IB/iser: Fix responder resources advertisement
[linux-2.6-block.git] / drivers / infiniband / ulp / iser / iser_verbs.c
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_CQ_LEN      (ISER_QP_MAX_RECV_DTOS * ISCSI_ISER_MAX_CONN)
43 #define ISER_MAX_TX_CQ_LEN      (ISER_QP_MAX_REQ_DTOS  * ISCSI_ISER_MAX_CONN)
44
45 static void iser_cq_tasklet_fn(unsigned long data);
46 static void iser_cq_callback(struct ib_cq *cq, void *cq_context);
47
48 static void iser_cq_event_callback(struct ib_event *cause, void *context)
49 {
50         iser_err("got cq event %d \n", cause->event);
51 }
52
53 static void iser_qp_event_callback(struct ib_event *cause, void *context)
54 {
55         iser_err("got qp event %d\n",cause->event);
56 }
57
58 static void iser_event_handler(struct ib_event_handler *handler,
59                                 struct ib_event *event)
60 {
61         iser_err("async event %d on device %s port %d\n", event->event,
62                 event->device->name, event->element.port_num);
63 }
64
65 /**
66  * iser_create_device_ib_res - creates Protection Domain (PD), Completion
67  * Queue (CQ), DMA Memory Region (DMA MR) with the device associated with
68  * the adapator.
69  *
70  * returns 0 on success, -1 on failure
71  */
72 static int iser_create_device_ib_res(struct iser_device *device)
73 {
74         struct iser_cq_desc *cq_desc;
75         struct ib_device_attr *dev_attr = &device->dev_attr;
76         int ret, i, j;
77
78         ret = ib_query_device(device->ib_device, dev_attr);
79         if (ret) {
80                 pr_warn("Query device failed for %s\n", device->ib_device->name);
81                 return ret;
82         }
83
84         /* Assign function handles  - based on FMR support */
85         if (device->ib_device->alloc_fmr && device->ib_device->dealloc_fmr &&
86             device->ib_device->map_phys_fmr && device->ib_device->unmap_fmr) {
87                 iser_info("FMR supported, using FMR for registration\n");
88                 device->iser_alloc_rdma_reg_res = iser_create_fmr_pool;
89                 device->iser_free_rdma_reg_res = iser_free_fmr_pool;
90                 device->iser_reg_rdma_mem = iser_reg_rdma_mem_fmr;
91                 device->iser_unreg_rdma_mem = iser_unreg_mem_fmr;
92         } else
93         if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
94                 iser_info("FastReg supported, using FastReg for registration\n");
95                 device->iser_alloc_rdma_reg_res = iser_create_fastreg_pool;
96                 device->iser_free_rdma_reg_res = iser_free_fastreg_pool;
97                 device->iser_reg_rdma_mem = iser_reg_rdma_mem_fastreg;
98                 device->iser_unreg_rdma_mem = iser_unreg_mem_fastreg;
99         } else {
100                 iser_err("IB device does not support FMRs nor FastRegs, can't register memory\n");
101                 return -1;
102         }
103
104         device->cqs_used = min(ISER_MAX_CQ, device->ib_device->num_comp_vectors);
105         iser_info("using %d CQs, device %s supports %d vectors\n",
106                   device->cqs_used, device->ib_device->name,
107                   device->ib_device->num_comp_vectors);
108
109         device->cq_desc = kmalloc(sizeof(struct iser_cq_desc) * device->cqs_used,
110                                   GFP_KERNEL);
111         if (device->cq_desc == NULL)
112                 goto cq_desc_err;
113         cq_desc = device->cq_desc;
114
115         device->pd = ib_alloc_pd(device->ib_device);
116         if (IS_ERR(device->pd))
117                 goto pd_err;
118
119         for (i = 0; i < device->cqs_used; i++) {
120                 cq_desc[i].device   = device;
121                 cq_desc[i].cq_index = i;
122
123                 device->rx_cq[i] = ib_create_cq(device->ib_device,
124                                           iser_cq_callback,
125                                           iser_cq_event_callback,
126                                           (void *)&cq_desc[i],
127                                           ISER_MAX_RX_CQ_LEN, i);
128                 if (IS_ERR(device->rx_cq[i]))
129                         goto cq_err;
130
131                 device->tx_cq[i] = ib_create_cq(device->ib_device,
132                                           NULL, iser_cq_event_callback,
133                                           (void *)&cq_desc[i],
134                                           ISER_MAX_TX_CQ_LEN, i);
135
136                 if (IS_ERR(device->tx_cq[i]))
137                         goto cq_err;
138
139                 if (ib_req_notify_cq(device->rx_cq[i], IB_CQ_NEXT_COMP))
140                         goto cq_err;
141
142                 tasklet_init(&device->cq_tasklet[i],
143                              iser_cq_tasklet_fn,
144                         (unsigned long)&cq_desc[i]);
145         }
146
147         device->mr = ib_get_dma_mr(device->pd, IB_ACCESS_LOCAL_WRITE |
148                                    IB_ACCESS_REMOTE_WRITE |
149                                    IB_ACCESS_REMOTE_READ);
150         if (IS_ERR(device->mr))
151                 goto dma_mr_err;
152
153         INIT_IB_EVENT_HANDLER(&device->event_handler, device->ib_device,
154                                 iser_event_handler);
155         if (ib_register_event_handler(&device->event_handler))
156                 goto handler_err;
157
158         return 0;
159
160 handler_err:
161         ib_dereg_mr(device->mr);
162 dma_mr_err:
163         for (j = 0; j < device->cqs_used; j++)
164                 tasklet_kill(&device->cq_tasklet[j]);
165 cq_err:
166         for (j = 0; j < i; j++) {
167                 if (device->tx_cq[j])
168                         ib_destroy_cq(device->tx_cq[j]);
169                 if (device->rx_cq[j])
170                         ib_destroy_cq(device->rx_cq[j]);
171         }
172         ib_dealloc_pd(device->pd);
173 pd_err:
174         kfree(device->cq_desc);
175 cq_desc_err:
176         iser_err("failed to allocate an IB resource\n");
177         return -1;
178 }
179
180 /**
181  * iser_free_device_ib_res - destroy/dealloc/dereg the DMA MR,
182  * CQ and PD created with the device associated with the adapator.
