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
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
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
32 * $Id: ib_srp.c 3932 2005-11-01 17:19:29Z roland $
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 <asm/atomic.h>
46 #include <scsi/scsi.h>
47 #include <scsi/scsi_device.h>
48 #include <scsi/scsi_dbg.h>
51 #include <rdma/ib_cache.h>
55 #define DRV_NAME "ib_srp"
56 #define PFX DRV_NAME ": "
57 #define DRV_VERSION "0.2"
58 #define DRV_RELDATE "November 1, 2005"
60 MODULE_AUTHOR("Roland Dreier");
61 MODULE_DESCRIPTION("InfiniBand SCSI RDMA Protocol initiator "
62 "v" DRV_VERSION " (" DRV_RELDATE ")");
63 MODULE_LICENSE("Dual BSD/GPL");
65 static int srp_sg_tablesize = SRP_DEF_SG_TABLESIZE;
66 static int srp_max_iu_len;
68 module_param(srp_sg_tablesize, int, 0444);
69 MODULE_PARM_DESC(srp_sg_tablesize,
70 "Max number of gather/scatter entries per I/O (default is 12)");
72 static int topspin_workarounds = 1;
74 module_param(topspin_workarounds, int, 0444);
75 MODULE_PARM_DESC(topspin_workarounds,
76 "Enable workarounds for Topspin/Cisco SRP target bugs if != 0");
78 static const u8 topspin_oui[3] = { 0x00, 0x05, 0xad };
80 static int mellanox_workarounds = 1;
82 module_param(mellanox_workarounds, int, 0444);
83 MODULE_PARM_DESC(mellanox_workarounds,
84 "Enable workarounds for Mellanox SRP target bugs if != 0");
86 static const u8 mellanox_oui[3] = { 0x00, 0x02, 0xc9 };
88 static void srp_add_one(struct ib_device *device);
89 static void srp_remove_one(struct ib_device *device);
90 static void srp_completion(struct ib_cq *cq, void *target_ptr);
91 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event);
93 static struct ib_client srp_client = {
96 .remove = srp_remove_one
99 static struct ib_sa_client srp_sa_client;
101 static inline struct srp_target_port *host_to_target(struct Scsi_Host *host)
103 return (struct srp_target_port *) host->hostdata;
106 static const char *srp_target_info(struct Scsi_Host *host)
108 return host_to_target(host)->target_name;
111 static struct srp_iu *srp_alloc_iu(struct srp_host *host, size_t size,
113 enum dma_data_direction direction)
117 iu = kmalloc(sizeof *iu, gfp_mask);
121 iu->buf = kzalloc(size, gfp_mask);
125 iu->dma = ib_dma_map_single(host->dev->dev, iu->buf, size, direction);
126 if (ib_dma_mapping_error(host->dev->dev, iu->dma))
130 iu->direction = direction;
142 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
147 ib_dma_unmap_single(host->dev->dev, iu->dma, iu->size, iu->direction);
152 static void srp_qp_event(struct ib_event *event, void *context)
154 printk(KERN_ERR PFX "QP event %d\n", event->event);
157 static int srp_init_qp(struct srp_target_port *target,
160 struct ib_qp_attr *attr;
163 attr = kmalloc(sizeof *attr, GFP_KERNEL);
167 ret = ib_find_cached_pkey(target->srp_host->dev->dev,
168 target->srp_host->port,
169 be16_to_cpu(target->path.pkey),
174 attr->qp_state = IB_QPS_INIT;
175 attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
176 IB_ACCESS_REMOTE_WRITE);
177 attr->port_num = target->srp_host->port;
179 ret = ib_modify_qp(qp, attr,
190 static int srp_create_target_ib(struct srp_target_port *target)
192 struct ib_qp_init_attr *init_attr;
195 init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
199 target->cq = ib_create_cq(target->srp_host->dev->dev, srp_completion,
200 NULL, target, SRP_CQ_SIZE);
201 if (IS_ERR(target->cq)) {
202 ret = PTR_ERR(target->cq);
206 ib_req_notify_cq(target->cq, IB_CQ_NEXT_COMP);
208 init_attr->event_handler = srp_qp_event;
209 init_attr->cap.max_send_wr = SRP_SQ_SIZE;
210 init_attr->cap.max_recv_wr = SRP_RQ_SIZE;
211 init_attr->cap.max_recv_sge = 1;
212 init_attr->cap.max_send_sge = 1;
213 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
214 init_attr->qp_type = IB_QPT_RC;
215 init_attr->send_cq = target->cq;
216 init_attr->recv_cq = target->cq;
218 target->qp = ib_create_qp(target->srp_host->dev->pd, init_attr);
219 if (IS_ERR(target->qp)) {
220 ret = PTR_ERR(target->qp);
221 ib_destroy_cq(target->cq);
225 ret = srp_init_qp(target, target->qp);
227 ib_destroy_qp(target->qp);
228 ib_destroy_cq(target->cq);
237 static void srp_free_target_ib(struct srp_target_port *target)
241 ib_destroy_qp(target->qp);
242 ib_destroy_cq(target->cq);
244 for (i = 0; i < SRP_RQ_SIZE; ++i)
245 srp_free_iu(target->srp_host, target->rx_ring[i]);
246 for (i = 0; i < SRP_SQ_SIZE + 1; ++i)
247 srp_free_iu(target->srp_host, target->tx_ring[i]);
250 static void srp_path_rec_completion(int status,
251 struct ib_sa_path_rec *pathrec,
254 struct srp_target_port *target = target_ptr;
256 target->status = status;
258 printk(KERN_ERR PFX "Got failed path rec status %d\n", status);
260 target->path = *pathrec;
261 complete(&target->done);
264 static int srp_lookup_path(struct srp_target_port *target)
266 target->path.numb_path = 1;
268 init_completion(&target->done);
270 target->path_query_id = ib_sa_path_rec_get(&srp_sa_client,
271 target->srp_host->dev->dev,
272 target->srp_host->port,
274 IB_SA_PATH_REC_DGID |
275 IB_SA_PATH_REC_SGID |
276 IB_SA_PATH_REC_NUMB_PATH |
278 SRP_PATH_REC_TIMEOUT_MS,
280 srp_path_rec_completion,
281 target, &target->path_query);
282 if (target->path_query_id < 0)
283 return target->path_query_id;
285 wait_for_completion(&target->done);
287 if (target->status < 0)
288 printk(KERN_WARNING PFX "Path record query failed\n");
290 return target->status;
293 static int srp_send_req(struct srp_target_port *target)
296 struct ib_cm_req_param param;
297 struct srp_login_req priv;
301 req = kzalloc(sizeof *req, GFP_KERNEL);
305 req->param.primary_path = &target->path;
306 req->param.alternate_path = NULL;
307 req->param.service_id = target->service_id;
308 req->param.qp_num = target->qp->qp_num;
309 req->param.qp_type = target->qp->qp_type;
310 req->param.private_data = &req->priv;
311 req->param.private_data_len = sizeof req->priv;
312 req->param.flow_control = 1;
314 get_random_bytes(&req->param.starting_psn, 4);
315 req->param.starting_psn &= 0xffffff;
318 * Pick some arbitrary defaults here; we could make these
319 * module parameters if anyone cared about setting them.
