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


95590a38e88a2cf0c973370df8154809ead5cd55
[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 int ret;
226
227 init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
228 if (!init_attr)
229 return -ENOMEM;
230
231 target->recv_cq = ib_create_cq(target->srp_host->srp_dev->dev,
232 srp_recv_completion, NULL, target, SRP_RQ_SIZE, 0);
233 if (IS_ERR(target->recv_cq)) {
234 ret = PTR_ERR(target->recv_cq);
235 goto err;
236 }
237
238 target->send_cq = ib_create_cq(target->srp_host->srp_dev->dev,
239 srp_send_completion, NULL, target, SRP_SQ_SIZE, 0);
240 if (IS_ERR(target->send_cq)) {
241 ret = PTR_ERR(target->send_cq);
242 goto err_recv_cq;
243 }
244
245 ib_req_notify_cq(target->recv_cq, IB_CQ_NEXT_COMP);
246
247 init_attr->event_handler = srp_qp_event;
248 init_attr->cap.max_send_wr = SRP_SQ_SIZE;
249 init_attr->cap.max_recv_wr = SRP_RQ_SIZE;
250 init_attr->cap.max_recv_sge = 1;
251 init_attr->cap.max_send_sge = 1;
252 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
253 init_attr->qp_type = IB_QPT_RC;
254 init_attr->send_cq = target->send_cq;
255 init_attr->recv_cq = target->recv_cq;
256
257 target->qp = ib_create_qp(target->srp_host->srp_dev->pd, init_attr);
258 if (IS_ERR(target->qp)) {
259 ret = PTR_ERR(target->qp);
260 goto err_send_cq;
261 }
262
263 ret = srp_init_qp(target, target->qp);
264 if (ret)
265 goto err_qp;
266
267 kfree(init_attr);
268 return 0;
269
270 err_qp:
271 ib_destroy_qp(target->qp);
272
273 err_send_cq:
274 ib_destroy_cq(target->send_cq);
275
276 err_recv_cq:
277 ib_destroy_cq(target->recv_cq);
278
279 err:
280 kfree(init_attr);
281 return ret;
282 }
283
284 static void srp_free_target_ib(struct srp_target_port *target)
285 {
286 int i;
287
288 ib_destroy_qp(target->qp);
289 ib_destroy_cq(target->send_cq);
290 ib_destroy_cq(target->recv_cq);
291
292 for (i = 0; i < SRP_RQ_SIZE; ++i)
293 srp_free_iu(target->srp_host, target->rx_ring[i]);
294 for (i = 0; i < SRP_SQ_SIZE; ++i)
295 srp_free_iu(target->srp_host, target->tx_ring[i]);
296 }
297
298 static void srp_path_rec_completion(int status,
299 struct ib_sa_path_rec *pathrec,
300 void *target_ptr)
301 {
302 struct srp_target_port *target = target_ptr;
303
304 target->status = status;
305 if (status)
306 shost_printk(KERN_ERR, target->scsi_host,
307 PFX "Got failed path rec status %d\n", status);
308 else
309 target->path = *pathrec;
310 complete(&target->done);
311 }
312
313 static int srp_lookup_path(struct srp_target_port *target)
314 {
315 target->path.numb_path = 1;
316
317 init_completion(&target->done);
318
319 target->path_query_id = ib_sa_path_rec_get(&srp_sa_client,
320 target->srp_host->srp_dev->dev,
321 target->srp_host->port,
322 &target->path,
323 IB_SA_PATH_REC_SERVICE_ID |
324 IB_SA_PATH_REC_DGID |
325 IB_SA_PATH_REC_SGID |
326 IB_SA_PATH_REC_NUMB_PATH |
327 IB_SA_PATH_REC_PKEY,
328 SRP_PATH_REC_TIMEOUT_MS,
329 GFP_KERNEL,
330 srp_path_rec_completion,
331 target, &target->path_query);
332 if (target->path_query_id < 0)
333 return target->path_query_id;
334
335 wait_for_completion(&target->done);
336
337 if (target->status < 0)
338 shost_printk(KERN_WARNING, target->scsi_host,
339 PFX "Path record query failed\n");
340
341 return target->status;
342 }
343
344 static int srp_send_req(struct srp_target_port *target)
345 {
346 struct {
347 struct ib_cm_req_param param;
348 struct srp_login_req priv;
349 } *req = NULL;
350 int status;
351
352 req = kzalloc(sizeof *req, GFP_KERNEL);
353 if (!req)
354 return -ENOMEM;
355
356 req->param.primary_path = &target->path;
357 req->param.alternate_path = NULL;
358 req->param.service_id = target->service_id;
359 req->param.qp_num = target->qp->qp_num;
360 req->param.qp_type = target->qp->qp_type;
361 req->param.private_data = &req->priv;
362 req->param.private_data_len = sizeof req->priv;
363 req->param.flow_control = 1;
364
365 get_random_bytes(&req->param.starting_psn, 4);
366 req->param.starting_psn &= 0xffffff;
367
368 /*
369 * Pick some arbitrary defaults here; we could make these
370 * module parameters if anyone cared about setting them.
371 */
372 req->param.responder_resources = 4;
373 req->param.remote_cm_response_timeout = 20;
374 req->param.local_cm_response_timeout = 20;
375 req->param.retry_count = 7;
376 req->param.rnr_retry_count = 7;
377 req->param.max_cm_retries = 15;
378
379 req->priv.opcode = SRP_LOGIN_REQ;
380 req->priv.tag = 0;
381 req->priv.req_it_iu_len = cpu_to_be32(target->max_iu_len);
382 req->priv.req_buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
383 SRP_BUF_FORMAT_INDIRECT);
384 /*
385 * In the published SRP specification (draft rev. 16a), the
386 * port identifier format is 8 bytes of ID extension followed
387 * by 8 bytes of GUID. Older drafts put the two halves in the
388 * opposite order, so that the GUID comes first.
389 *
390 * Targets conforming to these obsolete drafts can be
391 * recognized by the I/O Class they report.
392 */
393 if (target->io_class == SRP_REV10_IB_IO_CLASS) {
394 memcpy(req->priv.initiator_port_id,
395 &target->path.sgid.global.interface_id, 8);
396 memcpy(req->priv.initiator_port_id + 8,
397 &target->initiator_ext, 8);
398 memcpy(req->priv.target_port_id, &target->ioc_guid, 8);
399 memcpy(req->priv.target_port_id + 8, &target->id_ext, 8);
400 } else {
401 memcpy(req->priv.initiator_port_id,
402 &target->initiator_ext, 8);
403 memcpy(req->priv.initiator_port_id + 8,
404 &target->path.sgid.global.interface_id, 8);
405 memcpy(req->priv.target_port_id, &target->id_ext, 8);
406 memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
407 }
408
409 /*
410 * Topspin/Cisco SRP targets will reject our login unless we
411 * zero out the first 8 bytes of our initiator port ID and set
412 * the second 8 bytes to the local node GUID.
413 */
414 if (srp_target_is_topspin(target)) {
415 shost_printk(KERN_DEBUG, target->scsi_host,
416 PFX "Topspin/Cisco initiator port ID workaround "
417 "activated for target GUID %016llx\n",
418 (unsigned long long) be64_to_cpu(target->ioc_guid));
419 memset(req->priv.initiator_port_id, 0, 8);
420 memcpy(req->priv.initiator_port_id + 8,
421 &target->srp_host->srp_dev->dev->node_guid, 8);
422 }
423
424 status = ib_send_cm_req(target->cm_id, &req->param);
425
426 kfree(req);
427
428 return status;
429 }
430
431 static bool srp_queue_remove_work(struct srp_target_port *target)
432 {
433 bool changed = false;
434
435 spin_lock_irq(&target->lock);
436 if (target->state != SRP_TARGET_REMOVED) {
437 target->state = SRP_TARGET_REMOVED;
438 changed = true;
439 }
440 spin_unlock_irq(&target->lock);
441
442 if (changed)
443 queue_work(system_long_wq, &target->remove_work);
444
445 return changed;
446 }
447
448 static bool srp_change_conn_state(struct srp_target_port *target,
449 bool connected)
450 {
451 bool changed = false;
452
453 spin_lock_irq(&target->lock);
454 if (target->connected != connected) {
455 target->connected = connected;
456 changed = true;
457 }
458 spin_unlock_irq(&target->lock);
459
460 return changed;
461 }
462
463 static void srp_disconnect_target(struct srp_target_port *target)
464 {
465 if (srp_change_conn_state(target, false)) {
466 /* XXX should send SRP_I_LOGOUT request */
467
468 init_completion(&target->done);
469 if (ib_send_cm_dreq(target->cm_id, NULL, 0)) {
470 shost_printk(KERN_DEBUG, target->scsi_host,
471 PFX "Sending CM DREQ failed\n");
472 } else {
473 wait_for_completion(&target->done);
474 }
475 }
476 }
477
478 static void srp_free_req_data(struct srp_target_port *target)
479 {
480 struct ib_device *ibdev = target->srp_host->srp_dev->dev;
481 struct srp_request *req;
482 int i;
483
484 for (i = 0, req = target->req_ring; i < SRP_CMD_SQ_SIZE; ++i, ++req) {
485 kfree(req->fmr_list);
486 kfree(req->map_page);
487 if (req->indirect_dma_addr) {
488 ib_dma_unmap_single(ibdev, req->indirect_dma_addr,
489 target->indirect_size,
490 DMA_TO_DEVICE);
491 }
492 kfree(req->indirect_desc);
493 }
494 }
495
496 /**
497 * srp_del_scsi_host_attr() - Remove attributes defined in the host template.
498 * @shost: SCSI host whose attributes to remove from sysfs.
499 *
500 * Note: Any attributes defined in the host template and that did not exist
501 * before invocation of this function will be ignored.
502 */
503 static void srp_del_scsi_host_attr(struct Scsi_Host *shost)
504 {
505 struct device_attribute **attr;
506
507 for (attr = shost->hostt->shost_attrs; attr && *attr; ++attr)
508 device_remove_file(&shost->shost_dev, *attr);
509 }
510
511 static void srp_remove_target(struct srp_target_port *target)
512 {
513 WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
514
515 srp_del_scsi_host_attr(target->scsi_host);
516 srp_remove_host(target->scsi_host);
517 scsi_remove_host(target->scsi_host);
518 srp_disconnect_target(target);
519 ib_destroy_cm_id(target->cm_id);
520 srp_free_target_ib(target);
521 srp_free_req_data(target);
522 scsi_host_put(target->scsi_host);
523 }
524
525 static void srp_remove_work(struct work_struct *work)
526 {
527 struct srp_target_port *target =
528 container_of(work, struct srp_target_port, remove_work);
529
530 WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
531
532 spin_lock(&target->srp_host->target_lock);
533 list_del(&target->list);
534 spin_unlock(&target->srp_host->target_lock);
535
536 srp_remove_target(target);
537 }
538
539 static int srp_connect_target(struct srp_target_port *target)
540 {
541 int retries = 3;
542 int ret;
543
544 WARN_ON_ONCE(target->connected);
545
546 target->qp_in_error = false;
547
548 ret = srp_lookup_path(target);
549 if (ret)
550 return ret;
551
552 while (1) {
553 init_completion(&target->done);
554 ret = srp_send_req(target);
555 if (ret)
556 return ret;
557 wait_for_completion(&target->done);
558
559 /*
560 * The CM event handling code will set status to
561 * SRP_PORT_REDIRECT if we get a port redirect REJ
562 * back, or SRP_DLID_REDIRECT if we get a lid/qp
563 * redirect REJ back.
