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);
109 iser_info("using %d CQs, device %s supports %d vectors\n",
110 device->comps_used, device->ib_device->name,
111 device->ib_device->num_comp_vectors);
113 device->pd = ib_alloc_pd(device->ib_device);
114 if (IS_ERR(device->pd))
117 for (i = 0; i < device->comps_used; i++) {
118 struct iser_comp *comp = &device->comps[i];
120 comp->device = device;
121 comp->cq = ib_create_cq(device->ib_device,
123 iser_cq_event_callback,
126 if (IS_ERR(comp->cq)) {
131 if (ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP))
134 tasklet_init(&comp->tasklet, iser_cq_tasklet_fn,
135 (unsigned long)comp);
138 device->mr = ib_get_dma_mr(device->pd, IB_ACCESS_LOCAL_WRITE |
139 IB_ACCESS_REMOTE_WRITE |
140 IB_ACCESS_REMOTE_READ);
141 if (IS_ERR(device->mr))
144 INIT_IB_EVENT_HANDLER(&device->event_handler, device->ib_device,
146 if (ib_register_event_handler(&device->event_handler))
152 ib_dereg_mr(device->mr);
154 for (i = 0; i < device->comps_used; i++)
155 tasklet_kill(&device->comps[i].tasklet);
157 for (i = 0; i < device->comps_used; i++) {
158 struct iser_comp *comp = &device->comps[i];
161 ib_destroy_cq(comp->cq);
163 ib_dealloc_pd(device->pd);
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);
195 * iser_create_fmr_pool - Creates FMR pool and page_vector
197 * returns 0 on success, or errno code on failure
199 int iser_create_fmr_pool(struct ib_conn *ib_conn, unsigned cmds_max)
201 struct iser_device *device = ib_conn->device;
202 struct ib_fmr_pool_param params;
205 ib_conn->fmr.page_vec = kmalloc(sizeof(*ib_conn->fmr.page_vec) +
206 (sizeof(u64)*(ISCSI_ISER_SG_TABLESIZE + 1)),
208 if (!ib_conn->fmr.page_vec)
211 ib_conn->fmr.page_vec->pages = (u64 *)(ib_conn->fmr.page_vec + 1);
213 params.page_shift = SHIFT_4K;
214 /* when the first/last SG element are not start/end *
215 * page aligned, the map whould be of N+1 pages */
216 params.max_pages_per_fmr = ISCSI_ISER_SG_TABLESIZE + 1;
217 /* make the pool size twice the max number of SCSI commands *
218 * the ML is expected to queue, watermark for unmap at 50% */
219 params.pool_size = cmds_max * 2;
220 params.dirty_watermark = cmds_max;
222 params.flush_function = NULL;
223 params.access = (IB_ACCESS_LOCAL_WRITE |
224 IB_ACCESS_REMOTE_WRITE |
225 IB_ACCESS_REMOTE_READ);
227 ib_conn->fmr.pool = ib_create_fmr_pool(device->pd, ¶ms);
228 if (!IS_ERR(ib_conn->fmr.pool))
231 /* no FMR => no need for page_vec */
232 kfree(ib_conn->fmr.page_vec);
233 ib_conn->fmr.page_vec = NULL;
235 ret = PTR_ERR(ib_conn->fmr.pool);
236 ib_conn->fmr.pool = NULL;
237 if (ret != -ENOSYS) {
238 iser_err("FMR allocation failed, err %d\n", ret);
241 iser_warn("FMRs are not supported, using unaligned mode\n");
247 * iser_free_fmr_pool - releases the FMR pool and page vec
249 void iser_free_fmr_pool(struct ib_conn *ib_conn)
251 iser_info("freeing conn %p fmr pool %p\n",
252 ib_conn, ib_conn->fmr.pool);
254 if (ib_conn->fmr.pool != NULL)
255 ib_destroy_fmr_pool(ib_conn->fmr.pool);
257 ib_conn->fmr.pool = NULL;
259 kfree(ib_conn->fmr.page_vec);
260 ib_conn->fmr.page_vec = NULL;
264 iser_create_fastreg_desc(struct ib_device *ib_device, struct ib_pd *pd,
265 bool pi_enable, struct fast_reg_descriptor *desc)
269 desc->data_frpl = ib_alloc_fast_reg_page_list(ib_device,
270 ISCSI_ISER_SG_TABLESIZE + 1);
271 if (IS_ERR(desc->data_frpl)) {
272 ret = PTR_ERR(desc->data_frpl);
273 iser_err("Failed to allocate ib_fast_reg_page_list err=%d\n",
275 return PTR_ERR(desc->data_frpl);
278 desc->data_mr = ib_alloc_fast_reg_mr(pd, ISCSI_ISER_SG_TABLESIZE + 1);