321 req->param.responder_resources = 4;
322 req->param.remote_cm_response_timeout = 20;
323 req->param.local_cm_response_timeout = 20;
324 req->param.retry_count = 7;
325 req->param.rnr_retry_count = 7;
326 req->param.max_cm_retries = 15;
328 req->priv.opcode = SRP_LOGIN_REQ;
330 req->priv.req_it_iu_len = cpu_to_be32(srp_max_iu_len);
331 req->priv.req_buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
332 SRP_BUF_FORMAT_INDIRECT);
334 * In the published SRP specification (draft rev. 16a), the
335 * port identifier format is 8 bytes of ID extension followed
336 * by 8 bytes of GUID. Older drafts put the two halves in the
337 * opposite order, so that the GUID comes first.
339 * Targets conforming to these obsolete drafts can be
340 * recognized by the I/O Class they report.
342 if (target->io_class == SRP_REV10_IB_IO_CLASS) {
343 memcpy(req->priv.initiator_port_id,
344 &target->path.sgid.global.interface_id, 8);
345 memcpy(req->priv.initiator_port_id + 8,
346 &target->initiator_ext, 8);
347 memcpy(req->priv.target_port_id, &target->ioc_guid, 8);
348 memcpy(req->priv.target_port_id + 8, &target->id_ext, 8);
350 memcpy(req->priv.initiator_port_id,
351 &target->initiator_ext, 8);
352 memcpy(req->priv.initiator_port_id + 8,
353 &target->path.sgid.global.interface_id, 8);
354 memcpy(req->priv.target_port_id, &target->id_ext, 8);
355 memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
359 * Topspin/Cisco SRP targets will reject our login unless we
360 * zero out the first 8 bytes of our initiator port ID and set
361 * the second 8 bytes to the local node GUID.
363 if (topspin_workarounds && !memcmp(&target->ioc_guid, topspin_oui, 3)) {
364 printk(KERN_DEBUG PFX "Topspin/Cisco initiator port ID workaround "
365 "activated for target GUID %016llx\n",
366 (unsigned long long) be64_to_cpu(target->ioc_guid));
367 memset(req->priv.initiator_port_id, 0, 8);
368 memcpy(req->priv.initiator_port_id + 8,
369 &target->srp_host->dev->dev->node_guid, 8);
372 status = ib_send_cm_req(target->cm_id, &req->param);
379 static void srp_disconnect_target(struct srp_target_port *target)
381 /* XXX should send SRP_I_LOGOUT request */
383 init_completion(&target->done);
384 if (ib_send_cm_dreq(target->cm_id, NULL, 0)) {
385 printk(KERN_DEBUG PFX "Sending CM DREQ failed\n");
388 wait_for_completion(&target->done);
391 static void srp_remove_work(struct work_struct *work)
393 struct srp_target_port *target =
394 container_of(work, struct srp_target_port, work);
396 spin_lock_irq(target->scsi_host->host_lock);
397 if (target->state != SRP_TARGET_DEAD) {
398 spin_unlock_irq(target->scsi_host->host_lock);
401 target->state = SRP_TARGET_REMOVED;
402 spin_unlock_irq(target->scsi_host->host_lock);
404 spin_lock(&target->srp_host->target_lock);
405 list_del(&target->list);
406 spin_unlock(&target->srp_host->target_lock);
408 scsi_remove_host(target->scsi_host);
409 ib_destroy_cm_id(target->cm_id);
410 srp_free_target_ib(target);
411 scsi_host_put(target->scsi_host);
414 static int srp_connect_target(struct srp_target_port *target)
418 ret = srp_lookup_path(target);
423 init_completion(&target->done);
424 ret = srp_send_req(target);
427 wait_for_completion(&target->done);
430 * The CM event handling code will set status to
431 * SRP_PORT_REDIRECT if we get a port redirect REJ
432 * back, or SRP_DLID_REDIRECT if we get a lid/qp
435 switch (target->status) {
439 case SRP_PORT_REDIRECT:
440 ret = srp_lookup_path(target);
445 case SRP_DLID_REDIRECT:
449 return target->status;
454 static void srp_unmap_data(struct scsi_cmnd *scmnd,
455 struct srp_target_port *target,
456 struct srp_request *req)
458 struct scatterlist *scat;
461 if (!scmnd->request_buffer ||
462 (scmnd->sc_data_direction != DMA_TO_DEVICE &&
463 scmnd->sc_data_direction != DMA_FROM_DEVICE))
467 ib_fmr_pool_unmap(req->fmr);
472 * This handling of non-SG commands can be killed when the
473 * SCSI midlayer no longer generates non-SG commands.
475 if (likely(scmnd->use_sg)) {
476 nents = scmnd->use_sg;
477 scat = scmnd->request_buffer;
480 scat = &req->fake_sg;
483 ib_dma_unmap_sg(target->srp_host->dev->dev, scat, nents,
484 scmnd->sc_data_direction);
487 static void srp_remove_req(struct srp_target_port *target, struct srp_request *req)
489 srp_unmap_data(req->scmnd, target, req);
490 list_move_tail(&req->list, &target->free_reqs);
493 static void srp_reset_req(struct srp_target_port *target, struct srp_request *req)
495 req->scmnd->result = DID_RESET << 16;
496 req->scmnd->scsi_done(req->scmnd);
497 srp_remove_req(target, req);
500 static int srp_reconnect_target(struct srp_target_port *target)
502 struct ib_cm_id *new_cm_id;
503 struct ib_qp_attr qp_attr;
504 struct srp_request *req, *tmp;
508 spin_lock_irq(target->scsi_host->host_lock);
509 if (target->state != SRP_TARGET_LIVE) {
510 spin_unlock_irq(target->scsi_host->host_lock);
513 target->state = SRP_TARGET_CONNECTING;
514 spin_unlock_irq(target->scsi_host->host_lock);
516 srp_disconnect_target(target);
518 * Now get a new local CM ID so that we avoid confusing the
519 * target in case things are really fouled up.
