IB/srp: Maintain a single connection per I_T nexus
[linux/fpc-iii.git] / drivers / infiniband / ulp / srp / ib_srp.c
blob830ef0037087fd47c2e370c59cc02d80c9fa1bbb
1 /*
2 * Copyright (c) 2005 Cisco Systems. All rights reserved.
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:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
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.
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.
33 #define pr_fmt(fmt) PFX fmt
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>
44 #include <linux/atomic.h>
46 #include <scsi/scsi.h>
47 #include <scsi/scsi_device.h>
48 #include <scsi/scsi_dbg.h>
49 #include <scsi/srp.h>
50 #include <scsi/scsi_transport_srp.h>
52 #include "ib_srp.h"
54 #define DRV_NAME "ib_srp"
55 #define PFX DRV_NAME ": "
56 #define DRV_VERSION "0.2"
57 #define DRV_RELDATE "November 1, 2005"
59 MODULE_AUTHOR("Roland Dreier");
60 MODULE_DESCRIPTION("InfiniBand SCSI RDMA Protocol initiator "
61 "v" DRV_VERSION " (" DRV_RELDATE ")");
62 MODULE_LICENSE("Dual BSD/GPL");
64 static unsigned int srp_sg_tablesize;
65 static unsigned int cmd_sg_entries;
66 static unsigned int indirect_sg_entries;
67 static bool allow_ext_sg;
68 static int topspin_workarounds = 1;
70 module_param(srp_sg_tablesize, uint, 0444);
71 MODULE_PARM_DESC(srp_sg_tablesize, "Deprecated name for cmd_sg_entries");
73 module_param(cmd_sg_entries, uint, 0444);
74 MODULE_PARM_DESC(cmd_sg_entries,
75 "Default number of gather/scatter entries in the SRP command (default is 12, max 255)");
77 module_param(indirect_sg_entries, uint, 0444);
78 MODULE_PARM_DESC(indirect_sg_entries,
79 "Default max number of gather/scatter entries (default is 12, max is " __stringify(SCSI_MAX_SG_CHAIN_SEGMENTS) ")");
81 module_param(allow_ext_sg, bool, 0444);
82 MODULE_PARM_DESC(allow_ext_sg,
83 "Default behavior when there are more than cmd_sg_entries S/G entries after mapping; fails the request when false (default false)");
85 module_param(topspin_workarounds, int, 0444);
86 MODULE_PARM_DESC(topspin_workarounds,
87 "Enable workarounds for Topspin/Cisco SRP target bugs if != 0");
89 static void srp_add_one(struct ib_device *device);
90 static void srp_remove_one(struct ib_device *device);
91 static void srp_recv_completion(struct ib_cq *cq, void *target_ptr);
92 static void srp_send_completion(struct ib_cq *cq, void *target_ptr);
93 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event);
95 static struct scsi_transport_template *ib_srp_transport_template;
97 static struct ib_client srp_client = {
98 .name = "srp",
99 .add = srp_add_one,
100 .remove = srp_remove_one
103 static struct ib_sa_client srp_sa_client;
105 static inline struct srp_target_port *host_to_target(struct Scsi_Host *host)
107 return (struct srp_target_port *) host->hostdata;
110 static const char *srp_target_info(struct Scsi_Host *host)
112 return host_to_target(host)->target_name;
115 static int srp_target_is_topspin(struct srp_target_port *target)
117 static const u8 topspin_oui[3] = { 0x00, 0x05, 0xad };
118 static const u8 cisco_oui[3] = { 0x00, 0x1b, 0x0d };
120 return topspin_workarounds &&
121 (!memcmp(&target->ioc_guid, topspin_oui, sizeof topspin_oui) ||
122 !memcmp(&target->ioc_guid, cisco_oui, sizeof cisco_oui));
125 static struct srp_iu *srp_alloc_iu(struct srp_host *host, size_t size,
126 gfp_t gfp_mask,
127 enum dma_data_direction direction)
129 struct srp_iu *iu;
131 iu = kmalloc(sizeof *iu, gfp_mask);
132 if (!iu)
133 goto out;
135 iu->buf = kzalloc(size, gfp_mask);
136 if (!iu->buf)
137 goto out_free_iu;
139 iu->dma = ib_dma_map_single(host->srp_dev->dev, iu->buf, size,
140 direction);
141 if (ib_dma_mapping_error(host->srp_dev->dev, iu->dma))
142 goto out_free_buf;
144 iu->size = size;
145 iu->direction = direction;
147 return iu;
149 out_free_buf:
150 kfree(iu->buf);
151 out_free_iu:
152 kfree(iu);
153 out:
154 return NULL;
157 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
159 if (!iu)
160 return;
162 ib_dma_unmap_single(host->srp_dev->dev, iu->dma, iu->size,
163 iu->direction);
164 kfree(iu->buf);
165 kfree(iu);
168 static void srp_qp_event(struct ib_event *event, void *context)
170 pr_debug("QP event %d\n", event->event);
173 static int srp_init_qp(struct srp_target_port *target,
174 struct ib_qp *qp)
176 struct ib_qp_attr *attr;
177 int ret;
179 attr = kmalloc(sizeof *attr, GFP_KERNEL);
180 if (!attr)
181 return -ENOMEM;
183 ret = ib_find_pkey(target->srp_host->srp_dev->dev,
184 target->srp_host->port,
185 be16_to_cpu(target->path.pkey),
186 &attr->pkey_index);
187 if (ret)
188 goto out;
190 attr->qp_state = IB_QPS_INIT;
191 attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
192 IB_ACCESS_REMOTE_WRITE);
193 attr->port_num = target->srp_host->port;
195 ret = ib_modify_qp(qp, attr,
196 IB_QP_STATE |
197 IB_QP_PKEY_INDEX |
198 IB_QP_ACCESS_FLAGS |
199 IB_QP_PORT);
201 out:
202 kfree(attr);
203 return ret;
206 static int srp_new_cm_id(struct srp_target_port *target)
208 struct ib_cm_id *new_cm_id;
210 new_cm_id = ib_create_cm_id(target->srp_host->srp_dev->dev,
211 srp_cm_handler, target);
212 if (IS_ERR(new_cm_id))
213 return PTR_ERR(new_cm_id);
215 if (target->cm_id)
216 ib_destroy_cm_id(target->cm_id);
217 target->cm_id = new_cm_id;
219 return 0;
222 static int srp_create_target_ib(struct srp_target_port *target)
224 struct ib_qp_init_attr *init_attr;
225 struct ib_cq *recv_cq, *send_cq;
226 struct ib_qp *qp;
227 int ret;
229 init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
230 if (!init_attr)
231 return -ENOMEM;
233 recv_cq = ib_create_cq(target->srp_host->srp_dev->dev,
234 srp_recv_completion, NULL, target, SRP_RQ_SIZE, 0);
235 if (IS_ERR(recv_cq)) {
236 ret = PTR_ERR(recv_cq);
237 goto err;
240 send_cq = ib_create_cq(target->srp_host->srp_dev->dev,
241 srp_send_completion, NULL, target, SRP_SQ_SIZE, 0);
242 if (IS_ERR(send_cq)) {
243 ret = PTR_ERR(send_cq);
244 goto err_recv_cq;
247 ib_req_notify_cq(recv_cq, IB_CQ_NEXT_COMP);
249 init_attr->event_handler = srp_qp_event;
250 init_attr->cap.max_send_wr = SRP_SQ_SIZE;
251 init_attr->cap.max_recv_wr = SRP_RQ_SIZE;
252 init_attr->cap.max_recv_sge = 1;
253 init_attr->cap.max_send_sge = 1;
254 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
255 init_attr->qp_type = IB_QPT_RC;
256 init_attr->send_cq = send_cq;
257 init_attr->recv_cq = recv_cq;
259 qp = ib_create_qp(target->srp_host->srp_dev->pd, init_attr);
260 if (IS_ERR(qp)) {
261 ret = PTR_ERR(qp);
262 goto err_send_cq;
265 ret = srp_init_qp(target, qp);
266 if (ret)
267 goto err_qp;
269 if (target->qp)
270 ib_destroy_qp(target->qp);
271 if (target->recv_cq)
272 ib_destroy_cq(target->recv_cq);
273 if (target->send_cq)
274 ib_destroy_cq(target->send_cq);
276 target->qp = qp;
277 target->recv_cq = recv_cq;
278 target->send_cq = send_cq;
280 kfree(init_attr);
281 return 0;
283 err_qp:
284 ib_destroy_qp(qp);
286 err_send_cq:
287 ib_destroy_cq(send_cq);
289 err_recv_cq:
290 ib_destroy_cq(recv_cq);
292 err:
293 kfree(init_attr);
294 return ret;
297 static void srp_free_target_ib(struct srp_target_port *target)
299 int i;
301 ib_destroy_qp(target->qp);
302 ib_destroy_cq(target->send_cq);
303 ib_destroy_cq(target->recv_cq);
305 target->qp = NULL;
306 target->send_cq = target->recv_cq = NULL;
308 for (i = 0; i < SRP_RQ_SIZE; ++i)
309 srp_free_iu(target->srp_host, target->rx_ring[i]);
310 for (i = 0; i < SRP_SQ_SIZE; ++i)
311 srp_free_iu(target->srp_host, target->tx_ring[i]);
314 static void srp_path_rec_completion(int status,
315 struct ib_sa_path_rec *pathrec,
316 void *target_ptr)
318 struct srp_target_port *target = target_ptr;
320 target->status = status;
321 if (status)
322 shost_printk(KERN_ERR, target->scsi_host,
323 PFX "Got failed path rec status %d\n", status);
324 else
325 target->path = *pathrec;
326 complete(&target->done);
329 static int srp_lookup_path(struct srp_target_port *target)
331 target->path.numb_path = 1;
333 init_completion(&target->done);
335 target->path_query_id = ib_sa_path_rec_get(&srp_sa_client,
336 target->srp_host->srp_dev->dev,
337 target->srp_host->port,
338 &target->path,
339 IB_SA_PATH_REC_SERVICE_ID |
340 IB_SA_PATH_REC_DGID |
341 IB_SA_PATH_REC_SGID |
342 IB_SA_PATH_REC_NUMB_PATH |
343 IB_SA_PATH_REC_PKEY,
344 SRP_PATH_REC_TIMEOUT_MS,
345 GFP_KERNEL,
346 srp_path_rec_completion,
347 target, &target->path_query);
348 if (target->path_query_id < 0)
349 return target->path_query_id;
351 wait_for_completion(&target->done);
353 if (target->status < 0)
354 shost_printk(KERN_WARNING, target->scsi_host,
355 PFX "Path record query failed\n");
357 return target->status;
360 static int srp_send_req(struct srp_target_port *target)
362 struct {
363 struct ib_cm_req_param param;
364 struct srp_login_req priv;
365 } *req = NULL;
366 int status;
368 req = kzalloc(sizeof *req, GFP_KERNEL);
369 if (!req)
370 return -ENOMEM;
372 req->param.primary_path = &target->path;
373 req->param.alternate_path = NULL;
374 req->param.service_id = target->service_id;
375 req->param.qp_num = target->qp->qp_num;
376 req->param.qp_type = target->qp->qp_type;
377 req->param.private_data = &req->priv;
378 req->param.private_data_len = sizeof req->priv;
379 req->param.flow_control = 1;
381 get_random_bytes(&req->param.starting_psn, 4);
382 req->param.starting_psn &= 0xffffff;
385 * Pick some arbitrary defaults here; we could make these
386 * module parameters if anyone cared about setting them.
