0bfa545675b8e116832deeda2b9234c16e13a73c
[linux-2.6.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 #include <linux/module.h>
34 #include <linux/init.h>
35 #include <linux/slab.h>
36 #include <linux/err.h>
37 #include <linux/string.h>
38 #include <linux/parser.h>
39 #include <linux/random.h>
40 #include <linux/jiffies.h>
41
42 #include <linux/atomic.h>
43
44 #include <scsi/scsi.h>
45 #include <scsi/scsi_device.h>
46 #include <scsi/scsi_dbg.h>
47 #include <scsi/srp.h>
48 #include <scsi/scsi_transport_srp.h>
49
50 #include "ib_srp.h"
51
52 #define DRV_NAME        "ib_srp"
53 #define PFX             DRV_NAME ": "
54 #define DRV_VERSION     "0.2"
55 #define DRV_RELDATE     "November 1, 2005"
56
57 MODULE_AUTHOR("Roland Dreier");
58 MODULE_DESCRIPTION("InfiniBand SCSI RDMA Protocol initiator "
59                    "v" DRV_VERSION " (" DRV_RELDATE ")");
60 MODULE_LICENSE("Dual BSD/GPL");
61
62 static unsigned int srp_sg_tablesize;
63 static unsigned int cmd_sg_entries;
64 static unsigned int indirect_sg_entries;
65 static bool allow_ext_sg;
66 static int topspin_workarounds = 1;
67
68 module_param(srp_sg_tablesize, uint, 0444);
69 MODULE_PARM_DESC(srp_sg_tablesize, "Deprecated name for cmd_sg_entries");
70
71 module_param(cmd_sg_entries, uint, 0444);
72 MODULE_PARM_DESC(cmd_sg_entries,
73                  "Default number of gather/scatter entries in the SRP command (default is 12, max 255)");
74
75 module_param(indirect_sg_entries, uint, 0444);
76 MODULE_PARM_DESC(indirect_sg_entries,
77                  "Default max number of gather/scatter entries (default is 12, max is " __stringify(SCSI_MAX_SG_CHAIN_SEGMENTS) ")");
78
79 module_param(allow_ext_sg, bool, 0444);
80 MODULE_PARM_DESC(allow_ext_sg,
81                   "Default behavior when there are more than cmd_sg_entries S/G entries after mapping; fails the request when false (default false)");
82
83 module_param(topspin_workarounds, int, 0444);
84 MODULE_PARM_DESC(topspin_workarounds,
85                  "Enable workarounds for Topspin/Cisco SRP target bugs if != 0");
86
87 static void srp_add_one(struct ib_device *device);
88 static void srp_remove_one(struct ib_device *device);
89 static void srp_recv_completion(struct ib_cq *cq, void *target_ptr);
90 static void srp_send_completion(struct ib_cq *cq, void *target_ptr);
91 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event);
92
93 static struct scsi_transport_template *ib_srp_transport_template;
94
95 static struct ib_client srp_client = {
96         .name   = "srp",
97         .add    = srp_add_one,
98         .remove = srp_remove_one
99 };
100
101 static struct ib_sa_client srp_sa_client;
102
103 static inline struct srp_target_port *host_to_target(struct Scsi_Host *host)
104 {
105         return (struct srp_target_port *) host->hostdata;
106 }
107
108 static const char *srp_target_info(struct Scsi_Host *host)
109 {
110         return host_to_target(host)->target_name;
111 }
112
113 static int srp_target_is_topspin(struct srp_target_port *target)
114 {
115         static const u8 topspin_oui[3] = { 0x00, 0x05, 0xad };
116         static const u8 cisco_oui[3]   = { 0x00, 0x1b, 0x0d };
117
118         return topspin_workarounds &&
119                 (!memcmp(&target->ioc_guid, topspin_oui, sizeof topspin_oui) ||
120                  !memcmp(&target->ioc_guid, cisco_oui, sizeof cisco_oui));
121 }
122
123 static struct srp_iu *srp_alloc_iu(struct srp_host *host, size_t size,
124                                    gfp_t gfp_mask,
125                                    enum dma_data_direction direction)
126 {
127         struct srp_iu *iu;
128
129         iu = kmalloc(sizeof *iu, gfp_mask);
130         if (!iu)
131                 goto out;
132
133         iu->buf = kzalloc(size, gfp_mask);
134         if (!iu->buf)
135                 goto out_free_iu;
136
137         iu->dma = ib_dma_map_single(host->srp_dev->dev, iu->buf, size,
138                                     direction);
139         if (ib_dma_mapping_error(host->srp_dev->dev, iu->dma))
140                 goto out_free_buf;
141
142         iu->size      = size;
143         iu->direction = direction;
144
145         return iu;
146
147 out_free_buf:
148         kfree(iu->buf);
149 out_free_iu:
150         kfree(iu);
151 out:
152         return NULL;
153 }
154
155 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
156 {
157         if (!iu)
158                 return;
159
160         ib_dma_unmap_single(host->srp_dev->dev, iu->dma, iu->size,
161                             iu->direction);
162         kfree(iu->buf);
163         kfree(iu);
164 }
165
166 static void srp_qp_event(struct ib_event *event, void *context)
167 {
168         printk(KERN_ERR PFX "QP event %d\n", event->event);
169 }
170
171 static int srp_init_qp(struct srp_target_port *target,
172                        struct ib_qp *qp)
173 {
174         struct ib_qp_attr *attr;
175         int ret;
176
177         attr = kmalloc(sizeof *attr, GFP_KERNEL);
178         if (!attr)
179                 return -ENOMEM;
180
181         ret = ib_find_pkey(target->srp_host->srp_dev->dev,
182                            target->srp_host->port,
183                            be16_to_cpu(target->path.pkey),
184                            &attr->pkey_index);
185         if (ret)
186                 goto out;
187
188         attr->qp_state        = IB_QPS_INIT;
189         attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
190                                     IB_ACCESS_REMOTE_WRITE);
191         attr->port_num        = target->srp_host->port;
192
193         ret = ib_modify_qp(qp, attr,
194                            IB_QP_STATE          |
195                            IB_QP_PKEY_INDEX     |
196                            IB_QP_ACCESS_FLAGS   |
197                            IB_QP_PORT);
198
199 out:
200         kfree(attr);
201         return ret;
202 }
203
204 static int srp_new_cm_id(struct srp_target_port *target)
205 {
206         struct ib_cm_id *new_cm_id;
207
208         new_cm_id = ib_create_cm_id(target->srp_host->srp_dev->dev,
209                                     srp_cm_handler, target);
210         if (IS_ERR(new_cm_id))
211                 return PTR_ERR(new_cm_id);
212
213         if (target->cm_id)
214                 ib_destroy_cm_id(target->cm_id);
215         target->cm_id = new_cm_id;
216
217         return 0;
218 }
219
220 static int srp_create_target_ib(struct srp_target_port *target)
221 {
222         struct ib_qp_init_attr *init_attr;
223         int ret;
224
225         init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
226         if (!init_attr)
227                 return -ENOMEM;
228
229         target->recv_cq = ib_create_cq(target->srp_host->srp_dev->dev,
230                                        srp_recv_completion, NULL, target, SRP_RQ_SIZE, 0);
231         if (IS_ERR(target->recv_cq)) {
232                 ret = PTR_ERR(target->recv_cq);
233                 goto err;
234         }
235
236         target->send_cq = ib_create_cq(target->srp_host->srp_dev->dev,
237                                        srp_send_completion, NULL, target, SRP_SQ_SIZE, 0);
238         if (IS_ERR(target->send_cq)) {
239                 ret = PTR_ERR(target->send_cq);
240                 goto err_recv_cq;
241         }
242
243         ib_req_notify_cq(target->recv_cq, IB_CQ_NEXT_COMP);
244
245         init_attr->event_handler       = srp_qp_event;
246         init_attr->cap.max_send_wr     = SRP_SQ_SIZE;
247         init_attr->cap.max_recv_wr     = SRP_RQ_SIZE;
248         init_attr->cap.max_recv_sge    = 1;
249         init_attr->cap.max_send_sge    = 1;
250         init_attr->sq_sig_type         = IB_SIGNAL_ALL_WR;
251         init_attr->qp_type             = IB_QPT_RC;
252         init_attr->send_cq             = target->send_cq;
253         init_attr->recv_cq             = target->recv_cq;
254
255         target->qp = ib_create_qp(target->srp_host->srp_dev->pd, init_attr);
256         if (IS_ERR(target->qp)) {
257                 ret = PTR_ERR(target->qp);
258                 goto err_send_cq;
259         }
260
261         ret = srp_init_qp(target, target->qp);
262         if (ret)
263                 goto err_qp;
264
265         kfree(init_attr);
266         return 0;
267
268 err_qp:
269         ib_destroy_qp(target->qp);
270
271 err_send_cq:
272         ib_destroy_cq(target->send_cq);
273
274 err_recv_cq:
275         ib_destroy_cq(target->recv_cq);
276
277 err:
278         kfree(init_attr);
279         return ret;
280 }
281
282 static void srp_free_target_ib(struct srp_target_port *target)
283 {
284         int i;
285
286         ib_destroy_qp(target->qp);
287         ib_destroy_cq(target->send_cq);
288         ib_destroy_cq(target->recv_cq);
289
290         for (i = 0; i < SRP_RQ_SIZE; ++i)
291                 srp_free_iu(target->srp_host, target->rx_ring[i]);
292         for (i = 0; i < SRP_SQ_SIZE; ++i)
293                 srp_free_iu(target->srp_host, target->tx_ring[i]);
294 }
295
296 static void srp_path_rec_completion(int status,
297                                     struct ib_sa_path_rec *pathrec,
298                                     void *target_ptr)
299 {
300         struct srp_target_port *target = target_ptr;
301
302         target->status = status;
303         if (status)
304                 shost_printk(KERN_ERR, target->scsi_host,
305                              PFX "Got failed path rec status %d\n", status);
306         else
307                 target->path = *pathrec;
308         complete(&target->done);
309 }
310
311 static int srp_lookup_path(struct srp_target_port *target)
312 {
313         target->path.numb_path = 1;
314
315         init_completion(&target->done);
316
317         target->path_query_id = ib_sa_path_rec_get(&srp_sa_client,
318                                                    target->srp_host->srp_dev->dev,
319                                                    target->srp_host->port,
320                                                    &target->path,
321                                                    IB_SA_PATH_REC_SERVICE_ID    |
322                                                    IB_SA_PATH_REC_DGID          |
323                                                    IB_SA_PATH_REC_SGID          |
324                                                    IB_SA_PATH_REC_NUMB_PATH     |
325                                                    IB_SA_PATH_REC_PKEY,
326                                                    SRP_PATH_REC_TIMEOUT_MS,
327                                                    GFP_KERNEL,
328                                                    srp_path_rec_completion,
329                                                    target, &target->path_query);
330         if (target->path_query_id < 0)
331                 return target->path_query_id;
332
333         wait_for_completion(&target->done);
334
335         if (target->status < 0)
336                 shost_printk(KERN_WARNING, target->scsi_host,
337                              PFX "Path record query failed\n");
338
339         return target->status;
340 }
341
342 static int srp_send_req(struct srp_target_port *target)
343 {
344         struct {
345                 struct ib_cm_req_param param;
346                 struct srp_login_req   priv;
347         } *req = NULL;
348         int status;
349
350         req = kzalloc(sizeof *req, GFP_KERNEL);
351         if (!req)
352                 return -ENOMEM;
353
354         req->param.primary_path               = &target->path;
355         req->param.alternate_path             = NULL;
356         req->param.service_id                 = target->service_id;
357         req->param.qp_num                     = target->qp->qp_num;
358         req->param.qp_type                    = target->qp->qp_type;
359         req->param.private_data               = &req->priv;
360         req->param.private_data_len           = sizeof req->priv;
361         req->param.flow_control               = 1;
362
363         get_random_bytes(&req->param.starting_psn, 4);
364         req->param.starting_psn              &= 0xffffff;
365
366         /*
367          * Pick some arbitrary defaults here; we could make these
368          * module parameters if anyone cared about setting them.