564 */
565 switch (target->status) {
566 case 0:
567 srp_change_conn_state(target, true);
568 return 0;
569
570 case SRP_PORT_REDIRECT:
571 ret = srp_lookup_path(target);
572 if (ret)
573 return ret;
574 break;
575
576 case SRP_DLID_REDIRECT:
577 break;
578
579 case SRP_STALE_CONN:
580 /* Our current CM id was stale, and is now in timewait.
581 * Try to reconnect with a new one.
582 */
583 if (!retries-- || srp_new_cm_id(target)) {
584 shost_printk(KERN_ERR, target->scsi_host, PFX
585 "giving up on stale connection\n");
586 target->status = -ECONNRESET;
587 return target->status;
588 }
589
590 shost_printk(KERN_ERR, target->scsi_host, PFX
591 "retrying stale connection\n");
592 break;
593
594 default:
595 return target->status;
596 }
597 }
598 }
599
600 static void srp_unmap_data(struct scsi_cmnd *scmnd,
601 struct srp_target_port *target,
602 struct srp_request *req)
603 {
604 struct ib_device *ibdev = target->srp_host->srp_dev->dev;
605 struct ib_pool_fmr **pfmr;
606
607 if (!scsi_sglist(scmnd) ||
608 (scmnd->sc_data_direction != DMA_TO_DEVICE &&
609 scmnd->sc_data_direction != DMA_FROM_DEVICE))
610 return;
611
612 pfmr = req->fmr_list;
613 while (req->nfmr--)
614 ib_fmr_pool_unmap(*pfmr++);
615
616 ib_dma_unmap_sg(ibdev, scsi_sglist(scmnd), scsi_sg_count(scmnd),
617 scmnd->sc_data_direction);
618 }
619
620 /**
621 * srp_claim_req - Take ownership of the scmnd associated with a request.
622 * @target: SRP target port.
623 * @req: SRP request.
624 * @scmnd: If NULL, take ownership of @req->scmnd. If not NULL, only take
625 * ownership of @req->scmnd if it equals @scmnd.
626 *
627 * Return value:
628 * Either NULL or a pointer to the SCSI command the caller became owner of.
629 */
630 static struct scsi_cmnd *srp_claim_req(struct srp_target_port *target,
631 struct srp_request *req,
632 struct scsi_cmnd *scmnd)
633 {
634 unsigned long flags;
635
636 spin_lock_irqsave(&target->lock, flags);
637 if (!scmnd) {
638 scmnd = req->scmnd;
639 req->scmnd = NULL;
640 } else if (req->scmnd == scmnd) {
641 req->scmnd = NULL;
642 } else {
643 scmnd = NULL;
644 }
645 spin_unlock_irqrestore(&target->lock, flags);
646
647 return scmnd;
648 }
649
650 /**
651 * srp_free_req() - Unmap data and add request to the free request list.
652 */
653 static void srp_free_req(struct srp_target_port *target,
654 struct srp_request *req, struct scsi_cmnd *scmnd,
655 s32 req_lim_delta)
656 {
657 unsigned long flags;
658
659 srp_unmap_data(scmnd, target, req);
660
661 spin_lock_irqsave(&target->lock, flags);
662 target->req_lim += req_lim_delta;
663 list_add_tail(&req->list, &target->free_reqs);
664 spin_unlock_irqrestore(&target->lock, flags);
665 }
666
667 static void srp_reset_req(struct srp_target_port *target, struct srp_request *req)
668 {
669 struct scsi_cmnd *scmnd = srp_claim_req(target, req, NULL);
670
671 if (scmnd) {
672 srp_free_req(target, req, scmnd, 0);
673 scmnd->result = DID_RESET << 16;
674 scmnd->scsi_done(scmnd);
675 }
676 }
677
678 static int srp_reconnect_target(struct srp_target_port *target)
679 {
680 struct Scsi_Host *shost = target->scsi_host;
681 struct ib_qp_attr qp_attr;
682 struct ib_wc wc;
683 int i, ret;
684
685 if (target->state != SRP_TARGET_LIVE)
686 return -EAGAIN;
687
688 scsi_target_block(&shost->shost_gendev);
689
690 srp_disconnect_target(target);
691 /*
692 * Now get a new local CM ID so that we avoid confusing the
693 * target in case things are really fouled up.
694 */
695 ret = srp_new_cm_id(target);
696 if (ret)
697 goto unblock;
698
699 qp_attr.qp_state = IB_QPS_RESET;
700 ret = ib_modify_qp(target->qp, &qp_attr, IB_QP_STATE);
701 if (ret)
702 goto unblock;
703
704 ret = srp_init_qp(target, target->qp);
705 if (ret)
706 goto unblock;
707
708 while (ib_poll_cq(target->recv_cq, 1, &wc) > 0)
709 ; /* nothing */
710 while (ib_poll_cq(target->send_cq, 1, &wc) > 0)
711 ; /* nothing */
712
713 for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
714 struct srp_request *req = &target->req_ring[i];
715 if (req->scmnd)
716 srp_reset_req(target, req);
717 }
718
719 INIT_LIST_HEAD(&target->free_tx);
720 for (i = 0; i < SRP_SQ_SIZE; ++i)
721 list_add(&target->tx_ring[i]->list, &target->free_tx);
722
723 ret = srp_connect_target(target);
724
725 unblock:
726 scsi_target_unblock(&shost->shost_gendev, ret == 0 ? SDEV_RUNNING :
727 SDEV_TRANSPORT_OFFLINE);
728
729 if (ret)
730 goto err;
731
732 shost_printk(KERN_INFO, target->scsi_host, PFX "reconnect succeeded\n");
733
734 return ret;
735
736 err:
737 shost_printk(KERN_ERR, target->scsi_host,
738 PFX "reconnect failed (%d), removing target port.\n", ret);
739
740 /*
741 * We couldn't reconnect, so kill our target port off.
742 * However, we have to defer the real removal because we
743 * are in the context of the SCSI error handler now, which
744 * will deadlock if we call scsi_remove_host().
745 */
746 srp_queue_remove_work(target);
747
748 return ret;
749 }
750
751 static void srp_map_desc(struct srp_map_state *state, dma_addr_t dma_addr,
752 unsigned int dma_len, u32 rkey)
753 {
754 struct srp_direct_buf *desc = state->desc;
755
756 desc->va = cpu_to_be64(dma_addr);
757 desc->key = cpu_to_be32(rkey);
758 desc->len = cpu_to_be32(dma_len);
759
760 state->total_len += dma_len;
761 state->desc++;
762 state->ndesc++;
763 }
764
765 static int srp_map_finish_fmr(struct srp_map_state *state,
766 struct srp_target_port *target)
767 {
768 struct srp_device *dev = target->srp_host->srp_dev;
769 struct ib_pool_fmr *fmr;
770 u64 io_addr = 0;
771
772 if (!state->npages)
773 return 0;
774
775 if (state->npages == 1) {
776 srp_map_desc(state, state->base_dma_addr, state->fmr_len,
777 target->rkey);
778 state->npages = state->fmr_len = 0;
779 return 0;
780 }
781
782 fmr = ib_fmr_pool_map_phys(dev->fmr_pool, state->pages,
783 state->npages, io_addr);
784 if (IS_ERR(fmr))
785 return PTR_ERR(fmr);
786
787 *state->next_fmr++ = fmr;
788 state->nfmr++;
789
790 srp_map_desc(state, 0, state->fmr_len, fmr->fmr->rkey);
791 state->npages = state->fmr_len = 0;
792 return 0;
793 }
794
795 static void srp_map_update_start(struct srp_map_state *state,
796 struct scatterlist *sg, int sg_index,
797 dma_addr_t dma_addr)
798 {
799 state->unmapped_sg = sg;
800 state->unmapped_index = sg_index;
801 state->unmapped_addr = dma_addr;
802 }
803
804 static int srp_map_sg_entry(struct srp_map_state *state,
805 struct srp_target_port *target,
806 struct scatterlist *sg, int sg_index,
807 int use_fmr)
808 {
809 struct srp_device *dev = target->srp_host->srp_dev;
810 struct ib_device *ibdev = dev->dev;
811 dma_addr_t dma_addr = ib_sg_dma_address(ibdev, sg);
812 unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
813 unsigned int len;
814 int ret;
815
816 if (!dma_len)
817 return 0;
818
819 if (use_fmr == SRP_MAP_NO_FMR) {
820 /* Once we're in direct map mode for a request, we don't
821 * go back to FMR mode, so no need to update anything
822 * other than the descriptor.
823 */
824 srp_map_desc(state, dma_addr, dma_len, target->rkey);
825 return 0;
826 }
827
828 /* If we start at an offset into the FMR page, don't merge into
829 * the current FMR. Finish it out, and use the kernel's MR for this
830 * sg entry. This is to avoid potential bugs on some SRP targets
831 * that were never quite defined, but went away when the initiator
832 * avoided using FMR on such page fragments.
833 */
834 if (dma_addr & ~dev->fmr_page_mask || dma_len > dev->fmr_max_size) {
835 ret = srp_map_finish_fmr(state, target);
836 if (ret)
837 return ret;
838
839 srp_map_desc(state, dma_addr, dma_len, target->rkey);
840 srp_map_update_start(state, NULL, 0, 0);
841 return 0;
842 }
843
844 /* If this is the first sg to go into the FMR, save our position.
845 * We need to know the first unmapped entry, its index, and the
846 * first unmapped address within that entry to be able to restart
847 * mapping after an error.
848 */
849 if (!state->unmapped_sg)
850 srp_map_update_start(state, sg, sg_index, dma_addr);
851
852 while (dma_len) {
853 if (state->npages == SRP_FMR_SIZE) {
854 ret = srp_map_finish_fmr(state, target);
855 if (ret)
856 return ret;
857
858 srp_map_update_start(state, sg, sg_index, dma_addr);
859 }
860
861 len = min_t(unsigned int, dma_len, dev->fmr_page_size);
862
863 if (!state->npages)
864 state->base_dma_addr = dma_addr;
865 state->pages[state->npages++] = dma_addr;
866 state->fmr_len += len;
867 dma_addr += len;
868 dma_len -= len;
869 }
870
871 /* If the last entry of the FMR wasn't a full page, then we need to
872 * close it out and start a new one -- we can only merge at page
873 * boundries.
