IB/srp: Allow sg_tablesize to be adjusted
[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  * $Id: ib_srp.c 3932 2005-11-01 17:19:29Z roland $
33  */
34
35 #include <linux/module.h>
36 #include <linux/init.h>
37 #include <linux/slab.h>
38 #include <linux/err.h>
39 #include <linux/string.h>
40 #include <linux/parser.h>
41 #include <linux/random.h>
42 #include <linux/jiffies.h>
43
44 #include <asm/atomic.h>
45
46 #include <scsi/scsi.h>
47 #include <scsi/scsi_device.h>
48 #include <scsi/scsi_dbg.h>
49 #include <scsi/srp.h>
50
51 #include <rdma/ib_cache.h>
52
53 #include "ib_srp.h"
54
55 #define DRV_NAME        "ib_srp"
56 #define PFX             DRV_NAME ": "
57 #define DRV_VERSION     "0.2"
58 #define DRV_RELDATE     "November 1, 2005"
59
60 MODULE_AUTHOR("Roland Dreier");
61 MODULE_DESCRIPTION("InfiniBand SCSI RDMA Protocol initiator "
62                    "v" DRV_VERSION " (" DRV_RELDATE ")");
63 MODULE_LICENSE("Dual BSD/GPL");
64
65 static int srp_sg_tablesize = SRP_DEF_SG_TABLESIZE;
66 static int srp_max_iu_len;
67
68 module_param(srp_sg_tablesize, int, 0444);
69 MODULE_PARM_DESC(srp_sg_tablesize,
70                  "Max number of gather/scatter entries per I/O (default is 12)");
71
72 static int topspin_workarounds = 1;
73
74 module_param(topspin_workarounds, int, 0444);
75 MODULE_PARM_DESC(topspin_workarounds,
76                  "Enable workarounds for Topspin/Cisco SRP target bugs if != 0");
77
78 static const u8 topspin_oui[3] = { 0x00, 0x05, 0xad };
79
80 static void srp_add_one(struct ib_device *device);
81 static void srp_remove_one(struct ib_device *device);
82 static void srp_completion(struct ib_cq *cq, void *target_ptr);
83 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event);
84
85 static struct ib_client srp_client = {
86         .name   = "srp",
87         .add    = srp_add_one,
88         .remove = srp_remove_one
89 };
90
91 static inline struct srp_target_port *host_to_target(struct Scsi_Host *host)
92 {
93         return (struct srp_target_port *) host->hostdata;
94 }
95
96 static const char *srp_target_info(struct Scsi_Host *host)
97 {
98         return host_to_target(host)->target_name;
99 }
100
101 static struct srp_iu *srp_alloc_iu(struct srp_host *host, size_t size,
102                                    gfp_t gfp_mask,
103                                    enum dma_data_direction direction)
104 {
105         struct srp_iu *iu;
106
107         iu = kmalloc(sizeof *iu, gfp_mask);
108         if (!iu)
109                 goto out;
110
111         iu->buf = kzalloc(size, gfp_mask);
112         if (!iu->buf)
113                 goto out_free_iu;
114
115         iu->dma = dma_map_single(host->dev->dev->dma_device,
116                                  iu->buf, size, direction);
117         if (dma_mapping_error(iu->dma))
118                 goto out_free_buf;
119
120         iu->size      = size;
121         iu->direction = direction;
122
123         return iu;
124
125 out_free_buf:
126         kfree(iu->buf);
127 out_free_iu:
128         kfree(iu);
129 out:
130         return NULL;
131 }
132
133 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
134 {
135         if (!iu)
136                 return;
137
138         dma_unmap_single(host->dev->dev->dma_device,
139                          iu->dma, iu->size, iu->direction);
140         kfree(iu->buf);
141         kfree(iu);
142 }
143
144 static void srp_qp_event(struct ib_event *event, void *context)
145 {
146         printk(KERN_ERR PFX "QP event %d\n", event->event);
147 }
148
149 static int srp_init_qp(struct srp_target_port *target,
150                        struct ib_qp *qp)
151 {
152         struct ib_qp_attr *attr;
153         int ret;
154
155         attr = kmalloc(sizeof *attr, GFP_KERNEL);
156         if (!attr)
157                 return -ENOMEM;
158
159         ret = ib_find_cached_pkey(target->srp_host->dev->dev,
160                                   target->srp_host->port,
161                                   be16_to_cpu(target->path.pkey),
162                                   &attr->pkey_index);
163         if (ret)
164                 goto out;
165
166         attr->qp_state        = IB_QPS_INIT;
167         attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
168                                     IB_ACCESS_REMOTE_WRITE);
169         attr->port_num        = target->srp_host->port;
170
171         ret = ib_modify_qp(qp, attr,
172                            IB_QP_STATE          |
173                            IB_QP_PKEY_INDEX     |
174                            IB_QP_ACCESS_FLAGS   |
175                            IB_QP_PORT);
176
177 out:
178         kfree(attr);
179         return ret;
180 }
181
182 static int srp_create_target_ib(struct srp_target_port *target)
183 {
184         struct ib_qp_init_attr *init_attr;
185         int ret;
186
187         init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
188         if (!init_attr)
189                 return -ENOMEM;
190
191         target->cq = ib_create_cq(target->srp_host->dev->dev, srp_completion,
192                                   NULL, target, SRP_CQ_SIZE);
193         if (IS_ERR(target->cq)) {
194                 ret = PTR_ERR(target->cq);
195                 goto out;
196         }
197
198         ib_req_notify_cq(target->cq, IB_CQ_NEXT_COMP);
199
200         init_attr->event_handler       = srp_qp_event;
201         init_attr->cap.max_send_wr     = SRP_SQ_SIZE;
202         init_attr->cap.max_recv_wr     = SRP_RQ_SIZE;
203         init_attr->cap.max_recv_sge    = 1;
204         init_attr->cap.max_send_sge    = 1;
205         init_attr->sq_sig_type         = IB_SIGNAL_ALL_WR;
206         init_attr->qp_type             = IB_QPT_RC;
207         init_attr->send_cq             = target->cq;
208         init_attr->recv_cq             = target->cq;
209
210         target->qp = ib_create_qp(target->srp_host->dev->pd, init_attr);
211         if (IS_ERR(target->qp)) {
212                 ret = PTR_ERR(target->qp);
213                 ib_destroy_cq(target->cq);
214                 goto out;
215         }
216
217         ret = srp_init_qp(target, target->qp);
218         if (ret) {
219                 ib_destroy_qp(target->qp);
220                 ib_destroy_cq(target->cq);
221                 goto out;
222         }
223
224 out:
225         kfree(init_attr);
226         return ret;
227 }
228
229 static void srp_free_target_ib(struct srp_target_port *target)
230 {
231         int i;
232
233         ib_destroy_qp(target->qp);
234         ib_destroy_cq(target->cq);
235
236         for (i = 0; i < SRP_RQ_SIZE; ++i)
237                 srp_free_iu(target->srp_host, target->rx_ring[i]);
238         for (i = 0; i < SRP_SQ_SIZE + 1; ++i)
239                 srp_free_iu(target->srp_host, target->tx_ring[i]);
240 }
241
242 static void srp_path_rec_completion(int status,
243                                     struct ib_sa_path_rec *pathrec,
244                                     void *target_ptr)
245 {
246         struct srp_target_port *target = target_ptr;
247
248         target->status = status;
249         if (status)
250                 printk(KERN_ERR PFX "Got failed path rec status %d\n", status);
251         else
252                 target->path = *pathrec;
253         complete(&target->done);
254 }
255
256 static int srp_lookup_path(struct srp_target_port *target)
257 {
258         target->path.numb_path = 1;
259
260         init_completion(&target->done);
261
262         target->path_query_id = ib_sa_path_rec_get(target->srp_host->dev->dev,
263                                                    target->srp_host->port,
264                                                    &target->path,
265                                                    IB_SA_PATH_REC_DGID          |
266                                                    IB_SA_PATH_REC_SGID          |
267                                                    IB_SA_PATH_REC_NUMB_PATH     |
268                                                    IB_SA_PATH_REC_PKEY,
269                                                    SRP_PATH_REC_TIMEOUT_MS,
270                                                    GFP_KERNEL,
271                                                    srp_path_rec_completion,
272                                                    target, &target->path_query);
273         if (target->path_query_id < 0)
274                 return target->path_query_id;
275
276         wait_for_completion(&target->done);
277
278         if (target->status < 0)
279                 printk(KERN_WARNING PFX "Path record query failed\n");
280
281         return target->status;
282 }
283
284 static int srp_send_req(struct srp_target_port *target)
285 {
286         struct {
287                 struct ib_cm_req_param param;
288                 struct srp_login_req   priv;
289         } *req = NULL;
290         int status;
291
292         req = kzalloc(sizeof *req, GFP_KERNEL);
293         if (!req)
294                 return -ENOMEM;
295
296         req->param.primary_path               = &target->path;
297         req->param.alternate_path             = NULL;
298         req->param.service_id                 = target->service_id;
299         req->param.qp_num                     = target->qp->qp_num;
300         req->param.qp_type                    = target->qp->qp_type;
301         req->param.private_data               = &req->priv;
302         req->param.private_data_len           = sizeof req->priv;
303         req->param.flow_control               = 1;
304
305         get_random_bytes(&req->param.starting_psn, 4);
306         req->param.starting_psn              &= 0xffffff;
307
308         /*
309          * Pick some arbitrary defaults here; we could make these
310          * module parameters if anyone cared about setting them.
