1/*
2 * Copyright (c) 2004, 2005, 2006 Voltaire, Inc. All rights reserved.
3 * Copyright (c) 2005, 2006 Cisco Systems.  All rights reserved.
4 * Copyright (c) 2013-2014 Mellanox Technologies. All rights reserved.
5 *
6 * This software is available to you under a choice of one of two
7 * licenses.  You may choose to be licensed under the terms of the GNU
8 * General Public License (GPL) Version 2, available from the file
9 * COPYING in the main directory of this source tree, or the
10 * OpenIB.org BSD license below:
11 *
12 *     Redistribution and use in source and binary forms, with or
13 *     without modification, are permitted provided that the following
14 *     conditions are met:
15 *
16 *	- Redistributions of source code must retain the above
17 *	  copyright notice, this list of conditions and the following
18 *	  disclaimer.
19 *
20 *	- Redistributions in binary form must reproduce the above
21 *	  copyright notice, this list of conditions and the following
22 *	  disclaimer in the documentation and/or other materials
23 *	  provided with the distribution.
24 *
25 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32 * SOFTWARE.
33 */
34#include <linux/kernel.h>
35#include <linux/module.h>
36#include <linux/slab.h>
37#include <linux/delay.h>
38
39#include "iscsi_iser.h"
40
41#define ISCSI_ISER_MAX_CONN	8
42#define ISER_MAX_RX_LEN		(ISER_QP_MAX_RECV_DTOS * ISCSI_ISER_MAX_CONN)
43#define ISER_MAX_TX_LEN		(ISER_QP_MAX_REQ_DTOS  * ISCSI_ISER_MAX_CONN)
44#define ISER_MAX_CQ_LEN		(ISER_MAX_RX_LEN + ISER_MAX_TX_LEN + \
45				 ISCSI_ISER_MAX_CONN)
46
47static int iser_cq_poll_limit = 512;
48
49static void iser_cq_tasklet_fn(unsigned long data);
50static void iser_cq_callback(struct ib_cq *cq, void *cq_context);
51
52static void iser_cq_event_callback(struct ib_event *cause, void *context)
53{
54	iser_err("got cq event %d \n", cause->event);
55}
56
57static void iser_qp_event_callback(struct ib_event *cause, void *context)
58{
59	iser_err("got qp event %d\n",cause->event);
60}
61
62static void iser_event_handler(struct ib_event_handler *handler,
63				struct ib_event *event)
64{
65	iser_err("async event %d on device %s port %d\n", event->event,
66		event->device->name, event->element.port_num);
67}
68
69/**
70 * iser_create_device_ib_res - creates Protection Domain (PD), Completion
71 * Queue (CQ), DMA Memory Region (DMA MR) with the device associated with
72 * the adapator.
73 *
74 * returns 0 on success, -1 on failure
75 */
76static int iser_create_device_ib_res(struct iser_device *device)
77{
78	struct ib_device_attr *dev_attr = &device->dev_attr;
79	int ret, i, max_cqe;
80
81	ret = ib_query_device(device->ib_device, dev_attr);
82	if (ret) {
83		pr_warn("Query device failed for %s\n", device->ib_device->name);
84		return ret;
85	}
86
87	/* Assign function handles  - based on FMR support */
88	if (device->ib_device->alloc_fmr && device->ib_device->dealloc_fmr &&
89	    device->ib_device->map_phys_fmr && device->ib_device->unmap_fmr) {
90		iser_info("FMR supported, using FMR for registration\n");
91		device->iser_alloc_rdma_reg_res = iser_create_fmr_pool;
92		device->iser_free_rdma_reg_res = iser_free_fmr_pool;
93		device->iser_reg_rdma_mem = iser_reg_rdma_mem_fmr;
94		device->iser_unreg_rdma_mem = iser_unreg_mem_fmr;
95	} else
96	if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
97		iser_info("FastReg supported, using FastReg for registration\n");
98		device->iser_alloc_rdma_reg_res = iser_create_fastreg_pool;
99		device->iser_free_rdma_reg_res = iser_free_fastreg_pool;
100		device->iser_reg_rdma_mem = iser_reg_rdma_mem_fastreg;
101		device->iser_unreg_rdma_mem = iser_unreg_mem_fastreg;
102	} else {
103		iser_err("IB device does not support FMRs nor FastRegs, can't register memory\n");
104		return -1;
105	}
106
107	device->comps_used = min_t(int, num_online_cpus(),
108				 device->ib_device->num_comp_vectors);
109
110	device->comps = kcalloc(device->comps_used, sizeof(*device->comps),
111				GFP_KERNEL);
112	if (!device->comps)
113		goto comps_err;
114
115	max_cqe = min(ISER_MAX_CQ_LEN, dev_attr->max_cqe);
116
117	iser_info("using %d CQs, device %s supports %d vectors max_cqe %d\n",
118		  device->comps_used, device->ib_device->name,
119		  device->ib_device->num_comp_vectors, max_cqe);
120
121	device->pd = ib_alloc_pd(device->ib_device);
122	if (IS_ERR(device->pd))
123		goto pd_err;
124
125	for (i = 0; i < device->comps_used; i++) {
126		struct iser_comp *comp = &device->comps[i];
127
128		comp->device = device;
129		comp->cq = ib_create_cq(device->ib_device,
130					iser_cq_callback,
131					iser_cq_event_callback,
132					(void *)comp,
133					max_cqe, i);
134		if (IS_ERR(comp->cq)) {
135			comp->cq = NULL;
136			goto cq_err;
137		}
138
139		if (ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP))
140			goto cq_err;
141
142		tasklet_init(&comp->tasklet, iser_cq_tasklet_fn,
143			     (unsigned long)comp);
144	}
145
146	device->mr = ib_get_dma_mr(device->pd, IB_ACCESS_LOCAL_WRITE |
147				   IB_ACCESS_REMOTE_WRITE |
148				   IB_ACCESS_REMOTE_READ);
149	if (IS_ERR(device->mr))
150		goto dma_mr_err;
151
152	INIT_IB_EVENT_HANDLER(&device->event_handler, device->ib_device,
153				iser_event_handler);
154	if (ib_register_event_handler(&device->event_handler))
155		goto handler_err;
156
157	return 0;
158
159handler_err:
160	ib_dereg_mr(device->mr);
161dma_mr_err:
162	for (i = 0; i < device->comps_used; i++)
163		tasklet_kill(&device->comps[i].tasklet);
164cq_err:
165	for (i = 0; i < device->comps_used; i++) {
166		struct iser_comp *comp = &device->comps[i];
167
168		if (comp->cq)
169			ib_destroy_cq(comp->cq);
170	}
171	ib_dealloc_pd(device->pd);
172pd_err:
173	kfree(device->comps);
174comps_err:
175	iser_err("failed to allocate an IB resource\n");
176	return -1;
177}
178
179/**
180 * iser_free_device_ib_res - destroy/dealloc/dereg the DMA MR,
181 * CQ and PD created with the device associated with the adapator.
