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 if (!iser_always_reg) {
137 int access = IB_ACCESS_LOCAL_WRITE |
138 IB_ACCESS_REMOTE_WRITE |
139 IB_ACCESS_REMOTE_READ;
141 device->mr = ib_get_dma_mr(device->pd, access);
142 if (IS_ERR(device->mr))
146 INIT_IB_EVENT_HANDLER(&device->event_handler, device->ib_device,
148 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 kfree(device->comps);
170 iser_err("failed to allocate an IB resource\n");
175 * iser_free_device_ib_res - destroy/dealloc/dereg the DMA MR,
176 * CQ and PD created with the device associated with the adapator.
178 static void iser_free_device_ib_res(struct iser_device *device)
182 for (i = 0; i < device->comps_used; i++) {
183 struct iser_comp *comp = &device->comps[i];
185 tasklet_kill(&comp->tasklet);
186 ib_destroy_cq(comp->cq);
190 (void)ib_unregister_event_handler(&device->event_handler);
192 (void)ib_dereg_mr(device->mr);
193 ib_dealloc_pd(device->pd);
195 kfree(device->comps);
196 device->comps = NULL;
203 * iser_alloc_fmr_pool - Creates FMR pool and page_vector
205 * returns 0 on success, or errno code on failure
207 int iser_alloc_fmr_pool(struct ib_conn *ib_conn,
211 struct iser_device *device = ib_conn->device;
212 struct iser_fr_pool *fr_pool = &ib_conn->fr_pool;
213 struct iser_page_vec *page_vec;
214 struct iser_fr_desc *desc;
215 struct ib_fmr_pool *fmr_pool;
216 struct ib_fmr_pool_param params;
219 INIT_LIST_HEAD(&fr_pool->list);
220 spin_lock_init(&fr_pool->lock);
222 desc = kzalloc(sizeof(*desc), GFP_KERNEL);
226 page_vec = kmalloc(sizeof(*page_vec) + (sizeof(u64) * size),
233 page_vec->pages = (u64 *)(page_vec + 1);
235 params.page_shift = SHIFT_4K;
236 params.max_pages_per_fmr = size;
237 /* make the pool size twice the max number of SCSI commands *
238 * the ML is expected to queue, watermark for unmap at 50% */
239 params.pool_size = cmds_max * 2;
240 params.dirty_watermark = cmds_max;
242 params.flush_function = NULL;
243 params.access = (IB_ACCESS_LOCAL_WRITE |
244 IB_ACCESS_REMOTE_WRITE |
245 IB_ACCESS_REMOTE_READ);
247 fmr_pool = ib_create_fmr_pool(device->pd, ¶ms);
248 if (IS_ERR(fmr_pool)) {
249 ret = PTR_ERR(fmr_pool);
250 iser_err("FMR allocation failed, err %d\n", ret);
254 desc->rsc.page_vec = page_vec;
255 desc->rsc.fmr_pool = fmr_pool;
256 list_add(&desc->list, &fr_pool->list);
269 * iser_free_fmr_pool - releases the FMR pool and page vec
271 void iser_free_fmr_pool(struct ib_conn *ib_conn)
273 struct iser_fr_pool *fr_pool = &ib_conn->fr_pool;
274 struct iser_fr_desc *desc;
276 desc = list_first_entry(&fr_pool->list,
277 struct iser_fr_desc, list);
278 list_del(&desc->list);
280 iser_info("freeing conn %p fmr pool %p\n",
281 ib_conn, desc->rsc.fmr_pool);
283 ib_destroy_fmr_pool(desc->rsc.fmr_pool);
284 kfree(desc->rsc.page_vec);
289 iser_alloc_reg_res(struct ib_device *ib_device,
291 struct iser_reg_resources *res,
296 res->mr = ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG, size);
297 if (IS_ERR(res->mr)) {
298 ret = PTR_ERR(res->mr);
299 iser_err("Failed to allocate ib_fast_reg_mr err=%d\n", ret);
308 iser_free_reg_res(struct iser_reg_resources *rsc)
310 ib_dereg_mr(rsc->mr);
314 iser_alloc_pi_ctx(struct ib_device *ib_device,
316 struct iser_fr_desc *desc,
319 struct iser_pi_context *pi_ctx = NULL;
322 desc->pi_ctx = kzalloc(sizeof(*desc->pi_ctx), GFP_KERNEL);
326 pi_ctx = desc->pi_ctx;
328 ret = iser_alloc_reg_res(ib_device, pd, &pi_ctx->rsc, size);