279 if (IS_ERR(desc->data_mr)) {
280 ret = PTR_ERR(desc->data_mr);
281 iser_err("Failed to allocate ib_fast_reg_mr err=%d\n", ret);
282 goto fast_reg_mr_failure;
284 desc->reg_indicators |= ISER_DATA_KEY_VALID;
287 struct ib_mr_init_attr mr_init_attr = {0};
288 struct iser_pi_context *pi_ctx = NULL;
290 desc->pi_ctx = kzalloc(sizeof(*desc->pi_ctx), GFP_KERNEL);
292 iser_err("Failed to allocate pi context\n");
294 goto pi_ctx_alloc_failure;
296 pi_ctx = desc->pi_ctx;
298 pi_ctx->prot_frpl = ib_alloc_fast_reg_page_list(ib_device,
299 ISCSI_ISER_SG_TABLESIZE);
300 if (IS_ERR(pi_ctx->prot_frpl)) {
301 ret = PTR_ERR(pi_ctx->prot_frpl);
302 iser_err("Failed to allocate prot frpl ret=%d\n",
304 goto prot_frpl_failure;
307 pi_ctx->prot_mr = ib_alloc_fast_reg_mr(pd,
308 ISCSI_ISER_SG_TABLESIZE + 1);
309 if (IS_ERR(pi_ctx->prot_mr)) {
310 ret = PTR_ERR(pi_ctx->prot_mr);
311 iser_err("Failed to allocate prot frmr ret=%d\n",
313 goto prot_mr_failure;
315 desc->reg_indicators |= ISER_PROT_KEY_VALID;
317 mr_init_attr.max_reg_descriptors = 2;
318 mr_init_attr.flags |= IB_MR_SIGNATURE_EN;
319 pi_ctx->sig_mr = ib_create_mr(pd, &mr_init_attr);
320 if (IS_ERR(pi_ctx->sig_mr)) {
321 ret = PTR_ERR(pi_ctx->sig_mr);
322 iser_err("Failed to allocate signature enabled mr err=%d\n",
326 desc->reg_indicators |= ISER_SIG_KEY_VALID;
328 desc->reg_indicators &= ~ISER_FASTREG_PROTECTED;
330 iser_dbg("Create fr_desc %p page_list %p\n",
331 desc, desc->data_frpl->page_list);
335 ib_dereg_mr(desc->pi_ctx->prot_mr);
337 ib_free_fast_reg_page_list(desc->pi_ctx->prot_frpl);
340 pi_ctx_alloc_failure:
341 ib_dereg_mr(desc->data_mr);
343 ib_free_fast_reg_page_list(desc->data_frpl);
349 * iser_create_fastreg_pool - Creates pool of fast_reg descriptors
350 * for fast registration work requests.
351 * returns 0 on success, or errno code on failure
353 int iser_create_fastreg_pool(struct ib_conn *ib_conn, unsigned cmds_max)
355 struct iser_device *device = ib_conn->device;
356 struct fast_reg_descriptor *desc;
359 INIT_LIST_HEAD(&ib_conn->fastreg.pool);
360 ib_conn->fastreg.pool_size = 0;
361 for (i = 0; i < cmds_max; i++) {
362 desc = kzalloc(sizeof(*desc), GFP_KERNEL);
364 iser_err("Failed to allocate a new fast_reg descriptor\n");
369 ret = iser_create_fastreg_desc(device->ib_device, device->pd,
370 ib_conn->pi_support, desc);
372 iser_err("Failed to create fastreg descriptor err=%d\n",
378 list_add_tail(&desc->list, &ib_conn->fastreg.pool);
379 ib_conn->fastreg.pool_size++;
385 iser_free_fastreg_pool(ib_conn);
390 * iser_free_fastreg_pool - releases the pool of fast_reg descriptors
392 void iser_free_fastreg_pool(struct ib_conn *ib_conn)
394 struct fast_reg_descriptor *desc, *tmp;
397 if (list_empty(&ib_conn->fastreg.pool))
400 iser_info("freeing conn %p fr pool\n", ib_conn);
402 list_for_each_entry_safe(desc, tmp, &ib_conn->fastreg.pool, list) {
403 list_del(&desc->list);
404 ib_free_fast_reg_page_list(desc->data_frpl);
405 ib_dereg_mr(desc->data_mr);
407 ib_free_fast_reg_page_list(desc->pi_ctx->prot_frpl);
408 ib_dereg_mr(desc->pi_ctx->prot_mr);
409 ib_destroy_mr(desc->pi_ctx->sig_mr);
416 if (i < ib_conn->fastreg.pool_size)
417 iser_warn("pool still has %d regions registered\n",
418 ib_conn->fastreg.pool_size - i);
422 * iser_create_ib_conn_res - Queue-Pair (QP)
424 * returns 0 on success, -1 on failure
426 static int iser_create_ib_conn_res(struct ib_conn *ib_conn)
428 struct iser_device *device;
429 struct ib_qp_init_attr init_attr;
431 int index, min_index = 0;
433 BUG_ON(ib_conn->device == NULL);
435 device = ib_conn->device;
437 memset(&init_attr, 0, sizeof init_attr);
