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_agent *mad_agent;
553 	struct ib_mad_reg_req reg_req;
554 	struct ib_port_modify port_modify;
555 	struct ib_port_attr port_attr;
556 	int ret;
557 
558 	ret = ib_query_port(sport->sdev->device, sport->port, &port_attr);
559 	if (ret)
560 		return ret;
561 
562 	sport->sm_lid = port_attr.sm_lid;
563 	sport->lid = port_attr.lid;
564 
565 	ret = rdma_query_gid(sport->sdev->device, sport->port, 0, &sport->gid);
566 	if (ret)
567 		return ret;
568 
569 	srpt_format_guid(sport->guid_name, ARRAY_SIZE(sport->guid_name),
570 			 &sport->gid.global.interface_id);
571 	snprintf(sport->gid_name, ARRAY_SIZE(sport->gid_name),
572 		 "0x%016llx%016llx",
573 		 be64_to_cpu(sport->gid.global.subnet_prefix),
574 		 be64_to_cpu(sport->gid.global.interface_id));
575 
576 	if (rdma_protocol_iwarp(sport->sdev->device, sport->port))
577 		return 0;
578 
579 	memset(&port_modify, 0, sizeof(port_modify));
580 	port_modify.set_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP;
581 	port_modify.clr_port_cap_mask = 0;
582 
583 	ret = ib_modify_port(sport->sdev->device, sport->port, 0, &port_modify);
584 	if (ret) {
585 		pr_warn("%s-%d: enabling device management failed (%d). Note: this is expected if SR-IOV is enabled.\n",
586 			dev_name(&sport->sdev->device->dev), sport->port, ret);
587 		return 0;
588 	}
589 
590 	if (!sport->mad_agent) {
591 		memset(&reg_req, 0, sizeof(reg_req));
592 		reg_req.mgmt_class = IB_MGMT_CLASS_DEVICE_MGMT;
593 		reg_req.mgmt_class_version = IB_MGMT_BASE_VERSION;
594 		set_bit(IB_MGMT_METHOD_GET, reg_req.method_mask);
595 		set_bit(IB_MGMT_METHOD_SET, reg_req.method_mask);
596 
597 		mad_agent = ib_register_mad_agent(sport->sdev->device,
598 						  sport->port,
599 						  IB_QPT_GSI,
600 						  &reg_req, 0,
601 						  srpt_mad_send_handler,
602 						  srpt_mad_recv_handler,
603 						  sport, 0);
604 		if (IS_ERR(mad_agent)) {
605 			pr_err("%s-%d: MAD agent registration failed (%ld). Note: this is expected if SR-IOV is enabled.\n",
606 			       dev_name(&sport->sdev->device->dev), sport->port,
607 			       PTR_ERR(mad_agent));
608 			sport->mad_agent = NULL;
609 			memset(&port_modify, 0, sizeof(port_modify));
610 			port_modify.clr_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP;
611 			ib_modify_port(sport->sdev->device, sport->port, 0,
612 				       &port_modify);
613 			return 0;
614 		}
615 
616 		sport->mad_agent = mad_agent;
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->data, 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_init_cmd(cmd, ch->sess, &send_ioctx->sense_data[0],
1532 			     scsilun_to_int(&srp_cmd->lun), data_len,
1533 			     TCM_SIMPLE_TAG, dir, TARGET_SCF_ACK_KREF);
1534 	if (rc != 0) {
1535 		pr_debug("target_submit_cmd() returned %d for tag %#llx\n", rc,
1536 			 srp_cmd->tag);
1537 		goto busy;
1538 	}
1539 
1540 	if (target_submit_prep(cmd, srp_cmd->cdb, sg, sg_cnt, NULL, 0, NULL, 0,
1541 			       GFP_KERNEL))
1542 		return;
1543 
1544 	target_submit(cmd);
1545 	return;
1546 
1547 busy:
1548 	target_send_busy(cmd);
1549 }
1550 
1551 static int srp_tmr_to_tcm(int fn)
1552 {
1553 	switch (fn) {
1554 	case SRP_TSK_ABORT_TASK:
1555 		return TMR_ABORT_TASK;
1556 	case SRP_TSK_ABORT_TASK_SET:
1557 		return TMR_ABORT_TASK_SET;
1558 	case SRP_TSK_CLEAR_TASK_SET:
1559 		return TMR_CLEAR_TASK_SET;
1560 	case SRP_TSK_LUN_RESET:
1561 		return TMR_LUN_RESET;
1562 	case SRP_TSK_CLEAR_ACA:
1563 		return TMR_CLEAR_ACA;
1564 	default:
1565 		return -1;
1566 	}
1567 }
1568 
1569 /**
1570  * srpt_handle_tsk_mgmt - process a SRP_TSK_MGMT information unit
1571  * @ch: SRPT RDMA channel.
1572  * @recv_ioctx: Receive I/O context.
1573  * @send_ioctx: Send I/O context.
1574  *
1575  * Returns 0 if and only if the request will be processed by the target core.
1576  *
1577  * For more information about SRP_TSK_MGMT information units, see also section
1578  * 6.7 in the SRP r16a document.
1579  */
1580 static void srpt_handle_tsk_mgmt(struct srpt_rdma_ch *ch,
1581 				 struct srpt_recv_ioctx *recv_ioctx,
1582 				 struct srpt_send_ioctx *send_ioctx)
1583 {
1584 	struct srp_tsk_mgmt *srp_tsk;
1585 	struct se_cmd *cmd;
1586 	struct se_session *sess = ch->sess;
1587 	int tcm_tmr;
1588 	int rc;
1589 
1590 	BUG_ON(!send_ioctx);
1591 
1592 	srp_tsk = recv_ioctx->ioctx.buf + recv_ioctx->ioctx.offset;
1593 	cmd = &send_ioctx->cmd;
1594 
1595 	pr_debug("recv tsk_mgmt fn %d for task_tag %lld and cmd tag %lld ch %p sess %p\n",
1596 		 srp_tsk->tsk_mgmt_func, srp_tsk->task_tag, srp_tsk->tag, ch,
1597 		 ch->sess);
1598 
1599 	srpt_set_cmd_state(send_ioctx, SRPT_STATE_MGMT);
1600 	send_ioctx->cmd.tag = srp_tsk->tag;
1601 	tcm_tmr = srp_tmr_to_tcm(srp_tsk->tsk_mgmt_func);
1602 	rc = target_submit_tmr(&send_ioctx->cmd, sess, NULL,
1603 			       scsilun_to_int(&srp_tsk->lun), srp_tsk, tcm_tmr,
1604 			       GFP_KERNEL, srp_tsk->task_tag,
1605 			       TARGET_SCF_ACK_KREF);
1606 	if (rc != 0) {
1607 		send_ioctx->cmd.se_tmr_req->response = TMR_FUNCTION_REJECTED;
1608 		cmd->se_tfo->queue_tm_rsp(cmd);
1609 	}
1610 	return;
1611 }
1612 
1613 /**
1614  * srpt_handle_new_iu - process a newly received information unit
1615  * @ch:    RDMA channel through which the information unit has been received.
1616  * @recv_ioctx: Receive I/O context associated with the information unit.
1617  */
1618 static bool
1619 srpt_handle_new_iu(struct srpt_rdma_ch *ch, struct srpt_recv_ioctx *recv_ioctx)
1620 {
1621 	struct srpt_send_ioctx *send_ioctx = NULL;
1622 	struct srp_cmd *srp_cmd;
1623 	bool res = false;
1624 	u8 opcode;
1625 
1626 	BUG_ON(!ch);
1627 	BUG_ON(!recv_ioctx);
1628 
1629 	if (unlikely(ch->state == CH_CONNECTING))
1630 		goto push;
1631 
1632 	ib_dma_sync_single_for_cpu(ch->sport->sdev->device,
1633 				   recv_ioctx->ioctx.dma,
1634 				   recv_ioctx->ioctx.offset + srp_max_req_size,
1635 				   DMA_FROM_DEVICE);
1636 
1637 	srp_cmd = recv_ioctx->ioctx.buf + recv_ioctx->ioctx.offset;
1638 	opcode = srp_cmd->opcode;
1639 	if (opcode == SRP_CMD || opcode == SRP_TSK_MGMT) {
1640 		send_ioctx = srpt_get_send_ioctx(ch);
1641 		if (unlikely(!send_ioctx))
1642 			goto push;
1643 	}
1644 
1645 	if (!list_empty(&recv_ioctx->wait_list)) {
1646 		WARN_ON_ONCE(!ch->processing_wait_list);
1647 		list_del_init(&recv_ioctx->wait_list);
1648 	}
1649 
1650 	switch (opcode) {
1651 	case SRP_CMD:
1652 		srpt_handle_cmd(ch, recv_ioctx, send_ioctx);
1653 		break;
1654 	case SRP_TSK_MGMT:
1655 		srpt_handle_tsk_mgmt(ch, recv_ioctx, send_ioctx);
1656 		break;
1657 	case SRP_I_LOGOUT:
1658 		pr_err("Not yet implemented: SRP_I_LOGOUT\n");
1659 		break;
1660 	case SRP_CRED_RSP:
1661 		pr_debug("received SRP_CRED_RSP\n");
1662 		break;
1663 	case SRP_AER_RSP:
1664 		pr_debug("received SRP_AER_RSP\n");
1665 		break;
1666 	case SRP_RSP:
1667 		pr_err("Received SRP_RSP\n");
1668 		break;
1669 	default:
1670 		pr_err("received IU with unknown opcode 0x%x\n", opcode);
1671 		break;
1672 	}
1673 
1674 	if (!send_ioctx || !send_ioctx->recv_ioctx)
1675 		srpt_post_recv(ch->sport->sdev, ch, recv_ioctx);
1676 	res = true;
1677 
1678 out:
1679 	return res;
1680 
1681 push:
1682 	if (list_empty(&recv_ioctx->wait_list)) {
1683 		WARN_ON_ONCE(ch->processing_wait_list);
1684 		list_add_tail(&recv_ioctx->wait_list, &ch->cmd_wait_list);
1685 	}
1686 	goto out;
1687 }
1688 
1689 static void srpt_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1690 {
1691 	struct srpt_rdma_ch *ch = wc->qp->qp_context;
1692 	struct srpt_recv_ioctx *ioctx =
1693 		container_of(wc->wr_cqe, struct srpt_recv_ioctx, ioctx.cqe);
1694 
1695 	if (wc->status == IB_WC_SUCCESS) {
1696 		int req_lim;
1697 
1698 		req_lim = atomic_dec_return(&ch->req_lim);
1699 		if (unlikely(req_lim < 0))
1700 			pr_err("req_lim = %d < 0\n", req_lim);
1701 		ioctx->byte_len = wc->byte_len;
1702 		srpt_handle_new_iu(ch, ioctx);
1703 	} else {
1704 		pr_info_ratelimited("receiving failed for ioctx %p with status %d\n",
1705 				    ioctx, wc->status);
1706 	}
1707 }
1708 
1709 /*
1710  * This function must be called from the context in which RDMA completions are
1711  * processed because it accesses the wait list without protection against
1712  * access from other threads.
1713  */
1714 static void srpt_process_wait_list(struct srpt_rdma_ch *ch)
1715 {
1716 	struct srpt_recv_ioctx *recv_ioctx, *tmp;
1717 
1718 	WARN_ON_ONCE(ch->state == CH_CONNECTING);
1719 
1720 	if (list_empty(&ch->cmd_wait_list))
1721 		return;
1722 
1723 	WARN_ON_ONCE(ch->processing_wait_list);
1724 	ch->processing_wait_list = true;
1725 	list_for_each_entry_safe(recv_ioctx, tmp, &ch->cmd_wait_list,
1726 				 wait_list) {
1727 		if (!srpt_handle_new_iu(ch, recv_ioctx))
1728 			break;
1729 	}
1730 	ch->processing_wait_list = false;
1731 }
1732 
1733 /**
1734  * srpt_send_done - send completion callback
1735  * @cq: Completion queue.
1736  * @wc: Work completion.
1737  *
1738  * Note: Although this has not yet been observed during tests, at least in
1739  * theory it is possible that the srpt_get_send_ioctx() call invoked by
1740  * srpt_handle_new_iu() fails. This is possible because the req_lim_delta
1741  * value in each response is set to one, and it is possible that this response
1742  * makes the initiator send a new request before the send completion for that
1743  * response has been processed. This could e.g. happen if the call to
1744  * srpt_put_send_iotcx() is delayed because of a higher priority interrupt or
1745  * if IB retransmission causes generation of the send completion to be
1746  * delayed. Incoming information units for which srpt_get_send_ioctx() fails
1747  * are queued on cmd_wait_list. The code below processes these delayed
1748  * requests one at a time.