330 iser_err("failed to allocate reg_resources\n");
331 goto alloc_reg_res_err;
334 pi_ctx->sig_mr = ib_alloc_mr(pd, IB_MR_TYPE_SIGNATURE, 2);
335 if (IS_ERR(pi_ctx->sig_mr)) {
336 ret = PTR_ERR(pi_ctx->sig_mr);
339 pi_ctx->sig_mr_valid = 1;
340 desc->pi_ctx->sig_protected = 0;
345 iser_free_reg_res(&pi_ctx->rsc);
353 iser_free_pi_ctx(struct iser_pi_context *pi_ctx)
355 iser_free_reg_res(&pi_ctx->rsc);
356 ib_dereg_mr(pi_ctx->sig_mr);
360 static struct iser_fr_desc *
361 iser_create_fastreg_desc(struct ib_device *ib_device,
366 struct iser_fr_desc *desc;
369 desc = kzalloc(sizeof(*desc), GFP_KERNEL);
371 return ERR_PTR(-ENOMEM);
373 ret = iser_alloc_reg_res(ib_device, pd, &desc->rsc, size);
375 goto reg_res_alloc_failure;
378 ret = iser_alloc_pi_ctx(ib_device, pd, desc, size);
380 goto pi_ctx_alloc_failure;
385 pi_ctx_alloc_failure:
386 iser_free_reg_res(&desc->rsc);
387 reg_res_alloc_failure:
394 * iser_alloc_fastreg_pool - Creates pool of fast_reg descriptors
395 * for fast registration work requests.
396 * returns 0 on success, or errno code on failure
398 int iser_alloc_fastreg_pool(struct ib_conn *ib_conn,
402 struct iser_device *device = ib_conn->device;
403 struct iser_fr_pool *fr_pool = &ib_conn->fr_pool;
404 struct iser_fr_desc *desc;
407 INIT_LIST_HEAD(&fr_pool->list);
408 spin_lock_init(&fr_pool->lock);
410 for (i = 0; i < cmds_max; i++) {
411 desc = iser_create_fastreg_desc(device->ib_device, device->pd,
412 ib_conn->pi_support, size);
418 list_add_tail(&desc->list, &fr_pool->list);
425 iser_free_fastreg_pool(ib_conn);
430 * iser_free_fastreg_pool - releases the pool of fast_reg descriptors
432 void iser_free_fastreg_pool(struct ib_conn *ib_conn)
434 struct iser_fr_pool *fr_pool = &ib_conn->fr_pool;
435 struct iser_fr_desc *desc, *tmp;
438 if (list_empty(&fr_pool->list))
441 iser_info("freeing conn %p fr pool\n", ib_conn);
443 list_for_each_entry_safe(desc, tmp, &fr_pool->list, list) {
444 list_del(&desc->list);
445 iser_free_reg_res(&desc->rsc);
447 iser_free_pi_ctx(desc->pi_ctx);
452 if (i < fr_pool->size)
453 iser_warn("pool still has %d regions registered\n",
458 * iser_create_ib_conn_res - Queue-Pair (QP)
460 * returns 0 on success, -1 on failure
462 static int iser_create_ib_conn_res(struct ib_conn *ib_conn)
464 struct iser_conn *iser_conn = to_iser_conn(ib_conn);
465 struct iser_device *device;
466 struct ib_device_attr *dev_attr;
467 struct ib_qp_init_attr init_attr;
469 int index, min_index = 0;
471 BUG_ON(ib_conn->device == NULL);
473 device = ib_conn->device;
474 dev_attr = &device->dev_attr;
476 memset(&init_attr, 0, sizeof init_attr);
478 mutex_lock(&ig.connlist_mutex);
479 /* select the CQ with the minimal number of usages */
480 for (index = 0; index < device->comps_used; index++) {
481 if (device->comps[index].active_qps <
482 device->comps[min_index].active_qps)
485 ib_conn->comp = &device->comps[min_index];
486 ib_conn->comp->active_qps++;
487 mutex_unlock(&ig.connlist_mutex);
488 iser_info("cq index %d used for ib_conn %p\n", min_index, ib_conn);
490 init_attr.event_handler = iser_qp_event_callback;
491 init_attr.qp_context = (void *)ib_conn;
492 init_attr.send_cq = ib_conn->comp->cq;
493 init_attr.recv_cq = ib_conn->comp->cq;
494 init_attr.cap.max_recv_wr = ISER_QP_MAX_RECV_DTOS;
495 init_attr.cap.max_send_sge = 2;
496 init_attr.cap.max_recv_sge = 1;