183  */
184 static void iser_free_device_ib_res(struct iser_device *device)
185 {
186         int i;
187         BUG_ON(device->mr == NULL);
188
189         for (i = 0; i < device->cqs_used; i++) {
190                 tasklet_kill(&device->cq_tasklet[i]);
191                 (void)ib_destroy_cq(device->tx_cq[i]);
192                 (void)ib_destroy_cq(device->rx_cq[i]);
193                 device->tx_cq[i] = NULL;
194                 device->rx_cq[i] = NULL;
195         }
196
197         (void)ib_unregister_event_handler(&device->event_handler);
198         (void)ib_dereg_mr(device->mr);
199         (void)ib_dealloc_pd(device->pd);
200
201         kfree(device->cq_desc);
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  */
212 int iser_create_fmr_pool(struct iser_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  */
262 void iser_free_fmr_pool(struct iser_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
276 static int
277 iser_create_fastreg_desc(struct ib_device *ib_device, struct ib_pd *pd,
278                          bool pi_enable, struct fast_reg_descriptor *desc)
279 {
280         int ret;
281
282         desc->data_frpl = ib_alloc_fast_reg_page_list(ib_device,
283                                                       ISCSI_ISER_SG_TABLESIZE + 1);
284         if (IS_ERR(desc->data_frpl)) {
285                 ret = PTR_ERR(desc->data_frpl);
286                 iser_err("Failed to allocate ib_fast_reg_page_list err=%d\n",
287                          ret);
288                 return PTR_ERR(desc->data_frpl);
289         }
290
291         desc->data_mr = ib_alloc_fast_reg_mr(pd, ISCSI_ISER_SG_TABLESIZE + 1);
292         if (IS_ERR(desc->data_mr)) {
293                 ret = PTR_ERR(desc->data_mr);
294                 iser_err("Failed to allocate ib_fast_reg_mr err=%d\n", ret);
295                 goto fast_reg_mr_failure;
296         }
297         desc->reg_indicators |= ISER_DATA_KEY_VALID;
298
299         if (pi_enable) {
300                 struct ib_mr_init_attr mr_init_attr = {0};
301                 struct iser_pi_context *pi_ctx = NULL;
302
303                 desc->pi_ctx = kzalloc(sizeof(*desc->pi_ctx), GFP_KERNEL);
304                 if (!desc->pi_ctx) {
305                         iser_err("Failed to allocate pi context\n");
306                         ret = -ENOMEM;
307                         goto pi_ctx_alloc_failure;
308                 }
309                 pi_ctx = desc->pi_ctx;
310
311                 pi_ctx->prot_frpl = ib_alloc_fast_reg_page_list(ib_device,
312                                                     ISCSI_ISER_SG_TABLESIZE);
313                 if (IS_ERR(pi_ctx->prot_frpl)) {
314                         ret = PTR_ERR(pi_ctx->prot_frpl);
315                         iser_err("Failed to allocate prot frpl ret=%d\n",
316                                  ret);
317                         goto prot_frpl_failure;
318                 }
319
320                 pi_ctx->prot_mr = ib_alloc_fast_reg_mr(pd,
321                                                 ISCSI_ISER_SG_TABLESIZE + 1);
322                 if (IS_ERR(pi_ctx->prot_mr)) {
323                         ret = PTR_ERR(pi_ctx->prot_mr);
324                         iser_err("Failed to allocate prot frmr ret=%d\n",
325                                  ret);
326                         goto prot_mr_failure;
327                 }
328                 desc->reg_indicators |= ISER_PROT_KEY_VALID;
329
330                 mr_init_attr.max_reg_descriptors = 2;
331                 mr_init_attr.flags |= IB_MR_SIGNATURE_EN;
332                 pi_ctx->sig_mr = ib_create_mr(pd, &mr_init_attr);
333                 if (IS_ERR(pi_ctx->sig_mr)) {
334                         ret = PTR_ERR(pi_ctx->sig_mr);
335                         iser_err("Failed to allocate signature enabled mr err=%d\n",
336                                  ret);
337                         goto sig_mr_failure;
338                 }
339                 desc->reg_indicators |= ISER_SIG_KEY_VALID;
340         }
341         desc->reg_indicators &= ~ISER_FASTREG_PROTECTED;
342
343         iser_dbg("Create fr_desc %p page_list %p\n",
344                  desc, desc->data_frpl->page_list);
345
346         return 0;
347 sig_mr_failure:
348         ib_dereg_mr(desc->pi_ctx->prot_mr);
349 prot_mr_failure:
350         ib_free_fast_reg_page_list(desc->pi_ctx->prot_frpl);
351 prot_frpl_failure:
352         kfree(desc->pi_ctx);
353 pi_ctx_alloc_failure:
354         ib_dereg_mr(desc->data_mr);
355 fast_reg_mr_failure:
356         ib_free_fast_reg_page_list(desc->data_frpl);
357
358         return ret;
359 }
360
361 /**
362  * iser_create_fastreg_pool - Creates pool of fast_reg descriptors
363  * for fast registration work requests.
