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 <scsi/scsi_tcq.h>
45 #include <target/configfs_macros.h>
46 #include <target/target_core_base.h>
47 #include <target/target_core_fabric_configfs.h>
48 #include <target/target_core_fabric.h>
49 #include <target/target_core_configfs.h>
50 #include "ib_srpt.h"
51 
52 /* Name of this kernel module. */
53 #define DRV_NAME		"ib_srpt"
54 #define DRV_VERSION		"2.0.0"
55 #define DRV_RELDATE		"2011-02-14"
56 
57 #define SRPT_ID_STRING	"Linux SRP target"
58 
59 #undef pr_fmt
60 #define pr_fmt(fmt) DRV_NAME " " fmt
61 
62 MODULE_AUTHOR("Vu Pham and Bart Van Assche");
63 MODULE_DESCRIPTION("InfiniBand SCSI RDMA Protocol target "
64 		   "v" DRV_VERSION " (" DRV_RELDATE ")");
65 MODULE_LICENSE("Dual BSD/GPL");
66 
67 /*
68  * Global Variables
69  */
70 
71 static u64 srpt_service_guid;
72 static DEFINE_SPINLOCK(srpt_dev_lock);	/* Protects srpt_dev_list. */
73 static LIST_HEAD(srpt_dev_list);	/* List of srpt_device structures. */
74 
75 static unsigned srp_max_req_size = DEFAULT_MAX_REQ_SIZE;
76 module_param(srp_max_req_size, int, 0444);
77 MODULE_PARM_DESC(srp_max_req_size,
78 		 "Maximum size of SRP request messages in bytes.");
79 
80 static int srpt_srq_size = DEFAULT_SRPT_SRQ_SIZE;
81 module_param(srpt_srq_size, int, 0444);
82 MODULE_PARM_DESC(srpt_srq_size,
83 		 "Shared receive queue (SRQ) size.");
84 
85 static int srpt_get_u64_x(char *buffer, struct kernel_param *kp)
86 {
87 	return sprintf(buffer, "0x%016llx", *(u64 *)kp->arg);
88 }
89 module_param_call(srpt_service_guid, NULL, srpt_get_u64_x, &srpt_service_guid,
90 		  0444);
91 MODULE_PARM_DESC(srpt_service_guid,
92 		 "Using this value for ioc_guid, id_ext, and cm_listen_id"
93 		 " instead of using the node_guid of the first HCA.");
94 
95 static struct ib_client srpt_client;
96 static struct target_fabric_configfs *srpt_target;
97 static void srpt_release_channel(struct srpt_rdma_ch *ch);
98 static int srpt_queue_status(struct se_cmd *cmd);
99 
100 /**
101  * opposite_dma_dir() - Swap DMA_TO_DEVICE and DMA_FROM_DEVICE.
102  */
103 static inline
104 enum dma_data_direction opposite_dma_dir(enum dma_data_direction dir)
105 {
106 	switch (dir) {
107 	case DMA_TO_DEVICE:	return DMA_FROM_DEVICE;
108 	case DMA_FROM_DEVICE:	return DMA_TO_DEVICE;
109 	default:		return dir;
110 	}
111 }
112 
113 /**
114  * srpt_sdev_name() - Return the name associated with the HCA.
115  *
116  * Examples are ib0, ib1, ...
117  */
118 static inline const char *srpt_sdev_name(struct srpt_device *sdev)
119 {
120 	return sdev->device->name;
121 }
122 
123 static enum rdma_ch_state srpt_get_ch_state(struct srpt_rdma_ch *ch)
124 {
125 	unsigned long flags;
126 	enum rdma_ch_state state;
127 
128 	spin_lock_irqsave(&ch->spinlock, flags);
129 	state = ch->state;
130 	spin_unlock_irqrestore(&ch->spinlock, flags);
131 	return state;
132 }
133 
134 static enum rdma_ch_state
135 srpt_set_ch_state(struct srpt_rdma_ch *ch, enum rdma_ch_state new_state)
136 {
137 	unsigned long flags;
138 	enum rdma_ch_state prev;
139 
140 	spin_lock_irqsave(&ch->spinlock, flags);
141 	prev = ch->state;
142 	ch->state = new_state;
143 	spin_unlock_irqrestore(&ch->spinlock, flags);
144 	return prev;
145 }
146 
147 /**
148  * srpt_test_and_set_ch_state() - Test and set the channel state.
149  *
150  * Returns true if and only if the channel state has been set to the new state.
151  */
152 static bool
153 srpt_test_and_set_ch_state(struct srpt_rdma_ch *ch, enum rdma_ch_state old,
154 			   enum rdma_ch_state new)
155 {
156 	unsigned long flags;
157 	enum rdma_ch_state prev;
158 
159 	spin_lock_irqsave(&ch->spinlock, flags);
160 	prev = ch->state;
161 	if (prev == old)
162 		ch->state = new;
163 	spin_unlock_irqrestore(&ch->spinlock, flags);
164 	return prev == old;
165 }
166 
167 /**
168  * srpt_event_handler() - Asynchronous IB event callback function.
169  *
170  * Callback function called by the InfiniBand core when an asynchronous IB
171  * event occurs. This callback may occur in interrupt context. See also
172  * section 11.5.2, Set Asynchronous Event Handler in the InfiniBand
173  * Architecture Specification.
174  */
175 static void srpt_event_handler(struct ib_event_handler *handler,
176 			       struct ib_event *event)
177 {
178 	struct srpt_device *sdev;
179 	struct srpt_port *sport;
180 
181 	sdev = ib_get_client_data(event->device, &srpt_client);
182 	if (!sdev || sdev->device != event->device)
183 		return;
184 
185 	pr_debug("ASYNC event= %d on device= %s\n", event->event,
186 		 srpt_sdev_name(sdev));
187 
188 	switch (event->event) {
189 	case IB_EVENT_PORT_ERR:
190 		if (event->element.port_num <= sdev->device->phys_port_cnt) {
191 			sport = &sdev->port[event->element.port_num - 1];
192 			sport->lid = 0;
193 			sport->sm_lid = 0;
194 		}
195 		break;
196 	case IB_EVENT_PORT_ACTIVE:
197 	case IB_EVENT_LID_CHANGE:
198 	case IB_EVENT_PKEY_CHANGE:
199 	case IB_EVENT_SM_CHANGE:
200 	case IB_EVENT_CLIENT_REREGISTER:
201 		/* Refresh port data asynchronously. */
202 		if (event->element.port_num <= sdev->device->phys_port_cnt) {
203 			sport = &sdev->port[event->element.port_num - 1];
204 			if (!sport->lid && !sport->sm_lid)
205 				schedule_work(&sport->work);
206 		}
207 		break;
208 	default:
209 		printk(KERN_ERR "received unrecognized IB event %d\n",
210 		       event->event);
211 		break;
212 	}
213 }
214 
215 /**
216  * srpt_srq_event() - SRQ event callback function.
217  */
218 static void srpt_srq_event(struct ib_event *event, void *ctx)
219 {
220 	printk(KERN_INFO "SRQ event %d\n", event->event);
221 }
222 
223 /**
224  * srpt_qp_event() - QP event callback function.
225  */
226 static void srpt_qp_event(struct ib_event *event, struct srpt_rdma_ch *ch)
227 {
228 	pr_debug("QP event %d on cm_id=%p sess_name=%s state=%d\n",
229 		 event->event, ch->cm_id, ch->sess_name, srpt_get_ch_state(ch));
230 
231 	switch (event->event) {
232 	case IB_EVENT_COMM_EST:
233 		ib_cm_notify(ch->cm_id, event->event);
234 		break;
235 	case IB_EVENT_QP_LAST_WQE_REACHED:
236 		if (srpt_test_and_set_ch_state(ch, CH_DRAINING,
237 					       CH_RELEASING))
238 			srpt_release_channel(ch);
239 		else
240 			pr_debug("%s: state %d - ignored LAST_WQE.\n",
241 				 ch->sess_name, srpt_get_ch_state(ch));
242 		break;
243 	default:
244 		printk(KERN_ERR "received unrecognized IB QP event %d\n",
245 		       event->event);
246 		break;
247 	}
248 }
249 
250 /**
251  * srpt_set_ioc() - Helper function for initializing an IOUnitInfo structure.
252  *
253  * @slot: one-based slot number.
254  * @value: four-bit value.
255  *
256  * Copies the lowest four bits of value in element slot of the array of four
257  * bit elements called c_list (controller list). The index slot is one-based.
258  */
259 static void srpt_set_ioc(u8 *c_list, u32 slot, u8 value)
260 {
261 	u16 id;
262 	u8 tmp;
263 
264 	id = (slot - 1) / 2;
265 	if (slot & 0x1) {
266 		tmp = c_list[id] & 0xf;
267 		c_list[id] = (value << 4) | tmp;
268 	} else {
269 		tmp = c_list[id] & 0xf0;
270 		c_list[id] = (value & 0xf) | tmp;
271 	}
272 }
273 
274 /**
275  * srpt_get_class_port_info() - Copy ClassPortInfo to a management datagram.
276  *
277  * See also section 16.3.3.1 ClassPortInfo in the InfiniBand Architecture
278  * Specification.
279  */
280 static void srpt_get_class_port_info(struct ib_dm_mad *mad)
281 {
282 	struct ib_class_port_info *cif;
283 
284 	cif = (struct ib_class_port_info *)mad->data;
285 	memset(cif, 0, sizeof *cif);
286 	cif->base_version = 1;
287 	cif->class_version = 1;
288 	cif->resp_time_value = 20;
289 
290 	mad->mad_hdr.status = 0;
291 }
292 
293 /**
294  * srpt_get_iou() - Write IOUnitInfo to a management datagram.
295  *
296  * See also section 16.3.3.3 IOUnitInfo in the InfiniBand Architecture
297  * Specification. See also section B.7, table B.6 in the SRP r16a document.
298  */
299 static void srpt_get_iou(struct ib_dm_mad *mad)
300 {
301 	struct ib_dm_iou_info *ioui;
302 	u8 slot;
303 	int i;
304 
305 	ioui = (struct ib_dm_iou_info *)mad->data;
306 	ioui->change_id = __constant_cpu_to_be16(1);
307 	ioui->max_controllers = 16;
308 
309 	/* set present for slot 1 and empty for the rest */
310 	srpt_set_ioc(ioui->controller_list, 1, 1);
311 	for (i = 1, slot = 2; i < 16; i++, slot++)
312 		srpt_set_ioc(ioui->controller_list, slot, 0);
313 
314 	mad->mad_hdr.status = 0;
315 }
316 
317 /**
318  * srpt_get_ioc() - Write IOControllerprofile to a management datagram.
319  *
320  * See also section 16.3.3.4 IOControllerProfile in the InfiniBand
321  * Architecture Specification. See also section B.7, table B.7 in the SRP
322  * r16a document.
323  */
324 static void srpt_get_ioc(struct srpt_port *sport, u32 slot,
325 			 struct ib_dm_mad *mad)
326 {
327 	struct srpt_device *sdev = sport->sdev;
328 	struct ib_dm_ioc_profile *iocp;
329 
330 	iocp = (struct ib_dm_ioc_profile *)mad->data;
331 
332 	if (!slot || slot > 16) {
333 		mad->mad_hdr.status
334 			= __constant_cpu_to_be16(DM_MAD_STATUS_INVALID_FIELD);
335 		return;
336 	}
337 
338 	if (slot > 2) {
339 		mad->mad_hdr.status
340 			= __constant_cpu_to_be16(DM_MAD_STATUS_NO_IOC);
341 		return;
342 	}
343 
344 	memset(iocp, 0, sizeof *iocp);
345 	strcpy(iocp->id_string, SRPT_ID_STRING);
346 	iocp->guid = cpu_to_be64(srpt_service_guid);
347 	iocp->vendor_id = cpu_to_be32(sdev->dev_attr.vendor_id);
348 	iocp->device_id = cpu_to_be32(sdev->dev_attr.vendor_part_id);
349 	iocp->device_version = cpu_to_be16(sdev->dev_attr.hw_ver);
350 	iocp->subsys_vendor_id = cpu_to_be32(sdev->dev_attr.vendor_id);
351 	iocp->subsys_device_id = 0x0;
352 	iocp->io_class = __constant_cpu_to_be16(SRP_REV16A_IB_IO_CLASS);
353 	iocp->io_subclass = __constant_cpu_to_be16(SRP_IO_SUBCLASS);
354 	iocp->protocol = __constant_cpu_to_be16(SRP_PROTOCOL);
355 	iocp->protocol_version = __constant_cpu_to_be16(SRP_PROTOCOL_VERSION);
356 	iocp->send_queue_depth = cpu_to_be16(sdev->srq_size);
357 	iocp->rdma_read_depth = 4;
358 	iocp->send_size = cpu_to_be32(srp_max_req_size);
359 	iocp->rdma_size = cpu_to_be32(min(sport->port_attrib.srp_max_rdma_size,
360 					  1U << 24));
361 	iocp->num_svc_entries = 1;
362 	iocp->op_cap_mask = SRP_SEND_TO_IOC | SRP_SEND_FROM_IOC |
363 		SRP_RDMA_READ_FROM_IOC | SRP_RDMA_WRITE_FROM_IOC;
364 
365 	mad->mad_hdr.status = 0;
366 }
367 
368 /**
369  * srpt_get_svc_entries() - Write ServiceEntries to a management datagram.
370  *
371  * See also section 16.3.3.5 ServiceEntries in the InfiniBand Architecture
372  * Specification. See also section B.7, table B.8 in the SRP r16a document.
373  */
374 static void srpt_get_svc_entries(u64 ioc_guid,
375 				 u16 slot, u8 hi, u8 lo, struct ib_dm_mad *mad)
376 {
377 	struct ib_dm_svc_entries *svc_entries;
378 
379 	WARN_ON(!ioc_guid);
380 
381 	if (!slot || slot > 16) {
382 		mad->mad_hdr.status
383 			= __constant_cpu_to_be16(DM_MAD_STATUS_INVALID_FIELD);
384 		return;
385 	}
386 
387 	if (slot > 2 || lo > hi || hi > 1) {
388 		mad->mad_hdr.status
389 			= __constant_cpu_to_be16(DM_MAD_STATUS_NO_IOC);
390 		return;
391 	}
392 
393 	svc_entries = (struct ib_dm_svc_entries *)mad->data;
394 	memset(svc_entries, 0, sizeof *svc_entries);
395 	svc_entries->service_entries[0].id = cpu_to_be64(ioc_guid);
396 	snprintf(svc_entries->service_entries[0].name,
397 		 sizeof(svc_entries->service_entries[0].name),
398 		 "%s%016llx",
399 		 SRP_SERVICE_NAME_PREFIX,
400 		 ioc_guid);
401 
402 	mad->mad_hdr.status = 0;
403 }
404 
405 /**
406  * srpt_mgmt_method_get() - Process a received management datagram.
407  * @sp:      source port through which the MAD has been received.
408  * @rq_mad:  received MAD.
409  * @rsp_mad: response MAD.
410  */
411 static void srpt_mgmt_method_get(struct srpt_port *sp, struct ib_mad *rq_mad,
412 				 struct ib_dm_mad *rsp_mad)
413 {
414 	u16 attr_id;
415 	u32 slot;
416 	u8 hi, lo;
417 
418 	attr_id = be16_to_cpu(rq_mad->mad_hdr.attr_id);
419 	switch (attr_id) {
420 	case DM_ATTR_CLASS_PORT_INFO:
421 		srpt_get_class_port_info(rsp_mad);
422 		break;
423 	case DM_ATTR_IOU_INFO:
424 		srpt_get_iou(rsp_mad);
425 		break;
426 	case DM_ATTR_IOC_PROFILE:
427 		slot = be32_to_cpu(rq_mad->mad_hdr.attr_mod);
428 		srpt_get_ioc(sp, slot, rsp_mad);
429 		break;
430 	case DM_ATTR_SVC_ENTRIES:
431 		slot = be32_to_cpu(rq_mad->mad_hdr.attr_mod);
432 		hi = (u8) ((slot >> 8) & 0xff);
433 		lo = (u8) (slot & 0xff);
434 		slot = (u16) ((slot >> 16) & 0xffff);
435 		srpt_get_svc_entries(srpt_service_guid,
436 				     slot, hi, lo, rsp_mad);
437 		break;
438 	default:
439 		rsp_mad->mad_hdr.status =
440 		    __constant_cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD_ATTR);
441 		break;
442 	}
443 }
444 
445 /**
446  * srpt_mad_send_handler() - Post MAD-send callback function.
447  */
448 static void srpt_mad_send_handler(struct ib_mad_agent *mad_agent,
449 				  struct ib_mad_send_wc *mad_wc)
450 {
451 	ib_destroy_ah(mad_wc->send_buf->ah);
452 	ib_free_send_mad(mad_wc->send_buf);
453 }
454 
455 /**
456  * srpt_mad_recv_handler() - MAD reception callback function.
457  */
458 static void srpt_mad_recv_handler(struct ib_mad_agent *mad_agent,
459 				  struct ib_mad_recv_wc *mad_wc)
460 {
461 	struct srpt_port *sport = (struct srpt_port *)mad_agent->context;
462 	struct ib_ah *ah;
463 	struct ib_mad_send_buf *rsp;
464 	struct ib_dm_mad *dm_mad;
465 
466 	if (!mad_wc || !mad_wc->recv_buf.mad)
467 		return;
468 
469 	ah = ib_create_ah_from_wc(mad_agent->qp->pd, mad_wc->wc,
470 				  mad_wc->recv_buf.grh, mad_agent->port_num);
471 	if (IS_ERR(ah))
472 		goto err;
473 
474 	BUILD_BUG_ON(offsetof(struct ib_dm_mad, data) != IB_MGMT_DEVICE_HDR);
475 
476 	rsp = ib_create_send_mad(mad_agent, mad_wc->wc->src_qp,
477 				 mad_wc->wc->pkey_index, 0,
478 				 IB_MGMT_DEVICE_HDR, IB_MGMT_DEVICE_DATA,
479 				 GFP_KERNEL);
480 	if (IS_ERR(rsp))
481 		goto err_rsp;
482 
483 	rsp->ah = ah;
484 
485 	dm_mad = rsp->mad;
486 	memcpy(dm_mad, mad_wc->recv_buf.mad, sizeof *dm_mad);
487 	dm_mad->mad_hdr.method = IB_MGMT_METHOD_GET_RESP;
488 	dm_mad->mad_hdr.status = 0;
489 
490 	switch (mad_wc->recv_buf.mad->mad_hdr.method) {
491 	case IB_MGMT_METHOD_GET:
492 		srpt_mgmt_method_get(sport, mad_wc->recv_buf.mad, dm_mad);
493 		break;
494 	case IB_MGMT_METHOD_SET:
495 		dm_mad->mad_hdr.status =
496 		    __constant_cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD_ATTR);
497 		break;
498 	default:
499 		dm_mad->mad_hdr.status =
500 		    __constant_cpu_to_be16(DM_MAD_STATUS_UNSUP_METHOD);
501 		break;
502 	}
503 
504 	if (!ib_post_send_mad(rsp, NULL)) {
505 		ib_free_recv_mad(mad_wc);
506 		/* will destroy_ah & free_send_mad in send completion */
507 		return;
508 	}
509 
510 	ib_free_send_mad(rsp);
511 
512 err_rsp:
513 	ib_destroy_ah(ah);
514 err:
515 	ib_free_recv_mad(mad_wc);
516 }
517 
518 /**
519  * srpt_refresh_port() - Configure a HCA port.
520  *
521  * Enable InfiniBand management datagram processing, update the cached sm_lid,
522  * lid and gid values, and register a callback function for processing MADs
523  * on the specified port.
524  *
525  * Note: It is safe to call this function more than once for the same port.
526  */
527 static int srpt_refresh_port(struct srpt_port *sport)
528 {
529 	struct ib_mad_reg_req reg_req;
530 	struct ib_port_modify port_modify;
531 	struct ib_port_attr port_attr;
532 	int ret;
533 
534 	memset(&port_modify, 0, sizeof port_modify);
535 	port_modify.set_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP;
536 	port_modify.clr_port_cap_mask = 0;
537 
538 	ret = ib_modify_port(sport->sdev->device, sport->port, 0, &port_modify);
539 	if (ret)
540 		goto err_mod_port;
541 
542 	ret = ib_query_port(sport->sdev->device, sport->port, &port_attr);
543 	if (ret)
544 		goto err_query_port;
545 
546 	sport->sm_lid = port_attr.sm_lid;
547 	sport->lid = port_attr.lid;
548 
549 	ret = ib_query_gid(sport->sdev->device, sport->port, 0, &sport->gid);
550 	if (ret)
551 		goto err_query_port;
552 
553 	if (!sport->mad_agent) {
554 		memset(&reg_req, 0, sizeof reg_req);
555 		reg_req.mgmt_class = IB_MGMT_CLASS_DEVICE_MGMT;
556 		reg_req.mgmt_class_version = IB_MGMT_BASE_VERSION;
557 		set_bit(IB_MGMT_METHOD_GET, reg_req.method_mask);
558 		set_bit(IB_MGMT_METHOD_SET, reg_req.method_mask);
559 
560 		sport->mad_agent = ib_register_mad_agent(sport->sdev->device,
561 							 sport->port,
562 							 IB_QPT_GSI,
563 							 &reg_req, 0,
564 							 srpt_mad_send_handler,
565 							 srpt_mad_recv_handler,
566 							 sport);
567 		if (IS_ERR(sport->mad_agent)) {
568 			ret = PTR_ERR(sport->mad_agent);
569 			sport->mad_agent = NULL;
570 			goto err_query_port;
571 		}
572 	}
573 
574 	return 0;
575 
576 err_query_port:
577 
578 	port_modify.set_port_cap_mask = 0;
579 	port_modify.clr_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP;
580 	ib_modify_port(sport->sdev->device, sport->port, 0, &port_modify);
581 
582 err_mod_port:
583 
584 	return ret;
585 }
586 
587 /**
588  * srpt_unregister_mad_agent() - Unregister MAD callback functions.
