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("cq event %s (%d)\n",
55 ib_event_msg(cause->event), cause->event);
58 static void iser_qp_event_callback(struct ib_event *cause, void *context)
60 iser_err("qp event %s (%d)\n",
61 ib_event_msg(cause->event), cause->event);
64 static void iser_event_handler(struct ib_event_handler *handler,
65 struct ib_event *event)
67 iser_err("async event %s (%d) on device %s port %d\n",
68 ib_event_msg(event->event), event->event,
69 event->device->name, event->element.port_num);
73 * iser_create_device_ib_res - creates Protection Domain (PD), Completion
74 * Queue (CQ), DMA Memory Region (DMA MR) with the device associated with
77 * returns 0 on success, -1 on failure
79 static int iser_create_device_ib_res(struct iser_device *device)
81 struct ib_device_attr *dev_attr = &device->dev_attr;
84 ret = ib_query_device(device->ib_device, dev_attr);
86 pr_warn("Query device failed for %s\n", device->ib_device->name);
90 ret = iser_assign_reg_ops(device);
94 device->comps_used = min_t(int, num_online_cpus(),
95 device->ib_device->num_comp_vectors);
97 device->comps = kcalloc(device->comps_used, sizeof(*device->comps),
102 max_cqe = min(ISER_MAX_CQ_LEN, dev_attr->max_cqe);
104 iser_info("using %d CQs, device %s supports %d vectors max_cqe %d\n",
105 device->comps_used, device->ib_device->name,
106 device->ib_device->num_comp_vectors, max_cqe);
108 device->pd = ib_alloc_pd(device->ib_device);
109 if (IS_ERR(device->pd))
112 for (i = 0; i < device->comps_used; i++) {
113 struct ib_cq_init_attr cq_attr = {};
114 struct iser_comp *comp = &device->comps[i];
116 comp->device = device;
117 cq_attr.cqe = max_cqe;
118 cq_attr.comp_vector = i;
119 comp->cq = ib_create_cq(device->ib_device,
121 iser_cq_event_callback,
124 if (IS_ERR(comp->cq)) {
129 if (ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP))
132 tasklet_init(&comp->tasklet, iser_cq_tasklet_fn,
133 (unsigned long)comp);
136 device->mr = ib_get_dma_mr(device->pd, IB_ACCESS_LOCAL_WRITE |
137 IB_ACCESS_REMOTE_WRITE |
138 IB_ACCESS_REMOTE_READ);
139 if (IS_ERR(device->mr))
142 INIT_IB_EVENT_HANDLER(&device->event_handler, device->ib_device,
144 if (ib_register_event_handler(&device->event_handler))
150 ib_dereg_mr(device->mr);
152 for (i = 0; i < device->comps_used; i++)
153 tasklet_kill(&device->comps[i].tasklet);
155 for (i = 0; i < device->comps_used; i++) {
156 struct iser_comp *comp = &device->comps[i];
159 ib_destroy_cq(comp->cq);
161 ib_dealloc_pd(device->pd);
163 kfree(device->comps);
165 iser_err("failed to allocate an IB resource\n");
170 * iser_free_device_ib_res - destroy/dealloc/dereg the DMA MR,
171 * CQ and PD created with the device associated with the adapator.
