projects / deliverable / linux.git / blob
? search:


8a7eb9f98a0c89501d4342d1cd15f3d0d00599a1
[deliverable/linux.git] / drivers / infiniband / ulp / srp / ib_srp.c
1 /*
2 * Copyright (c) 2005 Cisco Systems. All rights reserved.
3 *
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 * SOFTWARE.
31 */
32
33 #define pr_fmt(fmt) PFX fmt
34
35 #include <linux/module.h>
36 #include <linux/init.h>
37 #include <linux/slab.h>
38 #include <linux/err.h>
39 #include <linux/string.h>
40 #include <linux/parser.h>
41 #include <linux/random.h>
42 #include <linux/jiffies.h>
43
44 #include <linux/atomic.h>
45
46 #include <scsi/scsi.h>
47 #include <scsi/scsi_device.h>
48 #include <scsi/scsi_dbg.h>
49 #include <scsi/srp.h>
50 #include <scsi/scsi_transport_srp.h>
51
52 #include "ib_srp.h"
53
54 #define DRV_NAME "ib_srp"
55 #define PFX DRV_NAME ": "
56 #define DRV_VERSION "0.2"
57 #define DRV_RELDATE "November 1, 2005"
58
59 MODULE_AUTHOR("Roland Dreier");
60 MODULE_DESCRIPTION("InfiniBand SCSI RDMA Protocol initiator "
61 "v" DRV_VERSION " (" DRV_RELDATE ")");
62 MODULE_LICENSE("Dual BSD/GPL");
63
64 static unsigned int srp_sg_tablesize;
65 static unsigned int cmd_sg_entries;
66 static unsigned int indirect_sg_entries;
67 static bool allow_ext_sg;
68 static int topspin_workarounds = 1;
69
70 module_param(srp_sg_tablesize, uint, 0444);
71 MODULE_PARM_DESC(srp_sg_tablesize, "Deprecated name for cmd_sg_entries");
72
73 module_param(cmd_sg_entries, uint, 0444);
74 MODULE_PARM_DESC(cmd_sg_entries,
75 "Default number of gather/scatter entries in the SRP command (default is 12, max 255)");
76
77 module_param(indirect_sg_entries, uint, 0444);
78 MODULE_PARM_DESC(indirect_sg_entries,
79 "Default max number of gather/scatter entries (default is 12, max is " __stringify(SCSI_MAX_SG_CHAIN_SEGMENTS) ")");
80
81 module_param(allow_ext_sg, bool, 0444);
82 MODULE_PARM_DESC(allow_ext_sg,
83 "Default behavior when there are more than cmd_sg_entries S/G entries after mapping; fails the request when false (default false)");
84
85 module_param(topspin_workarounds, int, 0444);
86 MODULE_PARM_DESC(topspin_workarounds,
87 "Enable workarounds for Topspin/Cisco SRP target bugs if != 0");
88
89 static void srp_add_one(struct ib_device *device);
90 static void srp_remove_one(struct ib_device *device);
91 static void srp_recv_completion(struct ib_cq *cq, void *target_ptr);
92 static void srp_send_completion(struct ib_cq *cq, void *target_ptr);
93 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event);
94
95 static struct scsi_transport_template *ib_srp_transport_template;
96
97 static struct ib_client srp_client = {
98 .name = "srp",
99 .add = srp_add_one,
100 .remove = srp_remove_one
101 };
102
103 static struct ib_sa_client srp_sa_client;
104
105 static inline struct srp_target_port *host_to_target(struct Scsi_Host *host)
106 {
107 return (struct srp_target_port *) host->hostdata;
108 }
109
110 static const char *srp_target_info(struct Scsi_Host *host)
111 {
112 return host_to_target(host)->target_name;
113 }
114
115 static int srp_target_is_topspin(struct srp_target_port *target)
116 {
117 static const u8 topspin_oui[3] = { 0x00, 0x05, 0xad };
118 static const u8 cisco_oui[3] = { 0x00, 0x1b, 0x0d };
119
120 return topspin_workarounds &&
121 (!memcmp(&target->ioc_guid, topspin_oui, sizeof topspin_oui) ||
122 !memcmp(&target->ioc_guid, cisco_oui, sizeof cisco_oui));
123 }
124
125 static struct srp_iu *srp_alloc_iu(struct srp_host *host, size_t size,
126 gfp_t gfp_mask,
127 enum dma_data_direction direction)
128 {
129 struct srp_iu *iu;
130
131 iu = kmalloc(sizeof *iu, gfp_mask);
132 if (!iu)
133 goto out;
134
135 iu->buf = kzalloc(size, gfp_mask);
136 if (!iu->buf)
137 goto out_free_iu;
138
139 iu->dma = ib_dma_map_single(host->srp_dev->dev, iu->buf, size,
140 direction);
141 if (ib_dma_mapping_error(host->srp_dev->dev, iu->dma))
142 goto out_free_buf;
143
144 iu->size = size;
145 iu->direction = direction;
146
147 return iu;
148
149 out_free_buf:
150 kfree(iu->buf);
151 out_free_iu:
152 kfree(iu);
153 out:
154 return NULL;
155 }
156
157 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
158 {
159 if (!iu)
160 return;
161
162 ib_dma_unmap_single(host->srp_dev->dev, iu->dma, iu->size,
163 iu->direction);
164 kfree(iu->buf);
165 kfree(iu);
166 }
167
168 static void srp_qp_event(struct ib_event *event, void *context)
169 {
170 pr_debug("QP event %d\n", event->event);
171 }
172
173 static int srp_init_qp(struct srp_target_port *target,
174 struct ib_qp *qp)
175 {
176 struct ib_qp_attr *attr;
177 int ret;
178
179 attr = kmalloc(sizeof *attr, GFP_KERNEL);
180 if (!attr)
181 return -ENOMEM;
182
183 ret = ib_find_pkey(target->srp_host->srp_dev->dev,
184 target->srp_host->port,
185 be16_to_cpu(target->path.pkey),
186 &attr->pkey_index);
187 if (ret)
188 goto out;
189
190 attr->qp_state = IB_QPS_INIT;
191 attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
192 IB_ACCESS_REMOTE_WRITE);
193 attr->port_num = target->srp_host->port;
194
195 ret = ib_modify_qp(qp, attr,
196 IB_QP_STATE |
197 IB_QP_PKEY_INDEX |
198 IB_QP_ACCESS_FLAGS |
199 IB_QP_PORT);
200
201 out:
202 kfree(attr);
203 return ret;
204 }
205
206 static int srp_new_cm_id(struct srp_target_port *target)
207 {
208 struct ib_cm_id *new_cm_id;
209
210 new_cm_id = ib_create_cm_id(target->srp_host->srp_dev->dev,
211 srp_cm_handler, target);
212 if (IS_ERR(new_cm_id))
213 return PTR_ERR(new_cm_id);
214
215 if (target->cm_id)
216 ib_destroy_cm_id(target->cm_id);
217 target->cm_id = new_cm_id;
218
219 return 0;
220 }
221
222 static int srp_create_target_ib(struct srp_target_port *target)
223 {
224 struct ib_qp_init_attr *init_attr;
225 struct ib_cq *recv_cq, *send_cq;
226 struct ib_qp *qp;
227 int ret;
228
229 init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
230 if (!init_attr)
231 return -ENOMEM;
232
233 recv_cq = ib_create_cq(target->srp_host->srp_dev->dev,
234 srp_recv_completion, NULL, target, SRP_RQ_SIZE, 0);
235 if (IS_ERR(recv_cq)) {
236 ret = PTR_ERR(recv_cq);
237 goto err;
238 }
239
240 send_cq = ib_create_cq(target->srp_host->srp_dev->dev,
241 srp_send_completion, NULL, target, SRP_SQ_SIZE, 0);
242 if (IS_ERR(send_cq)) {
243 ret = PTR_ERR(send_cq);
244 goto err_recv_cq;
245 }
246
247 ib_req_notify_cq(recv_cq, IB_CQ_NEXT_COMP);
248
249 init_attr->event_handler = srp_qp_event;
250 init_attr->cap.max_send_wr = SRP_SQ_SIZE;
251 init_attr->cap.max_recv_wr = SRP_RQ_SIZE;
252 init_attr->cap.max_recv_sge = 1;
253 init_attr->cap.max_send_sge = 1;
254 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
255 init_attr->qp_type = IB_QPT_RC;
256 init_attr->send_cq = send_cq;
257 init_attr->recv_cq = recv_cq;
258
259 qp = ib_create_qp(target->srp_host->srp_dev->pd, init_attr);
260 if (IS_ERR(qp)) {
261 ret = PTR_ERR(qp);
262 goto err_send_cq;
263 }
264
265 ret = srp_init_qp(target, qp);
266 if (ret)
267 goto err_qp;
268
269 if (target->qp)
270 ib_destroy_qp(target->qp);
271 if (target->recv_cq)
272 ib_destroy_cq(target->recv_cq);
273 if (target->send_cq)
274 ib_destroy_cq(target->send_cq);
275
276 target->qp = qp;
277 target->recv_cq = recv_cq;
278 target->send_cq = send_cq;
279
280 kfree(init_attr);
281 return 0;
282
283 err_qp:
284 ib_destroy_qp(qp);
285
286 err_send_cq:
287 ib_destroy_cq(send_cq);
288
289 err_recv_cq:
290 ib_destroy_cq(recv_cq);
291
292 err:
293 kfree(init_attr);
294 return ret;
295 }
296
297 static void srp_free_target_ib(struct srp_target_port *target)
298 {
299 int i;
300
301 ib_destroy_qp(target->qp);
302 ib_destroy_cq(target->send_cq);
303 ib_destroy_cq(target->recv_cq);
304
305 target->qp = NULL;
306 target->send_cq = target->recv_cq = NULL;
307
308 for (i = 0; i < SRP_RQ_SIZE; ++i)
309 srp_free_iu(target->srp_host, target->rx_ring[i]);
310 for (i = 0; i < SRP_SQ_SIZE; ++i)
311 srp_free_iu(target->srp_host, target->tx_ring[i]);
312 }
313
314 static void srp_path_rec_completion(int status,
315 struct ib_sa_path_rec *pathrec,
316 void *target_ptr)
317 {
318 struct srp_target_port *target = target_ptr;
319
320 target->status = status;
321 if (status)
322 shost_printk(KERN_ERR, target->scsi_host,
323 PFX "Got failed path rec status %d\n", status);
324 else
325 target->path = *pathrec;
326 complete(&target->done);
327 }
328
329 static int srp_lookup_path(struct srp_target_port *target)
330 {
331 target->path.numb_path = 1;
332
333 init_completion(&target->done);
334
335 target->path_query_id = ib_sa_path_rec_get(&srp_sa_client,
336 target->srp_host->srp_dev->dev,
337 target->srp_host->port,
338 &target->path,
339 IB_SA_PATH_REC_SERVICE_ID |
340 IB_SA_PATH_REC_DGID |
341 IB_SA_PATH_REC_SGID |
342 IB_SA_PATH_REC_NUMB_PATH |
343 IB_SA_PATH_REC_PKEY,
344 SRP_PATH_REC_TIMEOUT_MS,
345 GFP_KERNEL,
346 srp_path_rec_completion,
347 target, &target->path_query);
348 if (target->path_query_id < 0)
349 return target->path_query_id;
350
351 wait_for_completion(&target->done);
352
353 if (target->status < 0)
354 shost_printk(KERN_WARNING, target->scsi_host,
355 PFX "Path record query failed\n");
356
357 return target->status;
358 }
359
360 static int srp_send_req(struct srp_target_port *target)
361 {
362 struct {
363 struct ib_cm_req_param param;
364 struct srp_login_req priv;
365 } *req = NULL;
366 int status;
367
368 req = kzalloc(sizeof *req, GFP_KERNEL);
369 if (!req)
370 return -ENOMEM;
371
372 req->param.primary_path = &target->path;
373 req->param.alternate_path = NULL;
374 req->param.service_id = target->service_id;
375 req->param.qp_num = target->qp->qp_num;
376 req->param.qp_type = target->qp->qp_type;
377 req->param.private_data = &req->priv;
378 req->param.private_data_len = sizeof req->priv;
379 req->param.flow_control = 1;
380
381 get_random_bytes(&req->param.starting_psn, 4);
382 req->param.starting_psn &= 0xffffff;
383
384 /*
385 * Pick some arbitrary defaults here; we could make these
386 * module parameters if anyone cared about setting them.
