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 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)
75 struct iser_cq_desc *cq_desc;
76 struct ib_device_attr *dev_attr;
78 dev_attr = kmalloc(sizeof(*dev_attr), GFP_KERNEL);
82 if (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("FRWR supported, using FRWR for registration\n");
98 device->iser_alloc_rdma_reg_res = iser_create_frwr_pool;
99 device->iser_free_rdma_reg_res = iser_free_frwr_pool;
100 device->iser_reg_rdma_mem = iser_reg_rdma_mem_frwr;
101 device->iser_unreg_rdma_mem = iser_unreg_mem_frwr;
103 iser_err("IB device does not support FMRs nor FRWRs, can't register memory\n");
107 device->cqs_used = min(ISER_MAX_CQ, device->ib_device->num_comp_vectors);
108 iser_info("using %d CQs, device %s supports %d vectors\n",
109 device->cqs_used, device->ib_device->name,
110 device->ib_device->num_comp_vectors);
112 device->cq_desc = kmalloc(sizeof(struct iser_cq_desc) * device->cqs_used,
114 if (device->cq_desc == NULL)
116 cq_desc = device->cq_desc;
118 device->pd = ib_alloc_pd(device->ib_device);
119 if (IS_ERR(device->pd))
122 for (i = 0; i < device->cqs_used; i++) {
123 cq_desc[i].device = device;
124 cq_desc[i].cq_index = i;
126 device->rx_cq[i] = ib_create_cq(device->ib_device,
128 iser_cq_event_callback,
130 ISER_MAX_RX_CQ_LEN, i);
131 if (IS_ERR(device->rx_cq[i]))
134 device->tx_cq[i] = ib_create_cq(device->ib_device,
135 NULL, iser_cq_event_callback,
137 ISER_MAX_TX_CQ_LEN, i);
139 if (IS_ERR(device->tx_cq[i]))
142 if (ib_req_notify_cq(device->rx_cq[i], IB_CQ_NEXT_COMP))
145 tasklet_init(&device->cq_tasklet[i],
147 (unsigned long)&cq_desc[i]);
150 device->mr = ib_get_dma_mr(device->pd, IB_ACCESS_LOCAL_WRITE |
151 IB_ACCESS_REMOTE_WRITE |
152 IB_ACCESS_REMOTE_READ);
153 if (IS_ERR(device->mr))
156 INIT_IB_EVENT_HANDLER(&device->event_handler, device->ib_device,
158 if (ib_register_event_handler(&device->event_handler))
165 ib_dereg_mr(device->mr);
167 for (j = 0; j < device->cqs_used; j++)
168 tasklet_kill(&device->cq_tasklet[j]);
170 for (j = 0; j < i; j++) {
171 if (device->tx_cq[j])
172 ib_destroy_cq(device->tx_cq[j]);
173 if (device->rx_cq[j])
174 ib_destroy_cq(device->rx_cq[j]);
176 ib_dealloc_pd(device->pd);
178 kfree(device->cq_desc);
180 iser_err("failed to allocate an IB resource\n");
187 * iser_free_device_ib_res - destroy/dealloc/dereg the DMA MR,
188 * CQ and PD created with the device associated with the adapator.