173 static void iser_free_device_ib_res(struct iser_device *device)
176 BUG_ON(device->mr == NULL);
178 for (i = 0; i < device->comps_used; i++) {
179 struct iser_comp *comp = &device->comps[i];
181 tasklet_kill(&comp->tasklet);
182 ib_destroy_cq(comp->cq);
186 (void)ib_unregister_event_handler(&device->event_handler);
187 (void)ib_dereg_mr(device->mr);
188 (void)ib_dealloc_pd(device->pd);
190 kfree(device->comps);
191 device->comps = NULL;
198 * iser_alloc_fmr_pool - Creates FMR pool and page_vector
200 * returns 0 on success, or errno code on failure
202 int iser_alloc_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)) {
232 ret = PTR_ERR(ib_conn->fmr.pool);
233 iser_err("FMR allocation failed, err %d\n", ret);
240 kfree(ib_conn->fmr.page_vec);
241 ib_conn->fmr.page_vec = NULL;
246 * iser_free_fmr_pool - releases the FMR pool and page vec
248 void iser_free_fmr_pool(struct ib_conn *ib_conn)
250 iser_info("freeing conn %p fmr pool %p\n",
251 ib_conn, ib_conn->fmr.pool);
253 ib_destroy_fmr_pool(ib_conn->fmr.pool);
254 ib_conn->fmr.pool = NULL;
256 kfree(ib_conn->fmr.page_vec);
257 ib_conn->fmr.page_vec = NULL;
261 iser_alloc_reg_res(struct ib_device *ib_device, struct ib_pd *pd,
262 struct iser_reg_resources *res)
266 res->frpl = ib_alloc_fast_reg_page_list(ib_device,
267 ISCSI_ISER_SG_TABLESIZE + 1);
268 if (IS_ERR(res->frpl)) {
269 ret = PTR_ERR(res->frpl);
270 iser_err("Failed to allocate ib_fast_reg_page_list err=%d\n",
272 return PTR_ERR(res->frpl);
275 res->mr = ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG,
276 ISCSI_ISER_SG_TABLESIZE + 1);
277 if (IS_ERR(res->mr)) {
278 ret = PTR_ERR(res->mr);
279 iser_err("Failed to allocate ib_fast_reg_mr err=%d\n", ret);
280 goto fast_reg_mr_failure;
287 ib_free_fast_reg_page_list(res->frpl);
293 iser_free_reg_res(struct iser_reg_resources *rsc)
295 ib_dereg_mr(rsc->mr);
296 ib_free_fast_reg_page_list(rsc->frpl);
300 iser_alloc_pi_ctx(struct ib_device *ib_device, struct ib_pd *pd,
301 struct iser_fr_desc *desc)
303 struct iser_pi_context *pi_ctx = NULL;
306 desc->pi_ctx = kzalloc(sizeof(*desc->pi_ctx), GFP_KERNEL);
310 pi_ctx = desc->pi_ctx;
312 ret = iser_alloc_reg_res(ib_device, pd, &pi_ctx->rsc);
314 iser_err("failed to allocate reg_resources\n");
315 goto alloc_reg_res_err;
318 pi_ctx->sig_mr = ib_alloc_mr(pd, IB_MR_TYPE_SIGNATURE, 2);
319 if (IS_ERR(pi_ctx->sig_mr)) {
320 ret = PTR_ERR(pi_ctx->sig_mr);
323 pi_ctx->sig_mr_valid = 1;
324 desc->pi_ctx->sig_protected = 0;
329 iser_free_reg_res(&pi_ctx->rsc);
337 iser_free_pi_ctx(struct iser_pi_context *pi_ctx)
339 iser_free_reg_res(&pi_ctx->rsc);
340 ib_dereg_mr(pi_ctx->sig_mr);
344 static struct iser_fr_desc *
345 iser_create_fastreg_desc(struct ib_device *ib_device, struct ib_pd *pd,
348 struct iser_fr_desc *desc;
351 desc = kzalloc(sizeof(*desc), GFP_KERNEL);
353 return ERR_PTR(-ENOMEM);
355 ret = iser_alloc_reg_res(ib_device, pd, &desc->rsc);
357 goto reg_res_alloc_failure;
360 ret = iser_alloc_pi_ctx(ib_device, pd, desc);
362 goto pi_ctx_alloc_failure;
367 pi_ctx_alloc_failure:
368 iser_free_reg_res(&desc->rsc);
369 reg_res_alloc_failure:
376 * iser_alloc_fastreg_pool - Creates pool of fast_reg descriptors
377 * for fast registration work requests.