369          */
370         req->param.responder_resources        = 4;
371         req->param.remote_cm_response_timeout = 20;
372         req->param.local_cm_response_timeout  = 20;
373         req->param.retry_count                = 7;
374         req->param.rnr_retry_count            = 7;
375         req->param.max_cm_retries             = 15;
376
377         req->priv.opcode        = SRP_LOGIN_REQ;
378         req->priv.tag           = 0;
379         req->priv.req_it_iu_len = cpu_to_be32(target->max_iu_len);
380         req->priv.req_buf_fmt   = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
381                                               SRP_BUF_FORMAT_INDIRECT);
382         /*
383          * In the published SRP specification (draft rev. 16a), the
384          * port identifier format is 8 bytes of ID extension followed
385          * by 8 bytes of GUID.  Older drafts put the two halves in the
386          * opposite order, so that the GUID comes first.
387          *
388          * Targets conforming to these obsolete drafts can be
389          * recognized by the I/O Class they report.
390          */
391         if (target->io_class == SRP_REV10_IB_IO_CLASS) {
392                 memcpy(req->priv.initiator_port_id,
393                        &target->path.sgid.global.interface_id, 8);
394                 memcpy(req->priv.initiator_port_id + 8,
395                        &target->initiator_ext, 8);
396                 memcpy(req->priv.target_port_id,     &target->ioc_guid, 8);
397                 memcpy(req->priv.target_port_id + 8, &target->id_ext, 8);
398         } else {
399                 memcpy(req->priv.initiator_port_id,
400                        &target->initiator_ext, 8);
401                 memcpy(req->priv.initiator_port_id + 8,
402                        &target->path.sgid.global.interface_id, 8);
403                 memcpy(req->priv.target_port_id,     &target->id_ext, 8);
404                 memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
405         }
406
407         /*
408          * Topspin/Cisco SRP targets will reject our login unless we
409          * zero out the first 8 bytes of our initiator port ID and set
410          * the second 8 bytes to the local node GUID.
411          */
412         if (srp_target_is_topspin(target)) {
413                 shost_printk(KERN_DEBUG, target->scsi_host,
414                              PFX "Topspin/Cisco initiator port ID workaround "
415                              "activated for target GUID %016llx\n",
416                              (unsigned long long) be64_to_cpu(target->ioc_guid));
417                 memset(req->priv.initiator_port_id, 0, 8);
418                 memcpy(req->priv.initiator_port_id + 8,
419                        &target->srp_host->srp_dev->dev->node_guid, 8);
420         }
421
422         status = ib_send_cm_req(target->cm_id, &req->param);
423
424         kfree(req);
425
426         return status;
427 }
428
429 static void srp_disconnect_target(struct srp_target_port *target)
430 {
431         /* XXX should send SRP_I_LOGOUT request */
432
433         init_completion(&target->done);
434         if (ib_send_cm_dreq(target->cm_id, NULL, 0)) {
435                 shost_printk(KERN_DEBUG, target->scsi_host,
436                              PFX "Sending CM DREQ failed\n");
437                 return;
438         }
439         wait_for_completion(&target->done);
440 }
441
442 static bool srp_change_state(struct srp_target_port *target,
443                             enum srp_target_state old,
444                             enum srp_target_state new)
445 {
446         bool changed = false;
447
448         spin_lock_irq(&target->lock);
449         if (target->state == old) {
450                 target->state = new;
451                 changed = true;
452         }
453         spin_unlock_irq(&target->lock);
454         return changed;
455 }
456
457 static void srp_free_req_data(struct srp_target_port *target)
458 {
459         struct ib_device *ibdev = target->srp_host->srp_dev->dev;
460         struct srp_request *req;
461         int i;
462
463         for (i = 0, req = target->req_ring; i < SRP_CMD_SQ_SIZE; ++i, ++req) {
464                 kfree(req->fmr_list);
465                 kfree(req->map_page);
466                 if (req->indirect_dma_addr) {
467                         ib_dma_unmap_single(ibdev, req->indirect_dma_addr,
468                                             target->indirect_size,
469                                             DMA_TO_DEVICE);
470                 }
471                 kfree(req->indirect_desc);
472         }
473 }
474
475 static void srp_remove_work(struct work_struct *work)
476 {
477         struct srp_target_port *target =
478                 container_of(work, struct srp_target_port, work);
479
480         if (!srp_change_state(target, SRP_TARGET_DEAD, SRP_TARGET_REMOVED))
481                 return;
482
483         spin_lock(&target->srp_host->target_lock);
484         list_del(&target->list);
485         spin_unlock(&target->srp_host->target_lock);
486
487         srp_remove_host(target->scsi_host);
488         scsi_remove_host(target->scsi_host);
489         ib_destroy_cm_id(target->cm_id);
490         srp_free_target_ib(target);
491         srp_free_req_data(target);
492         scsi_host_put(target->scsi_host);
493 }
494
495 static int srp_connect_target(struct srp_target_port *target)
496 {
497         int retries = 3;
498         int ret;
499
500         ret = srp_lookup_path(target);
501         if (ret)
502                 return ret;
503
504         while (1) {
505                 init_completion(&target->done);
506                 ret = srp_send_req(target);
507                 if (ret)
508                         return ret;
509                 wait_for_completion(&target->done);
510
511                 /*
512                  * The CM event handling code will set status to
513                  * SRP_PORT_REDIRECT if we get a port redirect REJ
514                  * back, or SRP_DLID_REDIRECT if we get a lid/qp
515                  * redirect REJ back.
516                  */
517                 switch (target->status) {
518                 case 0:
519                         return 0;
520
521                 case SRP_PORT_REDIRECT:
522                         ret = srp_lookup_path(target);
523                         if (ret)
524                                 return ret;
525                         break;
526
527                 case SRP_DLID_REDIRECT:
528                         break;
529
530                 case SRP_STALE_CONN:
531                         /* Our current CM id was stale, and is now in timewait.
532                          * Try to reconnect with a new one.
533                          */
534                         if (!retries-- || srp_new_cm_id(target)) {
535                                 shost_printk(KERN_ERR, target->scsi_host, PFX
536                                              "giving up on stale connection\n");
537                                 target->status = -ECONNRESET;
538                                 return target->status;
539                         }
540
541                         shost_printk(KERN_ERR, target->scsi_host, PFX
542                                      "retrying stale connection\n");
543                         break;
544
545                 default:
546                         return target->status;
547                 }
548         }
549 }
550
551 static void srp_unmap_data(struct scsi_cmnd *scmnd,
552                            struct srp_target_port *target,
553                            struct srp_request *req)
554 {
555         struct ib_device *ibdev = target->srp_host->srp_dev->dev;
556         struct ib_pool_fmr **pfmr;
557
558         if (!scsi_sglist(scmnd) ||
559             (scmnd->sc_data_direction != DMA_TO_DEVICE &&
560              scmnd->sc_data_direction != DMA_FROM_DEVICE))
561                 return;
562
563         pfmr = req->fmr_list;
564         while (req->nfmr--)
565                 ib_fmr_pool_unmap(*pfmr++);
566
567         ib_dma_unmap_sg(ibdev, scsi_sglist(scmnd), scsi_sg_count(scmnd),
568                         scmnd->sc_data_direction);
569 }
570
571 static void srp_remove_req(struct srp_target_port *target,
572                            struct srp_request *req, s32 req_lim_delta)
573 {
574         unsigned long flags;
575
576         srp_unmap_data(req->scmnd, target, req);
577         spin_lock_irqsave(&target->lock, flags);
578         target->req_lim += req_lim_delta;
579         req->scmnd = NULL;
580         list_add_tail(&req->list, &target->free_reqs);
581         spin_unlock_irqrestore(&target->lock, flags);
582 }
583
584 static void srp_reset_req(struct srp_target_port *target, struct srp_request *req)
585 {
586         req->scmnd->result = DID_RESET << 16;
587         req->scmnd->scsi_done(req->scmnd);
588         srp_remove_req(target, req, 0);
589 }
590
591 static int srp_reconnect_target(struct srp_target_port *target)
592 {
593         struct ib_qp_attr qp_attr;
594         struct ib_wc wc;
595         int i, ret;
596
597         if (!srp_change_state(target, SRP_TARGET_LIVE, SRP_TARGET_CONNECTING))
598                 return -EAGAIN;
599
600         srp_disconnect_target(target);
601         /*
602          * Now get a new local CM ID so that we avoid confusing the
603          * target in case things are really fouled up.
604          */
605         ret = srp_new_cm_id(target);
606         if (ret)
607                 goto err;
608
609         qp_attr.qp_state = IB_QPS_RESET;
610         ret = ib_modify_qp(target->qp, &qp_attr, IB_QP_STATE);
611         if (ret)
612                 goto err;
613
614         ret = srp_init_qp(target, target->qp);
615         if (ret)
616                 goto err;
617
618         while (ib_poll_cq(target->recv_cq, 1, &wc) > 0)
619                 ; /* nothing */
620         while (ib_poll_cq(target->send_cq, 1, &wc) > 0)
621                 ; /* nothing */
622
623         for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
624                 struct srp_request *req = &target->req_ring[i];
625                 if (req->scmnd)
626                         srp_reset_req(target, req);
627         }
628
629         INIT_LIST_HEAD(&target->free_tx);
630         for (i = 0; i < SRP_SQ_SIZE; ++i)
631                 list_add(&target->tx_ring[i]->list, &target->free_tx);
632
633         target->qp_in_error = 0;
634         ret = srp_connect_target(target);
635         if (ret)
636                 goto err;
637
638         if (!srp_change_state(target, SRP_TARGET_CONNECTING, SRP_TARGET_LIVE))
639                 ret = -EAGAIN;
640
641         return ret;
642
643 err:
644         shost_printk(KERN_ERR, target->scsi_host,
645                      PFX "reconnect failed (%d), removing target port.\n", ret);
646
647         /*
648          * We couldn't reconnect, so kill our target port off.
649          * However, we have to defer the real removal because we
650          * are in the context of the SCSI error handler now, which
651          * will deadlock if we call scsi_remove_host().
652          *
653          * Schedule our work inside the lock to avoid a race with
654          * the flush_scheduled_work() in srp_remove_one().
