cpu_to_be16(SRP_PROTOCOL_VERSION);
357 	iocp->send_queue_depth = cpu_to_be16(send_queue_depth);
358 	iocp->rdma_read_depth = 4;
359 	iocp->send_size = cpu_to_be32(srp_max_req_size);
360 	iocp->rdma_size = cpu_to_be32(min(sport->port_attrib.srp_max_rdma_size,
361 					  1U << 24));
362 	iocp->num_svc_entries = 1;
363 	iocp->op_cap_mask = SRP_SEND_TO_IOC | SRP_SEND_FROM_IOC |
364 		SRP_RDMA_READ_FROM_IOC | SRP_RDMA_WRITE_FROM_IOC;
365 
366 	mad->mad_hdr.status = 0;
367 }
368 
369 /**
370  * srpt_get_svc_entries - write ServiceEntries to a management datagram
371  * @ioc_guid: I/O controller GUID to use in reply.
372  * @slot: I/O controller number.
373  * @hi: End of the range of service entries to be specified in the reply.
374  * @lo: Start of the range of service entries to be specified in the reply..
375  * @mad: Datagram that will be sent as response to DM_ATTR_SVC_ENTRIES.
376  *
377  * See also section 16.3.3.5 ServiceEntries in the InfiniBand Architecture
378  * Specification. See also section B.7, table B.8 in the SRP r16a document.
379  */
380 static void srpt_get_svc_entries(u64 ioc_guid,
381 				 u16 slot, u8 hi, u8 lo, struct ib_dm_mad *mad)
382 {
383 	struct ib_dm_svc_entries *svc_entries;
384 
385 	WARN_ON(!ioc_guid);
386 
387 	if (!slot || slot > 16) {
388 		mad->mad_hdr.status
389 			= cpu_to_be16(DM_MAD_STATUS_INVALID_FIELD);
390 		return;
391 	}
392 
393 	if (slot > 2 || lo > hi || hi > 1) {
394 		mad->mad_hdr.status
395 			= cpu_to_be16(DM_MAD_STATUS_NO_IOC);
396 		return;
397 	}
398 
399 	svc_entries = (struct ib_dm_svc_entries *)mad->data;
400 	memset(svc_entries, 0, sizeof(*svc_entries));
401 	svc_entries->service_entries[0].id = cpu_to_be64(ioc_guid);
402 	snprintf(svc_entries->service_entries[0].name,
403 		 sizeof(svc_entries->service_entries[0].name),
404 		 "%s%016llx",
405 		 SRP_SERVICE_NAME_PREFIX,
406 		 ioc_guid);
407 
408 	mad->mad_hdr.status = 0;
409 }
410 
411 /**
412  * srpt_mgmt_method_get - process a received management datagram
413  * @sp:      HCA port through which the MAD has been received.
414  * @rq_mad:  received MAD.
415  * @rsp_mad: response MAD.
416  */
417 static void srpt_mgmt_method_get(struct srpt_port *sp, struct ib_mad *rq_mad,
418 				 struct ib_dm_mad *rsp_mad)
419 {
420 	u16 attr_id;
421 	u32 slot;
422 	u8 hi, lo;
423 
424 	attr_id = be16_to_cpu(rq_mad->mad_hdr.attr_id);
425 	switch (attr_id) {
426 	case DM_ATTR_CLASS_PORT_INFO:
427 		srpt_get_class_port_info(rsp_mad);
428 		break;
429 	case DM_ATTR_IOU_INFO:
430 		srpt_get_iou(rsp_mad);
431 		break;
432 	case DM_ATTR_IOC_PROFILE:
433 		slot = be32_to_cpu(rq_mad->mad_hdr.attr_mod);
434 		srpt_get_ioc(sp, slot, rsp_mad);
435 		break;
436 	case DM_ATTR_SVC_ENTRIES:
437 		slot = be32_to_cpu(rq_mad->mad_hdr.attr_mod);
438 		hi = (u8) ((slot >> 8) & 0xff);
439 		lo = (u8) (slot & 0xff);
440 		slot = (u16) ((slot >> 16) & 0xffff);
441 		srpt_get_svc_entries(srpt_service_guid,
442 				     slot, hi, lo, rsp_mad);
443 		break;
444 	default:
445 		rsp_mad->mad_hdr.status =
446 		    cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD_ATTR);
447 		break;
448 	}
449 }
450 
451 /**
452  * srpt_mad_send_handler - MAD send completion callback
453  * @mad_agent: Return value of ib_register_mad_agent().
454  * @mad_wc: Work completion reporting that the MAD has been sent.
455  */
456 static void srpt_mad_send_handler(struct ib_mad_agent *mad_agent,
457 				  struct ib_mad_send_wc *mad_wc)
458 {
459 	rdma_destroy_ah(mad_wc->send_buf->ah, RDMA_DESTROY_AH_SLEEPABLE);
460 	ib_free_send_mad(mad_wc->send_buf);
461 }
462 
463 /**
464  * srpt_mad_recv_handler - MAD reception callback function
465  * @mad_agent: Return value of ib_register_mad_agent().
466  * @send_buf: Not used.
467  * @mad_wc: Work completion reporting that a MAD has been received.
468  */
469 static void srpt_mad_recv_handler(struct ib_mad_agent *mad_agent,
470 				  struct ib_mad_send_buf *send_buf,
471 				  struct ib_mad_recv_wc *mad_wc)
472 {
473 	struct srpt_port *sport = (struct srpt_port *)mad_agent->context;
474 	struct ib_ah *ah;
475 	struct ib_mad_send_buf *rsp;
476 	struct ib_dm_mad *dm_mad;
477 
478 	if (!mad_wc || !mad_wc->recv_buf.mad)
479 		return;
480 
481 	ah = ib_create_ah_from_wc(mad_agent->qp->pd, mad_wc->wc,
482 				  mad_wc->recv_buf.grh, mad_agent->port_num);
483 	if (IS_ERR(ah))
484 		goto err;
485 
486 	BUILD_BUG_ON(offsetof(struct ib_dm_mad, data) != IB_MGMT_DEVICE_HDR);
487 
488 	rsp = ib_create_send_mad(mad_agent, mad_wc->wc->src_qp,
489 				 mad_wc->wc->pkey_index, 0,
490 				 IB_MGMT_DEVICE_HDR, IB_MGMT_DEVICE_DATA,
491 				 GFP_KERNEL,
492 				 IB_MGMT_BASE_VERSION);
493 	if (IS_ERR(rsp))
494 		goto err_rsp;
495 
496 	rsp->ah = ah;
497 
498 	dm_mad = rsp->mad;
499 	memcpy(dm_mad, mad_wc->recv_buf.mad, sizeof(*dm_mad));
500 	dm_mad->mad_hdr.method = IB_MGMT_METHOD_GET_RESP;
501 	dm_mad->mad_hdr.status = 0;
502 
503 	switch (mad_wc->recv_buf.mad->mad_hdr.method) {
504 	case IB_MGMT_METHOD_GET:
505 		srpt_mgmt_method_get(sport, mad_wc->recv_buf.mad, dm_mad);
506 		break;
507 	case IB_MGMT_METHOD_SET:
508 		dm_mad->mad_hdr.status =
509 		    cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD_ATTR);
510 		break;
511 	default:
512 		dm_mad->mad_hdr.status =
513 		    cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD);
514 		break;
515 	}
516 
517 	if (!ib_post_send_mad(rsp, NULL)) {
518 		ib_free_recv_mad(mad_wc);
519 		/* will destroy_ah & free_send_mad in send completion */
520 		return;
521 	}
522 
523 	ib_free_send_mad(rsp);
524 
525 err_rsp:
526 	rdma_destroy_ah(ah, RDMA_DESTROY_AH_SLEEPABLE);
527 err:
528 	ib_free_recv_mad(mad_wc);
529 }
530 
531 static int srpt_format_guid(char *buf, unsigned int size, const __be64 *guid)
532 {
533 	const __be16 *g = (const __be16 *)guid;
534 
535 	return snprintf(buf, size, "%04x:%04x:%04x:%04x",
536 			be16_to_cpu(g[0]), be16_to_cpu(g[1]),
537 			be16_to_cpu(g[2]), be16_to_cpu(g[3]));
538 }
539 
540 /**
541  * srpt_refresh_port - configure a HCA port
542  * @sport: SRPT HCA port.
543  *
544  * Enable InfiniBand management datagram processing, update the cached sm_lid,
545  * lid and gid values, and register a callback function for processing MADs
546  * on the specified port.
547  *
548  * Note: It is safe to call this function more than once for the same port.
549  */
550 static int srpt_refresh_port(struct srpt_port *sport)
551 {
552 	struct ib_mad_reg_req reg_req;
553 	struct ib_port_modify port_modify;
554 	struct ib_port_attr port_attr;
555 	int ret;
556 
557 	ret = ib_query_port(sport->sdev->device, sport->port, &port_attr);
558 	if (ret)
559 		return ret;
560 
561 	sport->sm_lid = port_attr.sm_lid;
562 	sport->lid = port_attr.lid;
563 
564 	ret = rdma_query_gid(sport->sdev->device, sport->port, 0, &sport->gid);
565 	if (ret)
566 		return ret;
567 
568 	sport->port_guid_id.wwn.priv = sport;
569 	srpt_format_guid(sport->port_guid_id.name,
570 			 sizeof(sport->port_guid_id.name),
571 			 &sport->gid.global.interface_id);
572 	sport->port_gid_id.wwn.priv = sport;
573 	snprintf(sport->port_gid_id.name, sizeof(sport->port_gid_id.name),
574 		 "0x%016llx%016llx",
575 		 be64_to_cpu(sport->gid.global.subnet_prefix),
576 		 be64_to_cpu(sport->gid.global.interface_id));
577 
578 	if (rdma_protocol_iwarp(sport->sdev->device, sport->port))
579 		return 0;
580 
581 	memset(&port_modify, 0, sizeof(port_modify));
582 	port_modify.set_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP;
583 	port_modify.clr_port_cap_mask = 0;
584 
585 	ret = ib_modify_port(sport->sdev->device, sport->port, 0, &port_modify);
586 	if (ret) {
587 		pr_warn("%s-%d: enabling device management failed (%d). Note: this is expected if SR-IOV is enabled.\n",
588 			dev_name(&sport->sdev->device->dev), sport->port, ret);
589 		return 0;
590 	}
591 
592 	if (!sport->mad_agent) {
593 		memset(&reg_req, 0, sizeof(reg_req));
594 		reg_req.mgmt_class = IB_MGMT_CLASS_DEVICE_MGMT;
595 		reg_req.mgmt_class_version = IB_MGMT_BASE_VERSION;
596 		set_bit(IB_MGMT_METHOD_GET, reg_req.method_mask);
597 		set_bit(IB_MGMT_METHOD_SET, reg_req.method_mask);
598 
599 		sport->mad_agent = ib_register_mad_agent(sport->sdev->device,
600 							 sport->port,
601 							 IB_QPT_GSI,
602 							 &reg_req, 0,
603 							 srpt_mad_send_handler,
604 							 srpt_mad_recv_handler,
605 							 sport, 0);
606 		if (IS_ERR(sport->mad_agent)) {
607 			pr_err("%s-%d: MAD agent registration failed (%ld). Note: this is expected if SR-IOV is enabled.\n",
608 			       dev_name(&sport->sdev->device->dev), sport->port,
609 			       PTR_ERR(sport->mad_agent));
610 			sport->mad_agent = NULL;
611 			memset(&port_modify, 0, sizeof(port_modify));
612 			port_modify.clr_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP;
613 			ib_modify_port(sport->sdev->device, sport->port, 0,
614 				       &port_modify);
615 
616 		}
617 	}
618 
619 	return 0;
620 }
621 
622 /**
623  * srpt_unregister_mad_agent - unregister MAD callback functions
624  * @sdev: SRPT HCA pointer.
625  * @port_cnt: number of ports with registered MAD
626  *
627  * Note: It is safe to call this function more than once for the same device.
628  */
629 static void srpt_unregister_mad_agent(struct srpt_device *sdev, int port_cnt)
630 {
631 	struct ib_port_modify port_modify = {
632 		.clr_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP,
633 	};
634 	struct srpt_port *sport;
635 	int i;
636 
637 	for (i = 1; i <= port_cnt; i++) {
638 		sport = &sdev->port[i - 1];
639 		WARN_ON(sport->port != i);
640 		if (sport->mad_agent) {
641 			ib_modify_port(sdev->device, i, 0, &port_modify);
642 			ib_unregister_mad_agent(sport->mad_agent);
643 			sport->mad_agent = NULL;
644 		}
645 	}
646 }
647 
648 /**
649  * srpt_alloc_ioctx - allocate a SRPT I/O context structure
650  * @sdev: SRPT HCA pointer.
651  * @ioctx_size: I/O context size.
652  * @buf_cache: I/O buffer cache.
653  * @dir: DMA data direction.
654  */
655 static struct srpt_ioctx *srpt_alloc_ioctx(struct srpt_device *sdev,
656 					   int ioctx_size,
657 					   struct kmem_cache *buf_cache,
658 					   enum dma_data_direction dir)
659 {
660 	struct srpt_ioctx *ioctx;
661 
662 	ioctx = kzalloc(ioctx_size, GFP_KERNEL);
663 	if (!ioctx)
664 		goto err;
665 
666 	ioctx->buf = kmem_cache_alloc(buf_cache, GFP_KERNEL);
667 	if (!ioctx->buf)
668 		goto err_free_ioctx;
669 
670 	ioctx->dma = ib_dma_map_single(sdev->device, ioctx->buf,
671 				       kmem_cache_size(buf_cache), dir);
672 	if (ib_dma_mapping_error(sdev->device, ioctx->dma))
673 		goto err_free_buf;
674 
675 	return ioctx;
676 
677 err_free_buf:
678 	kmem_cache_free(buf_cache, ioctx->buf);
679 err_free_ioctx:
680 	kfree(ioctx);
681 err:
682 	return NULL;
683 }
684 
685 /**
686  * srpt_free_ioctx - free a SRPT I/O context structure
687  * @sdev: SRPT HCA pointer.
688  * @ioctx: I/O context pointer.
689  * @buf_cache: I/O buffer cache.
690  * @dir: DMA data direction.
691  */
692 static void srpt_free_ioctx(struct srpt_device *sdev, struct srpt_ioctx *ioctx,
693 			    struct kmem_cache *buf_cache,
694 			    enum dma_data_direction dir)
695 {
696 	if (!ioctx)
697 		return;
698 
699 	ib_dma_unmap_single(sdev->device, ioctx->dma,
700 			    kmem_cache_size(buf_cache), dir);
701 	kmem_cache_free(buf_cache, ioctx->buf);
702 	kfree(ioctx);
703 }
704 
705 /**
706  * srpt_alloc_ioctx_ring - allocate a ring of SRPT I/O context structures
707  * @sdev:       Device to allocate the I/O context ring for.
708  * @ring_size:  Number of elements in the I/O context ring.
709  * @ioctx_size: I/O context size.
710  * @buf_cache:  I/O buffer cache.
711  * @alignment_offset: Offset in each ring buffer at which the SRP information
712  *		unit starts.
713  * @dir:        DMA data direction.
714  */
715 static struct srpt_ioctx **srpt_alloc_ioctx_ring(struct srpt_device *sdev,
716 				int ring_size, int ioctx_size,
717 				struct kmem_cache *buf_cache,
718 				int alignment_offset,
719 				enum dma_data_direction dir)
720 {
721 	struct srpt_ioctx **ring;
722 	int i;
723 
724 	WARN_ON(ioctx_size != sizeof(struct srpt_recv_ioctx) &&
725 		ioctx_size != sizeof(struct srpt_send_ioctx));
726 
727 	ring = kvmalloc_array(ring_size, sizeof(ring[0]), GFP_KERNEL);
728 	if (!ring)
729 		goto out;
730 	for (i = 0; i < ring_size; ++i) {
731 		ring[i] = srpt_alloc_ioctx(sdev, ioctx_size, buf_cache, dir);
732 		if (!ring[i])
733 			goto err;
734 		ring[i]->index = i;
735 		ring[i]->offset = alignment_offset;
736 	}
737 	goto out;
738 
739 err:
740 	while (--i >= 0)
741 		srpt_free_ioctx(sdev, ring[i], buf_cache, dir);
742 	kvfree(ring);
743 	ring = NULL;
744 out:
745 	return ring;
746 }
747 
748 /**
749  * srpt_free_ioctx_ring - free the ring of SRPT I/O context structures
750  * @ioctx_ring: I/O context ring to be freed.
751  * @sdev: SRPT HCA pointer.
752  * @ring_size: Number of ring elements.
753  * @buf_cache: I/O buffer cache.
754  * @dir: DMA data direction.
755  */
756 static void srpt_free_ioctx_ring(struct srpt_ioctx **ioctx_ring,
757 				 struct srpt_device *sdev, int ring_size,
758 				 struct kmem_cache *buf_cache,
759 				 enum dma_data_direction dir)
760 {
761 	int i;
762 
763 	if (!ioctx_ring)
764 		return;
765 
766 	for (i = 0; i < ring_size; ++i)
767 		srpt_free_ioctx(sdev, ioctx_ring[i], buf_cache, dir);
768 	kvfree(ioctx_ring);
769 }
770 
771 /**
772  * srpt_set_cmd_state - set the state of a SCSI command
773  * @ioctx: Send I/O context.
774  * @new: New I/O context state.
775  *
776  * Does not modify the state of aborted commands. Returns the previous command
777  * state.
778  */
779 static enum srpt_command_state srpt_set_cmd_state(struct srpt_send_ioctx *ioctx,
780 						  enum srpt_command_state new)
781 {
782 	enum srpt_command_state previous;
783 
784 	previous = ioctx->state;
785 	if (previous != SRPT_STATE_DONE)
786 		ioctx->state = new;
787 
788 	return previous;
789 }
790 
791 /**
792  * srpt_test_and_set_cmd_state - test and set the state of a command
793  * @ioctx: Send I/O context.
794  * @old: Current I/O context state.
795  * @new: New I/O context state.
796  *
797  * Returns true if and only if the previous command state was equal to 'old'.
798  */
799 static bool srpt_test_and_set_cmd_state(struct srpt_send_ioctx *ioctx,
800 					enum srpt_command_state old,
801 					enum srpt_command_state new)
802 {
803 	enum srpt_command_state previous;
804 
805 	WARN_ON(!ioctx);
806 	WARN_ON(old == SRPT_STATE_DONE);
807 	WARN_ON(new == SRPT_STATE_NEW);
808 
809 	previous = ioctx->state;
810 	if (previous == old)
811 		ioctx->state = new;
812 
813 	return previous == old;
814 }
815 
816 /**
817  * srpt_post_recv - post an IB receive request
818  * @sdev: SRPT HCA pointer.
819  * @ch: SRPT RDMA channel.
820  * @ioctx: Receive I/O context pointer.