378 * returns 0 on success, or errno code on failure
380 int iser_alloc_fastreg_pool(struct ib_conn *ib_conn, unsigned cmds_max)
382 struct iser_device *device = ib_conn->device;
383 struct iser_fr_desc *desc;
386 INIT_LIST_HEAD(&ib_conn->fastreg.pool);
387 ib_conn->fastreg.pool_size = 0;
388 for (i = 0; i < cmds_max; i++) {
389 desc = iser_create_fastreg_desc(device->ib_device, device->pd,
390 ib_conn->pi_support);
396 list_add_tail(&desc->list, &ib_conn->fastreg.pool);
397 ib_conn->fastreg.pool_size++;
403 iser_free_fastreg_pool(ib_conn);
408 * iser_free_fastreg_pool - releases the pool of fast_reg descriptors
410 void iser_free_fastreg_pool(struct ib_conn *ib_conn)
412 struct iser_fr_desc *desc, *tmp;
415 if (list_empty(&ib_conn->fastreg.pool))
418 iser_info("freeing conn %p fr pool\n", ib_conn);
420 list_for_each_entry_safe(desc, tmp, &ib_conn->fastreg.pool, list) {
421 list_del(&desc->list);
422 iser_free_reg_res(&desc->rsc);
424 iser_free_pi_ctx(desc->pi_ctx);
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);
435 * iser_create_ib_conn_res - Queue-Pair (QP)
437 * returns 0 on success, -1 on failure
439 static int iser_create_ib_conn_res(struct ib_conn *ib_conn)
441 struct iser_conn *iser_conn = container_of(ib_conn, struct iser_conn,
443 struct iser_device *device;
444 struct ib_device_attr *dev_attr;
445 struct ib_qp_init_attr init_attr;
447 int index, min_index = 0;
449 BUG_ON(ib_conn->device == NULL);
451 device = ib_conn->device;
452 dev_attr = &device->dev_attr;
454 memset(&init_attr, 0, sizeof init_attr);
456 mutex_lock(&ig.connlist_mutex);
457 /* select the CQ with the minimal number of usages */
458 for (index = 0; index < device->comps_used; index++) {
459 if (device->comps[index].active_qps <
460 device->comps[min_index].active_qps)
463 ib_conn->comp = &device->comps[min_index];
464 ib_conn->comp->active_qps++;
465 mutex_unlock(&ig.connlist_mutex);
466 iser_info("cq index %d used for ib_conn %p\n", min_index, ib_conn);
468 init_attr.event_handler = iser_qp_event_callback;
469 init_attr.qp_context = (void *)ib_conn;
470 init_attr.send_cq = ib_conn->comp->cq;
471 init_attr.recv_cq = ib_conn->comp->cq;
472 init_attr.cap.max_recv_wr = ISER_QP_MAX_RECV_DTOS;
473 init_attr.cap.max_send_sge = 2;
474 init_attr.cap.max_recv_sge = 1;
475 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
476 init_attr.qp_type = IB_QPT_RC;
477 if (ib_conn->pi_support) {
478 init_attr.cap.max_send_wr = ISER_QP_SIG_MAX_REQ_DTOS + 1;
479 init_attr.create_flags |= IB_QP_CREATE_SIGNATURE_EN;
480 iser_conn->max_cmds =
481 ISER_GET_MAX_XMIT_CMDS(ISER_QP_SIG_MAX_REQ_DTOS);
483 if (dev_attr->max_qp_wr > ISER_QP_MAX_REQ_DTOS) {
484 init_attr.cap.max_send_wr = ISER_QP_MAX_REQ_DTOS + 1;
485 iser_conn->max_cmds =
486 ISER_GET_MAX_XMIT_CMDS(ISER_QP_MAX_REQ_DTOS);
488 init_attr.cap.max_send_wr = dev_attr->max_qp_wr;
489 iser_conn->max_cmds =
490 ISER_GET_MAX_XMIT_CMDS(dev_attr->max_qp_wr);
491 iser_dbg("device %s supports max_send_wr %d\n",
492 device->ib_device->name, dev_attr->max_qp_wr);
496 ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
500 ib_conn->qp = ib_conn->cma_id->qp;
501 iser_info("setting conn %p cma_id %p qp %p\n",
502 ib_conn, ib_conn->cma_id,
503 ib_conn->cma_id->qp);
507 mutex_lock(&ig.connlist_mutex);
508 ib_conn->comp->active_qps--;
509 mutex_unlock(&ig.connlist_mutex);
510 iser_err("unable to alloc mem or create resource, err %d\n", ret);
516 * based on the resolved device node GUID see if there already allocated
517 * device for this device. If there's no such, create one.