311          */
312         req->param.responder_resources        = 4;
313         req->param.remote_cm_response_timeout = 20;
314         req->param.local_cm_response_timeout  = 20;
315         req->param.retry_count                = 7;
316         req->param.rnr_retry_count            = 7;
317         req->param.max_cm_retries             = 15;
318
319         req->priv.opcode        = SRP_LOGIN_REQ;
320         req->priv.tag           = 0;
321         req->priv.req_it_iu_len = cpu_to_be32(srp_max_iu_len);
322         req->priv.req_buf_fmt   = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
323                                               SRP_BUF_FORMAT_INDIRECT);
324         memcpy(req->priv.initiator_port_id, target->srp_host->initiator_port_id, 16);
325         /*
326          * Topspin/Cisco SRP targets will reject our login unless we
327          * zero out the first 8 bytes of our initiator port ID.  The
328          * second 8 bytes must be our local node GUID, but we always
329          * use that anyway.
330          */
331         if (topspin_workarounds && !memcmp(&target->ioc_guid, topspin_oui, 3)) {
332                 printk(KERN_DEBUG PFX "Topspin/Cisco initiator port ID workaround "
333                        "activated for target GUID %016llx\n",
334                        (unsigned long long) be64_to_cpu(target->ioc_guid));
335                 memset(req->priv.initiator_port_id, 0, 8);
336         }
337         memcpy(req->priv.target_port_id,     &target->id_ext, 8);
338         memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
339
340         status = ib_send_cm_req(target->cm_id, &req->param);
341
342         kfree(req);
343
344         return status;
345 }
346
347 static void srp_disconnect_target(struct srp_target_port *target)
348 {
349         /* XXX should send SRP_I_LOGOUT request */
350
351         init_completion(&target->done);
352         if (ib_send_cm_dreq(target->cm_id, NULL, 0)) {
353                 printk(KERN_DEBUG PFX "Sending CM DREQ failed\n");
354                 return;
355         }
356         wait_for_completion(&target->done);
357 }
358
359 static void srp_remove_work(void *target_ptr)
360 {
361         struct srp_target_port *target = target_ptr;
362
363         spin_lock_irq(target->scsi_host->host_lock);
364         if (target->state != SRP_TARGET_DEAD) {
365                 spin_unlock_irq(target->scsi_host->host_lock);
366                 return;
367         }
368         target->state = SRP_TARGET_REMOVED;
369         spin_unlock_irq(target->scsi_host->host_lock);
370
371         spin_lock(&target->srp_host->target_lock);
372         list_del(&target->list);
373         spin_unlock(&target->srp_host->target_lock);
374
375         scsi_remove_host(target->scsi_host);
376         ib_destroy_cm_id(target->cm_id);
377         srp_free_target_ib(target);
378         scsi_host_put(target->scsi_host);
379 }
380
381 static int srp_connect_target(struct srp_target_port *target)
382 {
383         int ret;
384
385         ret = srp_lookup_path(target);
386         if (ret)
387                 return ret;
388
389         while (1) {
390                 init_completion(&target->done);
391                 ret = srp_send_req(target);
392                 if (ret)
393                         return ret;
394                 wait_for_completion(&target->done);
395
396                 /*
397                  * The CM event handling code will set status to
398                  * SRP_PORT_REDIRECT if we get a port redirect REJ
399                  * back, or SRP_DLID_REDIRECT if we get a lid/qp
400                  * redirect REJ back.
401                  */
402                 switch (target->status) {
403                 case 0:
404                         return 0;
405
406                 case SRP_PORT_REDIRECT:
407                         ret = srp_lookup_path(target);
408                         if (ret)
409                                 return ret;
410                         break;
411
412                 case SRP_DLID_REDIRECT:
413                         break;
414
415                 default:
416                         return target->status;
417                 }
418         }
419 }
420
421 static void srp_unmap_data(struct scsi_cmnd *scmnd,
422                            struct srp_target_port *target,
423                            struct srp_request *req)
424 {
425         struct scatterlist *scat;
426         int nents;
427
428         if (!scmnd->request_buffer ||
429             (scmnd->sc_data_direction != DMA_TO_DEVICE &&
430              scmnd->sc_data_direction != DMA_FROM_DEVICE))
431                 return;
432
433         if (req->fmr) {
434                 ib_fmr_pool_unmap(req->fmr);
435                 req->fmr = NULL;
436         }
437
438         /*
439          * This handling of non-SG commands can be killed when the
440          * SCSI midlayer no longer generates non-SG commands.
441          */
442         if (likely(scmnd->use_sg)) {
443                 nents = scmnd->use_sg;
444                 scat  = scmnd->request_buffer;
445         } else {
446                 nents = 1;
447                 scat  = &req->fake_sg;
448         }
449
450         dma_unmap_sg(target->srp_host->dev->dev->dma_device, scat, nents,
451                      scmnd->sc_data_direction);
452 }
453
454 static int srp_reconnect_target(struct srp_target_port *target)
455 {
456         struct ib_cm_id *new_cm_id;
457         struct ib_qp_attr qp_attr;
458         struct srp_request *req;
459         struct ib_wc wc;
460         int ret;
461         int i;
462
463         spin_lock_irq(target->scsi_host->host_lock);
464         if (target->state != SRP_TARGET_LIVE) {
465                 spin_unlock_irq(target->scsi_host->host_lock);
466                 return -EAGAIN;
467         }
468         target->state = SRP_TARGET_CONNECTING;
469         spin_unlock_irq(target->scsi_host->host_lock);
470
471         srp_disconnect_target(target);
472         /*
473          * Now get a new local CM ID so that we avoid confusing the
474          * target in case things are really fouled up.
475          */
476         new_cm_id = ib_create_cm_id(target->srp_host->dev->dev,
477                                     srp_cm_handler, target);
478         if (IS_ERR(new_cm_id)) {
479                 ret = PTR_ERR(new_cm_id);
480                 goto err;
481         }
482         ib_destroy_cm_id(target->cm_id);
483         target->cm_id = new_cm_id;
484
485         qp_attr.qp_state = IB_QPS_RESET;
486         ret = ib_modify_qp(target->qp, &qp_attr, IB_QP_STATE);
487         if (ret)
488                 goto err;
489
490         ret = srp_init_qp(target, target->qp);
491         if (ret)
492                 goto err;
493
494         while (ib_poll_cq(target->cq, 1, &wc) > 0)
495                 ; /* nothing */
496
497         list_for_each_entry(req, &target->req_queue, list) {
498                 req->scmnd->result = DID_RESET << 16;
499                 req->scmnd->scsi_done(req->scmnd);
500                 srp_unmap_data(req->scmnd, target, req);
501         }
502
503         target->rx_head  = 0;
504         target->tx_head  = 0;
505         target->tx_tail  = 0;
506         INIT_LIST_HEAD(&target->free_reqs);
507         INIT_LIST_HEAD(&target->req_queue);
508         for (i = 0; i < SRP_SQ_SIZE; ++i)
509                 list_add_tail(&target->req_ring[i].list, &target->free_reqs);
510
511         ret = srp_connect_target(target);
512         if (ret)
513                 goto err;
514
515         spin_lock_irq(target->scsi_host->host_lock);
516         if (target->state == SRP_TARGET_CONNECTING) {
517                 ret = 0;
518                 target->state = SRP_TARGET_LIVE;
519         } else
520                 ret = -EAGAIN;
521         spin_unlock_irq(target->scsi_host->host_lock);
522
523         return ret;
524
525 err:
526         printk(KERN_ERR PFX "reconnect failed (%d), removing target port.\n", ret);
527
528         /*
529          * We couldn't reconnect, so kill our target port off.