190 static void iser_free_device_ib_res(struct iser_device *device)
193 BUG_ON(device->mr == NULL);
195 for (i = 0; i < device->cqs_used; i++) {
196 tasklet_kill(&device->cq_tasklet[i]);
197 (void)ib_destroy_cq(device->tx_cq[i]);
198 (void)ib_destroy_cq(device->rx_cq[i]);
199 device->tx_cq[i] = NULL;
200 device->rx_cq[i] = NULL;
203 (void)ib_unregister_event_handler(&device->event_handler);
204 (void)ib_dereg_mr(device->mr);
205 (void)ib_dealloc_pd(device->pd);
207 kfree(device->cq_desc);
214 * iser_create_fmr_pool - Creates FMR pool and page_vector
216 * returns 0 on success, or errno code on failure
218 int iser_create_fmr_pool(struct iser_conn *ib_conn, unsigned cmds_max)
220 struct iser_device *device = ib_conn->device;
221 struct ib_fmr_pool_param params;
224 ib_conn->fastreg.fmr.page_vec = kmalloc(sizeof(struct iser_page_vec) +
225 (sizeof(u64)*(ISCSI_ISER_SG_TABLESIZE + 1)),
227 if (!ib_conn->fastreg.fmr.page_vec)
230 ib_conn->fastreg.fmr.page_vec->pages = (u64 *)(ib_conn->fastreg.fmr.page_vec + 1);
232 params.page_shift = SHIFT_4K;
233 /* when the first/last SG element are not start/end *
234 * page aligned, the map whould be of N+1 pages */
235 params.max_pages_per_fmr = ISCSI_ISER_SG_TABLESIZE + 1;
236 /* make the pool size twice the max number of SCSI commands *
237 * the ML is expected to queue, watermark for unmap at 50% */
238 params.pool_size = cmds_max * 2;
239 params.dirty_watermark = cmds_max;
241 params.flush_function = NULL;
242 params.access = (IB_ACCESS_LOCAL_WRITE |
243 IB_ACCESS_REMOTE_WRITE |
244 IB_ACCESS_REMOTE_READ);
246 ib_conn->fastreg.fmr.pool = ib_create_fmr_pool(device->pd, ¶ms);
247 if (!IS_ERR(ib_conn->fastreg.fmr.pool))
250 /* no FMR => no need for page_vec */
251 kfree(ib_conn->fastreg.fmr.page_vec);
252 ib_conn->fastreg.fmr.page_vec = NULL;
254 ret = PTR_ERR(ib_conn->fastreg.fmr.pool);
255 ib_conn->fastreg.fmr.pool = NULL;
256 if (ret != -ENOSYS) {
257 iser_err("FMR allocation failed, err %d\n", ret);
260 iser_warn("FMRs are not supported, using unaligned mode\n");
266 * iser_free_fmr_pool - releases the FMR pool and page vec
268 void iser_free_fmr_pool(struct iser_conn *ib_conn)
270 iser_info("freeing conn %p fmr pool %p\n",
271 ib_conn, ib_conn->fastreg.fmr.pool);
273 if (ib_conn->fastreg.fmr.pool != NULL)
274 ib_destroy_fmr_pool(ib_conn->fastreg.fmr.pool);
276 ib_conn->fastreg.fmr.pool = NULL;
278 kfree(ib_conn->fastreg.fmr.page_vec);
279 ib_conn->fastreg.fmr.page_vec = NULL;
283 * iser_create_frwr_pool - Creates pool of fast_reg descriptors
284 * for fast registration work requests.
285 * returns 0 on success, or errno code on failure
287 int iser_create_frwr_pool(struct iser_conn *ib_conn, unsigned cmds_max)
289 struct iser_device *device = ib_conn->device;
290 struct fast_reg_descriptor *desc;
293 INIT_LIST_HEAD(&ib_conn->fastreg.frwr.pool);
294 ib_conn->fastreg.frwr.pool_size = 0;
295 for (i = 0; i < cmds_max; i++) {
296 desc = kmalloc(sizeof(*desc), GFP_KERNEL);
298 iser_err("Failed to allocate a new fast_reg descriptor\n");
303 desc->data_frpl = ib_alloc_fast_reg_page_list(device->ib_device,
304 ISCSI_ISER_SG_TABLESIZE + 1);
305 if (IS_ERR(desc->data_frpl)) {
306 ret = PTR_ERR(desc->data_frpl);
307 iser_err("Failed to allocate ib_fast_reg_page_list err=%d\n", ret);