520 struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
522 struct iser_device *device;
524 mutex_lock(&ig.device_list_mutex);
526 list_for_each_entry(device, &ig.device_list, ig_list)
527 /* find if there's a match using the node GUID */
528 if (device->ib_device->node_guid == cma_id->device->node_guid)
531 device = kzalloc(sizeof *device, GFP_KERNEL);
535 /* assign this device to the device */
536 device->ib_device = cma_id->device;
537 /* init the device and link it into ig device list */
538 if (iser_create_device_ib_res(device)) {
543 list_add(&device->ig_list, &ig.device_list);
548 mutex_unlock(&ig.device_list_mutex);
552 /* if there's no demand for this device, release it */
553 static void iser_device_try_release(struct iser_device *device)
555 mutex_lock(&ig.device_list_mutex);
557 iser_info("device %p refcount %d\n", device, device->refcount);
558 if (!device->refcount) {
559 iser_free_device_ib_res(device);
560 list_del(&device->ig_list);
563 mutex_unlock(&ig.device_list_mutex);
567 * Called with state mutex held
569 static int iser_conn_state_comp_exch(struct iser_conn *iser_conn,
570 enum iser_conn_state comp,
571 enum iser_conn_state exch)
575 ret = (iser_conn->state == comp);
577 iser_conn->state = exch;
582 void iser_release_work(struct work_struct *work)
584 struct iser_conn *iser_conn;
586 iser_conn = container_of(work, struct iser_conn, release_work);
588 /* Wait for conn_stop to complete */
589 wait_for_completion(&iser_conn->stop_completion);
590 /* Wait for IB resouces cleanup to complete */
591 wait_for_completion(&iser_conn->ib_completion);
593 mutex_lock(&iser_conn->state_mutex);
594 iser_conn->state = ISER_CONN_DOWN;
595 mutex_unlock(&iser_conn->state_mutex);
597 iser_conn_release(iser_conn);
601 * iser_free_ib_conn_res - release IB related resources
602 * @iser_conn: iser connection struct
603 * @destroy: indicator if we need to try to release the
604 * iser device and memory regoins pool (only iscsi
605 * shutdown and DEVICE_REMOVAL will use this).
607 * This routine is called with the iser state mutex held
608 * so the cm_id removal is out of here. It is Safe to
609 * be invoked multiple times.
611 static void iser_free_ib_conn_res(struct iser_conn *iser_conn,
614 struct ib_conn *ib_conn = &iser_conn->ib_conn;
615 struct iser_device *device = ib_conn->device;
617 iser_info("freeing conn %p cma_id %p qp %p\n",
618 iser_conn, ib_conn->cma_id, ib_conn->qp);
620 if (ib_conn->qp != NULL) {
621 ib_conn->comp->active_qps--;
622 rdma_destroy_qp(ib_conn->cma_id);
627 if (iser_conn->rx_descs)
628 iser_free_rx_descriptors(iser_conn);
630 if (device != NULL) {
631 iser_device_try_release(device);
632 ib_conn->device = NULL;
638 * Frees all conn objects and deallocs conn descriptor
640 void iser_conn_release(struct iser_conn *iser_conn)
642 struct ib_conn *ib_conn = &iser_conn->ib_conn;
644 mutex_lock(&ig.connlist_mutex);
645 list_del(&iser_conn->conn_list);
646 mutex_unlock(&ig.connlist_mutex);
648 mutex_lock(&iser_conn->state_mutex);
649 /* In case we endup here without ep_disconnect being invoked. */
650 if (iser_conn->state != ISER_CONN_DOWN) {
651 iser_warn("iser conn %p state %d, expected state down.\n",
652 iser_conn, iser_conn->state);
653 iscsi_destroy_endpoint(iser_conn->ep);
654 iser_conn->state = ISER_CONN_DOWN;
657 * In case we never got to bind stage, we still need to
658 * release IB resources (which is safe to call more than once).
660 iser_free_ib_conn_res(iser_conn, true);
661 mutex_unlock(&iser_conn->state_mutex);
663 if (ib_conn->cma_id != NULL) {
664 rdma_destroy_id(ib_conn->cma_id);
665 ib_conn->cma_id = NULL;
672 * triggers start of the disconnect procedures and wait for them to be done
673 * Called with state mutex held
675 int iser_conn_terminate(struct iser_conn *iser_conn)
677 struct ib_conn *ib_conn = &iser_conn->ib_conn;
678 struct ib_send_wr *bad_wr;
681 /* terminate the iser conn only if the conn state is UP */
682 if (!iser_conn_state_comp_exch(iser_conn, ISER_CONN_UP,
683 ISER_CONN_TERMINATING))
686 iser_info("iser_conn %p state %d\n", iser_conn, iser_conn->state);
688 /* suspend queuing of new iscsi commands */
689 if (iser_conn->iscsi_conn)
690 iscsi_suspend_queue(iser_conn->iscsi_conn);
693 * In case we didn't already clean up the cma_id (peer initiated
694 * a disconnection), we need to Cause the CMA to change the QP
697 if (ib_conn->cma_id) {
698 err = rdma_disconnect(ib_conn->cma_id);
700 iser_err("Failed to disconnect, conn: 0x%p err %d\n",
703 /* post an indication that all flush errors were consumed */
704 err = ib_post_send(ib_conn->qp, &ib_conn->beacon, &bad_wr);