388 req->param.responder_resources = 4;
389 req->param.remote_cm_response_timeout = 20;
390 req->param.local_cm_response_timeout = 20;
391 req->param.retry_count = 7;
392 req->param.rnr_retry_count = 7;
393 req->param.max_cm_retries = 15;
395 req->priv.opcode = SRP_LOGIN_REQ;
396 req->priv.tag = 0;
397 req->priv.req_it_iu_len = cpu_to_be32(target->max_iu_len);
398 req->priv.req_buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
399 SRP_BUF_FORMAT_INDIRECT);
401 * In the published SRP specification (draft rev. 16a), the
402 * port identifier format is 8 bytes of ID extension followed
403 * by 8 bytes of GUID. Older drafts put the two halves in the
404 * opposite order, so that the GUID comes first.
406 * Targets conforming to these obsolete drafts can be
407 * recognized by the I/O Class they report.
409 if (target->io_class == SRP_REV10_IB_IO_CLASS) {
410 memcpy(req->priv.initiator_port_id,
411 &target->path.sgid.global.interface_id, 8);
412 memcpy(req->priv.initiator_port_id + 8,
413 &target->initiator_ext, 8);
414 memcpy(req->priv.target_port_id, &target->ioc_guid, 8);
415 memcpy(req->priv.target_port_id + 8, &target->id_ext, 8);
416 } else {
417 memcpy(req->priv.initiator_port_id,
418 &target->initiator_ext, 8);
419 memcpy(req->priv.initiator_port_id + 8,
420 &target->path.sgid.global.interface_id, 8);
421 memcpy(req->priv.target_port_id, &target->id_ext, 8);
422 memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
426 * Topspin/Cisco SRP targets will reject our login unless we
427 * zero out the first 8 bytes of our initiator port ID and set
428 * the second 8 bytes to the local node GUID.
430 if (srp_target_is_topspin(target)) {
431 shost_printk(KERN_DEBUG, target->scsi_host,
432 PFX "Topspin/Cisco initiator port ID workaround "
433 "activated for target GUID %016llx\n",
434 (unsigned long long) be64_to_cpu(target->ioc_guid));
435 memset(req->priv.initiator_port_id, 0, 8);
436 memcpy(req->priv.initiator_port_id + 8,
437 &target->srp_host->srp_dev->dev->node_guid, 8);
440 status = ib_send_cm_req(target->cm_id, &req->param);
442 kfree(req);
444 return status;
447 static bool srp_queue_remove_work(struct srp_target_port *target)
449 bool changed = false;
451 spin_lock_irq(&target->lock);
452 if (target->state != SRP_TARGET_REMOVED) {
453 target->state = SRP_TARGET_REMOVED;
454 changed = true;
456 spin_unlock_irq(&target->lock);
458 if (changed)
459 queue_work(system_long_wq, &target->remove_work);
461 return changed;
464 static bool srp_change_conn_state(struct srp_target_port *target,
465 bool connected)
467 bool changed = false;
469 spin_lock_irq(&target->lock);
470 if (target->connected != connected) {
471 target->connected = connected;
472 changed = true;
474 spin_unlock_irq(&target->lock);
476 return changed;
479 static void srp_disconnect_target(struct srp_target_port *target)
481 if (srp_change_conn_state(target, false)) {
482 /* XXX should send SRP_I_LOGOUT request */
484 if (ib_send_cm_dreq(target->cm_id, NULL, 0)) {
485 shost_printk(KERN_DEBUG, target->scsi_host,
486 PFX "Sending CM DREQ failed\n");
491 static void srp_free_req_data(struct srp_target_port *target)
493 struct ib_device *ibdev = target->srp_host->srp_dev->dev;
494 struct srp_request *req;
495 int i;
497 for (i = 0, req = target->req_ring; i < SRP_CMD_SQ_SIZE; ++i, ++req) {
498 kfree(req->fmr_list);
499 kfree(req->map_page);
500 if (req->indirect_dma_addr) {
501 ib_dma_unmap_single(ibdev, req->indirect_dma_addr,
502 target->indirect_size,
503 DMA_TO_DEVICE);
505 kfree(req->indirect_desc);
510 * srp_del_scsi_host_attr() - Remove attributes defined in the host template.
511 * @shost: SCSI host whose attributes to remove from sysfs.
513 * Note: Any attributes defined in the host template and that did not exist
514 * before invocation of this function will be ignored.
516 static void srp_del_scsi_host_attr(struct Scsi_Host *shost)
518 struct device_attribute **attr;
520 for (attr = shost->hostt->shost_attrs; attr && *attr; ++attr)
521 device_remove_file(&shost->shost_dev, *attr);
524 static void srp_remove_target(struct srp_target_port *target)
526 WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
528 srp_del_scsi_host_attr(target->scsi_host);
529 srp_remove_host(target->scsi_host);
530 scsi_remove_host(target->scsi_host);
531 srp_disconnect_target(target);
532 ib_destroy_cm_id(target->cm_id);
533 srp_free_target_ib(target);
534 srp_free_req_data(target);
535 scsi_host_put(target->scsi_host);
538 static void srp_remove_work(struct work_struct *work)
540 struct srp_target_port *target =
541 container_of(work, struct srp_target_port, remove_work);
543 WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
545 srp_remove_target(target);
547 spin_lock(&target->srp_host->target_lock);
548 list_del(&target->list);
549 spin_unlock(&target->srp_host->target_lock);
552 static void srp_rport_delete(struct srp_rport *rport)
554 struct srp_target_port *target = rport->lld_data;
556 srp_queue_remove_work(target);
559 static int srp_connect_target(struct srp_target_port *target)
561 int retries = 3;
562 int ret;
564 WARN_ON_ONCE(target->connected);
566 target->qp_in_error = false;
568 ret = srp_lookup_path(target);
569 if (ret)
570 return ret;
572 while (1) {
573 init_completion(&target->done);
574 ret = srp_send_req(target);
575 if (ret)
576 return ret;
577 wait_for_completion(&target->done);
580 * The CM event handling code will set status to
581 * SRP_PORT_REDIRECT if we get a port redirect REJ
582 * back, or SRP_DLID_REDIRECT if we get a lid/qp
583 * redirect REJ back.
585 switch (target->status) {
586 case 0:
587 srp_change_conn_state(target, true);
588 return 0;
590 case SRP_PORT_REDIRECT:
591 ret = srp_lookup_path(target);
592 if (ret)
593 return ret;
594 break;
596 case SRP_DLID_REDIRECT:
597 break;
599 case SRP_STALE_CONN:
600 /* Our current CM id was stale, and is now in timewait.
601 * Try to reconnect with a new one.
603 if (!retries-- || srp_new_cm_id(target)) {
604 shost_printk(KERN_ERR, target->scsi_host, PFX
605 "giving up on stale connection\n");
606 target->status = -ECONNRESET;
607 return target->status;
610 shost_printk(KERN_ERR, target->scsi_host, PFX
611 "retrying stale connection\n");
612 break;
614 default:
615 return target->status;
620 static void srp_unmap_data(struct scsi_cmnd *scmnd,
621 struct srp_target_port *target,
622 struct srp_request *req)
624 struct ib_device *ibdev = target->srp_host->srp_dev->dev;
625 struct ib_pool_fmr **pfmr;
627 if (!scsi_sglist(scmnd) ||
628 (scmnd->sc_data_direction != DMA_TO_DEVICE &&
629 scmnd->sc_data_direction != DMA_FROM_DEVICE))
630 return;
632 pfmr = req->fmr_list;
633 while (req->nfmr--)
634 ib_fmr_pool_unmap(*pfmr++);
636 ib_dma_unmap_sg(ibdev, scsi_sglist(scmnd), scsi_sg_count(scmnd),
637 scmnd->sc_data_direction);
641 * srp_claim_req - Take ownership of the scmnd associated with a request.
642 * @target: SRP target port.
643 * @req: SRP request.
644 * @scmnd: If NULL, take ownership of @req->scmnd. If not NULL, only take
645 * ownership of @req->scmnd if it equals @scmnd.
647 * Return value:
648 * Either NULL or a pointer to the SCSI command the caller became owner of.
650 static struct scsi_cmnd *srp_claim_req(struct srp_target_port *target,
651 struct srp_request *req,
652 struct scsi_cmnd *scmnd)
654 unsigned long flags;
656 spin_lock_irqsave(&target->lock, flags);
657 if (!scmnd) {
658 scmnd = req->scmnd;
659 req->scmnd = NULL;
660 } else if (req->scmnd == scmnd) {
661 req->scmnd = NULL;
662 } else {
663 scmnd = NULL;
665 spin_unlock_irqrestore(&target->lock, flags);
667 return scmnd;
671 * srp_free_req() - Unmap data and add request to the free request list.
673 static void srp_free_req(struct srp_target_port *target,
674 struct srp_request *req, struct scsi_cmnd *scmnd,
675 s32 req_lim_delta)
677 unsigned long flags;
679 srp_unmap_data(scmnd, target, req);
681 spin_lock_irqsave(&target->lock, flags);
682 target->req_lim += req_lim_delta;
683 list_add_tail(&req->list, &target->free_reqs);
684 spin_unlock_irqrestore(&target->lock, flags);
687 static void srp_reset_req(struct srp_target_port *target, struct srp_request *req)
689 struct scsi_cmnd *scmnd = srp_claim_req(target, req, NULL);
691 if (scmnd) {
692 srp_free_req(target, req, scmnd, 0);
693 scmnd->result = DID_RESET << 16;
694 scmnd->scsi_done(scmnd);
698 static int srp_reconnect_target(struct srp_target_port *target)
700 struct Scsi_Host *shost = target->scsi_host;
701 int i, ret;
703 scsi_target_block(&shost->shost_gendev);
705 srp_disconnect_target(target);
707 * Now get a new local CM ID so that we avoid confusing the target in
708 * case things are really fouled up. Doing so also ensures that all CM
709 * callbacks will have finished before a new QP is allocated.
711 ret = srp_new_cm_id(target);
713 * Whether or not creating a new CM ID succeeded, create a new
714 * QP. This guarantees that all completion callback function
715 * invocations have finished before request resetting starts.