182 */
183static void iser_free_device_ib_res(struct iser_device *device)
184{
185	int i;
186	BUG_ON(device->mr == NULL);
187
188	for (i = 0; i < device->comps_used; i++) {
189		struct iser_comp *comp = &device->comps[i];
190
191		tasklet_kill(&comp->tasklet);
192		ib_destroy_cq(comp->cq);
193		comp->cq = NULL;
194	}
195
196	(void)ib_unregister_event_handler(&device->event_handler);
197	(void)ib_dereg_mr(device->mr);
198	(void)ib_dealloc_pd(device->pd);
199
200	kfree(device->comps);
201	device->comps = NULL;
202
203	device->mr = NULL;
204	device->pd = NULL;
205}
206
207/**
208 * iser_create_fmr_pool - Creates FMR pool and page_vector
209 *
210 * returns 0 on success, or errno code on failure
211 */
212int iser_create_fmr_pool(struct ib_conn *ib_conn, unsigned cmds_max)
213{
214	struct iser_device *device = ib_conn->device;
215	struct ib_fmr_pool_param params;
216	int ret = -ENOMEM;
217
218	ib_conn->fmr.page_vec = kmalloc(sizeof(*ib_conn->fmr.page_vec) +
219					(sizeof(u64)*(ISCSI_ISER_SG_TABLESIZE + 1)),
220					GFP_KERNEL);
221	if (!ib_conn->fmr.page_vec)
222		return ret;
223
224	ib_conn->fmr.page_vec->pages = (u64 *)(ib_conn->fmr.page_vec + 1);
225
226	params.page_shift        = SHIFT_4K;
227	/* when the first/last SG element are not start/end *
228	 * page aligned, the map whould be of N+1 pages     */
229	params.max_pages_per_fmr = ISCSI_ISER_SG_TABLESIZE + 1;
230	/* make the pool size twice the max number of SCSI commands *
231	 * the ML is expected to queue, watermark for unmap at 50%  */
232	params.pool_size	 = cmds_max * 2;
233	params.dirty_watermark	 = cmds_max;
234	params.cache		 = 0;
235	params.flush_function	 = NULL;
236	params.access		 = (IB_ACCESS_LOCAL_WRITE  |
237				    IB_ACCESS_REMOTE_WRITE |
238				    IB_ACCESS_REMOTE_READ);
239
240	ib_conn->fmr.pool = ib_create_fmr_pool(device->pd, &params);
241	if (!IS_ERR(ib_conn->fmr.pool))
242		return 0;
243
244	/* no FMR => no need for page_vec */
245	kfree(ib_conn->fmr.page_vec);
246	ib_conn->fmr.page_vec = NULL;
247
248	ret = PTR_ERR(ib_conn->fmr.pool);
249	ib_conn->fmr.pool = NULL;
250	if (ret != -ENOSYS) {
251		iser_err("FMR allocation failed, err %d\n", ret);
252		return ret;
253	} else {
254		iser_warn("FMRs are not supported, using unaligned mode\n");
255		return 0;
256	}
257}
258
259/**
260 * iser_free_fmr_pool - releases the FMR pool and page vec
261 */
262void iser_free_fmr_pool(struct ib_conn *ib_conn)
263{
264	iser_info("freeing conn %p fmr pool %p\n",
265		  ib_conn, ib_conn->fmr.pool);
266
267	if (ib_conn->fmr.pool != NULL)
268		ib_destroy_fmr_pool(ib_conn->fmr.pool);
269
270	ib_conn->fmr.pool = NULL;
271
272	kfree(ib_conn->fmr.page_vec);
273	ib_conn->fmr.page_vec = NULL;
274}
275
276static int
277iser_alloc_pi_ctx(struct ib_device *ib_device, struct ib_pd *pd,
278		  struct fast_reg_descriptor *desc)
279{
280	struct iser_pi_context *pi_ctx = NULL;
281	struct ib_mr_init_attr mr_init_attr = {.max_reg_descriptors = 2,
282					       .flags = IB_MR_SIGNATURE_EN};
283	int ret = 0;
284
285	desc->pi_ctx = kzalloc(sizeof(*desc->pi_ctx), GFP_KERNEL);
286	if (!desc->pi_ctx)
287		return -ENOMEM;
288
289	pi_ctx = desc->pi_ctx;
290
291	pi_ctx->prot_frpl = ib_alloc_fast_reg_page_list(ib_device,
292					    ISCSI_ISER_SG_TABLESIZE);
293	if (IS_ERR(pi_ctx->prot_frpl)) {
294		ret = PTR_ERR(pi_ctx->prot_frpl);
295		goto prot_frpl_failure;
296	}
297
298	pi_ctx->prot_mr = ib_alloc_fast_reg_mr(pd,
299					ISCSI_ISER_SG_TABLESIZE + 1);
300	if (IS_ERR(pi_ctx->prot_mr)) {
301		ret = PTR_ERR(pi_ctx->prot_mr);
302		goto prot_mr_failure;
303	}
304	desc->reg_indicators |= ISER_PROT_KEY_VALID;
305
306	pi_ctx->sig_mr = ib_create_mr(pd, &mr_init_attr);
307	if (IS_ERR(pi_ctx->sig_mr)) {
308		ret = PTR_ERR(pi_ctx->sig_mr);
309		goto sig_mr_failure;
310	}
311	desc->reg_indicators |= ISER_SIG_KEY_VALID;
312	desc->reg_indicators &= ~ISER_FASTREG_PROTECTED;
313
314	return 0;
315
316sig_mr_failure:
317	ib_dereg_mr(desc->pi_ctx->prot_mr);
318prot_mr_failure:
319	ib_free_fast_reg_page_list(desc->pi_ctx->prot_frpl);
320prot_frpl_failure:
321	kfree(desc->pi_ctx);
322
323	return ret;
324}
325
326static void
327iser_free_pi_ctx(struct iser_pi_context *pi_ctx)
328{
329	ib_free_fast_reg_page_list(pi_ctx->prot_frpl);