706 iser_err("conn %p failed to post beacon", ib_conn);
710 wait_for_completion(&ib_conn->flush_comp);
717 * Called with state mutex held
719 static void iser_connect_error(struct rdma_cm_id *cma_id)
721 struct iser_conn *iser_conn;
723 iser_conn = (struct iser_conn *)cma_id->context;
724 iser_conn->state = ISER_CONN_TERMINATING;
728 * Called with state mutex held
730 static void iser_addr_handler(struct rdma_cm_id *cma_id)
732 struct iser_device *device;
733 struct iser_conn *iser_conn;
734 struct ib_conn *ib_conn;
737 iser_conn = (struct iser_conn *)cma_id->context;
738 if (iser_conn->state != ISER_CONN_PENDING)
742 ib_conn = &iser_conn->ib_conn;
743 device = iser_device_find_by_ib_device(cma_id);
745 iser_err("device lookup/creation failed\n");
746 iser_connect_error(cma_id);
750 ib_conn->device = device;
752 /* connection T10-PI support */
753 if (iser_pi_enable) {
754 if (!(device->dev_attr.device_cap_flags &
755 IB_DEVICE_SIGNATURE_HANDOVER)) {
756 iser_warn("T10-PI requested but not supported on %s, "
757 "continue without T10-PI\n",
758 ib_conn->device->ib_device->name);
759 ib_conn->pi_support = false;
761 ib_conn->pi_support = true;
765 ret = rdma_resolve_route(cma_id, 1000);
767 iser_err("resolve route failed: %d\n", ret);
768 iser_connect_error(cma_id);
774 * Called with state mutex held
776 static void iser_route_handler(struct rdma_cm_id *cma_id)
778 struct rdma_conn_param conn_param;
780 struct iser_cm_hdr req_hdr;
781 struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
782 struct ib_conn *ib_conn = &iser_conn->ib_conn;
783 struct iser_device *device = ib_conn->device;
785 if (iser_conn->state != ISER_CONN_PENDING)
789 ret = iser_create_ib_conn_res(ib_conn);
793 memset(&conn_param, 0, sizeof conn_param);
794 conn_param.responder_resources = device->dev_attr.max_qp_rd_atom;
795 conn_param.initiator_depth = 1;
796 conn_param.retry_count = 7;
797 conn_param.rnr_retry_count = 6;
799 memset(&req_hdr, 0, sizeof(req_hdr));
800 req_hdr.flags = (ISER_ZBVA_NOT_SUPPORTED |
801 ISER_SEND_W_INV_NOT_SUPPORTED);
802 conn_param.private_data = (void *)&req_hdr;
803 conn_param.private_data_len = sizeof(struct iser_cm_hdr);
805 ret = rdma_connect(cma_id, &conn_param);
807 iser_err("failure connecting: %d\n", ret);
813 iser_connect_error(cma_id);
816 static void iser_connected_handler(struct rdma_cm_id *cma_id)
818 struct iser_conn *iser_conn;
819 struct ib_qp_attr attr;
820 struct ib_qp_init_attr init_attr;
822 iser_conn = (struct iser_conn *)cma_id->context;
823 if (iser_conn->state != ISER_CONN_PENDING)
827 (void)ib_query_qp(cma_id->qp, &attr, ~0, &init_attr);
828 iser_info("remote qpn:%x my qpn:%x\n", attr.dest_qp_num, cma_id->qp->qp_num);
830 iser_conn->state = ISER_CONN_UP;
831 complete(&iser_conn->up_completion);
834 static void iser_disconnected_handler(struct rdma_cm_id *cma_id)
836 struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
838 if (iser_conn_terminate(iser_conn)) {
839 if (iser_conn->iscsi_conn)
840 iscsi_conn_failure(iser_conn->iscsi_conn,
841 ISCSI_ERR_CONN_FAILED);
843 iser_err("iscsi_iser connection isn't bound\n");
847 static void iser_cleanup_handler(struct rdma_cm_id *cma_id,
850 struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
853 * We are not guaranteed that we visited disconnected_handler
854 * by now, call it here to be safe that we handle CM drep
857 iser_disconnected_handler(cma_id);
858 iser_free_ib_conn_res(iser_conn, destroy);
859 complete(&iser_conn->ib_completion);
862 static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
864 struct iser_conn *iser_conn;
867 iser_conn = (struct iser_conn *)cma_id->context;
868 iser_info("%s (%d): status %d conn %p id %p\n",
869 rdma_event_msg(event->event), event->event,
870 event->status, cma_id->context, cma_id);
872 mutex_lock(&iser_conn->state_mutex);
873 switch (event->event) {
874 case RDMA_CM_EVENT_ADDR_RESOLVED:
875 iser_addr_handler(cma_id);
877 case RDMA_CM_EVENT_ROUTE_RESOLVED:
878 iser_route_handler(cma_id);
880 case RDMA_CM_EVENT_ESTABLISHED:
881 iser_connected_handler(cma_id);
883 case RDMA_CM_EVENT_ADDR_ERROR:
884 case RDMA_CM_EVENT_ROUTE_ERROR:
885 case RDMA_CM_EVENT_CONNECT_ERROR:
886 case RDMA_CM_EVENT_UNREACHABLE:
887 case RDMA_CM_EVENT_REJECTED:
888 iser_connect_error(cma_id);
890 case RDMA_CM_EVENT_DISCONNECTED:
891 case RDMA_CM_EVENT_ADDR_CHANGE:
892 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
893 iser_cleanup_handler(cma_id, false);
895 case RDMA_CM_EVENT_DEVICE_REMOVAL:
897 * we *must* destroy the device as we cannot rely
898 * on iscsid to be around to initiate error handling.
