Linux 4.2.1
[linux/fpc-iii.git] / drivers / infiniband / ulp / iser / iser_verbs.c
blob5c9f565ea0e88840c3c3c7a031e20180c3c8025d
1 /*
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
14 * conditions are met:
16 * - Redistributions of source code must retain the above
17 * copyright notice, this list of conditions and the following
18 * disclaimer.
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
32 * SOFTWARE.
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 + \
45 ISCSI_ISER_MAX_CONN)
47 static int iser_cq_poll_limit = 512;
49 static void iser_cq_tasklet_fn(unsigned long data);
50 static void iser_cq_callback(struct ib_cq *cq, void *cq_context);
52 static void iser_cq_event_callback(struct ib_event *cause, void *context)
54 iser_err("cq event %s (%d)\n",
55 ib_event_msg(cause->event), cause->event);
58 static void iser_qp_event_callback(struct ib_event *cause, void *context)
60 iser_err("qp event %s (%d)\n",
61 ib_event_msg(cause->event), cause->event);
64 static void iser_event_handler(struct ib_event_handler *handler,
65 struct ib_event *event)
67 iser_err("async event %s (%d) on device %s port %d\n",
68 ib_event_msg(event->event), event->event,
69 event->device->name, event->element.port_num);
72 /**
73 * iser_create_device_ib_res - creates Protection Domain (PD), Completion
74 * Queue (CQ), DMA Memory Region (DMA MR) with the device associated with
75 * the adapator.
77 * returns 0 on success, -1 on failure
79 static int iser_create_device_ib_res(struct iser_device *device)
81 struct ib_device_attr *dev_attr = &device->dev_attr;
82 int ret, i, max_cqe;
84 ret = ib_query_device(device->ib_device, dev_attr);
85 if (ret) {
86 pr_warn("Query device failed for %s\n", device->ib_device->name);
87 return ret;
90 /* Assign function handles - based on FMR support */
91 if (device->ib_device->alloc_fmr && device->ib_device->dealloc_fmr &&
92 device->ib_device->map_phys_fmr && device->ib_device->unmap_fmr) {
93 iser_info("FMR supported, using FMR for registration\n");
94 device->iser_alloc_rdma_reg_res = iser_create_fmr_pool;
95 device->iser_free_rdma_reg_res = iser_free_fmr_pool;
96 device->iser_reg_rdma_mem = iser_reg_rdma_mem_fmr;
97 device->iser_unreg_rdma_mem = iser_unreg_mem_fmr;
98 } else
99 if (dev_attr->device_cap_flags & IB_DEVICE_MEM_MGT_EXTENSIONS) {
100 iser_info("FastReg supported, using FastReg for registration\n");
101 device->iser_alloc_rdma_reg_res = iser_create_fastreg_pool;
102 device->iser_free_rdma_reg_res = iser_free_fastreg_pool;
103 device->iser_reg_rdma_mem = iser_reg_rdma_mem_fastreg;
104 device->iser_unreg_rdma_mem = iser_unreg_mem_fastreg;
105 } else {
106 iser_err("IB device does not support FMRs nor FastRegs, can't register memory\n");
107 return -1;
110 device->comps_used = min_t(int, num_online_cpus(),
111 device->ib_device->num_comp_vectors);
113 device->comps = kcalloc(device->comps_used, sizeof(*device->comps),
114 GFP_KERNEL);
115 if (!device->comps)
116 goto comps_err;
118 max_cqe = min(ISER_MAX_CQ_LEN, dev_attr->max_cqe);
120 iser_info("using %d CQs, device %s supports %d vectors max_cqe %d\n",
121 device->comps_used, device->ib_device->name,
122 device->ib_device->num_comp_vectors, max_cqe);
124 device->pd = ib_alloc_pd(device->ib_device);
125 if (IS_ERR(device->pd))
126 goto pd_err;
128 for (i = 0; i < device->comps_used; i++) {
129 struct ib_cq_init_attr cq_attr = {};
130 struct iser_comp *comp = &device->comps[i];
132 comp->device = device;
133 cq_attr.cqe = max_cqe;
134 cq_attr.comp_vector = i;
135 comp->cq = ib_create_cq(device->ib_device,
136 iser_cq_callback,
137 iser_cq_event_callback,
138 (void *)comp,
139 &cq_attr);
140 if (IS_ERR(comp->cq)) {
141 comp->cq = NULL;
142 goto cq_err;
145 if (ib_req_notify_cq(comp->cq, IB_CQ_NEXT_COMP))
146 goto cq_err;
148 tasklet_init(&comp->tasklet, iser_cq_tasklet_fn,
149 (unsigned long)comp);
152 device->mr = ib_get_dma_mr(device->pd, IB_ACCESS_LOCAL_WRITE |
153 IB_ACCESS_REMOTE_WRITE |
154 IB_ACCESS_REMOTE_READ);
155 if (IS_ERR(device->mr))
156 goto dma_mr_err;
158 INIT_IB_EVENT_HANDLER(&device->event_handler, device->ib_device,
159 iser_event_handler);
160 if (ib_register_event_handler(&device->event_handler))
161 goto handler_err;
163 return 0;
165 handler_err:
166 ib_dereg_mr(device->mr);
167 dma_mr_err:
168 for (i = 0; i < device->comps_used; i++)
169 tasklet_kill(&device->comps[i].tasklet);
170 cq_err:
171 for (i = 0; i < device->comps_used; i++) {
172 struct iser_comp *comp = &device->comps[i];
174 if (comp->cq)
175 ib_destroy_cq(comp->cq);
177 ib_dealloc_pd(device->pd);
178 pd_err:
179 kfree(device->comps);
180 comps_err:
181 iser_err("failed to allocate an IB resource\n");
182 return -1;
186 * iser_free_device_ib_res - destroy/dealloc/dereg the DMA MR,
187 * CQ and PD created with the device associated with the adapator.