439 mutex_lock(&ig.connlist_mutex);
440 /* select the CQ with the minimal number of usages */
441 for (index = 0; index < device->comps_used; index++) {
442 if (device->comps[index].active_qps <
443 device->comps[min_index].active_qps)
446 ib_conn->comp = &device->comps[min_index];
447 ib_conn->comp->active_qps++;
448 mutex_unlock(&ig.connlist_mutex);
449 iser_info("cq index %d used for ib_conn %p\n", min_index, ib_conn);
451 init_attr.event_handler = iser_qp_event_callback;
452 init_attr.qp_context = (void *)ib_conn;
453 init_attr.send_cq = ib_conn->comp->cq;
454 init_attr.recv_cq = ib_conn->comp->cq;
455 init_attr.cap.max_recv_wr = ISER_QP_MAX_RECV_DTOS;
456 init_attr.cap.max_send_sge = 2;
457 init_attr.cap.max_recv_sge = 1;
458 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
459 init_attr.qp_type = IB_QPT_RC;
460 if (ib_conn->pi_support) {
461 init_attr.cap.max_send_wr = ISER_QP_SIG_MAX_REQ_DTOS + 1;
462 init_attr.create_flags |= IB_QP_CREATE_SIGNATURE_EN;
464 init_attr.cap.max_send_wr = ISER_QP_MAX_REQ_DTOS + 1;
467 ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
471 ib_conn->qp = ib_conn->cma_id->qp;
472 iser_info("setting conn %p cma_id %p qp %p\n",
473 ib_conn, ib_conn->cma_id,
474 ib_conn->cma_id->qp);
478 iser_err("unable to alloc mem or create resource, err %d\n", ret);
483 * based on the resolved device node GUID see if there already allocated
484 * device for this device. If there's no such, create one.
487 struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
489 struct iser_device *device;
491 mutex_lock(&ig.device_list_mutex);
493 list_for_each_entry(device, &ig.device_list, ig_list)
494 /* find if there's a match using the node GUID */
495 if (device->ib_device->node_guid == cma_id->device->node_guid)
498 device = kzalloc(sizeof *device, GFP_KERNEL);
502 /* assign this device to the device */
503 device->ib_device = cma_id->device;
504 /* init the device and link it into ig device list */
505 if (iser_create_device_ib_res(device)) {
510 list_add(&device->ig_list, &ig.device_list);
515 mutex_unlock(&ig.device_list_mutex);
519 /* if there's no demand for this device, release it */
520 static void iser_device_try_release(struct iser_device *device)
522 mutex_lock(&ig.device_list_mutex);
524 iser_info("device %p refcount %d\n", device, device->refcount);
525 if (!device->refcount) {
526 iser_free_device_ib_res(device);
527 list_del(&device->ig_list);
530 mutex_unlock(&ig.device_list_mutex);
534 * Called with state mutex held
536 static int iser_conn_state_comp_exch(struct iser_conn *iser_conn,
537 enum iser_conn_state comp,
538 enum iser_conn_state exch)
542 ret = (iser_conn->state == comp);
544 iser_conn->state = exch;
549 void iser_release_work(struct work_struct *work)
551 struct iser_conn *iser_conn;
553 iser_conn = container_of(work, struct iser_conn, release_work);
555 /* Wait for conn_stop to complete */
556 wait_for_completion(&iser_conn->stop_completion);
557 /* Wait for IB resouces cleanup to complete */
558 wait_for_completion(&iser_conn->ib_completion);
560 mutex_lock(&iser_conn->state_mutex);
561 iser_conn->state = ISER_CONN_DOWN;
562 mutex_unlock(&iser_conn->state_mutex);
564 iser_conn_release(iser_conn);
568 * iser_free_ib_conn_res - release IB related resources
569 * @iser_conn: iser connection struct
570 * @destroy_device: indicator if we need to try to release
571 * the iser device (only iscsi shutdown and DEVICE_REMOVAL
574 * This routine is called with the iser state mutex held
575 * so the cm_id removal is out of here. It is Safe to
576 * be invoked multiple times.
