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.
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:
12 * Redistribution and use in source and binary forms, with or
13 * without modification, are permitted provided that the following
16 * - Redistributions of source code must retain the above
17 * copyright notice, this list of conditions and the following
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.
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
34 #include <linux/kernel.h>
35 #include <linux/module.h>
36 #include <linux/slab.h>
37 #include <linux/delay.h>
39 #include "iscsi_iser.h"
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 + \
47 static int iser_cq_poll_limit = 512;
49 static void iser_cq_tasklet_fn(unsigned long data);
50 static void iser_cq_callback(struct ib_cq *cq, void *cq_context);
52 static void iser_cq_event_callback(struct ib_event *cause, void *context)
54 iser_err("got cq event %d \n", cause->event);
57 static void iser_qp_event_callback(struct ib_event *cause, void *context)
59 iser_err("got qp event %d\n",cause->event);
62 static void iser_event_handler(struct ib_event_handler *handler,
63 struct ib_event *event)
65 iser_err("async event %d on device %s port %d\n", event->event,
66 event->device->name, event->element.port_num);
70 * iser_create_device_ib_res - creates Protection Domain (PD), Completion
71 * Queue (CQ), DMA Memory Region (DMA MR) with the device associated with
74 * returns 0 on success, -1 on failure
76 static int iser_create_device_ib_res(struct iser_device *device)
78 struct ib_device_attr *dev_attr = &device->dev_attr;
81 ret = ib_query_device(device->ib_device, dev_attr);
83 pr_warn("Query device failed for %s\n", device->ib_device->name);
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;
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;
103 iser_err("IB device does not support FMRs nor FastRegs, can't register memory\n");
107 device->comps_used = min(ISER_MAX_CQ,
108 device->ib_device->num_comp_vectors);
110 max_cqe = min(ISER_MAX_CQ_LEN, dev_attr->max_cqe);
112 iser_info("using %d CQs, device %s supports %d vectors max_cqe %d\n",
113 device->comps_used, device->ib_device->name,
114 device->ib_device->num_comp_vectors, max_cqe);
116 device->pd = ib_alloc_pd(device->ib_device);
117 if (IS_ERR(device->pd))
120 for (i = 0; i < device->comps_used; i++) {
121 struct iser_comp *comp = &device->comps[i];
123 comp->device = device;
124 comp->cq = ib_create_cq(device->ib_device,
126 iser_cq_event_callback,
129 if (IS_ERR(comp->cq)) {
134 if (ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP))
137 tasklet_init(&comp->tasklet, iser_cq_tasklet_fn,
138 (unsigned long)comp);
141 device->mr = ib_get_dma_mr(device->pd, IB_ACCESS_LOCAL_WRITE |
142 IB_ACCESS_REMOTE_WRITE |
143 IB_ACCESS_REMOTE_READ);
144 if (IS_ERR(device->mr))
147 INIT_IB_EVENT_HANDLER(&device->event_handler, device->ib_device,
149 if (ib_register_event_handler(&device->event_handler))
155 ib_dereg_mr(device->mr);
157 for (i = 0; i < device->comps_used; i++)
158 tasklet_kill(&device->comps[i].tasklet);
160 for (i = 0; i < device->comps_used; i++) {
161 struct iser_comp *comp = &device->comps[i];
164 ib_destroy_cq(comp->cq);
166 ib_dealloc_pd(device->pd);
168 iser_err("failed to allocate an IB resource\n");
173 * iser_free_device_ib_res - destroy/dealloc/dereg the DMA MR,
174 * CQ and PD created with the device associated with the adapator.
176 static void iser_free_device_ib_res(struct iser_device *device)
179 BUG_ON(device->mr == NULL);
181 for (i = 0; i < device->comps_used; i++) {
182 struct iser_comp *comp = &device->comps[i];
184 tasklet_kill(&comp->tasklet);
185 ib_destroy_cq(comp->cq);
189 (void)ib_unregister_event_handler(&device->event_handler);
190 (void)ib_dereg_mr(device->mr);
191 (void)ib_dealloc_pd(device->pd);
198 * iser_create_fmr_pool - Creates FMR pool and page_vector
200 * returns 0 on success, or errno code on failure