717 if (ret == 0)
718 ret = srp_create_target_ib(target);
719 else
720 srp_create_target_ib(target);
722 for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
723 struct srp_request *req = &target->req_ring[i];
724 if (req->scmnd)
725 srp_reset_req(target, req);
728 INIT_LIST_HEAD(&target->free_tx);
729 for (i = 0; i < SRP_SQ_SIZE; ++i)
730 list_add(&target->tx_ring[i]->list, &target->free_tx);
732 if (ret == 0)
733 ret = srp_connect_target(target);
735 scsi_target_unblock(&shost->shost_gendev, ret == 0 ? SDEV_RUNNING :
736 SDEV_TRANSPORT_OFFLINE);
737 target->transport_offline = !!ret;
739 if (ret)
740 goto err;
742 shost_printk(KERN_INFO, target->scsi_host, PFX "reconnect succeeded\n");
744 return ret;
746 err:
747 shost_printk(KERN_ERR, target->scsi_host,
748 PFX "reconnect failed (%d), removing target port.\n", ret);
751 * We couldn't reconnect, so kill our target port off.
752 * However, we have to defer the real removal because we
753 * are in the context of the SCSI error handler now, which
754 * will deadlock if we call scsi_remove_host().
756 srp_queue_remove_work(target);
758 return ret;
761 static void srp_map_desc(struct srp_map_state *state, dma_addr_t dma_addr,
762 unsigned int dma_len, u32 rkey)
764 struct srp_direct_buf *desc = state->desc;
766 desc->va = cpu_to_be64(dma_addr);
767 desc->key = cpu_to_be32(rkey);
768 desc->len = cpu_to_be32(dma_len);
770 state->total_len += dma_len;
771 state->desc++;
772 state->ndesc++;
775 static int srp_map_finish_fmr(struct srp_map_state *state,
776 struct srp_target_port *target)
778 struct srp_device *dev = target->srp_host->srp_dev;
779 struct ib_pool_fmr *fmr;
780 u64 io_addr = 0;
782 if (!state->npages)
783 return 0;
785 if (state->npages == 1) {
786 srp_map_desc(state, state->base_dma_addr, state->fmr_len,
787 target->rkey);
788 state->npages = state->fmr_len = 0;
789 return 0;
792 fmr = ib_fmr_pool_map_phys(dev->fmr_pool, state->pages,
793 state->npages, io_addr);
794 if (IS_ERR(fmr))
795 return PTR_ERR(fmr);
797 *state->next_fmr++ = fmr;
798 state->nfmr++;
800 srp_map_desc(state, 0, state->fmr_len, fmr->fmr->rkey);
801 state->npages = state->fmr_len = 0;
802 return 0;
805 static void srp_map_update_start(struct srp_map_state *state,
806 struct scatterlist *sg, int sg_index,
807 dma_addr_t dma_addr)
809 state->unmapped_sg = sg;
810 state->unmapped_index = sg_index;
811 state->unmapped_addr = dma_addr;
814 static int srp_map_sg_entry(struct srp_map_state *state,
815 struct srp_target_port *target,
816 struct scatterlist *sg, int sg_index,
817 int use_fmr)
819 struct srp_device *dev = target->srp_host->srp_dev;
820 struct ib_device *ibdev = dev->dev;
821 dma_addr_t dma_addr = ib_sg_dma_address(ibdev, sg);
822 unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
823 unsigned int len;
824 int ret;
826 if (!dma_len)
827 return 0;
829 if (use_fmr == SRP_MAP_NO_FMR) {
830 /* Once we're in direct map mode for a request, we don't
831 * go back to FMR mode, so no need to update anything
832 * other than the descriptor.
834 srp_map_desc(state, dma_addr, dma_len, target->rkey);
835 return 0;
838 /* If we start at an offset into the FMR page, don't merge into
839 * the current FMR. Finish it out, and use the kernel's MR for this
840 * sg entry. This is to avoid potential bugs on some SRP targets
841 * that were never quite defined, but went away when the initiator
842 * avoided using FMR on such page fragments.
844 if (dma_addr & ~dev->fmr_page_mask || dma_len > dev->fmr_max_size) {
845 ret = srp_map_finish_fmr(state, target);
846 if (ret)
847 return ret;
849 srp_map_desc(state, dma_addr, dma_len, target->rkey);
850 srp_map_update_start(state, NULL, 0, 0);
851 return 0;
854 /* If this is the first sg to go into the FMR, save our position.
855 * We need to know the first unmapped entry, its index, and the
856 * first unmapped address within that entry to be able to restart
857 * mapping after an error.
859 if (!state->unmapped_sg)
860 srp_map_update_start(state, sg, sg_index, dma_addr);
862 while (dma_len) {
863 if (state->npages == SRP_FMR_SIZE) {
864 ret = srp_map_finish_fmr(state, target);
865 if (ret)
866 return ret;
868 srp_map_update_start(state, sg, sg_index, dma_addr);
871 len = min_t(unsigned int, dma_len, dev->fmr_page_size);
873 if (!state->npages)
874 state->base_dma_addr = dma_addr;
875 state->pages[state->npages++] = dma_addr;
876 state->fmr_len += len;
877 dma_addr += len;
878 dma_len -= len;
881 /* If the last entry of the FMR wasn't a full page, then we need to
882 * close it out and start a new one -- we can only merge at page
883 * boundries.
885 ret = 0;
886 if (len != dev->fmr_page_size) {
887 ret = srp_map_finish_fmr(state, target);
888 if (!ret)
889 srp_map_update_start(state, NULL, 0, 0);
891 return ret;
894 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_target_port *target,
895 struct srp_request *req)
897 struct scatterlist *scat, *sg;
898 struct srp_cmd *cmd = req->cmd->buf;
899 int i, len, nents, count, use_fmr;
900 struct srp_device *dev;
901 struct ib_device *ibdev;
902 struct srp_map_state state;
903 struct srp_indirect_buf *indirect_hdr;
904 u32 table_len;
905 u8 fmt;
907 if (!scsi_sglist(scmnd) || scmnd->sc_data_direction == DMA_NONE)
908 return sizeof (struct srp_cmd);
910 if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
911 scmnd->sc_data_direction != DMA_TO_DEVICE) {
912 shost_printk(KERN_WARNING, target->scsi_host,
913 PFX "Unhandled data direction %d\n",
914 scmnd->sc_data_direction);
915 return -EINVAL;
918 nents = scsi_sg_count(scmnd);
919 scat = scsi_sglist(scmnd);
921 dev = target->srp_host->srp_dev;
922 ibdev = dev->dev;
924 count = ib_dma_map_sg(ibdev, scat, nents, scmnd->sc_data_direction);
925 if (unlikely(count == 0))
926 return -EIO;
928 fmt = SRP_DATA_DESC_DIRECT;
929 len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
931 if (count == 1) {
933 * The midlayer only generated a single gather/scatter
934 * entry, or DMA mapping coalesced everything to a
935 * single entry. So a direct descriptor along with
936 * the DMA MR suffices.
938 struct srp_direct_buf *buf = (void *) cmd->add_data;
940 buf->va = cpu_to_be64(ib_sg_dma_address(ibdev, scat));
941 buf->key = cpu_to_be32(target->rkey);
942 buf->len = cpu_to_be32(ib_sg_dma_len(ibdev, scat));
944 req->nfmr = 0;
945 goto map_complete;
948 /* We have more than one scatter/gather entry, so build our indirect
949 * descriptor table, trying to merge as many entries with FMR as we
950 * can.
952 indirect_hdr = (void *) cmd->add_data;
954 ib_dma_sync_single_for_cpu(ibdev, req->indirect_dma_addr,
955 target->indirect_size, DMA_TO_DEVICE);
957 memset(&state, 0, sizeof(state));
958 state.desc = req->indirect_desc;
959 state.pages = req->map_page;
960 state.next_fmr = req->fmr_list;
962 use_fmr = dev->fmr_pool ? SRP_MAP_ALLOW_FMR : SRP_MAP_NO_FMR;
964 for_each_sg(scat, sg, count, i) {
965 if (srp_map_sg_entry(&state, target, sg, i, use_fmr)) {
966 /* FMR mapping failed, so backtrack to the first
967 * unmapped entry and continue on without using FMR.
969 dma_addr_t dma_addr;
970 unsigned int dma_len;
972 backtrack:
973 sg = state.unmapped_sg;
974 i = state.unmapped_index;
976 dma_addr = ib_sg_dma_address(ibdev, sg);
977 dma_len = ib_sg_dma_len(ibdev, sg);
978 dma_len -= (state.unmapped_addr - dma_addr);
979 dma_addr = state.unmapped_addr;
980 use_fmr = SRP_MAP_NO_FMR;
981 srp_map_desc(&state, dma_addr, dma_len, target->rkey);
985 if (use_fmr == SRP_MAP_ALLOW_FMR && srp_map_finish_fmr(&state, target))
986 goto backtrack;
988 /* We've mapped the request, now pull as much of the indirect
989 * descriptor table as we can into the command buffer. If this
990 * target is not using an external indirect table, we are
991 * guaranteed to fit into the command, as the SCSI layer won't
992 * give us more S/G entries than we allow.
994 req->nfmr = state.nfmr;
995 if (state.ndesc == 1) {
996 /* FMR mapping was able to collapse this to one entry,
997 * so use a direct descriptor.
999 struct srp_direct_buf *buf = (void *) cmd->add_data;
1001 *buf = req->indirect_desc[0];
1002 goto map_complete;
1005 if (unlikely(target->cmd_sg_cnt < state.ndesc &&
1006 !target->allow_ext_sg)) {
1007 shost_printk(KERN_ERR, target->scsi_host,
1008 "Could not fit S/G list into SRP_CMD\n");
1009 return -EIO;
1012 count = min(state.ndesc, target->cmd_sg_cnt);
1013 table_len = state.ndesc * sizeof (struct srp_direct_buf);
1015 fmt = SRP_DATA_DESC_INDIRECT;
1016 len = sizeof(struct srp_cmd) + sizeof (struct srp_indirect_buf);
1017 len += count * sizeof (struct srp_direct_buf);
1019 memcpy(indirect_hdr->desc_list, req->indirect_desc,
1020 count * sizeof (struct srp_direct_buf));
1022 indirect_hdr->table_desc.va = cpu_to_be64(req->indirect_dma_addr);
1023 indirect_hdr->table_desc.key = cpu_to_be32(target->rkey);
1024 indirect_hdr->table_desc.len = cpu_to_be32(table_len);
1025 indirect_hdr->len = cpu_to_be32(state.total_len);
1027 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1028 cmd->data_out_desc_cnt = count;
1029 else
1030 cmd->data_in_desc_cnt = count;
1032 ib_dma_sync_single_for_device(ibdev, req->indirect_dma_addr, table_len,
1033 DMA_TO_DEVICE);
1035 map_complete:
1036 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1037 cmd->buf_fmt = fmt << 4;
1038 else
1039 cmd->buf_fmt = fmt;
1041 return len;
1045 * Return an IU and possible credit to the free pool
1047 static void srp_put_tx_iu(struct srp_target_port *target, struct srp_iu *iu,
1048 enum srp_iu_type iu_type)
1050 unsigned long flags;
1052 spin_lock_irqsave(&target->lock, flags);
1053 list_add(&iu->list, &target->free_tx);
1054 if (iu_type != SRP_IU_RSP)
1055 ++target->req_lim;
1056 spin_unlock_irqrestore(&target->lock, flags);
1060 * Must be called with target->lock held to protect req_lim and free_tx.
1061 * If IU is not sent, it must be returned using srp_put_tx_iu().
