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_CQ_LEN (ISER_QP_MAX_RECV_DTOS * ISCSI_ISER_MAX_CONN)
43 #define ISER_MAX_TX_CQ_LEN (ISER_QP_MAX_REQ_DTOS * ISCSI_ISER_MAX_CONN)
45 static void iser_cq_tasklet_fn(unsigned long data);
46 static void iser_cq_callback(struct ib_cq *cq, void *cq_context);
48 static void iser_cq_event_callback(struct ib_event *cause, void *context)
50 iser_err("got cq event %d \n", cause->event);
53 static void iser_qp_event_callback(struct ib_event *cause, void *context)
55 iser_err("got qp event %d\n",cause->event);
58 static void iser_event_handler(struct ib_event_handler *handler,
59 struct ib_event *event)
61 iser_err("async event %d on device %s port %d\n", event->event,
62 event->device->name, event->element.port_num);
66 * iser_create_device_ib_res - creates Protection Domain (PD), Completion
67 * Queue (CQ), DMA Memory Region (DMA MR) with the device associated with
70 * returns 0 on success, -1 on failure
72 static int iser_create_device_ib_res(struct iser_device *device)
74 struct iser_cq_desc *cq_desc;
75 struct ib_device_attr *dev_attr = &device->dev_attr;
78 ret = ib_query_device(device->ib_device, dev_attr);
80 pr_warn("Query device failed for %s\n", device->ib_device->name);
84 /* Assign function handles - based on FMR support */
85 if (device->ib_device->alloc_fmr && device->ib_device->dealloc_fmr &&
86 device->ib_device->map_phys_fmr && device->ib_device->unmap_fmr) {
87 iser_info("FMR supported, using FMR for registration\n");
88 device->iser_alloc_rdma_reg_res = iser_create_fmr_pool;
89 device->iser_free_rdma_reg_res = iser_free_fmr_pool;
90 device->iser_reg_rdma_mem = iser_reg_rdma_mem_fmr;
91 device->iser_unreg_rdma_mem = iser_unreg_mem_fmr;
93 if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
94 iser_info("FastReg supported, using FastReg for registration\n");
95 device->iser_alloc_rdma_reg_res = iser_create_fastreg_pool;
96 device->iser_free_rdma_reg_res = iser_free_fastreg_pool;
97 device->iser_reg_rdma_mem = iser_reg_rdma_mem_fastreg;
98 device->iser_unreg_rdma_mem = iser_unreg_mem_fastreg;
100 iser_err("IB device does not support FMRs nor FastRegs, can't register memory\n");
104 device->cqs_used = min(ISER_MAX_CQ, device->ib_device->num_comp_vectors);
105 iser_info("using %d CQs, device %s supports %d vectors\n",
106 device->cqs_used, device->ib_device->name,
107 device->ib_device->num_comp_vectors);
109 device->cq_desc = kmalloc(sizeof(struct iser_cq_desc) * device->cqs_used,
111 if (device->cq_desc == NULL)
113 cq_desc = device->cq_desc;
115 device->pd = ib_alloc_pd(device->ib_device);
116 if (IS_ERR(device->pd))
119 for (i = 0; i < device->cqs_used; i++) {
120 cq_desc[i].device = device;
121 cq_desc[i].cq_index = i;
123 device->rx_cq[i] = ib_create_cq(device->ib_device,
125 iser_cq_event_callback,
127 ISER_MAX_RX_CQ_LEN, i);
128 if (IS_ERR(device->rx_cq[i]))
131 device->tx_cq[i] = ib_create_cq(device->ib_device,
132 NULL, iser_cq_event_callback,
134 ISER_MAX_TX_CQ_LEN, i);
136 if (IS_ERR(device->tx_cq[i]))
139 if (ib_req_notify_cq(device->rx_cq[i], IB_CQ_NEXT_COMP))
142 tasklet_init(&device->cq_tasklet[i],
144 (unsigned long)&cq_desc[i]);
147 device->mr = ib_get_dma_mr(device->pd, IB_ACCESS_LOCAL_WRITE |
148 IB_ACCESS_REMOTE_WRITE |
149 IB_ACCESS_REMOTE_READ);
150 if (IS_ERR(device->mr))
153 INIT_IB_EVENT_HANDLER(&device->event_handler, device->ib_device,
155 if (ib_register_event_handler(&device->event_handler))
161 ib_dereg_mr(device->mr);
163 for (j = 0; j < device->cqs_used; j++)
164 tasklet_kill(&device->cq_tasklet[j]);
166 for (j = 0; j < i; j++) {
167 if (device->tx_cq[j])
168 ib_destroy_cq(device->tx_cq[j]);
169 if (device->rx_cq[j])
170 ib_destroy_cq(device->rx_cq[j]);
172 ib_dealloc_pd(device->pd);
174 kfree(device->cq_desc);
176 iser_err("failed to allocate an IB resource\n");
181 * iser_free_device_ib_res - destroy/dealloc/dereg the DMA MR,
182 * CQ and PD created with the device associated with the adapator.