497 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
498 init_attr.qp_type = IB_QPT_RC;
499 if (ib_conn->pi_support) {
500 init_attr.cap.max_send_wr = ISER_QP_SIG_MAX_REQ_DTOS + 1;
501 init_attr.create_flags |= IB_QP_CREATE_SIGNATURE_EN;
502 iser_conn->max_cmds =
503 ISER_GET_MAX_XMIT_CMDS(ISER_QP_SIG_MAX_REQ_DTOS);
505 if (dev_attr->max_qp_wr > ISER_QP_MAX_REQ_DTOS) {
506 init_attr.cap.max_send_wr = ISER_QP_MAX_REQ_DTOS + 1;
507 iser_conn->max_cmds =
508 ISER_GET_MAX_XMIT_CMDS(ISER_QP_MAX_REQ_DTOS);
510 init_attr.cap.max_send_wr = dev_attr->max_qp_wr;
511 iser_conn->max_cmds =
512 ISER_GET_MAX_XMIT_CMDS(dev_attr->max_qp_wr);
513 iser_dbg("device %s supports max_send_wr %d\n",
514 device->ib_device->name, dev_attr->max_qp_wr);
518 ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
522 ib_conn->qp = ib_conn->cma_id->qp;
523 iser_info("setting conn %p cma_id %p qp %p\n",
524 ib_conn, ib_conn->cma_id,
525 ib_conn->cma_id->qp);
529 mutex_lock(&ig.connlist_mutex);
530 ib_conn->comp->active_qps--;
531 mutex_unlock(&ig.connlist_mutex);
532 iser_err("unable to alloc mem or create resource, err %d\n", ret);
538 * based on the resolved device node GUID see if there already allocated
539 * device for this device. If there's no such, create one.
542 struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
544 struct iser_device *device;
546 mutex_lock(&ig.device_list_mutex);
548 list_for_each_entry(device, &ig.device_list, ig_list)
549 /* find if there's a match using the node GUID */
550 if (device->ib_device->node_guid == cma_id->device->node_guid)
553 device = kzalloc(sizeof *device, GFP_KERNEL);
557 /* assign this device to the device */
558 device->ib_device = cma_id->device;
559 /* init the device and link it into ig device list */
560 if (iser_create_device_ib_res(device)) {
565 list_add(&device->ig_list, &ig.device_list);
570 mutex_unlock(&ig.device_list_mutex);
574 /* if there's no demand for this device, release it */
575 static void iser_device_try_release(struct iser_device *device)
577 mutex_lock(&ig.device_list_mutex);
579 iser_info("device %p refcount %d\n", device, device->refcount);
580 if (!device->refcount) {
581 iser_free_device_ib_res(device);
582 list_del(&device->ig_list);
585 mutex_unlock(&ig.device_list_mutex);
589 * Called with state mutex held
591 static int iser_conn_state_comp_exch(struct iser_conn *iser_conn,
592 enum iser_conn_state comp,
593 enum iser_conn_state exch)
597 ret = (iser_conn->state == comp);
599 iser_conn->state = exch;
604 void iser_release_work(struct work_struct *work)
606 struct iser_conn *iser_conn;
608 iser_conn = container_of(work, struct iser_conn, release_work);
610 /* Wait for conn_stop to complete */
611 wait_for_completion(&iser_conn->stop_completion);
612 /* Wait for IB resouces cleanup to complete */
613 wait_for_completion(&iser_conn->ib_completion);
615 mutex_lock(&iser_conn->state_mutex);
616 iser_conn->state = ISER_CONN_DOWN;
617 mutex_unlock(&iser_conn->state_mutex);
619 iser_conn_release(iser_conn);
623 * iser_free_ib_conn_res - release IB related resources
624 * @iser_conn: iser connection struct
625 * @destroy: indicator if we need to try to release the
626 * iser device and memory regoins pool (only iscsi
627 * shutdown and DEVICE_REMOVAL will use this).
629 * This routine is called with the iser state mutex held
630 * so the cm_id removal is out of here. It is Safe to
631 * be invoked multiple times.