202 int iser_create_fmr_pool(struct ib_conn *ib_conn, unsigned cmds_max)
204 struct iser_device *device = ib_conn->device;
205 struct ib_fmr_pool_param params;
208 ib_conn->fmr.page_vec = kmalloc(sizeof(*ib_conn->fmr.page_vec) +
209 (sizeof(u64)*(ISCSI_ISER_SG_TABLESIZE + 1)),
211 if (!ib_conn->fmr.page_vec)
214 ib_conn->fmr.page_vec->pages = (u64 *)(ib_conn->fmr.page_vec + 1);
216 params.page_shift = SHIFT_4K;
217 /* when the first/last SG element are not start/end *
218 * page aligned, the map whould be of N+1 pages */
219 params.max_pages_per_fmr = ISCSI_ISER_SG_TABLESIZE + 1;
220 /* make the pool size twice the max number of SCSI commands *
221 * the ML is expected to queue, watermark for unmap at 50% */
222 params.pool_size = cmds_max * 2;
223 params.dirty_watermark = cmds_max;
225 params.flush_function = NULL;
226 params.access = (IB_ACCESS_LOCAL_WRITE |
227 IB_ACCESS_REMOTE_WRITE |
228 IB_ACCESS_REMOTE_READ);
230 ib_conn->fmr.pool = ib_create_fmr_pool(device->pd, ¶ms);
231 if (!IS_ERR(ib_conn->fmr.pool))
234 /* no FMR => no need for page_vec */
235 kfree(ib_conn->fmr.page_vec);
236 ib_conn->fmr.page_vec = NULL;
238 ret = PTR_ERR(ib_conn->fmr.pool);
239 ib_conn->fmr.pool = NULL;
240 if (ret != -ENOSYS) {
241 iser_err("FMR allocation failed, err %d\n", ret);
244 iser_warn("FMRs are not supported, using unaligned mode\n");
250 * iser_free_fmr_pool - releases the FMR pool and page vec
252 void iser_free_fmr_pool(struct ib_conn *ib_conn)
254 iser_info("freeing conn %p fmr pool %p\n",
255 ib_conn, ib_conn->fmr.pool);
257 if (ib_conn->fmr.pool != NULL)
258 ib_destroy_fmr_pool(ib_conn->fmr.pool);
260 ib_conn->fmr.pool = NULL;
262 kfree(ib_conn->fmr.page_vec);
263 ib_conn->fmr.page_vec = NULL;
267 iser_create_fastreg_desc(struct ib_device *ib_device, struct ib_pd *pd,
268 bool pi_enable, struct fast_reg_descriptor *desc)
272 desc->data_frpl = ib_alloc_fast_reg_page_list(ib_device,
273 ISCSI_ISER_SG_TABLESIZE + 1);
274 if (IS_ERR(desc->data_frpl)) {
275 ret = PTR_ERR(desc->data_frpl);
276 iser_err("Failed to allocate ib_fast_reg_page_list err=%d\n",
278 return PTR_ERR(desc->data_frpl);
281 desc->data_mr = ib_alloc_fast_reg_mr(pd, ISCSI_ISER_SG_TABLESIZE + 1);
282 if (IS_ERR(desc->data_mr)) {
283 ret = PTR_ERR(desc->data_mr);
284 iser_err("Failed to allocate ib_fast_reg_mr err=%d\n", ret);
285 goto fast_reg_mr_failure;
287 desc->reg_indicators |= ISER_DATA_KEY_VALID;
290 struct ib_mr_init_attr mr_init_attr = {0};
291 struct iser_pi_context *pi_ctx = NULL;
293 desc->pi_ctx = kzalloc(sizeof(*desc->pi_ctx), GFP_KERNEL);
295 iser_err("Failed to allocate pi context\n");
297 goto pi_ctx_alloc_failure;
299 pi_ctx = desc->pi_ctx;
301 pi_ctx->prot_frpl = ib_alloc_fast_reg_page_list(ib_device,
302 ISCSI_ISER_SG_TABLESIZE);
303 if (IS_ERR(pi_ctx->prot_frpl)) {
304 ret = PTR_ERR(pi_ctx->prot_frpl);
305 iser_err("Failed to allocate prot frpl ret=%d\n",
307 goto prot_frpl_failure;
310 pi_ctx->prot_mr = ib_alloc_fast_reg_mr(pd,
311 ISCSI_ISER_SG_TABLESIZE + 1);
312 if (IS_ERR(pi_ctx->prot_mr)) {
313 ret = PTR_ERR(pi_ctx->prot_mr);
314 iser_err("Failed to allocate prot frmr ret=%d\n",
316 goto prot_mr_failure;
318 desc->reg_indicators |= ISER_PROT_KEY_VALID;
320 mr_init_attr.max_reg_descriptors = 2;
321 mr_init_attr.flags |= IB_MR_SIGNATURE_EN;
322 pi_ctx->sig_mr = ib_create_mr(pd, &mr_init_attr);
323 if (IS_ERR(pi_ctx->sig_mr)) {
324 ret = PTR_ERR(pi_ctx->sig_mr);
325 iser_err("Failed to allocate signature enabled mr err=%d\n",
329 desc->reg_indicators |= ISER_SIG_KEY_VALID;
331 desc->reg_indicators &= ~ISER_FASTREG_PROTECTED;
333 iser_dbg("Create fr_desc %p page_list %p\n",
334 desc, desc->data_frpl->page_list);
338 ib_dereg_mr(desc->pi_ctx->prot_mr);
340 ib_free_fast_reg_page_list(desc->pi_ctx->prot_frpl);
343 pi_ctx_alloc_failure:
344 ib_dereg_mr(desc->data_mr);
346 ib_free_fast_reg_page_list(desc->data_frpl);
352 * iser_create_fastreg_pool - Creates pool of fast_reg descriptors
353 * for fast registration work requests.