184 static void iser_free_device_ib_res(struct iser_device *device)
187 BUG_ON(device->mr == NULL);
189 for (i = 0; i < device->cqs_used; i++) {
190 tasklet_kill(&device->cq_tasklet[i]);
191 (void)ib_destroy_cq(device->tx_cq[i]);
192 (void)ib_destroy_cq(device->rx_cq[i]);
193 device->tx_cq[i] = NULL;
194 device->rx_cq[i] = NULL;
197 (void)ib_unregister_event_handler(&device->event_handler);
198 (void)ib_dereg_mr(device->mr);
199 (void)ib_dealloc_pd(device->pd);
201 kfree(device->cq_desc);
208 * iser_create_fmr_pool - Creates FMR pool and page_vector
210 * returns 0 on success, or errno code on failure
212 int iser_create_fmr_pool(struct iser_conn *ib_conn, unsigned cmds_max)
214 struct iser_device *device = ib_conn->device;
215 struct ib_fmr_pool_param params;
218 ib_conn->fmr.page_vec = kmalloc(sizeof(*ib_conn->fmr.page_vec) +
219 (sizeof(u64)*(ISCSI_ISER_SG_TABLESIZE + 1)),
221 if (!ib_conn->fmr.page_vec)
224 ib_conn->fmr.page_vec->pages = (u64 *)(ib_conn->fmr.page_vec + 1);
226 params.page_shift = SHIFT_4K;
227 /* when the first/last SG element are not start/end *
228 * page aligned, the map whould be of N+1 pages */
229 params.max_pages_per_fmr = ISCSI_ISER_SG_TABLESIZE + 1;
230 /* make the pool size twice the max number of SCSI commands *
231 * the ML is expected to queue, watermark for unmap at 50% */
232 params.pool_size = cmds_max * 2;
233 params.dirty_watermark = cmds_max;
235 params.flush_function = NULL;
236 params.access = (IB_ACCESS_LOCAL_WRITE |
237 IB_ACCESS_REMOTE_WRITE |
238 IB_ACCESS_REMOTE_READ);
240 ib_conn->fmr.pool = ib_create_fmr_pool(device->pd, ¶ms);
241 if (!IS_ERR(ib_conn->fmr.pool))
244 /* no FMR => no need for page_vec */
245 kfree(ib_conn->fmr.page_vec);
246 ib_conn->fmr.page_vec = NULL;
248 ret = PTR_ERR(ib_conn->fmr.pool);
249 ib_conn->fmr.pool = NULL;
250 if (ret != -ENOSYS) {
251 iser_err("FMR allocation failed, err %d\n", ret);
254 iser_warn("FMRs are not supported, using unaligned mode\n");
260 * iser_free_fmr_pool - releases the FMR pool and page vec
262 void iser_free_fmr_pool(struct iser_conn *ib_conn)
264 iser_info("freeing conn %p fmr pool %p\n",
265 ib_conn, ib_conn->fmr.pool);
267 if (ib_conn->fmr.pool != NULL)
268 ib_destroy_fmr_pool(ib_conn->fmr.pool);
270 ib_conn->fmr.pool = NULL;
272 kfree(ib_conn->fmr.page_vec);
273 ib_conn->fmr.page_vec = NULL;
277 iser_create_fastreg_desc(struct ib_device *ib_device, struct ib_pd *pd,
278 bool pi_enable, struct fast_reg_descriptor *desc)
282 desc->data_frpl = ib_alloc_fast_reg_page_list(ib_device,
283 ISCSI_ISER_SG_TABLESIZE + 1);
284 if (IS_ERR(desc->data_frpl)) {
285 ret = PTR_ERR(desc->data_frpl);
286 iser_err("Failed to allocate ib_fast_reg_page_list err=%d\n",
288 return PTR_ERR(desc->data_frpl);
291 desc->data_mr = ib_alloc_fast_reg_mr(pd, ISCSI_ISER_SG_TABLESIZE + 1);
292 if (IS_ERR(desc->data_mr)) {
293 ret = PTR_ERR(desc->data_mr);
294 iser_err("Failed to allocate ib_fast_reg_mr err=%d\n", ret);
295 goto fast_reg_mr_failure;
297 desc->reg_indicators |= ISER_DATA_KEY_VALID;
300 struct ib_mr_init_attr mr_init_attr = {0};
301 struct iser_pi_context *pi_ctx = NULL;
303 desc->pi_ctx = kzalloc(sizeof(*desc->pi_ctx), GFP_KERNEL);
305 iser_err("Failed to allocate pi context\n");
307 goto pi_ctx_alloc_failure;
309 pi_ctx = desc->pi_ctx;
311 pi_ctx->prot_frpl = ib_alloc_fast_reg_page_list(ib_device,
312 ISCSI_ISER_SG_TABLESIZE);
313 if (IS_ERR(pi_ctx->prot_frpl)) {
314 ret = PTR_ERR(pi_ctx->prot_frpl);
315 iser_err("Failed to allocate prot frpl ret=%d\n",
317 goto prot_frpl_failure;
320 pi_ctx->prot_mr = ib_alloc_fast_reg_mr(pd,
321 ISCSI_ISER_SG_TABLESIZE + 1);
322 if (IS_ERR(pi_ctx->prot_mr)) {
323 ret = PTR_ERR(pi_ctx->prot_mr);
324 iser_err("Failed to allocate prot frmr ret=%d\n",
326 goto prot_mr_failure;
328 desc->reg_indicators |= ISER_PROT_KEY_VALID;
330 mr_init_attr.max_reg_descriptors = 2;
331 mr_init_attr.flags |= IB_MR_SIGNATURE_EN;
332 pi_ctx->sig_mr = ib_create_mr(pd, &mr_init_attr);
333 if (IS_ERR(pi_ctx->sig_mr)) {
334 ret = PTR_ERR(pi_ctx->sig_mr);
