projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
IB/srp: Initialize dma_length in srp_map_idb
[deliverable/linux.git] / drivers / infiniband / ulp / srp / ib_srp.c
1 /*
2 * Copyright (c) 2005 Cisco Systems. All rights reserved.
3 *
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 * SOFTWARE.
31 */
32
33 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
34
35 #include <linux/module.h>
36 #include <linux/init.h>
37 #include <linux/slab.h>
38 #include <linux/err.h>
39 #include <linux/string.h>
40 #include <linux/parser.h>
41 #include <linux/random.h>
42 #include <linux/jiffies.h>
43 #include <rdma/ib_cache.h>
44
45 #include <linux/atomic.h>
46
47 #include <scsi/scsi.h>
48 #include <scsi/scsi_device.h>
49 #include <scsi/scsi_dbg.h>
50 #include <scsi/scsi_tcq.h>
51 #include <scsi/srp.h>
52 #include <scsi/scsi_transport_srp.h>
53
54 #include "ib_srp.h"
55
56 #define DRV_NAME "ib_srp"
57 #define PFX DRV_NAME ": "
58 #define DRV_VERSION "2.0"
59 #define DRV_RELDATE "July 26, 2015"
60
61 MODULE_AUTHOR("Roland Dreier");
62 MODULE_DESCRIPTION("InfiniBand SCSI RDMA Protocol initiator");
63 MODULE_LICENSE("Dual BSD/GPL");
64 MODULE_VERSION(DRV_VERSION);
65 MODULE_INFO(release_date, DRV_RELDATE);
66
67 static unsigned int srp_sg_tablesize;
68 static unsigned int cmd_sg_entries;
69 static unsigned int indirect_sg_entries;
70 static bool allow_ext_sg;
71 static bool prefer_fr = true;
72 static bool register_always = true;
73 static int topspin_workarounds = 1;
74
75 module_param(srp_sg_tablesize, uint, 0444);
76 MODULE_PARM_DESC(srp_sg_tablesize, "Deprecated name for cmd_sg_entries");
77
78 module_param(cmd_sg_entries, uint, 0444);
79 MODULE_PARM_DESC(cmd_sg_entries,
80 "Default number of gather/scatter entries in the SRP command (default is 12, max 255)");
81
82 module_param(indirect_sg_entries, uint, 0444);
83 MODULE_PARM_DESC(indirect_sg_entries,
84 "Default max number of gather/scatter entries (default is 12, max is " __stringify(SCSI_MAX_SG_CHAIN_SEGMENTS) ")");
85
86 module_param(allow_ext_sg, bool, 0444);
87 MODULE_PARM_DESC(allow_ext_sg,
88 "Default behavior when there are more than cmd_sg_entries S/G entries after mapping; fails the request when false (default false)");
89
90 module_param(topspin_workarounds, int, 0444);
91 MODULE_PARM_DESC(topspin_workarounds,
92 "Enable workarounds for Topspin/Cisco SRP target bugs if != 0");
93
94 module_param(prefer_fr, bool, 0444);
95 MODULE_PARM_DESC(prefer_fr,
96 "Whether to use fast registration if both FMR and fast registration are supported");
97
98 module_param(register_always, bool, 0444);
99 MODULE_PARM_DESC(register_always,
100 "Use memory registration even for contiguous memory regions");
101
102 static const struct kernel_param_ops srp_tmo_ops;
103
104 static int srp_reconnect_delay = 10;
105 module_param_cb(reconnect_delay, &srp_tmo_ops, &srp_reconnect_delay,
106 S_IRUGO | S_IWUSR);
107 MODULE_PARM_DESC(reconnect_delay, "Time between successive reconnect attempts");
108
109 static int srp_fast_io_fail_tmo = 15;
110 module_param_cb(fast_io_fail_tmo, &srp_tmo_ops, &srp_fast_io_fail_tmo,
111 S_IRUGO | S_IWUSR);
112 MODULE_PARM_DESC(fast_io_fail_tmo,
113 "Number of seconds between the observation of a transport"
114 " layer error and failing all I/O. \"off\" means that this"
115 " functionality is disabled.");
116
117 static int srp_dev_loss_tmo = 600;
118 module_param_cb(dev_loss_tmo, &srp_tmo_ops, &srp_dev_loss_tmo,
119 S_IRUGO | S_IWUSR);
120 MODULE_PARM_DESC(dev_loss_tmo,
121 "Maximum number of seconds that the SRP transport should"
122 " insulate transport layer errors. After this time has been"
123 " exceeded the SCSI host is removed. Should be"
124 " between 1 and " __stringify(SCSI_DEVICE_BLOCK_MAX_TIMEOUT)
125 " if fast_io_fail_tmo has not been set. \"off\" means that"
126 " this functionality is disabled.");
127
128 static unsigned ch_count;
129 module_param(ch_count, uint, 0444);
130 MODULE_PARM_DESC(ch_count,
131 "Number of RDMA channels to use for communication with an SRP target. Using more than one channel improves performance if the HCA supports multiple completion vectors. The default value is the minimum of four times the number of online CPU sockets and the number of completion vectors supported by the HCA.");
132
133 static void srp_add_one(struct ib_device *device);
134 static void srp_remove_one(struct ib_device *device, void *client_data);
135 static void srp_recv_completion(struct ib_cq *cq, void *ch_ptr);
136 static void srp_send_completion(struct ib_cq *cq, void *ch_ptr);
137 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event);
138
139 static struct scsi_transport_template *ib_srp_transport_template;
140 static struct workqueue_struct *srp_remove_wq;
141
142 static struct ib_client srp_client = {
143 .name = "srp",
144 .add = srp_add_one,
145 .remove = srp_remove_one
146 };
147
148 static struct ib_sa_client srp_sa_client;
149
150 static int srp_tmo_get(char *buffer, const struct kernel_param *kp)
151 {
152 int tmo = *(int *)kp->arg;
153
154 if (tmo >= 0)
155 return sprintf(buffer, "%d", tmo);
156 else
157 return sprintf(buffer, "off");
158 }
159
160 static int srp_tmo_set(const char *val, const struct kernel_param *kp)
161 {
162 int tmo, res;
163
164 res = srp_parse_tmo(&tmo, val);
165 if (res)
166 goto out;
167
168 if (kp->arg == &srp_reconnect_delay)
169 res = srp_tmo_valid(tmo, srp_fast_io_fail_tmo,
170 srp_dev_loss_tmo);
171 else if (kp->arg == &srp_fast_io_fail_tmo)
172 res = srp_tmo_valid(srp_reconnect_delay, tmo, srp_dev_loss_tmo);
173 else
174 res = srp_tmo_valid(srp_reconnect_delay, srp_fast_io_fail_tmo,
175 tmo);
176 if (res)
177 goto out;
178 *(int *)kp->arg = tmo;
179
180 out:
181 return res;
182 }
183
184 static const struct kernel_param_ops srp_tmo_ops = {
185 .get = srp_tmo_get,
186 .set = srp_tmo_set,
187 };
188
189 static inline struct srp_target_port *host_to_target(struct Scsi_Host *host)
190 {
191 return (struct srp_target_port *) host->hostdata;
192 }
193
194 static const char *srp_target_info(struct Scsi_Host *host)
195 {
196 return host_to_target(host)->target_name;
197 }
198
199 static int srp_target_is_topspin(struct srp_target_port *target)
200 {
201 static const u8 topspin_oui[3] = { 0x00, 0x05, 0xad };
202 static const u8 cisco_oui[3] = { 0x00, 0x1b, 0x0d };
203
204 return topspin_workarounds &&
205 (!memcmp(&target->ioc_guid, topspin_oui, sizeof topspin_oui) ||
206 !memcmp(&target->ioc_guid, cisco_oui, sizeof cisco_oui));
207 }
208
209 static struct srp_iu *srp_alloc_iu(struct srp_host *host, size_t size,
210 gfp_t gfp_mask,
211 enum dma_data_direction direction)
212 {
213 struct srp_iu *iu;
214
215 iu = kmalloc(sizeof *iu, gfp_mask);
216 if (!iu)
217 goto out;
218
219 iu->buf = kzalloc(size, gfp_mask);
220 if (!iu->buf)
221 goto out_free_iu;
222
223 iu->dma = ib_dma_map_single(host->srp_dev->dev, iu->buf, size,
224 direction);
225 if (ib_dma_mapping_error(host->srp_dev->dev, iu->dma))
226 goto out_free_buf;
227
228 iu->size = size;
229 iu->direction = direction;
230
231 return iu;
232
233 out_free_buf:
234 kfree(iu->buf);
235 out_free_iu:
236 kfree(iu);
237 out:
238 return NULL;
239 }
240
241 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
242 {
243 if (!iu)
244 return;
245
246 ib_dma_unmap_single(host->srp_dev->dev, iu->dma, iu->size,
247 iu->direction);
248 kfree(iu->buf);
249 kfree(iu);
250 }
251
252 static void srp_qp_event(struct ib_event *event, void *context)
253 {
254 pr_debug("QP event %s (%d)\n",
255 ib_event_msg(event->event), event->event);
256 }
257
258 static int srp_init_qp(struct srp_target_port *target,
259 struct ib_qp *qp)
260 {
261 struct ib_qp_attr *attr;
262 int ret;
263
264 attr = kmalloc(sizeof *attr, GFP_KERNEL);
265 if (!attr)
266 return -ENOMEM;
267
268 ret = ib_find_cached_pkey(target->srp_host->srp_dev->dev,
269 target->srp_host->port,
270 be16_to_cpu(target->pkey),
271 &attr->pkey_index);
272 if (ret)
273 goto out;
274
275 attr->qp_state = IB_QPS_INIT;
276 attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
277 IB_ACCESS_REMOTE_WRITE);
278 attr->port_num = target->srp_host->port;
279
280 ret = ib_modify_qp(qp, attr,
281 IB_QP_STATE |
282 IB_QP_PKEY_INDEX |
283 IB_QP_ACCESS_FLAGS |
284 IB_QP_PORT);
285
286 out:
287 kfree(attr);
288 return ret;
289 }
290
291 static int srp_new_cm_id(struct srp_rdma_ch *ch)
292 {
293 struct srp_target_port *target = ch->target;
294 struct ib_cm_id *new_cm_id;
295
296 new_cm_id = ib_create_cm_id(target->srp_host->srp_dev->dev,
297 srp_cm_handler, ch);
298 if (IS_ERR(new_cm_id))
299 return PTR_ERR(new_cm_id);
300
301 if (ch->cm_id)
302 ib_destroy_cm_id(ch->cm_id);
303 ch->cm_id = new_cm_id;
304 ch->path.sgid = target->sgid;
305 ch->path.dgid = target->orig_dgid;
306 ch->path.pkey = target->pkey;
307 ch->path.service_id = target->service_id;
308
309 return 0;
310 }
311
312 static struct ib_fmr_pool *srp_alloc_fmr_pool(struct srp_target_port *target)
313 {
314 struct srp_device *dev = target->srp_host->srp_dev;
315 struct ib_fmr_pool_param fmr_param;
316
317 memset(&fmr_param, 0, sizeof(fmr_param));
318 fmr_param.pool_size = target->scsi_host->can_queue;
319 fmr_param.dirty_watermark = fmr_param.pool_size / 4;
320 fmr_param.cache = 1;
321 fmr_param.max_pages_per_fmr = dev->max_pages_per_mr;
322 fmr_param.page_shift = ilog2(dev->mr_page_size);
323 fmr_param.access = (IB_ACCESS_LOCAL_WRITE |
324 IB_ACCESS_REMOTE_WRITE |
325 IB_ACCESS_REMOTE_READ);
326
327 return ib_create_fmr_pool(dev->pd, &fmr_param);
328 }
329
330 /**
331 * srp_destroy_fr_pool() - free the resources owned by a pool
332 * @pool: Fast registration pool to be destroyed.
333 */
334 static void srp_destroy_fr_pool(struct srp_fr_pool *pool)
335 {
336 int i;
337 struct srp_fr_desc *d;
338
339 if (!pool)
340 return;
341
342 for (i = 0, d = &pool->desc[0]; i < pool->size; i++, d++) {
343 if (d->mr)
344 ib_dereg_mr(d->mr);
345 }
346 kfree(pool);
347 }
348
349 /**
350 * srp_create_fr_pool() - allocate and initialize a pool for fast registration
351 * @device: IB device to allocate fast registration descriptors for.
352 * @pd: Protection domain associated with the FR descriptors.
353 * @pool_size: Number of descriptors to allocate.
354 * @max_page_list_len: Maximum fast registration work request page list length.
355 */
356 static struct srp_fr_pool *srp_create_fr_pool(struct ib_device *device,
357 struct ib_pd *pd, int pool_size,
358 int max_page_list_len)
359 {
360 struct srp_fr_pool *pool;
361 struct srp_fr_desc *d;
362 struct ib_mr *mr;
363 int i, ret = -EINVAL;
364
365 if (pool_size <= 0)
366 goto err;
367 ret = -ENOMEM;
368 pool = kzalloc(sizeof(struct srp_fr_pool) +
369 pool_size * sizeof(struct srp_fr_desc), GFP_KERNEL);
370 if (!pool)
371 goto err;
372 pool->size = pool_size;
373 pool->max_page_list_len = max_page_list_len;
374 spin_lock_init(&pool->lock);
375 INIT_LIST_HEAD(&pool->free_list);
376
377 for (i = 0, d = &pool->desc[0]; i < pool->size; i++, d++) {
378 mr = ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG,
379 max_page_list_len);
380 if (IS_ERR(mr)) {
381 ret = PTR_ERR(mr);
382 goto destroy_pool;
383 }
384 d->mr = mr;
385 list_add_tail(&d->entry, &pool->free_list);
386 }
387
388 out:
389 return pool;
390
391 destroy_pool:
392 srp_destroy_fr_pool(pool);
393
394 err:
395 pool = ERR_PTR(ret);
396 goto out;
397 }
398
399 /**
400 * srp_fr_pool_get() - obtain a descriptor suitable for fast registration
401 * @pool: Pool to obtain descriptor from.
402 */
403 static struct srp_fr_desc *srp_fr_pool_get(struct srp_fr_pool *pool)
404 {
405 struct srp_fr_desc *d = NULL;
406 unsigned long flags;
407
408 spin_lock_irqsave(&pool->lock, flags);
409 if (!list_empty(&pool->free_list)) {
410 d = list_first_entry(&pool->free_list, typeof(*d), entry);
411 list_del(&d->entry);
412 }
413 spin_unlock_irqrestore(&pool->lock, flags);
414
415 return d;
416 }
417
418 /**
419 * srp_fr_pool_put() - put an FR descriptor back in the free list
420 * @pool: Pool the descriptor was allocated from.
421 * @desc: Pointer to an array of fast registration descriptor pointers.
422 * @n: Number of descriptors to put back.
423 *
424 * Note: The caller must already have queued an invalidation request for
425 * desc->mr->rkey before calling this function.
426 */
427 static void srp_fr_pool_put(struct srp_fr_pool *pool, struct srp_fr_desc **desc,
428 int n)
429 {
430 unsigned long flags;
431 int i;
432
433 spin_lock_irqsave(&pool->lock, flags);
434 for (i = 0; i < n; i++)
435 list_add(&desc[i]->entry, &pool->free_list);
436 spin_unlock_irqrestore(&pool->lock, flags);
437 }
438
439 static struct srp_fr_pool *srp_alloc_fr_pool(struct srp_target_port *target)
440 {
441 struct srp_device *dev = target->srp_host->srp_dev;
442
443 return srp_create_fr_pool(dev->dev, dev->pd,
444 target->scsi_host->can_queue,
445 dev->max_pages_per_mr);
446 }
447
448 /**
449 * srp_destroy_qp() - destroy an RDMA queue pair
450 * @ch: SRP RDMA channel.
451 *
452 * Change a queue pair into the error state and wait until all receive
453 * completions have been processed before destroying it. This avoids that
454 * the receive completion handler can access the queue pair while it is
455 * being destroyed.