589  *
590  * Note: It is safe to call this function more than once for the same device.
591  */
592 static void srpt_unregister_mad_agent(struct srpt_device *sdev)
593 {
594 	struct ib_port_modify port_modify = {
595 		.clr_port_cap_mask = IB_PORT_DEVICE_MGMT_SUP,
596 	};
597 	struct srpt_port *sport;
598 	int i;
599 
600 	for (i = 1; i <= sdev->device->phys_port_cnt; i++) {
601 		sport = &sdev->port[i - 1];
602 		WARN_ON(sport->port != i);
603 		if (ib_modify_port(sdev->device, i, 0, &port_modify) < 0)
604 			printk(KERN_ERR "disabling MAD processing failed.\n");
605 		if (sport->mad_agent) {
606 			ib_unregister_mad_agent(sport->mad_agent);
607 			sport->mad_agent = NULL;
608 		}
609 	}
610 }
611 
612 /**
613  * srpt_alloc_ioctx() - Allocate an SRPT I/O context structure.
614  */
615 static struct srpt_ioctx *srpt_alloc_ioctx(struct srpt_device *sdev,
616 					   int ioctx_size, int dma_size,
617 					   enum dma_data_direction dir)
618 {
619 	struct srpt_ioctx *ioctx;
620 
621 	ioctx = kmalloc(ioctx_size, GFP_KERNEL);
622 	if (!ioctx)
623 		goto err;
624 
625 	ioctx->buf = kmalloc(dma_size, GFP_KERNEL);
626 	if (!ioctx->buf)
627 		goto err_free_ioctx;
628 
629 	ioctx->dma = ib_dma_map_single(sdev->device, ioctx->buf, dma_size, dir);
630 	if (ib_dma_mapping_error(sdev->device, ioctx->dma))
631 		goto err_free_buf;
632 
633 	return ioctx;
634 
635 err_free_buf:
636 	kfree(ioctx->buf);
637 err_free_ioctx:
638 	kfree(ioctx);
639 err:
640 	return NULL;
641 }
642 
643 /**
644  * srpt_free_ioctx() - Free an SRPT I/O context structure.
645  */
646 static void srpt_free_ioctx(struct srpt_device *sdev, struct srpt_ioctx *ioctx,
647 			    int dma_size, enum dma_data_direction dir)
648 {
649 	if (!ioctx)
650 		return;
651 
652 	ib_dma_unmap_single(sdev->device, ioctx->dma, dma_size, dir);
653 	kfree(ioctx->buf);
654 	kfree(ioctx);
655 }
656 
657 /**
658  * srpt_alloc_ioctx_ring() - Allocate a ring of SRPT I/O context structures.
659  * @sdev:       Device to allocate the I/O context ring for.
660  * @ring_size:  Number of elements in the I/O context ring.
661  * @ioctx_size: I/O context size.
662  * @dma_size:   DMA buffer size.
663  * @dir:        DMA data direction.
664  */
665 static struct srpt_ioctx **srpt_alloc_ioctx_ring(struct srpt_device *sdev,
666 				int ring_size, int ioctx_size,
667 				int dma_size, enum dma_data_direction dir)
668 {
669 	struct srpt_ioctx **ring;
670 	int i;
671 
672 	WARN_ON(ioctx_size != sizeof(struct srpt_recv_ioctx)
673 		&& ioctx_size != sizeof(struct srpt_send_ioctx));
674 
675 	ring = kmalloc(ring_size * sizeof(ring[0]), GFP_KERNEL);
676 	if (!ring)
677 		goto out;
678 	for (i = 0; i < ring_size; ++i) {
679 		ring[i] = srpt_alloc_ioctx(sdev, ioctx_size, dma_size, dir);
680 		if (!ring[i])
681 			goto err;
682 		ring[i]->index = i;
683 	}
684 	goto out;
685 
686 err:
687 	while (--i >= 0)
688 		srpt_free_ioctx(sdev, ring[i], dma_size, dir);
689 	kfree(ring);
690 	ring = NULL;
691 out:
692 	return ring;
693 }
694 
695 /**
696  * srpt_free_ioctx_ring() - Free the ring of SRPT I/O context structures.
697  */
698 static void srpt_free_ioctx_ring(struct srpt_ioctx **ioctx_ring,
699 				 struct srpt_device *sdev, int ring_size,
700 				 int dma_size, enum dma_data_direction dir)
701 {
702 	int i;
703 
704 	for (i = 0; i < ring_size; ++i)
705 		srpt_free_ioctx(sdev, ioctx_ring[i], dma_size, dir);
706 	kfree(ioctx_ring);
707 }
708 
709 /**
710  * srpt_get_cmd_state() - Get the state of a SCSI command.
711  */
712 static enum srpt_command_state srpt_get_cmd_state(struct srpt_send_ioctx *ioctx)
713 {
714 	enum srpt_command_state state;
715 	unsigned long flags;
716 
717 	BUG_ON(!ioctx);
718 
719 	spin_lock_irqsave(&ioctx->spinlock, flags);
720 	state = ioctx->state;
721 	spin_unlock_irqrestore(&ioctx->spinlock, flags);
722 	return state;
723 }
724 
725 /**
726  * srpt_set_cmd_state() - Set the state of a SCSI command.
727  *
728  * Does not modify the state of aborted commands. Returns the previous command
729  * state.
730  */
731 static enum srpt_command_state srpt_set_cmd_state(struct srpt_send_ioctx *ioctx,
732 						  enum srpt_command_state new)
733 {
734 	enum srpt_command_state previous;
735 	unsigned long flags;
736 
737 	BUG_ON(!ioctx);
738 
739 	spin_lock_irqsave(&ioctx->spinlock, flags);
740 	previous = ioctx->state;
741 	if (previous != SRPT_STATE_DONE)
742 		ioctx->state = new;
743 	spin_unlock_irqrestore(&ioctx->spinlock, flags);
744 
745 	return previous;
746 }
747 
748 /**
749  * srpt_test_and_set_cmd_state() - Test and set the state of a command.
750  *
751  * Returns true if and only if the previous command state was equal to 'old'.
752  */
753 static bool srpt_test_and_set_cmd_state(struct srpt_send_ioctx *ioctx,
754 					enum srpt_command_state old,
755 					enum srpt_command_state new)
756 {
757 	enum srpt_command_state previous;
758 	unsigned long flags;
759 
760 	WARN_ON(!ioctx);
761 	WARN_ON(old == SRPT_STATE_DONE);
762 	WARN_ON(new == SRPT_STATE_NEW);
763 
764 	spin_lock_irqsave(&ioctx->spinlock, flags);
765 	previous = ioctx->state;
766 	if (previous == old)
767 		ioctx->state = new;
768 	spin_unlock_irqrestore(&ioctx->spinlock, flags);
769 	return previous == old;
770 }
771 
772 /**
773  * srpt_post_recv() - Post an IB receive request.
774  */
775 static int srpt_post_recv(struct srpt_device *sdev,
776 			  struct srpt_recv_ioctx *ioctx)
777 {
778 	struct ib_sge list;
779 	struct ib_recv_wr wr, *bad_wr;
780 
781 	BUG_ON(!sdev);
782 	wr.wr_id = encode_wr_id(SRPT_RECV, ioctx->ioctx.index);
783 
784 	list.addr = ioctx->ioctx.dma;
785 	list.length = srp_max_req_size;
786 	list.lkey = sdev->mr->lkey;
787 
788 	wr.next = NULL;
789 	wr.sg_list = &list;
790 	wr.num_sge = 1;
791 
792 	return ib_post_srq_recv(sdev->srq, &wr, &bad_wr);
793 }
794 
795 /**
796  * srpt_post_send() - Post an IB send request.
797  *
798  * Returns zero upon success and a non-zero value upon failure.
799  */
800 static int srpt_post_send(struct srpt_rdma_ch *ch,
801 			  struct srpt_send_ioctx *ioctx, int len)
802 {
803 	struct ib_sge list;
804 	struct ib_send_wr wr, *bad_wr;
805 	struct srpt_device *sdev = ch->sport->sdev;
806 	int ret;
807 
808 	atomic_inc(&ch->req_lim);
809 
810 	ret = -ENOMEM;
811 	if (unlikely(atomic_dec_return(&ch->sq_wr_avail) < 0)) {
812 		printk(KERN_WARNING "IB send queue full (needed 1)\n");
813 		goto out;
814 	}
815 
816 	ib_dma_sync_single_for_device(sdev->device, ioctx->ioctx.dma, len,
817 				      DMA_TO_DEVICE);
818 
819 	list.addr = ioctx->ioctx.dma;
820 	list.length = len;
821 	list.lkey = sdev->mr->lkey;
822 
823 	wr.next = NULL;
824 	wr.wr_id = encode_wr_id(SRPT_SEND, ioctx->ioctx.index);
825 	wr.sg_list = &list;
826 	wr.num_sge = 1;
827 	wr.opcode = IB_WR_SEND;
828 	wr.send_flags = IB_SEND_SIGNALED;
829 
830 	ret = ib_post_send(ch->qp, &wr, &bad_wr);
831 
832 out:
833 	if (ret < 0) {
834 		atomic_inc(&ch->sq_wr_avail);
835 		atomic_dec(&ch->req_lim);
836 	}
837 	return ret;
838 }
839 
840 /**
841  * srpt_get_desc_tbl() - Parse the data descriptors of an SRP_CMD request.
842  * @ioctx: Pointer to the I/O context associated with the request.
843  * @srp_cmd: Pointer to the SRP_CMD request data.
844  * @dir: Pointer to the variable to which the transfer direction will be
845  *   written.
846  * @data_len: Pointer to the variable to which the total data length of all
847  *   descriptors in the SRP_CMD request will be written.
848  *
849  * This function initializes ioctx->nrbuf and ioctx->r_bufs.
850  *
851  * Returns -EINVAL when the SRP_CMD request contains inconsistent descriptors;
852  * -ENOMEM when memory allocation fails and zero upon success.
853  */
854 static int srpt_get_desc_tbl(struct srpt_send_ioctx *ioctx,
855 			     struct srp_cmd *srp_cmd,
856 			     enum dma_data_direction *dir, u64 *data_len)
857 {
858 	struct srp_indirect_buf *idb;
859 	struct srp_direct_buf *db;
860 	unsigned add_cdb_offset;
861 	int ret;
862 
863 	/*
864 	 * The pointer computations below will only be compiled correctly
865 	 * if srp_cmd::add_data is declared as s8*, u8*, s8[] or u8[], so check
866 	 * whether srp_cmd::add_data has been declared as a byte pointer.
867 	 */
868 	BUILD_BUG_ON(!__same_type(srp_cmd->add_data[0], (s8)0)
869 		     && !__same_type(srp_cmd->add_data[0], (u8)0));
870 
871 	BUG_ON(!dir);
872 	BUG_ON(!data_len);
873 
874 	ret = 0;
875 	*data_len = 0;
876 
877 	/*
878 	 * The lower four bits of the buffer format field contain the DATA-IN
879 	 * buffer descriptor format, and the highest four bits contain the
880 	 * DATA-OUT buffer descriptor format.
881 	 */
882 	*dir = DMA_NONE;
883 	if (srp_cmd->buf_fmt & 0xf)
884 		/* DATA-IN: transfer data from target to initiator (read). */
885 		*dir = DMA_FROM_DEVICE;
886 	else if (srp_cmd->buf_fmt >> 4)
887 		/* DATA-OUT: transfer data from initiator to target (write). */
888 		*dir = DMA_TO_DEVICE;
889 
890 	/*
891 	 * According to the SRP spec, the lower two bits of the 'ADDITIONAL
892 	 * CDB LENGTH' field are reserved and the size in bytes of this field
893 	 * is four times the value specified in bits 3..7. Hence the "& ~3".
894 	 */
895 	add_cdb_offset = srp_cmd->add_cdb_len & ~3;
896 	if (((srp_cmd->buf_fmt & 0xf) == SRP_DATA_DESC_DIRECT) ||
897 	    ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_DIRECT)) {
898 		ioctx->n_rbuf = 1;
899 		ioctx->rbufs = &ioctx->single_rbuf;
900 
901 		db = (struct srp_direct_buf *)(srp_cmd->add_data
902 					       + add_cdb_offset);
903 		memcpy(ioctx->rbufs, db, sizeof *db);
904 		*data_len = be32_to_cpu(db->len);
905 	} else if (((srp_cmd->buf_fmt & 0xf) == SRP_DATA_DESC_INDIRECT) ||
906 		   ((srp_cmd->buf_fmt >> 4) == SRP_DATA_DESC_INDIRECT)) {
907 		idb = (struct srp_indirect_buf *)(srp_cmd->add_data
908 						  + add_cdb_offset);
909 
910 		ioctx->n_rbuf = be32_to_cpu(idb->table_desc.len) / sizeof *db;
911 
912 		if (ioctx->n_rbuf >
913 		    (srp_cmd->data_out_desc_cnt + srp_cmd->data_in_desc_cnt)) {
914 			printk(KERN_ERR "received unsupported SRP_CMD request"
915 			       " type (%u out + %u in != %u / %zu)\n",
916 			       srp_cmd->data_out_desc_cnt,
917 			       srp_cmd->data_in_desc_cnt,
918 			       be32_to_cpu(idb->table_desc.len),
919 			       sizeof(*db));
920 			ioctx->n_rbuf = 0;
921 			ret = -EINVAL;
922 			goto out;
923 		}
924 
925 		if (ioctx->n_rbuf == 1)
926 			ioctx->rbufs = &ioctx->single_rbuf;
927 		else {
928 			ioctx->rbufs =
929 				kmalloc(ioctx->n_rbuf * sizeof *db, GFP_ATOMIC);
930 			if (!ioctx->rbufs) {
931 				ioctx->n_rbuf = 0;
932 				ret = -ENOMEM;
933 				goto out;
934 			}
935 		}
936 
937 		db = idb->desc_list;
938 		memcpy(ioctx->rbufs, db, ioctx->n_rbuf * sizeof *db);
939 		*data_len = be32_to_cpu(idb->len);
940 	}
941 out:
942 	return ret;
943 }
944 
945 /**
946  * srpt_init_ch_qp() - Initialize queue pair attributes.
947  *
948  * Initialized the attributes of queue pair 'qp' by allowing local write,
949  * remote read and remote write. Also transitions 'qp' to state IB_QPS_INIT.
950  */
951 static int srpt_init_ch_qp(struct srpt_rdma_ch *ch, struct ib_qp *qp)
952 {
953 	struct ib_qp_attr *attr;
954 	int ret;
955 
956 	attr = kzalloc(sizeof *attr, GFP_KERNEL);
957 	if (!attr)
958 		return -ENOMEM;
959 
960 	attr->qp_state = IB_QPS_INIT;
961 	attr->qp_access_flags = IB_ACCESS_LOCAL_WRITE | IB_ACCESS_REMOTE_READ |
962 	    IB_ACCESS_REMOTE_WRITE;
963 	attr->port_num = ch->sport->port;
964 	attr->pkey_index = 0;
965 
966 	ret = ib_modify_qp(qp, attr,
967 			   IB_QP_STATE | IB_QP_ACCESS_FLAGS | IB_QP_PORT |
968 			   IB_QP_PKEY_INDEX);
969 
970 	kfree(attr);
971 	return ret;
972 }
973 
974 /**
975  * srpt_ch_qp_rtr() - Change the state of a channel to 'ready to receive' (RTR).
976  * @ch: channel of the queue pair.
977  * @qp: queue pair to change the state of.
978  *
979  * Returns zero upon success and a negative value upon failure.
980  *
981  * Note: currently a struct ib_qp_attr takes 136 bytes on a 64-bit system.
982  * If this structure ever becomes larger, it might be necessary to allocate
983  * it dynamically instead of on the stack.
984  */
985 static int srpt_ch_qp_rtr(struct srpt_rdma_ch *ch, struct ib_qp *qp)
986 {
987 	struct ib_qp_attr qp_attr;
988 	int attr_mask;
989 	int ret;
990 
991 	qp_attr.qp_state = IB_QPS_RTR;
992 	ret = ib_cm_init_qp_attr(ch->cm_id, &qp_attr, &attr_mask);
993 	if (ret)
994 		goto out;
995 
996 	qp_attr.max_dest_rd_atomic = 4;
997 
998 	ret = ib_modify_qp(qp, &qp_attr, attr_mask);
999 
1000 out:
1001 	return ret;
1002 }
1003 
1004 /**
1005  * srpt_ch_qp_rts() - Change the state of a channel to 'ready to send' (RTS).
1006  * @ch: channel of the queue pair.
1007  * @qp: queue pair to change the state of.
1008  *
1009  * Returns zero upon success and a negative value upon failure.
1010  *
1011  * Note: currently a struct ib_qp_attr takes 136 bytes on a 64-bit system.
1012  * If this structure ever becomes larger, it might be necessary to allocate
1013  * it dynamically instead of on the stack.
1014  */
1015 static int srpt_ch_qp_rts(struct srpt_rdma_ch *ch, struct ib_qp *qp)
1016 {
1017 	struct ib_qp_attr qp_attr;
1018 	int attr_mask;
1019 	int ret;
1020 
1021 	qp_attr.qp_state = IB_QPS_RTS;
1022 	ret = ib_cm_init_qp_attr(ch->cm_id, &qp_attr, &attr_mask);
1023 	if (ret)
1024 		goto out;
1025 
1026 	qp_attr.max_rd_atomic = 4;
1027 
1028 	ret = ib_modify_qp(qp, &qp_attr, attr_mask);
1029 
1030 out:
1031 	return ret;
1032 }
1033 
1034 /**
1035  * srpt_ch_qp_err() - Set the channel queue pair state to 'error'.
1036  */
1037 static int srpt_ch_qp_err(struct srpt_rdma_ch *ch)
1038 {
1039 	struct ib_qp_attr qp_attr;
1040 
1041 	qp_attr.qp_state = IB_QPS_ERR;
1042 	return ib_modify_qp(ch->qp, &qp_attr, IB_QP_STATE);
1043 }
1044 
1045 /**
1046  * srpt_unmap_sg_to_ib_sge() - Unmap an IB SGE list.
1047  */
1048 static void srpt_unmap_sg_to_ib_sge(struct srpt_rdma_ch *ch,
1049 				    struct srpt_send_ioctx *ioctx)
1050 {
1051 	struct scatterlist *sg;
1052 	enum dma_data_direction dir;
1053 
1054 	BUG_ON(!ch);
1055 	BUG_ON(!ioctx);
1056 	BUG_ON(ioctx->n_rdma && !ioctx->rdma_ius);
1057 
1058 	while (ioctx->n_rdma)
1059 		kfree(ioctx->rdma_ius[--ioctx->n_rdma].sge);
1060 
1061 	kfree(ioctx->rdma_ius);
1062 	ioctx->rdma_ius = NULL;
1063 
1064 	if (ioctx->mapped_sg_count) {
1065 		sg = ioctx->sg;
1066 		WARN_ON(!sg);
1067 		dir = ioctx->cmd.data_direction;
1068 		BUG_ON(dir == DMA_NONE);
1069 		ib_dma_unmap_sg(ch->sport->sdev->device, sg, ioctx->sg_cnt,
1070 				opposite_dma_dir(dir));
1071 		ioctx->mapped_sg_count = 0;
1072 	}
1073 }
1074 
1075 /**
1076  * srpt_map_sg_to_ib_sge() - Map an SG list to an IB SGE list.