308 goto fast_reg_page_failure;
311 desc->data_mr = ib_alloc_fast_reg_mr(device->pd,
312 ISCSI_ISER_SG_TABLESIZE + 1);
313 if (IS_ERR(desc->data_mr)) {
314 ret = PTR_ERR(desc->data_mr);
315 iser_err("Failed to allocate ib_fast_reg_mr err=%d\n", ret);
316 goto fast_reg_mr_failure;
319 list_add_tail(&desc->list, &ib_conn->fastreg.frwr.pool);
320 ib_conn->fastreg.frwr.pool_size++;
326 ib_free_fast_reg_page_list(desc->data_frpl);
327 fast_reg_page_failure:
330 iser_free_frwr_pool(ib_conn);
335 * iser_free_frwr_pool - releases the pool of fast_reg descriptors
337 void iser_free_frwr_pool(struct iser_conn *ib_conn)
339 struct fast_reg_descriptor *desc, *tmp;
342 if (list_empty(&ib_conn->fastreg.frwr.pool))
345 iser_info("freeing conn %p frwr pool\n", ib_conn);
347 list_for_each_entry_safe(desc, tmp, &ib_conn->fastreg.frwr.pool, list) {
348 list_del(&desc->list);
349 ib_free_fast_reg_page_list(desc->data_frpl);
350 ib_dereg_mr(desc->data_mr);
355 if (i < ib_conn->fastreg.frwr.pool_size)
356 iser_warn("pool still has %d regions registered\n",
357 ib_conn->fastreg.frwr.pool_size - i);
361 * iser_create_ib_conn_res - Queue-Pair (QP)
363 * returns 0 on success, -1 on failure
365 static int iser_create_ib_conn_res(struct iser_conn *ib_conn)
367 struct iser_device *device;
368 struct ib_qp_init_attr init_attr;
370 int index, min_index = 0;
372 BUG_ON(ib_conn->device == NULL);
374 device = ib_conn->device;
376 memset(&init_attr, 0, sizeof init_attr);
378 mutex_lock(&ig.connlist_mutex);
379 /* select the CQ with the minimal number of usages */
380 for (index = 0; index < device->cqs_used; index++)
381 if (device->cq_active_qps[index] <
382 device->cq_active_qps[min_index])
384 device->cq_active_qps[min_index]++;
385 mutex_unlock(&ig.connlist_mutex);
386 iser_info("cq index %d used for ib_conn %p\n", min_index, ib_conn);
388 init_attr.event_handler = iser_qp_event_callback;
389 init_attr.qp_context = (void *)ib_conn;
390 init_attr.send_cq = device->tx_cq[min_index];
391 init_attr.recv_cq = device->rx_cq[min_index];
392 init_attr.cap.max_send_wr = ISER_QP_MAX_REQ_DTOS;
393 init_attr.cap.max_recv_wr = ISER_QP_MAX_RECV_DTOS;
394 init_attr.cap.max_send_sge = 2;
395 init_attr.cap.max_recv_sge = 1;
396 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
397 init_attr.qp_type = IB_QPT_RC;
399 ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
403 ib_conn->qp = ib_conn->cma_id->qp;
404 iser_info("setting conn %p cma_id %p qp %p\n",
405 ib_conn, ib_conn->cma_id,
406 ib_conn->cma_id->qp);
410 iser_err("unable to alloc mem or create resource, err %d\n", ret);
415 * releases the QP objects, returns 0 on success,
418 static int iser_free_ib_conn_res(struct iser_conn *ib_conn)
421 BUG_ON(ib_conn == NULL);
423 iser_info("freeing conn %p cma_id %p qp %p\n",
424 ib_conn, ib_conn->cma_id,
427 /* qp is created only once both addr & route are resolved */
429 if (ib_conn->qp != NULL) {
430 cq_index = ((struct iser_cq_desc *)ib_conn->qp->recv_cq->cq_context)->cq_index;
431 ib_conn->device->cq_active_qps[cq_index]--;
433 rdma_destroy_qp(ib_conn->cma_id);
442 * based on the resolved device node GUID see if there already allocated
443 * device for this device. If there's no such, create one.