456 */
457 static void srp_destroy_qp(struct srp_rdma_ch *ch)
458 {
459 static struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
460 static struct ib_recv_wr wr = { .wr_id = SRP_LAST_WR_ID };
461 struct ib_recv_wr *bad_wr;
462 int ret;
463
464 /* Destroying a QP and reusing ch->done is only safe if not connected */
465 WARN_ON_ONCE(ch->connected);
466
467 ret = ib_modify_qp(ch->qp, &attr, IB_QP_STATE);
468 WARN_ONCE(ret, "ib_cm_init_qp_attr() returned %d\n", ret);
469 if (ret)
470 goto out;
471
472 init_completion(&ch->done);
473 ret = ib_post_recv(ch->qp, &wr, &bad_wr);
474 WARN_ONCE(ret, "ib_post_recv() returned %d\n", ret);
475 if (ret == 0)
476 wait_for_completion(&ch->done);
477
478 out:
479 ib_destroy_qp(ch->qp);
480 }
481
482 static int srp_create_ch_ib(struct srp_rdma_ch *ch)
483 {
484 struct srp_target_port *target = ch->target;
485 struct srp_device *dev = target->srp_host->srp_dev;
486 struct ib_qp_init_attr *init_attr;
487 struct ib_cq *recv_cq, *send_cq;
488 struct ib_qp *qp;
489 struct ib_fmr_pool *fmr_pool = NULL;
490 struct srp_fr_pool *fr_pool = NULL;
491 const int m = dev->use_fast_reg ? 3 : 1;
492 struct ib_cq_init_attr cq_attr = {};
493 int ret;
494
495 init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
496 if (!init_attr)
497 return -ENOMEM;
498
499 /* + 1 for SRP_LAST_WR_ID */
500 cq_attr.cqe = target->queue_size + 1;
501 cq_attr.comp_vector = ch->comp_vector;
502 recv_cq = ib_create_cq(dev->dev, srp_recv_completion, NULL, ch,
503 &cq_attr);
504 if (IS_ERR(recv_cq)) {
505 ret = PTR_ERR(recv_cq);
506 goto err;
507 }
508
509 cq_attr.cqe = m * target->queue_size;
510 cq_attr.comp_vector = ch->comp_vector;
511 send_cq = ib_create_cq(dev->dev, srp_send_completion, NULL, ch,
512 &cq_attr);
513 if (IS_ERR(send_cq)) {
514 ret = PTR_ERR(send_cq);
515 goto err_recv_cq;
516 }
517
518 ib_req_notify_cq(recv_cq, IB_CQ_NEXT_COMP);
519
520 init_attr->event_handler = srp_qp_event;
521 init_attr->cap.max_send_wr = m * target->queue_size;
522 init_attr->cap.max_recv_wr = target->queue_size + 1;
523 init_attr->cap.max_recv_sge = 1;
524 init_attr->cap.max_send_sge = 1;
525 init_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
526 init_attr->qp_type = IB_QPT_RC;
527 init_attr->send_cq = send_cq;
528 init_attr->recv_cq = recv_cq;
529
530 qp = ib_create_qp(dev->pd, init_attr);
531 if (IS_ERR(qp)) {
532 ret = PTR_ERR(qp);
533 goto err_send_cq;
534 }
535
536 ret = srp_init_qp(target, qp);
537 if (ret)
538 goto err_qp;
539
540 if (dev->use_fast_reg) {
541 fr_pool = srp_alloc_fr_pool(target);
542 if (IS_ERR(fr_pool)) {
543 ret = PTR_ERR(fr_pool);
544 shost_printk(KERN_WARNING, target->scsi_host, PFX
545 "FR pool allocation failed (%d)\n", ret);
546 goto err_qp;
547 }
548 } else if (dev->use_fmr) {
549 fmr_pool = srp_alloc_fmr_pool(target);
550 if (IS_ERR(fmr_pool)) {
551 ret = PTR_ERR(fmr_pool);
552 shost_printk(KERN_WARNING, target->scsi_host, PFX
553 "FMR pool allocation failed (%d)\n", ret);
554 goto err_qp;
555 }
556 }
557
558 if (ch->qp)
559 srp_destroy_qp(ch);
560 if (ch->recv_cq)
561 ib_destroy_cq(ch->recv_cq);
562 if (ch->send_cq)
563 ib_destroy_cq(ch->send_cq);
564
565 ch->qp = qp;
566 ch->recv_cq = recv_cq;
567 ch->send_cq = send_cq;
568
569 if (dev->use_fast_reg) {
570 if (ch->fr_pool)
571 srp_destroy_fr_pool(ch->fr_pool);
572 ch->fr_pool = fr_pool;
573 } else if (dev->use_fmr) {
574 if (ch->fmr_pool)
575 ib_destroy_fmr_pool(ch->fmr_pool);
576 ch->fmr_pool = fmr_pool;
577 }
578
579 kfree(init_attr);
580 return 0;
581
582 err_qp:
583 ib_destroy_qp(qp);
584
585 err_send_cq:
586 ib_destroy_cq(send_cq);
587
588 err_recv_cq:
589 ib_destroy_cq(recv_cq);
590
591 err:
592 kfree(init_attr);
593 return ret;
594 }
595
596 /*
597 * Note: this function may be called without srp_alloc_iu_bufs() having been
598 * invoked. Hence the ch->[rt]x_ring checks.
599 */
600 static void srp_free_ch_ib(struct srp_target_port *target,
601 struct srp_rdma_ch *ch)
602 {
603 struct srp_device *dev = target->srp_host->srp_dev;
604 int i;
605
606 if (!ch->target)
607 return;
608
609 if (ch->cm_id) {
610 ib_destroy_cm_id(ch->cm_id);
611 ch->cm_id = NULL;
612 }
613
614 /* If srp_new_cm_id() succeeded but srp_create_ch_ib() not, return. */
615 if (!ch->qp)
616 return;
617
618 if (dev->use_fast_reg) {
619 if (ch->fr_pool)
620 srp_destroy_fr_pool(ch->fr_pool);
621 } else if (dev->use_fmr) {
622 if (ch->fmr_pool)
623 ib_destroy_fmr_pool(ch->fmr_pool);
624 }
625 srp_destroy_qp(ch);
626 ib_destroy_cq(ch->send_cq);
627 ib_destroy_cq(ch->recv_cq);
628
629 /*
630 * Avoid that the SCSI error handler tries to use this channel after
631 * it has been freed. The SCSI error handler can namely continue
632 * trying to perform recovery actions after scsi_remove_host()
633 * returned.
634 */
635 ch->target = NULL;
636
637 ch->qp = NULL;
638 ch->send_cq = ch->recv_cq = NULL;
639
640 if (ch->rx_ring) {
641 for (i = 0; i < target->queue_size; ++i)
642 srp_free_iu(target->srp_host, ch->rx_ring[i]);
643 kfree(ch->rx_ring);
644 ch->rx_ring = NULL;
645 }
646 if (ch->tx_ring) {
647 for (i = 0; i < target->queue_size; ++i)
648 srp_free_iu(target->srp_host, ch->tx_ring[i]);
649 kfree(ch->tx_ring);
650 ch->tx_ring = NULL;
651 }
652 }
653
654 static void srp_path_rec_completion(int status,
655 struct ib_sa_path_rec *pathrec,
656 void *ch_ptr)
657 {
658 struct srp_rdma_ch *ch = ch_ptr;
659 struct srp_target_port *target = ch->target;
660
661 ch->status = status;
662 if (status)
663 shost_printk(KERN_ERR, target->scsi_host,
664 PFX "Got failed path rec status %d\n", status);
665 else
666 ch->path = *pathrec;
667 complete(&ch->done);
668 }
669
670 static int srp_lookup_path(struct srp_rdma_ch *ch)
671 {
672 struct srp_target_port *target = ch->target;
673 int ret;
674
675 ch->path.numb_path = 1;
676
677 init_completion(&ch->done);
678
679 ch->path_query_id = ib_sa_path_rec_get(&srp_sa_client,
680 target->srp_host->srp_dev->dev,
681 target->srp_host->port,
682 &ch->path,
683 IB_SA_PATH_REC_SERVICE_ID |
684 IB_SA_PATH_REC_DGID |
685 IB_SA_PATH_REC_SGID |
686 IB_SA_PATH_REC_NUMB_PATH |
687 IB_SA_PATH_REC_PKEY,
688 SRP_PATH_REC_TIMEOUT_MS,
689 GFP_KERNEL,
690 srp_path_rec_completion,
691 ch, &ch->path_query);
692 if (ch->path_query_id < 0)
693 return ch->path_query_id;
694
695 ret = wait_for_completion_interruptible(&ch->done);
696 if (ret < 0)
697 return ret;
698
699 if (ch->status < 0)
700 shost_printk(KERN_WARNING, target->scsi_host,
701 PFX "Path record query failed\n");
702
703 return ch->status;
704 }
705
706 static int srp_send_req(struct srp_rdma_ch *ch, bool multich)
707 {
708 struct srp_target_port *target = ch->target;
709 struct {
710 struct ib_cm_req_param param;
711 struct srp_login_req priv;
712 } *req = NULL;
713 int status;
714
715 req = kzalloc(sizeof *req, GFP_KERNEL);
716 if (!req)
717 return -ENOMEM;
718
719 req->param.primary_path = &ch->path;
720 req->param.alternate_path = NULL;
721 req->param.service_id = target->service_id;
722 req->param.qp_num = ch->qp->qp_num;
723 req->param.qp_type = ch->qp->qp_type;
724 req->param.private_data = &req->priv;
725 req->param.private_data_len = sizeof req->priv;
726 req->param.flow_control = 1;
727
728 get_random_bytes(&req->param.starting_psn, 4);
729 req->param.starting_psn &= 0xffffff;
730
731 /*
732 * Pick some arbitrary defaults here; we could make these
733 * module parameters if anyone cared about setting them.
734 */
735 req->param.responder_resources = 4;
736 req->param.remote_cm_response_timeout = 20;
737 req->param.local_cm_response_timeout = 20;
738 req->param.retry_count = target->tl_retry_count;
739 req->param.rnr_retry_count = 7;
740 req->param.max_cm_retries = 15;
741
742 req->priv.opcode = SRP_LOGIN_REQ;
743 req->priv.tag = 0;
744 req->priv.req_it_iu_len = cpu_to_be32(target->max_iu_len);
745 req->priv.req_buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
746 SRP_BUF_FORMAT_INDIRECT);
747 req->priv.req_flags = (multich ? SRP_MULTICHAN_MULTI :
748 SRP_MULTICHAN_SINGLE);
749 /*
750 * In the published SRP specification (draft rev. 16a), the
751 * port identifier format is 8 bytes of ID extension followed
752 * by 8 bytes of GUID. Older drafts put the two halves in the
753 * opposite order, so that the GUID comes first.
754 *
755 * Targets conforming to these obsolete drafts can be
756 * recognized by the I/O Class they report.
757 */
758 if (target->io_class == SRP_REV10_IB_IO_CLASS) {
759 memcpy(req->priv.initiator_port_id,
760 &target->sgid.global.interface_id, 8);
761 memcpy(req->priv.initiator_port_id + 8,
762 &target->initiator_ext, 8);
763 memcpy(req->priv.target_port_id, &target->ioc_guid, 8);
764 memcpy(req->priv.target_port_id + 8, &target->id_ext, 8);
765 } else {
766 memcpy(req->priv.initiator_port_id,
767 &target->initiator_ext, 8);
768 memcpy(req->priv.initiator_port_id + 8,
769 &target->sgid.global.interface_id, 8);
770 memcpy(req->priv.target_port_id, &target->id_ext, 8);
771 memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
772 }
773
774 /*
775 * Topspin/Cisco SRP targets will reject our login unless we
776 * zero out the first 8 bytes of our initiator port ID and set
777 * the second 8 bytes to the local node GUID.
778 */
779 if (srp_target_is_topspin(target)) {
780 shost_printk(KERN_DEBUG, target->scsi_host,
781 PFX "Topspin/Cisco initiator port ID workaround "
782 "activated for target GUID %016llx\n",
783 be64_to_cpu(target->ioc_guid));
784 memset(req->priv.initiator_port_id, 0, 8);
785 memcpy(req->priv.initiator_port_id + 8,
786 &target->srp_host->srp_dev->dev->node_guid, 8);
787 }
788
789 status = ib_send_cm_req(ch->cm_id, &req->param);
790
791 kfree(req);
792
793 return status;
794 }
795
796 static bool srp_queue_remove_work(struct srp_target_port *target)
797 {
798 bool changed = false;
799
800 spin_lock_irq(&target->lock);
801 if (target->state != SRP_TARGET_REMOVED) {
802 target->state = SRP_TARGET_REMOVED;
803 changed = true;
804 }
805 spin_unlock_irq(&target->lock);
806
807 if (changed)
808 queue_work(srp_remove_wq, &target->remove_work);
809
810 return changed;
811 }
812
813 static void srp_disconnect_target(struct srp_target_port *target)
814 {
815 struct srp_rdma_ch *ch;
816 int i;
817
818 /* XXX should send SRP_I_LOGOUT request */
819
820 for (i = 0; i < target->ch_count; i++) {
821 ch = &target->ch[i];
822 ch->connected = false;
823 if (ch->cm_id && ib_send_cm_dreq(ch->cm_id, NULL, 0)) {
824 shost_printk(KERN_DEBUG, target->scsi_host,
825 PFX "Sending CM DREQ failed\n");
826 }
827 }
828 }
829
830 static void srp_free_req_data(struct srp_target_port *target,
831 struct srp_rdma_ch *ch)
832 {
833 struct srp_device *dev = target->srp_host->srp_dev;
834 struct ib_device *ibdev = dev->dev;
835 struct srp_request *req;
836 int i;
837
838 if (!ch->req_ring)
839 return;
840
841 for (i = 0; i < target->req_ring_size; ++i) {
842 req = &ch->req_ring[i];
843 if (dev->use_fast_reg) {
844 kfree(req->fr_list);
845 } else {
846 kfree(req->fmr_list);
847 kfree(req->map_page);
848 }
849 if (req->indirect_dma_addr) {
850 ib_dma_unmap_single(ibdev, req->indirect_dma_addr,
851 target->indirect_size,
852 DMA_TO_DEVICE);
853 }
854 kfree(req->indirect_desc);
855 }
856
857 kfree(ch->req_ring);
858 ch->req_ring = NULL;
859 }
860
861 static int srp_alloc_req_data(struct srp_rdma_ch *ch)
862 {
863 struct srp_target_port *target = ch->target;
864 struct srp_device *srp_dev = target->srp_host->srp_dev;
865 struct ib_device *ibdev = srp_dev->dev;
866 struct srp_request *req;
867 void *mr_list;
868 dma_addr_t dma_addr;
869 int i, ret = -ENOMEM;
870
871 ch->req_ring = kcalloc(target->req_ring_size, sizeof(*ch->req_ring),
872 GFP_KERNEL);
873 if (!ch->req_ring)
874 goto out;
875
876 for (i = 0; i < target->req_ring_size; ++i) {
877 req = &ch->req_ring[i];
878 mr_list = kmalloc(target->cmd_sg_cnt * sizeof(void *),
879 GFP_KERNEL);
880 if (!mr_list)
881 goto out;
882 if (srp_dev->use_fast_reg) {
883 req->fr_list = mr_list;
884 } else {
885 req->fmr_list = mr_list;
886 req->map_page = kmalloc(srp_dev->max_pages_per_mr *
887 sizeof(void *), GFP_KERNEL);
888 if (!req->map_page)
889 goto out;
890 }
891 req->indirect_desc = kmalloc(target->indirect_size, GFP_KERNEL);
892 if (!req->indirect_desc)
893 goto out;
894
895 dma_addr = ib_dma_map_single(ibdev, req->indirect_desc,
896 target->indirect_size,
897 DMA_TO_DEVICE);
898 if (ib_dma_mapping_error(ibdev, dma_addr))
899 goto out;
900
901 req->indirect_dma_addr = dma_addr;
902 }
903 ret = 0;
904
905 out:
906 return ret;
907 }
908
909 /**
910 * srp_del_scsi_host_attr() - Remove attributes defined in the host template.
911 * @shost: SCSI host whose attributes to remove from sysfs.
912 *
913 * Note: Any attributes defined in the host template and that did not exist
914 * before invocation of this function will be ignored.
915 */
916 static void srp_del_scsi_host_attr(struct Scsi_Host *shost)
917 {
918 struct device_attribute **attr;
919
920 for (attr = shost->hostt->shost_attrs; attr && *attr; ++attr)
921 device_remove_file(&shost->shost_dev, *attr);
922 }
923
924 static void srp_remove_target(struct srp_target_port *target)
925 {
926 struct srp_rdma_ch *ch;
927 int i;
928
929 WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
930
931 srp_del_scsi_host_attr(target->scsi_host);
932 srp_rport_get(target->rport);
933 srp_remove_host(target->scsi_host);
934 scsi_remove_host(target->scsi_host);
935 srp_stop_rport_timers(target->rport);
936 srp_disconnect_target(target);
937 for (i = 0; i < target->ch_count; i++) {
938 ch = &target->ch[i];
939 srp_free_ch_ib(target, ch);
940 }
941 cancel_work_sync(&target->tl_err_work);
942 srp_rport_put(target->rport);
943 for (i = 0; i < target->ch_count; i++) {
944 ch = &target->ch[i];
945 srp_free_req_data(target, ch);
946 }
947 kfree(target->ch);
948 target->ch = NULL;
949
950 spin_lock(&target->srp_host->target_lock);
951 list_del(&target->list);
952 spin_unlock(&target->srp_host->target_lock);
953
954 scsi_host_put(target->scsi_host);
955 }
956
957 static void srp_remove_work(struct work_struct *work)
958 {
959 struct srp_target_port *target =
960 container_of(work, struct srp_target_port, remove_work);
961
962 WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
963
964 srp_remove_target(target);
965 }
966
967 static void srp_rport_delete(struct srp_rport *rport)
968 {
969 struct srp_target_port *target = rport->lld_data;
970
971 srp_queue_remove_work(target);
972 }
973
974 /**
975 * srp_connected_ch() - number of connected channels
976 * @target: SRP target port.