521 new_cm_id = ib_create_cm_id(target->srp_host->dev->dev,
522 srp_cm_handler, target);
523 if (IS_ERR(new_cm_id)) {
524 ret = PTR_ERR(new_cm_id);
527 ib_destroy_cm_id(target->cm_id);
528 target->cm_id = new_cm_id;
530 qp_attr.qp_state = IB_QPS_RESET;
531 ret = ib_modify_qp(target->qp, &qp_attr, IB_QP_STATE);
535 ret = srp_init_qp(target, target->qp);
539 while (ib_poll_cq(target->cq, 1, &wc) > 0)
542 spin_lock_irq(target->scsi_host->host_lock);
543 list_for_each_entry_safe(req, tmp, &target->req_queue, list)
544 srp_reset_req(target, req);
545 spin_unlock_irq(target->scsi_host->host_lock);
551 ret = srp_connect_target(target);
555 spin_lock_irq(target->scsi_host->host_lock);
556 if (target->state == SRP_TARGET_CONNECTING) {
558 target->state = SRP_TARGET_LIVE;
561 spin_unlock_irq(target->scsi_host->host_lock);
566 printk(KERN_ERR PFX "reconnect failed (%d), removing target port.\n", ret);
569 * We couldn't reconnect, so kill our target port off.
570 * However, we have to defer the real removal because we might
571 * be in the context of the SCSI error handler now, which
572 * would deadlock if we call scsi_remove_host().
574 spin_lock_irq(target->scsi_host->host_lock);
575 if (target->state == SRP_TARGET_CONNECTING) {
576 target->state = SRP_TARGET_DEAD;
577 INIT_WORK(&target->work, srp_remove_work);
578 schedule_work(&target->work);
580 spin_unlock_irq(target->scsi_host->host_lock);
585 static int srp_map_fmr(struct srp_target_port *target, struct scatterlist *scat,
586 int sg_cnt, struct srp_request *req,
587 struct srp_direct_buf *buf)
595 struct srp_device *dev = target->srp_host->dev;
596 struct ib_device *ibdev = dev->dev;
601 if ((ib_sg_dma_address(ibdev, &scat[0]) & ~dev->fmr_page_mask) &&
602 mellanox_workarounds && !memcmp(&target->ioc_guid, mellanox_oui, 3))
606 for (i = 0; i < sg_cnt; ++i) {
607 unsigned int dma_len = ib_sg_dma_len(ibdev, &scat[i]);
609 if (ib_sg_dma_address(ibdev, &scat[i]) & ~dev->fmr_page_mask) {
615 if ((ib_sg_dma_address(ibdev, &scat[i]) + dma_len) &
616 ~dev->fmr_page_mask) {
626 page_cnt += len >> dev->fmr_page_shift;
627 if (page_cnt > SRP_FMR_SIZE)
630 dma_pages = kmalloc(sizeof (u64) * page_cnt, GFP_ATOMIC);
635 for (i = 0; i < sg_cnt; ++i) {
636 unsigned int dma_len = ib_sg_dma_len(ibdev, &scat[i]);
638 for (j = 0; j < dma_len; j += dev->fmr_page_size)
639 dma_pages[page_cnt++] =
640 (ib_sg_dma_address(ibdev, &scat[i]) &
641 dev->fmr_page_mask) + j;
644 req->fmr = ib_fmr_pool_map_phys(dev->fmr_pool,
645 dma_pages, page_cnt, io_addr);
646 if (IS_ERR(req->fmr)) {
647 ret = PTR_ERR(req->fmr);
652 buf->va = cpu_to_be64(ib_sg_dma_address(ibdev, &scat[0]) &
653 ~dev->fmr_page_mask);
654 buf->key = cpu_to_be32(req->fmr->fmr->rkey);
655 buf->len = cpu_to_be32(len);
665 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_target_port *target,
666 struct srp_request *req)
668 struct scatterlist *scat;
669 struct srp_cmd *cmd = req->cmd->buf;
670 int len, nents, count;
671 u8 fmt = SRP_DATA_DESC_DIRECT;
672 struct srp_device *dev;
673 struct ib_device *ibdev;
675 if (!scmnd->request_buffer || scmnd->sc_data_direction == DMA_NONE)
676 return sizeof (struct srp_cmd);
678 if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
679 scmnd->sc_data_direction != DMA_TO_DEVICE) {
680 printk(KERN_WARNING PFX "Unhandled data direction %d\n",
681 scmnd->sc_data_direction);
686 * This handling of non-SG commands can be killed when the
687 * SCSI midlayer no longer generates non-SG commands.
689 if (likely(scmnd->use_sg)) {
690 nents = scmnd->use_sg;
691 scat = scmnd->request_buffer;
694 scat = &req->fake_sg;
695 sg_init_one(scat, scmnd->request_buffer, scmnd->request_bufflen);
698 dev = target->srp_host->dev;
701 count = ib_dma_map_sg(ibdev, scat, nents, scmnd->sc_data_direction);
703 fmt = SRP_DATA_DESC_DIRECT;
704 len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
708 * The midlayer only generated a single gather/scatter
709 * entry, or DMA mapping coalesced everything to a
710 * single entry. So a direct descriptor along with
711 * the DMA MR suffices.
