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