1749  */
1750 static void srpt_send_done(struct ib_cq *cq, struct ib_wc *wc)
1751 {
1752 	struct srpt_rdma_ch *ch = wc->qp->qp_context;
1753 	struct srpt_send_ioctx *ioctx =
1754 		container_of(wc->wr_cqe, struct srpt_send_ioctx, ioctx.cqe);
1755 	enum srpt_command_state state;
1756 
1757 	state = srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
1758 
1759 	WARN_ON(state != SRPT_STATE_CMD_RSP_SENT &&
1760 		state != SRPT_STATE_MGMT_RSP_SENT);
1761 
1762 	atomic_add(1 + ioctx->n_rdma, &ch->sq_wr_avail);
1763 
1764 	if (wc->status != IB_WC_SUCCESS)
1765 		pr_info("sending response for ioctx 0x%p failed with status %d\n",
1766 			ioctx, wc->status);
1767 
1768 	if (state != SRPT_STATE_DONE) {
1769 		transport_generic_free_cmd(&ioctx->cmd, 0);
1770 	} else {
1771 		pr_err("IB completion has been received too late for wr_id = %u.\n",
1772 		       ioctx->ioctx.index);
1773 	}
1774 
1775 	srpt_process_wait_list(ch);
1776 }
1777 
1778 /**
1779  * srpt_create_ch_ib - create receive and send completion queues
1780  * @ch: SRPT RDMA channel.
1781  */
1782 static int srpt_create_ch_ib(struct srpt_rdma_ch *ch)
1783 {
1784 	struct ib_qp_init_attr *qp_init;
1785 	struct srpt_port *sport = ch->sport;
1786 	struct srpt_device *sdev = sport->sdev;
1787 	const struct ib_device_attr *attrs = &sdev->device->attrs;
1788 	int sq_size = sport->port_attrib.srp_sq_size;
1789 	int i, ret;
1790 
1791 	WARN_ON(ch->rq_size < 1);
1792 
1793 	ret = -ENOMEM;
1794 	qp_init = kzalloc(sizeof(*qp_init), GFP_KERNEL);
1795 	if (!qp_init)
1796 		goto out;
1797 
1798 retry:
1799 	ch->cq = ib_cq_pool_get(sdev->device, ch->rq_size + sq_size, -1,
1800 				 IB_POLL_WORKQUEUE);
1801 	if (IS_ERR(ch->cq)) {
1802 		ret = PTR_ERR(ch->cq);
1803 		pr_err("failed to create CQ cqe= %d ret= %d\n",
1804 		       ch->rq_size + sq_size, ret);
1805 		goto out;
1806 	}
1807 	ch->cq_size = ch->rq_size + sq_size;
1808 
1809 	qp_init->qp_context = (void *)ch;
1810 	qp_init->event_handler
1811 		= (void(*)(struct ib_event *, void*))srpt_qp_event;
1812 	qp_init->send_cq = ch->cq;
1813 	qp_init->recv_cq = ch->cq;
1814 	qp_init->sq_sig_type = IB_SIGNAL_REQ_WR;
1815 	qp_init->qp_type = IB_QPT_RC;
1816 	/*
1817 	 * We divide up our send queue size into half SEND WRs to send the
1818 	 * completions, and half R/W contexts to actually do the RDMA
1819 	 * READ/WRITE transfers.  Note that we need to allocate CQ slots for
1820 	 * both both, as RDMA contexts will also post completions for the
1821 	 * RDMA READ case.
1822 	 */
1823 	qp_init->cap.max_send_wr = min(sq_size / 2, attrs->max_qp_wr);
1824 	qp_init->cap.max_rdma_ctxs = sq_size / 2;
1825 	qp_init->cap.max_send_sge = attrs->max_send_sge;
1826 	qp_init->cap.max_recv_sge = 1;
1827 	qp_init->port_num = ch->sport->port;
1828 	if (sdev->use_srq)
1829 		qp_init->srq = sdev->srq;
1830 	else
1831 		qp_init->cap.max_recv_wr = ch->rq_size;
1832 
1833 	if (ch->using_rdma_cm) {
1834 		ret = rdma_create_qp(ch->rdma_cm.cm_id, sdev->pd, qp_init);
1835 		ch->qp = ch->rdma_cm.cm_id->qp;
1836 	} else {
1837 		ch->qp = ib_create_qp(sdev->pd, qp_init);
1838 		if (!IS_ERR(ch->qp)) {
1839 			ret = srpt_init_ch_qp(ch, ch->qp);
1840 			if (ret)
1841 				ib_destroy_qp(ch->qp);
1842 		} else {
1843 			ret = PTR_ERR(ch->qp);
1844 		}
1845 	}
1846 	if (ret) {
1847 		bool retry = sq_size > MIN_SRPT_SQ_SIZE;
1848 
1849 		if (retry) {
1850 			pr_debug("failed to create queue pair with sq_size = %d (%d) - retrying\n",
1851 				 sq_size, ret);
1852 			ib_cq_pool_put(ch->cq, ch->cq_size);
1853 			sq_size = max(sq_size / 2, MIN_SRPT_SQ_SIZE);
1854 			goto retry;
1855 		} else {
1856 			pr_err("failed to create queue pair with sq_size = %d (%d)\n",
1857 			       sq_size, ret);
1858 			goto err_destroy_cq;
1859 		}
1860 	}
1861 
1862 	atomic_set(&ch->sq_wr_avail, qp_init->cap.max_send_wr);
1863 
1864 	pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d ch= %p\n",
1865 		 __func__, ch->cq->cqe, qp_init->cap.max_send_sge,
1866 		 qp_init->cap.max_send_wr, ch);
1867 
1868 	if (!sdev->use_srq)
1869 		for (i = 0; i < ch->rq_size; i++)
1870 			srpt_post_recv(sdev, ch, ch->ioctx_recv_ring[i]);
1871 
1872 out:
1873 	kfree(qp_init);
1874 	return ret;
1875 
1876 err_destroy_cq:
1877 	ch->qp = NULL;
1878 	ib_cq_pool_put(ch->cq, ch->cq_size);
1879 	goto out;
1880 }
1881 
1882 static void srpt_destroy_ch_ib(struct srpt_rdma_ch *ch)
1883 {
1884 	ib_destroy_qp(ch->qp);
1885 	ib_cq_pool_put(ch->cq, ch->cq_size);
1886 }
1887 
1888 /**
1889  * srpt_close_ch - close a RDMA channel
1890  * @ch: SRPT RDMA channel.
1891  *
1892  * Make sure all resources associated with the channel will be deallocated at
1893  * an appropriate time.
1894  *
1895  * Returns true if and only if the channel state has been modified into
1896  * CH_DRAINING.
1897  */
1898 static bool srpt_close_ch(struct srpt_rdma_ch *ch)
1899 {
1900 	int ret;
1901 
1902 	if (!srpt_set_ch_state(ch, CH_DRAINING)) {
1903 		pr_debug("%s: already closed\n", ch->sess_name);
1904 		return false;
1905 	}
1906 
1907 	kref_get(&ch->kref);
1908 
1909 	ret = srpt_ch_qp_err(ch);
1910 	if (ret < 0)
1911 		pr_err("%s-%d: changing queue pair into error state failed: %d\n",
1912 		       ch->sess_name, ch->qp->qp_num, ret);
1913 
1914 	ret = srpt_zerolength_write(ch);
1915 	if (ret < 0) {
1916 		pr_err("%s-%d: queuing zero-length write failed: %d\n",
1917 		       ch->sess_name, ch->qp->qp_num, ret);
1918 		if (srpt_set_ch_state(ch, CH_DISCONNECTED))
1919 			schedule_work(&ch->release_work);
1920 		else
1921 			WARN_ON_ONCE(true);
1922 	}
1923 
1924 	kref_put(&ch->kref, srpt_free_ch);
1925 
1926 	return true;
1927 }
1928 
1929 /*
1930  * Change the channel state into CH_DISCONNECTING. If a channel has not yet
1931  * reached the connected state, close it. If a channel is in the connected
1932  * state, send a DREQ. If a DREQ has been received, send a DREP. Note: it is
1933  * the responsibility of the caller to ensure that this function is not
1934  * invoked concurrently with the code that accepts a connection. This means
1935  * that this function must either be invoked from inside a CM callback
1936  * function or that it must be invoked with the srpt_port.mutex held.
1937  */
1938 static int srpt_disconnect_ch(struct srpt_rdma_ch *ch)
1939 {
1940 	int ret;
1941 
1942 	if (!srpt_set_ch_state(ch, CH_DISCONNECTING))
1943 		return -ENOTCONN;
1944 
1945 	if (ch->using_rdma_cm) {
1946 		ret = rdma_disconnect(ch->rdma_cm.cm_id);
1947 	} else {
1948 		ret = ib_send_cm_dreq(ch->ib_cm.cm_id, NULL, 0);
1949 		if (ret < 0)
1950 			ret = ib_send_cm_drep(ch->ib_cm.cm_id, NULL, 0);
1951 	}
1952 
1953 	if (ret < 0 && srpt_close_ch(ch))
1954 		ret = 0;
1955 
1956 	return ret;
1957 }
1958 
1959 /* Send DREQ and wait for DREP. */
1960 static void srpt_disconnect_ch_sync(struct srpt_rdma_ch *ch)
1961 {
1962 	DECLARE_COMPLETION_ONSTACK(closed);
1963 	struct srpt_port *sport = ch->sport;
1964 
1965 	pr_debug("ch %s-%d state %d\n", ch->sess_name, ch->qp->qp_num,
1966 		 ch->state);
1967 
1968 	ch->closed = &closed;
1969 
1970 	mutex_lock(&sport->mutex);
1971 	srpt_disconnect_ch(ch);
1972 	mutex_unlock(&sport->mutex);
1973 
1974 	while (wait_for_completion_timeout(&closed, 5 * HZ) == 0)
1975 		pr_info("%s(%s-%d state %d): still waiting ...\n", __func__,
1976 			ch->sess_name, ch->qp->qp_num, ch->state);
1977 
1978 }
1979 
1980 static void __srpt_close_all_ch(struct srpt_port *sport)
1981 {
1982 	struct srpt_nexus *nexus;
1983 	struct srpt_rdma_ch *ch;
1984 
1985 	lockdep_assert_held(&sport->mutex);
1986 
1987 	list_for_each_entry(nexus, &sport->nexus_list, entry) {
1988 		list_for_each_entry(ch, &nexus->ch_list, list) {
1989 			if (srpt_disconnect_ch(ch) >= 0)
1990 				pr_info("Closing channel %s-%d because target %s_%d has been disabled\n",
1991 					ch->sess_name, ch->qp->qp_num,
1992 					dev_name(&sport->sdev->device->dev),
1993 					sport->port);
1994 			srpt_close_ch(ch);
1995 		}
1996 	}
1997 }
1998 
1999 /*
2000  * Look up (i_port_id, t_port_id) in sport->nexus_list. Create an entry if
2001  * it does not yet exist.
2002  */
2003 static struct srpt_nexus *srpt_get_nexus(struct srpt_port *sport,
2004 					 const u8 i_port_id[16],
2005 					 const u8 t_port_id[16])
2006 {
2007 	struct srpt_nexus *nexus = NULL, *tmp_nexus = NULL, *n;
2008 
2009 	for (;;) {
2010 		mutex_lock(&sport->mutex);
2011 		list_for_each_entry(n, &sport->nexus_list, entry) {
2012 			if (memcmp(n->i_port_id, i_port_id, 16) == 0 &&
2013 			    memcmp(n->t_port_id, t_port_id, 16) == 0) {
2014 				nexus = n;
2015 				break;
2016 			}
2017 		}
2018 		if (!nexus && tmp_nexus) {
2019 			list_add_tail_rcu(&tmp_nexus->entry,
2020 					  &sport->nexus_list);
2021 			swap(nexus, tmp_nexus);
2022 		}
2023 		mutex_unlock(&sport->mutex);
2024 
2025 		if (nexus)
2026 			break;
2027 		tmp_nexus = kzalloc(sizeof(*nexus), GFP_KERNEL);
2028 		if (!tmp_nexus) {
2029 			nexus = ERR_PTR(-ENOMEM);
2030 			break;
2031 		}
2032 		INIT_LIST_HEAD(&tmp_nexus->ch_list);
2033 		memcpy(tmp_nexus->i_port_id, i_port_id, 16);
2034 		memcpy(tmp_nexus->t_port_id, t_port_id, 16);
2035 	}
2036 
2037 	kfree(tmp_nexus);
2038 
2039 	return nexus;
2040 }
2041 
2042 static void srpt_set_enabled(struct srpt_port *sport, bool enabled)
2043 	__must_hold(&sport->mutex)
2044 {
2045 	lockdep_assert_held(&sport->mutex);
2046 
2047 	if (sport->enabled == enabled)
2048 		return;
2049 	sport->enabled = enabled;
2050 	if (!enabled)
2051 		__srpt_close_all_ch(sport);
2052 }
2053 
2054 static void srpt_drop_sport_ref(struct srpt_port *sport)
2055 {
2056 	if (atomic_dec_return(&sport->refcount) == 0 && sport->freed_channels)
2057 		complete(sport->freed_channels);
2058 }
2059 
2060 static void srpt_free_ch(struct kref *kref)
2061 {
2062 	struct srpt_rdma_ch *ch = container_of(kref, struct srpt_rdma_ch, kref);
2063 
2064 	srpt_drop_sport_ref(ch->sport);
2065 	kfree_rcu(ch, rcu);
2066 }
2067 
2068 /*
2069  * Shut down the SCSI target session, tell the connection manager to
2070  * disconnect the associated RDMA channel, transition the QP to the error
2071  * state and remove the channel from the channel list. This function is
2072  * typically called from inside srpt_zerolength_write_done(). Concurrent
2073  * srpt_zerolength_write() calls from inside srpt_close_ch() are possible
2074  * as long as the channel is on sport->nexus_list.