1077  */
1078 static int srpt_map_sg_to_ib_sge(struct srpt_rdma_ch *ch,
1079 				 struct srpt_send_ioctx *ioctx)
1080 {
1081 	struct se_cmd *cmd;
1082 	struct scatterlist *sg, *sg_orig;
1083 	int sg_cnt;
1084 	enum dma_data_direction dir;
1085 	struct rdma_iu *riu;
1086 	struct srp_direct_buf *db;
1087 	dma_addr_t dma_addr;
1088 	struct ib_sge *sge;
1089 	u64 raddr;
1090 	u32 rsize;
1091 	u32 tsize;
1092 	u32 dma_len;
1093 	int count, nrdma;
1094 	int i, j, k;
1095 
1096 	BUG_ON(!ch);
1097 	BUG_ON(!ioctx);
1098 	cmd = &ioctx->cmd;
1099 	dir = cmd->data_direction;
1100 	BUG_ON(dir == DMA_NONE);
1101 
1102 	transport_do_task_sg_chain(cmd);
1103 	ioctx->sg = sg = sg_orig = cmd->t_tasks_sg_chained;
1104 	ioctx->sg_cnt = sg_cnt = cmd->t_tasks_sg_chained_no;
1105 
1106 	count = ib_dma_map_sg(ch->sport->sdev->device, sg, sg_cnt,
1107 			      opposite_dma_dir(dir));
1108 	if (unlikely(!count))
1109 		return -EAGAIN;
1110 
1111 	ioctx->mapped_sg_count = count;
1112 
1113 	if (ioctx->rdma_ius && ioctx->n_rdma_ius)
1114 		nrdma = ioctx->n_rdma_ius;
1115 	else {
1116 		nrdma = (count + SRPT_DEF_SG_PER_WQE - 1) / SRPT_DEF_SG_PER_WQE
1117 			+ ioctx->n_rbuf;
1118 
1119 		ioctx->rdma_ius = kzalloc(nrdma * sizeof *riu, GFP_KERNEL);
1120 		if (!ioctx->rdma_ius)
1121 			goto free_mem;
1122 
1123 		ioctx->n_rdma_ius = nrdma;
1124 	}
1125 
1126 	db = ioctx->rbufs;
1127 	tsize = cmd->data_length;
1128 	dma_len = sg_dma_len(&sg[0]);
1129 	riu = ioctx->rdma_ius;
1130 
1131 	/*
1132 	 * For each remote desc - calculate the #ib_sge.
1133 	 * If #ib_sge < SRPT_DEF_SG_PER_WQE per rdma operation then
1134 	 *      each remote desc rdma_iu is required a rdma wr;
1135 	 * else
1136 	 *      we need to allocate extra rdma_iu to carry extra #ib_sge in
1137 	 *      another rdma wr
1138 	 */
1139 	for (i = 0, j = 0;
1140 	     j < count && i < ioctx->n_rbuf && tsize > 0; ++i, ++riu, ++db) {
1141 		rsize = be32_to_cpu(db->len);
1142 		raddr = be64_to_cpu(db->va);
1143 		riu->raddr = raddr;
1144 		riu->rkey = be32_to_cpu(db->key);
1145 		riu->sge_cnt = 0;
1146 
1147 		/* calculate how many sge required for this remote_buf */
1148 		while (rsize > 0 && tsize > 0) {
1149 
1150 			if (rsize >= dma_len) {
1151 				tsize -= dma_len;
1152 				rsize -= dma_len;
1153 				raddr += dma_len;
1154 
1155 				if (tsize > 0) {
1156 					++j;
1157 					if (j < count) {
1158 						sg = sg_next(sg);
1159 						dma_len = sg_dma_len(sg);
1160 					}
1161 				}
1162 			} else {
1163 				tsize -= rsize;
1164 				dma_len -= rsize;
1165 				rsize = 0;
1166 			}
1167 
1168 			++riu->sge_cnt;
1169 
1170 			if (rsize > 0 && riu->sge_cnt == SRPT_DEF_SG_PER_WQE) {
1171 				++ioctx->n_rdma;
1172 				riu->sge =
1173 				    kmalloc(riu->sge_cnt * sizeof *riu->sge,
1174 					    GFP_KERNEL);
1175 				if (!riu->sge)
1176 					goto free_mem;
1177 
1178 				++riu;
1179 				riu->sge_cnt = 0;
1180 				riu->raddr = raddr;
1181 				riu->rkey = be32_to_cpu(db->key);
1182 			}
1183 		}
1184 
1185 		++ioctx->n_rdma;
1186 		riu->sge = kmalloc(riu->sge_cnt * sizeof *riu->sge,
1187 				   GFP_KERNEL);
1188 		if (!riu->sge)
1189 			goto free_mem;
1190 	}
1191 
1192 	db = ioctx->rbufs;
1193 	tsize = cmd->data_length;
1194 	riu = ioctx->rdma_ius;
1195 	sg = sg_orig;
1196 	dma_len = sg_dma_len(&sg[0]);
1197 	dma_addr = sg_dma_address(&sg[0]);
1198 
1199 	/* this second loop is really mapped sg_addres to rdma_iu->ib_sge */
1200 	for (i = 0, j = 0;
1201 	     j < count && i < ioctx->n_rbuf && tsize > 0; ++i, ++riu, ++db) {
1202 		rsize = be32_to_cpu(db->len);
1203 		sge = riu->sge;
1204 		k = 0;
1205 
1206 		while (rsize > 0 && tsize > 0) {
1207 			sge->addr = dma_addr;
1208 			sge->lkey = ch->sport->sdev->mr->lkey;
1209 
1210 			if (rsize >= dma_len) {
1211 				sge->length =
1212 					(tsize < dma_len) ? tsize : dma_len;
1213 				tsize -= dma_len;
1214 				rsize -= dma_len;
1215 
1216 				if (tsize > 0) {
1217 					++j;
1218 					if (j < count) {
1219 						sg = sg_next(sg);
1220 						dma_len = sg_dma_len(sg);
1221 						dma_addr = sg_dma_address(sg);
1222 					}
1223 				}
1224 			} else {
1225 				sge->length = (tsize < rsize) ? tsize : rsize;
1226 				tsize -= rsize;
1227 				dma_len -= rsize;
1228 				dma_addr += rsize;
1229 				rsize = 0;
1230 			}
1231 
1232 			++k;
1233 			if (k == riu->sge_cnt && rsize > 0 && tsize > 0) {
1234 				++riu;
1235 				sge = riu->sge;
1236 				k = 0;
1237 			} else if (rsize > 0 && tsize > 0)
1238 				++sge;
1239 		}
1240 	}
1241 
1242 	return 0;
1243 
1244 free_mem:
1245 	srpt_unmap_sg_to_ib_sge(ch, ioctx);
1246 
1247 	return -ENOMEM;
1248 }
1249 
1250 /**
1251  * srpt_get_send_ioctx() - Obtain an I/O context for sending to the initiator.
1252  */
1253 static struct srpt_send_ioctx *srpt_get_send_ioctx(struct srpt_rdma_ch *ch)
1254 {
1255 	struct srpt_send_ioctx *ioctx;
1256 	unsigned long flags;
1257 
1258 	BUG_ON(!ch);
1259 
1260 	ioctx = NULL;
1261 	spin_lock_irqsave(&ch->spinlock, flags);
1262 	if (!list_empty(&ch->free_list)) {
1263 		ioctx = list_first_entry(&ch->free_list,
1264 					 struct srpt_send_ioctx, free_list);
1265 		list_del(&ioctx->free_list);
1266 	}
1267 	spin_unlock_irqrestore(&ch->spinlock, flags);
1268 
1269 	if (!ioctx)
1270 		return ioctx;
1271 
1272 	BUG_ON(ioctx->ch != ch);
1273 	kref_init(&ioctx->kref);
1274 	spin_lock_init(&ioctx->spinlock);
1275 	ioctx->state = SRPT_STATE_NEW;
1276 	ioctx->n_rbuf = 0;
1277 	ioctx->rbufs = NULL;
1278 	ioctx->n_rdma = 0;
1279 	ioctx->n_rdma_ius = 0;
1280 	ioctx->rdma_ius = NULL;
1281 	ioctx->mapped_sg_count = 0;
1282 	init_completion(&ioctx->tx_done);
1283 	ioctx->queue_status_only = false;
1284 	/*
1285 	 * transport_init_se_cmd() does not initialize all fields, so do it
1286 	 * here.
1287 	 */
1288 	memset(&ioctx->cmd, 0, sizeof(ioctx->cmd));
1289 	memset(&ioctx->sense_data, 0, sizeof(ioctx->sense_data));
1290 
1291 	return ioctx;
1292 }
1293 
1294 /**
1295  * srpt_put_send_ioctx() - Free up resources.
1296  */
1297 static void srpt_put_send_ioctx(struct srpt_send_ioctx *ioctx)
1298 {
1299 	struct srpt_rdma_ch *ch;
1300 	unsigned long flags;
1301 
1302 	BUG_ON(!ioctx);
1303 	ch = ioctx->ch;
1304 	BUG_ON(!ch);
1305 
1306 	WARN_ON(srpt_get_cmd_state(ioctx) != SRPT_STATE_DONE);
1307 
1308 	srpt_unmap_sg_to_ib_sge(ioctx->ch, ioctx);
1309 	transport_generic_free_cmd(&ioctx->cmd, 0);
1310 
1311 	if (ioctx->n_rbuf > 1) {
1312 		kfree(ioctx->rbufs);
1313 		ioctx->rbufs = NULL;
1314 		ioctx->n_rbuf = 0;
1315 	}
1316 
1317 	spin_lock_irqsave(&ch->spinlock, flags);
1318 	list_add(&ioctx->free_list, &ch->free_list);
1319 	spin_unlock_irqrestore(&ch->spinlock, flags);
1320 }
1321 
1322 static void srpt_put_send_ioctx_kref(struct kref *kref)
1323 {
1324 	srpt_put_send_ioctx(container_of(kref, struct srpt_send_ioctx, kref));
1325 }
1326 
1327 /**
1328  * srpt_abort_cmd() - Abort a SCSI command.
1329  * @ioctx:   I/O context associated with the SCSI command.
1330  * @context: Preferred execution context.
1331  */
1332 static int srpt_abort_cmd(struct srpt_send_ioctx *ioctx)
1333 {
1334 	enum srpt_command_state state;
1335 	unsigned long flags;
1336 
1337 	BUG_ON(!ioctx);
1338 
1339 	/*
1340 	 * If the command is in a state where the target core is waiting for
1341 	 * the ib_srpt driver, change the state to the next state. Changing
1342 	 * the state of the command from SRPT_STATE_NEED_DATA to
1343 	 * SRPT_STATE_DATA_IN ensures that srpt_xmit_response() will call this
1344 	 * function a second time.
1345 	 */
1346 
1347 	spin_lock_irqsave(&ioctx->spinlock, flags);
1348 	state = ioctx->state;
1349 	switch (state) {
1350 	case SRPT_STATE_NEED_DATA:
1351 		ioctx->state = SRPT_STATE_DATA_IN;
1352 		break;
1353 	case SRPT_STATE_DATA_IN:
1354 	case SRPT_STATE_CMD_RSP_SENT:
1355 	case SRPT_STATE_MGMT_RSP_SENT:
1356 		ioctx->state = SRPT_STATE_DONE;
1357 		break;
1358 	default:
1359 		break;
1360 	}
1361 	spin_unlock_irqrestore(&ioctx->spinlock, flags);
1362 
1363 	if (state == SRPT_STATE_DONE)
1364 		goto out;
1365 
1366 	pr_debug("Aborting cmd with state %d and tag %lld\n", state,
1367 		 ioctx->tag);
1368 
1369 	switch (state) {
1370 	case SRPT_STATE_NEW:
1371 	case SRPT_STATE_DATA_IN:
1372 	case SRPT_STATE_MGMT:
1373 		/*
1374 		 * Do nothing - defer abort processing until
1375 		 * srpt_queue_response() is invoked.
1376 		 */
1377 		WARN_ON(!transport_check_aborted_status(&ioctx->cmd, false));
1378 		break;
1379 	case SRPT_STATE_NEED_DATA:
1380 		/* DMA_TO_DEVICE (write) - RDMA read error. */
1381 		spin_lock_irqsave(&ioctx->cmd.t_state_lock, flags);
1382 		ioctx->cmd.transport_state |= CMD_T_LUN_STOP;
1383 		spin_unlock_irqrestore(&ioctx->cmd.t_state_lock, flags);
1384 		transport_generic_handle_data(&ioctx->cmd);
1385 		break;
1386 	case SRPT_STATE_CMD_RSP_SENT:
1387 		/*
1388 		 * SRP_RSP sending failed or the SRP_RSP send completion has
1389 		 * not been received in time.
1390 		 */
1391 		srpt_unmap_sg_to_ib_sge(ioctx->ch, ioctx);
1392 		spin_lock_irqsave(&ioctx->cmd.t_state_lock, flags);
1393 		ioctx->cmd.transport_state |= CMD_T_LUN_STOP;
1394 		spin_unlock_irqrestore(&ioctx->cmd.t_state_lock, flags);
1395 		kref_put(&ioctx->kref, srpt_put_send_ioctx_kref);
1396 		break;
1397 	case SRPT_STATE_MGMT_RSP_SENT:
1398 		srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
1399 		kref_put(&ioctx->kref, srpt_put_send_ioctx_kref);
1400 		break;
1401 	default:
1402 		WARN_ON("ERROR: unexpected command state");
1403 		break;
1404 	}
1405 
1406 out:
1407 	return state;
1408 }
1409 
1410 /**
1411  * srpt_handle_send_err_comp() - Process an IB_WC_SEND error completion.
1412  */
1413 static void srpt_handle_send_err_comp(struct srpt_rdma_ch *ch, u64 wr_id)
1414 {
1415 	struct srpt_send_ioctx *ioctx;
1416 	enum srpt_command_state state;
1417 	struct se_cmd *cmd;
1418 	u32 index;
1419 
1420 	atomic_inc(&ch->sq_wr_avail);
1421 
1422 	index = idx_from_wr_id(wr_id);
1423 	ioctx = ch->ioctx_ring[index];
1424 	state = srpt_get_cmd_state(ioctx);
1425 	cmd = &ioctx->cmd;
1426 
1427 	WARN_ON(state != SRPT_STATE_CMD_RSP_SENT
1428 		&& state != SRPT_STATE_MGMT_RSP_SENT
1429 		&& state != SRPT_STATE_NEED_DATA
1430 		&& state != SRPT_STATE_DONE);
1431 
1432 	/* If SRP_RSP sending failed, undo the ch->req_lim change. */
1433 	if (state == SRPT_STATE_CMD_RSP_SENT
1434 	    || state == SRPT_STATE_MGMT_RSP_SENT)
1435 		atomic_dec(&ch->req_lim);
1436 
1437 	srpt_abort_cmd(ioctx);
1438 }
1439 
1440 /**
1441  * srpt_handle_send_comp() - Process an IB send completion notification.
1442  */
1443 static void srpt_handle_send_comp(struct srpt_rdma_ch *ch,
1444 				  struct srpt_send_ioctx *ioctx)
1445 {
1446 	enum srpt_command_state state;
1447 
1448 	atomic_inc(&ch->sq_wr_avail);
1449 
1450 	state = srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
1451 
1452 	if (WARN_ON(state != SRPT_STATE_CMD_RSP_SENT
1453 		    && state != SRPT_STATE_MGMT_RSP_SENT
1454 		    && state != SRPT_STATE_DONE))
1455 		pr_debug("state = %d\n", state);
1456 
1457 	if (state != SRPT_STATE_DONE)
1458 		kref_put(&ioctx->kref, srpt_put_send_ioctx_kref);
1459 	else
1460 		printk(KERN_ERR "IB completion has been received too late for"
1461 		       " wr_id = %u.\n", ioctx->ioctx.index);
1462 }
1463 
1464 /**
1465  * srpt_handle_rdma_comp() - Process an IB RDMA completion notification.
1466  *
1467  * Note: transport_generic_handle_data() is asynchronous so unmapping the
1468  * data that has been transferred via IB RDMA must be postponed until the
1469  * check_stop_free() callback.
1470  */
1471 static void srpt_handle_rdma_comp(struct srpt_rdma_ch *ch,
1472 				  struct srpt_send_ioctx *ioctx,
1473 				  enum srpt_opcode opcode)
1474 {
1475 	WARN_ON(ioctx->n_rdma <= 0);
1476 	atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
1477 
1478 	if (opcode == SRPT_RDMA_READ_LAST) {
1479 		if (srpt_test_and_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA,
1480 						SRPT_STATE_DATA_IN))
1481 			transport_generic_handle_data(&ioctx->cmd);
1482 		else
1483 			printk(KERN_ERR "%s[%d]: wrong state = %d\n", __func__,
1484 			       __LINE__, srpt_get_cmd_state(ioctx));
1485 	} else if (opcode == SRPT_RDMA_ABORT) {
1486 		ioctx->rdma_aborted = true;
1487 	} else {
1488 		WARN(true, "unexpected opcode %d\n", opcode);
1489 	}
1490 }
1491 
1492 /**
1493  * srpt_handle_rdma_err_comp() - Process an IB RDMA error completion.
1494  */
1495 static void srpt_handle_rdma_err_comp(struct srpt_rdma_ch *ch,
1496 				      struct srpt_send_ioctx *ioctx,
1497 				      enum srpt_opcode opcode)
1498 {
1499 	struct se_cmd *cmd;
1500 	enum srpt_command_state state;
1501 	unsigned long flags;
1502 
1503 	cmd = &ioctx->cmd;
1504 	state = srpt_get_cmd_state(ioctx);
1505 	switch (opcode) {
1506 	case SRPT_RDMA_READ_LAST:
1507 		if (ioctx->n_rdma <= 0) {
1508 			printk(KERN_ERR "Received invalid RDMA read"
1509 			       " error completion with idx %d\n",
1510 			       ioctx->ioctx.index);
1511 			break;
1512 		}
1513 		atomic_add(ioctx->n_rdma, &ch->sq_wr_avail);
1514 		if (state == SRPT_STATE_NEED_DATA)
1515 			srpt_abort_cmd(ioctx);
1516 		else
1517 			printk(KERN_ERR "%s[%d]: wrong state = %d\n",
1518 			       __func__, __LINE__, state);
1519 		break;
1520 	case SRPT_RDMA_WRITE_LAST:
1521 		spin_lock_irqsave(&ioctx->cmd.t_state_lock, flags);
1522 		ioctx->cmd.transport_state |= CMD_T_LUN_STOP;
1523 		spin_unlock_irqrestore(&ioctx->cmd.t_state_lock, flags);
1524 		break;
1525 	default:
1526 		printk(KERN_ERR "%s[%d]: opcode = %u\n", __func__,
1527 		       __LINE__, opcode);
1528 		break;
1529 	}
1530 }
1531 
1532 /**
1533  * srpt_build_cmd_rsp() - Build an SRP_RSP response.
1534  * @ch: RDMA channel through which the request has been received.
1535  * @ioctx: I/O context associated with the SRP_CMD request. The response will
1536  *   be built in the buffer ioctx->buf points at and hence this function will
1537  *   overwrite the request data.
1538  * @tag: tag of the request for which this response is being generated.
1539  * @status: value for the STATUS field of the SRP_RSP information unit.
1540  *
1541  * Returns the size in bytes of the SRP_RSP response.
1542  *
1543  * An SRP_RSP response contains a SCSI status or service response. See also
1544  * section 6.9 in the SRP r16a document for the format of an SRP_RSP
1545  * response. See also SPC-2 for more information about sense data.
1546  */
1547 static int srpt_build_cmd_rsp(struct srpt_rdma_ch *ch,
1548 			      struct srpt_send_ioctx *ioctx, u64 tag,
1549 			      int status)
1550 {
1551 	struct srp_rsp *srp_rsp;
1552 	const u8 *sense_data;
1553 	int sense_data_len, max_sense_len;
1554 
1555 	/*
1556 	 * The lowest bit of all SAM-3 status codes is zero (see also
1557 	 * paragraph 5.3 in SAM-3).
1558 	 */
1559 	WARN_ON(status & 1);
1560 
1561 	srp_rsp = ioctx->ioctx.buf;
1562 	BUG_ON(!srp_rsp);
1563 
1564 	sense_data = ioctx->sense_data;
1565 	sense_data_len = ioctx->cmd.scsi_sense_length;
1566 	WARN_ON(sense_data_len > sizeof(ioctx->sense_data));
1567 
1568 	memset(srp_rsp, 0, sizeof *srp_rsp);
1569 	srp_rsp->opcode = SRP_RSP;
1570 	srp_rsp->req_lim_delta =
1571 		__constant_cpu_to_be32(1 + atomic_xchg(&ch->req_lim_delta, 0));
1572 	srp_rsp->tag = tag;
1573 	srp_rsp->status = status;
1574 
1575 	if (sense_data_len) {
1576 		BUILD_BUG_ON(MIN_MAX_RSP_SIZE <= sizeof(*srp_rsp));
1577 		max_sense_len = ch->max_ti_iu_len - sizeof(*srp_rsp);
1578 		if (sense_data_len > max_sense_len) {
1579 			printk(KERN_WARNING "truncated sense data from %d to %d"
1580 			       " bytes\n", sense_data_len, max_sense_len);
1581 			sense_data_len = max_sense_len;
1582 		}
1583 
1584 		srp_rsp->flags |= SRP_RSP_FLAG_SNSVALID;
1585 		srp_rsp->sense_data_len = cpu_to_be32(sense_data_len);
1586 		memcpy(srp_rsp + 1, sense_data, sense_data_len);
1587 	}
1588 
1589 	return sizeof(*srp_rsp) + sense_data_len;
1590 }
1591 
1592 /**
1593  * srpt_build_tskmgmt_rsp() - Build a task management response.