633 static void iser_free_ib_conn_res(struct iser_conn *iser_conn,
636 struct ib_conn *ib_conn = &iser_conn->ib_conn;
637 struct iser_device *device = ib_conn->device;
639 iser_info("freeing conn %p cma_id %p qp %p\n",
640 iser_conn, ib_conn->cma_id, ib_conn->qp);
642 if (ib_conn->qp != NULL) {
643 ib_conn->comp->active_qps--;
644 rdma_destroy_qp(ib_conn->cma_id);
649 if (iser_conn->rx_descs)
650 iser_free_rx_descriptors(iser_conn);
652 if (device != NULL) {
653 iser_device_try_release(device);
654 ib_conn->device = NULL;
660 * Frees all conn objects and deallocs conn descriptor
662 void iser_conn_release(struct iser_conn *iser_conn)
664 struct ib_conn *ib_conn = &iser_conn->ib_conn;
666 mutex_lock(&ig.connlist_mutex);
667 list_del(&iser_conn->conn_list);
668 mutex_unlock(&ig.connlist_mutex);
670 mutex_lock(&iser_conn->state_mutex);
671 /* In case we endup here without ep_disconnect being invoked. */
672 if (iser_conn->state != ISER_CONN_DOWN) {
673 iser_warn("iser conn %p state %d, expected state down.\n",
674 iser_conn, iser_conn->state);
675 iscsi_destroy_endpoint(iser_conn->ep);
676 iser_conn->state = ISER_CONN_DOWN;
679 * In case we never got to bind stage, we still need to
680 * release IB resources (which is safe to call more than once).
682 iser_free_ib_conn_res(iser_conn, true);
683 mutex_unlock(&iser_conn->state_mutex);
685 if (ib_conn->cma_id != NULL) {
686 rdma_destroy_id(ib_conn->cma_id);
687 ib_conn->cma_id = NULL;
694 * triggers start of the disconnect procedures and wait for them to be done
695 * Called with state mutex held
697 int iser_conn_terminate(struct iser_conn *iser_conn)
699 struct ib_conn *ib_conn = &iser_conn->ib_conn;
700 struct ib_send_wr *bad_wr;
703 /* terminate the iser conn only if the conn state is UP */
704 if (!iser_conn_state_comp_exch(iser_conn, ISER_CONN_UP,
705 ISER_CONN_TERMINATING))
708 iser_info("iser_conn %p state %d\n", iser_conn, iser_conn->state);
710 /* suspend queuing of new iscsi commands */
711 if (iser_conn->iscsi_conn)
712 iscsi_suspend_queue(iser_conn->iscsi_conn);
715 * In case we didn't already clean up the cma_id (peer initiated
716 * a disconnection), we need to Cause the CMA to change the QP
719 if (ib_conn->cma_id) {
720 err = rdma_disconnect(ib_conn->cma_id);
722 iser_err("Failed to disconnect, conn: 0x%p err %d\n",
725 /* post an indication that all flush errors were consumed */
726 err = ib_post_send(ib_conn->qp, &ib_conn->beacon, &bad_wr);
728 iser_err("conn %p failed to post beacon", ib_conn);
732 wait_for_completion(&ib_conn->flush_comp);
739 * Called with state mutex held
741 static void iser_connect_error(struct rdma_cm_id *cma_id)
743 struct iser_conn *iser_conn;
745 iser_conn = (struct iser_conn *)cma_id->context;
746 iser_conn->state = ISER_CONN_TERMINATING;
750 iser_calc_scsi_params(struct iser_conn *iser_conn,
751 unsigned int max_sectors)
753 struct iser_device *device = iser_conn->ib_conn.device;
754 unsigned short sg_tablesize, sup_sg_tablesize;
756 sg_tablesize = DIV_ROUND_UP(max_sectors * 512, SIZE_4K);
757 sup_sg_tablesize = min_t(unsigned, ISCSI_ISER_MAX_SG_TABLESIZE,
758 device->dev_attr.max_fast_reg_page_list_len);
760 if (sg_tablesize > sup_sg_tablesize) {
761 sg_tablesize = sup_sg_tablesize;
762 iser_conn->scsi_max_sectors = sg_tablesize * SIZE_4K / 512;
764 iser_conn->scsi_max_sectors = max_sectors;
767 iser_conn->scsi_sg_tablesize = sg_tablesize;
769 iser_dbg("iser_conn %p, sg_tablesize %u, max_sectors %u\n",
770 iser_conn, iser_conn->scsi_sg_tablesize,
771 iser_conn->scsi_max_sectors);
775 * Called with state mutex held
777 static void iser_addr_handler(struct rdma_cm_id *cma_id)
779 struct iser_device *device;
780 struct iser_conn *iser_conn;
781 struct ib_conn *ib_conn;
784 iser_conn = (struct iser_conn *)cma_id->context;
785 if (iser_conn->state != ISER_CONN_PENDING)
789 ib_conn = &iser_conn->ib_conn;
790 device = iser_device_find_by_ib_device(cma_id);