446 struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
448 struct iser_device *device;
450 mutex_lock(&ig.device_list_mutex);
452 list_for_each_entry(device, &ig.device_list, ig_list)
453 /* find if there's a match using the node GUID */
454 if (device->ib_device->node_guid == cma_id->device->node_guid)
457 device = kzalloc(sizeof *device, GFP_KERNEL);
461 /* assign this device to the device */
462 device->ib_device = cma_id->device;
463 /* init the device and link it into ig device list */
464 if (iser_create_device_ib_res(device)) {
469 list_add(&device->ig_list, &ig.device_list);
474 mutex_unlock(&ig.device_list_mutex);
478 /* if there's no demand for this device, release it */
479 static void iser_device_try_release(struct iser_device *device)
481 mutex_lock(&ig.device_list_mutex);
483 iser_info("device %p refcount %d\n", device, device->refcount);
484 if (!device->refcount) {
485 iser_free_device_ib_res(device);
486 list_del(&device->ig_list);
489 mutex_unlock(&ig.device_list_mutex);
492 static int iser_conn_state_comp_exch(struct iser_conn *ib_conn,
493 enum iser_ib_conn_state comp,
494 enum iser_ib_conn_state exch)
498 spin_lock_bh(&ib_conn->lock);
499 if ((ret = (ib_conn->state == comp)))
500 ib_conn->state = exch;
501 spin_unlock_bh(&ib_conn->lock);
506 * Frees all conn objects and deallocs conn descriptor
508 static void iser_conn_release(struct iser_conn *ib_conn, int can_destroy_id)
510 struct iser_device *device = ib_conn->device;
512 BUG_ON(ib_conn->state != ISER_CONN_DOWN);
514 mutex_lock(&ig.connlist_mutex);
515 list_del(&ib_conn->conn_list);
516 mutex_unlock(&ig.connlist_mutex);
517 iser_free_rx_descriptors(ib_conn);
518 iser_free_ib_conn_res(ib_conn);
519 ib_conn->device = NULL;
520 /* on EVENT_ADDR_ERROR there's no device yet for this conn */
522 iser_device_try_release(device);
523 /* if cma handler context, the caller actually destroy the id */
524 if (ib_conn->cma_id != NULL && can_destroy_id) {
525 rdma_destroy_id(ib_conn->cma_id);
526 ib_conn->cma_id = NULL;
528 iscsi_destroy_endpoint(ib_conn->ep);
531 void iser_conn_get(struct iser_conn *ib_conn)
533 atomic_inc(&ib_conn->refcount);
536 int iser_conn_put(struct iser_conn *ib_conn, int can_destroy_id)
538 if (atomic_dec_and_test(&ib_conn->refcount)) {
539 iser_conn_release(ib_conn, can_destroy_id);
546 * triggers start of the disconnect procedures and wait for them to be done
548 void iser_conn_terminate(struct iser_conn *ib_conn)
552 /* change the ib conn state only if the conn is UP, however always call
553 * rdma_disconnect since this is the only way to cause the CMA to change
554 * the QP state to ERROR
557 iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP, ISER_CONN_TERMINATING);
558 err = rdma_disconnect(ib_conn->cma_id);
560 iser_err("Failed to disconnect, conn: 0x%p err %d\n",
563 wait_event_interruptible(ib_conn->wait,
564 ib_conn->state == ISER_CONN_DOWN);
566 iser_conn_put(ib_conn, 1); /* deref ib conn deallocate */
569 static int iser_connect_error(struct rdma_cm_id *cma_id)
571 struct iser_conn *ib_conn;
572 ib_conn = (struct iser_conn *)cma_id->context;
574 ib_conn->state = ISER_CONN_DOWN;
