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])) {
129 device->rx_cq[i] = NULL;
133 device->tx_cq[i] = ib_create_cq(device->ib_device,
134 NULL, iser_cq_event_callback,
136 ISER_MAX_TX_CQ_LEN, i);
138 if (IS_ERR(device->tx_cq[i])) {
139 device->tx_cq[i] = NULL;
143 if (ib_req_notify_cq(device->rx_cq[i], IB_CQ_NEXT_COMP))
146 tasklet_init(&device->cq_tasklet[i],
148 (unsigned long)&cq_desc[i]);
151 device->mr = ib_get_dma_mr(device->pd, IB_ACCESS_LOCAL_WRITE |
152 IB_ACCESS_REMOTE_WRITE |
153 IB_ACCESS_REMOTE_READ);
154 if (IS_ERR(device->mr))
157 INIT_IB_EVENT_HANDLER(&device->event_handler, device->ib_device,
159 if (ib_register_event_handler(&device->event_handler))
165 ib_dereg_mr(device->mr);
167 for (i = 0; i < device->cqs_used; i++)
168 tasklet_kill(&device->cq_tasklet[i]);
170 for (i = 0; i < device->cqs_used; i++) {
171 if (device->tx_cq[i])
172 ib_destroy_cq(device->tx_cq[i]);
173 if (device->rx_cq[i])
174 ib_destroy_cq(device->rx_cq[i]);
176 ib_dealloc_pd(device->pd);
178 kfree(device->cq_desc);
180 iser_err("failed to allocate an IB resource\n");
185 * iser_free_device_ib_res - destroy/dealloc/dereg the DMA MR,
186 * CQ and PD created with the device associated with the adapator.
188 static void iser_free_device_ib_res(struct iser_device *device)
191 BUG_ON(device->mr == NULL);
193 for (i = 0; i < device->cqs_used; i++) {
194 tasklet_kill(&device->cq_tasklet[i]);
195 (void)ib_destroy_cq(device->tx_cq[i]);
196 (void)ib_destroy_cq(device->rx_cq[i]);
197 device->tx_cq[i] = NULL;
198 device->rx_cq[i] = NULL;
201 (void)ib_unregister_event_handler(&device->event_handler);
202 (void)ib_dereg_mr(device->mr);
203 (void)ib_dealloc_pd(device->pd);
205 kfree(device->cq_desc);
212 * iser_create_fmr_pool - Creates FMR pool and page_vector
214 * returns 0 on success, or errno code on failure
216 int iser_create_fmr_pool(struct iser_conn *ib_conn, unsigned cmds_max)
218 struct iser_device *device = ib_conn->device;
219 struct ib_fmr_pool_param params;
222 ib_conn->fmr.page_vec = kmalloc(sizeof(*ib_conn->fmr.page_vec) +
223 (sizeof(u64)*(ISCSI_ISER_SG_TABLESIZE + 1)),
225 if (!ib_conn->fmr.page_vec)
228 ib_conn->fmr.page_vec->pages = (u64 *)(ib_conn->fmr.page_vec + 1);
230 params.page_shift = SHIFT_4K;
231 /* when the first/last SG element are not start/end *
232 * page aligned, the map whould be of N+1 pages */
233 params.max_pages_per_fmr = ISCSI_ISER_SG_TABLESIZE + 1;
234 /* make the pool size twice the max number of SCSI commands *
235 * the ML is expected to queue, watermark for unmap at 50% */
236 params.pool_size = cmds_max * 2;
237 params.dirty_watermark = cmds_max;
239 params.flush_function = NULL;
240 params.access = (IB_ACCESS_LOCAL_WRITE |
241 IB_ACCESS_REMOTE_WRITE |
242 IB_ACCESS_REMOTE_READ);
244 ib_conn->fmr.pool = ib_create_fmr_pool(device->pd, ¶ms);
245 if (!IS_ERR(ib_conn->fmr.pool))
248 /* no FMR => no need for page_vec */