530          * However, we have to defer the real removal because we might
531          * be in the context of the SCSI error handler now, which
532          * would deadlock if we call scsi_remove_host().
533          */
534         spin_lock_irq(target->scsi_host->host_lock);
535         if (target->state == SRP_TARGET_CONNECTING) {
536                 target->state = SRP_TARGET_DEAD;
537                 INIT_WORK(&target->work, srp_remove_work, target);
538                 schedule_work(&target->work);
539         }
540         spin_unlock_irq(target->scsi_host->host_lock);
541
542         return ret;
543 }
544
545 static int srp_map_fmr(struct srp_device *dev, struct scatterlist *scat,
546                        int sg_cnt, struct srp_request *req,
547                        struct srp_direct_buf *buf)
548 {
549         u64 io_addr = 0;
550         u64 *dma_pages;
551         u32 len;
552         int page_cnt;
553         int i, j;
554         int ret;
555
556         if (!dev->fmr_pool)
557                 return -ENODEV;
558
559         len = page_cnt = 0;
560         for (i = 0; i < sg_cnt; ++i) {
561                 if (sg_dma_address(&scat[i]) & ~dev->fmr_page_mask) {
562                         if (i > 0)
563                                 return -EINVAL;
564                         else
565                                 ++page_cnt;
566                 }
567                 if ((sg_dma_address(&scat[i]) + sg_dma_len(&scat[i])) &
568                     ~dev->fmr_page_mask) {
569                         if (i < sg_cnt - 1)
570                                 return -EINVAL;
571                         else
572                                 ++page_cnt;
573                 }
574
575                 len += sg_dma_len(&scat[i]);
576         }
577
578         page_cnt += len >> dev->fmr_page_shift;
579         if (page_cnt > SRP_FMR_SIZE)
580                 return -ENOMEM;
581
582         dma_pages = kmalloc(sizeof (u64) * page_cnt, GFP_ATOMIC);
583         if (!dma_pages)
584                 return -ENOMEM;
585
586         page_cnt = 0;
587         for (i = 0; i < sg_cnt; ++i)
588                 for (j = 0; j < sg_dma_len(&scat[i]); j += dev->fmr_page_size)
589                         dma_pages[page_cnt++] =
590                                 (sg_dma_address(&scat[i]) & dev->fmr_page_mask) + j;
591
592         req->fmr = ib_fmr_pool_map_phys(dev->fmr_pool,
593                                         dma_pages, page_cnt, &io_addr);
594         if (IS_ERR(req->fmr)) {
595                 ret = PTR_ERR(req->fmr);
596                 goto out;
597         }
598
599         buf->va  = cpu_to_be64(sg_dma_address(&scat[0]) & ~dev->fmr_page_mask);
600         buf->key = cpu_to_be32(req->fmr->fmr->rkey);
601         buf->len = cpu_to_be32(len);
602
603         ret = 0;
604
605 out:
606         kfree(dma_pages);
607
608         return ret;
609 }
610
611 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_target_port *target,
612                         struct srp_request *req)
613 {
614         struct scatterlist *scat;
615         struct srp_cmd *cmd = req->cmd->buf;
616         int len, nents, count;
617         u8 fmt = SRP_DATA_DESC_DIRECT;
618
619         if (!scmnd->request_buffer || scmnd->sc_data_direction == DMA_NONE)
620                 return sizeof (struct srp_cmd);
621
622         if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
623             scmnd->sc_data_direction != DMA_TO_DEVICE) {
624                 printk(KERN_WARNING PFX "Unhandled data direction %d\n",
625                        scmnd->sc_data_direction);
626                 return -EINVAL;
627         }
628
629         /*
630          * This handling of non-SG commands can be killed when the
631          * SCSI midlayer no longer generates non-SG commands.
632          */
633         if (likely(scmnd->use_sg)) {
634                 nents = scmnd->use_sg;
635                 scat  = scmnd->request_buffer;
636         } else {
637                 nents = 1;
638                 scat  = &req->fake_sg;
639                 sg_init_one(scat, scmnd->request_buffer, scmnd->request_bufflen);
640         }
641
642         count = dma_map_sg(target->srp_host->dev->dev->dma_device,
643                            scat, nents, scmnd->sc_data_direction);
644
645         fmt = SRP_DATA_DESC_DIRECT;
646         len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
647
648         if (count == 1) {
649                 /*
650                  * The midlayer only generated a single gather/scatter
651                  * entry, or DMA mapping coalesced everything to a
652                  * single entry.  So a direct descriptor along with
653                  * the DMA MR suffices.
654                  */
655                 struct srp_direct_buf *buf = (void *) cmd->add_data;
656
657                 buf->va  = cpu_to_be64(sg_dma_address(scat));
658                 buf->key = cpu_to_be32(target->srp_host->dev->mr->rkey);
659                 buf->len = cpu_to_be32(sg_dma_len(scat));
660         } else if (srp_map_fmr(target->srp_host->dev, scat, count, req,
661                                (void *) cmd->add_data)) {
662                 /*
663                  * FMR mapping failed, and the scatterlist has more
664                  * than one entry.  Generate an indirect memory
665                  * descriptor.
666                  */
667                 struct srp_indirect_buf *buf = (void *) cmd->add_data;
668                 u32 datalen = 0;
669                 int i;
670
671                 fmt = SRP_DATA_DESC_INDIRECT;
672                 len = sizeof (struct srp_cmd) +
673                         sizeof (struct srp_indirect_buf) +
674                         count * sizeof (struct srp_direct_buf);
675
676                 for (i = 0; i < count; ++i) {
677                         buf->desc_list[i].va  =
678                                 cpu_to_be64(sg_dma_address(&scat[i]));
679                         buf->desc_list[i].key =
680                                 cpu_to_be32(target->srp_host->dev->mr->rkey);
681                         buf->desc_list[i].len =
682                                 cpu_to_be32(sg_dma_len(&scat[i]));
683                         datalen += sg_dma_len(&scat[i]);
684                 }
685
686                 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
687                         cmd->data_out_desc_cnt = count;
688                 else
689                         cmd->data_in_desc_cnt = count;
690
691                 buf->table_desc.va  =
692                         cpu_to_be64(req->cmd->dma + sizeof *cmd + sizeof *buf);
693                 buf->table_desc.key =
694                         cpu_to_be32(target->srp_host->dev->mr->rkey);
695                 buf->table_desc.len =
696                         cpu_to_be32(count * sizeof (struct srp_direct_buf));
697
698                 buf->len = cpu_to_be32(datalen);
699         }
700
701         if (scmnd->sc_data_direction == DMA_TO_DEVICE)
702                 cmd->buf_fmt = fmt << 4;
703         else
704                 cmd->buf_fmt = fmt;
705
706         return len;
707 }
708
709 static void srp_remove_req(struct srp_target_port *target, struct srp_request *req)
710 {
711         srp_unmap_data(req->scmnd, target, req);
712         list_move_tail(&req->list, &target->free_reqs);
713 }
714
715 static void srp_process_rsp(struct srp_target_port *target, struct srp_rsp *rsp)
716 {
717         struct srp_request *req;
718         struct scsi_cmnd *scmnd;
719         unsigned long flags;
720         s32 delta;
721
722         delta = (s32) be32_to_cpu(rsp->req_lim_delta);
723
724         spin_lock_irqsave(target->scsi_host->host_lock, flags);
725
726         target->req_lim += delta;
727
728         req = &target->req_ring[rsp->tag & ~SRP_TAG_TSK_MGMT];
729
730         if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
731                 if (be32_to_cpu(rsp->resp_data_len) < 4)
732                         req->tsk_status = -1;
733                 else
734                         req->tsk_status = rsp->data[3];
735                 complete(&req->done);
736         } else {
737                 scmnd = req->scmnd;
738                 if (!scmnd)
739                         printk(KERN_ERR "Null scmnd for RSP w/tag %016llx\n",
740                                (unsigned long long) rsp->tag);
741                 scmnd->result = rsp->status;
742
743                 if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
744                         memcpy(scmnd->sense_buffer, rsp->data +
745                                be32_to_cpu(rsp->resp_data_len),
746                                min_t(int, be32_to_cpu(rsp->sense_data_len),
747                                      SCSI_SENSE_BUFFERSIZE));
748                 }
749
750                 if (rsp->flags & (SRP_RSP_FLAG_DOOVER | SRP_RSP_FLAG_DOUNDER))
751                         scmnd->resid = be32_to_cpu(rsp->data_out_res_cnt);
752                 else if (rsp->flags & (SRP_RSP_FLAG_DIOVER | SRP_RSP_FLAG_DIUNDER))
753                         scmnd->resid = be32_to_cpu(rsp->data_in_res_cnt);
754
755                 if (!req->tsk_mgmt) {
756                         scmnd->host_scribble = (void *) -1L;
757                         scmnd->scsi_done(scmnd);
758
759                         srp_remove_req(target, req);
760                 } else
761                         req->cmd_done = 1;
762         }
763
764         spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
765 }
766
767 static void srp_reconnect_work(void *target_ptr)
768 {
769         struct srp_target_port *target = target_ptr;
770
771         srp_reconnect_target(target);
772 }
773
774 static void srp_handle_recv(struct srp_target_port *target, struct ib_wc *wc)
775 {
776         struct srp_iu *iu;
777         u8 opcode;
778
779         iu = target->rx_ring[wc->wr_id & ~SRP_OP_RECV];
780