655          */
656         spin_lock_irq(&target->lock);
657         if (target->state == SRP_TARGET_CONNECTING) {
658                 target->state = SRP_TARGET_DEAD;
659                 INIT_WORK(&target->work, srp_remove_work);
660                 queue_work(ib_wq, &target->work);
661         }
662         spin_unlock_irq(&target->lock);
663
664         return ret;
665 }
666
667 static void srp_map_desc(struct srp_map_state *state, dma_addr_t dma_addr,
668                          unsigned int dma_len, u32 rkey)
669 {
670         struct srp_direct_buf *desc = state->desc;
671
672         desc->va = cpu_to_be64(dma_addr);
673         desc->key = cpu_to_be32(rkey);
674         desc->len = cpu_to_be32(dma_len);
675
676         state->total_len += dma_len;
677         state->desc++;
678         state->ndesc++;
679 }
680
681 static int srp_map_finish_fmr(struct srp_map_state *state,
682                               struct srp_target_port *target)
683 {
684         struct srp_device *dev = target->srp_host->srp_dev;
685         struct ib_pool_fmr *fmr;
686         u64 io_addr = 0;
687
688         if (!state->npages)
689                 return 0;
690
691         if (state->npages == 1) {
692                 srp_map_desc(state, state->base_dma_addr, state->fmr_len,
693                              target->rkey);
694                 state->npages = state->fmr_len = 0;
695                 return 0;
696         }
697
698         fmr = ib_fmr_pool_map_phys(dev->fmr_pool, state->pages,
699                                    state->npages, io_addr);
700         if (IS_ERR(fmr))
701                 return PTR_ERR(fmr);
702
703         *state->next_fmr++ = fmr;
704         state->nfmr++;
705
706         srp_map_desc(state, 0, state->fmr_len, fmr->fmr->rkey);
707         state->npages = state->fmr_len = 0;
708         return 0;
709 }
710
711 static void srp_map_update_start(struct srp_map_state *state,
712                                  struct scatterlist *sg, int sg_index,
713                                  dma_addr_t dma_addr)
714 {
715         state->unmapped_sg = sg;
716         state->unmapped_index = sg_index;
717         state->unmapped_addr = dma_addr;
718 }
719
720 static int srp_map_sg_entry(struct srp_map_state *state,
721                             struct srp_target_port *target,
722                             struct scatterlist *sg, int sg_index,
723                             int use_fmr)
724 {
725         struct srp_device *dev = target->srp_host->srp_dev;
726         struct ib_device *ibdev = dev->dev;
727         dma_addr_t dma_addr = ib_sg_dma_address(ibdev, sg);
728         unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
729         unsigned int len;
730         int ret;
731
732         if (!dma_len)
733                 return 0;
734
735         if (use_fmr == SRP_MAP_NO_FMR) {
736                 /* Once we're in direct map mode for a request, we don't
737                  * go back to FMR mode, so no need to update anything
738                  * other than the descriptor.
739                  */
740                 srp_map_desc(state, dma_addr, dma_len, target->rkey);
741                 return 0;
742         }
743
744         /* If we start at an offset into the FMR page, don't merge into
745          * the current FMR. Finish it out, and use the kernel's MR for this
746          * sg entry. This is to avoid potential bugs on some SRP targets
747          * that were never quite defined, but went away when the initiator
748          * avoided using FMR on such page fragments.
749          */
750         if (dma_addr & ~dev->fmr_page_mask || dma_len > dev->fmr_max_size) {
751                 ret = srp_map_finish_fmr(state, target);
752                 if (ret)
753                         return ret;
754
755                 srp_map_desc(state, dma_addr, dma_len, target->rkey);
756                 srp_map_update_start(state, NULL, 0, 0);
757                 return 0;
758         }
759
760         /* If this is the first sg to go into the FMR, save our position.
761          * We need to know the first unmapped entry, its index, and the
762          * first unmapped address within that entry to be able to restart
763          * mapping after an error.
764          */
765         if (!state->unmapped_sg)
766                 srp_map_update_start(state, sg, sg_index, dma_addr);
767
768         while (dma_len) {
769                 if (state->npages == SRP_FMR_SIZE) {
770                         ret = srp_map_finish_fmr(state, target);
771                         if (ret)
772                                 return ret;
773
774                         srp_map_update_start(state, sg, sg_index, dma_addr);
775                 }
776
777                 len = min_t(unsigned int, dma_len, dev->fmr_page_size);
778
779                 if (!state->npages)
780                         state->base_dma_addr = dma_addr;
781                 state->pages[state->npages++] = dma_addr;
782                 state->fmr_len += len;
783                 dma_addr += len;
784                 dma_len -= len;
785         }
786
787         /* If the last entry of the FMR wasn't a full page, then we need to
788          * close it out and start a new one -- we can only merge at page
789          * boundries.
790          */
791         ret = 0;
792         if (len != dev->fmr_page_size) {
793                 ret = srp_map_finish_fmr(state, target);
794                 if (!ret)
795                         srp_map_update_start(state, NULL, 0, 0);
796         }
797         return ret;
798 }
799
800 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_target_port *target,
801                         struct srp_request *req)
802 {
803         struct scatterlist *scat, *sg;
804         struct srp_cmd *cmd = req->cmd->buf;
805         int i, len, nents, count, use_fmr;
806         struct srp_device *dev;
807         struct ib_device *ibdev;
808         struct srp_map_state state;
809         struct srp_indirect_buf *indirect_hdr;
810         u32 table_len;
811         u8 fmt;
812
813         if (!scsi_sglist(scmnd) || scmnd->sc_data_direction == DMA_NONE)
814                 return sizeof (struct srp_cmd);
815
816         if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
817             scmnd->sc_data_direction != DMA_TO_DEVICE) {
818                 shost_printk(KERN_WARNING, target->scsi_host,
819                              PFX "Unhandled data direction %d\n",
820                              scmnd->sc_data_direction);
821                 return -EINVAL;
822         }
823
824         nents = scsi_sg_count(scmnd);
825         scat  = scsi_sglist(scmnd);
826
827         dev = target->srp_host->srp_dev;
828         ibdev = dev->dev;
829
830         count = ib_dma_map_sg(ibdev, scat, nents, scmnd->sc_data_direction);
831         if (unlikely(count == 0))
832                 return -EIO;
833
834         fmt = SRP_DATA_DESC_DIRECT;
835         len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
836
837         if (count == 1) {
838                 /*
839                  * The midlayer only generated a single gather/scatter
840                  * entry, or DMA mapping coalesced everything to a
841                  * single entry.  So a direct descriptor along with
842                  * the DMA MR suffices.
843                  */
844                 struct srp_direct_buf *buf = (void *) cmd->add_data;
845
846                 buf->va  = cpu_to_be64(ib_sg_dma_address(ibdev, scat));
847                 buf->key = cpu_to_be32(target->rkey);
848                 buf->len = cpu_to_be32(ib_sg_dma_len(ibdev, scat));
849
850                 req->nfmr = 0;
851                 goto map_complete;
852         }
853
854         /* We have more than one scatter/gather entry, so build our indirect
855          * descriptor table, trying to merge as many entries with FMR as we
856          * can.
857          */
858         indirect_hdr = (void *) cmd->add_data;
859
860         ib_dma_sync_single_for_cpu(ibdev, req->indirect_dma_addr,
861                                    target->indirect_size, DMA_TO_DEVICE);
862
863         memset(&state, 0, sizeof(state));
864         state.desc      = req->indirect_desc;
865         state.pages     = req->map_page;
866         state.next_fmr  = req->fmr_list;
867
868         use_fmr = dev->fmr_pool ? SRP_MAP_ALLOW_FMR : SRP_MAP_NO_FMR;
869
870         for_each_sg(scat, sg, count, i) {
871                 if (srp_map_sg_entry(&state, target, sg, i, use_fmr)) {
872                         /* FMR mapping failed, so backtrack to the first
873                          * unmapped entry and continue on without using FMR.
874                          */
875                         dma_addr_t dma_addr;
876                         unsigned int dma_len;
877
878 backtrack:
879                         sg = state.unmapped_sg;
880                         i = state.unmapped_index;
881
882                         dma_addr = ib_sg_dma_address(ibdev, sg);
883                         dma_len = ib_sg_dma_len(ibdev, sg);
884                         dma_len -= (state.unmapped_addr - dma_addr);
885                         dma_addr = state.unmapped_addr;
886                         use_fmr = SRP_MAP_NO_FMR;
887                         srp_map_desc(&state, dma_addr, dma_len, target->rkey);
888                 }
889         }
890
891         if (use_fmr == SRP_MAP_ALLOW_FMR && srp_map_finish_fmr(&state, target))
892                 goto backtrack;
893
894         /* We've mapped the request, now pull as much of the indirect
895          * descriptor table as we can into the command buffer. If this
896          * target is not using an external indirect table, we are
897          * guaranteed to fit into the command, as the SCSI layer won't
898          * give us more S/G entries than we allow.
899          */
900         req->nfmr = state.nfmr;
901         if (state.ndesc == 1) {
902                 /* FMR mapping was able to collapse this to one entry,
903                  * so use a direct descriptor.
904                  */
905                 struct srp_direct_buf *buf = (void *) cmd->add_data;
906
907                 *buf = req->indirect_desc[0];
908                 goto map_complete;
909         }
910
911         if (unlikely(target->cmd_sg_cnt < state.ndesc &&
912                                                 !target->allow_ext_sg)) {
913                 shost_printk(KERN_ERR, target->scsi_host,
914                              "Could not fit S/G list into SRP_CMD\n");
915                 return -EIO;
916         }
917
918         count = min(state.ndesc, target->cmd_sg_cnt);
919         table_len = state.ndesc * sizeof (struct srp_direct_buf);
920
921         fmt = SRP_DATA_DESC_INDIRECT;
922         len = sizeof(struct srp_cmd) + sizeof (struct srp_indirect_buf);
923         len += count * sizeof (struct srp_direct_buf);
924
925         memcpy(indirect_hdr->desc_list, req->indirect_desc,
926                count * sizeof (struct srp_direct_buf));
927
928         indirect_hdr->table_desc.va = cpu_to_be64(req->indirect_dma_addr);
929         indirect_hdr->table_desc.key = cpu_to_be32(target->rkey);
930         indirect_hdr->table_desc.len = cpu_to_be32(table_len);
931         indirect_hdr->len = cpu_to_be32(state.total_len);
932
933         if (scmnd->sc_data_direction == DMA_TO_DEVICE)
934                 cmd->data_out_desc_cnt = count;
935         else
936                 cmd->data_in_desc_cnt = count;
937
938         ib_dma_sync_single_for_device(ibdev, req->indirect_dma_addr, table_len,
939                                       DMA_TO_DEVICE);
940
941 map_complete:
942         if (scmnd->sc_data_direction == DMA_TO_DEVICE)
943                 cmd->buf_fmt = fmt << 4;
944         else
945                 cmd->buf_fmt = fmt;
946
947         return len;
948 }
949
950 /*
951  * Return an IU and possible credit to the free pool
952  */
953 static void srp_put_tx_iu(struct srp_target_port *target, struct srp_iu *iu,
954                           enum srp_iu_type iu_type)
955 {
956         unsigned long flags;
957
958         spin_lock_irqsave(&target->lock, flags);
959         list_add(&iu->list, &target->free_tx);
960         if (iu_type != SRP_IU_RSP)
961                 ++target->req_lim;
962         spin_unlock_irqrestore(&target->lock, flags);
963 }
964
965 /*
966  * Must be called with target->lock held to protect req_lim and free_tx.
967  * If IU is not sent, it must be returned using srp_put_tx_iu().
968  *
969  * Note:
970  * An upper limit for the number of allocated information units for each
971  * request type is:
972  * - SRP_IU_CMD: SRP_CMD_SQ_SIZE, since the SCSI mid-layer never queues
973  *   more than Scsi_Host.can_queue requests.
974  * - SRP_IU_TSK_MGMT: SRP_TSK_MGMT_SQ_SIZE.
975  * - SRP_IU_RSP: 1, since a conforming SRP target never sends more than
976  *   one unanswered SRP request to an initiator.