335 iser_err("Failed to allocate signature enabled mr err=%d\n",
339 desc->reg_indicators |= ISER_SIG_KEY_VALID;
341 desc->reg_indicators &= ~ISER_FASTREG_PROTECTED;
343 iser_dbg("Create fr_desc %p page_list %p\n",
344 desc, desc->data_frpl->page_list);
348 ib_dereg_mr(desc->pi_ctx->prot_mr);
350 ib_free_fast_reg_page_list(desc->pi_ctx->prot_frpl);
353 pi_ctx_alloc_failure:
354 ib_dereg_mr(desc->data_mr);
356 ib_free_fast_reg_page_list(desc->data_frpl);
362 * iser_create_fastreg_pool - Creates pool of fast_reg descriptors
363 * for fast registration work requests.
364 * returns 0 on success, or errno code on failure
366 int iser_create_fastreg_pool(struct iser_conn *ib_conn, unsigned cmds_max)
368 struct iser_device *device = ib_conn->device;
369 struct fast_reg_descriptor *desc;
372 INIT_LIST_HEAD(&ib_conn->fastreg.pool);
373 ib_conn->fastreg.pool_size = 0;
374 for (i = 0; i < cmds_max; i++) {
375 desc = kzalloc(sizeof(*desc), GFP_KERNEL);
377 iser_err("Failed to allocate a new fast_reg descriptor\n");
382 ret = iser_create_fastreg_desc(device->ib_device, device->pd,
383 ib_conn->pi_support, desc);
385 iser_err("Failed to create fastreg descriptor err=%d\n",
391 list_add_tail(&desc->list, &ib_conn->fastreg.pool);
392 ib_conn->fastreg.pool_size++;
398 iser_free_fastreg_pool(ib_conn);
403 * iser_free_fastreg_pool - releases the pool of fast_reg descriptors
405 void iser_free_fastreg_pool(struct iser_conn *ib_conn)
407 struct fast_reg_descriptor *desc, *tmp;
410 if (list_empty(&ib_conn->fastreg.pool))
413 iser_info("freeing conn %p fr pool\n", ib_conn);
415 list_for_each_entry_safe(desc, tmp, &ib_conn->fastreg.pool, list) {
416 list_del(&desc->list);
417 ib_free_fast_reg_page_list(desc->data_frpl);
418 ib_dereg_mr(desc->data_mr);
420 ib_free_fast_reg_page_list(desc->pi_ctx->prot_frpl);
421 ib_dereg_mr(desc->pi_ctx->prot_mr);
422 ib_destroy_mr(desc->pi_ctx->sig_mr);
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 iser_conn *ib_conn)
441 struct iser_device *device;
442 struct ib_qp_init_attr init_attr;
444 int index, min_index = 0;
446 BUG_ON(ib_conn->device == NULL);
448 device = ib_conn->device;
450 memset(&init_attr, 0, sizeof init_attr);
452 mutex_lock(&ig.connlist_mutex);
453 /* select the CQ with the minimal number of usages */
454 for (index = 0; index < device->cqs_used; index++)
455 if (device->cq_active_qps[index] <
456 device->cq_active_qps[min_index])
458 device->cq_active_qps[min_index]++;
459 mutex_unlock(&ig.connlist_mutex);
460 iser_info("cq index %d used for ib_conn %p\n", min_index, ib_conn);
462 init_attr.event_handler = iser_qp_event_callback;
463 init_attr.qp_context = (void *)ib_conn;
464 init_attr.send_cq = device->tx_cq[min_index];
465 init_attr.recv_cq = device->rx_cq[min_index];
466 init_attr.cap.max_recv_wr = ISER_QP_MAX_RECV_DTOS;
467 init_attr.cap.max_send_sge = 2;
468 init_attr.cap.max_recv_sge = 1;
469 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
470 init_attr.qp_type = IB_QPT_RC;
471 if (ib_conn->pi_support) {
472 init_attr.cap.max_send_wr = ISER_QP_SIG_MAX_REQ_DTOS;
473 init_attr.create_flags |= IB_QP_CREATE_SIGNATURE_EN;
475 init_attr.cap.max_send_wr = ISER_QP_MAX_REQ_DTOS;
478 ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
482 ib_conn->qp = ib_conn->cma_id->qp;
483 iser_info("setting conn %p cma_id %p qp %p\n",
484 ib_conn, ib_conn->cma_id,
485 ib_conn->cma_id->qp);
489 iser_err("unable to alloc mem or create resource, err %d\n", ret);
494 * releases the QP object
496 static void iser_free_ib_conn_res(struct iser_conn *ib_conn)
499 BUG_ON(ib_conn == NULL);
501 iser_info("freeing conn %p cma_id %p qp %p\n",
502 ib_conn, ib_conn->cma_id,
505 /* qp is created only once both addr & route are resolved */
507 if (ib_conn->qp != NULL) {
508 cq_index = ((struct iser_cq_desc *)ib_conn->qp->recv_cq->cq_context)->cq_index;
509 ib_conn->device->cq_active_qps[cq_index]--;
511 rdma_destroy_qp(ib_conn->cma_id);
518 * based on the resolved device node GUID see if there already allocated
519 * device for this device. If there's no such, create one.