575 wake_up_interruptible(&ib_conn->wait);
576 return iser_conn_put(ib_conn, 0); /* deref ib conn's cma id */
579 static int iser_addr_handler(struct rdma_cm_id *cma_id)
581 struct iser_device *device;
582 struct iser_conn *ib_conn;
585 device = iser_device_find_by_ib_device(cma_id);
587 iser_err("device lookup/creation failed\n");
588 return iser_connect_error(cma_id);
591 ib_conn = (struct iser_conn *)cma_id->context;
592 ib_conn->device = device;
594 ret = rdma_resolve_route(cma_id, 1000);
596 iser_err("resolve route failed: %d\n", ret);
597 return iser_connect_error(cma_id);
603 static int iser_route_handler(struct rdma_cm_id *cma_id)
605 struct rdma_conn_param conn_param;
607 struct iser_cm_hdr req_hdr;
609 ret = iser_create_ib_conn_res((struct iser_conn *)cma_id->context);
613 memset(&conn_param, 0, sizeof conn_param);
614 conn_param.responder_resources = 4;
615 conn_param.initiator_depth = 1;
616 conn_param.retry_count = 7;
617 conn_param.rnr_retry_count = 6;
619 memset(&req_hdr, 0, sizeof(req_hdr));
620 req_hdr.flags = (ISER_ZBVA_NOT_SUPPORTED |
621 ISER_SEND_W_INV_NOT_SUPPORTED);
622 conn_param.private_data = (void *)&req_hdr;
623 conn_param.private_data_len = sizeof(struct iser_cm_hdr);
625 ret = rdma_connect(cma_id, &conn_param);
627 iser_err("failure connecting: %d\n", ret);
633 return iser_connect_error(cma_id);
636 static void iser_connected_handler(struct rdma_cm_id *cma_id)
638 struct iser_conn *ib_conn;
640 ib_conn = (struct iser_conn *)cma_id->context;
641 ib_conn->state = ISER_CONN_UP;
642 wake_up_interruptible(&ib_conn->wait);
645 static int iser_disconnected_handler(struct rdma_cm_id *cma_id)
647 struct iser_conn *ib_conn;
650 ib_conn = (struct iser_conn *)cma_id->context;
652 /* getting here when the state is UP means that the conn is being *
653 * terminated asynchronously from the iSCSI layer's perspective. */
654 if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
655 ISER_CONN_TERMINATING))
656 iscsi_conn_failure(ib_conn->iser_conn->iscsi_conn,
657 ISCSI_ERR_CONN_FAILED);
659 /* Complete the termination process if no posts are pending */
660 if (ib_conn->post_recv_buf_count == 0 &&
661 (atomic_read(&ib_conn->post_send_buf_count) == 0)) {
662 ib_conn->state = ISER_CONN_DOWN;
663 wake_up_interruptible(&ib_conn->wait);
666 ret = iser_conn_put(ib_conn, 0); /* deref ib conn's cma id */
670 static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
674 iser_info("event %d status %d conn %p id %p\n",
675 event->event, event->status, cma_id->context, cma_id);
677 switch (event->event) {
678 case RDMA_CM_EVENT_ADDR_RESOLVED:
679 ret = iser_addr_handler(cma_id);
681 case RDMA_CM_EVENT_ROUTE_RESOLVED:
682 ret = iser_route_handler(cma_id);
684 case RDMA_CM_EVENT_ESTABLISHED:
685 iser_connected_handler(cma_id);
687 case RDMA_CM_EVENT_ADDR_ERROR:
688 case RDMA_CM_EVENT_ROUTE_ERROR:
689 case RDMA_CM_EVENT_CONNECT_ERROR:
690 case RDMA_CM_EVENT_UNREACHABLE:
691 case RDMA_CM_EVENT_REJECTED:
692 ret = iser_connect_error(cma_id);
694 case RDMA_CM_EVENT_DISCONNECTED:
695 case RDMA_CM_EVENT_DEVICE_REMOVAL:
696 case RDMA_CM_EVENT_ADDR_CHANGE:
697 ret = iser_disconnected_handler(cma_id);