330	ib_dereg_mr(pi_ctx->prot_mr);
331	ib_destroy_mr(pi_ctx->sig_mr);
332	kfree(pi_ctx);
333}
334
335static int
336iser_create_fastreg_desc(struct ib_device *ib_device, struct ib_pd *pd,
337			 bool pi_enable, struct fast_reg_descriptor *desc)
338{
339	int ret;
340
341	desc->data_frpl = ib_alloc_fast_reg_page_list(ib_device,
342						      ISCSI_ISER_SG_TABLESIZE + 1);
343	if (IS_ERR(desc->data_frpl)) {
344		ret = PTR_ERR(desc->data_frpl);
345		iser_err("Failed to allocate ib_fast_reg_page_list err=%d\n",
346			 ret);
347		return PTR_ERR(desc->data_frpl);
348	}
349
350	desc->data_mr = ib_alloc_fast_reg_mr(pd, ISCSI_ISER_SG_TABLESIZE + 1);
351	if (IS_ERR(desc->data_mr)) {
352		ret = PTR_ERR(desc->data_mr);
353		iser_err("Failed to allocate ib_fast_reg_mr err=%d\n", ret);
354		goto fast_reg_mr_failure;
355	}
356	desc->reg_indicators |= ISER_DATA_KEY_VALID;
357
358	if (pi_enable) {
359		ret = iser_alloc_pi_ctx(ib_device, pd, desc);
360		if (ret)
361			goto pi_ctx_alloc_failure;
362	}
363
364	return 0;
365pi_ctx_alloc_failure:
366	ib_dereg_mr(desc->data_mr);
367fast_reg_mr_failure:
368	ib_free_fast_reg_page_list(desc->data_frpl);
369
370	return ret;
371}
372
373/**
374 * iser_create_fastreg_pool - Creates pool of fast_reg descriptors
375 * for fast registration work requests.
376 * returns 0 on success, or errno code on failure
377 */
378int iser_create_fastreg_pool(struct ib_conn *ib_conn, unsigned cmds_max)
379{
380	struct iser_device *device = ib_conn->device;
381	struct fast_reg_descriptor *desc;
382	int i, ret;
383
384	INIT_LIST_HEAD(&ib_conn->fastreg.pool);
385	ib_conn->fastreg.pool_size = 0;
386	for (i = 0; i < cmds_max; i++) {
387		desc = kzalloc(sizeof(*desc), GFP_KERNEL);
388		if (!desc) {
389			iser_err("Failed to allocate a new fast_reg descriptor\n");
390			ret = -ENOMEM;
391			goto err;
392		}
393
394		ret = iser_create_fastreg_desc(device->ib_device, device->pd,
395					       ib_conn->pi_support, desc);
396		if (ret) {
397			iser_err("Failed to create fastreg descriptor err=%d\n",
398				 ret);
399			kfree(desc);
400			goto err;
401		}
402
403		list_add_tail(&desc->list, &ib_conn->fastreg.pool);
404		ib_conn->fastreg.pool_size++;
405	}
406
407	return 0;
408
409err:
410	iser_free_fastreg_pool(ib_conn);
411	return ret;
412}
413
414/**
415 * iser_free_fastreg_pool - releases the pool of fast_reg descriptors
416 */
417void iser_free_fastreg_pool(struct ib_conn *ib_conn)
418{
419	struct fast_reg_descriptor *desc, *tmp;
420	int i = 0;
421
422	if (list_empty(&ib_conn->fastreg.pool))
423		return;
424
425	iser_info("freeing conn %p fr pool\n", ib_conn);
426
427	list_for_each_entry_safe(desc, tmp, &ib_conn->fastreg.pool, list) {
428		list_del(&desc->list);
429		ib_free_fast_reg_page_list(desc->data_frpl);
430		ib_dereg_mr(desc->data_mr);
431		if (desc->pi_ctx)
432			iser_free_pi_ctx(desc->pi_ctx);
433		kfree(desc);
434		++i;
435	}
436
437	if (i < ib_conn->fastreg.pool_size)
438		iser_warn("pool still has %d regions registered\n",
439			  ib_conn->fastreg.pool_size - i);
440}
441
442/**
443 * iser_create_ib_conn_res - Queue-Pair (QP)
444 *
445 * returns 0 on success, -1 on failure
446 */
447static int iser_create_ib_conn_res(struct ib_conn *ib_conn)
448{
449	struct iser_conn *iser_conn = container_of(ib_conn, struct iser_conn,
450						   ib_conn);
451	struct iser_device	*device;
452	struct ib_device_attr *dev_attr;
453	struct ib_qp_init_attr	init_attr;
454	int			ret = -ENOMEM;
455	int index, min_index = 0;
456
457	BUG_ON(ib_conn->device == NULL);
458
459	device = ib_conn->device;
460	dev_attr = &device->dev_attr;
461
462	memset(&init_attr, 0, sizeof init_attr);
463
464	mutex_lock(&ig.connlist_mutex);
465	/* select the CQ with the minimal number of usages */
466	for (index = 0; index < device->comps_used; index++) {
467		if (device->comps[index].active_qps <
468		    device->comps[min_index].active_qps)
469			min_index = index;
470	}
471	ib_conn->comp = &device->comps[min_index];
472	ib_conn->comp->active_qps++;
473	mutex_unlock(&ig.connlist_mutex);
474	iser_info("cq index %d used for ib_conn %p\n", min_index, ib_conn);
475
476	init_attr.event_handler = iser_qp_event_callback;
477	init_attr.qp_context	= (void *)ib_conn;
478	init_attr.send_cq	= ib_conn->comp->cq;