977 */
978 static int srp_connected_ch(struct srp_target_port *target)
979 {
980 int i, c = 0;
981
982 for (i = 0; i < target->ch_count; i++)
983 c += target->ch[i].connected;
984
985 return c;
986 }
987
988 static int srp_connect_ch(struct srp_rdma_ch *ch, bool multich)
989 {
990 struct srp_target_port *target = ch->target;
991 int ret;
992
993 WARN_ON_ONCE(!multich && srp_connected_ch(target) > 0);
994
995 ret = srp_lookup_path(ch);
996 if (ret)
997 goto out;
998
999 while (1) {
1000 init_completion(&ch->done);
1001 ret = srp_send_req(ch, multich);
1002 if (ret)
1003 goto out;
1004 ret = wait_for_completion_interruptible(&ch->done);
1005 if (ret < 0)
1006 goto out;
1007
1008 /*
1009 * The CM event handling code will set status to
1010 * SRP_PORT_REDIRECT if we get a port redirect REJ
1011 * back, or SRP_DLID_REDIRECT if we get a lid/qp
1012 * redirect REJ back.
1013 */
1014 ret = ch->status;
1015 switch (ret) {
1016 case 0:
1017 ch->connected = true;
1018 goto out;
1019
1020 case SRP_PORT_REDIRECT:
1021 ret = srp_lookup_path(ch);
1022 if (ret)
1023 goto out;
1024 break;
1025
1026 case SRP_DLID_REDIRECT:
1027 break;
1028
1029 case SRP_STALE_CONN:
1030 shost_printk(KERN_ERR, target->scsi_host, PFX
1031 "giving up on stale connection\n");
1032 ret = -ECONNRESET;
1033 goto out;
1034
1035 default:
1036 goto out;
1037 }
1038 }
1039
1040 out:
1041 return ret <= 0 ? ret : -ENODEV;
1042 }
1043
1044 static int srp_inv_rkey(struct srp_rdma_ch *ch, u32 rkey)
1045 {
1046 struct ib_send_wr *bad_wr;
1047 struct ib_send_wr wr = {
1048 .opcode = IB_WR_LOCAL_INV,
1049 .wr_id = LOCAL_INV_WR_ID_MASK,
1050 .next = NULL,
1051 .num_sge = 0,
1052 .send_flags = 0,
1053 .ex.invalidate_rkey = rkey,
1054 };
1055
1056 return ib_post_send(ch->qp, &wr, &bad_wr);
1057 }
1058
1059 static void srp_unmap_data(struct scsi_cmnd *scmnd,
1060 struct srp_rdma_ch *ch,
1061 struct srp_request *req)
1062 {
1063 struct srp_target_port *target = ch->target;
1064 struct srp_device *dev = target->srp_host->srp_dev;
1065 struct ib_device *ibdev = dev->dev;
1066 int i, res;
1067
1068 if (!scsi_sglist(scmnd) ||
1069 (scmnd->sc_data_direction != DMA_TO_DEVICE &&
1070 scmnd->sc_data_direction != DMA_FROM_DEVICE))
1071 return;
1072
1073 if (dev->use_fast_reg) {
1074 struct srp_fr_desc **pfr;
1075
1076 for (i = req->nmdesc, pfr = req->fr_list; i > 0; i--, pfr++) {
1077 res = srp_inv_rkey(ch, (*pfr)->mr->rkey);
1078 if (res < 0) {
1079 shost_printk(KERN_ERR, target->scsi_host, PFX
1080 "Queueing INV WR for rkey %#x failed (%d)\n",
1081 (*pfr)->mr->rkey, res);
1082 queue_work(system_long_wq,
1083 &target->tl_err_work);
1084 }
1085 }
1086 if (req->nmdesc)
1087 srp_fr_pool_put(ch->fr_pool, req->fr_list,
1088 req->nmdesc);
1089 } else if (dev->use_fmr) {
1090 struct ib_pool_fmr **pfmr;
1091
1092 for (i = req->nmdesc, pfmr = req->fmr_list; i > 0; i--, pfmr++)
1093 ib_fmr_pool_unmap(*pfmr);
1094 }
1095
1096 ib_dma_unmap_sg(ibdev, scsi_sglist(scmnd), scsi_sg_count(scmnd),
1097 scmnd->sc_data_direction);
1098 }
1099
1100 /**
1101 * srp_claim_req - Take ownership of the scmnd associated with a request.
1102 * @ch: SRP RDMA channel.
1103 * @req: SRP request.
1104 * @sdev: If not NULL, only take ownership for this SCSI device.
1105 * @scmnd: If NULL, take ownership of @req->scmnd. If not NULL, only take
1106 * ownership of @req->scmnd if it equals @scmnd.
1107 *
1108 * Return value:
1109 * Either NULL or a pointer to the SCSI command the caller became owner of.
1110 */
1111 static struct scsi_cmnd *srp_claim_req(struct srp_rdma_ch *ch,
1112 struct srp_request *req,
1113 struct scsi_device *sdev,
1114 struct scsi_cmnd *scmnd)
1115 {
1116 unsigned long flags;
1117
1118 spin_lock_irqsave(&ch->lock, flags);
1119 if (req->scmnd &&
1120 (!sdev || req->scmnd->device == sdev) &&
1121 (!scmnd || req->scmnd == scmnd)) {
1122 scmnd = req->scmnd;
1123 req->scmnd = NULL;
1124 } else {
1125 scmnd = NULL;
1126 }
1127 spin_unlock_irqrestore(&ch->lock, flags);
1128
1129 return scmnd;
1130 }
1131
1132 /**
1133 * srp_free_req() - Unmap data and add request to the free request list.
1134 * @ch: SRP RDMA channel.
1135 * @req: Request to be freed.
1136 * @scmnd: SCSI command associated with @req.
1137 * @req_lim_delta: Amount to be added to @target->req_lim.
1138 */
1139 static void srp_free_req(struct srp_rdma_ch *ch, struct srp_request *req,
1140 struct scsi_cmnd *scmnd, s32 req_lim_delta)
1141 {
1142 unsigned long flags;
1143
1144 srp_unmap_data(scmnd, ch, req);
1145
1146 spin_lock_irqsave(&ch->lock, flags);
1147 ch->req_lim += req_lim_delta;
1148 spin_unlock_irqrestore(&ch->lock, flags);
1149 }
1150
1151 static void srp_finish_req(struct srp_rdma_ch *ch, struct srp_request *req,
1152 struct scsi_device *sdev, int result)
1153 {
1154 struct scsi_cmnd *scmnd = srp_claim_req(ch, req, sdev, NULL);
1155
1156 if (scmnd) {
1157 srp_free_req(ch, req, scmnd, 0);
1158 scmnd->result = result;
1159 scmnd->scsi_done(scmnd);
1160 }
1161 }
1162
1163 static void srp_terminate_io(struct srp_rport *rport)
1164 {
1165 struct srp_target_port *target = rport->lld_data;
1166 struct srp_rdma_ch *ch;
1167 struct Scsi_Host *shost = target->scsi_host;
1168 struct scsi_device *sdev;
1169 int i, j;
1170
1171 /*
1172 * Invoking srp_terminate_io() while srp_queuecommand() is running
1173 * is not safe. Hence the warning statement below.
1174 */
1175 shost_for_each_device(sdev, shost)
1176 WARN_ON_ONCE(sdev->request_queue->request_fn_active);
1177
1178 for (i = 0; i < target->ch_count; i++) {
1179 ch = &target->ch[i];
1180
1181 for (j = 0; j < target->req_ring_size; ++j) {
1182 struct srp_request *req = &ch->req_ring[j];
1183
1184 srp_finish_req(ch, req, NULL,
1185 DID_TRANSPORT_FAILFAST << 16);
1186 }
1187 }
1188 }
1189
1190 /*
1191 * It is up to the caller to ensure that srp_rport_reconnect() calls are
1192 * serialized and that no concurrent srp_queuecommand(), srp_abort(),
1193 * srp_reset_device() or srp_reset_host() calls will occur while this function
1194 * is in progress. One way to realize that is not to call this function
1195 * directly but to call srp_reconnect_rport() instead since that last function
1196 * serializes calls of this function via rport->mutex and also blocks
1197 * srp_queuecommand() calls before invoking this function.
1198 */
1199 static int srp_rport_reconnect(struct srp_rport *rport)
1200 {
1201 struct srp_target_port *target = rport->lld_data;
1202 struct srp_rdma_ch *ch;
1203 int i, j, ret = 0;
1204 bool multich = false;
1205
1206 srp_disconnect_target(target);
1207
1208 if (target->state == SRP_TARGET_SCANNING)
1209 return -ENODEV;
1210
1211 /*
1212 * Now get a new local CM ID so that we avoid confusing the target in
1213 * case things are really fouled up. Doing so also ensures that all CM
1214 * callbacks will have finished before a new QP is allocated.
1215 */
1216 for (i = 0; i < target->ch_count; i++) {
1217 ch = &target->ch[i];
1218 ret += srp_new_cm_id(ch);
1219 }
1220 for (i = 0; i < target->ch_count; i++) {
1221 ch = &target->ch[i];
1222 for (j = 0; j < target->req_ring_size; ++j) {
1223 struct srp_request *req = &ch->req_ring[j];
1224
1225 srp_finish_req(ch, req, NULL, DID_RESET << 16);
1226 }
1227 }
1228 for (i = 0; i < target->ch_count; i++) {
1229 ch = &target->ch[i];
1230 /*
1231 * Whether or not creating a new CM ID succeeded, create a new
1232 * QP. This guarantees that all completion callback function
1233 * invocations have finished before request resetting starts.
1234 */
1235 ret += srp_create_ch_ib(ch);
1236
1237 INIT_LIST_HEAD(&ch->free_tx);
1238 for (j = 0; j < target->queue_size; ++j)
1239 list_add(&ch->tx_ring[j]->list, &ch->free_tx);
1240 }
1241
1242 target->qp_in_error = false;
1243
1244 for (i = 0; i < target->ch_count; i++) {
1245 ch = &target->ch[i];
1246 if (ret)
1247 break;
1248 ret = srp_connect_ch(ch, multich);
1249 multich = true;
1250 }
1251
1252 if (ret == 0)
1253 shost_printk(KERN_INFO, target->scsi_host,
1254 PFX "reconnect succeeded\n");
1255
1256 return ret;
1257 }
1258
1259 static void srp_map_desc(struct srp_map_state *state, dma_addr_t dma_addr,
1260 unsigned int dma_len, u32 rkey)
1261 {
1262 struct srp_direct_buf *desc = state->desc;
1263
1264 WARN_ON_ONCE(!dma_len);
1265
1266 desc->va = cpu_to_be64(dma_addr);
1267 desc->key = cpu_to_be32(rkey);
1268 desc->len = cpu_to_be32(dma_len);
1269
1270 state->total_len += dma_len;
1271 state->desc++;
1272 state->ndesc++;
1273 }
1274
1275 static int srp_map_finish_fmr(struct srp_map_state *state,
1276 struct srp_rdma_ch *ch)
1277 {
1278 struct srp_target_port *target = ch->target;
1279 struct srp_device *dev = target->srp_host->srp_dev;
1280 struct ib_pool_fmr *fmr;
1281 u64 io_addr = 0;
1282
1283 if (state->fmr.next >= state->fmr.end)
1284 return -ENOMEM;
1285
1286 WARN_ON_ONCE(!dev->use_fmr);
1287
1288 if (state->npages == 0)
1289 return 0;
1290
1291 if (state->npages == 1 && target->global_mr) {
1292 srp_map_desc(state, state->base_dma_addr, state->dma_len,
1293 target->global_mr->rkey);
1294 goto reset_state;
1295 }
1296
1297 fmr = ib_fmr_pool_map_phys(ch->fmr_pool, state->pages,
1298 state->npages, io_addr);
1299 if (IS_ERR(fmr))
1300 return PTR_ERR(fmr);
1301
1302 *state->fmr.next++ = fmr;
1303 state->nmdesc++;
1304
1305 srp_map_desc(state, state->base_dma_addr & ~dev->mr_page_mask,
1306 state->dma_len, fmr->fmr->rkey);
1307
1308 reset_state:
1309 state->npages = 0;
1310 state->dma_len = 0;
1311
1312 return 0;
1313 }
1314
1315 static int srp_map_finish_fr(struct srp_map_state *state,
1316 struct srp_rdma_ch *ch)
1317 {
1318 struct srp_target_port *target = ch->target;
1319 struct srp_device *dev = target->srp_host->srp_dev;
1320 struct ib_send_wr *bad_wr;
1321 struct ib_reg_wr wr;
1322 struct srp_fr_desc *desc;
1323 u32 rkey;
1324 int n, err;
1325
1326 if (state->fr.next >= state->fr.end)
1327 return -ENOMEM;
1328
1329 WARN_ON_ONCE(!dev->use_fast_reg);
1330
1331 if (state->sg_nents == 0)
1332 return 0;
1333
1334 if (state->sg_nents == 1 && target->global_mr) {
1335 srp_map_desc(state, sg_dma_address(state->sg),
1336 sg_dma_len(state->sg),
1337 target->global_mr->rkey);
1338 return 1;
1339 }
1340
1341 desc = srp_fr_pool_get(ch->fr_pool);
1342 if (!desc)
1343 return -ENOMEM;
1344
1345 rkey = ib_inc_rkey(desc->mr->rkey);
1346 ib_update_fast_reg_key(desc->mr, rkey);
1347
1348 n = ib_map_mr_sg(desc->mr, state->sg, state->sg_nents,
1349 dev->mr_page_size);
1350 if (unlikely(n < 0))
1351 return n;
1352
1353 wr.wr.next = NULL;
1354 wr.wr.opcode = IB_WR_REG_MR;
1355 wr.wr.wr_id = FAST_REG_WR_ID_MASK;
1356 wr.wr.num_sge = 0;
1357 wr.wr.send_flags = 0;
1358 wr.mr = desc->mr;
1359 wr.key = desc->mr->rkey;
1360 wr.access = (IB_ACCESS_LOCAL_WRITE |
1361 IB_ACCESS_REMOTE_READ |
1362 IB_ACCESS_REMOTE_WRITE);
1363
1364 *state->fr.next++ = desc;
1365 state->nmdesc++;
1366
1367 srp_map_desc(state, desc->mr->iova,
1368 desc->mr->length, desc->mr->rkey);
1369
1370 err = ib_post_send(ch->qp, &wr.wr, &bad_wr);
1371 if (unlikely(err))
1372 return err;
1373
1374 return n;
1375 }
1376
1377 static int srp_map_sg_entry(struct srp_map_state *state,
1378 struct srp_rdma_ch *ch,
1379 struct scatterlist *sg, int sg_index)
1380 {
1381 struct srp_target_port *target = ch->target;
1382 struct srp_device *dev = target->srp_host->srp_dev;
1383 struct ib_device *ibdev = dev->dev;
1384 dma_addr_t dma_addr = ib_sg_dma_address(ibdev, sg);
1385 unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
1386 unsigned int len = 0;
1387 int ret;
1388
1389 WARN_ON_ONCE(!dma_len);
1390
1391 while (dma_len) {
1392 unsigned offset = dma_addr & ~dev->mr_page_mask;
1393 if (state->npages == dev->max_pages_per_mr || offset != 0) {
1394 ret = srp_map_finish_fmr(state, ch);
1395 if (ret)
1396 return ret;
1397 }
1398
1399 len = min_t(unsigned int, dma_len, dev->mr_page_size - offset);
1400
1401 if (!state->npages)
1402 state->base_dma_addr = dma_addr;
1403 state->pages[state->npages++] = dma_addr & dev->mr_page_mask;
1404 state->dma_len += len;
1405 dma_addr += len;
1406 dma_len -= len;
1407 }
1408
1409 /*
1410 * If the last entry of the MR wasn't a full page, then we need to
1411 * close it out and start a new one -- we can only merge at page
1412 * boundries.