700 iser_err("Unexpected RDMA CM event (%d)\n", event->event);
706 void iser_conn_init(struct iser_conn *ib_conn)
708 ib_conn->state = ISER_CONN_INIT;
709 init_waitqueue_head(&ib_conn->wait);
710 ib_conn->post_recv_buf_count = 0;
711 atomic_set(&ib_conn->post_send_buf_count, 0);
712 atomic_set(&ib_conn->refcount, 1); /* ref ib conn allocation */
713 INIT_LIST_HEAD(&ib_conn->conn_list);
714 spin_lock_init(&ib_conn->lock);
718 * starts the process of connecting to the target
719 * sleeps until the connection is established or rejected
721 int iser_connect(struct iser_conn *ib_conn,
722 struct sockaddr_in *src_addr,
723 struct sockaddr_in *dst_addr,
726 struct sockaddr *src, *dst;
729 sprintf(ib_conn->name, "%pI4:%d",
730 &dst_addr->sin_addr.s_addr, dst_addr->sin_port);
732 /* the device is known only --after-- address resolution */
733 ib_conn->device = NULL;
735 iser_info("connecting to: %pI4, port 0x%x\n",
736 &dst_addr->sin_addr, dst_addr->sin_port);
738 ib_conn->state = ISER_CONN_PENDING;
740 iser_conn_get(ib_conn); /* ref ib conn's cma id */
741 ib_conn->cma_id = rdma_create_id(iser_cma_handler,
743 RDMA_PS_TCP, IB_QPT_RC);
744 if (IS_ERR(ib_conn->cma_id)) {
745 err = PTR_ERR(ib_conn->cma_id);
746 iser_err("rdma_create_id failed: %d\n", err);
750 src = (struct sockaddr *)src_addr;
751 dst = (struct sockaddr *)dst_addr;
752 err = rdma_resolve_addr(ib_conn->cma_id, src, dst, 1000);
754 iser_err("rdma_resolve_addr failed: %d\n", err);
759 wait_event_interruptible(ib_conn->wait,
760 (ib_conn->state != ISER_CONN_PENDING));
762 if (ib_conn->state != ISER_CONN_UP) {
764 goto connect_failure;
768 mutex_lock(&ig.connlist_mutex);
769 list_add(&ib_conn->conn_list, &ig.connlist);
770 mutex_unlock(&ig.connlist_mutex);
774 ib_conn->cma_id = NULL;
776 ib_conn->state = ISER_CONN_DOWN;
777 iser_conn_put(ib_conn, 1); /* deref ib conn's cma id */
779 iser_conn_put(ib_conn, 1); /* deref ib conn deallocate */
784 * iser_reg_page_vec - Register physical memory
786 * returns: 0 on success, errno code on failure
788 int iser_reg_page_vec(struct iser_conn *ib_conn,
789 struct iser_page_vec *page_vec,
790 struct iser_mem_reg *mem_reg)
792 struct ib_pool_fmr *mem;
797 page_list = page_vec->pages;
798 io_addr = page_list[0];
800 mem = ib_fmr_pool_map_phys(ib_conn->fastreg.fmr.pool,
806 status = (int)PTR_ERR(mem);
807 iser_err("ib_fmr_pool_map_phys failed: %d\n", status);
811 mem_reg->lkey = mem->fmr->lkey;
812 mem_reg->rkey = mem->fmr->rkey;
813 mem_reg->len = page_vec->length * SIZE_4K;
814 mem_reg->va = io_addr;
816 mem_reg->mem_h = (void *)mem;
818 mem_reg->va += page_vec->offset;
819 mem_reg->len = page_vec->data_size;
821 iser_dbg("PHYSICAL Mem.register, [PHYS p_array: 0x%p, sz: %d, "
822 "entry[0]: (0x%08lx,%ld)] -> "
823 "[lkey: 0x%08X mem_h: 0x%p va: 0x%08lX sz: %ld]\n",
824 page_vec, page_vec->length,
825 (unsigned long)page_vec->pages[0],
826 (unsigned long)page_vec->data_size,
827 (unsigned int)mem_reg->lkey, mem_reg->mem_h,
828 (unsigned long)mem_reg->va, (unsigned long)mem_reg->len);
833 * Unregister (previosuly registered using FMR) memory.
834 * If memory is non-FMR does nothing.