2075  */
2076 static void srpt_release_channel_work(struct work_struct *w)
2077 {
2078 	struct srpt_rdma_ch *ch;
2079 	struct srpt_device *sdev;
2080 	struct srpt_port *sport;
2081 	struct se_session *se_sess;
2082 
2083 	ch = container_of(w, struct srpt_rdma_ch, release_work);
2084 	pr_debug("%s-%d\n", ch->sess_name, ch->qp->qp_num);
2085 
2086 	sdev = ch->sport->sdev;
2087 	BUG_ON(!sdev);
2088 
2089 	se_sess = ch->sess;
2090 	BUG_ON(!se_sess);
2091 
2092 	target_stop_session(se_sess);
2093 	target_wait_for_sess_cmds(se_sess);
2094 
2095 	target_remove_session(se_sess);
2096 	ch->sess = NULL;
2097 
2098 	if (ch->using_rdma_cm)
2099 		rdma_destroy_id(ch->rdma_cm.cm_id);
2100 	else
2101 		ib_destroy_cm_id(ch->ib_cm.cm_id);
2102 
2103 	sport = ch->sport;
2104 	mutex_lock(&sport->mutex);
2105 	list_del_rcu(&ch->list);
2106 	mutex_unlock(&sport->mutex);
2107 
2108 	if (ch->closed)
2109 		complete(ch->closed);
2110 
2111 	srpt_destroy_ch_ib(ch);
2112 
2113 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
2114 			     ch->sport->sdev, ch->rq_size,
2115 			     ch->rsp_buf_cache, DMA_TO_DEVICE);
2116 
2117 	kmem_cache_destroy(ch->rsp_buf_cache);
2118 
2119 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_recv_ring,
2120 			     sdev, ch->rq_size,
2121 			     ch->req_buf_cache, DMA_FROM_DEVICE);
2122 
2123 	kmem_cache_destroy(ch->req_buf_cache);
2124 
2125 	kref_put(&ch->kref, srpt_free_ch);
2126 }
2127 
2128 /**
2129  * srpt_cm_req_recv - process the event IB_CM_REQ_RECEIVED
2130  * @sdev: HCA through which the login request was received.
2131  * @ib_cm_id: IB/CM connection identifier in case of IB/CM.
2132  * @rdma_cm_id: RDMA/CM connection identifier in case of RDMA/CM.
2133  * @port_num: Port through which the REQ message was received.
2134  * @pkey: P_Key of the incoming connection.
2135  * @req: SRP login request.
2136  * @src_addr: GID (IB/CM) or IP address (RDMA/CM) of the port that submitted
2137  * the login request.
2138  *
2139  * Ownership of the cm_id is transferred to the target session if this
2140  * function returns zero. Otherwise the caller remains the owner of cm_id.
2141  */
2142 static int srpt_cm_req_recv(struct srpt_device *const sdev,
2143 			    struct ib_cm_id *ib_cm_id,
2144 			    struct rdma_cm_id *rdma_cm_id,
2145 			    u8 port_num, __be16 pkey,
2146 			    const struct srp_login_req *req,
2147 			    const char *src_addr)
2148 {
2149 	struct srpt_port *sport = &sdev->port[port_num - 1];
2150 	struct srpt_nexus *nexus;
2151 	struct srp_login_rsp *rsp = NULL;
2152 	struct srp_login_rej *rej = NULL;
2153 	union {
2154 		struct rdma_conn_param rdma_cm;
2155 		struct ib_cm_rep_param ib_cm;
2156 	} *rep_param = NULL;
2157 	struct srpt_rdma_ch *ch = NULL;
2158 	char i_port_id[36];
2159 	u32 it_iu_len;
2160 	int i, tag_num, tag_size, ret;
2161 	struct srpt_tpg *stpg;
2162 
2163 	WARN_ON_ONCE(irqs_disabled());
2164 
2165 	it_iu_len = be32_to_cpu(req->req_it_iu_len);
2166 
2167 	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",
2168 		req->initiator_port_id, req->target_port_id, it_iu_len,
2169 		port_num, &sport->gid, be16_to_cpu(pkey));
2170 
2171 	nexus = srpt_get_nexus(sport, req->initiator_port_id,
2172 			       req->target_port_id);
2173 	if (IS_ERR(nexus)) {
2174 		ret = PTR_ERR(nexus);
2175 		goto out;
2176 	}
2177 
2178 	ret = -ENOMEM;
2179 	rsp = kzalloc(sizeof(*rsp), GFP_KERNEL);
2180 	rej = kzalloc(sizeof(*rej), GFP_KERNEL);
2181 	rep_param = kzalloc(sizeof(*rep_param), GFP_KERNEL);
2182 	if (!rsp || !rej || !rep_param)
2183 		goto out;
2184 
2185 	ret = -EINVAL;
2186 	if (it_iu_len > srp_max_req_size || it_iu_len < 64) {
2187 		rej->reason = cpu_to_be32(
2188 				SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE);
2189 		pr_err("rejected SRP_LOGIN_REQ because its length (%d bytes) is out of range (%d .. %d)\n",
2190 		       it_iu_len, 64, srp_max_req_size);
2191 		goto reject;
2192 	}
2193 
2194 	if (!sport->enabled) {
2195 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2196 		pr_info("rejected SRP_LOGIN_REQ because target port %s_%d has not yet been enabled\n",
2197 			dev_name(&sport->sdev->device->dev), port_num);
2198 		goto reject;
2199 	}
2200 
2201 	if (*(__be64 *)req->target_port_id != cpu_to_be64(srpt_service_guid)
2202 	    || *(__be64 *)(req->target_port_id + 8) !=
2203 	       cpu_to_be64(srpt_service_guid)) {
2204 		rej->reason = cpu_to_be32(
2205 				SRP_LOGIN_REJ_UNABLE_ASSOCIATE_CHANNEL);
2206 		pr_err("rejected SRP_LOGIN_REQ because it has an invalid target port identifier.\n");
2207 		goto reject;
2208 	}
2209 
2210 	ret = -ENOMEM;
2211 	ch = kzalloc(sizeof(*ch), GFP_KERNEL);
2212 	if (!ch) {
2213 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2214 		pr_err("rejected SRP_LOGIN_REQ because out of memory.\n");
2215 		goto reject;
2216 	}
2217 
2218 	kref_init(&ch->kref);
2219 	ch->pkey = be16_to_cpu(pkey);
2220 	ch->nexus = nexus;
2221 	ch->zw_cqe.done = srpt_zerolength_write_done;
2222 	INIT_WORK(&ch->release_work, srpt_release_channel_work);
2223 	ch->sport = sport;
2224 	if (rdma_cm_id) {
2225 		ch->using_rdma_cm = true;
2226 		ch->rdma_cm.cm_id = rdma_cm_id;
2227 		rdma_cm_id->context = ch;
2228 	} else {
2229 		ch->ib_cm.cm_id = ib_cm_id;
2230 		ib_cm_id->context = ch;
2231 	}
2232 	/*
2233 	 * ch->rq_size should be at least as large as the initiator queue
2234 	 * depth to avoid that the initiator driver has to report QUEUE_FULL
2235 	 * to the SCSI mid-layer.
2236 	 */
2237 	ch->rq_size = min(MAX_SRPT_RQ_SIZE, sdev->device->attrs.max_qp_wr);
2238 	spin_lock_init(&ch->spinlock);
2239 	ch->state = CH_CONNECTING;
2240 	INIT_LIST_HEAD(&ch->cmd_wait_list);
2241 	ch->max_rsp_size = ch->sport->port_attrib.srp_max_rsp_size;
2242 
2243 	ch->rsp_buf_cache = kmem_cache_create("srpt-rsp-buf", ch->max_rsp_size,
2244 					      512, 0, NULL);
2245 	if (!ch->rsp_buf_cache)
2246 		goto free_ch;
2247 
2248 	ch->ioctx_ring = (struct srpt_send_ioctx **)
2249 		srpt_alloc_ioctx_ring(ch->sport->sdev, ch->rq_size,
2250 				      sizeof(*ch->ioctx_ring[0]),
2251 				      ch->rsp_buf_cache, 0, DMA_TO_DEVICE);
2252 	if (!ch->ioctx_ring) {
2253 		pr_err("rejected SRP_LOGIN_REQ because creating a new QP SQ ring failed.\n");
2254 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2255 		goto free_rsp_cache;
2256 	}
2257 
2258 	for (i = 0; i < ch->rq_size; i++)
2259 		ch->ioctx_ring[i]->ch = ch;
2260 	if (!sdev->use_srq) {
2261 		u16 imm_data_offset = req->req_flags & SRP_IMMED_REQUESTED ?
2262 			be16_to_cpu(req->imm_data_offset) : 0;
2263 		u16 alignment_offset;
2264 		u32 req_sz;
2265 
2266 		if (req->req_flags & SRP_IMMED_REQUESTED)
2267 			pr_debug("imm_data_offset = %d\n",
2268 				 be16_to_cpu(req->imm_data_offset));
2269 		if (imm_data_offset >= sizeof(struct srp_cmd)) {
2270 			ch->imm_data_offset = imm_data_offset;
2271 			rsp->rsp_flags |= SRP_LOGIN_RSP_IMMED_SUPP;
2272 		} else {
2273 			ch->imm_data_offset = 0;
2274 		}
2275 		alignment_offset = round_up(imm_data_offset, 512) -
2276 			imm_data_offset;
2277 		req_sz = alignment_offset + imm_data_offset + srp_max_req_size;
2278 		ch->req_buf_cache = kmem_cache_create("srpt-req-buf", req_sz,
2279 						      512, 0, NULL);
2280 		if (!ch->req_buf_cache)
2281 			goto free_rsp_ring;
2282 
2283 		ch->ioctx_recv_ring = (struct srpt_recv_ioctx **)
2284 			srpt_alloc_ioctx_ring(ch->sport->sdev, ch->rq_size,
2285 					      sizeof(*ch->ioctx_recv_ring[0]),
2286 					      ch->req_buf_cache,
2287 					      alignment_offset,
2288 					      DMA_FROM_DEVICE);
2289 		if (!ch->ioctx_recv_ring) {
2290 			pr_err("rejected SRP_LOGIN_REQ because creating a new QP RQ ring failed.\n");
2291 			rej->reason =
2292 			    cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2293 			goto free_recv_cache;
2294 		}
2295 		for (i = 0; i < ch->rq_size; i++)
2296 			INIT_LIST_HEAD(&ch->ioctx_recv_ring[i]->wait_list);
2297 	}
2298 
2299 	ret = srpt_create_ch_ib(ch);
2300 	if (ret) {
2301 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2302 		pr_err("rejected SRP_LOGIN_REQ because creating a new RDMA channel failed.\n");
2303 		goto free_recv_ring;
2304 	}
2305 
2306 	strscpy(ch->sess_name, src_addr, sizeof(ch->sess_name));
2307 	snprintf(i_port_id, sizeof(i_port_id), "0x%016llx%016llx",
2308 			be64_to_cpu(*(__be64 *)nexus->i_port_id),
2309 			be64_to_cpu(*(__be64 *)(nexus->i_port_id + 8)));
2310 
2311 	pr_debug("registering src addr %s or i_port_id %s\n", ch->sess_name,
2312 		 i_port_id);
2313 
2314 	tag_num = ch->rq_size;
2315 	tag_size = 1; /* ib_srpt does not use se_sess->sess_cmd_map */
2316 
2317 	if (sport->guid_id) {
2318 		mutex_lock(&sport->guid_id->mutex);
2319 		list_for_each_entry(stpg, &sport->guid_id->tpg_list, entry) {
2320 			if (!IS_ERR_OR_NULL(ch->sess))
2321 				break;
2322 			ch->sess = target_setup_session(&stpg->tpg, tag_num,
2323 						tag_size, TARGET_PROT_NORMAL,
2324 						ch->sess_name, ch, NULL);
2325 		}
2326 		mutex_unlock(&sport->guid_id->mutex);
2327 	}
2328 
2329 	if (sport->gid_id) {
2330 		mutex_lock(&sport->gid_id->mutex);
2331 		list_for_each_entry(stpg, &sport->gid_id->tpg_list, entry) {
2332 			if (!IS_ERR_OR_NULL(ch->sess))
2333 				break;
2334 			ch->sess = target_setup_session(&stpg->tpg, tag_num,
2335 					tag_size, TARGET_PROT_NORMAL, i_port_id,
2336 					ch, NULL);
2337 			if (!IS_ERR_OR_NULL(ch->sess))
2338 				break;
2339 			/* Retry without leading "0x" */
2340 			ch->sess = target_setup_session(&stpg->tpg, tag_num,
2341 						tag_size, TARGET_PROT_NORMAL,
2342 						i_port_id + 2, ch, NULL);
2343 		}
2344 		mutex_unlock(&sport->gid_id->mutex);
2345 	}
2346 
2347 	if (IS_ERR_OR_NULL(ch->sess)) {
2348 		WARN_ON_ONCE(ch->sess == NULL);
2349 		ret = PTR_ERR(ch->sess);
2350 		ch->sess = NULL;
2351 		pr_info("Rejected login for initiator %s: ret = %d.\n",
2352 			ch->sess_name, ret);
2353 		rej->reason = cpu_to_be32(ret == -ENOMEM ?