522 struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
524 struct iser_device *device;
526 mutex_lock(&ig.device_list_mutex);
528 list_for_each_entry(device, &ig.device_list, ig_list)
529 /* find if there's a match using the node GUID */
530 if (device->ib_device->node_guid == cma_id->device->node_guid)
533 device = kzalloc(sizeof *device, GFP_KERNEL);
537 /* assign this device to the device */
538 device->ib_device = cma_id->device;
539 /* init the device and link it into ig device list */
540 if (iser_create_device_ib_res(device)) {
545 list_add(&device->ig_list, &ig.device_list);
550 mutex_unlock(&ig.device_list_mutex);
554 /* if there's no demand for this device, release it */
555 static void iser_device_try_release(struct iser_device *device)
557 mutex_lock(&ig.device_list_mutex);
559 iser_info("device %p refcount %d\n", device, device->refcount);
560 if (!device->refcount) {
561 iser_free_device_ib_res(device);
562 list_del(&device->ig_list);
565 mutex_unlock(&ig.device_list_mutex);
569 * Called with state mutex held
571 static int iser_conn_state_comp_exch(struct iser_conn *ib_conn,
572 enum iser_ib_conn_state comp,
573 enum iser_ib_conn_state exch)
577 if ((ret = (ib_conn->state == comp)))
578 ib_conn->state = exch;
582 void iser_release_work(struct work_struct *work)
584 struct iser_conn *ib_conn;
587 ib_conn = container_of(work, struct iser_conn, release_work);
589 /* wait for .conn_stop callback */
590 rc = wait_for_completion_timeout(&ib_conn->stop_completion, 30 * HZ);
593 /* wait for the qp`s post send and post receive buffers to empty */
594 rc = wait_for_completion_timeout(&ib_conn->flush_completion, 30 * HZ);
597 ib_conn->state = ISER_CONN_DOWN;
599 mutex_lock(&ib_conn->state_mutex);
600 ib_conn->state = ISER_CONN_DOWN;
601 mutex_unlock(&ib_conn->state_mutex);
603 iser_conn_release(ib_conn);
607 * Frees all conn objects and deallocs conn descriptor
609 void iser_conn_release(struct iser_conn *ib_conn)
611 struct iser_device *device = ib_conn->device;
613 mutex_lock(&ig.connlist_mutex);
614 list_del(&ib_conn->conn_list);
615 mutex_unlock(&ig.connlist_mutex);
617 mutex_lock(&ib_conn->state_mutex);
618 BUG_ON(ib_conn->state != ISER_CONN_DOWN);
620 iser_free_rx_descriptors(ib_conn);
621 iser_free_ib_conn_res(ib_conn);
622 ib_conn->device = NULL;
623 /* on EVENT_ADDR_ERROR there's no device yet for this conn */
625 iser_device_try_release(device);
626 mutex_unlock(&ib_conn->state_mutex);
628 /* if cma handler context, the caller actually destroy the id */
629 if (ib_conn->cma_id != NULL) {
630 rdma_destroy_id(ib_conn->cma_id);
631 ib_conn->cma_id = NULL;
637 * triggers start of the disconnect procedures and wait for them to be done
639 void iser_conn_terminate(struct iser_conn *ib_conn)
643 /* change the ib conn state only if the conn is UP, however always call
644 * rdma_disconnect since this is the only way to cause the CMA to change
645 * the QP state to ERROR
648 iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP, ISER_CONN_TERMINATING);
649 err = rdma_disconnect(ib_conn->cma_id);
651 iser_err("Failed to disconnect, conn: 0x%p err %d\n",
656 * Called with state mutex held
658 static void iser_connect_error(struct rdma_cm_id *cma_id)
660 struct iser_conn *ib_conn;
662 ib_conn = (struct iser_conn *)cma_id->context;
663 ib_conn->state = ISER_CONN_DOWN;
667 * Called with state mutex held
669 static void iser_addr_handler(struct rdma_cm_id *cma_id)
671 struct iser_device *device;