874 */
875 ret = 0;
876 if (len != dev->fmr_page_size) {
877 ret = srp_map_finish_fmr(state, target);
878 if (!ret)
879 srp_map_update_start(state, NULL, 0, 0);
880 }
881 return ret;
882 }
883
884 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_target_port *target,
885 struct srp_request *req)
886 {
887 struct scatterlist *scat, *sg;
888 struct srp_cmd *cmd = req->cmd->buf;
889 int i, len, nents, count, use_fmr;
890 struct srp_device *dev;
891 struct ib_device *ibdev;
892 struct srp_map_state state;
893 struct srp_indirect_buf *indirect_hdr;
894 u32 table_len;
895 u8 fmt;
896
897 if (!scsi_sglist(scmnd) || scmnd->sc_data_direction == DMA_NONE)
898 return sizeof (struct srp_cmd);
899
900 if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
901 scmnd->sc_data_direction != DMA_TO_DEVICE) {
902 shost_printk(KERN_WARNING, target->scsi_host,
903 PFX "Unhandled data direction %d\n",
904 scmnd->sc_data_direction);
905 return -EINVAL;
906 }
907
908 nents = scsi_sg_count(scmnd);
909 scat = scsi_sglist(scmnd);
910
911 dev = target->srp_host->srp_dev;
912 ibdev = dev->dev;
913
914 count = ib_dma_map_sg(ibdev, scat, nents, scmnd->sc_data_direction);
915 if (unlikely(count == 0))
916 return -EIO;
917
918 fmt = SRP_DATA_DESC_DIRECT;
919 len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
920
921 if (count == 1) {
922 /*
923 * The midlayer only generated a single gather/scatter
924 * entry, or DMA mapping coalesced everything to a
925 * single entry. So a direct descriptor along with
926 * the DMA MR suffices.
927 */
928 struct srp_direct_buf *buf = (void *) cmd->add_data;
929
930 buf->va = cpu_to_be64(ib_sg_dma_address(ibdev, scat));
931 buf->key = cpu_to_be32(target->rkey);
932 buf->len = cpu_to_be32(ib_sg_dma_len(ibdev, scat));
933
934 req->nfmr = 0;
935 goto map_complete;
936 }
937
938 /* We have more than one scatter/gather entry, so build our indirect
939 * descriptor table, trying to merge as many entries with FMR as we
940 * can.
941 */
942 indirect_hdr = (void *) cmd->add_data;
943
944 ib_dma_sync_single_for_cpu(ibdev, req->indirect_dma_addr,
945 target->indirect_size, DMA_TO_DEVICE);
946
947 memset(&state, 0, sizeof(state));
948 state.desc = req->indirect_desc;
949 state.pages = req->map_page;
950 state.next_fmr = req->fmr_list;
951
952 use_fmr = dev->fmr_pool ? SRP_MAP_ALLOW_FMR : SRP_MAP_NO_FMR;
953
954 for_each_sg(scat, sg, count, i) {
955 if (srp_map_sg_entry(&state, target, sg, i, use_fmr)) {
956 /* FMR mapping failed, so backtrack to the first
957 * unmapped entry and continue on without using FMR.
958 */
959 dma_addr_t dma_addr;
960 unsigned int dma_len;
961
962 backtrack:
963 sg = state.unmapped_sg;
964 i = state.unmapped_index;
965
966 dma_addr = ib_sg_dma_address(ibdev, sg);
967 dma_len = ib_sg_dma_len(ibdev, sg);
968 dma_len -= (state.unmapped_addr - dma_addr);
969 dma_addr = state.unmapped_addr;
970 use_fmr = SRP_MAP_NO_FMR;
971 srp_map_desc(&state, dma_addr, dma_len, target->rkey);
972 }
973 }
974
975 if (use_fmr == SRP_MAP_ALLOW_FMR && srp_map_finish_fmr(&state, target))
976 goto backtrack;
977
978 /* We've mapped the request, now pull as much of the indirect
979 * descriptor table as we can into the command buffer. If this
980 * target is not using an external indirect table, we are
981 * guaranteed to fit into the command, as the SCSI layer won't
982 * give us more S/G entries than we allow.
983 */
984 req->nfmr = state.nfmr;
985 if (state.ndesc == 1) {
986 /* FMR mapping was able to collapse this to one entry,
987 * so use a direct descriptor.
988 */
989 struct srp_direct_buf *buf = (void *) cmd->add_data;
990
991 *buf = req->indirect_desc[0];
992 goto map_complete;
993 }
994
995 if (unlikely(target->cmd_sg_cnt < state.ndesc &&
996 !target->allow_ext_sg)) {
997 shost_printk(KERN_ERR, target->scsi_host,
998 "Could not fit S/G list into SRP_CMD\n");
999 return -EIO;
1000 }
1001
1002 count = min(state.ndesc, target->cmd_sg_cnt);
1003 table_len = state.ndesc * sizeof (struct srp_direct_buf);
1004
1005 fmt = SRP_DATA_DESC_INDIRECT;
1006 len = sizeof(struct srp_cmd) + sizeof (struct srp_indirect_buf);
1007 len += count * sizeof (struct srp_direct_buf);
1008
1009 memcpy(indirect_hdr->desc_list, req->indirect_desc,
1010 count * sizeof (struct srp_direct_buf));
1011
1012 indirect_hdr->table_desc.va = cpu_to_be64(req->indirect_dma_addr);
1013 indirect_hdr->table_desc.key = cpu_to_be32(target->rkey);
1014 indirect_hdr->table_desc.len = cpu_to_be32(table_len);
1015 indirect_hdr->len = cpu_to_be32(state.total_len);
1016
1017 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1018 cmd->data_out_desc_cnt = count;
1019 else
1020 cmd->data_in_desc_cnt = count;
1021
1022 ib_dma_sync_single_for_device(ibdev, req->indirect_dma_addr, table_len,
1023 DMA_TO_DEVICE);
1024
1025 map_complete:
1026 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1027 cmd->buf_fmt = fmt << 4;
1028 else
1029 cmd->buf_fmt = fmt;
1030
1031 return len;
1032 }
1033
1034 /*
1035 * Return an IU and possible credit to the free pool
1036 */
1037 static void srp_put_tx_iu(struct srp_target_port *target, struct srp_iu *iu,
1038 enum srp_iu_type iu_type)
1039 {
1040 unsigned long flags;
1041
1042 spin_lock_irqsave(&target->lock, flags);
1043 list_add(&iu->list, &target->free_tx);
1044 if (iu_type != SRP_IU_RSP)
1045 ++target->req_lim;
1046 spin_unlock_irqrestore(&target->lock, flags);
1047 }
1048
1049 /*
1050 * Must be called with target->lock held to protect req_lim and free_tx.
1051 * If IU is not sent, it must be returned using srp_put_tx_iu().
1052 *
1053 * Note:
1054 * An upper limit for the number of allocated information units for each
1055 * request type is:
1056 * - SRP_IU_CMD: SRP_CMD_SQ_SIZE, since the SCSI mid-layer never queues
1057 * more than Scsi_Host.can_queue requests.
1058 * - SRP_IU_TSK_MGMT: SRP_TSK_MGMT_SQ_SIZE.
1059 * - SRP_IU_RSP: 1, since a conforming SRP target never sends more than
1060 * one unanswered SRP request to an initiator.