189 static void iser_free_device_ib_res(struct iser_device *device)
191 int i;
192 BUG_ON(device->mr == NULL);
194 for (i = 0; i < device->comps_used; i++) {
195 struct iser_comp *comp = &device->comps[i];
197 tasklet_kill(&comp->tasklet);
198 ib_destroy_cq(comp->cq);
199 comp->cq = NULL;
202 (void)ib_unregister_event_handler(&device->event_handler);
203 (void)ib_dereg_mr(device->mr);
204 (void)ib_dealloc_pd(device->pd);
206 kfree(device->comps);
207 device->comps = NULL;
209 device->mr = NULL;
210 device->pd = NULL;
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 ib_conn *ib_conn, unsigned cmds_max)
220 struct iser_device *device = ib_conn->device;
221 struct ib_fmr_pool_param params;
222 int ret = -ENOMEM;
224 ib_conn->fmr.page_vec = kmalloc(sizeof(*ib_conn->fmr.page_vec) +
225 (sizeof(u64)*(ISCSI_ISER_SG_TABLESIZE + 1)),
226 GFP_KERNEL);
227 if (!ib_conn->fmr.page_vec)
228 return ret;
230 ib_conn->fmr.page_vec->pages = (u64 *)(ib_conn->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;
240 params.cache = 0;
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->fmr.pool = ib_create_fmr_pool(device->pd, &params);
247 if (!IS_ERR(ib_conn->fmr.pool))
248 return 0;
250 /* no FMR => no need for page_vec */
251 kfree(ib_conn->fmr.page_vec);
252 ib_conn->fmr.page_vec = NULL;
254 ret = PTR_ERR(ib_conn->fmr.pool);
255 ib_conn->fmr.pool = NULL;
256 if (ret != -ENOSYS) {
257 iser_err("FMR allocation failed, err %d\n", ret);
258 return ret;
259 } else {
260 iser_warn("FMRs are not supported, using unaligned mode\n");
261 return 0;
266 * iser_free_fmr_pool - releases the FMR pool and page vec
268 void iser_free_fmr_pool(struct ib_conn *ib_conn)
270 iser_info("freeing conn %p fmr pool %p\n",
271 ib_conn, ib_conn->fmr.pool);
273 if (ib_conn->fmr.pool != NULL)
274 ib_destroy_fmr_pool(ib_conn->fmr.pool);
276 ib_conn->fmr.pool = NULL;
278 kfree(ib_conn->fmr.page_vec);
279 ib_conn->fmr.page_vec = NULL;
282 static int
283 iser_alloc_pi_ctx(struct ib_device *ib_device, struct ib_pd *pd,
284 struct fast_reg_descriptor *desc)
286 struct iser_pi_context *pi_ctx = NULL;
287 struct ib_mr_init_attr mr_init_attr = {.max_reg_descriptors = 2,
288 .flags = IB_MR_SIGNATURE_EN};
289 int ret = 0;
291 desc->pi_ctx = kzalloc(sizeof(*desc->pi_ctx), GFP_KERNEL);
292 if (!desc->pi_ctx)
293 return -ENOMEM;
295 pi_ctx = desc->pi_ctx;
297 pi_ctx->prot_frpl = ib_alloc_fast_reg_page_list(ib_device,
298 ISCSI_ISER_SG_TABLESIZE);
299 if (IS_ERR(pi_ctx->prot_frpl)) {
300 ret = PTR_ERR(pi_ctx->prot_frpl);
301 goto prot_frpl_failure;
304 pi_ctx->prot_mr = ib_alloc_fast_reg_mr(pd,
305 ISCSI_ISER_SG_TABLESIZE + 1);
306 if (IS_ERR(pi_ctx->prot_mr)) {
307 ret = PTR_ERR(pi_ctx->prot_mr);
308 goto prot_mr_failure;
310 desc->reg_indicators |= ISER_PROT_KEY_VALID;
312 pi_ctx->sig_mr = ib_create_mr(pd, &mr_init_attr);
313 if (IS_ERR(pi_ctx->sig_mr)) {
314 ret = PTR_ERR(pi_ctx->sig_mr);
315 goto sig_mr_failure;
317 desc->reg_indicators |= ISER_SIG_KEY_VALID;
318 desc->reg_indicators &= ~ISER_FASTREG_PROTECTED;
320 return 0;
322 sig_mr_failure:
323 ib_dereg_mr(desc->pi_ctx->prot_mr);
324 prot_mr_failure:
325 ib_free_fast_reg_page_list(desc->pi_ctx->prot_frpl);
326 prot_frpl_failure:
327 kfree(desc->pi_ctx);
329 return ret;
332 static void
333 iser_free_pi_ctx(struct iser_pi_context *pi_ctx)
335 ib_free_fast_reg_page_list(pi_ctx->prot_frpl);
336 ib_dereg_mr(pi_ctx->prot_mr);
337 ib_destroy_mr(pi_ctx->sig_mr);
338 kfree(pi_ctx);
341 static int
342 iser_create_fastreg_desc(struct ib_device *ib_device, struct ib_pd *pd,
343 bool pi_enable, struct fast_reg_descriptor *desc)
345 int ret;
347 desc->data_frpl = ib_alloc_fast_reg_page_list(ib_device,
348 ISCSI_ISER_SG_TABLESIZE + 1);
349 if (IS_ERR(desc->data_frpl)) {
350 ret = PTR_ERR(desc->data_frpl);
351 iser_err("Failed to allocate ib_fast_reg_page_list err=%d\n",
352 ret);
353 return PTR_ERR(desc->data_frpl);
356 desc->data_mr = ib_alloc_fast_reg_mr(pd, ISCSI_ISER_SG_TABLESIZE + 1);
357 if (IS_ERR(desc->data_mr)) {
358 ret = PTR_ERR(desc->data_mr);
359 iser_err("Failed to allocate ib_fast_reg_mr err=%d\n", ret);
360 goto fast_reg_mr_failure;
362 desc->reg_indicators |= ISER_DATA_KEY_VALID;
364 if (pi_enable) {
365 ret = iser_alloc_pi_ctx(ib_device, pd, desc);
366 if (ret)
367 goto pi_ctx_alloc_failure;
370 return 0;
371 pi_ctx_alloc_failure:
372 ib_dereg_mr(desc->data_mr);
373 fast_reg_mr_failure:
374 ib_free_fast_reg_page_list(desc->data_frpl);
376 return ret;
380 * iser_create_fastreg_pool - Creates pool of fast_reg descriptors
381 * for fast registration work requests.