672 struct iser_conn *ib_conn;
675 ib_conn = (struct iser_conn *)cma_id->context;
676 if (ib_conn->state != ISER_CONN_PENDING)
680 device = iser_device_find_by_ib_device(cma_id);
682 iser_err("device lookup/creation failed\n");
683 iser_connect_error(cma_id);
687 ib_conn->device = device;
689 /* connection T10-PI support */
690 if (iser_pi_enable) {
691 if (!(device->dev_attr.device_cap_flags &
692 IB_DEVICE_SIGNATURE_HANDOVER)) {
693 iser_warn("T10-PI requested but not supported on %s, "
694 "continue without T10-PI\n",
695 ib_conn->device->ib_device->name);
696 ib_conn->pi_support = false;
698 ib_conn->pi_support = true;
702 ret = rdma_resolve_route(cma_id, 1000);
704 iser_err("resolve route failed: %d\n", ret);
705 iser_connect_error(cma_id);
711 * Called with state mutex held
713 static void iser_route_handler(struct rdma_cm_id *cma_id)
715 struct rdma_conn_param conn_param;
717 struct iser_cm_hdr req_hdr;
718 struct iser_conn *ib_conn = (struct iser_conn *)cma_id->context;
719 struct iser_device *device = ib_conn->device;
721 if (ib_conn->state != ISER_CONN_PENDING)
725 ret = iser_create_ib_conn_res((struct iser_conn *)cma_id->context);
729 memset(&conn_param, 0, sizeof conn_param);
730 conn_param.responder_resources = device->dev_attr.max_qp_rd_atom;
731 conn_param.initiator_depth = 1;
732 conn_param.retry_count = 7;
733 conn_param.rnr_retry_count = 6;
735 memset(&req_hdr, 0, sizeof(req_hdr));
736 req_hdr.flags = (ISER_ZBVA_NOT_SUPPORTED |
737 ISER_SEND_W_INV_NOT_SUPPORTED);
738 conn_param.private_data = (void *)&req_hdr;
739 conn_param.private_data_len = sizeof(struct iser_cm_hdr);
741 ret = rdma_connect(cma_id, &conn_param);
743 iser_err("failure connecting: %d\n", ret);
749 iser_connect_error(cma_id);
752 static void iser_connected_handler(struct rdma_cm_id *cma_id)
754 struct iser_conn *ib_conn;
755 struct ib_qp_attr attr;
756 struct ib_qp_init_attr init_attr;
758 ib_conn = (struct iser_conn *)cma_id->context;
759 if (ib_conn->state != ISER_CONN_PENDING)
763 (void)ib_query_qp(cma_id->qp, &attr, ~0, &init_attr);
764 iser_info("remote qpn:%x my qpn:%x\n", attr.dest_qp_num, cma_id->qp->qp_num);
766 ib_conn->state = ISER_CONN_UP;
767 complete(&ib_conn->up_completion);
770 static void iser_disconnected_handler(struct rdma_cm_id *cma_id)
772 struct iser_conn *ib_conn;
774 ib_conn = (struct iser_conn *)cma_id->context;
776 /* getting here when the state is UP means that the conn is being *
777 * terminated asynchronously from the iSCSI layer's perspective. */
778 if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
779 ISER_CONN_TERMINATING)){
780 if (ib_conn->iscsi_conn)
781 iscsi_conn_failure(ib_conn->iscsi_conn, ISCSI_ERR_CONN_FAILED);
783 iser_err("iscsi_iser connection isn't bound\n");
786 /* Complete the termination process if no posts are pending */
787 if (ib_conn->post_recv_buf_count == 0 &&
788 (atomic_read(&ib_conn->post_send_buf_count) == 0)) {
789 complete(&ib_conn->flush_completion);
793 static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
795 struct iser_conn *ib_conn;
797 ib_conn = (struct iser_conn *)cma_id->context;
798 iser_info("event %d status %d conn %p id %p\n",
799 event->event, event->status, cma_id->context, cma_id);
801 mutex_lock(&ib_conn->state_mutex);
802 switch (event->event) {
803 case RDMA_CM_EVENT_ADDR_RESOLVED:
804 iser_addr_handler(cma_id);
806 case RDMA_CM_EVENT_ROUTE_RESOLVED:
807 iser_route_handler(cma_id);
809 case RDMA_CM_EVENT_ESTABLISHED:
810 iser_connected_handler(cma_id);
812 case RDMA_CM_EVENT_ADDR_ERROR:
813 case RDMA_CM_EVENT_ROUTE_ERROR:
814 case RDMA_CM_EVENT_CONNECT_ERROR:
815 case RDMA_CM_EVENT_UNREACHABLE:
816 case RDMA_CM_EVENT_REJECTED:
817 iser_connect_error(cma_id);
819 case RDMA_CM_EVENT_DISCONNECTED:
820 case RDMA_CM_EVENT_DEVICE_REMOVAL:
821 case RDMA_CM_EVENT_ADDR_CHANGE:
822 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
823 iser_disconnected_handler(cma_id);
826 iser_err("Unexpected RDMA CM event (%d)\n", event->event);
829 mutex_unlock(&ib_conn->state_mutex);
833 void iser_conn_init(struct iser_conn *ib_conn)
835 ib_conn->state = ISER_CONN_INIT;
836 ib_conn->post_recv_buf_count = 0;
837 atomic_set(&ib_conn->post_send_buf_count, 0);
838 init_completion(&ib_conn->stop_completion);
839 init_completion(&ib_conn->flush_completion);
840 init_completion(&ib_conn->up_completion);
841 INIT_LIST_HEAD(&ib_conn->conn_list);
842 spin_lock_init(&ib_conn->lock);
843 mutex_init(&ib_conn->state_mutex);
847 * starts the process of connecting to the target
848 * sleeps until the connection is established or rejected
850 int iser_connect(struct iser_conn *ib_conn,
851 struct sockaddr *src_addr,
852 struct sockaddr *dst_addr,
857 mutex_lock(&ib_conn->state_mutex);
859 sprintf(ib_conn->name, "%pISp", dst_addr);
861 iser_info("connecting to: %s\n", ib_conn->name);
863 /* the device is known only --after-- address resolution */
864 ib_conn->device = NULL;
866 ib_conn->state = ISER_CONN_PENDING;
868 ib_conn->cma_id = rdma_create_id(iser_cma_handler,
870 RDMA_PS_TCP, IB_QPT_RC);
871 if (IS_ERR(ib_conn->cma_id)) {
872 err = PTR_ERR(ib_conn->cma_id);
873 iser_err("rdma_create_id failed: %d\n", err);
877 err = rdma_resolve_addr(ib_conn->cma_id, src_addr, dst_addr, 1000);
879 iser_err("rdma_resolve_addr failed: %d\n", err);
884 wait_for_completion_interruptible(&ib_conn->up_completion);
886 if (ib_conn->state != ISER_CONN_UP) {
888 goto connect_failure;
891 mutex_unlock(&ib_conn->state_mutex);
893 mutex_lock(&ig.connlist_mutex);
894 list_add(&ib_conn->conn_list, &ig.connlist);
895 mutex_unlock(&ig.connlist_mutex);
899 ib_conn->cma_id = NULL;
901 ib_conn->state = ISER_CONN_DOWN;
903 mutex_unlock(&ib_conn->state_mutex);
904 iser_conn_release(ib_conn);
909 * iser_reg_page_vec - Register physical memory
911 * returns: 0 on success, errno code on failure
913 int iser_reg_page_vec(struct iser_conn *ib_conn,
914 struct iser_page_vec *page_vec,
915 struct iser_mem_reg *mem_reg)
917 struct ib_pool_fmr *mem;
922 page_list = page_vec->pages;
923 io_addr = page_list[0];
925 mem = ib_fmr_pool_map_phys(ib_conn->fmr.pool,
931 status = (int)PTR_ERR(mem);
932 iser_err("ib_fmr_pool_map_phys failed: %d\n", status);
936 mem_reg->lkey = mem->fmr->lkey;
937 mem_reg->rkey = mem->fmr->rkey;
938 mem_reg->len = page_vec->length * SIZE_4K;
939 mem_reg->va = io_addr;
941 mem_reg->mem_h = (void *)mem;
943 mem_reg->va += page_vec->offset;
944 mem_reg->len = page_vec->data_size;
946 iser_dbg("PHYSICAL Mem.register, [PHYS p_array: 0x%p, sz: %d, "
947 "entry[0]: (0x%08lx,%ld)] -> "
948 "[lkey: 0x%08X mem_h: 0x%p va: 0x%08lX sz: %ld]\n",
949 page_vec, page_vec->length,
950 (unsigned long)page_vec->pages[0],
951 (unsigned long)page_vec->data_size,
952 (unsigned int)mem_reg->lkey, mem_reg->mem_h,
953 (unsigned long)mem_reg->va, (unsigned long)mem_reg->len);
958 * Unregister (previosuly registered using FMR) memory.
959 * If memory is non-FMR does nothing.