1061 */
1062 static struct srp_iu *__srp_get_tx_iu(struct srp_target_port *target,
1063 enum srp_iu_type iu_type)
1064 {
1065 s32 rsv = (iu_type == SRP_IU_TSK_MGMT) ? 0 : SRP_TSK_MGMT_SQ_SIZE;
1066 struct srp_iu *iu;
1067
1068 srp_send_completion(target->send_cq, target);
1069
1070 if (list_empty(&target->free_tx))
1071 return NULL;
1072
1073 /* Initiator responses to target requests do not consume credits */
1074 if (iu_type != SRP_IU_RSP) {
1075 if (target->req_lim <= rsv) {
1076 ++target->zero_req_lim;
1077 return NULL;
1078 }
1079
1080 --target->req_lim;
1081 }
1082
1083 iu = list_first_entry(&target->free_tx, struct srp_iu, list);
1084 list_del(&iu->list);
1085 return iu;
1086 }
1087
1088 static int srp_post_send(struct srp_target_port *target,
1089 struct srp_iu *iu, int len)
1090 {
1091 struct ib_sge list;
1092 struct ib_send_wr wr, *bad_wr;
1093
1094 list.addr = iu->dma;
1095 list.length = len;
1096 list.lkey = target->lkey;
1097
1098 wr.next = NULL;
1099 wr.wr_id = (uintptr_t) iu;
1100 wr.sg_list = &list;
1101 wr.num_sge = 1;
1102 wr.opcode = IB_WR_SEND;
1103 wr.send_flags = IB_SEND_SIGNALED;
1104
1105 return ib_post_send(target->qp, &wr, &bad_wr);
1106 }
1107
1108 static int srp_post_recv(struct srp_target_port *target, struct srp_iu *iu)
1109 {
1110 struct ib_recv_wr wr, *bad_wr;
1111 struct ib_sge list;
1112
1113 list.addr = iu->dma;
1114 list.length = iu->size;
1115 list.lkey = target->lkey;
1116
1117 wr.next = NULL;
1118 wr.wr_id = (uintptr_t) iu;
1119 wr.sg_list = &list;
1120 wr.num_sge = 1;
1121
1122 return ib_post_recv(target->qp, &wr, &bad_wr);
1123 }
1124
1125 static void srp_process_rsp(struct srp_target_port *target, struct srp_rsp *rsp)
1126 {
1127 struct srp_request *req;
1128 struct scsi_cmnd *scmnd;
1129 unsigned long flags;
1130
1131 if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
1132 spin_lock_irqsave(&target->lock, flags);
1133 target->req_lim += be32_to_cpu(rsp->req_lim_delta);
1134 spin_unlock_irqrestore(&target->lock, flags);
1135
1136 target->tsk_mgmt_status = -1;
1137 if (be32_to_cpu(rsp->resp_data_len) >= 4)
1138 target->tsk_mgmt_status = rsp->data[3];
1139 complete(&target->tsk_mgmt_done);
1140 } else {
1141 req = &target->req_ring[rsp->tag];
1142 scmnd = srp_claim_req(target, req, NULL);
1143 if (!scmnd) {
1144 shost_printk(KERN_ERR, target->scsi_host,
1145 "Null scmnd for RSP w/tag %016llx\n",
1146 (unsigned long long) rsp->tag);
1147
1148 spin_lock_irqsave(&target->lock, flags);
1149 target->req_lim += be32_to_cpu(rsp->req_lim_delta);
1150 spin_unlock_irqrestore(&target->lock, flags);
1151
1152 return;
1153 }
1154 scmnd->result = rsp->status;
1155
1156 if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
1157 memcpy(scmnd->sense_buffer, rsp->data +
1158 be32_to_cpu(rsp->resp_data_len),
1159 min_t(int, be32_to_cpu(rsp->sense_data_len),
1160 SCSI_SENSE_BUFFERSIZE));
1161 }
1162
1163 if (rsp->flags & (SRP_RSP_FLAG_DOOVER | SRP_RSP_FLAG_DOUNDER))
1164 scsi_set_resid(scmnd, be32_to_cpu(rsp->data_out_res_cnt));
1165 else if (rsp->flags & (SRP_RSP_FLAG_DIOVER | SRP_RSP_FLAG_DIUNDER))
1166 scsi_set_resid(scmnd, be32_to_cpu(rsp->data_in_res_cnt));
1167
1168 srp_free_req(target, req, scmnd,
1169 be32_to_cpu(rsp->req_lim_delta));
1170
1171 scmnd->host_scribble = NULL;
1172 scmnd->scsi_done(scmnd);
1173 }
1174 }
1175
1176 static int srp_response_common(struct srp_target_port *target, s32 req_delta,
1177 void *rsp, int len)
1178 {
1179 struct ib_device *dev = target->srp_host->srp_dev->dev;
1180 unsigned long flags;
1181 struct srp_iu *iu;
1182 int err;
1183
1184 spin_lock_irqsave(&target->lock, flags);
1185 target->req_lim += req_delta;
1186 iu = __srp_get_tx_iu(target, SRP_IU_RSP);
1187 spin_unlock_irqrestore(&target->lock, flags);
1188
1189 if (!iu) {
1190 shost_printk(KERN_ERR, target->scsi_host, PFX
1191 "no IU available to send response\n");
1192 return 1;
1193 }
1194
1195 ib_dma_sync_single_for_cpu(dev, iu->dma, len, DMA_TO_DEVICE);
1196 memcpy(iu->buf, rsp, len);
1197 ib_dma_sync_single_for_device(dev, iu->dma, len, DMA_TO_DEVICE);
1198
1199 err = srp_post_send(target, iu, len);
1200 if (err) {
1201 shost_printk(KERN_ERR, target->scsi_host, PFX
1202 "unable to post response: %d\n", err);
1203 srp_put_tx_iu(target, iu, SRP_IU_RSP);
1204 }
1205
1206 return err;
1207 }
1208
1209 static void srp_process_cred_req(struct srp_target_port *target,
1210 struct srp_cred_req *req)
1211 {
1212 struct srp_cred_rsp rsp = {
1213 .opcode = SRP_CRED_RSP,
1214 .tag = req->tag,
1215 };
1216 s32 delta = be32_to_cpu(req->req_lim_delta);
1217
1218 if (srp_response_common(target, delta, &rsp, sizeof rsp))
1219 shost_printk(KERN_ERR, target->scsi_host, PFX
1220 "problems processing SRP_CRED_REQ\n");
1221 }
1222
1223 static void srp_process_aer_req(struct srp_target_port *target,
1224 struct srp_aer_req *req)
1225 {
1226 struct srp_aer_rsp rsp = {
1227 .opcode = SRP_AER_RSP,
1228 .tag = req->tag,
1229 };
1230 s32 delta = be32_to_cpu(req->req_lim_delta);
1231
1232 shost_printk(KERN_ERR, target->scsi_host, PFX
1233 "ignoring AER for LUN %llu\n", be64_to_cpu(req->lun));
1234
1235 if (srp_response_common(target, delta, &rsp, sizeof rsp))
1236 shost_printk(KERN_ERR, target->scsi_host, PFX
1237 "problems processing SRP_AER_REQ\n");
1238 }
1239
1240 static void srp_handle_recv(struct srp_target_port *target, struct ib_wc *wc)
1241 {
1242 struct ib_device *dev = target->srp_host->srp_dev->dev;
1243 struct srp_iu *iu = (struct srp_iu *) (uintptr_t) wc->wr_id;
1244 int res;
1245 u8 opcode;
1246
1247 ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_ti_iu_len,
1248 DMA_FROM_DEVICE);
1249
1250 opcode = *(u8 *) iu->buf;
1251
1252 if (0) {
1253 shost_printk(KERN_ERR, target->scsi_host,
1254 PFX "recv completion, opcode 0x%02x\n", opcode);
1255 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 8, 1,
1256 iu->buf, wc->byte_len, true);
1257 }
1258
1259 switch (opcode) {
1260 case SRP_RSP:
1261 srp_process_rsp(target, iu->buf);
1262 break;
1263
1264 case SRP_CRED_REQ:
1265 srp_process_cred_req(target, iu->buf);
1266 break;
1267
1268 case SRP_AER_REQ:
1269 srp_process_aer_req(target, iu->buf);
1270 break;
1271
1272 case SRP_T_LOGOUT:
1273 /* XXX Handle target logout */
1274 shost_printk(KERN_WARNING, target->scsi_host,
1275 PFX "Got target logout request\n");
1276 break;
1277
1278 default:
1279 shost_printk(KERN_WARNING, target->scsi_host,
1280 PFX "Unhandled SRP opcode 0x%02x\n", opcode);
1281 break;
1282 }
1283
1284 ib_dma_sync_single_for_device(dev, iu->dma, target->max_ti_iu_len,
1285 DMA_FROM_DEVICE);
1286
1287 res = srp_post_recv(target, iu);
1288 if (res != 0)
1289 shost_printk(KERN_ERR, target->scsi_host,
1290 PFX "Recv failed with error code %d\n", res);
1291 }
1292
1293 static void srp_handle_qp_err(enum ib_wc_status wc_status,
1294 enum ib_wc_opcode wc_opcode,
1295 struct srp_target_port *target)
1296 {
1297 if (target->connected && !target->qp_in_error) {
1298 shost_printk(KERN_ERR, target->scsi_host,
1299 PFX "failed %s status %d\n",
1300 wc_opcode & IB_WC_RECV ? "receive" : "send",
1301 wc_status);
1302 }
1303 target->qp_in_error = true;
1304 }
1305
1306 static void srp_recv_completion(struct ib_cq *cq, void *target_ptr)
1307 {
1308 struct srp_target_port *target = target_ptr;
1309 struct ib_wc wc;
1310
1311 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1312 while (ib_poll_cq(cq, 1, &wc) > 0) {
1313 if (likely(wc.status == IB_WC_SUCCESS)) {
1314 srp_handle_recv(target, &wc);
1315 } else {
1316 srp_handle_qp_err(wc.status, wc.opcode, target);
1317 }
1318 }
1319 }
1320
1321 static void srp_send_completion(struct ib_cq *cq, void *target_ptr)
1322 {
1323 struct srp_target_port *target = target_ptr;
1324 struct ib_wc wc;
1325 struct srp_iu *iu;
1326
1327 while (ib_poll_cq(cq, 1, &wc) > 0) {
1328 if (likely(wc.status == IB_WC_SUCCESS)) {
1329 iu = (struct srp_iu *) (uintptr_t) wc.wr_id;
1330 list_add(&iu->list, &target->free_tx);
1331 } else {
1332 srp_handle_qp_err(wc.status, wc.opcode, target);
1333 }
1334 }
1335 }
1336
1337 static int srp_queuecommand(struct Scsi_Host *shost, struct scsi_cmnd *scmnd)
1338 {
1339 struct srp_target_port *target = host_to_target(shost);
1340 struct srp_request *req;
1341 struct srp_iu *iu;
1342 struct srp_cmd *cmd;
1343 struct ib_device *dev;
1344 unsigned long flags;
1345 int len;
1346
1347 spin_lock_irqsave(&target->lock, flags);
1348 iu = __srp_get_tx_iu(target, SRP_IU_CMD);
1349 if (!iu)
1350 goto err_unlock;
1351
1352 req = list_first_entry(&target->free_reqs, struct srp_request, list);
1353 list_del(&req->list);
1354 spin_unlock_irqrestore(&target->lock, flags);
1355
1356 dev = target->srp_host->srp_dev->dev;
1357 ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_iu_len,
1358 DMA_TO_DEVICE);
1359
1360 scmnd->result = 0;
1361 scmnd->host_scribble = (void *) req;
1362
1363 cmd = iu->buf;
1364 memset(cmd, 0, sizeof *cmd);
1365
1366 cmd->opcode = SRP_CMD;
1367 cmd->lun = cpu_to_be64((u64) scmnd->device->lun << 48);
1368 cmd->tag = req->index;
1369 memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
1370
1371 req->scmnd = scmnd;
1372 req->cmd = iu;
1373
1374 len = srp_map_data(scmnd, target, req);
1375 if (len < 0) {
1376 shost_printk(KERN_ERR, target->scsi_host,
1377 PFX "Failed to map data\n");
1378 goto err_iu;
1379 }
1380
1381 ib_dma_sync_single_for_device(dev, iu->dma, target->max_iu_len,
1382 DMA_TO_DEVICE);
1383
1384 if (srp_post_send(target, iu, len)) {
1385 shost_printk(KERN_ERR, target->scsi_host, PFX "Send failed\n");
1386 goto err_unmap;
1387 }
1388
1389 return 0;
1390
1391 err_unmap:
1392 srp_unmap_data(scmnd, target, req);
1393
1394 err_iu:
1395 srp_put_tx_iu(target, iu, SRP_IU_CMD);
1396
1397 spin_lock_irqsave(&target->lock, flags);
1398 list_add(&req->list, &target->free_reqs);
1399
1400 err_unlock:
1401 spin_unlock_irqrestore(&target->lock, flags);
1402
1403 return SCSI_MLQUEUE_HOST_BUSY;
1404 }
1405
1406 static int srp_alloc_iu_bufs(struct srp_target_port *target)
1407 {
1408 int i;
1409
1410 for (i = 0; i < SRP_RQ_SIZE; ++i) {
1411 target->rx_ring[i] = srp_alloc_iu(target->srp_host,
1412 target->max_ti_iu_len,
1413 GFP_KERNEL, DMA_FROM_DEVICE);
1414 if (!target->rx_ring[i])
1415 goto err;
1416 }
1417
1418 for (i = 0; i < SRP_SQ_SIZE; ++i) {
1419 target->tx_ring[i] = srp_alloc_iu(target->srp_host,
1420 target->max_iu_len,
1421 GFP_KERNEL, DMA_TO_DEVICE);
1422 if (!target->tx_ring[i])
1423 goto err;
1424
1425 list_add(&target->tx_ring[i]->list, &target->free_tx);
1426 }
1427
1428 return 0;
1429
1430 err:
1431 for (i = 0; i < SRP_RQ_SIZE; ++i) {
1432 srp_free_iu(target->srp_host, target->rx_ring[i]);
1433 target->rx_ring[i] = NULL;
1434 }
1435
1436 for (i = 0; i < SRP_SQ_SIZE; ++i) {
1437 srp_free_iu(target->srp_host, target->tx_ring[i]);
1438 target->tx_ring[i] = NULL;
1439 }
1440
1441 return -ENOMEM;
1442 }
1443
1444 static uint32_t srp_compute_rq_tmo(struct ib_qp_attr *qp_attr, int attr_mask)
1445 {
1446 uint64_t T_tr_ns, max_compl_time_ms;
1447 uint32_t rq_tmo_jiffies;
1448
1449 /*
1450 * According to section 11.2.4.2 in the IBTA spec (Modify Queue Pair,
1451 * table 91), both the QP timeout and the retry count have to be set
1452 * for RC QP's during the RTR to RTS transition.