382 * returns 0 on success, or errno code on failure
384 int iser_create_fastreg_pool(struct ib_conn *ib_conn, unsigned cmds_max)
386 struct iser_device *device = ib_conn->device;
387 struct fast_reg_descriptor *desc;
388 int i, ret;
390 INIT_LIST_HEAD(&ib_conn->fastreg.pool);
391 ib_conn->fastreg.pool_size = 0;
392 for (i = 0; i < cmds_max; i++) {
393 desc = kzalloc(sizeof(*desc), GFP_KERNEL);
394 if (!desc) {
395 iser_err("Failed to allocate a new fast_reg descriptor\n");
396 ret = -ENOMEM;
397 goto err;
400 ret = iser_create_fastreg_desc(device->ib_device, device->pd,
401 ib_conn->pi_support, desc);
402 if (ret) {
403 iser_err("Failed to create fastreg descriptor err=%d\n",
404 ret);
405 kfree(desc);
406 goto err;
409 list_add_tail(&desc->list, &ib_conn->fastreg.pool);
410 ib_conn->fastreg.pool_size++;
413 return 0;
415 err:
416 iser_free_fastreg_pool(ib_conn);
417 return ret;
421 * iser_free_fastreg_pool - releases the pool of fast_reg descriptors
423 void iser_free_fastreg_pool(struct ib_conn *ib_conn)
425 struct fast_reg_descriptor *desc, *tmp;
426 int i = 0;
428 if (list_empty(&ib_conn->fastreg.pool))
429 return;
431 iser_info("freeing conn %p fr pool\n", ib_conn);
433 list_for_each_entry_safe(desc, tmp, &ib_conn->fastreg.pool, list) {
434 list_del(&desc->list);
435 ib_free_fast_reg_page_list(desc->data_frpl);
436 ib_dereg_mr(desc->data_mr);
437 if (desc->pi_ctx)
438 iser_free_pi_ctx(desc->pi_ctx);
439 kfree(desc);
440 ++i;
443 if (i < ib_conn->fastreg.pool_size)
444 iser_warn("pool still has %d regions registered\n",
445 ib_conn->fastreg.pool_size - i);
449 * iser_create_ib_conn_res - Queue-Pair (QP)
451 * returns 0 on success, -1 on failure
453 static int iser_create_ib_conn_res(struct ib_conn *ib_conn)
455 struct iser_conn *iser_conn = container_of(ib_conn, struct iser_conn,
456 ib_conn);
457 struct iser_device *device;
458 struct ib_device_attr *dev_attr;
459 struct ib_qp_init_attr init_attr;
460 int ret = -ENOMEM;
461 int index, min_index = 0;
463 BUG_ON(ib_conn->device == NULL);
465 device = ib_conn->device;
466 dev_attr = &device->dev_attr;
468 memset(&init_attr, 0, sizeof init_attr);
470 mutex_lock(&ig.connlist_mutex);
471 /* select the CQ with the minimal number of usages */
472 for (index = 0; index < device->comps_used; index++) {
473 if (device->comps[index].active_qps <
474 device->comps[min_index].active_qps)
475 min_index = index;
477 ib_conn->comp = &device->comps[min_index];
478 ib_conn->comp->active_qps++;
479 mutex_unlock(&ig.connlist_mutex);
480 iser_info("cq index %d used for ib_conn %p\n", min_index, ib_conn);
482 init_attr.event_handler = iser_qp_event_callback;
483 init_attr.qp_context = (void *)ib_conn;
484 init_attr.send_cq = ib_conn->comp->cq;
485 init_attr.recv_cq = ib_conn->comp->cq;
486 init_attr.cap.max_recv_wr = ISER_QP_MAX_RECV_DTOS;
487 init_attr.cap.max_send_sge = 2;
488 init_attr.cap.max_recv_sge = 1;
489 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
490 init_attr.qp_type = IB_QPT_RC;
491 if (ib_conn->pi_support) {
492 init_attr.cap.max_send_wr = ISER_QP_SIG_MAX_REQ_DTOS + 1;
493 init_attr.create_flags |= IB_QP_CREATE_SIGNATURE_EN;
494 iser_conn->max_cmds =
495 ISER_GET_MAX_XMIT_CMDS(ISER_QP_SIG_MAX_REQ_DTOS);
496 } else {
497 if (dev_attr->max_qp_wr > ISER_QP_MAX_REQ_DTOS) {
498 init_attr.cap.max_send_wr = ISER_QP_MAX_REQ_DTOS + 1;
499 iser_conn->max_cmds =
500 ISER_GET_MAX_XMIT_CMDS(ISER_QP_MAX_REQ_DTOS);
501 } else {
502 init_attr.cap.max_send_wr = dev_attr->max_qp_wr;
503 iser_conn->max_cmds =
504 ISER_GET_MAX_XMIT_CMDS(dev_attr->max_qp_wr);
505 iser_dbg("device %s supports max_send_wr %d\n",
506 device->ib_device->name, dev_attr->max_qp_wr);
510 ret = rdma_create_qp(ib_conn->cma_id, device->pd, &init_attr);
511 if (ret)
512 goto out_err;
514 ib_conn->qp = ib_conn->cma_id->qp;
515 iser_info("setting conn %p cma_id %p qp %p\n",
516 ib_conn, ib_conn->cma_id,
517 ib_conn->cma_id->qp);
518 return ret;
520 out_err:
521 mutex_lock(&ig.connlist_mutex);
522 ib_conn->comp->active_qps--;
523 mutex_unlock(&ig.connlist_mutex);
524 iser_err("unable to alloc mem or create resource, err %d\n", ret);
526 return ret;
530 * based on the resolved device node GUID see if there already allocated
531 * device for this device. If there's no such, create one.