479	init_attr.recv_cq	= ib_conn->comp->cq;
480	init_attr.cap.max_recv_wr  = ISER_QP_MAX_RECV_DTOS;
481	init_attr.cap.max_send_sge = 2;
482	init_attr.cap.max_recv_sge = 1;
483	init_attr.sq_sig_type	= IB_SIGNAL_REQ_WR;
484	init_attr.qp_type	= IB_QPT_RC;
485	if (ib_conn->pi_support) {
486		init_attr.cap.max_send_wr = ISER_QP_SIG_MAX_REQ_DTOS + 1;
487		init_attr.create_flags |= IB_QP_CREATE_SIGNATURE_EN;
488		iser_conn->max_cmds =
489			ISER_GET_MAX_XMIT_CMDS(ISER_QP_SIG_MAX_REQ_DTOS);
490	} else {
491		if (dev_attr->max_qp_wr > ISER_QP_MAX_REQ_DTOS) {
492			init_attr.cap.max_send_wr  = ISER_QP_MAX_REQ_DTOS + 1;
493			iser_conn->max_cmds =
494				ISER_GET_MAX_XMIT_CMDS(ISER_QP_MAX_REQ_DTOS);
495		} else {
496			init_attr.cap.max_send_wr = dev_attr->max_qp_wr;
497			iser_conn->max_cmds =
498				ISER_GET_MAX_XMIT_CMDS(dev_attr->max_qp_wr);
499			iser_dbg("device %s supports max_send_wr %d\n",
500				 device->ib_device->name, dev_attr->max_qp_wr);
501		}
502	}
503
504	ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
505	if (ret)
506		goto out_err;
507
508	ib_conn->qp = ib_conn->cma_id->qp;
509	iser_info("setting conn %p cma_id %p qp %p\n",
510		  ib_conn, ib_conn->cma_id,
511		  ib_conn->cma_id->qp);
512	return ret;
513
514out_err:
515	mutex_lock(&ig.connlist_mutex);
516	ib_conn->comp->active_qps--;
517	mutex_unlock(&ig.connlist_mutex);
518	iser_err("unable to alloc mem or create resource, err %d\n", ret);
519
520	return ret;
521}
522
523/**
524 * based on the resolved device node GUID see if there already allocated
525 * device for this device. If there's no such, create one.
526 */
527static
528struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
529{
530	struct iser_device *device;
531
532	mutex_lock(&ig.device_list_mutex);
533
534	list_for_each_entry(device, &ig.device_list, ig_list)
535		/* find if there's a match using the node GUID */
536		if (device->ib_device->node_guid == cma_id->device->node_guid)
537			goto inc_refcnt;
538
539	device = kzalloc(sizeof *device, GFP_KERNEL);
540	if (device == NULL)
541		goto out;
542
543	/* assign this device to the device */
544	device->ib_device = cma_id->device;
545	/* init the device and link it into ig device list */
546	if (iser_create_device_ib_res(device)) {
547		kfree(device);
548		device = NULL;
549		goto out;
550	}
551	list_add(&device->ig_list, &ig.device_list);
552
553inc_refcnt:
554	device->refcount++;
555out:
556	mutex_unlock(&ig.device_list_mutex);
557	return device;
558}
559
560/* if there's no demand for this device, release it */
561static void iser_device_try_release(struct iser_device *device)
562{
563	mutex_lock(&ig.device_list_mutex);
564	device->refcount--;
565	iser_info("device %p refcount %d\n", device, device->refcount);
566	if (!device->refcount) {
567		iser_free_device_ib_res(device);
568		list_del(&device->ig_list);
569		kfree(device);
570	}
571	mutex_unlock(&ig.device_list_mutex);
572}
573
574/**
575 * Called with state mutex held
576 **/
577static int iser_conn_state_comp_exch(struct iser_conn *iser_conn,
578				     enum iser_conn_state comp,
579				     enum iser_conn_state exch)
580{
581	int ret;
582
583	ret = (iser_conn->state == comp);
584	if (ret)
585		iser_conn->state = exch;
586
587	return ret;
588}
589
590void iser_release_work(struct work_struct *work)
591{
592	struct iser_conn *iser_conn;
593
594	iser_conn = container_of(work, struct iser_conn, release_work);
595
596	/* Wait for conn_stop to complete */
597	wait_for_completion(&iser_conn->stop_completion);
598	/* Wait for IB resouces cleanup to complete */
599	wait_for_completion(&iser_conn->ib_completion);
600
601	mutex_lock(&iser_conn->state_mutex);
602	iser_conn->state = ISER_CONN_DOWN;
603	mutex_unlock(&iser_conn->state_mutex);
604
605	iser_conn_release(iser_conn);
606}
607
608/**
609 * iser_free_ib_conn_res - release IB related resources
610 * @iser_conn: iser connection struct
611 * @destroy: indicator if we need to try to release the
612 *     iser device and memory regoins pool (only iscsi
613 *     shutdown and DEVICE_REMOVAL will use this).
614 *
615 * This routine is called with the iser state mutex held
616 * so the cm_id removal is out of here. It is Safe to
617 * be invoked multiple times.