578 static void iser_free_ib_conn_res(struct iser_conn *iser_conn,
581 struct ib_conn *ib_conn = &iser_conn->ib_conn;
582 struct iser_device *device = ib_conn->device;
584 iser_info("freeing conn %p cma_id %p qp %p\n",
585 iser_conn, ib_conn->cma_id, ib_conn->qp);
587 iser_free_rx_descriptors(iser_conn);
589 if (ib_conn->qp != NULL) {
590 ib_conn->comp->active_qps--;
591 rdma_destroy_qp(ib_conn->cma_id);
595 if (destroy_device && device != NULL) {
596 iser_device_try_release(device);
597 ib_conn->device = NULL;
602 * Frees all conn objects and deallocs conn descriptor
604 void iser_conn_release(struct iser_conn *iser_conn)
606 struct ib_conn *ib_conn = &iser_conn->ib_conn;
608 mutex_lock(&ig.connlist_mutex);
609 list_del(&iser_conn->conn_list);
610 mutex_unlock(&ig.connlist_mutex);
612 mutex_lock(&iser_conn->state_mutex);
613 if (iser_conn->state != ISER_CONN_DOWN)
614 iser_warn("iser conn %p state %d, expected state down.\n",
615 iser_conn, iser_conn->state);
617 * In case we never got to bind stage, we still need to
618 * release IB resources (which is safe to call more than once).
620 iser_free_ib_conn_res(iser_conn, true);
621 mutex_unlock(&iser_conn->state_mutex);
623 if (ib_conn->cma_id != NULL) {
624 rdma_destroy_id(ib_conn->cma_id);
625 ib_conn->cma_id = NULL;
632 * triggers start of the disconnect procedures and wait for them to be done
633 * Called with state mutex held
635 int iser_conn_terminate(struct iser_conn *iser_conn)
637 struct ib_conn *ib_conn = &iser_conn->ib_conn;
638 struct ib_send_wr *bad_wr;
641 /* terminate the iser conn only if the conn state is UP */
642 if (!iser_conn_state_comp_exch(iser_conn, ISER_CONN_UP,
643 ISER_CONN_TERMINATING))
646 iser_info("iser_conn %p state %d\n", iser_conn, iser_conn->state);
648 /* suspend queuing of new iscsi commands */
649 if (iser_conn->iscsi_conn)
650 iscsi_suspend_queue(iser_conn->iscsi_conn);
653 * In case we didn't already clean up the cma_id (peer initiated
654 * a disconnection), we need to Cause the CMA to change the QP
657 if (ib_conn->cma_id) {
658 err = rdma_disconnect(ib_conn->cma_id);
660 iser_err("Failed to disconnect, conn: 0x%p err %d\n",
663 /* post an indication that all flush errors were consumed */
664 err = ib_post_send(ib_conn->qp, &ib_conn->beacon, &bad_wr);
666 iser_err("conn %p failed to post beacon", ib_conn);
668 wait_for_completion(&ib_conn->flush_comp);
675 * Called with state mutex held
677 static void iser_connect_error(struct rdma_cm_id *cma_id)
679 struct iser_conn *iser_conn;
681 iser_conn = (struct iser_conn *)cma_id->context;
682 iser_conn->state = ISER_CONN_DOWN;
686 * Called with state mutex held
688 static void iser_addr_handler(struct rdma_cm_id *cma_id)
690 struct iser_device *device;
691 struct iser_conn *iser_conn;
692 struct ib_conn *ib_conn;
695 iser_conn = (struct iser_conn *)cma_id->context;
696 if (iser_conn->state != ISER_CONN_PENDING)
700 ib_conn = &iser_conn->ib_conn;
701 device = iser_device_find_by_ib_device(cma_id);
703 iser_err("device lookup/creation failed\n");
704 iser_connect_error(cma_id);
708 ib_conn->device = device;
710 /* connection T10-PI support */
711 if (iser_pi_enable) {
712 if (!(device->dev_attr.device_cap_flags &
713 IB_DEVICE_SIGNATURE_HANDOVER)) {
714 iser_warn("T10-PI requested but not supported on %s, "
715 "continue without T10-PI\n",
716 ib_conn->device->ib_device->name);
717 ib_conn->pi_support = false;