533 static
534 struct iser_device *iser_device_find_by_ib_device(struct rdma_cm_id *cma_id)
536 struct iser_device *device;
538 mutex_lock(&ig.device_list_mutex);
540 list_for_each_entry(device, &ig.device_list, ig_list)
541 /* find if there's a match using the node GUID */
542 if (device->ib_device->node_guid == cma_id->device->node_guid)
543 goto inc_refcnt;
545 device = kzalloc(sizeof *device, GFP_KERNEL);
546 if (device == NULL)
547 goto out;
549 /* assign this device to the device */
550 device->ib_device = cma_id->device;
551 /* init the device and link it into ig device list */
552 if (iser_create_device_ib_res(device)) {
553 kfree(device);
554 device = NULL;
555 goto out;
557 list_add(&device->ig_list, &ig.device_list);
559 inc_refcnt:
560 device->refcount++;
561 out:
562 mutex_unlock(&ig.device_list_mutex);
563 return device;
566 /* if there's no demand for this device, release it */
567 static void iser_device_try_release(struct iser_device *device)
569 mutex_lock(&ig.device_list_mutex);
570 device->refcount--;
571 iser_info("device %p refcount %d\n", device, device->refcount);
572 if (!device->refcount) {
573 iser_free_device_ib_res(device);
574 list_del(&device->ig_list);
575 kfree(device);
577 mutex_unlock(&ig.device_list_mutex);
581 * Called with state mutex held
583 static int iser_conn_state_comp_exch(struct iser_conn *iser_conn,
584 enum iser_conn_state comp,
585 enum iser_conn_state exch)
587 int ret;
589 ret = (iser_conn->state == comp);
590 if (ret)
591 iser_conn->state = exch;
593 return ret;
596 void iser_release_work(struct work_struct *work)
598 struct iser_conn *iser_conn;
600 iser_conn = container_of(work, struct iser_conn, release_work);
602 /* Wait for conn_stop to complete */
603 wait_for_completion(&iser_conn->stop_completion);
604 /* Wait for IB resouces cleanup to complete */
605 wait_for_completion(&iser_conn->ib_completion);
607 mutex_lock(&iser_conn->state_mutex);
608 iser_conn->state = ISER_CONN_DOWN;
609 mutex_unlock(&iser_conn->state_mutex);
611 iser_conn_release(iser_conn);
615 * iser_free_ib_conn_res - release IB related resources
616 * @iser_conn: iser connection struct
617 * @destroy: indicator if we need to try to release the
618 * iser device and memory regoins pool (only iscsi
619 * shutdown and DEVICE_REMOVAL will use this).
621 * This routine is called with the iser state mutex held
622 * so the cm_id removal is out of here. It is Safe to
623 * be invoked multiple times.
625 static void iser_free_ib_conn_res(struct iser_conn *iser_conn,
626 bool destroy)
628 struct ib_conn *ib_conn = &iser_conn->ib_conn;
629 struct iser_device *device = ib_conn->device;
631 iser_info("freeing conn %p cma_id %p qp %p\n",
632 iser_conn, ib_conn->cma_id, ib_conn->qp);
634 if (ib_conn->qp != NULL) {
635 ib_conn->comp->active_qps--;
636 rdma_destroy_qp(ib_conn->cma_id);
637 ib_conn->qp = NULL;
640 if (destroy) {
641 if (iser_conn->rx_descs)
642 iser_free_rx_descriptors(iser_conn);
644 if (device != NULL) {
645 iser_device_try_release(device);
646 ib_conn->device = NULL;
652 * Frees all conn objects and deallocs conn descriptor
654 void iser_conn_release(struct iser_conn *iser_conn)
656 struct ib_conn *ib_conn = &iser_conn->ib_conn;
658 mutex_lock(&ig.connlist_mutex);
659 list_del(&iser_conn->conn_list);
660 mutex_unlock(&ig.connlist_mutex);
662 mutex_lock(&iser_conn->state_mutex);
663 /* In case we endup here without ep_disconnect being invoked. */
664 if (iser_conn->state != ISER_CONN_DOWN) {
665 iser_warn("iser conn %p state %d, expected state down.\n",
666 iser_conn, iser_conn->state);
667 iscsi_destroy_endpoint(iser_conn->ep);
668 iser_conn->state = ISER_CONN_DOWN;
671 * In case we never got to bind stage, we still need to
672 * release IB resources (which is safe to call more than once).