1594  * @ch:       RDMA channel through which the request has been received.
1595  * @ioctx:    I/O context in which the SRP_RSP response will be built.
1596  * @rsp_code: RSP_CODE that will be stored in the response.
1597  * @tag:      Tag of the request for which this response is being generated.
1598  *
1599  * Returns the size in bytes of the SRP_RSP response.
1600  *
1601  * An SRP_RSP response contains a SCSI status or service response. See also
1602  * section 6.9 in the SRP r16a document for the format of an SRP_RSP
1603  * response.
1604  */
1605 static int srpt_build_tskmgmt_rsp(struct srpt_rdma_ch *ch,
1606 				  struct srpt_send_ioctx *ioctx,
1607 				  u8 rsp_code, u64 tag)
1608 {
1609 	struct srp_rsp *srp_rsp;
1610 	int resp_data_len;
1611 	int resp_len;
1612 
1613 	resp_data_len = (rsp_code == SRP_TSK_MGMT_SUCCESS) ? 0 : 4;
1614 	resp_len = sizeof(*srp_rsp) + resp_data_len;
1615 
1616 	srp_rsp = ioctx->ioctx.buf;
1617 	BUG_ON(!srp_rsp);
1618 	memset(srp_rsp, 0, sizeof *srp_rsp);
1619 
1620 	srp_rsp->opcode = SRP_RSP;
1621 	srp_rsp->req_lim_delta = __constant_cpu_to_be32(1
1622 				    + atomic_xchg(&ch->req_lim_delta, 0));
1623 	srp_rsp->tag = tag;
1624 
1625 	if (rsp_code != SRP_TSK_MGMT_SUCCESS) {
1626 		srp_rsp->flags |= SRP_RSP_FLAG_RSPVALID;
1627 		srp_rsp->resp_data_len = cpu_to_be32(resp_data_len);
1628 		srp_rsp->data[3] = rsp_code;
1629 	}
1630 
1631 	return resp_len;
1632 }
1633 
1634 #define NO_SUCH_LUN ((uint64_t)-1LL)
1635 
1636 /*
1637  * SCSI LUN addressing method. See also SAM-2 and the section about
1638  * eight byte LUNs.
1639  */
1640 enum scsi_lun_addr_method {
1641 	SCSI_LUN_ADDR_METHOD_PERIPHERAL   = 0,
1642 	SCSI_LUN_ADDR_METHOD_FLAT         = 1,
1643 	SCSI_LUN_ADDR_METHOD_LUN          = 2,
1644 	SCSI_LUN_ADDR_METHOD_EXTENDED_LUN = 3,
1645 };
1646 
1647 /*
1648  * srpt_unpack_lun() - Convert from network LUN to linear LUN.
1649  *
1650  * Convert an 2-byte, 4-byte, 6-byte or 8-byte LUN structure in network byte
1651  * order (big endian) to a linear LUN. Supports three LUN addressing methods:
1652  * peripheral, flat and logical unit. See also SAM-2, section 4.9.4 (page 40).
1653  */
1654 static uint64_t srpt_unpack_lun(const uint8_t *lun, int len)
1655 {
1656 	uint64_t res = NO_SUCH_LUN;
1657 	int addressing_method;
1658 
1659 	if (unlikely(len < 2)) {
1660 		printk(KERN_ERR "Illegal LUN length %d, expected 2 bytes or "
1661 		       "more", len);
1662 		goto out;
1663 	}
1664 
1665 	switch (len) {
1666 	case 8:
1667 		if ((*((__be64 *)lun) &
1668 		     __constant_cpu_to_be64(0x0000FFFFFFFFFFFFLL)) != 0)
1669 			goto out_err;
1670 		break;
1671 	case 4:
1672 		if (*((__be16 *)&lun[2]) != 0)
1673 			goto out_err;
1674 		break;
1675 	case 6:
1676 		if (*((__be32 *)&lun[2]) != 0)
1677 			goto out_err;
1678 		break;
1679 	case 2:
1680 		break;
1681 	default:
1682 		goto out_err;
1683 	}
1684 
1685 	addressing_method = (*lun) >> 6; /* highest two bits of byte 0 */
1686 	switch (addressing_method) {
1687 	case SCSI_LUN_ADDR_METHOD_PERIPHERAL:
1688 	case SCSI_LUN_ADDR_METHOD_FLAT:
1689 	case SCSI_LUN_ADDR_METHOD_LUN:
1690 		res = *(lun + 1) | (((*lun) & 0x3f) << 8);
1691 		break;
1692 
1693 	case SCSI_LUN_ADDR_METHOD_EXTENDED_LUN:
1694 	default:
1695 		printk(KERN_ERR "Unimplemented LUN addressing method %u",
1696 		       addressing_method);
1697 		break;
1698 	}
1699 
1700 out:
1701 	return res;
1702 
1703 out_err:
1704 	printk(KERN_ERR "Support for multi-level LUNs has not yet been"
1705 	       " implemented");
1706 	goto out;
1707 }
1708 
1709 static int srpt_check_stop_free(struct se_cmd *cmd)
1710 {
1711 	struct srpt_send_ioctx *ioctx;
1712 
1713 	ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
1714 	return kref_put(&ioctx->kref, srpt_put_send_ioctx_kref);
1715 }
1716 
1717 /**
1718  * srpt_handle_cmd() - Process SRP_CMD.
1719  */
1720 static int srpt_handle_cmd(struct srpt_rdma_ch *ch,
1721 			   struct srpt_recv_ioctx *recv_ioctx,
1722 			   struct srpt_send_ioctx *send_ioctx)
1723 {
1724 	struct se_cmd *cmd;
1725 	struct srp_cmd *srp_cmd;
1726 	uint64_t unpacked_lun;
1727 	u64 data_len;
1728 	enum dma_data_direction dir;
1729 	int ret;
1730 
1731 	BUG_ON(!send_ioctx);
1732 
1733 	srp_cmd = recv_ioctx->ioctx.buf;
1734 	kref_get(&send_ioctx->kref);
1735 	cmd = &send_ioctx->cmd;
1736 	send_ioctx->tag = srp_cmd->tag;
1737 
1738 	switch (srp_cmd->task_attr) {
1739 	case SRP_CMD_SIMPLE_Q:
1740 		cmd->sam_task_attr = MSG_SIMPLE_TAG;
1741 		break;
1742 	case SRP_CMD_ORDERED_Q:
1743 	default:
1744 		cmd->sam_task_attr = MSG_ORDERED_TAG;
1745 		break;
1746 	case SRP_CMD_HEAD_OF_Q:
1747 		cmd->sam_task_attr = MSG_HEAD_TAG;
1748 		break;
1749 	case SRP_CMD_ACA:
1750 		cmd->sam_task_attr = MSG_ACA_TAG;
1751 		break;
1752 	}
1753 
1754 	ret = srpt_get_desc_tbl(send_ioctx, srp_cmd, &dir, &data_len);
1755 	if (ret) {
1756 		printk(KERN_ERR "0x%llx: parsing SRP descriptor table failed.\n",
1757 		       srp_cmd->tag);
1758 		cmd->se_cmd_flags |= SCF_SCSI_CDB_EXCEPTION;
1759 		cmd->scsi_sense_reason = TCM_INVALID_CDB_FIELD;
1760 		kref_put(&send_ioctx->kref, srpt_put_send_ioctx_kref);
1761 		goto send_sense;
1762 	}
1763 
1764 	cmd->data_length = data_len;
1765 	cmd->data_direction = dir;
1766 	unpacked_lun = srpt_unpack_lun((uint8_t *)&srp_cmd->lun,
1767 				       sizeof(srp_cmd->lun));
1768 	if (transport_lookup_cmd_lun(cmd, unpacked_lun) < 0) {
1769 		kref_put(&send_ioctx->kref, srpt_put_send_ioctx_kref);
1770 		goto send_sense;
1771 	}
1772 	ret = transport_generic_allocate_tasks(cmd, srp_cmd->cdb);
1773 	if (ret < 0) {
1774 		kref_put(&send_ioctx->kref, srpt_put_send_ioctx_kref);
1775 		if (cmd->se_cmd_flags & SCF_SCSI_RESERVATION_CONFLICT) {
1776 			srpt_queue_status(cmd);
1777 			return 0;
1778 		} else
1779 			goto send_sense;
1780 	}
1781 
1782 	transport_handle_cdb_direct(cmd);
1783 	return 0;
1784 
1785 send_sense:
1786 	transport_send_check_condition_and_sense(cmd, cmd->scsi_sense_reason,
1787 						 0);
1788 	return -1;
1789 }
1790 
1791 /**
1792  * srpt_rx_mgmt_fn_tag() - Process a task management function by tag.
1793  * @ch: RDMA channel of the task management request.
1794  * @fn: Task management function to perform.
1795  * @req_tag: Tag of the SRP task management request.
1796  * @mgmt_ioctx: I/O context of the task management request.
1797  *
1798  * Returns zero if the target core will process the task management
1799  * request asynchronously.
1800  *
1801  * Note: It is assumed that the initiator serializes tag-based task management
1802  * requests.
1803  */
1804 static int srpt_rx_mgmt_fn_tag(struct srpt_send_ioctx *ioctx, u64 tag)
1805 {
1806 	struct srpt_device *sdev;
1807 	struct srpt_rdma_ch *ch;
1808 	struct srpt_send_ioctx *target;
1809 	int ret, i;
1810 
1811 	ret = -EINVAL;
1812 	ch = ioctx->ch;
1813 	BUG_ON(!ch);
1814 	BUG_ON(!ch->sport);
1815 	sdev = ch->sport->sdev;
1816 	BUG_ON(!sdev);
1817 	spin_lock_irq(&sdev->spinlock);
1818 	for (i = 0; i < ch->rq_size; ++i) {
1819 		target = ch->ioctx_ring[i];
1820 		if (target->cmd.se_lun == ioctx->cmd.se_lun &&
1821 		    target->tag == tag &&
1822 		    srpt_get_cmd_state(target) != SRPT_STATE_DONE) {
1823 			ret = 0;
1824 			/* now let the target core abort &target->cmd; */
1825 			break;
1826 		}
1827 	}
1828 	spin_unlock_irq(&sdev->spinlock);
1829 	return ret;
1830 }
1831 
1832 static int srp_tmr_to_tcm(int fn)
1833 {
1834 	switch (fn) {
1835 	case SRP_TSK_ABORT_TASK:
1836 		return TMR_ABORT_TASK;
1837 	case SRP_TSK_ABORT_TASK_SET:
1838 		return TMR_ABORT_TASK_SET;
1839 	case SRP_TSK_CLEAR_TASK_SET:
1840 		return TMR_CLEAR_TASK_SET;
1841 	case SRP_TSK_LUN_RESET:
1842 		return TMR_LUN_RESET;
1843 	case SRP_TSK_CLEAR_ACA:
1844 		return TMR_CLEAR_ACA;
1845 	default:
1846 		return -1;
1847 	}
1848 }
1849 
1850 /**
1851  * srpt_handle_tsk_mgmt() - Process an SRP_TSK_MGMT information unit.
1852  *
1853  * Returns 0 if and only if the request will be processed by the target core.
1854  *
1855  * For more information about SRP_TSK_MGMT information units, see also section
1856  * 6.7 in the SRP r16a document.
1857  */
1858 static void srpt_handle_tsk_mgmt(struct srpt_rdma_ch *ch,
1859 				 struct srpt_recv_ioctx *recv_ioctx,
1860 				 struct srpt_send_ioctx *send_ioctx)
1861 {
1862 	struct srp_tsk_mgmt *srp_tsk;
1863 	struct se_cmd *cmd;
1864 	uint64_t unpacked_lun;
1865 	int tcm_tmr;
1866 	int res;
1867 
1868 	BUG_ON(!send_ioctx);
1869 
1870 	srp_tsk = recv_ioctx->ioctx.buf;
1871 	cmd = &send_ioctx->cmd;
1872 
1873 	pr_debug("recv tsk_mgmt fn %d for task_tag %lld and cmd tag %lld"
1874 		 " cm_id %p sess %p\n", srp_tsk->tsk_mgmt_func,
1875 		 srp_tsk->task_tag, srp_tsk->tag, ch->cm_id, ch->sess);
1876 
1877 	srpt_set_cmd_state(send_ioctx, SRPT_STATE_MGMT);
1878 	send_ioctx->tag = srp_tsk->tag;
1879 	tcm_tmr = srp_tmr_to_tcm(srp_tsk->tsk_mgmt_func);
1880 	if (tcm_tmr < 0) {
1881 		send_ioctx->cmd.se_cmd_flags |= SCF_SCSI_CDB_EXCEPTION;
1882 		send_ioctx->cmd.se_tmr_req->response =
1883 			TMR_TASK_MGMT_FUNCTION_NOT_SUPPORTED;
1884 		goto process_tmr;
1885 	}
1886 	res = core_tmr_alloc_req(cmd, NULL, tcm_tmr, GFP_KERNEL);
1887 	if (res < 0) {
1888 		send_ioctx->cmd.se_cmd_flags |= SCF_SCSI_CDB_EXCEPTION;
1889 		send_ioctx->cmd.se_tmr_req->response = TMR_FUNCTION_REJECTED;
1890 		goto process_tmr;
1891 	}
1892 
1893 	unpacked_lun = srpt_unpack_lun((uint8_t *)&srp_tsk->lun,
1894 				       sizeof(srp_tsk->lun));
1895 	res = transport_lookup_tmr_lun(&send_ioctx->cmd, unpacked_lun);
1896 	if (res) {
1897 		pr_debug("rejecting TMR for LUN %lld\n", unpacked_lun);
1898 		send_ioctx->cmd.se_cmd_flags |= SCF_SCSI_CDB_EXCEPTION;
1899 		send_ioctx->cmd.se_tmr_req->response = TMR_LUN_DOES_NOT_EXIST;
1900 		goto process_tmr;
1901 	}
1902 
1903 	if (srp_tsk->tsk_mgmt_func == SRP_TSK_ABORT_TASK)
1904 		srpt_rx_mgmt_fn_tag(send_ioctx, srp_tsk->task_tag);
1905 
1906 process_tmr:
1907 	kref_get(&send_ioctx->kref);
1908 	if (!(send_ioctx->cmd.se_cmd_flags & SCF_SCSI_CDB_EXCEPTION))
1909 		transport_generic_handle_tmr(&send_ioctx->cmd);
1910 	else
1911 		transport_send_check_condition_and_sense(cmd,
1912 						cmd->scsi_sense_reason, 0);
1913 
1914 }
1915 
1916 /**
1917  * srpt_handle_new_iu() - Process a newly received information unit.
1918  * @ch:    RDMA channel through which the information unit has been received.
1919  * @ioctx: SRPT I/O context associated with the information unit.
1920  */
1921 static void srpt_handle_new_iu(struct srpt_rdma_ch *ch,
1922 			       struct srpt_recv_ioctx *recv_ioctx,
1923 			       struct srpt_send_ioctx *send_ioctx)
1924 {
1925 	struct srp_cmd *srp_cmd;
1926 	enum rdma_ch_state ch_state;
1927 
1928 	BUG_ON(!ch);
1929 	BUG_ON(!recv_ioctx);
1930 
1931 	ib_dma_sync_single_for_cpu(ch->sport->sdev->device,
1932 				   recv_ioctx->ioctx.dma, srp_max_req_size,
1933 				   DMA_FROM_DEVICE);
1934 
1935 	ch_state = srpt_get_ch_state(ch);
1936 	if (unlikely(ch_state == CH_CONNECTING)) {
1937 		list_add_tail(&recv_ioctx->wait_list, &ch->cmd_wait_list);
1938 		goto out;
1939 	}
1940 
1941 	if (unlikely(ch_state != CH_LIVE))
1942 		goto out;
1943 
1944 	srp_cmd = recv_ioctx->ioctx.buf;
1945 	if (srp_cmd->opcode == SRP_CMD || srp_cmd->opcode == SRP_TSK_MGMT) {
1946 		if (!send_ioctx)
1947 			send_ioctx = srpt_get_send_ioctx(ch);
1948 		if (unlikely(!send_ioctx)) {
1949 			list_add_tail(&recv_ioctx->wait_list,
1950 				      &ch->cmd_wait_list);
1951 			goto out;
1952 		}
1953 	}
1954 
1955 	transport_init_se_cmd(&send_ioctx->cmd, &srpt_target->tf_ops, ch->sess,
1956 			      0, DMA_NONE, MSG_SIMPLE_TAG,
1957 			      send_ioctx->sense_data);
1958 
1959 	switch (srp_cmd->opcode) {
1960 	case SRP_CMD:
1961 		srpt_handle_cmd(ch, recv_ioctx, send_ioctx);
1962 		break;
1963 	case SRP_TSK_MGMT:
1964 		srpt_handle_tsk_mgmt(ch, recv_ioctx, send_ioctx);
1965 		break;
1966 	case SRP_I_LOGOUT:
1967 		printk(KERN_ERR "Not yet implemented: SRP_I_LOGOUT\n");
1968 		break;
1969 	case SRP_CRED_RSP:
1970 		pr_debug("received SRP_CRED_RSP\n");
1971 		break;
1972 	case SRP_AER_RSP:
1973 		pr_debug("received SRP_AER_RSP\n");
1974 		break;
1975 	case SRP_RSP:
1976 		printk(KERN_ERR "Received SRP_RSP\n");
1977 		break;
1978 	default:
1979 		printk(KERN_ERR "received IU with unknown opcode 0x%x\n",
1980 		       srp_cmd->opcode);
1981 		break;
1982 	}
1983 
1984 	srpt_post_recv(ch->sport->sdev, recv_ioctx);
1985 out:
1986 	return;
1987 }
1988 
1989 static void srpt_process_rcv_completion(struct ib_cq *cq,
1990 					struct srpt_rdma_ch *ch,
1991 					struct ib_wc *wc)
1992 {
1993 	struct srpt_device *sdev = ch->sport->sdev;
1994 	struct srpt_recv_ioctx *ioctx;
1995 	u32 index;
1996 
1997 	index = idx_from_wr_id(wc->wr_id);
1998 	if (wc->status == IB_WC_SUCCESS) {
1999 		int req_lim;
2000 
2001 		req_lim = atomic_dec_return(&ch->req_lim);
2002 		if (unlikely(req_lim < 0))
2003 			printk(KERN_ERR "req_lim = %d < 0\n", req_lim);
2004 		ioctx = sdev->ioctx_ring[index];
2005 		srpt_handle_new_iu(ch, ioctx, NULL);
2006 	} else {
2007 		printk(KERN_INFO "receiving failed for idx %u with status %d\n",
2008 		       index, wc->status);
2009 	}
2010 }
2011 
2012 /**
2013  * srpt_process_send_completion() - Process an IB send completion.
2014  *
2015  * Note: Although this has not yet been observed during tests, at least in
2016  * theory it is possible that the srpt_get_send_ioctx() call invoked by
2017  * srpt_handle_new_iu() fails. This is possible because the req_lim_delta
2018  * value in each response is set to one, and it is possible that this response
2019  * makes the initiator send a new request before the send completion for that
2020  * response has been processed. This could e.g. happen if the call to
2021  * srpt_put_send_iotcx() is delayed because of a higher priority interrupt or
2022  * if IB retransmission causes generation of the send completion to be
2023  * delayed. Incoming information units for which srpt_get_send_ioctx() fails
2024  * are queued on cmd_wait_list. The code below processes these delayed
2025  * requests one at a time.