387 */
388 req->param.responder_resources = 4;
389 req->param.remote_cm_response_timeout = 20;
390 req->param.local_cm_response_timeout = 20;
391 req->param.retry_count = 7;
392 req->param.rnr_retry_count = 7;
393 req->param.max_cm_retries = 15;
394
395 req->priv.opcode = SRP_LOGIN_REQ;
396 req->priv.tag = 0;
397 req->priv.req_it_iu_len = cpu_to_be32(target->max_iu_len);
398 req->priv.req_buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
399 SRP_BUF_FORMAT_INDIRECT);
400 /*
401 * In the published SRP specification (draft rev. 16a), the
402 * port identifier format is 8 bytes of ID extension followed
403 * by 8 bytes of GUID. Older drafts put the two halves in the
404 * opposite order, so that the GUID comes first.
405 *
406 * Targets conforming to these obsolete drafts can be
407 * recognized by the I/O Class they report.
408 */
409 if (target->io_class == SRP_REV10_IB_IO_CLASS) {
410 memcpy(req->priv.initiator_port_id,
411 &target->path.sgid.global.interface_id, 8);
412 memcpy(req->priv.initiator_port_id + 8,
413 &target->initiator_ext, 8);
414 memcpy(req->priv.target_port_id, &target->ioc_guid, 8);
415 memcpy(req->priv.target_port_id + 8, &target->id_ext, 8);
416 } else {
417 memcpy(req->priv.initiator_port_id,
418 &target->initiator_ext, 8);
419 memcpy(req->priv.initiator_port_id + 8,
420 &target->path.sgid.global.interface_id, 8);
421 memcpy(req->priv.target_port_id, &target->id_ext, 8);
422 memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
423 }
424
425 /*
426 * Topspin/Cisco SRP targets will reject our login unless we
427 * zero out the first 8 bytes of our initiator port ID and set
428 * the second 8 bytes to the local node GUID.
429 */
430 if (srp_target_is_topspin(target)) {
431 shost_printk(KERN_DEBUG, target->scsi_host,
432 PFX "Topspin/Cisco initiator port ID workaround "
433 "activated for target GUID %016llx\n",
434 (unsigned long long) be64_to_cpu(target->ioc_guid));
435 memset(req->priv.initiator_port_id, 0, 8);
436 memcpy(req->priv.initiator_port_id + 8,
437 &target->srp_host->srp_dev->dev->node_guid, 8);
438 }
439
440 status = ib_send_cm_req(target->cm_id, &req->param);
441
442 kfree(req);
443
444 return status;
445 }
446
447 static bool srp_queue_remove_work(struct srp_target_port *target)
448 {
449 bool changed = false;
450
451 spin_lock_irq(&target->lock);
452 if (target->state != SRP_TARGET_REMOVED) {
453 target->state = SRP_TARGET_REMOVED;
454 changed = true;
455 }
456 spin_unlock_irq(&target->lock);
457
458 if (changed)
459 queue_work(system_long_wq, &target->remove_work);
460
461 return changed;
462 }
463
464 static bool srp_change_conn_state(struct srp_target_port *target,
465 bool connected)
466 {
467 bool changed = false;
468
469 spin_lock_irq(&target->lock);
470 if (target->connected != connected) {
471 target->connected = connected;
472 changed = true;
473 }
474 spin_unlock_irq(&target->lock);
475
476 return changed;
477 }
478
479 static void srp_disconnect_target(struct srp_target_port *target)
480 {
481 if (srp_change_conn_state(target, false)) {
482 /* XXX should send SRP_I_LOGOUT request */
483
484 if (ib_send_cm_dreq(target->cm_id, NULL, 0)) {
485 shost_printk(KERN_DEBUG, target->scsi_host,
486 PFX "Sending CM DREQ failed\n");
487 }
488 }
489 }
490
491 static void srp_free_req_data(struct srp_target_port *target)
492 {
493 struct ib_device *ibdev = target->srp_host->srp_dev->dev;
494 struct srp_request *req;
495 int i;
496
497 for (i = 0, req = target->req_ring; i < SRP_CMD_SQ_SIZE; ++i, ++req) {
498 kfree(req->fmr_list);
499 kfree(req->map_page);
500 if (req->indirect_dma_addr) {
501 ib_dma_unmap_single(ibdev, req->indirect_dma_addr,
502 target->indirect_size,
503 DMA_TO_DEVICE);
504 }
505 kfree(req->indirect_desc);
506 }
507 }
508
509 /**
510 * srp_del_scsi_host_attr() - Remove attributes defined in the host template.
511 * @shost: SCSI host whose attributes to remove from sysfs.
512 *
513 * Note: Any attributes defined in the host template and that did not exist
514 * before invocation of this function will be ignored.
515 */
516 static void srp_del_scsi_host_attr(struct Scsi_Host *shost)
517 {
518 struct device_attribute **attr;
519
520 for (attr = shost->hostt->shost_attrs; attr && *attr; ++attr)
521 device_remove_file(&shost->shost_dev, *attr);
522 }
523
524 static void srp_remove_target(struct srp_target_port *target)
525 {
526 WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
527
528 srp_del_scsi_host_attr(target->scsi_host);
529 srp_remove_host(target->scsi_host);
530 scsi_remove_host(target->scsi_host);
531 srp_disconnect_target(target);
532 ib_destroy_cm_id(target->cm_id);
533 srp_free_target_ib(target);
534 srp_free_req_data(target);
535 scsi_host_put(target->scsi_host);
536 }
537
538 static void srp_remove_work(struct work_struct *work)
539 {
540 struct srp_target_port *target =
541 container_of(work, struct srp_target_port, remove_work);
542
543 WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
544
545 spin_lock(&target->srp_host->target_lock);
546 list_del(&target->list);
547 spin_unlock(&target->srp_host->target_lock);
548
549 srp_remove_target(target);
550 }
551
552 static void srp_rport_delete(struct srp_rport *rport)
553 {
554 struct srp_target_port *target = rport->lld_data;
555
556 srp_queue_remove_work(target);
557 }
558
559 static int srp_connect_target(struct srp_target_port *target)
560 {
561 int retries = 3;
562 int ret;
563
564 WARN_ON_ONCE(target->connected);
565
566 target->qp_in_error = false;
567
568 ret = srp_lookup_path(target);
569 if (ret)
570 return ret;
571
572 while (1) {
573 init_completion(&target->done);
574 ret = srp_send_req(target);
575 if (ret)
576 return ret;
577 wait_for_completion(&target->done);
578
579 /*
580 * The CM event handling code will set status to
581 * SRP_PORT_REDIRECT if we get a port redirect REJ
582 * back, or SRP_DLID_REDIRECT if we get a lid/qp
583 * redirect REJ back.
584 */
585 switch (target->status) {
586 case 0:
587 srp_change_conn_state(target, true);
588 return 0;
589
590 case SRP_PORT_REDIRECT:
591 ret = srp_lookup_path(target);
592 if (ret)
593 return ret;
594 break;
595
596 case SRP_DLID_REDIRECT:
597 break;
598
599 case SRP_STALE_CONN:
600 /* Our current CM id was stale, and is now in timewait.
601 * Try to reconnect with a new one.
602 */
603 if (!retries-- || srp_new_cm_id(target)) {
604 shost_printk(KERN_ERR, target->scsi_host, PFX
605 "giving up on stale connection\n");
606 target->status = -ECONNRESET;
607 return target->status;
608 }
609
610 shost_printk(KERN_ERR, target->scsi_host, PFX
611 "retrying stale connection\n");
612 break;
613
614 default:
615 return target->status;
616 }
617 }
618 }
619
620 static void srp_unmap_data(struct scsi_cmnd *scmnd,
621 struct srp_target_port *target,
622 struct srp_request *req)
623 {
624 struct ib_device *ibdev = target->srp_host->srp_dev->dev;
625 struct ib_pool_fmr **pfmr;
626
627 if (!scsi_sglist(scmnd) ||
628 (scmnd->sc_data_direction != DMA_TO_DEVICE &&
629 scmnd->sc_data_direction != DMA_FROM_DEVICE))
630 return;
631
632 pfmr = req->fmr_list;
633 while (req->nfmr--)
634 ib_fmr_pool_unmap(*pfmr++);
635
636 ib_dma_unmap_sg(ibdev, scsi_sglist(scmnd), scsi_sg_count(scmnd),
637 scmnd->sc_data_direction);
638 }
639
640 /**
641 * srp_claim_req - Take ownership of the scmnd associated with a request.
642 * @target: SRP target port.
643 * @req: SRP request.
644 * @scmnd: If NULL, take ownership of @req->scmnd. If not NULL, only take
645 * ownership of @req->scmnd if it equals @scmnd.
646 *
647 * Return value:
648 * Either NULL or a pointer to the SCSI command the caller became owner of.
649 */
650 static struct scsi_cmnd *srp_claim_req(struct srp_target_port *target,
651 struct srp_request *req,
652 struct scsi_cmnd *scmnd)
653 {
654 unsigned long flags;
655
656 spin_lock_irqsave(&target->lock, flags);
657 if (!scmnd) {
658 scmnd = req->scmnd;
659 req->scmnd = NULL;
660 } else if (req->scmnd == scmnd) {
661 req->scmnd = NULL;
662 } else {
663 scmnd = NULL;
664 }
665 spin_unlock_irqrestore(&target->lock, flags);
666
667 return scmnd;
668 }
669
670 /**
671 * srp_free_req() - Unmap data and add request to the free request list.
672 */
673 static void srp_free_req(struct srp_target_port *target,
674 struct srp_request *req, struct scsi_cmnd *scmnd,
675 s32 req_lim_delta)
676 {
677 unsigned long flags;
678
679 srp_unmap_data(scmnd, target, req);
680
681 spin_lock_irqsave(&target->lock, flags);
682 target->req_lim += req_lim_delta;
683 list_add_tail(&req->list, &target->free_reqs);
684 spin_unlock_irqrestore(&target->lock, flags);
685 }
686
687 static void srp_reset_req(struct srp_target_port *target, struct srp_request *req)
688 {
689 struct scsi_cmnd *scmnd = srp_claim_req(target, req, NULL);
690
691 if (scmnd) {
692 srp_free_req(target, req, scmnd, 0);
693 scmnd->result = DID_RESET << 16;
694 scmnd->scsi_done(scmnd);
695 }
696 }
697
698 static int srp_reconnect_target(struct srp_target_port *target)
699 {
700 struct Scsi_Host *shost = target->scsi_host;
701 int i, ret;
702
703 scsi_target_block(&shost->shost_gendev);
704
705 srp_disconnect_target(target);
706 /*
707 * Now get a new local CM ID so that we avoid confusing the target in
708 * case things are really fouled up. Doing so also ensures that all CM
709 * callbacks will have finished before a new QP is allocated.
710 */
711 ret = srp_new_cm_id(target);
712 /*
713 * Whether or not creating a new CM ID succeeded, create a new
714 * QP. This guarantees that all completion callback function
715 * invocations have finished before request resetting starts.