1063 * Note:
1064 * An upper limit for the number of allocated information units for each
1065 * request type is:
1066 * - SRP_IU_CMD: SRP_CMD_SQ_SIZE, since the SCSI mid-layer never queues
1067 * more than Scsi_Host.can_queue requests.
1068 * - SRP_IU_TSK_MGMT: SRP_TSK_MGMT_SQ_SIZE.
1069 * - SRP_IU_RSP: 1, since a conforming SRP target never sends more than
1070 * one unanswered SRP request to an initiator.
1072 static struct srp_iu *__srp_get_tx_iu(struct srp_target_port *target,
1073 enum srp_iu_type iu_type)
1075 s32 rsv = (iu_type == SRP_IU_TSK_MGMT) ? 0 : SRP_TSK_MGMT_SQ_SIZE;
1076 struct srp_iu *iu;
1078 srp_send_completion(target->send_cq, target);
1080 if (list_empty(&target->free_tx))
1081 return NULL;
1083 /* Initiator responses to target requests do not consume credits */
1084 if (iu_type != SRP_IU_RSP) {
1085 if (target->req_lim <= rsv) {
1086 ++target->zero_req_lim;
1087 return NULL;
1090 --target->req_lim;
1093 iu = list_first_entry(&target->free_tx, struct srp_iu, list);
1094 list_del(&iu->list);
1095 return iu;
1098 static int srp_post_send(struct srp_target_port *target,
1099 struct srp_iu *iu, int len)
1101 struct ib_sge list;
1102 struct ib_send_wr wr, *bad_wr;
1104 list.addr = iu->dma;
1105 list.length = len;
1106 list.lkey = target->lkey;
1108 wr.next = NULL;
1109 wr.wr_id = (uintptr_t) iu;
1110 wr.sg_list = &list;
1111 wr.num_sge = 1;
1112 wr.opcode = IB_WR_SEND;
1113 wr.send_flags = IB_SEND_SIGNALED;
1115 return ib_post_send(target->qp, &wr, &bad_wr);
1118 static int srp_post_recv(struct srp_target_port *target, struct srp_iu *iu)
1120 struct ib_recv_wr wr, *bad_wr;
1121 struct ib_sge list;
1123 list.addr = iu->dma;
1124 list.length = iu->size;
1125 list.lkey = target->lkey;
1127 wr.next = NULL;
1128 wr.wr_id = (uintptr_t) iu;
1129 wr.sg_list = &list;
1130 wr.num_sge = 1;
1132 return ib_post_recv(target->qp, &wr, &bad_wr);
1135 static void srp_process_rsp(struct srp_target_port *target, struct srp_rsp *rsp)
1137 struct srp_request *req;
1138 struct scsi_cmnd *scmnd;
1139 unsigned long flags;
1141 if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
1142 spin_lock_irqsave(&target->lock, flags);
1143 target->req_lim += be32_to_cpu(rsp->req_lim_delta);
1144 spin_unlock_irqrestore(&target->lock, flags);
1146 target->tsk_mgmt_status = -1;
1147 if (be32_to_cpu(rsp->resp_data_len) >= 4)
1148 target->tsk_mgmt_status = rsp->data[3];
1149 complete(&target->tsk_mgmt_done);
1150 } else {
1151 req = &target->req_ring[rsp->tag];
1152 scmnd = srp_claim_req(target, req, NULL);
1153 if (!scmnd) {
1154 shost_printk(KERN_ERR, target->scsi_host,
1155 "Null scmnd for RSP w/tag %016llx\n",
1156 (unsigned long long) rsp->tag);
1158 spin_lock_irqsave(&target->lock, flags);
1159 target->req_lim += be32_to_cpu(rsp->req_lim_delta);
1160 spin_unlock_irqrestore(&target->lock, flags);
1162 return;
1164 scmnd->result = rsp->status;
1166 if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
1167 memcpy(scmnd->sense_buffer, rsp->data +
1168 be32_to_cpu(rsp->resp_data_len),
1169 min_t(int, be32_to_cpu(rsp->sense_data_len),
1170 SCSI_SENSE_BUFFERSIZE));
1173 if (rsp->flags & (SRP_RSP_FLAG_DOOVER | SRP_RSP_FLAG_DOUNDER))
1174 scsi_set_resid(scmnd, be32_to_cpu(rsp->data_out_res_cnt));
1175 else if (rsp->flags & (SRP_RSP_FLAG_DIOVER | SRP_RSP_FLAG_DIUNDER))
1176 scsi_set_resid(scmnd, be32_to_cpu(rsp->data_in_res_cnt));
1178 srp_free_req(target, req, scmnd,
1179 be32_to_cpu(rsp->req_lim_delta));
1181 scmnd->host_scribble = NULL;
1182 scmnd->scsi_done(scmnd);
1186 static int srp_response_common(struct srp_target_port *target, s32 req_delta,
1187 void *rsp, int len)
1189 struct ib_device *dev = target->srp_host->srp_dev->dev;
1190 unsigned long flags;
1191 struct srp_iu *iu;
1192 int err;
1194 spin_lock_irqsave(&target->lock, flags);
1195 target->req_lim += req_delta;
1196 iu = __srp_get_tx_iu(target, SRP_IU_RSP);
1197 spin_unlock_irqrestore(&target->lock, flags);
1199 if (!iu) {
1200 shost_printk(KERN_ERR, target->scsi_host, PFX
1201 "no IU available to send response\n");
1202 return 1;
1205 ib_dma_sync_single_for_cpu(dev, iu->dma, len, DMA_TO_DEVICE);
1206 memcpy(iu->buf, rsp, len);
1207 ib_dma_sync_single_for_device(dev, iu->dma, len, DMA_TO_DEVICE);
1209 err = srp_post_send(target, iu, len);
1210 if (err) {
1211 shost_printk(KERN_ERR, target->scsi_host, PFX
1212 "unable to post response: %d\n", err);
1213 srp_put_tx_iu(target, iu, SRP_IU_RSP);
1216 return err;
1219 static void srp_process_cred_req(struct srp_target_port *target,
1220 struct srp_cred_req *req)
1222 struct srp_cred_rsp rsp = {
1223 .opcode = SRP_CRED_RSP,
1224 .tag = req->tag,
1226 s32 delta = be32_to_cpu(req->req_lim_delta);
1228 if (srp_response_common(target, delta, &rsp, sizeof rsp))
1229 shost_printk(KERN_ERR, target->scsi_host, PFX
1230 "problems processing SRP_CRED_REQ\n");
1233 static void srp_process_aer_req(struct srp_target_port *target,
1234 struct srp_aer_req *req)
1236 struct srp_aer_rsp rsp = {
1237 .opcode = SRP_AER_RSP,
1238 .tag = req->tag,
1240 s32 delta = be32_to_cpu(req->req_lim_delta);
1242 shost_printk(KERN_ERR, target->scsi_host, PFX
1243 "ignoring AER for LUN %llu\n", be64_to_cpu(req->lun));
1245 if (srp_response_common(target, delta, &rsp, sizeof rsp))
1246 shost_printk(KERN_ERR, target->scsi_host, PFX
1247 "problems processing SRP_AER_REQ\n");
1250 static void srp_handle_recv(struct srp_target_port *target, struct ib_wc *wc)
1252 struct ib_device *dev = target->srp_host->srp_dev->dev;
1253 struct srp_iu *iu = (struct srp_iu *) (uintptr_t) wc->wr_id;
1254 int res;
1255 u8 opcode;
1257 ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_ti_iu_len,
1258 DMA_FROM_DEVICE);
1260 opcode = *(u8 *) iu->buf;
1262 if (0) {
1263 shost_printk(KERN_ERR, target->scsi_host,
1264 PFX "recv completion, opcode 0x%02x\n", opcode);
1265 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 8, 1,
1266 iu->buf, wc->byte_len, true);
1269 switch (opcode) {
1270 case SRP_RSP:
1271 srp_process_rsp(target, iu->buf);
1272 break;
1274 case SRP_CRED_REQ:
1275 srp_process_cred_req(target, iu->buf);
1276 break;
1278 case SRP_AER_REQ:
1279 srp_process_aer_req(target, iu->buf);
1280 break;
1282 case SRP_T_LOGOUT:
1283 /* XXX Handle target logout */
1284 shost_printk(KERN_WARNING, target->scsi_host,
1285 PFX "Got target logout request\n");
1286 break;
1288 default:
1289 shost_printk(KERN_WARNING, target->scsi_host,
1290 PFX "Unhandled SRP opcode 0x%02x\n", opcode);
1291 break;
1294 ib_dma_sync_single_for_device(dev, iu->dma, target->max_ti_iu_len,
1295 DMA_FROM_DEVICE);
1297 res = srp_post_recv(target, iu);
1298 if (res != 0)
1299 shost_printk(KERN_ERR, target->scsi_host,
1300 PFX "Recv failed with error code %d\n", res);
1303 static void srp_handle_qp_err(enum ib_wc_status wc_status,
1304 enum ib_wc_opcode wc_opcode,
1305 struct srp_target_port *target)
1307 if (target->connected && !target->qp_in_error) {
1308 shost_printk(KERN_ERR, target->scsi_host,
1309 PFX "failed %s status %d\n",
1310 wc_opcode & IB_WC_RECV ? "receive" : "send",
1311 wc_status);
1313 target->qp_in_error = true;
1316 static void srp_recv_completion(struct ib_cq *cq, void *target_ptr)
1318 struct srp_target_port *target = target_ptr;
1319 struct ib_wc wc;
1321 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1322 while (ib_poll_cq(cq, 1, &wc) > 0) {
1323 if (likely(wc.status == IB_WC_SUCCESS)) {
1324 srp_handle_recv(target, &wc);
1325 } else {
1326 srp_handle_qp_err(wc.status, wc.opcode, target);
1331 static void srp_send_completion(struct ib_cq *cq, void *target_ptr)
1333 struct srp_target_port *target = target_ptr;
1334 struct ib_wc wc;
1335 struct srp_iu *iu;
1337 while (ib_poll_cq(cq, 1, &wc) > 0) {
1338 if (likely(wc.status == IB_WC_SUCCESS)) {
1339 iu = (struct srp_iu *) (uintptr_t) wc.wr_id;
1340 list_add(&iu->list, &target->free_tx);
1341 } else {
1342 srp_handle_qp_err(wc.status, wc.opcode, target);
1347 static int srp_queuecommand(struct Scsi_Host *shost, struct scsi_cmnd *scmnd)
1349 struct srp_target_port *target = host_to_target(shost);
1350 struct srp_request *req;
1351 struct srp_iu *iu;
1352 struct srp_cmd *cmd;
1353 struct ib_device *dev;
1354 unsigned long flags;
1355 int len;
1357 if (unlikely(target->transport_offline)) {
1358 scmnd->result = DID_NO_CONNECT << 16;
1359 scmnd->scsi_done(scmnd);
1360 return 0;
1363 spin_lock_irqsave(&target->lock, flags);
1364 iu = __srp_get_tx_iu(target, SRP_IU_CMD);
1365 if (!iu)
1366 goto err_unlock;
1368 req = list_first_entry(&target->free_reqs, struct srp_request, list);
1369 list_del(&req->list);
1370 spin_unlock_irqrestore(&target->lock, flags);
1372 dev = target->srp_host->srp_dev->dev;
1373 ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_iu_len,
1374 DMA_TO_DEVICE);
1376 scmnd->result = 0;
1377 scmnd->host_scribble = (void *) req;
1379 cmd = iu->buf;
1380 memset(cmd, 0, sizeof *cmd);
1382 cmd->opcode = SRP_CMD;
1383 cmd->lun = cpu_to_be64((u64) scmnd->device->lun << 48);
1384 cmd->tag = req->index;
1385 memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
1387 req->scmnd = scmnd;
1388 req->cmd = iu;
1390 len = srp_map_data(scmnd, target, req);