674 iser_free_ib_conn_res(iser_conn, true);
675 mutex_unlock(&iser_conn->state_mutex);
677 if (ib_conn->cma_id != NULL) {
678 rdma_destroy_id(ib_conn->cma_id);
679 ib_conn->cma_id = NULL;
682 kfree(iser_conn);
686 * triggers start of the disconnect procedures and wait for them to be done
687 * Called with state mutex held
689 int iser_conn_terminate(struct iser_conn *iser_conn)
691 struct ib_conn *ib_conn = &iser_conn->ib_conn;
692 struct ib_send_wr *bad_wr;
693 int err = 0;
695 /* terminate the iser conn only if the conn state is UP */
696 if (!iser_conn_state_comp_exch(iser_conn, ISER_CONN_UP,
697 ISER_CONN_TERMINATING))
698 return 0;
700 iser_info("iser_conn %p state %d\n", iser_conn, iser_conn->state);
702 /* suspend queuing of new iscsi commands */
703 if (iser_conn->iscsi_conn)
704 iscsi_suspend_queue(iser_conn->iscsi_conn);
707 * In case we didn't already clean up the cma_id (peer initiated
708 * a disconnection), we need to Cause the CMA to change the QP
709 * state to ERROR.
711 if (ib_conn->cma_id) {
712 err = rdma_disconnect(ib_conn->cma_id);
713 if (err)
714 iser_err("Failed to disconnect, conn: 0x%p err %d\n",
715 iser_conn, err);
717 /* post an indication that all flush errors were consumed */
718 err = ib_post_send(ib_conn->qp, &ib_conn->beacon, &bad_wr);
719 if (err) {
720 iser_err("conn %p failed to post beacon", ib_conn);
721 return 1;
724 wait_for_completion(&ib_conn->flush_comp);
727 return 1;
731 * Called with state mutex held
733 static void iser_connect_error(struct rdma_cm_id *cma_id)
735 struct iser_conn *iser_conn;
737 iser_conn = (struct iser_conn *)cma_id->context;
738 iser_conn->state = ISER_CONN_TERMINATING;
742 * Called with state mutex held
744 static void iser_addr_handler(struct rdma_cm_id *cma_id)
746 struct iser_device *device;
747 struct iser_conn *iser_conn;
748 struct ib_conn *ib_conn;
749 int ret;
751 iser_conn = (struct iser_conn *)cma_id->context;
752 if (iser_conn->state != ISER_CONN_PENDING)
753 /* bailout */
754 return;
756 ib_conn = &iser_conn->ib_conn;
757 device = iser_device_find_by_ib_device(cma_id);
758 if (!device) {
759 iser_err("device lookup/creation failed\n");
760 iser_connect_error(cma_id);
761 return;
764 ib_conn->device = device;
766 /* connection T10-PI support */
767 if (iser_pi_enable) {
768 if (!(device->dev_attr.device_cap_flags &
769 IB_DEVICE_SIGNATURE_HANDOVER)) {
770 iser_warn("T10-PI requested but not supported on %s, "
771 "continue without T10-PI\n",
772 ib_conn->device->ib_device->name);
773 ib_conn->pi_support = false;
774 } else {
775 ib_conn->pi_support = true;
779 ret = rdma_resolve_route(cma_id, 1000);
780 if (ret) {
781 iser_err("resolve route failed: %d\n", ret);
782 iser_connect_error(cma_id);
783 return;
788 * Called with state mutex held
790 static void iser_route_handler(struct rdma_cm_id *cma_id)
792 struct rdma_conn_param conn_param;
793 int ret;
794 struct iser_cm_hdr req_hdr;
795 struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
796 struct ib_conn *ib_conn = &iser_conn->ib_conn;
797 struct iser_device *device = ib_conn->device;
799 if (iser_conn->state != ISER_CONN_PENDING)
800 /* bailout */
801 return;
803 ret = iser_create_ib_conn_res(ib_conn);
804 if (ret)
805 goto failure;
807 memset(&conn_param, 0, sizeof conn_param);
808 conn_param.responder_resources = device->dev_attr.max_qp_rd_atom;
809 conn_param.initiator_depth = 1;
810 conn_param.retry_count = 7;
811 conn_param.rnr_retry_count = 6;
813 memset(&req_hdr, 0, sizeof(req_hdr));
814 req_hdr.flags = (ISER_ZBVA_NOT_SUPPORTED |
815 ISER_SEND_W_INV_NOT_SUPPORTED);
816 conn_param.private_data = (void *)&req_hdr;
817 conn_param.private_data_len = sizeof(struct iser_cm_hdr);
819 ret = rdma_connect(cma_id, &conn_param);
820 if (ret) {
821 iser_err("failure connecting: %d\n", ret);
822 goto failure;
825 return;
826 failure:
827 iser_connect_error(cma_id);
830 static void iser_connected_handler(struct rdma_cm_id *cma_id)
832 struct iser_conn *iser_conn;
833 struct ib_qp_attr attr;
834 struct ib_qp_init_attr init_attr;
836 iser_conn = (struct iser_conn *)cma_id->context;
837 if (iser_conn->state != ISER_CONN_PENDING)
838 /* bailout */
839 return;
841 (void)ib_query_qp(cma_id->qp, &attr, ~0, &init_attr);
842 iser_info("remote qpn:%x my qpn:%x\n", attr.dest_qp_num, cma_id->qp->qp_num);
844 iser_conn->state = ISER_CONN_UP;
845 complete(&iser_conn->up_completion);
848 static void iser_disconnected_handler(struct rdma_cm_id *cma_id)
850 struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
852 if (iser_conn_terminate(iser_conn)) {
853 if (iser_conn->iscsi_conn)
854 iscsi_conn_failure(iser_conn->iscsi_conn,
855 ISCSI_ERR_CONN_FAILED);
856 else
857 iser_err("iscsi_iser connection isn't bound\n");
861 static void iser_cleanup_handler(struct rdma_cm_id *cma_id,
862 bool destroy)
864 struct iser_conn *iser_conn = (struct iser_conn *)cma_id->context;
867 * We are not guaranteed that we visited disconnected_handler
868 * by now, call it here to be safe that we handle CM drep
869 * and flush errors.