899 * also if we are not in state DOWN implicitly destroy
902 iser_cleanup_handler(cma_id, true);
903 if (iser_conn->state != ISER_CONN_DOWN) {
904 iser_conn->ib_conn.cma_id = NULL;
909 iser_err("Unexpected RDMA CM event: %s (%d)\n",
910 rdma_event_msg(event->event), event->event);
913 mutex_unlock(&iser_conn->state_mutex);
918 void iser_conn_init(struct iser_conn *iser_conn)
920 iser_conn->state = ISER_CONN_INIT;
921 iser_conn->ib_conn.post_recv_buf_count = 0;
922 init_completion(&iser_conn->ib_conn.flush_comp);
923 init_completion(&iser_conn->stop_completion);
924 init_completion(&iser_conn->ib_completion);
925 init_completion(&iser_conn->up_completion);
926 INIT_LIST_HEAD(&iser_conn->conn_list);
927 spin_lock_init(&iser_conn->ib_conn.lock);
928 mutex_init(&iser_conn->state_mutex);
932 * starts the process of connecting to the target
933 * sleeps until the connection is established or rejected
935 int iser_connect(struct iser_conn *iser_conn,
936 struct sockaddr *src_addr,
937 struct sockaddr *dst_addr,
940 struct ib_conn *ib_conn = &iser_conn->ib_conn;
943 mutex_lock(&iser_conn->state_mutex);
945 sprintf(iser_conn->name, "%pISp", dst_addr);
947 iser_info("connecting to: %s\n", iser_conn->name);
949 /* the device is known only --after-- address resolution */
950 ib_conn->device = NULL;
952 iser_conn->state = ISER_CONN_PENDING;
954 ib_conn->beacon.wr_id = ISER_BEACON_WRID;
955 ib_conn->beacon.opcode = IB_WR_SEND;
957 ib_conn->cma_id = rdma_create_id(iser_cma_handler,
959 RDMA_PS_TCP, IB_QPT_RC);
960 if (IS_ERR(ib_conn->cma_id)) {
961 err = PTR_ERR(ib_conn->cma_id);
962 iser_err("rdma_create_id failed: %d\n", err);
966 err = rdma_resolve_addr(ib_conn->cma_id, src_addr, dst_addr, 1000);
968 iser_err("rdma_resolve_addr failed: %d\n", err);
973 wait_for_completion_interruptible(&iser_conn->up_completion);
975 if (iser_conn->state != ISER_CONN_UP) {
977 goto connect_failure;
980 mutex_unlock(&iser_conn->state_mutex);
982 mutex_lock(&ig.connlist_mutex);
983 list_add(&iser_conn->conn_list, &ig.connlist);
984 mutex_unlock(&ig.connlist_mutex);
988 ib_conn->cma_id = NULL;
990 iser_conn->state = ISER_CONN_DOWN;
992 mutex_unlock(&iser_conn->state_mutex);
993 iser_conn_release(iser_conn);
997 int iser_post_recvl(struct iser_conn *iser_conn)
999 struct ib_recv_wr rx_wr, *rx_wr_failed;
1000 struct ib_conn *ib_conn = &iser_conn->ib_conn;
1004 sge.addr = iser_conn->login_resp_dma;
1005 sge.length = ISER_RX_LOGIN_SIZE;
1006 sge.lkey = ib_conn->device->mr->lkey;
1008 rx_wr.wr_id = (uintptr_t)iser_conn->login_resp_buf;
1009 rx_wr.sg_list = &sge;
1013 ib_conn->post_recv_buf_count++;
1014 ib_ret = ib_post_recv(ib_conn->qp, &rx_wr, &rx_wr_failed);
1016 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1017 ib_conn->post_recv_buf_count--;
1022 int iser_post_recvm(struct iser_conn *iser_conn, int count)
1024 struct ib_recv_wr *rx_wr, *rx_wr_failed;
1026 struct ib_conn *ib_conn = &iser_conn->ib_conn;
1027 unsigned int my_rx_head = iser_conn->rx_desc_head;
1028 struct iser_rx_desc *rx_desc;
1030 for (rx_wr = ib_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