716 */
717 if (ret == 0)
718 ret = srp_create_target_ib(target);
719 else
720 srp_create_target_ib(target);
721
722 for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
723 struct srp_request *req = &target->req_ring[i];
724 if (req->scmnd)
725 srp_reset_req(target, req);
726 }
727
728 INIT_LIST_HEAD(&target->free_tx);
729 for (i = 0; i < SRP_SQ_SIZE; ++i)
730 list_add(&target->tx_ring[i]->list, &target->free_tx);
731
732 if (ret == 0)
733 ret = srp_connect_target(target);
734
735 scsi_target_unblock(&shost->shost_gendev, ret == 0 ? SDEV_RUNNING :
736 SDEV_TRANSPORT_OFFLINE);
737
738 if (ret)
739 goto err;
740
741 shost_printk(KERN_INFO, target->scsi_host, PFX "reconnect succeeded\n");
742
743 return ret;
744
745 err:
746 shost_printk(KERN_ERR, target->scsi_host,
747 PFX "reconnect failed (%d), removing target port.\n", ret);
748
749 /*
750 * We couldn't reconnect, so kill our target port off.
751 * However, we have to defer the real removal because we
752 * are in the context of the SCSI error handler now, which
753 * will deadlock if we call scsi_remove_host().
754 */
755 srp_queue_remove_work(target);
756
757 return ret;
758 }
759
760 static void srp_map_desc(struct srp_map_state *state, dma_addr_t dma_addr,
761 unsigned int dma_len, u32 rkey)
762 {
763 struct srp_direct_buf *desc = state->desc;
764
765 desc->va = cpu_to_be64(dma_addr);
766 desc->key = cpu_to_be32(rkey);
767 desc->len = cpu_to_be32(dma_len);
768
769 state->total_len += dma_len;
770 state->desc++;
771 state->ndesc++;
772 }
773
774 static int srp_map_finish_fmr(struct srp_map_state *state,
775 struct srp_target_port *target)
776 {
777 struct srp_device *dev = target->srp_host->srp_dev;
778 struct ib_pool_fmr *fmr;
779 u64 io_addr = 0;
780
781 if (!state->npages)
782 return 0;
783
784 if (state->npages == 1) {
785 srp_map_desc(state, state->base_dma_addr, state->fmr_len,
786 target->rkey);
787 state->npages = state->fmr_len = 0;
788 return 0;
789 }
790
791 fmr = ib_fmr_pool_map_phys(dev->fmr_pool, state->pages,
792 state->npages, io_addr);
793 if (IS_ERR(fmr))
794 return PTR_ERR(fmr);
795
796 *state->next_fmr++ = fmr;
797 state->nfmr++;
798
799 srp_map_desc(state, 0, state->fmr_len, fmr->fmr->rkey);
800 state->npages = state->fmr_len = 0;
801 return 0;
802 }
803
804 static void srp_map_update_start(struct srp_map_state *state,
805 struct scatterlist *sg, int sg_index,
806 dma_addr_t dma_addr)
807 {
808 state->unmapped_sg = sg;
809 state->unmapped_index = sg_index;
810 state->unmapped_addr = dma_addr;
811 }
812
813 static int srp_map_sg_entry(struct srp_map_state *state,
814 struct srp_target_port *target,
815 struct scatterlist *sg, int sg_index,
816 int use_fmr)
817 {
818 struct srp_device *dev = target->srp_host->srp_dev;
819 struct ib_device *ibdev = dev->dev;
820 dma_addr_t dma_addr = ib_sg_dma_address(ibdev, sg);
821 unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
822 unsigned int len;
823 int ret;
824
825 if (!dma_len)
826 return 0;
827
828 if (use_fmr == SRP_MAP_NO_FMR) {
829 /* Once we're in direct map mode for a request, we don't
830 * go back to FMR mode, so no need to update anything
831 * other than the descriptor.
832 */
833 srp_map_desc(state, dma_addr, dma_len, target->rkey);
834 return 0;
835 }
836
837 /* If we start at an offset into the FMR page, don't merge into
838 * the current FMR. Finish it out, and use the kernel's MR for this
839 * sg entry. This is to avoid potential bugs on some SRP targets
840 * that were never quite defined, but went away when the initiator
841 * avoided using FMR on such page fragments.
842 */
843 if (dma_addr & ~dev->fmr_page_mask || dma_len > dev->fmr_max_size) {
844 ret = srp_map_finish_fmr(state, target);
845 if (ret)
846 return ret;
847
848 srp_map_desc(state, dma_addr, dma_len, target->rkey);
849 srp_map_update_start(state, NULL, 0, 0);
850 return 0;
851 }
852
853 /* If this is the first sg to go into the FMR, save our position.
854 * We need to know the first unmapped entry, its index, and the
855 * first unmapped address within that entry to be able to restart
856 * mapping after an error.
857 */
858 if (!state->unmapped_sg)
859 srp_map_update_start(state, sg, sg_index, dma_addr);
860
861 while (dma_len) {
862 if (state->npages == SRP_FMR_SIZE) {
863 ret = srp_map_finish_fmr(state, target);
864 if (ret)
865 return ret;
866
867 srp_map_update_start(state, sg, sg_index, dma_addr);
868 }
869
870 len = min_t(unsigned int, dma_len, dev->fmr_page_size);
871
872 if (!state->npages)
873 state->base_dma_addr = dma_addr;
874 state->pages[state->npages++] = dma_addr;
875 state->fmr_len += len;
876 dma_addr += len;
877 dma_len -= len;
878 }
879
880 /* If the last entry of the FMR wasn't a full page, then we need to
881 * close it out and start a new one -- we can only merge at page
882 * boundries.
883 */
884 ret = 0;
885 if (len != dev->fmr_page_size) {
886 ret = srp_map_finish_fmr(state, target);
887 if (!ret)
888 srp_map_update_start(state, NULL, 0, 0);
889 }
890 return ret;
891 }
892
893 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_target_port *target,
894 struct srp_request *req)
895 {
896 struct scatterlist *scat, *sg;
897 struct srp_cmd *cmd = req->cmd->buf;
898 int i, len, nents, count, use_fmr;
899 struct srp_device *dev;
900 struct ib_device *ibdev;
901 struct srp_map_state state;
902 struct srp_indirect_buf *indirect_hdr;
903 u32 table_len;
904 u8 fmt;
905
906 if (!scsi_sglist(scmnd) || scmnd->sc_data_direction == DMA_NONE)
907 return sizeof (struct srp_cmd);
908
909 if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
910 scmnd->sc_data_direction != DMA_TO_DEVICE) {
911 shost_printk(KERN_WARNING, target->scsi_host,
912 PFX "Unhandled data direction %d\n",
913 scmnd->sc_data_direction);
914 return -EINVAL;
915 }
916
917 nents = scsi_sg_count(scmnd);
918 scat = scsi_sglist(scmnd);
919
920 dev = target->srp_host->srp_dev;
921 ibdev = dev->dev;
922
923 count = ib_dma_map_sg(ibdev, scat, nents, scmnd->sc_data_direction);
924 if (unlikely(count == 0))
925 return -EIO;
926
927 fmt = SRP_DATA_DESC_DIRECT;
928 len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
929
930 if (count == 1) {
931 /*
932 * The midlayer only generated a single gather/scatter
933 * entry, or DMA mapping coalesced everything to a
934 * single entry. So a direct descriptor along with
935 * the DMA MR suffices.
936 */
937 struct srp_direct_buf *buf = (void *) cmd->add_data;
938
939 buf->va = cpu_to_be64(ib_sg_dma_address(ibdev, scat));
940 buf->key = cpu_to_be32(target->rkey);
941 buf->len = cpu_to_be32(ib_sg_dma_len(ibdev, scat));
942
943 req->nfmr = 0;
944 goto map_complete;
945 }
946
947 /* We have more than one scatter/gather entry, so build our indirect
948 * descriptor table, trying to merge as many entries with FMR as we
949 * can.
950 */
951 indirect_hdr = (void *) cmd->add_data;
952
953 ib_dma_sync_single_for_cpu(ibdev, req->indirect_dma_addr,
954 target->indirect_size, DMA_TO_DEVICE);
955
956 memset(&state, 0, sizeof(state));
957 state.desc = req->indirect_desc;
958 state.pages = req->map_page;
959 state.next_fmr = req->fmr_list;
960
961 use_fmr = dev->fmr_pool ? SRP_MAP_ALLOW_FMR : SRP_MAP_NO_FMR;
962
963 for_each_sg(scat, sg, count, i) {
964 if (srp_map_sg_entry(&state, target, sg, i, use_fmr)) {
965 /* FMR mapping failed, so backtrack to the first
966 * unmapped entry and continue on without using FMR.
967 */
968 dma_addr_t dma_addr;
969 unsigned int dma_len;
970
971 backtrack:
972 sg = state.unmapped_sg;
973 i = state.unmapped_index;
974
975 dma_addr = ib_sg_dma_address(ibdev, sg);
976 dma_len = ib_sg_dma_len(ibdev, sg);
977 dma_len -= (state.unmapped_addr - dma_addr);
978 dma_addr = state.unmapped_addr;
979 use_fmr = SRP_MAP_NO_FMR;
980 srp_map_desc(&state, dma_addr, dma_len, target->rkey);
981 }
982 }
983
984 if (use_fmr == SRP_MAP_ALLOW_FMR && srp_map_finish_fmr(&state, target))
985 goto backtrack;
986
987 /* We've mapped the request, now pull as much of the indirect
988 * descriptor table as we can into the command buffer. If this
989 * target is not using an external indirect table, we are
990 * guaranteed to fit into the command, as the SCSI layer won't
991 * give us more S/G entries than we allow.
992 */
993 req->nfmr = state.nfmr;
994 if (state.ndesc == 1) {
995 /* FMR mapping was able to collapse this to one entry,
996 * so use a direct descriptor.
997 */
998 struct srp_direct_buf *buf = (void *) cmd->add_data;
999
1000 *buf = req->indirect_desc[0];
1001 goto map_complete;
1002 }
1003
1004 if (unlikely(target->cmd_sg_cnt < state.ndesc &&
1005 !target->allow_ext_sg)) {
1006 shost_printk(KERN_ERR, target->scsi_host,
1007 "Could not fit S/G list into SRP_CMD\n");
1008 return -EIO;
1009 }
1010
1011 count = min(state.ndesc, target->cmd_sg_cnt);
1012 table_len = state.ndesc * sizeof (struct srp_direct_buf);
1013
1014 fmt = SRP_DATA_DESC_INDIRECT;
1015 len = sizeof(struct srp_cmd) + sizeof (struct srp_indirect_buf);
1016 len += count * sizeof (struct srp_direct_buf);
1017
1018 memcpy(indirect_hdr->desc_list, req->indirect_desc,
1019 count * sizeof (struct srp_direct_buf));
1020
1021 indirect_hdr->table_desc.va = cpu_to_be64(req->indirect_dma_addr);
1022 indirect_hdr->table_desc.key = cpu_to_be32(target->rkey);
1023 indirect_hdr->table_desc.len = cpu_to_be32(table_len);
1024 indirect_hdr->len = cpu_to_be32(state.total_len);
1025
1026 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1027 cmd->data_out_desc_cnt = count;
1028 else
1029 cmd->data_in_desc_cnt = count;
1030
1031 ib_dma_sync_single_for_device(ibdev, req->indirect_dma_addr, table_len,
1032 DMA_TO_DEVICE);
1033
1034 map_complete:
1035 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1036 cmd->buf_fmt = fmt << 4;
1037 else
1038 cmd->buf_fmt = fmt;
1039
1040 return len;
1041 }
1042
1043 /*
1044 * Return an IU and possible credit to the free pool
1045 */
1046 static void srp_put_tx_iu(struct srp_target_port *target, struct srp_iu *iu,
1047 enum srp_iu_type iu_type)
1048 {
1049 unsigned long flags;
1050
1051 spin_lock_irqsave(&target->lock, flags);
1052 list_add(&iu->list, &target->free_tx);
1053 if (iu_type != SRP_IU_RSP)
1054 ++target->req_lim;
1055 spin_unlock_irqrestore(&target->lock, flags);
1056 }
1057
1058 /*
1059 * Must be called with target->lock held to protect req_lim and free_tx.