1413 */
1414 ret = 0;
1415 if (len != dev->mr_page_size)
1416 ret = srp_map_finish_fmr(state, ch);
1417 return ret;
1418 }
1419
1420 static int srp_map_sg_fmr(struct srp_map_state *state, struct srp_rdma_ch *ch,
1421 struct srp_request *req, struct scatterlist *scat,
1422 int count)
1423 {
1424 struct scatterlist *sg;
1425 int i, ret;
1426
1427 state->desc = req->indirect_desc;
1428 state->pages = req->map_page;
1429 state->fmr.next = req->fmr_list;
1430 state->fmr.end = req->fmr_list + ch->target->cmd_sg_cnt;
1431
1432 for_each_sg(scat, sg, count, i) {
1433 ret = srp_map_sg_entry(state, ch, sg, i);
1434 if (ret)
1435 return ret;
1436 }
1437
1438 ret = srp_map_finish_fmr(state, ch);
1439 if (ret)
1440 return ret;
1441
1442 req->nmdesc = state->nmdesc;
1443
1444 return 0;
1445 }
1446
1447 static int srp_map_sg_fr(struct srp_map_state *state, struct srp_rdma_ch *ch,
1448 struct srp_request *req, struct scatterlist *scat,
1449 int count)
1450 {
1451 state->desc = req->indirect_desc;
1452 state->fr.next = req->fr_list;
1453 state->fr.end = req->fr_list + ch->target->cmd_sg_cnt;
1454 state->sg = scat;
1455 state->sg_nents = scsi_sg_count(req->scmnd);
1456
1457 while (state->sg_nents) {
1458 int i, n;
1459
1460 n = srp_map_finish_fr(state, ch);
1461 if (unlikely(n < 0))
1462 return n;
1463
1464 state->sg_nents -= n;
1465 for (i = 0; i < n; i++)
1466 state->sg = sg_next(state->sg);
1467 }
1468
1469 req->nmdesc = state->nmdesc;
1470
1471 return 0;
1472 }
1473
1474 static int srp_map_sg_dma(struct srp_map_state *state, struct srp_rdma_ch *ch,
1475 struct srp_request *req, struct scatterlist *scat,
1476 int count)
1477 {
1478 struct srp_target_port *target = ch->target;
1479 struct srp_device *dev = target->srp_host->srp_dev;
1480 struct scatterlist *sg;
1481 int i;
1482
1483 state->desc = req->indirect_desc;
1484 for_each_sg(scat, sg, count, i) {
1485 srp_map_desc(state, ib_sg_dma_address(dev->dev, sg),
1486 ib_sg_dma_len(dev->dev, sg),
1487 target->global_mr->rkey);
1488 }
1489
1490 req->nmdesc = state->nmdesc;
1491
1492 return 0;
1493 }
1494
1495 /*
1496 * Register the indirect data buffer descriptor with the HCA.
1497 *
1498 * Note: since the indirect data buffer descriptor has been allocated with
1499 * kmalloc() it is guaranteed that this buffer is a physically contiguous
1500 * memory buffer.
1501 */
1502 static int srp_map_idb(struct srp_rdma_ch *ch, struct srp_request *req,
1503 void **next_mr, void **end_mr, u32 idb_len,
1504 __be32 *idb_rkey)
1505 {
1506 struct srp_target_port *target = ch->target;
1507 struct srp_device *dev = target->srp_host->srp_dev;
1508 struct srp_map_state state;
1509 struct srp_direct_buf idb_desc;
1510 u64 idb_pages[1];
1511 struct scatterlist idb_sg[1];
1512 int ret;
1513
1514 memset(&state, 0, sizeof(state));
1515 memset(&idb_desc, 0, sizeof(idb_desc));
1516 state.gen.next = next_mr;
1517 state.gen.end = end_mr;
1518 state.desc = &idb_desc;
1519 state.base_dma_addr = req->indirect_dma_addr;
1520 state.dma_len = idb_len;
1521
1522 if (dev->use_fast_reg) {
1523 state.sg = idb_sg;
1524 state.sg_nents = 1;
1525 sg_set_buf(idb_sg, req->indirect_desc, idb_len);
1526 idb_sg->dma_address = req->indirect_dma_addr; /* hack! */
1527 #ifdef CONFIG_NEED_SG_DMA_LENGTH
1528 idb_sg->dma_length = idb_sg->length; /* hack^2 */
1529 #endif
1530 ret = srp_map_finish_fr(&state, ch);
1531 if (ret < 0)
1532 return ret;
1533 } else if (dev->use_fmr) {
1534 state.pages = idb_pages;
1535 state.pages[0] = (req->indirect_dma_addr &
1536 dev->mr_page_mask);
1537 state.npages = 1;
1538 ret = srp_map_finish_fmr(&state, ch);
1539 if (ret < 0)
1540 return ret;
1541 } else {
1542 return -EINVAL;
1543 }
1544
1545 *idb_rkey = idb_desc.key;
1546
1547 return 0;
1548 }
1549
1550 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_rdma_ch *ch,
1551 struct srp_request *req)
1552 {
1553 struct srp_target_port *target = ch->target;
1554 struct scatterlist *scat;
1555 struct srp_cmd *cmd = req->cmd->buf;
1556 int len, nents, count, ret;
1557 struct srp_device *dev;
1558 struct ib_device *ibdev;
1559 struct srp_map_state state;
1560 struct srp_indirect_buf *indirect_hdr;
1561 u32 idb_len, table_len;
1562 __be32 idb_rkey;
1563 u8 fmt;
1564
1565 if (!scsi_sglist(scmnd) || scmnd->sc_data_direction == DMA_NONE)
1566 return sizeof (struct srp_cmd);
1567
1568 if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
1569 scmnd->sc_data_direction != DMA_TO_DEVICE) {
1570 shost_printk(KERN_WARNING, target->scsi_host,
1571 PFX "Unhandled data direction %d\n",
1572 scmnd->sc_data_direction);
1573 return -EINVAL;
1574 }
1575
1576 nents = scsi_sg_count(scmnd);
1577 scat = scsi_sglist(scmnd);
1578
1579 dev = target->srp_host->srp_dev;
1580 ibdev = dev->dev;
1581
1582 count = ib_dma_map_sg(ibdev, scat, nents, scmnd->sc_data_direction);
1583 if (unlikely(count == 0))
1584 return -EIO;
1585
1586 fmt = SRP_DATA_DESC_DIRECT;
1587 len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
1588
1589 if (count == 1 && target->global_mr) {
1590 /*
1591 * The midlayer only generated a single gather/scatter
1592 * entry, or DMA mapping coalesced everything to a
1593 * single entry. So a direct descriptor along with
1594 * the DMA MR suffices.
1595 */
1596 struct srp_direct_buf *buf = (void *) cmd->add_data;
1597
1598 buf->va = cpu_to_be64(ib_sg_dma_address(ibdev, scat));
1599 buf->key = cpu_to_be32(target->global_mr->rkey);
1600 buf->len = cpu_to_be32(ib_sg_dma_len(ibdev, scat));
1601
1602 req->nmdesc = 0;
1603 goto map_complete;
1604 }
1605
1606 /*
1607 * We have more than one scatter/gather entry, so build our indirect
1608 * descriptor table, trying to merge as many entries as we can.
1609 */
1610 indirect_hdr = (void *) cmd->add_data;
1611
1612 ib_dma_sync_single_for_cpu(ibdev, req->indirect_dma_addr,
1613 target->indirect_size, DMA_TO_DEVICE);
1614
1615 memset(&state, 0, sizeof(state));
1616 if (dev->use_fast_reg)
1617 srp_map_sg_fr(&state, ch, req, scat, count);
1618 else if (dev->use_fmr)
1619 srp_map_sg_fmr(&state, ch, req, scat, count);
1620 else
1621 srp_map_sg_dma(&state, ch, req, scat, count);
1622
1623 /* We've mapped the request, now pull as much of the indirect
1624 * descriptor table as we can into the command buffer. If this
1625 * target is not using an external indirect table, we are
1626 * guaranteed to fit into the command, as the SCSI layer won't
1627 * give us more S/G entries than we allow.
1628 */
1629 if (state.ndesc == 1) {
1630 /*
1631 * Memory registration collapsed the sg-list into one entry,
1632 * so use a direct descriptor.
1633 */
1634 struct srp_direct_buf *buf = (void *) cmd->add_data;
1635
1636 *buf = req->indirect_desc[0];
1637 goto map_complete;
1638 }
1639
1640 if (unlikely(target->cmd_sg_cnt < state.ndesc &&
1641 !target->allow_ext_sg)) {
1642 shost_printk(KERN_ERR, target->scsi_host,
1643 "Could not fit S/G list into SRP_CMD\n");
1644 return -EIO;
1645 }
1646
1647 count = min(state.ndesc, target->cmd_sg_cnt);
1648 table_len = state.ndesc * sizeof (struct srp_direct_buf);
1649 idb_len = sizeof(struct srp_indirect_buf) + table_len;
1650
1651 fmt = SRP_DATA_DESC_INDIRECT;
1652 len = sizeof(struct srp_cmd) + sizeof (struct srp_indirect_buf);
1653 len += count * sizeof (struct srp_direct_buf);
1654
1655 memcpy(indirect_hdr->desc_list, req->indirect_desc,
1656 count * sizeof (struct srp_direct_buf));
1657
1658 if (!target->global_mr) {
1659 ret = srp_map_idb(ch, req, state.gen.next, state.gen.end,
1660 idb_len, &idb_rkey);
1661 if (ret < 0)
1662 return ret;
1663 req->nmdesc++;
1664 } else {
1665 idb_rkey = target->global_mr->rkey;
1666 }
1667
1668 indirect_hdr->table_desc.va = cpu_to_be64(req->indirect_dma_addr);
1669 indirect_hdr->table_desc.key = idb_rkey;
1670 indirect_hdr->table_desc.len = cpu_to_be32(table_len);
1671 indirect_hdr->len = cpu_to_be32(state.total_len);
1672
1673 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1674 cmd->data_out_desc_cnt = count;
1675 else
1676 cmd->data_in_desc_cnt = count;
1677
1678 ib_dma_sync_single_for_device(ibdev, req->indirect_dma_addr, table_len,
1679 DMA_TO_DEVICE);
1680
1681 map_complete:
1682 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1683 cmd->buf_fmt = fmt << 4;
1684 else
1685 cmd->buf_fmt = fmt;
1686
1687 return len;
1688 }
1689
1690 /*
1691 * Return an IU and possible credit to the free pool
1692 */
1693 static void srp_put_tx_iu(struct srp_rdma_ch *ch, struct srp_iu *iu,
1694 enum srp_iu_type iu_type)
1695 {
1696 unsigned long flags;
1697
1698 spin_lock_irqsave(&ch->lock, flags);
1699 list_add(&iu->list, &ch->free_tx);
1700 if (iu_type != SRP_IU_RSP)
1701 ++ch->req_lim;
1702 spin_unlock_irqrestore(&ch->lock, flags);
1703 }
1704
1705 /*
1706 * Must be called with ch->lock held to protect req_lim and free_tx.
1707 * If IU is not sent, it must be returned using srp_put_tx_iu().
1708 *
1709 * Note:
1710 * An upper limit for the number of allocated information units for each
1711 * request type is:
1712 * - SRP_IU_CMD: SRP_CMD_SQ_SIZE, since the SCSI mid-layer never queues
1713 * more than Scsi_Host.can_queue requests.
1714 * - SRP_IU_TSK_MGMT: SRP_TSK_MGMT_SQ_SIZE.
1715 * - SRP_IU_RSP: 1, since a conforming SRP target never sends more than
1716 * one unanswered SRP request to an initiator.
1717 */
1718 static struct srp_iu *__srp_get_tx_iu(struct srp_rdma_ch *ch,
1719 enum srp_iu_type iu_type)
1720 {
1721 struct srp_target_port *target = ch->target;
1722 s32 rsv = (iu_type == SRP_IU_TSK_MGMT) ? 0 : SRP_TSK_MGMT_SQ_SIZE;
1723 struct srp_iu *iu;
1724
1725 srp_send_completion(ch->send_cq, ch);
1726
1727 if (list_empty(&ch->free_tx))
1728 return NULL;
1729
1730 /* Initiator responses to target requests do not consume credits */
1731 if (iu_type != SRP_IU_RSP) {
1732 if (ch->req_lim <= rsv) {
1733 ++target->zero_req_lim;
1734 return NULL;
1735 }
1736
1737 --ch->req_lim;
1738 }
1739
1740 iu = list_first_entry(&ch->free_tx, struct srp_iu, list);
1741 list_del(&iu->list);
1742 return iu;
1743 }
1744
1745 static int srp_post_send(struct srp_rdma_ch *ch, struct srp_iu *iu, int len)
1746 {
1747 struct srp_target_port *target = ch->target;
1748 struct ib_sge list;
1749 struct ib_send_wr wr, *bad_wr;
1750
1751 list.addr = iu->dma;
1752 list.length = len;
1753 list.lkey = target->lkey;
1754
1755 wr.next = NULL;
1756 wr.wr_id = (uintptr_t) iu;
1757 wr.sg_list = &list;
1758 wr.num_sge = 1;
1759 wr.opcode = IB_WR_SEND;
1760 wr.send_flags = IB_SEND_SIGNALED;
1761
1762 return ib_post_send(ch->qp, &wr, &bad_wr);
1763 }
1764
1765 static int srp_post_recv(struct srp_rdma_ch *ch, struct srp_iu *iu)
1766 {
1767 struct srp_target_port *target = ch->target;
1768 struct ib_recv_wr wr, *bad_wr;
1769 struct ib_sge list;
1770
1771 list.addr = iu->dma;
1772 list.length = iu->size;
1773 list.lkey = target->lkey;
1774
1775 wr.next = NULL;
1776 wr.wr_id = (uintptr_t) iu;
1777 wr.sg_list = &list;
1778 wr.num_sge = 1;
1779
1780 return ib_post_recv(ch->qp, &wr, &bad_wr);
1781 }
1782
1783 static void srp_process_rsp(struct srp_rdma_ch *ch, struct srp_rsp *rsp)
1784 {
1785 struct srp_target_port *target = ch->target;
1786 struct srp_request *req;
1787 struct scsi_cmnd *scmnd;
1788 unsigned long flags;
1789
1790 if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
1791 spin_lock_irqsave(&ch->lock, flags);
1792 ch->req_lim += be32_to_cpu(rsp->req_lim_delta);
1793 spin_unlock_irqrestore(&ch->lock, flags);
1794
1795 ch->tsk_mgmt_status = -1;
1796 if (be32_to_cpu(rsp->resp_data_len) >= 4)
1797 ch->tsk_mgmt_status = rsp->data[3];
1798 complete(&ch->tsk_mgmt_done);
1799 } else {
1800 scmnd = scsi_host_find_tag(target->scsi_host, rsp->tag);
1801 if (scmnd) {
1802 req = (void *)scmnd->host_scribble;
1803 scmnd = srp_claim_req(ch, req, NULL, scmnd);
1804 }
1805 if (!scmnd) {
1806 shost_printk(KERN_ERR, target->scsi_host,
1807 "Null scmnd for RSP w/tag %#016llx received on ch %td / QP %#x\n",
1808 rsp->tag, ch - target->ch, ch->qp->qp_num);
1809
1810 spin_lock_irqsave(&ch->lock, flags);
1811 ch->req_lim += be32_to_cpu(rsp->req_lim_delta);
1812 spin_unlock_irqrestore(&ch->lock, flags);
1813
1814 return;
1815 }
1816 scmnd->result = rsp->status;
1817
1818 if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
1819 memcpy(scmnd->sense_buffer, rsp->data +
1820 be32_to_cpu(rsp->resp_data_len),
1821 min_t(int, be32_to_cpu(rsp->sense_data_len),
1822 SCSI_SENSE_BUFFERSIZE));
1823 }
1824
1825 if (unlikely(rsp->flags & SRP_RSP_FLAG_DIUNDER))
1826 scsi_set_resid(scmnd, be32_to_cpu(rsp->data_in_res_cnt));
1827 else if (unlikely(rsp->flags & SRP_RSP_FLAG_DIOVER))
1828 scsi_set_resid(scmnd, -be32_to_cpu(rsp->data_in_res_cnt));
1829 else if (unlikely(rsp->flags & SRP_RSP_FLAG_DOUNDER))
1830 scsi_set_resid(scmnd, be32_to_cpu(rsp->data_out_res_cnt));
1831 else if (unlikely(rsp->flags & SRP_RSP_FLAG_DOOVER))
1832 scsi_set_resid(scmnd, -be32_to_cpu(rsp->data_out_res_cnt));
1833
1834 srp_free_req(ch, req, scmnd,
1835 be32_to_cpu(rsp->req_lim_delta));
1836
1837 scmnd->host_scribble = NULL;
1838 scmnd->scsi_done(scmnd);
1839 }
1840 }
1841
1842 static int srp_response_common(struct srp_rdma_ch *ch, s32 req_delta,
1843 void *rsp, int len)
1844 {
1845 struct srp_target_port *target = ch->target;
1846 struct ib_device *dev = target->srp_host->srp_dev->dev;
1847 unsigned long flags;
1848 struct srp_iu *iu;
1849 int err;
1850
1851 spin_lock_irqsave(&ch->lock, flags);
1852 ch->req_lim += req_delta;
1853 iu = __srp_get_tx_iu(ch, SRP_IU_RSP);
1854 spin_unlock_irqrestore(&ch->lock, flags);
1855
1856 if (!iu) {
1857 shost_printk(KERN_ERR, target->scsi_host, PFX
1858 "no IU available to send response\n");
1859 return 1;
1860 }
1861
1862 ib_dma_sync_single_for_cpu(dev, iu->dma, len, DMA_TO_DEVICE);
1863 memcpy(iu->buf, rsp, len);
1864 ib_dma_sync_single_for_device(dev, iu->dma, len, DMA_TO_DEVICE);
1865
1866 err = srp_post_send(ch, iu, len);
1867 if (err) {
1868 shost_printk(KERN_ERR, target->scsi_host, PFX
1869 "unable to post response: %d\n", err);
1870 srp_put_tx_iu(ch, iu, SRP_IU_RSP);
1871 }
1872
1873 return err;
1874 }
1875
1876 static void srp_process_cred_req(struct srp_rdma_ch *ch,
1877 struct srp_cred_req *req)
1878 {
1879 struct srp_cred_rsp rsp = {
1880 .opcode = SRP_CRED_RSP,
1881 .tag = req->tag,
1882 };
1883 s32 delta = be32_to_cpu(req->req_lim_delta);
1884
1885 if (srp_response_common(ch, delta, &rsp, sizeof(rsp)))
1886 shost_printk(KERN_ERR, ch->target->scsi_host, PFX
1887 "problems processing SRP_CRED_REQ\n");
1888 }
1889
1890 static void srp_process_aer_req(struct srp_rdma_ch *ch,
1891 struct srp_aer_req *req)
1892 {
1893 struct srp_target_port *target = ch->target;
1894 struct srp_aer_rsp rsp = {
1895 .opcode = SRP_AER_RSP,
1896 .tag = req->tag,
1897 };
1898 s32 delta = be32_to_cpu(req->req_lim_delta);
1899
1900 shost_printk(KERN_ERR, target->scsi_host, PFX
1901 "ignoring AER for LUN %llu\n", scsilun_to_int(&req->lun));
1902
1903 if (srp_response_common(ch, delta, &rsp, sizeof(rsp)))
1904 shost_printk(KERN_ERR, target->scsi_host, PFX
1905 "problems processing SRP_AER_REQ\n");
1906 }
1907
1908 static void srp_handle_recv(struct srp_rdma_ch *ch, struct ib_wc *wc)
1909 {
1910 struct srp_target_port *target = ch->target;
1911 struct ib_device *dev = target->srp_host->srp_dev->dev;
1912 struct srp_iu *iu = (struct srp_iu *) (uintptr_t) wc->wr_id;
1913 int res;
1914 u8 opcode;
1915
1916 ib_dma_sync_single_for_cpu(dev, iu->dma, ch->max_ti_iu_len,
1917 DMA_FROM_DEVICE);
1918
1919 opcode = *(u8 *) iu->buf;
1920
1921 if (0) {
1922 shost_printk(KERN_ERR, target->scsi_host,
1923 PFX "recv completion, opcode 0x%02x\n", opcode);
1924 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 8, 1,
1925 iu->buf, wc->byte_len, true);
1926 }
1927
1928 switch (opcode) {
1929 case SRP_RSP:
1930 srp_process_rsp(ch, iu->buf);
1931 break;
1932
1933 case SRP_CRED_REQ:
1934 srp_process_cred_req(ch, iu->buf);
1935 break;
1936
1937 case SRP_AER_REQ:
1938 srp_process_aer_req(ch, iu->buf);
1939 break;
1940
1941 case SRP_T_LOGOUT:
1942 /* XXX Handle target logout */
1943 shost_printk(KERN_WARNING, target->scsi_host,
1944 PFX "Got target logout request\n");
1945 break;
1946
1947 default:
1948 shost_printk(KERN_WARNING, target->scsi_host,
1949 PFX "Unhandled SRP opcode 0x%02x\n", opcode);
1950 break;
1951 }
1952
1953 ib_dma_sync_single_for_device(dev, iu->dma, ch->max_ti_iu_len,
1954 DMA_FROM_DEVICE);
1955
1956 res = srp_post_recv(ch, iu);
1957 if (res != 0)
1958 shost_printk(KERN_ERR, target->scsi_host,
1959 PFX "Recv failed with error code %d\n", res);
1960 }
1961
1962 /**
1963 * srp_tl_err_work() - handle a transport layer error
1964 * @work: Work structure embedded in an SRP target port.