1453 */
1454 WARN_ON_ONCE((attr_mask & (IB_QP_TIMEOUT | IB_QP_RETRY_CNT)) !=
1455 (IB_QP_TIMEOUT | IB_QP_RETRY_CNT));
1456
1457 /*
1458 * Set target->rq_tmo_jiffies to one second more than the largest time
1459 * it can take before an error completion is generated. See also
1460 * C9-140..142 in the IBTA spec for more information about how to
1461 * convert the QP Local ACK Timeout value to nanoseconds.
1462 */
1463 T_tr_ns = 4096 * (1ULL << qp_attr->timeout);
1464 max_compl_time_ms = qp_attr->retry_cnt * 4 * T_tr_ns;
1465 do_div(max_compl_time_ms, NSEC_PER_MSEC);
1466 rq_tmo_jiffies = msecs_to_jiffies(max_compl_time_ms + 1000);
1467
1468 return rq_tmo_jiffies;
1469 }
1470
1471 static void srp_cm_rep_handler(struct ib_cm_id *cm_id,
1472 struct srp_login_rsp *lrsp,
1473 struct srp_target_port *target)
1474 {
1475 struct ib_qp_attr *qp_attr = NULL;
1476 int attr_mask = 0;
1477 int ret;
1478 int i;
1479
1480 if (lrsp->opcode == SRP_LOGIN_RSP) {
1481 target->max_ti_iu_len = be32_to_cpu(lrsp->max_ti_iu_len);
1482 target->req_lim = be32_to_cpu(lrsp->req_lim_delta);
1483
1484 /*
1485 * Reserve credits for task management so we don't
1486 * bounce requests back to the SCSI mid-layer.
1487 */
1488 target->scsi_host->can_queue
1489 = min(target->req_lim - SRP_TSK_MGMT_SQ_SIZE,
1490 target->scsi_host->can_queue);
1491 } else {
1492 shost_printk(KERN_WARNING, target->scsi_host,
1493 PFX "Unhandled RSP opcode %#x\n", lrsp->opcode);
1494 ret = -ECONNRESET;
1495 goto error;
1496 }
1497
1498 if (!target->rx_ring[0]) {
1499 ret = srp_alloc_iu_bufs(target);
1500 if (ret)
1501 goto error;
1502 }
1503
1504 ret = -ENOMEM;
1505 qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
1506 if (!qp_attr)
1507 goto error;
1508
1509 qp_attr->qp_state = IB_QPS_RTR;
1510 ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1511 if (ret)
1512 goto error_free;
1513
1514 ret = ib_modify_qp(target->qp, qp_attr, attr_mask);
1515 if (ret)
1516 goto error_free;
1517
1518 for (i = 0; i < SRP_RQ_SIZE; i++) {
1519 struct srp_iu *iu = target->rx_ring[i];
1520 ret = srp_post_recv(target, iu);
1521 if (ret)
1522 goto error_free;
1523 }
1524
1525 qp_attr->qp_state = IB_QPS_RTS;
1526 ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1527 if (ret)
1528 goto error_free;
1529
1530 target->rq_tmo_jiffies = srp_compute_rq_tmo(qp_attr, attr_mask);
1531
1532 ret = ib_modify_qp(target->qp, qp_attr, attr_mask);
1533 if (ret)
1534 goto error_free;
1535
1536 ret = ib_send_cm_rtu(cm_id, NULL, 0);
1537
1538 error_free:
1539 kfree(qp_attr);
1540
1541 error:
1542 target->status = ret;
1543 }
1544
1545 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
1546 struct ib_cm_event *event,
1547 struct srp_target_port *target)
1548 {
1549 struct Scsi_Host *shost = target->scsi_host;
1550 struct ib_class_port_info *cpi;
1551 int opcode;
1552
1553 switch (event->param.rej_rcvd.reason) {
1554 case IB_CM_REJ_PORT_CM_REDIRECT:
1555 cpi = event->param.rej_rcvd.ari;
1556 target->path.dlid = cpi->redirect_lid;
1557 target->path.pkey = cpi->redirect_pkey;
1558 cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
1559 memcpy(target->path.dgid.raw, cpi->redirect_gid, 16);
1560
1561 target->status = target->path.dlid ?
1562 SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
1563 break;
1564
1565 case IB_CM_REJ_PORT_REDIRECT:
1566 if (srp_target_is_topspin(target)) {
1567 /*
1568 * Topspin/Cisco SRP gateways incorrectly send
1569 * reject reason code 25 when they mean 24
1570 * (port redirect).
1571 */
1572 memcpy(target->path.dgid.raw,
1573 event->param.rej_rcvd.ari, 16);
1574
1575 shost_printk(KERN_DEBUG, shost,
1576 PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
1577 (unsigned long long) be64_to_cpu(target->path.dgid.global.subnet_prefix),
1578 (unsigned long long) be64_to_cpu(target->path.dgid.global.interface_id));
1579
1580 target->status = SRP_PORT_REDIRECT;
1581 } else {
1582 shost_printk(KERN_WARNING, shost,
1583 " REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
1584 target->status = -ECONNRESET;
1585 }
1586 break;
1587
1588 case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
1589 shost_printk(KERN_WARNING, shost,
1590 " REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
1591 target->status = -ECONNRESET;
1592 break;
1593
1594 case IB_CM_REJ_CONSUMER_DEFINED:
1595 opcode = *(u8 *) event->private_data;
1596 if (opcode == SRP_LOGIN_REJ) {
1597 struct srp_login_rej *rej = event->private_data;
1598 u32 reason = be32_to_cpu(rej->reason);
1599
1600 if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
1601 shost_printk(KERN_WARNING, shost,
1602 PFX "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
1603 else
1604 shost_printk(KERN_WARNING, shost,
1605 PFX "SRP LOGIN REJECTED, reason 0x%08x\n", reason);
1606 } else
1607 shost_printk(KERN_WARNING, shost,
1608 " REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
1609 " opcode 0x%02x\n", opcode);
1610 target->status = -ECONNRESET;
1611 break;
1612
1613 case IB_CM_REJ_STALE_CONN:
1614 shost_printk(KERN_WARNING, shost, " REJ reason: stale connection\n");
1615 target->status = SRP_STALE_CONN;
1616 break;
1617
1618 default:
1619 shost_printk(KERN_WARNING, shost, " REJ reason 0x%x\n",
1620 event->param.rej_rcvd.reason);
1621 target->status = -ECONNRESET;
1622 }
1623 }
1624
1625 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
1626 {
1627 struct srp_target_port *target = cm_id->context;
1628 int comp = 0;
1629
1630 switch (event->event) {
1631 case IB_CM_REQ_ERROR:
1632 shost_printk(KERN_DEBUG, target->scsi_host,
1633 PFX "Sending CM REQ failed\n");
1634 comp = 1;
1635 target->status = -ECONNRESET;
1636 break;
1637
1638 case IB_CM_REP_RECEIVED:
1639 comp = 1;
1640 srp_cm_rep_handler(cm_id, event->private_data, target);
1641 break;
1642
1643 case IB_CM_REJ_RECEIVED:
1644 shost_printk(KERN_DEBUG, target->scsi_host, PFX "REJ received\n");
1645 comp = 1;
1646
1647 srp_cm_rej_handler(cm_id, event, target);
1648 break;
1649
1650 case IB_CM_DREQ_RECEIVED:
1651 shost_printk(KERN_WARNING, target->scsi_host,
1652 PFX "DREQ received - connection closed\n");
1653 srp_change_conn_state(target, false);
1654 if (ib_send_cm_drep(cm_id, NULL, 0))
1655 shost_printk(KERN_ERR, target->scsi_host,
1656 PFX "Sending CM DREP failed\n");
1657 break;
1658
1659 case IB_CM_TIMEWAIT_EXIT:
1660 shost_printk(KERN_ERR, target->scsi_host,
1661 PFX "connection closed\n");
1662
1663 comp = 1;
1664 target->status = 0;
1665 break;
1666
1667 case IB_CM_MRA_RECEIVED:
1668 case IB_CM_DREQ_ERROR:
1669 case IB_CM_DREP_RECEIVED:
1670 break;
1671
1672 default:
1673 shost_printk(KERN_WARNING, target->scsi_host,
1674 PFX "Unhandled CM event %d\n", event->event);
1675 break;
1676 }
1677
1678 if (comp)
1679 complete(&target->done);
1680
1681 return 0;
1682 }
1683
1684 static int srp_send_tsk_mgmt(struct srp_target_port *target,
1685 u64 req_tag, unsigned int lun, u8 func)
1686 {
1687 struct ib_device *dev = target->srp_host->srp_dev->dev;
1688 struct srp_iu *iu;
1689 struct srp_tsk_mgmt *tsk_mgmt;
1690
1691 init_completion(&target->tsk_mgmt_done);
1692
1693 spin_lock_irq(&target->lock);
1694 iu = __srp_get_tx_iu(target, SRP_IU_TSK_MGMT);
1695 spin_unlock_irq(&target->lock);
1696
1697 if (!iu)
1698 return -1;
1699
1700 ib_dma_sync_single_for_cpu(dev, iu->dma, sizeof *tsk_mgmt,
1701 DMA_TO_DEVICE);
1702 tsk_mgmt = iu->buf;
1703 memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
1704
1705 tsk_mgmt->opcode = SRP_TSK_MGMT;
1706 tsk_mgmt->lun = cpu_to_be64((u64) lun << 48);
1707 tsk_mgmt->tag = req_tag | SRP_TAG_TSK_MGMT;
1708 tsk_mgmt->tsk_mgmt_func = func;
1709 tsk_mgmt->task_tag = req_tag;
1710
1711 ib_dma_sync_single_for_device(dev, iu->dma, sizeof *tsk_mgmt,
1712 DMA_TO_DEVICE);
1713 if (srp_post_send(target, iu, sizeof *tsk_mgmt)) {
1714 srp_put_tx_iu(target, iu, SRP_IU_TSK_MGMT);
1715 return -1;
1716 }
1717
1718 if (!wait_for_completion_timeout(&target->tsk_mgmt_done,
1719 msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
1720 return -1;
1721
1722 return 0;
1723 }
1724
1725 static int srp_abort(struct scsi_cmnd *scmnd)
1726 {
1727 struct srp_target_port *target = host_to_target(scmnd->device->host);