354 * returns 0 on success, or errno code on failure
356 int iser_create_fastreg_pool(struct ib_conn *ib_conn, unsigned cmds_max)
358 struct iser_device *device = ib_conn->device;
359 struct fast_reg_descriptor *desc;
362 INIT_LIST_HEAD(&ib_conn->fastreg.pool);
363 ib_conn->fastreg.pool_size = 0;
364 for (i = 0; i < cmds_max; i++) {
365 desc = kzalloc(sizeof(*desc), GFP_KERNEL);
367 iser_err("Failed to allocate a new fast_reg descriptor\n");
372 ret = iser_create_fastreg_desc(device->ib_device, device->pd,
373 ib_conn->pi_support, desc);
375 iser_err("Failed to create fastreg descriptor err=%d\n",
381 list_add_tail(&desc->list, &ib_conn->fastreg.pool);
382 ib_conn->fastreg.pool_size++;
388 iser_free_fastreg_pool(ib_conn);
393 * iser_free_fastreg_pool - releases the pool of fast_reg descriptors
395 void iser_free_fastreg_pool(struct ib_conn *ib_conn)
397 struct fast_reg_descriptor *desc, *tmp;
400 if (list_empty(&ib_conn->fastreg.pool))
403 iser_info("freeing conn %p fr pool\n", ib_conn);
405 list_for_each_entry_safe(desc, tmp, &ib_conn->fastreg.pool, list) {
406 list_del(&desc->list);
407 ib_free_fast_reg_page_list(desc->data_frpl);
408 ib_dereg_mr(desc->data_mr);
410 ib_free_fast_reg_page_list(desc->pi_ctx->prot_frpl);
411 ib_dereg_mr(desc->pi_ctx->prot_mr);
412 ib_destroy_mr(desc->pi_ctx->sig_mr);
419 if (i < ib_conn->fastreg.pool_size)
420 iser_warn("pool still has %d regions registered\n",
421 ib_conn->fastreg.pool_size - i);
425 * iser_create_ib_conn_res - Queue-Pair (QP)
427 * returns 0 on success, -1 on failure
429 static int iser_create_ib_conn_res(struct ib_conn *ib_conn)
431 struct iser_conn *iser_conn = container_of(ib_conn, struct iser_conn,
433 struct iser_device *device;
434 struct ib_device_attr *dev_attr;
435 struct ib_qp_init_attr init_attr;
437 int index, min_index = 0;
439 BUG_ON(ib_conn->device == NULL);
441 device = ib_conn->device;
442 dev_attr = &device->dev_attr;
444 memset(&init_attr, 0, sizeof init_attr);
446 mutex_lock(&ig.connlist_mutex);
447 /* select the CQ with the minimal number of usages */
448 for (index = 0; index < device->comps_used; index++) {
449 if (device->comps[index].active_qps <
450 device->comps[min_index].active_qps)
453 ib_conn->comp = &device->comps[min_index];
454 ib_conn->comp->active_qps++;
455 mutex_unlock(&ig.connlist_mutex);
456 iser_info("cq index %d used for ib_conn %p\n", min_index, ib_conn);
458 init_attr.event_handler = iser_qp_event_callback;
459 init_attr.qp_context = (void *)ib_conn;
460 init_attr.send_cq = ib_conn->comp->cq;
461 init_attr.recv_cq = ib_conn->comp->cq;
462 init_attr.cap.max_recv_wr = ISER_QP_MAX_RECV_DTOS;
463 init_attr.cap.max_send_sge = 2;
464 init_attr.cap.max_recv_sge = 1;
465 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
466 init_attr.qp_type = IB_QPT_RC;
467 if (ib_conn->pi_support) {
468 init_attr.cap.max_send_wr = ISER_QP_SIG_MAX_REQ_DTOS + 1;
469 init_attr.create_flags |= IB_QP_CREATE_SIGNATURE_EN;
470 iser_conn->max_cmds =
471 ISER_GET_MAX_XMIT_CMDS(ISER_QP_SIG_MAX_REQ_DTOS);
473 if (dev_attr->max_qp_wr > ISER_QP_MAX_REQ_DTOS) {
474 init_attr.cap.max_send_wr = ISER_QP_MAX_REQ_DTOS + 1;
475 iser_conn->max_cmds =
476 ISER_GET_MAX_XMIT_CMDS(ISER_QP_MAX_REQ_DTOS);
478 init_attr.cap.max_send_wr = dev_attr->max_qp_wr;
479 iser_conn->max_cmds =
480 ISER_GET_MAX_XMIT_CMDS(dev_attr->max_qp_wr);
481 iser_dbg("device %s supports max_send_wr %d\n",
482 device->ib_device->name, dev_attr->max_qp_wr);
486 ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
490 ib_conn->qp = ib_conn->cma_id->qp;
491 iser_info("setting conn %p cma_id %p qp %p\n",
492 ib_conn, ib_conn->cma_id,
493 ib_conn->cma_id->qp);
497 mutex_lock(&ig.connlist_mutex);
498 ib_conn->comp->active_qps--;
499 mutex_unlock(&ig.connlist_mutex);
500 iser_err("unable to alloc mem or create resource, err %d\n", ret);
506 * based on the resolved device node GUID see if there already allocated
507 * device for this device. If there's no such, create one.