719 ib_conn->pi_support = true;
723 ret = rdma_resolve_route(cma_id, 1000);
725 iser_err("resolve route failed: %d\n", ret);
726 iser_connect_error(cma_id);
732 * Called with state mutex held
734 static void iser_route_handler(struct rdma_cm_id *cma_id)
736 struct rdma_conn_param conn_param;
738 struct iser_cm_hdr req_hdr;
739 struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
740 struct ib_conn *ib_conn = &iser_conn->ib_conn;
741 struct iser_device *device = ib_conn->device;
743 if (iser_conn->state != ISER_CONN_PENDING)
747 ret = iser_create_ib_conn_res(ib_conn);
751 memset(&conn_param, 0, sizeof conn_param);
752 conn_param.responder_resources = device->dev_attr.max_qp_rd_atom;
753 conn_param.initiator_depth = 1;
754 conn_param.retry_count = 7;
755 conn_param.rnr_retry_count = 6;
757 memset(&req_hdr, 0, sizeof(req_hdr));
758 req_hdr.flags = (ISER_ZBVA_NOT_SUPPORTED |
759 ISER_SEND_W_INV_NOT_SUPPORTED);
760 conn_param.private_data = (void *)&req_hdr;
761 conn_param.private_data_len = sizeof(struct iser_cm_hdr);
763 ret = rdma_connect(cma_id, &conn_param);
765 iser_err("failure connecting: %d\n", ret);
771 iser_connect_error(cma_id);
774 static void iser_connected_handler(struct rdma_cm_id *cma_id)
776 struct iser_conn *iser_conn;
777 struct ib_qp_attr attr;
778 struct ib_qp_init_attr init_attr;
780 iser_conn = (struct iser_conn *)cma_id->context;
781 if (iser_conn->state != ISER_CONN_PENDING)
785 (void)ib_query_qp(cma_id->qp, &attr, ~0, &init_attr);
786 iser_info("remote qpn:%x my qpn:%x\n", attr.dest_qp_num, cma_id->qp->qp_num);
788 iser_conn->state = ISER_CONN_UP;
789 complete(&iser_conn->up_completion);
792 static void iser_disconnected_handler(struct rdma_cm_id *cma_id)
794 struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
796 if (iser_conn_terminate(iser_conn)) {
797 if (iser_conn->iscsi_conn)
798 iscsi_conn_failure(iser_conn->iscsi_conn,
799 ISCSI_ERR_CONN_FAILED);
801 iser_err("iscsi_iser connection isn't bound\n");
805 static void iser_cleanup_handler(struct rdma_cm_id *cma_id,
808 struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
811 * We are not guaranteed that we visited disconnected_handler
812 * by now, call it here to be safe that we handle CM drep
815 iser_disconnected_handler(cma_id);
816 iser_free_ib_conn_res(iser_conn, destroy_device);
817 complete(&iser_conn->ib_completion);
820 static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
822 struct iser_conn *iser_conn;
825 iser_conn = (struct iser_conn *)cma_id->context;
826 iser_info("event %d status %d conn %p id %p\n",
827 event->event, event->status, cma_id->context, cma_id);
829 mutex_lock(&iser_conn->state_mutex);
830 switch (event->event) {
831 case RDMA_CM_EVENT_ADDR_RESOLVED:
832 iser_addr_handler(cma_id);
834 case RDMA_CM_EVENT_ROUTE_RESOLVED:
835 iser_route_handler(cma_id);
837 case RDMA_CM_EVENT_ESTABLISHED:
838 iser_connected_handler(cma_id);
840 case RDMA_CM_EVENT_ADDR_ERROR:
841 case RDMA_CM_EVENT_ROUTE_ERROR:
842 case RDMA_CM_EVENT_CONNECT_ERROR:
843 case RDMA_CM_EVENT_UNREACHABLE:
844 case RDMA_CM_EVENT_REJECTED:
845 iser_connect_error(cma_id);
847 case RDMA_CM_EVENT_DISCONNECTED:
848 case RDMA_CM_EVENT_ADDR_CHANGE:
849 iser_disconnected_handler(cma_id);
851 case RDMA_CM_EVENT_DEVICE_REMOVAL:
853 * we *must* destroy the device as we cannot rely
854 * on iscsid to be around to initiate error handling.
855 * also implicitly destroy the cma_id.