977  */
978 static struct srp_iu *__srp_get_tx_iu(struct srp_target_port *target,
979                                       enum srp_iu_type iu_type)
980 {
981         s32 rsv = (iu_type == SRP_IU_TSK_MGMT) ? 0 : SRP_TSK_MGMT_SQ_SIZE;
982         struct srp_iu *iu;
983
984         srp_send_completion(target->send_cq, target);
985
986         if (list_empty(&target->free_tx))
987                 return NULL;
988
989         /* Initiator responses to target requests do not consume credits */
990         if (iu_type != SRP_IU_RSP) {
991                 if (target->req_lim <= rsv) {
992                         ++target->zero_req_lim;
993                         return NULL;
994                 }
995
996                 --target->req_lim;
997         }
998
999         iu = list_first_entry(&target->free_tx, struct srp_iu, list);
1000         list_del(&iu->list);
1001         return iu;
1002 }
1003
1004 static int srp_post_send(struct srp_target_port *target,
1005                          struct srp_iu *iu, int len)
1006 {
1007         struct ib_sge list;
1008         struct ib_send_wr wr, *bad_wr;
1009
1010         list.addr   = iu->dma;
1011         list.length = len;
1012         list.lkey   = target->lkey;
1013
1014         wr.next       = NULL;
1015         wr.wr_id      = (uintptr_t) iu;
1016         wr.sg_list    = &list;
1017         wr.num_sge    = 1;
1018         wr.opcode     = IB_WR_SEND;
1019         wr.send_flags = IB_SEND_SIGNALED;
1020
1021         return ib_post_send(target->qp, &wr, &bad_wr);
1022 }
1023
1024 static int srp_post_recv(struct srp_target_port *target, struct srp_iu *iu)
1025 {
1026         struct ib_recv_wr wr, *bad_wr;
1027         struct ib_sge list;
1028
1029         list.addr   = iu->dma;
1030         list.length = iu->size;
1031         list.lkey   = target->lkey;
1032
1033         wr.next     = NULL;
1034         wr.wr_id    = (uintptr_t) iu;
1035         wr.sg_list  = &list;
1036         wr.num_sge  = 1;
1037
1038         return ib_post_recv(target->qp, &wr, &bad_wr);
1039 }
1040
1041 static void srp_process_rsp(struct srp_target_port *target, struct srp_rsp *rsp)
1042 {
1043         struct srp_request *req;
1044         struct scsi_cmnd *scmnd;
1045         unsigned long flags;
1046
1047         if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
1048                 spin_lock_irqsave(&target->lock, flags);
1049                 target->req_lim += be32_to_cpu(rsp->req_lim_delta);
1050                 spin_unlock_irqrestore(&target->lock, flags);
1051
1052                 target->tsk_mgmt_status = -1;
1053                 if (be32_to_cpu(rsp->resp_data_len) >= 4)
1054                         target->tsk_mgmt_status = rsp->data[3];
1055                 complete(&target->tsk_mgmt_done);
1056         } else {
1057                 req = &target->req_ring[rsp->tag];
1058                 scmnd = req->scmnd;
1059                 if (!scmnd)
1060                         shost_printk(KERN_ERR, target->scsi_host,
1061                                      "Null scmnd for RSP w/tag %016llx\n",
1062                                      (unsigned long long) rsp->tag);
1063                 scmnd->result = rsp->status;
1064
1065                 if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
1066                         memcpy(scmnd->sense_buffer, rsp->data +
1067                                be32_to_cpu(rsp->resp_data_len),
1068                                min_t(int, be32_to_cpu(rsp->sense_data_len),
1069                                      SCSI_SENSE_BUFFERSIZE));
1070                 }
1071
1072                 if (rsp->flags & (SRP_RSP_FLAG_DOOVER | SRP_RSP_FLAG_DOUNDER))
1073                         scsi_set_resid(scmnd, be32_to_cpu(rsp->data_out_res_cnt));
1074                 else if (rsp->flags & (SRP_RSP_FLAG_DIOVER | SRP_RSP_FLAG_DIUNDER))
1075                         scsi_set_resid(scmnd, be32_to_cpu(rsp->data_in_res_cnt));
1076
1077                 srp_remove_req(target, req, be32_to_cpu(rsp->req_lim_delta));
1078                 scmnd->host_scribble = NULL;
1079                 scmnd->scsi_done(scmnd);
1080         }
1081 }
1082
1083 static int srp_response_common(struct srp_target_port *target, s32 req_delta,
1084                                void *rsp, int len)
1085 {
1086         struct ib_device *dev = target->srp_host->srp_dev->dev;
1087         unsigned long flags;
1088         struct srp_iu *iu;
1089         int err;
1090
1091         spin_lock_irqsave(&target->lock, flags);
1092         target->req_lim += req_delta;
1093         iu = __srp_get_tx_iu(target, SRP_IU_RSP);
1094         spin_unlock_irqrestore(&target->lock, flags);
1095
1096         if (!iu) {
1097                 shost_printk(KERN_ERR, target->scsi_host, PFX
1098                              "no IU available to send response\n");
1099                 return 1;
1100         }
1101
1102         ib_dma_sync_single_for_cpu(dev, iu->dma, len, DMA_TO_DEVICE);
1103         memcpy(iu->buf, rsp, len);
1104         ib_dma_sync_single_for_device(dev, iu->dma, len, DMA_TO_DEVICE);
1105
1106         err = srp_post_send(target, iu, len);
1107         if (err) {
1108                 shost_printk(KERN_ERR, target->scsi_host, PFX
1109                              "unable to post response: %d\n", err);
1110                 srp_put_tx_iu(target, iu, SRP_IU_RSP);
1111         }
1112
1113         return err;
1114 }
1115
1116 static void srp_process_cred_req(struct srp_target_port *target,
1117                                  struct srp_cred_req *req)
1118 {
1119         struct srp_cred_rsp rsp = {
1120                 .opcode = SRP_CRED_RSP,
1121                 .tag = req->tag,
1122         };
1123         s32 delta = be32_to_cpu(req->req_lim_delta);
1124
1125         if (srp_response_common(target, delta, &rsp, sizeof rsp))
1126                 shost_printk(KERN_ERR, target->scsi_host, PFX
1127                              "problems processing SRP_CRED_REQ\n");
1128 }
1129
1130 static void srp_process_aer_req(struct srp_target_port *target,
1131                                 struct srp_aer_req *req)
1132 {
1133         struct srp_aer_rsp rsp = {
1134                 .opcode = SRP_AER_RSP,
1135                 .tag = req->tag,
1136         };
1137         s32 delta = be32_to_cpu(req->req_lim_delta);
1138
1139         shost_printk(KERN_ERR, target->scsi_host, PFX
1140                      "ignoring AER for LUN %llu\n", be64_to_cpu(req->lun));
1141
1142         if (srp_response_common(target, delta, &rsp, sizeof rsp))
1143                 shost_printk(KERN_ERR, target->scsi_host, PFX
1144                              "problems processing SRP_AER_REQ\n");
1145 }
1146
1147 static void srp_handle_recv(struct srp_target_port *target, struct ib_wc *wc)
1148 {
1149         struct ib_device *dev = target->srp_host->srp_dev->dev;
1150         struct srp_iu *iu = (struct srp_iu *) (uintptr_t) wc->wr_id;
1151         int res;
1152         u8 opcode;
1153
1154         ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_ti_iu_len,
1155                                    DMA_FROM_DEVICE);
1156
1157         opcode = *(u8 *) iu->buf;
1158
1159         if (0) {
1160                 shost_printk(KERN_ERR, target->scsi_host,
1161                              PFX "recv completion, opcode 0x%02x\n", opcode);
1162                 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 8, 1,
1163                                iu->buf, wc->byte_len, true);
1164         }
1165
1166         switch (opcode) {
1167         case SRP_RSP:
1168                 srp_process_rsp(target, iu->buf);
1169                 break;
1170
1171         case SRP_CRED_REQ:
1172                 srp_process_cred_req(target, iu->buf);
1173                 break;
1174
1175         case SRP_AER_REQ:
1176                 srp_process_aer_req(target, iu->buf);
1177                 break;
1178
1179         case SRP_T_LOGOUT:
1180                 /* XXX Handle target logout */
1181                 shost_printk(KERN_WARNING, target->scsi_host,
1182                              PFX "Got target logout request\n");
1183                 break;
1184
1185         default:
1186                 shost_printk(KERN_WARNING, target->scsi_host,
1187                              PFX "Unhandled SRP opcode 0x%02x\n", opcode);
1188                 break;
1189         }
1190
1191         ib_dma_sync_single_for_device(dev, iu->dma, target->max_ti_iu_len,
1192                                       DMA_FROM_DEVICE);
1193
1194         res = srp_post_recv(target, iu);
1195         if (res != 0)
1196                 shost_printk(KERN_ERR, target->scsi_host,
1197                              PFX "Recv failed with error code %d\n", res);
1198 }
1199
1200 static void srp_recv_completion(struct ib_cq *cq, void *target_ptr)
1201 {
1202         struct srp_target_port *target = target_ptr;
1203         struct ib_wc wc;
1204
1205         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1206         while (ib_poll_cq(cq, 1, &wc) > 0) {
1207                 if (wc.status) {
1208                         shost_printk(KERN_ERR, target->scsi_host,
1209                                      PFX "failed receive status %d\n",
1210                                      wc.status);
1211                         target->qp_in_error = 1;
1212                         break;
1213                 }
1214
1215                 srp_handle_recv(target, &wc);
1216         }
1217 }
1218
1219 static void srp_send_completion(struct ib_cq *cq, void *target_ptr)
1220 {
1221         struct srp_target_port *target = target_ptr;
1222         struct ib_wc wc;
1223         struct srp_iu *iu;
1224
1225         while (ib_poll_cq(cq, 1, &wc) > 0) {
1226                 if (wc.status) {
1227                         shost_printk(KERN_ERR, target->scsi_host,
1228                                      PFX "failed send status %d\n",
1229                                      wc.status);
1230                         target->qp_in_error = 1;
1231                         break;
1232                 }
1233
1234                 iu = (struct srp_iu *) (uintptr_t) wc.wr_id;
1235                 list_add(&iu->list, &target->free_tx);
1236         }
1237 }
1238
1239 static int srp_queuecommand(struct Scsi_Host *shost, struct scsi_cmnd *scmnd)
1240 {
1241         struct srp_target_port *target = host_to_target(shost);
1242         struct srp_request *req;
1243         struct srp_iu *iu;
1244         struct srp_cmd *cmd;
1245         struct ib_device *dev;
1246         unsigned long flags;
1247         int len;
1248
1249         if (target->state == SRP_TARGET_CONNECTING)
1250                 goto err;
1251
1252         if (target->state == SRP_TARGET_DEAD ||
1253             target->state == SRP_TARGET_REMOVED) {
1254                 scmnd->result = DID_BAD_TARGET << 16;
1255                 scmnd->scsi_done(scmnd);
1256                 return 0;
1257         }
1258
1259         spin_lock_irqsave(&target->lock, flags);
1260         iu = __srp_get_tx_iu(target, SRP_IU_CMD);
1261         if (!iu)
1262                 goto err_unlock;
1263
1264         req = list_first_entry(&target->free_reqs, struct srp_request, list);
1265         list_del(&req->list);
1266         spin_unlock_irqrestore(&target->lock, flags);
1267
1268         dev = target->srp_host->srp_dev->dev;
1269         ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_iu_len,
1270                                    DMA_TO_DEVICE);
1271
1272         scmnd->result        = 0;
1273         scmnd->host_scribble = (void *) req;
1274
1275         cmd = iu->buf;
1276         memset(cmd, 0, sizeof *cmd);
1277
1278         cmd->opcode = SRP_CMD;
1279         cmd->lun    = cpu_to_be64((u64) scmnd->device->lun << 48);
1280         cmd->tag    = req->index;
1281         memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
1282
1283         req->scmnd    = scmnd;
1284         req->cmd      = iu;
1285
1286         len = srp_map_data(scmnd, target, req);
1287         if (len < 0) {
1288                 shost_printk(KERN_ERR, target->scsi_host,
1289                              PFX "Failed to map data\n");
1290                 goto err_iu;
1291         }
1292
1293         ib_dma_sync_single_for_device(dev, iu->dma, target->max_iu_len,
1294                                       DMA_TO_DEVICE);
1295
1296         if (srp_post_send(target, iu, len)) {
1297                 shost_printk(KERN_ERR, target->scsi_host, PFX "Send failed\n");
1298                 goto err_unmap;
1299         }
1300
1301         return 0;
1302
1303 err_unmap:
1304         srp_unmap_data(scmnd, target, req);
1305
1306 err_iu:
1307         srp_put_tx_iu(target, iu, SRP_IU_CMD);
1308
1309         spin_lock_irqsave(&target->lock, flags);
1310         list_add(&req->list, &target->free_reqs);
1311
1312 err_unlock:
1313         spin_unlock_irqrestore(&target->lock, flags);
1314
1315 err:
1316         return SCSI_MLQUEUE_HOST_BUSY;
1317 }
1318
1319 static int srp_alloc_iu_bufs(struct srp_target_port *target)
1320 {
1321         int i;
1322
1323         for (i = 0; i < SRP_RQ_SIZE; ++i) {
1324                 target->rx_ring[i] = srp_alloc_iu(target->srp_host,
1325                                                   target->max_ti_iu_len,
1326                                                   GFP_KERNEL, DMA_FROM_DEVICE);
1327                 if (!target->rx_ring[i])
1328                         goto err;
1329         }
1330
1331         for (i = 0; i < SRP_SQ_SIZE; ++i) {
1332                 target->tx_ring[i] = srp_alloc_iu(target->srp_host,
1333                                                   target->max_iu_len,
1334                                                   GFP_KERNEL, DMA_TO_DEVICE);
1335                 if (!target->tx_ring[i])
1336                         goto err;
1337
1338                 list_add(&target->tx_ring[i]->list, &target->free_tx);
1339         }
1340
1341         return 0;
1342
1343 err:
1344         for (i = 0; i < SRP_RQ_SIZE; ++i) {
1345                 srp_free_iu(target->srp_host, target->rx_ring[i]);
1346                 target->rx_ring[i] = NULL;
1347         }
1348
1349         for (i = 0; i < SRP_SQ_SIZE; ++i) {
1350                 srp_free_iu(target->srp_host, target->tx_ring[i]);
1351                 target->tx_ring[i] = NULL;
1352         }
1353
1354         return -ENOMEM;
1355 }
1356
1357 static void srp_cm_rep_handler(struct ib_cm_id *cm_id,
1358                                struct srp_login_rsp *lrsp,
1359                                struct srp_target_port *target)
1360 {
1361         struct ib_qp_attr *qp_attr = NULL;
1362         int attr_mask = 0;
1363         int ret;
1364         int i;
1365
1366         if (lrsp->opcode == SRP_LOGIN_RSP) {
1367                 target->max_ti_iu_len = be32_to_cpu(lrsp->max_ti_iu_len);
1368                 target->req_lim       = be32_to_cpu(lrsp->req_lim_delta);
1369
1370                 /*
1371                  * Reserve credits for task management so we don't
1372                  * bounce requests back to the SCSI mid-layer.