364  * returns 0 on success, or errno code on failure
365  */
366 int iser_create_fastreg_pool(struct iser_conn *ib_conn, unsigned cmds_max)
367 {
368         struct iser_device      *device = ib_conn->device;
369         struct fast_reg_descriptor      *desc;
370         int i, ret;
371
372         INIT_LIST_HEAD(&ib_conn->fastreg.pool);
373         ib_conn->fastreg.pool_size = 0;
374         for (i = 0; i < cmds_max; i++) {
375                 desc = kzalloc(sizeof(*desc), GFP_KERNEL);
376                 if (!desc) {
377                         iser_err("Failed to allocate a new fast_reg descriptor\n");
378                         ret = -ENOMEM;
379                         goto err;
380                 }
381
382                 ret = iser_create_fastreg_desc(device->ib_device, device->pd,
383                                                ib_conn->pi_support, desc);
384                 if (ret) {
385                         iser_err("Failed to create fastreg descriptor err=%d\n",
386                                  ret);
387                         kfree(desc);
388                         goto err;
389                 }
390
391                 list_add_tail(&desc->list, &ib_conn->fastreg.pool);
392                 ib_conn->fastreg.pool_size++;
393         }
394
395         return 0;
396
397 err:
398         iser_free_fastreg_pool(ib_conn);
399         return ret;
400 }
401
402 /**
403  * iser_free_fastreg_pool - releases the pool of fast_reg descriptors
404  */
405 void iser_free_fastreg_pool(struct iser_conn *ib_conn)
406 {
407         struct fast_reg_descriptor *desc, *tmp;
408         int i = 0;
409
410         if (list_empty(&ib_conn->fastreg.pool))
411                 return;
412
413         iser_info("freeing conn %p fr pool\n", ib_conn);
414
415         list_for_each_entry_safe(desc, tmp, &ib_conn->fastreg.pool, list) {
416                 list_del(&desc->list);
417                 ib_free_fast_reg_page_list(desc->data_frpl);
418                 ib_dereg_mr(desc->data_mr);
419                 if (desc->pi_ctx) {
420                         ib_free_fast_reg_page_list(desc->pi_ctx->prot_frpl);
421                         ib_dereg_mr(desc->pi_ctx->prot_mr);
422                         ib_destroy_mr(desc->pi_ctx->sig_mr);
423                         kfree(desc->pi_ctx);
424                 }
425                 kfree(desc);
426                 ++i;
427         }
428
429         if (i < ib_conn->fastreg.pool_size)
430                 iser_warn("pool still has %d regions registered\n",
431                           ib_conn->fastreg.pool_size - i);
432 }
433
434 /**
435  * iser_create_ib_conn_res - Queue-Pair (QP)
436  *
437  * returns 0 on success, -1 on failure
438  */
439 static int iser_create_ib_conn_res(struct iser_conn *ib_conn)
440 {
441         struct iser_device      *device;
442         struct ib_qp_init_attr  init_attr;
443         int                     ret = -ENOMEM;
444         int index, min_index = 0;
445
446         BUG_ON(ib_conn->device == NULL);
447
448         device = ib_conn->device;
449
450         memset(&init_attr, 0, sizeof init_attr);
451
452         mutex_lock(&ig.connlist_mutex);
453         /* select the CQ with the minimal number of usages */
454         for (index = 0; index < device->cqs_used; index++)
455                 if (device->cq_active_qps[index] <
456                     device->cq_active_qps[min_index])
457                         min_index = index;
458         device->cq_active_qps[min_index]++;
459         mutex_unlock(&ig.connlist_mutex);
460         iser_info("cq index %d used for ib_conn %p\n", min_index, ib_conn);
461
462         init_attr.event_handler = iser_qp_event_callback;
463         init_attr.qp_context    = (void *)ib_conn;
464         init_attr.send_cq       = device->tx_cq[min_index];
465         init_attr.recv_cq       = device->rx_cq[min_index];
466         init_attr.cap.max_recv_wr  = ISER_QP_MAX_RECV_DTOS;
467         init_attr.cap.max_send_sge = 2;
468         init_attr.cap.max_recv_sge = 1;
469         init_attr.sq_sig_type   = IB_SIGNAL_REQ_WR;
470         init_attr.qp_type       = IB_QPT_RC;
471         if (ib_conn->pi_support) {
472                 init_attr.cap.max_send_wr = ISER_QP_SIG_MAX_REQ_DTOS;
473                 init_attr.create_flags |= IB_QP_CREATE_SIGNATURE_EN;
474         } else {
475                 init_attr.cap.max_send_wr  = ISER_QP_MAX_REQ_DTOS;
476         }
477
478         ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
479         if (ret)
480                 goto out_err;
481
482         ib_conn->qp = ib_conn->cma_id->qp;
483         iser_info("setting conn %p cma_id %p qp %p\n",
484                   ib_conn, ib_conn->cma_id,
485                   ib_conn->cma_id->qp);
486         return ret;
487
488 out_err:
489         iser_err("unable to alloc mem or create resource, err %d\n", ret);
490         return ret;
491 }
492
493 /**
494  * releases the QP objects, returns 0 on success,
495  * -1 on failure
496  */
497 static int iser_free_ib_conn_res(struct iser_conn *ib_conn)
498 {
499         int cq_index;
500         BUG_ON(ib_conn == NULL);
501
502         iser_info("freeing conn %p cma_id %p qp %p\n",
503                   ib_conn, ib_conn->cma_id,
504                   ib_conn->qp);
505
506         /* qp is created only once both addr & route are resolved */
507
508         if (ib_conn->qp != NULL) {
509                 cq_index = ((struct iser_cq_desc *)ib_conn->qp->recv_cq->cq_context)->cq_index;
510                 ib_conn->device->cq_active_qps[cq_index]--;
511
512                 rdma_destroy_qp(ib_conn->cma_id);
513         }
514
515         ib_conn->qp       = NULL;
516
517         return 0;
518 }
519
520 /**
521  * based on the resolved device node GUID see if there already allocated
522  * device for this device. If there's no such, create one.