1060 * If IU is not sent, it must be returned using srp_put_tx_iu().
1061 *
1062 * Note:
1063 * An upper limit for the number of allocated information units for each
1064 * request type is:
1065 * - SRP_IU_CMD: SRP_CMD_SQ_SIZE, since the SCSI mid-layer never queues
1066 * more than Scsi_Host.can_queue requests.
1067 * - SRP_IU_TSK_MGMT: SRP_TSK_MGMT_SQ_SIZE.
1068 * - SRP_IU_RSP: 1, since a conforming SRP target never sends more than
1069 * one unanswered SRP request to an initiator.
1070 */
1071 static struct srp_iu *__srp_get_tx_iu(struct srp_target_port *target,
1072 enum srp_iu_type iu_type)
1073 {
1074 s32 rsv = (iu_type == SRP_IU_TSK_MGMT) ? 0 : SRP_TSK_MGMT_SQ_SIZE;
1075 struct srp_iu *iu;
1076
1077 srp_send_completion(target->send_cq, target);
1078
1079 if (list_empty(&target->free_tx))
1080 return NULL;
1081
1082 /* Initiator responses to target requests do not consume credits */
1083 if (iu_type != SRP_IU_RSP) {
1084 if (target->req_lim <= rsv) {
1085 ++target->zero_req_lim;
1086 return NULL;
1087 }
1088
1089 --target->req_lim;
1090 }
1091
1092 iu = list_first_entry(&target->free_tx, struct srp_iu, list);
1093 list_del(&iu->list);
1094 return iu;
1095 }
1096
1097 static int srp_post_send(struct srp_target_port *target,
1098 struct srp_iu *iu, int len)
1099 {
1100 struct ib_sge list;
1101 struct ib_send_wr wr, *bad_wr;
1102
1103 list.addr = iu->dma;
1104 list.length = len;
1105 list.lkey = target->lkey;
1106
1107 wr.next = NULL;
1108 wr.wr_id = (uintptr_t) iu;
1109 wr.sg_list = &list;
1110 wr.num_sge = 1;
1111 wr.opcode = IB_WR_SEND;
1112 wr.send_flags = IB_SEND_SIGNALED;
1113
1114 return ib_post_send(target->qp, &wr, &bad_wr);
1115 }
1116
1117 static int srp_post_recv(struct srp_target_port *target, struct srp_iu *iu)
1118 {
1119 struct ib_recv_wr wr, *bad_wr;
1120 struct ib_sge list;
1121
1122 list.addr = iu->dma;
1123 list.length = iu->size;
1124 list.lkey = target->lkey;
1125
1126 wr.next = NULL;
1127 wr.wr_id = (uintptr_t) iu;
1128 wr.sg_list = &list;
1129 wr.num_sge = 1;
1130
1131 return ib_post_recv(target->qp, &wr, &bad_wr);
1132 }
1133
1134 static void srp_process_rsp(struct srp_target_port *target, struct srp_rsp *rsp)
1135 {
1136 struct srp_request *req;
1137 struct scsi_cmnd *scmnd;
1138 unsigned long flags;
1139
1140 if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
1141 spin_lock_irqsave(&target->lock, flags);
1142 target->req_lim += be32_to_cpu(rsp->req_lim_delta);
1143 spin_unlock_irqrestore(&target->lock, flags);
1144
1145 target->tsk_mgmt_status = -1;
1146 if (be32_to_cpu(rsp->resp_data_len) >= 4)
1147 target->tsk_mgmt_status = rsp->data[3];
1148 complete(&target->tsk_mgmt_done);
1149 } else {
1150 req = &target->req_ring[rsp->tag];
1151 scmnd = srp_claim_req(target, req, NULL);
1152 if (!scmnd) {
1153 shost_printk(KERN_ERR, target->scsi_host,
1154 "Null scmnd for RSP w/tag %016llx\n",
1155 (unsigned long long) rsp->tag);
1156
1157 spin_lock_irqsave(&target->lock, flags);
1158 target->req_lim += be32_to_cpu(rsp->req_lim_delta);
1159 spin_unlock_irqrestore(&target->lock, flags);
1160
1161 return;
1162 }
1163 scmnd->result = rsp->status;
1164
1165 if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
1166 memcpy(scmnd->sense_buffer, rsp->data +
1167 be32_to_cpu(rsp->resp_data_len),
1168 min_t(int, be32_to_cpu(rsp->sense_data_len),
1169 SCSI_SENSE_BUFFERSIZE));
1170 }
1171
1172 if (rsp->flags & (SRP_RSP_FLAG_DOOVER | SRP_RSP_FLAG_DOUNDER))
1173 scsi_set_resid(scmnd, be32_to_cpu(rsp->data_out_res_cnt));
1174 else if (rsp->flags & (SRP_RSP_FLAG_DIOVER | SRP_RSP_FLAG_DIUNDER))
1175 scsi_set_resid(scmnd, be32_to_cpu(rsp->data_in_res_cnt));
1176
1177 srp_free_req(target, req, scmnd,
1178 be32_to_cpu(rsp->req_lim_delta));
1179
1180 scmnd->host_scribble = NULL;
1181 scmnd->scsi_done(scmnd);
1182 }
1183 }
1184
1185 static int srp_response_common(struct srp_target_port *target, s32 req_delta,
1186 void *rsp, int len)
1187 {
1188 struct ib_device *dev = target->srp_host->srp_dev->dev;
1189 unsigned long flags;
1190 struct srp_iu *iu;
1191 int err;
1192
1193 spin_lock_irqsave(&target->lock, flags);
1194 target->req_lim += req_delta;
1195 iu = __srp_get_tx_iu(target, SRP_IU_RSP);
1196 spin_unlock_irqrestore(&target->lock, flags);
1197
1198 if (!iu) {
1199 shost_printk(KERN_ERR, target->scsi_host, PFX
1200 "no IU available to send response\n");
1201 return 1;
1202 }
1203
1204 ib_dma_sync_single_for_cpu(dev, iu->dma, len, DMA_TO_DEVICE);
1205 memcpy(iu->buf, rsp, len);
1206 ib_dma_sync_single_for_device(dev, iu->dma, len, DMA_TO_DEVICE);
1207
1208 err = srp_post_send(target, iu, len);
1209 if (err) {
1210 shost_printk(KERN_ERR, target->scsi_host, PFX
1211 "unable to post response: %d\n", err);
1212 srp_put_tx_iu(target, iu, SRP_IU_RSP);
1213 }
1214
1215 return err;
1216 }
1217
1218 static void srp_process_cred_req(struct srp_target_port *target,
1219 struct srp_cred_req *req)
1220 {
1221 struct srp_cred_rsp rsp = {
1222 .opcode = SRP_CRED_RSP,
1223 .tag = req->tag,
1224 };
1225 s32 delta = be32_to_cpu(req->req_lim_delta);
1226
1227 if (srp_response_common(target, delta, &rsp, sizeof rsp))
1228 shost_printk(KERN_ERR, target->scsi_host, PFX
1229 "problems processing SRP_CRED_REQ\n");
1230 }
1231
1232 static void srp_process_aer_req(struct srp_target_port *target,
1233 struct srp_aer_req *req)
1234 {
1235 struct srp_aer_rsp rsp = {
1236 .opcode = SRP_AER_RSP,
1237 .tag = req->tag,
1238 };
1239 s32 delta = be32_to_cpu(req->req_lim_delta);
1240
1241 shost_printk(KERN_ERR, target->scsi_host, PFX
1242 "ignoring AER for LUN %llu\n", be64_to_cpu(req->lun));
1243
1244 if (srp_response_common(target, delta, &rsp, sizeof rsp))
1245 shost_printk(KERN_ERR, target->scsi_host, PFX
1246 "problems processing SRP_AER_REQ\n");
1247 }
1248
1249 static void srp_handle_recv(struct srp_target_port *target, struct ib_wc *wc)
1250 {
1251 struct ib_device *dev = target->srp_host->srp_dev->dev;
1252 struct srp_iu *iu = (struct srp_iu *) (uintptr_t) wc->wr_id;
1253 int res;
1254 u8 opcode;
1255
1256 ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_ti_iu_len,
1257 DMA_FROM_DEVICE);
1258
1259 opcode = *(u8 *) iu->buf;
1260
1261 if (0) {
1262 shost_printk(KERN_ERR, target->scsi_host,
1263 PFX "recv completion, opcode 0x%02x\n", opcode);
1264 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 8, 1,
1265 iu->buf, wc->byte_len, true);
1266 }
1267
1268 switch (opcode) {
1269 case SRP_RSP:
1270 srp_process_rsp(target, iu->buf);
1271 break;
1272
1273 case SRP_CRED_REQ:
1274 srp_process_cred_req(target, iu->buf);
1275 break;
1276
1277 case SRP_AER_REQ:
1278 srp_process_aer_req(target, iu->buf);
1279 break;
1280
1281 case SRP_T_LOGOUT:
1282 /* XXX Handle target logout */
1283 shost_printk(KERN_WARNING, target->scsi_host,
1284 PFX "Got target logout request\n");
1285 break;
1286
1287 default:
1288 shost_printk(KERN_WARNING, target->scsi_host,
1289 PFX "Unhandled SRP opcode 0x%02x\n", opcode);
1290 break;
1291 }
1292
1293 ib_dma_sync_single_for_device(dev, iu->dma, target->max_ti_iu_len,
1294 DMA_FROM_DEVICE);
1295
1296 res = srp_post_recv(target, iu);
1297 if (res != 0)
1298 shost_printk(KERN_ERR, target->scsi_host,
1299 PFX "Recv failed with error code %d\n", res);
1300 }
1301
1302 static void srp_handle_qp_err(enum ib_wc_status wc_status,
1303 enum ib_wc_opcode wc_opcode,
1304 struct srp_target_port *target)
1305 {
1306 if (target->connected && !target->qp_in_error) {
1307 shost_printk(KERN_ERR, target->scsi_host,
1308 PFX "failed %s status %d\n",
1309 wc_opcode & IB_WC_RECV ? "receive" : "send",
1310 wc_status);
1311 }
1312 target->qp_in_error = true;