836 void iser_unreg_mem_fmr(struct iscsi_iser_task *iser_task,
837 enum iser_data_dir cmd_dir)
839 struct iser_mem_reg *reg = &iser_task->rdma_regd[cmd_dir].reg;
845 iser_dbg("PHYSICAL Mem.Unregister mem_h %p\n",reg->mem_h);
847 ret = ib_fmr_pool_unmap((struct ib_pool_fmr *)reg->mem_h);
849 iser_err("ib_fmr_pool_unmap failed %d\n", ret);
854 void iser_unreg_mem_frwr(struct iscsi_iser_task *iser_task,
855 enum iser_data_dir cmd_dir)
857 struct iser_mem_reg *reg = &iser_task->rdma_regd[cmd_dir].reg;
858 struct iser_conn *ib_conn = iser_task->iser_conn->ib_conn;
859 struct fast_reg_descriptor *desc = reg->mem_h;
866 spin_lock_bh(&ib_conn->lock);
867 list_add_tail(&desc->list, &ib_conn->fastreg.frwr.pool);
868 spin_unlock_bh(&ib_conn->lock);
871 int iser_post_recvl(struct iser_conn *ib_conn)
873 struct ib_recv_wr rx_wr, *rx_wr_failed;
877 sge.addr = ib_conn->login_resp_dma;
878 sge.length = ISER_RX_LOGIN_SIZE;
879 sge.lkey = ib_conn->device->mr->lkey;
881 rx_wr.wr_id = (unsigned long)ib_conn->login_resp_buf;
882 rx_wr.sg_list = &sge;
886 ib_conn->post_recv_buf_count++;
887 ib_ret = ib_post_recv(ib_conn->qp, &rx_wr, &rx_wr_failed);
889 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
890 ib_conn->post_recv_buf_count--;
895 int iser_post_recvm(struct iser_conn *ib_conn, int count)
897 struct ib_recv_wr *rx_wr, *rx_wr_failed;
899 unsigned int my_rx_head = ib_conn->rx_desc_head;
900 struct iser_rx_desc *rx_desc;
902 for (rx_wr = ib_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
903 rx_desc = &ib_conn->rx_descs[my_rx_head];
904 rx_wr->wr_id = (unsigned long)rx_desc;
905 rx_wr->sg_list = &rx_desc->rx_sg;
907 rx_wr->next = rx_wr + 1;
908 my_rx_head = (my_rx_head + 1) & ib_conn->qp_max_recv_dtos_mask;
912 rx_wr->next = NULL; /* mark end of work requests list */
914 ib_conn->post_recv_buf_count += count;
915 ib_ret = ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &rx_wr_failed);
917 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
918 ib_conn->post_recv_buf_count -= count;
920 ib_conn->rx_desc_head = my_rx_head;
926 * iser_start_send - Initiate a Send DTO operation
928 * returns 0 on success, -1 on failure
930 int iser_post_send(struct iser_conn *ib_conn, struct iser_tx_desc *tx_desc)
933 struct ib_send_wr send_wr, *send_wr_failed;
935 ib_dma_sync_single_for_device(ib_conn->device->ib_device,
936 tx_desc->dma_addr, ISER_HEADERS_LEN, DMA_TO_DEVICE);
939 send_wr.wr_id = (unsigned long)tx_desc;
940 send_wr.sg_list = tx_desc->tx_sg;
941 send_wr.num_sge = tx_desc->num_sge;
942 send_wr.opcode = IB_WR_SEND;
943 send_wr.send_flags = IB_SEND_SIGNALED;
945 atomic_inc(&ib_conn->post_send_buf_count);
947 ib_ret = ib_post_send(ib_conn->qp, &send_wr, &send_wr_failed);
949 iser_err("ib_post_send failed, ret:%d\n", ib_ret);
950 atomic_dec(&ib_conn->post_send_buf_count);
955 static void iser_handle_comp_error(struct iser_tx_desc *desc,
956 struct iser_conn *ib_conn)
958 if (desc && desc->type == ISCSI_TX_DATAOUT)
959 kmem_cache_free(ig.desc_cache, desc);
961 if (ib_conn->post_recv_buf_count == 0 &&
962 atomic_read(&ib_conn->post_send_buf_count) == 0) {