1728 struct srp_request *req = (struct srp_request *) scmnd->host_scribble;
1729
1730 shost_printk(KERN_ERR, target->scsi_host, "SRP abort called\n");
1731
1732 if (!req || target->qp_in_error || !srp_claim_req(target, req, scmnd))
1733 return FAILED;
1734 srp_send_tsk_mgmt(target, req->index, scmnd->device->lun,
1735 SRP_TSK_ABORT_TASK);
1736 srp_free_req(target, req, scmnd, 0);
1737 scmnd->result = DID_ABORT << 16;
1738 scmnd->scsi_done(scmnd);
1739
1740 return SUCCESS;
1741 }
1742
1743 static int srp_reset_device(struct scsi_cmnd *scmnd)
1744 {
1745 struct srp_target_port *target = host_to_target(scmnd->device->host);
1746 int i;
1747
1748 shost_printk(KERN_ERR, target->scsi_host, "SRP reset_device called\n");
1749
1750 if (target->qp_in_error)
1751 return FAILED;
1752 if (srp_send_tsk_mgmt(target, SRP_TAG_NO_REQ, scmnd->device->lun,
1753 SRP_TSK_LUN_RESET))
1754 return FAILED;
1755 if (target->tsk_mgmt_status)
1756 return FAILED;
1757
1758 for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
1759 struct srp_request *req = &target->req_ring[i];
1760 if (req->scmnd && req->scmnd->device == scmnd->device)
1761 srp_reset_req(target, req);
1762 }
1763
1764 return SUCCESS;
1765 }
1766
1767 static int srp_reset_host(struct scsi_cmnd *scmnd)
1768 {
1769 struct srp_target_port *target = host_to_target(scmnd->device->host);
1770 int ret = FAILED;
1771
1772 shost_printk(KERN_ERR, target->scsi_host, PFX "SRP reset_host called\n");
1773
1774 if (!srp_reconnect_target(target))
1775 ret = SUCCESS;
1776
1777 return ret;
1778 }
1779
1780 static int srp_slave_configure(struct scsi_device *sdev)
1781 {
1782 struct Scsi_Host *shost = sdev->host;
1783 struct srp_target_port *target = host_to_target(shost);
1784 struct request_queue *q = sdev->request_queue;
1785 unsigned long timeout;
1786
1787 if (sdev->type == TYPE_DISK) {
1788 timeout = max_t(unsigned, 30 * HZ, target->rq_tmo_jiffies);
1789 blk_queue_rq_timeout(q, timeout);
1790 }
1791
1792 return 0;
1793 }
1794
1795 static ssize_t show_id_ext(struct device *dev, struct device_attribute *attr,
1796 char *buf)
1797 {
1798 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1799
1800 return sprintf(buf, "0x%016llx\n",
1801 (unsigned long long) be64_to_cpu(target->id_ext));
1802 }
1803
1804 static ssize_t show_ioc_guid(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->ioc_guid));
1811 }
1812
1813 static ssize_t show_service_id(struct device *dev,
1814 struct device_attribute *attr, 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->service_id));
1820 }
1821
1822 static ssize_t show_pkey(struct device *dev, struct device_attribute *attr,
1823 char *buf)
1824 {
1825 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1826
1827 return sprintf(buf, "0x%04x\n", be16_to_cpu(target->path.pkey));
1828 }
1829
1830 static ssize_t show_dgid(struct device *dev, struct device_attribute *attr,
1831 char *buf)
1832 {
1833 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1834
1835 return sprintf(buf, "%pI6\n", target->path.dgid.raw);
1836 }
1837
1838 static ssize_t show_orig_dgid(struct device *dev,
1839 struct device_attribute *attr, char *buf)
1840 {
1841 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1842
1843 return sprintf(buf, "%pI6\n", target->orig_dgid);
1844 }
1845
1846 static ssize_t show_req_lim(struct device *dev,
1847 struct device_attribute *attr, char *buf)
1848 {
1849 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1850
1851 return sprintf(buf, "%d\n", target->req_lim);
1852 }
1853
1854 static ssize_t show_zero_req_lim(struct device *dev,
1855 struct device_attribute *attr, char *buf)
1856 {
1857 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1858
1859 return sprintf(buf, "%d\n", target->zero_req_lim);
1860 }
1861
1862 static ssize_t show_local_ib_port(struct device *dev,
1863 struct device_attribute *attr, char *buf)
1864 {
1865 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1866
1867 return sprintf(buf, "%d\n", target->srp_host->port);
1868 }
1869
1870 static ssize_t show_local_ib_device(struct device *dev,
1871 struct device_attribute *attr, char *buf)
1872 {
1873 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1874
1875 return sprintf(buf, "%s\n", target->srp_host->srp_dev->dev->name);
1876 }
1877
1878 static ssize_t show_cmd_sg_entries(struct device *dev,
1879 struct device_attribute *attr, char *buf)
1880 {
1881 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1882
1883 return sprintf(buf, "%u\n", target->cmd_sg_cnt);
1884 }
1885
1886 static ssize_t show_allow_ext_sg(struct device *dev,
1887 struct device_attribute *attr, char *buf)
1888 {
1889 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1890
1891 return sprintf(buf, "%s\n", target->allow_ext_sg ? "true" : "false");
1892 }
1893
1894 static DEVICE_ATTR(id_ext, S_IRUGO, show_id_ext, NULL);
1895 static DEVICE_ATTR(ioc_guid, S_IRUGO, show_ioc_guid, NULL);
1896 static DEVICE_ATTR(service_id, S_IRUGO, show_service_id, NULL);
1897 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1898 static DEVICE_ATTR(dgid, S_IRUGO, show_dgid, NULL);
1899 static DEVICE_ATTR(orig_dgid, S_IRUGO, show_orig_dgid, NULL);
1900 static DEVICE_ATTR(req_lim, S_IRUGO, show_req_lim, NULL);
1901 static DEVICE_ATTR(zero_req_lim, S_IRUGO, show_zero_req_lim, NULL);
1902 static DEVICE_ATTR(local_ib_port, S_IRUGO, show_local_ib_port, NULL);
1903 static DEVICE_ATTR(local_ib_device, S_IRUGO, show_local_ib_device, NULL);
1904 static DEVICE_ATTR(cmd_sg_entries, S_IRUGO, show_cmd_sg_entries, NULL);
1905 static DEVICE_ATTR(allow_ext_sg, S_IRUGO, show_allow_ext_sg, NULL);
1906
1907 static struct device_attribute *srp_host_attrs[] = {
1908 &dev_attr_id_ext,
1909 &dev_attr_ioc_guid,
1910 &dev_attr_service_id,
1911 &dev_attr_pkey,
1912 &dev_attr_dgid,
1913 &dev_attr_orig_dgid,
1914 &dev_attr_req_lim,
1915 &dev_attr_zero_req_lim,
1916 &dev_attr_local_ib_port,
1917 &dev_attr_local_ib_device,
1918 &dev_attr_cmd_sg_entries,
1919 &dev_attr_allow_ext_sg,
1920 NULL
1921 };
1922
1923 static struct scsi_host_template srp_template = {
1924 .module = THIS_MODULE,
1925 .name = "InfiniBand SRP initiator",
1926 .proc_name = DRV_NAME,
1927 .slave_configure = srp_slave_configure,
1928 .info = srp_target_info,
1929 .queuecommand = srp_queuecommand,
1930 .eh_abort_handler = srp_abort,
1931 .eh_device_reset_handler = srp_reset_device,
1932 .eh_host_reset_handler = srp_reset_host,
1933 .sg_tablesize = SRP_DEF_SG_TABLESIZE,
1934 .can_queue = SRP_CMD_SQ_SIZE,
1935 .this_id = -1,
1936 .cmd_per_lun = SRP_CMD_SQ_SIZE,
1937 .use_clustering = ENABLE_CLUSTERING,
1938 .shost_attrs = srp_host_attrs
1939 };
1940
1941 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
1942 {
1943 struct srp_rport_identifiers ids;
1944 struct srp_rport *rport;
1945
1946 sprintf(target->target_name, "SRP.T10:%016llX",
1947 (unsigned long long) be64_to_cpu(target->id_ext));
1948
1949 if (scsi_add_host(target->scsi_host, host->srp_dev->dev->dma_device))
1950 return -ENODEV;
1951
1952 memcpy(ids.port_id, &target->id_ext, 8);
1953 memcpy(ids.port_id + 8, &target->ioc_guid, 8);
1954 ids.roles = SRP_RPORT_ROLE_TARGET;
1955 rport = srp_rport_add(target->scsi_host, &ids);
1956 if (IS_ERR(rport)) {
1957 scsi_remove_host(target->scsi_host);
1958 return PTR_ERR(rport);
1959 }
1960
1961 spin_lock(&host->target_lock);
1962 list_add_tail(&target->list, &host->target_list);
1963 spin_unlock(&host->target_lock);
1964
1965 target->state = SRP_TARGET_LIVE;
1966 target->connected = false;
1967
1968 scsi_scan_target(&target->scsi_host->shost_gendev,
1969 0, target->scsi_id, SCAN_WILD_CARD, 0);
1970
1971 return 0;
1972 }
1973
1974 static void srp_release_dev(struct device *dev)
1975 {
1976 struct srp_host *host =
1977 container_of(dev, struct srp_host, dev);
1978
1979 complete(&host->released);
1980 }
1981
1982 static struct class srp_class = {
1983 .name = "infiniband_srp",
1984 .dev_release = srp_release_dev
1985 };
1986
1987 /*
1988 * Target ports are added by writing
1989 *
1990 * id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
1991 * pkey=<P_Key>,service_id=<service ID>
1992 *
1993 * to the add_target sysfs attribute.