1313 }
1314
1315 static void srp_recv_completion(struct ib_cq *cq, void *target_ptr)
1316 {
1317 struct srp_target_port *target = target_ptr;
1318 struct ib_wc wc;
1319
1320 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1321 while (ib_poll_cq(cq, 1, &wc) > 0) {
1322 if (likely(wc.status == IB_WC_SUCCESS)) {
1323 srp_handle_recv(target, &wc);
1324 } else {
1325 srp_handle_qp_err(wc.status, wc.opcode, target);
1326 }
1327 }
1328 }
1329
1330 static void srp_send_completion(struct ib_cq *cq, void *target_ptr)
1331 {
1332 struct srp_target_port *target = target_ptr;
1333 struct ib_wc wc;
1334 struct srp_iu *iu;
1335
1336 while (ib_poll_cq(cq, 1, &wc) > 0) {
1337 if (likely(wc.status == IB_WC_SUCCESS)) {
1338 iu = (struct srp_iu *) (uintptr_t) wc.wr_id;
1339 list_add(&iu->list, &target->free_tx);
1340 } else {
1341 srp_handle_qp_err(wc.status, wc.opcode, target);
1342 }
1343 }
1344 }
1345
1346 static int srp_queuecommand(struct Scsi_Host *shost, struct scsi_cmnd *scmnd)
1347 {
1348 struct srp_target_port *target = host_to_target(shost);
1349 struct srp_request *req;
1350 struct srp_iu *iu;
1351 struct srp_cmd *cmd;
1352 struct ib_device *dev;
1353 unsigned long flags;
1354 int len;
1355
1356 spin_lock_irqsave(&target->lock, flags);
1357 iu = __srp_get_tx_iu(target, SRP_IU_CMD);
1358 if (!iu)
1359 goto err_unlock;
1360
1361 req = list_first_entry(&target->free_reqs, struct srp_request, list);
1362 list_del(&req->list);
1363 spin_unlock_irqrestore(&target->lock, flags);
1364
1365 dev = target->srp_host->srp_dev->dev;
1366 ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_iu_len,
1367 DMA_TO_DEVICE);
1368
1369 scmnd->result = 0;
1370 scmnd->host_scribble = (void *) req;
1371
1372 cmd = iu->buf;
1373 memset(cmd, 0, sizeof *cmd);
1374
1375 cmd->opcode = SRP_CMD;
1376 cmd->lun = cpu_to_be64((u64) scmnd->device->lun << 48);
1377 cmd->tag = req->index;
1378 memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
1379
1380 req->scmnd = scmnd;
1381 req->cmd = iu;
1382
1383 len = srp_map_data(scmnd, target, req);
1384 if (len < 0) {
1385 shost_printk(KERN_ERR, target->scsi_host,
1386 PFX "Failed to map data\n");
1387 goto err_iu;
1388 }
1389
1390 ib_dma_sync_single_for_device(dev, iu->dma, target->max_iu_len,
1391 DMA_TO_DEVICE);
1392
1393 if (srp_post_send(target, iu, len)) {
1394 shost_printk(KERN_ERR, target->scsi_host, PFX "Send failed\n");
1395 goto err_unmap;
1396 }
1397
1398 return 0;
1399
1400 err_unmap:
1401 srp_unmap_data(scmnd, target, req);
1402
1403 err_iu:
1404 srp_put_tx_iu(target, iu, SRP_IU_CMD);
1405
1406 spin_lock_irqsave(&target->lock, flags);
1407 list_add(&req->list, &target->free_reqs);
1408
1409 err_unlock:
1410 spin_unlock_irqrestore(&target->lock, flags);
1411
1412 return SCSI_MLQUEUE_HOST_BUSY;
1413 }
1414
1415 static int srp_alloc_iu_bufs(struct srp_target_port *target)
1416 {
1417 int i;
1418
1419 for (i = 0; i < SRP_RQ_SIZE; ++i) {
1420 target->rx_ring[i] = srp_alloc_iu(target->srp_host,
1421 target->max_ti_iu_len,
1422 GFP_KERNEL, DMA_FROM_DEVICE);
1423 if (!target->rx_ring[i])
1424 goto err;
1425 }
1426
1427 for (i = 0; i < SRP_SQ_SIZE; ++i) {
1428 target->tx_ring[i] = srp_alloc_iu(target->srp_host,
1429 target->max_iu_len,
1430 GFP_KERNEL, DMA_TO_DEVICE);
1431 if (!target->tx_ring[i])
1432 goto err;
1433
1434 list_add(&target->tx_ring[i]->list, &target->free_tx);
1435 }
1436
1437 return 0;
1438
1439 err:
1440 for (i = 0; i < SRP_RQ_SIZE; ++i) {
1441 srp_free_iu(target->srp_host, target->rx_ring[i]);
1442 target->rx_ring[i] = NULL;
1443 }
1444
1445 for (i = 0; i < SRP_SQ_SIZE; ++i) {
1446 srp_free_iu(target->srp_host, target->tx_ring[i]);
1447 target->tx_ring[i] = NULL;
1448 }
1449
1450 return -ENOMEM;
1451 }
1452
1453 static uint32_t srp_compute_rq_tmo(struct ib_qp_attr *qp_attr, int attr_mask)
1454 {
1455 uint64_t T_tr_ns, max_compl_time_ms;
1456 uint32_t rq_tmo_jiffies;
1457
1458 /*
1459 * According to section 11.2.4.2 in the IBTA spec (Modify Queue Pair,
1460 * table 91), both the QP timeout and the retry count have to be set
1461 * for RC QP's during the RTR to RTS transition.
1462 */
1463 WARN_ON_ONCE((attr_mask & (IB_QP_TIMEOUT | IB_QP_RETRY_CNT)) !=
1464 (IB_QP_TIMEOUT | IB_QP_RETRY_CNT));
1465
1466 /*
1467 * Set target->rq_tmo_jiffies to one second more than the largest time
1468 * it can take before an error completion is generated. See also
1469 * C9-140..142 in the IBTA spec for more information about how to
1470 * convert the QP Local ACK Timeout value to nanoseconds.
1471 */
1472 T_tr_ns = 4096 * (1ULL << qp_attr->timeout);
1473 max_compl_time_ms = qp_attr->retry_cnt * 4 * T_tr_ns;
1474 do_div(max_compl_time_ms, NSEC_PER_MSEC);
1475 rq_tmo_jiffies = msecs_to_jiffies(max_compl_time_ms + 1000);
1476
1477 return rq_tmo_jiffies;
1478 }
1479
1480 static void srp_cm_rep_handler(struct ib_cm_id *cm_id,
1481 struct srp_login_rsp *lrsp,
1482 struct srp_target_port *target)
1483 {
1484 struct ib_qp_attr *qp_attr = NULL;
1485 int attr_mask = 0;
1486 int ret;
1487 int i;
1488
1489 if (lrsp->opcode == SRP_LOGIN_RSP) {
1490 target->max_ti_iu_len = be32_to_cpu(lrsp->max_ti_iu_len);
1491 target->req_lim = be32_to_cpu(lrsp->req_lim_delta);
1492
1493 /*
1494 * Reserve credits for task management so we don't
1495 * bounce requests back to the SCSI mid-layer.
1496 */
1497 target->scsi_host->can_queue
1498 = min(target->req_lim - SRP_TSK_MGMT_SQ_SIZE,
1499 target->scsi_host->can_queue);
1500 } else {
1501 shost_printk(KERN_WARNING, target->scsi_host,
1502 PFX "Unhandled RSP opcode %#x\n", lrsp->opcode);
1503 ret = -ECONNRESET;
1504 goto error;
1505 }
1506
1507 if (!target->rx_ring[0]) {
1508 ret = srp_alloc_iu_bufs(target);
1509 if (ret)
1510 goto error;
1511 }
1512
1513 ret = -ENOMEM;
1514 qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
1515 if (!qp_attr)
1516 goto error;
1517
1518 qp_attr->qp_state = IB_QPS_RTR;
1519 ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1520 if (ret)
1521 goto error_free;
1522
1523 ret = ib_modify_qp(target->qp, qp_attr, attr_mask);
1524 if (ret)
1525 goto error_free;
1526
1527 for (i = 0; i < SRP_RQ_SIZE; i++) {
1528 struct srp_iu *iu = target->rx_ring[i];
1529 ret = srp_post_recv(target, iu);
1530 if (ret)
1531 goto error_free;
1532 }
1533
1534 qp_attr->qp_state = IB_QPS_RTS;
1535 ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1536 if (ret)
1537 goto error_free;
1538
1539 target->rq_tmo_jiffies = srp_compute_rq_tmo(qp_attr, attr_mask);
1540
1541 ret = ib_modify_qp(target->qp, qp_attr, attr_mask);
1542 if (ret)
1543 goto error_free;
1544
1545 ret = ib_send_cm_rtu(cm_id, NULL, 0);
1546
1547 error_free:
1548 kfree(qp_attr);
1549
1550 error:
1551 target->status = ret;
1552 }
1553
1554 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
1555 struct ib_cm_event *event,
1556 struct srp_target_port *target)
1557 {
1558 struct Scsi_Host *shost = target->scsi_host;
1559 struct ib_class_port_info *cpi;
1560 int opcode;
1561
1562 switch (event->param.rej_rcvd.reason) {
1563 case IB_CM_REJ_PORT_CM_REDIRECT:
1564 cpi = event->param.rej_rcvd.ari;
1565 target->path.dlid = cpi->redirect_lid;
1566 target->path.pkey = cpi->redirect_pkey;
1567 cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
1568 memcpy(target->path.dgid.raw, cpi->redirect_gid, 16);
1569
1570 target->status = target->path.dlid ?
1571 SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
1572 break;
1573
1574 case IB_CM_REJ_PORT_REDIRECT:
1575 if (srp_target_is_topspin(target)) {
1576 /*
1577 * Topspin/Cisco SRP gateways incorrectly send
1578 * reject reason code 25 when they mean 24
1579 * (port redirect).