781         dma_sync_single_for_cpu(target->srp_host->dev->dev->dma_device, iu->dma,
782                                 target->max_ti_iu_len, DMA_FROM_DEVICE);
783
784         opcode = *(u8 *) iu->buf;
785
786         if (0) {
787                 int i;
788
789                 printk(KERN_ERR PFX "recv completion, opcode 0x%02x\n", opcode);
790
791                 for (i = 0; i < wc->byte_len; ++i) {
792                         if (i % 8 == 0)
793                                 printk(KERN_ERR "  [%02x] ", i);
794                         printk(" %02x", ((u8 *) iu->buf)[i]);
795                         if ((i + 1) % 8 == 0)
796                                 printk("\n");
797                 }
798
799                 if (wc->byte_len % 8)
800                         printk("\n");
801         }
802
803         switch (opcode) {
804         case SRP_RSP:
805                 srp_process_rsp(target, iu->buf);
806                 break;
807
808         case SRP_T_LOGOUT:
809                 /* XXX Handle target logout */
810                 printk(KERN_WARNING PFX "Got target logout request\n");
811                 break;
812
813         default:
814                 printk(KERN_WARNING PFX "Unhandled SRP opcode 0x%02x\n", opcode);
815                 break;
816         }
817
818         dma_sync_single_for_device(target->srp_host->dev->dev->dma_device, iu->dma,
819                                    target->max_ti_iu_len, DMA_FROM_DEVICE);
820 }
821
822 static void srp_completion(struct ib_cq *cq, void *target_ptr)
823 {
824         struct srp_target_port *target = target_ptr;
825         struct ib_wc wc;
826         unsigned long flags;
827
828         ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
829         while (ib_poll_cq(cq, 1, &wc) > 0) {
830                 if (wc.status) {
831                         printk(KERN_ERR PFX "failed %s status %d\n",
832                                wc.wr_id & SRP_OP_RECV ? "receive" : "send",
833                                wc.status);
834                         spin_lock_irqsave(target->scsi_host->host_lock, flags);
835                         if (target->state == SRP_TARGET_LIVE)
836                                 schedule_work(&target->work);
837                         spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
838                         break;
839                 }
840
841                 if (wc.wr_id & SRP_OP_RECV)
842                         srp_handle_recv(target, &wc);
843                 else
844                         ++target->tx_tail;
845         }
846 }
847
848 static int __srp_post_recv(struct srp_target_port *target)
849 {
850         struct srp_iu *iu;
851         struct ib_sge list;
852         struct ib_recv_wr wr, *bad_wr;
853         unsigned int next;
854         int ret;
855
856         next     = target->rx_head & (SRP_RQ_SIZE - 1);
857         wr.wr_id = next | SRP_OP_RECV;
858         iu       = target->rx_ring[next];
859
860         list.addr   = iu->dma;
861         list.length = iu->size;
862         list.lkey   = target->srp_host->dev->mr->lkey;
863
864         wr.next     = NULL;
865         wr.sg_list  = &list;
866         wr.num_sge  = 1;
867
868         ret = ib_post_recv(target->qp, &wr, &bad_wr);
869         if (!ret)
870                 ++target->rx_head;
871
872         return ret;
873 }
874
875 static int srp_post_recv(struct srp_target_port *target)
876 {
877         unsigned long flags;
878         int ret;
879
880         spin_lock_irqsave(target->scsi_host->host_lock, flags);
881         ret = __srp_post_recv(target);
882         spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
883
884         return ret;
885 }
886
887 /*
888  * Must be called with target->scsi_host->host_lock held to protect
889  * req_lim and tx_head.  Lock cannot be dropped between call here and
890  * call to __srp_post_send().
891  */
892 static struct srp_iu *__srp_get_tx_iu(struct srp_target_port *target)
893 {
894         if (target->tx_head - target->tx_tail >= SRP_SQ_SIZE)
895                 return NULL;
896
897         if (unlikely(target->req_lim < 1)) {
898                 if (printk_ratelimit())
899                         printk(KERN_DEBUG PFX "Target has req_lim %d\n",
900                                target->req_lim);
901                 return NULL;
902         }
903
904         return target->tx_ring[target->tx_head & SRP_SQ_SIZE];
905 }
906
907 /*
908  * Must be called with target->scsi_host->host_lock held to protect
909  * req_lim and tx_head.
910  */
911 static int __srp_post_send(struct srp_target_port *target,
912                            struct srp_iu *iu, int len)
913 {
914         struct ib_sge list;
915         struct ib_send_wr wr, *bad_wr;
916         int ret = 0;
917
918         list.addr   = iu->dma;
919         list.length = len;
920         list.lkey   = target->srp_host->dev->mr->lkey;
921
922         wr.next       = NULL;
923         wr.wr_id      = target->tx_head & SRP_SQ_SIZE;
924         wr.sg_list    = &list;
925         wr.num_sge    = 1;
926         wr.opcode     = IB_WR_SEND;
927         wr.send_flags = IB_SEND_SIGNALED;
928
929         ret = ib_post_send(target->qp, &wr, &bad_wr);
930
931         if (!ret) {
932                 ++target->tx_head;
933                 --target->req_lim;
934         }
935
936         return ret;
937 }
938
939 static int srp_queuecommand(struct scsi_cmnd *scmnd,
940                             void (*done)(struct scsi_cmnd *))
941 {
942         struct srp_target_port *target = host_to_target(scmnd->device->host);
943         struct srp_request *req;
944         struct srp_iu *iu;
945         struct srp_cmd *cmd;
946         int len;
947
948         if (target->state == SRP_TARGET_CONNECTING)
949                 goto err;
950
951         if (target->state == SRP_TARGET_DEAD ||
952             target->state == SRP_TARGET_REMOVED) {
953                 scmnd->result = DID_BAD_TARGET << 16;
954                 done(scmnd);
955                 return 0;
956         }
957
958         iu = __srp_get_tx_iu(target);
959         if (!iu)
960                 goto err;
961
962         dma_sync_single_for_cpu(target->srp_host->dev->dev->dma_device, iu->dma,
963                                 srp_max_iu_len, DMA_TO_DEVICE);
964
965         req = list_entry(target->free_reqs.next, struct srp_request, list);
966
967         scmnd->scsi_done     = done;
968         scmnd->result        = 0;
969         scmnd->host_scribble = (void *) (long) req->index;
970
971         cmd = iu->buf;
972         memset(cmd, 0, sizeof *cmd);
973
974         cmd->opcode = SRP_CMD;
975         cmd->lun    = cpu_to_be64((u64) scmnd->device->lun << 48);
976         cmd->tag    = req->index;
977         memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
978
979         req->scmnd    = scmnd;
980         req->cmd      = iu;
981         req->cmd_done = 0;
982         req->tsk_mgmt = NULL;
983
984         len = srp_map_data(scmnd, target, req);
985         if (len < 0) {
986                 printk(KERN_ERR PFX "Failed to map data\n");
987                 goto err;
988         }
989
990         if (__srp_post_recv(target)) {
991                 printk(KERN_ERR PFX "Recv failed\n");
992                 goto err_unmap;
993         }
994
995         dma_sync_single_for_device(target->srp_host->dev->dev->dma_device, iu->dma,
996                                    srp_max_iu_len, DMA_TO_DEVICE);
997
998         if (__srp_post_send(target, iu, len)) {
999                 printk(KERN_ERR PFX "Send failed\n");
1000                 goto err_unmap;
1001         }
1002
1003         list_move_tail(&req->list, &target->req_queue);
1004
1005         return 0;
1006
1007 err_unmap:
1008         srp_unmap_data(scmnd, target, req);
1009
1010 err:
1011         return SCSI_MLQUEUE_HOST_BUSY;
1012 }
1013
1014 static int srp_alloc_iu_bufs(struct srp_target_port *target)
1015 {
1016         int i;
1017
1018         for (i = 0; i < SRP_RQ_SIZE; ++i) {
1019                 target->rx_ring[i] = srp_alloc_iu(target->srp_host,
1020                                                   target->max_ti_iu_len,
1021                                                   GFP_KERNEL, DMA_FROM_DEVICE);
1022                 if (!target->rx_ring[i])
1023                         goto err;
1024         }
1025
1026         for (i = 0; i < SRP_SQ_SIZE + 1; ++i) {
1027                 target->tx_ring[i] = srp_alloc_iu(target->srp_host,
1028                                                   srp_max_iu_len,
1029                                                   GFP_KERNEL, DMA_TO_DEVICE);
1030                 if (!target->tx_ring[i])
1031                         goto err;
1032         }
1033
1034         return 0;
1035
1036 err:
1037         for (i = 0; i < SRP_RQ_SIZE; ++i) {
1038                 srp_free_iu(target->srp_host, target->rx_ring[i]);
1039                 target->rx_ring[i] = NULL;
1040         }
1041
1042         for (i = 0; i < SRP_SQ_SIZE + 1; ++i) {
1043                 srp_free_iu(target->srp_host, target->tx_ring[i]);
1044                 target->tx_ring[i] = NULL;
1045         }
1046
1047         return -ENOMEM;
1048 }
1049
1050 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
1051                                struct ib_cm_event *event,
1052                                struct srp_target_port *target)
1053 {
1054         struct ib_class_port_info *cpi;
1055         int opcode;
1056
1057         switch (event->param.rej_rcvd.reason) {
1058         case IB_CM_REJ_PORT_CM_REDIRECT:
1059                 cpi = event->param.rej_rcvd.ari;
1060                 target->path.dlid = cpi->redirect_lid;
1061                 target->path.pkey = cpi->redirect_pkey;
1062                 cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
1063                 memcpy(target->path.dgid.raw, cpi->redirect_gid, 16);
1064
1065                 target->status = target->path.dlid ?