963 /* getting here when the state is UP means that the conn is *
964 * being terminated asynchronously from the iSCSI layer's *
966 if (iser_conn_state_comp_exch(ib_conn, ISER_CONN_UP,
967 ISER_CONN_TERMINATING))
968 iscsi_conn_failure(ib_conn->iser_conn->iscsi_conn,
969 ISCSI_ERR_CONN_FAILED);
971 /* no more non completed posts to the QP, complete the
972 * termination process w.o worrying on disconnect event */
973 ib_conn->state = ISER_CONN_DOWN;
974 wake_up_interruptible(&ib_conn->wait);
978 static int iser_drain_tx_cq(struct iser_device *device, int cq_index)
980 struct ib_cq *cq = device->tx_cq[cq_index];
982 struct iser_tx_desc *tx_desc;
983 struct iser_conn *ib_conn;
984 int completed_tx = 0;
986 while (ib_poll_cq(cq, 1, &wc) == 1) {
987 tx_desc = (struct iser_tx_desc *) (unsigned long) wc.wr_id;
988 ib_conn = wc.qp->qp_context;
989 if (wc.status == IB_WC_SUCCESS) {
990 if (wc.opcode == IB_WC_SEND)
991 iser_snd_completion(tx_desc, ib_conn);
992 else if (wc.opcode == IB_WC_LOCAL_INV ||
993 wc.opcode == IB_WC_FAST_REG_MR) {
994 atomic_dec(&ib_conn->post_send_buf_count);
997 iser_err("expected opcode %d got %d\n",
998 IB_WC_SEND, wc.opcode);
1000 iser_err("tx id %llx status %d vend_err %x\n",
1001 wc.wr_id, wc.status, wc.vendor_err);
1002 atomic_dec(&ib_conn->post_send_buf_count);
1003 iser_handle_comp_error(tx_desc, ib_conn);
1007 return completed_tx;
1011 static void iser_cq_tasklet_fn(unsigned long data)
1013 struct iser_cq_desc *cq_desc = (struct iser_cq_desc *)data;
1014 struct iser_device *device = cq_desc->device;
1015 int cq_index = cq_desc->cq_index;
1016 struct ib_cq *cq = device->rx_cq[cq_index];
1018 struct iser_rx_desc *desc;
1019 unsigned long xfer_len;
1020 struct iser_conn *ib_conn;
1021 int completed_tx, completed_rx;
1022 completed_tx = completed_rx = 0;
1024 while (ib_poll_cq(cq, 1, &wc) == 1) {
1025 desc = (struct iser_rx_desc *) (unsigned long) wc.wr_id;
1026 BUG_ON(desc == NULL);
1027 ib_conn = wc.qp->qp_context;
1028 if (wc.status == IB_WC_SUCCESS) {
1029 if (wc.opcode == IB_WC_RECV) {
1030 xfer_len = (unsigned long)wc.byte_len;
1031 iser_rcv_completion(desc, xfer_len, ib_conn);
1033 iser_err("expected opcode %d got %d\n",
1034 IB_WC_RECV, wc.opcode);
1036 if (wc.status != IB_WC_WR_FLUSH_ERR)
1037 iser_err("rx id %llx status %d vend_err %x\n",
1038 wc.wr_id, wc.status, wc.vendor_err);
1039 ib_conn->post_recv_buf_count--;
1040 iser_handle_comp_error(NULL, ib_conn);
1043 if (!(completed_rx & 63))
1044 completed_tx += iser_drain_tx_cq(device, cq_index);
1046 /* #warning "it is assumed here that arming CQ only once its empty" *
1047 * " would not cause interrupts to be missed" */
1048 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1050 completed_tx += iser_drain_tx_cq(device, cq_index);
1051 iser_dbg("got %d rx %d tx completions\n", completed_rx, completed_tx);
1054 static void iser_cq_callback(struct ib_cq *cq, void *cq_context)
1056 struct iser_cq_desc *cq_desc = (struct iser_cq_desc *)cq_context;
1057 struct iser_device *device = cq_desc->device;
1058 int cq_index = cq_desc->cq_index;
1060 tasklet_schedule(&device->cq_tasklet[cq_index]);