1373                  */
1374                 target->scsi_host->can_queue
1375                         = min(target->req_lim - SRP_TSK_MGMT_SQ_SIZE,
1376                               target->scsi_host->can_queue);
1377         } else {
1378                 shost_printk(KERN_WARNING, target->scsi_host,
1379                              PFX "Unhandled RSP opcode %#x\n", lrsp->opcode);
1380                 ret = -ECONNRESET;
1381                 goto error;
1382         }
1383
1384         if (!target->rx_ring[0]) {
1385                 ret = srp_alloc_iu_bufs(target);
1386                 if (ret)
1387                         goto error;
1388         }
1389
1390         ret = -ENOMEM;
1391         qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
1392         if (!qp_attr)
1393                 goto error;
1394
1395         qp_attr->qp_state = IB_QPS_RTR;
1396         ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1397         if (ret)
1398                 goto error_free;
1399
1400         ret = ib_modify_qp(target->qp, qp_attr, attr_mask);
1401         if (ret)
1402                 goto error_free;
1403
1404         for (i = 0; i < SRP_RQ_SIZE; i++) {
1405                 struct srp_iu *iu = target->rx_ring[i];
1406                 ret = srp_post_recv(target, iu);
1407                 if (ret)
1408                         goto error_free;
1409         }
1410
1411         qp_attr->qp_state = IB_QPS_RTS;
1412         ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1413         if (ret)
1414                 goto error_free;
1415
1416         ret = ib_modify_qp(target->qp, qp_attr, attr_mask);
1417         if (ret)
1418                 goto error_free;
1419
1420         ret = ib_send_cm_rtu(cm_id, NULL, 0);
1421
1422 error_free:
1423         kfree(qp_attr);
1424
1425 error:
1426         target->status = ret;
1427 }
1428
1429 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
1430                                struct ib_cm_event *event,
1431                                struct srp_target_port *target)
1432 {
1433         struct Scsi_Host *shost = target->scsi_host;
1434         struct ib_class_port_info *cpi;
1435         int opcode;
1436
1437         switch (event->param.rej_rcvd.reason) {
1438         case IB_CM_REJ_PORT_CM_REDIRECT:
1439                 cpi = event->param.rej_rcvd.ari;
1440                 target->path.dlid = cpi->redirect_lid;
1441                 target->path.pkey = cpi->redirect_pkey;
1442                 cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
1443                 memcpy(target->path.dgid.raw, cpi->redirect_gid, 16);
1444
1445                 target->status = target->path.dlid ?
1446                         SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
1447                 break;
1448
1449         case IB_CM_REJ_PORT_REDIRECT:
1450                 if (srp_target_is_topspin(target)) {
1451                         /*
1452                          * Topspin/Cisco SRP gateways incorrectly send
1453                          * reject reason code 25 when they mean 24
1454                          * (port redirect).
1455                          */
1456                         memcpy(target->path.dgid.raw,
1457                                event->param.rej_rcvd.ari, 16);
1458
1459                         shost_printk(KERN_DEBUG, shost,
1460                                      PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
1461                                      (unsigned long long) be64_to_cpu(target->path.dgid.global.subnet_prefix),
1462                                      (unsigned long long) be64_to_cpu(target->path.dgid.global.interface_id));
1463
1464                         target->status = SRP_PORT_REDIRECT;
1465                 } else {
1466                         shost_printk(KERN_WARNING, shost,
1467                                      "  REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
1468                         target->status = -ECONNRESET;
1469                 }
1470                 break;
1471
1472         case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
1473                 shost_printk(KERN_WARNING, shost,
1474                             "  REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
1475                 target->status = -ECONNRESET;
1476                 break;
1477
1478         case IB_CM_REJ_CONSUMER_DEFINED:
1479                 opcode = *(u8 *) event->private_data;
1480                 if (opcode == SRP_LOGIN_REJ) {
1481                         struct srp_login_rej *rej = event->private_data;
1482                         u32 reason = be32_to_cpu(rej->reason);
1483
1484                         if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
1485                                 shost_printk(KERN_WARNING, shost,
1486                                              PFX "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
1487                         else
1488                                 shost_printk(KERN_WARNING, shost,
1489                                             PFX "SRP LOGIN REJECTED, reason 0x%08x\n", reason);
1490                 } else
1491                         shost_printk(KERN_WARNING, shost,
1492                                      "  REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
1493                                      " opcode 0x%02x\n", opcode);
1494                 target->status = -ECONNRESET;
1495                 break;
1496
1497         case IB_CM_REJ_STALE_CONN:
1498                 shost_printk(KERN_WARNING, shost, "  REJ reason: stale connection\n");
1499                 target->status = SRP_STALE_CONN;
1500                 break;
1501
1502         default:
1503                 shost_printk(KERN_WARNING, shost, "  REJ reason 0x%x\n",
1504                              event->param.rej_rcvd.reason);
1505                 target->status = -ECONNRESET;
1506         }
1507 }
1508
1509 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
1510 {
1511         struct srp_target_port *target = cm_id->context;
1512         int comp = 0;
1513
1514         switch (event->event) {
1515         case IB_CM_REQ_ERROR:
1516                 shost_printk(KERN_DEBUG, target->scsi_host,
1517                              PFX "Sending CM REQ failed\n");
1518                 comp = 1;
1519                 target->status = -ECONNRESET;
1520                 break;
1521
1522         case IB_CM_REP_RECEIVED:
1523                 comp = 1;
1524                 srp_cm_rep_handler(cm_id, event->private_data, target);
1525                 break;
1526
1527         case IB_CM_REJ_RECEIVED:
1528                 shost_printk(KERN_DEBUG, target->scsi_host, PFX "REJ received\n");
1529                 comp = 1;
1530
1531                 srp_cm_rej_handler(cm_id, event, target);
1532                 break;
1533
1534         case IB_CM_DREQ_RECEIVED:
1535                 shost_printk(KERN_WARNING, target->scsi_host,
1536                              PFX "DREQ received - connection closed\n");
1537                 if (ib_send_cm_drep(cm_id, NULL, 0))
1538                         shost_printk(KERN_ERR, target->scsi_host,
1539                                      PFX "Sending CM DREP failed\n");
1540                 break;
1541
1542         case IB_CM_TIMEWAIT_EXIT:
1543                 shost_printk(KERN_ERR, target->scsi_host,
1544                              PFX "connection closed\n");
1545
1546                 comp = 1;
1547                 target->status = 0;
1548                 break;
1549
1550         case IB_CM_MRA_RECEIVED:
1551         case IB_CM_DREQ_ERROR:
1552         case IB_CM_DREP_RECEIVED:
1553                 break;
1554
1555         default:
1556                 shost_printk(KERN_WARNING, target->scsi_host,
1557                              PFX "Unhandled CM event %d\n", event->event);
1558                 break;
1559         }
1560
1561         if (comp)
1562                 complete(&target->done);
1563
1564         return 0;
1565 }
1566
1567 static int srp_send_tsk_mgmt(struct srp_target_port *target,
1568                              u64 req_tag, unsigned int lun, u8 func)
1569 {
1570         struct ib_device *dev = target->srp_host->srp_dev->dev;
1571         struct srp_iu *iu;
1572         struct srp_tsk_mgmt *tsk_mgmt;
1573
1574         if (target->state == SRP_TARGET_DEAD ||
1575             target->state == SRP_TARGET_REMOVED)
1576                 return -1;
1577
1578         init_completion(&target->tsk_mgmt_done);
1579
1580         spin_lock_irq(&target->lock);
1581         iu = __srp_get_tx_iu(target, SRP_IU_TSK_MGMT);
1582         spin_unlock_irq(&target->lock);
1583
1584         if (!iu)
1585                 return -1;
1586
1587         ib_dma_sync_single_for_cpu(dev, iu->dma, sizeof *tsk_mgmt,
1588                                    DMA_TO_DEVICE);
1589         tsk_mgmt = iu->buf;
1590         memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
1591
1592         tsk_mgmt->opcode        = SRP_TSK_MGMT;
1593         tsk_mgmt->lun           = cpu_to_be64((u64) lun << 48);
1594         tsk_mgmt->tag           = req_tag | SRP_TAG_TSK_MGMT;
1595         tsk_mgmt->tsk_mgmt_func = func;
1596         tsk_mgmt->task_tag      = req_tag;
1597
1598         ib_dma_sync_single_for_device(dev, iu->dma, sizeof *tsk_mgmt,
1599                                       DMA_TO_DEVICE);
1600         if (srp_post_send(target, iu, sizeof *tsk_mgmt)) {
1601                 srp_put_tx_iu(target, iu, SRP_IU_TSK_MGMT);
1602                 return -1;
1603         }
1604
1605         if (!wait_for_completion_timeout(&target->tsk_mgmt_done,
1606                                          msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
1607                 return -1;
1608
1609         return 0;
1610 }
1611
1612 static int srp_abort(struct scsi_cmnd *scmnd)
1613 {
1614         struct srp_target_port *target = host_to_target(scmnd->device->host);
1615         struct srp_request *req = (struct srp_request *) scmnd->host_scribble;
1616         int ret = SUCCESS;
1617
1618         shost_printk(KERN_ERR, target->scsi_host, "SRP abort called\n");
1619
1620         if (!req || target->qp_in_error)
1621                 return FAILED;
1622         if (srp_send_tsk_mgmt(target, req->index, scmnd->device->lun,
1623                               SRP_TSK_ABORT_TASK))
1624                 return FAILED;
1625
1626         if (req->scmnd) {
1627                 if (!target->tsk_mgmt_status) {
1628                         srp_remove_req(target, req, 0);