510 struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
512 struct iser_device *device;
514 mutex_lock(&ig.device_list_mutex);
516 list_for_each_entry(device, &ig.device_list, ig_list)
517 /* find if there's a match using the node GUID */
518 if (device->ib_device->node_guid == cma_id->device->node_guid)
521 device = kzalloc(sizeof *device, GFP_KERNEL);
525 /* assign this device to the device */
526 device->ib_device = cma_id->device;
527 /* init the device and link it into ig device list */
528 if (iser_create_device_ib_res(device)) {
533 list_add(&device->ig_list, &ig.device_list);
538 mutex_unlock(&ig.device_list_mutex);
542 /* if there's no demand for this device, release it */
543 static void iser_device_try_release(struct iser_device *device)
545 mutex_lock(&ig.device_list_mutex);
547 iser_info("device %p refcount %d\n", device, device->refcount);
548 if (!device->refcount) {
549 iser_free_device_ib_res(device);
550 list_del(&device->ig_list);
553 mutex_unlock(&ig.device_list_mutex);
557 * Called with state mutex held
559 static int iser_conn_state_comp_exch(struct iser_conn *iser_conn,
560 enum iser_conn_state comp,
561 enum iser_conn_state exch)
565 ret = (iser_conn->state == comp);
567 iser_conn->state = exch;
572 void iser_release_work(struct work_struct *work)
574 struct iser_conn *iser_conn;
576 iser_conn = container_of(work, struct iser_conn, release_work);
578 /* Wait for conn_stop to complete */
579 wait_for_completion(&iser_conn->stop_completion);
580 /* Wait for IB resouces cleanup to complete */
581 wait_for_completion(&iser_conn->ib_completion);
583 mutex_lock(&iser_conn->state_mutex);
584 iser_conn->state = ISER_CONN_DOWN;
585 mutex_unlock(&iser_conn->state_mutex);
587 iser_conn_release(iser_conn);
591 * iser_free_ib_conn_res - release IB related resources
592 * @iser_conn: iser connection struct
593 * @destroy_device: indicator if we need to try to release
594 * the iser device (only iscsi shutdown and DEVICE_REMOVAL
597 * This routine is called with the iser state mutex held
598 * so the cm_id removal is out of here. It is Safe to
599 * be invoked multiple times.
601 static void iser_free_ib_conn_res(struct iser_conn *iser_conn,
604 struct ib_conn *ib_conn = &iser_conn->ib_conn;
605 struct iser_device *device = ib_conn->device;
607 iser_info("freeing conn %p cma_id %p qp %p\n",
608 iser_conn, ib_conn->cma_id, ib_conn->qp);
610 iser_free_rx_descriptors(iser_conn);
612 if (ib_conn->qp != NULL) {
613 ib_conn->comp->active_qps--;
614 rdma_destroy_qp(ib_conn->cma_id);
618 if (destroy_device && device != NULL) {
619 iser_device_try_release(device);
620 ib_conn->device = NULL;
625 * Frees all conn objects and deallocs conn descriptor
627 void iser_conn_release(struct iser_conn *iser_conn)
629 struct ib_conn *ib_conn = &iser_conn->ib_conn;
631 mutex_lock(&ig.connlist_mutex);
632 list_del(&iser_conn->conn_list);
633 mutex_unlock(&ig.connlist_mutex);
635 mutex_lock(&iser_conn->state_mutex);
636 if (iser_conn->state != ISER_CONN_DOWN) {
637 iser_warn("iser conn %p state %d, expected state down.\n",
638 iser_conn, iser_conn->state);
639 iser_conn->state = ISER_CONN_DOWN;
642 * In case we never got to bind stage, we still need to
643 * release IB resources (which is safe to call more than once).
645 iser_free_ib_conn_res(iser_conn, true);
646 mutex_unlock(&iser_conn->state_mutex);
648 if (ib_conn->cma_id != NULL) {
649 rdma_destroy_id(ib_conn->cma_id);
650 ib_conn->cma_id = NULL;
657 * triggers start of the disconnect procedures and wait for them to be done
658 * Called with state mutex held
660 int iser_conn_terminate(struct iser_conn *iser_conn)
662 struct ib_conn *ib_conn = &iser_conn->ib_conn;
663 struct ib_send_wr *bad_wr;
666 /* terminate the iser conn only if the conn state is UP */
667 if (!iser_conn_state_comp_exch(iser_conn, ISER_CONN_UP,
668 ISER_CONN_TERMINATING))
671 iser_info("iser_conn %p state %d\n", iser_conn, iser_conn->state);
673 /* suspend queuing of new iscsi commands */
674 if (iser_conn->iscsi_conn)
675 iscsi_suspend_queue(iser_conn->iscsi_conn);
678 * In case we didn't already clean up the cma_id (peer initiated
679 * a disconnection), we need to Cause the CMA to change the QP
682 if (ib_conn->cma_id) {
683 err = rdma_disconnect(ib_conn->cma_id);
685 iser_err("Failed to disconnect, conn: 0x%p err %d\n",
688 /* post an indication that all flush errors were consumed */
689 err = ib_post_send(ib_conn->qp, &ib_conn->beacon, &bad_wr);
691 iser_err("conn %p failed to post beacon", ib_conn);