1031 rx_desc = &iser_conn->rx_descs[my_rx_head];
1032 rx_wr->wr_id = (uintptr_t)rx_desc;
1033 rx_wr->sg_list = &rx_desc->rx_sg;
1035 rx_wr->next = rx_wr + 1;
1036 my_rx_head = (my_rx_head + 1) & iser_conn->qp_max_recv_dtos_mask;
1040 rx_wr->next = NULL; /* mark end of work requests list */
1042 ib_conn->post_recv_buf_count += count;
1043 ib_ret = ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &rx_wr_failed);
1045 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1046 ib_conn->post_recv_buf_count -= count;
1048 iser_conn->rx_desc_head = my_rx_head;
1054 * iser_start_send - Initiate a Send DTO operation
1056 * returns 0 on success, -1 on failure
1058 int iser_post_send(struct ib_conn *ib_conn, struct iser_tx_desc *tx_desc,
1062 struct ib_send_wr send_wr, *send_wr_failed;
1064 ib_dma_sync_single_for_device(ib_conn->device->ib_device,
1065 tx_desc->dma_addr, ISER_HEADERS_LEN,
1068 send_wr.next = NULL;
1069 send_wr.wr_id = (uintptr_t)tx_desc;
1070 send_wr.sg_list = tx_desc->tx_sg;
1071 send_wr.num_sge = tx_desc->num_sge;
1072 send_wr.opcode = IB_WR_SEND;
1073 send_wr.send_flags = signal ? IB_SEND_SIGNALED : 0;
1075 ib_ret = ib_post_send(ib_conn->qp, &send_wr, &send_wr_failed);
1077 iser_err("ib_post_send failed, ret:%d\n", ib_ret);
1083 * is_iser_tx_desc - Indicate if the completion wr_id
1084 * is a TX descriptor or not.
1085 * @iser_conn: iser connection
1086 * @wr_id: completion WR identifier
1088 * Since we cannot rely on wc opcode in FLUSH errors
1089 * we must work around it by checking if the wr_id address
1090 * falls in the iser connection rx_descs buffer. If so
1091 * it is an RX descriptor, otherwize it is a TX.
1094 is_iser_tx_desc(struct iser_conn *iser_conn, void *wr_id)
1096 void *start = iser_conn->rx_descs;
1097 int len = iser_conn->num_rx_descs * sizeof(*iser_conn->rx_descs);
1099 if (wr_id >= start && wr_id < start + len)
1106 * iser_handle_comp_error() - Handle error completion
1107 * @ib_conn: connection RDMA resources
1108 * @wc: work completion
1110 * Notes: We may handle a FLUSH error completion and in this case
1111 * we only cleanup in case TX type was DATAOUT. For non-FLUSH
1112 * error completion we should also notify iscsi layer that
1113 * connection is failed (in case we passed bind stage).
1116 iser_handle_comp_error(struct ib_conn *ib_conn,
1119 void *wr_id = (void *)(uintptr_t)wc->wr_id;
1120 struct iser_conn *iser_conn = container_of(ib_conn, struct iser_conn,
1123 if (wc->status != IB_WC_WR_FLUSH_ERR)
1124 if (iser_conn->iscsi_conn)
1125 iscsi_conn_failure(iser_conn->iscsi_conn,
1126 ISCSI_ERR_CONN_FAILED);
1128 if (wc->wr_id == ISER_FASTREG_LI_WRID)
1131 if (is_iser_tx_desc(iser_conn, wr_id)) {
1132 struct iser_tx_desc *desc = wr_id;
1134 if (desc->type == ISCSI_TX_DATAOUT)
1135 kmem_cache_free(ig.desc_cache, desc);
1137 ib_conn->post_recv_buf_count--;
1142 * iser_handle_wc - handle a single work completion
1143 * @wc: work completion
1145 * Soft-IRQ context, work completion can be either
1146 * SEND or RECV, and can turn out successful or
1147 * with error (or flush error).