2026  */
2027 static void srpt_process_send_completion(struct ib_cq *cq,
2028 					 struct srpt_rdma_ch *ch,
2029 					 struct ib_wc *wc)
2030 {
2031 	struct srpt_send_ioctx *send_ioctx;
2032 	uint32_t index;
2033 	enum srpt_opcode opcode;
2034 
2035 	index = idx_from_wr_id(wc->wr_id);
2036 	opcode = opcode_from_wr_id(wc->wr_id);
2037 	send_ioctx = ch->ioctx_ring[index];
2038 	if (wc->status == IB_WC_SUCCESS) {
2039 		if (opcode == SRPT_SEND)
2040 			srpt_handle_send_comp(ch, send_ioctx);
2041 		else {
2042 			WARN_ON(opcode != SRPT_RDMA_ABORT &&
2043 				wc->opcode != IB_WC_RDMA_READ);
2044 			srpt_handle_rdma_comp(ch, send_ioctx, opcode);
2045 		}
2046 	} else {
2047 		if (opcode == SRPT_SEND) {
2048 			printk(KERN_INFO "sending response for idx %u failed"
2049 			       " with status %d\n", index, wc->status);
2050 			srpt_handle_send_err_comp(ch, wc->wr_id);
2051 		} else if (opcode != SRPT_RDMA_MID) {
2052 			printk(KERN_INFO "RDMA t %d for idx %u failed with"
2053 				" status %d", opcode, index, wc->status);
2054 			srpt_handle_rdma_err_comp(ch, send_ioctx, opcode);
2055 		}
2056 	}
2057 
2058 	while (unlikely(opcode == SRPT_SEND
2059 			&& !list_empty(&ch->cmd_wait_list)
2060 			&& srpt_get_ch_state(ch) == CH_LIVE
2061 			&& (send_ioctx = srpt_get_send_ioctx(ch)) != NULL)) {
2062 		struct srpt_recv_ioctx *recv_ioctx;
2063 
2064 		recv_ioctx = list_first_entry(&ch->cmd_wait_list,
2065 					      struct srpt_recv_ioctx,
2066 					      wait_list);
2067 		list_del(&recv_ioctx->wait_list);
2068 		srpt_handle_new_iu(ch, recv_ioctx, send_ioctx);
2069 	}
2070 }
2071 
2072 static void srpt_process_completion(struct ib_cq *cq, struct srpt_rdma_ch *ch)
2073 {
2074 	struct ib_wc *const wc = ch->wc;
2075 	int i, n;
2076 
2077 	WARN_ON(cq != ch->cq);
2078 
2079 	ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
2080 	while ((n = ib_poll_cq(cq, ARRAY_SIZE(ch->wc), wc)) > 0) {
2081 		for (i = 0; i < n; i++) {
2082 			if (opcode_from_wr_id(wc[i].wr_id) == SRPT_RECV)
2083 				srpt_process_rcv_completion(cq, ch, &wc[i]);
2084 			else
2085 				srpt_process_send_completion(cq, ch, &wc[i]);
2086 		}
2087 	}
2088 }
2089 
2090 /**
2091  * srpt_completion() - IB completion queue callback function.
2092  *
2093  * Notes:
2094  * - It is guaranteed that a completion handler will never be invoked
2095  *   concurrently on two different CPUs for the same completion queue. See also
2096  *   Documentation/infiniband/core_locking.txt and the implementation of
2097  *   handle_edge_irq() in kernel/irq/chip.c.
2098  * - When threaded IRQs are enabled, completion handlers are invoked in thread
2099  *   context instead of interrupt context.
2100  */
2101 static void srpt_completion(struct ib_cq *cq, void *ctx)
2102 {
2103 	struct srpt_rdma_ch *ch = ctx;
2104 
2105 	wake_up_interruptible(&ch->wait_queue);
2106 }
2107 
2108 static int srpt_compl_thread(void *arg)
2109 {
2110 	struct srpt_rdma_ch *ch;
2111 
2112 	/* Hibernation / freezing of the SRPT kernel thread is not supported. */
2113 	current->flags |= PF_NOFREEZE;
2114 
2115 	ch = arg;
2116 	BUG_ON(!ch);
2117 	printk(KERN_INFO "Session %s: kernel thread %s (PID %d) started\n",
2118 	       ch->sess_name, ch->thread->comm, current->pid);
2119 	while (!kthread_should_stop()) {
2120 		wait_event_interruptible(ch->wait_queue,
2121 			(srpt_process_completion(ch->cq, ch),
2122 			 kthread_should_stop()));
2123 	}
2124 	printk(KERN_INFO "Session %s: kernel thread %s (PID %d) stopped\n",
2125 	       ch->sess_name, ch->thread->comm, current->pid);
2126 	return 0;
2127 }
2128 
2129 /**
2130  * srpt_create_ch_ib() - Create receive and send completion queues.
2131  */
2132 static int srpt_create_ch_ib(struct srpt_rdma_ch *ch)
2133 {
2134 	struct ib_qp_init_attr *qp_init;
2135 	struct srpt_port *sport = ch->sport;
2136 	struct srpt_device *sdev = sport->sdev;
2137 	u32 srp_sq_size = sport->port_attrib.srp_sq_size;
2138 	int ret;
2139 
2140 	WARN_ON(ch->rq_size < 1);
2141 
2142 	ret = -ENOMEM;
2143 	qp_init = kzalloc(sizeof *qp_init, GFP_KERNEL);
2144 	if (!qp_init)
2145 		goto out;
2146 
2147 	ch->cq = ib_create_cq(sdev->device, srpt_completion, NULL, ch,
2148 			      ch->rq_size + srp_sq_size, 0);
2149 	if (IS_ERR(ch->cq)) {
2150 		ret = PTR_ERR(ch->cq);
2151 		printk(KERN_ERR "failed to create CQ cqe= %d ret= %d\n",
2152 		       ch->rq_size + srp_sq_size, ret);
2153 		goto out;
2154 	}
2155 
2156 	qp_init->qp_context = (void *)ch;
2157 	qp_init->event_handler
2158 		= (void(*)(struct ib_event *, void*))srpt_qp_event;
2159 	qp_init->send_cq = ch->cq;
2160 	qp_init->recv_cq = ch->cq;
2161 	qp_init->srq = sdev->srq;
2162 	qp_init->sq_sig_type = IB_SIGNAL_REQ_WR;
2163 	qp_init->qp_type = IB_QPT_RC;
2164 	qp_init->cap.max_send_wr = srp_sq_size;
2165 	qp_init->cap.max_send_sge = SRPT_DEF_SG_PER_WQE;
2166 
2167 	ch->qp = ib_create_qp(sdev->pd, qp_init);
2168 	if (IS_ERR(ch->qp)) {
2169 		ret = PTR_ERR(ch->qp);
2170 		printk(KERN_ERR "failed to create_qp ret= %d\n", ret);
2171 		goto err_destroy_cq;
2172 	}
2173 
2174 	atomic_set(&ch->sq_wr_avail, qp_init->cap.max_send_wr);
2175 
2176 	pr_debug("%s: max_cqe= %d max_sge= %d sq_size = %d cm_id= %p\n",
2177 		 __func__, ch->cq->cqe, qp_init->cap.max_send_sge,
2178 		 qp_init->cap.max_send_wr, ch->cm_id);
2179 
2180 	ret = srpt_init_ch_qp(ch, ch->qp);
2181 	if (ret)
2182 		goto err_destroy_qp;
2183 
2184 	init_waitqueue_head(&ch->wait_queue);
2185 
2186 	pr_debug("creating thread for session %s\n", ch->sess_name);
2187 
2188 	ch->thread = kthread_run(srpt_compl_thread, ch, "ib_srpt_compl");
2189 	if (IS_ERR(ch->thread)) {
2190 		printk(KERN_ERR "failed to create kernel thread %ld\n",
2191 		       PTR_ERR(ch->thread));
2192 		ch->thread = NULL;
2193 		goto err_destroy_qp;
2194 	}
2195 
2196 out:
2197 	kfree(qp_init);
2198 	return ret;
2199 
2200 err_destroy_qp:
2201 	ib_destroy_qp(ch->qp);
2202 err_destroy_cq:
2203 	ib_destroy_cq(ch->cq);
2204 	goto out;
2205 }
2206 
2207 static void srpt_destroy_ch_ib(struct srpt_rdma_ch *ch)
2208 {
2209 	if (ch->thread)
2210 		kthread_stop(ch->thread);
2211 
2212 	ib_destroy_qp(ch->qp);
2213 	ib_destroy_cq(ch->cq);
2214 }
2215 
2216 /**
2217  * __srpt_close_ch() - Close an RDMA channel by setting the QP error state.
2218  *
2219  * Reset the QP and make sure all resources associated with the channel will
2220  * be deallocated at an appropriate time.
2221  *
2222  * Note: The caller must hold ch->sport->sdev->spinlock.
2223  */
2224 static void __srpt_close_ch(struct srpt_rdma_ch *ch)
2225 {
2226 	struct srpt_device *sdev;
2227 	enum rdma_ch_state prev_state;
2228 	unsigned long flags;
2229 
2230 	sdev = ch->sport->sdev;
2231 
2232 	spin_lock_irqsave(&ch->spinlock, flags);
2233 	prev_state = ch->state;
2234 	switch (prev_state) {
2235 	case CH_CONNECTING:
2236 	case CH_LIVE:
2237 		ch->state = CH_DISCONNECTING;
2238 		break;
2239 	default:
2240 		break;
2241 	}
2242 	spin_unlock_irqrestore(&ch->spinlock, flags);
2243 
2244 	switch (prev_state) {
2245 	case CH_CONNECTING:
2246 		ib_send_cm_rej(ch->cm_id, IB_CM_REJ_NO_RESOURCES, NULL, 0,
2247 			       NULL, 0);
2248 		/* fall through */
2249 	case CH_LIVE:
2250 		if (ib_send_cm_dreq(ch->cm_id, NULL, 0) < 0)
2251 			printk(KERN_ERR "sending CM DREQ failed.\n");
2252 		break;
2253 	case CH_DISCONNECTING:
2254 		break;
2255 	case CH_DRAINING:
2256 	case CH_RELEASING:
2257 		break;
2258 	}
2259 }
2260 
2261 /**
2262  * srpt_close_ch() - Close an RDMA channel.
2263  */
2264 static void srpt_close_ch(struct srpt_rdma_ch *ch)
2265 {
2266 	struct srpt_device *sdev;
2267 
2268 	sdev = ch->sport->sdev;
2269 	spin_lock_irq(&sdev->spinlock);
2270 	__srpt_close_ch(ch);
2271 	spin_unlock_irq(&sdev->spinlock);
2272 }
2273 
2274 /**
2275  * srpt_drain_channel() - Drain a channel by resetting the IB queue pair.
2276  * @cm_id: Pointer to the CM ID of the channel to be drained.
2277  *
2278  * Note: Must be called from inside srpt_cm_handler to avoid a race between
2279  * accessing sdev->spinlock and the call to kfree(sdev) in srpt_remove_one()
2280  * (the caller of srpt_cm_handler holds the cm_id spinlock; srpt_remove_one()
2281  * waits until all target sessions for the associated IB device have been
2282  * unregistered and target session registration involves a call to
2283  * ib_destroy_cm_id(), which locks the cm_id spinlock and hence waits until
2284  * this function has finished).
2285  */
2286 static void srpt_drain_channel(struct ib_cm_id *cm_id)
2287 {
2288 	struct srpt_device *sdev;
2289 	struct srpt_rdma_ch *ch;
2290 	int ret;
2291 	bool do_reset = false;
2292 
2293 	WARN_ON_ONCE(irqs_disabled());
2294 
2295 	sdev = cm_id->context;
2296 	BUG_ON(!sdev);
2297 	spin_lock_irq(&sdev->spinlock);
2298 	list_for_each_entry(ch, &sdev->rch_list, list) {
2299 		if (ch->cm_id == cm_id) {
2300 			do_reset = srpt_test_and_set_ch_state(ch,
2301 					CH_CONNECTING, CH_DRAINING) ||
2302 				   srpt_test_and_set_ch_state(ch,
2303 					CH_LIVE, CH_DRAINING) ||
2304 				   srpt_test_and_set_ch_state(ch,
2305 					CH_DISCONNECTING, CH_DRAINING);
2306 			break;
2307 		}
2308 	}
2309 	spin_unlock_irq(&sdev->spinlock);
2310 
2311 	if (do_reset) {
2312 		ret = srpt_ch_qp_err(ch);
2313 		if (ret < 0)
2314 			printk(KERN_ERR "Setting queue pair in error state"
2315 			       " failed: %d\n", ret);
2316 	}
2317 }
2318 
2319 /**
2320  * srpt_find_channel() - Look up an RDMA channel.
2321  * @cm_id: Pointer to the CM ID of the channel to be looked up.
2322  *
2323  * Return NULL if no matching RDMA channel has been found.
2324  */
2325 static struct srpt_rdma_ch *srpt_find_channel(struct srpt_device *sdev,
2326 					      struct ib_cm_id *cm_id)
2327 {
2328 	struct srpt_rdma_ch *ch;
2329 	bool found;
2330 
2331 	WARN_ON_ONCE(irqs_disabled());
2332 	BUG_ON(!sdev);
2333 
2334 	found = false;
2335 	spin_lock_irq(&sdev->spinlock);
2336 	list_for_each_entry(ch, &sdev->rch_list, list) {
2337 		if (ch->cm_id == cm_id) {
2338 			found = true;
2339 			break;
2340 		}
2341 	}
2342 	spin_unlock_irq(&sdev->spinlock);
2343 
2344 	return found ? ch : NULL;
2345 }
2346 
2347 /**
2348  * srpt_release_channel() - Release channel resources.
2349  *
2350  * Schedules the actual release because:
2351  * - Calling the ib_destroy_cm_id() call from inside an IB CM callback would
2352  *   trigger a deadlock.
2353  * - It is not safe to call TCM transport_* functions from interrupt context.
2354  */
2355 static void srpt_release_channel(struct srpt_rdma_ch *ch)
2356 {
2357 	schedule_work(&ch->release_work);
2358 }
2359 
2360 static void srpt_release_channel_work(struct work_struct *w)
2361 {
2362 	struct srpt_rdma_ch *ch;
2363 	struct srpt_device *sdev;
2364 
2365 	ch = container_of(w, struct srpt_rdma_ch, release_work);
2366 	pr_debug("ch = %p; ch->sess = %p; release_done = %p\n", ch, ch->sess,
2367 		 ch->release_done);
2368 
2369 	sdev = ch->sport->sdev;
2370 	BUG_ON(!sdev);
2371 
2372 	transport_deregister_session_configfs(ch->sess);
2373 	transport_deregister_session(ch->sess);
2374 	ch->sess = NULL;
2375 
2376 	srpt_destroy_ch_ib(ch);
2377 
2378 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
2379 			     ch->sport->sdev, ch->rq_size,
2380 			     ch->rsp_size, DMA_TO_DEVICE);
2381 
2382 	spin_lock_irq(&sdev->spinlock);
2383 	list_del(&ch->list);
2384 	spin_unlock_irq(&sdev->spinlock);
2385 
2386 	ib_destroy_cm_id(ch->cm_id);
2387 
2388 	if (ch->release_done)
2389 		complete(ch->release_done);
2390 
2391 	wake_up(&sdev->ch_releaseQ);
2392 
2393 	kfree(ch);
2394 }
2395 
2396 static struct srpt_node_acl *__srpt_lookup_acl(struct srpt_port *sport,
2397 					       u8 i_port_id[16])
2398 {
2399 	struct srpt_node_acl *nacl;
2400 
2401 	list_for_each_entry(nacl, &sport->port_acl_list, list)
2402 		if (memcmp(nacl->i_port_id, i_port_id,
2403 			   sizeof(nacl->i_port_id)) == 0)
2404 			return nacl;
2405 
2406 	return NULL;
2407 }
2408 
2409 static struct srpt_node_acl *srpt_lookup_acl(struct srpt_port *sport,
2410 					     u8 i_port_id[16])
2411 {
2412 	struct srpt_node_acl *nacl;
2413 
2414 	spin_lock_irq(&sport->port_acl_lock);
2415 	nacl = __srpt_lookup_acl(sport, i_port_id);
2416 	spin_unlock_irq(&sport->port_acl_lock);
2417 
2418 	return nacl;
2419 }
2420 
2421 /**
2422  * srpt_cm_req_recv() - Process the event IB_CM_REQ_RECEIVED.
2423  *
2424  * Ownership of the cm_id is transferred to the target session if this
2425  * functions returns zero. Otherwise the caller remains the owner of cm_id.
2426  */
2427 static int srpt_cm_req_recv(struct ib_cm_id *cm_id,
2428 			    struct ib_cm_req_event_param *param,
2429 			    void *private_data)
2430 {
2431 	struct srpt_device *sdev = cm_id->context;
2432 	struct srpt_port *sport = &sdev->port[param->port - 1];
2433 	struct srp_login_req *req;
2434 	struct srp_login_rsp *rsp;
2435 	struct srp_login_rej *rej;
2436 	struct ib_cm_rep_param *rep_param;
2437 	struct srpt_rdma_ch *ch, *tmp_ch;
2438 	struct srpt_node_acl *nacl;
2439 	u32 it_iu_len;
2440 	int i;
2441 	int ret = 0;
2442 
2443 	WARN_ON_ONCE(irqs_disabled());
2444 
2445 	if (WARN_ON(!sdev || !private_data))
2446 		return -EINVAL;
2447 
2448 	req = (struct srp_login_req *)private_data;
2449 
2450 	it_iu_len = be32_to_cpu(req->req_it_iu_len);
2451 
2452 	printk(KERN_INFO "Received SRP_LOGIN_REQ with i_port_id 0x%llx:0x%llx,"
2453 	       " t_port_id 0x%llx:0x%llx and it_iu_len %d on port %d"
2454 	       " (guid=0x%llx:0x%llx)\n",
2455 	       be64_to_cpu(*(__be64 *)&req->initiator_port_id[0]),
2456 	       be64_to_cpu(*(__be64 *)&req->initiator_port_id[8]),
2457 	       be64_to_cpu(*(__be64 *)&req->target_port_id[0]),
2458 	       be64_to_cpu(*(__be64 *)&req->target_port_id[8]),
2459 	       it_iu_len,
2460 	       param->port,
2461 	       be64_to_cpu(*(__be64 *)&sdev->port[param->port - 1].gid.raw[0]),
2462 	       be64_to_cpu(*(__be64 *)&sdev->port[param->port - 1].gid.raw[8]));
2463 
2464 	rsp = kzalloc(sizeof *rsp, GFP_KERNEL);
2465 	rej = kzalloc(sizeof *rej, GFP_KERNEL);
2466 	rep_param = kzalloc(sizeof *rep_param, GFP_KERNEL);
2467 
2468 	if (!rsp || !rej || !rep_param) {
2469 		ret = -ENOMEM;
2470 		goto out;
2471 	}
2472 
2473 	if (it_iu_len > srp_max_req_size || it_iu_len < 64) {
2474 		rej->reason = __constant_cpu_to_be32(
2475 				SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE);
2476 		ret = -EINVAL;
2477 		printk(KERN_ERR "rejected SRP_LOGIN_REQ because its"
2478 		       " length (%d bytes) is out of range (%d .. %d)\n",
2479 		       it_iu_len, 64, srp_max_req_size);
2480 		goto reject;
2481 	}
2482 
2483 	if (!sport->enabled) {
2484 		rej->reason = __constant_cpu_to_be32(
2485 			     SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2486 		ret = -EINVAL;
2487 		printk(KERN_ERR "rejected SRP_LOGIN_REQ because the target port"
2488 		       " has not yet been enabled\n");
2489 		goto reject;
2490 	}
2491 
2492 	if ((req->req_flags & SRP_MTCH_ACTION) == SRP_MULTICHAN_SINGLE) {
2493 		rsp->rsp_flags = SRP_LOGIN_RSP_MULTICHAN_NO_CHAN;
2494 
2495 		spin_lock_irq(&sdev->spinlock);
2496 
2497 		list_for_each_entry_safe(ch, tmp_ch, &sdev->rch_list, list) {
2498 			if (!memcmp(ch->i_port_id, req->initiator_port_id, 16)
2499 			    && !memcmp(ch->t_port_id, req->target_port_id, 16)
2500 			    && param->port == ch->sport->port
2501 			    && param->listen_id == ch->sport->sdev->cm_id
2502 			    && ch->cm_id) {
2503 				enum rdma_ch_state ch_state;
2504 
2505 				ch_state = srpt_get_ch_state(ch);
2506 				if (ch_state != CH_CONNECTING
2507 				    && ch_state != CH_LIVE)
2508 					continue;
2509 
2510 				/* found an existing channel */
2511 				pr_debug("Found existing channel %s"
2512 					 " cm_id= %p state= %d\n",
2513 					 ch->sess_name, ch->cm_id, ch_state);
2514 
2515 				__srpt_close_ch(ch);
2516 
2517 				rsp->rsp_flags =
2518 					SRP_LOGIN_RSP_MULTICHAN_TERMINATED;
2519 			}
2520 		}
2521 
2522 		spin_unlock_irq(&sdev->spinlock);
2523 
2524 	} else
2525 		rsp->rsp_flags = SRP_LOGIN_RSP_MULTICHAN_MAINTAINED;
2526 
2527 	if (*(__be64 *)req->target_port_id != cpu_to_be64(srpt_service_guid)
2528 	    || *(__be64 *)(req->target_port_id + 8) !=
2529 	       cpu_to_be64(srpt_service_guid)) {
2530 		rej->reason = __constant_cpu_to_be32(
2531 				SRP_LOGIN_REJ_UNABLE_ASSOCIATE_CHANNEL);
2532 		ret = -ENOMEM;
2533 		printk(KERN_ERR "rejected SRP_LOGIN_REQ because it"
2534 		       " has an invalid target port identifier.\n");
2535 		goto reject;
2536 	}
2537 
2538 	ch = kzalloc(sizeof *ch, GFP_KERNEL);
2539 	if (!ch) {
2540 		rej->reason = __constant_cpu_to_be32(
2541 					SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2542 		printk(KERN_ERR "rejected SRP_LOGIN_REQ because no memory.\n");
2543 		ret = -ENOMEM;
2544 		goto reject;
2545 	}
2546 
2547 	INIT_WORK(&ch->release_work, srpt_release_channel_work);
2548 	memcpy(ch->i_port_id, req->initiator_port_id, 16);
2549 	memcpy(ch->t_port_id, req->target_port_id, 16);
2550 	ch->sport = &sdev->port[param->port - 1];
2551 	ch->cm_id = cm_id;
2552 	/*
2553 	 * Avoid QUEUE_FULL conditions by limiting the number of buffers used
2554 	 * for the SRP protocol to the command queue size.