695 wait_for_completion(&ib_conn->flush_comp);
702 * Called with state mutex held
704 static void iser_connect_error(struct rdma_cm_id *cma_id)
706 struct iser_conn *iser_conn;
708 iser_conn = (struct iser_conn *)cma_id->context;
709 iser_conn->state = ISER_CONN_DOWN;
713 * Called with state mutex held
715 static void iser_addr_handler(struct rdma_cm_id *cma_id)
717 struct iser_device *device;
718 struct iser_conn *iser_conn;
719 struct ib_conn *ib_conn;
722 iser_conn = (struct iser_conn *)cma_id->context;
723 if (iser_conn->state != ISER_CONN_PENDING)
727 ib_conn = &iser_conn->ib_conn;
728 device = iser_device_find_by_ib_device(cma_id);
730 iser_err("device lookup/creation failed\n");
731 iser_connect_error(cma_id);
735 ib_conn->device = device;
737 /* connection T10-PI support */
738 if (iser_pi_enable) {
739 if (!(device->dev_attr.device_cap_flags &
740 IB_DEVICE_SIGNATURE_HANDOVER)) {
741 iser_warn("T10-PI requested but not supported on %s, "
742 "continue without T10-PI\n",
743 ib_conn->device->ib_device->name);
744 ib_conn->pi_support = false;
746 ib_conn->pi_support = true;
750 ret = rdma_resolve_route(cma_id, 1000);
752 iser_err("resolve route failed: %d\n", ret);
753 iser_connect_error(cma_id);
759 * Called with state mutex held
761 static void iser_route_handler(struct rdma_cm_id *cma_id)
763 struct rdma_conn_param conn_param;
765 struct iser_cm_hdr req_hdr;
766 struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
767 struct ib_conn *ib_conn = &iser_conn->ib_conn;
768 struct iser_device *device = ib_conn->device;
770 if (iser_conn->state != ISER_CONN_PENDING)
774 ret = iser_create_ib_conn_res(ib_conn);
778 memset(&conn_param, 0, sizeof conn_param);
779 conn_param.responder_resources = device->dev_attr.max_qp_rd_atom;
780 conn_param.initiator_depth = 1;
781 conn_param.retry_count = 7;
782 conn_param.rnr_retry_count = 6;
784 memset(&req_hdr, 0, sizeof(req_hdr));
785 req_hdr.flags = (ISER_ZBVA_NOT_SUPPORTED |
786 ISER_SEND_W_INV_NOT_SUPPORTED);
787 conn_param.private_data = (void *)&req_hdr;
788 conn_param.private_data_len = sizeof(struct iser_cm_hdr);
790 ret = rdma_connect(cma_id, &conn_param);
792 iser_err("failure connecting: %d\n", ret);
798 iser_connect_error(cma_id);
801 static void iser_connected_handler(struct rdma_cm_id *cma_id)
803 struct iser_conn *iser_conn;
804 struct ib_qp_attr attr;
805 struct ib_qp_init_attr init_attr;
807 iser_conn = (struct iser_conn *)cma_id->context;
808 if (iser_conn->state != ISER_CONN_PENDING)
812 (void)ib_query_qp(cma_id->qp, &attr, ~0, &init_attr);
813 iser_info("remote qpn:%x my qpn:%x\n", attr.dest_qp_num, cma_id->qp->qp_num);
815 iser_conn->state = ISER_CONN_UP;
816 complete(&iser_conn->up_completion);
819 static void iser_disconnected_handler(struct rdma_cm_id *cma_id)
821 struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
823 if (iser_conn_terminate(iser_conn)) {
824 if (iser_conn->iscsi_conn)
825 iscsi_conn_failure(iser_conn->iscsi_conn,
826 ISCSI_ERR_CONN_FAILED);
828 iser_err("iscsi_iser connection isn't bound\n");
832 static void iser_cleanup_handler(struct rdma_cm_id *cma_id,
835 struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
838 * We are not guaranteed that we visited disconnected_handler
839 * by now, call it here to be safe that we handle CM drep
842 iser_disconnected_handler(cma_id);
843 iser_free_ib_conn_res(iser_conn, destroy_device);
844 complete(&iser_conn->ib_completion);
847 static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
849 struct iser_conn *iser_conn;
852 iser_conn = (struct iser_conn *)cma_id->context;
853 iser_info("event %d status %d conn %p id %p\n",
854 event->event, event->status, cma_id->context, cma_id);
856 mutex_lock(&iser_conn->state_mutex);
857 switch (event->event) {
858 case RDMA_CM_EVENT_ADDR_RESOLVED:
859 iser_addr_handler(cma_id);
861 case RDMA_CM_EVENT_ROUTE_RESOLVED:
862 iser_route_handler(cma_id);
864 case RDMA_CM_EVENT_ESTABLISHED:
865 iser_connected_handler(cma_id);
867 case RDMA_CM_EVENT_ADDR_ERROR:
868 case RDMA_CM_EVENT_ROUTE_ERROR:
869 case RDMA_CM_EVENT_CONNECT_ERROR:
870 case RDMA_CM_EVENT_UNREACHABLE:
871 case RDMA_CM_EVENT_REJECTED:
872 iser_connect_error(cma_id);
874 case RDMA_CM_EVENT_DISCONNECTED:
875 case RDMA_CM_EVENT_ADDR_CHANGE:
876 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
877 iser_cleanup_handler(cma_id, false);
879 case RDMA_CM_EVENT_DEVICE_REMOVAL:
881 * we *must* destroy the device as we cannot rely
882 * on iscsid to be around to initiate error handling.