523  */
524 static
525 struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
526 {
527         struct iser_device *device;
528
529         mutex_lock(&ig.device_list_mutex);
530
531         list_for_each_entry(device, &ig.device_list, ig_list)
532                 /* find if there's a match using the node GUID */
533                 if (device->ib_device->node_guid == cma_id->device->node_guid)
534                         goto inc_refcnt;
535
536         device = kzalloc(sizeof *device, GFP_KERNEL);
537         if (device == NULL)
538                 goto out;
539
540         /* assign this device to the device */
541         device->ib_device = cma_id->device;
542         /* init the device and link it into ig device list */
543         if (iser_create_device_ib_res(device)) {
544                 kfree(device);
545                 device = NULL;
546                 goto out;
547         }
548         list_add(&device->ig_list, &ig.device_list);
549
550 inc_refcnt:
551         device->refcount++;
552 out:
553         mutex_unlock(&ig.device_list_mutex);
554         return device;
555 }
556
557 /* if there's no demand for this device, release it */
558 static void iser_device_try_release(struct iser_device *device)
559 {
560         mutex_lock(&ig.device_list_mutex);
561         device->refcount--;
562         iser_info("device %p refcount %d\n", device, device->refcount);
563         if (!device->refcount) {
564                 iser_free_device_ib_res(device);
565                 list_del(&device->ig_list);
566                 kfree(device);
567         }
568         mutex_unlock(&ig.device_list_mutex);
569 }
570
571 static int iser_conn_state_comp_exch(struct iser_conn *ib_conn,
572                                      enum iser_ib_conn_state comp,
573                                      enum iser_ib_conn_state exch)
574 {
575         int ret;
576
577         spin_lock_bh(&ib_conn->lock);
578         if ((ret = (ib_conn->state == comp)))
579                 ib_conn->state = exch;
580         spin_unlock_bh(&ib_conn->lock);
581         return ret;
582 }
583
584 void iser_release_work(struct work_struct *work)
585 {
586         struct iser_conn *ib_conn;
587
588         ib_conn = container_of(work, struct iser_conn, release_work);
589
590         /* wait for .conn_stop callback */
591         wait_for_completion(&ib_conn->stop_completion);
592
593         /* wait for the qp`s post send and post receive buffers to empty */
594         wait_event_interruptible(ib_conn->wait,
595                                  ib_conn->state == ISER_CONN_DOWN);
596
597         iser_conn_release(ib_conn);
598 }
599
600 /**
601  * Frees all conn objects and deallocs conn descriptor
602  */
603 void iser_conn_release(struct iser_conn *ib_conn)
604 {
605         struct iser_device  *device = ib_conn->device;
606
607         BUG_ON(ib_conn->state == ISER_CONN_UP);
608
609         mutex_lock(&ig.connlist_mutex);
610         list_del(&ib_conn->conn_list);
611         mutex_unlock(&ig.connlist_mutex);
612         iser_free_rx_descriptors(ib_conn);
613         iser_free_ib_conn_res(ib_conn);
614         ib_conn->device = NULL;
615         /* on EVENT_ADDR_ERROR there's no device yet for this conn */
616         if (device != NULL)
617                 iser_device_try_release(device);
618         /* if cma handler context, the caller actually destroy the id */
619         if (ib_conn->cma_id != NULL) {
620                 rdma_destroy_id(ib_conn->cma_id);
621                 ib_conn->cma_id = NULL;
622         }
623         iscsi_destroy_endpoint(ib_conn->ep);
624 }
625
626 /**
627  * triggers start of the disconnect procedures and wait for them to be done
628  */
629 void iser_conn_terminate(struct iser_conn *ib_conn)
630 {
631         int err = 0;
632
633         /* change the ib conn state only if the conn is UP, however always call
634          * rdma_disconnect since this is the only way to cause the CMA to change
635          * the QP state to ERROR
636          */
637
638         iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP, ISER_CONN_TERMINATING);
639         err = rdma_disconnect(ib_conn->cma_id);
640         if (err)
641                 iser_err("Failed to disconnect, conn: 0x%p err %d\n",
642                          ib_conn,err);
643 }
644
645 static void iser_connect_error(struct rdma_cm_id *cma_id)
646 {
647         struct iser_conn *ib_conn;
648
649         ib_conn = (struct iser_conn *)cma_id->context;
650
651         ib_conn->state = ISER_CONN_DOWN;
652         wake_up_interruptible(&ib_conn->wait);
653 }
654
655 static void iser_addr_handler(struct rdma_cm_id *cma_id)
656 {
657         struct iser_device *device;
658         struct iser_conn   *ib_conn;
659         int    ret;
660
661         device = iser_device_find_by_ib_device(cma_id);
662         if (!device) {
663                 iser_err("device lookup/creation failed\n");