1391 if (len < 0) {
1392 shost_printk(KERN_ERR, target->scsi_host,
1393 PFX "Failed to map data\n");
1394 goto err_iu;
1397 ib_dma_sync_single_for_device(dev, iu->dma, target->max_iu_len,
1398 DMA_TO_DEVICE);
1400 if (srp_post_send(target, iu, len)) {
1401 shost_printk(KERN_ERR, target->scsi_host, PFX "Send failed\n");
1402 goto err_unmap;
1405 return 0;
1407 err_unmap:
1408 srp_unmap_data(scmnd, target, req);
1410 err_iu:
1411 srp_put_tx_iu(target, iu, SRP_IU_CMD);
1413 spin_lock_irqsave(&target->lock, flags);
1414 list_add(&req->list, &target->free_reqs);
1416 err_unlock:
1417 spin_unlock_irqrestore(&target->lock, flags);
1419 return SCSI_MLQUEUE_HOST_BUSY;
1422 static int srp_alloc_iu_bufs(struct srp_target_port *target)
1424 int i;
1426 for (i = 0; i < SRP_RQ_SIZE; ++i) {
1427 target->rx_ring[i] = srp_alloc_iu(target->srp_host,
1428 target->max_ti_iu_len,
1429 GFP_KERNEL, DMA_FROM_DEVICE);
1430 if (!target->rx_ring[i])
1431 goto err;
1434 for (i = 0; i < SRP_SQ_SIZE; ++i) {
1435 target->tx_ring[i] = srp_alloc_iu(target->srp_host,
1436 target->max_iu_len,
1437 GFP_KERNEL, DMA_TO_DEVICE);
1438 if (!target->tx_ring[i])
1439 goto err;
1441 list_add(&target->tx_ring[i]->list, &target->free_tx);
1444 return 0;
1446 err:
1447 for (i = 0; i < SRP_RQ_SIZE; ++i) {
1448 srp_free_iu(target->srp_host, target->rx_ring[i]);
1449 target->rx_ring[i] = NULL;
1452 for (i = 0; i < SRP_SQ_SIZE; ++i) {
1453 srp_free_iu(target->srp_host, target->tx_ring[i]);
1454 target->tx_ring[i] = NULL;
1457 return -ENOMEM;
1460 static uint32_t srp_compute_rq_tmo(struct ib_qp_attr *qp_attr, int attr_mask)
1462 uint64_t T_tr_ns, max_compl_time_ms;
1463 uint32_t rq_tmo_jiffies;
1466 * According to section 11.2.4.2 in the IBTA spec (Modify Queue Pair,
1467 * table 91), both the QP timeout and the retry count have to be set
1468 * for RC QP's during the RTR to RTS transition.
1470 WARN_ON_ONCE((attr_mask & (IB_QP_TIMEOUT | IB_QP_RETRY_CNT)) !=
1471 (IB_QP_TIMEOUT | IB_QP_RETRY_CNT));
1474 * Set target->rq_tmo_jiffies to one second more than the largest time
1475 * it can take before an error completion is generated. See also
1476 * C9-140..142 in the IBTA spec for more information about how to
1477 * convert the QP Local ACK Timeout value to nanoseconds.
1479 T_tr_ns = 4096 * (1ULL << qp_attr->timeout);
1480 max_compl_time_ms = qp_attr->retry_cnt * 4 * T_tr_ns;
1481 do_div(max_compl_time_ms, NSEC_PER_MSEC);
1482 rq_tmo_jiffies = msecs_to_jiffies(max_compl_time_ms + 1000);
1484 return rq_tmo_jiffies;
1487 static void srp_cm_rep_handler(struct ib_cm_id *cm_id,
1488 struct srp_login_rsp *lrsp,
1489 struct srp_target_port *target)
1491 struct ib_qp_attr *qp_attr = NULL;
1492 int attr_mask = 0;
1493 int ret;
1494 int i;
1496 if (lrsp->opcode == SRP_LOGIN_RSP) {
1497 target->max_ti_iu_len = be32_to_cpu(lrsp->max_ti_iu_len);
1498 target->req_lim = be32_to_cpu(lrsp->req_lim_delta);
1501 * Reserve credits for task management so we don't
1502 * bounce requests back to the SCSI mid-layer.
1504 target->scsi_host->can_queue
1505 = min(target->req_lim - SRP_TSK_MGMT_SQ_SIZE,
1506 target->scsi_host->can_queue);
1507 } else {
1508 shost_printk(KERN_WARNING, target->scsi_host,
1509 PFX "Unhandled RSP opcode %#x\n", lrsp->opcode);
1510 ret = -ECONNRESET;
1511 goto error;
1514 if (!target->rx_ring[0]) {
1515 ret = srp_alloc_iu_bufs(target);
1516 if (ret)
1517 goto error;
1520 ret = -ENOMEM;
1521 qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
1522 if (!qp_attr)
1523 goto error;
1525 qp_attr->qp_state = IB_QPS_RTR;
1526 ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1527 if (ret)
1528 goto error_free;
1530 ret = ib_modify_qp(target->qp, qp_attr, attr_mask);
1531 if (ret)
1532 goto error_free;
1534 for (i = 0; i < SRP_RQ_SIZE; i++) {
1535 struct srp_iu *iu = target->rx_ring[i];
1536 ret = srp_post_recv(target, iu);
1537 if (ret)
1538 goto error_free;
1541 qp_attr->qp_state = IB_QPS_RTS;
1542 ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1543 if (ret)
1544 goto error_free;
1546 target->rq_tmo_jiffies = srp_compute_rq_tmo(qp_attr, attr_mask);
1548 ret = ib_modify_qp(target->qp, qp_attr, attr_mask);
1549 if (ret)
1550 goto error_free;
1552 ret = ib_send_cm_rtu(cm_id, NULL, 0);
1554 error_free:
1555 kfree(qp_attr);
1557 error:
1558 target->status = ret;
1561 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
1562 struct ib_cm_event *event,
1563 struct srp_target_port *target)
1565 struct Scsi_Host *shost = target->scsi_host;
1566 struct ib_class_port_info *cpi;
1567 int opcode;
1569 switch (event->param.rej_rcvd.reason) {
1570 case IB_CM_REJ_PORT_CM_REDIRECT:
1571 cpi = event->param.rej_rcvd.ari;
1572 target->path.dlid = cpi->redirect_lid;
1573 target->path.pkey = cpi->redirect_pkey;
1574 cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
1575 memcpy(target->path.dgid.raw, cpi->redirect_gid, 16);
1577 target->status = target->path.dlid ?
1578 SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
1579 break;
1581 case IB_CM_REJ_PORT_REDIRECT:
1582 if (srp_target_is_topspin(target)) {
1584 * Topspin/Cisco SRP gateways incorrectly send
1585 * reject reason code 25 when they mean 24
1586 * (port redirect).
1588 memcpy(target->path.dgid.raw,
1589 event->param.rej_rcvd.ari, 16);
1591 shost_printk(KERN_DEBUG, shost,
1592 PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
1593 (unsigned long long) be64_to_cpu(target->path.dgid.global.subnet_prefix),
1594 (unsigned long long) be64_to_cpu(target->path.dgid.global.interface_id));
1596 target->status = SRP_PORT_REDIRECT;
1597 } else {
1598 shost_printk(KERN_WARNING, shost,
1599 " REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
1600 target->status = -ECONNRESET;
1602 break;
1604 case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
1605 shost_printk(KERN_WARNING, shost,
1606 " REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
1607 target->status = -ECONNRESET;
1608 break;
1610 case IB_CM_REJ_CONSUMER_DEFINED:
1611 opcode = *(u8 *) event->private_data;
1612 if (opcode == SRP_LOGIN_REJ) {
1613 struct srp_login_rej *rej = event->private_data;
1614 u32 reason = be32_to_cpu(rej->reason);
1616 if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
1617 shost_printk(KERN_WARNING, shost,
1618 PFX "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
1619 else
1620 shost_printk(KERN_WARNING, shost,
1621 PFX "SRP LOGIN REJECTED, reason 0x%08x\n", reason);
1622 } else
1623 shost_printk(KERN_WARNING, shost,
1624 " REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
1625 " opcode 0x%02x\n", opcode);
1626 target->status = -ECONNRESET;
1627 break;
1629 case IB_CM_REJ_STALE_CONN:
1630 shost_printk(KERN_WARNING, shost, " REJ reason: stale connection\n");
1631 target->status = SRP_STALE_CONN;
1632 break;
1634 default:
1635 shost_printk(KERN_WARNING, shost, " REJ reason 0x%x\n",
1636 event->param.rej_rcvd.reason);
1637 target->status = -ECONNRESET;
1641 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
1643 struct srp_target_port *target = cm_id->context;
1644 int comp = 0;
1646 switch (event->event) {
1647 case IB_CM_REQ_ERROR:
1648 shost_printk(KERN_DEBUG, target->scsi_host,
1649 PFX "Sending CM REQ failed\n");
1650 comp = 1;
1651 target->status = -ECONNRESET;
1652 break;
1654 case IB_CM_REP_RECEIVED:
1655 comp = 1;
1656 srp_cm_rep_handler(cm_id, event->private_data, target);
1657 break;
1659 case IB_CM_REJ_RECEIVED:
1660 shost_printk(KERN_DEBUG, target->scsi_host, PFX "REJ received\n");
1661 comp = 1;
1663 srp_cm_rej_handler(cm_id, event, target);
1664 break;
1666 case IB_CM_DREQ_RECEIVED:
1667 shost_printk(KERN_WARNING, target->scsi_host,
1668 PFX "DREQ received - connection closed\n");
1669 srp_change_conn_state(target, false);
1670 if (ib_send_cm_drep(cm_id, NULL, 0))
1671 shost_printk(KERN_ERR, target->scsi_host,
1672 PFX "Sending CM DREP failed\n");
1673 break;
1675 case IB_CM_TIMEWAIT_EXIT:
1676 shost_printk(KERN_ERR, target->scsi_host,
1677 PFX "connection closed\n");
1679 target->status = 0;
1680 break;
1682 case IB_CM_MRA_RECEIVED:
1683 case IB_CM_DREQ_ERROR:
1684 case IB_CM_DREP_RECEIVED:
1685 break;
1687 default:
1688 shost_printk(KERN_WARNING, target->scsi_host,
1689 PFX "Unhandled CM event %d\n", event->event);
1690 break;
1693 if (comp)
1694 complete(&target->done);
1696 return 0;
1699 static int srp_send_tsk_mgmt(struct srp_target_port *target,
1700 u64 req_tag, unsigned int lun, u8 func)
1702 struct ib_device *dev = target->srp_host->srp_dev->dev;
1703 struct srp_iu *iu;
1704 struct srp_tsk_mgmt *tsk_mgmt;
1706 if (!target->connected || target->qp_in_error)
1707 return -1;
1709 init_completion(&target->tsk_mgmt_done);
1711 spin_lock_irq(&target->lock);
1712 iu = __srp_get_tx_iu(target, SRP_IU_TSK_MGMT);
1713 spin_unlock_irq(&target->lock);
1715 if (!iu)
1716 return -1;
1718 ib_dma_sync_single_for_cpu(dev, iu->dma, sizeof *tsk_mgmt,
1719 DMA_TO_DEVICE);
1720 tsk_mgmt = iu->buf;
1721 memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
1723 tsk_mgmt->opcode = SRP_TSK_MGMT;
1724 tsk_mgmt->lun = cpu_to_be64((u64) lun << 48);
1725 tsk_mgmt->tag = req_tag | SRP_TAG_TSK_MGMT;
1726 tsk_mgmt->tsk_mgmt_func = func;