1629                         scmnd->result = DID_ABORT << 16;
1630                 } else
1631                         ret = FAILED;
1632         }
1633
1634         return ret;
1635 }
1636
1637 static int srp_reset_device(struct scsi_cmnd *scmnd)
1638 {
1639         struct srp_target_port *target = host_to_target(scmnd->device->host);
1640         int i;
1641
1642         shost_printk(KERN_ERR, target->scsi_host, "SRP reset_device called\n");
1643
1644         if (target->qp_in_error)
1645                 return FAILED;
1646         if (srp_send_tsk_mgmt(target, SRP_TAG_NO_REQ, scmnd->device->lun,
1647                               SRP_TSK_LUN_RESET))
1648                 return FAILED;
1649         if (target->tsk_mgmt_status)
1650                 return FAILED;
1651
1652         for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
1653                 struct srp_request *req = &target->req_ring[i];
1654                 if (req->scmnd && req->scmnd->device == scmnd->device)
1655                         srp_reset_req(target, req);
1656         }
1657
1658         return SUCCESS;
1659 }
1660
1661 static int srp_reset_host(struct scsi_cmnd *scmnd)
1662 {
1663         struct srp_target_port *target = host_to_target(scmnd->device->host);
1664         int ret = FAILED;
1665
1666         shost_printk(KERN_ERR, target->scsi_host, PFX "SRP reset_host called\n");
1667
1668         if (!srp_reconnect_target(target))
1669                 ret = SUCCESS;
1670
1671         return ret;
1672 }
1673
1674 static ssize_t show_id_ext(struct device *dev, struct device_attribute *attr,
1675                            char *buf)
1676 {
1677         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1678
1679         if (target->state == SRP_TARGET_DEAD ||
1680             target->state == SRP_TARGET_REMOVED)
1681                 return -ENODEV;
1682
1683         return sprintf(buf, "0x%016llx\n",
1684                        (unsigned long long) be64_to_cpu(target->id_ext));
1685 }
1686
1687 static ssize_t show_ioc_guid(struct device *dev, struct device_attribute *attr,
1688                              char *buf)
1689 {
1690         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1691
1692         if (target->state == SRP_TARGET_DEAD ||
1693             target->state == SRP_TARGET_REMOVED)
1694                 return -ENODEV;
1695
1696         return sprintf(buf, "0x%016llx\n",
1697                        (unsigned long long) be64_to_cpu(target->ioc_guid));
1698 }
1699
1700 static ssize_t show_service_id(struct device *dev,
1701                                struct device_attribute *attr, char *buf)
1702 {
1703         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1704
1705         if (target->state == SRP_TARGET_DEAD ||
1706             target->state == SRP_TARGET_REMOVED)
1707                 return -ENODEV;
1708
1709         return sprintf(buf, "0x%016llx\n",
1710                        (unsigned long long) be64_to_cpu(target->service_id));
1711 }
1712
1713 static ssize_t show_pkey(struct device *dev, struct device_attribute *attr,
1714                          char *buf)
1715 {
1716         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1717
1718         if (target->state == SRP_TARGET_DEAD ||
1719             target->state == SRP_TARGET_REMOVED)
1720                 return -ENODEV;
1721
1722         return sprintf(buf, "0x%04x\n", be16_to_cpu(target->path.pkey));
1723 }
1724
1725 static ssize_t show_dgid(struct device *dev, struct device_attribute *attr,
1726                          char *buf)
1727 {
1728         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1729
1730         if (target->state == SRP_TARGET_DEAD ||
1731             target->state == SRP_TARGET_REMOVED)
1732                 return -ENODEV;
1733
1734         return sprintf(buf, "%pI6\n", target->path.dgid.raw);
1735 }
1736
1737 static ssize_t show_orig_dgid(struct device *dev,
1738                               struct device_attribute *attr, char *buf)
1739 {
1740         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1741
1742         if (target->state == SRP_TARGET_DEAD ||
1743             target->state == SRP_TARGET_REMOVED)
1744                 return -ENODEV;
1745
1746         return sprintf(buf, "%pI6\n", target->orig_dgid);
1747 }
1748
1749 static ssize_t show_req_lim(struct device *dev,
1750                             struct device_attribute *attr, char *buf)
1751 {
1752         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1753
1754         if (target->state == SRP_TARGET_DEAD ||
1755             target->state == SRP_TARGET_REMOVED)
1756                 return -ENODEV;
1757
1758         return sprintf(buf, "%d\n", target->req_lim);
1759 }
1760
1761 static ssize_t show_zero_req_lim(struct device *dev,
1762                                  struct device_attribute *attr, char *buf)
1763 {
1764         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1765
1766         if (target->state == SRP_TARGET_DEAD ||
1767             target->state == SRP_TARGET_REMOVED)
1768                 return -ENODEV;
1769
1770         return sprintf(buf, "%d\n", target->zero_req_lim);
1771 }
1772
1773 static ssize_t show_local_ib_port(struct device *dev,
1774                                   struct device_attribute *attr, char *buf)
1775 {
1776         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1777
1778         return sprintf(buf, "%d\n", target->srp_host->port);
1779 }
1780
1781 static ssize_t show_local_ib_device(struct device *dev,
1782                                     struct device_attribute *attr, char *buf)
1783 {
1784         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1785
1786         return sprintf(buf, "%s\n", target->srp_host->srp_dev->dev->name);
1787 }
1788
1789 static ssize_t show_cmd_sg_entries(struct device *dev,
1790                                    struct device_attribute *attr, char *buf)
1791 {
1792         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1793
1794         return sprintf(buf, "%u\n", target->cmd_sg_cnt);
1795 }
1796
1797 static ssize_t show_allow_ext_sg(struct device *dev,
1798                                  struct device_attribute *attr, char *buf)
1799 {
1800         struct srp_target_port *target = host_to_target(class_to_shost(dev));
1801
1802         return sprintf(buf, "%s\n", target->allow_ext_sg ? "true" : "false");
1803 }
1804
1805 static DEVICE_ATTR(id_ext,          S_IRUGO, show_id_ext,          NULL);
1806 static DEVICE_ATTR(ioc_guid,        S_IRUGO, show_ioc_guid,        NULL);
1807 static DEVICE_ATTR(service_id,      S_IRUGO, show_service_id,      NULL);
1808 static DEVICE_ATTR(pkey,            S_IRUGO, show_pkey,            NULL);
1809 static DEVICE_ATTR(dgid,            S_IRUGO, show_dgid,            NULL);
1810 static DEVICE_ATTR(orig_dgid,       S_IRUGO, show_orig_dgid,       NULL);
1811 static DEVICE_ATTR(req_lim,         S_IRUGO, show_req_lim,         NULL);
1812 static DEVICE_ATTR(zero_req_lim,    S_IRUGO, show_zero_req_lim,    NULL);
1813 static DEVICE_ATTR(local_ib_port,   S_IRUGO, show_local_ib_port,   NULL);
1814 static DEVICE_ATTR(local_ib_device, S_IRUGO, show_local_ib_device, NULL);
1815 static DEVICE_ATTR(cmd_sg_entries,  S_IRUGO, show_cmd_sg_entries,  NULL);
1816 static DEVICE_ATTR(allow_ext_sg,    S_IRUGO, show_allow_ext_sg,    NULL);
1817
1818 static struct device_attribute *srp_host_attrs[] = {
1819         &dev_attr_id_ext,
1820         &dev_attr_ioc_guid,
1821         &dev_attr_service_id,
1822         &dev_attr_pkey,
1823         &dev_attr_dgid,
1824         &dev_attr_orig_dgid,
1825         &dev_attr_req_lim,
1826         &dev_attr_zero_req_lim,
1827         &dev_attr_local_ib_port,
1828         &dev_attr_local_ib_device,
1829         &dev_attr_cmd_sg_entries,
1830         &dev_attr_allow_ext_sg,
1831         NULL
1832 };
1833
1834 static struct scsi_host_template srp_template = {
1835         .module                         = THIS_MODULE,
1836         .name                           = "InfiniBand SRP initiator",
1837         .proc_name                      = DRV_NAME,
1838         .info                           = srp_target_info,
1839         .queuecommand                   = srp_queuecommand,
1840         .eh_abort_handler               = srp_abort,
1841         .eh_device_reset_handler        = srp_reset_device,
1842         .eh_host_reset_handler          = srp_reset_host,
1843         .sg_tablesize                   = SRP_DEF_SG_TABLESIZE,
1844         .can_queue                      = SRP_CMD_SQ_SIZE,
1845         .this_id                        = -1,
1846         .cmd_per_lun                    = SRP_CMD_SQ_SIZE,
1847         .use_clustering                 = ENABLE_CLUSTERING,
1848         .shost_attrs                    = srp_host_attrs
1849 };
1850
1851 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
1852 {
1853         struct srp_rport_identifiers ids;
1854         struct srp_rport *rport;
1855
1856         sprintf(target->target_name, "SRP.T10:%016llX",
1857                  (unsigned long long) be64_to_cpu(target->id_ext));
1858
1859         if (scsi_add_host(target->scsi_host, host->srp_dev->dev->dma_device))
1860                 return -ENODEV;
1861
1862         memcpy(ids.port_id, &target->id_ext, 8);
1863         memcpy(ids.port_id + 8, &target->ioc_guid, 8);
1864         ids.roles = SRP_RPORT_ROLE_TARGET;
1865         rport = srp_rport_add(target->scsi_host, &ids);
1866         if (IS_ERR(rport)) {
1867                 scsi_remove_host(target->scsi_host);
1868                 return PTR_ERR(rport);
1869         }
1870
1871         spin_lock(&host->target_lock);
1872         list_add_tail(&target->list, &host->target_list);
1873         spin_unlock(&host->target_lock);
1874
1875         target->state = SRP_TARGET_LIVE;
1876
1877         scsi_scan_target(&target->scsi_host->shost_gendev,
1878                          0, target->scsi_id, SCAN_WILD_CARD, 0);
1879
1880         return 0;
1881 }
1882
1883 static void srp_release_dev(struct device *dev)
1884 {
1885         struct srp_host *host =
1886                 container_of(dev, struct srp_host, dev);
1887
1888         complete(&host->released);
1889 }
1890
1891 static struct class srp_class = {
1892         .name    = "infiniband_srp",
1893         .dev_release = srp_release_dev
1894 };
1895
1896 /*
1897  * Target ports are added by writing
1898  *
1899  *     id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
1900  *     pkey=<P_Key>,service_id=<service ID>
1901  *
1902  * to the add_target sysfs attribute.