664                 iser_connect_error(cma_id);
665                 return;
666         }
667
668         ib_conn = (struct iser_conn *)cma_id->context;
669         ib_conn->device = device;
670
671         /* connection T10-PI support */
672         if (iser_pi_enable) {
673                 if (!(device->dev_attr.device_cap_flags &
674                       IB_DEVICE_SIGNATURE_HANDOVER)) {
675                         iser_warn("T10-PI requested but not supported on %s, "
676                                   "continue without T10-PI\n",
677                                   ib_conn->device->ib_device->name);
678                         ib_conn->pi_support = false;
679                 } else {
680                         ib_conn->pi_support = true;
681                 }
682         }
683
684         ret = rdma_resolve_route(cma_id, 1000);
685         if (ret) {
686                 iser_err("resolve route failed: %d\n", ret);
687                 iser_connect_error(cma_id);
688                 return;
689         }
690 }
691
692 static void iser_route_handler(struct rdma_cm_id *cma_id)
693 {
694         struct rdma_conn_param conn_param;
695         int    ret;
696         struct iser_cm_hdr req_hdr;
697         struct iser_conn *ib_conn = (struct iser_conn *)cma_id->context;
698         struct iser_device *device = ib_conn->device;
699
700         ret = iser_create_ib_conn_res((struct iser_conn *)cma_id->context);
701         if (ret)
702                 goto failure;
703
704         memset(&conn_param, 0, sizeof conn_param);
705         conn_param.responder_resources = device->dev_attr.max_qp_rd_atom;
706         conn_param.initiator_depth     = 1;
707         conn_param.retry_count         = 7;
708         conn_param.rnr_retry_count     = 6;
709
710         memset(&req_hdr, 0, sizeof(req_hdr));
711         req_hdr.flags = (ISER_ZBVA_NOT_SUPPORTED |
712                         ISER_SEND_W_INV_NOT_SUPPORTED);
713         conn_param.private_data         = (void *)&req_hdr;
714         conn_param.private_data_len     = sizeof(struct iser_cm_hdr);
715
716         ret = rdma_connect(cma_id, &conn_param);
717         if (ret) {
718                 iser_err("failure connecting: %d\n", ret);
719                 goto failure;
720         }
721
722         return;
723 failure:
724         iser_connect_error(cma_id);
725 }
726
727 static void iser_connected_handler(struct rdma_cm_id *cma_id)
728 {
729         struct iser_conn *ib_conn;
730         struct ib_qp_attr attr;
731         struct ib_qp_init_attr init_attr;
732
733         (void)ib_query_qp(cma_id->qp, &attr, ~0, &init_attr);
734         iser_info("remote qpn:%x my qpn:%x\n", attr.dest_qp_num, cma_id->qp->qp_num);
735
736         ib_conn = (struct iser_conn *)cma_id->context;
737         if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_PENDING, ISER_CONN_UP))
738                 wake_up_interruptible(&ib_conn->wait);
739 }
740
741 static void iser_disconnected_handler(struct rdma_cm_id *cma_id)
742 {
743         struct iser_conn *ib_conn;
744
745         ib_conn = (struct iser_conn *)cma_id->context;
746
747         /* getting here when the state is UP means that the conn is being *
748          * terminated asynchronously from the iSCSI layer's perspective.  */
749         if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
750                                         ISER_CONN_TERMINATING)){
751                 if (ib_conn->iscsi_conn)
752                         iscsi_conn_failure(ib_conn->iscsi_conn, ISCSI_ERR_CONN_FAILED);
753                 else
754                         iser_err("iscsi_iser connection isn't bound\n");
755         }
756
757         /* Complete the termination process if no posts are pending */
758         if (ib_conn->post_recv_buf_count == 0 &&
759             (atomic_read(&ib_conn->post_send_buf_count) == 0)) {
760                 ib_conn->state = ISER_CONN_DOWN;
761                 wake_up_interruptible(&ib_conn->wait);
762         }
763 }
764
765 static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
766 {
767         iser_info("event %d status %d conn %p id %p\n",
768                   event->event, event->status, cma_id->context, cma_id);
769
770         switch (event->event) {
771         case RDMA_CM_EVENT_ADDR_RESOLVED:
772                 iser_addr_handler(cma_id);
773                 break;
774         case RDMA_CM_EVENT_ROUTE_RESOLVED:
775                 iser_route_handler(cma_id);
776                 break;
777         case RDMA_CM_EVENT_ESTABLISHED:
778                 iser_connected_handler(cma_id);
779                 break;
780         case RDMA_CM_EVENT_ADDR_ERROR:
781         case RDMA_CM_EVENT_ROUTE_ERROR:
782         case RDMA_CM_EVENT_CONNECT_ERROR:
783         case RDMA_CM_EVENT_UNREACHABLE:
784         case RDMA_CM_EVENT_REJECTED:
785                 iser_connect_error(cma_id);
786                 break;