618 */
619static void iser_free_ib_conn_res(struct iser_conn *iser_conn,
620				  bool destroy)
621{
622	struct ib_conn *ib_conn = &iser_conn->ib_conn;
623	struct iser_device *device = ib_conn->device;
624
625	iser_info("freeing conn %p cma_id %p qp %p\n",
626		  iser_conn, ib_conn->cma_id, ib_conn->qp);
627
628	if (ib_conn->qp != NULL) {
629		ib_conn->comp->active_qps--;
630		rdma_destroy_qp(ib_conn->cma_id);
631		ib_conn->qp = NULL;
632	}
633
634	if (destroy) {
635		if (iser_conn->rx_descs)
636			iser_free_rx_descriptors(iser_conn);
637
638		if (device != NULL) {
639			iser_device_try_release(device);
640			ib_conn->device = NULL;
641		}
642	}
643}
644
645/**
646 * Frees all conn objects and deallocs conn descriptor
647 */
648void iser_conn_release(struct iser_conn *iser_conn)
649{
650	struct ib_conn *ib_conn = &iser_conn->ib_conn;
651
652	mutex_lock(&ig.connlist_mutex);
653	list_del(&iser_conn->conn_list);
654	mutex_unlock(&ig.connlist_mutex);
655
656	mutex_lock(&iser_conn->state_mutex);
657	/* In case we endup here without ep_disconnect being invoked. */
658	if (iser_conn->state != ISER_CONN_DOWN) {
659		iser_warn("iser conn %p state %d, expected state down.\n",
660			  iser_conn, iser_conn->state);
661		iscsi_destroy_endpoint(iser_conn->ep);
662		iser_conn->state = ISER_CONN_DOWN;
663	}
664	/*
665	 * In case we never got to bind stage, we still need to
666	 * release IB resources (which is safe to call more than once).
667	 */
668	iser_free_ib_conn_res(iser_conn, true);
669	mutex_unlock(&iser_conn->state_mutex);
670
671	if (ib_conn->cma_id != NULL) {
672		rdma_destroy_id(ib_conn->cma_id);
673		ib_conn->cma_id = NULL;
674	}
675
676	kfree(iser_conn);
677}
678
679/**
680 * triggers start of the disconnect procedures and wait for them to be done
681 * Called with state mutex held
682 */
683int iser_conn_terminate(struct iser_conn *iser_conn)
684{
685	struct ib_conn *ib_conn = &iser_conn->ib_conn;
686	struct ib_send_wr *bad_wr;
687	int err = 0;
688
689	/* terminate the iser conn only if the conn state is UP */
690	if (!iser_conn_state_comp_exch(iser_conn, ISER_CONN_UP,
691				       ISER_CONN_TERMINATING))
692		return 0;
693
694	iser_info("iser_conn %p state %d\n", iser_conn, iser_conn->state);
695
696	/* suspend queuing of new iscsi commands */
697	if (iser_conn->iscsi_conn)
698		iscsi_suspend_queue(iser_conn->iscsi_conn);
699
700	/*
701	 * In case we didn't already clean up the cma_id (peer initiated
702	 * a disconnection), we need to Cause the CMA to change the QP
703	 * state to ERROR.
704	 */
705	if (ib_conn->cma_id) {
706		err = rdma_disconnect(ib_conn->cma_id);
707		if (err)
708			iser_err("Failed to disconnect, conn: 0x%p err %d\n",
709				 iser_conn, err);
710
711		/* post an indication that all flush errors were consumed */
712		err = ib_post_send(ib_conn->qp, &ib_conn->beacon, &bad_wr);
713		if (err) {
714			iser_err("conn %p failed to post beacon", ib_conn);
715			return 1;
716		}
717
718		wait_for_completion(&ib_conn->flush_comp);
719	}
720
721	return 1;
722}
723
724/**
725 * Called with state mutex held
726 **/
727static void iser_connect_error(struct rdma_cm_id *cma_id)
728{
729	struct iser_conn *iser_conn;
730
731	iser_conn = (struct iser_conn *)cma_id->context;
732	iser_conn->state = ISER_CONN_TERMINATING;
733}
734
735/**
736 * Called with state mutex held
737 **/
738static void iser_addr_handler(struct rdma_cm_id *cma_id)
739{
740	struct iser_device *device;
741	struct iser_conn   *iser_conn;
742	struct ib_conn   *ib_conn;
743	int    ret;
744
745	iser_conn = (struct iser_conn *)cma_id->context;
746	if (iser_conn->state != ISER_CONN_PENDING)
747		/* bailout */
748		return;
749
750	ib_conn = &iser_conn->ib_conn;
751	device = iser_device_find_by_ib_device(cma_id);
752	if (!device) {
753		iser_err("device lookup/creation failed\n");
754		iser_connect_error(cma_id);
755		return;
756	}
757
758	ib_conn->device = device;
759
760	/* connection T10-PI support */
761	if (iser_pi_enable) {
762		if (!(device->dev_attr.device_cap_flags &
763		      IB_DEVICE_SIGNATURE_HANDOVER)) {
764			iser_warn("T10-PI requested but not supported on %s, "
765				  "continue without T10-PI\n",
766				  ib_conn->device->ib_device->name);
767			ib_conn->pi_support = false;
768		} else {
769			ib_conn->pi_support = true;
770		}
771	}
772
773	ret = rdma_resolve_route(cma_id, 1000);
774	if (ret) {
775		iser_err("resolve route failed: %d\n", ret);
776		iser_connect_error(cma_id);
777		return;
778	}
779}
780
781/**
782 * Called with state mutex held
783 **/
784static void iser_route_handler(struct rdma_cm_id *cma_id)
785{
786	struct rdma_conn_param conn_param;
787	int    ret;
788	struct iser_cm_hdr req_hdr;
789	struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
790	struct ib_conn *ib_conn = &iser_conn->ib_conn;
791	struct iser_device *device = ib_conn->device;
792
793	if (iser_conn->state != ISER_CONN_PENDING)
794		/* bailout */
795		return;
796