883 * also if we are not in state DOWN implicitly destroy
886 iser_cleanup_handler(cma_id, true);
887 if (iser_conn->state != ISER_CONN_DOWN) {
888 iser_conn->ib_conn.cma_id = NULL;
893 iser_err("Unexpected RDMA CM event (%d)\n", event->event);
896 mutex_unlock(&iser_conn->state_mutex);
901 void iser_conn_init(struct iser_conn *iser_conn)
903 iser_conn->state = ISER_CONN_INIT;
904 iser_conn->ib_conn.post_recv_buf_count = 0;
905 init_completion(&iser_conn->ib_conn.flush_comp);
906 init_completion(&iser_conn->stop_completion);
907 init_completion(&iser_conn->ib_completion);
908 init_completion(&iser_conn->up_completion);
909 INIT_LIST_HEAD(&iser_conn->conn_list);
910 spin_lock_init(&iser_conn->ib_conn.lock);
911 mutex_init(&iser_conn->state_mutex);
915 * starts the process of connecting to the target
916 * sleeps until the connection is established or rejected
918 int iser_connect(struct iser_conn *iser_conn,
919 struct sockaddr *src_addr,
920 struct sockaddr *dst_addr,
923 struct ib_conn *ib_conn = &iser_conn->ib_conn;
926 mutex_lock(&iser_conn->state_mutex);
928 sprintf(iser_conn->name, "%pISp", dst_addr);
930 iser_info("connecting to: %s\n", iser_conn->name);
932 /* the device is known only --after-- address resolution */
933 ib_conn->device = NULL;
935 iser_conn->state = ISER_CONN_PENDING;
937 ib_conn->beacon.wr_id = ISER_BEACON_WRID;
938 ib_conn->beacon.opcode = IB_WR_SEND;
940 ib_conn->cma_id = rdma_create_id(iser_cma_handler,
942 RDMA_PS_TCP, IB_QPT_RC);
943 if (IS_ERR(ib_conn->cma_id)) {
944 err = PTR_ERR(ib_conn->cma_id);
945 iser_err("rdma_create_id failed: %d\n", err);
949 err = rdma_resolve_addr(ib_conn->cma_id, src_addr, dst_addr, 1000);
951 iser_err("rdma_resolve_addr failed: %d\n", err);
956 wait_for_completion_interruptible(&iser_conn->up_completion);
958 if (iser_conn->state != ISER_CONN_UP) {
960 goto connect_failure;
963 mutex_unlock(&iser_conn->state_mutex);
965 mutex_lock(&ig.connlist_mutex);
966 list_add(&iser_conn->conn_list, &ig.connlist);
967 mutex_unlock(&ig.connlist_mutex);
971 ib_conn->cma_id = NULL;
973 iser_conn->state = ISER_CONN_DOWN;
975 mutex_unlock(&iser_conn->state_mutex);
976 iser_conn_release(iser_conn);
981 * iser_reg_page_vec - Register physical memory
983 * returns: 0 on success, errno code on failure
985 int iser_reg_page_vec(struct ib_conn *ib_conn,
986 struct iser_page_vec *page_vec,
987 struct iser_mem_reg *mem_reg)
989 struct ib_pool_fmr *mem;
994 page_list = page_vec->pages;
995 io_addr = page_list[0];
997 mem = ib_fmr_pool_map_phys(ib_conn->fmr.pool,
1003 status = (int)PTR_ERR(mem);
1004 iser_err("ib_fmr_pool_map_phys failed: %d\n", status);
1008 mem_reg->lkey = mem->fmr->lkey;
1009 mem_reg->rkey = mem->fmr->rkey;
1010 mem_reg->len = page_vec->length * SIZE_4K;
1011 mem_reg->va = io_addr;
1012 mem_reg->mem_h = (void *)mem;
1014 mem_reg->va += page_vec->offset;
1015 mem_reg->len = page_vec->data_size;
1017 iser_dbg("PHYSICAL Mem.register, [PHYS p_array: 0x%p, sz: %d, "
1018 "entry[0]: (0x%08lx,%ld)] -> "
1019 "[lkey: 0x%08X mem_h: 0x%p va: 0x%08lX sz: %ld]\n",
1020 page_vec, page_vec->length,
1021 (unsigned long)page_vec->pages[0],
1022 (unsigned long)page_vec->data_size,
1023 (unsigned int)mem_reg->lkey, mem_reg->mem_h,
1024 (unsigned long)mem_reg->va, (unsigned long)mem_reg->len);
1029 * Unregister (previosuly registered using FMR) memory.
1030 * If memory is non-FMR does nothing.