1903  */
1904 enum {
1905         SRP_OPT_ERR             = 0,
1906         SRP_OPT_ID_EXT          = 1 << 0,
1907         SRP_OPT_IOC_GUID        = 1 << 1,
1908         SRP_OPT_DGID            = 1 << 2,
1909         SRP_OPT_PKEY            = 1 << 3,
1910         SRP_OPT_SERVICE_ID      = 1 << 4,
1911         SRP_OPT_MAX_SECT        = 1 << 5,
1912         SRP_OPT_MAX_CMD_PER_LUN = 1 << 6,
1913         SRP_OPT_IO_CLASS        = 1 << 7,
1914         SRP_OPT_INITIATOR_EXT   = 1 << 8,
1915         SRP_OPT_CMD_SG_ENTRIES  = 1 << 9,
1916         SRP_OPT_ALLOW_EXT_SG    = 1 << 10,
1917         SRP_OPT_SG_TABLESIZE    = 1 << 11,
1918         SRP_OPT_ALL             = (SRP_OPT_ID_EXT       |
1919                                    SRP_OPT_IOC_GUID     |
1920                                    SRP_OPT_DGID         |
1921                                    SRP_OPT_PKEY         |
1922                                    SRP_OPT_SERVICE_ID),
1923 };
1924
1925 static const match_table_t srp_opt_tokens = {
1926         { SRP_OPT_ID_EXT,               "id_ext=%s"             },
1927         { SRP_OPT_IOC_GUID,             "ioc_guid=%s"           },
1928         { SRP_OPT_DGID,                 "dgid=%s"               },
1929         { SRP_OPT_PKEY,                 "pkey=%x"               },
1930         { SRP_OPT_SERVICE_ID,           "service_id=%s"         },
1931         { SRP_OPT_MAX_SECT,             "max_sect=%d"           },
1932         { SRP_OPT_MAX_CMD_PER_LUN,      "max_cmd_per_lun=%d"    },
1933         { SRP_OPT_IO_CLASS,             "io_class=%x"           },
1934         { SRP_OPT_INITIATOR_EXT,        "initiator_ext=%s"      },
1935         { SRP_OPT_CMD_SG_ENTRIES,       "cmd_sg_entries=%u"     },
1936         { SRP_OPT_ALLOW_EXT_SG,         "allow_ext_sg=%u"       },
1937         { SRP_OPT_SG_TABLESIZE,         "sg_tablesize=%u"       },
1938         { SRP_OPT_ERR,                  NULL                    }
1939 };
1940
1941 static int srp_parse_options(const char *buf, struct srp_target_port *target)
1942 {
1943         char *options, *sep_opt;
1944         char *p;
1945         char dgid[3];
1946         substring_t args[MAX_OPT_ARGS];
1947         int opt_mask = 0;
1948         int token;
1949         int ret = -EINVAL;
1950         int i;
1951
1952         options = kstrdup(buf, GFP_KERNEL);
1953         if (!options)
1954                 return -ENOMEM;
1955
1956         sep_opt = options;
1957         while ((p = strsep(&sep_opt, ",")) != NULL) {
1958                 if (!*p)
1959                         continue;
1960
1961                 token = match_token(p, srp_opt_tokens, args);
1962                 opt_mask |= token;
1963
1964                 switch (token) {
1965                 case SRP_OPT_ID_EXT:
1966                         p = match_strdup(args);
1967                         if (!p) {
1968                                 ret = -ENOMEM;
1969                                 goto out;
1970                         }
1971                         target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
1972                         kfree(p);
1973                         break;
1974
1975                 case SRP_OPT_IOC_GUID:
1976                         p = match_strdup(args);
1977                         if (!p) {
1978                                 ret = -ENOMEM;
1979                                 goto out;
1980                         }
1981                         target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
1982                         kfree(p);
1983                         break;
1984
1985                 case SRP_OPT_DGID:
1986                         p = match_strdup(args);
1987                         if (!p) {
1988                                 ret = -ENOMEM;
1989                                 goto out;
1990                         }
1991                         if (strlen(p) != 32) {
1992                                 printk(KERN_WARNING PFX "bad dest GID parameter '%s'\n", p);
1993                                 kfree(p);
1994                                 goto out;
1995                         }
1996
1997                         for (i = 0; i < 16; ++i) {
1998                                 strlcpy(dgid, p + i * 2, 3);
1999                                 target->path.dgid.raw[i] = simple_strtoul(dgid, NULL, 16);
2000                         }
2001                         kfree(p);
2002                         memcpy(target->orig_dgid, target->path.dgid.raw, 16);
2003                         break;
2004
2005                 case SRP_OPT_PKEY:
2006                         if (match_hex(args, &token)) {
2007                                 printk(KERN_WARNING PFX "bad P_Key parameter '%s'\n", p);
2008                                 goto out;
2009                         }
2010                         target->path.pkey = cpu_to_be16(token);
2011                         break;
2012
2013                 case SRP_OPT_SERVICE_ID:
2014                         p = match_strdup(args);
2015                         if (!p) {
2016                                 ret = -ENOMEM;
2017                                 goto out;
2018                         }
2019                         target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
2020                         target->path.service_id = target->service_id;
2021                         kfree(p);
2022                         break;
2023
2024                 case SRP_OPT_MAX_SECT:
2025                         if (match_int(args, &token)) {
2026                                 printk(KERN_WARNING PFX "bad max sect parameter '%s'\n", p);
2027                                 goto out;
2028                         }
2029                         target->scsi_host->max_sectors = token;
2030                         break;
2031
2032                 case SRP_OPT_MAX_CMD_PER_LUN:
2033                         if (match_int(args, &token)) {
2034                                 printk(KERN_WARNING PFX "bad max cmd_per_lun parameter '%s'\n", p);
2035                                 goto out;
2036                         }
2037                         target->scsi_host->cmd_per_lun = min(token, SRP_CMD_SQ_SIZE);
2038                         break;
2039
2040                 case SRP_OPT_IO_CLASS:
2041                         if (match_hex(args, &token)) {
2042                                 printk(KERN_WARNING PFX "bad  IO class parameter '%s' \n", p);
2043                                 goto out;
2044                         }
2045                         if (token != SRP_REV10_IB_IO_CLASS &&
2046                             token != SRP_REV16A_IB_IO_CLASS) {
2047                                 printk(KERN_WARNING PFX "unknown IO class parameter value"
2048                                        " %x specified (use %x or %x).\n",
2049                                        token, SRP_REV10_IB_IO_CLASS, SRP_REV16A_IB_IO_CLASS);
2050                                 goto out;
2051                         }
2052                         target->io_class = token;
2053                         break;
2054
2055                 case SRP_OPT_INITIATOR_EXT:
2056                         p = match_strdup(args);
2057                         if (!p) {
2058                                 ret = -ENOMEM;
2059                                 goto out;
2060                         }
2061                         target->initiator_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2062                         kfree(p);
2063                         break;
2064
2065                 case SRP_OPT_CMD_SG_ENTRIES:
2066                         if (match_int(args, &token) || token < 1 || token > 255) {
2067                                 printk(KERN_WARNING PFX "bad max cmd_sg_entries parameter '%s'\n", p);
2068                                 goto out;
2069                         }
2070                         target->cmd_sg_cnt = token;
2071                         break;
2072
2073                 case SRP_OPT_ALLOW_EXT_SG:
2074                         if (match_int(args, &token)) {
2075                                 printk(KERN_WARNING PFX "bad allow_ext_sg parameter '%s'\n", p);
2076                                 goto out;
2077                         }
2078                         target->allow_ext_sg = !!token;
2079                         break;
2080
2081                 case SRP_OPT_SG_TABLESIZE:
2082                         if (match_int(args, &token) || token < 1 ||
2083                                         token > SCSI_MAX_SG_CHAIN_SEGMENTS) {
2084                                 printk(KERN_WARNING PFX "bad max sg_tablesize parameter '%s'\n", p);
2085                                 goto out;
2086                         }
2087                         target->sg_tablesize = token;
2088                         break;
2089
2090                 default:
2091                         printk(KERN_WARNING PFX "unknown parameter or missing value "
2092                                "'%s' in target creation request\n", p);
2093                         goto out;
2094                 }
2095         }
2096
2097         if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
2098                 ret = 0;
2099         else
2100                 for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
2101                         if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
2102                             !(srp_opt_tokens[i].token & opt_mask))
2103                                 printk(KERN_WARNING PFX "target creation request is "
2104                                        "missing parameter '%s'\n",
2105                                        srp_opt_tokens[i].pattern);
2106
2107 out:
2108         kfree(options);
2109         return ret;
2110 }
2111
2112 static ssize_t srp_create_target(struct device *dev,
2113                                  struct device_attribute *attr,
2114                                  const char *buf, size_t count)
2115 {
2116         struct srp_host *host =
2117                 container_of(dev, struct srp_host, dev);
2118         struct Scsi_Host *target_host;
2119         struct srp_target_port *target;
2120         struct ib_device *ibdev = host->srp_dev->dev;
2121         dma_addr_t dma_addr;
2122         int i, ret;
2123
2124         target_host = scsi_host_alloc(&srp_template,
2125                                       sizeof (struct srp_target_port));
2126         if (!target_host)
2127                 return -ENOMEM;
2128
2129         target_host->transportt  = ib_srp_transport_template;
2130         target_host->max_channel = 0;
2131         target_host->max_id      = 1;
2132         target_host->max_lun     = SRP_MAX_LUN;
2133         target_host->max_cmd_len = sizeof ((struct srp_cmd *) (void *) 0L)->cdb;
2134
2135         target = host_to_target(target_host);
2136
2137         target->io_class        = SRP_REV16A_IB_IO_CLASS;
2138         target->scsi_host       = target_host;
2139         target->srp_host        = host;
2140         target->lkey            = host->srp_dev->mr->lkey;
2141         target->rkey            = host->srp_dev->mr->rkey;
2142         target->cmd_sg_cnt      = cmd_sg_entries;
2143         target->sg_tablesize    = indirect_sg_entries ? : cmd_sg_entries;
2144         target->allow_ext_sg    = allow_ext_sg;
2145
2146         ret = srp_parse_options(buf, target);
2147         if (ret)
2148                 goto err;
2149
2150         if (!host->srp_dev->fmr_pool && !target->allow_ext_sg &&
2151                                 target->cmd_sg_cnt < target->sg_tablesize) {
2152                 printk(KERN_WARNING PFX "No FMR pool and no external indirect descriptors, limiting sg_tablesize to cmd_sg_cnt\n");
2153                 target->sg_tablesize = target->cmd_sg_cnt;
2154         }
2155
2156         target_host->sg_tablesize = target->sg_tablesize;
2157         target->indirect_size = target->sg_tablesize *
2158                                 sizeof (struct srp_direct_buf);
2159         target->max_iu_len = sizeof (struct srp_cmd) +
2160                              sizeof (struct srp_indirect_buf) +
2161                              target->cmd_sg_cnt * sizeof (struct srp_direct_buf);
2162
2163         spin_lock_init(&target->lock);
2164         INIT_LIST_HEAD(&target->free_tx);
2165         INIT_LIST_HEAD(&target->free_reqs);
2166         for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
2167                 struct srp_request *req = &target->req_ring[i];
2168
2169                 req->fmr_list = kmalloc(target->cmd_sg_cnt * sizeof (void *),
2170                                         GFP_KERNEL);
2171                 req->map_page = kmalloc(SRP_FMR_SIZE * sizeof (void *),
2172                                         GFP_KERNEL);
2173                 req->indirect_desc = kmalloc(target->indirect_size, GFP_KERNEL);