1149 static void iser_handle_wc(struct ib_wc *wc)
1151 struct ib_conn *ib_conn;
1152 struct iser_tx_desc *tx_desc;
1153 struct iser_rx_desc *rx_desc;
1155 ib_conn = wc->qp->qp_context;
1156 if (likely(wc->status == IB_WC_SUCCESS)) {
1157 if (wc->opcode == IB_WC_RECV) {
1158 rx_desc = (struct iser_rx_desc *)(uintptr_t)wc->wr_id;
1159 iser_rcv_completion(rx_desc, wc->byte_len,
1162 if (wc->opcode == IB_WC_SEND) {
1163 tx_desc = (struct iser_tx_desc *)(uintptr_t)wc->wr_id;
1164 iser_snd_completion(tx_desc, ib_conn);
1166 iser_err("Unknown wc opcode %d\n", wc->opcode);
1169 if (wc->status != IB_WC_WR_FLUSH_ERR)
1170 iser_err("%s (%d): wr id %llx vend_err %x\n",
1171 ib_wc_status_msg(wc->status), wc->status,
1172 wc->wr_id, wc->vendor_err);
1174 iser_dbg("%s (%d): wr id %llx\n",
1175 ib_wc_status_msg(wc->status), wc->status,
1178 if (wc->wr_id == ISER_BEACON_WRID)
1179 /* all flush errors were consumed */
1180 complete(&ib_conn->flush_comp);
1182 iser_handle_comp_error(ib_conn, wc);
1187 * iser_cq_tasklet_fn - iSER completion polling loop
1188 * @data: iSER completion context
1190 * Soft-IRQ context, polling connection CQ until
1191 * either CQ was empty or we exausted polling budget
1193 static void iser_cq_tasklet_fn(unsigned long data)
1195 struct iser_comp *comp = (struct iser_comp *)data;
1196 struct ib_cq *cq = comp->cq;
1197 struct ib_wc *const wcs = comp->wcs;
1198 int i, n, completed = 0;
1200 while ((n = ib_poll_cq(cq, ARRAY_SIZE(comp->wcs), wcs)) > 0) {
1201 for (i = 0; i < n; i++)
1202 iser_handle_wc(&wcs[i]);
1205 if (completed >= iser_cq_poll_limit)
1210 * It is assumed here that arming CQ only once its empty
1211 * would not cause interrupts to be missed.
1213 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1215 iser_dbg("got %d completions\n", completed);
1218 static void iser_cq_callback(struct ib_cq *cq, void *cq_context)
1220 struct iser_comp *comp = cq_context;
1222 tasklet_schedule(&comp->tasklet);
1225 u8 iser_check_task_pi_status(struct iscsi_iser_task *iser_task,
1226 enum iser_data_dir cmd_dir, sector_t *sector)
1228 struct iser_mem_reg *reg = &iser_task->rdma_reg[cmd_dir];
1229 struct iser_fr_desc *desc = reg->mem_h;
1230 unsigned long sector_size = iser_task->sc->device->sector_size;
1231 struct ib_mr_status mr_status;
1234 if (desc && desc->pi_ctx->sig_protected) {
1235 desc->pi_ctx->sig_protected = 0;
1236 ret = ib_check_mr_status(desc->pi_ctx->sig_mr,
1237 IB_MR_CHECK_SIG_STATUS, &mr_status);
1239 pr_err("ib_check_mr_status failed, ret %d\n", ret);
1243 if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1244 sector_t sector_off = mr_status.sig_err.sig_err_offset;
1246 do_div(sector_off, sector_size + 8);
1247 *sector = scsi_get_lba(iser_task->sc) + sector_off;
1249 pr_err("PI error found type %d at sector %llx "
1250 "expected %x vs actual %x\n",
1251 mr_status.sig_err.err_type,
1252 (unsigned long long)*sector,
1253 mr_status.sig_err.expected,
1254 mr_status.sig_err.actual);
1256 switch (mr_status.sig_err.err_type) {
1257 case IB_SIG_BAD_GUARD:
1259 case IB_SIG_BAD_REFTAG:
1261 case IB_SIG_BAD_APPTAG:
1269 /* Not alot we can do here, return ambiguous guard error */