1994 */
1995 enum {
1996 SRP_OPT_ERR = 0,
1997 SRP_OPT_ID_EXT = 1 << 0,
1998 SRP_OPT_IOC_GUID = 1 << 1,
1999 SRP_OPT_DGID = 1 << 2,
2000 SRP_OPT_PKEY = 1 << 3,
2001 SRP_OPT_SERVICE_ID = 1 << 4,
2002 SRP_OPT_MAX_SECT = 1 << 5,
2003 SRP_OPT_MAX_CMD_PER_LUN = 1 << 6,
2004 SRP_OPT_IO_CLASS = 1 << 7,
2005 SRP_OPT_INITIATOR_EXT = 1 << 8,
2006 SRP_OPT_CMD_SG_ENTRIES = 1 << 9,
2007 SRP_OPT_ALLOW_EXT_SG = 1 << 10,
2008 SRP_OPT_SG_TABLESIZE = 1 << 11,
2009 SRP_OPT_ALL = (SRP_OPT_ID_EXT |
2010 SRP_OPT_IOC_GUID |
2011 SRP_OPT_DGID |
2012 SRP_OPT_PKEY |
2013 SRP_OPT_SERVICE_ID),
2014 };
2015
2016 static const match_table_t srp_opt_tokens = {
2017 { SRP_OPT_ID_EXT, "id_ext=%s" },
2018 { SRP_OPT_IOC_GUID, "ioc_guid=%s" },
2019 { SRP_OPT_DGID, "dgid=%s" },
2020 { SRP_OPT_PKEY, "pkey=%x" },
2021 { SRP_OPT_SERVICE_ID, "service_id=%s" },
2022 { SRP_OPT_MAX_SECT, "max_sect=%d" },
2023 { SRP_OPT_MAX_CMD_PER_LUN, "max_cmd_per_lun=%d" },
2024 { SRP_OPT_IO_CLASS, "io_class=%x" },
2025 { SRP_OPT_INITIATOR_EXT, "initiator_ext=%s" },
2026 { SRP_OPT_CMD_SG_ENTRIES, "cmd_sg_entries=%u" },
2027 { SRP_OPT_ALLOW_EXT_SG, "allow_ext_sg=%u" },
2028 { SRP_OPT_SG_TABLESIZE, "sg_tablesize=%u" },
2029 { SRP_OPT_ERR, NULL }
2030 };
2031
2032 static int srp_parse_options(const char *buf, struct srp_target_port *target)
2033 {
2034 char *options, *sep_opt;
2035 char *p;
2036 char dgid[3];
2037 substring_t args[MAX_OPT_ARGS];
2038 int opt_mask = 0;
2039 int token;
2040 int ret = -EINVAL;
2041 int i;
2042
2043 options = kstrdup(buf, GFP_KERNEL);
2044 if (!options)
2045 return -ENOMEM;
2046
2047 sep_opt = options;
2048 while ((p = strsep(&sep_opt, ",")) != NULL) {
2049 if (!*p)
2050 continue;
2051
2052 token = match_token(p, srp_opt_tokens, args);
2053 opt_mask |= token;
2054
2055 switch (token) {
2056 case SRP_OPT_ID_EXT:
2057 p = match_strdup(args);
2058 if (!p) {
2059 ret = -ENOMEM;
2060 goto out;
2061 }
2062 target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2063 kfree(p);
2064 break;
2065
2066 case SRP_OPT_IOC_GUID:
2067 p = match_strdup(args);
2068 if (!p) {
2069 ret = -ENOMEM;
2070 goto out;
2071 }
2072 target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
2073 kfree(p);
2074 break;
2075
2076 case SRP_OPT_DGID:
2077 p = match_strdup(args);
2078 if (!p) {
2079 ret = -ENOMEM;
2080 goto out;
2081 }
2082 if (strlen(p) != 32) {
2083 pr_warn("bad dest GID parameter '%s'\n", p);
2084 kfree(p);
2085 goto out;
2086 }
2087
2088 for (i = 0; i < 16; ++i) {
2089 strlcpy(dgid, p + i * 2, 3);
2090 target->path.dgid.raw[i] = simple_strtoul(dgid, NULL, 16);
2091 }
2092 kfree(p);
2093 memcpy(target->orig_dgid, target->path.dgid.raw, 16);
2094 break;
2095
2096 case SRP_OPT_PKEY:
2097 if (match_hex(args, &token)) {
2098 pr_warn("bad P_Key parameter '%s'\n", p);
2099 goto out;
2100 }
2101 target->path.pkey = cpu_to_be16(token);
2102 break;
2103
2104 case SRP_OPT_SERVICE_ID:
2105 p = match_strdup(args);
2106 if (!p) {
2107 ret = -ENOMEM;
2108 goto out;
2109 }
2110 target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
2111 target->path.service_id = target->service_id;
2112 kfree(p);
2113 break;
2114
2115 case SRP_OPT_MAX_SECT:
2116 if (match_int(args, &token)) {
2117 pr_warn("bad max sect parameter '%s'\n", p);
2118 goto out;
2119 }
2120 target->scsi_host->max_sectors = token;
2121 break;
2122
2123 case SRP_OPT_MAX_CMD_PER_LUN:
2124 if (match_int(args, &token)) {
2125 pr_warn("bad max cmd_per_lun parameter '%s'\n",
2126 p);
2127 goto out;
2128 }
2129 target->scsi_host->cmd_per_lun = min(token, SRP_CMD_SQ_SIZE);
2130 break;
2131
2132 case SRP_OPT_IO_CLASS:
2133 if (match_hex(args, &token)) {
2134 pr_warn("bad IO class parameter '%s'\n", p);
2135 goto out;
2136 }
2137 if (token != SRP_REV10_IB_IO_CLASS &&
2138 token != SRP_REV16A_IB_IO_CLASS) {
2139 pr_warn("unknown IO class parameter value %x specified (use %x or %x).\n",
2140 token, SRP_REV10_IB_IO_CLASS,
2141 SRP_REV16A_IB_IO_CLASS);
2142 goto out;
2143 }
2144 target->io_class = token;
2145 break;
2146
2147 case SRP_OPT_INITIATOR_EXT:
2148 p = match_strdup(args);
2149 if (!p) {
2150 ret = -ENOMEM;
2151 goto out;
2152 }
2153 target->initiator_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2154 kfree(p);
2155 break;
2156
2157 case SRP_OPT_CMD_SG_ENTRIES:
2158 if (match_int(args, &token) || token < 1 || token > 255) {
2159 pr_warn("bad max cmd_sg_entries parameter '%s'\n",
2160 p);
2161 goto out;
2162 }
2163 target->cmd_sg_cnt = token;
2164 break;
2165
2166 case SRP_OPT_ALLOW_EXT_SG:
2167 if (match_int(args, &token)) {
2168 pr_warn("bad allow_ext_sg parameter '%s'\n", p);
2169 goto out;
2170 }
2171 target->allow_ext_sg = !!token;
2172 break;
2173
2174 case SRP_OPT_SG_TABLESIZE:
2175 if (match_int(args, &token) || token < 1 ||
2176 token > SCSI_MAX_SG_CHAIN_SEGMENTS) {
2177 pr_warn("bad max sg_tablesize parameter '%s'\n",
2178 p);
2179 goto out;
2180 }
2181 target->sg_tablesize = token;
2182 break;
2183
2184 default:
2185 pr_warn("unknown parameter or missing value '%s' in target creation request\n",
2186 p);
2187 goto out;
2188 }
2189 }
2190
2191 if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
2192 ret = 0;
2193 else
2194 for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
2195 if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
2196 !(srp_opt_tokens[i].token & opt_mask))
2197 pr_warn("target creation request is missing parameter '%s'\n",
2198 srp_opt_tokens[i].pattern);
2199
2200 out:
2201 kfree(options);
2202 return ret;
2203 }
2204
2205 static ssize_t srp_create_target(struct device *dev,
2206 struct device_attribute *attr,
2207 const char *buf, size_t count)
2208 {
2209 struct srp_host *host =
2210 container_of(dev, struct srp_host, dev);
2211 struct Scsi_Host *target_host;
2212 struct srp_target_port *target;
2213 struct ib_device *ibdev = host->srp_dev->dev;
2214 dma_addr_t dma_addr;
2215 int i, ret;
2216
2217 target_host = scsi_host_alloc(&srp_template,
2218 sizeof (struct srp_target_port));
2219 if (!target_host)
2220 return -ENOMEM;
2221
2222 target_host->transportt = ib_srp_transport_template;
2223 target_host->max_channel = 0;
2224 target_host->max_id = 1;
2225 target_host->max_lun = SRP_MAX_LUN;
2226 target_host->max_cmd_len = sizeof ((struct srp_cmd *) (void *) 0L)->cdb;
2227
2228 target = host_to_target(target_host);
2229
2230 target->io_class = SRP_REV16A_IB_IO_CLASS;
2231 target->scsi_host = target_host;
2232 target->srp_host = host;
2233 target->lkey = host->srp_dev->mr->lkey;
2234 target->rkey = host->srp_dev->mr->rkey;
2235 target->cmd_sg_cnt = cmd_sg_entries;
2236 target->sg_tablesize = indirect_sg_entries ? : cmd_sg_entries;
2237 target->allow_ext_sg = allow_ext_sg;
2238
2239 ret = srp_parse_options(buf, target);
2240 if (ret)
2241 goto err;
2242
2243 if (!host->srp_dev->fmr_pool && !target->allow_ext_sg &&
2244 target->cmd_sg_cnt < target->sg_tablesize) {
2245 pr_warn("No FMR pool and no external indirect descriptors, limiting sg_tablesize to cmd_sg_cnt\n");
2246 target->sg_tablesize = target->cmd_sg_cnt;
2247 }
2248
2249 target_host->sg_tablesize = target->sg_tablesize;
2250 target->indirect_size = target->sg_tablesize *
2251 sizeof (struct srp_direct_buf);
2252 target->max_iu_len = sizeof (struct srp_cmd) +
2253 sizeof (struct srp_indirect_buf) +
2254 target->cmd_sg_cnt * sizeof (struct srp_direct_buf);
2255
2256 INIT_WORK(&target->remove_work, srp_remove_work);
2257 spin_lock_init(&target->lock);
2258 INIT_LIST_HEAD(&target->free_tx);
2259 INIT_LIST_HEAD(&target->free_reqs);
2260 for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
2261 struct srp_request *req = &target->req_ring[i];
2262
2263 req->fmr_list = kmalloc(target->cmd_sg_cnt * sizeof (void *),
2264 GFP_KERNEL);
2265 req->map_page = kmalloc(SRP_FMR_SIZE * sizeof (void *),
2266 GFP_KERNEL);
2267 req->indirect_desc = kmalloc(target->indirect_size, GFP_KERNEL);
2268 if (!req->fmr_list || !req->map_page || !req->indirect_desc)
2269 goto err_free_mem;
2270
2271 dma_addr = ib_dma_map_single(ibdev, req->indirect_desc,
2272 target->indirect_size,
2273 DMA_TO_DEVICE);
2274 if (ib_dma_mapping_error(ibdev, dma_addr))
2275 goto err_free_mem;
2276
2277 req->indirect_dma_addr = dma_addr;
2278 req->index = i;
2279 list_add_tail(&req->list, &target->free_reqs);
2280 }
2281
2282 ib_query_gid(ibdev, host->port, 0, &target->path.sgid);
2283
2284 shost_printk(KERN_DEBUG, target->scsi_host, PFX