1066                         SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
1067                 break;
1068
1069         case IB_CM_REJ_PORT_REDIRECT:
1070                 if (topspin_workarounds &&
1071                     !memcmp(&target->ioc_guid, topspin_oui, 3)) {
1072                         /*
1073                          * Topspin/Cisco SRP gateways incorrectly send
1074                          * reject reason code 25 when they mean 24
1075                          * (port redirect).
1076                          */
1077                         memcpy(target->path.dgid.raw,
1078                                event->param.rej_rcvd.ari, 16);
1079
1080                         printk(KERN_DEBUG PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
1081                                (unsigned long long) be64_to_cpu(target->path.dgid.global.subnet_prefix),
1082                                (unsigned long long) be64_to_cpu(target->path.dgid.global.interface_id));
1083
1084                         target->status = SRP_PORT_REDIRECT;
1085                 } else {
1086                         printk(KERN_WARNING "  REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
1087                         target->status = -ECONNRESET;
1088                 }
1089                 break;
1090
1091         case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
1092                 printk(KERN_WARNING "  REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
1093                 target->status = -ECONNRESET;
1094                 break;
1095
1096         case IB_CM_REJ_CONSUMER_DEFINED:
1097                 opcode = *(u8 *) event->private_data;
1098                 if (opcode == SRP_LOGIN_REJ) {
1099                         struct srp_login_rej *rej = event->private_data;
1100                         u32 reason = be32_to_cpu(rej->reason);
1101
1102                         if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
1103                                 printk(KERN_WARNING PFX
1104                                        "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
1105                         else
1106                                 printk(KERN_WARNING PFX
1107                                        "SRP LOGIN REJECTED, reason 0x%08x\n", reason);
1108                 } else
1109                         printk(KERN_WARNING "  REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
1110                                " opcode 0x%02x\n", opcode);
1111                 target->status = -ECONNRESET;
1112                 break;
1113
1114         default:
1115                 printk(KERN_WARNING "  REJ reason 0x%x\n",
1116                        event->param.rej_rcvd.reason);
1117                 target->status = -ECONNRESET;
1118         }
1119 }
1120
1121 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
1122 {
1123         struct srp_target_port *target = cm_id->context;
1124         struct ib_qp_attr *qp_attr = NULL;
1125         int attr_mask = 0;
1126         int comp = 0;
1127         int opcode = 0;
1128
1129         switch (event->event) {
1130         case IB_CM_REQ_ERROR:
1131                 printk(KERN_DEBUG PFX "Sending CM REQ failed\n");
1132                 comp = 1;
1133                 target->status = -ECONNRESET;
1134                 break;
1135
1136         case IB_CM_REP_RECEIVED:
1137                 comp = 1;
1138                 opcode = *(u8 *) event->private_data;
1139
1140                 if (opcode == SRP_LOGIN_RSP) {
1141                         struct srp_login_rsp *rsp = event->private_data;
1142
1143                         target->max_ti_iu_len = be32_to_cpu(rsp->max_ti_iu_len);
1144                         target->req_lim       = be32_to_cpu(rsp->req_lim_delta);
1145
1146                         target->scsi_host->can_queue = min(target->req_lim,
1147                                                            target->scsi_host->can_queue);
1148                 } else {
1149                         printk(KERN_WARNING PFX "Unhandled RSP opcode %#x\n", opcode);
1150                         target->status = -ECONNRESET;
1151                         break;
1152                 }
1153
1154                 target->status = srp_alloc_iu_bufs(target);
1155                 if (target->status)
1156                         break;
1157
1158                 qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
1159                 if (!qp_attr) {
1160                         target->status = -ENOMEM;
1161                         break;
1162                 }
1163
1164                 qp_attr->qp_state = IB_QPS_RTR;
1165                 target->status = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1166                 if (target->status)
1167                         break;
1168
1169                 target->status = ib_modify_qp(target->qp, qp_attr, attr_mask);
1170                 if (target->status)
1171                         break;
1172
1173                 target->status = srp_post_recv(target);
1174                 if (target->status)
1175                         break;
1176
1177                 qp_attr->qp_state = IB_QPS_RTS;
1178                 target->status = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1179                 if (target->status)
1180                         break;
1181
1182                 target->status = ib_modify_qp(target->qp, qp_attr, attr_mask);
1183                 if (target->status)
1184                         break;
1185
1186                 target->status = ib_send_cm_rtu(cm_id, NULL, 0);
1187                 if (target->status)
1188                         break;
1189
1190                 break;
1191
1192         case IB_CM_REJ_RECEIVED:
1193                 printk(KERN_DEBUG PFX "REJ received\n");
1194                 comp = 1;
1195
1196                 srp_cm_rej_handler(cm_id, event, target);
1197                 break;
1198
1199         case IB_CM_MRA_RECEIVED:
1200                 printk(KERN_ERR PFX "MRA received\n");
1201                 break;
1202
1203         case IB_CM_DREP_RECEIVED:
1204                 break;
1205
1206         case IB_CM_TIMEWAIT_EXIT:
1207                 printk(KERN_ERR PFX "connection closed\n");
1208
1209                 comp = 1;
1210                 target->status = 0;
1211                 break;
1212
1213         default:
1214                 printk(KERN_WARNING PFX "Unhandled CM event %d\n", event->event);
1215                 break;
1216         }
1217
1218         if (comp)
1219                 complete(&target->done);
1220
1221         kfree(qp_attr);
1222
1223         return 0;
1224 }
1225
1226 static int srp_send_tsk_mgmt(struct srp_target_port *target,
1227                              struct srp_request *req, u8 func)
1228 {
1229         struct srp_iu *iu;
1230         struct srp_tsk_mgmt *tsk_mgmt;
1231
1232         spin_lock_irq(target->scsi_host->host_lock);
1233
1234         if (target->state == SRP_TARGET_DEAD ||
1235             target->state == SRP_TARGET_REMOVED) {
1236                 req->scmnd->result = DID_BAD_TARGET << 16;
1237                 goto out;
1238         }
1239
1240         init_completion(&req->done);
1241
1242         iu = __srp_get_tx_iu(target);
1243         if (!iu)
1244                 goto out;
1245
1246         tsk_mgmt = iu->buf;
1247         memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
1248
1249         tsk_mgmt->opcode        = SRP_TSK_MGMT;
1250         tsk_mgmt->lun           = cpu_to_be64((u64) req->scmnd->device->lun << 48);
1251         tsk_mgmt->tag           = req->index | SRP_TAG_TSK_MGMT;
1252         tsk_mgmt->tsk_mgmt_func = func;
1253         tsk_mgmt->task_tag      = req->index;
1254
1255         if (__srp_post_send(target, iu, sizeof *tsk_mgmt))
1256                 goto out;
1257
1258         req->tsk_mgmt = iu;
1259
1260         spin_unlock_irq(target->scsi_host->host_lock);
1261
1262         if (!wait_for_completion_timeout(&req->done,
1263                                          msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
1264                 return -1;
1265
1266         return 0;
1267
1268 out:
1269         spin_unlock_irq(target->scsi_host->host_lock);
1270         return -1;
1271 }
1272
1273 static int srp_find_req(struct srp_target_port *target,