821  */
822 static int srpt_post_recv(struct srpt_device *sdev, struct srpt_rdma_ch *ch,
823 			  struct srpt_recv_ioctx *ioctx)
824 {
825 	struct ib_sge list;
826 	struct ib_recv_wr wr;
827 
828 	BUG_ON(!sdev);
829 	list.addr = ioctx->ioctx.dma + ioctx->ioctx.offset;
830 	list.length = srp_max_req_size;
831 	list.lkey = sdev->lkey;
832 
833 	ioctx->ioctx.cqe.done = srpt_recv_done;
834 	wr.wr_cqe = &ioctx->ioctx.cqe;
835 	wr.next = NULL;
836 	wr.sg_list = &list;
837 	wr.num_sge = 1;
838 
839 	if (sdev->use_srq)
840 		return ib_post_srq_recv(sdev->srq, &wr, NULL);
841 	else
842 		return ib_post_recv(ch->qp, &wr, NULL);
843 }
844 
845 /**
846  * srpt_zerolength_write - perform a zero-length RDMA write
847  * @ch: SRPT RDMA channel.
848  *
849  * A quote from the InfiniBand specification: C9-88: For an HCA responder
850  * using Reliable Connection service, for each zero-length RDMA READ or WRITE
851  * request, the R_Key shall not be validated, even if the request includes
852  * Immediate data.
853  */
854 static int srpt_zerolength_write(struct srpt_rdma_ch *ch)
855 {
856 	struct ib_rdma_wr wr = {
857 		.wr = {
858 			.next		= NULL,
859 			{ .wr_cqe	= &ch->zw_cqe, },
860 			.opcode		= IB_WR_RDMA_WRITE,
861 			.send_flags	= IB_SEND_SIGNALED,
862 		}
863 	};
864 
865 	pr_debug("%s-%d: queued zerolength write\n", ch->sess_name,
866 		 ch->qp->qp_num);
867 
868 	return ib_post_send(ch->qp, &wr.wr, NULL);
869 }
870 
871 static void srpt_zerolength_write_done(struct ib_cq *cq, struct ib_wc *wc)
872 {
873 	struct srpt_rdma_ch *ch = wc->qp->qp_context;
874 
875 	pr_debug("%s-%d wc->status %d\n", ch->sess_name, ch->qp->qp_num,
876 		 wc->status);
877 
878 	if (wc->status == IB_WC_SUCCESS) {
879 		srpt_process_wait_list(ch);
880 	} else {
881 		if (srpt_set_ch_state(ch, CH_DISCONNECTED))
882 			schedule_work(&ch->release_work);
883 		else
884 			pr_debug("%s-%d: already disconnected.\n",
885 				 ch->sess_name, ch->qp->qp_num);
886 	}
887 }
888 
889 static int srpt_alloc_rw_ctxs(struct srpt_send_ioctx *ioctx,
890 		struct srp_direct_buf *db, int nbufs, struct scatterlist **sg,
891 		unsigned *sg_cnt)
892 {
893 	enum dma_data_direction dir = target_reverse_dma_direction(&ioctx->cmd);
894 	struct srpt_rdma_ch *ch = ioctx->ch;
895 	struct scatterlist *prev = NULL;
896 	unsigned prev_nents;
897 	int ret, i;
898 
899 	if (nbufs == 1) {
900 		ioctx->rw_ctxs = &ioctx->s_rw_ctx;
901 	} else {
902 		ioctx->rw_ctxs = kmalloc_array(nbufs, sizeof(*ioctx->rw_ctxs),
903 			GFP_KERNEL);
904 		if (!ioctx->rw_ctxs)
905 			return -ENOMEM;
906 	}
907 
908 	for (i = ioctx->n_rw_ctx; i < nbufs; i++, db++) {
909 		struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i];
910 		u64 remote_addr = be64_to_cpu(db->va);
911 		u32 size = be32_to_cpu(db->len);
912 		u32 rkey = be32_to_cpu(db->key);
913 
914 		ret = target_alloc_sgl(&ctx->sg, &ctx->nents, size, false,
915 				i < nbufs - 1);
916 		if (ret)
917 			goto unwind;
918 
919 		ret = rdma_rw_ctx_init(&ctx->rw, ch->qp, ch->sport->port,
920 				ctx->sg, ctx->nents, 0, remote_addr, rkey, dir);
921 		if (ret < 0) {
922 			target_free_sgl(ctx->sg, ctx->nents);
923 			goto unwind;
924 		}
925 
926 		ioctx->n_rdma += ret;
927 		ioctx->n_rw_ctx++;
928 
929 		if (prev) {
930 			sg_unmark_end(&prev[prev_nents - 1]);
931 			sg_chain(prev, prev_nents + 1, ctx->sg);
932 		} else {
933 			*sg = ctx->sg;
934 		}
935 
936 		prev = ctx->sg;
937 		prev_nents = ctx->nents;
938 
939 		*sg_cnt += ctx->nents;
940 	}
941 
942 	return 0;
943 
944 unwind:
945 	while (--i >= 0) {
946 		struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i];
947 
948 		rdma_rw_ctx_destroy(&ctx->rw, ch->qp, ch->sport->port,
949 				ctx->sg, ctx->nents, dir);
950 		target_free_sgl(ctx->sg, ctx->nents);
951 	}
952 	if (ioctx->rw_ctxs != &ioctx->s_rw_ctx)
953 		kfree(ioctx->rw_ctxs);
954 	return ret;
955 }
956 
957 static void srpt_free_rw_ctxs(struct srpt_rdma_ch *ch,
958 				    struct srpt_send_ioctx *ioctx)
959 {
960 	enum dma_data_direction dir = target_reverse_dma_direction(&ioctx->cmd);
961 	int i;
962 
963 	for (i = 0; i < ioctx->n_rw_ctx; i++) {
964 		struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i];
965 
966 		rdma_rw_ctx_destroy(&ctx->rw, ch->qp, ch->sport->port,
967 				ctx->sg, ctx->nents, dir);
968 		target_free_sgl(ctx->sg, ctx->nents);
969 	}
970 
971 	if (ioctx->rw_ctxs != &ioctx->s_rw_ctx)
972 		kfree(ioctx->rw_ctxs);
973 }
974 
975 static inline void *srpt_get_desc_buf(struct srp_cmd *srp_cmd)
976 {
977 	/*
978 	 * The pointer computations below will only be compiled correctly
979 	 * if srp_cmd::add_data is declared as s8*, u8*, s8[] or u8[], so check
980 	 * whether srp_cmd::add_data has been declared as a byte pointer.
981 	 */
982 	BUILD_BUG_ON(!__same_type(srp_cmd->add_data[0], (s8)0) &&
983 		     !__same_type(srp_cmd->add_data[0], (u8)0));
984 
985 	/*
986 	 * According to the SRP spec, the lower two bits of the 'ADDITIONAL
987 	 * CDB LENGTH' field are reserved and the size in bytes of this field
988 	 * is four times the value specified in bits 3..7. Hence the "& ~3".
989 	 */
990 	return srp_cmd->add_data + (srp_cmd->add_cdb_len & ~3);
991 }
992 
993 /**
994  * srpt_get_desc_tbl - parse the data descriptors of a SRP_CMD request
995  * @recv_ioctx: I/O context associated with the received command @srp_cmd.
996  * @ioctx: I/O context that will be used for responding to the initiator.
997  * @srp_cmd: Pointer to the SRP_CMD request data.
998  * @dir: Pointer to the variable to which the transfer direction will be
999  *   written.
1000  * @sg: [out] scatterlist for the parsed SRP_CMD.
1001  * @sg_cnt: [out] length of @sg.
1002  * @data_len: Pointer to the variable to which the total data length of all
1003  *   descriptors in the SRP_CMD request will be written.
1004  * @imm_data_offset: [in] Offset in SRP_CMD requests at which immediate data
1005  *   starts.
1006  *
1007  * This function initializes ioctx->nrbuf and ioctx->r_bufs.
1008  *
1009  * Returns -EINVAL when the SRP_CMD request contains inconsistent descriptors;
1010  * -ENOMEM when memory allocation fails and zero upon success.
1011  */
1012 static int srpt_get_desc_tbl(struct srpt_recv_ioctx *recv_ioctx,
1013 		struct srpt_send_ioctx *ioctx,
1014 		struct srp_cmd *srp_cmd, enum dma_data_direction *dir,
1015 		struct scatterlist **sg, unsigned int *sg_cnt, u64 *data_len,
1016 		u16 imm_data_offset)
1017 {
1018 	BUG_ON(!dir);
1019 	BUG_ON(!data_len);
1020 
1021 	/*
1022 	 * The lower four bits of the buffer format field contain the DATA-IN
1023 	 * buffer descriptor format, and the highest four bits contain the
1024 	 * DATA-OUT buffer descriptor format.
1025 	 */
1026 	if (srp_cmd->buf_fmt & 0xf)
1027 		/* DATA-IN: transfer data from target to initiator (read). */
1028 		*dir = DMA_FROM_DEVICE;
1029 	else if (srp_cmd->buf_fmt >> 4)
1030 		/* DATA-OUT: transfer data from initiator to target (write). */
1031 		*dir = DMA_TO_DEVICE;
1032 	else
1033 		*dir = DMA_NONE;
1034 
1035 	/* initialize data_direction early as srpt_alloc_rw_ctxs needs it */
1036 	ioctx->cmd.data_direction = *dir;
1037 
1038 	if (((srp_cmd->buf_fmt & 0xf) == SRP_DATA_DESC_DIRECT) ||
1039 	    ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_DIRECT)) {
1040 		struct srp_direct_buf *db = srpt_get_desc_buf(srp_cmd);
1041 
1042 		*data_len = be32_to_cpu(db->len);
1043 		return srpt_alloc_rw_ctxs(ioctx, db, 1, sg, sg_cnt);
1044 	} else if (((srp_cmd->buf_fmt & 0xf) == SRP_DATA_DESC_INDIRECT) ||
1045 		   ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_INDIRECT)) {
1046 		struct srp_indirect_buf *idb = srpt_get_desc_buf(srp_cmd);
1047 		int nbufs = be32_to_cpu(idb->table_desc.len) /
1048 				sizeof(struct srp_direct_buf);
1049 
1050 		if (nbufs >
1051 		    (srp_cmd->data_out_desc_cnt + srp_cmd->data_in_desc_cnt)) {
1052 			pr_err("received unsupported SRP_CMD request type (%u out + %u in != %u / %zu)\n",
1053 			       srp_cmd->data_out_desc_cnt,
1054 			       srp_cmd->data_in_desc_cnt,
1055 			       be32_to_cpu(idb->table_desc.len),
1056 			       sizeof(struct srp_direct_buf));
1057 			return -EINVAL;
1058 		}
1059 
1060 		*data_len = be32_to_cpu(idb->len);
1061 		return srpt_alloc_rw_ctxs(ioctx, idb->desc_list, nbufs,
1062 				sg, sg_cnt);
1063 	} else if ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_IMM) {
1064 		struct srp_imm_buf *imm_buf = srpt_get_desc_buf(srp_cmd);
1065 		void *data = (void *)srp_cmd + imm_data_offset;
1066 		uint32_t len = be32_to_cpu(imm_buf->len);
1067 		uint32_t req_size = imm_data_offset + len;
1068 
1069 		if (req_size > srp_max_req_size) {
1070 			pr_err("Immediate data (length %d + %d) exceeds request size %d\n",
1071 			       imm_data_offset, len, srp_max_req_size);
1072 			return -EINVAL;
1073 		}
1074 		if (recv_ioctx->byte_len < req_size) {
1075 			pr_err("Received too few data - %d < %d\n",
1076 			       recv_ioctx->byte_len, req_size);
1077 			return -EIO;
1078 		}
1079 		/*
1080 		 * The immediate data buffer descriptor must occur before the
1081 		 * immediate data itself.
1082 		 */
1083 		if ((void *)(imm_buf + 1) > (void *)data) {
1084 			pr_err("Received invalid write request\n");
1085 			return -EINVAL;
1086 		}
1087 		*data_len = len;
1088 		ioctx->recv_ioctx = recv_ioctx;
1089 		if ((uintptr_t)data & 511) {
1090 			pr_warn_once("Internal error - the receive buffers are not aligned properly.\n");
1091 			return -EINVAL;
1092 		}
1093 		sg_init_one(&ioctx->imm_sg, data, len);
1094 		*sg = &ioctx->imm_sg;
1095 		*sg_cnt = 1;
1096 		return 0;
1097 	} else {
1098 		*data_len = 0;
1099 		return 0;
1100 	}
1101 }
1102 
1103 /**
1104  * srpt_init_ch_qp - initialize queue pair attributes
1105  * @ch: SRPT RDMA channel.
1106  * @qp: Queue pair pointer.
1107  *
1108  * Initialized the attributes of queue pair 'qp' by allowing local write,
1109  * remote read and remote write. Also transitions 'qp' to state IB_QPS_INIT.
1110  */
1111 static int srpt_init_ch_qp(struct srpt_rdma_ch *ch, struct ib_qp *qp)
1112 {
1113 	struct ib_qp_attr *attr;
1114 	int ret;
1115 
1116 	WARN_ON_ONCE(ch->using_rdma_cm);
1117 
1118 	attr = kzalloc(sizeof(*attr), GFP_KERNEL);
1119 	if (!attr)
1120 		return -ENOMEM;
1121 
1122 	attr->qp_state = IB_QPS_INIT;
1123 	attr->qp_access_flags = IB_ACCESS_LOCAL_WRITE;
1124 	attr->port_num = ch->sport->port;
1125 
1126 	ret = ib_find_cached_pkey(ch->sport->sdev->device, ch->sport->port,
1127 				  ch->pkey, &attr->pkey_index);
1128 	if (ret < 0)
1129 		pr_err("Translating pkey %#x failed (%d) - using index 0\n",
1130 		       ch->pkey, ret);
1131 
1132 	ret = ib_modify_qp(qp, attr,
1133 			   IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PORT |
1134 			   IB_QP_PKEY_INDEX);
1135 
1136 	kfree(attr);
1137 	return ret;
1138 }
1139 
1140 /**
1141  * srpt_ch_qp_rtr - change the state of a channel to 'ready to receive' (RTR)
1142  * @ch: channel of the queue pair.
1143  * @qp: queue pair to change the state of.
1144  *
1145  * Returns zero upon success and a negative value upon failure.
1146  *
1147  * Note: currently a struct ib_qp_attr takes 136 bytes on a 64-bit system.
1148  * If this structure ever becomes larger, it might be necessary to allocate
1149  * it dynamically instead of on the stack.
1150  */
1151 static int srpt_ch_qp_rtr(struct srpt_rdma_ch *ch, struct ib_qp *qp)
1152 {
1153 	struct ib_qp_attr qp_attr;
1154 	int attr_mask;
1155 	int ret;
1156 
1157 	WARN_ON_ONCE(ch->using_rdma_cm);
1158 
1159 	qp_attr.qp_state = IB_QPS_RTR;
1160 	ret = ib_cm_init_qp_attr(ch->ib_cm.cm_id, &qp_attr, &attr_mask);
1161 	if (ret)
1162 		goto out;
1163 
1164 	qp_attr.max_dest_rd_atomic = 4;
1165 
1166 	ret = ib_modify_qp(qp, &qp_attr, attr_mask);
1167 
1168 out:
1169 	return ret;
1170 }
1171 
1172 /**
1173  * srpt_ch_qp_rts - change the state of a channel to 'ready to send' (RTS)
1174  * @ch: channel of the queue pair.
1175  * @qp: queue pair to change the state of.
1176  *
1177  * Returns zero upon success and a negative value upon failure.
1178  *
1179  * Note: currently a struct ib_qp_attr takes 136 bytes on a 64-bit system.
1180  * If this structure ever becomes larger, it might be necessary to allocate
1181  * it dynamically instead of on the stack.
1182  */
1183 static int srpt_ch_qp_rts(struct srpt_rdma_ch *ch, struct ib_qp *qp)
1184 {
1185 	struct ib_qp_attr qp_attr;
1186 	int attr_mask;
1187 	int ret;
1188 
1189 	qp_attr.qp_state = IB_QPS_RTS;
1190 	ret = ib_cm_init_qp_attr(ch->ib_cm.cm_id, &qp_attr, &attr_mask);
1191 	if (ret)
1192 		goto out;
1193 
1194 	qp_attr.max_rd_atomic = 4;
1195 
1196 	ret = ib_modify_qp(qp, &qp_attr, attr_mask);
1197 
1198 out:
1199 	return ret;
1200 }
1201 
1202 /**
1203  * srpt_ch_qp_err - set the channel queue pair state to 'error'
1204  * @ch: SRPT RDMA channel.
1205  */
1206 static int srpt_ch_qp_err(struct srpt_rdma_ch *ch)
1207 {
1208 	struct ib_qp_attr qp_attr;
1209 
1210 	qp_attr.qp_state = IB_QPS_ERR;
1211 	return ib_modify_qp(ch->qp, &qp_attr, IB_QP_STATE);
1212 }
1213 
1214 /**
1215  * srpt_get_send_ioctx - obtain an I/O context for sending to the initiator
1216  * @ch: SRPT RDMA channel.
1217  */
1218 static struct srpt_send_ioctx *srpt_get_send_ioctx(struct srpt_rdma_ch *ch)
1219 {
1220 	struct srpt_send_ioctx *ioctx;
1221 	int tag, cpu;
1222 
1223 	BUG_ON(!ch);
1224 
1225 	tag = sbitmap_queue_get(&ch->sess->sess_tag_pool, &cpu);
1226 	if (tag < 0)
1227 		return NULL;
1228 
1229 	ioctx = ch->ioctx_ring[tag];
1230 	BUG_ON(ioctx->ch != ch);
1231 	ioctx->state = SRPT_STATE_NEW;
1232 	WARN_ON_ONCE(ioctx->recv_ioctx);
1233 	ioctx->n_rdma = 0;
1234 	ioctx->n_rw_ctx = 0;
1235 	ioctx->queue_status_only = false;
1236 	/*
1237 	 * transport_init_se_cmd() does not initialize all fields, so do it
1238 	 * here.
1239 	 */
1240 	memset(&ioctx->cmd, 0, sizeof(ioctx->cmd));
1241 	memset(&ioctx->sense_data, 0, sizeof(ioctx->sense_data));
1242 	ioctx->cmd.map_tag = tag;
1243 	ioctx->cmd.map_cpu = cpu;
1244 
1245 	return ioctx;
1246 }
1247 
1248 /**
1249  * srpt_abort_cmd - abort a SCSI command
1250  * @ioctx:   I/O context associated with the SCSI command.
1251  */
1252 static int srpt_abort_cmd(struct srpt_send_ioctx *ioctx)
1253 {
1254 	enum srpt_command_state state;
1255 
1256 	BUG_ON(!ioctx);
1257 
1258 	/*
1259 	 * If the command is in a state where the target core is waiting for
1260 	 * the ib_srpt driver, change the state to the next state.