1965 *
1966 * Note: This function may get invoked before the rport has been created,
1967 * hence the target->rport test.
1968 */
1969 static void srp_tl_err_work(struct work_struct *work)
1970 {
1971 struct srp_target_port *target;
1972
1973 target = container_of(work, struct srp_target_port, tl_err_work);
1974 if (target->rport)
1975 srp_start_tl_fail_timers(target->rport);
1976 }
1977
1978 static void srp_handle_qp_err(u64 wr_id, enum ib_wc_status wc_status,
1979 bool send_err, struct srp_rdma_ch *ch)
1980 {
1981 struct srp_target_port *target = ch->target;
1982
1983 if (wr_id == SRP_LAST_WR_ID) {
1984 complete(&ch->done);
1985 return;
1986 }
1987
1988 if (ch->connected && !target->qp_in_error) {
1989 if (wr_id & LOCAL_INV_WR_ID_MASK) {
1990 shost_printk(KERN_ERR, target->scsi_host, PFX
1991 "LOCAL_INV failed with status %s (%d)\n",
1992 ib_wc_status_msg(wc_status), wc_status);
1993 } else if (wr_id & FAST_REG_WR_ID_MASK) {
1994 shost_printk(KERN_ERR, target->scsi_host, PFX
1995 "FAST_REG_MR failed status %s (%d)\n",
1996 ib_wc_status_msg(wc_status), wc_status);
1997 } else {
1998 shost_printk(KERN_ERR, target->scsi_host,
1999 PFX "failed %s status %s (%d) for iu %p\n",
2000 send_err ? "send" : "receive",
2001 ib_wc_status_msg(wc_status), wc_status,
2002 (void *)(uintptr_t)wr_id);
2003 }
2004 queue_work(system_long_wq, &target->tl_err_work);
2005 }
2006 target->qp_in_error = true;
2007 }
2008
2009 static void srp_recv_completion(struct ib_cq *cq, void *ch_ptr)
2010 {
2011 struct srp_rdma_ch *ch = ch_ptr;
2012 struct ib_wc wc;
2013
2014 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
2015 while (ib_poll_cq(cq, 1, &wc) > 0) {
2016 if (likely(wc.status == IB_WC_SUCCESS)) {
2017 srp_handle_recv(ch, &wc);
2018 } else {
2019 srp_handle_qp_err(wc.wr_id, wc.status, false, ch);
2020 }
2021 }
2022 }
2023
2024 static void srp_send_completion(struct ib_cq *cq, void *ch_ptr)
2025 {
2026 struct srp_rdma_ch *ch = ch_ptr;
2027 struct ib_wc wc;
2028 struct srp_iu *iu;
2029
2030 while (ib_poll_cq(cq, 1, &wc) > 0) {
2031 if (likely(wc.status == IB_WC_SUCCESS)) {
2032 iu = (struct srp_iu *) (uintptr_t) wc.wr_id;
2033 list_add(&iu->list, &ch->free_tx);
2034 } else {
2035 srp_handle_qp_err(wc.wr_id, wc.status, true, ch);
2036 }
2037 }
2038 }
2039
2040 static int srp_queuecommand(struct Scsi_Host *shost, struct scsi_cmnd *scmnd)
2041 {
2042 struct srp_target_port *target = host_to_target(shost);
2043 struct srp_rport *rport = target->rport;
2044 struct srp_rdma_ch *ch;
2045 struct srp_request *req;
2046 struct srp_iu *iu;
2047 struct srp_cmd *cmd;
2048 struct ib_device *dev;
2049 unsigned long flags;
2050 u32 tag;
2051 u16 idx;
2052 int len, ret;
2053 const bool in_scsi_eh = !in_interrupt() && current == shost->ehandler;
2054
2055 /*
2056 * The SCSI EH thread is the only context from which srp_queuecommand()
2057 * can get invoked for blocked devices (SDEV_BLOCK /
2058 * SDEV_CREATED_BLOCK). Avoid racing with srp_reconnect_rport() by
2059 * locking the rport mutex if invoked from inside the SCSI EH.
2060 */
2061 if (in_scsi_eh)
2062 mutex_lock(&rport->mutex);
2063
2064 scmnd->result = srp_chkready(target->rport);
2065 if (unlikely(scmnd->result))
2066 goto err;
2067
2068 WARN_ON_ONCE(scmnd->request->tag < 0);
2069 tag = blk_mq_unique_tag(scmnd->request);
2070 ch = &target->ch[blk_mq_unique_tag_to_hwq(tag)];
2071 idx = blk_mq_unique_tag_to_tag(tag);
2072 WARN_ONCE(idx >= target->req_ring_size, "%s: tag %#x: idx %d >= %d\n",
2073 dev_name(&shost->shost_gendev), tag, idx,
2074 target->req_ring_size);
2075
2076 spin_lock_irqsave(&ch->lock, flags);
2077 iu = __srp_get_tx_iu(ch, SRP_IU_CMD);
2078 spin_unlock_irqrestore(&ch->lock, flags);
2079
2080 if (!iu)
2081 goto err;
2082
2083 req = &ch->req_ring[idx];
2084 dev = target->srp_host->srp_dev->dev;
2085 ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_iu_len,
2086 DMA_TO_DEVICE);
2087
2088 scmnd->host_scribble = (void *) req;
2089
2090 cmd = iu->buf;
2091 memset(cmd, 0, sizeof *cmd);
2092
2093 cmd->opcode = SRP_CMD;
2094 int_to_scsilun(scmnd->device->lun, &cmd->lun);
2095 cmd->tag = tag;
2096 memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
2097
2098 req->scmnd = scmnd;
2099 req->cmd = iu;
2100
2101 len = srp_map_data(scmnd, ch, req);
2102 if (len < 0) {
2103 shost_printk(KERN_ERR, target->scsi_host,
2104 PFX "Failed to map data (%d)\n", len);
2105 /*
2106 * If we ran out of memory descriptors (-ENOMEM) because an
2107 * application is queuing many requests with more than
2108 * max_pages_per_mr sg-list elements, tell the SCSI mid-layer
2109 * to reduce queue depth temporarily.
2110 */
2111 scmnd->result = len == -ENOMEM ?
2112 DID_OK << 16 | QUEUE_FULL << 1 : DID_ERROR << 16;
2113 goto err_iu;
2114 }
2115
2116 ib_dma_sync_single_for_device(dev, iu->dma, target->max_iu_len,
2117 DMA_TO_DEVICE);
2118
2119 if (srp_post_send(ch, iu, len)) {
2120 shost_printk(KERN_ERR, target->scsi_host, PFX "Send failed\n");
2121 goto err_unmap;
2122 }
2123
2124 ret = 0;
2125
2126 unlock_rport:
2127 if (in_scsi_eh)
2128 mutex_unlock(&rport->mutex);
2129
2130 return ret;
2131
2132 err_unmap:
2133 srp_unmap_data(scmnd, ch, req);
2134
2135 err_iu:
2136 srp_put_tx_iu(ch, iu, SRP_IU_CMD);
2137
2138 /*
2139 * Avoid that the loops that iterate over the request ring can
2140 * encounter a dangling SCSI command pointer.
2141 */
2142 req->scmnd = NULL;
2143
2144 err:
2145 if (scmnd->result) {
2146 scmnd->scsi_done(scmnd);
2147 ret = 0;
2148 } else {
2149 ret = SCSI_MLQUEUE_HOST_BUSY;
2150 }
2151
2152 goto unlock_rport;
2153 }
2154
2155 /*
2156 * Note: the resources allocated in this function are freed in
2157 * srp_free_ch_ib().
2158 */
2159 static int srp_alloc_iu_bufs(struct srp_rdma_ch *ch)
2160 {
2161 struct srp_target_port *target = ch->target;
2162 int i;
2163
2164 ch->rx_ring = kcalloc(target->queue_size, sizeof(*ch->rx_ring),
2165 GFP_KERNEL);
2166 if (!ch->rx_ring)
2167 goto err_no_ring;
2168 ch->tx_ring = kcalloc(target->queue_size, sizeof(*ch->tx_ring),
2169 GFP_KERNEL);
2170 if (!ch->tx_ring)
2171 goto err_no_ring;
2172
2173 for (i = 0; i < target->queue_size; ++i) {
2174 ch->rx_ring[i] = srp_alloc_iu(target->srp_host,
2175 ch->max_ti_iu_len,
2176 GFP_KERNEL, DMA_FROM_DEVICE);
2177 if (!ch->rx_ring[i])
2178 goto err;
2179 }
2180
2181 for (i = 0; i < target->queue_size; ++i) {
2182 ch->tx_ring[i] = srp_alloc_iu(target->srp_host,
2183 target->max_iu_len,
2184 GFP_KERNEL, DMA_TO_DEVICE);
2185 if (!ch->tx_ring[i])
2186 goto err;
2187
2188 list_add(&ch->tx_ring[i]->list, &ch->free_tx);
2189 }
2190
2191 return 0;
2192
2193 err:
2194 for (i = 0; i < target->queue_size; ++i) {
2195 srp_free_iu(target->srp_host, ch->rx_ring[i]);
2196 srp_free_iu(target->srp_host, ch->tx_ring[i]);
2197 }
2198
2199
2200 err_no_ring:
2201 kfree(ch->tx_ring);
2202 ch->tx_ring = NULL;
2203 kfree(ch->rx_ring);
2204 ch->rx_ring = NULL;
2205
2206 return -ENOMEM;
2207 }
2208
2209 static uint32_t srp_compute_rq_tmo(struct ib_qp_attr *qp_attr, int attr_mask)
2210 {
2211 uint64_t T_tr_ns, max_compl_time_ms;
2212 uint32_t rq_tmo_jiffies;
2213
2214 /*
2215 * According to section 11.2.4.2 in the IBTA spec (Modify Queue Pair,
2216 * table 91), both the QP timeout and the retry count have to be set
2217 * for RC QP's during the RTR to RTS transition.
2218 */
2219 WARN_ON_ONCE((attr_mask & (IB_QP_TIMEOUT | IB_QP_RETRY_CNT)) !=
2220 (IB_QP_TIMEOUT | IB_QP_RETRY_CNT));
2221
2222 /*
2223 * Set target->rq_tmo_jiffies to one second more than the largest time
2224 * it can take before an error completion is generated. See also
2225 * C9-140..142 in the IBTA spec for more information about how to
2226 * convert the QP Local ACK Timeout value to nanoseconds.
2227 */
2228 T_tr_ns = 4096 * (1ULL << qp_attr->timeout);
2229 max_compl_time_ms = qp_attr->retry_cnt * 4 * T_tr_ns;
2230 do_div(max_compl_time_ms, NSEC_PER_MSEC);
2231 rq_tmo_jiffies = msecs_to_jiffies(max_compl_time_ms + 1000);
2232
2233 return rq_tmo_jiffies;
2234 }
2235
2236 static void srp_cm_rep_handler(struct ib_cm_id *cm_id,
2237 const struct srp_login_rsp *lrsp,
2238 struct srp_rdma_ch *ch)
2239 {
2240 struct srp_target_port *target = ch->target;
2241 struct ib_qp_attr *qp_attr = NULL;
2242 int attr_mask = 0;
2243 int ret;
2244 int i;
2245
2246 if (lrsp->opcode == SRP_LOGIN_RSP) {
2247 ch->max_ti_iu_len = be32_to_cpu(lrsp->max_ti_iu_len);
2248 ch->req_lim = be32_to_cpu(lrsp->req_lim_delta);
2249
2250 /*
2251 * Reserve credits for task management so we don't
2252 * bounce requests back to the SCSI mid-layer.
2253 */
2254 target->scsi_host->can_queue
2255 = min(ch->req_lim - SRP_TSK_MGMT_SQ_SIZE,
2256 target->scsi_host->can_queue);
2257 target->scsi_host->cmd_per_lun
2258 = min_t(int, target->scsi_host->can_queue,
2259 target->scsi_host->cmd_per_lun);
2260 } else {
2261 shost_printk(KERN_WARNING, target->scsi_host,
2262 PFX "Unhandled RSP opcode %#x\n", lrsp->opcode);
2263 ret = -ECONNRESET;
2264 goto error;
2265 }
2266
2267 if (!ch->rx_ring) {
2268 ret = srp_alloc_iu_bufs(ch);
2269 if (ret)
2270 goto error;
2271 }
2272
2273 ret = -ENOMEM;
2274 qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
2275 if (!qp_attr)
2276 goto error;
2277
2278 qp_attr->qp_state = IB_QPS_RTR;
2279 ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
2280 if (ret)
2281 goto error_free;
2282
2283 ret = ib_modify_qp(ch->qp, qp_attr, attr_mask);
2284 if (ret)
2285 goto error_free;
2286
2287 for (i = 0; i < target->queue_size; i++) {
2288 struct srp_iu *iu = ch->rx_ring[i];
2289
2290 ret = srp_post_recv(ch, iu);
2291 if (ret)
2292 goto error_free;
2293 }
2294
2295 qp_attr->qp_state = IB_QPS_RTS;
2296 ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
2297 if (ret)
2298 goto error_free;
2299
2300 target->rq_tmo_jiffies = srp_compute_rq_tmo(qp_attr, attr_mask);
2301
2302 ret = ib_modify_qp(ch->qp, qp_attr, attr_mask);
2303 if (ret)
2304 goto error_free;
2305
2306 ret = ib_send_cm_rtu(cm_id, NULL, 0);
2307
2308 error_free:
2309 kfree(qp_attr);
2310
2311 error:
2312 ch->status = ret;
2313 }
2314
2315 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
2316 struct ib_cm_event *event,
2317 struct srp_rdma_ch *ch)
2318 {
2319 struct srp_target_port *target = ch->target;
2320 struct Scsi_Host *shost = target->scsi_host;
2321 struct ib_class_port_info *cpi;
2322 int opcode;
2323
2324 switch (event->param.rej_rcvd.reason) {
2325 case IB_CM_REJ_PORT_CM_REDIRECT:
2326 cpi = event->param.rej_rcvd.ari;
2327 ch->path.dlid = cpi->redirect_lid;
2328 ch->path.pkey = cpi->redirect_pkey;
2329 cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
2330 memcpy(ch->path.dgid.raw, cpi->redirect_gid, 16);
2331
2332 ch->status = ch->path.dlid ?