1580 */
1581 memcpy(target->path.dgid.raw,
1582 event->param.rej_rcvd.ari, 16);
1583
1584 shost_printk(KERN_DEBUG, shost,
1585 PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
1586 (unsigned long long) be64_to_cpu(target->path.dgid.global.subnet_prefix),
1587 (unsigned long long) be64_to_cpu(target->path.dgid.global.interface_id));
1588
1589 target->status = SRP_PORT_REDIRECT;
1590 } else {
1591 shost_printk(KERN_WARNING, shost,
1592 " REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
1593 target->status = -ECONNRESET;
1594 }
1595 break;
1596
1597 case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
1598 shost_printk(KERN_WARNING, shost,
1599 " REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
1600 target->status = -ECONNRESET;
1601 break;
1602
1603 case IB_CM_REJ_CONSUMER_DEFINED:
1604 opcode = *(u8 *) event->private_data;
1605 if (opcode == SRP_LOGIN_REJ) {
1606 struct srp_login_rej *rej = event->private_data;
1607 u32 reason = be32_to_cpu(rej->reason);
1608
1609 if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
1610 shost_printk(KERN_WARNING, shost,
1611 PFX "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
1612 else
1613 shost_printk(KERN_WARNING, shost,
1614 PFX "SRP LOGIN REJECTED, reason 0x%08x\n", reason);
1615 } else
1616 shost_printk(KERN_WARNING, shost,
1617 " REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
1618 " opcode 0x%02x\n", opcode);
1619 target->status = -ECONNRESET;
1620 break;
1621
1622 case IB_CM_REJ_STALE_CONN:
1623 shost_printk(KERN_WARNING, shost, " REJ reason: stale connection\n");
1624 target->status = SRP_STALE_CONN;
1625 break;
1626
1627 default:
1628 shost_printk(KERN_WARNING, shost, " REJ reason 0x%x\n",
1629 event->param.rej_rcvd.reason);
1630 target->status = -ECONNRESET;
1631 }
1632 }
1633
1634 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
1635 {
1636 struct srp_target_port *target = cm_id->context;
1637 int comp = 0;
1638
1639 switch (event->event) {
1640 case IB_CM_REQ_ERROR:
1641 shost_printk(KERN_DEBUG, target->scsi_host,
1642 PFX "Sending CM REQ failed\n");
1643 comp = 1;
1644 target->status = -ECONNRESET;
1645 break;
1646
1647 case IB_CM_REP_RECEIVED:
1648 comp = 1;
1649 srp_cm_rep_handler(cm_id, event->private_data, target);
1650 break;
1651
1652 case IB_CM_REJ_RECEIVED:
1653 shost_printk(KERN_DEBUG, target->scsi_host, PFX "REJ received\n");
1654 comp = 1;
1655
1656 srp_cm_rej_handler(cm_id, event, target);
1657 break;
1658
1659 case IB_CM_DREQ_RECEIVED:
1660 shost_printk(KERN_WARNING, target->scsi_host,
1661 PFX "DREQ received - connection closed\n");
1662 srp_change_conn_state(target, false);
1663 if (ib_send_cm_drep(cm_id, NULL, 0))
1664 shost_printk(KERN_ERR, target->scsi_host,
1665 PFX "Sending CM DREP failed\n");
1666 break;
1667
1668 case IB_CM_TIMEWAIT_EXIT:
1669 shost_printk(KERN_ERR, target->scsi_host,
1670 PFX "connection closed\n");
1671
1672 target->status = 0;
1673 break;
1674
1675 case IB_CM_MRA_RECEIVED:
1676 case IB_CM_DREQ_ERROR:
1677 case IB_CM_DREP_RECEIVED:
1678 break;
1679
1680 default:
1681 shost_printk(KERN_WARNING, target->scsi_host,
1682 PFX "Unhandled CM event %d\n", event->event);
1683 break;
1684 }
1685
1686 if (comp)
1687 complete(&target->done);
1688
1689 return 0;
1690 }
1691
1692 static int srp_send_tsk_mgmt(struct srp_target_port *target,
1693 u64 req_tag, unsigned int lun, u8 func)
1694 {
1695 struct ib_device *dev = target->srp_host->srp_dev->dev;
1696 struct srp_iu *iu;
1697 struct srp_tsk_mgmt *tsk_mgmt;
1698
1699 if (!target->connected || target->qp_in_error)
1700 return -1;
1701
1702 init_completion(&target->tsk_mgmt_done);
1703
1704 spin_lock_irq(&target->lock);
1705 iu = __srp_get_tx_iu(target, SRP_IU_TSK_MGMT);
1706 spin_unlock_irq(&target->lock);
1707
1708 if (!iu)
1709 return -1;
1710
1711 ib_dma_sync_single_for_cpu(dev, iu->dma, sizeof *tsk_mgmt,
1712 DMA_TO_DEVICE);
1713 tsk_mgmt = iu->buf;
1714 memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
1715
1716 tsk_mgmt->opcode = SRP_TSK_MGMT;
1717 tsk_mgmt->lun = cpu_to_be64((u64) lun << 48);
1718 tsk_mgmt->tag = req_tag | SRP_TAG_TSK_MGMT;
1719 tsk_mgmt->tsk_mgmt_func = func;
1720 tsk_mgmt->task_tag = req_tag;
1721
1722 ib_dma_sync_single_for_device(dev, iu->dma, sizeof *tsk_mgmt,
1723 DMA_TO_DEVICE);
1724 if (srp_post_send(target, iu, sizeof *tsk_mgmt)) {
1725 srp_put_tx_iu(target, iu, SRP_IU_TSK_MGMT);
1726 return -1;
1727 }
1728
1729 if (!wait_for_completion_timeout(&target->tsk_mgmt_done,
1730 msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
1731 return -1;
1732
1733 return 0;
1734 }
1735
1736 static int srp_abort(struct scsi_cmnd *scmnd)
1737 {
1738 struct srp_target_port *target = host_to_target(scmnd->device->host);
1739 struct srp_request *req = (struct srp_request *) scmnd->host_scribble;
1740
1741 shost_printk(KERN_ERR, target->scsi_host, "SRP abort called\n");
1742
1743 if (!req || !srp_claim_req(target, req, scmnd))
1744 return FAILED;
1745 srp_send_tsk_mgmt(target, req->index, scmnd->device->lun,
1746 SRP_TSK_ABORT_TASK);
1747 srp_free_req(target, req, scmnd, 0);
1748 scmnd->result = DID_ABORT << 16;
1749 scmnd->scsi_done(scmnd);
1750
1751 return SUCCESS;
1752 }
1753
1754 static int srp_reset_device(struct scsi_cmnd *scmnd)
1755 {
1756 struct srp_target_port *target = host_to_target(scmnd->device->host);
1757 int i;
1758
1759 shost_printk(KERN_ERR, target->scsi_host, "SRP reset_device called\n");
1760
1761 if (srp_send_tsk_mgmt(target, SRP_TAG_NO_REQ, scmnd->device->lun,
1762 SRP_TSK_LUN_RESET))
1763 return FAILED;
1764 if (target->tsk_mgmt_status)
1765 return FAILED;
1766
1767 for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
1768 struct srp_request *req = &target->req_ring[i];
1769 if (req->scmnd && req->scmnd->device == scmnd->device)
1770 srp_reset_req(target, req);
1771 }
1772
1773 return SUCCESS;
1774 }
1775
1776 static int srp_reset_host(struct scsi_cmnd *scmnd)
1777 {
1778 struct srp_target_port *target = host_to_target(scmnd->device->host);
1779 int ret = FAILED;
1780
1781 shost_printk(KERN_ERR, target->scsi_host, PFX "SRP reset_host called\n");
1782
1783 if (!srp_reconnect_target(target))
1784 ret = SUCCESS;
1785
1786 return ret;
1787 }
1788
1789 static int srp_slave_configure(struct scsi_device *sdev)
1790 {
1791 struct Scsi_Host *shost = sdev->host;
1792 struct srp_target_port *target = host_to_target(shost);
1793 struct request_queue *q = sdev->request_queue;
1794 unsigned long timeout;
1795
1796 if (sdev->type == TYPE_DISK) {
1797 timeout = max_t(unsigned, 30 * HZ, target->rq_tmo_jiffies);
1798 blk_queue_rq_timeout(q, timeout);
1799 }
1800
1801 return 0;
1802 }
1803
1804 static ssize_t show_id_ext(struct device *dev, struct device_attribute *attr,
1805 char *buf)
1806 {
1807 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1808
1809 return sprintf(buf, "0x%016llx\n",
1810 (unsigned long long) be64_to_cpu(target->id_ext));
1811 }
1812
1813 static ssize_t show_ioc_guid(struct device *dev, struct device_attribute *attr,
1814 char *buf)
1815 {
1816 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1817
1818 return sprintf(buf, "0x%016llx\n",
1819 (unsigned long long) be64_to_cpu(target->ioc_guid));
1820 }
1821
1822 static ssize_t show_service_id(struct device *dev,
1823 struct device_attribute *attr, char *buf)
1824 {
1825 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1826
1827 return sprintf(buf, "0x%016llx\n",
1828 (unsigned long long) be64_to_cpu(target->service_id));
1829 }
1830
1831 static ssize_t show_pkey(struct device *dev, struct device_attribute *attr,
1832 char *buf)
1833 {
1834 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1835
1836 return sprintf(buf, "0x%04x\n", be16_to_cpu(target->path.pkey));
1837 }
1838
1839 static ssize_t show_dgid(struct device *dev, struct device_attribute *attr,
1840 char *buf)
1841 {
1842 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1843
1844 return sprintf(buf, "%pI6\n", target->path.dgid.raw);
1845 }
1846
1847 static ssize_t show_orig_dgid(struct device *dev,
1848 struct device_attribute *attr, char *buf)
1849 {
1850 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1851
1852 return sprintf(buf, "%pI6\n", target->orig_dgid);
1853 }
1854
1855 static ssize_t show_req_lim(struct device *dev,
1856 struct device_attribute *attr, char *buf)
1857 {
1858 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1859
1860 return sprintf(buf, "%d\n", target->req_lim);
1861 }
1862
1863 static ssize_t show_zero_req_lim(struct device *dev,
1864 struct device_attribute *attr, char *buf)
1865 {
1866 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1867
1868 return sprintf(buf, "%d\n", target->zero_req_lim);
1869 }
1870
1871 static ssize_t show_local_ib_port(struct device *dev,
1872 struct device_attribute *attr, char *buf)
1873 {
1874 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1875
1876 return sprintf(buf, "%d\n", target->srp_host->port);
1877 }
1878
1879 static ssize_t show_local_ib_device(struct device *dev,
1880 struct device_attribute *attr, char *buf)
1881 {
1882 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1883
1884 return sprintf(buf, "%s\n", target->srp_host->srp_dev->dev->name);
1885 }
1886
1887 static ssize_t show_cmd_sg_entries(struct device *dev,
1888 struct device_attribute *attr, char *buf)
1889 {
1890 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1891
1892 return sprintf(buf, "%u\n", target->cmd_sg_cnt);
1893 }
1894
1895 static ssize_t show_allow_ext_sg(struct device *dev,
1896 struct device_attribute *attr, char *buf)
1897 {
1898 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1899
1900 return sprintf(buf, "%s\n", target->allow_ext_sg ? "true" : "false");
1901 }
1902
1903 static DEVICE_ATTR(id_ext, S_IRUGO, show_id_ext, NULL);
1904 static DEVICE_ATTR(ioc_guid, S_IRUGO, show_ioc_guid, NULL);
1905 static DEVICE_ATTR(service_id, S_IRUGO, show_service_id, NULL);
1906 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1907 static DEVICE_ATTR(dgid, S_IRUGO, show_dgid, NULL);
1908 static DEVICE_ATTR(orig_dgid, S_IRUGO, show_orig_dgid, NULL);
1909 static DEVICE_ATTR(req_lim, S_IRUGO, show_req_lim, NULL);
1910 static DEVICE_ATTR(zero_req_lim, S_IRUGO, show_zero_req_lim, NULL);
1911 static DEVICE_ATTR(local_ib_port, S_IRUGO, show_local_ib_port, NULL);
1912 static DEVICE_ATTR(local_ib_device, S_IRUGO, show_local_ib_device, NULL);
1913 static DEVICE_ATTR(cmd_sg_entries, S_IRUGO, show_cmd_sg_entries, NULL);
1914 static DEVICE_ATTR(allow_ext_sg, S_IRUGO, show_allow_ext_sg, NULL);
1915
1916 static struct device_attribute *srp_host_attrs[] = {
1917 &dev_attr_id_ext,
1918 &dev_attr_ioc_guid,
1919 &dev_attr_service_id,
1920 &dev_attr_pkey,
1921 &dev_attr_dgid,
1922 &dev_attr_orig_dgid,
1923 &dev_attr_req_lim,
1924 &dev_attr_zero_req_lim,
1925 &dev_attr_local_ib_port,
1926 &dev_attr_local_ib_device,
1927 &dev_attr_cmd_sg_entries,
1928 &dev_attr_allow_ext_sg,
1929 NULL
1930 };
1931
1932 static struct scsi_host_template srp_template = {
1933 .module = THIS_MODULE,
1934 .name = "InfiniBand SRP initiator",
1935 .proc_name = DRV_NAME,
1936 .slave_configure = srp_slave_configure,
1937 .info = srp_target_info,
1938 .queuecommand = srp_queuecommand,
1939 .eh_abort_handler = srp_abort,
1940 .eh_device_reset_handler = srp_reset_device,
1941 .eh_host_reset_handler = srp_reset_host,
1942 .sg_tablesize = SRP_DEF_SG_TABLESIZE,
1943 .can_queue = SRP_CMD_SQ_SIZE,
1944 .this_id = -1,
1945 .cmd_per_lun = SRP_CMD_SQ_SIZE,
1946 .use_clustering = ENABLE_CLUSTERING,
1947 .shost_attrs = srp_host_attrs
1948 };
1949
1950 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
1951 {
1952 struct srp_rport_identifiers ids;
1953 struct srp_rport *rport;
1954
1955 sprintf(target->target_name, "SRP.T10:%016llX",
1956 (unsigned long long) be64_to_cpu(target->id_ext));
1957
1958 if (scsi_add_host(target->scsi_host, host->srp_dev->dev->dma_device))
1959 return -ENODEV;
1960
1961 memcpy(ids.port_id, &target->id_ext, 8);
1962 memcpy(ids.port_id + 8, &target->ioc_guid, 8);
1963 ids.roles = SRP_RPORT_ROLE_TARGET;
1964 rport = srp_rport_add(target->scsi_host, &ids);
1965 if (IS_ERR(rport)) {
1966 scsi_remove_host(target->scsi_host);
1967 return PTR_ERR(rport);
1968 }
1969
1970 rport->lld_data = target;
1971
1972 spin_lock(&host->target_lock);
1973 list_add_tail(&target->list, &host->target_list);
1974 spin_unlock(&host->target_lock);
1975
1976 target->state = SRP_TARGET_LIVE;
1977
1978 scsi_scan_target(&target->scsi_host->shost_gendev,
1979 0, target->scsi_id, SCAN_WILD_CARD, 0);
1980
1981 return 0;
1982 }
1983
1984 static void srp_release_dev(struct device *dev)
1985 {
1986 struct srp_host *host =
1987 container_of(dev, struct srp_host, dev);
1988
1989 complete(&host->released);
1990 }
1991
1992 static struct class srp_class = {
1993 .name = "infiniband_srp",
1994 .dev_release = srp_release_dev
1995 };
1996
1997 /*
1998 * Target ports are added by writing
1999 *
2000 * id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
2001 * pkey=<P_Key>,service_id=<service ID>
2002 *
2003 * to the add_target sysfs attribute.