792 iser_err("device lookup/creation failed\n");
793 iser_connect_error(cma_id);
797 ib_conn->device = device;
799 /* connection T10-PI support */
800 if (iser_pi_enable) {
801 if (!(device->dev_attr.device_cap_flags &
802 IB_DEVICE_SIGNATURE_HANDOVER)) {
803 iser_warn("T10-PI requested but not supported on %s, "
804 "continue without T10-PI\n",
805 ib_conn->device->ib_device->name);
806 ib_conn->pi_support = false;
808 ib_conn->pi_support = true;
812 iser_calc_scsi_params(iser_conn, iser_max_sectors);
814 ret = rdma_resolve_route(cma_id, 1000);
816 iser_err("resolve route failed: %d\n", ret);
817 iser_connect_error(cma_id);
823 * Called with state mutex held
825 static void iser_route_handler(struct rdma_cm_id *cma_id)
827 struct rdma_conn_param conn_param;
829 struct iser_cm_hdr req_hdr;
830 struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
831 struct ib_conn *ib_conn = &iser_conn->ib_conn;
832 struct iser_device *device = ib_conn->device;
834 if (iser_conn->state != ISER_CONN_PENDING)
838 ret = iser_create_ib_conn_res(ib_conn);
842 memset(&conn_param, 0, sizeof conn_param);
843 conn_param.responder_resources = device->dev_attr.max_qp_rd_atom;
844 conn_param.initiator_depth = 1;
845 conn_param.retry_count = 7;
846 conn_param.rnr_retry_count = 6;
848 memset(&req_hdr, 0, sizeof(req_hdr));
849 req_hdr.flags = (ISER_ZBVA_NOT_SUPPORTED |
850 ISER_SEND_W_INV_NOT_SUPPORTED);
851 conn_param.private_data = (void *)&req_hdr;
852 conn_param.private_data_len = sizeof(struct iser_cm_hdr);
854 ret = rdma_connect(cma_id, &conn_param);
856 iser_err("failure connecting: %d\n", ret);
862 iser_connect_error(cma_id);
865 static void iser_connected_handler(struct rdma_cm_id *cma_id)
867 struct iser_conn *iser_conn;
868 struct ib_qp_attr attr;
869 struct ib_qp_init_attr init_attr;
871 iser_conn = (struct iser_conn *)cma_id->context;
872 if (iser_conn->state != ISER_CONN_PENDING)
876 (void)ib_query_qp(cma_id->qp, &attr, ~0, &init_attr);
877 iser_info("remote qpn:%x my qpn:%x\n", attr.dest_qp_num, cma_id->qp->qp_num);
879 iser_conn->state = ISER_CONN_UP;
880 complete(&iser_conn->up_completion);
883 static void iser_disconnected_handler(struct rdma_cm_id *cma_id)
885 struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
887 if (iser_conn_terminate(iser_conn)) {
888 if (iser_conn->iscsi_conn)
889 iscsi_conn_failure(iser_conn->iscsi_conn,
890 ISCSI_ERR_CONN_FAILED);
892 iser_err("iscsi_iser connection isn't bound\n");
896 static void iser_cleanup_handler(struct rdma_cm_id *cma_id,
899 struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
902 * We are not guaranteed that we visited disconnected_handler
903 * by now, call it here to be safe that we handle CM drep
906 iser_disconnected_handler(cma_id);
907 iser_free_ib_conn_res(iser_conn, destroy);
908 complete(&iser_conn->ib_completion);
911 static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
913 struct iser_conn *iser_conn;
916 iser_conn = (struct iser_conn *)cma_id->context;
917 iser_info("%s (%d): status %d conn %p id %p\n",
918 rdma_event_msg(event->event), event->event,
919 event->status, cma_id->context, cma_id);
921 mutex_lock(&iser_conn->state_mutex);
922 switch (event->event) {
923 case RDMA_CM_EVENT_ADDR_RESOLVED:
924 iser_addr_handler(cma_id);
926 case RDMA_CM_EVENT_ROUTE_RESOLVED:
927 iser_route_handler(cma_id);
929 case RDMA_CM_EVENT_ESTABLISHED:
930 iser_connected_handler(cma_id);
932 case RDMA_CM_EVENT_ADDR_ERROR:
933 case RDMA_CM_EVENT_ROUTE_ERROR:
934 case RDMA_CM_EVENT_CONNECT_ERROR:
935 case RDMA_CM_EVENT_UNREACHABLE:
936 case RDMA_CM_EVENT_REJECTED:
937 iser_connect_error(cma_id);
939 case RDMA_CM_EVENT_DISCONNECTED:
940 case RDMA_CM_EVENT_ADDR_CHANGE:
941 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
942 iser_cleanup_handler(cma_id, false);
944 case RDMA_CM_EVENT_DEVICE_REMOVAL:
946 * we *must* destroy the device as we cannot rely
947 * on iscsid to be around to initiate error handling.