2555 	 */
2556 	ch->rq_size = SRPT_RQ_SIZE;
2557 	spin_lock_init(&ch->spinlock);
2558 	ch->state = CH_CONNECTING;
2559 	INIT_LIST_HEAD(&ch->cmd_wait_list);
2560 	ch->rsp_size = ch->sport->port_attrib.srp_max_rsp_size;
2561 
2562 	ch->ioctx_ring = (struct srpt_send_ioctx **)
2563 		srpt_alloc_ioctx_ring(ch->sport->sdev, ch->rq_size,
2564 				      sizeof(*ch->ioctx_ring[0]),
2565 				      ch->rsp_size, DMA_TO_DEVICE);
2566 	if (!ch->ioctx_ring)
2567 		goto free_ch;
2568 
2569 	INIT_LIST_HEAD(&ch->free_list);
2570 	for (i = 0; i < ch->rq_size; i++) {
2571 		ch->ioctx_ring[i]->ch = ch;
2572 		list_add_tail(&ch->ioctx_ring[i]->free_list, &ch->free_list);
2573 	}
2574 
2575 	ret = srpt_create_ch_ib(ch);
2576 	if (ret) {
2577 		rej->reason = __constant_cpu_to_be32(
2578 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2579 		printk(KERN_ERR "rejected SRP_LOGIN_REQ because creating"
2580 		       " a new RDMA channel failed.\n");
2581 		goto free_ring;
2582 	}
2583 
2584 	ret = srpt_ch_qp_rtr(ch, ch->qp);
2585 	if (ret) {
2586 		rej->reason = __constant_cpu_to_be32(
2587 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2588 		printk(KERN_ERR "rejected SRP_LOGIN_REQ because enabling"
2589 		       " RTR failed (error code = %d)\n", ret);
2590 		goto destroy_ib;
2591 	}
2592 	/*
2593 	 * Use the initator port identifier as the session name.
2594 	 */
2595 	snprintf(ch->sess_name, sizeof(ch->sess_name), "0x%016llx%016llx",
2596 			be64_to_cpu(*(__be64 *)ch->i_port_id),
2597 			be64_to_cpu(*(__be64 *)(ch->i_port_id + 8)));
2598 
2599 	pr_debug("registering session %s\n", ch->sess_name);
2600 
2601 	nacl = srpt_lookup_acl(sport, ch->i_port_id);
2602 	if (!nacl) {
2603 		printk(KERN_INFO "Rejected login because no ACL has been"
2604 		       " configured yet for initiator %s.\n", ch->sess_name);
2605 		rej->reason = __constant_cpu_to_be32(
2606 				SRP_LOGIN_REJ_CHANNEL_LIMIT_REACHED);
2607 		goto destroy_ib;
2608 	}
2609 
2610 	ch->sess = transport_init_session();
2611 	if (IS_ERR(ch->sess)) {
2612 		rej->reason = __constant_cpu_to_be32(
2613 				SRP_LOGIN_REJ_INSUFFICIENT_RESOURCES);
2614 		pr_debug("Failed to create session\n");
2615 		goto deregister_session;
2616 	}
2617 	ch->sess->se_node_acl = &nacl->nacl;
2618 	transport_register_session(&sport->port_tpg_1, &nacl->nacl, ch->sess, ch);
2619 
2620 	pr_debug("Establish connection sess=%p name=%s cm_id=%p\n", ch->sess,
2621 		 ch->sess_name, ch->cm_id);
2622 
2623 	/* create srp_login_response */
2624 	rsp->opcode = SRP_LOGIN_RSP;
2625 	rsp->tag = req->tag;
2626 	rsp->max_it_iu_len = req->req_it_iu_len;
2627 	rsp->max_ti_iu_len = req->req_it_iu_len;
2628 	ch->max_ti_iu_len = it_iu_len;
2629 	rsp->buf_fmt = __constant_cpu_to_be16(SRP_BUF_FORMAT_DIRECT
2630 					      | SRP_BUF_FORMAT_INDIRECT);
2631 	rsp->req_lim_delta = cpu_to_be32(ch->rq_size);
2632 	atomic_set(&ch->req_lim, ch->rq_size);
2633 	atomic_set(&ch->req_lim_delta, 0);
2634 
2635 	/* create cm reply */
2636 	rep_param->qp_num = ch->qp->qp_num;
2637 	rep_param->private_data = (void *)rsp;
2638 	rep_param->private_data_len = sizeof *rsp;
2639 	rep_param->rnr_retry_count = 7;
2640 	rep_param->flow_control = 1;
2641 	rep_param->failover_accepted = 0;
2642 	rep_param->srq = 1;
2643 	rep_param->responder_resources = 4;
2644 	rep_param->initiator_depth = 4;
2645 
2646 	ret = ib_send_cm_rep(cm_id, rep_param);
2647 	if (ret) {
2648 		printk(KERN_ERR "sending SRP_LOGIN_REQ response failed"
2649 		       " (error code = %d)\n", ret);
2650 		goto release_channel;
2651 	}
2652 
2653 	spin_lock_irq(&sdev->spinlock);
2654 	list_add_tail(&ch->list, &sdev->rch_list);
2655 	spin_unlock_irq(&sdev->spinlock);
2656 
2657 	goto out;
2658 
2659 release_channel:
2660 	srpt_set_ch_state(ch, CH_RELEASING);
2661 	transport_deregister_session_configfs(ch->sess);
2662 
2663 deregister_session:
2664 	transport_deregister_session(ch->sess);
2665 	ch->sess = NULL;
2666 
2667 destroy_ib:
2668 	srpt_destroy_ch_ib(ch);
2669 
2670 free_ring:
2671 	srpt_free_ioctx_ring((struct srpt_ioctx **)ch->ioctx_ring,
2672 			     ch->sport->sdev, ch->rq_size,
2673 			     ch->rsp_size, DMA_TO_DEVICE);
2674 free_ch:
2675 	kfree(ch);
2676 
2677 reject:
2678 	rej->opcode = SRP_LOGIN_REJ;
2679 	rej->tag = req->tag;
2680 	rej->buf_fmt = __constant_cpu_to_be16(SRP_BUF_FORMAT_DIRECT
2681 					      | SRP_BUF_FORMAT_INDIRECT);
2682 
2683 	ib_send_cm_rej(cm_id, IB_CM_REJ_CONSUMER_DEFINED, NULL, 0,
2684 			     (void *)rej, sizeof *rej);
2685 
2686 out:
2687 	kfree(rep_param);
2688 	kfree(rsp);
2689 	kfree(rej);
2690 
2691 	return ret;
2692 }
2693 
2694 static void srpt_cm_rej_recv(struct ib_cm_id *cm_id)
2695 {
2696 	printk(KERN_INFO "Received IB REJ for cm_id %p.\n", cm_id);
2697 	srpt_drain_channel(cm_id);
2698 }
2699 
2700 /**
2701  * srpt_cm_rtu_recv() - Process an IB_CM_RTU_RECEIVED or USER_ESTABLISHED event.
2702  *
2703  * An IB_CM_RTU_RECEIVED message indicates that the connection is established
2704  * and that the recipient may begin transmitting (RTU = ready to use).
2705  */
2706 static void srpt_cm_rtu_recv(struct ib_cm_id *cm_id)
2707 {
2708 	struct srpt_rdma_ch *ch;
2709 	int ret;
2710 
2711 	ch = srpt_find_channel(cm_id->context, cm_id);
2712 	BUG_ON(!ch);
2713 
2714 	if (srpt_test_and_set_ch_state(ch, CH_CONNECTING, CH_LIVE)) {
2715 		struct srpt_recv_ioctx *ioctx, *ioctx_tmp;
2716 
2717 		ret = srpt_ch_qp_rts(ch, ch->qp);
2718 
2719 		list_for_each_entry_safe(ioctx, ioctx_tmp, &ch->cmd_wait_list,
2720 					 wait_list) {
2721 			list_del(&ioctx->wait_list);
2722 			srpt_handle_new_iu(ch, ioctx, NULL);
2723 		}
2724 		if (ret)
2725 			srpt_close_ch(ch);
2726 	}
2727 }
2728 
2729 static void srpt_cm_timewait_exit(struct ib_cm_id *cm_id)
2730 {
2731 	printk(KERN_INFO "Received IB TimeWait exit for cm_id %p.\n", cm_id);
2732 	srpt_drain_channel(cm_id);
2733 }
2734 
2735 static void srpt_cm_rep_error(struct ib_cm_id *cm_id)
2736 {
2737 	printk(KERN_INFO "Received IB REP error for cm_id %p.\n", cm_id);
2738 	srpt_drain_channel(cm_id);
2739 }
2740 
2741 /**
2742  * srpt_cm_dreq_recv() - Process reception of a DREQ message.
2743  */
2744 static void srpt_cm_dreq_recv(struct ib_cm_id *cm_id)
2745 {
2746 	struct srpt_rdma_ch *ch;
2747 	unsigned long flags;
2748 	bool send_drep = false;
2749 
2750 	ch = srpt_find_channel(cm_id->context, cm_id);
2751 	BUG_ON(!ch);
2752 
2753 	pr_debug("cm_id= %p ch->state= %d\n", cm_id, srpt_get_ch_state(ch));
2754 
2755 	spin_lock_irqsave(&ch->spinlock, flags);
2756 	switch (ch->state) {
2757 	case CH_CONNECTING:
2758 	case CH_LIVE:
2759 		send_drep = true;
2760 		ch->state = CH_DISCONNECTING;
2761 		break;
2762 	case CH_DISCONNECTING:
2763 	case CH_DRAINING:
2764 	case CH_RELEASING:
2765 		WARN(true, "unexpected channel state %d\n", ch->state);
2766 		break;
2767 	}
2768 	spin_unlock_irqrestore(&ch->spinlock, flags);
2769 
2770 	if (send_drep) {
2771 		if (ib_send_cm_drep(ch->cm_id, NULL, 0) < 0)
2772 			printk(KERN_ERR "Sending IB DREP failed.\n");
2773 		printk(KERN_INFO "Received DREQ and sent DREP for session %s.\n",
2774 		       ch->sess_name);
2775 	}
2776 }
2777 
2778 /**
2779  * srpt_cm_drep_recv() - Process reception of a DREP message.
2780  */
2781 static void srpt_cm_drep_recv(struct ib_cm_id *cm_id)
2782 {
2783 	printk(KERN_INFO "Received InfiniBand DREP message for cm_id %p.\n",
2784 	       cm_id);
2785 	srpt_drain_channel(cm_id);
2786 }
2787 
2788 /**
2789  * srpt_cm_handler() - IB connection manager callback function.
2790  *
2791  * A non-zero return value will cause the caller destroy the CM ID.
2792  *
2793  * Note: srpt_cm_handler() must only return a non-zero value when transferring
2794  * ownership of the cm_id to a channel by srpt_cm_req_recv() failed. Returning
2795  * a non-zero value in any other case will trigger a race with the
2796  * ib_destroy_cm_id() call in srpt_release_channel().
2797  */
2798 static int srpt_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
2799 {
2800 	int ret;
2801 
2802 	ret = 0;
2803 	switch (event->event) {
2804 	case IB_CM_REQ_RECEIVED:
2805 		ret = srpt_cm_req_recv(cm_id, &event->param.req_rcvd,
2806 				       event->private_data);
2807 		break;
2808 	case IB_CM_REJ_RECEIVED:
2809 		srpt_cm_rej_recv(cm_id);
2810 		break;
2811 	case IB_CM_RTU_RECEIVED:
2812 	case IB_CM_USER_ESTABLISHED:
2813 		srpt_cm_rtu_recv(cm_id);
2814 		break;
2815 	case IB_CM_DREQ_RECEIVED:
2816 		srpt_cm_dreq_recv(cm_id);
2817 		break;
2818 	case IB_CM_DREP_RECEIVED:
2819 		srpt_cm_drep_recv(cm_id);
2820 		break;
2821 	case IB_CM_TIMEWAIT_EXIT:
2822 		srpt_cm_timewait_exit(cm_id);
2823 		break;
2824 	case IB_CM_REP_ERROR:
2825 		srpt_cm_rep_error(cm_id);
2826 		break;
2827 	case IB_CM_DREQ_ERROR:
2828 		printk(KERN_INFO "Received IB DREQ ERROR event.\n");
2829 		break;
2830 	case IB_CM_MRA_RECEIVED:
2831 		printk(KERN_INFO "Received IB MRA event\n");
2832 		break;
2833 	default:
2834 		printk(KERN_ERR "received unrecognized IB CM event %d\n",
2835 		       event->event);
2836 		break;
2837 	}
2838 
2839 	return ret;
2840 }
2841 
2842 /**
2843  * srpt_perform_rdmas() - Perform IB RDMA.
2844  *
2845  * Returns zero upon success or a negative number upon failure.
2846  */
2847 static int srpt_perform_rdmas(struct srpt_rdma_ch *ch,
2848 			      struct srpt_send_ioctx *ioctx)
2849 {
2850 	struct ib_send_wr wr;
2851 	struct ib_send_wr *bad_wr;
2852 	struct rdma_iu *riu;
2853 	int i;
2854 	int ret;
2855 	int sq_wr_avail;
2856 	enum dma_data_direction dir;
2857 	const int n_rdma = ioctx->n_rdma;
2858 
2859 	dir = ioctx->cmd.data_direction;
2860 	if (dir == DMA_TO_DEVICE) {
2861 		/* write */
2862 		ret = -ENOMEM;
2863 		sq_wr_avail = atomic_sub_return(n_rdma, &ch->sq_wr_avail);
2864 		if (sq_wr_avail < 0) {
2865 			printk(KERN_WARNING "IB send queue full (needed %d)\n",
2866 			       n_rdma);
2867 			goto out;
2868 		}
2869 	}
2870 
2871 	ioctx->rdma_aborted = false;
2872 	ret = 0;
2873 	riu = ioctx->rdma_ius;
2874 	memset(&wr, 0, sizeof wr);
2875 
2876 	for (i = 0; i < n_rdma; ++i, ++riu) {
2877 		if (dir == DMA_FROM_DEVICE) {
2878 			wr.opcode = IB_WR_RDMA_WRITE;
2879 			wr.wr_id = encode_wr_id(i == n_rdma - 1 ?
2880 						SRPT_RDMA_WRITE_LAST :
2881 						SRPT_RDMA_MID,
2882 						ioctx->ioctx.index);
2883 		} else {
2884 			wr.opcode = IB_WR_RDMA_READ;
2885 			wr.wr_id = encode_wr_id(i == n_rdma - 1 ?
2886 						SRPT_RDMA_READ_LAST :
2887 						SRPT_RDMA_MID,
2888 						ioctx->ioctx.index);
2889 		}
2890 		wr.next = NULL;
2891 		wr.wr.rdma.remote_addr = riu->raddr;
2892 		wr.wr.rdma.rkey = riu->rkey;
2893 		wr.num_sge = riu->sge_cnt;
2894 		wr.sg_list = riu->sge;
2895 
2896 		/* only get completion event for the last rdma write */
2897 		if (i == (n_rdma - 1) && dir == DMA_TO_DEVICE)
2898 			wr.send_flags = IB_SEND_SIGNALED;
2899 
2900 		ret = ib_post_send(ch->qp, &wr, &bad_wr);
2901 		if (ret)
2902 			break;
2903 	}
2904 
2905 	if (ret)
2906 		printk(KERN_ERR "%s[%d]: ib_post_send() returned %d for %d/%d",
2907 				 __func__, __LINE__, ret, i, n_rdma);
2908 	if (ret && i > 0) {
2909 		wr.num_sge = 0;
2910 		wr.wr_id = encode_wr_id(SRPT_RDMA_ABORT, ioctx->ioctx.index);
2911 		wr.send_flags = IB_SEND_SIGNALED;
2912 		while (ch->state == CH_LIVE &&
2913 			ib_post_send(ch->qp, &wr, &bad_wr) != 0) {
2914 			printk(KERN_INFO "Trying to abort failed RDMA transfer [%d]",
2915 				ioctx->ioctx.index);
2916 			msleep(1000);
2917 		}
2918 		while (ch->state != CH_RELEASING && !ioctx->rdma_aborted) {
2919 			printk(KERN_INFO "Waiting until RDMA abort finished [%d]",
2920 				ioctx->ioctx.index);
2921 			msleep(1000);
2922 		}
2923 	}
2924 out:
2925 	if (unlikely(dir == DMA_TO_DEVICE && ret < 0))
2926 		atomic_add(n_rdma, &ch->sq_wr_avail);
2927 	return ret;
2928 }
2929 
2930 /**
2931  * srpt_xfer_data() - Start data transfer from initiator to target.
2932  */
2933 static int srpt_xfer_data(struct srpt_rdma_ch *ch,
2934 			  struct srpt_send_ioctx *ioctx)
2935 {
2936 	int ret;
2937 
2938 	ret = srpt_map_sg_to_ib_sge(ch, ioctx);
2939 	if (ret) {
2940 		printk(KERN_ERR "%s[%d] ret=%d\n", __func__, __LINE__, ret);
2941 		goto out;
2942 	}
2943 
2944 	ret = srpt_perform_rdmas(ch, ioctx);
2945 	if (ret) {
2946 		if (ret == -EAGAIN || ret == -ENOMEM)
2947 			printk(KERN_INFO "%s[%d] queue full -- ret=%d\n",
2948 				   __func__, __LINE__, ret);
2949 		else
2950 			printk(KERN_ERR "%s[%d] fatal error -- ret=%d\n",
2951 			       __func__, __LINE__, ret);
2952 		goto out_unmap;
2953 	}
2954 
2955 out:
2956 	return ret;
2957 out_unmap:
2958 	srpt_unmap_sg_to_ib_sge(ch, ioctx);
2959 	goto out;
2960 }
2961 
2962 static int srpt_write_pending_status(struct se_cmd *se_cmd)
2963 {
2964 	struct srpt_send_ioctx *ioctx;
2965 
2966 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
2967 	return srpt_get_cmd_state(ioctx) == SRPT_STATE_NEED_DATA;
2968 }
2969 
2970 /*
2971  * srpt_write_pending() - Start data transfer from initiator to target (write).
2972  */
2973 static int srpt_write_pending(struct se_cmd *se_cmd)
2974 {
2975 	struct srpt_rdma_ch *ch;
2976 	struct srpt_send_ioctx *ioctx;
2977 	enum srpt_command_state new_state;
2978 	enum rdma_ch_state ch_state;
2979 	int ret;
2980 
2981 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
2982 
2983 	new_state = srpt_set_cmd_state(ioctx, SRPT_STATE_NEED_DATA);
2984 	WARN_ON(new_state == SRPT_STATE_DONE);
2985 
2986 	ch = ioctx->ch;
2987 	BUG_ON(!ch);
2988 
2989 	ch_state = srpt_get_ch_state(ch);
2990 	switch (ch_state) {
2991 	case CH_CONNECTING:
2992 		WARN(true, "unexpected channel state %d\n", ch_state);
2993 		ret = -EINVAL;
2994 		goto out;
2995 	case CH_LIVE:
2996 		break;
2997 	case CH_DISCONNECTING:
2998 	case CH_DRAINING:
2999 	case CH_RELEASING:
3000 		pr_debug("cmd with tag %lld: channel disconnecting\n",
3001 			 ioctx->tag);
3002 		srpt_set_cmd_state(ioctx, SRPT_STATE_DATA_IN);
3003 		ret = -EINVAL;
3004 		goto out;
3005 	}
3006 	ret = srpt_xfer_data(ch, ioctx);
3007 
3008 out:
3009 	return ret;
3010 }
3011 
3012 static u8 tcm_to_srp_tsk_mgmt_status(const int tcm_mgmt_status)
3013 {
3014 	switch (tcm_mgmt_status) {
3015 	case TMR_FUNCTION_COMPLETE:
3016 		return SRP_TSK_MGMT_SUCCESS;
3017 	case TMR_FUNCTION_REJECTED:
3018 		return SRP_TSK_MGMT_FUNC_NOT_SUPP;
3019 	}
3020 	return SRP_TSK_MGMT_FAILED;
3021 }
3022 
3023 /**
3024  * srpt_queue_response() - Transmits the response to a SCSI command.
3025  *
3026  * Callback function called by the TCM core. Must not block since it can be
3027  * invoked on the context of the IB completion handler.