2174                 if (!req->fmr_list || !req->map_page || !req->indirect_desc)
2175                         goto err_free_mem;
2176
2177                 dma_addr = ib_dma_map_single(ibdev, req->indirect_desc,
2178                                              target->indirect_size,
2179                                              DMA_TO_DEVICE);
2180                 if (ib_dma_mapping_error(ibdev, dma_addr))
2181                         goto err_free_mem;
2182
2183                 req->indirect_dma_addr = dma_addr;
2184                 req->index = i;
2185                 list_add_tail(&req->list, &target->free_reqs);
2186         }
2187
2188         ib_query_gid(ibdev, host->port, 0, &target->path.sgid);
2189
2190         shost_printk(KERN_DEBUG, target->scsi_host, PFX
2191                      "new target: id_ext %016llx ioc_guid %016llx pkey %04x "
2192                      "service_id %016llx dgid %pI6\n",
2193                (unsigned long long) be64_to_cpu(target->id_ext),
2194                (unsigned long long) be64_to_cpu(target->ioc_guid),
2195                be16_to_cpu(target->path.pkey),
2196                (unsigned long long) be64_to_cpu(target->service_id),
2197                target->path.dgid.raw);
2198
2199         ret = srp_create_target_ib(target);
2200         if (ret)
2201                 goto err_free_mem;
2202
2203         ret = srp_new_cm_id(target);
2204         if (ret)
2205                 goto err_free_ib;
2206
2207         target->qp_in_error = 0;
2208         ret = srp_connect_target(target);
2209         if (ret) {
2210                 shost_printk(KERN_ERR, target->scsi_host,
2211                              PFX "Connection failed\n");
2212                 goto err_cm_id;
2213         }
2214
2215         ret = srp_add_target(host, target);
2216         if (ret)
2217                 goto err_disconnect;
2218
2219         return count;
2220
2221 err_disconnect:
2222         srp_disconnect_target(target);
2223
2224 err_cm_id:
2225         ib_destroy_cm_id(target->cm_id);
2226
2227 err_free_ib:
2228         srp_free_target_ib(target);
2229
2230 err_free_mem:
2231         srp_free_req_data(target);
2232
2233 err:
2234         scsi_host_put(target_host);
2235
2236         return ret;
2237 }
2238
2239 static DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
2240
2241 static ssize_t show_ibdev(struct device *dev, struct device_attribute *attr,
2242                           char *buf)
2243 {
2244         struct srp_host *host = container_of(dev, struct srp_host, dev);
2245
2246         return sprintf(buf, "%s\n", host->srp_dev->dev->name);
2247 }
2248
2249 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
2250
2251 static ssize_t show_port(struct device *dev, struct device_attribute *attr,
2252                          char *buf)
2253 {
2254         struct srp_host *host = container_of(dev, struct srp_host, dev);
2255
2256         return sprintf(buf, "%d\n", host->port);
2257 }
2258
2259 static DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
2260
2261 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
2262 {
2263         struct srp_host *host;
2264
2265         host = kzalloc(sizeof *host, GFP_KERNEL);
2266         if (!host)
2267                 return NULL;
2268
2269         INIT_LIST_HEAD(&host->target_list);
2270         spin_lock_init(&host->target_lock);
2271         init_completion(&host->released);
2272         host->srp_dev = device;
2273         host->port = port;
2274
2275         host->dev.class = &srp_class;
2276         host->dev.parent = device->dev->dma_device;
2277         dev_set_name(&host->dev, "srp-%s-%d", device->dev->name, port);
2278
2279         if (device_register(&host->dev))
2280                 goto free_host;
2281         if (device_create_file(&host->dev, &dev_attr_add_target))
2282                 goto err_class;
2283         if (device_create_file(&host->dev, &dev_attr_ibdev))
2284                 goto err_class;
2285         if (device_create_file(&host->dev, &dev_attr_port))
2286                 goto err_class;
2287
2288         return host;
2289
2290 err_class:
2291         device_unregister(&host->dev);
2292
2293 free_host:
2294         kfree(host);
2295
2296         return NULL;
2297 }
2298
2299 static void srp_add_one(struct ib_device *device)
2300 {
2301         struct srp_device *srp_dev;
2302         struct ib_device_attr *dev_attr;
2303         struct ib_fmr_pool_param fmr_param;
2304         struct srp_host *host;
2305         int max_pages_per_fmr, fmr_page_shift, s, e, p;
2306
2307         dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
2308         if (!dev_attr)
2309                 return;
2310
2311         if (ib_query_device(device, dev_attr)) {
2312                 printk(KERN_WARNING PFX "Query device failed for %s\n",
2313                        device->name);
2314                 goto free_attr;
2315         }
2316
2317         srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
2318         if (!srp_dev)
2319                 goto free_attr;
2320
2321         /*
2322          * Use the smallest page size supported by the HCA, down to a
2323          * minimum of 4096 bytes. We're unlikely to build large sglists
2324          * out of smaller entries.
2325          */
2326         fmr_page_shift          = max(12, ffs(dev_attr->page_size_cap) - 1);
2327         srp_dev->fmr_page_size  = 1 << fmr_page_shift;
2328         srp_dev->fmr_page_mask  = ~((u64) srp_dev->fmr_page_size - 1);
2329         srp_dev->fmr_max_size   = srp_dev->fmr_page_size * SRP_FMR_SIZE;
2330
2331         INIT_LIST_HEAD(&srp_dev->dev_list);
2332
2333         srp_dev->dev = device;
2334         srp_dev->pd  = ib_alloc_pd(device);
2335         if (IS_ERR(srp_dev->pd))
2336                 goto free_dev;
2337
2338         srp_dev->mr = ib_get_dma_mr(srp_dev->pd,
2339                                     IB_ACCESS_LOCAL_WRITE |
2340                                     IB_ACCESS_REMOTE_READ |
2341                                     IB_ACCESS_REMOTE_WRITE);
2342         if (IS_ERR(srp_dev->mr))
2343                 goto err_pd;
2344
2345         for (max_pages_per_fmr = SRP_FMR_SIZE;
2346                         max_pages_per_fmr >= SRP_FMR_MIN_SIZE;
2347                         max_pages_per_fmr /= 2, srp_dev->fmr_max_size /= 2) {
2348                 memset(&fmr_param, 0, sizeof fmr_param);
2349                 fmr_param.pool_size         = SRP_FMR_POOL_SIZE;
2350                 fmr_param.dirty_watermark   = SRP_FMR_DIRTY_SIZE;
2351                 fmr_param.cache             = 1;
2352                 fmr_param.max_pages_per_fmr = max_pages_per_fmr;
2353                 fmr_param.page_shift        = fmr_page_shift;
2354                 fmr_param.access            = (IB_ACCESS_LOCAL_WRITE |
2355                                                IB_ACCESS_REMOTE_WRITE |
2356                                                IB_ACCESS_REMOTE_READ);
2357
2358                 srp_dev->fmr_pool = ib_create_fmr_pool(srp_dev->pd, &fmr_param);
2359                 if (!IS_ERR(srp_dev->fmr_pool))
2360                         break;
2361         }
2362
2363         if (IS_ERR(srp_dev->fmr_pool))
2364                 srp_dev->fmr_pool = NULL;
2365
2366         if (device->node_type == RDMA_NODE_IB_SWITCH) {
2367                 s = 0;
2368                 e = 0;
2369         } else {
2370                 s = 1;
2371                 e = device->phys_port_cnt;
2372         }
2373
2374         for (p = s; p <= e; ++p) {
2375                 host = srp_add_port(srp_dev, p);
2376                 if (host)
2377                         list_add_tail(&host->list, &srp_dev->dev_list);
2378         }
2379
2380         ib_set_client_data(device, &srp_client, srp_dev);
2381
2382         goto free_attr;
2383
2384 err_pd:
2385         ib_dealloc_pd(srp_dev->pd);
2386
2387 free_dev:
2388         kfree(srp_dev);
2389
2390 free_attr:
2391         kfree(dev_attr);
2392 }
2393
2394 static void srp_remove_one(struct ib_device *device)
2395 {
2396         struct srp_device *srp_dev;
2397         struct srp_host *host, *tmp_host;
2398         LIST_HEAD(target_list);
2399         struct srp_target_port *target, *tmp_target;
2400
2401         srp_dev = ib_get_client_data(device, &srp_client);
2402
2403         list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
2404                 device_unregister(&host->dev);
2405                 /*
2406                  * Wait for the sysfs entry to go away, so that no new
2407                  * target ports can be created.
2408                  */
2409                 wait_for_completion(&host->released);
2410
2411                 /*
2412                  * Mark all target ports as removed, so we stop queueing
2413                  * commands and don't try to reconnect.
2414                  */
2415                 spin_lock(&host->target_lock);
2416                 list_for_each_entry(target, &host->target_list, list) {
2417                         spin_lock_irq(&target->lock);
2418                         target->state = SRP_TARGET_REMOVED;
2419                         spin_unlock_irq(&target->lock);
2420                 }
2421                 spin_unlock(&host->target_lock);
2422
2423                 /*
2424                  * Wait for any reconnection tasks that may have
2425                  * started before we marked our target ports as
2426                  * removed, and any target port removal tasks.
2427                  */
2428                 flush_workqueue(ib_wq);
2429
2430                 list_for_each_entry_safe(target, tmp_target,
2431                                          &host->target_list, list) {
2432                         srp_remove_host(target->scsi_host);
2433                         scsi_remove_host(target->scsi_host);
2434                         srp_disconnect_target(target);
2435                         ib_destroy_cm_id(target->cm_id);
2436                         srp_free_target_ib(target);
2437                         srp_free_req_data(target);
2438                         scsi_host_put(target->scsi_host);
2439                 }
2440
2441                 kfree(host);
2442         }
2443
2444         if (srp_dev->fmr_pool)
2445                 ib_destroy_fmr_pool(srp_dev->fmr_pool);
2446         ib_dereg_mr(srp_dev->mr);
2447         ib_dealloc_pd(srp_dev->pd);
2448
2449         kfree(srp_dev);
2450 }
2451
2452 static struct srp_function_template ib_srp_transport_functions = {
2453 };
2454
2455 static int __init srp_init_module(void)
2456 {
2457         int ret;
2458
2459         BUILD_BUG_ON(FIELD_SIZEOF(struct ib_wc, wr_id) < sizeof(void *));
2460
2461         if (srp_sg_tablesize) {
2462                 printk(KERN_WARNING PFX "srp_sg_tablesize is deprecated, please use cmd_sg_entries\n");
2463                 if (!cmd_sg_entries)
2464                         cmd_sg_entries = srp_sg_tablesize;
2465         }
2466
2467         if (!cmd_sg_entries)
2468                 cmd_sg_entries = SRP_DEF_SG_TABLESIZE;
2469
2470         if (cmd_sg_entries > 255) {
2471                 printk(KERN_WARNING PFX "Clamping cmd_sg_entries to 255\n");
2472                 cmd_sg_entries = 255;
2473         }
2474
2475         if (!indirect_sg_entries)
2476                 indirect_sg_entries = cmd_sg_entries;
2477         else if (indirect_sg_entries < cmd_sg_entries) {
2478                 printk(KERN_WARNING PFX "Bumping up indirect_sg_entries to match cmd_sg_entries (%u)\n", cmd_sg_entries);
2479                 indirect_sg_entries = cmd_sg_entries;
2480         }
2481
2482         ib_srp_transport_template =
2483                 srp_attach_transport(&ib_srp_transport_functions);
2484         if (!ib_srp_transport_template)
2485                 return -ENOMEM;
2486
2487         ret = class_register(&srp_class);
2488         if (ret) {
2489                 printk(KERN_ERR PFX "couldn't register class infiniband_srp\n");
2490                 srp_release_transport(ib_srp_transport_template);
2491                 return ret;
2492         }
2493
2494         ib_sa_register_client(&srp_sa_client);
2495
2496         ret = ib_register_client(&srp_client);
2497         if (ret) {
2498                 printk(KERN_ERR PFX "couldn't register IB client\n");
2499                 srp_release_transport(ib_srp_transport_template);
2500                 ib_sa_unregister_client(&srp_sa_client);
2501                 class_unregister(&srp_class);
2502                 return ret;
2503         }
2504
2505         return 0;
2506 }
2507
2508 static void __exit srp_cleanup_module(void)
2509 {
2510         ib_unregister_client(&srp_client);
2511         ib_sa_unregister_client(&srp_sa_client);
2512         class_unregister(&srp_class);
2513         srp_release_transport(ib_srp_transport_template);
2514 }
2515
2516 module_init(srp_init_module);
2517 module_exit(srp_cleanup_module);