249 kfree(ib_conn->fmr.page_vec);
250 ib_conn->fmr.page_vec = NULL;
252 ret = PTR_ERR(ib_conn->fmr.pool);
253 ib_conn->fmr.pool = NULL;
254 if (ret != -ENOSYS) {
255 iser_err("FMR allocation failed, err %d\n", ret);
258 iser_warn("FMRs are not supported, using unaligned mode\n");
264 * iser_free_fmr_pool - releases the FMR pool and page vec
266 void iser_free_fmr_pool(struct iser_conn *ib_conn)
268 iser_info("freeing conn %p fmr pool %p\n",
269 ib_conn, ib_conn->fmr.pool);
271 if (ib_conn->fmr.pool != NULL)
272 ib_destroy_fmr_pool(ib_conn->fmr.pool);
274 ib_conn->fmr.pool = NULL;
276 kfree(ib_conn->fmr.page_vec);
277 ib_conn->fmr.page_vec = NULL;
281 iser_create_fastreg_desc(struct ib_device *ib_device, struct ib_pd *pd,
282 bool pi_enable, struct fast_reg_descriptor *desc)
286 desc->data_frpl = ib_alloc_fast_reg_page_list(ib_device,
287 ISCSI_ISER_SG_TABLESIZE + 1);
288 if (IS_ERR(desc->data_frpl)) {
289 ret = PTR_ERR(desc->data_frpl);
290 iser_err("Failed to allocate ib_fast_reg_page_list err=%d\n",
292 return PTR_ERR(desc->data_frpl);
295 desc->data_mr = ib_alloc_fast_reg_mr(pd, ISCSI_ISER_SG_TABLESIZE + 1);
296 if (IS_ERR(desc->data_mr)) {
297 ret = PTR_ERR(desc->data_mr);
298 iser_err("Failed to allocate ib_fast_reg_mr err=%d\n", ret);
299 goto fast_reg_mr_failure;
301 desc->reg_indicators |= ISER_DATA_KEY_VALID;
304 struct ib_mr_init_attr mr_init_attr = {0};
305 struct iser_pi_context *pi_ctx = NULL;
307 desc->pi_ctx = kzalloc(sizeof(*desc->pi_ctx), GFP_KERNEL);
309 iser_err("Failed to allocate pi context\n");
311 goto pi_ctx_alloc_failure;
313 pi_ctx = desc->pi_ctx;
315 pi_ctx->prot_frpl = ib_alloc_fast_reg_page_list(ib_device,
316 ISCSI_ISER_SG_TABLESIZE);
317 if (IS_ERR(pi_ctx->prot_frpl)) {
318 ret = PTR_ERR(pi_ctx->prot_frpl);
319 iser_err("Failed to allocate prot frpl ret=%d\n",
321 goto prot_frpl_failure;
324 pi_ctx->prot_mr = ib_alloc_fast_reg_mr(pd,
325 ISCSI_ISER_SG_TABLESIZE + 1);
326 if (IS_ERR(pi_ctx->prot_mr)) {
327 ret = PTR_ERR(pi_ctx->prot_mr);
328 iser_err("Failed to allocate prot frmr ret=%d\n",
330 goto prot_mr_failure;
332 desc->reg_indicators |= ISER_PROT_KEY_VALID;
334 mr_init_attr.max_reg_descriptors = 2;
335 mr_init_attr.flags |= IB_MR_SIGNATURE_EN;
336 pi_ctx->sig_mr = ib_create_mr(pd, &mr_init_attr);
337 if (IS_ERR(pi_ctx->sig_mr)) {
338 ret = PTR_ERR(pi_ctx->sig_mr);
339 iser_err("Failed to allocate signature enabled mr err=%d\n",
343 desc->reg_indicators |= ISER_SIG_KEY_VALID;
345 desc->reg_indicators &= ~ISER_FASTREG_PROTECTED;
347 iser_dbg("Create fr_desc %p page_list %p\n",
348 desc, desc->data_frpl->page_list);
352 ib_dereg_mr(desc->pi_ctx->prot_mr);
354 ib_free_fast_reg_page_list(desc->pi_ctx->prot_frpl);
357 pi_ctx_alloc_failure:
358 ib_dereg_mr(desc->data_mr);
360 ib_free_fast_reg_page_list(desc->data_frpl);
366 * iser_create_fastreg_pool - Creates pool of fast_reg descriptors
367 * for fast registration work requests.