3028  */
3029 static int srpt_queue_response(struct se_cmd *cmd)
3030 {
3031 	struct srpt_rdma_ch *ch;
3032 	struct srpt_send_ioctx *ioctx;
3033 	enum srpt_command_state state;
3034 	unsigned long flags;
3035 	int ret;
3036 	enum dma_data_direction dir;
3037 	int resp_len;
3038 	u8 srp_tm_status;
3039 
3040 	ret = 0;
3041 
3042 	ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
3043 	ch = ioctx->ch;
3044 	BUG_ON(!ch);
3045 
3046 	spin_lock_irqsave(&ioctx->spinlock, flags);
3047 	state = ioctx->state;
3048 	switch (state) {
3049 	case SRPT_STATE_NEW:
3050 	case SRPT_STATE_DATA_IN:
3051 		ioctx->state = SRPT_STATE_CMD_RSP_SENT;
3052 		break;
3053 	case SRPT_STATE_MGMT:
3054 		ioctx->state = SRPT_STATE_MGMT_RSP_SENT;
3055 		break;
3056 	default:
3057 		WARN(true, "ch %p; cmd %d: unexpected command state %d\n",
3058 			ch, ioctx->ioctx.index, ioctx->state);
3059 		break;
3060 	}
3061 	spin_unlock_irqrestore(&ioctx->spinlock, flags);
3062 
3063 	if (unlikely(transport_check_aborted_status(&ioctx->cmd, false)
3064 		     || WARN_ON_ONCE(state == SRPT_STATE_CMD_RSP_SENT))) {
3065 		atomic_inc(&ch->req_lim_delta);
3066 		srpt_abort_cmd(ioctx);
3067 		goto out;
3068 	}
3069 
3070 	dir = ioctx->cmd.data_direction;
3071 
3072 	/* For read commands, transfer the data to the initiator. */
3073 	if (dir == DMA_FROM_DEVICE && ioctx->cmd.data_length &&
3074 	    !ioctx->queue_status_only) {
3075 		ret = srpt_xfer_data(ch, ioctx);
3076 		if (ret) {
3077 			printk(KERN_ERR "xfer_data failed for tag %llu\n",
3078 			       ioctx->tag);
3079 			goto out;
3080 		}
3081 	}
3082 
3083 	if (state != SRPT_STATE_MGMT)
3084 		resp_len = srpt_build_cmd_rsp(ch, ioctx, ioctx->tag,
3085 					      cmd->scsi_status);
3086 	else {
3087 		srp_tm_status
3088 			= tcm_to_srp_tsk_mgmt_status(cmd->se_tmr_req->response);
3089 		resp_len = srpt_build_tskmgmt_rsp(ch, ioctx, srp_tm_status,
3090 						 ioctx->tag);
3091 	}
3092 	ret = srpt_post_send(ch, ioctx, resp_len);
3093 	if (ret) {
3094 		printk(KERN_ERR "sending cmd response failed for tag %llu\n",
3095 		       ioctx->tag);
3096 		srpt_unmap_sg_to_ib_sge(ch, ioctx);
3097 		srpt_set_cmd_state(ioctx, SRPT_STATE_DONE);
3098 		kref_put(&ioctx->kref, srpt_put_send_ioctx_kref);
3099 	}
3100 
3101 out:
3102 	return ret;
3103 }
3104 
3105 static int srpt_queue_status(struct se_cmd *cmd)
3106 {
3107 	struct srpt_send_ioctx *ioctx;
3108 
3109 	ioctx = container_of(cmd, struct srpt_send_ioctx, cmd);
3110 	BUG_ON(ioctx->sense_data != cmd->sense_buffer);
3111 	if (cmd->se_cmd_flags &
3112 	    (SCF_TRANSPORT_TASK_SENSE | SCF_EMULATED_TASK_SENSE))
3113 		WARN_ON(cmd->scsi_status != SAM_STAT_CHECK_CONDITION);
3114 	ioctx->queue_status_only = true;
3115 	return srpt_queue_response(cmd);
3116 }
3117 
3118 static void srpt_refresh_port_work(struct work_struct *work)
3119 {
3120 	struct srpt_port *sport = container_of(work, struct srpt_port, work);
3121 
3122 	srpt_refresh_port(sport);
3123 }
3124 
3125 static int srpt_ch_list_empty(struct srpt_device *sdev)
3126 {
3127 	int res;
3128 
3129 	spin_lock_irq(&sdev->spinlock);
3130 	res = list_empty(&sdev->rch_list);
3131 	spin_unlock_irq(&sdev->spinlock);
3132 
3133 	return res;
3134 }
3135 
3136 /**
3137  * srpt_release_sdev() - Free the channel resources associated with a target.
3138  */
3139 static int srpt_release_sdev(struct srpt_device *sdev)
3140 {
3141 	struct srpt_rdma_ch *ch, *tmp_ch;
3142 	int res;
3143 
3144 	WARN_ON_ONCE(irqs_disabled());
3145 
3146 	BUG_ON(!sdev);
3147 
3148 	spin_lock_irq(&sdev->spinlock);
3149 	list_for_each_entry_safe(ch, tmp_ch, &sdev->rch_list, list)
3150 		__srpt_close_ch(ch);
3151 	spin_unlock_irq(&sdev->spinlock);
3152 
3153 	res = wait_event_interruptible(sdev->ch_releaseQ,
3154 				       srpt_ch_list_empty(sdev));
3155 	if (res)
3156 		printk(KERN_ERR "%s: interrupted.\n", __func__);
3157 
3158 	return 0;
3159 }
3160 
3161 static struct srpt_port *__srpt_lookup_port(const char *name)
3162 {
3163 	struct ib_device *dev;
3164 	struct srpt_device *sdev;
3165 	struct srpt_port *sport;
3166 	int i;
3167 
3168 	list_for_each_entry(sdev, &srpt_dev_list, list) {
3169 		dev = sdev->device;
3170 		if (!dev)
3171 			continue;
3172 
3173 		for (i = 0; i < dev->phys_port_cnt; i++) {
3174 			sport = &sdev->port[i];
3175 
3176 			if (!strcmp(sport->port_guid, name))
3177 				return sport;
3178 		}
3179 	}
3180 
3181 	return NULL;
3182 }
3183 
3184 static struct srpt_port *srpt_lookup_port(const char *name)
3185 {
3186 	struct srpt_port *sport;
3187 
3188 	spin_lock(&srpt_dev_lock);
3189 	sport = __srpt_lookup_port(name);
3190 	spin_unlock(&srpt_dev_lock);
3191 
3192 	return sport;
3193 }
3194 
3195 /**
3196  * srpt_add_one() - Infiniband device addition callback function.
3197  */
3198 static void srpt_add_one(struct ib_device *device)
3199 {
3200 	struct srpt_device *sdev;
3201 	struct srpt_port *sport;
3202 	struct ib_srq_init_attr srq_attr;
3203 	int i;
3204 
3205 	pr_debug("device = %p, device->dma_ops = %p\n", device,
3206 		 device->dma_ops);
3207 
3208 	sdev = kzalloc(sizeof *sdev, GFP_KERNEL);
3209 	if (!sdev)
3210 		goto err;
3211 
3212 	sdev->device = device;
3213 	INIT_LIST_HEAD(&sdev->rch_list);
3214 	init_waitqueue_head(&sdev->ch_releaseQ);
3215 	spin_lock_init(&sdev->spinlock);
3216 
3217 	if (ib_query_device(device, &sdev->dev_attr))
3218 		goto free_dev;
3219 
3220 	sdev->pd = ib_alloc_pd(device);
3221 	if (IS_ERR(sdev->pd))
3222 		goto free_dev;
3223 
3224 	sdev->mr = ib_get_dma_mr(sdev->pd, IB_ACCESS_LOCAL_WRITE);
3225 	if (IS_ERR(sdev->mr))
3226 		goto err_pd;
3227 
3228 	sdev->srq_size = min(srpt_srq_size, sdev->dev_attr.max_srq_wr);
3229 
3230 	srq_attr.event_handler = srpt_srq_event;
3231 	srq_attr.srq_context = (void *)sdev;
3232 	srq_attr.attr.max_wr = sdev->srq_size;
3233 	srq_attr.attr.max_sge = 1;
3234 	srq_attr.attr.srq_limit = 0;
3235 	srq_attr.srq_type = IB_SRQT_BASIC;
3236 
3237 	sdev->srq = ib_create_srq(sdev->pd, &srq_attr);
3238 	if (IS_ERR(sdev->srq))
3239 		goto err_mr;
3240 
3241 	pr_debug("%s: create SRQ #wr= %d max_allow=%d dev= %s\n",
3242 		 __func__, sdev->srq_size, sdev->dev_attr.max_srq_wr,
3243 		 device->name);
3244 
3245 	if (!srpt_service_guid)
3246 		srpt_service_guid = be64_to_cpu(device->node_guid);
3247 
3248 	sdev->cm_id = ib_create_cm_id(device, srpt_cm_handler, sdev);
3249 	if (IS_ERR(sdev->cm_id))
3250 		goto err_srq;
3251 
3252 	/* print out target login information */
3253 	pr_debug("Target login info: id_ext=%016llx,ioc_guid=%016llx,"
3254 		 "pkey=ffff,service_id=%016llx\n", srpt_service_guid,
3255 		 srpt_service_guid, srpt_service_guid);
3256 
3257 	/*
3258 	 * We do not have a consistent service_id (ie. also id_ext of target_id)
3259 	 * to identify this target. We currently use the guid of the first HCA
3260 	 * in the system as service_id; therefore, the target_id will change
3261 	 * if this HCA is gone bad and replaced by different HCA
3262 	 */
3263 	if (ib_cm_listen(sdev->cm_id, cpu_to_be64(srpt_service_guid), 0, NULL))
3264 		goto err_cm;
3265 
3266 	INIT_IB_EVENT_HANDLER(&sdev->event_handler, sdev->device,
3267 			      srpt_event_handler);
3268 	if (ib_register_event_handler(&sdev->event_handler))
3269 		goto err_cm;
3270 
3271 	sdev->ioctx_ring = (struct srpt_recv_ioctx **)
3272 		srpt_alloc_ioctx_ring(sdev, sdev->srq_size,
3273 				      sizeof(*sdev->ioctx_ring[0]),
3274 				      srp_max_req_size, DMA_FROM_DEVICE);
3275 	if (!sdev->ioctx_ring)
3276 		goto err_event;
3277 
3278 	for (i = 0; i < sdev->srq_size; ++i)
3279 		srpt_post_recv(sdev, sdev->ioctx_ring[i]);
3280 
3281 	WARN_ON(sdev->device->phys_port_cnt > ARRAY_SIZE(sdev->port));
3282 
3283 	for (i = 1; i <= sdev->device->phys_port_cnt; i++) {
3284 		sport = &sdev->port[i - 1];
3285 		sport->sdev = sdev;
3286 		sport->port = i;
3287 		sport->port_attrib.srp_max_rdma_size = DEFAULT_MAX_RDMA_SIZE;
3288 		sport->port_attrib.srp_max_rsp_size = DEFAULT_MAX_RSP_SIZE;
3289 		sport->port_attrib.srp_sq_size = DEF_SRPT_SQ_SIZE;
3290 		INIT_WORK(&sport->work, srpt_refresh_port_work);
3291 		INIT_LIST_HEAD(&sport->port_acl_list);
3292 		spin_lock_init(&sport->port_acl_lock);
3293 
3294 		if (srpt_refresh_port(sport)) {
3295 			printk(KERN_ERR "MAD registration failed for %s-%d.\n",
3296 			       srpt_sdev_name(sdev), i);
3297 			goto err_ring;
3298 		}
3299 		snprintf(sport->port_guid, sizeof(sport->port_guid),
3300 			"0x%016llx%016llx",
3301 			be64_to_cpu(sport->gid.global.subnet_prefix),
3302 			be64_to_cpu(sport->gid.global.interface_id));
3303 	}
3304 
3305 	spin_lock(&srpt_dev_lock);
3306 	list_add_tail(&sdev->list, &srpt_dev_list);
3307 	spin_unlock(&srpt_dev_lock);
3308 
3309 out:
3310 	ib_set_client_data(device, &srpt_client, sdev);
3311 	pr_debug("added %s.\n", device->name);
3312 	return;
3313 
3314 err_ring:
3315 	srpt_free_ioctx_ring((struct srpt_ioctx **)sdev->ioctx_ring, sdev,
3316 			     sdev->srq_size, srp_max_req_size,
3317 			     DMA_FROM_DEVICE);
3318 err_event:
3319 	ib_unregister_event_handler(&sdev->event_handler);
3320 err_cm:
3321 	ib_destroy_cm_id(sdev->cm_id);
3322 err_srq:
3323 	ib_destroy_srq(sdev->srq);
3324 err_mr:
3325 	ib_dereg_mr(sdev->mr);
3326 err_pd:
3327 	ib_dealloc_pd(sdev->pd);
3328 free_dev:
3329 	kfree(sdev);
3330 err:
3331 	sdev = NULL;
3332 	printk(KERN_INFO "%s(%s) failed.\n", __func__, device->name);
3333 	goto out;
3334 }
3335 
3336 /**
3337  * srpt_remove_one() - InfiniBand device removal callback function.
3338  */
3339 static void srpt_remove_one(struct ib_device *device)
3340 {
3341 	struct srpt_device *sdev;
3342 	int i;
3343 
3344 	sdev = ib_get_client_data(device, &srpt_client);
3345 	if (!sdev) {
3346 		printk(KERN_INFO "%s(%s): nothing to do.\n", __func__,
3347 		       device->name);
3348 		return;
3349 	}
3350 
3351 	srpt_unregister_mad_agent(sdev);
3352 
3353 	ib_unregister_event_handler(&sdev->event_handler);
3354 
3355 	/* Cancel any work queued by the just unregistered IB event handler. */
3356 	for (i = 0; i < sdev->device->phys_port_cnt; i++)
3357 		cancel_work_sync(&sdev->port[i].work);
3358 
3359 	ib_destroy_cm_id(sdev->cm_id);
3360 
3361 	/*
3362 	 * Unregistering a target must happen after destroying sdev->cm_id
3363 	 * such that no new SRP_LOGIN_REQ information units can arrive while
3364 	 * destroying the target.
3365 	 */
3366 	spin_lock(&srpt_dev_lock);
3367 	list_del(&sdev->list);
3368 	spin_unlock(&srpt_dev_lock);
3369 	srpt_release_sdev(sdev);
3370 
3371 	ib_destroy_srq(sdev->srq);
3372 	ib_dereg_mr(sdev->mr);
3373 	ib_dealloc_pd(sdev->pd);
3374 
3375 	srpt_free_ioctx_ring((struct srpt_ioctx **)sdev->ioctx_ring, sdev,
3376 			     sdev->srq_size, srp_max_req_size, DMA_FROM_DEVICE);
3377 	sdev->ioctx_ring = NULL;
3378 	kfree(sdev);
3379 }
3380 
3381 static struct ib_client srpt_client = {
3382 	.name = DRV_NAME,
3383 	.add = srpt_add_one,
3384 	.remove = srpt_remove_one
3385 };
3386 
3387 static int srpt_check_true(struct se_portal_group *se_tpg)
3388 {
3389 	return 1;
3390 }
3391 
3392 static int srpt_check_false(struct se_portal_group *se_tpg)
3393 {
3394 	return 0;
3395 }
3396 
3397 static char *srpt_get_fabric_name(void)
3398 {
3399 	return "srpt";
3400 }
3401 
3402 static u8 srpt_get_fabric_proto_ident(struct se_portal_group *se_tpg)
3403 {
3404 	return SCSI_TRANSPORTID_PROTOCOLID_SRP;
3405 }
3406 
3407 static char *srpt_get_fabric_wwn(struct se_portal_group *tpg)
3408 {
3409 	struct srpt_port *sport = container_of(tpg, struct srpt_port, port_tpg_1);
3410 
3411 	return sport->port_guid;
3412 }
3413 
3414 static u16 srpt_get_tag(struct se_portal_group *tpg)
3415 {
3416 	return 1;
3417 }
3418 
3419 static u32 srpt_get_default_depth(struct se_portal_group *se_tpg)
3420 {
3421 	return 1;
3422 }
3423 
3424 static u32 srpt_get_pr_transport_id(struct se_portal_group *se_tpg,
3425 				    struct se_node_acl *se_nacl,
3426 				    struct t10_pr_registration *pr_reg,
3427 				    int *format_code, unsigned char *buf)
3428 {
3429 	struct srpt_node_acl *nacl;
3430 	struct spc_rdma_transport_id *tr_id;
3431 
3432 	nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3433 	tr_id = (void *)buf;
3434 	tr_id->protocol_identifier = SCSI_TRANSPORTID_PROTOCOLID_SRP;
3435 	memcpy(tr_id->i_port_id, nacl->i_port_id, sizeof(tr_id->i_port_id));
3436 	return sizeof(*tr_id);
3437 }
3438 
3439 static u32 srpt_get_pr_transport_id_len(struct se_portal_group *se_tpg,
3440 					struct se_node_acl *se_nacl,
3441 					struct t10_pr_registration *pr_reg,
3442 					int *format_code)
3443 {
3444 	*format_code = 0;
3445 	return sizeof(struct spc_rdma_transport_id);
3446 }
3447 
3448 static char *srpt_parse_pr_out_transport_id(struct se_portal_group *se_tpg,
3449 					    const char *buf, u32 *out_tid_len,
3450 					    char **port_nexus_ptr)
3451 {
3452 	struct spc_rdma_transport_id *tr_id;
3453 
3454 	*port_nexus_ptr = NULL;
3455 	*out_tid_len = sizeof(struct spc_rdma_transport_id);
3456 	tr_id = (void *)buf;
3457 	return (char *)tr_id->i_port_id;
3458 }
3459 
3460 static struct se_node_acl *srpt_alloc_fabric_acl(struct se_portal_group *se_tpg)
3461 {
3462 	struct srpt_node_acl *nacl;
3463 
3464 	nacl = kzalloc(sizeof(struct srpt_node_acl), GFP_KERNEL);
3465 	if (!nacl) {
3466 		printk(KERN_ERR "Unable to allocate struct srpt_node_acl\n");
3467 		return NULL;
3468 	}
3469 
3470 	return &nacl->nacl;
3471 }
3472 
3473 static void srpt_release_fabric_acl(struct se_portal_group *se_tpg,
3474 				    struct se_node_acl *se_nacl)
3475 {
3476 	struct srpt_node_acl *nacl;
3477 
3478 	nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3479 	kfree(nacl);
3480 }
3481 
3482 static u32 srpt_tpg_get_inst_index(struct se_portal_group *se_tpg)
3483 {
3484 	return 1;
3485 }
3486 
3487 static void srpt_release_cmd(struct se_cmd *se_cmd)
3488 {
3489 }
3490 
3491 /**
3492  * srpt_shutdown_session() - Whether or not a session may be shut down.
3493  */
3494 static int srpt_shutdown_session(struct se_session *se_sess)
3495 {
3496 	return true;
3497 }
3498 
3499 /**
3500  * srpt_close_session() - Forcibly close a session.
3501  *
3502  * Callback function invoked by the TCM core to clean up sessions associated
3503  * with a node ACL when the user invokes
3504  * rmdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3505  */
3506 static void srpt_close_session(struct se_session *se_sess)
3507 {
3508 	DECLARE_COMPLETION_ONSTACK(release_done);
3509 	struct srpt_rdma_ch *ch;
3510 	struct srpt_device *sdev;
3511 	int res;
3512 
3513 	ch = se_sess->fabric_sess_ptr;
3514 	WARN_ON(ch->sess != se_sess);
3515 
3516 	pr_debug("ch %p state %d\n", ch, srpt_get_ch_state(ch));
3517 
3518 	sdev = ch->sport->sdev;
3519 	spin_lock_irq(&sdev->spinlock);
3520 	BUG_ON(ch->release_done);
3521 	ch->release_done = &release_done;
3522 	__srpt_close_ch(ch);
3523 	spin_unlock_irq(&sdev->spinlock);
3524 
3525 	res = wait_for_completion_timeout(&release_done, 60 * HZ);
3526 	WARN_ON(res <= 0);
3527 }
3528 
3529 /**
3530  * srpt_sess_get_index() - Return the value of scsiAttIntrPortIndex (SCSI-MIB).
3531  *
3532  * A quote from RFC 4455 (SCSI-MIB) about this MIB object:
3533  * This object represents an arbitrary integer used to uniquely identify a
3534  * particular attached remote initiator port to a particular SCSI target port
3535  * within a particular SCSI target device within a particular SCSI instance.
3536  */
3537 static u32 srpt_sess_get_index(struct se_session *se_sess)
3538 {
3539 	return 0;
3540 }
3541 
3542 static void srpt_set_default_node_attrs(struct se_node_acl *nacl)
3543 {
3544 }
3545 
3546 static u32 srpt_get_task_tag(struct se_cmd *se_cmd)
3547 {
3548 	struct srpt_send_ioctx *ioctx;
3549 
3550 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
3551 	return ioctx->tag;
3552 }
3553 
3554 /* Note: only used from inside debug printk's by the TCM core. */
3555 static int srpt_get_tcm_cmd_state(struct se_cmd *se_cmd)
3556 {
3557 	struct srpt_send_ioctx *ioctx;
3558 
3559 	ioctx = container_of(se_cmd, struct srpt_send_ioctx, cmd);
3560 	return srpt_get_cmd_state(ioctx);
3561 }
3562 
3563 static u16 srpt_set_fabric_sense_len(struct se_cmd *cmd, u32 sense_length)
3564 {
3565 	return 0;
3566 }
3567 
3568 static u16 srpt_get_fabric_sense_len(void)
3569 {
3570 	return 0;
3571 }
3572 
3573 /**
3574  * srpt_parse_i_port_id() - Parse an initiator port ID.