787         case RDMA_CM_EVENT_DISCONNECTED:
788         case RDMA_CM_EVENT_DEVICE_REMOVAL:
789         case RDMA_CM_EVENT_ADDR_CHANGE:
790         case RDMA_CM_EVENT_TIMEWAIT_EXIT:
791                 iser_disconnected_handler(cma_id);
792                 break;
793         default:
794                 iser_err("Unexpected RDMA CM event (%d)\n", event->event);
795                 break;
796         }
797         return 0;
798 }
799
800 void iser_conn_init(struct iser_conn *ib_conn)
801 {
802         ib_conn->state = ISER_CONN_INIT;
803         init_waitqueue_head(&ib_conn->wait);
804         ib_conn->post_recv_buf_count = 0;
805         atomic_set(&ib_conn->post_send_buf_count, 0);
806         init_completion(&ib_conn->stop_completion);
807         INIT_LIST_HEAD(&ib_conn->conn_list);
808         spin_lock_init(&ib_conn->lock);
809 }
810
811  /**
812  * starts the process of connecting to the target
813  * sleeps until the connection is established or rejected
814  */
815 int iser_connect(struct iser_conn   *ib_conn,
816                  struct sockaddr    *src_addr,
817                  struct sockaddr    *dst_addr,
818                  int                 non_blocking)
819 {
820         int err = 0;
821
822         sprintf(ib_conn->name, "%pISp", dst_addr);
823
824         iser_info("connecting to: %s\n", ib_conn->name);
825
826         /* the device is known only --after-- address resolution */
827         ib_conn->device = NULL;
828
829         ib_conn->state = ISER_CONN_PENDING;
830
831         ib_conn->cma_id = rdma_create_id(iser_cma_handler,
832                                              (void *)ib_conn,
833                                              RDMA_PS_TCP, IB_QPT_RC);
834         if (IS_ERR(ib_conn->cma_id)) {
835                 err = PTR_ERR(ib_conn->cma_id);
836                 iser_err("rdma_create_id failed: %d\n", err);
837                 goto id_failure;
838         }
839
840         err = rdma_resolve_addr(ib_conn->cma_id, src_addr, dst_addr, 1000);
841         if (err) {
842                 iser_err("rdma_resolve_addr failed: %d\n", err);
843                 goto addr_failure;
844         }
845
846         if (!non_blocking) {
847                 wait_event_interruptible(ib_conn->wait,
848                                          (ib_conn->state != ISER_CONN_PENDING));
849
850                 if (ib_conn->state != ISER_CONN_UP) {
851                         err =  -EIO;
852                         goto connect_failure;
853                 }
854         }
855
856         mutex_lock(&ig.connlist_mutex);
857         list_add(&ib_conn->conn_list, &ig.connlist);
858         mutex_unlock(&ig.connlist_mutex);
859         return 0;
860
861 id_failure:
862         ib_conn->cma_id = NULL;
863 addr_failure:
864         ib_conn->state = ISER_CONN_DOWN;
865 connect_failure:
866         iser_conn_release(ib_conn);
867         return err;
868 }
869
870 /**
871  * iser_reg_page_vec - Register physical memory
872  *
873  * returns: 0 on success, errno code on failure
874  */
875 int iser_reg_page_vec(struct iser_conn     *ib_conn,
876                       struct iser_page_vec *page_vec,
877                       struct iser_mem_reg  *mem_reg)
878 {
879         struct ib_pool_fmr *mem;
880         u64                io_addr;
881         u64                *page_list;
882         int                status;
883
884         page_list = page_vec->pages;
885         io_addr   = page_list[0];
886
887         mem  = ib_fmr_pool_map_phys(ib_conn->fmr.pool,
888                                     page_list,
889                                     page_vec->length,
890                                     io_addr);
891
892         if (IS_ERR(mem)) {
893                 status = (int)PTR_ERR(mem);
894                 iser_err("ib_fmr_pool_map_phys failed: %d\n", status);
895                 return status;
896         }
897
898         mem_reg->lkey  = mem->fmr->lkey;
899         mem_reg->rkey  = mem->fmr->rkey;
900         mem_reg->len   = page_vec->length * SIZE_4K;
901         mem_reg->va    = io_addr;
902         mem_reg->is_mr = 1;
903         mem_reg->mem_h = (void *)mem;
904
905         mem_reg->va   += page_vec->offset;
906         mem_reg->len   = page_vec->data_size;
907
908         iser_dbg("PHYSICAL Mem.register, [PHYS p_array: 0x%p, sz: %d, "
909                  "entry[0]: (0x%08lx,%ld)] -> "
910                  "[lkey: 0x%08X mem_h: 0x%p va: 0x%08lX sz: %ld]\n",
911                  page_vec, page_vec->length,
912                  (unsigned long)page_vec->pages[0],
913                  (unsigned long)page_vec->data_size,
914                  (unsigned int)mem_reg->lkey, mem_reg->mem_h,
915                  (unsigned long)mem_reg->va, (unsigned long)mem_reg->len);
916         return 0;
917 }
918
919 /**
920  * Unregister (previosuly registered using FMR) memory.
921  * If memory is non-FMR does nothing.