1032 void iser_unreg_mem_fmr(struct iscsi_iser_task *iser_task,
1033 enum iser_data_dir cmd_dir)
1035 struct iser_mem_reg *reg = &iser_task->rdma_regd[cmd_dir].reg;
1041 iser_dbg("PHYSICAL Mem.Unregister mem_h %p\n",reg->mem_h);
1043 ret = ib_fmr_pool_unmap((struct ib_pool_fmr *)reg->mem_h);
1045 iser_err("ib_fmr_pool_unmap failed %d\n", ret);
1050 void iser_unreg_mem_fastreg(struct iscsi_iser_task *iser_task,
1051 enum iser_data_dir cmd_dir)
1053 struct iser_mem_reg *reg = &iser_task->rdma_regd[cmd_dir].reg;
1054 struct iser_conn *iser_conn = iser_task->iser_conn;
1055 struct ib_conn *ib_conn = &iser_conn->ib_conn;
1056 struct fast_reg_descriptor *desc = reg->mem_h;
1062 spin_lock_bh(&ib_conn->lock);
1063 list_add_tail(&desc->list, &ib_conn->fastreg.pool);
1064 spin_unlock_bh(&ib_conn->lock);
1067 int iser_post_recvl(struct iser_conn *iser_conn)
1069 struct ib_recv_wr rx_wr, *rx_wr_failed;
1070 struct ib_conn *ib_conn = &iser_conn->ib_conn;
1074 sge.addr = iser_conn->login_resp_dma;
1075 sge.length = ISER_RX_LOGIN_SIZE;
1076 sge.lkey = ib_conn->device->mr->lkey;
1078 rx_wr.wr_id = (uintptr_t)iser_conn->login_resp_buf;
1079 rx_wr.sg_list = &sge;
1083 ib_conn->post_recv_buf_count++;
1084 ib_ret = ib_post_recv(ib_conn->qp, &rx_wr, &rx_wr_failed);
1086 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1087 ib_conn->post_recv_buf_count--;
1092 int iser_post_recvm(struct iser_conn *iser_conn, int count)
1094 struct ib_recv_wr *rx_wr, *rx_wr_failed;
1096 struct ib_conn *ib_conn = &iser_conn->ib_conn;
1097 unsigned int my_rx_head = iser_conn->rx_desc_head;
1098 struct iser_rx_desc *rx_desc;
1100 for (rx_wr = ib_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
1101 rx_desc = &iser_conn->rx_descs[my_rx_head];
1102 rx_wr->wr_id = (uintptr_t)rx_desc;
1103 rx_wr->sg_list = &rx_desc->rx_sg;
1105 rx_wr->next = rx_wr + 1;
1106 my_rx_head = (my_rx_head + 1) & iser_conn->qp_max_recv_dtos_mask;
1110 rx_wr->next = NULL; /* mark end of work requests list */
1112 ib_conn->post_recv_buf_count += count;
1113 ib_ret = ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &rx_wr_failed);
1115 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1116 ib_conn->post_recv_buf_count -= count;
1118 iser_conn->rx_desc_head = my_rx_head;
1124 * iser_start_send - Initiate a Send DTO operation
1126 * returns 0 on success, -1 on failure
1128 int iser_post_send(struct ib_conn *ib_conn, struct iser_tx_desc *tx_desc,
1132 struct ib_send_wr send_wr, *send_wr_failed;
1134 ib_dma_sync_single_for_device(ib_conn->device->ib_device,
1135 tx_desc->dma_addr, ISER_HEADERS_LEN,
1138 send_wr.next = NULL;
1139 send_wr.wr_id = (uintptr_t)tx_desc;
1140 send_wr.sg_list = tx_desc->tx_sg;
1141 send_wr.num_sge = tx_desc->num_sge;
1142 send_wr.opcode = IB_WR_SEND;
1143 send_wr.send_flags = signal ? IB_SEND_SIGNALED : 0;
1145 ib_ret = ib_post_send(ib_conn->qp, &send_wr, &send_wr_failed);
1147 iser_err("ib_post_send failed, ret:%d\n", ib_ret);
1153 * is_iser_tx_desc - Indicate if the completion wr_id
1154 * is a TX descriptor or not.
1155 * @iser_conn: iser connection
1156 * @wr_id: completion WR identifier
1158 * Since we cannot rely on wc opcode in FLUSH errors
1159 * we must work around it by checking if the wr_id address
1160 * falls in the iser connection rx_descs buffer. If so
1161 * it is an RX descriptor, otherwize it is a TX.
1164 is_iser_tx_desc(struct iser_conn *iser_conn, void *wr_id)
1166 void *start = iser_conn->rx_descs;
1167 int len = iser_conn->num_rx_descs * sizeof(*iser_conn->rx_descs);
1169 if (wr_id >= start && wr_id < start + len)
1176 * iser_handle_comp_error() - Handle error completion
1177 * @ib_conn: connection RDMA resources
1178 * @wc: work completion
1180 * Notes: We may handle a FLUSH error completion and in this case
1181 * we only cleanup in case TX type was DATAOUT. For non-FLUSH
1182 * error completion we should also notify iscsi layer that
1183 * connection is failed (in case we passed bind stage).