713 struct srp_direct_buf *buf = (void *) cmd->add_data;
715 buf->va = cpu_to_be64(ib_sg_dma_address(ibdev, scat));
716 buf->key = cpu_to_be32(dev->mr->rkey);
717 buf->len = cpu_to_be32(ib_sg_dma_len(ibdev, scat));
718 } else if (srp_map_fmr(target, scat, count, req,
719 (void *) cmd->add_data)) {
721 * FMR mapping failed, and the scatterlist has more
722 * than one entry. Generate an indirect memory
725 struct srp_indirect_buf *buf = (void *) cmd->add_data;
729 fmt = SRP_DATA_DESC_INDIRECT;
730 len = sizeof (struct srp_cmd) +
731 sizeof (struct srp_indirect_buf) +
732 count * sizeof (struct srp_direct_buf);
734 for (i = 0; i < count; ++i) {
735 unsigned int dma_len = ib_sg_dma_len(ibdev, &scat[i]);
737 buf->desc_list[i].va =
738 cpu_to_be64(ib_sg_dma_address(ibdev, &scat[i]));
739 buf->desc_list[i].key =
740 cpu_to_be32(dev->mr->rkey);
741 buf->desc_list[i].len = cpu_to_be32(dma_len);
745 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
746 cmd->data_out_desc_cnt = count;
748 cmd->data_in_desc_cnt = count;
751 cpu_to_be64(req->cmd->dma + sizeof *cmd + sizeof *buf);
752 buf->table_desc.key =
753 cpu_to_be32(target->srp_host->dev->mr->rkey);
754 buf->table_desc.len =
755 cpu_to_be32(count * sizeof (struct srp_direct_buf));
757 buf->len = cpu_to_be32(datalen);
760 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
761 cmd->buf_fmt = fmt << 4;
768 static void srp_process_rsp(struct srp_target_port *target, struct srp_rsp *rsp)
770 struct srp_request *req;
771 struct scsi_cmnd *scmnd;
775 delta = (s32) be32_to_cpu(rsp->req_lim_delta);
777 spin_lock_irqsave(target->scsi_host->host_lock, flags);
779 target->req_lim += delta;
781 req = &target->req_ring[rsp->tag & ~SRP_TAG_TSK_MGMT];
783 if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
784 if (be32_to_cpu(rsp->resp_data_len) < 4)
785 req->tsk_status = -1;
787 req->tsk_status = rsp->data[3];
788 complete(&req->done);
792 printk(KERN_ERR "Null scmnd for RSP w/tag %016llx\n",
793 (unsigned long long) rsp->tag);
794 scmnd->result = rsp->status;
796 if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
797 memcpy(scmnd->sense_buffer, rsp->data +
798 be32_to_cpu(rsp->resp_data_len),
799 min_t(int, be32_to_cpu(rsp->sense_data_len),
800 SCSI_SENSE_BUFFERSIZE));
803 if (rsp->flags & (SRP_RSP_FLAG_DOOVER | SRP_RSP_FLAG_DOUNDER))
804 scmnd->resid = be32_to_cpu(rsp->data_out_res_cnt);
805 else if (rsp->flags & (SRP_RSP_FLAG_DIOVER | SRP_RSP_FLAG_DIUNDER))
806 scmnd->resid = be32_to_cpu(rsp->data_in_res_cnt);
808 if (!req->tsk_mgmt) {
809 scmnd->host_scribble = (void *) -1L;
810 scmnd->scsi_done(scmnd);
812 srp_remove_req(target, req);
817 spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
820 static void srp_handle_recv(struct srp_target_port *target, struct ib_wc *wc)
822 struct ib_device *dev;
826 iu = target->rx_ring[wc->wr_id & ~SRP_OP_RECV];
828 dev = target->srp_host->dev->dev;
829 ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_ti_iu_len,
832 opcode = *(u8 *) iu->buf;
837 printk(KERN_ERR PFX "recv completion, opcode 0x%02x\n", opcode);
839 for (i = 0; i < wc->byte_len; ++i) {
841 printk(KERN_ERR " [%02x] ", i);
842 printk(" %02x", ((u8 *) iu->buf)[i]);
843 if ((i + 1) % 8 == 0)
847 if (wc->byte_len % 8)
853 srp_process_rsp(target, iu->buf);
857 /* XXX Handle target logout */
858 printk(KERN_WARNING PFX "Got target logout request\n");
862 printk(KERN_WARNING PFX "Unhandled SRP opcode 0x%02x\n", opcode);
866 ib_dma_sync_single_for_device(dev, iu->dma, target->max_ti_iu_len,
870 static void srp_completion(struct ib_cq *cq, void *target_ptr)
872 struct srp_target_port *target = target_ptr;
875 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
876 while (ib_poll_cq(cq, 1, &wc) > 0) {
878 printk(KERN_ERR PFX "failed %s status %d\n",
879 wc.wr_id & SRP_OP_RECV ? "receive" : "send",
884 if (wc.wr_id & SRP_OP_RECV)
885 srp_handle_recv(target, &wc);
891 static int __srp_post_recv(struct srp_target_port *target)
895 struct ib_recv_wr wr, *bad_wr;
899 next = target->rx_head & (SRP_RQ_SIZE - 1);
900 wr.wr_id = next | SRP_OP_RECV;
901 iu = target->rx_ring[next];
904 list.length = iu->size;
905 list.lkey = target->srp_host->dev->mr->lkey;
911 ret = ib_post_recv(target->qp, &wr, &bad_wr);
918 static int srp_post_recv(struct srp_target_port *target)
923 spin_lock_irqsave(target->scsi_host->host_lock, flags);
924 ret = __srp_post_recv(target);
925 spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
931 * Must be called with target->scsi_host->host_lock held to protect
932 * req_lim and tx_head. Lock cannot be dropped between call here and
933 * call to __srp_post_send().
935 static struct srp_iu *__srp_get_tx_iu(struct srp_target_port *target)
937 if (target->tx_head - target->tx_tail >= SRP_SQ_SIZE)
940 if (unlikely(target->req_lim < 1))
941 ++target->zero_req_lim;
943 return target->tx_ring[target->tx_head & SRP_SQ_SIZE];
947 * Must be called with target->scsi_host->host_lock held to protect
948 * req_lim and tx_head.