922  */
923 void iser_unreg_mem_fmr(struct iscsi_iser_task *iser_task,
924                         enum iser_data_dir cmd_dir)
925 {
926         struct iser_mem_reg *reg = &iser_task->rdma_regd[cmd_dir].reg;
927         int ret;
928
929         if (!reg->is_mr)
930                 return;
931
932         iser_dbg("PHYSICAL Mem.Unregister mem_h %p\n",reg->mem_h);
933
934         ret = ib_fmr_pool_unmap((struct ib_pool_fmr *)reg->mem_h);
935         if (ret)
936                 iser_err("ib_fmr_pool_unmap failed %d\n", ret);
937
938         reg->mem_h = NULL;
939 }
940
941 void iser_unreg_mem_fastreg(struct iscsi_iser_task *iser_task,
942                             enum iser_data_dir cmd_dir)
943 {
944         struct iser_mem_reg *reg = &iser_task->rdma_regd[cmd_dir].reg;
945         struct iser_conn *ib_conn = iser_task->ib_conn;
946         struct fast_reg_descriptor *desc = reg->mem_h;
947
948         if (!reg->is_mr)
949                 return;
950
951         reg->mem_h = NULL;
952         reg->is_mr = 0;
953         spin_lock_bh(&ib_conn->lock);
954         list_add_tail(&desc->list, &ib_conn->fastreg.pool);
955         spin_unlock_bh(&ib_conn->lock);
956 }
957
958 int iser_post_recvl(struct iser_conn *ib_conn)
959 {
960         struct ib_recv_wr rx_wr, *rx_wr_failed;
961         struct ib_sge     sge;
962         int ib_ret;
963
964         sge.addr   = ib_conn->login_resp_dma;
965         sge.length = ISER_RX_LOGIN_SIZE;
966         sge.lkey   = ib_conn->device->mr->lkey;
967
968         rx_wr.wr_id   = (unsigned long)ib_conn->login_resp_buf;
969         rx_wr.sg_list = &sge;
970         rx_wr.num_sge = 1;
971         rx_wr.next    = NULL;
972
973         ib_conn->post_recv_buf_count++;
974         ib_ret  = ib_post_recv(ib_conn->qp, &rx_wr, &rx_wr_failed);
975         if (ib_ret) {
976                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
977                 ib_conn->post_recv_buf_count--;
978         }
979         return ib_ret;
980 }
981
982 int iser_post_recvm(struct iser_conn *ib_conn, int count)
983 {
984         struct ib_recv_wr *rx_wr, *rx_wr_failed;
985         int i, ib_ret;
986         unsigned int my_rx_head = ib_conn->rx_desc_head;
987         struct iser_rx_desc *rx_desc;
988
989         for (rx_wr = ib_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
990                 rx_desc         = &ib_conn->rx_descs[my_rx_head];
991                 rx_wr->wr_id    = (unsigned long)rx_desc;
992                 rx_wr->sg_list  = &rx_desc->rx_sg;
993                 rx_wr->num_sge  = 1;
994                 rx_wr->next     = rx_wr + 1;
995                 my_rx_head = (my_rx_head + 1) & ib_conn->qp_max_recv_dtos_mask;
996         }
997
998         rx_wr--;
999         rx_wr->next = NULL; /* mark end of work requests list */
1000
1001         ib_conn->post_recv_buf_count += count;
1002         ib_ret  = ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &rx_wr_failed);
1003         if (ib_ret) {
1004                 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1005                 ib_conn->post_recv_buf_count -= count;
1006         } else
1007                 ib_conn->rx_desc_head = my_rx_head;
1008         return ib_ret;
1009 }
1010
1011
1012 /**
1013  * iser_start_send - Initiate a Send DTO operation
1014  *
1015  * returns 0 on success, -1 on failure
1016  */
1017 int iser_post_send(struct iser_conn *ib_conn, struct iser_tx_desc *tx_desc)
1018 {
1019         int               ib_ret;
1020         struct ib_send_wr send_wr, *send_wr_failed;
1021
1022         ib_dma_sync_single_for_device(ib_conn->device->ib_device,
1023                 tx_desc->dma_addr, ISER_HEADERS_LEN, DMA_TO_DEVICE);
1024
1025         send_wr.next       = NULL;
1026         send_wr.wr_id      = (unsigned long)tx_desc;
1027         send_wr.sg_list    = tx_desc->tx_sg;
1028         send_wr.num_sge    = tx_desc->num_sge;
1029         send_wr.opcode     = IB_WR_SEND;
1030         send_wr.send_flags = IB_SEND_SIGNALED;
1031
1032         atomic_inc(&ib_conn->post_send_buf_count);
1033
1034         ib_ret = ib_post_send(ib_conn->qp, &send_wr, &send_wr_failed);
1035         if (ib_ret) {
1036                 iser_err("ib_post_send failed, ret:%d\n", ib_ret);
1037                 atomic_dec(&ib_conn->post_send_buf_count);
1038         }
1039         return ib_ret;
1040 }
1041
1042 static void iser_handle_comp_error(struct iser_tx_desc *desc,
1043                                 struct iser_conn *ib_conn)
1044 {
1045         if (desc && desc->type == ISCSI_TX_DATAOUT)
1046                 kmem_cache_free(ig.desc_cache, desc);
1047
1048         if (ib_conn->post_recv_buf_count == 0 &&
1049             atomic_read(&ib_conn->post_send_buf_count) == 0) {
1050                 /* getting here when the state is UP means that the conn is *
1051                  * being terminated asynchronously from the iSCSI layer's   *
1052                  * perspective.                                             */
1053                 if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
1054                     ISER_CONN_TERMINATING))
1055                         iscsi_conn_failure(ib_conn->iscsi_conn,
1056                                            ISCSI_ERR_CONN_FAILED);
1057
1058                 /* no more non completed posts to the QP, complete the
1059                  * termination process w.o worrying on disconnect event */
1060                 ib_conn->state = ISER_CONN_DOWN;
1061                 wake_up_interruptible(&ib_conn->wait);
1062         }
1063 }
1064
1065 static int iser_drain_tx_cq(struct iser_device  *device, int cq_index)
1066 {
1067         struct ib_cq  *cq = device->tx_cq[cq_index];
1068         struct ib_wc  wc;
1069         struct iser_tx_desc *tx_desc;
1070         struct iser_conn *ib_conn;
1071         int completed_tx = 0;
1072
1073         while (ib_poll_cq(cq, 1, &wc) == 1) {
1074                 tx_desc = (struct iser_tx_desc *) (unsigned long) wc.wr_id;
1075                 ib_conn = wc.qp->qp_context;
1076                 if (wc.status == IB_WC_SUCCESS) {
1077                         if (wc.opcode == IB_WC_SEND)
1078                                 iser_snd_completion(tx_desc, ib_conn);
1079                         else
1080                                 iser_err("expected opcode %d got %d\n",
1081                                         IB_WC_SEND, wc.opcode);
1082                 } else {
1083                         iser_err("tx id %llx status %d vend_err %x\n",
1084                                  wc.wr_id, wc.status, wc.vendor_err);
1085                         if (wc.wr_id != ISER_FASTREG_LI_WRID) {
1086                                 atomic_dec(&ib_conn->post_send_buf_count);
1087                                 iser_handle_comp_error(tx_desc, ib_conn);
1088                         }
1089                 }
1090                 completed_tx++;
1091         }
1092         return completed_tx;
1093 }
1094
1095
1096 static void iser_cq_tasklet_fn(unsigned long data)
1097 {
1098         struct iser_cq_desc *cq_desc = (struct iser_cq_desc *)data;
1099         struct iser_device  *device = cq_desc->device;
1100         int cq_index = cq_desc->cq_index;
1101         struct ib_cq         *cq = device->rx_cq[cq_index];
1102          struct ib_wc        wc;
1103          struct iser_rx_desc *desc;
1104          unsigned long       xfer_len;
1105         struct iser_conn *ib_conn;
1106         int completed_tx, completed_rx = 0;
1107
1108         /* First do tx drain, so in a case where we have rx flushes and a successful
1109          * tx completion we will still go through completion error handling.