857 iser_cleanup_handler(cma_id, true);
858 iser_conn->ib_conn.cma_id = NULL;
861 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
862 iser_cleanup_handler(cma_id, false);
865 iser_err("Unexpected RDMA CM event (%d)\n", event->event);
868 mutex_unlock(&iser_conn->state_mutex);
873 void iser_conn_init(struct iser_conn *iser_conn)
875 iser_conn->state = ISER_CONN_INIT;
876 iser_conn->ib_conn.post_recv_buf_count = 0;
877 init_completion(&iser_conn->ib_conn.flush_comp);
878 init_completion(&iser_conn->stop_completion);
879 init_completion(&iser_conn->ib_completion);
880 init_completion(&iser_conn->up_completion);
881 INIT_LIST_HEAD(&iser_conn->conn_list);
882 spin_lock_init(&iser_conn->ib_conn.lock);
883 mutex_init(&iser_conn->state_mutex);
887 * starts the process of connecting to the target
888 * sleeps until the connection is established or rejected
890 int iser_connect(struct iser_conn *iser_conn,
891 struct sockaddr *src_addr,
892 struct sockaddr *dst_addr,
895 struct ib_conn *ib_conn = &iser_conn->ib_conn;
898 mutex_lock(&iser_conn->state_mutex);
900 sprintf(iser_conn->name, "%pISp", dst_addr);
902 iser_info("connecting to: %s\n", iser_conn->name);
904 /* the device is known only --after-- address resolution */
905 ib_conn->device = NULL;
907 iser_conn->state = ISER_CONN_PENDING;
909 ib_conn->beacon.wr_id = ISER_BEACON_WRID;
910 ib_conn->beacon.opcode = IB_WR_SEND;
912 ib_conn->cma_id = rdma_create_id(iser_cma_handler,
914 RDMA_PS_TCP, IB_QPT_RC);
915 if (IS_ERR(ib_conn->cma_id)) {
916 err = PTR_ERR(ib_conn->cma_id);
917 iser_err("rdma_create_id failed: %d\n", err);
921 err = rdma_resolve_addr(ib_conn->cma_id, src_addr, dst_addr, 1000);
923 iser_err("rdma_resolve_addr failed: %d\n", err);
928 wait_for_completion_interruptible(&iser_conn->up_completion);
930 if (iser_conn->state != ISER_CONN_UP) {
932 goto connect_failure;
935 mutex_unlock(&iser_conn->state_mutex);
937 mutex_lock(&ig.connlist_mutex);
938 list_add(&iser_conn->conn_list, &ig.connlist);
939 mutex_unlock(&ig.connlist_mutex);
943 ib_conn->cma_id = NULL;
945 iser_conn->state = ISER_CONN_DOWN;
947 mutex_unlock(&iser_conn->state_mutex);
948 iser_conn_release(iser_conn);
953 * iser_reg_page_vec - Register physical memory
955 * returns: 0 on success, errno code on failure
957 int iser_reg_page_vec(struct ib_conn *ib_conn,
958 struct iser_page_vec *page_vec,
959 struct iser_mem_reg *mem_reg)
961 struct ib_pool_fmr *mem;
966 page_list = page_vec->pages;
967 io_addr = page_list[0];
969 mem = ib_fmr_pool_map_phys(ib_conn->fmr.pool,
975 status = (int)PTR_ERR(mem);
976 iser_err("ib_fmr_pool_map_phys failed: %d\n", status);
980 mem_reg->lkey = mem->fmr->lkey;
981 mem_reg->rkey = mem->fmr->rkey;
982 mem_reg->len = page_vec->length * SIZE_4K;
983 mem_reg->va = io_addr;
985 mem_reg->mem_h = (void *)mem;
987 mem_reg->va += page_vec->offset;
988 mem_reg->len = page_vec->data_size;
990 iser_dbg("PHYSICAL Mem.register, [PHYS p_array: 0x%p, sz: %d, "
991 "entry[0]: (0x%08lx,%ld)] -> "
992 "[lkey: 0x%08X mem_h: 0x%p va: 0x%08lX sz: %ld]\n",
993 page_vec, page_vec->length,
994 (unsigned long)page_vec->pages[0],
995 (unsigned long)page_vec->data_size,
996 (unsigned int)mem_reg->lkey, mem_reg->mem_h,
997 (unsigned long)mem_reg->va, (unsigned long)mem_reg->len);
1002 * Unregister (previosuly registered using FMR) memory.
1003 * If memory is non-FMR does nothing.