368 * returns 0 on success, or errno code on failure
370 int iser_create_fastreg_pool(struct iser_conn *ib_conn, unsigned cmds_max)
372 struct iser_device *device = ib_conn->device;
373 struct fast_reg_descriptor *desc;
376 INIT_LIST_HEAD(&ib_conn->fastreg.pool);
377 ib_conn->fastreg.pool_size = 0;
378 for (i = 0; i < cmds_max; i++) {
379 desc = kzalloc(sizeof(*desc), GFP_KERNEL);
381 iser_err("Failed to allocate a new fast_reg descriptor\n");
386 ret = iser_create_fastreg_desc(device->ib_device, device->pd,
387 ib_conn->pi_support, desc);
389 iser_err("Failed to create fastreg descriptor err=%d\n",
395 list_add_tail(&desc->list, &ib_conn->fastreg.pool);
396 ib_conn->fastreg.pool_size++;
402 iser_free_fastreg_pool(ib_conn);
407 * iser_free_fastreg_pool - releases the pool of fast_reg descriptors
409 void iser_free_fastreg_pool(struct iser_conn *ib_conn)
411 struct fast_reg_descriptor *desc, *tmp;
414 if (list_empty(&ib_conn->fastreg.pool))
417 iser_info("freeing conn %p fr pool\n", ib_conn);
419 list_for_each_entry_safe(desc, tmp, &ib_conn->fastreg.pool, list) {
420 list_del(&desc->list);
421 ib_free_fast_reg_page_list(desc->data_frpl);
422 ib_dereg_mr(desc->data_mr);
424 ib_free_fast_reg_page_list(desc->pi_ctx->prot_frpl);
425 ib_dereg_mr(desc->pi_ctx->prot_mr);
426 ib_destroy_mr(desc->pi_ctx->sig_mr);
433 if (i < ib_conn->fastreg.pool_size)
434 iser_warn("pool still has %d regions registered\n",
435 ib_conn->fastreg.pool_size - i);
439 * iser_create_ib_conn_res - Queue-Pair (QP)
441 * returns 0 on success, -1 on failure
443 static int iser_create_ib_conn_res(struct iser_conn *ib_conn)
445 struct iser_device *device;
446 struct ib_qp_init_attr init_attr;
448 int index, min_index = 0;
450 BUG_ON(ib_conn->device == NULL);
452 device = ib_conn->device;
454 memset(&init_attr, 0, sizeof init_attr);
456 mutex_lock(&ig.connlist_mutex);
457 /* select the CQ with the minimal number of usages */
458 for (index = 0; index < device->cqs_used; index++)
459 if (device->cq_active_qps[index] <
460 device->cq_active_qps[min_index])
462 device->cq_active_qps[min_index]++;
463 mutex_unlock(&ig.connlist_mutex);
464 iser_info("cq index %d used for ib_conn %p\n", min_index, ib_conn);
466 init_attr.event_handler = iser_qp_event_callback;
467 init_attr.qp_context = (void *)ib_conn;
468 init_attr.send_cq = device->tx_cq[min_index];
469 init_attr.recv_cq = device->rx_cq[min_index];
470 init_attr.cap.max_recv_wr = ISER_QP_MAX_RECV_DTOS;
471 init_attr.cap.max_send_sge = 2;
472 init_attr.cap.max_recv_sge = 1;
473 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
474 init_attr.qp_type = IB_QPT_RC;
475 if (ib_conn->pi_support) {
476 init_attr.cap.max_send_wr = ISER_QP_SIG_MAX_REQ_DTOS;
477 init_attr.create_flags |= IB_QP_CREATE_SIGNATURE_EN;
479 init_attr.cap.max_send_wr = ISER_QP_MAX_REQ_DTOS;
482 ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
486 ib_conn->qp = ib_conn->cma_id->qp;
487 iser_info("setting conn %p cma_id %p qp %p\n",
488 ib_conn, ib_conn->cma_id,
489 ib_conn->cma_id->qp);
493 iser_err("unable to alloc mem or create resource, err %d\n", ret);
498 * releases the QP object
500 static void iser_free_ib_conn_res(struct iser_conn *ib_conn)
503 BUG_ON(ib_conn == NULL);
505 iser_info("freeing conn %p cma_id %p qp %p\n",
506 ib_conn, ib_conn->cma_id,
509 /* qp is created only once both addr & route are resolved */
511 if (ib_conn->qp != NULL) {
512 cq_index = ((struct iser_cq_desc *)ib_conn->qp->recv_cq->cq_context)->cq_index;
513 ib_conn->device->cq_active_qps[cq_index]--;
515 rdma_destroy_qp(ib_conn->cma_id);
522 * based on the resolved device node GUID see if there already allocated
523 * device for this device. If there's no such, create one.