2285 "new target: id_ext %016llx ioc_guid %016llx pkey %04x "
2286 "service_id %016llx dgid %pI6\n",
2287 (unsigned long long) be64_to_cpu(target->id_ext),
2288 (unsigned long long) be64_to_cpu(target->ioc_guid),
2289 be16_to_cpu(target->path.pkey),
2290 (unsigned long long) be64_to_cpu(target->service_id),
2291 target->path.dgid.raw);
2292
2293 ret = srp_create_target_ib(target);
2294 if (ret)
2295 goto err_free_mem;
2296
2297 ret = srp_new_cm_id(target);
2298 if (ret)
2299 goto err_free_ib;
2300
2301 ret = srp_connect_target(target);
2302 if (ret) {
2303 shost_printk(KERN_ERR, target->scsi_host,
2304 PFX "Connection failed\n");
2305 goto err_cm_id;
2306 }
2307
2308 ret = srp_add_target(host, target);
2309 if (ret)
2310 goto err_disconnect;
2311
2312 return count;
2313
2314 err_disconnect:
2315 srp_disconnect_target(target);
2316
2317 err_cm_id:
2318 ib_destroy_cm_id(target->cm_id);
2319
2320 err_free_ib:
2321 srp_free_target_ib(target);
2322
2323 err_free_mem:
2324 srp_free_req_data(target);
2325
2326 err:
2327 scsi_host_put(target_host);
2328
2329 return ret;
2330 }
2331
2332 static DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
2333
2334 static ssize_t show_ibdev(struct device *dev, struct device_attribute *attr,
2335 char *buf)
2336 {
2337 struct srp_host *host = container_of(dev, struct srp_host, dev);
2338
2339 return sprintf(buf, "%s\n", host->srp_dev->dev->name);
2340 }
2341
2342 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
2343
2344 static ssize_t show_port(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, "%d\n", host->port);
2350 }
2351
2352 static DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
2353
2354 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
2355 {
2356 struct srp_host *host;
2357
2358 host = kzalloc(sizeof *host, GFP_KERNEL);
2359 if (!host)
2360 return NULL;
2361
2362 INIT_LIST_HEAD(&host->target_list);
2363 spin_lock_init(&host->target_lock);
2364 init_completion(&host->released);
2365 host->srp_dev = device;
2366 host->port = port;
2367
2368 host->dev.class = &srp_class;
2369 host->dev.parent = device->dev->dma_device;
2370 dev_set_name(&host->dev, "srp-%s-%d", device->dev->name, port);
2371
2372 if (device_register(&host->dev))
2373 goto free_host;
2374 if (device_create_file(&host->dev, &dev_attr_add_target))
2375 goto err_class;
2376 if (device_create_file(&host->dev, &dev_attr_ibdev))
2377 goto err_class;
2378 if (device_create_file(&host->dev, &dev_attr_port))
2379 goto err_class;
2380
2381 return host;
2382
2383 err_class:
2384 device_unregister(&host->dev);
2385
2386 free_host:
2387 kfree(host);
2388
2389 return NULL;
2390 }
2391
2392 static void srp_add_one(struct ib_device *device)
2393 {
2394 struct srp_device *srp_dev;
2395 struct ib_device_attr *dev_attr;
2396 struct ib_fmr_pool_param fmr_param;
2397 struct srp_host *host;
2398 int max_pages_per_fmr, fmr_page_shift, s, e, p;
2399
2400 dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
2401 if (!dev_attr)
2402 return;
2403
2404 if (ib_query_device(device, dev_attr)) {
2405 pr_warn("Query device failed for %s\n", device->name);
2406 goto free_attr;
2407 }
2408
2409 srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
2410 if (!srp_dev)
2411 goto free_attr;
2412
2413 /*
2414 * Use the smallest page size supported by the HCA, down to a
2415 * minimum of 4096 bytes. We're unlikely to build large sglists
2416 * out of smaller entries.
2417 */
2418 fmr_page_shift = max(12, ffs(dev_attr->page_size_cap) - 1);
2419 srp_dev->fmr_page_size = 1 << fmr_page_shift;
2420 srp_dev->fmr_page_mask = ~((u64) srp_dev->fmr_page_size - 1);
2421 srp_dev->fmr_max_size = srp_dev->fmr_page_size * SRP_FMR_SIZE;
2422
2423 INIT_LIST_HEAD(&srp_dev->dev_list);
2424
2425 srp_dev->dev = device;
2426 srp_dev->pd = ib_alloc_pd(device);
2427 if (IS_ERR(srp_dev->pd))
2428 goto free_dev;
2429
2430 srp_dev->mr = ib_get_dma_mr(srp_dev->pd,
2431 IB_ACCESS_LOCAL_WRITE |
2432 IB_ACCESS_REMOTE_READ |
2433 IB_ACCESS_REMOTE_WRITE);
2434 if (IS_ERR(srp_dev->mr))
2435 goto err_pd;
2436
2437 for (max_pages_per_fmr = SRP_FMR_SIZE;
2438 max_pages_per_fmr >= SRP_FMR_MIN_SIZE;
2439 max_pages_per_fmr /= 2, srp_dev->fmr_max_size /= 2) {
2440 memset(&fmr_param, 0, sizeof fmr_param);
2441 fmr_param.pool_size = SRP_FMR_POOL_SIZE;
2442 fmr_param.dirty_watermark = SRP_FMR_DIRTY_SIZE;
2443 fmr_param.cache = 1;
2444 fmr_param.max_pages_per_fmr = max_pages_per_fmr;
2445 fmr_param.page_shift = fmr_page_shift;
2446 fmr_param.access = (IB_ACCESS_LOCAL_WRITE |
2447 IB_ACCESS_REMOTE_WRITE |
2448 IB_ACCESS_REMOTE_READ);
2449
2450 srp_dev->fmr_pool = ib_create_fmr_pool(srp_dev->pd, &fmr_param);
2451 if (!IS_ERR(srp_dev->fmr_pool))
2452 break;
2453 }
2454
2455 if (IS_ERR(srp_dev->fmr_pool))
2456 srp_dev->fmr_pool = NULL;
2457
2458 if (device->node_type == RDMA_NODE_IB_SWITCH) {
2459 s = 0;
2460 e = 0;
2461 } else {
2462 s = 1;
2463 e = device->phys_port_cnt;
2464 }
2465
2466 for (p = s; p <= e; ++p) {
2467 host = srp_add_port(srp_dev, p);
2468 if (host)
2469 list_add_tail(&host->list, &srp_dev->dev_list);
2470 }
2471
2472 ib_set_client_data(device, &srp_client, srp_dev);
2473
2474 goto free_attr;
2475
2476 err_pd:
2477 ib_dealloc_pd(srp_dev->pd);
2478
2479 free_dev:
2480 kfree(srp_dev);
2481
2482 free_attr:
2483 kfree(dev_attr);
2484 }
2485
2486 static void srp_remove_one(struct ib_device *device)
2487 {
2488 struct srp_device *srp_dev;
2489 struct srp_host *host, *tmp_host;
2490 struct srp_target_port *target;
2491
2492 srp_dev = ib_get_client_data(device, &srp_client);
2493
2494 list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
2495 device_unregister(&host->dev);
2496 /*
2497 * Wait for the sysfs entry to go away, so that no new
2498 * target ports can be created.
2499 */
2500 wait_for_completion(&host->released);
2501
2502 /*
2503 * Remove all target ports.
2504 */
2505 spin_lock(&host->target_lock);
2506 list_for_each_entry(target, &host->target_list, list)
2507 srp_queue_remove_work(target);
2508 spin_unlock(&host->target_lock);
2509
2510 /*
2511 * Wait for target port removal tasks.
2512 */
2513 flush_workqueue(system_long_wq);
2514
2515 kfree(host);
2516 }
2517
2518 if (srp_dev->fmr_pool)
2519 ib_destroy_fmr_pool(srp_dev->fmr_pool);
2520 ib_dereg_mr(srp_dev->mr);
2521 ib_dealloc_pd(srp_dev->pd);
2522
2523 kfree(srp_dev);
2524 }
2525
2526 static struct srp_function_template ib_srp_transport_functions = {
2527 };
2528
2529 static int __init srp_init_module(void)
2530 {
2531 int ret;
2532
2533 BUILD_BUG_ON(FIELD_SIZEOF(struct ib_wc, wr_id) < sizeof(void *));
2534
2535 if (srp_sg_tablesize) {
2536 pr_warn("srp_sg_tablesize is deprecated, please use cmd_sg_entries\n");
2537 if (!cmd_sg_entries)
2538 cmd_sg_entries = srp_sg_tablesize;
2539 }
2540
2541 if (!cmd_sg_entries)
2542 cmd_sg_entries = SRP_DEF_SG_TABLESIZE;
2543
2544 if (cmd_sg_entries > 255) {
2545 pr_warn("Clamping cmd_sg_entries to 255\n");
2546 cmd_sg_entries = 255;
2547 }
2548
2549 if (!indirect_sg_entries)
2550 indirect_sg_entries = cmd_sg_entries;
2551 else if (indirect_sg_entries < cmd_sg_entries) {
2552 pr_warn("Bumping up indirect_sg_entries to match cmd_sg_entries (%u)\n",
2553 cmd_sg_entries);
2554 indirect_sg_entries = cmd_sg_entries;
2555 }
2556
2557 ib_srp_transport_template =
2558 srp_attach_transport(&ib_srp_transport_functions);
2559 if (!ib_srp_transport_template)
2560 return -ENOMEM;
2561
2562 ret = class_register(&srp_class);
2563 if (ret) {
2564 pr_err("couldn't register class infiniband_srp\n");
2565 srp_release_transport(ib_srp_transport_template);
2566 return ret;
2567 }
2568
2569 ib_sa_register_client(&srp_sa_client);
2570
2571 ret = ib_register_client(&srp_client);
2572 if (ret) {
2573 pr_err("couldn't register IB client\n");
2574 srp_release_transport(ib_srp_transport_template);
2575 ib_sa_unregister_client(&srp_sa_client);
2576 class_unregister(&srp_class);
2577 return ret;
2578 }
2579
2580 return 0;
2581 }
2582
2583 static void __exit srp_cleanup_module(void)
2584 {
2585 ib_unregister_client(&srp_client);
2586 ib_sa_unregister_client(&srp_sa_client);
2587 class_unregister(&srp_class);
2588 srp_release_transport(ib_srp_transport_template);
2589 }
2590
2591 module_init(srp_init_module);
2592 module_exit(srp_cleanup_module);
Linux kernel - Deliverables
This page took 0.080169 seconds and 4 git commands to generate.