1110          */
1111         completed_tx = iser_drain_tx_cq(device, cq_index);
1112
1113         while (ib_poll_cq(cq, 1, &wc) == 1) {
1114                 desc     = (struct iser_rx_desc *) (unsigned long) wc.wr_id;
1115                 BUG_ON(desc == NULL);
1116                 ib_conn = wc.qp->qp_context;
1117                 if (wc.status == IB_WC_SUCCESS) {
1118                         if (wc.opcode == IB_WC_RECV) {
1119                                 xfer_len = (unsigned long)wc.byte_len;
1120                                 iser_rcv_completion(desc, xfer_len, ib_conn);
1121                         } else
1122                                 iser_err("expected opcode %d got %d\n",
1123                                         IB_WC_RECV, wc.opcode);
1124                 } else {
1125                         if (wc.status != IB_WC_WR_FLUSH_ERR)
1126                                 iser_err("rx id %llx status %d vend_err %x\n",
1127                                         wc.wr_id, wc.status, wc.vendor_err);
1128                         ib_conn->post_recv_buf_count--;
1129                         iser_handle_comp_error(NULL, ib_conn);
1130                 }
1131                 completed_rx++;
1132                 if (!(completed_rx & 63))
1133                         completed_tx += iser_drain_tx_cq(device, cq_index);
1134         }
1135         /* #warning "it is assumed here that arming CQ only once its empty" *
1136          * " would not cause interrupts to be missed"                       */
1137         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1138
1139         iser_dbg("got %d rx %d tx completions\n", completed_rx, completed_tx);
1140 }
1141
1142 static void iser_cq_callback(struct ib_cq *cq, void *cq_context)
1143 {
1144         struct iser_cq_desc *cq_desc = (struct iser_cq_desc *)cq_context;
1145         struct iser_device  *device = cq_desc->device;
1146         int cq_index = cq_desc->cq_index;
1147
1148         tasklet_schedule(&device->cq_tasklet[cq_index]);
1149 }
1150
1151 u8 iser_check_task_pi_status(struct iscsi_iser_task *iser_task,
1152                              enum iser_data_dir cmd_dir, sector_t *sector)
1153 {
1154         struct iser_mem_reg *reg = &iser_task->rdma_regd[cmd_dir].reg;
1155         struct fast_reg_descriptor *desc = reg->mem_h;
1156         unsigned long sector_size = iser_task->sc->device->sector_size;
1157         struct ib_mr_status mr_status;
1158         int ret;
1159
1160         if (desc && desc->reg_indicators & ISER_FASTREG_PROTECTED) {
1161                 desc->reg_indicators &= ~ISER_FASTREG_PROTECTED;
1162                 ret = ib_check_mr_status(desc->pi_ctx->sig_mr,
1163                                          IB_MR_CHECK_SIG_STATUS, &mr_status);
1164                 if (ret) {
1165                         pr_err("ib_check_mr_status failed, ret %d\n", ret);
1166                         goto err;
1167                 }
1168
1169                 if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1170                         sector_t sector_off = mr_status.sig_err.sig_err_offset;
1171
1172                         do_div(sector_off, sector_size + 8);
1173                         *sector = scsi_get_lba(iser_task->sc) + sector_off;
1174
1175                         pr_err("PI error found type %d at sector %llx "
1176                                "expected %x vs actual %x\n",
1177                                mr_status.sig_err.err_type,
1178                                (unsigned long long)*sector,
1179                                mr_status.sig_err.expected,
1180                                mr_status.sig_err.actual);
1181
1182                         switch (mr_status.sig_err.err_type) {
1183                         case IB_SIG_BAD_GUARD:
1184                                 return 0x1;
1185                         case IB_SIG_BAD_REFTAG:
1186                                 return 0x3;
1187                         case IB_SIG_BAD_APPTAG:
1188                                 return 0x2;
1189                         }
1190                 }
1191         }
1192
1193         return 0;
1194 err:
1195         /* Not alot we can do here, return ambiguous guard error */
1196         return 0x1;
1197 }