950 static int __srp_post_send(struct srp_target_port *target,
951 struct srp_iu *iu, int len)
954 struct ib_send_wr wr, *bad_wr;
959 list.lkey = target->srp_host->dev->mr->lkey;
962 wr.wr_id = target->tx_head & SRP_SQ_SIZE;
965 wr.opcode = IB_WR_SEND;
966 wr.send_flags = IB_SEND_SIGNALED;
968 ret = ib_post_send(target->qp, &wr, &bad_wr);
978 static int srp_queuecommand(struct scsi_cmnd *scmnd,
979 void (*done)(struct scsi_cmnd *))
981 struct srp_target_port *target = host_to_target(scmnd->device->host);
982 struct srp_request *req;
985 struct ib_device *dev;
988 if (target->state == SRP_TARGET_CONNECTING)
991 if (target->state == SRP_TARGET_DEAD ||
992 target->state == SRP_TARGET_REMOVED) {
993 scmnd->result = DID_BAD_TARGET << 16;
998 iu = __srp_get_tx_iu(target);
1002 dev = target->srp_host->dev->dev;
1003 ib_dma_sync_single_for_cpu(dev, iu->dma, srp_max_iu_len,
1006 req = list_entry(target->free_reqs.next, struct srp_request, list);
1008 scmnd->scsi_done = done;
1010 scmnd->host_scribble = (void *) (long) req->index;
1013 memset(cmd, 0, sizeof *cmd);
1015 cmd->opcode = SRP_CMD;
1016 cmd->lun = cpu_to_be64((u64) scmnd->device->lun << 48);
1017 cmd->tag = req->index;
1018 memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
1023 req->tsk_mgmt = NULL;
1025 len = srp_map_data(scmnd, target, req);
1027 printk(KERN_ERR PFX "Failed to map data\n");
1031 if (__srp_post_recv(target)) {
1032 printk(KERN_ERR PFX "Recv failed\n");
1036 ib_dma_sync_single_for_device(dev, iu->dma, srp_max_iu_len,
1039 if (__srp_post_send(target, iu, len)) {
1040 printk(KERN_ERR PFX "Send failed\n");
1044 list_move_tail(&req->list, &target->req_queue);
1049 srp_unmap_data(scmnd, target, req);
1052 return SCSI_MLQUEUE_HOST_BUSY;
1055 static int srp_alloc_iu_bufs(struct srp_target_port *target)
1059 for (i = 0; i < SRP_RQ_SIZE; ++i) {
1060 target->rx_ring[i] = srp_alloc_iu(target->srp_host,
1061 target->max_ti_iu_len,
1062 GFP_KERNEL, DMA_FROM_DEVICE);
1063 if (!target->rx_ring[i])
1067 for (i = 0; i < SRP_SQ_SIZE + 1; ++i) {
1068 target->tx_ring[i] = srp_alloc_iu(target->srp_host,
1070 GFP_KERNEL, DMA_TO_DEVICE);
1071 if (!target->tx_ring[i])
1078 for (i = 0; i < SRP_RQ_SIZE; ++i) {
1079 srp_free_iu(target->srp_host, target->rx_ring[i]);
1080 target->rx_ring[i] = NULL;
1083 for (i = 0; i < SRP_SQ_SIZE + 1; ++i) {
1084 srp_free_iu(target->srp_host, target->tx_ring[i]);
1085 target->tx_ring[i] = NULL;
1091 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
1092 struct ib_cm_event *event,
1093 struct srp_target_port *target)
1095 struct ib_class_port_info *cpi;
1098 switch (event->param.rej_rcvd.reason) {
1099 case IB_CM_REJ_PORT_CM_REDIRECT:
1100 cpi = event->param.rej_rcvd.ari;
1101 target->path.dlid = cpi->redirect_lid;
1102 target->path.pkey = cpi->redirect_pkey;
1103 cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
1104 memcpy(target->path.dgid.raw, cpi->redirect_gid, 16);
1106 target->status = target->path.dlid ?
1107 SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
1110 case IB_CM_REJ_PORT_REDIRECT:
1111 if (topspin_workarounds &&
1112 !memcmp(&target->ioc_guid, topspin_oui, 3)) {
1114 * Topspin/Cisco SRP gateways incorrectly send
1115 * reject reason code 25 when they mean 24
1118 memcpy(target->path.dgid.raw,
1119 event->param.rej_rcvd.ari, 16);
1121 printk(KERN_DEBUG PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
1122 (unsigned long long) be64_to_cpu(target->path.dgid.global.subnet_prefix),
1123 (unsigned long long) be64_to_cpu(target->path.dgid.global.interface_id));
1125 target->status = SRP_PORT_REDIRECT;
1127 printk(KERN_WARNING " REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
1128 target->status = -ECONNRESET;
1132 case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
1133 printk(KERN_WARNING " REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
1134 target->status = -ECONNRESET;
1137 case IB_CM_REJ_CONSUMER_DEFINED:
1138 opcode = *(u8 *) event->private_data;
1139 if (opcode == SRP_LOGIN_REJ) {
1140 struct srp_login_rej *rej = event->private_data;
1141 u32 reason = be32_to_cpu(rej->reason);
1143 if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
1144 printk(KERN_WARNING PFX
1145 "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
1147 printk(KERN_WARNING PFX
1148 "SRP LOGIN REJECTED, reason 0x%08x\n", reason);
1150 printk(KERN_WARNING " REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
1151 " opcode 0x%02x\n", opcode);
1152 target->status = -ECONNRESET;
1156 printk(KERN_WARNING " REJ reason 0x%x\n",
1157 event->param.rej_rcvd.reason);
1158 target->status = -ECONNRESET;
1162 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
1164 struct srp_target_port *target = cm_id->context;
1165 struct ib_qp_attr *qp_attr = NULL;
1170 switch (event->event) {
1171 case IB_CM_REQ_ERROR:
1172 printk(KERN_DEBUG PFX "Sending CM REQ failed\n");
1174 target->status = -ECONNRESET;
1177 case IB_CM_REP_RECEIVED:
1179 opcode = *(u8 *) event->private_data;
1181 if (opcode == SRP_LOGIN_RSP) {
1182 struct srp_login_rsp *rsp = event->private_data;
1184 target->max_ti_iu_len = be32_to_cpu(rsp->max_ti_iu_len);
1185 target->req_lim = be32_to_cpu(rsp->req_lim_delta);
1187 target->scsi_host->can_queue = min(target->req_lim,
1188 target->scsi_host->can_queue);
1190 printk(KERN_WARNING PFX "Unhandled RSP opcode %#x\n", opcode);
1191 target->status = -ECONNRESET;
1195 if (!target->rx_ring[0]) {
1196 target->status = srp_alloc_iu_bufs(target);
1201 qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
1203 target->status = -ENOMEM;
1207 qp_attr->qp_state = IB_QPS_RTR;