1261 	 */
1262 
1263 	state = ioctx->state;
1264 	switch (state) {
1265 	case SRPT_STATE_NEED_DATA:
1266 		ioctx->state = SRPT_STATE_DATA_IN;
1267 		break;
1268 	case SRPT_STATE_CMD_RSP_SENT:
1269 	case SRPT_STATE_MGMT_RSP_SENT:
1270 		ioctx->state = SRPT_STATE_DONE;
1271 		break;
1272 	default:
1273 		WARN_ONCE(true, "%s: unexpected I/O context state %d\n",
1274 			  __func__, state);
1275 		break;
1276 	}
1277 
1278 	pr_debug("Aborting cmd with state %d -> %d and tag %lld\n", state,
1279 		 ioctx->state, ioctx->cmd.tag);
1280 
1281 	switch (state) {
1282 	case SRPT_STATE_NEW:
1283 	case SRPT_STATE_DATA_IN:
1284 	case SRPT_STATE_MGMT:
1285 	case SRPT_STATE_DONE:
1286 		/*
1287 		 * Do nothing - defer abort processing until
1288 		 * srpt_queue_response() is invoked.
1289 		 */
1290 		break;
1291 	case SRPT_STATE_NEED_DATA:
1292 		pr_debug("tag %#llx: RDMA read error\n", ioctx->cmd.tag);
1293 		transport_generic_request_failure(&ioctx->cmd,
1294 					TCM_CHECK_CONDITION_ABORT_CMD);
1295 		break;
1296 	case SRPT_STATE_CMD_RSP_SENT:
1297 		/*
1298 		 * SRP_RSP sending failed or the SRP_RSP send completion has
1299 		 * not been received in time.
1300 		 */
1301 		transport_generic_free_cmd(&ioctx->cmd, 0);
1302 		break;
1303 	case SRPT_STATE_MGMT_RSP_SENT:
1304 		transport_generic_free_cmd(&ioctx->cmd, 0);
1305 		break;
1306 	default:
1307 		WARN(1, "Unexpected command state (%d)", state);
1308 		break;
1309 	}
1310 
1311 	return state;
1312 }
1313 
1314 /**
1315  * srpt_rdma_read_done - RDMA read completion callback
1316  * @cq: Completion queue.
1317  * @wc: Work completion.
1318  *
1319  * XXX: what is now target_execute_cmd used to be asynchronous, and unmapping
1320  * the data that has been transferred via IB RDMA had to be postponed until the
1321  * check_stop_free() callback.  None of this is necessary anymore and needs to
1322  * be cleaned up.
1323  */
1324 static void srpt_rdma_read_done(struct ib_cq *cq, struct ib_wc *wc)
1325 {
1326 	struct srpt_rdma_ch *ch = wc->qp->qp_context;
1327 	struct srpt_send_ioctx *ioctx =
1328 		container_of(wc->wr_cqe, struct srpt_send_ioctx, rdma_cqe);
1329 
1330 	WARN_ON(ioctx->n_rdma <= 0);
1331 	atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
1332 	ioctx->n_rdma = 0;
1333 
1334 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
1335 		pr_info("RDMA_READ for ioctx 0x%p failed with status %d\n",
1336 			ioctx, wc->status);
1337 		srpt_abort_cmd(ioctx);
1338 		return;
1339 	}
1340 
1341 	if (srpt_test_and_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA,
1342 					SRPT_STATE_DATA_IN))
1343 		target_execute_cmd(&ioctx->cmd);
1344 	else
1345 		pr_err("%s[%d]: wrong state = %d\n", __func__,
1346 		       __LINE__, ioctx->state);
1347 }
1348 
1349 /**
1350  * srpt_build_cmd_rsp - build a SRP_RSP response
1351  * @ch: RDMA channel through which the request has been received.
1352  * @ioctx: I/O context associated with the SRP_CMD request. The response will
1353  *   be built in the buffer ioctx->buf points at and hence this function will
1354  *   overwrite the request data.
1355  * @tag: tag of the request for which this response is being generated.
1356  * @status: value for the STATUS field of the SRP_RSP information unit.
1357  *
1358  * Returns the size in bytes of the SRP_RSP response.
1359  *
1360  * An SRP_RSP response contains a SCSI status or service response. See also
1361  * section 6.9 in the SRP r16a document for the format of an SRP_RSP
1362  * response. See also SPC-2 for more information about sense data.
1363  */
1364 static int srpt_build_cmd_rsp(struct srpt_rdma_ch *ch,
1365 			      struct srpt_send_ioctx *ioctx, u64 tag,
1366 			      int status)
1367 {
1368 	struct se_cmd *cmd = &ioctx->cmd;
1369 	struct srp_rsp *srp_rsp;
1370 	const u8 *sense_data;
1371 	int sense_data_len, max_sense_len;
1372 	u32 resid = cmd->residual_count;
1373 
1374 	/*
1375 	 * The lowest bit of all SAM-3 status codes is zero (see also
1376 	 * paragraph 5.3 in SAM-3).
1377 	 */
1378 	WARN_ON(status & 1);
1379 
1380 	srp_rsp = ioctx->ioctx.buf;
1381 	BUG_ON(!srp_rsp);
1382 
1383 	sense_data = ioctx->sense_data;
1384 	sense_data_len = ioctx->cmd.scsi_sense_length;
1385 	WARN_ON(sense_data_len > sizeof(ioctx->sense_data));
1386 
1387 	memset(srp_rsp, 0, sizeof(*srp_rsp));
1388 	srp_rsp->opcode = SRP_RSP;
1389 	srp_rsp->req_lim_delta =
1390 		cpu_to_be32(1 + atomic_xchg(&ch->req_lim_delta, 0));
1391 	srp_rsp->tag = tag;
1392 	srp_rsp->status = status;
1393 
1394 	if (cmd->se_cmd_flags & SCF_UNDERFLOW_BIT) {
1395 		if (cmd->data_direction == DMA_TO_DEVICE) {
1396 			/* residual data from an underflow write */
1397 			srp_rsp->flags = SRP_RSP_FLAG_DOUNDER;
1398 			srp_rsp->data_out_res_cnt = cpu_to_be32(resid);
1399 		} else if (cmd->data_direction == DMA_FROM_DEVICE) {
1400 			/* residual data from an underflow read */
1401 			srp_rsp->flags = SRP_RSP_FLAG_DIUNDER;
1402 			srp_rsp->data_in_res_cnt = cpu_to_be32(resid);
1403 		}
1404 	} else if (cmd->se_cmd_flags & SCF_OVERFLOW_BIT) {
1405 		if (cmd->data_direction == DMA_TO_DEVICE) {
1406 			/* residual data from an overflow write */
1407 			srp_rsp->flags = SRP_RSP_FLAG_DOOVER;
1408 			srp_rsp->data_out_res_cnt = cpu_to_be32(resid);
1409 		} else if (cmd->data_direction == DMA_FROM_DEVICE) {
1410 			/* residual data from an overflow read */
1411 			srp_rsp->flags = SRP_RSP_FLAG_DIOVER;
1412 			srp_rsp->data_in_res_cnt = cpu_to_be32(resid);
1413 		}
1414 	}
1415 
1416 	if (sense_data_len) {
1417 		BUILD_BUG_ON(MIN_MAX_RSP_SIZE <= sizeof(*srp_rsp));
1418 		max_sense_len = ch->max_ti_iu_len - sizeof(*srp_rsp);
1419 		if (sense_data_len > max_sense_len) {
1420 			pr_warn("truncated sense data from %d to %d bytes\n",
1421 				sense_data_len, max_sense_len);
1422 			sense_data_len = max_sense_len;
1423 		}
1424 
1425 		srp_rsp->flags |= SRP_RSP_FLAG_SNSVALID;
1426 		srp_rsp->sense_data_len = cpu_to_be32(sense_data_len);
1427 		memcpy(srp_rsp + 1, sense_data, sense_data_len);
1428 	}
1429 
1430 	return sizeof(*srp_rsp) + sense_data_len;
1431 }
1432 
1433 /**
1434  * srpt_build_tskmgmt_rsp - build a task management response
1435  * @ch:       RDMA channel through which the request has been received.
1436  * @ioctx:    I/O context in which the SRP_RSP response will be built.
1437  * @rsp_code: RSP_CODE that will be stored in the response.
1438  * @tag:      Tag of the request for which this response is being generated.
1439  *
1440  * Returns the size in bytes of the SRP_RSP response.
1441  *
1442  * An SRP_RSP response contains a SCSI status or service response. See also
1443  * section 6.9 in the SRP r16a document for the format of an SRP_RSP
1444  * response.
1445  */
1446 static int srpt_build_tskmgmt_rsp(struct srpt_rdma_ch *ch,
1447 				  struct srpt_send_ioctx *ioctx,
1448 				  u8 rsp_code, u64 tag)
1449 {
1450 	struct srp_rsp *srp_rsp;
1451 	int resp_data_len;
1452 	int resp_len;
1453 
1454 	resp_data_len = 4;
1455 	resp_len = sizeof(*srp_rsp) + resp_data_len;
1456 
1457 	srp_rsp = ioctx->ioctx.buf;
1458 	BUG_ON(!srp_rsp);
1459 	memset(srp_rsp, 0, sizeof(*srp_rsp));
1460 
1461 	srp_rsp->opcode = SRP_RSP;
1462 	srp_rsp->req_lim_delta =
1463 		cpu_to_be32(1 + atomic_xchg(&ch->req_lim_delta, 0));
1464 	srp_rsp->tag = tag;
1465 
1466 	srp_rsp->flags |= SRP_RSP_FLAG_RSPVALID;
1467 	srp_rsp->resp_data_len = cpu_to_be32(resp_data_len);
1468 	srp_rsp->data[3] = rsp_code;
1469 
1470 	return resp_len;
1471 }
1472 
1473 static int srpt_check_stop_free(struct se_cmd *cmd)
1474 {
1475 	struct srpt_send_ioctx *ioctx = container_of(cmd,
1476 				struct srpt_send_ioctx, cmd);
1477 
1478 	return target_put_sess_cmd(&ioctx->cmd);
1479 }
1480 
1481 /**
1482  * srpt_handle_cmd - process a SRP_CMD information unit
1483  * @ch: SRPT RDMA channel.
1484  * @recv_ioctx: Receive I/O context.
1485  * @send_ioctx: Send I/O context.
1486  */
1487 static void srpt_handle_cmd(struct srpt_rdma_ch *ch,
1488 			    struct srpt_recv_ioctx *recv_ioctx,
1489 			    struct srpt_send_ioctx *send_ioctx)
1490 {
1491 	struct se_cmd *cmd;
1492 	struct srp_cmd *srp_cmd;
1493 	struct scatterlist *sg = NULL;
1494 	unsigned sg_cnt = 0;
1495 	u64 data_len;
1496 	enum dma_data_direction dir;
1497 	int rc;
1498 
1499 	BUG_ON(!send_ioctx);
1500 
1501 	srp_cmd = recv_ioctx->ioctx.buf + recv_ioctx->ioctx.offset;
1502 	cmd = &send_ioctx->cmd;
1503 	cmd->tag = srp_cmd->tag;
1504 
1505 	switch (srp_cmd->task_attr) {
1506 	case SRP_CMD_SIMPLE_Q:
1507 		cmd->sam_task_attr = TCM_SIMPLE_TAG;
1508 		break;
1509 	case SRP_CMD_ORDERED_Q:
1510 	default:
1511 		cmd->sam_task_attr = TCM_ORDERED_TAG;
1512 		break;
1513 	case SRP_CMD_HEAD_OF_Q:
1514 		cmd->sam_task_attr = TCM_HEAD_TAG;
1515 		break;
1516 	case SRP_CMD_ACA:
1517 		cmd->sam_task_attr = TCM_ACA_TAG;
1518 		break;
1519 	}
1520 
1521 	rc = srpt_get_desc_tbl(recv_ioctx, send_ioctx, srp_cmd, &dir,
1522 			       &sg, &sg_cnt, &data_len, ch->imm_data_offset);
1523 	if (rc) {
1524 		if (rc != -EAGAIN) {
1525 			pr_err("0x%llx: parsing SRP descriptor table failed.\n",
1526 			       srp_cmd->tag);
1527 		}
1528 		goto busy;
1529 	}
1530 
1531 	rc = target_submit_cmd_map_sgls(cmd, ch->sess, srp_cmd->cdb,
1532 			       &send_ioctx->sense_data[0],
1533 			       scsilun_to_int(&srp_cmd->lun), data_len,
1534 			       TCM_SIMPLE_TAG, dir, TARGET_SCF_ACK_KREF,
1535 			       sg, sg_cnt, NULL, 0, NULL, 0);
1536 	if (rc != 0) {
1537 		pr_debug("target_submit_cmd() returned %d for tag %#llx\n", rc,
1538 			 srp_cmd->tag);
1539 		goto busy;
1540 	}
1541 	return;
1542 
1543 busy:
1544 	target_send_busy(cmd);
1545 }
1546 
1547 static int srp_tmr_to_tcm(int fn)
1548 {
1549 	switch (fn) {
1550 	case SRP_TSK_ABORT_TASK:
1551 		return TMR_ABORT_TASK;
1552 	case SRP_TSK_ABORT_TASK_SET:
1553 		return TMR_ABORT_TASK_SET;
1554 	case SRP_TSK_CLEAR_TASK_SET:
1555 		return TMR_CLEAR_TASK_SET;
1556 	case SRP_TSK_LUN_RESET:
1557 		return TMR_LUN_RESET;
1558 	case SRP_TSK_CLEAR_ACA:
1559 		return TMR_CLEAR_ACA;
1560 	default:
1561 		return -1;
1562 	}
1563 }
1564 
1565 /**
1566  * srpt_handle_tsk_mgmt - process a SRP_TSK_MGMT information unit
1567  * @ch: SRPT RDMA channel.
1568  * @recv_ioctx: Receive I/O context.
1569  * @send_ioctx: Send I/O context.
1570  *
1571  * Returns 0 if and only if the request will be processed by the target core.
1572  *
1573  * For more information about SRP_TSK_MGMT information units, see also section
1574  * 6.7 in the SRP r16a document.
1575  */
1576 static void srpt_handle_tsk_mgmt(struct srpt_rdma_ch *ch,
1577 				 struct srpt_recv_ioctx *recv_ioctx,
1578 				 struct srpt_send_ioctx *send_ioctx)
1579 {
1580 	struct srp_tsk_mgmt *srp_tsk;
1581 	struct se_cmd *cmd;
1582 	struct se_session *sess = ch->sess;
1583 	int tcm_tmr;
1584 	int rc;
1585 
1586 	BUG_ON(!send_ioctx);
1587 
1588 	srp_tsk = recv_ioctx->ioctx.buf + recv_ioctx->ioctx.offset;
1589 	cmd = &send_ioctx->cmd;
1590 
1591 	pr_debug("recv tsk_mgmt fn %d for task_tag %lld and cmd tag %lld ch %p sess %p\n",
1592 		 srp_tsk->tsk_mgmt_func, srp_tsk->task_tag, srp_tsk->tag, ch,
1593 		 ch->sess);
1594 
1595 	srpt_set_cmd_state(send_ioctx, SRPT_STATE_MGMT);
1596 	send_ioctx->cmd.tag = srp_tsk->tag;
1597 	tcm_tmr = srp_tmr_to_tcm(srp_tsk->tsk_mgmt_func);
1598 	rc = target_submit_tmr(&send_ioctx->cmd, sess, NULL,
1599 			       scsilun_to_int(&srp_tsk->lun), srp_tsk, tcm_tmr,
1600 			       GFP_KERNEL, srp_tsk->task_tag,
1601 			       TARGET_SCF_ACK_KREF);
1602 	if (rc != 0) {
1603 		send_ioctx->cmd.se_tmr_req->response = TMR_FUNCTION_REJECTED;
1604 		cmd->se_tfo->queue_tm_rsp(cmd);
1605 	}
1606 	return;
1607 }
1608 
1609 /**
1610  * srpt_handle_new_iu - process a newly received information unit
1611  * @ch:    RDMA channel through which the information unit has been received.
1612  * @recv_ioctx: Receive I/O context associated with the information unit.
1613  */
1614 static bool
1615 srpt_handle_new_iu(struct srpt_rdma_ch *ch, struct srpt_recv_ioctx *recv_ioctx)
1616 {
1617 	struct srpt_send_ioctx *send_ioctx = NULL;
1618 	struct srp_cmd *srp_cmd;
1619 	bool res = false;
1620 	u8 opcode;
1621 
1622 	BUG_ON(!ch);
1623 	BUG_ON(!recv_ioctx);
1624 
1625 	if (unlikely(ch->state == CH_CONNECTING))
1626 		goto push;
1627 
1628 	ib_dma_sync_single_for_cpu(ch->sport->sdev->device,
1629 				   recv_ioctx->ioctx.dma,
1630 				   recv_ioctx->ioctx.offset + srp_max_req_size,
1631 				   DMA_FROM_DEVICE);
1632 
1633 	srp_cmd = recv_ioctx->ioctx.buf + recv_ioctx->ioctx.offset;
1634 	opcode = srp_cmd->opcode;
1635 	if (opcode == SRP_CMD || opcode == SRP_TSK_MGMT) {
1636 		send_ioctx = srpt_get_send_ioctx(ch);
1637 		if (unlikely(!send_ioctx))
1638 			goto push;
1639 	}
1640 
1641 	if (!list_empty(&recv_ioctx->wait_list)) {
1642 		WARN_ON_ONCE(!ch->processing_wait_list);
1643 		list_del_init(&recv_ioctx->wait_list);
1644 	}
1645 
1646 	switch (opcode) {
1647 	case SRP_CMD:
1648 		srpt_handle_cmd(ch, recv_ioctx, send_ioctx);
1649 		break;
1650 	case SRP_TSK_MGMT:
1651 		srpt_handle_tsk_mgmt(ch, recv_ioctx, send_ioctx);
1652 		break;
1653 	case SRP_I_LOGOUT:
1654 		pr_err("Not yet implemented: SRP_I_LOGOUT\n");
1655 		break;
1656 	case SRP_CRED_RSP:
1657 		pr_debug("received SRP_CRED_RSP\n");
1658 		break;
1659 	case SRP_AER_RSP:
1660 		pr_debug("received SRP_AER_RSP\n");
1661 		break;
1662 	case SRP_RSP:
1663 		pr_err("Received SRP_RSP\n");
1664 		break;
1665 	default:
1666 		pr_err("received IU with unknown opcode 0x%x\n", opcode);
1667 		break;
1668 	}
1669 
1670 	if (!send_ioctx || !send_ioctx->recv_ioctx)
1671 		srpt_post_recv(ch->sport->sdev, ch, recv_ioctx);
1672 	res = true;
1673 
1674 out:
1675 	return res;
1676 
1677 push:
1678 	if (list_empty(&recv_ioctx->wait_list)) {
1679 		WARN_ON_ONCE(ch->processing_wait_list);
1680 		list_add_tail(&recv_ioctx->wait_list, &ch->cmd_wait_list);
1681 	}
1682 	goto out;
1683 }
1684 
1685 static void srpt_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1686 {
1687 	struct srpt_rdma_ch *ch = wc->qp->qp_context;
1688 	struct srpt_recv_ioctx *ioctx =
1689 		container_of(wc->wr_cqe, struct srpt_recv_ioctx, ioctx.cqe);
1690 
1691 	if (wc->status == IB_WC_SUCCESS) {
1692 		int req_lim;
1693 
1694 		req_lim = atomic_dec_return(&ch->req_lim);
1695 		if (unlikely(req_lim < 0))
1696 			pr_err("req_lim = %d < 0\n", req_lim);
1697 		ioctx->byte_len = wc->byte_len;
1698 		srpt_handle_new_iu(ch, ioctx);
1699 	} else {
1700 		pr_info_ratelimited("receiving failed for ioctx %p with status %d\n",
1701 				    ioctx, wc->status);
1702 	}
1703 }
1704 
1705 /*
1706  * This function must be called from the context in which RDMA completions are
1707  * processed because it accesses the wait list without protection against
1708  * access from other threads.