871 iser_disconnected_handler(cma_id);
872 iser_free_ib_conn_res(iser_conn, destroy);
873 complete(&iser_conn->ib_completion);
876 static int iser_cma_handler(struct rdma_cm_id *cma_id, struct rdma_cm_event *event)
878 struct iser_conn *iser_conn;
879 int ret = 0;
881 iser_conn = (struct iser_conn *)cma_id->context;
882 iser_info("%s (%d): status %d conn %p id %p\n",
883 rdma_event_msg(event->event), event->event,
884 event->status, cma_id->context, cma_id);
886 mutex_lock(&iser_conn->state_mutex);
887 switch (event->event) {
888 case RDMA_CM_EVENT_ADDR_RESOLVED:
889 iser_addr_handler(cma_id);
890 break;
891 case RDMA_CM_EVENT_ROUTE_RESOLVED:
892 iser_route_handler(cma_id);
893 break;
894 case RDMA_CM_EVENT_ESTABLISHED:
895 iser_connected_handler(cma_id);
896 break;
897 case RDMA_CM_EVENT_ADDR_ERROR:
898 case RDMA_CM_EVENT_ROUTE_ERROR:
899 case RDMA_CM_EVENT_CONNECT_ERROR:
900 case RDMA_CM_EVENT_UNREACHABLE:
901 case RDMA_CM_EVENT_REJECTED:
902 iser_connect_error(cma_id);
903 break;
904 case RDMA_CM_EVENT_DISCONNECTED:
905 case RDMA_CM_EVENT_ADDR_CHANGE:
906 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
907 iser_cleanup_handler(cma_id, false);
908 break;
909 case RDMA_CM_EVENT_DEVICE_REMOVAL:
911 * we *must* destroy the device as we cannot rely
912 * on iscsid to be around to initiate error handling.
913 * also if we are not in state DOWN implicitly destroy
914 * the cma_id.
916 iser_cleanup_handler(cma_id, true);
917 if (iser_conn->state != ISER_CONN_DOWN) {
918 iser_conn->ib_conn.cma_id = NULL;
919 ret = 1;
921 break;
922 default:
923 iser_err("Unexpected RDMA CM event: %s (%d)\n",
924 rdma_event_msg(event->event), event->event);
925 break;
927 mutex_unlock(&iser_conn->state_mutex);
929 return ret;
932 void iser_conn_init(struct iser_conn *iser_conn)
934 iser_conn->state = ISER_CONN_INIT;
935 iser_conn->ib_conn.post_recv_buf_count = 0;
936 init_completion(&iser_conn->ib_conn.flush_comp);
937 init_completion(&iser_conn->stop_completion);
938 init_completion(&iser_conn->ib_completion);
939 init_completion(&iser_conn->up_completion);
940 INIT_LIST_HEAD(&iser_conn->conn_list);
941 spin_lock_init(&iser_conn->ib_conn.lock);
942 mutex_init(&iser_conn->state_mutex);
946 * starts the process of connecting to the target
947 * sleeps until the connection is established or rejected
949 int iser_connect(struct iser_conn *iser_conn,
950 struct sockaddr *src_addr,
951 struct sockaddr *dst_addr,
952 int non_blocking)
954 struct ib_conn *ib_conn = &iser_conn->ib_conn;
955 int err = 0;
957 mutex_lock(&iser_conn->state_mutex);
959 sprintf(iser_conn->name, "%pISp", dst_addr);
961 iser_info("connecting to: %s\n", iser_conn->name);
963 /* the device is known only --after-- address resolution */
964 ib_conn->device = NULL;
966 iser_conn->state = ISER_CONN_PENDING;
968 ib_conn->beacon.wr_id = ISER_BEACON_WRID;
969 ib_conn->beacon.opcode = IB_WR_SEND;
971 ib_conn->cma_id = rdma_create_id(iser_cma_handler,
972 (void *)iser_conn,
973 RDMA_PS_TCP, IB_QPT_RC);
974 if (IS_ERR(ib_conn->cma_id)) {
975 err = PTR_ERR(ib_conn->cma_id);
976 iser_err("rdma_create_id failed: %d\n", err);
977 goto id_failure;
980 err = rdma_resolve_addr(ib_conn->cma_id, src_addr, dst_addr, 1000);
981 if (err) {
982 iser_err("rdma_resolve_addr failed: %d\n", err);
983 goto addr_failure;
986 if (!non_blocking) {
987 wait_for_completion_interruptible(&iser_conn->up_completion);
989 if (iser_conn->state != ISER_CONN_UP) {
990 err = -EIO;
991 goto connect_failure;
994 mutex_unlock(&iser_conn->state_mutex);
996 mutex_lock(&ig.connlist_mutex);
997 list_add(&iser_conn->conn_list, &ig.connlist);
998 mutex_unlock(&ig.connlist_mutex);
999 return 0;
1001 id_failure:
1002 ib_conn->cma_id = NULL;