2354 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES :
2355 				SRP_LOGIN_REJ_CHANNEL_LIMIT_REACHED);
2356 		goto destroy_ib;
2357 	}
2358 
2359 	/*
2360 	 * Once a session has been created destruction of srpt_rdma_ch objects
2361 	 * will decrement sport->refcount. Hence increment sport->refcount now.
2362 	 */
2363 	atomic_inc(&sport->refcount);
2364 
2365 	mutex_lock(&sport->mutex);
2366 
2367 	if ((req->req_flags & SRP_MTCH_ACTION) == SRP_MULTICHAN_SINGLE) {
2368 		struct srpt_rdma_ch *ch2;
2369 
2370 		list_for_each_entry(ch2, &nexus->ch_list, list) {
2371 			if (srpt_disconnect_ch(ch2) < 0)
2372 				continue;
2373 			pr_info("Relogin - closed existing channel %s\n",
2374 				ch2->sess_name);
2375 			rsp->rsp_flags |= SRP_LOGIN_RSP_MULTICHAN_TERMINATED;
2376 		}
2377 	} else {
2378 		rsp->rsp_flags |= SRP_LOGIN_RSP_MULTICHAN_MAINTAINED;
2379 	}
2380 
2381 	list_add_tail_rcu(&ch->list, &nexus->ch_list);
2382 
2383 	if (!sport->enabled) {
2384 		rej->reason = cpu_to_be32(
2385 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2386 		pr_info("rejected SRP_LOGIN_REQ because target %s_%d is not enabled\n",
2387 			dev_name(&sdev->device->dev), port_num);
2388 		mutex_unlock(&sport->mutex);
2389 		ret = -EINVAL;
2390 		goto reject;
2391 	}
2392 
2393 	mutex_unlock(&sport->mutex);
2394 
2395 	ret = ch->using_rdma_cm ? 0 : srpt_ch_qp_rtr(ch, ch->qp);
2396 	if (ret) {
2397 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2398 		pr_err("rejected SRP_LOGIN_REQ because enabling RTR failed (error code = %d)\n",
2399 		       ret);
2400 		goto reject;
2401 	}
2402 
2403 	pr_debug("Establish connection sess=%p name=%s ch=%p\n", ch->sess,
2404 		 ch->sess_name, ch);
2405 
2406 	/* create srp_login_response */
2407 	rsp->opcode = SRP_LOGIN_RSP;
2408 	rsp->tag = req->tag;
2409 	rsp->max_it_iu_len = cpu_to_be32(srp_max_req_size);
2410 	rsp->max_ti_iu_len = req->req_it_iu_len;
2411 	ch->max_ti_iu_len = it_iu_len;
2412 	rsp->buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
2413 				   SRP_BUF_FORMAT_INDIRECT);
2414 	rsp->req_lim_delta = cpu_to_be32(ch->rq_size);
2415 	atomic_set(&ch->req_lim, ch->rq_size);
2416 	atomic_set(&ch->req_lim_delta, 0);
2417 
2418 	/* create cm reply */
2419 	if (ch->using_rdma_cm) {
2420 		rep_param->rdma_cm.private_data = (void *)rsp;
2421 		rep_param->rdma_cm.private_data_len = sizeof(*rsp);
2422 		rep_param->rdma_cm.rnr_retry_count = 7;
2423 		rep_param->rdma_cm.flow_control = 1;
2424 		rep_param->rdma_cm.responder_resources = 4;
2425 		rep_param->rdma_cm.initiator_depth = 4;
2426 	} else {
2427 		rep_param->ib_cm.qp_num = ch->qp->qp_num;
2428 		rep_param->ib_cm.private_data = (void *)rsp;
2429 		rep_param->ib_cm.private_data_len = sizeof(*rsp);
2430 		rep_param->ib_cm.rnr_retry_count = 7;
2431 		rep_param->ib_cm.flow_control = 1;
2432 		rep_param->ib_cm.failover_accepted = 0;
2433 		rep_param->ib_cm.srq = 1;
2434 		rep_param->ib_cm.responder_resources = 4;
2435 		rep_param->ib_cm.initiator_depth = 4;
2436 	}
2437 
2438 	/*
2439 	 * Hold the sport mutex while accepting a connection to avoid that
2440 	 * srpt_disconnect_ch() is invoked concurrently with this code.
2441 	 */
2442 	mutex_lock(&sport->mutex);
2443 	if (sport->enabled && ch->state == CH_CONNECTING) {
2444 		if (ch->using_rdma_cm)
2445 			ret = rdma_accept(rdma_cm_id, &rep_param->rdma_cm);
2446 		else
2447 			ret = ib_send_cm_rep(ib_cm_id, &rep_param->ib_cm);
2448 	} else {
2449 		ret = -EINVAL;
2450 	}
2451 	mutex_unlock(&sport->mutex);
2452 
2453 	switch (ret) {
2454 	case 0:
2455 		break;
2456 	case -EINVAL:
2457 		goto reject;
2458 	default:
2459 		rej->reason = cpu_to_be32(SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2460 		pr_err("sending SRP_LOGIN_REQ response failed (error code = %d)\n",
2461 		       ret);
2462 		goto reject;
2463 	}
2464 
2465 	goto out;
2466 
2467 destroy_ib:
2468 	srpt_destroy_ch_ib(ch);
2469 
2470 free_recv_ring:
2471 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_recv_ring,
2472 			     ch->sport->sdev, ch->rq_size,
2473 			     ch->req_buf_cache, DMA_FROM_DEVICE);
2474 
2475 free_recv_cache:
2476 	kmem_cache_destroy(ch->req_buf_cache);
2477 
2478 free_rsp_ring:
2479 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
2480 			     ch->sport->sdev, ch->rq_size,
2481 			     ch->rsp_buf_cache, DMA_TO_DEVICE);
2482 
2483 free_rsp_cache:
2484 	kmem_cache_destroy(ch->rsp_buf_cache);
2485 
2486 free_ch:
2487 	if (rdma_cm_id)
2488 		rdma_cm_id->context = NULL;
2489 	else
2490 		ib_cm_id->context = NULL;
2491 	kfree(ch);
2492 	ch = NULL;
2493 
2494 	WARN_ON_ONCE(ret == 0);
2495 
2496 reject:
2497 	pr_info("Rejecting login with reason %#x\n", be32_to_cpu(rej->reason));
2498 	rej->opcode = SRP_LOGIN_REJ;
2499 	rej->tag = req->tag;
2500 	rej->buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
2501 				   SRP_BUF_FORMAT_INDIRECT);
2502 
2503 	if (rdma_cm_id)
2504 		rdma_reject(rdma_cm_id, rej, sizeof(*rej),
2505 			    IB_CM_REJ_CONSUMER_DEFINED);
2506 	else
2507 		ib_send_cm_rej(ib_cm_id, IB_CM_REJ_CONSUMER_DEFINED, NULL, 0,
2508 			       rej, sizeof(*rej));
2509 
2510 	if (ch && ch->sess) {
2511 		srpt_close_ch(ch);
2512 		/*
2513 		 * Tell the caller not to free cm_id since
2514 		 * srpt_release_channel_work() will do that.
2515 		 */
2516 		ret = 0;
2517 	}
2518 
2519 out:
2520 	kfree(rep_param);
2521 	kfree(rsp);
2522 	kfree(rej);
2523 
2524 	return ret;
2525 }
2526 
2527 static int srpt_ib_cm_req_recv(struct ib_cm_id *cm_id,
2528 			       const struct ib_cm_req_event_param *param,
2529 			       void *private_data)
2530 {
2531 	char sguid[40];
2532 
2533 	srpt_format_guid(sguid, sizeof(sguid),
2534 			 &param->primary_path->dgid.global.interface_id);
2535 
2536 	return srpt_cm_req_recv(cm_id->context, cm_id, NULL, param->port,
2537 				param->primary_path->pkey,
2538 				private_data, sguid);
2539 }
2540 
2541 static int srpt_rdma_cm_req_recv(struct rdma_cm_id *cm_id,
2542 				 struct rdma_cm_event *event)
2543 {
2544 	struct srpt_device *sdev;
2545 	struct srp_login_req req;
2546 	const struct srp_login_req_rdma *req_rdma;
2547 	struct sa_path_rec *path_rec = cm_id->route.path_rec;
2548 	char src_addr[40];
2549 
2550 	sdev = ib_get_client_data(cm_id->device, &srpt_client);
2551 	if (!sdev)
2552 		return -ECONNREFUSED;
2553 
2554 	if (event->param.conn.private_data_len < sizeof(*req_rdma))
2555 		return -EINVAL;
2556 
2557 	/* Transform srp_login_req_rdma into srp_login_req. */
2558 	req_rdma = event->param.conn.private_data;
2559 	memset(&req, 0, sizeof(req));
2560 	req.opcode		= req_rdma->opcode;
2561 	req.tag			= req_rdma->tag;
2562 	req.req_it_iu_len	= req_rdma->req_it_iu_len;
2563 	req.req_buf_fmt		= req_rdma->req_buf_fmt;
2564 	req.req_flags		= req_rdma->req_flags;
2565 	memcpy(req.initiator_port_id, req_rdma->initiator_port_id, 16);
2566 	memcpy(req.target_port_id, req_rdma->target_port_id, 16);
2567 	req.imm_data_offset	= req_rdma->imm_data_offset;
2568 
2569 	snprintf(src_addr, sizeof(src_addr), "%pIS",
2570 		 &cm_id->route.addr.src_addr);
2571 
2572 	return srpt_cm_req_recv(sdev, NULL, cm_id, cm_id->port_num,
2573 				path_rec ? path_rec->pkey : 0, &req, src_addr);
2574 }
2575 
2576 static void srpt_cm_rej_recv(struct srpt_rdma_ch *ch,
2577 			     enum ib_cm_rej_reason reason,
2578 			     const u8 *private_data,
2579 			     u8 private_data_len)
2580 {
2581 	char *priv = NULL;
2582 	int i;
2583 
2584 	if (private_data_len && (priv = kmalloc(private_data_len * 3 + 1,
2585 						GFP_KERNEL))) {
2586 		for (i = 0; i < private_data_len; i++)
2587 			sprintf(priv + 3 * i, " %02x", private_data[i]);
2588 	}
2589 	pr_info("Received CM REJ for ch %s-%d; reason %d%s%s.\n",
2590 		ch->sess_name, ch->qp->qp_num, reason, private_data_len ?
2591 		"; private data" : "", priv ? priv : " (?)");
2592 	kfree(priv);
2593 }
2594 
2595 /**
2596  * srpt_cm_rtu_recv - process an IB_CM_RTU_RECEIVED or USER_ESTABLISHED event
2597  * @ch: SRPT RDMA channel.
2598  *
2599  * An RTU (ready to use) message indicates that the connection has been
2600  * established and that the recipient may begin transmitting.
2601  */
2602 static void srpt_cm_rtu_recv(struct srpt_rdma_ch *ch)
2603 {
2604 	int ret;
2605 
2606 	ret = ch->using_rdma_cm ? 0 : srpt_ch_qp_rts(ch, ch->qp);
2607 	if (ret < 0) {
2608 		pr_err("%s-%d: QP transition to RTS failed\n", ch->sess_name,
2609 		       ch->qp->qp_num);
2610 		srpt_close_ch(ch);
2611 		return;
2612 	}
2613 
2614 	/*
2615 	 * Note: calling srpt_close_ch() if the transition to the LIVE state
2616 	 * fails is not necessary since that means that that function has
2617 	 * already been invoked from another thread.
2618 	 */
2619 	if (!srpt_set_ch_state(ch, CH_LIVE)) {
2620 		pr_err("%s-%d: channel transition to LIVE state failed\n",
2621 		       ch->sess_name, ch->qp->qp_num);
2622 		return;
2623 	}
2624 
2625 	/* Trigger wait list processing. */
2626 	ret = srpt_zerolength_write(ch);
2627 	WARN_ONCE(ret < 0, "%d\n", ret);
2628 }
2629 
2630 /**
2631  * srpt_cm_handler - IB connection manager callback function
2632  * @cm_id: IB/CM connection identifier.
2633  * @event: IB/CM event.
2634  *
2635  * A non-zero return value will cause the caller destroy the CM ID.
2636  *
2637  * Note: srpt_cm_handler() must only return a non-zero value when transferring
2638  * ownership of the cm_id to a channel by srpt_cm_req_recv() failed. Returning
2639  * a non-zero value in any other case will trigger a race with the
2640  * ib_destroy_cm_id() call in srpt_release_channel().