1186 iser_handle_comp_error(struct ib_conn *ib_conn,
1189 void *wr_id = (void *)(uintptr_t)wc->wr_id;
1190 struct iser_conn *iser_conn = container_of(ib_conn, struct iser_conn,
1193 if (wc->status != IB_WC_WR_FLUSH_ERR)
1194 if (iser_conn->iscsi_conn)
1195 iscsi_conn_failure(iser_conn->iscsi_conn,
1196 ISCSI_ERR_CONN_FAILED);
1198 if (is_iser_tx_desc(iser_conn, wr_id)) {
1199 struct iser_tx_desc *desc = wr_id;
1201 if (desc->type == ISCSI_TX_DATAOUT)
1202 kmem_cache_free(ig.desc_cache, desc);
1204 ib_conn->post_recv_buf_count--;
1209 * iser_handle_wc - handle a single work completion
1210 * @wc: work completion
1212 * Soft-IRQ context, work completion can be either
1213 * SEND or RECV, and can turn out successful or
1214 * with error (or flush error).
1216 static void iser_handle_wc(struct ib_wc *wc)
1218 struct ib_conn *ib_conn;
1219 struct iser_tx_desc *tx_desc;
1220 struct iser_rx_desc *rx_desc;
1222 ib_conn = wc->qp->qp_context;
1223 if (wc->status == IB_WC_SUCCESS) {
1224 if (wc->opcode == IB_WC_RECV) {
1225 rx_desc = (struct iser_rx_desc *)(uintptr_t)wc->wr_id;
1226 iser_rcv_completion(rx_desc, wc->byte_len,
1229 if (wc->opcode == IB_WC_SEND) {
1230 tx_desc = (struct iser_tx_desc *)(uintptr_t)wc->wr_id;
1231 iser_snd_completion(tx_desc, ib_conn);
1233 iser_err("Unknown wc opcode %d\n", wc->opcode);
1236 if (wc->status != IB_WC_WR_FLUSH_ERR)
1237 iser_err("wr id %llx status %d vend_err %x\n",
1238 wc->wr_id, wc->status, wc->vendor_err);
1240 iser_dbg("flush error: wr id %llx\n", wc->wr_id);
1242 if (wc->wr_id != ISER_FASTREG_LI_WRID &&
1243 wc->wr_id != ISER_BEACON_WRID)
1244 iser_handle_comp_error(ib_conn, wc);
1246 /* complete in case all flush errors were consumed */
1247 if (wc->wr_id == ISER_BEACON_WRID)
1248 complete(&ib_conn->flush_comp);
1253 * iser_cq_tasklet_fn - iSER completion polling loop
1254 * @data: iSER completion context
1256 * Soft-IRQ context, polling connection CQ until
1257 * either CQ was empty or we exausted polling budget
1259 static void iser_cq_tasklet_fn(unsigned long data)
1261 struct iser_comp *comp = (struct iser_comp *)data;
1262 struct ib_cq *cq = comp->cq;
1263 struct ib_wc *const wcs = comp->wcs;
1264 int i, n, completed = 0;
1266 while ((n = ib_poll_cq(cq, ARRAY_SIZE(comp->wcs), wcs)) > 0) {
1267 for (i = 0; i < n; i++)
1268 iser_handle_wc(&wcs[i]);
1271 if (completed >= iser_cq_poll_limit)
1276 * It is assumed here that arming CQ only once its empty
1277 * would not cause interrupts to be missed.
1279 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1281 iser_dbg("got %d completions\n", completed);
1284 static void iser_cq_callback(struct ib_cq *cq, void *cq_context)
1286 struct iser_comp *comp = cq_context;
1288 tasklet_schedule(&comp->tasklet);
1291 u8 iser_check_task_pi_status(struct iscsi_iser_task *iser_task,
1292 enum iser_data_dir cmd_dir, sector_t *sector)
1294 struct iser_mem_reg *reg = &iser_task->rdma_regd[cmd_dir].reg;
1295 struct fast_reg_descriptor *desc = reg->mem_h;
1296 unsigned long sector_size = iser_task->sc->device->sector_size;
1297 struct ib_mr_status mr_status;
1300 if (desc && desc->reg_indicators & ISER_FASTREG_PROTECTED) {
1301 desc->reg_indicators &= ~ISER_FASTREG_PROTECTED;
1302 ret = ib_check_mr_status(desc->pi_ctx->sig_mr,
1303 IB_MR_CHECK_SIG_STATUS, &mr_status);
1305 pr_err("ib_check_mr_status failed, ret %d\n", ret);
1309 if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1310 sector_t sector_off = mr_status.sig_err.sig_err_offset;
1312 do_div(sector_off, sector_size + 8);
1313 *sector = scsi_get_lba(iser_task->sc) + sector_off;
1315 pr_err("PI error found type %d at sector %llx "
1316 "expected %x vs actual %x\n",
1317 mr_status.sig_err.err_type,
1318 (unsigned long long)*sector,
1319 mr_status.sig_err.expected,
1320 mr_status.sig_err.actual);
1322 switch (mr_status.sig_err.err_type) {
1323 case IB_SIG_BAD_GUARD:
1325 case IB_SIG_BAD_REFTAG:
1327 case IB_SIG_BAD_APPTAG:
1335 /* Not alot we can do here, return ambiguous guard error */