1709  */
1710 static void srpt_process_wait_list(struct srpt_rdma_ch *ch)
1711 {
1712 	struct srpt_recv_ioctx *recv_ioctx, *tmp;
1713 
1714 	WARN_ON_ONCE(ch->state == CH_CONNECTING);
1715 
1716 	if (list_empty(&ch->cmd_wait_list))
1717 		return;
1718 
1719 	WARN_ON_ONCE(ch->processing_wait_list);
1720 	ch->processing_wait_list = true;
1721 	list_for_each_entry_safe(recv_ioctx, tmp, &ch->cmd_wait_list,
1722 				 wait_list) {
1723 		if (!srpt_handle_new_iu(ch, recv_ioctx))
1724 			break;
1725 	}
1726 	ch->processing_wait_list = false;
1727 }
1728 
1729 /**
1730  * srpt_send_done - send completion callback
1731  * @cq: Completion queue.
1732  * @wc: Work completion.
1733  *
1734  * Note: Although this has not yet been observed during tests, at least in
1735  * theory it is possible that the srpt_get_send_ioctx() call invoked by
1736  * srpt_handle_new_iu() fails. This is possible because the req_lim_delta
1737  * value in each response is set to one, and it is possible that this response
1738  * makes the initiator send a new request before the send completion for that
1739  * response has been processed. This could e.g. happen if the call to
1740  * srpt_put_send_iotcx() is delayed because of a higher priority interrupt or
1741  * if IB retransmission causes generation of the send completion to be
1742  * delayed. Incoming information units for which srpt_get_send_ioctx() fails
1743  * are queued on cmd_wait_list. The code below processes these delayed
1744  * requests one at a time.
1745  */
1746 static void srpt_send_done(struct ib_cq *cq, struct ib_wc *wc)
1747 {
1748 	struct srpt_rdma_ch *ch = wc->qp->qp_context;
1749 	struct srpt_send_ioctx *ioctx =
1750 		container_of(wc->wr_cqe, struct srpt_send_ioctx, ioctx.cqe);
1751 	enum srpt_command_state state;
1752 
1753 	state = srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
1754 
1755 	WARN_ON(state != SRPT_STATE_CMD_RSP_SENT &&
1756 		state != SRPT_STATE_MGMT_RSP_SENT);
1757 
1758 	atomic_add(1 + ioctx->n_rdma, &ch->sq_wr_avail);
1759 
1760 	if (wc->status != IB_WC_SUCCESS)
1761 		pr_info("sending response for ioctx 0x%p failed with status %d\n",
1762 			ioctx, wc->status);
1763 
1764 	if (state != SRPT_STATE_DONE) {
1765 		transport_generic_free_cmd(&ioctx->cmd, 0);
1766 	} else {
1767 		pr_err("IB completion has been received too late for wr_id = %u.\n",
1768 		       ioctx->ioctx.index);
1769 	}
1770 
1771 	srpt_process_wait_list(ch);
1772 }
1773 
1774 /**
1775  * srpt_create_ch_ib - create receive and send completion queues
1776  * @ch: SRPT RDMA channel.
1777  */
1778 static int srpt_create_ch_ib(struct srpt_rdma_ch *ch)
1779 {
1780 	struct ib_qp_init_attr *qp_init;
1781 	struct srpt_port *sport = ch->sport;
1782 	struct srpt_device *sdev = sport->sdev;
1783 	const struct ib_device_attr *attrs = &sdev->device->attrs;
1784 	int sq_size = sport->port_attrib.srp_sq_size;
1785 	int i, ret;
1786 
1787 	WARN_ON(ch->rq_size < 1);
1788 
1789 	ret = -ENOMEM;
1790 	qp_init = kzalloc(sizeof(*qp_init), GFP_KERNEL);
1791 	if (!qp_init)
1792 		goto out;
1793 
1794 retry:
1795 	ch->cq = ib_cq_pool_get(sdev->device, ch->rq_size + sq_size, -1,
1796 				 IB_POLL_WORKQUEUE);
1797 	if (IS_ERR(ch->cq)) {
1798 		ret = PTR_ERR(ch->cq);
1799 		pr_err("failed to create CQ cqe= %d ret= %d\n",
1800 		       ch->rq_size + sq_size, ret);
1801 		goto out;
1802 	}
1803 	ch->cq_size = ch->rq_size + sq_size;
1804 
1805 	qp_init->qp_context = (void *)ch;
1806 	qp_init->event_handler
1807 		= (void(*)(struct ib_event *, void*))srpt_qp_event;
1808 	qp_init->send_cq = ch->cq;
1809 	qp_init->recv_cq = ch->cq;
1810 	qp_init->sq_sig_type = IB_SIGNAL_REQ_WR;
1811 	qp_init->qp_type = IB_QPT_RC;
1812 	/*
1813 	 * We divide up our send queue size into half SEND WRs to send the
1814 	 * completions, and half R/W contexts to actually do the RDMA
1815 	 * READ/WRITE transfers.  Note that we need to allocate CQ slots for
1816 	 * both both, as RDMA contexts will also post completions for the
1817 	 * RDMA READ case.
1818 	 */
1819 	qp_init->cap.max_send_wr = min(sq_size / 2, attrs->max_qp_wr);
1820 	qp_init->cap.max_rdma_ctxs = sq_size / 2;
1821 	qp_init->cap.max_send_sge = attrs->max_send_sge;
1822 	qp_init->cap.max_recv_sge = 1;
1823 	qp_init->port_num = ch->sport->port;
1824 	if (sdev->use_srq)
1825 		qp_init->srq = sdev->srq;
1826 	else
1827 		qp_init->cap.max_recv_wr = ch->rq_size;
1828 
1829 	if (ch->using_rdma_cm) {
1830 		ret = rdma_create_qp(ch->rdma_cm.cm_id, sdev->pd, qp_init);
1831 		ch->qp = ch->rdma_cm.cm_id->qp;
1832 	} else {
1833 		ch->qp = ib_create_qp(sdev->pd, qp_init);
1834 		if (!IS_ERR(ch->qp)) {
1835 			ret = srpt_init_ch_qp(ch, ch->qp);
1836 			if (ret)
1837 				ib_destroy_qp(ch->qp);
1838 		} else {
1839 			ret = PTR_ERR(ch->qp);
1840 		}
1841 	}
1842 	if (ret) {
1843 		bool retry = sq_size > MIN_SRPT_SQ_SIZE;
1844 
1845 		if (retry) {
1846 			pr_debug("failed to create queue pair with sq_size = %d (%d) - retrying\n",
1847 				 sq_size, ret);
1848 			ib_cq_pool_put(ch->cq, ch->cq_size);
1849 			sq_size = max(sq_size / 2, MIN_SRPT_SQ_SIZE);
1850 			goto retry;
1851 		} else {
1852 			pr_err("failed to create queue pair with sq_size = %d (%d)\n",
1853 			       sq_size, ret);
1854 			goto err_destroy_cq;
1855 		}
1856 	}
1857 
1858 	atomic_set(&ch->sq_wr_avail, qp_init->cap.max_send_wr);
1859 
1860 	pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d ch= %p\n",
1861 		 __func__, ch->cq->cqe, qp_init->cap.max_send_sge,
1862 		 qp_init->cap.max_send_wr, ch);
1863 
1864 	if (!sdev->use_srq)
1865 		for (i = 0; i < ch->rq_size; i++)
1866 			srpt_post_recv(sdev, ch, ch->ioctx_recv_ring[i]);
1867 
1868 out:
1869 	kfree(qp_init);
1870 	return ret;
1871 
1872 err_destroy_cq:
1873 	ch->qp = NULL;
1874 	ib_cq_pool_put(ch->cq, ch->cq_size);
1875 	goto out;
1876 }
1877 
1878 static void srpt_destroy_ch_ib(struct srpt_rdma_ch *ch)
1879 {
1880 	ib_destroy_qp(ch->qp);
1881 	ib_cq_pool_put(ch->cq, ch->cq_size);
1882 }
1883 
1884 /**
1885  * srpt_close_ch - close a RDMA channel
1886  * @ch: SRPT RDMA channel.
1887  *
1888  * Make sure all resources associated with the channel will be deallocated at
1889  * an appropriate time.
1890  *
1891  * Returns true if and only if the channel state has been modified into
1892  * CH_DRAINING.
1893  */
1894 static bool srpt_close_ch(struct srpt_rdma_ch *ch)
1895 {
1896 	int ret;
1897 
1898 	if (!srpt_set_ch_state(ch, CH_DRAINING)) {
1899 		pr_debug("%s: already closed\n", ch->sess_name);
1900 		return false;
1901 	}
1902 
1903 	kref_get(&ch->kref);
1904 
1905 	ret = srpt_ch_qp_err(ch);
1906 	if (ret < 0)
1907 		pr_err("%s-%d: changing queue pair into error state failed: %d\n",
1908 		       ch->sess_name, ch->qp->qp_num, ret);
1909 
1910 	ret = srpt_zerolength_write(ch);
1911 	if (ret < 0) {
1912 		pr_err("%s-%d: queuing zero-length write failed: %d\n",
1913 		       ch->sess_name, ch->qp->qp_num, ret);
1914 		if (srpt_set_ch_state(ch, CH_DISCONNECTED))
1915 			schedule_work(&ch->release_work);
1916 		else
1917 			WARN_ON_ONCE(true);
1918 	}
1919 
1920 	kref_put(&ch->kref, srpt_free_ch);
1921 
1922 	return true;
1923 }
1924 
1925 /*
1926  * Change the channel state into CH_DISCONNECTING. If a channel has not yet
1927  * reached the connected state, close it. If a channel is in the connected
1928  * state, send a DREQ. If a DREQ has been received, send a DREP. Note: it is
1929  * the responsibility of the caller to ensure that this function is not
1930  * invoked concurrently with the code that accepts a connection. This means
1931  * that this function must either be invoked from inside a CM callback
1932  * function or that it must be invoked with the srpt_port.mutex held.
1933  */
1934 static int srpt_disconnect_ch(struct srpt_rdma_ch *ch)
1935 {
1936 	int ret;
1937 
1938 	if (!srpt_set_ch_state(ch, CH_DISCONNECTING))
1939 		return -ENOTCONN;
1940 
1941 	if (ch->using_rdma_cm) {
1942 		ret = rdma_disconnect(ch->rdma_cm.cm_id);
1943 	} else {
1944 		ret = ib_send_cm_dreq(ch->ib_cm.cm_id, NULL, 0);
1945 		if (ret < 0)
1946 			ret = ib_send_cm_drep(ch->ib_cm.cm_id, NULL, 0);
1947 	}
1948 
1949 	if (ret < 0 && srpt_close_ch(ch))
1950 		ret = 0;
1951 
1952 	return ret;
1953 }
1954 
1955 /* Send DREQ and wait for DREP. */
1956 static void srpt_disconnect_ch_sync(struct srpt_rdma_ch *ch)
1957 {
1958 	DECLARE_COMPLETION_ONSTACK(closed);
1959 	struct srpt_port *sport = ch->sport;
1960 
1961 	pr_debug("ch %s-%d state %d\n", ch->sess_name, ch->qp->qp_num,
1962 		 ch->state);
1963 
1964 	ch->closed = &closed;
1965 
1966 	mutex_lock(&sport->mutex);
1967 	srpt_disconnect_ch(ch);
1968 	mutex_unlock(&sport->mutex);
1969 
1970 	while (wait_for_completion_timeout(&closed, 5 * HZ) == 0)
1971 		pr_info("%s(%s-%d state %d): still waiting ...\n", __func__,
1972 			ch->sess_name, ch->qp->qp_num, ch->state);
1973 
1974 }
1975 
1976 static void __srpt_close_all_ch(struct srpt_port *sport)
1977 {
1978 	struct srpt_nexus *nexus;
1979 	struct srpt_rdma_ch *ch;
1980 
1981 	lockdep_assert_held(&sport->mutex);
1982 
1983 	list_for_each_entry(nexus, &sport->nexus_list, entry) {
1984 		list_for_each_entry(ch, &nexus->ch_list, list) {
1985 			if (srpt_disconnect_ch(ch) >= 0)
1986 				pr_info("Closing channel %s-%d because target %s_%d has been disabled\n",
1987 					ch->sess_name, ch->qp->qp_num,
1988 					dev_name(&sport->sdev->device->dev),
1989 					sport->port);
1990 			srpt_close_ch(ch);
1991 		}
1992 	}
1993 }
1994 
1995 /*
1996  * Look up (i_port_id, t_port_id) in sport->nexus_list. Create an entry if
1997  * it does not yet exist.
1998  */
1999 static struct srpt_nexus *srpt_get_nexus(struct srpt_port *sport,
2000 					 const u8 i_port_id[16],
2001 					 const u8 t_port_id[16])
2002 {
2003 	struct srpt_nexus *nexus = NULL, *tmp_nexus = NULL, *n;
2004 
2005 	for (;;) {
2006 		mutex_lock(&sport->mutex);
2007 		list_for_each_entry(n, &sport->nexus_list, entry) {
2008 			if (memcmp(n->i_port_id, i_port_id, 16) == 0 &&
2009 			    memcmp(n->t_port_id, t_port_id, 16) == 0) {
2010 				nexus = n;
2011 				break;
2012 			}
2013 		}
2014 		if (!nexus && tmp_nexus) {
2015 			list_add_tail_rcu(&tmp_nexus->entry,
2016 					  &sport->nexus_list);
2017 			swap(nexus, tmp_nexus);
2018 		}
2019 		mutex_unlock(&sport->mutex);
2020 
2021 		if (nexus)
2022 			break;
2023 		tmp_nexus = kzalloc(sizeof(*nexus), GFP_KERNEL);
2024 		if (!tmp_nexus) {
2025 			nexus = ERR_PTR(-ENOMEM);
2026 			break;
2027 		}
2028 		INIT_LIST_HEAD(&tmp_nexus->ch_list);
2029 		memcpy(tmp_nexus->i_port_id, i_port_id, 16);
2030 		memcpy(tmp_nexus->t_port_id, t_port_id, 16);
2031 	}
2032 
2033 	kfree(tmp_nexus);
2034 
2035 	return nexus;
2036 }
2037 
2038 static void srpt_set_enabled(struct srpt_port *sport, bool enabled)
2039 	__must_hold(&sport->mutex)
2040 {
2041 	lockdep_assert_held(&sport->mutex);
2042 
2043 	if (sport->enabled == enabled)
2044 		return;
2045 	sport->enabled = enabled;
2046 	if (!enabled)
2047 		__srpt_close_all_ch(sport);
2048 }
2049 
2050 static void srpt_drop_sport_ref(struct srpt_port *sport)
2051 {
2052 	if (atomic_dec_return(&sport->refcount) == 0 && sport->freed_channels)
2053 		complete(sport->freed_channels);
2054 }
2055 
2056 static void srpt_free_ch(struct kref *kref)
2057 {
2058 	struct srpt_rdma_ch *ch = container_of(kref, struct srpt_rdma_ch, kref);
2059 
2060 	srpt_drop_sport_ref(ch->sport);
2061 	kfree_rcu(ch, rcu);
2062 }
2063 
2064 /*
2065  * Shut down the SCSI target session, tell the connection manager to
2066  * disconnect the associated RDMA channel, transition the QP to the error
2067  * state and remove the channel from the channel list. This function is
2068  * typically called from inside srpt_zerolength_write_done(). Concurrent
2069  * srpt_zerolength_write() calls from inside srpt_close_ch() are possible
2070  * as long as the channel is on sport->nexus_list.
2071  */
2072 static void srpt_release_channel_work(struct work_struct *w)
2073 {
2074 	struct srpt_rdma_ch *ch;
2075 	struct srpt_device *sdev;
2076 	struct srpt_port *sport;
2077 	struct se_session *se_sess;
2078 
2079 	ch = container_of(w, struct srpt_rdma_ch, release_work);
2080 	pr_debug("%s-%d\n", ch->sess_name, ch->qp->qp_num);
2081 
2082 	sdev = ch->sport->sdev;
2083 	BUG_ON(!sdev);
2084 
2085 	se_sess = ch->sess;
2086 	BUG_ON(!se_sess);
2087 
2088 	target_stop_session(se_sess);
2089 	target_wait_for_sess_cmds(se_sess);
2090 
2091 	target_remove_session(se_sess);
2092 	ch->sess = NULL;
2093 
2094 	if (ch->using_rdma_cm)
2095 		rdma_destroy_id(ch->rdma_cm.cm_id);
2096 	else
2097 		ib_destroy_cm_id(ch->ib_cm.cm_id);
2098 
2099 	sport = ch->sport;
2100 	mutex_lock(&sport->mutex);
2101 	list_del_rcu(&ch->list);
2102 	mutex_unlock(&sport->mutex);
2103 
2104 	if (ch->closed)
2105 		complete(ch->closed);
2106 
2107 	srpt_destroy_ch_ib(ch);
2108 
2109 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
2110 			     ch->sport->sdev, ch->rq_size,
2111 			     ch->rsp_buf_cache, DMA_TO_DEVICE);
2112 
2113 	kmem_cache_destroy(ch->rsp_buf_cache);
2114 
2115 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_recv_ring,
2116 			     sdev, ch->rq_size,
2117 			     ch->req_buf_cache, DMA_FROM_DEVICE);
2118 
2119 	kmem_cache_destroy(ch->req_buf_cache);
2120 
2121 	kref_put(&ch->kref, srpt_free_ch);
2122 }
2123 
2124 /**
2125  * srpt_cm_req_recv - process the event IB_CM_REQ_RECEIVED
2126  * @sdev: HCA through which the login request was received.
2127  * @ib_cm_id: IB/CM connection identifier in case of IB/CM.
2128  * @rdma_cm_id: RDMA/CM connection identifier in case of RDMA/CM.