1274                         struct scsi_cmnd *scmnd,
1275                         struct srp_request **req)
1276 {
1277         if (scmnd->host_scribble == (void *) -1L)
1278                 return -1;
1279
1280         *req = &target->req_ring[(long) scmnd->host_scribble];
1281
1282         return 0;
1283 }
1284
1285 static int srp_abort(struct scsi_cmnd *scmnd)
1286 {
1287         struct srp_target_port *target = host_to_target(scmnd->device->host);
1288         struct srp_request *req;
1289         int ret = SUCCESS;
1290
1291         printk(KERN_ERR "SRP abort called\n");
1292
1293         if (srp_find_req(target, scmnd, &req))
1294                 return FAILED;
1295         if (srp_send_tsk_mgmt(target, req, SRP_TSK_ABORT_TASK))
1296                 return FAILED;
1297
1298         spin_lock_irq(target->scsi_host->host_lock);
1299
1300         if (req->cmd_done) {
1301                 srp_remove_req(target, req);
1302                 scmnd->scsi_done(scmnd);
1303         } else if (!req->tsk_status) {
1304                 srp_remove_req(target, req);
1305                 scmnd->result = DID_ABORT << 16;
1306         } else
1307                 ret = FAILED;
1308
1309         spin_unlock_irq(target->scsi_host->host_lock);
1310
1311         return ret;
1312 }
1313
1314 static int srp_reset_device(struct scsi_cmnd *scmnd)
1315 {
1316         struct srp_target_port *target = host_to_target(scmnd->device->host);
1317         struct srp_request *req, *tmp;
1318
1319         printk(KERN_ERR "SRP reset_device called\n");
1320
1321         if (srp_find_req(target, scmnd, &req))
1322                 return FAILED;
1323         if (srp_send_tsk_mgmt(target, req, SRP_TSK_LUN_RESET))
1324                 return FAILED;
1325         if (req->tsk_status)
1326                 return FAILED;
1327
1328         spin_lock_irq(target->scsi_host->host_lock);
1329
1330         list_for_each_entry_safe(req, tmp, &target->req_queue, list)
1331                 if (req->scmnd->device == scmnd->device) {
1332                         req->scmnd->result = DID_RESET << 16;
1333                         req->scmnd->scsi_done(req->scmnd);
1334                         srp_remove_req(target, req);
1335                 }
1336
1337         spin_unlock_irq(target->scsi_host->host_lock);
1338
1339         return SUCCESS;
1340 }
1341
1342 static int srp_reset_host(struct scsi_cmnd *scmnd)
1343 {
1344         struct srp_target_port *target = host_to_target(scmnd->device->host);
1345         int ret = FAILED;
1346
1347         printk(KERN_ERR PFX "SRP reset_host called\n");
1348
1349         if (!srp_reconnect_target(target))
1350                 ret = SUCCESS;
1351
1352         return ret;
1353 }
1354
1355 static ssize_t show_id_ext(struct class_device *cdev, char *buf)
1356 {
1357         struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1358
1359         if (target->state == SRP_TARGET_DEAD ||
1360             target->state == SRP_TARGET_REMOVED)
1361                 return -ENODEV;
1362
1363         return sprintf(buf, "0x%016llx\n",
1364                        (unsigned long long) be64_to_cpu(target->id_ext));
1365 }
1366
1367 static ssize_t show_ioc_guid(struct class_device *cdev, char *buf)
1368 {
1369         struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1370
1371         if (target->state == SRP_TARGET_DEAD ||
1372             target->state == SRP_TARGET_REMOVED)
1373                 return -ENODEV;
1374
1375         return sprintf(buf, "0x%016llx\n",
1376                        (unsigned long long) be64_to_cpu(target->ioc_guid));
1377 }
1378
1379 static ssize_t show_service_id(struct class_device *cdev, char *buf)
1380 {
1381         struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1382
1383         if (target->state == SRP_TARGET_DEAD ||
1384             target->state == SRP_TARGET_REMOVED)
1385                 return -ENODEV;
1386
1387         return sprintf(buf, "0x%016llx\n",
1388                        (unsigned long long) be64_to_cpu(target->service_id));
1389 }
1390
1391 static ssize_t show_pkey(struct class_device *cdev, char *buf)
1392 {
1393         struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1394
1395         if (target->state == SRP_TARGET_DEAD ||
1396             target->state == SRP_TARGET_REMOVED)
1397                 return -ENODEV;
1398
1399         return sprintf(buf, "0x%04x\n", be16_to_cpu(target->path.pkey));
1400 }
1401
1402 static ssize_t show_dgid(struct class_device *cdev, char *buf)
1403 {
1404         struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1405
1406         if (target->state == SRP_TARGET_DEAD ||
1407             target->state == SRP_TARGET_REMOVED)
1408                 return -ENODEV;
1409
1410         return sprintf(buf, "%04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x\n",
1411                        be16_to_cpu(((__be16 *) target->path.dgid.raw)[0]),
1412                        be16_to_cpu(((__be16 *) target->path.dgid.raw)[1]),
1413                        be16_to_cpu(((__be16 *) target->path.dgid.raw)[2]),
1414                        be16_to_cpu(((__be16 *) target->path.dgid.raw)[3]),
1415                        be16_to_cpu(((__be16 *) target->path.dgid.raw)[4]),
1416                        be16_to_cpu(((__be16 *) target->path.dgid.raw)[5]),
1417                        be16_to_cpu(((__be16 *) target->path.dgid.raw)[6]),
1418                        be16_to_cpu(((__be16 *) target->path.dgid.raw)[7]));
1419 }
1420
1421 static CLASS_DEVICE_ATTR(id_ext,        S_IRUGO, show_id_ext,           NULL);
1422 static CLASS_DEVICE_ATTR(ioc_guid,      S_IRUGO, show_ioc_guid,         NULL);
1423 static CLASS_DEVICE_ATTR(service_id,    S_IRUGO, show_service_id,       NULL);
1424 static CLASS_DEVICE_ATTR(pkey,          S_IRUGO, show_pkey,             NULL);
1425 static CLASS_DEVICE_ATTR(dgid,          S_IRUGO, show_dgid,             NULL);
1426
1427 static struct class_device_attribute *srp_host_attrs[] = {
1428         &class_device_attr_id_ext,
1429         &class_device_attr_ioc_guid,
1430         &class_device_attr_service_id,
1431         &class_device_attr_pkey,
1432         &class_device_attr_dgid,
1433         NULL
1434 };
1435
1436 static struct scsi_host_template srp_template = {
1437         .module                         = THIS_MODULE,
1438         .name                           = DRV_NAME,
1439         .info                           = srp_target_info,
1440         .queuecommand                   = srp_queuecommand,
1441         .eh_abort_handler               = srp_abort,
1442         .eh_device_reset_handler        = srp_reset_device,
1443         .eh_host_reset_handler          = srp_reset_host,
1444         .can_queue                      = SRP_SQ_SIZE,
1445         .this_id                        = -1,
1446         .cmd_per_lun                    = SRP_SQ_SIZE,
1447         .use_clustering                 = ENABLE_CLUSTERING,
1448         .shost_attrs                    = srp_host_attrs
1449 };
1450
1451 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
1452 {
1453         sprintf(target->target_name, "SRP.T10:%016llX",
1454                  (unsigned long long) be64_to_cpu(target->id_ext));
1455
1456         if (scsi_add_host(target->scsi_host, host->dev->dev->dma_device))
1457                 return -ENODEV;
1458
1459         spin_lock(&host->target_lock);
1460         list_add_tail(&target->list, &host->target_list);
1461         spin_unlock(&host->target_lock);
1462
1463         target->state = SRP_TARGET_LIVE;
1464
1465         scsi_scan_target(&target->scsi_host->shost_gendev,
1466                          0, target->scsi_id, SCAN_WILD_CARD, 0);
1467
1468         return 0;
1469 }
1470
1471 static void srp_release_class_dev(struct class_device *class_dev)
1472 {
1473         struct srp_host *host =
1474                 container_of(class_dev, struct srp_host, class_dev);
1475
1476         complete(&host->released);
1477 }
1478
1479 static struct class srp_class = {
1480         .name    = "infiniband_srp",
1481         .release = srp_release_class_dev
1482 };
1483
1484 /*
1485  * Target ports are added by writing
1486  *
1487  *     id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
1488  *     pkey=<P_Key>,service_id=<service ID>
1489  *
1490  * to the add_target sysfs attribute.