1727 tsk_mgmt->task_tag = req_tag;
1729 ib_dma_sync_single_for_device(dev, iu->dma, sizeof *tsk_mgmt,
1730 DMA_TO_DEVICE);
1731 if (srp_post_send(target, iu, sizeof *tsk_mgmt)) {
1732 srp_put_tx_iu(target, iu, SRP_IU_TSK_MGMT);
1733 return -1;
1736 if (!wait_for_completion_timeout(&target->tsk_mgmt_done,
1737 msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
1738 return -1;
1740 return 0;
1743 static int srp_abort(struct scsi_cmnd *scmnd)
1745 struct srp_target_port *target = host_to_target(scmnd->device->host);
1746 struct srp_request *req = (struct srp_request *) scmnd->host_scribble;
1747 int ret;
1749 shost_printk(KERN_ERR, target->scsi_host, "SRP abort called\n");
1751 if (!req || !srp_claim_req(target, req, scmnd))
1752 return FAILED;
1753 if (srp_send_tsk_mgmt(target, req->index, scmnd->device->lun,
1754 SRP_TSK_ABORT_TASK) == 0 ||
1755 target->transport_offline)
1756 ret = SUCCESS;
1757 else if (target->transport_offline)
1758 ret = FAST_IO_FAIL;
1759 else
1760 ret = FAILED;
1761 srp_free_req(target, req, scmnd, 0);
1762 scmnd->result = DID_ABORT << 16;
1763 scmnd->scsi_done(scmnd);
1765 return ret;
1768 static int srp_reset_device(struct scsi_cmnd *scmnd)
1770 struct srp_target_port *target = host_to_target(scmnd->device->host);
1771 int i;
1773 shost_printk(KERN_ERR, target->scsi_host, "SRP reset_device called\n");
1775 if (srp_send_tsk_mgmt(target, SRP_TAG_NO_REQ, scmnd->device->lun,
1776 SRP_TSK_LUN_RESET))
1777 return FAILED;
1778 if (target->tsk_mgmt_status)
1779 return FAILED;
1781 for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
1782 struct srp_request *req = &target->req_ring[i];
1783 if (req->scmnd && req->scmnd->device == scmnd->device)
1784 srp_reset_req(target, req);
1787 return SUCCESS;
1790 static int srp_reset_host(struct scsi_cmnd *scmnd)
1792 struct srp_target_port *target = host_to_target(scmnd->device->host);
1793 int ret = FAILED;
1795 shost_printk(KERN_ERR, target->scsi_host, PFX "SRP reset_host called\n");
1797 if (!srp_reconnect_target(target))
1798 ret = SUCCESS;
1800 return ret;
1803 static int srp_slave_configure(struct scsi_device *sdev)
1805 struct Scsi_Host *shost = sdev->host;
1806 struct srp_target_port *target = host_to_target(shost);
1807 struct request_queue *q = sdev->request_queue;
1808 unsigned long timeout;
1810 if (sdev->type == TYPE_DISK) {
1811 timeout = max_t(unsigned, 30 * HZ, target->rq_tmo_jiffies);
1812 blk_queue_rq_timeout(q, timeout);
1815 return 0;
1818 static ssize_t show_id_ext(struct device *dev, struct device_attribute *attr,
1819 char *buf)
1821 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1823 return sprintf(buf, "0x%016llx\n",
1824 (unsigned long long) be64_to_cpu(target->id_ext));
1827 static ssize_t show_ioc_guid(struct device *dev, struct device_attribute *attr,
1828 char *buf)
1830 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1832 return sprintf(buf, "0x%016llx\n",
1833 (unsigned long long) be64_to_cpu(target->ioc_guid));
1836 static ssize_t show_service_id(struct device *dev,
1837 struct device_attribute *attr, char *buf)
1839 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1841 return sprintf(buf, "0x%016llx\n",
1842 (unsigned long long) be64_to_cpu(target->service_id));
1845 static ssize_t show_pkey(struct device *dev, struct device_attribute *attr,
1846 char *buf)
1848 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1850 return sprintf(buf, "0x%04x\n", be16_to_cpu(target->path.pkey));
1853 static ssize_t show_dgid(struct device *dev, struct device_attribute *attr,
1854 char *buf)
1856 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1858 return sprintf(buf, "%pI6\n", target->path.dgid.raw);
1861 static ssize_t show_orig_dgid(struct device *dev,
1862 struct device_attribute *attr, char *buf)
1864 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1866 return sprintf(buf, "%pI6\n", target->orig_dgid);
1869 static ssize_t show_req_lim(struct device *dev,
1870 struct device_attribute *attr, char *buf)
1872 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1874 return sprintf(buf, "%d\n", target->req_lim);
1877 static ssize_t show_zero_req_lim(struct device *dev,
1878 struct device_attribute *attr, char *buf)
1880 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1882 return sprintf(buf, "%d\n", target->zero_req_lim);
1885 static ssize_t show_local_ib_port(struct device *dev,
1886 struct device_attribute *attr, char *buf)
1888 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1890 return sprintf(buf, "%d\n", target->srp_host->port);
1893 static ssize_t show_local_ib_device(struct device *dev,
1894 struct device_attribute *attr, char *buf)
1896 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1898 return sprintf(buf, "%s\n", target->srp_host->srp_dev->dev->name);
1901 static ssize_t show_cmd_sg_entries(struct device *dev,
1902 struct device_attribute *attr, char *buf)
1904 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1906 return sprintf(buf, "%u\n", target->cmd_sg_cnt);
1909 static ssize_t show_allow_ext_sg(struct device *dev,
1910 struct device_attribute *attr, char *buf)
1912 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1914 return sprintf(buf, "%s\n", target->allow_ext_sg ? "true" : "false");
1917 static DEVICE_ATTR(id_ext, S_IRUGO, show_id_ext, NULL);
1918 static DEVICE_ATTR(ioc_guid, S_IRUGO, show_ioc_guid, NULL);
1919 static DEVICE_ATTR(service_id, S_IRUGO, show_service_id, NULL);
1920 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1921 static DEVICE_ATTR(dgid, S_IRUGO, show_dgid, NULL);
1922 static DEVICE_ATTR(orig_dgid, S_IRUGO, show_orig_dgid, NULL);
1923 static DEVICE_ATTR(req_lim, S_IRUGO, show_req_lim, NULL);
1924 static DEVICE_ATTR(zero_req_lim, S_IRUGO, show_zero_req_lim, NULL);
1925 static DEVICE_ATTR(local_ib_port, S_IRUGO, show_local_ib_port, NULL);
1926 static DEVICE_ATTR(local_ib_device, S_IRUGO, show_local_ib_device, NULL);
1927 static DEVICE_ATTR(cmd_sg_entries, S_IRUGO, show_cmd_sg_entries, NULL);
1928 static DEVICE_ATTR(allow_ext_sg, S_IRUGO, show_allow_ext_sg, NULL);
1930 static struct device_attribute *srp_host_attrs[] = {
1931 &dev_attr_id_ext,
1932 &dev_attr_ioc_guid,
1933 &dev_attr_service_id,
1934 &dev_attr_pkey,
1935 &dev_attr_dgid,
1936 &dev_attr_orig_dgid,
1937 &dev_attr_req_lim,
1938 &dev_attr_zero_req_lim,
1939 &dev_attr_local_ib_port,
1940 &dev_attr_local_ib_device,
1941 &dev_attr_cmd_sg_entries,
1942 &dev_attr_allow_ext_sg,
1943 NULL
1946 static struct scsi_host_template srp_template = {
1947 .module = THIS_MODULE,
1948 .name = "InfiniBand SRP initiator",
1949 .proc_name = DRV_NAME,
1950 .slave_configure = srp_slave_configure,
1951 .info = srp_target_info,
1952 .queuecommand = srp_queuecommand,
1953 .eh_abort_handler = srp_abort,
1954 .eh_device_reset_handler = srp_reset_device,
1955 .eh_host_reset_handler = srp_reset_host,
1956 .skip_settle_delay = true,
1957 .sg_tablesize = SRP_DEF_SG_TABLESIZE,
1958 .can_queue = SRP_CMD_SQ_SIZE,
1959 .this_id = -1,
1960 .cmd_per_lun = SRP_CMD_SQ_SIZE,
1961 .use_clustering = ENABLE_CLUSTERING,
1962 .shost_attrs = srp_host_attrs
1965 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
1967 struct srp_rport_identifiers ids;
1968 struct srp_rport *rport;
1970 sprintf(target->target_name, "SRP.T10:%016llX",
1971 (unsigned long long) be64_to_cpu(target->id_ext));
1973 if (scsi_add_host(target->scsi_host, host->srp_dev->dev->dma_device))
1974 return -ENODEV;
1976 memcpy(ids.port_id, &target->id_ext, 8);
1977 memcpy(ids.port_id + 8, &target->ioc_guid, 8);
1978 ids.roles = SRP_RPORT_ROLE_TARGET;
1979 rport = srp_rport_add(target->scsi_host, &ids);
1980 if (IS_ERR(rport)) {
1981 scsi_remove_host(target->scsi_host);
1982 return PTR_ERR(rport);
1985 rport->lld_data = target;
1987 spin_lock(&host->target_lock);
1988 list_add_tail(&target->list, &host->target_list);
1989 spin_unlock(&host->target_lock);
1991 target->state = SRP_TARGET_LIVE;
1993 scsi_scan_target(&target->scsi_host->shost_gendev,
1994 0, target->scsi_id, SCAN_WILD_CARD, 0);
1996 return 0;
1999 static void srp_release_dev(struct device *dev)
2001 struct srp_host *host =
2002 container_of(dev, struct srp_host, dev);
2004 complete(&host->released);
2007 static struct class srp_class = {
2008 .name = "infiniband_srp",
2009 .dev_release = srp_release_dev
2013 * srp_conn_unique() - check whether the connection to a target is unique
2015 static bool srp_conn_unique(struct srp_host *host,
2016 struct srp_target_port *target)
2018 struct srp_target_port *t;
2019 bool ret = false;
2021 if (target->state == SRP_TARGET_REMOVED)
2022 goto out;
2024 ret = true;
2026 spin_lock(&host->target_lock);
2027 list_for_each_entry(t, &host->target_list, list) {
2028 if (t != target &&
2029 target->id_ext == t->id_ext &&
2030 target->ioc_guid == t->ioc_guid &&
2031 target->initiator_ext == t->initiator_ext) {
2032 ret = false;
2033 break;
2036 spin_unlock(&host->target_lock);
2038 out:
2039 return ret;
2043 * Target ports are added by writing
2045 * id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
2046 * pkey=<P_Key>,service_id=<service ID>
2048 * to the add_target sysfs attribute.