2641  */
2642 static int srpt_cm_handler(struct ib_cm_id *cm_id,
2643 			   const struct ib_cm_event *event)
2644 {
2645 	struct srpt_rdma_ch *ch = cm_id->context;
2646 	int ret;
2647 
2648 	ret = 0;
2649 	switch (event->event) {
2650 	case IB_CM_REQ_RECEIVED:
2651 		ret = srpt_ib_cm_req_recv(cm_id, &event->param.req_rcvd,
2652 					  event->private_data);
2653 		break;
2654 	case IB_CM_REJ_RECEIVED:
2655 		srpt_cm_rej_recv(ch, event->param.rej_rcvd.reason,
2656 				 event->private_data,
2657 				 IB_CM_REJ_PRIVATE_DATA_SIZE);
2658 		break;
2659 	case IB_CM_RTU_RECEIVED:
2660 	case IB_CM_USER_ESTABLISHED:
2661 		srpt_cm_rtu_recv(ch);
2662 		break;
2663 	case IB_CM_DREQ_RECEIVED:
2664 		srpt_disconnect_ch(ch);
2665 		break;
2666 	case IB_CM_DREP_RECEIVED:
2667 		pr_info("Received CM DREP message for ch %s-%d.\n",
2668 			ch->sess_name, ch->qp->qp_num);
2669 		srpt_close_ch(ch);
2670 		break;
2671 	case IB_CM_TIMEWAIT_EXIT:
2672 		pr_info("Received CM TimeWait exit for ch %s-%d.\n",
2673 			ch->sess_name, ch->qp->qp_num);
2674 		srpt_close_ch(ch);
2675 		break;
2676 	case IB_CM_REP_ERROR:
2677 		pr_info("Received CM REP error for ch %s-%d.\n", ch->sess_name,
2678 			ch->qp->qp_num);
2679 		break;
2680 	case IB_CM_DREQ_ERROR:
2681 		pr_info("Received CM DREQ ERROR event.\n");
2682 		break;
2683 	case IB_CM_MRA_RECEIVED:
2684 		pr_info("Received CM MRA event\n");
2685 		break;
2686 	default:
2687 		pr_err("received unrecognized CM event %d\n", event->event);
2688 		break;
2689 	}
2690 
2691 	return ret;
2692 }
2693 
2694 static int srpt_rdma_cm_handler(struct rdma_cm_id *cm_id,
2695 				struct rdma_cm_event *event)
2696 {
2697 	struct srpt_rdma_ch *ch = cm_id->context;
2698 	int ret = 0;
2699 
2700 	switch (event->event) {
2701 	case RDMA_CM_EVENT_CONNECT_REQUEST:
2702 		ret = srpt_rdma_cm_req_recv(cm_id, event);
2703 		break;
2704 	case RDMA_CM_EVENT_REJECTED:
2705 		srpt_cm_rej_recv(ch, event->status,
2706 				 event->param.conn.private_data,
2707 				 event->param.conn.private_data_len);
2708 		break;
2709 	case RDMA_CM_EVENT_ESTABLISHED:
2710 		srpt_cm_rtu_recv(ch);
2711 		break;
2712 	case RDMA_CM_EVENT_DISCONNECTED:
2713 		if (ch->state < CH_DISCONNECTING)
2714 			srpt_disconnect_ch(ch);
2715 		else
2716 			srpt_close_ch(ch);
2717 		break;
2718 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
2719 		srpt_close_ch(ch);
2720 		break;
2721 	case RDMA_CM_EVENT_UNREACHABLE:
2722 		pr_info("Received CM REP error for ch %s-%d.\n", ch->sess_name,
2723 			ch->qp->qp_num);
2724 		break;
2725 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
2726 	case RDMA_CM_EVENT_ADDR_CHANGE:
2727 		break;
2728 	default:
2729 		pr_err("received unrecognized RDMA CM event %d\n",
2730 		       event->event);
2731 		break;
2732 	}
2733 
2734 	return ret;
2735 }
2736 
2737 /*
2738  * srpt_write_pending - Start data transfer from initiator to target (write).
2739  */
2740 static int srpt_write_pending(struct se_cmd *se_cmd)
2741 {
2742 	struct srpt_send_ioctx *ioctx =
2743 		container_of(se_cmd, struct srpt_send_ioctx, cmd);
2744 	struct srpt_rdma_ch *ch = ioctx->ch;
2745 	struct ib_send_wr *first_wr = NULL;
2746 	struct ib_cqe *cqe = &ioctx->rdma_cqe;
2747 	enum srpt_command_state new_state;
2748 	int ret, i;
2749 
2750 	if (ioctx->recv_ioctx) {
2751 		srpt_set_cmd_state(ioctx, SRPT_STATE_DATA_IN);
2752 		target_execute_cmd(&ioctx->cmd);
2753 		return 0;
2754 	}
2755 
2756 	new_state = srpt_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA);
2757 	WARN_ON(new_state == SRPT_STATE_DONE);
2758 
2759 	if (atomic_sub_return(ioctx->n_rdma, &ch->sq_wr_avail) < 0) {
2760 		pr_warn("%s: IB send queue full (needed %d)\n",
2761 				__func__, ioctx->n_rdma);
2762 		ret = -ENOMEM;
2763 		goto out_undo;
2764 	}
2765 
2766 	cqe->done = srpt_rdma_read_done;
2767 	for (i = ioctx->n_rw_ctx - 1; i >= 0; i--) {
2768 		struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i];
2769 
2770 		first_wr = rdma_rw_ctx_wrs(&ctx->rw, ch->qp, ch->sport->port,
2771 				cqe, first_wr);
2772 		cqe = NULL;
2773 	}
2774 
2775 	ret = ib_post_send(ch->qp, first_wr, NULL);
2776 	if (ret) {
2777 		pr_err("%s: ib_post_send() returned %d for %d (avail: %d)\n",
2778 			 __func__, ret, ioctx->n_rdma,
2779 			 atomic_read(&ch->sq_wr_avail));
2780 		goto out_undo;
2781 	}
2782 
2783 	return 0;
2784 out_undo:
2785 	atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
2786 	return ret;
2787 }
2788 
2789 static u8 tcm_to_srp_tsk_mgmt_status(const int tcm_mgmt_status)
2790 {
2791 	switch (tcm_mgmt_status) {
2792 	case TMR_FUNCTION_COMPLETE:
2793 		return SRP_TSK_MGMT_SUCCESS;
2794 	case TMR_FUNCTION_REJECTED:
2795 		return SRP_TSK_MGMT_FUNC_NOT_SUPP;
2796 	}
2797 	return SRP_TSK_MGMT_FAILED;
2798 }
2799 
2800 /**
2801  * srpt_queue_response - transmit the response to a SCSI command
2802  * @cmd: SCSI target command.
2803  *
2804  * Callback function called by the TCM core. Must not block since it can be
2805  * invoked on the context of the IB completion handler.
2806  */
2807 static void srpt_queue_response(struct se_cmd *cmd)
2808 {
2809 	struct srpt_send_ioctx *ioctx =
2810 		container_of(cmd, struct srpt_send_ioctx, cmd);
2811 	struct srpt_rdma_ch *ch = ioctx->ch;
2812 	struct srpt_device *sdev = ch->sport->sdev;
2813 	struct ib_send_wr send_wr, *first_wr = &send_wr;
2814 	struct ib_sge sge;
2815 	enum srpt_command_state state;
2816 	int resp_len, ret, i;
2817 	u8 srp_tm_status;
2818 
2819 	state = ioctx->state;
2820 	switch (state) {
2821 	case SRPT_STATE_NEW:
2822 	case SRPT_STATE_DATA_IN:
2823 		ioctx->state = SRPT_STATE_CMD_RSP_SENT;
2824 		break;
2825 	case SRPT_STATE_MGMT:
2826 		ioctx->state = SRPT_STATE_MGMT_RSP_SENT;
2827 		break;
2828 	default:
2829 		WARN(true, "ch %p; cmd %d: unexpected command state %d\n",
2830 			ch, ioctx->ioctx.index, ioctx->state);
2831 		break;
2832 	}
2833 
2834 	if (WARN_ON_ONCE(state == SRPT_STATE_CMD_RSP_SENT))
2835 		return;
2836 
2837 	/* For read commands, transfer the data to the initiator. */
2838 	if (ioctx->cmd.data_direction == DMA_FROM_DEVICE &&
2839 	    ioctx->cmd.data_length &&
2840 	    !ioctx->queue_status_only) {
2841 		for (i = ioctx->n_rw_ctx - 1; i >= 0; i--) {
2842 			struct srpt_rw_ctx *ctx = &ioctx->rw_ctxs[i];
2843 
2844 			first_wr = rdma_rw_ctx_wrs(&ctx->rw, ch->qp,
2845 					ch->sport->port, NULL, first_wr);
2846 		}
2847 	}
2848 
2849 	if (state != SRPT_STATE_MGMT)
2850 		resp_len = srpt_build_cmd_rsp(ch, ioctx, ioctx->cmd.tag,
2851 					      cmd->scsi_status);
2852 	else {
2853 		srp_tm_status
2854 			= tcm_to_srp_tsk_mgmt_status(cmd->se_tmr_req->response);
2855 		resp_len = srpt_build_tskmgmt_rsp(ch, ioctx, srp_tm_status,
2856 						 ioctx->cmd.tag);
2857 	}
2858 
2859 	atomic_inc(&ch->req_lim);
2860 
2861 	if (unlikely(atomic_sub_return(1 + ioctx->n_rdma,
2862 			&ch->sq_wr_avail) < 0)) {
2863 		pr_warn("%s: IB send queue full (needed %d)\n",
2864 				__func__, ioctx->n_rdma);
2865 		goto out;
2866 	}
2867 
2868 	ib_dma_sync_single_for_device(sdev->device, ioctx->ioctx.dma, resp_len,
2869 				      DMA_TO_DEVICE);
2870 
2871 	sge.addr = ioctx->ioctx.dma;
2872 	sge.length = resp_len;
2873 	sge.lkey = sdev->lkey;
2874 
2875 	ioctx->ioctx.cqe.done = srpt_send_done;
2876 	send_wr.next = NULL;
2877 	send_wr.wr_cqe = &ioctx->ioctx.cqe;
2878 	send_wr.sg_list = &sge;
2879 	send_wr.num_sge = 1;
2880 	send_wr.opcode = IB_WR_SEND;
2881 	send_wr.send_flags = IB_SEND_SIGNALED;
2882 
2883 	ret = ib_post_send(ch->qp, first_wr, NULL);
2884 	if (ret < 0) {
2885 		pr_err("%s: sending cmd response failed for tag %llu (%d)\n",
2886 			__func__, ioctx->cmd.tag, ret);
2887 		goto out;
2888 	}
2889 
2890 	return;
2891 
2892 out:
2893 	atomic_add(1 + ioctx->n_rdma, &ch->sq_wr_avail);
2894 	atomic_dec(&ch->req_lim);
2895 	srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
2896 	target_put_sess_cmd(&ioctx->cmd);
2897 }
2898 
2899 static int srpt_queue_data_in(struct se_cmd *cmd)
2900 {
2901 	srpt_queue_response(cmd);
2902 	return 0;
2903 }
2904 
2905 static void srpt_queue_tm_rsp(struct se_cmd *cmd)
2906 {
2907 	srpt_queue_response(cmd);
2908 }
2909 
2910 /*
2911  * This function is called for aborted commands if no response is sent to the
2912  * initiator. Make sure that the credits freed by aborting a command are
2913  * returned to the initiator the next time a response is sent by incrementing
2914  * ch->req_lim_delta.
2915  */
2916 static void srpt_aborted_task(struct se_cmd *cmd)
2917 {
2918 	struct srpt_send_ioctx *ioctx = container_of(cmd,
2919 				struct srpt_send_ioctx, cmd);
2920 	struct srpt_rdma_ch *ch = ioctx->ch;
2921 
2922 	atomic_inc(&ch->req_lim_delta);
2923 }
2924 
2925 static int srpt_queue_status(struct se_cmd *cmd)
2926 {
2927 	struct srpt_send_ioctx *ioctx;
2928 
2929 	ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
2930 	BUG_ON(ioctx->sense_data != cmd->sense_buffer);
2931 	if (cmd->se_cmd_flags &
2932 	    (SCF_TRANSPORT_TASK_SENSE | SCF_EMULATED_TASK_SENSE))
2933 		WARN_ON(cmd->scsi_status != SAM_STAT_CHECK_CONDITION);
2934 	ioctx->queue_status_only = true;
2935 	srpt_queue_response(cmd);
2936 	return 0;
2937 }
2938 
2939 static void srpt_refresh_port_work(struct work_struct *work)
2940 {
2941 	struct srpt_port *sport = container_of(work, struct srpt_port, work);
2942 
2943 	srpt_refresh_port(sport);
2944 }
2945 
2946 /**
2947  * srpt_release_sport - disable login and wait for associated channels
2948  * @sport: SRPT HCA port.