797	ret = iser_create_ib_conn_res(ib_conn);
798	if (ret)
799		goto failure;
800
801	memset(&conn_param, 0, sizeof conn_param);
802	conn_param.responder_resources = device->dev_attr.max_qp_rd_atom;
803	conn_param.initiator_depth     = 1;
804	conn_param.retry_count	       = 7;
805	conn_param.rnr_retry_count     = 6;
806
807	memset(&req_hdr, 0, sizeof(req_hdr));
808	req_hdr.flags = (ISER_ZBVA_NOT_SUPPORTED |
809			ISER_SEND_W_INV_NOT_SUPPORTED);
810	conn_param.private_data		= (void *)&req_hdr;
811	conn_param.private_data_len	= sizeof(struct iser_cm_hdr);
812
813	ret = rdma_connect(cma_id, &conn_param);
814	if (ret) {
815		iser_err("failure connecting: %d\n", ret);
816		goto failure;
817	}
818
819	return;
820failure:
821	iser_connect_error(cma_id);
822}
823
824static void iser_connected_handler(struct rdma_cm_id *cma_id)
825{
826	struct iser_conn *iser_conn;
827	struct ib_qp_attr attr;
828	struct ib_qp_init_attr init_attr;
829
830	iser_conn = (struct iser_conn *)cma_id->context;
831	if (iser_conn->state != ISER_CONN_PENDING)
832		/* bailout */
833		return;
834
835	(void)ib_query_qp(cma_id->qp, &attr, ~0, &init_attr);
836	iser_info("remote qpn:%x my qpn:%x\n", attr.dest_qp_num, cma_id->qp->qp_num);
837
838	iser_conn->state = ISER_CONN_UP;
839	complete(&iser_conn->up_completion);
840}
841
842static void iser_disconnected_handler(struct rdma_cm_id *cma_id)
843{
844	struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
845
846	if (iser_conn_terminate(iser_conn)) {
847		if (iser_conn->iscsi_conn)
848			iscsi_conn_failure(iser_conn->iscsi_conn,
849					   ISCSI_ERR_CONN_FAILED);
850		else
851			iser_err("iscsi_iser connection isn't bound\n");
852	}
853}
854
855static void iser_cleanup_handler(struct rdma_cm_id *cma_id,
856				 bool destroy)
857{
858	struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
859
860	/*
861	 * We are not guaranteed that we visited disconnected_handler
862	 * by now, call it here to be safe that we handle CM drep
863	 * and flush errors.
864	 */
865	iser_disconnected_handler(cma_id);
866	iser_free_ib_conn_res(iser_conn, destroy);
867	complete(&iser_conn->ib_completion);
868};
869
870static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
871{
872	struct iser_conn *iser_conn;
873	int ret = 0;
874
875	iser_conn = (struct iser_conn *)cma_id->context;
876	iser_info("event %d status %d conn %p id %p\n",
877		  event->event, event->status, cma_id->context, cma_id);
878
879	mutex_lock(&iser_conn->state_mutex);
880	switch (event->event) {
881	case RDMA_CM_EVENT_ADDR_RESOLVED:
882		iser_addr_handler(cma_id);
883		break;
884	case RDMA_CM_EVENT_ROUTE_RESOLVED:
885		iser_route_handler(cma_id);
886		break;
887	case RDMA_CM_EVENT_ESTABLISHED:
888		iser_connected_handler(cma_id);
889		break;
890	case RDMA_CM_EVENT_ADDR_ERROR:
891	case RDMA_CM_EVENT_ROUTE_ERROR:
892	case RDMA_CM_EVENT_CONNECT_ERROR:
893	case RDMA_CM_EVENT_UNREACHABLE:
894	case RDMA_CM_EVENT_REJECTED:
895		iser_connect_error(cma_id);
896		break;
897	case RDMA_CM_EVENT_DISCONNECTED:
898	case RDMA_CM_EVENT_ADDR_CHANGE:
899	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
900		iser_cleanup_handler(cma_id, false);
901		break;
902	case RDMA_CM_EVENT_DEVICE_REMOVAL:
903		/*
904		 * we *must* destroy the device as we cannot rely
905		 * on iscsid to be around to initiate error handling.
906		 * also if we are not in state DOWN implicitly destroy
907		 * the cma_id.
908		 */
909		iser_cleanup_handler(cma_id, true);
910		if (iser_conn->state != ISER_CONN_DOWN) {
911			iser_conn->ib_conn.cma_id = NULL;
912			ret = 1;
913		}
914		break;
915	default:
916		iser_err("Unexpected RDMA CM event (%d)\n", event->event);
917		break;
918	}
919	mutex_unlock(&iser_conn->state_mutex);
920
921	return ret;
922}
923
924void iser_conn_init(struct iser_conn *iser_conn)
925{
926	iser_conn->state = ISER_CONN_INIT;
927	iser_conn->ib_conn.post_recv_buf_count = 0;
928	init_completion(&iser_conn->ib_conn.flush_comp);
929	init_completion(&iser_conn->stop_completion);
930	init_completion(&iser_conn->ib_completion);
931	init_completion(&iser_conn->up_completion);
932	INIT_LIST_HEAD(&iser_conn->conn_list);
933	spin_lock_init(&iser_conn->ib_conn.lock);
934	mutex_init(&iser_conn->state_mutex);
935}
936
937 /**
938 * starts the process of connecting to the target
939 * sleeps until the connection is established or rejected
940 */
941int iser_connect(struct iser_conn   *iser_conn,
942		 struct sockaddr    *src_addr,
943		 struct sockaddr    *dst_addr,
944		 int                 non_blocking)
945{
946	struct ib_conn *ib_conn = &iser_conn->ib_conn;
947	int err = 0;
948
949	mutex_lock(&iser_conn->state_mutex);
950
951	sprintf(iser_conn->name, "%pISp", dst_addr);
952
953	iser_info("connecting to: %s\n", iser_conn->name);
954
955	/* the device is known only --after-- address resolution */
956	ib_conn->device = NULL;
957
958	iser_conn->state = ISER_CONN_PENDING;
959
960	ib_conn->beacon.wr_id = ISER_BEACON_WRID;