2333 SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
2334 break;
2335
2336 case IB_CM_REJ_PORT_REDIRECT:
2337 if (srp_target_is_topspin(target)) {
2338 /*
2339 * Topspin/Cisco SRP gateways incorrectly send
2340 * reject reason code 25 when they mean 24
2341 * (port redirect).
2342 */
2343 memcpy(ch->path.dgid.raw,
2344 event->param.rej_rcvd.ari, 16);
2345
2346 shost_printk(KERN_DEBUG, shost,
2347 PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
2348 be64_to_cpu(ch->path.dgid.global.subnet_prefix),
2349 be64_to_cpu(ch->path.dgid.global.interface_id));
2350
2351 ch->status = SRP_PORT_REDIRECT;
2352 } else {
2353 shost_printk(KERN_WARNING, shost,
2354 " REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
2355 ch->status = -ECONNRESET;
2356 }
2357 break;
2358
2359 case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
2360 shost_printk(KERN_WARNING, shost,
2361 " REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
2362 ch->status = -ECONNRESET;
2363 break;
2364
2365 case IB_CM_REJ_CONSUMER_DEFINED:
2366 opcode = *(u8 *) event->private_data;
2367 if (opcode == SRP_LOGIN_REJ) {
2368 struct srp_login_rej *rej = event->private_data;
2369 u32 reason = be32_to_cpu(rej->reason);
2370
2371 if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
2372 shost_printk(KERN_WARNING, shost,
2373 PFX "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
2374 else
2375 shost_printk(KERN_WARNING, shost, PFX
2376 "SRP LOGIN from %pI6 to %pI6 REJECTED, reason 0x%08x\n",
2377 target->sgid.raw,
2378 target->orig_dgid.raw, reason);
2379 } else
2380 shost_printk(KERN_WARNING, shost,
2381 " REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
2382 " opcode 0x%02x\n", opcode);
2383 ch->status = -ECONNRESET;
2384 break;
2385
2386 case IB_CM_REJ_STALE_CONN:
2387 shost_printk(KERN_WARNING, shost, " REJ reason: stale connection\n");
2388 ch->status = SRP_STALE_CONN;
2389 break;
2390
2391 default:
2392 shost_printk(KERN_WARNING, shost, " REJ reason 0x%x\n",
2393 event->param.rej_rcvd.reason);
2394 ch->status = -ECONNRESET;
2395 }
2396 }
2397
2398 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
2399 {
2400 struct srp_rdma_ch *ch = cm_id->context;
2401 struct srp_target_port *target = ch->target;
2402 int comp = 0;
2403
2404 switch (event->event) {
2405 case IB_CM_REQ_ERROR:
2406 shost_printk(KERN_DEBUG, target->scsi_host,
2407 PFX "Sending CM REQ failed\n");
2408 comp = 1;
2409 ch->status = -ECONNRESET;
2410 break;
2411
2412 case IB_CM_REP_RECEIVED:
2413 comp = 1;
2414 srp_cm_rep_handler(cm_id, event->private_data, ch);
2415 break;
2416
2417 case IB_CM_REJ_RECEIVED:
2418 shost_printk(KERN_DEBUG, target->scsi_host, PFX "REJ received\n");
2419 comp = 1;
2420
2421 srp_cm_rej_handler(cm_id, event, ch);
2422 break;
2423
2424 case IB_CM_DREQ_RECEIVED:
2425 shost_printk(KERN_WARNING, target->scsi_host,
2426 PFX "DREQ received - connection closed\n");
2427 ch->connected = false;
2428 if (ib_send_cm_drep(cm_id, NULL, 0))
2429 shost_printk(KERN_ERR, target->scsi_host,
2430 PFX "Sending CM DREP failed\n");
2431 queue_work(system_long_wq, &target->tl_err_work);
2432 break;
2433
2434 case IB_CM_TIMEWAIT_EXIT:
2435 shost_printk(KERN_ERR, target->scsi_host,
2436 PFX "connection closed\n");
2437 comp = 1;
2438
2439 ch->status = 0;
2440 break;
2441
2442 case IB_CM_MRA_RECEIVED:
2443 case IB_CM_DREQ_ERROR:
2444 case IB_CM_DREP_RECEIVED:
2445 break;
2446
2447 default:
2448 shost_printk(KERN_WARNING, target->scsi_host,
2449 PFX "Unhandled CM event %d\n", event->event);
2450 break;
2451 }
2452
2453 if (comp)
2454 complete(&ch->done);
2455
2456 return 0;
2457 }
2458
2459 /**
2460 * srp_change_queue_depth - setting device queue depth
2461 * @sdev: scsi device struct
2462 * @qdepth: requested queue depth
2463 *
2464 * Returns queue depth.
2465 */
2466 static int
2467 srp_change_queue_depth(struct scsi_device *sdev, int qdepth)
2468 {
2469 if (!sdev->tagged_supported)
2470 qdepth = 1;
2471 return scsi_change_queue_depth(sdev, qdepth);
2472 }
2473
2474 static int srp_send_tsk_mgmt(struct srp_rdma_ch *ch, u64 req_tag, u64 lun,
2475 u8 func)
2476 {
2477 struct srp_target_port *target = ch->target;
2478 struct srp_rport *rport = target->rport;
2479 struct ib_device *dev = target->srp_host->srp_dev->dev;
2480 struct srp_iu *iu;
2481 struct srp_tsk_mgmt *tsk_mgmt;
2482
2483 if (!ch->connected || target->qp_in_error)
2484 return -1;
2485
2486 init_completion(&ch->tsk_mgmt_done);
2487
2488 /*
2489 * Lock the rport mutex to avoid that srp_create_ch_ib() is
2490 * invoked while a task management function is being sent.
2491 */
2492 mutex_lock(&rport->mutex);
2493 spin_lock_irq(&ch->lock);
2494 iu = __srp_get_tx_iu(ch, SRP_IU_TSK_MGMT);
2495 spin_unlock_irq(&ch->lock);
2496
2497 if (!iu) {
2498 mutex_unlock(&rport->mutex);
2499
2500 return -1;
2501 }
2502
2503 ib_dma_sync_single_for_cpu(dev, iu->dma, sizeof *tsk_mgmt,
2504 DMA_TO_DEVICE);
2505 tsk_mgmt = iu->buf;
2506 memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
2507
2508 tsk_mgmt->opcode = SRP_TSK_MGMT;
2509 int_to_scsilun(lun, &tsk_mgmt->lun);
2510 tsk_mgmt->tag = req_tag | SRP_TAG_TSK_MGMT;
2511 tsk_mgmt->tsk_mgmt_func = func;
2512 tsk_mgmt->task_tag = req_tag;
2513
2514 ib_dma_sync_single_for_device(dev, iu->dma, sizeof *tsk_mgmt,
2515 DMA_TO_DEVICE);
2516 if (srp_post_send(ch, iu, sizeof(*tsk_mgmt))) {
2517 srp_put_tx_iu(ch, iu, SRP_IU_TSK_MGMT);
2518 mutex_unlock(&rport->mutex);
2519
2520 return -1;
2521 }
2522 mutex_unlock(&rport->mutex);
2523
2524 if (!wait_for_completion_timeout(&ch->tsk_mgmt_done,
2525 msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
2526 return -1;
2527
2528 return 0;
2529 }
2530
2531 static int srp_abort(struct scsi_cmnd *scmnd)
2532 {
2533 struct srp_target_port *target = host_to_target(scmnd->device->host);
2534 struct srp_request *req = (struct srp_request *) scmnd->host_scribble;
2535 u32 tag;
2536 u16 ch_idx;
2537 struct srp_rdma_ch *ch;
2538 int ret;
2539
2540 shost_printk(KERN_ERR, target->scsi_host, "SRP abort called\n");
2541
2542 if (!req)
2543 return SUCCESS;
2544 tag = blk_mq_unique_tag(scmnd->request);
2545 ch_idx = blk_mq_unique_tag_to_hwq(tag);
2546 if (WARN_ON_ONCE(ch_idx >= target->ch_count))
2547 return SUCCESS;
2548 ch = &target->ch[ch_idx];
2549 if (!srp_claim_req(ch, req, NULL, scmnd))
2550 return SUCCESS;
2551 shost_printk(KERN_ERR, target->scsi_host,
2552 "Sending SRP abort for tag %#x\n", tag);
2553 if (srp_send_tsk_mgmt(ch, tag, scmnd->device->lun,
2554 SRP_TSK_ABORT_TASK) == 0)
2555 ret = SUCCESS;
2556 else if (target->rport->state == SRP_RPORT_LOST)
2557 ret = FAST_IO_FAIL;
2558 else
2559 ret = FAILED;
2560 srp_free_req(ch, req, scmnd, 0);
2561 scmnd->result = DID_ABORT << 16;
2562 scmnd->scsi_done(scmnd);
2563
2564 return ret;
2565 }
2566
2567 static int srp_reset_device(struct scsi_cmnd *scmnd)
2568 {
2569 struct srp_target_port *target = host_to_target(scmnd->device->host);
2570 struct srp_rdma_ch *ch;
2571 int i;
2572
2573 shost_printk(KERN_ERR, target->scsi_host, "SRP reset_device called\n");
2574
2575 ch = &target->ch[0];
2576 if (srp_send_tsk_mgmt(ch, SRP_TAG_NO_REQ, scmnd->device->lun,
2577 SRP_TSK_LUN_RESET))
2578 return FAILED;
2579 if (ch->tsk_mgmt_status)
2580 return FAILED;
2581
2582 for (i = 0; i < target->ch_count; i++) {
2583 ch = &target->ch[i];
2584 for (i = 0; i < target->req_ring_size; ++i) {
2585 struct srp_request *req = &ch->req_ring[i];
2586
2587 srp_finish_req(ch, req, scmnd->device, DID_RESET << 16);
2588 }
2589 }
2590
2591 return SUCCESS;
2592 }
2593
2594 static int srp_reset_host(struct scsi_cmnd *scmnd)
2595 {
2596 struct srp_target_port *target = host_to_target(scmnd->device->host);
2597
2598 shost_printk(KERN_ERR, target->scsi_host, PFX "SRP reset_host called\n");
2599
2600 return srp_reconnect_rport(target->rport) == 0 ? SUCCESS : FAILED;
2601 }
2602
2603 static int srp_slave_configure(struct scsi_device *sdev)
2604 {
2605 struct Scsi_Host *shost = sdev->host;
2606 struct srp_target_port *target = host_to_target(shost);
2607 struct request_queue *q = sdev->request_queue;
2608 unsigned long timeout;
2609
2610 if (sdev->type == TYPE_DISK) {
2611 timeout = max_t(unsigned, 30 * HZ, target->rq_tmo_jiffies);
2612 blk_queue_rq_timeout(q, timeout);
2613 }
2614
2615 return 0;
2616 }
2617
2618 static ssize_t show_id_ext(struct device *dev, struct device_attribute *attr,
2619 char *buf)
2620 {
2621 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2622
2623 return sprintf(buf, "0x%016llx\n", be64_to_cpu(target->id_ext));
2624 }
2625
2626 static ssize_t show_ioc_guid(struct device *dev, struct device_attribute *attr,
2627 char *buf)
2628 {
2629 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2630
2631 return sprintf(buf, "0x%016llx\n", be64_to_cpu(target->ioc_guid));
2632 }
2633
2634 static ssize_t show_service_id(struct device *dev,
2635 struct device_attribute *attr, char *buf)
2636 {
2637 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2638
2639 return sprintf(buf, "0x%016llx\n", be64_to_cpu(target->service_id));
2640 }
2641
2642 static ssize_t show_pkey(struct device *dev, struct device_attribute *attr,
2643 char *buf)
2644 {
2645 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2646
2647 return sprintf(buf, "0x%04x\n", be16_to_cpu(target->pkey));
2648 }
2649
2650 static ssize_t show_sgid(struct device *dev, struct device_attribute *attr,
2651 char *buf)
2652 {
2653 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2654
2655 return sprintf(buf, "%pI6\n", target->sgid.raw);
2656 }
2657
2658 static ssize_t show_dgid(struct device *dev, struct device_attribute *attr,
2659 char *buf)
2660 {
2661 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2662 struct srp_rdma_ch *ch = &target->ch[0];
2663
2664 return sprintf(buf, "%pI6\n", ch->path.dgid.raw);
2665 }
2666
2667 static ssize_t show_orig_dgid(struct device *dev,
2668 struct device_attribute *attr, char *buf)
2669 {
2670 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2671
2672 return sprintf(buf, "%pI6\n", target->orig_dgid.raw);
2673 }
2674
2675 static ssize_t show_req_lim(struct device *dev,
2676 struct device_attribute *attr, char *buf)
2677 {
2678 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2679 struct srp_rdma_ch *ch;
2680 int i, req_lim = INT_MAX;
2681
2682 for (i = 0; i < target->ch_count; i++) {
2683 ch = &target->ch[i];
2684 req_lim = min(req_lim, ch->req_lim);
2685 }
2686 return sprintf(buf, "%d\n", req_lim);
2687 }
2688
2689 static ssize_t show_zero_req_lim(struct device *dev,
2690 struct device_attribute *attr, char *buf)
2691 {
2692 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2693
2694 return sprintf(buf, "%d\n", target->zero_req_lim);
2695 }
2696
2697 static ssize_t show_local_ib_port(struct device *dev,
2698 struct device_attribute *attr, char *buf)
2699 {
2700 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2701
2702 return sprintf(buf, "%d\n", target->srp_host->port);
2703 }
2704
2705 static ssize_t show_local_ib_device(struct device *dev,
2706 struct device_attribute *attr, char *buf)
2707 {
2708 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2709
2710 return sprintf(buf, "%s\n", target->srp_host->srp_dev->dev->name);
2711 }
2712
2713 static ssize_t show_ch_count(struct device *dev, struct device_attribute *attr,
2714 char *buf)
2715 {
2716 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2717
2718 return sprintf(buf, "%d\n", target->ch_count);
2719 }
2720
2721 static ssize_t show_comp_vector(struct device *dev,
2722 struct device_attribute *attr, char *buf)
2723 {
2724 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2725
2726 return sprintf(buf, "%d\n", target->comp_vector);
2727 }
2728
2729 static ssize_t show_tl_retry_count(struct device *dev,
2730 struct device_attribute *attr, char *buf)
2731 {
2732 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2733
2734 return sprintf(buf, "%d\n", target->tl_retry_count);
2735 }
2736
2737 static ssize_t show_cmd_sg_entries(struct device *dev,
2738 struct device_attribute *attr, char *buf)
2739 {
2740 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2741
2742 return sprintf(buf, "%u\n", target->cmd_sg_cnt);
2743 }
2744
2745 static ssize_t show_allow_ext_sg(struct device *dev,
2746 struct device_attribute *attr, char *buf)
2747 {
2748 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2749
2750 return sprintf(buf, "%s\n", target->allow_ext_sg ? "true" : "false");
2751 }
2752
2753 static DEVICE_ATTR(id_ext, S_IRUGO, show_id_ext, NULL);
2754 static DEVICE_ATTR(ioc_guid, S_IRUGO, show_ioc_guid, NULL);
2755 static DEVICE_ATTR(service_id, S_IRUGO, show_service_id, NULL);
2756 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
2757 static DEVICE_ATTR(sgid, S_IRUGO, show_sgid, NULL);
2758 static DEVICE_ATTR(dgid, S_IRUGO, show_dgid, NULL);
2759 static DEVICE_ATTR(orig_dgid, S_IRUGO, show_orig_dgid, NULL);
2760 static DEVICE_ATTR(req_lim, S_IRUGO, show_req_lim, NULL);
2761 static DEVICE_ATTR(zero_req_lim, S_IRUGO, show_zero_req_lim, NULL);
2762 static DEVICE_ATTR(local_ib_port, S_IRUGO, show_local_ib_port, NULL);
2763 static DEVICE_ATTR(local_ib_device, S_IRUGO, show_local_ib_device, NULL);
2764 static DEVICE_ATTR(ch_count, S_IRUGO, show_ch_count, NULL);
2765 static DEVICE_ATTR(comp_vector, S_IRUGO, show_comp_vector, NULL);
2766 static DEVICE_ATTR(tl_retry_count, S_IRUGO, show_tl_retry_count, NULL);
2767 static DEVICE_ATTR(cmd_sg_entries, S_IRUGO, show_cmd_sg_entries, NULL);
2768 static DEVICE_ATTR(allow_ext_sg, S_IRUGO, show_allow_ext_sg, NULL);
2769
2770 static struct device_attribute *srp_host_attrs[] = {
2771 &dev_attr_id_ext,
2772 &dev_attr_ioc_guid,
2773 &dev_attr_service_id,
2774 &dev_attr_pkey,
2775 &dev_attr_sgid,
2776 &dev_attr_dgid,
2777 &dev_attr_orig_dgid,
2778 &dev_attr_req_lim,
2779 &dev_attr_zero_req_lim,
2780 &dev_attr_local_ib_port,
2781 &dev_attr_local_ib_device,
2782 &dev_attr_ch_count,
2783 &dev_attr_comp_vector,
2784 &dev_attr_tl_retry_count,
2785 &dev_attr_cmd_sg_entries,
2786 &dev_attr_allow_ext_sg,
2787 NULL
2788 };
2789
2790 static struct scsi_host_template srp_template = {
2791 .module = THIS_MODULE,
2792 .name = "InfiniBand SRP initiator",
2793 .proc_name = DRV_NAME,
2794 .slave_configure = srp_slave_configure,
2795 .info = srp_target_info,
2796 .queuecommand = srp_queuecommand,
2797 .change_queue_depth = srp_change_queue_depth,
2798 .eh_abort_handler = srp_abort,
2799 .eh_device_reset_handler = srp_reset_device,
2800 .eh_host_reset_handler = srp_reset_host,
2801 .skip_settle_delay = true,
2802 .sg_tablesize = SRP_DEF_SG_TABLESIZE,
2803 .can_queue = SRP_DEFAULT_CMD_SQ_SIZE,
2804 .this_id = -1,
2805 .cmd_per_lun = SRP_DEFAULT_CMD_SQ_SIZE,
2806 .use_clustering = ENABLE_CLUSTERING,
2807 .shost_attrs = srp_host_attrs,
2808 .track_queue_depth = 1,
2809 };
2810
2811 static int srp_sdev_count(struct Scsi_Host *host)
2812 {
2813 struct scsi_device *sdev;
2814 int c = 0;
2815
2816 shost_for_each_device(sdev, host)
2817 c++;
2818
2819 return c;
2820 }
2821
2822 /*
2823 * Return values:
2824 * < 0 upon failure. Caller is responsible for SRP target port cleanup.