526 struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
528 struct iser_device *device;
530 mutex_lock(&ig.device_list_mutex);
532 list_for_each_entry(device, &ig.device_list, ig_list)
533 /* find if there's a match using the node GUID */
534 if (device->ib_device->node_guid == cma_id->device->node_guid)
537 device = kzalloc(sizeof *device, GFP_KERNEL);
541 /* assign this device to the device */
542 device->ib_device = cma_id->device;
543 /* init the device and link it into ig device list */
544 if (iser_create_device_ib_res(device)) {
549 list_add(&device->ig_list, &ig.device_list);
554 mutex_unlock(&ig.device_list_mutex);
558 /* if there's no demand for this device, release it */
559 static void iser_device_try_release(struct iser_device *device)
561 mutex_lock(&ig.device_list_mutex);
563 iser_info("device %p refcount %d\n", device, device->refcount);
564 if (!device->refcount) {
565 iser_free_device_ib_res(device);
566 list_del(&device->ig_list);
569 mutex_unlock(&ig.device_list_mutex);
573 * Called with state mutex held
575 static int iser_conn_state_comp_exch(struct iser_conn *ib_conn,
576 enum iser_ib_conn_state comp,
577 enum iser_ib_conn_state exch)
581 if ((ret = (ib_conn->state == comp)))
582 ib_conn->state = exch;
586 void iser_release_work(struct work_struct *work)
588 struct iser_conn *ib_conn;
591 ib_conn = container_of(work, struct iser_conn, release_work);
593 /* wait for .conn_stop callback */
594 rc = wait_for_completion_timeout(&ib_conn->stop_completion, 30 * HZ);
597 /* wait for the qp`s post send and post receive buffers to empty */
598 rc = wait_for_completion_timeout(&ib_conn->flush_completion, 30 * HZ);
601 ib_conn->state = ISER_CONN_DOWN;
603 mutex_lock(&ib_conn->state_mutex);
604 ib_conn->state = ISER_CONN_DOWN;
605 mutex_unlock(&ib_conn->state_mutex);
607 iser_conn_release(ib_conn);
611 * Frees all conn objects and deallocs conn descriptor
613 void iser_conn_release(struct iser_conn *ib_conn)
615 struct iser_device *device = ib_conn->device;
617 mutex_lock(&ig.connlist_mutex);
618 list_del(&ib_conn->conn_list);
619 mutex_unlock(&ig.connlist_mutex);
621 mutex_lock(&ib_conn->state_mutex);
622 BUG_ON(ib_conn->state != ISER_CONN_DOWN);
624 iser_free_rx_descriptors(ib_conn);
625 iser_free_ib_conn_res(ib_conn);
626 ib_conn->device = NULL;
627 /* on EVENT_ADDR_ERROR there's no device yet for this conn */
629 iser_device_try_release(device);
630 mutex_unlock(&ib_conn->state_mutex);
632 /* if cma handler context, the caller actually destroy the id */
633 if (ib_conn->cma_id != NULL) {
634 rdma_destroy_id(ib_conn->cma_id);
635 ib_conn->cma_id = NULL;
641 * triggers start of the disconnect procedures and wait for them to be done
643 void iser_conn_terminate(struct iser_conn *ib_conn)
647 /* change the ib conn state only if the conn is UP, however always call
648 * rdma_disconnect since this is the only way to cause the CMA to change
649 * the QP state to ERROR
652 iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP, ISER_CONN_TERMINATING);
653 err = rdma_disconnect(ib_conn->cma_id);
655 iser_err("Failed to disconnect, conn: 0x%p err %d\n",
660 * Called with state mutex held
662 static void iser_connect_error(struct rdma_cm_id *cma_id)
664 struct iser_conn *ib_conn;
666 ib_conn = (struct iser_conn *)cma_id->context;
667 ib_conn->state = ISER_CONN_DOWN;
671 * Called with state mutex held
673 static void iser_addr_handler(struct rdma_cm_id *cma_id)
675 struct iser_device *device;
676 struct iser_conn *ib_conn;