2949  */
2950 static int srpt_release_sport(struct srpt_port *sport)
2951 {
2952 	DECLARE_COMPLETION_ONSTACK(c);
2953 	struct srpt_nexus *nexus, *next_n;
2954 	struct srpt_rdma_ch *ch;
2955 
2956 	WARN_ON_ONCE(irqs_disabled());
2957 
2958 	sport->freed_channels = &c;
2959 
2960 	mutex_lock(&sport->mutex);
2961 	srpt_set_enabled(sport, false);
2962 	mutex_unlock(&sport->mutex);
2963 
2964 	while (atomic_read(&sport->refcount) > 0 &&
2965 	       wait_for_completion_timeout(&c, 5 * HZ) <= 0) {
2966 		pr_info("%s_%d: waiting for unregistration of %d sessions ...\n",
2967 			dev_name(&sport->sdev->device->dev), sport->port,
2968 			atomic_read(&sport->refcount));
2969 		rcu_read_lock();
2970 		list_for_each_entry(nexus, &sport->nexus_list, entry) {
2971 			list_for_each_entry(ch, &nexus->ch_list, list) {
2972 				pr_info("%s-%d: state %s\n",
2973 					ch->sess_name, ch->qp->qp_num,
2974 					get_ch_state_name(ch->state));
2975 			}
2976 		}
2977 		rcu_read_unlock();
2978 	}
2979 
2980 	mutex_lock(&sport->mutex);
2981 	list_for_each_entry_safe(nexus, next_n, &sport->nexus_list, entry) {
2982 		list_del(&nexus->entry);
2983 		kfree_rcu(nexus, rcu);
2984 	}
2985 	mutex_unlock(&sport->mutex);
2986 
2987 	return 0;
2988 }
2989 
2990 struct port_and_port_id {
2991 	struct srpt_port *sport;
2992 	struct srpt_port_id **port_id;
2993 };
2994 
2995 static struct port_and_port_id __srpt_lookup_port(const char *name)
2996 {
2997 	struct ib_device *dev;
2998 	struct srpt_device *sdev;
2999 	struct srpt_port *sport;
3000 	int i;
3001 
3002 	list_for_each_entry(sdev, &srpt_dev_list, list) {
3003 		dev = sdev->device;
3004 		if (!dev)
3005 			continue;
3006 
3007 		for (i = 0; i < dev->phys_port_cnt; i++) {
3008 			sport = &sdev->port[i];
3009 
3010 			if (strcmp(sport->guid_name, name) == 0) {
3011 				kref_get(&sdev->refcnt);
3012 				return (struct port_and_port_id){
3013 					sport, &sport->guid_id};
3014 			}
3015 			if (strcmp(sport->gid_name, name) == 0) {
3016 				kref_get(&sdev->refcnt);
3017 				return (struct port_and_port_id){
3018 					sport, &sport->gid_id};
3019 			}
3020 		}
3021 	}
3022 
3023 	return (struct port_and_port_id){};
3024 }
3025 
3026 /**
3027  * srpt_lookup_port() - Look up an RDMA port by name
3028  * @name: ASCII port name
3029  *
3030  * Increments the RDMA port reference count if an RDMA port pointer is returned.
3031  * The caller must drop that reference count by calling srpt_port_put_ref().
3032  */
3033 static struct port_and_port_id srpt_lookup_port(const char *name)
3034 {
3035 	struct port_and_port_id papi;
3036 
3037 	spin_lock(&srpt_dev_lock);
3038 	papi = __srpt_lookup_port(name);
3039 	spin_unlock(&srpt_dev_lock);
3040 
3041 	return papi;
3042 }
3043 
3044 static void srpt_free_srq(struct srpt_device *sdev)
3045 {
3046 	if (!sdev->srq)
3047 		return;
3048 
3049 	ib_destroy_srq(sdev->srq);
3050 	srpt_free_ioctx_ring((struct srpt_ioctx **)sdev->ioctx_ring, sdev,
3051 			     sdev->srq_size, sdev->req_buf_cache,
3052 			     DMA_FROM_DEVICE);
3053 	kmem_cache_destroy(sdev->req_buf_cache);
3054 	sdev->srq = NULL;
3055 }
3056 
3057 static int srpt_alloc_srq(struct srpt_device *sdev)
3058 {
3059 	struct ib_srq_init_attr srq_attr = {
3060 		.event_handler = srpt_srq_event,
3061 		.srq_context = (void *)sdev,
3062 		.attr.max_wr = sdev->srq_size,
3063 		.attr.max_sge = 1,
3064 		.srq_type = IB_SRQT_BASIC,
3065 	};
3066 	struct ib_device *device = sdev->device;
3067 	struct ib_srq *srq;
3068 	int i;
3069 
3070 	WARN_ON_ONCE(sdev->srq);
3071 	srq = ib_create_srq(sdev->pd, &srq_attr);
3072 	if (IS_ERR(srq)) {
3073 		pr_debug("ib_create_srq() failed: %ld\n", PTR_ERR(srq));
3074 		return PTR_ERR(srq);
3075 	}
3076 
3077 	pr_debug("create SRQ #wr= %d max_allow=%d dev= %s\n", sdev->srq_size,
3078 		 sdev->device->attrs.max_srq_wr, dev_name(&device->dev));
3079 
3080 	sdev->req_buf_cache = kmem_cache_create("srpt-srq-req-buf",
3081 						srp_max_req_size, 0, 0, NULL);
3082 	if (!sdev->req_buf_cache)
3083 		goto free_srq;
3084 
3085 	sdev->ioctx_ring = (struct srpt_recv_ioctx **)
3086 		srpt_alloc_ioctx_ring(sdev, sdev->srq_size,
3087 				      sizeof(*sdev->ioctx_ring[0]),
3088 				      sdev->req_buf_cache, 0, DMA_FROM_DEVICE);
3089 	if (!sdev->ioctx_ring)
3090 		goto free_cache;
3091 
3092 	sdev->use_srq = true;
3093 	sdev->srq = srq;
3094 
3095 	for (i = 0; i < sdev->srq_size; ++i) {
3096 		INIT_LIST_HEAD(&sdev->ioctx_ring[i]->wait_list);
3097 		srpt_post_recv(sdev, NULL, sdev->ioctx_ring[i]);
3098 	}
3099 
3100 	return 0;
3101 
3102 free_cache:
3103 	kmem_cache_destroy(sdev->req_buf_cache);
3104 
3105 free_srq:
3106 	ib_destroy_srq(srq);
3107 	return -ENOMEM;
3108 }
3109 
3110 static int srpt_use_srq(struct srpt_device *sdev, bool use_srq)
3111 {
3112 	struct ib_device *device = sdev->device;
3113 	int ret = 0;
3114 
3115 	if (!use_srq) {
3116 		srpt_free_srq(sdev);
3117 		sdev->use_srq = false;
3118 	} else if (use_srq && !sdev->srq) {
3119 		ret = srpt_alloc_srq(sdev);
3120 	}
3121 	pr_debug("%s(%s): use_srq = %d; ret = %d\n", __func__,
3122 		 dev_name(&device->dev), sdev->use_srq, ret);
3123 	return ret;
3124 }
3125 
3126 static void srpt_free_sdev(struct kref *refcnt)
3127 {
3128 	struct srpt_device *sdev = container_of(refcnt, typeof(*sdev), refcnt);
3129 
3130 	kfree(sdev);
3131 }
3132 
3133 static void srpt_sdev_put(struct srpt_device *sdev)
3134 {
3135 	kref_put(&sdev->refcnt, srpt_free_sdev);
3136 }
3137 
3138 /**
3139  * srpt_add_one - InfiniBand device addition callback function
3140  * @device: Describes a HCA.
3141  */
3142 static int srpt_add_one(struct ib_device *device)
3143 {
3144 	struct srpt_device *sdev;
3145 	struct srpt_port *sport;
3146 	int ret;
3147 	u32 i;
3148 
3149 	pr_debug("device = %p\n", device);
3150 
3151 	sdev = kzalloc(struct_size(sdev, port, device->phys_port_cnt),
3152 		       GFP_KERNEL);
3153 	if (!sdev)
3154 		return -ENOMEM;
3155 
3156 	kref_init(&sdev->refcnt);
3157 	sdev->device = device;
3158 	mutex_init(&sdev->sdev_mutex);
3159 
3160 	sdev->pd = ib_alloc_pd(device, 0);
3161 	if (IS_ERR(sdev->pd)) {
3162 		ret = PTR_ERR(sdev->pd);
3163 		goto free_dev;
3164 	}
3165 
3166 	sdev->lkey = sdev->pd->local_dma_lkey;
3167 
3168 	sdev->srq_size = min(srpt_srq_size, sdev->device->attrs.max_srq_wr);
3169 
3170 	srpt_use_srq(sdev, sdev->port[0].port_attrib.use_srq);
3171 
3172 	if (!srpt_service_guid)
3173 		srpt_service_guid = be64_to_cpu(device->node_guid);
3174 
3175 	if (rdma_port_get_link_layer(device, 1) == IB_LINK_LAYER_INFINIBAND)
3176 		sdev->cm_id = ib_create_cm_id(device, srpt_cm_handler, sdev);
3177 	if (IS_ERR(sdev->cm_id)) {
3178 		pr_info("ib_create_cm_id() failed: %ld\n",
3179 			PTR_ERR(sdev->cm_id));
3180 		ret = PTR_ERR(sdev->cm_id);
3181 		sdev->cm_id = NULL;
3182 		if (!rdma_cm_id)
3183 			goto err_ring;
3184 	}
3185 
3186 	/* print out target login information */
3187 	pr_debug("Target login info: id_ext=%016llx,ioc_guid=%016llx,pkey=ffff,service_id=%016llx\n",
3188 		 srpt_service_guid, srpt_service_guid, srpt_service_guid);
3189 
3190 	/*
3191 	 * We do not have a consistent service_id (ie. also id_ext of target_id)
3192 	 * to identify this target. We currently use the guid of the first HCA
3193 	 * in the system as service_id; therefore, the target_id will change
3194 	 * if this HCA is gone bad and replaced by different HCA
3195 	 */
3196 	ret = sdev->cm_id ?
3197 		ib_cm_listen(sdev->cm_id, cpu_to_be64(srpt_service_guid)) :
3198 		0;
3199 	if (ret < 0) {
3200 		pr_err("ib_cm_listen() failed: %d (cm_id state = %d)\n", ret,
3201 		       sdev->cm_id->state);
3202 		goto err_cm;
3203 	}
3204 
3205 	INIT_IB_EVENT_HANDLER(&sdev->event_handler, sdev->device,
3206 			      srpt_event_handler);
3207 	ib_register_event_handler(&sdev->event_handler);
3208 
3209 	for (i = 1; i <= sdev->device->phys_port_cnt; i++) {
3210 		sport = &sdev->port[i - 1];
3211 		INIT_LIST_HEAD(&sport->nexus_list);
3212 		mutex_init(&sport->mutex);
3213 		sport->sdev = sdev;
3214 		sport->port = i;
3215 		sport->port_attrib.srp_max_rdma_size = DEFAULT_MAX_RDMA_SIZE;
3216 		sport->port_attrib.srp_max_rsp_size = DEFAULT_MAX_RSP_SIZE;
3217 		sport->port_attrib.srp_sq_size = DEF_SRPT_SQ_SIZE;
3218 		sport->port_attrib.use_srq = false;
3219 		INIT_WORK(&sport->work, srpt_refresh_port_work);
3220 
3221 		ret = srpt_refresh_port(sport);
3222 		if (ret) {
3223 			pr_err("MAD registration failed for %s-%d.\n",
3224 			       dev_name(&sdev->device->dev), i);
3225 			i--;
3226 			goto err_port;
3227 		}
3228 	}
3229 
3230 	spin_lock(&srpt_dev_lock);
3231 	list_add_tail(&sdev->list, &srpt_dev_list);
3232 	spin_unlock(&srpt_dev_lock);
3233 
3234 	ib_set_client_data(device, &srpt_client, sdev);
3235 	pr_debug("added %s.\n", dev_name(&device->dev));
3236 	return 0;
3237 
3238 err_port:
3239 	srpt_unregister_mad_agent(sdev, i);
3240 	ib_unregister_event_handler(&sdev->event_handler);
3241 err_cm:
3242 	if (sdev->cm_id)
3243 		ib_destroy_cm_id(sdev->cm_id);
3244 err_ring:
3245 	srpt_free_srq(sdev);
3246 	ib_dealloc_pd(sdev->pd);
3247 free_dev:
3248 	srpt_sdev_put(sdev);
3249 	pr_info("%s(%s) failed.\n", __func__, dev_name(&device->dev));
3250 	return ret;
3251 }
3252 
3253 /**
3254  * srpt_remove_one - InfiniBand device removal callback function
3255  * @device: Describes a HCA.
3256  * @client_data: The value passed as the third argument to ib_set_client_data().
3257  */
3258 static void srpt_remove_one(struct ib_device *device, void *client_data)
3259 {
3260 	struct srpt_device *sdev = client_data;
3261 	int i;
3262 
3263 	srpt_unregister_mad_agent(sdev, sdev->device->phys_port_cnt);
3264 
3265 	ib_unregister_event_handler(&sdev->event_handler);
3266 
3267 	/* Cancel any work queued by the just unregistered IB event handler. */
3268 	for (i = 0; i < sdev->device->phys_port_cnt; i++)
3269 		cancel_work_sync(&sdev->port[i].work);
3270 
3271 	if (sdev->cm_id)
3272 		ib_destroy_cm_id(sdev->cm_id);
3273 
3274 	ib_set_client_data(device, &srpt_client, NULL);
3275 
3276 	/*
3277 	 * Unregistering a target must happen after destroying sdev->cm_id
3278 	 * such that no new SRP_LOGIN_REQ information units can arrive while
3279 	 * destroying the target.