948 * also if we are not in state DOWN implicitly destroy
951 iser_cleanup_handler(cma_id, true);
952 if (iser_conn->state != ISER_CONN_DOWN) {
953 iser_conn->ib_conn.cma_id = NULL;
958 iser_err("Unexpected RDMA CM event: %s (%d)\n",
959 rdma_event_msg(event->event), event->event);
962 mutex_unlock(&iser_conn->state_mutex);
967 void iser_conn_init(struct iser_conn *iser_conn)
969 iser_conn->state = ISER_CONN_INIT;
970 iser_conn->ib_conn.post_recv_buf_count = 0;
971 init_completion(&iser_conn->ib_conn.flush_comp);
972 init_completion(&iser_conn->stop_completion);
973 init_completion(&iser_conn->ib_completion);
974 init_completion(&iser_conn->up_completion);
975 INIT_LIST_HEAD(&iser_conn->conn_list);
976 mutex_init(&iser_conn->state_mutex);
980 * starts the process of connecting to the target
981 * sleeps until the connection is established or rejected
983 int iser_connect(struct iser_conn *iser_conn,
984 struct sockaddr *src_addr,
985 struct sockaddr *dst_addr,
988 struct ib_conn *ib_conn = &iser_conn->ib_conn;
991 mutex_lock(&iser_conn->state_mutex);
993 sprintf(iser_conn->name, "%pISp", dst_addr);
995 iser_info("connecting to: %s\n", iser_conn->name);
997 /* the device is known only --after-- address resolution */
998 ib_conn->device = NULL;
1000 iser_conn->state = ISER_CONN_PENDING;
1002 ib_conn->beacon.wr_id = ISER_BEACON_WRID;
1003 ib_conn->beacon.opcode = IB_WR_SEND;
1005 ib_conn->cma_id = rdma_create_id(&init_net, iser_cma_handler,
1007 RDMA_PS_TCP, IB_QPT_RC);
1008 if (IS_ERR(ib_conn->cma_id)) {
1009 err = PTR_ERR(ib_conn->cma_id);
1010 iser_err("rdma_create_id failed: %d\n", err);
1014 err = rdma_resolve_addr(ib_conn->cma_id, src_addr, dst_addr, 1000);
1016 iser_err("rdma_resolve_addr failed: %d\n", err);
1020 if (!non_blocking) {
1021 wait_for_completion_interruptible(&iser_conn->up_completion);
1023 if (iser_conn->state != ISER_CONN_UP) {
1025 goto connect_failure;
1028 mutex_unlock(&iser_conn->state_mutex);
1030 mutex_lock(&ig.connlist_mutex);
1031 list_add(&iser_conn->conn_list, &ig.connlist);
1032 mutex_unlock(&ig.connlist_mutex);
1036 ib_conn->cma_id = NULL;
1038 iser_conn->state = ISER_CONN_DOWN;
1040 mutex_unlock(&iser_conn->state_mutex);
1041 iser_conn_release(iser_conn);
1045 int iser_post_recvl(struct iser_conn *iser_conn)
1047 struct ib_recv_wr rx_wr, *rx_wr_failed;
1048 struct ib_conn *ib_conn = &iser_conn->ib_conn;
1049 struct iser_login_desc *desc = &iser_conn->login_desc;
1052 desc->sge.addr = desc->rsp_dma;
1053 desc->sge.length = ISER_RX_LOGIN_SIZE;
1054 desc->sge.lkey = ib_conn->device->pd->local_dma_lkey;
1056 rx_wr.wr_id = (uintptr_t)desc;
1057 rx_wr.sg_list = &desc->sge;
1061 ib_conn->post_recv_buf_count++;
1062 ib_ret = ib_post_recv(ib_conn->qp, &rx_wr, &rx_wr_failed);
1064 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1065 ib_conn->post_recv_buf_count--;
1070 int iser_post_recvm(struct iser_conn *iser_conn, int count)
1072 struct ib_recv_wr *rx_wr, *rx_wr_failed;
1074 struct ib_conn *ib_conn = &iser_conn->ib_conn;
1075 unsigned int my_rx_head = iser_conn->rx_desc_head;
1076 struct iser_rx_desc *rx_desc;