2050 enum {
2051 SRP_OPT_ERR = 0,
2052 SRP_OPT_ID_EXT = 1 << 0,
2053 SRP_OPT_IOC_GUID = 1 << 1,
2054 SRP_OPT_DGID = 1 << 2,
2055 SRP_OPT_PKEY = 1 << 3,
2056 SRP_OPT_SERVICE_ID = 1 << 4,
2057 SRP_OPT_MAX_SECT = 1 << 5,
2058 SRP_OPT_MAX_CMD_PER_LUN = 1 << 6,
2059 SRP_OPT_IO_CLASS = 1 << 7,
2060 SRP_OPT_INITIATOR_EXT = 1 << 8,
2061 SRP_OPT_CMD_SG_ENTRIES = 1 << 9,
2062 SRP_OPT_ALLOW_EXT_SG = 1 << 10,
2063 SRP_OPT_SG_TABLESIZE = 1 << 11,
2064 SRP_OPT_ALL = (SRP_OPT_ID_EXT |
2065 SRP_OPT_IOC_GUID |
2066 SRP_OPT_DGID |
2067 SRP_OPT_PKEY |
2068 SRP_OPT_SERVICE_ID),
2071 static const match_table_t srp_opt_tokens = {
2072 { SRP_OPT_ID_EXT, "id_ext=%s" },
2073 { SRP_OPT_IOC_GUID, "ioc_guid=%s" },
2074 { SRP_OPT_DGID, "dgid=%s" },
2075 { SRP_OPT_PKEY, "pkey=%x" },
2076 { SRP_OPT_SERVICE_ID, "service_id=%s" },
2077 { SRP_OPT_MAX_SECT, "max_sect=%d" },
2078 { SRP_OPT_MAX_CMD_PER_LUN, "max_cmd_per_lun=%d" },
2079 { SRP_OPT_IO_CLASS, "io_class=%x" },
2080 { SRP_OPT_INITIATOR_EXT, "initiator_ext=%s" },
2081 { SRP_OPT_CMD_SG_ENTRIES, "cmd_sg_entries=%u" },
2082 { SRP_OPT_ALLOW_EXT_SG, "allow_ext_sg=%u" },
2083 { SRP_OPT_SG_TABLESIZE, "sg_tablesize=%u" },
2084 { SRP_OPT_ERR, NULL }
2087 static int srp_parse_options(const char *buf, struct srp_target_port *target)
2089 char *options, *sep_opt;
2090 char *p;
2091 char dgid[3];
2092 substring_t args[MAX_OPT_ARGS];
2093 int opt_mask = 0;
2094 int token;
2095 int ret = -EINVAL;
2096 int i;
2098 options = kstrdup(buf, GFP_KERNEL);
2099 if (!options)
2100 return -ENOMEM;
2102 sep_opt = options;
2103 while ((p = strsep(&sep_opt, ",")) != NULL) {
2104 if (!*p)
2105 continue;
2107 token = match_token(p, srp_opt_tokens, args);
2108 opt_mask |= token;
2110 switch (token) {
2111 case SRP_OPT_ID_EXT:
2112 p = match_strdup(args);
2113 if (!p) {
2114 ret = -ENOMEM;
2115 goto out;
2117 target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2118 kfree(p);
2119 break;
2121 case SRP_OPT_IOC_GUID:
2122 p = match_strdup(args);
2123 if (!p) {
2124 ret = -ENOMEM;
2125 goto out;
2127 target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
2128 kfree(p);
2129 break;
2131 case SRP_OPT_DGID:
2132 p = match_strdup(args);
2133 if (!p) {
2134 ret = -ENOMEM;
2135 goto out;
2137 if (strlen(p) != 32) {
2138 pr_warn("bad dest GID parameter '%s'\n", p);
2139 kfree(p);
2140 goto out;
2143 for (i = 0; i < 16; ++i) {
2144 strlcpy(dgid, p + i * 2, 3);
2145 target->path.dgid.raw[i] = simple_strtoul(dgid, NULL, 16);
2147 kfree(p);
2148 memcpy(target->orig_dgid, target->path.dgid.raw, 16);
2149 break;
2151 case SRP_OPT_PKEY:
2152 if (match_hex(args, &token)) {
2153 pr_warn("bad P_Key parameter '%s'\n", p);
2154 goto out;
2156 target->path.pkey = cpu_to_be16(token);
2157 break;
2159 case SRP_OPT_SERVICE_ID:
2160 p = match_strdup(args);
2161 if (!p) {
2162 ret = -ENOMEM;
2163 goto out;
2165 target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
2166 target->path.service_id = target->service_id;
2167 kfree(p);
2168 break;
2170 case SRP_OPT_MAX_SECT:
2171 if (match_int(args, &token)) {
2172 pr_warn("bad max sect parameter '%s'\n", p);
2173 goto out;
2175 target->scsi_host->max_sectors = token;
2176 break;
2178 case SRP_OPT_MAX_CMD_PER_LUN:
2179 if (match_int(args, &token)) {
2180 pr_warn("bad max cmd_per_lun parameter '%s'\n",
2182 goto out;
2184 target->scsi_host->cmd_per_lun = min(token, SRP_CMD_SQ_SIZE);
2185 break;
2187 case SRP_OPT_IO_CLASS:
2188 if (match_hex(args, &token)) {
2189 pr_warn("bad IO class parameter '%s'\n", p);
2190 goto out;
2192 if (token != SRP_REV10_IB_IO_CLASS &&
2193 token != SRP_REV16A_IB_IO_CLASS) {
2194 pr_warn("unknown IO class parameter value %x specified (use %x or %x).\n",
2195 token, SRP_REV10_IB_IO_CLASS,
2196 SRP_REV16A_IB_IO_CLASS);
2197 goto out;
2199 target->io_class = token;
2200 break;
2202 case SRP_OPT_INITIATOR_EXT:
2203 p = match_strdup(args);
2204 if (!p) {
2205 ret = -ENOMEM;
2206 goto out;
2208 target->initiator_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2209 kfree(p);
2210 break;
2212 case SRP_OPT_CMD_SG_ENTRIES:
2213 if (match_int(args, &token) || token < 1 || token > 255) {
2214 pr_warn("bad max cmd_sg_entries parameter '%s'\n",
2216 goto out;
2218 target->cmd_sg_cnt = token;
2219 break;
2221 case SRP_OPT_ALLOW_EXT_SG:
2222 if (match_int(args, &token)) {
2223 pr_warn("bad allow_ext_sg parameter '%s'\n", p);
2224 goto out;
2226 target->allow_ext_sg = !!token;
2227 break;
2229 case SRP_OPT_SG_TABLESIZE:
2230 if (match_int(args, &token) || token < 1 ||
2231 token > SCSI_MAX_SG_CHAIN_SEGMENTS) {
2232 pr_warn("bad max sg_tablesize parameter '%s'\n",
2234 goto out;
2236 target->sg_tablesize = token;
2237 break;
2239 default:
2240 pr_warn("unknown parameter or missing value '%s' in target creation request\n",
2242 goto out;
2246 if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
2247 ret = 0;
2248 else
2249 for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
2250 if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
2251 !(srp_opt_tokens[i].token & opt_mask))
2252 pr_warn("target creation request is missing parameter '%s'\n",
2253 srp_opt_tokens[i].pattern);
2255 out:
2256 kfree(options);
2257 return ret;
2260 static ssize_t srp_create_target(struct device *dev,
2261 struct device_attribute *attr,
2262 const char *buf, size_t count)
2264 struct srp_host *host =
2265 container_of(dev, struct srp_host, dev);
2266 struct Scsi_Host *target_host;
2267 struct srp_target_port *target;
2268 struct ib_device *ibdev = host->srp_dev->dev;
2269 dma_addr_t dma_addr;
2270 int i, ret;
2272 target_host = scsi_host_alloc(&srp_template,
2273 sizeof (struct srp_target_port));
2274 if (!target_host)
2275 return -ENOMEM;
2277 target_host->transportt = ib_srp_transport_template;
2278 target_host->max_channel = 0;
2279 target_host->max_id = 1;
2280 target_host->max_lun = SRP_MAX_LUN;
2281 target_host->max_cmd_len = sizeof ((struct srp_cmd *) (void *) 0L)->cdb;
2283 target = host_to_target(target_host);
2285 target->io_class = SRP_REV16A_IB_IO_CLASS;
2286 target->scsi_host = target_host;
2287 target->srp_host = host;
2288 target->lkey = host->srp_dev->mr->lkey;
2289 target->rkey = host->srp_dev->mr->rkey;
2290 target->cmd_sg_cnt = cmd_sg_entries;
2291 target->sg_tablesize = indirect_sg_entries ? : cmd_sg_entries;
2292 target->allow_ext_sg = allow_ext_sg;
2294 ret = srp_parse_options(buf, target);
2295 if (ret)
2296 goto err;
2298 if (!srp_conn_unique(target->srp_host, target)) {
2299 shost_printk(KERN_INFO, target->scsi_host,
2300 PFX "Already connected to target port with id_ext=%016llx;ioc_guid=%016llx;initiator_ext=%016llx\n",
2301 be64_to_cpu(target->id_ext),
2302 be64_to_cpu(target->ioc_guid),
2303 be64_to_cpu(target->initiator_ext));
2304 ret = -EEXIST;
2305 goto err;
2308 if (!host->srp_dev->fmr_pool && !target->allow_ext_sg &&
2309 target->cmd_sg_cnt < target->sg_tablesize) {
2310 pr_warn("No FMR pool and no external indirect descriptors, limiting sg_tablesize to cmd_sg_cnt\n");
2311 target->sg_tablesize = target->cmd_sg_cnt;
2314 target_host->sg_tablesize = target->sg_tablesize;
2315 target->indirect_size = target->sg_tablesize *
2316 sizeof (struct srp_direct_buf);
2317 target->max_iu_len = sizeof (struct srp_cmd) +
2318 sizeof (struct srp_indirect_buf) +
2319 target->cmd_sg_cnt * sizeof (struct srp_direct_buf);
2321 INIT_WORK(&target->remove_work, srp_remove_work);
2322 spin_lock_init(&target->lock);
2323 INIT_LIST_HEAD(&target->free_tx);
2324 INIT_LIST_HEAD(&target->free_reqs);
2325 for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
2326 struct srp_request *req = &target->req_ring[i];
2328 req->fmr_list = kmalloc(target->cmd_sg_cnt * sizeof (void *),
2329 GFP_KERNEL);
2330 req->map_page = kmalloc(SRP_FMR_SIZE * sizeof (void *),
2331 GFP_KERNEL);
2332 req->indirect_desc = kmalloc(target->indirect_size, GFP_KERNEL);
2333 if (!req->fmr_list || !req->map_page || !req->indirect_desc)
2334 goto err_free_mem;
2336 dma_addr = ib_dma_map_single(ibdev, req->indirect_desc,
2337 target->indirect_size,
2338 DMA_TO_DEVICE);
2339 if (ib_dma_mapping_error(ibdev, dma_addr))