2825 * 0 and target->state == SRP_TARGET_REMOVED if asynchronous target port
2826 * removal has been scheduled.
2827 * 0 and target->state != SRP_TARGET_REMOVED upon success.
2828 */
2829 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
2830 {
2831 struct srp_rport_identifiers ids;
2832 struct srp_rport *rport;
2833
2834 target->state = SRP_TARGET_SCANNING;
2835 sprintf(target->target_name, "SRP.T10:%016llX",
2836 be64_to_cpu(target->id_ext));
2837
2838 if (scsi_add_host(target->scsi_host, host->srp_dev->dev->dma_device))
2839 return -ENODEV;
2840
2841 memcpy(ids.port_id, &target->id_ext, 8);
2842 memcpy(ids.port_id + 8, &target->ioc_guid, 8);
2843 ids.roles = SRP_RPORT_ROLE_TARGET;
2844 rport = srp_rport_add(target->scsi_host, &ids);
2845 if (IS_ERR(rport)) {
2846 scsi_remove_host(target->scsi_host);
2847 return PTR_ERR(rport);
2848 }
2849
2850 rport->lld_data = target;
2851 target->rport = rport;
2852
2853 spin_lock(&host->target_lock);
2854 list_add_tail(&target->list, &host->target_list);
2855 spin_unlock(&host->target_lock);
2856
2857 scsi_scan_target(&target->scsi_host->shost_gendev,
2858 0, target->scsi_id, SCAN_WILD_CARD, 0);
2859
2860 if (srp_connected_ch(target) < target->ch_count ||
2861 target->qp_in_error) {
2862 shost_printk(KERN_INFO, target->scsi_host,
2863 PFX "SCSI scan failed - removing SCSI host\n");
2864 srp_queue_remove_work(target);
2865 goto out;
2866 }
2867
2868 pr_debug(PFX "%s: SCSI scan succeeded - detected %d LUNs\n",
2869 dev_name(&target->scsi_host->shost_gendev),
2870 srp_sdev_count(target->scsi_host));
2871
2872 spin_lock_irq(&target->lock);
2873 if (target->state == SRP_TARGET_SCANNING)
2874 target->state = SRP_TARGET_LIVE;
2875 spin_unlock_irq(&target->lock);
2876
2877 out:
2878 return 0;
2879 }
2880
2881 static void srp_release_dev(struct device *dev)
2882 {
2883 struct srp_host *host =
2884 container_of(dev, struct srp_host, dev);
2885
2886 complete(&host->released);
2887 }
2888
2889 static struct class srp_class = {
2890 .name = "infiniband_srp",
2891 .dev_release = srp_release_dev
2892 };
2893
2894 /**
2895 * srp_conn_unique() - check whether the connection to a target is unique
2896 * @host: SRP host.
2897 * @target: SRP target port.
2898 */
2899 static bool srp_conn_unique(struct srp_host *host,
2900 struct srp_target_port *target)
2901 {
2902 struct srp_target_port *t;
2903 bool ret = false;
2904
2905 if (target->state == SRP_TARGET_REMOVED)
2906 goto out;
2907
2908 ret = true;
2909
2910 spin_lock(&host->target_lock);
2911 list_for_each_entry(t, &host->target_list, list) {
2912 if (t != target &&
2913 target->id_ext == t->id_ext &&
2914 target->ioc_guid == t->ioc_guid &&
2915 target->initiator_ext == t->initiator_ext) {
2916 ret = false;
2917 break;
2918 }
2919 }
2920 spin_unlock(&host->target_lock);
2921
2922 out:
2923 return ret;
2924 }
2925
2926 /*
2927 * Target ports are added by writing
2928 *
2929 * id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
2930 * pkey=<P_Key>,service_id=<service ID>
2931 *
2932 * to the add_target sysfs attribute.
2933 */
2934 enum {
2935 SRP_OPT_ERR = 0,
2936 SRP_OPT_ID_EXT = 1 << 0,
2937 SRP_OPT_IOC_GUID = 1 << 1,
2938 SRP_OPT_DGID = 1 << 2,
2939 SRP_OPT_PKEY = 1 << 3,
2940 SRP_OPT_SERVICE_ID = 1 << 4,
2941 SRP_OPT_MAX_SECT = 1 << 5,
2942 SRP_OPT_MAX_CMD_PER_LUN = 1 << 6,
2943 SRP_OPT_IO_CLASS = 1 << 7,
2944 SRP_OPT_INITIATOR_EXT = 1 << 8,
2945 SRP_OPT_CMD_SG_ENTRIES = 1 << 9,
2946 SRP_OPT_ALLOW_EXT_SG = 1 << 10,
2947 SRP_OPT_SG_TABLESIZE = 1 << 11,
2948 SRP_OPT_COMP_VECTOR = 1 << 12,
2949 SRP_OPT_TL_RETRY_COUNT = 1 << 13,
2950 SRP_OPT_QUEUE_SIZE = 1 << 14,
2951 SRP_OPT_ALL = (SRP_OPT_ID_EXT |
2952 SRP_OPT_IOC_GUID |
2953 SRP_OPT_DGID |
2954 SRP_OPT_PKEY |
2955 SRP_OPT_SERVICE_ID),
2956 };
2957
2958 static const match_table_t srp_opt_tokens = {
2959 { SRP_OPT_ID_EXT, "id_ext=%s" },
2960 { SRP_OPT_IOC_GUID, "ioc_guid=%s" },
2961 { SRP_OPT_DGID, "dgid=%s" },
2962 { SRP_OPT_PKEY, "pkey=%x" },
2963 { SRP_OPT_SERVICE_ID, "service_id=%s" },
2964 { SRP_OPT_MAX_SECT, "max_sect=%d" },
2965 { SRP_OPT_MAX_CMD_PER_LUN, "max_cmd_per_lun=%d" },
2966 { SRP_OPT_IO_CLASS, "io_class=%x" },
2967 { SRP_OPT_INITIATOR_EXT, "initiator_ext=%s" },
2968 { SRP_OPT_CMD_SG_ENTRIES, "cmd_sg_entries=%u" },
2969 { SRP_OPT_ALLOW_EXT_SG, "allow_ext_sg=%u" },
2970 { SRP_OPT_SG_TABLESIZE, "sg_tablesize=%u" },
2971 { SRP_OPT_COMP_VECTOR, "comp_vector=%u" },
2972 { SRP_OPT_TL_RETRY_COUNT, "tl_retry_count=%u" },
2973 { SRP_OPT_QUEUE_SIZE, "queue_size=%d" },
2974 { SRP_OPT_ERR, NULL }
2975 };
2976
2977 static int srp_parse_options(const char *buf, struct srp_target_port *target)
2978 {
2979 char *options, *sep_opt;
2980 char *p;
2981 char dgid[3];
2982 substring_t args[MAX_OPT_ARGS];
2983 int opt_mask = 0;
2984 int token;
2985 int ret = -EINVAL;
2986 int i;
2987
2988 options = kstrdup(buf, GFP_KERNEL);
2989 if (!options)
2990 return -ENOMEM;
2991
2992 sep_opt = options;
2993 while ((p = strsep(&sep_opt, ",\n")) != NULL) {
2994 if (!*p)
2995 continue;
2996
2997 token = match_token(p, srp_opt_tokens, args);
2998 opt_mask |= token;
2999
3000 switch (token) {
3001 case SRP_OPT_ID_EXT:
3002 p = match_strdup(args);
3003 if (!p) {
3004 ret = -ENOMEM;
3005 goto out;
3006 }
3007 target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
3008 kfree(p);
3009 break;
3010
3011 case SRP_OPT_IOC_GUID:
3012 p = match_strdup(args);
3013 if (!p) {
3014 ret = -ENOMEM;
3015 goto out;
3016 }
3017 target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
3018 kfree(p);
3019 break;
3020
3021 case SRP_OPT_DGID:
3022 p = match_strdup(args);
3023 if (!p) {
3024 ret = -ENOMEM;
3025 goto out;
3026 }
3027 if (strlen(p) != 32) {
3028 pr_warn("bad dest GID parameter '%s'\n", p);
3029 kfree(p);
3030 goto out;
3031 }
3032
3033 for (i = 0; i < 16; ++i) {
3034 strlcpy(dgid, p + i * 2, sizeof(dgid));
3035 if (sscanf(dgid, "%hhx",
3036 &target->orig_dgid.raw[i]) < 1) {
3037 ret = -EINVAL;
3038 kfree(p);
3039 goto out;
3040 }
3041 }
3042 kfree(p);
3043 break;
3044
3045 case SRP_OPT_PKEY:
3046 if (match_hex(args, &token)) {
3047 pr_warn("bad P_Key parameter '%s'\n", p);
3048 goto out;
3049 }
3050 target->pkey = cpu_to_be16(token);
3051 break;
3052
3053 case SRP_OPT_SERVICE_ID:
3054 p = match_strdup(args);
3055 if (!p) {
3056 ret = -ENOMEM;
3057 goto out;
3058 }
3059 target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
3060 kfree(p);
3061 break;
3062
3063 case SRP_OPT_MAX_SECT:
3064 if (match_int(args, &token)) {
3065 pr_warn("bad max sect parameter '%s'\n", p);
3066 goto out;
3067 }
3068 target->scsi_host->max_sectors = token;
3069 break;
3070
3071 case SRP_OPT_QUEUE_SIZE:
3072 if (match_int(args, &token) || token < 1) {
3073 pr_warn("bad queue_size parameter '%s'\n", p);
3074 goto out;
3075 }
3076 target->scsi_host->can_queue = token;
3077 target->queue_size = token + SRP_RSP_SQ_SIZE +
3078 SRP_TSK_MGMT_SQ_SIZE;
3079 if (!(opt_mask & SRP_OPT_MAX_CMD_PER_LUN))
3080 target->scsi_host->cmd_per_lun = token;
3081 break;
3082
3083 case SRP_OPT_MAX_CMD_PER_LUN:
3084 if (match_int(args, &token) || token < 1) {
3085 pr_warn("bad max cmd_per_lun parameter '%s'\n",
3086 p);
3087 goto out;
3088 }
3089 target->scsi_host->cmd_per_lun = token;
3090 break;
3091
3092 case SRP_OPT_IO_CLASS:
3093 if (match_hex(args, &token)) {
3094 pr_warn("bad IO class parameter '%s'\n", p);
3095 goto out;
3096 }
3097 if (token != SRP_REV10_IB_IO_CLASS &&
3098 token != SRP_REV16A_IB_IO_CLASS) {
3099 pr_warn("unknown IO class parameter value %x specified (use %x or %x).\n",
3100 token, SRP_REV10_IB_IO_CLASS,
3101 SRP_REV16A_IB_IO_CLASS);
3102 goto out;
3103 }
3104 target->io_class = token;
3105 break;
3106
3107 case SRP_OPT_INITIATOR_EXT:
3108 p = match_strdup(args);
3109 if (!p) {
3110 ret = -ENOMEM;
3111 goto out;
3112 }
3113 target->initiator_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
3114 kfree(p);
3115 break;
3116
3117 case SRP_OPT_CMD_SG_ENTRIES:
3118 if (match_int(args, &token) || token < 1 || token > 255) {
3119 pr_warn("bad max cmd_sg_entries parameter '%s'\n",
3120 p);
3121 goto out;
3122 }
3123 target->cmd_sg_cnt = token;
3124 break;
3125
3126 case SRP_OPT_ALLOW_EXT_SG:
3127 if (match_int(args, &token)) {
3128 pr_warn("bad allow_ext_sg parameter '%s'\n", p);
3129 goto out;
3130 }
3131 target->allow_ext_sg = !!token;
3132 break;
3133
3134 case SRP_OPT_SG_TABLESIZE:
3135 if (match_int(args, &token) || token < 1 ||
3136 token > SCSI_MAX_SG_CHAIN_SEGMENTS) {
3137 pr_warn("bad max sg_tablesize parameter '%s'\n",
3138 p);
3139 goto out;
3140 }
3141 target->sg_tablesize = token;
3142 break;
3143
3144 case SRP_OPT_COMP_VECTOR:
3145 if (match_int(args, &token) || token < 0) {
3146 pr_warn("bad comp_vector parameter '%s'\n", p);
3147 goto out;
3148 }
3149 target->comp_vector = token;
3150 break;
3151
3152 case SRP_OPT_TL_RETRY_COUNT:
3153 if (match_int(args, &token) || token < 2 || token > 7) {
3154 pr_warn("bad tl_retry_count parameter '%s' (must be a number between 2 and 7)\n",
3155 p);
3156 goto out;
3157 }
3158 target->tl_retry_count = token;
3159 break;
3160
3161 default:
3162 pr_warn("unknown parameter or missing value '%s' in target creation request\n",
3163 p);
3164 goto out;
3165 }
3166 }
3167
3168 if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
3169 ret = 0;
3170 else
3171 for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
3172 if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
3173 !(srp_opt_tokens[i].token & opt_mask))
3174 pr_warn("target creation request is missing parameter '%s'\n",
3175 srp_opt_tokens[i].pattern);
3176
3177 if (target->scsi_host->cmd_per_lun > target->scsi_host->can_queue
3178 && (opt_mask & SRP_OPT_MAX_CMD_PER_LUN))
3179 pr_warn("cmd_per_lun = %d > queue_size = %d\n",
3180 target->scsi_host->cmd_per_lun,
3181 target->scsi_host->can_queue);
3182
3183 out:
3184 kfree(options);
3185 return ret;
3186 }
3187
3188 static ssize_t srp_create_target(struct device *dev,
3189 struct device_attribute *attr,
3190 const char *buf, size_t count)
3191 {
3192 struct srp_host *host =
3193 container_of(dev, struct srp_host, dev);
3194 struct Scsi_Host *target_host;
3195 struct srp_target_port *target;
3196 struct srp_rdma_ch *ch;
3197 struct srp_device *srp_dev = host->srp_dev;
3198 struct ib_device *ibdev = srp_dev->dev;
3199 int ret, node_idx, node, cpu, i;
3200 bool multich = false;
3201
3202 target_host = scsi_host_alloc(&srp_template,
3203 sizeof (struct srp_target_port));
3204 if (!target_host)
3205 return -ENOMEM;
3206
3207 target_host->transportt = ib_srp_transport_template;
3208 target_host->max_channel = 0;
3209 target_host->max_id = 1;
3210 target_host->max_lun = -1LL;
3211 target_host->max_cmd_len = sizeof ((struct srp_cmd *) (void *) 0L)->cdb;
3212
3213 target = host_to_target(target_host);
3214
3215 target->io_class = SRP_REV16A_IB_IO_CLASS;
3216 target->scsi_host = target_host;
3217 target->srp_host = host;
3218 target->lkey = host->srp_dev->pd->local_dma_lkey;
3219 target->global_mr = host->srp_dev->global_mr;
3220 target->cmd_sg_cnt = cmd_sg_entries;
3221 target->sg_tablesize = indirect_sg_entries ? : cmd_sg_entries;
3222 target->allow_ext_sg = allow_ext_sg;
3223 target->tl_retry_count = 7;
3224 target->queue_size = SRP_DEFAULT_QUEUE_SIZE;
3225
3226 /*
3227 * Avoid that the SCSI host can be removed by srp_remove_target()
3228 * before this function returns.