1005 void iser_unreg_mem_fmr(struct iscsi_iser_task *iser_task,
1006 enum iser_data_dir cmd_dir)
1008 struct iser_mem_reg *reg = &iser_task->rdma_regd[cmd_dir].reg;
1014 iser_dbg("PHYSICAL Mem.Unregister mem_h %p\n",reg->mem_h);
1016 ret = ib_fmr_pool_unmap((struct ib_pool_fmr *)reg->mem_h);
1018 iser_err("ib_fmr_pool_unmap failed %d\n", ret);
1023 void iser_unreg_mem_fastreg(struct iscsi_iser_task *iser_task,
1024 enum iser_data_dir cmd_dir)
1026 struct iser_mem_reg *reg = &iser_task->rdma_regd[cmd_dir].reg;
1027 struct iser_conn *iser_conn = iser_task->iser_conn;
1028 struct ib_conn *ib_conn = &iser_conn->ib_conn;
1029 struct fast_reg_descriptor *desc = reg->mem_h;
1036 spin_lock_bh(&ib_conn->lock);
1037 list_add_tail(&desc->list, &ib_conn->fastreg.pool);
1038 spin_unlock_bh(&ib_conn->lock);
1041 int iser_post_recvl(struct iser_conn *iser_conn)
1043 struct ib_recv_wr rx_wr, *rx_wr_failed;
1044 struct ib_conn *ib_conn = &iser_conn->ib_conn;
1048 sge.addr = iser_conn->login_resp_dma;
1049 sge.length = ISER_RX_LOGIN_SIZE;
1050 sge.lkey = ib_conn->device->mr->lkey;
1052 rx_wr.wr_id = (unsigned long)iser_conn->login_resp_buf;
1053 rx_wr.sg_list = &sge;
1057 ib_conn->post_recv_buf_count++;
1058 ib_ret = ib_post_recv(ib_conn->qp, &rx_wr, &rx_wr_failed);
1060 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1061 ib_conn->post_recv_buf_count--;
1066 int iser_post_recvm(struct iser_conn *iser_conn, int count)
1068 struct ib_recv_wr *rx_wr, *rx_wr_failed;
1070 struct ib_conn *ib_conn = &iser_conn->ib_conn;
1071 unsigned int my_rx_head = iser_conn->rx_desc_head;
1072 struct iser_rx_desc *rx_desc;
1074 for (rx_wr = ib_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
1075 rx_desc = &iser_conn->rx_descs[my_rx_head];
1076 rx_wr->wr_id = (unsigned long)rx_desc;
1077 rx_wr->sg_list = &rx_desc->rx_sg;
1079 rx_wr->next = rx_wr + 1;
1080 my_rx_head = (my_rx_head + 1) & iser_conn->qp_max_recv_dtos_mask;
1084 rx_wr->next = NULL; /* mark end of work requests list */
1086 ib_conn->post_recv_buf_count += count;
1087 ib_ret = ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &rx_wr_failed);
1089 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1090 ib_conn->post_recv_buf_count -= count;
1092 iser_conn->rx_desc_head = my_rx_head;
1098 * iser_start_send - Initiate a Send DTO operation
1100 * returns 0 on success, -1 on failure
1102 int iser_post_send(struct ib_conn *ib_conn, struct iser_tx_desc *tx_desc,
1106 struct ib_send_wr send_wr, *send_wr_failed;
1108 ib_dma_sync_single_for_device(ib_conn->device->ib_device,
1109 tx_desc->dma_addr, ISER_HEADERS_LEN,
1112 send_wr.next = NULL;
1113 send_wr.wr_id = (unsigned long)tx_desc;
1114 send_wr.sg_list = tx_desc->tx_sg;
1115 send_wr.num_sge = tx_desc->num_sge;
1116 send_wr.opcode = IB_WR_SEND;
1117 send_wr.send_flags = signal ? IB_SEND_SIGNALED : 0;
1119 ib_ret = ib_post_send(ib_conn->qp, &send_wr, &send_wr_failed);
1121 iser_err("ib_post_send failed, ret:%d\n", ib_ret);
1127 * is_iser_tx_desc - Indicate if the completion wr_id
1128 * is a TX descriptor or not.
1129 * @iser_conn: iser connection
1130 * @wr_id: completion WR identifier
1132 * Since we cannot rely on wc opcode in FLUSH errors
1133 * we must work around it by checking if the wr_id address
1134 * falls in the iser connection rx_descs buffer. If so
1135 * it is an RX descriptor, otherwize it is a TX.
1138 is_iser_tx_desc(struct iser_conn *iser_conn, void *wr_id)
1140 void *start = iser_conn->rx_descs;
1141 int len = iser_conn->num_rx_descs * sizeof(*iser_conn->rx_descs);
1143 if (wr_id >= start && wr_id < start + len)
1150 * iser_handle_comp_error() - Handle error completion
1151 * @ib_conn: connection RDMA resources
1152 * @wc: work completion
1154 * Notes: We may handle a FLUSH error completion and in this case
1155 * we only cleanup in case TX type was DATAOUT. For non-FLUSH
1156 * error completion we should also notify iscsi layer that
1157 * connection is failed (in case we passed bind stage).
1160 iser_handle_comp_error(struct ib_conn *ib_conn,
1163 struct iser_conn *iser_conn = container_of(ib_conn, struct iser_conn,
1166 if (wc->status != IB_WC_WR_FLUSH_ERR)
1167 if (iser_conn->iscsi_conn)
1168 iscsi_conn_failure(iser_conn->iscsi_conn,
1169 ISCSI_ERR_CONN_FAILED);
1171 if (is_iser_tx_desc(iser_conn, (void *)wc->wr_id)) {
1172 struct iser_tx_desc *desc = (struct iser_tx_desc *)wc->wr_id;
1174 if (desc->type == ISCSI_TX_DATAOUT)
1175 kmem_cache_free(ig.desc_cache, desc);
1177 ib_conn->post_recv_buf_count--;
1182 * iser_handle_wc - handle a single work completion
1183 * @wc: work completion
1185 * Soft-IRQ context, work completion can be either
1186 * SEND or RECV, and can turn out successful or
1187 * with error (or flush error).