2004 */
2005 enum {
2006 SRP_OPT_ERR = 0,
2007 SRP_OPT_ID_EXT = 1 << 0,
2008 SRP_OPT_IOC_GUID = 1 << 1,
2009 SRP_OPT_DGID = 1 << 2,
2010 SRP_OPT_PKEY = 1 << 3,
2011 SRP_OPT_SERVICE_ID = 1 << 4,
2012 SRP_OPT_MAX_SECT = 1 << 5,
2013 SRP_OPT_MAX_CMD_PER_LUN = 1 << 6,
2014 SRP_OPT_IO_CLASS = 1 << 7,
2015 SRP_OPT_INITIATOR_EXT = 1 << 8,
2016 SRP_OPT_CMD_SG_ENTRIES = 1 << 9,
2017 SRP_OPT_ALLOW_EXT_SG = 1 << 10,
2018 SRP_OPT_SG_TABLESIZE = 1 << 11,
2019 SRP_OPT_ALL = (SRP_OPT_ID_EXT |
2020 SRP_OPT_IOC_GUID |
2021 SRP_OPT_DGID |
2022 SRP_OPT_PKEY |
2023 SRP_OPT_SERVICE_ID),
2024 };
2025
2026 static const match_table_t srp_opt_tokens = {
2027 { SRP_OPT_ID_EXT, "id_ext=%s" },
2028 { SRP_OPT_IOC_GUID, "ioc_guid=%s" },
2029 { SRP_OPT_DGID, "dgid=%s" },
2030 { SRP_OPT_PKEY, "pkey=%x" },
2031 { SRP_OPT_SERVICE_ID, "service_id=%s" },
2032 { SRP_OPT_MAX_SECT, "max_sect=%d" },
2033 { SRP_OPT_MAX_CMD_PER_LUN, "max_cmd_per_lun=%d" },
2034 { SRP_OPT_IO_CLASS, "io_class=%x" },
2035 { SRP_OPT_INITIATOR_EXT, "initiator_ext=%s" },
2036 { SRP_OPT_CMD_SG_ENTRIES, "cmd_sg_entries=%u" },
2037 { SRP_OPT_ALLOW_EXT_SG, "allow_ext_sg=%u" },
2038 { SRP_OPT_SG_TABLESIZE, "sg_tablesize=%u" },
2039 { SRP_OPT_ERR, NULL }
2040 };
2041
2042 static int srp_parse_options(const char *buf, struct srp_target_port *target)
2043 {
2044 char *options, *sep_opt;
2045 char *p;
2046 char dgid[3];
2047 substring_t args[MAX_OPT_ARGS];
2048 int opt_mask = 0;
2049 int token;
2050 int ret = -EINVAL;
2051 int i;
2052
2053 options = kstrdup(buf, GFP_KERNEL);
2054 if (!options)
2055 return -ENOMEM;
2056
2057 sep_opt = options;
2058 while ((p = strsep(&sep_opt, ",")) != NULL) {
2059 if (!*p)
2060 continue;
2061
2062 token = match_token(p, srp_opt_tokens, args);
2063 opt_mask |= token;
2064
2065 switch (token) {
2066 case SRP_OPT_ID_EXT:
2067 p = match_strdup(args);
2068 if (!p) {
2069 ret = -ENOMEM;
2070 goto out;
2071 }
2072 target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2073 kfree(p);
2074 break;
2075
2076 case SRP_OPT_IOC_GUID:
2077 p = match_strdup(args);
2078 if (!p) {
2079 ret = -ENOMEM;
2080 goto out;
2081 }
2082 target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
2083 kfree(p);
2084 break;
2085
2086 case SRP_OPT_DGID:
2087 p = match_strdup(args);
2088 if (!p) {
2089 ret = -ENOMEM;
2090 goto out;
2091 }
2092 if (strlen(p) != 32) {
2093 pr_warn("bad dest GID parameter '%s'\n", p);
2094 kfree(p);
2095 goto out;
2096 }
2097
2098 for (i = 0; i < 16; ++i) {
2099 strlcpy(dgid, p + i * 2, 3);
2100 target->path.dgid.raw[i] = simple_strtoul(dgid, NULL, 16);
2101 }
2102 kfree(p);
2103 memcpy(target->orig_dgid, target->path.dgid.raw, 16);
2104 break;
2105
2106 case SRP_OPT_PKEY:
2107 if (match_hex(args, &token)) {
2108 pr_warn("bad P_Key parameter '%s'\n", p);
2109 goto out;
2110 }
2111 target->path.pkey = cpu_to_be16(token);
2112 break;
2113
2114 case SRP_OPT_SERVICE_ID:
2115 p = match_strdup(args);
2116 if (!p) {
2117 ret = -ENOMEM;
2118 goto out;
2119 }
2120 target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
2121 target->path.service_id = target->service_id;
2122 kfree(p);
2123 break;
2124
2125 case SRP_OPT_MAX_SECT:
2126 if (match_int(args, &token)) {
2127 pr_warn("bad max sect parameter '%s'\n", p);
2128 goto out;
2129 }
2130 target->scsi_host->max_sectors = token;
2131 break;
2132
2133 case SRP_OPT_MAX_CMD_PER_LUN:
2134 if (match_int(args, &token)) {
2135 pr_warn("bad max cmd_per_lun parameter '%s'\n",
2136 p);
2137 goto out;
2138 }
2139 target->scsi_host->cmd_per_lun = min(token, SRP_CMD_SQ_SIZE);
2140 break;
2141
2142 case SRP_OPT_IO_CLASS:
2143 if (match_hex(args, &token)) {
2144 pr_warn("bad IO class parameter '%s'\n", p);
2145 goto out;
2146 }
2147 if (token != SRP_REV10_IB_IO_CLASS &&
2148 token != SRP_REV16A_IB_IO_CLASS) {
2149 pr_warn("unknown IO class parameter value %x specified (use %x or %x).\n",
2150 token, SRP_REV10_IB_IO_CLASS,
2151 SRP_REV16A_IB_IO_CLASS);
2152 goto out;
2153 }
2154 target->io_class = token;
2155 break;
2156
2157 case SRP_OPT_INITIATOR_EXT:
2158 p = match_strdup(args);
2159 if (!p) {
2160 ret = -ENOMEM;
2161 goto out;
2162 }
2163 target->initiator_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2164 kfree(p);
2165 break;
2166
2167 case SRP_OPT_CMD_SG_ENTRIES:
2168 if (match_int(args, &token) || token < 1 || token > 255) {
2169 pr_warn("bad max cmd_sg_entries parameter '%s'\n",
2170 p);
2171 goto out;
2172 }
2173 target->cmd_sg_cnt = token;
2174 break;
2175
2176 case SRP_OPT_ALLOW_EXT_SG:
2177 if (match_int(args, &token)) {
2178 pr_warn("bad allow_ext_sg parameter '%s'\n", p);
2179 goto out;
2180 }
2181 target->allow_ext_sg = !!token;
2182 break;
2183
2184 case SRP_OPT_SG_TABLESIZE:
2185 if (match_int(args, &token) || token < 1 ||
2186 token > SCSI_MAX_SG_CHAIN_SEGMENTS) {
2187 pr_warn("bad max sg_tablesize parameter '%s'\n",
2188 p);
2189 goto out;
2190 }
2191 target->sg_tablesize = token;
2192 break;
2193
2194 default:
2195 pr_warn("unknown parameter or missing value '%s' in target creation request\n",
2196 p);
2197 goto out;
2198 }
2199 }
2200
2201 if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
2202 ret = 0;
2203 else
2204 for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
2205 if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
2206 !(srp_opt_tokens[i].token & opt_mask))
2207 pr_warn("target creation request is missing parameter '%s'\n",
2208 srp_opt_tokens[i].pattern);
2209
2210 out:
2211 kfree(options);
2212 return ret;
2213 }
2214
2215 static ssize_t srp_create_target(struct device *dev,
2216 struct device_attribute *attr,
2217 const char *buf, size_t count)
2218 {
2219 struct srp_host *host =
2220 container_of(dev, struct srp_host, dev);
2221 struct Scsi_Host *target_host;
2222 struct srp_target_port *target;
2223 struct ib_device *ibdev = host->srp_dev->dev;
2224 dma_addr_t dma_addr;
2225 int i, ret;
2226
2227 target_host = scsi_host_alloc(&srp_template,
2228 sizeof (struct srp_target_port));
2229 if (!target_host)
2230 return -ENOMEM;
2231
2232 target_host->transportt = ib_srp_transport_template;
2233 target_host->max_channel = 0;
2234 target_host->max_id = 1;
2235 target_host->max_lun = SRP_MAX_LUN;
2236 target_host->max_cmd_len = sizeof ((struct srp_cmd *) (void *) 0L)->cdb;
2237
2238 target = host_to_target(target_host);
2239
2240 target->io_class = SRP_REV16A_IB_IO_CLASS;
2241 target->scsi_host = target_host;
2242 target->srp_host = host;
2243 target->lkey = host->srp_dev->mr->lkey;
2244 target->rkey = host->srp_dev->mr->rkey;
2245 target->cmd_sg_cnt = cmd_sg_entries;
2246 target->sg_tablesize = indirect_sg_entries ? : cmd_sg_entries;
2247 target->allow_ext_sg = allow_ext_sg;
2248
2249 ret = srp_parse_options(buf, target);
2250 if (ret)
2251 goto err;
2252
2253 if (!host->srp_dev->fmr_pool && !target->allow_ext_sg &&
2254 target->cmd_sg_cnt < target->sg_tablesize) {
2255 pr_warn("No FMR pool and no external indirect descriptors, limiting sg_tablesize to cmd_sg_cnt\n");
2256 target->sg_tablesize = target->cmd_sg_cnt;
2257 }
2258
2259 target_host->sg_tablesize = target->sg_tablesize;
2260 target->indirect_size = target->sg_tablesize *
2261 sizeof (struct srp_direct_buf);
2262 target->max_iu_len = sizeof (struct srp_cmd) +
2263 sizeof (struct srp_indirect_buf) +
2264 target->cmd_sg_cnt * sizeof (struct srp_direct_buf);
2265
2266 INIT_WORK(&target->remove_work, srp_remove_work);
2267 spin_lock_init(&target->lock);
2268 INIT_LIST_HEAD(&target->free_tx);
2269 INIT_LIST_HEAD(&target->free_reqs);
2270 for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
2271 struct srp_request *req = &target->req_ring[i];
2272
2273 req->fmr_list = kmalloc(target->cmd_sg_cnt * sizeof (void *),
2274 GFP_KERNEL);
2275 req->map_page = kmalloc(SRP_FMR_SIZE * sizeof (void *),
2276 GFP_KERNEL);
2277 req->indirect_desc = kmalloc(target->indirect_size, GFP_KERNEL);
2278 if (!req->fmr_list || !req->map_page || !req->indirect_desc)
2279 goto err_free_mem;
2280
2281 dma_addr = ib_dma_map_single(ibdev, req->indirect_desc,
2282 target->indirect_size,
2283 DMA_TO_DEVICE);
2284 if (ib_dma_mapping_error(ibdev, dma_addr))
2285 goto err_free_mem;
2286
2287 req->indirect_dma_addr = dma_addr;
2288 req->index = i;
2289 list_add_tail(&req->list, &target->free_reqs);
2290 }
2291
2292 ib_query_gid(ibdev, host->port, 0, &target->path.sgid);
2293
2294 shost_printk(KERN_DEBUG, target->scsi_host, PFX
2295 "new target: id_ext %016llx ioc_guid %016llx pkey %04x "