1491  */
1492 enum {
1493         SRP_OPT_ERR             = 0,
1494         SRP_OPT_ID_EXT          = 1 << 0,
1495         SRP_OPT_IOC_GUID        = 1 << 1,
1496         SRP_OPT_DGID            = 1 << 2,
1497         SRP_OPT_PKEY            = 1 << 3,
1498         SRP_OPT_SERVICE_ID      = 1 << 4,
1499         SRP_OPT_MAX_SECT        = 1 << 5,
1500         SRP_OPT_MAX_CMD_PER_LUN = 1 << 6,
1501         SRP_OPT_ALL             = (SRP_OPT_ID_EXT       |
1502                                    SRP_OPT_IOC_GUID     |
1503                                    SRP_OPT_DGID         |
1504                                    SRP_OPT_PKEY         |
1505                                    SRP_OPT_SERVICE_ID),
1506 };
1507
1508 static match_table_t srp_opt_tokens = {
1509         { SRP_OPT_ID_EXT,               "id_ext=%s"             },
1510         { SRP_OPT_IOC_GUID,             "ioc_guid=%s"           },
1511         { SRP_OPT_DGID,                 "dgid=%s"               },
1512         { SRP_OPT_PKEY,                 "pkey=%x"               },
1513         { SRP_OPT_SERVICE_ID,           "service_id=%s"         },
1514         { SRP_OPT_MAX_SECT,             "max_sect=%d"           },
1515         { SRP_OPT_MAX_CMD_PER_LUN,      "max_cmd_per_lun=%d"    },
1516         { SRP_OPT_ERR,                  NULL                    }
1517 };
1518
1519 static int srp_parse_options(const char *buf, struct srp_target_port *target)
1520 {
1521         char *options, *sep_opt;
1522         char *p;
1523         char dgid[3];
1524         substring_t args[MAX_OPT_ARGS];
1525         int opt_mask = 0;
1526         int token;
1527         int ret = -EINVAL;
1528         int i;
1529
1530         options = kstrdup(buf, GFP_KERNEL);
1531         if (!options)
1532                 return -ENOMEM;
1533
1534         sep_opt = options;
1535         while ((p = strsep(&sep_opt, ",")) != NULL) {
1536                 if (!*p)
1537                         continue;
1538
1539                 token = match_token(p, srp_opt_tokens, args);
1540                 opt_mask |= token;
1541
1542                 switch (token) {
1543                 case SRP_OPT_ID_EXT:
1544                         p = match_strdup(args);
1545                         target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
1546                         kfree(p);
1547                         break;
1548
1549                 case SRP_OPT_IOC_GUID:
1550                         p = match_strdup(args);
1551                         target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
1552                         kfree(p);
1553                         break;
1554
1555                 case SRP_OPT_DGID:
1556                         p = match_strdup(args);
1557                         if (strlen(p) != 32) {
1558                                 printk(KERN_WARNING PFX "bad dest GID parameter '%s'\n", p);
1559                                 kfree(p);
1560                                 goto out;
1561                         }
1562
1563                         for (i = 0; i < 16; ++i) {
1564                                 strlcpy(dgid, p + i * 2, 3);
1565                                 target->path.dgid.raw[i] = simple_strtoul(dgid, NULL, 16);
1566                         }
1567                         kfree(p);
1568                         break;
1569
1570                 case SRP_OPT_PKEY:
1571                         if (match_hex(args, &token)) {
1572                                 printk(KERN_WARNING PFX "bad P_Key parameter '%s'\n", p);
1573                                 goto out;
1574                         }
1575                         target->path.pkey = cpu_to_be16(token);
1576                         break;
1577
1578                 case SRP_OPT_SERVICE_ID:
1579                         p = match_strdup(args);
1580                         target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
1581                         kfree(p);
1582                         break;
1583
1584                 case SRP_OPT_MAX_SECT:
1585                         if (match_int(args, &token)) {
1586                                 printk(KERN_WARNING PFX "bad max sect parameter '%s'\n", p);
1587                                 goto out;
1588                         }
1589                         target->scsi_host->max_sectors = token;
1590                         break;
1591
1592                 case SRP_OPT_MAX_CMD_PER_LUN:
1593                         if (match_int(args, &token)) {
1594                                 printk(KERN_WARNING PFX "bad max cmd_per_lun parameter '%s'\n", p);
1595                                 goto out;
1596                         }
1597                         target->scsi_host->cmd_per_lun = min(token, SRP_SQ_SIZE);
1598                         break;
1599
1600                 default:
1601                         printk(KERN_WARNING PFX "unknown parameter or missing value "
1602                                "'%s' in target creation request\n", p);
1603                         goto out;
1604                 }
1605         }
1606
1607         if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
1608                 ret = 0;
1609         else
1610                 for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
1611                         if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
1612                             !(srp_opt_tokens[i].token & opt_mask))
1613                                 printk(KERN_WARNING PFX "target creation request is "
1614                                        "missing parameter '%s'\n",
1615                                        srp_opt_tokens[i].pattern);
1616
1617 out:
1618         kfree(options);
1619         return ret;
1620 }
1621
1622 static ssize_t srp_create_target(struct class_device *class_dev,
1623                                  const char *buf, size_t count)
1624 {
1625         struct srp_host *host =
1626                 container_of(class_dev, struct srp_host, class_dev);
1627         struct Scsi_Host *target_host;
1628         struct srp_target_port *target;
1629         int ret;
1630         int i;
1631
1632         target_host = scsi_host_alloc(&srp_template,
1633                                       sizeof (struct srp_target_port));
1634         if (!target_host)
1635                 return -ENOMEM;
1636
1637         target_host->max_lun = SRP_MAX_LUN;
1638
1639         target = host_to_target(target_host);
1640         memset(target, 0, sizeof *target);
1641
1642         target->scsi_host  = target_host;
1643         target->srp_host   = host;
1644
1645         INIT_WORK(&target->work, srp_reconnect_work, target);
1646
1647         INIT_LIST_HEAD(&target->free_reqs);
1648         INIT_LIST_HEAD(&target->req_queue);
1649         for (i = 0; i < SRP_SQ_SIZE; ++i) {
1650                 target->req_ring[i].index = i;
1651                 list_add_tail(&target->req_ring[i].list, &target->free_reqs);
1652         }
1653
1654         ret = srp_parse_options(buf, target);
1655         if (ret)
1656                 goto err;
1657
1658         ib_get_cached_gid(host->dev->dev, host->port, 0, &target->path.sgid);
1659
1660         printk(KERN_DEBUG PFX "new target: id_ext %016llx ioc_guid %016llx pkey %04x "
1661                "service_id %016llx dgid %04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x\n",
1662                (unsigned long long) be64_to_cpu(target->id_ext),
1663                (unsigned long long) be64_to_cpu(target->ioc_guid),
1664                be16_to_cpu(target->path.pkey),
1665                (unsigned long long) be64_to_cpu(target->service_id),
1666                (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[0]),
1667                (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[2]),
1668                (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[4]),
1669                (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[6]),
1670                (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[8]),
1671                (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[10]),
1672                (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[12]),
1673                (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[14]));
1674
1675         ret = srp_create_target_ib(target);
1676         if (ret)
1677                 goto err;
1678
1679         target->cm_id = ib_create_cm_id(host->dev->dev, srp_cm_handler, target);
1680         if (IS_ERR(target->cm_id)) {
1681                 ret = PTR_ERR(target->cm_id);
1682                 goto err_free;
1683         }
1684
1685         ret = srp_connect_target(target);
1686         if (ret) {
1687                 printk(KERN_ERR PFX "Connection failed\n");
1688                 goto err_cm_id;
1689         }
1690
1691         ret = srp_add_target(host, target);
1692         if (ret)
1693                 goto err_disconnect;
1694
1695         return count;
1696
1697 err_disconnect:
1698         srp_disconnect_target(target);
1699
1700 err_cm_id:
1701         ib_destroy_cm_id(target->cm_id);
1702
1703 err_free:
1704         srp_free_target_ib(target);
1705
1706 err:
1707         scsi_host_put(target_host);
1708
1709         return ret;
1710 }
1711
1712 static CLASS_DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
1713
1714 static ssize_t show_ibdev(struct class_device *class_dev, char *buf)
1715 {
1716         struct srp_host *host =
1717                 container_of(class_dev, struct srp_host, class_dev);
1718
1719         return sprintf(buf, "%s\n", host->dev->dev->name);
1720 }
1721
1722 static CLASS_DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
1723
1724 static ssize_t show_port(struct class_device *class_dev, char *buf)
1725 {
1726         struct srp_host *host =
1727                 container_of(class_dev, struct srp_host, class_dev);
1728
1729         return sprintf(buf, "%d\n", host->port);
1730 }
1731
1732 static CLASS_DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
1733
1734 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
1735 {
1736         struct srp_host *host;
1737
1738         host = kzalloc(sizeof *host, GFP_KERNEL);
1739         if (!host)
1740                 return NULL;
1741
1742         INIT_LIST_HEAD(&host->target_list);
1743         spin_lock_init(&host->target_lock);
1744         init_completion(&host->released);
1745         host->dev  = device;
1746         host->port = port;
1747
1748         host->initiator_port_id[7] = port;
1749         memcpy(host->initiator_port_id + 8, &device->dev->node_guid, 8);
1750
1751         host->class_dev.class = &srp_class;
1752         host->class_dev.dev   = device->dev->dma_device;
1753         snprintf(host->class_dev.class_id, BUS_ID_SIZE, "srp-%s-%d",
1754                  device->dev->name, port);
1755
1756         if (class_device_register(&host->class_dev))
1757                 goto free_host;
1758         if (class_device_create_file(&host->class_dev, &class_device_attr_add_target))
1759                 goto err_class;
1760         if (class_device_create_file(&host->class_dev, &class_device_attr_ibdev))
1761                 goto err_class;
1762         if (class_device_create_file(&host->class_dev, &class_device_attr_port))
1763                 goto err_class;
1764
1765         return host;
1766
1767 err_class:
1768         class_device_unregister(&host->class_dev);
1769
1770 free_host:
1771         kfree(host);
1772
1773         return NULL;
1774 }
1775
1776 static void srp_add_one(struct ib_device *device)
1777 {
1778         struct srp_device *srp_dev;
1779         struct ib_device_attr *dev_attr;
1780         struct ib_fmr_pool_param fmr_param;
1781         struct srp_host *host;
1782         int s, e, p;
1783
1784         dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
1785         if (!dev_attr)
1786                 return;
1787
1788         if (ib_query_device(device, dev_attr)) {
1789                 printk(KERN_WARNING PFX "Query device failed for %s\n",
1790                        device->name);
1791                 goto free_attr;
1792         }
1793
1794         srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
1795         if (!srp_dev)
1796                 goto free_attr;
1797
1798         /*
1799          * Use the smallest page size supported by the HCA, down to a
1800          * minimum of 512 bytes (which is the smallest sector that a
1801          * SCSI command will ever carry).
1802          */
1803         srp_dev->fmr_page_shift = max(9, ffs(dev_attr->page_size_cap) - 1);
1804         srp_dev->fmr_page_size  = 1 << srp_dev->fmr_page_shift;
1805         srp_dev->fmr_page_mask  = ~((unsigned long) srp_dev->fmr_page_size - 1);
1806
1807         INIT_LIST_HEAD(&srp_dev->dev_list);
1808
1809         srp_dev->dev = device;
1810         srp_dev->pd  = ib_alloc_pd(device);
1811         if (IS_ERR(srp_dev->pd))
1812                 goto free_dev;
1813
1814         srp_dev->mr = ib_get_dma_mr(srp_dev->pd,
1815                                     IB_ACCESS_LOCAL_WRITE |
1816                                     IB_ACCESS_REMOTE_READ |
1817                                     IB_ACCESS_REMOTE_WRITE);
1818         if (IS_ERR(srp_dev->mr))
1819                 goto err_pd;
1820
1821         memset(&fmr_param, 0, sizeof fmr_param);
1822         fmr_param.pool_size         = SRP_FMR_POOL_SIZE;
1823         fmr_param.dirty_watermark   = SRP_FMR_DIRTY_SIZE;
1824         fmr_param.cache             = 1;
1825         fmr_param.max_pages_per_fmr = SRP_FMR_SIZE;
1826         fmr_param.page_shift        = srp_dev->fmr_page_shift;
1827         fmr_param.access            = (IB_ACCESS_LOCAL_WRITE |
1828                                        IB_ACCESS_REMOTE_WRITE |
1829                                        IB_ACCESS_REMOTE_READ);
1830
1831         srp_dev->fmr_pool = ib_create_fmr_pool(srp_dev->pd, &fmr_param);
1832         if (IS_ERR(srp_dev->fmr_pool))
1833                 srp_dev->fmr_pool = NULL;
1834
1835         if (device->node_type == IB_NODE_SWITCH) {
1836                 s = 0;
1837                 e = 0;
1838         } else {
1839                 s = 1;
1840                 e = device->phys_port_cnt;
1841         }
1842
1843         for (p = s; p <= e; ++p) {
1844                 host = srp_add_port(srp_dev, p);
1845                 if (host)
1846                         list_add_tail(&host->list, &srp_dev->dev_list);
1847         }
1848
1849         ib_set_client_data(device, &srp_client, srp_dev);
1850
1851         goto free_attr;
1852
1853 err_pd:
1854         ib_dealloc_pd(srp_dev->pd);
1855
1856 free_dev:
1857         kfree(srp_dev);
1858
1859 free_attr:
1860         kfree(dev_attr);
1861 }
1862
1863 static void srp_remove_one(struct ib_device *device)
1864 {
1865         struct srp_device *srp_dev;
1866         struct srp_host *host, *tmp_host;
1867         LIST_HEAD(target_list);
1868         struct srp_target_port *target, *tmp_target;
1869
1870         srp_dev = ib_get_client_data(device, &srp_client);
1871
1872         list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
1873                 class_device_unregister(&host->class_dev);
1874                 /*
1875                  * Wait for the sysfs entry to go away, so that no new
1876                  * target ports can be created.
1877                  */
1878                 wait_for_completion(&host->released);
1879
1880                 /*
1881                  * Mark all target ports as removed, so we stop queueing
1882                  * commands and don't try to reconnect.
1883                  */
1884                 spin_lock(&host->target_lock);
1885                 list_for_each_entry(target, &host->target_list, list) {
1886                         spin_lock_irq(target->scsi_host->host_lock);
1887                         target->state = SRP_TARGET_REMOVED;
1888                         spin_unlock_irq(target->scsi_host->host_lock);
1889                 }
1890                 spin_unlock(&host->target_lock);
1891
1892                 /*
1893                  * Wait for any reconnection tasks that may have
1894                  * started before we marked our target ports as
1895                  * removed, and any target port removal tasks.
1896                  */
1897                 flush_scheduled_work();
1898
1899                 list_for_each_entry_safe(target, tmp_target,
1900                                          &host->target_list, list) {
1901                         scsi_remove_host(target->scsi_host);
1902                         srp_disconnect_target(target);
1903                         ib_destroy_cm_id(target->cm_id);
1904                         srp_free_target_ib(target);
1905                         scsi_host_put(target->scsi_host);
1906                 }
1907
1908                 kfree(host);
1909         }
1910
1911         if (srp_dev->fmr_pool)
1912                 ib_destroy_fmr_pool(srp_dev->fmr_pool);
1913         ib_dereg_mr(srp_dev->mr);
1914         ib_dealloc_pd(srp_dev->pd);
1915
1916         kfree(srp_dev);
1917 }
1918
1919 static int __init srp_init_module(void)
1920 {
1921         int ret;
1922
1923         srp_template.sg_tablesize = srp_sg_tablesize;
1924         srp_max_iu_len = (sizeof (struct srp_cmd) +
1925                           sizeof (struct srp_indirect_buf) +
1926                           srp_sg_tablesize * 16);
1927
1928         ret = class_register(&srp_class);
1929         if (ret) {
1930                 printk(KERN_ERR PFX "couldn't register class infiniband_srp\n");
1931                 return ret;
1932         }
1933
1934         ret = ib_register_client(&srp_client);
1935         if (ret) {
1936                 printk(KERN_ERR PFX "couldn't register IB client\n");
1937                 class_unregister(&srp_class);
1938                 return ret;
1939         }
1940
1941         return 0;
1942 }
1943
1944 static void __exit srp_cleanup_module(void)
1945 {
1946         ib_unregister_client(&srp_client);
1947         class_unregister(&srp_class);
1948 }
1949
1950 module_init(srp_init_module);
1951 module_exit(srp_cleanup_module);