2129  * @port_num: Port through which the REQ message was received.
2130  * @pkey: P_Key of the incoming connection.
2131  * @req: SRP login request.
2132  * @src_addr: GID (IB/CM) or IP address (RDMA/CM) of the port that submitted
2133  * the login request.
2134  *
2135  * Ownership of the cm_id is transferred to the target session if this
2136  * function returns zero. Otherwise the caller remains the owner of cm_id.
2137  */
2138 static int srpt_cm_req_recv(struct srpt_device *const sdev,
2139 			    struct ib_cm_id *ib_cm_id,
2140 			    struct rdma_cm_id *rdma_cm_id,
2141 			    u8 port_num, __be16 pkey,
2142 			    const struct srp_login_req *req,
2143 			    const char *src_addr)
2144 {
2145 	struct srpt_port *sport = &sdev->port[port_num - 1];
2146 	struct srpt_nexus *nexus;
2147 	struct srp_login_rsp *rsp = NULL;
2148 	struct srp_login_rej *rej = NULL;
2149 	union {
2150 		struct rdma_conn_param rdma_cm;
2151 		struct ib_cm_rep_param ib_cm;
2152 	} *rep_param = NULL;
2153 	struct srpt_rdma_ch *ch = NULL;
2154 	char i_port_id[36];
2155 	u32 it_iu_len;
2156 	int i, tag_num, tag_size, ret;
2157 	struct srpt_tpg *stpg;
2158 
2159 	WARN_ON_ONCE(irqs_disabled());
2160 
2161 	it_iu_len = be32_to_cpu(req->req_it_iu_len);
2162 
2163 	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",
2164 		req->initiator_port_id, req->target_port_id, it_iu_len,
2165 		port_num, &sport->gid, be16_to_cpu(pkey));
2166 
2167 	nexus = srpt_get_nexus(sport, req->initiator_port_id,
2168 			       req->target_port_id);
2169 	if (IS_ERR(nexus)) {
2170 		ret = PTR_ERR(nexus);
2171 		goto out;
2172 	}
2173 
2174 	ret = -ENOMEM;
2175 	rsp = kzalloc(sizeof(*rsp), GFP_KERNEL);
2176 	rej = kzalloc(sizeof(*rej), GFP_KERNEL);
2177 	rep_param = kzalloc(sizeof(*rep_param), GFP_KERNEL);
2178 	if (!rsp || !rej || !rep_param)
2179 		goto out;
2180 
2181 	ret = -EINVAL;
2182 	if (it_iu_len > srp_max_req_size || it_iu_len < 64) {
2183 		rej->reason = cpu_to_be32(
2184 				SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE);
2185 		pr_err("rejected SRP_LOGIN_REQ because its length (%d bytes) is out of range (%d .. %d)\n",
2186 		       it_iu_len, 64, srp_max_req_size);
2187 		goto reject;
2188 	}
2189 
2190 	if (!sport->enabled) {
2191 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2192 		pr_info("rejected SRP_LOGIN_REQ because target port %s_%d has not yet been enabled\n",
2193 			dev_name(&sport->sdev->device->dev), port_num);
2194 		goto reject;
2195 	}
2196 
2197 	if (*(__be64 *)req->target_port_id != cpu_to_be64(srpt_service_guid)
2198 	    || *(__be64 *)(req->target_port_id + 8) !=
2199 	       cpu_to_be64(srpt_service_guid)) {
2200 		rej->reason = cpu_to_be32(
2201 				SRP_LOGIN_REJ_UNABLE_ASSOCIATE_CHANNEL);
2202 		pr_err("rejected SRP_LOGIN_REQ because it has an invalid target port identifier.\n");
2203 		goto reject;
2204 	}
2205 
2206 	ret = -ENOMEM;
2207 	ch = kzalloc(sizeof(*ch), GFP_KERNEL);
2208 	if (!ch) {
2209 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2210 		pr_err("rejected SRP_LOGIN_REQ because out of memory.\n");
2211 		goto reject;
2212 	}
2213 
2214 	kref_init(&ch->kref);
2215 	ch->pkey = be16_to_cpu(pkey);
2216 	ch->nexus = nexus;
2217 	ch->zw_cqe.done = srpt_zerolength_write_done;
2218 	INIT_WORK(&ch->release_work, srpt_release_channel_work);
2219 	ch->sport = sport;
2220 	if (ib_cm_id) {
2221 		ch->ib_cm.cm_id = ib_cm_id;
2222 		ib_cm_id->context = ch;
2223 	} else {
2224 		ch->using_rdma_cm = true;
2225 		ch->rdma_cm.cm_id = rdma_cm_id;
2226 		rdma_cm_id->context = ch;
2227 	}
2228 	/*
2229 	 * ch->rq_size should be at least as large as the initiator queue
2230 	 * depth to avoid that the initiator driver has to report QUEUE_FULL
2231 	 * to the SCSI mid-layer.
2232 	 */
2233 	ch->rq_size = min(MAX_SRPT_RQ_SIZE, sdev->device->attrs.max_qp_wr);
2234 	spin_lock_init(&ch->spinlock);
2235 	ch->state = CH_CONNECTING;
2236 	INIT_LIST_HEAD(&ch->cmd_wait_list);
2237 	ch->max_rsp_size = ch->sport->port_attrib.srp_max_rsp_size;
2238 
2239 	ch->rsp_buf_cache = kmem_cache_create("srpt-rsp-buf", ch->max_rsp_size,
2240 					      512, 0, NULL);
2241 	if (!ch->rsp_buf_cache)
2242 		goto free_ch;
2243 
2244 	ch->ioctx_ring = (struct srpt_send_ioctx **)
2245 		srpt_alloc_ioctx_ring(ch->sport->sdev, ch->rq_size,
2246 				      sizeof(*ch->ioctx_ring[0]),
2247 				      ch->rsp_buf_cache, 0, DMA_TO_DEVICE);
2248 	if (!ch->ioctx_ring) {
2249 		pr_err("rejected SRP_LOGIN_REQ because creating a new QP SQ ring failed.\n");
2250 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2251 		goto free_rsp_cache;
2252 	}
2253 
2254 	for (i = 0; i < ch->rq_size; i++)
2255 		ch->ioctx_ring[i]->ch = ch;
2256 	if (!sdev->use_srq) {
2257 		u16 imm_data_offset = req->req_flags & SRP_IMMED_REQUESTED ?
2258 			be16_to_cpu(req->imm_data_offset) : 0;
2259 		u16 alignment_offset;
2260 		u32 req_sz;
2261 
2262 		if (req->req_flags & SRP_IMMED_REQUESTED)
2263 			pr_debug("imm_data_offset = %d\n",
2264 				 be16_to_cpu(req->imm_data_offset));
2265 		if (imm_data_offset >= sizeof(struct srp_cmd)) {
2266 			ch->imm_data_offset = imm_data_offset;
2267 			rsp->rsp_flags |= SRP_LOGIN_RSP_IMMED_SUPP;
2268 		} else {
2269 			ch->imm_data_offset = 0;
2270 		}
2271 		alignment_offset = round_up(imm_data_offset, 512) -
2272 			imm_data_offset;
2273 		req_sz = alignment_offset + imm_data_offset + srp_max_req_size;
2274 		ch->req_buf_cache = kmem_cache_create("srpt-req-buf", req_sz,
2275 						      512, 0, NULL);
2276 		if (!ch->req_buf_cache)
2277 			goto free_rsp_ring;
2278 
2279 		ch->ioctx_recv_ring = (struct srpt_recv_ioctx **)
2280 			srpt_alloc_ioctx_ring(ch->sport->sdev, ch->rq_size,
2281 					      sizeof(*ch->ioctx_recv_ring[0]),
2282 					      ch->req_buf_cache,
2283 					      alignment_offset,
2284 					      DMA_FROM_DEVICE);
2285 		if (!ch->ioctx_recv_ring) {
2286 			pr_err("rejected SRP_LOGIN_REQ because creating a new QP RQ ring failed.\n");
2287 			rej->reason =
2288 			    cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2289 			goto free_recv_cache;
2290 		}
2291 		for (i = 0; i < ch->rq_size; i++)
2292 			INIT_LIST_HEAD(&ch->ioctx_recv_ring[i]->wait_list);
2293 	}
2294 
2295 	ret = srpt_create_ch_ib(ch);
2296 	if (ret) {
2297 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2298 		pr_err("rejected SRP_LOGIN_REQ because creating a new RDMA channel failed.\n");
2299 		goto free_recv_ring;
2300 	}
2301 
2302 	strlcpy(ch->sess_name, src_addr, sizeof(ch->sess_name));
2303 	snprintf(i_port_id, sizeof(i_port_id), "0x%016llx%016llx",
2304 			be64_to_cpu(*(__be64 *)nexus->i_port_id),
2305 			be64_to_cpu(*(__be64 *)(nexus->i_port_id + 8)));
2306 
2307 	pr_debug("registering src addr %s or i_port_id %s\n", ch->sess_name,
2308 		 i_port_id);
2309 
2310 	tag_num = ch->rq_size;
2311 	tag_size = 1; /* ib_srpt does not use se_sess->sess_cmd_map */
2312 
2313 	mutex_lock(&sport->port_guid_id.mutex);
2314 	list_for_each_entry(stpg, &sport->port_guid_id.tpg_list, entry) {
2315 		if (!IS_ERR_OR_NULL(ch->sess))
2316 			break;
2317 		ch->sess = target_setup_session(&stpg->tpg, tag_num,
2318 						tag_size, TARGET_PROT_NORMAL,
2319 						ch->sess_name, ch, NULL);
2320 	}
2321 	mutex_unlock(&sport->port_guid_id.mutex);
2322 
2323 	mutex_lock(&sport->port_gid_id.mutex);
2324 	list_for_each_entry(stpg, &sport->port_gid_id.tpg_list, entry) {
2325 		if (!IS_ERR_OR_NULL(ch->sess))
2326 			break;
2327 		ch->sess = target_setup_session(&stpg->tpg, tag_num,
2328 					tag_size, TARGET_PROT_NORMAL, i_port_id,
2329 					ch, NULL);
2330 		if (!IS_ERR_OR_NULL(ch->sess))
2331 			break;
2332 		/* Retry without leading "0x" */
2333 		ch->sess = target_setup_session(&stpg->tpg, tag_num,
2334 						tag_size, TARGET_PROT_NORMAL,
2335 						i_port_id + 2, ch, NULL);
2336 	}
2337 	mutex_unlock(&sport->port_gid_id.mutex);
2338 
2339 	if (IS_ERR_OR_NULL(ch->sess)) {
2340 		WARN_ON_ONCE(ch->sess == NULL);
2341 		ret = PTR_ERR(ch->sess);
2342 		ch->sess = NULL;
2343 		pr_info("Rejected login for initiator %s: ret = %d.\n",
2344 			ch->sess_name, ret);
2345 		rej->reason = cpu_to_be32(ret == -ENOMEM ?
2346 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES :
2347 				SRP_LOGIN_REJ_CHANNEL_LIMIT_REACHED);
2348 		goto destroy_ib;
2349 	}
2350 
2351 	/*
2352 	 * Once a session has been created destruction of srpt_rdma_ch objects
2353 	 * will decrement sport->refcount. Hence increment sport->refcount now.
2354 	 */
2355 	atomic_inc(&sport->refcount);
2356 
2357 	mutex_lock(&sport->mutex);
2358 
2359 	if ((req->req_flags & SRP_MTCH_ACTION) == SRP_MULTICHAN_SINGLE) {
2360 		struct srpt_rdma_ch *ch2;
2361 
2362 		list_for_each_entry(ch2, &nexus->ch_list, list) {
2363 			if (srpt_disconnect_ch(ch2) < 0)
2364 				continue;
2365 			pr_info("Relogin - closed existing channel %s\n",
2366 				ch2->sess_name);
2367 			rsp->rsp_flags |= SRP_LOGIN_RSP_MULTICHAN_TERMINATED;
2368 		}
2369 	} else {
2370 		rsp->rsp_flags |= SRP_LOGIN_RSP_MULTICHAN_MAINTAINED;
2371 	}
2372 
2373 	list_add_tail_rcu(&ch->list, &nexus->ch_list);
2374 
2375 	if (!sport->enabled) {
2376 		rej->reason = cpu_to_be32(
2377 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2378 		pr_info("rejected SRP_LOGIN_REQ because target %s_%d is not enabled\n",
2379 			dev_name(&sdev->device->dev), port_num);
2380 		mutex_unlock(&sport->mutex);
2381 		goto reject;
2382 	}
2383 
2384 	mutex_unlock(&sport->mutex);
2385 
2386 	ret = ch->using_rdma_cm ? 0 : srpt_ch_qp_rtr(ch, ch->qp);
2387 	if (ret) {
2388 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2389 		pr_err("rejected SRP_LOGIN_REQ because enabling RTR failed (error code = %d)\n",
2390 		       ret);
2391 		goto reject;
2392 	}
2393 
2394 	pr_debug("Establish connection sess=%p name=%s ch=%p\n", ch->sess,
2395 		 ch->sess_name, ch);
2396 
2397 	/* create srp_login_response */
2398 	rsp->opcode = SRP_LOGIN_RSP;
2399 	rsp->tag = req->tag;
2400 	rsp->max_it_iu_len = cpu_to_be32(srp_max_req_size);
2401 	rsp->max_ti_iu_len = req->req_it_iu_len;
2402 	ch->max_ti_iu_len = it_iu_len;
2403 	rsp->buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
2404 				   SRP_BUF_FORMAT_INDIRECT);
2405 	rsp->req_lim_delta = cpu_to_be32(ch->rq_size);
2406 	atomic_set(&ch->req_lim, ch->rq_size);
2407 	atomic_set(&ch->req_lim_delta, 0);
2408 
2409 	/* create cm reply */
2410 	if (ch->using_rdma_cm) {
2411 		rep_param->rdma_cm.private_data = (void *)rsp;
2412 		rep_param->rdma_cm.private_data_len = sizeof(*rsp);
2413 		rep_param->rdma_cm.rnr_retry_count = 7;
2414 		rep_param->rdma_cm.flow_control = 1;
2415 		rep_param->rdma_cm.responder_resources = 4;
2416 		rep_param->rdma_cm.initiator_depth = 4;
2417 	} else {
2418 		rep_param->ib_cm.qp_num = ch->qp->qp_num;
2419 		rep_param->ib_cm.private_data = (void *)rsp;
2420 		rep_param->ib_cm.private_data_len = sizeof(*rsp);
2421 		rep_param->ib_cm.rnr_retry_count = 7;
2422 		rep_param->ib_cm.flow_control = 1;
2423 		rep_param->ib_cm.failover_accepted = 0;
2424 		rep_param->ib_cm.srq = 1;
2425 		rep_param->ib_cm.responder_resources = 4;
2426 		rep_param->ib_cm.initiator_depth = 4;
2427 	}
2428 
2429 	/*
2430 	 * Hold the sport mutex while accepting a connection to avoid that
2431 	 * srpt_disconnect_ch() is invoked concurrently with this code.
2432 	 */
2433 	mutex_lock(&sport->mutex);
2434 	if (sport->enabled && ch->state == CH_CONNECTING) {
2435 		if (ch->using_rdma_cm)
2436 			ret = rdma_accept(rdma_cm_id, &rep_param->rdma_cm);
2437 		else
2438 			ret = ib_send_cm_rep(ib_cm_id, &rep_param->ib_cm);
2439 	} else {
2440 		ret = -EINVAL;
2441 	}
2442 	mutex_unlock(&sport->mutex);
2443 
2444 	switch (ret) {
2445 	case 0:
2446 		break;
2447 	case -EINVAL:
2448 		goto reject;
2449 	default:
2450 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2451 		pr_err("sending SRP_LOGIN_REQ response failed (error code = %d)\n",
2452 		       ret);
2453 		goto reject;
2454 	}
2455 
2456 	goto out;
2457 
2458 destroy_ib:
2459 	srpt_destroy_ch_ib(ch);
2460 
2461 free_recv_ring:
2462 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_recv_ring,
2463 			     ch->sport->sdev, ch->rq_size,
2464 			     ch->req_buf_cache, DMA_FROM_DEVICE);
2465 
2466 free_recv_cache:
2467 	kmem_cache_destroy(ch->req_buf_cache);
2468 
2469 free_rsp_ring:
2470 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
2471 			     ch->sport->sdev, ch->rq_size,
2472 			     ch->rsp_buf_cache, DMA_TO_DEVICE);
2473 
2474 free_rsp_cache:
2475 	kmem_cache_destroy(ch->rsp_buf_cache);
2476 
2477 free_ch:
2478 	if (rdma_cm_id)
2479 		rdma_cm_id->context = NULL;
2480 	else
2481 		ib_cm_id->context = NULL;
2482 	kfree(ch);
2483 	ch = NULL;
2484 
2485 	WARN_ON_ONCE(ret == 0);
2486 
2487 reject:
2488 	pr_info("Rejecting login with reason %#x\n", be32_to_cpu(rej->reason));
2489 	rej->opcode = SRP_LOGIN_REJ;
2490 	rej->tag = req->tag;
2491 	rej->buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
2492 				   SRP_BUF_FORMAT_INDIRECT);
2493 
2494 	if (rdma_cm_id)
2495 		rdma_reject(rdma_cm_id, rej, sizeof(*rej),
2496 			    IB_CM_REJ_CONSUMER_DEFINED);
2497 	else
2498 		ib_send_cm_rej(ib_cm_id, IB_CM_REJ_CONSUMER_DEFINED, NULL, 0,
2499 			       rej, sizeof(*rej));
2500 
2501 	if (ch && ch->sess) {
2502 		srpt_close_ch(ch);
2503 		/*
2504 		 * Tell the caller not to free cm_id since
2505 		 * srpt_release_channel_work() will do that.