679 ib_conn = (struct iser_conn *)cma_id->context;
680 if (ib_conn->state != ISER_CONN_PENDING)
684 device = iser_device_find_by_ib_device(cma_id);
686 iser_err("device lookup/creation failed\n");
687 iser_connect_error(cma_id);
691 ib_conn->device = device;
693 /* connection T10-PI support */
694 if (iser_pi_enable) {
695 if (!(device->dev_attr.device_cap_flags &
696 IB_DEVICE_SIGNATURE_HANDOVER)) {
697 iser_warn("T10-PI requested but not supported on %s, "
698 "continue without T10-PI\n",
699 ib_conn->device->ib_device->name);
700 ib_conn->pi_support = false;
702 ib_conn->pi_support = true;
706 ret = rdma_resolve_route(cma_id, 1000);
708 iser_err("resolve route failed: %d\n", ret);
709 iser_connect_error(cma_id);
715 * Called with state mutex held
717 static void iser_route_handler(struct rdma_cm_id *cma_id)
719 struct rdma_conn_param conn_param;
721 struct iser_cm_hdr req_hdr;
722 struct iser_conn *ib_conn = (struct iser_conn *)cma_id->context;
723 struct iser_device *device = ib_conn->device;
725 if (ib_conn->state != ISER_CONN_PENDING)
729 ret = iser_create_ib_conn_res((struct iser_conn *)cma_id->context);
733 memset(&conn_param, 0, sizeof conn_param);
734 conn_param.responder_resources = device->dev_attr.max_qp_rd_atom;
735 conn_param.initiator_depth = 1;
736 conn_param.retry_count = 7;
737 conn_param.rnr_retry_count = 6;
739 memset(&req_hdr, 0, sizeof(req_hdr));
740 req_hdr.flags = (ISER_ZBVA_NOT_SUPPORTED |
741 ISER_SEND_W_INV_NOT_SUPPORTED);
742 conn_param.private_data = (void *)&req_hdr;
743 conn_param.private_data_len = sizeof(struct iser_cm_hdr);
745 ret = rdma_connect(cma_id, &conn_param);
747 iser_err("failure connecting: %d\n", ret);
753 iser_connect_error(cma_id);
756 static void iser_connected_handler(struct rdma_cm_id *cma_id)
758 struct iser_conn *ib_conn;
759 struct ib_qp_attr attr;
760 struct ib_qp_init_attr init_attr;
762 ib_conn = (struct iser_conn *)cma_id->context;
763 if (ib_conn->state != ISER_CONN_PENDING)
767 (void)ib_query_qp(cma_id->qp, &attr, ~0, &init_attr);
768 iser_info("remote qpn:%x my qpn:%x\n", attr.dest_qp_num, cma_id->qp->qp_num);
770 ib_conn->state = ISER_CONN_UP;
771 complete(&ib_conn->up_completion);
774 static void iser_disconnected_handler(struct rdma_cm_id *cma_id)
776 struct iser_conn *ib_conn;
778 ib_conn = (struct iser_conn *)cma_id->context;
780 /* getting here when the state is UP means that the conn is being *
781 * terminated asynchronously from the iSCSI layer's perspective. */
782 if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
783 ISER_CONN_TERMINATING)){
784 if (ib_conn->iscsi_conn)
785 iscsi_conn_failure(ib_conn->iscsi_conn, ISCSI_ERR_CONN_FAILED);
787 iser_err("iscsi_iser connection isn't bound\n");
790 /* Complete the termination process if no posts are pending. This code
791 * block also exists in iser_handle_comp_error(), but it is needed here
792 * for cases of no flushes at all, e.g. discovery over rdma.
794 if (ib_conn->post_recv_buf_count == 0 &&
795 (atomic_read(&ib_conn->post_send_buf_count) == 0)) {
796 complete(&ib_conn->flush_completion);
800 static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
802 struct iser_conn *ib_conn;
804 ib_conn = (struct iser_conn *)cma_id->context;
805 iser_info("event %d status %d conn %p id %p\n",
806 event->event, event->status, cma_id->context, cma_id);
808 mutex_lock(&ib_conn->state_mutex);
809 switch (event->event) {
810 case RDMA_CM_EVENT_ADDR_RESOLVED:
811 iser_addr_handler(cma_id);
813 case RDMA_CM_EVENT_ROUTE_RESOLVED:
814 iser_route_handler(cma_id);
816 case RDMA_CM_EVENT_ESTABLISHED:
817 iser_connected_handler(cma_id);
819 case RDMA_CM_EVENT_ADDR_ERROR:
820 case RDMA_CM_EVENT_ROUTE_ERROR:
821 case RDMA_CM_EVENT_CONNECT_ERROR:
822 case RDMA_CM_EVENT_UNREACHABLE:
823 case RDMA_CM_EVENT_REJECTED:
824 iser_connect_error(cma_id);
826 case RDMA_CM_EVENT_DISCONNECTED:
827 case RDMA_CM_EVENT_DEVICE_REMOVAL:
828 case RDMA_CM_EVENT_ADDR_CHANGE:
829 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
830 iser_disconnected_handler(cma_id);
833 iser_err("Unexpected RDMA CM event (%d)\n", event->event);
836 mutex_unlock(&ib_conn->state_mutex);
840 void iser_conn_init(struct iser_conn *ib_conn)
842 ib_conn->state = ISER_CONN_INIT;
843 ib_conn->post_recv_buf_count = 0;
844 atomic_set(&ib_conn->post_send_buf_count, 0);
845 init_completion(&ib_conn->stop_completion);
846 init_completion(&ib_conn->flush_completion);
847 init_completion(&ib_conn->up_completion);
848 INIT_LIST_HEAD(&ib_conn->conn_list);
849 spin_lock_init(&ib_conn->lock);
850 mutex_init(&ib_conn->state_mutex);
854 * starts the process of connecting to the target
855 * sleeps until the connection is established or rejected
857 int iser_connect(struct iser_conn *ib_conn,
858 struct sockaddr *src_addr,
859 struct sockaddr *dst_addr,
864 mutex_lock(&ib_conn->state_mutex);
866 sprintf(ib_conn->name, "%pISp", dst_addr);
868 iser_info("connecting to: %s\n", ib_conn->name);
870 /* the device is known only --after-- address resolution */
871 ib_conn->device = NULL;
873 ib_conn->state = ISER_CONN_PENDING;
875 ib_conn->cma_id = rdma_create_id(iser_cma_handler,
877 RDMA_PS_TCP, IB_QPT_RC);
878 if (IS_ERR(ib_conn->cma_id)) {
879 err = PTR_ERR(ib_conn->cma_id);
880 iser_err("rdma_create_id failed: %d\n", err);
884 err = rdma_resolve_addr(ib_conn->cma_id, src_addr, dst_addr, 1000);
886 iser_err("rdma_resolve_addr failed: %d\n", err);
891 wait_for_completion_interruptible(&ib_conn->up_completion);
893 if (ib_conn->state != ISER_CONN_UP) {
895 goto connect_failure;
898 mutex_unlock(&ib_conn->state_mutex);
900 mutex_lock(&ig.connlist_mutex);
901 list_add(&ib_conn->conn_list, &ig.connlist);
902 mutex_unlock(&ig.connlist_mutex);
906 ib_conn->cma_id = NULL;
908 ib_conn->state = ISER_CONN_DOWN;
910 mutex_unlock(&ib_conn->state_mutex);
911 iser_conn_release(ib_conn);
916 * iser_reg_page_vec - Register physical memory
918 * returns: 0 on success, errno code on failure
920 int iser_reg_page_vec(struct iser_conn *ib_conn,
921 struct iser_page_vec *page_vec,
922 struct iser_mem_reg *mem_reg)
924 struct ib_pool_fmr *mem;
929 page_list = page_vec->pages;
930 io_addr = page_list[0];
932 mem = ib_fmr_pool_map_phys(ib_conn->fmr.pool,
938 status = (int)PTR_ERR(mem);
939 iser_err("ib_fmr_pool_map_phys failed: %d\n", status);
943 mem_reg->lkey = mem->fmr->lkey;
944 mem_reg->rkey = mem->fmr->rkey;
945 mem_reg->len = page_vec->length * SIZE_4K;
946 mem_reg->va = io_addr;
948 mem_reg->mem_h = (void *)mem;
950 mem_reg->va += page_vec->offset;
951 mem_reg->len = page_vec->data_size;
953 iser_dbg("PHYSICAL Mem.register, [PHYS p_array: 0x%p, sz: %d, "
954 "entry[0]: (0x%08lx,%ld)] -> "
955 "[lkey: 0x%08X mem_h: 0x%p va: 0x%08lX sz: %ld]\n",
956 page_vec, page_vec->length,
957 (unsigned long)page_vec->pages[0],
958 (unsigned long)page_vec->data_size,
959 (unsigned int)mem_reg->lkey, mem_reg->mem_h,
960 (unsigned long)mem_reg->va, (unsigned long)mem_reg->len);
965 * Unregister (previosuly registered using FMR) memory.
966 * If memory is non-FMR does nothing.