961 void iser_unreg_mem_fmr(struct iscsi_iser_task *iser_task,
962 enum iser_data_dir cmd_dir)
964 struct iser_mem_reg *reg = &iser_task->rdma_regd[cmd_dir].reg;
970 iser_dbg("PHYSICAL Mem.Unregister mem_h %p\n",reg->mem_h);
972 ret = ib_fmr_pool_unmap((struct ib_pool_fmr *)reg->mem_h);
974 iser_err("ib_fmr_pool_unmap failed %d\n", ret);
979 void iser_unreg_mem_fastreg(struct iscsi_iser_task *iser_task,
980 enum iser_data_dir cmd_dir)
982 struct iser_mem_reg *reg = &iser_task->rdma_regd[cmd_dir].reg;
983 struct iser_conn *ib_conn = iser_task->ib_conn;
984 struct fast_reg_descriptor *desc = reg->mem_h;
991 spin_lock_bh(&ib_conn->lock);
992 list_add_tail(&desc->list, &ib_conn->fastreg.pool);
993 spin_unlock_bh(&ib_conn->lock);
996 int iser_post_recvl(struct iser_conn *ib_conn)
998 struct ib_recv_wr rx_wr, *rx_wr_failed;
1002 sge.addr = ib_conn->login_resp_dma;
1003 sge.length = ISER_RX_LOGIN_SIZE;
1004 sge.lkey = ib_conn->device->mr->lkey;
1006 rx_wr.wr_id = (unsigned long)ib_conn->login_resp_buf;
1007 rx_wr.sg_list = &sge;
1011 ib_conn->post_recv_buf_count++;
1012 ib_ret = ib_post_recv(ib_conn->qp, &rx_wr, &rx_wr_failed);
1014 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1015 ib_conn->post_recv_buf_count--;
1020 int iser_post_recvm(struct iser_conn *ib_conn, int count)
1022 struct ib_recv_wr *rx_wr, *rx_wr_failed;
1024 unsigned int my_rx_head = ib_conn->rx_desc_head;
1025 struct iser_rx_desc *rx_desc;
1027 for (rx_wr = ib_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
1028 rx_desc = &ib_conn->rx_descs[my_rx_head];
1029 rx_wr->wr_id = (unsigned long)rx_desc;
1030 rx_wr->sg_list = &rx_desc->rx_sg;
1032 rx_wr->next = rx_wr + 1;
1033 my_rx_head = (my_rx_head + 1) & ib_conn->qp_max_recv_dtos_mask;
1037 rx_wr->next = NULL; /* mark end of work requests list */
1039 ib_conn->post_recv_buf_count += count;
1040 ib_ret = ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &rx_wr_failed);
1042 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1043 ib_conn->post_recv_buf_count -= count;
1045 ib_conn->rx_desc_head = my_rx_head;
1051 * iser_start_send - Initiate a Send DTO operation
1053 * returns 0 on success, -1 on failure
1055 int iser_post_send(struct iser_conn *ib_conn, struct iser_tx_desc *tx_desc)
1058 struct ib_send_wr send_wr, *send_wr_failed;
1060 ib_dma_sync_single_for_device(ib_conn->device->ib_device,
1061 tx_desc->dma_addr, ISER_HEADERS_LEN, DMA_TO_DEVICE);
1063 send_wr.next = NULL;
1064 send_wr.wr_id = (unsigned long)tx_desc;
1065 send_wr.sg_list = tx_desc->tx_sg;
1066 send_wr.num_sge = tx_desc->num_sge;
1067 send_wr.opcode = IB_WR_SEND;
1068 send_wr.send_flags = IB_SEND_SIGNALED;
1070 atomic_inc(&ib_conn->post_send_buf_count);
1072 ib_ret = ib_post_send(ib_conn->qp, &send_wr, &send_wr_failed);
1074 iser_err("ib_post_send failed, ret:%d\n", ib_ret);
1075 atomic_dec(&ib_conn->post_send_buf_count);
1080 static void iser_handle_comp_error(struct iser_tx_desc *desc,
1081 struct iser_conn *ib_conn)
1083 if (desc && desc->type == ISCSI_TX_DATAOUT)
1084 kmem_cache_free(ig.desc_cache, desc);
1086 if (ib_conn->post_recv_buf_count == 0 &&
1087 atomic_read(&ib_conn->post_send_buf_count) == 0) {
1089 * getting here when the state is UP means that the conn is
1090 * being terminated asynchronously from the iSCSI layer's
1091 * perspective. It is safe to peek at the connection state
1092 * since iscsi_conn_failure is allowed to be called twice.