2340 goto err_free_mem;
2342 req->indirect_dma_addr = dma_addr;
2343 req->index = i;
2344 list_add_tail(&req->list, &target->free_reqs);
2347 ib_query_gid(ibdev, host->port, 0, &target->path.sgid);
2349 shost_printk(KERN_DEBUG, target->scsi_host, PFX
2350 "new target: id_ext %016llx ioc_guid %016llx pkey %04x "
2351 "service_id %016llx dgid %pI6\n",
2352 (unsigned long long) be64_to_cpu(target->id_ext),
2353 (unsigned long long) be64_to_cpu(target->ioc_guid),
2354 be16_to_cpu(target->path.pkey),
2355 (unsigned long long) be64_to_cpu(target->service_id),
2356 target->path.dgid.raw);
2358 ret = srp_create_target_ib(target);
2359 if (ret)
2360 goto err_free_mem;
2362 ret = srp_new_cm_id(target);
2363 if (ret)
2364 goto err_free_ib;
2366 ret = srp_connect_target(target);
2367 if (ret) {
2368 shost_printk(KERN_ERR, target->scsi_host,
2369 PFX "Connection failed\n");
2370 goto err_cm_id;
2373 ret = srp_add_target(host, target);
2374 if (ret)
2375 goto err_disconnect;
2377 return count;
2379 err_disconnect:
2380 srp_disconnect_target(target);
2382 err_cm_id:
2383 ib_destroy_cm_id(target->cm_id);
2385 err_free_ib:
2386 srp_free_target_ib(target);
2388 err_free_mem:
2389 srp_free_req_data(target);
2391 err:
2392 scsi_host_put(target_host);
2394 return ret;
2397 static DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
2399 static ssize_t show_ibdev(struct device *dev, struct device_attribute *attr,
2400 char *buf)
2402 struct srp_host *host = container_of(dev, struct srp_host, dev);
2404 return sprintf(buf, "%s\n", host->srp_dev->dev->name);
2407 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
2409 static ssize_t show_port(struct device *dev, struct device_attribute *attr,
2410 char *buf)
2412 struct srp_host *host = container_of(dev, struct srp_host, dev);
2414 return sprintf(buf, "%d\n", host->port);
2417 static DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
2419 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
2421 struct srp_host *host;
2423 host = kzalloc(sizeof *host, GFP_KERNEL);
2424 if (!host)
2425 return NULL;
2427 INIT_LIST_HEAD(&host->target_list);
2428 spin_lock_init(&host->target_lock);
2429 init_completion(&host->released);
2430 host->srp_dev = device;
2431 host->port = port;
2433 host->dev.class = &srp_class;
2434 host->dev.parent = device->dev->dma_device;
2435 dev_set_name(&host->dev, "srp-%s-%d", device->dev->name, port);
2437 if (device_register(&host->dev))
2438 goto free_host;
2439 if (device_create_file(&host->dev, &dev_attr_add_target))
2440 goto err_class;
2441 if (device_create_file(&host->dev, &dev_attr_ibdev))
2442 goto err_class;
2443 if (device_create_file(&host->dev, &dev_attr_port))
2444 goto err_class;
2446 return host;
2448 err_class:
2449 device_unregister(&host->dev);
2451 free_host:
2452 kfree(host);
2454 return NULL;
2457 static void srp_add_one(struct ib_device *device)
2459 struct srp_device *srp_dev;
2460 struct ib_device_attr *dev_attr;
2461 struct ib_fmr_pool_param fmr_param;
2462 struct srp_host *host;
2463 int max_pages_per_fmr, fmr_page_shift, s, e, p;
2465 dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
2466 if (!dev_attr)
2467 return;
2469 if (ib_query_device(device, dev_attr)) {
2470 pr_warn("Query device failed for %s\n", device->name);
2471 goto free_attr;
2474 srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
2475 if (!srp_dev)
2476 goto free_attr;
2479 * Use the smallest page size supported by the HCA, down to a
2480 * minimum of 4096 bytes. We're unlikely to build large sglists
2481 * out of smaller entries.
2483 fmr_page_shift = max(12, ffs(dev_attr->page_size_cap) - 1);
2484 srp_dev->fmr_page_size = 1 << fmr_page_shift;
2485 srp_dev->fmr_page_mask = ~((u64) srp_dev->fmr_page_size - 1);
2486 srp_dev->fmr_max_size = srp_dev->fmr_page_size * SRP_FMR_SIZE;
2488 INIT_LIST_HEAD(&srp_dev->dev_list);
2490 srp_dev->dev = device;
2491 srp_dev->pd = ib_alloc_pd(device);
2492 if (IS_ERR(srp_dev->pd))
2493 goto free_dev;
2495 srp_dev->mr = ib_get_dma_mr(srp_dev->pd,
2496 IB_ACCESS_LOCAL_WRITE |
2497 IB_ACCESS_REMOTE_READ |
2498 IB_ACCESS_REMOTE_WRITE);
2499 if (IS_ERR(srp_dev->mr))
2500 goto err_pd;
2502 for (max_pages_per_fmr = SRP_FMR_SIZE;
2503 max_pages_per_fmr >= SRP_FMR_MIN_SIZE;
2504 max_pages_per_fmr /= 2, srp_dev->fmr_max_size /= 2) {
2505 memset(&fmr_param, 0, sizeof fmr_param);
2506 fmr_param.pool_size = SRP_FMR_POOL_SIZE;
2507 fmr_param.dirty_watermark = SRP_FMR_DIRTY_SIZE;
2508 fmr_param.cache = 1;
2509 fmr_param.max_pages_per_fmr = max_pages_per_fmr;
2510 fmr_param.page_shift = fmr_page_shift;
2511 fmr_param.access = (IB_ACCESS_LOCAL_WRITE |
2512 IB_ACCESS_REMOTE_WRITE |
2513 IB_ACCESS_REMOTE_READ);
2515 srp_dev->fmr_pool = ib_create_fmr_pool(srp_dev->pd, &fmr_param);
2516 if (!IS_ERR(srp_dev->fmr_pool))
2517 break;
2520 if (IS_ERR(srp_dev->fmr_pool))
2521 srp_dev->fmr_pool = NULL;
2523 if (device->node_type == RDMA_NODE_IB_SWITCH) {
2524 s = 0;
2525 e = 0;
2526 } else {
2527 s = 1;
2528 e = device->phys_port_cnt;
2531 for (p = s; p <= e; ++p) {
2532 host = srp_add_port(srp_dev, p);
2533 if (host)
2534 list_add_tail(&host->list, &srp_dev->dev_list);
2537 ib_set_client_data(device, &srp_client, srp_dev);
2539 goto free_attr;
2541 err_pd:
2542 ib_dealloc_pd(srp_dev->pd);
2544 free_dev:
2545 kfree(srp_dev);
2547 free_attr:
2548 kfree(dev_attr);
2551 static void srp_remove_one(struct ib_device *device)
2553 struct srp_device *srp_dev;
2554 struct srp_host *host, *tmp_host;
2555 struct srp_target_port *target;
2557 srp_dev = ib_get_client_data(device, &srp_client);
2558 if (!srp_dev)
2559 return;
2561 list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
2562 device_unregister(&host->dev);
2564 * Wait for the sysfs entry to go away, so that no new
2565 * target ports can be created.
2567 wait_for_completion(&host->released);
2570 * Remove all target ports.
2572 spin_lock(&host->target_lock);
2573 list_for_each_entry(target, &host->target_list, list)
2574 srp_queue_remove_work(target);
2575 spin_unlock(&host->target_lock);
2578 * Wait for target port removal tasks.
2580 flush_workqueue(system_long_wq);
2582 kfree(host);
2585 if (srp_dev->fmr_pool)
2586 ib_destroy_fmr_pool(srp_dev->fmr_pool);
2587 ib_dereg_mr(srp_dev->mr);
2588 ib_dealloc_pd(srp_dev->pd);
2590 kfree(srp_dev);
2593 static struct srp_function_template ib_srp_transport_functions = {
2594 .rport_delete = srp_rport_delete,
2597 static int __init srp_init_module(void)
2599 int ret;
2601 BUILD_BUG_ON(FIELD_SIZEOF(struct ib_wc, wr_id) < sizeof(void *));
2603 if (srp_sg_tablesize) {
2604 pr_warn("srp_sg_tablesize is deprecated, please use cmd_sg_entries\n");
2605 if (!cmd_sg_entries)
2606 cmd_sg_entries = srp_sg_tablesize;
2609 if (!cmd_sg_entries)
2610 cmd_sg_entries = SRP_DEF_SG_TABLESIZE;
2612 if (cmd_sg_entries > 255) {
2613 pr_warn("Clamping cmd_sg_entries to 255\n");
2614 cmd_sg_entries = 255;
2617 if (!indirect_sg_entries)
2618 indirect_sg_entries = cmd_sg_entries;
2619 else if (indirect_sg_entries < cmd_sg_entries) {
2620 pr_warn("Bumping up indirect_sg_entries to match cmd_sg_entries (%u)\n",
2621 cmd_sg_entries);
2622 indirect_sg_entries = cmd_sg_entries;
2625 ib_srp_transport_template =
2626 srp_attach_transport(&ib_srp_transport_functions);
2627 if (!ib_srp_transport_template)
2628 return -ENOMEM;
2630 ret = class_register(&srp_class);
2631 if (ret) {
2632 pr_err("couldn't register class infiniband_srp\n");
2633 srp_release_transport(ib_srp_transport_template);
2634 return ret;
2637 ib_sa_register_client(&srp_sa_client);
2639 ret = ib_register_client(&srp_client);
2640 if (ret) {
2641 pr_err("couldn't register IB client\n");
2642 srp_release_transport(ib_srp_transport_template);
2643 ib_sa_unregister_client(&srp_sa_client);
2644 class_unregister(&srp_class);
2645 return ret;
2648 return 0;
2651 static void __exit srp_cleanup_module(void)
2653 ib_unregister_client(&srp_client);
2654 ib_sa_unregister_client(&srp_sa_client);
2655 class_unregister(&srp_class);
2656 srp_release_transport(ib_srp_transport_template);
2659 module_init(srp_init_module);
2660 module_exit(srp_cleanup_module);