2296 "service_id %016llx dgid %pI6\n",
2297 (unsigned long long) be64_to_cpu(target->id_ext),
2298 (unsigned long long) be64_to_cpu(target->ioc_guid),
2299 be16_to_cpu(target->path.pkey),
2300 (unsigned long long) be64_to_cpu(target->service_id),
2301 target->path.dgid.raw);
2302
2303 ret = srp_create_target_ib(target);
2304 if (ret)
2305 goto err_free_mem;
2306
2307 ret = srp_new_cm_id(target);
2308 if (ret)
2309 goto err_free_ib;
2310
2311 ret = srp_connect_target(target);
2312 if (ret) {
2313 shost_printk(KERN_ERR, target->scsi_host,
2314 PFX "Connection failed\n");
2315 goto err_cm_id;
2316 }
2317
2318 ret = srp_add_target(host, target);
2319 if (ret)
2320 goto err_disconnect;
2321
2322 return count;
2323
2324 err_disconnect:
2325 srp_disconnect_target(target);
2326
2327 err_cm_id:
2328 ib_destroy_cm_id(target->cm_id);
2329
2330 err_free_ib:
2331 srp_free_target_ib(target);
2332
2333 err_free_mem:
2334 srp_free_req_data(target);
2335
2336 err:
2337 scsi_host_put(target_host);
2338
2339 return ret;
2340 }
2341
2342 static DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
2343
2344 static ssize_t show_ibdev(struct device *dev, struct device_attribute *attr,
2345 char *buf)
2346 {
2347 struct srp_host *host = container_of(dev, struct srp_host, dev);
2348
2349 return sprintf(buf, "%s\n", host->srp_dev->dev->name);
2350 }
2351
2352 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
2353
2354 static ssize_t show_port(struct device *dev, struct device_attribute *attr,
2355 char *buf)
2356 {
2357 struct srp_host *host = container_of(dev, struct srp_host, dev);
2358
2359 return sprintf(buf, "%d\n", host->port);
2360 }
2361
2362 static DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
2363
2364 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
2365 {
2366 struct srp_host *host;
2367
2368 host = kzalloc(sizeof *host, GFP_KERNEL);
2369 if (!host)
2370 return NULL;
2371
2372 INIT_LIST_HEAD(&host->target_list);
2373 spin_lock_init(&host->target_lock);
2374 init_completion(&host->released);
2375 host->srp_dev = device;
2376 host->port = port;
2377
2378 host->dev.class = &srp_class;
2379 host->dev.parent = device->dev->dma_device;
2380 dev_set_name(&host->dev, "srp-%s-%d", device->dev->name, port);
2381
2382 if (device_register(&host->dev))
2383 goto free_host;
2384 if (device_create_file(&host->dev, &dev_attr_add_target))
2385 goto err_class;
2386 if (device_create_file(&host->dev, &dev_attr_ibdev))
2387 goto err_class;
2388 if (device_create_file(&host->dev, &dev_attr_port))
2389 goto err_class;
2390
2391 return host;
2392
2393 err_class:
2394 device_unregister(&host->dev);
2395
2396 free_host:
2397 kfree(host);
2398
2399 return NULL;
2400 }
2401
2402 static void srp_add_one(struct ib_device *device)
2403 {
2404 struct srp_device *srp_dev;
2405 struct ib_device_attr *dev_attr;
2406 struct ib_fmr_pool_param fmr_param;
2407 struct srp_host *host;
2408 int max_pages_per_fmr, fmr_page_shift, s, e, p;
2409
2410 dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
2411 if (!dev_attr)
2412 return;
2413
2414 if (ib_query_device(device, dev_attr)) {
2415 pr_warn("Query device failed for %s\n", device->name);
2416 goto free_attr;
2417 }
2418
2419 srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
2420 if (!srp_dev)
2421 goto free_attr;
2422
2423 /*
2424 * Use the smallest page size supported by the HCA, down to a
2425 * minimum of 4096 bytes. We're unlikely to build large sglists
2426 * out of smaller entries.
2427 */
2428 fmr_page_shift = max(12, ffs(dev_attr->page_size_cap) - 1);
2429 srp_dev->fmr_page_size = 1 << fmr_page_shift;
2430 srp_dev->fmr_page_mask = ~((u64) srp_dev->fmr_page_size - 1);
2431 srp_dev->fmr_max_size = srp_dev->fmr_page_size * SRP_FMR_SIZE;
2432
2433 INIT_LIST_HEAD(&srp_dev->dev_list);
2434
2435 srp_dev->dev = device;
2436 srp_dev->pd = ib_alloc_pd(device);
2437 if (IS_ERR(srp_dev->pd))
2438 goto free_dev;
2439
2440 srp_dev->mr = ib_get_dma_mr(srp_dev->pd,
2441 IB_ACCESS_LOCAL_WRITE |
2442 IB_ACCESS_REMOTE_READ |
2443 IB_ACCESS_REMOTE_WRITE);
2444 if (IS_ERR(srp_dev->mr))
2445 goto err_pd;
2446
2447 for (max_pages_per_fmr = SRP_FMR_SIZE;
2448 max_pages_per_fmr >= SRP_FMR_MIN_SIZE;
2449 max_pages_per_fmr /= 2, srp_dev->fmr_max_size /= 2) {
2450 memset(&fmr_param, 0, sizeof fmr_param);
2451 fmr_param.pool_size = SRP_FMR_POOL_SIZE;
2452 fmr_param.dirty_watermark = SRP_FMR_DIRTY_SIZE;
2453 fmr_param.cache = 1;
2454 fmr_param.max_pages_per_fmr = max_pages_per_fmr;
2455 fmr_param.page_shift = fmr_page_shift;
2456 fmr_param.access = (IB_ACCESS_LOCAL_WRITE |
2457 IB_ACCESS_REMOTE_WRITE |
2458 IB_ACCESS_REMOTE_READ);
2459
2460 srp_dev->fmr_pool = ib_create_fmr_pool(srp_dev->pd, &fmr_param);
2461 if (!IS_ERR(srp_dev->fmr_pool))
2462 break;
2463 }
2464
2465 if (IS_ERR(srp_dev->fmr_pool))
2466 srp_dev->fmr_pool = NULL;
2467
2468 if (device->node_type == RDMA_NODE_IB_SWITCH) {
2469 s = 0;
2470 e = 0;
2471 } else {
2472 s = 1;
2473 e = device->phys_port_cnt;
2474 }
2475
2476 for (p = s; p <= e; ++p) {
2477 host = srp_add_port(srp_dev, p);
2478 if (host)
2479 list_add_tail(&host->list, &srp_dev->dev_list);
2480 }
2481
2482 ib_set_client_data(device, &srp_client, srp_dev);
2483
2484 goto free_attr;
2485
2486 err_pd:
2487 ib_dealloc_pd(srp_dev->pd);
2488
2489 free_dev:
2490 kfree(srp_dev);
2491
2492 free_attr:
2493 kfree(dev_attr);
2494 }
2495
2496 static void srp_remove_one(struct ib_device *device)
2497 {
2498 struct srp_device *srp_dev;
2499 struct srp_host *host, *tmp_host;
2500 struct srp_target_port *target;
2501
2502 srp_dev = ib_get_client_data(device, &srp_client);
2503
2504 list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
2505 device_unregister(&host->dev);
2506 /*
2507 * Wait for the sysfs entry to go away, so that no new
2508 * target ports can be created.
2509 */
2510 wait_for_completion(&host->released);
2511
2512 /*
2513 * Remove all target ports.
2514 */
2515 spin_lock(&host->target_lock);
2516 list_for_each_entry(target, &host->target_list, list)
2517 srp_queue_remove_work(target);
2518 spin_unlock(&host->target_lock);
2519
2520 /*
2521 * Wait for target port removal tasks.
2522 */
2523 flush_workqueue(system_long_wq);
2524
2525 kfree(host);
2526 }
2527
2528 if (srp_dev->fmr_pool)
2529 ib_destroy_fmr_pool(srp_dev->fmr_pool);
2530 ib_dereg_mr(srp_dev->mr);
2531 ib_dealloc_pd(srp_dev->pd);
2532
2533 kfree(srp_dev);
2534 }
2535
2536 static struct srp_function_template ib_srp_transport_functions = {
2537 .rport_delete = srp_rport_delete,
2538 };
2539
2540 static int __init srp_init_module(void)
2541 {
2542 int ret;
2543
2544 BUILD_BUG_ON(FIELD_SIZEOF(struct ib_wc, wr_id) < sizeof(void *));
2545
2546 if (srp_sg_tablesize) {
2547 pr_warn("srp_sg_tablesize is deprecated, please use cmd_sg_entries\n");
2548 if (!cmd_sg_entries)
2549 cmd_sg_entries = srp_sg_tablesize;
2550 }
2551
2552 if (!cmd_sg_entries)
2553 cmd_sg_entries = SRP_DEF_SG_TABLESIZE;
2554
2555 if (cmd_sg_entries > 255) {
2556 pr_warn("Clamping cmd_sg_entries to 255\n");
2557 cmd_sg_entries = 255;
2558 }
2559
2560 if (!indirect_sg_entries)
2561 indirect_sg_entries = cmd_sg_entries;
2562 else if (indirect_sg_entries < cmd_sg_entries) {
2563 pr_warn("Bumping up indirect_sg_entries to match cmd_sg_entries (%u)\n",
2564 cmd_sg_entries);
2565 indirect_sg_entries = cmd_sg_entries;
2566 }
2567
2568 ib_srp_transport_template =
2569 srp_attach_transport(&ib_srp_transport_functions);
2570 if (!ib_srp_transport_template)
2571 return -ENOMEM;
2572
2573 ret = class_register(&srp_class);
2574 if (ret) {
2575 pr_err("couldn't register class infiniband_srp\n");
2576 srp_release_transport(ib_srp_transport_template);
2577 return ret;
2578 }
2579
2580 ib_sa_register_client(&srp_sa_client);
2581
2582 ret = ib_register_client(&srp_client);
2583 if (ret) {
2584 pr_err("couldn't register IB client\n");
2585 srp_release_transport(ib_srp_transport_template);
2586 ib_sa_unregister_client(&srp_sa_client);
2587 class_unregister(&srp_class);
2588 return ret;
2589 }
2590
2591 return 0;
2592 }
2593
2594 static void __exit srp_cleanup_module(void)
2595 {
2596 ib_unregister_client(&srp_client);
2597 ib_sa_unregister_client(&srp_sa_client);
2598 class_unregister(&srp_class);
2599 srp_release_transport(ib_srp_transport_template);
2600 }
2601
2602 module_init(srp_init_module);
2603 module_exit(srp_cleanup_module);
Linux kernel - Deliverables
This page took 0.238523 seconds and 4 git commands to generate.