1094 if (ib_conn->state == ISER_CONN_UP)
1095 iscsi_conn_failure(ib_conn->iscsi_conn,
1096 ISCSI_ERR_CONN_FAILED);
1098 /* no more non completed posts to the QP, complete the
1099 * termination process w.o worrying on disconnect event */
1100 complete(&ib_conn->flush_completion);
1104 static int iser_drain_tx_cq(struct iser_device *device, int cq_index)
1106 struct ib_cq *cq = device->tx_cq[cq_index];
1108 struct iser_tx_desc *tx_desc;
1109 struct iser_conn *ib_conn;
1110 int completed_tx = 0;
1112 while (ib_poll_cq(cq, 1, &wc) == 1) {
1113 tx_desc = (struct iser_tx_desc *) (unsigned long) wc.wr_id;
1114 ib_conn = wc.qp->qp_context;
1115 if (wc.status == IB_WC_SUCCESS) {
1116 if (wc.opcode == IB_WC_SEND)
1117 iser_snd_completion(tx_desc, ib_conn);
1119 iser_err("expected opcode %d got %d\n",
1120 IB_WC_SEND, wc.opcode);
1122 iser_err("tx id %llx status %d vend_err %x\n",
1123 wc.wr_id, wc.status, wc.vendor_err);
1124 if (wc.wr_id != ISER_FASTREG_LI_WRID) {
1125 atomic_dec(&ib_conn->post_send_buf_count);
1126 iser_handle_comp_error(tx_desc, ib_conn);
1131 return completed_tx;
1135 static void iser_cq_tasklet_fn(unsigned long data)
1137 struct iser_cq_desc *cq_desc = (struct iser_cq_desc *)data;
1138 struct iser_device *device = cq_desc->device;
1139 int cq_index = cq_desc->cq_index;
1140 struct ib_cq *cq = device->rx_cq[cq_index];
1142 struct iser_rx_desc *desc;
1143 unsigned long xfer_len;
1144 struct iser_conn *ib_conn;
1145 int completed_tx, completed_rx = 0;
1147 /* First do tx drain, so in a case where we have rx flushes and a successful
1148 * tx completion we will still go through completion error handling.
1150 completed_tx = iser_drain_tx_cq(device, cq_index);
1152 while (ib_poll_cq(cq, 1, &wc) == 1) {
1153 desc = (struct iser_rx_desc *) (unsigned long) wc.wr_id;
1154 BUG_ON(desc == NULL);
1155 ib_conn = wc.qp->qp_context;
1156 if (wc.status == IB_WC_SUCCESS) {
1157 if (wc.opcode == IB_WC_RECV) {
1158 xfer_len = (unsigned long)wc.byte_len;
1159 iser_rcv_completion(desc, xfer_len, ib_conn);
1161 iser_err("expected opcode %d got %d\n",
1162 IB_WC_RECV, wc.opcode);
1164 if (wc.status != IB_WC_WR_FLUSH_ERR)
1165 iser_err("rx id %llx status %d vend_err %x\n",
1166 wc.wr_id, wc.status, wc.vendor_err);
1167 ib_conn->post_recv_buf_count--;
1168 iser_handle_comp_error(NULL, ib_conn);
1171 if (!(completed_rx & 63))
1172 completed_tx += iser_drain_tx_cq(device, cq_index);
1174 /* #warning "it is assumed here that arming CQ only once its empty" *
1175 * " would not cause interrupts to be missed" */
1176 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1178 iser_dbg("got %d rx %d tx completions\n", completed_rx, completed_tx);
1181 static void iser_cq_callback(struct ib_cq *cq, void *cq_context)
1183 struct iser_cq_desc *cq_desc = (struct iser_cq_desc *)cq_context;
1184 struct iser_device *device = cq_desc->device;
1185 int cq_index = cq_desc->cq_index;
1187 tasklet_schedule(&device->cq_tasklet[cq_index]);
1190 u8 iser_check_task_pi_status(struct iscsi_iser_task *iser_task,
1191 enum iser_data_dir cmd_dir, sector_t *sector)
1193 struct iser_mem_reg *reg = &iser_task->rdma_regd[cmd_dir].reg;
1194 struct fast_reg_descriptor *desc = reg->mem_h;
1195 unsigned long sector_size = iser_task->sc->device->sector_size;
1196 struct ib_mr_status mr_status;
1199 if (desc && desc->reg_indicators & ISER_FASTREG_PROTECTED) {
1200 desc->reg_indicators &= ~ISER_FASTREG_PROTECTED;
1201 ret = ib_check_mr_status(desc->pi_ctx->sig_mr,
1202 IB_MR_CHECK_SIG_STATUS, &mr_status);
1204 pr_err("ib_check_mr_status failed, ret %d\n", ret);
1208 if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1209 sector_t sector_off = mr_status.sig_err.sig_err_offset;
1211 do_div(sector_off, sector_size + 8);
1212 *sector = scsi_get_lba(iser_task->sc) + sector_off;
1214 pr_err("PI error found type %d at sector %llx "
1215 "expected %x vs actual %x\n",
1216 mr_status.sig_err.err_type,
1217 (unsigned long long)*sector,
1218 mr_status.sig_err.expected,
1219 mr_status.sig_err.actual);
1221 switch (mr_status.sig_err.err_type) {
1222 case IB_SIG_BAD_GUARD:
1224 case IB_SIG_BAD_REFTAG:
1226 case IB_SIG_BAD_APPTAG:
1234 /* Not alot we can do here, return ambiguous guard error */