2506 		 */
2507 		ret = 0;
2508 	}
2509 
2510 out:
2511 	kfree(rep_param);
2512 	kfree(rsp);
2513 	kfree(rej);
2514 
2515 	return ret;
2516 }
2517 
2518 static int srpt_ib_cm_req_recv(struct ib_cm_id *cm_id,
2519 			       const struct ib_cm_req_event_param *param,
2520 			       void *private_data)
2521 {
2522 	char sguid[40];
2523 
2524 	srpt_format_guid(sguid, sizeof(sguid),
2525 			 &param->primary_path->dgid.global.interface_id);
2526 
2527 	return srpt_cm_req_recv(cm_id->context, cm_id, NULL, param->port,
2528 				param->primary_path->pkey,
2529 				private_data, sguid);
2530 }
2531 
2532 static int srpt_rdma_cm_req_recv(struct rdma_cm_id *cm_id,
2533 				 struct rdma_cm_event *event)
2534 {
2535 	struct srpt_device *sdev;
2536 	struct srp_login_req req;
2537 	const struct srp_login_req_rdma *req_rdma;
2538 	struct sa_path_rec *path_rec = cm_id->route.path_rec;
2539 	char src_addr[40];
2540 
2541 	sdev = ib_get_client_data(cm_id->device, &srpt_client);
2542 	if (!sdev)
2543 		return -ECONNREFUSED;
2544 
2545 	if (event->param.conn.private_data_len < sizeof(*req_rdma))
2546 		return -EINVAL;
2547 
2548 	/* Transform srp_login_req_rdma into srp_login_req. */
2549 	req_rdma = event->param.conn.private_data;
2550 	memset(&req, 0, sizeof(req));
2551 	req.opcode		= req_rdma->opcode;
2552 	req.tag			= req_rdma->tag;
2553 	req.req_it_iu_len	= req_rdma->req_it_iu_len;
2554 	req.req_buf_fmt		= req_rdma->req_buf_fmt;
2555 	req.req_flags		= req_rdma->req_flags;
2556 	memcpy(req.initiator_port_id, req_rdma->initiator_port_id, 16);
2557 	memcpy(req.target_port_id, req_rdma->target_port_id, 16);
2558 	req.imm_data_offset	= req_rdma->imm_data_offset;
2559 
2560 	snprintf(src_addr, sizeof(src_addr), "%pIS",
2561 		 &cm_id->route.addr.src_addr);
2562 
2563 	return srpt_cm_req_recv(sdev, NULL, cm_id, cm_id->port_num,
2564 				path_rec ? path_rec->pkey : 0, &req, src_addr);
2565 }
2566 
2567 static void srpt_cm_rej_recv(struct srpt_rdma_ch *ch,
2568 			     enum ib_cm_rej_reason reason,
2569 			     const u8 *private_data,
2570 			     u8 private_data_len)
2571 {
2572 	char *priv = NULL;
2573 	int i;
2574 
2575 	if (private_data_len && (priv = kmalloc(private_data_len * 3 + 1,
2576 						GFP_KERNEL))) {
2577 		for (i = 0; i < private_data_len; i++)
2578 			sprintf(priv + 3 * i, " %02x", private_data[i]);
2579 	}
2580 	pr_info("Received CM REJ for ch %s-%d; reason %d%s%s.\n",
2581 		ch->sess_name, ch->qp->qp_num, reason, private_data_len ?
2582 		"; private data" : "", priv ? priv : " (?)");
2583 	kfree(priv);
2584 }
2585 
2586 /**
2587  * srpt_cm_rtu_recv - process an IB_CM_RTU_RECEIVED or USER_ESTABLISHED event
2588  * @ch: SRPT RDMA channel.
2589  *
2590  * An RTU (ready to use) message indicates that the connection has been
2591  * established and that the recipient may begin transmitting.
2592  */
2593 static void srpt_cm_rtu_recv(struct srpt_rdma_ch *ch)
2594 {
2595 	int ret;
2596 
2597 	ret = ch->using_rdma_cm ? 0 : srpt_ch_qp_rts(ch, ch->qp);
2598 	if (ret < 0) {
2599 		pr_err("%s-%d: QP transition to RTS failed\n", ch->sess_name,
2600 		       ch->qp->qp_num);
2601 		srpt_close_ch(ch);
2602 		return;
2603 	}
2604 
2605 	/*
2606 	 * Note: calling srpt_close_ch() if the transition to the LIVE state
2607 	 * fails is not necessary since that means that that function has
2608 	 * already been invoked from another thread.
2609 	 */
2610 	if (!srpt_set_ch_state(ch, CH_LIVE)) {
2611 		pr_err("%s-%d: channel transition to LIVE state failed\n",
2612 		       ch->sess_name, ch->qp->qp_num);
2613 		return;
2614 	}
2615 
2616 	/* Trigger wait list processing. */
2617 	ret = srpt_zerolength_write(ch);
2618 	WARN_ONCE(ret < 0, "%d\n", ret);
2619 }
2620 
2621 /**
2622  * srpt_cm_handler - IB connection manager callback function
2623  * @cm_id: IB/CM connection identifier.
2624  * @event: IB/CM event.
2625  *
2626  * A non-zero return value will cause the caller destroy the CM ID.
2627  *
2628  * Note: srpt_cm_handler() must only return a non-zero value when transferring
2629  * ownership of the cm_id to a channel by srpt_cm_req_recv() failed. Returning
2630  * a non-zero value in any other case will trigger a race with the
2631  * ib_destroy_cm_id() call in srpt_release_channel().
2632  */
2633 static int srpt_cm_handler(struct ib_cm_id *cm_id,
2634 			   const struct ib_cm_event *event)
2635 {
2636 	struct srpt_rdma_ch *ch = cm_id->context;
2637 	int ret;
2638 
2639 	ret = 0;
2640 	switch (event->event) {
2641 	case IB_CM_REQ_RECEIVED:
2642 		ret = srpt_ib_cm_req_recv(cm_id, &event->param.req_rcvd,
2643 					  event->private_data);
2644 		break;
2645 	case IB_CM_REJ_RECEIVED:
2646 		srpt_cm_rej_recv(ch, event->param.rej_rcvd.reason,
2647 				 event->private_data,
2648 				 IB_CM_REJ_PRIVATE_DATA_SIZE);
2649 		break;
2650 	case IB_CM_RTU_RECEIVED:
2651 	case IB_CM_USER_ESTABLISHED:
2652 		srpt_cm_rtu_recv(ch);
2653 		break;
2654 	case IB_CM_DREQ_RECEIVED:
2655 		srpt_disconnect_ch(ch);
2656 		break;
2657 	case IB_CM_DREP_RECEIVED:
2658 		pr_info("Received CM DREP message for ch %s-%d.\n",
2659 			ch->sess_name, ch->qp->qp_num);
2660 		srpt_close_ch(ch);
2661 		break;
2662 	case IB_CM_TIMEWAIT_EXIT:
2663 		pr_info("Received CM TimeWait exit for ch %s-%d.\n",
2664 			ch->sess_name, ch->qp->qp_num);
2665 		srpt_close_ch(ch);
2666 		break;
2667 	case IB_CM_REP_ERROR:
2668 		pr_info("Received CM REP error for ch %s-%d.\n", ch->sess_name,
2669 			ch->qp->qp_num);
2670 		break;
2671 	case IB_CM_DREQ_ERROR:
2672 		pr_info("Received CM DREQ ERROR event.\n");
2673 		break;
2674 	case IB_CM_MRA_RECEIVED:
2675 		pr_info("Received CM MRA event\n");
2676 		break;
2677 	default:
2678 		pr_err("received unrecognized CM event %d\n", event->event);
2679 		break;
2680 	}
2681 
2682 	return ret;
2683 }
2684 
2685 static int srpt_rdma_cm_handler(struct rdma_cm_id *cm_id,
2686 				struct rdma_cm_event *event)
2687 {
2688 	struct srpt_rdma_ch *ch = cm_id->context;
2689 	int ret = 0;
2690 
2691 	switch (event->event) {
2692 	case RDMA_CM_EVENT_CONNECT_REQUEST:
2693 		ret = srpt_rdma_cm_req_recv(cm_id, event);
2694 		break;
2695 	case RDMA_CM_EVENT_REJECTED:
2696 		srpt_cm_rej_recv(ch, event->status,
2697 				 event->param.conn.private_data,
2698 				 event->param.conn.private_data_len);
2699 		break;
2700 	case RDMA_CM_EVENT_ESTABLISHED:
2701 		srpt_cm_rtu_recv(ch);
2702 		break;
2703 	case RDMA_CM_EVENT_DISCONNECTED:
2704 		if (ch->state < CH_DISCONNECTING)
2705 			srpt_disconnect_ch(ch);
2706 		else
2707 			srpt_close_ch(ch);
2708 		break;
2709 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
2710 		srpt_close_ch(ch);
2711 		break;
2712 	case RDMA_CM_EVENT_UNREACHABLE:
2713 		pr_info("Received CM REP error for ch %s-%d.\n", ch->sess_name,
2714 			ch->qp->qp_num);
2715 		break;
2716 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
2717 	case RDMA_CM_EVENT_ADDR_CHANGE:
2718 		break;
2719 	default:
2720 		pr_err("received unrecognized RDMA CM event %d\n",
2721 		       event->event);
2722 		break;
2723 	}
2724 
2725 	return ret;
2726 }
2727 
2728 /*
2729  * srpt_write_pending - Start data transfer from initiator to target (write).
2730  */
2731 static int srpt_write_pending(struct se_cmd *se_cmd)
2732 {
2733 	struct srpt_send_ioctx *ioctx =
2734 		container_of(se_cmd, struct srpt_send_ioctx, cmd);
2735 	struct srpt_rdma_ch *ch = ioctx->ch;
2736 	struct ib_send_wr *first_wr = NULL;
2737 	struct ib_cqe *cqe = &ioctx->rdma_cqe;
2738 	enum srpt_command_state new_state;
2739 	int ret, i;
2740 
2741 	if (ioctx->recv_ioctx) {
2742 		srpt_set_cmd_state(ioctx, SRPT_STATE_DATA_IN);
2743 		target_execute_cmd(&ioctx->cmd);
2744 		return 0;
2745 	}
2746 
2747 	new_state = srpt_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA);
2748 	WARN_ON(new_state == SRPT_STATE_DONE);
2749 
2750 	if (atomic_sub_return(ioctx->n_rdma, &ch->sq_wr_avail) < 0) {
2751 		pr_warn("%s: IB send queue full (needed %d)\n",
2752 				__func__, ioctx->n_rdma);
2753 		ret = -ENOMEM;
2754 		goto out_undo;
2755 	}
2756 
2757 	cqe->done = srpt_rdma_read_done;
2758 	for (i = ioctx->n_rw_ctx - 1; i >= 0; i--) {
2759 		struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i];
2760 
2761 		first_wr = rdma_rw_ctx_wrs(&ctx->rw, ch->qp, ch->sport->port,
2762 				cqe, first_wr);
2763 		cqe = NULL;
2764 	}
2765 
2766 	ret = ib_post_send(ch->qp, first_wr, NULL);
2767 	if (ret) {
2768 		pr_err("%s: ib_post_send() returned %d for %d (avail: %d)\n",
2769 			 __func__, ret, ioctx->n_rdma,
2770 			 atomic_read(&ch->sq_wr_avail));
2771 		goto out_undo;
2772 	}
2773 
2774 	return 0;
2775 out_undo:
2776 	atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
2777 	return ret;
2778 }
2779 
2780 static u8 tcm_to_srp_tsk_mgmt_status(const int tcm_mgmt_status)
2781 {
2782 	switch (tcm_mgmt_status) {
2783 	case TMR_FUNCTION_COMPLETE:
2784 		return SRP_TSK_MGMT_SUCCESS;
2785 	case TMR_FUNCTION_REJECTED:
2786 		return SRP_TSK_MGMT_FUNC_NOT_SUPP;
2787 	}
2788 	return SRP_TSK_MGMT_FAILED;
2789 }
2790 
2791 /**
2792  * srpt_queue_response - transmit the response to a SCSI command
2793  * @cmd: SCSI target command.
2794  *
2795  * Callback function called by the TCM core. Must not block since it can be
2796  * invoked on the context of the IB completion handler.
2797  */
2798 static void srpt_queue_response(struct se_cmd *cmd)
2799 {
2800 	struct srpt_send_ioctx *ioctx =
2801 		container_of(cmd, struct srpt_send_ioctx, cmd);
2802 	struct srpt_rdma_ch *ch = ioctx->ch;
2803 	struct srpt_device *sdev = ch->sport->sdev;
2804 	struct ib_send_wr send_wr, *first_wr = &send_wr;
2805 	struct ib_sge sge;
2806 	enum srpt_command_state state;
2807 	int resp_len, ret, i;
2808 	u8 srp_tm_status;
2809 
2810 	state = ioctx->state;
2811 	switch (state) {
2812 	case SRPT_STATE_NEW:
2813 	case SRPT_STATE_DATA_IN:
2814 		ioctx->state = SRPT_STATE_CMD_RSP_SENT;
2815 		break;
2816 	case SRPT_STATE_MGMT:
2817 		ioctx->state = SRPT_STATE_MGMT_RSP_SENT;
2818 		break;
2819 	default:
2820 		WARN(true, "ch %p; cmd %d: unexpected command state %d\n",
2821 			ch, ioctx->ioctx.index, ioctx->state);
2822 		break;
2823 	}
2824 
2825 	if (WARN_ON_ONCE(state == SRPT_STATE_CMD_RSP_SENT))
2826 		return;
2827 
2828 	/* For read commands, transfer the data to the initiator. */
2829 	if (ioctx->cmd.data_direction == DMA_FROM_DEVICE &&
2830 	    ioctx->cmd.data_length &&
2831 	    !ioctx->queue_status_only) {
2832 		for (i = ioctx->n_rw_ctx - 1; i >= 0; i--) {
2833 			struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i];
2834 
2835 			first_wr = rdma_rw_ctx_wrs(&ctx->rw, ch->qp,
2836 					ch->sport->port, NULL, first_wr);
2837 		}
2838 	}
2839 
2840 	if (state != SRPT_STATE_MGMT)
2841 		resp_len = srpt_build_cmd_rsp(ch, ioctx, ioctx->cmd.tag,
2842 					      cmd->scsi_status);
2843 	else {
2844 		srp_tm_status
2845 			= tcm_to_srp_tsk_mgmt_status(cmd->se_tmr_req->response);
2846 		resp_len = srpt_build_tskmgmt_rsp(ch, ioctx, srp_tm_status,
2847 						 ioctx->cmd.tag);
2848 	}
2849 
2850 	atomic_inc(&ch->req_lim);
2851 
2852 	if (unlikely(atomic_sub_return(1 + ioctx->n_rdma,
2853 			&ch->sq_wr_avail) < 0)) {
2854 		pr_warn("%s: IB send queue full (needed %d)\n",
2855 				__func__, ioctx->n_rdma);
2856 		ret = -ENOMEM;
2857 		goto out;
2858 	}
2859 
2860 	ib_dma_sync_single_for_device(sdev->device, ioctx->ioctx.dma, resp_len,
2861 				      DMA_TO_DEVICE);
2862 
2863 	sge.addr = ioctx->ioctx.dma;
2864 	sge.length = resp_len;
2865 	sge.lkey = sdev->lkey;
2866 
2867 	ioctx->ioctx.cqe.done = srpt_send_done;
2868 	send_wr.next = NULL;
2869 	send_wr.wr_cqe = &ioctx->ioctx.cqe;
2870 	send_wr.sg_list = &sge;
2871 	send_wr.num_sge = 1;
2872 	send_wr.opcode = IB_WR_SEND;
2873 	send_wr.send_flags = IB_SEND_SIGNALED;
2874 
2875 	ret = ib_post_send(ch->qp, first_wr, NULL);
2876 	if (ret < 0) {
2877 		pr_err("%s: sending cmd response failed for tag %llu (%d)\n",
2878 			__func__, ioctx->cmd.tag, ret);
2879 		goto out;
2880 	}
2881 
2882 	return;
2883 
2884 out:
2885 	atomic_add(1 + ioctx->n_rdma, &ch->sq_wr_avail);
2886 	atomic_dec(&ch->req_lim);
2887 	srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
2888 	target_put_sess_cmd(&ioctx->cmd);
2889 }
2890 
2891 static int srpt_queue_data_in(struct se_cmd *cmd)
2892 {
2893 	srpt_queue_response(cmd);
2894 	return 0;
2895 }
2896 
2897 static void srpt_queue_tm_rsp(struct se_cmd *cmd)
2898 {
2899 	srpt_queue_response(cmd);
2900 }
2901 
2902 /*
2903  * This function is called for aborted commands if no response is sent to the
2904  * initiator. Make sure that the credits freed by aborting a command are
2905  * returned to the initiator the next time a response is sent by incrementing
2906  * ch->req_lim_delta.
2907  */
2908 static void srpt_aborted_task(struct se_cmd *cmd)
2909 {
2910 	struct srpt_send_ioctx *ioctx = container_of(cmd,
2911 				struct srpt_send_ioctx, cmd);
2912 	struct srpt_rdma_ch *ch = ioctx->ch;
2913 
2914 	atomic_inc(&ch->req_lim_delta);
2915 }
2916 
2917 static int srpt_queue_status(struct se_cmd *cmd)
2918 {
2919 	struct srpt_send_ioctx *ioctx;
2920 
2921 	ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
2922 	BUG_ON(ioctx->sense_data != cmd->sense_buffer);
2923 	if (cmd->se_cmd_flags &
2924 	    (SCF_TRANSPORT_TASK_SENSE | SCF_EMULATED_TASK_SENSE))
2925 		WARN_ON(cmd->scsi_status != SAM_STAT_CHECK_CONDITION);
2926 	ioctx->queue_status_only = true;
2927 	srpt_queue_response(cmd);
2928 	return 0;
2929 }
2930 
2931 static void srpt_refresh_port_work(struct work_struct *work)
2932 {
2933 	struct srpt_port *sport = container_of(work, struct srpt_port, work);
2934 
2935 	srpt_refresh_port(sport);
2936 }
2937 
2938 /**
2939  * srpt_release_sport - disable login and wait for associated channels
2940  * @sport: SRPT HCA port.