1078 for (rx_wr = ib_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
1079 rx_desc = &iser_conn->rx_descs[my_rx_head];
1080 rx_wr->wr_id = (uintptr_t)rx_desc;
1081 rx_wr->sg_list = &rx_desc->rx_sg;
1083 rx_wr->next = rx_wr + 1;
1084 my_rx_head = (my_rx_head + 1) & iser_conn->qp_max_recv_dtos_mask;
1088 rx_wr->next = NULL; /* mark end of work requests list */
1090 ib_conn->post_recv_buf_count += count;
1091 ib_ret = ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &rx_wr_failed);
1093 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1094 ib_conn->post_recv_buf_count -= count;
1096 iser_conn->rx_desc_head = my_rx_head;
1102 * iser_start_send - Initiate a Send DTO operation
1104 * returns 0 on success, -1 on failure
1106 int iser_post_send(struct ib_conn *ib_conn, struct iser_tx_desc *tx_desc,
1109 struct ib_send_wr *bad_wr, *wr = iser_tx_next_wr(tx_desc);
1112 ib_dma_sync_single_for_device(ib_conn->device->ib_device,
1113 tx_desc->dma_addr, ISER_HEADERS_LEN,
1117 wr->wr_id = (uintptr_t)tx_desc;
1118 wr->sg_list = tx_desc->tx_sg;
1119 wr->num_sge = tx_desc->num_sge;
1120 wr->opcode = IB_WR_SEND;
1121 wr->send_flags = signal ? IB_SEND_SIGNALED : 0;
1123 ib_ret = ib_post_send(ib_conn->qp, &tx_desc->wrs[0].send, &bad_wr);
1125 iser_err("ib_post_send failed, ret:%d opcode:%d\n",
1126 ib_ret, bad_wr->opcode);
1132 * is_iser_tx_desc - Indicate if the completion wr_id
1133 * is a TX descriptor or not.
1134 * @iser_conn: iser connection
1135 * @wr_id: completion WR identifier
1137 * Since we cannot rely on wc opcode in FLUSH errors
1138 * we must work around it by checking if the wr_id address
1139 * falls in the iser connection rx_descs buffer. If so
1140 * it is an RX descriptor, otherwize it is a TX.
1143 is_iser_tx_desc(struct iser_conn *iser_conn, void *wr_id)
1145 void *start = iser_conn->rx_descs;
1146 int len = iser_conn->num_rx_descs * sizeof(*iser_conn->rx_descs);
1148 if (wr_id >= start && wr_id < start + len)
1155 * iser_handle_comp_error() - Handle error completion
1156 * @ib_conn: connection RDMA resources
1157 * @wc: work completion
1159 * Notes: We may handle a FLUSH error completion and in this case
1160 * we only cleanup in case TX type was DATAOUT. For non-FLUSH
1161 * error completion we should also notify iscsi layer that
1162 * connection is failed (in case we passed bind stage).
1165 iser_handle_comp_error(struct ib_conn *ib_conn,
1168 struct iser_conn *iser_conn = to_iser_conn(ib_conn);
1169 void *wr_id = (void *)(uintptr_t)wc->wr_id;
1171 if (wc->status != IB_WC_WR_FLUSH_ERR)
1172 if (iser_conn->iscsi_conn)
1173 iscsi_conn_failure(iser_conn->iscsi_conn,
1174 ISCSI_ERR_CONN_FAILED);
1176 if (wc->wr_id == ISER_FASTREG_LI_WRID)
1179 if (is_iser_tx_desc(iser_conn, wr_id)) {
1180 struct iser_tx_desc *desc = wr_id;
1182 if (desc->type == ISCSI_TX_DATAOUT)
1183 kmem_cache_free(ig.desc_cache, desc);
1185 ib_conn->post_recv_buf_count--;
1190 * iser_handle_wc - handle a single work completion
1191 * @wc: work completion
1193 * Soft-IRQ context, work completion can be either
1194 * SEND or RECV, and can turn out successful or
1195 * with error (or flush error).