3575  * @name: ASCII representation of a 128-bit initiator port ID.
3576  * @i_port_id: Binary 128-bit port ID.
3577  */
3578 static int srpt_parse_i_port_id(u8 i_port_id[16], const char *name)
3579 {
3580 	const char *p;
3581 	unsigned len, count, leading_zero_bytes;
3582 	int ret, rc;
3583 
3584 	p = name;
3585 	if (strnicmp(p, "0x", 2) == 0)
3586 		p += 2;
3587 	ret = -EINVAL;
3588 	len = strlen(p);
3589 	if (len % 2)
3590 		goto out;
3591 	count = min(len / 2, 16U);
3592 	leading_zero_bytes = 16 - count;
3593 	memset(i_port_id, 0, leading_zero_bytes);
3594 	rc = hex2bin(i_port_id + leading_zero_bytes, p, count);
3595 	if (rc < 0)
3596 		pr_debug("hex2bin failed for srpt_parse_i_port_id: %d\n", rc);
3597 	ret = 0;
3598 out:
3599 	return ret;
3600 }
3601 
3602 /*
3603  * configfs callback function invoked for
3604  * mkdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3605  */
3606 static struct se_node_acl *srpt_make_nodeacl(struct se_portal_group *tpg,
3607 					     struct config_group *group,
3608 					     const char *name)
3609 {
3610 	struct srpt_port *sport = container_of(tpg, struct srpt_port, port_tpg_1);
3611 	struct se_node_acl *se_nacl, *se_nacl_new;
3612 	struct srpt_node_acl *nacl;
3613 	int ret = 0;
3614 	u32 nexus_depth = 1;
3615 	u8 i_port_id[16];
3616 
3617 	if (srpt_parse_i_port_id(i_port_id, name) < 0) {
3618 		printk(KERN_ERR "invalid initiator port ID %s\n", name);
3619 		ret = -EINVAL;
3620 		goto err;
3621 	}
3622 
3623 	se_nacl_new = srpt_alloc_fabric_acl(tpg);
3624 	if (!se_nacl_new) {
3625 		ret = -ENOMEM;
3626 		goto err;
3627 	}
3628 	/*
3629 	 * nacl_new may be released by core_tpg_add_initiator_node_acl()
3630 	 * when converting a node ACL from demo mode to explict
3631 	 */
3632 	se_nacl = core_tpg_add_initiator_node_acl(tpg, se_nacl_new, name,
3633 						  nexus_depth);
3634 	if (IS_ERR(se_nacl)) {
3635 		ret = PTR_ERR(se_nacl);
3636 		goto err;
3637 	}
3638 	/* Locate our struct srpt_node_acl and set sdev and i_port_id. */
3639 	nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3640 	memcpy(&nacl->i_port_id[0], &i_port_id[0], 16);
3641 	nacl->sport = sport;
3642 
3643 	spin_lock_irq(&sport->port_acl_lock);
3644 	list_add_tail(&nacl->list, &sport->port_acl_list);
3645 	spin_unlock_irq(&sport->port_acl_lock);
3646 
3647 	return se_nacl;
3648 err:
3649 	return ERR_PTR(ret);
3650 }
3651 
3652 /*
3653  * configfs callback function invoked for
3654  * rmdir /sys/kernel/config/target/$driver/$port/$tpg/acls/$i_port_id
3655  */
3656 static void srpt_drop_nodeacl(struct se_node_acl *se_nacl)
3657 {
3658 	struct srpt_node_acl *nacl;
3659 	struct srpt_device *sdev;
3660 	struct srpt_port *sport;
3661 
3662 	nacl = container_of(se_nacl, struct srpt_node_acl, nacl);
3663 	sport = nacl->sport;
3664 	sdev = sport->sdev;
3665 	spin_lock_irq(&sport->port_acl_lock);
3666 	list_del(&nacl->list);
3667 	spin_unlock_irq(&sport->port_acl_lock);
3668 	core_tpg_del_initiator_node_acl(&sport->port_tpg_1, se_nacl, 1);
3669 	srpt_release_fabric_acl(NULL, se_nacl);
3670 }
3671 
3672 static ssize_t srpt_tpg_attrib_show_srp_max_rdma_size(
3673 	struct se_portal_group *se_tpg,
3674 	char *page)
3675 {
3676 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3677 
3678 	return sprintf(page, "%u\n", sport->port_attrib.srp_max_rdma_size);
3679 }
3680 
3681 static ssize_t srpt_tpg_attrib_store_srp_max_rdma_size(
3682 	struct se_portal_group *se_tpg,
3683 	const char *page,
3684 	size_t count)
3685 {
3686 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3687 	unsigned long val;
3688 	int ret;
3689 
3690 	ret = strict_strtoul(page, 0, &val);
3691 	if (ret < 0) {
3692 		pr_err("strict_strtoul() failed with ret: %d\n", ret);
3693 		return -EINVAL;
3694 	}
3695 	if (val > MAX_SRPT_RDMA_SIZE) {
3696 		pr_err("val: %lu exceeds MAX_SRPT_RDMA_SIZE: %d\n", val,
3697 			MAX_SRPT_RDMA_SIZE);
3698 		return -EINVAL;
3699 	}
3700 	if (val < DEFAULT_MAX_RDMA_SIZE) {
3701 		pr_err("val: %lu smaller than DEFAULT_MAX_RDMA_SIZE: %d\n",
3702 			val, DEFAULT_MAX_RDMA_SIZE);
3703 		return -EINVAL;
3704 	}
3705 	sport->port_attrib.srp_max_rdma_size = val;
3706 
3707 	return count;
3708 }
3709 
3710 TF_TPG_ATTRIB_ATTR(srpt, srp_max_rdma_size, S_IRUGO | S_IWUSR);
3711 
3712 static ssize_t srpt_tpg_attrib_show_srp_max_rsp_size(
3713 	struct se_portal_group *se_tpg,
3714 	char *page)
3715 {
3716 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3717 
3718 	return sprintf(page, "%u\n", sport->port_attrib.srp_max_rsp_size);
3719 }
3720 
3721 static ssize_t srpt_tpg_attrib_store_srp_max_rsp_size(
3722 	struct se_portal_group *se_tpg,
3723 	const char *page,
3724 	size_t count)
3725 {
3726 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3727 	unsigned long val;
3728 	int ret;
3729 
3730 	ret = strict_strtoul(page, 0, &val);
3731 	if (ret < 0) {
3732 		pr_err("strict_strtoul() failed with ret: %d\n", ret);
3733 		return -EINVAL;
3734 	}
3735 	if (val > MAX_SRPT_RSP_SIZE) {
3736 		pr_err("val: %lu exceeds MAX_SRPT_RSP_SIZE: %d\n", val,
3737 			MAX_SRPT_RSP_SIZE);
3738 		return -EINVAL;
3739 	}
3740 	if (val < MIN_MAX_RSP_SIZE) {
3741 		pr_err("val: %lu smaller than MIN_MAX_RSP_SIZE: %d\n", val,
3742 			MIN_MAX_RSP_SIZE);
3743 		return -EINVAL;
3744 	}
3745 	sport->port_attrib.srp_max_rsp_size = val;
3746 
3747 	return count;
3748 }
3749 
3750 TF_TPG_ATTRIB_ATTR(srpt, srp_max_rsp_size, S_IRUGO | S_IWUSR);
3751 
3752 static ssize_t srpt_tpg_attrib_show_srp_sq_size(
3753 	struct se_portal_group *se_tpg,
3754 	char *page)
3755 {
3756 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3757 
3758 	return sprintf(page, "%u\n", sport->port_attrib.srp_sq_size);
3759 }
3760 
3761 static ssize_t srpt_tpg_attrib_store_srp_sq_size(
3762 	struct se_portal_group *se_tpg,
3763 	const char *page,
3764 	size_t count)
3765 {
3766 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3767 	unsigned long val;
3768 	int ret;
3769 
3770 	ret = strict_strtoul(page, 0, &val);
3771 	if (ret < 0) {
3772 		pr_err("strict_strtoul() failed with ret: %d\n", ret);
3773 		return -EINVAL;
3774 	}
3775 	if (val > MAX_SRPT_SRQ_SIZE) {
3776 		pr_err("val: %lu exceeds MAX_SRPT_SRQ_SIZE: %d\n", val,
3777 			MAX_SRPT_SRQ_SIZE);
3778 		return -EINVAL;
3779 	}
3780 	if (val < MIN_SRPT_SRQ_SIZE) {
3781 		pr_err("val: %lu smaller than MIN_SRPT_SRQ_SIZE: %d\n", val,
3782 			MIN_SRPT_SRQ_SIZE);
3783 		return -EINVAL;
3784 	}
3785 	sport->port_attrib.srp_sq_size = val;
3786 
3787 	return count;
3788 }
3789 
3790 TF_TPG_ATTRIB_ATTR(srpt, srp_sq_size, S_IRUGO | S_IWUSR);
3791 
3792 static struct configfs_attribute *srpt_tpg_attrib_attrs[] = {
3793 	&srpt_tpg_attrib_srp_max_rdma_size.attr,
3794 	&srpt_tpg_attrib_srp_max_rsp_size.attr,
3795 	&srpt_tpg_attrib_srp_sq_size.attr,
3796 	NULL,
3797 };
3798 
3799 static ssize_t srpt_tpg_show_enable(
3800 	struct se_portal_group *se_tpg,
3801 	char *page)
3802 {
3803 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3804 
3805 	return snprintf(page, PAGE_SIZE, "%d\n", (sport->enabled) ? 1: 0);
3806 }
3807 
3808 static ssize_t srpt_tpg_store_enable(
3809 	struct se_portal_group *se_tpg,
3810 	const char *page,
3811 	size_t count)
3812 {
3813 	struct srpt_port *sport = container_of(se_tpg, struct srpt_port, port_tpg_1);
3814 	unsigned long tmp;
3815         int ret;
3816 
3817 	ret = strict_strtoul(page, 0, &tmp);
3818 	if (ret < 0) {
3819 		printk(KERN_ERR "Unable to extract srpt_tpg_store_enable\n");
3820 		return -EINVAL;
3821 	}
3822 
3823 	if ((tmp != 0) && (tmp != 1)) {
3824 		printk(KERN_ERR "Illegal value for srpt_tpg_store_enable: %lu\n", tmp);
3825 		return -EINVAL;
3826 	}
3827 	if (tmp == 1)
3828 		sport->enabled = true;
3829 	else
3830 		sport->enabled = false;
3831 
3832 	return count;
3833 }
3834 
3835 TF_TPG_BASE_ATTR(srpt, enable, S_IRUGO | S_IWUSR);
3836 
3837 static struct configfs_attribute *srpt_tpg_attrs[] = {
3838 	&srpt_tpg_enable.attr,
3839 	NULL,
3840 };
3841 
3842 /**
3843  * configfs callback invoked for
3844  * mkdir /sys/kernel/config/target/$driver/$port/$tpg
3845  */
3846 static struct se_portal_group *srpt_make_tpg(struct se_wwn *wwn,
3847 					     struct config_group *group,
3848 					     const char *name)
3849 {
3850 	struct srpt_port *sport = container_of(wwn, struct srpt_port, port_wwn);
3851 	int res;
3852 
3853 	/* Initialize sport->port_wwn and sport->port_tpg_1 */
3854 	res = core_tpg_register(&srpt_target->tf_ops, &sport->port_wwn,
3855 			&sport->port_tpg_1, sport, TRANSPORT_TPG_TYPE_NORMAL);
3856 	if (res)
3857 		return ERR_PTR(res);
3858 
3859 	return &sport->port_tpg_1;
3860 }
3861 
3862 /**
3863  * configfs callback invoked for
3864  * rmdir /sys/kernel/config/target/$driver/$port/$tpg
3865  */
3866 static void srpt_drop_tpg(struct se_portal_group *tpg)
3867 {
3868 	struct srpt_port *sport = container_of(tpg,
3869 				struct srpt_port, port_tpg_1);
3870 
3871 	sport->enabled = false;
3872 	core_tpg_deregister(&sport->port_tpg_1);
3873 }
3874 
3875 /**
3876  * configfs callback invoked for
3877  * mkdir /sys/kernel/config/target/$driver/$port
3878  */
3879 static struct se_wwn *srpt_make_tport(struct target_fabric_configfs *tf,
3880 				      struct config_group *group,
3881 				      const char *name)
3882 {
3883 	struct srpt_port *sport;
3884 	int ret;
3885 
3886 	sport = srpt_lookup_port(name);
3887 	pr_debug("make_tport(%s)\n", name);
3888 	ret = -EINVAL;
3889 	if (!sport)
3890 		goto err;
3891 
3892 	return &sport->port_wwn;
3893 
3894 err:
3895 	return ERR_PTR(ret);
3896 }
3897 
3898 /**
3899  * configfs callback invoked for
3900  * rmdir /sys/kernel/config/target/$driver/$port
3901  */
3902 static void srpt_drop_tport(struct se_wwn *wwn)
3903 {
3904 	struct srpt_port *sport = container_of(wwn, struct srpt_port, port_wwn);
3905 
3906 	pr_debug("drop_tport(%s\n", config_item_name(&sport->port_wwn.wwn_group.cg_item));
3907 }
3908 
3909 static ssize_t srpt_wwn_show_attr_version(struct target_fabric_configfs *tf,
3910 					      char *buf)
3911 {
3912 	return scnprintf(buf, PAGE_SIZE, "%s\n", DRV_VERSION);
3913 }
3914 
3915 TF_WWN_ATTR_RO(srpt, version);
3916 
3917 static struct configfs_attribute *srpt_wwn_attrs[] = {
3918 	&srpt_wwn_version.attr,
3919 	NULL,
3920 };
3921 
3922 static struct target_core_fabric_ops srpt_template = {
3923 	.get_fabric_name		= srpt_get_fabric_name,
3924 	.get_fabric_proto_ident		= srpt_get_fabric_proto_ident,
3925 	.tpg_get_wwn			= srpt_get_fabric_wwn,
3926 	.tpg_get_tag			= srpt_get_tag,
3927 	.tpg_get_default_depth		= srpt_get_default_depth,
3928 	.tpg_get_pr_transport_id	= srpt_get_pr_transport_id,
3929 	.tpg_get_pr_transport_id_len	= srpt_get_pr_transport_id_len,
3930 	.tpg_parse_pr_out_transport_id	= srpt_parse_pr_out_transport_id,
3931 	.tpg_check_demo_mode		= srpt_check_false,
3932 	.tpg_check_demo_mode_cache	= srpt_check_true,
3933 	.tpg_check_demo_mode_write_protect = srpt_check_true,
3934 	.tpg_check_prod_mode_write_protect = srpt_check_false,
3935 	.tpg_alloc_fabric_acl		= srpt_alloc_fabric_acl,
3936 	.tpg_release_fabric_acl		= srpt_release_fabric_acl,
3937 	.tpg_get_inst_index		= srpt_tpg_get_inst_index,
3938 	.release_cmd			= srpt_release_cmd,
3939 	.check_stop_free		= srpt_check_stop_free,
3940 	.shutdown_session		= srpt_shutdown_session,
3941 	.close_session			= srpt_close_session,
3942 	.sess_get_index			= srpt_sess_get_index,
3943 	.sess_get_initiator_sid		= NULL,
3944 	.write_pending			= srpt_write_pending,
3945 	.write_pending_status		= srpt_write_pending_status,
3946 	.set_default_node_attributes	= srpt_set_default_node_attrs,
3947 	.get_task_tag			= srpt_get_task_tag,
3948 	.get_cmd_state			= srpt_get_tcm_cmd_state,
3949 	.queue_data_in			= srpt_queue_response,
3950 	.queue_status			= srpt_queue_status,
3951 	.queue_tm_rsp			= srpt_queue_response,
3952 	.get_fabric_sense_len		= srpt_get_fabric_sense_len,
3953 	.set_fabric_sense_len		= srpt_set_fabric_sense_len,
3954 	/*
3955 	 * Setup function pointers for generic logic in
3956 	 * target_core_fabric_configfs.c
3957 	 */
3958 	.fabric_make_wwn		= srpt_make_tport,
3959 	.fabric_drop_wwn		= srpt_drop_tport,
3960 	.fabric_make_tpg		= srpt_make_tpg,
3961 	.fabric_drop_tpg		= srpt_drop_tpg,
3962 	.fabric_post_link		= NULL,
3963 	.fabric_pre_unlink		= NULL,
3964 	.fabric_make_np			= NULL,
3965 	.fabric_drop_np			= NULL,
3966 	.fabric_make_nodeacl		= srpt_make_nodeacl,
3967 	.fabric_drop_nodeacl		= srpt_drop_nodeacl,
3968 };
3969 
3970 /**
3971  * srpt_init_module() - Kernel module initialization.
3972  *
3973  * Note: Since ib_register_client() registers callback functions, and since at
3974  * least one of these callback functions (srpt_add_one()) calls target core
3975  * functions, this driver must be registered with the target core before
3976  * ib_register_client() is called.
3977  */
3978 static int __init srpt_init_module(void)
3979 {
3980 	int ret;
3981 
3982 	ret = -EINVAL;
3983 	if (srp_max_req_size < MIN_MAX_REQ_SIZE) {
3984 		printk(KERN_ERR "invalid value %d for kernel module parameter"
3985 		       " srp_max_req_size -- must be at least %d.\n",
3986 		       srp_max_req_size, MIN_MAX_REQ_SIZE);
3987 		goto out;
3988 	}
3989 
3990 	if (srpt_srq_size < MIN_SRPT_SRQ_SIZE
3991 	    || srpt_srq_size > MAX_SRPT_SRQ_SIZE) {
3992 		printk(KERN_ERR "invalid value %d for kernel module parameter"
3993 		       " srpt_srq_size -- must be in the range [%d..%d].\n",
3994 		       srpt_srq_size, MIN_SRPT_SRQ_SIZE, MAX_SRPT_SRQ_SIZE);
3995 		goto out;
3996 	}
3997 
3998 	srpt_target = target_fabric_configfs_init(THIS_MODULE, "srpt");
3999 	if (IS_ERR(srpt_target)) {
4000 		printk(KERN_ERR "couldn't register\n");
4001 		ret = PTR_ERR(srpt_target);
4002 		goto out;
4003 	}
4004 
4005 	srpt_target->tf_ops = srpt_template;
4006 
4007 	/* Enable SG chaining */
4008 	srpt_target->tf_ops.task_sg_chaining = true;
4009 
4010 	/*
4011 	 * Set up default attribute lists.
4012 	 */
4013 	srpt_target->tf_cit_tmpl.tfc_wwn_cit.ct_attrs = srpt_wwn_attrs;
4014 	srpt_target->tf_cit_tmpl.tfc_tpg_base_cit.ct_attrs = srpt_tpg_attrs;
4015 	srpt_target->tf_cit_tmpl.tfc_tpg_attrib_cit.ct_attrs = srpt_tpg_attrib_attrs;
4016 	srpt_target->tf_cit_tmpl.tfc_tpg_param_cit.ct_attrs = NULL;
4017 	srpt_target->tf_cit_tmpl.tfc_tpg_np_base_cit.ct_attrs = NULL;
4018 	srpt_target->tf_cit_tmpl.tfc_tpg_nacl_base_cit.ct_attrs = NULL;
4019 	srpt_target->tf_cit_tmpl.tfc_tpg_nacl_attrib_cit.ct_attrs = NULL;
4020 	srpt_target->tf_cit_tmpl.tfc_tpg_nacl_auth_cit.ct_attrs = NULL;
4021 	srpt_target->tf_cit_tmpl.tfc_tpg_nacl_param_cit.ct_attrs = NULL;
4022 
4023 	ret = target_fabric_configfs_register(srpt_target);
4024 	if (ret < 0) {
4025 		printk(KERN_ERR "couldn't register\n");
4026 		goto out_free_target;
4027 	}
4028 
4029 	ret = ib_register_client(&srpt_client);
4030 	if (ret) {
4031 		printk(KERN_ERR "couldn't register IB client\n");
4032 		goto out_unregister_target;
4033 	}
4034 
4035 	return 0;
4036 
4037 out_unregister_target:
4038 	target_fabric_configfs_deregister(srpt_target);
4039 	srpt_target = NULL;
4040 out_free_target:
4041 	if (srpt_target)
4042 		target_fabric_configfs_free(srpt_target);
4043 out:
4044 	return ret;
4045 }
4046 
4047 static void __exit srpt_cleanup_module(void)
4048 {
4049 	ib_unregister_client(&srpt_client);
4050 	target_fabric_configfs_deregister(srpt_target);
4051 	srpt_target = NULL;
4052 }
4053 
4054 module_init(srpt_init_module);
4055 module_exit(srpt_cleanup_module);
4056