968 void iser_unreg_mem_fmr(struct iscsi_iser_task *iser_task,
969 enum iser_data_dir cmd_dir)
971 struct iser_mem_reg *reg = &iser_task->rdma_regd[cmd_dir].reg;
977 iser_dbg("PHYSICAL Mem.Unregister mem_h %p\n",reg->mem_h);
979 ret = ib_fmr_pool_unmap((struct ib_pool_fmr *)reg->mem_h);
981 iser_err("ib_fmr_pool_unmap failed %d\n", ret);
986 void iser_unreg_mem_fastreg(struct iscsi_iser_task *iser_task,
987 enum iser_data_dir cmd_dir)
989 struct iser_mem_reg *reg = &iser_task->rdma_regd[cmd_dir].reg;
990 struct iser_conn *ib_conn = iser_task->ib_conn;
991 struct fast_reg_descriptor *desc = reg->mem_h;
998 spin_lock_bh(&ib_conn->lock);
999 list_add_tail(&desc->list, &ib_conn->fastreg.pool);
1000 spin_unlock_bh(&ib_conn->lock);
1003 int iser_post_recvl(struct iser_conn *ib_conn)
1005 struct ib_recv_wr rx_wr, *rx_wr_failed;
1009 sge.addr = ib_conn->login_resp_dma;
1010 sge.length = ISER_RX_LOGIN_SIZE;
1011 sge.lkey = ib_conn->device->mr->lkey;
1013 rx_wr.wr_id = (unsigned long)ib_conn->login_resp_buf;
1014 rx_wr.sg_list = &sge;
1018 ib_conn->post_recv_buf_count++;
1019 ib_ret = ib_post_recv(ib_conn->qp, &rx_wr, &rx_wr_failed);
1021 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1022 ib_conn->post_recv_buf_count--;
1027 int iser_post_recvm(struct iser_conn *ib_conn, int count)
1029 struct ib_recv_wr *rx_wr, *rx_wr_failed;
1031 unsigned int my_rx_head = ib_conn->rx_desc_head;
1032 struct iser_rx_desc *rx_desc;
1034 for (rx_wr = ib_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
1035 rx_desc = &ib_conn->rx_descs[my_rx_head];
1036 rx_wr->wr_id = (unsigned long)rx_desc;
1037 rx_wr->sg_list = &rx_desc->rx_sg;
1039 rx_wr->next = rx_wr + 1;
1040 my_rx_head = (my_rx_head + 1) & ib_conn->qp_max_recv_dtos_mask;
1044 rx_wr->next = NULL; /* mark end of work requests list */
1046 ib_conn->post_recv_buf_count += count;
1047 ib_ret = ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &rx_wr_failed);
1049 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1050 ib_conn->post_recv_buf_count -= count;
1052 ib_conn->rx_desc_head = my_rx_head;
1058 * iser_start_send - Initiate a Send DTO operation
1060 * returns 0 on success, -1 on failure
1062 int iser_post_send(struct iser_conn *ib_conn, struct iser_tx_desc *tx_desc)
1065 struct ib_send_wr send_wr, *send_wr_failed;
1067 ib_dma_sync_single_for_device(ib_conn->device->ib_device,
1068 tx_desc->dma_addr, ISER_HEADERS_LEN, DMA_TO_DEVICE);
1070 send_wr.next = NULL;
1071 send_wr.wr_id = (unsigned long)tx_desc;
1072 send_wr.sg_list = tx_desc->tx_sg;
1073 send_wr.num_sge = tx_desc->num_sge;
1074 send_wr.opcode = IB_WR_SEND;
1075 send_wr.send_flags = IB_SEND_SIGNALED;
1077 atomic_inc(&ib_conn->post_send_buf_count);
1079 ib_ret = ib_post_send(ib_conn->qp, &send_wr, &send_wr_failed);
1081 iser_err("ib_post_send failed, ret:%d\n", ib_ret);
1082 atomic_dec(&ib_conn->post_send_buf_count);
1087 static void iser_handle_comp_error(struct iser_tx_desc *desc,
1088 struct iser_conn *ib_conn)
1090 if (desc && desc->type == ISCSI_TX_DATAOUT)
1091 kmem_cache_free(ig.desc_cache, desc);
1093 if (ib_conn->post_recv_buf_count == 0 &&
1094 atomic_read(&ib_conn->post_send_buf_count) == 0) {
1096 * getting here when the state is UP means that the conn is
1097 * being terminated asynchronously from the iSCSI layer's
1098 * perspective. It is safe to peek at the connection state
1099 * since iscsi_conn_failure is allowed to be called twice.