1208 target->status = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1212 target->status = ib_modify_qp(target->qp, qp_attr, attr_mask);
1216 target->status = srp_post_recv(target);
1220 qp_attr->qp_state = IB_QPS_RTS;
1221 target->status = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1225 target->status = ib_modify_qp(target->qp, qp_attr, attr_mask);
1229 target->status = ib_send_cm_rtu(cm_id, NULL, 0);
1235 case IB_CM_REJ_RECEIVED:
1236 printk(KERN_DEBUG PFX "REJ received\n");
1239 srp_cm_rej_handler(cm_id, event, target);
1242 case IB_CM_DREQ_RECEIVED:
1243 printk(KERN_WARNING PFX "DREQ received - connection closed\n");
1244 if (ib_send_cm_drep(cm_id, NULL, 0))
1245 printk(KERN_ERR PFX "Sending CM DREP failed\n");
1248 case IB_CM_TIMEWAIT_EXIT:
1249 printk(KERN_ERR PFX "connection closed\n");
1255 case IB_CM_MRA_RECEIVED:
1256 case IB_CM_DREQ_ERROR:
1257 case IB_CM_DREP_RECEIVED:
1261 printk(KERN_WARNING PFX "Unhandled CM event %d\n", event->event);
1266 complete(&target->done);
1273 static int srp_send_tsk_mgmt(struct srp_target_port *target,
1274 struct srp_request *req, u8 func)
1277 struct srp_tsk_mgmt *tsk_mgmt;
1279 spin_lock_irq(target->scsi_host->host_lock);
1281 if (target->state == SRP_TARGET_DEAD ||
1282 target->state == SRP_TARGET_REMOVED) {
1283 req->scmnd->result = DID_BAD_TARGET << 16;
1287 init_completion(&req->done);
1289 iu = __srp_get_tx_iu(target);
1294 memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
1296 tsk_mgmt->opcode = SRP_TSK_MGMT;
1297 tsk_mgmt->lun = cpu_to_be64((u64) req->scmnd->device->lun << 48);
1298 tsk_mgmt->tag = req->index | SRP_TAG_TSK_MGMT;
1299 tsk_mgmt->tsk_mgmt_func = func;
1300 tsk_mgmt->task_tag = req->index;
1302 if (__srp_post_send(target, iu, sizeof *tsk_mgmt))
1307 spin_unlock_irq(target->scsi_host->host_lock);
1309 if (!wait_for_completion_timeout(&req->done,
1310 msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
1316 spin_unlock_irq(target->scsi_host->host_lock);
1320 static int srp_find_req(struct srp_target_port *target,
1321 struct scsi_cmnd *scmnd,
1322 struct srp_request **req)
1324 if (scmnd->host_scribble == (void *) -1L)
1327 *req = &target->req_ring[(long) scmnd->host_scribble];
1332 static int srp_abort(struct scsi_cmnd *scmnd)
1334 struct srp_target_port *target = host_to_target(scmnd->device->host);
1335 struct srp_request *req;
1338 printk(KERN_ERR "SRP abort called\n");
1340 if (srp_find_req(target, scmnd, &req))
1342 if (srp_send_tsk_mgmt(target, req, SRP_TSK_ABORT_TASK))
1345 spin_lock_irq(target->scsi_host->host_lock);
1347 if (req->cmd_done) {
1348 srp_remove_req(target, req);
1349 scmnd->scsi_done(scmnd);
1350 } else if (!req->tsk_status) {
1351 srp_remove_req(target, req);
1352 scmnd->result = DID_ABORT << 16;
1356 spin_unlock_irq(target->scsi_host->host_lock);
1361 static int srp_reset_device(struct scsi_cmnd *scmnd)
1363 struct srp_target_port *target = host_to_target(scmnd->device->host);
1364 struct srp_request *req, *tmp;
1366 printk(KERN_ERR "SRP reset_device called\n");
1368 if (srp_find_req(target, scmnd, &req))
1370 if (srp_send_tsk_mgmt(target, req, SRP_TSK_LUN_RESET))
1372 if (req->tsk_status)
1375 spin_lock_irq(target->scsi_host->host_lock);
1377 list_for_each_entry_safe(req, tmp, &target->req_queue, list)
1378 if (req->scmnd->device == scmnd->device)
1379 srp_reset_req(target, req);
1381 spin_unlock_irq(target->scsi_host->host_lock);
1386 static int srp_reset_host(struct scsi_cmnd *scmnd)
1388 struct srp_target_port *target = host_to_target(scmnd->device->host);
1391 printk(KERN_ERR PFX "SRP reset_host called\n");
1393 if (!srp_reconnect_target(target))
1399 static ssize_t show_id_ext(struct class_device *cdev, char *buf)
1401 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1403 if (target->state == SRP_TARGET_DEAD ||
1404 target->state == SRP_TARGET_REMOVED)
1407 return sprintf(buf, "0x%016llx\n",
1408 (unsigned long long) be64_to_cpu(target->id_ext));
1411 static ssize_t show_ioc_guid(struct class_device *cdev, char *buf)
1413 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1415 if (target->state == SRP_TARGET_DEAD ||
1416 target->state == SRP_TARGET_REMOVED)
1419 return sprintf(buf, "0x%016llx\n",
1420 (unsigned long long) be64_to_cpu(target->ioc_guid));
1423 static ssize_t show_service_id(struct class_device *cdev, char *buf)
1425 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1427 if (target->state == SRP_TARGET_DEAD ||
1428 target->state == SRP_TARGET_REMOVED)
1431 return sprintf(buf, "0x%016llx\n",
1432 (unsigned long long) be64_to_cpu(target->service_id));
1435 static ssize_t show_pkey(struct class_device *cdev, char *buf)
1437 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1439 if (target->state == SRP_TARGET_DEAD ||
1440 target->state == SRP_TARGET_REMOVED)
1443 return sprintf(buf, "0x%04x\n", be16_to_cpu(target->path.pkey));
1446 static ssize_t show_dgid(struct class_device *cdev, char *buf)
1448 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1450 if (target->state == SRP_TARGET_DEAD ||
1451 target->state == SRP_TARGET_REMOVED)
1454 return sprintf(buf, "%04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x\n",
1455 be16_to_cpu(((__be16 *) target->path.dgid.raw)[0]),
1456 be16_to_cpu(((__be16 *) target->path.dgid.raw)[1]),
1457 be16_to_cpu(((__be16 *) target->path.dgid.raw)[2]),
1458 be16_to_cpu(((__be16 *) target->path.dgid.raw)[3]),
1459 be16_to_cpu(((__be16 *) target->path.dgid.raw)[4]),
1460 be16_to_cpu(((__be16 *) target->path.dgid.raw)[5]),
1461 be16_to_cpu(((__be16 *) target->path.dgid.raw)[6]),
1462 be16_to_cpu(((__be16 *) target->path.dgid.raw)[7]));
1465 static ssize_t show_zero_req_lim(struct class_device *cdev, char *buf)