1197 static void iser_handle_wc(struct ib_wc *wc)
1199 struct ib_conn *ib_conn;
1200 struct iser_tx_desc *tx_desc;
1201 struct iser_rx_desc *rx_desc;
1203 ib_conn = wc->qp->qp_context;
1204 if (likely(wc->status == IB_WC_SUCCESS)) {
1205 if (wc->opcode == IB_WC_RECV) {
1206 rx_desc = (struct iser_rx_desc *)(uintptr_t)wc->wr_id;
1207 iser_rcv_completion(rx_desc, wc->byte_len,
1210 if (wc->opcode == IB_WC_SEND) {
1211 tx_desc = (struct iser_tx_desc *)(uintptr_t)wc->wr_id;
1212 iser_snd_completion(tx_desc, ib_conn);
1214 iser_err("Unknown wc opcode %d\n", wc->opcode);
1217 if (wc->status != IB_WC_WR_FLUSH_ERR)
1218 iser_err("%s (%d): wr id %llx vend_err %x\n",
1219 ib_wc_status_msg(wc->status), wc->status,
1220 wc->wr_id, wc->vendor_err);
1222 iser_dbg("%s (%d): wr id %llx\n",
1223 ib_wc_status_msg(wc->status), wc->status,
1226 if (wc->wr_id == ISER_BEACON_WRID)
1227 /* all flush errors were consumed */
1228 complete(&ib_conn->flush_comp);
1230 iser_handle_comp_error(ib_conn, wc);
1235 * iser_cq_tasklet_fn - iSER completion polling loop
1236 * @data: iSER completion context
1238 * Soft-IRQ context, polling connection CQ until
1239 * either CQ was empty or we exausted polling budget
1241 static void iser_cq_tasklet_fn(unsigned long data)
1243 struct iser_comp *comp = (struct iser_comp *)data;
1244 struct ib_cq *cq = comp->cq;
1245 struct ib_wc *const wcs = comp->wcs;
1246 int i, n, completed = 0;
1248 while ((n = ib_poll_cq(cq, ARRAY_SIZE(comp->wcs), wcs)) > 0) {
1249 for (i = 0; i < n; i++)
1250 iser_handle_wc(&wcs[i]);
1253 if (completed >= iser_cq_poll_limit)
1258 * It is assumed here that arming CQ only once its empty
1259 * would not cause interrupts to be missed.
1261 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1263 iser_dbg("got %d completions\n", completed);
1266 static void iser_cq_callback(struct ib_cq *cq, void *cq_context)
1268 struct iser_comp *comp = cq_context;
1270 tasklet_schedule(&comp->tasklet);
1273 u8 iser_check_task_pi_status(struct iscsi_iser_task *iser_task,
1274 enum iser_data_dir cmd_dir, sector_t *sector)
1276 struct iser_mem_reg *reg = &iser_task->rdma_reg[cmd_dir];
1277 struct iser_fr_desc *desc = reg->mem_h;
1278 unsigned long sector_size = iser_task->sc->device->sector_size;
1279 struct ib_mr_status mr_status;
1282 if (desc && desc->pi_ctx->sig_protected) {
1283 desc->pi_ctx->sig_protected = 0;
1284 ret = ib_check_mr_status(desc->pi_ctx->sig_mr,
1285 IB_MR_CHECK_SIG_STATUS, &mr_status);
1287 pr_err("ib_check_mr_status failed, ret %d\n", ret);
1291 if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1292 sector_t sector_off = mr_status.sig_err.sig_err_offset;
1294 do_div(sector_off, sector_size + 8);
1295 *sector = scsi_get_lba(iser_task->sc) + sector_off;
1297 pr_err("PI error found type %d at sector %llx "
1298 "expected %x vs actual %x\n",
1299 mr_status.sig_err.err_type,
1300 (unsigned long long)*sector,
1301 mr_status.sig_err.expected,
1302 mr_status.sig_err.actual);
1304 switch (mr_status.sig_err.err_type) {
1305 case IB_SIG_BAD_GUARD:
1307 case IB_SIG_BAD_REFTAG:
1309 case IB_SIG_BAD_APPTAG:
1317 /* Not alot we can do here, return ambiguous guard error */