1003 addr_failure:
1004 iser_conn->state = ISER_CONN_DOWN;
1005 connect_failure:
1006 mutex_unlock(&iser_conn->state_mutex);
1007 iser_conn_release(iser_conn);
1008 return err;
1011 int iser_post_recvl(struct iser_conn *iser_conn)
1013 struct ib_recv_wr rx_wr, *rx_wr_failed;
1014 struct ib_conn *ib_conn = &iser_conn->ib_conn;
1015 struct ib_sge sge;
1016 int ib_ret;
1018 sge.addr = iser_conn->login_resp_dma;
1019 sge.length = ISER_RX_LOGIN_SIZE;
1020 sge.lkey = ib_conn->device->mr->lkey;
1022 rx_wr.wr_id = (uintptr_t)iser_conn->login_resp_buf;
1023 rx_wr.sg_list = &sge;
1024 rx_wr.num_sge = 1;
1025 rx_wr.next = NULL;
1027 ib_conn->post_recv_buf_count++;
1028 ib_ret = ib_post_recv(ib_conn->qp, &rx_wr, &rx_wr_failed);
1029 if (ib_ret) {
1030 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1031 ib_conn->post_recv_buf_count--;
1033 return ib_ret;
1036 int iser_post_recvm(struct iser_conn *iser_conn, int count)
1038 struct ib_recv_wr *rx_wr, *rx_wr_failed;
1039 int i, ib_ret;
1040 struct ib_conn *ib_conn = &iser_conn->ib_conn;
1041 unsigned int my_rx_head = iser_conn->rx_desc_head;
1042 struct iser_rx_desc *rx_desc;
1044 for (rx_wr = ib_conn->rx_wr, i = 0; i < count; i++, rx_wr++) {
1045 rx_desc = &iser_conn->rx_descs[my_rx_head];
1046 rx_wr->wr_id = (uintptr_t)rx_desc;
1047 rx_wr->sg_list = &rx_desc->rx_sg;
1048 rx_wr->num_sge = 1;
1049 rx_wr->next = rx_wr + 1;
1050 my_rx_head = (my_rx_head + 1) & iser_conn->qp_max_recv_dtos_mask;
1053 rx_wr--;
1054 rx_wr->next = NULL; /* mark end of work requests list */
1056 ib_conn->post_recv_buf_count += count;
1057 ib_ret = ib_post_recv(ib_conn->qp, ib_conn->rx_wr, &rx_wr_failed);
1058 if (ib_ret) {
1059 iser_err("ib_post_recv failed ret=%d\n", ib_ret);
1060 ib_conn->post_recv_buf_count -= count;
1061 } else
1062 iser_conn->rx_desc_head = my_rx_head;
1063 return ib_ret;
1068 * iser_start_send - Initiate a Send DTO operation
1070 * returns 0 on success, -1 on failure
1072 int iser_post_send(struct ib_conn *ib_conn, struct iser_tx_desc *tx_desc,
1073 bool signal)
1075 int ib_ret;
1076 struct ib_send_wr send_wr, *send_wr_failed;
1078 ib_dma_sync_single_for_device(ib_conn->device->ib_device,
1079 tx_desc->dma_addr, ISER_HEADERS_LEN,
1080 DMA_TO_DEVICE);
1082 send_wr.next = NULL;
1083 send_wr.wr_id = (uintptr_t)tx_desc;
1084 send_wr.sg_list = tx_desc->tx_sg;
1085 send_wr.num_sge = tx_desc->num_sge;
1086 send_wr.opcode = IB_WR_SEND;
1087 send_wr.send_flags = signal ? IB_SEND_SIGNALED : 0;
1089 ib_ret = ib_post_send(ib_conn->qp, &send_wr, &send_wr_failed);
1090 if (ib_ret)
1091 iser_err("ib_post_send failed, ret:%d\n", ib_ret);
1093 return ib_ret;
1097 * is_iser_tx_desc - Indicate if the completion wr_id
1098 * is a TX descriptor or not.
1099 * @iser_conn: iser connection
1100 * @wr_id: completion WR identifier
1102 * Since we cannot rely on wc opcode in FLUSH errors
1103 * we must work around it by checking if the wr_id address
1104 * falls in the iser connection rx_descs buffer. If so
1105 * it is an RX descriptor, otherwize it is a TX.
1107 static inline bool
1108 is_iser_tx_desc(struct iser_conn *iser_conn, void *wr_id)
1110 void *start = iser_conn->rx_descs;
1111 int len = iser_conn->num_rx_descs * sizeof(*iser_conn->rx_descs);
1113 if (wr_id >= start && wr_id < start + len)
1114 return false;
1116 return true;
1120 * iser_handle_comp_error() - Handle error completion
1121 * @ib_conn: connection RDMA resources
1122 * @wc: work completion
1124 * Notes: We may handle a FLUSH error completion and in this case
1125 * we only cleanup in case TX type was DATAOUT. For non-FLUSH
1126 * error completion we should also notify iscsi layer that
1127 * connection is failed (in case we passed bind stage).