2941  */
2942 static int srpt_release_sport(struct srpt_port *sport)
2943 {
2944 	DECLARE_COMPLETION_ONSTACK(c);
2945 	struct srpt_nexus *nexus, *next_n;
2946 	struct srpt_rdma_ch *ch;
2947 
2948 	WARN_ON_ONCE(irqs_disabled());
2949 
2950 	sport->freed_channels = &c;
2951 
2952 	mutex_lock(&sport->mutex);
2953 	srpt_set_enabled(sport, false);
2954 	mutex_unlock(&sport->mutex);
2955 
2956 	while (atomic_read(&sport->refcount) > 0 &&
2957 	       wait_for_completion_timeout(&c, 5 * HZ) <= 0) {
2958 		pr_info("%s_%d: waiting for unregistration of %d sessions ...\n",
2959 			dev_name(&sport->sdev->device->dev), sport->port,
2960 			atomic_read(&sport->refcount));
2961 		rcu_read_lock();
2962 		list_for_each_entry(nexus, &sport->nexus_list, entry) {
2963 			list_for_each_entry(ch, &nexus->ch_list, list) {
2964 				pr_info("%s-%d: state %s\n",
2965 					ch->sess_name, ch->qp->qp_num,
2966 					get_ch_state_name(ch->state));
2967 			}
2968 		}
2969 		rcu_read_unlock();
2970 	}
2971 
2972 	mutex_lock(&sport->mutex);
2973 	list_for_each_entry_safe(nexus, next_n, &sport->nexus_list, entry) {
2974 		list_del(&nexus->entry);
2975 		kfree_rcu(nexus, rcu);
2976 	}
2977 	mutex_unlock(&sport->mutex);
2978 
2979 	return 0;
2980 }
2981 
2982 static struct se_wwn *__srpt_lookup_wwn(const char *name)
2983 {
2984 	struct ib_device *dev;
2985 	struct srpt_device *sdev;
2986 	struct srpt_port *sport;
2987 	int i;
2988 
2989 	list_for_each_entry(sdev, &srpt_dev_list, list) {
2990 		dev = sdev->device;
2991 		if (!dev)
2992 			continue;
2993 
2994 		for (i = 0; i < dev->phys_port_cnt; i++) {
2995 			sport = &sdev->port[i];
2996 
2997 			if (strcmp(sport->port_guid_id.name, name) == 0)
2998 				return &sport->port_guid_id.wwn;
2999 			if (strcmp(sport->port_gid_id.name, name) == 0)
3000 				return &sport->port_gid_id.wwn;
3001 		}
3002 	}
3003 
3004 	return NULL;
3005 }
3006 
3007 static struct se_wwn *srpt_lookup_wwn(const char *name)
3008 {
3009 	struct se_wwn *wwn;
3010 
3011 	spin_lock(&srpt_dev_lock);
3012 	wwn = __srpt_lookup_wwn(name);
3013 	spin_unlock(&srpt_dev_lock);
3014 
3015 	return wwn;
3016 }
3017 
3018 static void srpt_free_srq(struct srpt_device *sdev)
3019 {
3020 	if (!sdev->srq)
3021 		return;
3022 
3023 	ib_destroy_srq(sdev->srq);
3024 	srpt_free_ioctx_ring((struct srpt_ioctx **)sdev->ioctx_ring, sdev,
3025 			     sdev->srq_size, sdev->req_buf_cache,
3026 			     DMA_FROM_DEVICE);
3027 	kmem_cache_destroy(sdev->req_buf_cache);
3028 	sdev->srq = NULL;
3029 }
3030 
3031 static int srpt_alloc_srq(struct srpt_device *sdev)
3032 {
3033 	struct ib_srq_init_attr srq_attr = {
3034 		.event_handler = srpt_srq_event,
3035 		.srq_context = (void *)sdev,
3036 		.attr.max_wr = sdev->srq_size,
3037 		.attr.max_sge = 1,
3038 		.srq_type = IB_SRQT_BASIC,
3039 	};
3040 	struct ib_device *device = sdev->device;
3041 	struct ib_srq *srq;
3042 	int i;
3043 
3044 	WARN_ON_ONCE(sdev->srq);
3045 	srq = ib_create_srq(sdev->pd, &srq_attr);
3046 	if (IS_ERR(srq)) {
3047 		pr_debug("ib_create_srq() failed: %ld\n", PTR_ERR(srq));
3048 		return PTR_ERR(srq);
3049 	}
3050 
3051 	pr_debug("create SRQ #wr= %d max_allow=%d dev= %s\n", sdev->srq_size,
3052 		 sdev->device->attrs.max_srq_wr, dev_name(&device->dev));
3053 
3054 	sdev->req_buf_cache = kmem_cache_create("srpt-srq-req-buf",
3055 						srp_max_req_size, 0, 0, NULL);
3056 	if (!sdev->req_buf_cache)
3057 		goto free_srq;
3058 
3059 	sdev->ioctx_ring = (struct srpt_recv_ioctx **)
3060 		srpt_alloc_ioctx_ring(sdev, sdev->srq_size,
3061 				      sizeof(*sdev->ioctx_ring[0]),
3062 				      sdev->req_buf_cache, 0, DMA_FROM_DEVICE);
3063 	if (!sdev->ioctx_ring)
3064 		goto free_cache;
3065 
3066 	sdev->use_srq = true;
3067 	sdev->srq = srq;
3068 
3069 	for (i = 0; i < sdev->srq_size; ++i) {
3070 		INIT_LIST_HEAD(&sdev->ioctx_ring[i]->wait_list);
3071 		srpt_post_recv(sdev, NULL, sdev->ioctx_ring[i]);
3072 	}
3073 
3074 	return 0;
3075 
3076 free_cache:
3077 	kmem_cache_destroy(sdev->req_buf_cache);
3078 
3079 free_srq:
3080 	ib_destroy_srq(srq);
3081 	return -ENOMEM;
3082 }
3083 
3084 static int srpt_use_srq(struct srpt_device *sdev, bool use_srq)
3085 {
3086 	struct ib_device *device = sdev->device;
3087 	int ret = 0;
3088 
3089 	if (!use_srq) {
3090 		srpt_free_srq(sdev);
3091 		sdev->use_srq = false;
3092 	} else if (use_srq && !sdev->srq) {
3093 		ret = srpt_alloc_srq(sdev);
3094 	}
3095 	pr_debug("%s(%s): use_srq = %d; ret = %d\n", __func__,
3096 		 dev_name(&device->dev), sdev->use_srq, ret);
3097 	return ret;
3098 }
3099 
3100 /**
3101  * srpt_add_one - InfiniBand device addition callback function
3102  * @device: Describes a HCA.
3103  */
3104 static int srpt_add_one(struct ib_device *device)
3105 {
3106 	struct srpt_device *sdev;
3107 	struct srpt_port *sport;
3108 	int i, ret;
3109 
3110 	pr_debug("device = %p\n", device);
3111 
3112 	sdev = kzalloc(struct_size(sdev, port, device->phys_port_cnt),
3113 		       GFP_KERNEL);
3114 	if (!sdev)
3115 		return -ENOMEM;
3116 
3117 	sdev->device = device;
3118 	mutex_init(&sdev->sdev_mutex);
3119 
3120 	sdev->pd = ib_alloc_pd(device, 0);
3121 	if (IS_ERR(sdev->pd)) {
3122 		ret = PTR_ERR(sdev->pd);
3123 		goto free_dev;
3124 	}
3125 
3126 	sdev->lkey = sdev->pd->local_dma_lkey;
3127 
3128 	sdev->srq_size = min(srpt_srq_size, sdev->device->attrs.max_srq_wr);
3129 
3130 	srpt_use_srq(sdev, sdev->port[0].port_attrib.use_srq);
3131 
3132 	if (!srpt_service_guid)
3133 		srpt_service_guid = be64_to_cpu(device->node_guid);
3134 
3135 	if (rdma_port_get_link_layer(device, 1) == IB_LINK_LAYER_INFINIBAND)
3136 		sdev->cm_id = ib_create_cm_id(device, srpt_cm_handler, sdev);
3137 	if (IS_ERR(sdev->cm_id)) {
3138 		pr_info("ib_create_cm_id() failed: %ld\n",
3139 			PTR_ERR(sdev->cm_id));
3140 		ret = PTR_ERR(sdev->cm_id);
3141 		sdev->cm_id = NULL;
3142 		if (!rdma_cm_id)
3143 			goto err_ring;
3144 	}
3145 
3146 	/* print out target login information */
3147 	pr_debug("Target login info: id_ext=%016llx,ioc_guid=%016llx,pkey=ffff,service_id=%016llx\n",
3148 		 srpt_service_guid, srpt_service_guid, srpt_service_guid);
3149 
3150 	/*
3151 	 * We do not have a consistent service_id (ie. also id_ext of target_id)
3152 	 * to identify this target. We currently use the guid of the first HCA
3153 	 * in the system as service_id; therefore, the target_id will change
3154 	 * if this HCA is gone bad and replaced by different HCA
3155 	 */
3156 	ret = sdev->cm_id ?
3157 		ib_cm_listen(sdev->cm_id, cpu_to_be64(srpt_service_guid), 0) :
3158 		0;
3159 	if (ret < 0) {
3160 		pr_err("ib_cm_listen() failed: %d (cm_id state = %d)\n", ret,
3161 		       sdev->cm_id->state);
3162 		goto err_cm;
3163 	}
3164 
3165 	INIT_IB_EVENT_HANDLER(&sdev->event_handler, sdev->device,
3166 			      srpt_event_handler);
3167 	ib_register_event_handler(&sdev->event_handler);
3168 
3169 	for (i = 1; i <= sdev->device->phys_port_cnt; i++) {
3170 		sport = &sdev->port[i - 1];
3171 		INIT_LIST_HEAD(&sport->nexus_list);
3172 		mutex_init(&sport->mutex);
3173 		sport->sdev = sdev;
3174 		sport->port = i;
3175 		sport->port_attrib.srp_max_rdma_size = DEFAULT_MAX_RDMA_SIZE;
3176 		sport->port_attrib.srp_max_rsp_size = DEFAULT_MAX_RSP_SIZE;
3177 		sport->port_attrib.srp_sq_size = DEF_SRPT_SQ_SIZE;
3178 		sport->port_attrib.use_srq = false;
3179 		INIT_WORK(&sport->work, srpt_refresh_port_work);
3180 		mutex_init(&sport->port_guid_id.mutex);
3181 		INIT_LIST_HEAD(&sport->port_guid_id.tpg_list);
3182 		mutex_init(&sport->port_gid_id.mutex);
3183 		INIT_LIST_HEAD(&sport->port_gid_id.tpg_list);
3184 
3185 		ret = srpt_refresh_port(sport);
3186 		if (ret) {
3187 			pr_err("MAD registration failed for %s-%d.\n",
3188 			       dev_name(&sdev->device->dev), i);
3189 			i--;
3190 			goto err_port;
3191 		}
3192 	}
3193 
3194 	spin_lock(&srpt_dev_lock);
3195 	list_add_tail(&sdev->list, &srpt_dev_list);
3196 	spin_unlock(&srpt_dev_lock);
3197 
3198 	ib_set_client_data(device, &srpt_client, sdev);
3199 	pr_debug("added %s.\n", dev_name(&device->dev));
3200 	return 0;
3201 
3202 err_port:
3203 	srpt_unregister_mad_agent(sdev, i);
3204 	ib_unregister_event_handler(&sdev->event_handler);
3205 err_cm:
3206 	if (sdev->cm_id)
3207 		ib_destroy_cm_id(sdev->cm_id);
3208 err_ring:
3209 	srpt_free_srq(sdev);
3210 	ib_dealloc_pd(sdev->pd);
3211 free_dev:
3212 	kfree(sdev);
3213 	pr_info("%s(%s) failed.\n", __func__, dev_name(&device->dev));
3214 	return ret;
3215 }
3216 
3217 /**
3218  * srpt_remove_one - InfiniBand device removal callback function
3219  * @device: Describes a HCA.
3220  * @client_data: The value passed as the third argument to ib_set_client_data().
3221  */
3222 static void srpt_remove_one(struct ib_device *device, void *client_data)
3223 {
3224 	struct srpt_device *sdev = client_data;
3225 	int i;
3226 
3227 	srpt_unregister_mad_agent(sdev, sdev->device->phys_port_cnt);
3228 
3229 	ib_unregister_event_handler(&sdev->event_handler);
3230 
3231 	/* Cancel any work queued by the just unregistered IB event handler. */
3232 	for (i = 0; i < sdev->device->phys_port_cnt; i++)
3233 		cancel_work_sync(&sdev->port[i].work);
3234 
3235 	if (sdev->cm_id)
3236 		ib_destroy_cm_id(sdev->cm_id);
3237 
3238 	ib_set_client_data(device, &srpt_client, NULL);
3239 
3240 	/*
3241 	 * Unregistering a target must happen after destroying sdev->cm_id
3242 	 * such that no new SRP_LOGIN_REQ information units can arrive while
3243 	 * destroying the target.
3244 	 */
3245 	spin_lock(&srpt_dev_lock);
3246 	list_del(&sdev->list);
3247 	spin_unlock(&srpt_dev_lock);
3248 
3249 	for (i = 0; i < sdev->device->phys_port_cnt; i++)
3250 		srpt_release_sport(&sdev->port[i]);
3251 
3252 	srpt_free_srq(sdev);
3253 
3254 	ib_dealloc_pd(sdev->pd);
3255 
3256 	kfree(sdev);
3257 }
3258 
3259 static struct ib_client srpt_client = {
3260 	.name = DRV_NAME,
3261 	.add = srpt_add_one,
3262 	.remove = srpt_remove_one
3263 };
3264 
3265 static int srpt_check_true(struct se_portal_group *se_tpg)
3266 {
3267 	return 1;
3268 }
3269 
3270 static int srpt_check_false(struct se_portal_group *se_tpg)
3271 {
3272 	return 0;
3273 }
3274 
3275 static struct srpt_port *srpt_tpg_to_sport(struct se_portal_group *tpg)
3276 {
3277 	return tpg->se_tpg_wwn->priv;
3278 }
3279 
3280 static struct srpt_port_id *srpt_wwn_to_sport_id(struct se_wwn *wwn)
3281 {
3282 	struct srpt_port *sport = wwn->priv;
3283 
3284 	if (wwn == &sport->port_guid_id.wwn)
3285 		return &sport->port_guid_id;
3286 	if (wwn == &sport->port_gid_id.wwn)
3287 		return &sport->port_gid_id;
3288 	WARN_ON_ONCE(true);
3289 	return NULL;
3290 }
3291 
3292 static char *srpt_get_fabric_wwn(struct se_portal_group *tpg)
3293 {
3294 	struct srpt_tpg *stpg = container_of(tpg, typeof(*stpg), tpg);
3295 
3296 	return stpg->sport_id->name;
3297 }
3298 
3299 static u16 srpt_get_tag(struct se_portal_group *tpg)
3300 {
3301 	return 1;
3302 }
3303 
3304 static u32 srpt_tpg_get_inst_index(struct se_portal_group *se_tpg)
3305 {
3306 	return 1;
3307 }
3308 
3309 static void srpt_release_cmd(struct se_cmd *se_cmd)
3310 {
3311 	struct srpt_send_ioctx *ioctx = container_of(se_cmd,
3312 				struct srpt_send_ioctx, cmd);
3313 	struct srpt_rdma_ch *ch = ioctx->ch;
3314 	struct srpt_recv_ioctx *recv_ioctx = ioctx->recv_ioctx;
3315 
3316 	WARN_ON_ONCE(ioctx->state != SRPT_STATE_DONE &&
3317 		     !(ioctx->cmd.transport_state & CMD_T_ABORTED));
3318 
3319 	if (recv_ioctx) {
3320 		WARN_ON_ONCE(!list_empty(&recv_ioctx->wait_list));
3321 		ioctx->recv_ioctx = NULL;
3322 		srpt_post_recv(ch->sport->sdev, ch, recv_ioctx);
3323 	}
3324 
3325 	if (ioctx->n_rw_ctx) {
3326 		srpt_free_rw_ctxs(ch, ioctx);
3327 		ioctx->n_rw_ctx = 0;
3328 	}
3329 
3330 	target_free_tag(se_cmd->se_sess, se_cmd);
3331 }
3332 
3333 /**
3334  * srpt_close_session - forcibly close a session
3335  * @se_sess: SCSI target session.
3336  *
3337  * Callback function invoked by the TCM core to clean up sessions associated
3338  * with a node ACL when the user invokes
3339  * rmdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3340  */
3341 static void srpt_close_session(struct se_session *se_sess)
3342 {
3343 	struct srpt_rdma_ch *ch = se_sess->fabric_sess_ptr;
3344 
3345 	srpt_disconnect_ch_sync(ch);
3346 }
3347 
3348 /**
3349  * srpt_sess_get_index - return the value of scsiAttIntrPortIndex (SCSI-MIB)
3350  * @se_sess: SCSI target session.
3351  *
3352  * A quote from RFC 4455 (SCSI-MIB) about this MIB object:
3353  * This object represents an arbitrary integer used to uniquely identify a
3354  * particular attached remote initiator port to a particular SCSI target port
3355  * within a particular SCSI target device within a particular SCSI instance.
3356  */
3357 static u32 srpt_sess_get_index(struct se_session *se_sess)
3358 {
3359 	return 0;
3360 }
3361 
3362 static void srpt_set_default_node_attrs(struct se_node_acl *nacl)
3363 {
3364 }
3365 
3366 /* Note: only used from inside debug printk's by the TCM core. */
3367 static int srpt_get_tcm_cmd_state(struct se_cmd *se_cmd)
3368 {
3369 	struct srpt_send_ioctx *ioctx;
3370 
3371 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
3372 	return ioctx->state;
3373 }
3374 
3375 static int srpt_parse_guid(u64 *guid, const char *name)
3376 {
3377 	u16 w[4];
3378 	int ret = -EINVAL;
3379 
3380 	if (sscanf(name, "%hx:%hx:%hx:%hx", &w[0], &w[1], &w[2], &w[3]) != 4)
3381 		goto out;
3382 	*guid = get_unaligned_be64(w);
3383 	ret = 0;
3384 out:
3385 	return ret;
3386 }
3387 
3388 /**
3389  * srpt_parse_i_port_id - parse an initiator port ID
3390  * @name: ASCII representation of a 128-bit initiator port ID.
3391  * @i_port_id: Binary 128-bit port ID.