3229 */
3230 scsi_host_get(target->scsi_host);
3231
3232 mutex_lock(&host->add_target_mutex);
3233
3234 ret = srp_parse_options(buf, target);
3235 if (ret)
3236 goto out;
3237
3238 target->req_ring_size = target->queue_size - SRP_TSK_MGMT_SQ_SIZE;
3239
3240 if (!srp_conn_unique(target->srp_host, target)) {
3241 shost_printk(KERN_INFO, target->scsi_host,
3242 PFX "Already connected to target port with id_ext=%016llx;ioc_guid=%016llx;initiator_ext=%016llx\n",
3243 be64_to_cpu(target->id_ext),
3244 be64_to_cpu(target->ioc_guid),
3245 be64_to_cpu(target->initiator_ext));
3246 ret = -EEXIST;
3247 goto out;
3248 }
3249
3250 if (!srp_dev->has_fmr && !srp_dev->has_fr && !target->allow_ext_sg &&
3251 target->cmd_sg_cnt < target->sg_tablesize) {
3252 pr_warn("No MR pool and no external indirect descriptors, limiting sg_tablesize to cmd_sg_cnt\n");
3253 target->sg_tablesize = target->cmd_sg_cnt;
3254 }
3255
3256 target_host->sg_tablesize = target->sg_tablesize;
3257 target->indirect_size = target->sg_tablesize *
3258 sizeof (struct srp_direct_buf);
3259 target->max_iu_len = sizeof (struct srp_cmd) +
3260 sizeof (struct srp_indirect_buf) +
3261 target->cmd_sg_cnt * sizeof (struct srp_direct_buf);
3262
3263 INIT_WORK(&target->tl_err_work, srp_tl_err_work);
3264 INIT_WORK(&target->remove_work, srp_remove_work);
3265 spin_lock_init(&target->lock);
3266 ret = ib_query_gid(ibdev, host->port, 0, &target->sgid, NULL);
3267 if (ret)
3268 goto out;
3269
3270 ret = -ENOMEM;
3271 target->ch_count = max_t(unsigned, num_online_nodes(),
3272 min(ch_count ? :
3273 min(4 * num_online_nodes(),
3274 ibdev->num_comp_vectors),
3275 num_online_cpus()));
3276 target->ch = kcalloc(target->ch_count, sizeof(*target->ch),
3277 GFP_KERNEL);
3278 if (!target->ch)
3279 goto out;
3280
3281 node_idx = 0;
3282 for_each_online_node(node) {
3283 const int ch_start = (node_idx * target->ch_count /
3284 num_online_nodes());
3285 const int ch_end = ((node_idx + 1) * target->ch_count /
3286 num_online_nodes());
3287 const int cv_start = (node_idx * ibdev->num_comp_vectors /
3288 num_online_nodes() + target->comp_vector)
3289 % ibdev->num_comp_vectors;
3290 const int cv_end = ((node_idx + 1) * ibdev->num_comp_vectors /
3291 num_online_nodes() + target->comp_vector)
3292 % ibdev->num_comp_vectors;
3293 int cpu_idx = 0;
3294
3295 for_each_online_cpu(cpu) {
3296 if (cpu_to_node(cpu) != node)
3297 continue;
3298 if (ch_start + cpu_idx >= ch_end)
3299 continue;
3300 ch = &target->ch[ch_start + cpu_idx];
3301 ch->target = target;
3302 ch->comp_vector = cv_start == cv_end ? cv_start :
3303 cv_start + cpu_idx % (cv_end - cv_start);
3304 spin_lock_init(&ch->lock);
3305 INIT_LIST_HEAD(&ch->free_tx);
3306 ret = srp_new_cm_id(ch);
3307 if (ret)
3308 goto err_disconnect;
3309
3310 ret = srp_create_ch_ib(ch);
3311 if (ret)
3312 goto err_disconnect;
3313
3314 ret = srp_alloc_req_data(ch);
3315 if (ret)
3316 goto err_disconnect;
3317
3318 ret = srp_connect_ch(ch, multich);
3319 if (ret) {
3320 shost_printk(KERN_ERR, target->scsi_host,
3321 PFX "Connection %d/%d failed\n",
3322 ch_start + cpu_idx,
3323 target->ch_count);
3324 if (node_idx == 0 && cpu_idx == 0) {
3325 goto err_disconnect;
3326 } else {
3327 srp_free_ch_ib(target, ch);
3328 srp_free_req_data(target, ch);
3329 target->ch_count = ch - target->ch;
3330 goto connected;
3331 }
3332 }
3333
3334 multich = true;
3335 cpu_idx++;
3336 }
3337 node_idx++;
3338 }
3339
3340 connected:
3341 target->scsi_host->nr_hw_queues = target->ch_count;
3342
3343 ret = srp_add_target(host, target);
3344 if (ret)
3345 goto err_disconnect;
3346
3347 if (target->state != SRP_TARGET_REMOVED) {
3348 shost_printk(KERN_DEBUG, target->scsi_host, PFX
3349 "new target: id_ext %016llx ioc_guid %016llx pkey %04x service_id %016llx sgid %pI6 dgid %pI6\n",
3350 be64_to_cpu(target->id_ext),
3351 be64_to_cpu(target->ioc_guid),
3352 be16_to_cpu(target->pkey),
3353 be64_to_cpu(target->service_id),
3354 target->sgid.raw, target->orig_dgid.raw);
3355 }
3356
3357 ret = count;
3358
3359 out:
3360 mutex_unlock(&host->add_target_mutex);
3361
3362 scsi_host_put(target->scsi_host);
3363 if (ret < 0)
3364 scsi_host_put(target->scsi_host);
3365
3366 return ret;
3367
3368 err_disconnect:
3369 srp_disconnect_target(target);
3370
3371 for (i = 0; i < target->ch_count; i++) {
3372 ch = &target->ch[i];
3373 srp_free_ch_ib(target, ch);
3374 srp_free_req_data(target, ch);
3375 }
3376
3377 kfree(target->ch);
3378 goto out;
3379 }
3380
3381 static DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
3382
3383 static ssize_t show_ibdev(struct device *dev, struct device_attribute *attr,
3384 char *buf)
3385 {
3386 struct srp_host *host = container_of(dev, struct srp_host, dev);
3387
3388 return sprintf(buf, "%s\n", host->srp_dev->dev->name);
3389 }
3390
3391 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
3392
3393 static ssize_t show_port(struct device *dev, struct device_attribute *attr,
3394 char *buf)
3395 {
3396 struct srp_host *host = container_of(dev, struct srp_host, dev);
3397
3398 return sprintf(buf, "%d\n", host->port);
3399 }
3400
3401 static DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
3402
3403 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
3404 {
3405 struct srp_host *host;
3406
3407 host = kzalloc(sizeof *host, GFP_KERNEL);
3408 if (!host)
3409 return NULL;
3410
3411 INIT_LIST_HEAD(&host->target_list);
3412 spin_lock_init(&host->target_lock);
3413 init_completion(&host->released);
3414 mutex_init(&host->add_target_mutex);
3415 host->srp_dev = device;
3416 host->port = port;
3417
3418 host->dev.class = &srp_class;
3419 host->dev.parent = device->dev->dma_device;
3420 dev_set_name(&host->dev, "srp-%s-%d", device->dev->name, port);
3421
3422 if (device_register(&host->dev))
3423 goto free_host;
3424 if (device_create_file(&host->dev, &dev_attr_add_target))
3425 goto err_class;
3426 if (device_create_file(&host->dev, &dev_attr_ibdev))
3427 goto err_class;
3428 if (device_create_file(&host->dev, &dev_attr_port))
3429 goto err_class;
3430
3431 return host;
3432
3433 err_class:
3434 device_unregister(&host->dev);
3435
3436 free_host:
3437 kfree(host);
3438
3439 return NULL;
3440 }
3441
3442 static void srp_add_one(struct ib_device *device)
3443 {
3444 struct srp_device *srp_dev;
3445 struct ib_device_attr *dev_attr;
3446 struct srp_host *host;
3447 int mr_page_shift, p;
3448 u64 max_pages_per_mr;
3449
3450 dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
3451 if (!dev_attr)
3452 return;
3453
3454 if (ib_query_device(device, dev_attr)) {
3455 pr_warn("Query device failed for %s\n", device->name);
3456 goto free_attr;
3457 }
3458
3459 srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
3460 if (!srp_dev)
3461 goto free_attr;
3462
3463 srp_dev->has_fmr = (device->alloc_fmr && device->dealloc_fmr &&
3464 device->map_phys_fmr && device->unmap_fmr);
3465 srp_dev->has_fr = (dev_attr->device_cap_flags &
3466 IB_DEVICE_MEM_MGT_EXTENSIONS);
3467 if (!srp_dev->has_fmr && !srp_dev->has_fr)
3468 dev_warn(&device->dev, "neither FMR nor FR is supported\n");
3469
3470 srp_dev->use_fast_reg = (srp_dev->has_fr &&
3471 (!srp_dev->has_fmr || prefer_fr));
3472 srp_dev->use_fmr = !srp_dev->use_fast_reg && srp_dev->has_fmr;
3473
3474 /*
3475 * Use the smallest page size supported by the HCA, down to a
3476 * minimum of 4096 bytes. We're unlikely to build large sglists
3477 * out of smaller entries.
3478 */
3479 mr_page_shift = max(12, ffs(dev_attr->page_size_cap) - 1);
3480 srp_dev->mr_page_size = 1 << mr_page_shift;
3481 srp_dev->mr_page_mask = ~((u64) srp_dev->mr_page_size - 1);
3482 max_pages_per_mr = dev_attr->max_mr_size;
3483 do_div(max_pages_per_mr, srp_dev->mr_page_size);
3484 srp_dev->max_pages_per_mr = min_t(u64, SRP_MAX_PAGES_PER_MR,
3485 max_pages_per_mr);
3486 if (srp_dev->use_fast_reg) {
3487 srp_dev->max_pages_per_mr =
3488 min_t(u32, srp_dev->max_pages_per_mr,
3489 dev_attr->max_fast_reg_page_list_len);
3490 }
3491 srp_dev->mr_max_size = srp_dev->mr_page_size *
3492 srp_dev->max_pages_per_mr;
3493 pr_debug("%s: mr_page_shift = %d, dev_attr->max_mr_size = %#llx, dev_attr->max_fast_reg_page_list_len = %u, max_pages_per_mr = %d, mr_max_size = %#x\n",
3494 device->name, mr_page_shift, dev_attr->max_mr_size,
3495 dev_attr->max_fast_reg_page_list_len,
3496 srp_dev->max_pages_per_mr, srp_dev->mr_max_size);
3497
3498 INIT_LIST_HEAD(&srp_dev->dev_list);
3499
3500 srp_dev->dev = device;
3501 srp_dev->pd = ib_alloc_pd(device);
3502 if (IS_ERR(srp_dev->pd))
3503 goto free_dev;
3504
3505 if (!register_always || (!srp_dev->has_fmr && !srp_dev->has_fr)) {
3506 srp_dev->global_mr = ib_get_dma_mr(srp_dev->pd,
3507 IB_ACCESS_LOCAL_WRITE |
3508 IB_ACCESS_REMOTE_READ |
3509 IB_ACCESS_REMOTE_WRITE);
3510 if (IS_ERR(srp_dev->global_mr))
3511 goto err_pd;
3512 } else {
3513 srp_dev->global_mr = NULL;
3514 }
3515
3516 for (p = rdma_start_port(device); p <= rdma_end_port(device); ++p) {
3517 host = srp_add_port(srp_dev, p);
3518 if (host)
3519 list_add_tail(&host->list, &srp_dev->dev_list);
3520 }
3521
3522 ib_set_client_data(device, &srp_client, srp_dev);
3523
3524 goto free_attr;
3525
3526 err_pd:
3527 ib_dealloc_pd(srp_dev->pd);
3528
3529 free_dev:
3530 kfree(srp_dev);
3531
3532 free_attr:
3533 kfree(dev_attr);
3534 }
3535
3536 static void srp_remove_one(struct ib_device *device, void *client_data)
3537 {
3538 struct srp_device *srp_dev;
3539 struct srp_host *host, *tmp_host;
3540 struct srp_target_port *target;
3541
3542 srp_dev = client_data;
3543 if (!srp_dev)
3544 return;
3545
3546 list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
3547 device_unregister(&host->dev);
3548 /*
3549 * Wait for the sysfs entry to go away, so that no new
3550 * target ports can be created.
3551 */
3552 wait_for_completion(&host->released);
3553
3554 /*
3555 * Remove all target ports.
3556 */
3557 spin_lock(&host->target_lock);
3558 list_for_each_entry(target, &host->target_list, list)
3559 srp_queue_remove_work(target);
3560 spin_unlock(&host->target_lock);
3561
3562 /*
3563 * Wait for tl_err and target port removal tasks.
3564 */
3565 flush_workqueue(system_long_wq);
3566 flush_workqueue(srp_remove_wq);
3567
3568 kfree(host);
3569 }
3570
3571 if (srp_dev->global_mr)
3572 ib_dereg_mr(srp_dev->global_mr);
3573 ib_dealloc_pd(srp_dev->pd);
3574
3575 kfree(srp_dev);
3576 }
3577
3578 static struct srp_function_template ib_srp_transport_functions = {
3579 .has_rport_state = true,
3580 .reset_timer_if_blocked = true,
3581 .reconnect_delay = &srp_reconnect_delay,
3582 .fast_io_fail_tmo = &srp_fast_io_fail_tmo,
3583 .dev_loss_tmo = &srp_dev_loss_tmo,
3584 .reconnect = srp_rport_reconnect,
3585 .rport_delete = srp_rport_delete,
3586 .terminate_rport_io = srp_terminate_io,
3587 };
3588
3589 static int __init srp_init_module(void)
3590 {
3591 int ret;
3592
3593 BUILD_BUG_ON(FIELD_SIZEOF(struct ib_wc, wr_id) < sizeof(void *));
3594
3595 if (srp_sg_tablesize) {
3596 pr_warn("srp_sg_tablesize is deprecated, please use cmd_sg_entries\n");
3597 if (!cmd_sg_entries)
3598 cmd_sg_entries = srp_sg_tablesize;
3599 }
3600
3601 if (!cmd_sg_entries)
3602 cmd_sg_entries = SRP_DEF_SG_TABLESIZE;
3603
3604 if (cmd_sg_entries > 255) {
3605 pr_warn("Clamping cmd_sg_entries to 255\n");
3606 cmd_sg_entries = 255;
3607 }
3608
3609 if (!indirect_sg_entries)
3610 indirect_sg_entries = cmd_sg_entries;
3611 else if (indirect_sg_entries < cmd_sg_entries) {
3612 pr_warn("Bumping up indirect_sg_entries to match cmd_sg_entries (%u)\n",
3613 cmd_sg_entries);
3614 indirect_sg_entries = cmd_sg_entries;
3615 }
3616
3617 srp_remove_wq = create_workqueue("srp_remove");
3618 if (!srp_remove_wq) {
3619 ret = -ENOMEM;
3620 goto out;
3621 }
3622
3623 ret = -ENOMEM;
3624 ib_srp_transport_template =
3625 srp_attach_transport(&ib_srp_transport_functions);
3626 if (!ib_srp_transport_template)
3627 goto destroy_wq;
3628
3629 ret = class_register(&srp_class);
3630 if (ret) {
3631 pr_err("couldn't register class infiniband_srp\n");
3632 goto release_tr;
3633 }
3634
3635 ib_sa_register_client(&srp_sa_client);
3636
3637 ret = ib_register_client(&srp_client);
3638 if (ret) {
3639 pr_err("couldn't register IB client\n");
3640 goto unreg_sa;
3641 }
3642
3643 out:
3644 return ret;
3645
3646 unreg_sa:
3647 ib_sa_unregister_client(&srp_sa_client);
3648 class_unregister(&srp_class);
3649
3650 release_tr:
3651 srp_release_transport(ib_srp_transport_template);
3652
3653 destroy_wq:
3654 destroy_workqueue(srp_remove_wq);
3655 goto out;
3656 }
3657
3658 static void __exit srp_cleanup_module(void)
3659 {
3660 ib_unregister_client(&srp_client);
3661 ib_sa_unregister_client(&srp_sa_client);
3662 class_unregister(&srp_class);
3663 srp_release_transport(ib_srp_transport_template);
3664 destroy_workqueue(srp_remove_wq);
3665 }
3666
3667 module_init(srp_init_module);
3668 module_exit(srp_cleanup_module);
Linux kernel - Deliverables
This page took 0.110902 seconds and 5 git commands to generate.