1101 if (ib_conn->state == ISER_CONN_UP)
1102 iscsi_conn_failure(ib_conn->iscsi_conn,
1103 ISCSI_ERR_CONN_FAILED);
1105 /* no more non completed posts to the QP, complete the
1106 * termination process w.o worrying on disconnect event */
1107 complete(&ib_conn->flush_completion);
1111 static int iser_drain_tx_cq(struct iser_device *device, int cq_index)
1113 struct ib_cq *cq = device->tx_cq[cq_index];
1115 struct iser_tx_desc *tx_desc;
1116 struct iser_conn *ib_conn;
1117 int completed_tx = 0;
1119 while (ib_poll_cq(cq, 1, &wc) == 1) {
1120 tx_desc = (struct iser_tx_desc *) (unsigned long) wc.wr_id;
1121 ib_conn = wc.qp->qp_context;
1122 if (wc.status == IB_WC_SUCCESS) {
1123 if (wc.opcode == IB_WC_SEND)
1124 iser_snd_completion(tx_desc, ib_conn);
1126 iser_err("expected opcode %d got %d\n",
1127 IB_WC_SEND, wc.opcode);
1129 iser_err("tx id %llx status %d vend_err %x\n",
1130 wc.wr_id, wc.status, wc.vendor_err);
1131 if (wc.wr_id != ISER_FASTREG_LI_WRID) {
1132 atomic_dec(&ib_conn->post_send_buf_count);
1133 iser_handle_comp_error(tx_desc, ib_conn);
1138 return completed_tx;
1142 static void iser_cq_tasklet_fn(unsigned long data)
1144 struct iser_cq_desc *cq_desc = (struct iser_cq_desc *)data;
1145 struct iser_device *device = cq_desc->device;
1146 int cq_index = cq_desc->cq_index;
1147 struct ib_cq *cq = device->rx_cq[cq_index];
1149 struct iser_rx_desc *desc;
1150 unsigned long xfer_len;
1151 struct iser_conn *ib_conn;
1152 int completed_tx, completed_rx = 0;
1154 /* First do tx drain, so in a case where we have rx flushes and a successful
1155 * tx completion we will still go through completion error handling.
1157 completed_tx = iser_drain_tx_cq(device, cq_index);
1159 while (ib_poll_cq(cq, 1, &wc) == 1) {
1160 desc = (struct iser_rx_desc *) (unsigned long) wc.wr_id;
1161 BUG_ON(desc == NULL);
1162 ib_conn = wc.qp->qp_context;
1163 if (wc.status == IB_WC_SUCCESS) {
1164 if (wc.opcode == IB_WC_RECV) {
1165 xfer_len = (unsigned long)wc.byte_len;
1166 iser_rcv_completion(desc, xfer_len, ib_conn);
1168 iser_err("expected opcode %d got %d\n",
1169 IB_WC_RECV, wc.opcode);
1171 if (wc.status != IB_WC_WR_FLUSH_ERR)
1172 iser_err("rx id %llx status %d vend_err %x\n",
1173 wc.wr_id, wc.status, wc.vendor_err);
1174 ib_conn->post_recv_buf_count--;
1175 iser_handle_comp_error(NULL, ib_conn);
1178 if (!(completed_rx & 63))
1179 completed_tx += iser_drain_tx_cq(device, cq_index);
1181 /* #warning "it is assumed here that arming CQ only once its empty" *
1182 * " would not cause interrupts to be missed" */
1183 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1185 iser_dbg("got %d rx %d tx completions\n", completed_rx, completed_tx);
1188 static void iser_cq_callback(struct ib_cq *cq, void *cq_context)
1190 struct iser_cq_desc *cq_desc = (struct iser_cq_desc *)cq_context;
1191 struct iser_device *device = cq_desc->device;
1192 int cq_index = cq_desc->cq_index;
1194 tasklet_schedule(&device->cq_tasklet[cq_index]);
1197 u8 iser_check_task_pi_status(struct iscsi_iser_task *iser_task,
1198 enum iser_data_dir cmd_dir, sector_t *sector)
1200 struct iser_mem_reg *reg = &iser_task->rdma_regd[cmd_dir].reg;
1201 struct fast_reg_descriptor *desc = reg->mem_h;
1202 unsigned long sector_size = iser_task->sc->device->sector_size;
1203 struct ib_mr_status mr_status;
1206 if (desc && desc->reg_indicators & ISER_FASTREG_PROTECTED) {
1207 desc->reg_indicators &= ~ISER_FASTREG_PROTECTED;
1208 ret = ib_check_mr_status(desc->pi_ctx->sig_mr,
1209 IB_MR_CHECK_SIG_STATUS, &mr_status);
1211 pr_err("ib_check_mr_status failed, ret %d\n", ret);
1215 if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1216 sector_t sector_off = mr_status.sig_err.sig_err_offset;
1218 do_div(sector_off, sector_size + 8);
1219 *sector = scsi_get_lba(iser_task->sc) + sector_off;
1221 pr_err("PI error found type %d at sector %llx "
1222 "expected %x vs actual %x\n",
1223 mr_status.sig_err.err_type,
1224 (unsigned long long)*sector,
1225 mr_status.sig_err.expected,
1226 mr_status.sig_err.actual);
1228 switch (mr_status.sig_err.err_type) {
1229 case IB_SIG_BAD_GUARD:
1231 case IB_SIG_BAD_REFTAG:
1233 case IB_SIG_BAD_APPTAG:
1241 /* Not alot we can do here, return ambiguous guard error */