3392  */
3393 static int srpt_parse_i_port_id(u8 i_port_id[16], const char *name)
3394 {
3395 	const char *p;
3396 	unsigned len, count, leading_zero_bytes;
3397 	int ret;
3398 
3399 	p = name;
3400 	if (strncasecmp(p, "0x", 2) == 0)
3401 		p += 2;
3402 	ret = -EINVAL;
3403 	len = strlen(p);
3404 	if (len % 2)
3405 		goto out;
3406 	count = min(len / 2, 16U);
3407 	leading_zero_bytes = 16 - count;
3408 	memset(i_port_id, 0, leading_zero_bytes);
3409 	ret = hex2bin(i_port_id + leading_zero_bytes, p, count);
3410 
3411 out:
3412 	return ret;
3413 }
3414 
3415 /*
3416  * configfs callback function invoked for mkdir
3417  * /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3418  *
3419  * i_port_id must be an initiator port GUID, GID or IP address. See also the
3420  * target_alloc_session() calls in this driver. Examples of valid initiator
3421  * port IDs:
3422  * 0x0000000000000000505400fffe4a0b7b
3423  * 0000000000000000505400fffe4a0b7b
3424  * 5054:00ff:fe4a:0b7b
3425  * 192.168.122.76
3426  */
3427 static int srpt_init_nodeacl(struct se_node_acl *se_nacl, const char *name)
3428 {
3429 	struct sockaddr_storage sa;
3430 	u64 guid;
3431 	u8 i_port_id[16];
3432 	int ret;
3433 
3434 	ret = srpt_parse_guid(&guid, name);
3435 	if (ret < 0)
3436 		ret = srpt_parse_i_port_id(i_port_id, name);
3437 	if (ret < 0)
3438 		ret = inet_pton_with_scope(&init_net, AF_UNSPEC, name, NULL,
3439 					   &sa);
3440 	if (ret < 0)
3441 		pr_err("invalid initiator port ID %s\n", name);
3442 	return ret;
3443 }
3444 
3445 static ssize_t srpt_tpg_attrib_srp_max_rdma_size_show(struct config_item *item,
3446 		char *page)
3447 {
3448 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3449 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3450 
3451 	return sysfs_emit(page, "%u\n", sport->port_attrib.srp_max_rdma_size);
3452 }
3453 
3454 static ssize_t srpt_tpg_attrib_srp_max_rdma_size_store(struct config_item *item,
3455 		const char *page, size_t count)
3456 {
3457 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3458 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3459 	unsigned long val;
3460 	int ret;
3461 
3462 	ret = kstrtoul(page, 0, &val);
3463 	if (ret < 0) {
3464 		pr_err("kstrtoul() failed with ret: %d\n", ret);
3465 		return -EINVAL;
3466 	}
3467 	if (val > MAX_SRPT_RDMA_SIZE) {
3468 		pr_err("val: %lu exceeds MAX_SRPT_RDMA_SIZE: %d\n", val,
3469 			MAX_SRPT_RDMA_SIZE);
3470 		return -EINVAL;
3471 	}
3472 	if (val < DEFAULT_MAX_RDMA_SIZE) {
3473 		pr_err("val: %lu smaller than DEFAULT_MAX_RDMA_SIZE: %d\n",
3474 			val, DEFAULT_MAX_RDMA_SIZE);
3475 		return -EINVAL;
3476 	}
3477 	sport->port_attrib.srp_max_rdma_size = val;
3478 
3479 	return count;
3480 }
3481 
3482 static ssize_t srpt_tpg_attrib_srp_max_rsp_size_show(struct config_item *item,
3483 		char *page)
3484 {
3485 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3486 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3487 
3488 	return sysfs_emit(page, "%u\n", sport->port_attrib.srp_max_rsp_size);
3489 }
3490 
3491 static ssize_t srpt_tpg_attrib_srp_max_rsp_size_store(struct config_item *item,
3492 		const char *page, size_t count)
3493 {
3494 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3495 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3496 	unsigned long val;
3497 	int ret;
3498 
3499 	ret = kstrtoul(page, 0, &val);
3500 	if (ret < 0) {
3501 		pr_err("kstrtoul() failed with ret: %d\n", ret);
3502 		return -EINVAL;
3503 	}
3504 	if (val > MAX_SRPT_RSP_SIZE) {
3505 		pr_err("val: %lu exceeds MAX_SRPT_RSP_SIZE: %d\n", val,
3506 			MAX_SRPT_RSP_SIZE);
3507 		return -EINVAL;
3508 	}
3509 	if (val < MIN_MAX_RSP_SIZE) {
3510 		pr_err("val: %lu smaller than MIN_MAX_RSP_SIZE: %d\n", val,
3511 			MIN_MAX_RSP_SIZE);
3512 		return -EINVAL;
3513 	}
3514 	sport->port_attrib.srp_max_rsp_size = val;
3515 
3516 	return count;
3517 }
3518 
3519 static ssize_t srpt_tpg_attrib_srp_sq_size_show(struct config_item *item,
3520 		char *page)
3521 {
3522 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3523 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3524 
3525 	return sysfs_emit(page, "%u\n", sport->port_attrib.srp_sq_size);
3526 }
3527 
3528 static ssize_t srpt_tpg_attrib_srp_sq_size_store(struct config_item *item,
3529 		const char *page, size_t count)
3530 {
3531 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3532 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3533 	unsigned long val;
3534 	int ret;
3535 
3536 	ret = kstrtoul(page, 0, &val);
3537 	if (ret < 0) {
3538 		pr_err("kstrtoul() failed with ret: %d\n", ret);
3539 		return -EINVAL;
3540 	}
3541 	if (val > MAX_SRPT_SRQ_SIZE) {
3542 		pr_err("val: %lu exceeds MAX_SRPT_SRQ_SIZE: %d\n", val,
3543 			MAX_SRPT_SRQ_SIZE);
3544 		return -EINVAL;
3545 	}
3546 	if (val < MIN_SRPT_SRQ_SIZE) {
3547 		pr_err("val: %lu smaller than MIN_SRPT_SRQ_SIZE: %d\n", val,
3548 			MIN_SRPT_SRQ_SIZE);
3549 		return -EINVAL;
3550 	}
3551 	sport->port_attrib.srp_sq_size = val;
3552 
3553 	return count;
3554 }
3555 
3556 static ssize_t srpt_tpg_attrib_use_srq_show(struct config_item *item,
3557 					    char *page)
3558 {
3559 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3560 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3561 
3562 	return sysfs_emit(page, "%d\n", sport->port_attrib.use_srq);
3563 }
3564 
3565 static ssize_t srpt_tpg_attrib_use_srq_store(struct config_item *item,
3566 					     const char *page, size_t count)
3567 {
3568 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3569 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3570 	struct srpt_device *sdev = sport->sdev;
3571 	unsigned long val;
3572 	bool enabled;
3573 	int ret;
3574 
3575 	ret = kstrtoul(page, 0, &val);
3576 	if (ret < 0)
3577 		return ret;
3578 	if (val != !!val)
3579 		return -EINVAL;
3580 
3581 	ret = mutex_lock_interruptible(&sdev->sdev_mutex);
3582 	if (ret < 0)
3583 		return ret;
3584 	ret = mutex_lock_interruptible(&sport->mutex);
3585 	if (ret < 0)
3586 		goto unlock_sdev;
3587 	enabled = sport->enabled;
3588 	/* Log out all initiator systems before changing 'use_srq'. */
3589 	srpt_set_enabled(sport, false);
3590 	sport->port_attrib.use_srq = val;
3591 	srpt_use_srq(sdev, sport->port_attrib.use_srq);
3592 	srpt_set_enabled(sport, enabled);
3593 	ret = count;
3594 	mutex_unlock(&sport->mutex);
3595 unlock_sdev:
3596 	mutex_unlock(&sdev->sdev_mutex);
3597 
3598 	return ret;
3599 }
3600 
3601 CONFIGFS_ATTR(srpt_tpg_attrib_,  srp_max_rdma_size);
3602 CONFIGFS_ATTR(srpt_tpg_attrib_,  srp_max_rsp_size);
3603 CONFIGFS_ATTR(srpt_tpg_attrib_,  srp_sq_size);
3604 CONFIGFS_ATTR(srpt_tpg_attrib_,  use_srq);
3605 
3606 static struct configfs_attribute *srpt_tpg_attrib_attrs[] = {
3607 	&srpt_tpg_attrib_attr_srp_max_rdma_size,
3608 	&srpt_tpg_attrib_attr_srp_max_rsp_size,
3609 	&srpt_tpg_attrib_attr_srp_sq_size,
3610 	&srpt_tpg_attrib_attr_use_srq,
3611 	NULL,
3612 };
3613 
3614 static struct rdma_cm_id *srpt_create_rdma_id(struct sockaddr *listen_addr)
3615 {
3616 	struct rdma_cm_id *rdma_cm_id;
3617 	int ret;
3618 
3619 	rdma_cm_id = rdma_create_id(&init_net, srpt_rdma_cm_handler,
3620 				    NULL, RDMA_PS_TCP, IB_QPT_RC);
3621 	if (IS_ERR(rdma_cm_id)) {
3622 		pr_err("RDMA/CM ID creation failed: %ld\n",
3623 		       PTR_ERR(rdma_cm_id));
3624 		goto out;
3625 	}
3626 
3627 	ret = rdma_bind_addr(rdma_cm_id, listen_addr);
3628 	if (ret) {
3629 		char addr_str[64];
3630 
3631 		snprintf(addr_str, sizeof(addr_str), "%pISp", listen_addr);
3632 		pr_err("Binding RDMA/CM ID to address %s failed: %d\n",
3633 		       addr_str, ret);
3634 		rdma_destroy_id(rdma_cm_id);
3635 		rdma_cm_id = ERR_PTR(ret);
3636 		goto out;
3637 	}
3638 
3639 	ret = rdma_listen(rdma_cm_id, 128);
3640 	if (ret) {
3641 		pr_err("rdma_listen() failed: %d\n", ret);
3642 		rdma_destroy_id(rdma_cm_id);
3643 		rdma_cm_id = ERR_PTR(ret);
3644 	}
3645 
3646 out:
3647 	return rdma_cm_id;
3648 }
3649 
3650 static ssize_t srpt_rdma_cm_port_show(struct config_item *item, char *page)
3651 {
3652 	return sysfs_emit(page, "%d\n", rdma_cm_port);
3653 }
3654 
3655 static ssize_t srpt_rdma_cm_port_store(struct config_item *item,
3656 				       const char *page, size_t count)
3657 {
3658 	struct sockaddr_in  addr4 = { .sin_family  = AF_INET  };
3659 	struct sockaddr_in6 addr6 = { .sin6_family = AF_INET6 };
3660 	struct rdma_cm_id *new_id = NULL;
3661 	u16 val;
3662 	int ret;
3663 
3664 	ret = kstrtou16(page, 0, &val);
3665 	if (ret < 0)
3666 		return ret;
3667 	ret = count;
3668 	if (rdma_cm_port == val)
3669 		goto out;
3670 
3671 	if (val) {
3672 		addr6.sin6_port = cpu_to_be16(val);
3673 		new_id = srpt_create_rdma_id((struct sockaddr *)&addr6);
3674 		if (IS_ERR(new_id)) {
3675 			addr4.sin_port = cpu_to_be16(val);
3676 			new_id = srpt_create_rdma_id((struct sockaddr *)&addr4);
3677 			if (IS_ERR(new_id)) {
3678 				ret = PTR_ERR(new_id);
3679 				goto out;
3680 			}
3681 		}
3682 	}
3683 
3684 	mutex_lock(&rdma_cm_mutex);
3685 	rdma_cm_port = val;
3686 	swap(rdma_cm_id, new_id);
3687 	mutex_unlock(&rdma_cm_mutex);
3688 
3689 	if (new_id)
3690 		rdma_destroy_id(new_id);
3691 	ret = count;
3692 out:
3693 	return ret;
3694 }
3695 
3696 CONFIGFS_ATTR(srpt_, rdma_cm_port);
3697 
3698 static struct configfs_attribute *srpt_da_attrs[] = {
3699 	&srpt_attr_rdma_cm_port,
3700 	NULL,
3701 };
3702 
3703 static ssize_t srpt_tpg_enable_show(struct config_item *item, char *page)
3704 {
3705 	struct se_portal_group *se_tpg = to_tpg(item);
3706 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3707 
3708 	return sysfs_emit(page, "%d\n", sport->enabled);
3709 }
3710 
3711 static ssize_t srpt_tpg_enable_store(struct config_item *item,
3712 		const char *page, size_t count)
3713 {
3714 	struct se_portal_group *se_tpg = to_tpg(item);
3715 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3716 	unsigned long tmp;
3717 	int ret;
3718 
3719 	ret = kstrtoul(page, 0, &tmp);
3720 	if (ret < 0) {
3721 		pr_err("Unable to extract srpt_tpg_store_enable\n");
3722 		return -EINVAL;
3723 	}
3724 
3725 	if ((tmp != 0) && (tmp != 1)) {
3726 		pr_err("Illegal value for srpt_tpg_store_enable: %lu\n", tmp);
3727 		return -EINVAL;
3728 	}
3729 
3730 	mutex_lock(&sport->mutex);
3731 	srpt_set_enabled(sport, tmp);
3732 	mutex_unlock(&sport->mutex);
3733 
3734 	return count;
3735 }
3736 
3737 CONFIGFS_ATTR(srpt_tpg_, enable);
3738 
3739 static struct configfs_attribute *srpt_tpg_attrs[] = {
3740 	&srpt_tpg_attr_enable,
3741 	NULL,
3742 };
3743 
3744 /**
3745  * srpt_make_tpg - configfs callback invoked for mkdir /sys/kernel/config/target/$driver/$port/$tpg
3746  * @wwn: Corresponds to $driver/$port.
3747  * @name: $tpg.
3748  */
3749 static struct se_portal_group *srpt_make_tpg(struct se_wwn *wwn,
3750 					     const char *name)
3751 {
3752 	struct srpt_port_id *sport_id = srpt_wwn_to_sport_id(wwn);
3753 	struct srpt_tpg *stpg;
3754 	int res = -ENOMEM;
3755 
3756 	stpg = kzalloc(sizeof(*stpg), GFP_KERNEL);
3757 	if (!stpg)
3758 		return ERR_PTR(res);
3759 	stpg->sport_id = sport_id;
3760 	res = core_tpg_register(wwn, &stpg->tpg, SCSI_PROTOCOL_SRP);
3761 	if (res) {
3762 		kfree(stpg);
3763 		return ERR_PTR(res);
3764 	}
3765 
3766 	mutex_lock(&sport_id->mutex);
3767 	list_add_tail(&stpg->entry, &sport_id->tpg_list);
3768 	mutex_unlock(&sport_id->mutex);
3769 
3770 	return &stpg->tpg;
3771 }
3772 
3773 /**
3774  * srpt_drop_tpg - configfs callback invoked for rmdir /sys/kernel/config/target/$driver/$port/$tpg
3775  * @tpg: Target portal group to deregister.
3776  */
3777 static void srpt_drop_tpg(struct se_portal_group *tpg)
3778 {
3779 	struct srpt_tpg *stpg = container_of(tpg, typeof(*stpg), tpg);
3780 	struct srpt_port_id *sport_id = stpg->sport_id;
3781 	struct srpt_port *sport = srpt_tpg_to_sport(tpg);
3782 
3783 	mutex_lock(&sport_id->mutex);
3784 	list_del(&stpg->entry);
3785 	mutex_unlock(&sport_id->mutex);
3786 
3787 	sport->enabled = false;
3788 	core_tpg_deregister(tpg);
3789 	kfree(stpg);
3790 }
3791 
3792 /**
3793  * srpt_make_tport - configfs callback invoked for mkdir /sys/kernel/config/target/$driver/$port
3794  * @tf: Not used.
3795  * @group: Not used.
3796  * @name: $port.
3797  */
3798 static struct se_wwn *srpt_make_tport(struct target_fabric_configfs *tf,
3799 				      struct config_group *group,
3800 				      const char *name)
3801 {
3802 	return srpt_lookup_wwn(name) ? : ERR_PTR(-EINVAL);
3803 }
3804 
3805 /**
3806  * srpt_drop_tport - configfs callback invoked for rmdir /sys/kernel/config/target/$driver/$port
3807  * @wwn: $port.
3808  */
3809 static void srpt_drop_tport(struct se_wwn *wwn)
3810 {
3811 }
3812 
3813 static ssize_t srpt_wwn_version_show(struct config_item *item, char *buf)
3814 {
3815 	return sysfs_emit(buf, "\n");
3816 }
3817 
3818 CONFIGFS_ATTR_RO(srpt_wwn_, version);
3819 
3820 static struct configfs_attribute *srpt_wwn_attrs[] = {
3821 	&srpt_wwn_attr_version,
3822 	NULL,
3823 };
3824 
3825 static const struct target_core_fabric_ops srpt_template = {
3826 	.module				= THIS_MODULE,
3827 	.fabric_name			= "srpt",
3828 	.tpg_get_wwn			= srpt_get_fabric_wwn,
3829 	.tpg_get_tag			= srpt_get_tag,
3830 	.tpg_check_demo_mode		= srpt_check_false,
3831 	.tpg_check_demo_mode_cache	= srpt_check_true,
3832 	.tpg_check_demo_mode_write_protect = srpt_check_true,
3833 	.tpg_check_prod_mode_write_protect = srpt_check_false,
3834 	.tpg_get_inst_index		= srpt_tpg_get_inst_index,
3835 	.release_cmd			= srpt_release_cmd,
3836 	.check_stop_free		= srpt_check_stop_free,
3837 	.close_session			= srpt_close_session,
3838 	.sess_get_index			= srpt_sess_get_index,
3839 	.sess_get_initiator_sid		= NULL,
3840 	.write_pending			= srpt_write_pending,
3841 	.set_default_node_attributes	= srpt_set_default_node_attrs,
3842 	.get_cmd_state			= srpt_get_tcm_cmd_state,
3843 	.queue_data_in			= srpt_queue_data_in,
3844 	.queue_status			= srpt_queue_status,
3845 	.queue_tm_rsp			= srpt_queue_tm_rsp,
3846 	.aborted_task			= srpt_aborted_task,
3847 	/*
3848 	 * Setup function pointers for generic logic in
3849 	 * target_core_fabric_configfs.c
3850 	 */
3851 	.fabric_make_wwn		= srpt_make_tport,
3852 	.fabric_drop_wwn		= srpt_drop_tport,
3853 	.fabric_make_tpg		= srpt_make_tpg,
3854 	.fabric_drop_tpg		= srpt_drop_tpg,
3855 	.fabric_init_nodeacl		= srpt_init_nodeacl,
3856 
3857 	.tfc_discovery_attrs		= srpt_da_attrs,
3858 	.tfc_wwn_attrs			= srpt_wwn_attrs,
3859 	.tfc_tpg_base_attrs		= srpt_tpg_attrs,
3860 	.tfc_tpg_attrib_attrs		= srpt_tpg_attrib_attrs,
3861 };
3862 
3863 /**
3864  * srpt_init_module - kernel module initialization
3865  *
3866  * Note: Since ib_register_client() registers callback functions, and since at
3867  * least one of these callback functions (srpt_add_one()) calls target core
3868  * functions, this driver must be registered with the target core before
3869  * ib_register_client() is called.
3870  */
3871 static int __init srpt_init_module(void)
3872 {
3873 	int ret;
3874 
3875 	ret = -EINVAL;
3876 	if (srp_max_req_size < MIN_MAX_REQ_SIZE) {
3877 		pr_err("invalid value %d for kernel module parameter srp_max_req_size -- must be at least %d.\n",
3878 		       srp_max_req_size, MIN_MAX_REQ_SIZE);
3879 		goto out;
3880 	}
3881 
3882 	if (srpt_srq_size < MIN_SRPT_SRQ_SIZE
3883 	    || srpt_srq_size > MAX_SRPT_SRQ_SIZE) {
3884 		pr_err("invalid value %d for kernel module parameter srpt_srq_size -- must be in the range [%d..%d].\n",
3885 		       srpt_srq_size, MIN_SRPT_SRQ_SIZE, MAX_SRPT_SRQ_SIZE);
3886 		goto out;
3887 	}
3888 
3889 	ret = target_register_template(&srpt_template);
3890 	if (ret)
3891 		goto out;
3892 
3893 	ret = ib_register_client(&srpt_client);
3894 	if (ret) {
3895 		pr_err("couldn't register IB client\n");
3896 		goto out_unregister_target;
3897 	}
3898 
3899 	return 0;
3900 
3901 out_unregister_target:
3902 	target_unregister_template(&srpt_template);
3903 out:
3904 	return ret;
3905 }
3906 
3907 static void __exit srpt_cleanup_module(void)
3908 {
3909 	if (rdma_cm_id)
3910 		rdma_destroy_id(rdma_cm_id);
3911 	ib_unregister_client(&srpt_client);
3912 	target_unregister_template(&srpt_template);
3913 }
3914 
3915 module_init(srpt_init_module);
3916 module_exit(srpt_cleanup_module);
3917