961	ib_conn->beacon.opcode = IB_WR_SEND;
962
963	ib_conn->cma_id = rdma_create_id(iser_cma_handler,
964					 (void *)iser_conn,
965					 RDMA_PS_TCP, IB_QPT_RC);
966	if (IS_ERR(ib_conn->cma_id)) {
967		err = PTR_ERR(ib_conn->cma_id);
968		iser_err("rdma_create_id failed: %d\n", err);
969		goto id_failure;
970	}
971
972	err = rdma_resolve_addr(ib_conn->cma_id, src_addr, dst_addr, 1000);
973	if (err) {
974		iser_err("rdma_resolve_addr failed: %d\n", err);
975		goto addr_failure;
976	}
977
978	if (!non_blocking) {
979		wait_for_completion_interruptible(&iser_conn->up_completion);
980
981		if (iser_conn->state != ISER_CONN_UP) {
982			err =  -EIO;
983			goto connect_failure;
984		}
985	}
986	mutex_unlock(&iser_conn->state_mutex);
987
988	mutex_lock(&ig.connlist_mutex);
989	list_add(&iser_conn->conn_list, &ig.connlist);
990	mutex_unlock(&ig.connlist_mutex);
991	return 0;
992
993id_failure:
994	ib_conn->cma_id = NULL;
995addr_failure:
996	iser_conn->state = ISER_CONN_DOWN;
997connect_failure:
998	mutex_unlock(&iser_conn->state_mutex);
999	iser_conn_release(iser_conn);
1000	return err;
1001}
1002
1003int iser_post_recvl(struct iser_conn *iser_conn)
1004{
1005	struct ib_recv_wr rx_wr, *rx_wr_failed;
1006	struct ib_conn *ib_conn = &iser_conn->ib_conn;
1007	struct ib_sge	  sge;
1008	int ib_ret;
1009
1010	sge.addr   = iser_conn->login_resp_dma;
1011	sge.length = ISER_RX_LOGIN_SIZE;
1012	sge.lkey   = ib_conn->device->mr->lkey;
1013
1014	rx_wr.wr_id   = (uintptr_t)iser_conn->login_resp_buf;
1015	rx_wr.sg_list = &sge;
1016	rx_wr.num_sge = 1;
1017	rx_wr.next    = NULL;
1018
1019	ib_conn->post_recv_buf_count++;
1020	ib_ret	= ib_post_recv(ib_conn->qp, &rx_wr, &rx_wr_failed);
1021	if (ib_ret) {
1022		iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1023		ib_conn->post_recv_buf_count--;
1024	}
1025	return ib_ret;
1026}
1027
1028int iser_post_recvm(struct iser_conn *iser_conn, int count)
1029{
1030	struct ib_recv_wr *rx_wr, *rx_wr_failed;
1031	int i, ib_ret;
1032	struct ib_conn *ib_conn = &iser_conn->ib_conn;
1033	unsigned int my_rx_head = iser_conn->rx_desc_head;
1034	struct iser_rx_desc *rx_desc;
1035
1036	for (rx_wr = ib_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
1037		rx_desc		= &iser_conn->rx_descs[my_rx_head];
1038		rx_wr->wr_id	= (uintptr_t)rx_desc;
1039		rx_wr->sg_list	= &rx_desc->rx_sg;
1040		rx_wr->num_sge	= 1;
1041		rx_wr->next	= rx_wr + 1;
1042		my_rx_head = (my_rx_head + 1) & iser_conn->qp_max_recv_dtos_mask;
1043	}
1044
1045	rx_wr--;
1046	rx_wr->next = NULL; /* mark end of work requests list */
1047
1048	ib_conn->post_recv_buf_count += count;
1049	ib_ret	= ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &rx_wr_failed);
1050	if (ib_ret) {
1051		iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1052		ib_conn->post_recv_buf_count -= count;
1053	} else
1054		iser_conn->rx_desc_head = my_rx_head;
1055	return ib_ret;
1056}
1057
1058
1059/**
1060 * iser_start_send - Initiate a Send DTO operation
1061 *
1062 * returns 0 on success, -1 on failure
1063 */
1064int iser_post_send(struct ib_conn *ib_conn, struct iser_tx_desc *tx_desc,
1065		   bool signal)
1066{
1067	int		  ib_ret;
1068	struct ib_send_wr send_wr, *send_wr_failed;
1069
1070	ib_dma_sync_single_for_device(ib_conn->device->ib_device,
1071				      tx_desc->dma_addr, ISER_HEADERS_LEN,
1072				      DMA_TO_DEVICE);
1073
1074	send_wr.next	   = NULL;
1075	send_wr.wr_id	   = (uintptr_t)tx_desc;
1076	send_wr.sg_list	   = tx_desc->tx_sg;
1077	send_wr.num_sge	   = tx_desc->num_sge;
1078	send_wr.opcode	   = IB_WR_SEND;
1079	send_wr.send_flags = signal ? IB_SEND_SIGNALED : 0;
1080
1081	ib_ret = ib_post_send(ib_conn->qp, &send_wr, &send_wr_failed);
1082	if (ib_ret)
1083		iser_err("ib_post_send failed, ret:%d\n", ib_ret);
1084
1085	return ib_ret;
1086}
1087
1088/**
1089 * is_iser_tx_desc - Indicate if the completion wr_id
1090 *     is a TX descriptor or not.
1091 * @iser_conn: iser connection
1092 * @wr_id: completion WR identifier
1093 *
1094 * Since we cannot rely on wc opcode in FLUSH errors
1095 * we must work around it by checking if the wr_id address
1096 * falls in the iser connection rx_descs buffer. If so
1097 * it is an RX descriptor, otherwize it is a TX.
1098 */
1099static inline bool
1100is_iser_tx_desc(struct iser_conn *iser_conn, void *wr_id)
1101{
1102	void *start = iser_conn->rx_descs;
1103	int len = iser_conn->num_rx_descs * sizeof(*iser_conn->rx_descs);
1104
1105	if (wr_id >= start && wr_id < start + len)
1106		return false;
1107
1108	return true;
1109}
1110
1111/**
1112 * iser_handle_comp_error() - Handle error completion
1113 * @ib_conn:   connection RDMA resources
1114 * @wc:        work completion
1115 *
1116 * Notes: We may handle a FLUSH error completion and in this case
1117 *        we only cleanup in case TX type was DATAOUT. For non-FLUSH
1118 *        error completion we should also notify iscsi layer that
1119 *        connection is failed (in case we passed bind stage).