3280 	 */
3281 	spin_lock(&srpt_dev_lock);
3282 	list_del(&sdev->list);
3283 	spin_unlock(&srpt_dev_lock);
3284 
3285 	for (i = 0; i < sdev->device->phys_port_cnt; i++)
3286 		srpt_release_sport(&sdev->port[i]);
3287 
3288 	srpt_free_srq(sdev);
3289 
3290 	ib_dealloc_pd(sdev->pd);
3291 
3292 	srpt_sdev_put(sdev);
3293 }
3294 
3295 static struct ib_client srpt_client = {
3296 	.name = DRV_NAME,
3297 	.add = srpt_add_one,
3298 	.remove = srpt_remove_one
3299 };
3300 
3301 static int srpt_check_true(struct se_portal_group *se_tpg)
3302 {
3303 	return 1;
3304 }
3305 
3306 static struct srpt_port *srpt_tpg_to_sport(struct se_portal_group *tpg)
3307 {
3308 	return tpg->se_tpg_wwn->priv;
3309 }
3310 
3311 static struct srpt_port_id *srpt_wwn_to_sport_id(struct se_wwn *wwn)
3312 {
3313 	struct srpt_port *sport = wwn->priv;
3314 
3315 	if (sport->guid_id && &sport->guid_id->wwn == wwn)
3316 		return sport->guid_id;
3317 	if (sport->gid_id && &sport->gid_id->wwn == wwn)
3318 		return sport->gid_id;
3319 	WARN_ON_ONCE(true);
3320 	return NULL;
3321 }
3322 
3323 static char *srpt_get_fabric_wwn(struct se_portal_group *tpg)
3324 {
3325 	struct srpt_tpg *stpg = container_of(tpg, typeof(*stpg), tpg);
3326 
3327 	return stpg->sport_id->name;
3328 }
3329 
3330 static u16 srpt_get_tag(struct se_portal_group *tpg)
3331 {
3332 	return 1;
3333 }
3334 
3335 static void srpt_release_cmd(struct se_cmd *se_cmd)
3336 {
3337 	struct srpt_send_ioctx *ioctx = container_of(se_cmd,
3338 				struct srpt_send_ioctx, cmd);
3339 	struct srpt_rdma_ch *ch = ioctx->ch;
3340 	struct srpt_recv_ioctx *recv_ioctx = ioctx->recv_ioctx;
3341 
3342 	WARN_ON_ONCE(ioctx->state != SRPT_STATE_DONE &&
3343 		     !(ioctx->cmd.transport_state & CMD_T_ABORTED));
3344 
3345 	if (recv_ioctx) {
3346 		WARN_ON_ONCE(!list_empty(&recv_ioctx->wait_list));
3347 		ioctx->recv_ioctx = NULL;
3348 		srpt_post_recv(ch->sport->sdev, ch, recv_ioctx);
3349 	}
3350 
3351 	if (ioctx->n_rw_ctx) {
3352 		srpt_free_rw_ctxs(ch, ioctx);
3353 		ioctx->n_rw_ctx = 0;
3354 	}
3355 
3356 	target_free_tag(se_cmd->se_sess, se_cmd);
3357 }
3358 
3359 /**
3360  * srpt_close_session - forcibly close a session
3361  * @se_sess: SCSI target session.
3362  *
3363  * Callback function invoked by the TCM core to clean up sessions associated
3364  * with a node ACL when the user invokes
3365  * rmdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3366  */
3367 static void srpt_close_session(struct se_session *se_sess)
3368 {
3369 	struct srpt_rdma_ch *ch = se_sess->fabric_sess_ptr;
3370 
3371 	srpt_disconnect_ch_sync(ch);
3372 }
3373 
3374 /* Note: only used from inside debug printk's by the TCM core. */
3375 static int srpt_get_tcm_cmd_state(struct se_cmd *se_cmd)
3376 {
3377 	struct srpt_send_ioctx *ioctx;
3378 
3379 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
3380 	return ioctx->state;
3381 }
3382 
3383 static int srpt_parse_guid(u64 *guid, const char *name)
3384 {
3385 	u16 w[4];
3386 	int ret = -EINVAL;
3387 
3388 	if (sscanf(name, "%hx:%hx:%hx:%hx", &w[0], &w[1], &w[2], &w[3]) != 4)
3389 		goto out;
3390 	*guid = get_unaligned_be64(w);
3391 	ret = 0;
3392 out:
3393 	return ret;
3394 }
3395 
3396 /**
3397  * srpt_parse_i_port_id - parse an initiator port ID
3398  * @name: ASCII representation of a 128-bit initiator port ID.
3399  * @i_port_id: Binary 128-bit port ID.
3400  */
3401 static int srpt_parse_i_port_id(u8 i_port_id[16], const char *name)
3402 {
3403 	const char *p;
3404 	unsigned len, count, leading_zero_bytes;
3405 	int ret;
3406 
3407 	p = name;
3408 	if (strncasecmp(p, "0x", 2) == 0)
3409 		p += 2;
3410 	ret = -EINVAL;
3411 	len = strlen(p);
3412 	if (len % 2)
3413 		goto out;
3414 	count = min(len / 2, 16U);
3415 	leading_zero_bytes = 16 - count;
3416 	memset(i_port_id, 0, leading_zero_bytes);
3417 	ret = hex2bin(i_port_id + leading_zero_bytes, p, count);
3418 
3419 out:
3420 	return ret;
3421 }
3422 
3423 /*
3424  * configfs callback function invoked for mkdir
3425  * /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3426  *
3427  * i_port_id must be an initiator port GUID, GID or IP address. See also the
3428  * target_alloc_session() calls in this driver. Examples of valid initiator
3429  * port IDs:
3430  * 0x0000000000000000505400fffe4a0b7b
3431  * 0000000000000000505400fffe4a0b7b
3432  * 5054:00ff:fe4a:0b7b
3433  * 192.168.122.76
3434  */
3435 static int srpt_init_nodeacl(struct se_node_acl *se_nacl, const char *name)
3436 {
3437 	struct sockaddr_storage sa;
3438 	u64 guid;
3439 	u8 i_port_id[16];
3440 	int ret;
3441 
3442 	ret = srpt_parse_guid(&guid, name);
3443 	if (ret < 0)
3444 		ret = srpt_parse_i_port_id(i_port_id, name);
3445 	if (ret < 0)
3446 		ret = inet_pton_with_scope(&init_net, AF_UNSPEC, name, NULL,
3447 					   &sa);
3448 	if (ret < 0)
3449 		pr_err("invalid initiator port ID %s\n", name);
3450 	return ret;
3451 }
3452 
3453 static ssize_t srpt_tpg_attrib_srp_max_rdma_size_show(struct config_item *item,
3454 		char *page)
3455 {
3456 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3457 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3458 
3459 	return sysfs_emit(page, "%u\n", sport->port_attrib.srp_max_rdma_size);
3460 }
3461 
3462 static ssize_t srpt_tpg_attrib_srp_max_rdma_size_store(struct config_item *item,
3463 		const char *page, size_t count)
3464 {
3465 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3466 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3467 	unsigned long val;
3468 	int ret;
3469 
3470 	ret = kstrtoul(page, 0, &val);
3471 	if (ret < 0) {
3472 		pr_err("kstrtoul() failed with ret: %d\n", ret);
3473 		return -EINVAL;
3474 	}
3475 	if (val > MAX_SRPT_RDMA_SIZE) {
3476 		pr_err("val: %lu exceeds MAX_SRPT_RDMA_SIZE: %d\n", val,
3477 			MAX_SRPT_RDMA_SIZE);
3478 		return -EINVAL;
3479 	}
3480 	if (val < DEFAULT_MAX_RDMA_SIZE) {
3481 		pr_err("val: %lu smaller than DEFAULT_MAX_RDMA_SIZE: %d\n",
3482 			val, DEFAULT_MAX_RDMA_SIZE);
3483 		return -EINVAL;
3484 	}
3485 	sport->port_attrib.srp_max_rdma_size = val;
3486 
3487 	return count;
3488 }
3489 
3490 static ssize_t srpt_tpg_attrib_srp_max_rsp_size_show(struct config_item *item,
3491 		char *page)
3492 {
3493 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3494 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3495 
3496 	return sysfs_emit(page, "%u\n", sport->port_attrib.srp_max_rsp_size);
3497 }
3498 
3499 static ssize_t srpt_tpg_attrib_srp_max_rsp_size_store(struct config_item *item,
3500 		const char *page, size_t count)
3501 {
3502 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3503 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3504 	unsigned long val;
3505 	int ret;
3506 
3507 	ret = kstrtoul(page, 0, &val);
3508 	if (ret < 0) {
3509 		pr_err("kstrtoul() failed with ret: %d\n", ret);
3510 		return -EINVAL;
3511 	}
3512 	if (val > MAX_SRPT_RSP_SIZE) {
3513 		pr_err("val: %lu exceeds MAX_SRPT_RSP_SIZE: %d\n", val,
3514 			MAX_SRPT_RSP_SIZE);
3515 		return -EINVAL;
3516 	}
3517 	if (val < MIN_MAX_RSP_SIZE) {
3518 		pr_err("val: %lu smaller than MIN_MAX_RSP_SIZE: %d\n", val,
3519 			MIN_MAX_RSP_SIZE);
3520 		return -EINVAL;
3521 	}
3522 	sport->port_attrib.srp_max_rsp_size = val;
3523 
3524 	return count;
3525 }
3526 
3527 static ssize_t srpt_tpg_attrib_srp_sq_size_show(struct config_item *item,
3528 		char *page)
3529 {
3530 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3531 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3532 
3533 	return sysfs_emit(page, "%u\n", sport->port_attrib.srp_sq_size);
3534 }
3535 
3536 static ssize_t srpt_tpg_attrib_srp_sq_size_store(struct config_item *item,
3537 		const char *page, size_t count)
3538 {
3539 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3540 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3541 	unsigned long val;
3542 	int ret;
3543 
3544 	ret = kstrtoul(page, 0, &val);
3545 	if (ret < 0) {
3546 		pr_err("kstrtoul() failed with ret: %d\n", ret);
3547 		return -EINVAL;
3548 	}
3549 	if (val > MAX_SRPT_SRQ_SIZE) {
3550 		pr_err("val: %lu exceeds MAX_SRPT_SRQ_SIZE: %d\n", val,
3551 			MAX_SRPT_SRQ_SIZE);
3552 		return -EINVAL;
3553 	}
3554 	if (val < MIN_SRPT_SRQ_SIZE) {
3555 		pr_err("val: %lu smaller than MIN_SRPT_SRQ_SIZE: %d\n", val,
3556 			MIN_SRPT_SRQ_SIZE);
3557 		return -EINVAL;
3558 	}
3559 	sport->port_attrib.srp_sq_size = val;
3560 
3561 	return count;
3562 }
3563 
3564 static ssize_t srpt_tpg_attrib_use_srq_show(struct config_item *item,
3565 					    char *page)
3566 {
3567 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3568 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3569 
3570 	return sysfs_emit(page, "%d\n", sport->port_attrib.use_srq);
3571 }
3572 
3573 static ssize_t srpt_tpg_attrib_use_srq_store(struct config_item *item,
3574 					     const char *page, size_t count)
3575 {
3576 	struct se_portal_group *se_tpg = attrib_to_tpg(item);
3577 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3578 	struct srpt_device *sdev = sport->sdev;
3579 	unsigned long val;
3580 	bool enabled;
3581 	int ret;
3582 
3583 	ret = kstrtoul(page, 0, &val);
3584 	if (ret < 0)
3585 		return ret;
3586 	if (val != !!val)
3587 		return -EINVAL;
3588 
3589 	ret = mutex_lock_interruptible(&sdev->sdev_mutex);
3590 	if (ret < 0)
3591 		return ret;
3592 	ret = mutex_lock_interruptible(&sport->mutex);
3593 	if (ret < 0)
3594 		goto unlock_sdev;
3595 	enabled = sport->enabled;
3596 	/* Log out all initiator systems before changing 'use_srq'. */
3597 	srpt_set_enabled(sport, false);
3598 	sport->port_attrib.use_srq = val;
3599 	srpt_use_srq(sdev, sport->port_attrib.use_srq);
3600 	srpt_set_enabled(sport, enabled);
3601 	ret = count;
3602 	mutex_unlock(&sport->mutex);
3603 unlock_sdev:
3604 	mutex_unlock(&sdev->sdev_mutex);
3605 
3606 	return ret;