1189 static void iser_handle_wc(struct ib_wc *wc)
1191 struct ib_conn *ib_conn;
1192 struct iser_tx_desc *tx_desc;
1193 struct iser_rx_desc *rx_desc;
1195 ib_conn = wc->qp->qp_context;
1196 if (wc->status == IB_WC_SUCCESS) {
1197 if (wc->opcode == IB_WC_RECV) {
1198 rx_desc = (struct iser_rx_desc *)wc->wr_id;
1199 iser_rcv_completion(rx_desc, wc->byte_len,
1202 if (wc->opcode == IB_WC_SEND) {
1203 tx_desc = (struct iser_tx_desc *)wc->wr_id;
1204 iser_snd_completion(tx_desc, ib_conn);
1206 iser_err("Unknown wc opcode %d\n", wc->opcode);
1209 if (wc->status != IB_WC_WR_FLUSH_ERR)
1210 iser_err("wr id %llx status %d vend_err %x\n",
1211 wc->wr_id, wc->status, wc->vendor_err);
1213 iser_dbg("flush error: wr id %llx\n", wc->wr_id);
1215 if (wc->wr_id != ISER_FASTREG_LI_WRID &&
1216 wc->wr_id != ISER_BEACON_WRID)
1217 iser_handle_comp_error(ib_conn, wc);
1219 /* complete in case all flush errors were consumed */
1220 if (wc->wr_id == ISER_BEACON_WRID)
1221 complete(&ib_conn->flush_comp);
1226 * iser_cq_tasklet_fn - iSER completion polling loop
1227 * @data: iSER completion context
1229 * Soft-IRQ context, polling connection CQ until
1230 * either CQ was empty or we exausted polling budget
1232 static void iser_cq_tasklet_fn(unsigned long data)
1234 struct iser_comp *comp = (struct iser_comp *)data;
1235 struct ib_cq *cq = comp->cq;
1236 struct ib_wc *const wcs = comp->wcs;
1237 int i, n, completed = 0;
1239 while ((n = ib_poll_cq(cq, ARRAY_SIZE(comp->wcs), wcs)) > 0) {
1240 for (i = 0; i < n; i++)
1241 iser_handle_wc(&wcs[i]);
1244 if (completed >= iser_cq_poll_limit)
1249 * It is assumed here that arming CQ only once its empty
1250 * would not cause interrupts to be missed.
1252 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1254 iser_dbg("got %d completions\n", completed);
1257 static void iser_cq_callback(struct ib_cq *cq, void *cq_context)
1259 struct iser_comp *comp = cq_context;
1261 tasklet_schedule(&comp->tasklet);
1264 u8 iser_check_task_pi_status(struct iscsi_iser_task *iser_task,
1265 enum iser_data_dir cmd_dir, sector_t *sector)
1267 struct iser_mem_reg *reg = &iser_task->rdma_regd[cmd_dir].reg;
1268 struct fast_reg_descriptor *desc = reg->mem_h;
1269 unsigned long sector_size = iser_task->sc->device->sector_size;
1270 struct ib_mr_status mr_status;
1273 if (desc && desc->reg_indicators & ISER_FASTREG_PROTECTED) {
1274 desc->reg_indicators &= ~ISER_FASTREG_PROTECTED;
1275 ret = ib_check_mr_status(desc->pi_ctx->sig_mr,
1276 IB_MR_CHECK_SIG_STATUS, &mr_status);
1278 pr_err("ib_check_mr_status failed, ret %d\n", ret);
1282 if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1283 sector_t sector_off = mr_status.sig_err.sig_err_offset;
1285 do_div(sector_off, sector_size + 8);
1286 *sector = scsi_get_lba(iser_task->sc) + sector_off;
1288 pr_err("PI error found type %d at sector %llx "
1289 "expected %x vs actual %x\n",
1290 mr_status.sig_err.err_type,
1291 (unsigned long long)*sector,
1292 mr_status.sig_err.expected,
1293 mr_status.sig_err.actual);
1295 switch (mr_status.sig_err.err_type) {
1296 case IB_SIG_BAD_GUARD:
1298 case IB_SIG_BAD_REFTAG:
1300 case IB_SIG_BAD_APPTAG:
1308 /* Not alot we can do here, return ambiguous guard error */