1120 */
1121static void
1122iser_handle_comp_error(struct ib_conn *ib_conn,
1123		       struct ib_wc *wc)
1124{
1125	void *wr_id = (void *)(uintptr_t)wc->wr_id;
1126	struct iser_conn *iser_conn = container_of(ib_conn, struct iser_conn,
1127						   ib_conn);
1128
1129	if (wc->status != IB_WC_WR_FLUSH_ERR)
1130		if (iser_conn->iscsi_conn)
1131			iscsi_conn_failure(iser_conn->iscsi_conn,
1132					   ISCSI_ERR_CONN_FAILED);
1133
1134	if (wc->wr_id == ISER_FASTREG_LI_WRID)
1135		return;
1136
1137	if (is_iser_tx_desc(iser_conn, wr_id)) {
1138		struct iser_tx_desc *desc = wr_id;
1139
1140		if (desc->type == ISCSI_TX_DATAOUT)
1141			kmem_cache_free(ig.desc_cache, desc);
1142	} else {
1143		ib_conn->post_recv_buf_count--;
1144	}
1145}
1146
1147/**
1148 * iser_handle_wc - handle a single work completion
1149 * @wc: work completion
1150 *
1151 * Soft-IRQ context, work completion can be either
1152 * SEND or RECV, and can turn out successful or
1153 * with error (or flush error).
1154 */
1155static void iser_handle_wc(struct ib_wc *wc)
1156{
1157	struct ib_conn *ib_conn;
1158	struct iser_tx_desc *tx_desc;
1159	struct iser_rx_desc *rx_desc;
1160
1161	ib_conn = wc->qp->qp_context;
1162	if (likely(wc->status == IB_WC_SUCCESS)) {
1163		if (wc->opcode == IB_WC_RECV) {
1164			rx_desc = (struct iser_rx_desc *)(uintptr_t)wc->wr_id;
1165			iser_rcv_completion(rx_desc, wc->byte_len,
1166					    ib_conn);
1167		} else
1168		if (wc->opcode == IB_WC_SEND) {
1169			tx_desc = (struct iser_tx_desc *)(uintptr_t)wc->wr_id;
1170			iser_snd_completion(tx_desc, ib_conn);
1171		} else {
1172			iser_err("Unknown wc opcode %d\n", wc->opcode);
1173		}
1174	} else {
1175		if (wc->status != IB_WC_WR_FLUSH_ERR)
1176			iser_err("wr id %llx status %d vend_err %x\n",
1177				 wc->wr_id, wc->status, wc->vendor_err);
1178		else
1179			iser_dbg("flush error: wr id %llx\n", wc->wr_id);
1180
1181		if (wc->wr_id == ISER_BEACON_WRID)
1182			/* all flush errors were consumed */
1183			complete(&ib_conn->flush_comp);
1184		else
1185			iser_handle_comp_error(ib_conn, wc);
1186	}
1187}
1188
1189/**
1190 * iser_cq_tasklet_fn - iSER completion polling loop
1191 * @data: iSER completion context
1192 *
1193 * Soft-IRQ context, polling connection CQ until
1194 * either CQ was empty or we exausted polling budget
1195 */
1196static void iser_cq_tasklet_fn(unsigned long data)
1197{
1198	struct iser_comp *comp = (struct iser_comp *)data;
1199	struct ib_cq *cq = comp->cq;
1200	struct ib_wc *const wcs = comp->wcs;
1201	int i, n, completed = 0;
1202
1203	while ((n = ib_poll_cq(cq, ARRAY_SIZE(comp->wcs), wcs)) > 0) {
1204		for (i = 0; i < n; i++)
1205			iser_handle_wc(&wcs[i]);
1206
1207		completed += n;
1208		if (completed >= iser_cq_poll_limit)
1209			break;
1210	}
1211
1212	/*
1213	 * It is assumed here that arming CQ only once its empty
1214	 * would not cause interrupts to be missed.
1215	 */
1216	ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1217
1218	iser_dbg("got %d completions\n", completed);
1219}
1220
1221static void iser_cq_callback(struct ib_cq *cq, void *cq_context)
1222{
1223	struct iser_comp *comp = cq_context;
1224
1225	tasklet_schedule(&comp->tasklet);
1226}
1227
1228u8 iser_check_task_pi_status(struct iscsi_iser_task *iser_task,
1229			     enum iser_data_dir cmd_dir, sector_t *sector)
1230{
1231	struct iser_mem_reg *reg = &iser_task->rdma_reg[cmd_dir];
1232	struct fast_reg_descriptor *desc = reg->mem_h;
1233	unsigned long sector_size = iser_task->sc->device->sector_size;
1234	struct ib_mr_status mr_status;
1235	int ret;
1236
1237	if (desc && desc->reg_indicators & ISER_FASTREG_PROTECTED) {
1238		desc->reg_indicators &= ~ISER_FASTREG_PROTECTED;
1239		ret = ib_check_mr_status(desc->pi_ctx->sig_mr,
1240					 IB_MR_CHECK_SIG_STATUS, &mr_status);
1241		if (ret) {
1242			pr_err("ib_check_mr_status failed, ret %d\n", ret);
1243			goto err;
1244		}
1245
1246		if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1247			sector_t sector_off = mr_status.sig_err.sig_err_offset;
1248
1249			do_div(sector_off, sector_size + 8);
1250			*sector = scsi_get_lba(iser_task->sc) + sector_off;
1251
1252			pr_err("PI error found type %d at sector %llx "
1253			       "expected %x vs actual %x\n",
1254			       mr_status.sig_err.err_type,
1255			       (unsigned long long)*sector,
1256			       mr_status.sig_err.expected,
1257			       mr_status.sig_err.actual);
1258
1259			switch (mr_status.sig_err.err_type) {
1260			case IB_SIG_BAD_GUARD:
1261				return 0x1;
1262			case IB_SIG_BAD_REFTAG:
1263				return 0x3;
1264			case IB_SIG_BAD_APPTAG:
1265				return 0x2;
1266			}
1267		}
1268	}
1269
1270	return 0;
1271err:
1272	/* Not alot we can do here, return ambiguous guard error */
1273	return 0x1;
1274}
1275