3607 }
3608 
3609 CONFIGFS_ATTR(srpt_tpg_attrib_,  srp_max_rdma_size);
3610 CONFIGFS_ATTR(srpt_tpg_attrib_,  srp_max_rsp_size);
3611 CONFIGFS_ATTR(srpt_tpg_attrib_,  srp_sq_size);
3612 CONFIGFS_ATTR(srpt_tpg_attrib_,  use_srq);
3613 
3614 static struct configfs_attribute *srpt_tpg_attrib_attrs[] = {
3615 	&srpt_tpg_attrib_attr_srp_max_rdma_size,
3616 	&srpt_tpg_attrib_attr_srp_max_rsp_size,
3617 	&srpt_tpg_attrib_attr_srp_sq_size,
3618 	&srpt_tpg_attrib_attr_use_srq,
3619 	NULL,
3620 };
3621 
3622 static struct rdma_cm_id *srpt_create_rdma_id(struct sockaddr *listen_addr)
3623 {
3624 	struct rdma_cm_id *rdma_cm_id;
3625 	int ret;
3626 
3627 	rdma_cm_id = rdma_create_id(&init_net, srpt_rdma_cm_handler,
3628 				    NULL, RDMA_PS_TCP, IB_QPT_RC);
3629 	if (IS_ERR(rdma_cm_id)) {
3630 		pr_err("RDMA/CM ID creation failed: %ld\n",
3631 		       PTR_ERR(rdma_cm_id));
3632 		goto out;
3633 	}
3634 
3635 	ret = rdma_bind_addr(rdma_cm_id, listen_addr);
3636 	if (ret) {
3637 		char addr_str[64];
3638 
3639 		snprintf(addr_str, sizeof(addr_str), "%pISp", listen_addr);
3640 		pr_err("Binding RDMA/CM ID to address %s failed: %d\n",
3641 		       addr_str, ret);
3642 		rdma_destroy_id(rdma_cm_id);
3643 		rdma_cm_id = ERR_PTR(ret);
3644 		goto out;
3645 	}
3646 
3647 	ret = rdma_listen(rdma_cm_id, 128);
3648 	if (ret) {
3649 		pr_err("rdma_listen() failed: %d\n", ret);
3650 		rdma_destroy_id(rdma_cm_id);
3651 		rdma_cm_id = ERR_PTR(ret);
3652 	}
3653 
3654 out:
3655 	return rdma_cm_id;
3656 }
3657 
3658 static ssize_t srpt_rdma_cm_port_show(struct config_item *item, char *page)
3659 {
3660 	return sysfs_emit(page, "%d\n", rdma_cm_port);
3661 }
3662 
3663 static ssize_t srpt_rdma_cm_port_store(struct config_item *item,
3664 				       const char *page, size_t count)
3665 {
3666 	struct sockaddr_in  addr4 = { .sin_family  = AF_INET  };
3667 	struct sockaddr_in6 addr6 = { .sin6_family = AF_INET6 };
3668 	struct rdma_cm_id *new_id = NULL;
3669 	u16 val;
3670 	int ret;
3671 
3672 	ret = kstrtou16(page, 0, &val);
3673 	if (ret < 0)
3674 		return ret;
3675 	ret = count;
3676 	if (rdma_cm_port == val)
3677 		goto out;
3678 
3679 	if (val) {
3680 		addr6.sin6_port = cpu_to_be16(val);
3681 		new_id = srpt_create_rdma_id((struct sockaddr *)&addr6);
3682 		if (IS_ERR(new_id)) {
3683 			addr4.sin_port = cpu_to_be16(val);
3684 			new_id = srpt_create_rdma_id((struct sockaddr *)&addr4);
3685 			if (IS_ERR(new_id)) {
3686 				ret = PTR_ERR(new_id);
3687 				goto out;
3688 			}
3689 		}
3690 	}
3691 
3692 	mutex_lock(&rdma_cm_mutex);
3693 	rdma_cm_port = val;
3694 	swap(rdma_cm_id, new_id);
3695 	mutex_unlock(&rdma_cm_mutex);
3696 
3697 	if (new_id)
3698 		rdma_destroy_id(new_id);
3699 	ret = count;
3700 out:
3701 	return ret;
3702 }
3703 
3704 CONFIGFS_ATTR(srpt_, rdma_cm_port);
3705 
3706 static struct configfs_attribute *srpt_da_attrs[] = {
3707 	&srpt_attr_rdma_cm_port,
3708 	NULL,
3709 };
3710 
3711 static int srpt_enable_tpg(struct se_portal_group *se_tpg, bool enable)
3712 {
3713 	struct srpt_port *sport = srpt_tpg_to_sport(se_tpg);
3714 
3715 	mutex_lock(&sport->mutex);
3716 	srpt_set_enabled(sport, enable);
3717 	mutex_unlock(&sport->mutex);
3718 
3719 	return 0;
3720 }
3721 
3722 /**
3723  * srpt_make_tpg - configfs callback invoked for mkdir /sys/kernel/config/target/$driver/$port/$tpg
3724  * @wwn: Corresponds to $driver/$port.
3725  * @name: $tpg.
3726  */
3727 static struct se_portal_group *srpt_make_tpg(struct se_wwn *wwn,
3728 					     const char *name)
3729 {
3730 	struct srpt_port_id *sport_id = srpt_wwn_to_sport_id(wwn);
3731 	struct srpt_tpg *stpg;
3732 	int res = -ENOMEM;
3733 
3734 	stpg = kzalloc(sizeof(*stpg), GFP_KERNEL);
3735 	if (!stpg)
3736 		return ERR_PTR(res);
3737 	stpg->sport_id = sport_id;
3738 	res = core_tpg_register(wwn, &stpg->tpg, SCSI_PROTOCOL_SRP);
3739 	if (res) {
3740 		kfree(stpg);
3741 		return ERR_PTR(res);
3742 	}
3743 
3744 	mutex_lock(&sport_id->mutex);
3745 	list_add_tail(&stpg->entry, &sport_id->tpg_list);
3746 	mutex_unlock(&sport_id->mutex);
3747 
3748 	return &stpg->tpg;
3749 }
3750 
3751 /**
3752  * srpt_drop_tpg - configfs callback invoked for rmdir /sys/kernel/config/target/$driver/$port/$tpg
3753  * @tpg: Target portal group to deregister.
3754  */
3755 static void srpt_drop_tpg(struct se_portal_group *tpg)
3756 {
3757 	struct srpt_tpg *stpg = container_of(tpg, typeof(*stpg), tpg);
3758 	struct srpt_port_id *sport_id = stpg->sport_id;
3759 	struct srpt_port *sport = srpt_tpg_to_sport(tpg);
3760 
3761 	mutex_lock(&sport_id->mutex);
3762 	list_del(&stpg->entry);
3763 	mutex_unlock(&sport_id->mutex);
3764 
3765 	sport->enabled = false;
3766 	core_tpg_deregister(tpg);
3767 	kfree(stpg);
3768 }
3769 
3770 /**
3771  * srpt_make_tport - configfs callback invoked for mkdir /sys/kernel/config/target/$driver/$port
3772  * @tf: Not used.
3773  * @group: Not used.
3774  * @name: $port.
3775  */
3776 static struct se_wwn *srpt_make_tport(struct target_fabric_configfs *tf,
3777 				      struct config_group *group,
3778 				      const char *name)
3779 {
3780 	struct port_and_port_id papi = srpt_lookup_port(name);
3781 	struct srpt_port *sport = papi.sport;
3782 	struct srpt_port_id *port_id;
3783 
3784 	if (!papi.port_id)
3785 		return ERR_PTR(-EINVAL);
3786 	if (*papi.port_id) {
3787 		/* Attempt to create a directory that already exists. */
3788 		WARN_ON_ONCE(true);
3789 		return &(*papi.port_id)->wwn;
3790 	}
3791 	port_id = kzalloc(sizeof(*port_id), GFP_KERNEL);
3792 	if (!port_id) {
3793 		srpt_sdev_put(sport->sdev);
3794 		return ERR_PTR(-ENOMEM);
3795 	}
3796 	mutex_init(&port_id->mutex);
3797 	INIT_LIST_HEAD(&port_id->tpg_list);
3798 	port_id->wwn.priv = sport;
3799 	memcpy(port_id->name, port_id == sport->guid_id ? sport->guid_name :
3800 	       sport->gid_name, ARRAY_SIZE(port_id->name));
3801 
3802 	*papi.port_id = port_id;
3803 
3804 	return &port_id->wwn;
3805 }
3806 
3807 /**
3808  * srpt_drop_tport - configfs callback invoked for rmdir /sys/kernel/config/target/$driver/$port
3809  * @wwn: $port.
3810  */
3811 static void srpt_drop_tport(struct se_wwn *wwn)
3812 {
3813 	struct srpt_port_id *port_id = container_of(wwn, typeof(*port_id), wwn);
3814 	struct srpt_port *sport = wwn->priv;
3815 
3816 	if (sport->guid_id == port_id)
3817 		sport->guid_id = NULL;
3818 	else if (sport->gid_id == port_id)
3819 		sport->gid_id = NULL;
3820 	else
3821 		WARN_ON_ONCE(true);
3822 
3823 	srpt_sdev_put(sport->sdev);
3824 	kfree(port_id);
3825 }
3826 
3827 static ssize_t srpt_wwn_version_show(struct config_item *item, char *buf)
3828 {
3829 	return sysfs_emit(buf, "\n");
3830 }
3831 
3832 CONFIGFS_ATTR_RO(srpt_wwn_, version);
3833 
3834 static struct configfs_attribute *srpt_wwn_attrs[] = {
3835 	&srpt_wwn_attr_version,
3836 	NULL,
3837 };
3838 
3839 static const struct target_core_fabric_ops srpt_template = {
3840 	.module				= THIS_MODULE,
3841 	.fabric_name			= "srpt",
3842 	.tpg_get_wwn			= srpt_get_fabric_wwn,
3843 	.tpg_get_tag			= srpt_get_tag,
3844 	.tpg_check_demo_mode_cache	= srpt_check_true,
3845 	.tpg_check_demo_mode_write_protect = srpt_check_true,
3846 	.release_cmd			= srpt_release_cmd,
3847 	.check_stop_free		= srpt_check_stop_free,
3848 	.close_session			= srpt_close_session,
3849 	.sess_get_initiator_sid		= NULL,
3850 	.write_pending			= srpt_write_pending,
3851 	.get_cmd_state			= srpt_get_tcm_cmd_state,
3852 	.queue_data_in			= srpt_queue_data_in,
3853 	.queue_status			= srpt_queue_status,
3854 	.queue_tm_rsp			= srpt_queue_tm_rsp,
3855 	.aborted_task			= srpt_aborted_task,
3856 	/*
3857 	 * Setup function pointers for generic logic in
3858 	 * target_core_fabric_configfs.c
3859 	 */
3860 	.fabric_make_wwn		= srpt_make_tport,
3861 	.fabric_drop_wwn		= srpt_drop_tport,
3862 	.fabric_make_tpg		= srpt_make_tpg,
3863 	.fabric_enable_tpg		= srpt_enable_tpg,
3864 	.fabric_drop_tpg		= srpt_drop_tpg,
3865 	.fabric_init_nodeacl		= srpt_init_nodeacl,
3866 
3867 	.tfc_discovery_attrs		= srpt_da_attrs,
3868 	.tfc_wwn_attrs			= srpt_wwn_attrs,
3869 	.tfc_tpg_attrib_attrs		= srpt_tpg_attrib_attrs,
3870 };
3871 
3872 /**
3873  * srpt_init_module - kernel module initialization
3874  *
3875  * Note: Since ib_register_client() registers callback functions, and since at
3876  * least one of these callback functions (srpt_add_one()) calls target core
3877  * functions, this driver must be registered with the target core before
3878  * ib_register_client() is called.
3879  */
3880 static int __init srpt_init_module(void)
3881 {
3882 	int ret;
3883 
3884 	ret = -EINVAL;
3885 	if (srp_max_req_size < MIN_MAX_REQ_SIZE) {
3886 		pr_err("invalid value %d for kernel module parameter srp_max_req_size -- must be at least %d.\n",
3887 		       srp_max_req_size, MIN_MAX_REQ_SIZE);
3888 		goto out;
3889 	}
3890 
3891 	if (srpt_srq_size < MIN_SRPT_SRQ_SIZE
3892 	    || srpt_srq_size > MAX_SRPT_SRQ_SIZE) {
3893 		pr_err("invalid value %d for kernel module parameter srpt_srq_size -- must be in the range [%d..%d].\n",
3894 		       srpt_srq_size, MIN_SRPT_SRQ_SIZE, MAX_SRPT_SRQ_SIZE);
3895 		goto out;
3896 	}
3897 
3898 	ret = target_register_template(&srpt_template);
3899 	if (ret)
3900 		goto out;
3901 
3902 	ret = ib_register_client(&srpt_client);
3903 	if (ret) {
3904 		pr_err("couldn't register IB client\n");
3905 		goto out_unregister_target;
3906 	}
3907 
3908 	return 0;
3909 
3910 out_unregister_target:
3911 	target_unregister_template(&srpt_template);
3912 out:
3913 	return ret;
3914 }
3915 
3916 static void __exit srpt_cleanup_module(void)
3917 {
3918 	if (rdma_cm_id)
3919 		rdma_destroy_id(rdma_cm_id);
3920 	ib_unregister_client(&srpt_client);
3921 	target_unregister_template(&srpt_template);
3922 }
3923 
3924 module_init(srpt_init_module);
3925 module_exit(srpt_cleanup_module);
3926