1129 static void
1130 iser_handle_comp_error(struct ib_conn *ib_conn,
1131 struct ib_wc *wc)
1133 void *wr_id = (void *)(uintptr_t)wc->wr_id;
1134 struct iser_conn *iser_conn = container_of(ib_conn, struct iser_conn,
1135 ib_conn);
1137 if (wc->status != IB_WC_WR_FLUSH_ERR)
1138 if (iser_conn->iscsi_conn)
1139 iscsi_conn_failure(iser_conn->iscsi_conn,
1140 ISCSI_ERR_CONN_FAILED);
1142 if (wc->wr_id == ISER_FASTREG_LI_WRID)
1143 return;
1145 if (is_iser_tx_desc(iser_conn, wr_id)) {
1146 struct iser_tx_desc *desc = wr_id;
1148 if (desc->type == ISCSI_TX_DATAOUT)
1149 kmem_cache_free(ig.desc_cache, desc);
1150 } else {
1151 ib_conn->post_recv_buf_count--;
1156 * iser_handle_wc - handle a single work completion
1157 * @wc: work completion
1159 * Soft-IRQ context, work completion can be either
1160 * SEND or RECV, and can turn out successful or
1161 * with error (or flush error).
1163 static void iser_handle_wc(struct ib_wc *wc)
1165 struct ib_conn *ib_conn;
1166 struct iser_tx_desc *tx_desc;
1167 struct iser_rx_desc *rx_desc;
1169 ib_conn = wc->qp->qp_context;
1170 if (likely(wc->status == IB_WC_SUCCESS)) {
1171 if (wc->opcode == IB_WC_RECV) {
1172 rx_desc = (struct iser_rx_desc *)(uintptr_t)wc->wr_id;
1173 iser_rcv_completion(rx_desc, wc->byte_len,
1174 ib_conn);
1175 } else
1176 if (wc->opcode == IB_WC_SEND) {
1177 tx_desc = (struct iser_tx_desc *)(uintptr_t)wc->wr_id;
1178 iser_snd_completion(tx_desc, ib_conn);
1179 } else {
1180 iser_err("Unknown wc opcode %d\n", wc->opcode);
1182 } else {
1183 if (wc->status != IB_WC_WR_FLUSH_ERR)
1184 iser_err("%s (%d): wr id %llx vend_err %x\n",
1185 ib_wc_status_msg(wc->status), wc->status,
1186 wc->wr_id, wc->vendor_err);
1187 else
1188 iser_dbg("%s (%d): wr id %llx\n",
1189 ib_wc_status_msg(wc->status), wc->status,
1190 wc->wr_id);
1192 if (wc->wr_id == ISER_BEACON_WRID)
1193 /* all flush errors were consumed */
1194 complete(&ib_conn->flush_comp);
1195 else
1196 iser_handle_comp_error(ib_conn, wc);
1201 * iser_cq_tasklet_fn - iSER completion polling loop
1202 * @data: iSER completion context
1204 * Soft-IRQ context, polling connection CQ until
1205 * either CQ was empty or we exausted polling budget
1207 static void iser_cq_tasklet_fn(unsigned long data)
1209 struct iser_comp *comp = (struct iser_comp *)data;
1210 struct ib_cq *cq = comp->cq;
1211 struct ib_wc *const wcs = comp->wcs;
1212 int i, n, completed = 0;
1214 while ((n = ib_poll_cq(cq, ARRAY_SIZE(comp->wcs), wcs)) > 0) {
1215 for (i = 0; i < n; i++)
1216 iser_handle_wc(&wcs[i]);
1218 completed += n;
1219 if (completed >= iser_cq_poll_limit)
1220 break;
1224 * It is assumed here that arming CQ only once its empty
1225 * would not cause interrupts to be missed.
1227 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1229 iser_dbg("got %d completions\n", completed);
1232 static void iser_cq_callback(struct ib_cq *cq, void *cq_context)
1234 struct iser_comp *comp = cq_context;
1236 tasklet_schedule(&comp->tasklet);
1239 u8 iser_check_task_pi_status(struct iscsi_iser_task *iser_task,
1240 enum iser_data_dir cmd_dir, sector_t *sector)
1242 struct iser_mem_reg *reg = &iser_task->rdma_reg[cmd_dir];
1243 struct fast_reg_descriptor *desc = reg->mem_h;
1244 unsigned long sector_size = iser_task->sc->device->sector_size;
1245 struct ib_mr_status mr_status;
1246 int ret;
1248 if (desc && desc->reg_indicators & ISER_FASTREG_PROTECTED) {
1249 desc->reg_indicators &= ~ISER_FASTREG_PROTECTED;
1250 ret = ib_check_mr_status(desc->pi_ctx->sig_mr,
1251 IB_MR_CHECK_SIG_STATUS, &mr_status);
1252 if (ret) {
1253 pr_err("ib_check_mr_status failed, ret %d\n", ret);
1254 goto err;
1257 if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
1258 sector_t sector_off = mr_status.sig_err.sig_err_offset;
1260 do_div(sector_off, sector_size + 8);
1261 *sector = scsi_get_lba(iser_task->sc) + sector_off;
1263 pr_err("PI error found type %d at sector %llx "
1264 "expected %x vs actual %x\n",
1265 mr_status.sig_err.err_type,
1266 (unsigned long long)*sector,
1267 mr_status.sig_err.expected,
1268 mr_status.sig_err.actual);
1270 switch (mr_status.sig_err.err_type) {
1271 case IB_SIG_BAD_GUARD:
1272 return 0x1;
1273 case IB_SIG_BAD_REFTAG:
1274 return 0x3;
1275 case IB_SIG_BAD_APPTAG:
1276 return 0x2;
1281 return 0;
1282 err:
1283 /* Not alot we can do here, return ambiguous guard error */
1284 return 0x1;