1467 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1469 if (target->state == SRP_TARGET_DEAD ||
1470 target->state == SRP_TARGET_REMOVED)
1473 return sprintf(buf, "%d\n", target->zero_req_lim);
1476 static ssize_t show_local_ib_port(struct class_device *cdev, char *buf)
1478 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1480 return sprintf(buf, "%d\n", target->srp_host->port);
1483 static ssize_t show_local_ib_device(struct class_device *cdev, char *buf)
1485 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1487 return sprintf(buf, "%s\n", target->srp_host->dev->dev->name);
1490 static CLASS_DEVICE_ATTR(id_ext, S_IRUGO, show_id_ext, NULL);
1491 static CLASS_DEVICE_ATTR(ioc_guid, S_IRUGO, show_ioc_guid, NULL);
1492 static CLASS_DEVICE_ATTR(service_id, S_IRUGO, show_service_id, NULL);
1493 static CLASS_DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1494 static CLASS_DEVICE_ATTR(dgid, S_IRUGO, show_dgid, NULL);
1495 static CLASS_DEVICE_ATTR(zero_req_lim, S_IRUGO, show_zero_req_lim, NULL);
1496 static CLASS_DEVICE_ATTR(local_ib_port, S_IRUGO, show_local_ib_port, NULL);
1497 static CLASS_DEVICE_ATTR(local_ib_device, S_IRUGO, show_local_ib_device, NULL);
1499 static struct class_device_attribute *srp_host_attrs[] = {
1500 &class_device_attr_id_ext,
1501 &class_device_attr_ioc_guid,
1502 &class_device_attr_service_id,
1503 &class_device_attr_pkey,
1504 &class_device_attr_dgid,
1505 &class_device_attr_zero_req_lim,
1506 &class_device_attr_local_ib_port,
1507 &class_device_attr_local_ib_device,
1511 static struct scsi_host_template srp_template = {
1512 .module = THIS_MODULE,
1514 .info = srp_target_info,
1515 .queuecommand = srp_queuecommand,
1516 .eh_abort_handler = srp_abort,
1517 .eh_device_reset_handler = srp_reset_device,
1518 .eh_host_reset_handler = srp_reset_host,
1519 .can_queue = SRP_SQ_SIZE,
1521 .cmd_per_lun = SRP_SQ_SIZE,
1522 .use_clustering = ENABLE_CLUSTERING,
1523 .shost_attrs = srp_host_attrs
1526 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
1528 sprintf(target->target_name, "SRP.T10:%016llX",
1529 (unsigned long long) be64_to_cpu(target->id_ext));
1531 if (scsi_add_host(target->scsi_host, host->dev->dev->dma_device))
1534 spin_lock(&host->target_lock);
1535 list_add_tail(&target->list, &host->target_list);
1536 spin_unlock(&host->target_lock);
1538 target->state = SRP_TARGET_LIVE;
1540 scsi_scan_target(&target->scsi_host->shost_gendev,
1541 0, target->scsi_id, SCAN_WILD_CARD, 0);
1546 static void srp_release_class_dev(struct class_device *class_dev)
1548 struct srp_host *host =
1549 container_of(class_dev, struct srp_host, class_dev);
1551 complete(&host->released);
1554 static struct class srp_class = {
1555 .name = "infiniband_srp",
1556 .release = srp_release_class_dev
1560 * Target ports are added by writing
1562 * id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
1563 * pkey=<P_Key>,service_id=<service ID>
1565 * to the add_target sysfs attribute.
1569 SRP_OPT_ID_EXT = 1 << 0,
1570 SRP_OPT_IOC_GUID = 1 << 1,
1571 SRP_OPT_DGID = 1 << 2,
1572 SRP_OPT_PKEY = 1 << 3,
1573 SRP_OPT_SERVICE_ID = 1 << 4,
1574 SRP_OPT_MAX_SECT = 1 << 5,
1575 SRP_OPT_MAX_CMD_PER_LUN = 1 << 6,
1576 SRP_OPT_IO_CLASS = 1 << 7,
1577 SRP_OPT_INITIATOR_EXT = 1 << 8,
1578 SRP_OPT_ALL = (SRP_OPT_ID_EXT |
1582 SRP_OPT_SERVICE_ID),
1585 static match_table_t srp_opt_tokens = {
1586 { SRP_OPT_ID_EXT, "id_ext=%s" },
1587 { SRP_OPT_IOC_GUID, "ioc_guid=%s" },
1588 { SRP_OPT_DGID, "dgid=%s" },
1589 { SRP_OPT_PKEY, "pkey=%x" },
1590 { SRP_OPT_SERVICE_ID, "service_id=%s" },
1591 { SRP_OPT_MAX_SECT, "max_sect=%d" },
1592 { SRP_OPT_MAX_CMD_PER_LUN, "max_cmd_per_lun=%d" },
1593 { SRP_OPT_IO_CLASS, "io_class=%x" },
1594 { SRP_OPT_INITIATOR_EXT, "initiator_ext=%s" },
1595 { SRP_OPT_ERR, NULL }
1598 static int srp_parse_options(const char *buf, struct srp_target_port *target)
1600 char *options, *sep_opt;
1603 substring_t args[MAX_OPT_ARGS];
1609 options = kstrdup(buf, GFP_KERNEL);
1614 while ((p = strsep(&sep_opt, ",")) != NULL) {
1618 token = match_token(p, srp_opt_tokens, args);
1622 case SRP_OPT_ID_EXT:
1623 p = match_strdup(args);
1628 target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
1632 case SRP_OPT_IOC_GUID:
1633 p = match_strdup(args);
1638 target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
1643 p = match_strdup(args);
1648 if (strlen(p) != 32) {
1649 printk(KERN_WARNING PFX "bad dest GID parameter '%s'\n", p);
1654 for (i = 0; i < 16; ++i) {
1655 strlcpy(dgid, p + i * 2, 3);
1656 target->path.dgid.raw[i] = simple_strtoul(dgid, NULL, 16);
1662 if (match_hex(args, &token)) {
1663 printk(KERN_WARNING PFX "bad P_Key parameter '%s'\n", p);
1666 target->path.pkey = cpu_to_be16(token);
1669 case SRP_OPT_SERVICE_ID:
1670 p = match_strdup(args);
1675 target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
1679 case SRP_OPT_MAX_SECT:
1680 if (match_int(args, &token)) {
1681 printk(KERN_WARNING PFX "bad max sect parameter '%s'\n", p);
1684 target->scsi_host->max_sectors = token;
1687 case SRP_OPT_MAX_CMD_PER_LUN:
1688 if (match_int(args, &token)) {
1689 printk(KERN_WARNING PFX "bad max cmd_per_lun parameter '%s'\n", p);
1692 target->scsi_host->cmd_per_lun = min(token, SRP_SQ_SIZE);
1695 case SRP_OPT_IO_CLASS:
1696 if (match_hex(args, &token)) {
1697 printk(KERN_WARNING PFX "bad IO class parameter '%s' \n", p);
1700 if (token != SRP_REV10_IB_IO_CLASS &&
1701 token != SRP_REV16A_IB_IO_CLASS) {
1702 printk(KERN_WARNING PFX "unknown IO class parameter value"
1703 " %x specified (use %x or %x).\n",
1704 token, SRP_REV10_IB_IO_CLASS, SRP_REV16A_IB_IO_CLASS);
1707 target->io_class = token;
1710 case SRP_OPT_INITIATOR_EXT:
1711 p = match_strdup(args);
1716 target->initiator_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
1721 printk(KERN_WARNING PFX "unknown parameter or missing value "
1722 "'%s' in target creation request\n", p);
1727 if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
1730 for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
1731 if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
1732 !(srp_opt_tokens[i].token & opt_mask))
1733 printk(KERN_WARNING PFX "target creation request is "
1734 "missing parameter '%s'\n",
1735 srp_opt_tokens[i].pattern);
1742 static ssize_t srp_create_target(struct class_device *class_dev,
1743 const char *buf, size_t count)
1745 struct srp_host *host =
1746 container_of(class_dev, struct srp_host, class_dev);
1747 struct Scsi_Host *target_host;
1748 struct srp_target_port *target;
1752 target_host = scsi_host_alloc(&srp_template,
1753 sizeof (struct srp_target_port));
1757 target_host->max_lun = SRP_MAX_LUN;
1758 target_host->max_cmd_len = sizeof ((struct srp_cmd *) (void *) 0L)->cdb;
1760 target = host_to_target(target_host);
1762 target->io_class = SRP_REV16A_IB_IO_CLASS;
1763 target->scsi_host = target_host;
1764 target->srp_host = host;
1766 INIT_LIST_HEAD(&target->free_reqs);
1767 INIT_LIST_HEAD(&target->req_queue);
1768 for (i = 0; i < SRP_SQ_SIZE; ++i) {
1769 target->req_ring[i].index = i;
1770 list_add_tail(&target->req_ring[i].list, &target->free_reqs);
1773 ret = srp_parse_options(buf, target);
1777 ib_get_cached_gid(host->dev->dev, host->port, 0, &target->path.sgid);
1779 printk(KERN_DEBUG PFX "new target: id_ext %016llx ioc_guid %016llx pkey %04x "
1780 "service_id %016llx dgid %04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x\n",
1781 (unsigned long long) be64_to_cpu(target->id_ext),
1782 (unsigned long long) be64_to_cpu(target->ioc_guid),
1783 be16_to_cpu(target->path.pkey),
1784 (unsigned long long) be64_to_cpu(target->service_id),
1785 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[0]),
1786 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[2]),
1787 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[4]),
1788 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[6]),
1789 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[8]),
1790 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[10]),
1791 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[12]),
1792 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[14]));
1794 ret = srp_create_target_ib(target);
1798 target->cm_id = ib_create_cm_id(host->dev->dev, srp_cm_handler, target);
1799 if (IS_ERR(target->cm_id)) {
1800 ret = PTR_ERR(target->cm_id);
1804 ret = srp_connect_target(target);
1806 printk(KERN_ERR PFX "Connection failed\n");
1810 ret = srp_add_target(host, target);
1812 goto err_disconnect;
1817 srp_disconnect_target(target);
1820 ib_destroy_cm_id(target->cm_id);
1823 srp_free_target_ib(target);
1826 scsi_host_put(target_host);
1831 static CLASS_DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
1833 static ssize_t show_ibdev(struct class_device *class_dev, char *buf)
1835 struct srp_host *host =
1836 container_of(class_dev, struct srp_host, class_dev);
1838 return sprintf(buf, "%s\n", host->dev->dev->name);
1841 static CLASS_DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
1843 static ssize_t show_port(struct class_device *class_dev, char *buf)
1845 struct srp_host *host =
1846 container_of(class_dev, struct srp_host, class_dev);
1848 return sprintf(buf, "%d\n", host->port);
1851 static CLASS_DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
1853 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
1855 struct srp_host *host;
1857 host = kzalloc(sizeof *host, GFP_KERNEL);
1861 INIT_LIST_HEAD(&host->target_list);
1862 spin_lock_init(&host->target_lock);
1863 init_completion(&host->released);
1867 host->class_dev.class = &srp_class;
1868 host->class_dev.dev = device->dev->dma_device;
1869 snprintf(host->class_dev.class_id, BUS_ID_SIZE, "srp-%s-%d",
1870 device->dev->name, port);
1872 if (class_device_register(&host->class_dev))
1874 if (class_device_create_file(&host->class_dev, &class_device_attr_add_target))
1876 if (class_device_create_file(&host->class_dev, &class_device_attr_ibdev))
1878 if (class_device_create_file(&host->class_dev, &class_device_attr_port))
1884 class_device_unregister(&host->class_dev);
1892 static void srp_add_one(struct ib_device *device)
1894 struct srp_device *srp_dev;
1895 struct ib_device_attr *dev_attr;
1896 struct ib_fmr_pool_param fmr_param;
1897 struct srp_host *host;
1900 dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
1904 if (ib_query_device(device, dev_attr)) {
1905 printk(KERN_WARNING PFX "Query device failed for %s\n",
1910 srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
1915 * Use the smallest page size supported by the HCA, down to a
1916 * minimum of 512 bytes (which is the smallest sector that a
1917 * SCSI command will ever carry).
1919 srp_dev->fmr_page_shift = max(9, ffs(dev_attr->page_size_cap) - 1);
1920 srp_dev->fmr_page_size = 1 << srp_dev->fmr_page_shift;
1921 srp_dev->fmr_page_mask = ~((u64) srp_dev->fmr_page_size - 1);
1923 INIT_LIST_HEAD(&srp_dev->dev_list);
1925 srp_dev->dev = device;
1926 srp_dev->pd = ib_alloc_pd(device);
1927 if (IS_ERR(srp_dev->pd))
1930 srp_dev->mr = ib_get_dma_mr(srp_dev->pd,
1931 IB_ACCESS_LOCAL_WRITE |
1932 IB_ACCESS_REMOTE_READ |
1933 IB_ACCESS_REMOTE_WRITE);
1934 if (IS_ERR(srp_dev->mr))
1937 memset(&fmr_param, 0, sizeof fmr_param);
1938 fmr_param.pool_size = SRP_FMR_POOL_SIZE;
1939 fmr_param.dirty_watermark = SRP_FMR_DIRTY_SIZE;
1940 fmr_param.cache = 1;
1941 fmr_param.max_pages_per_fmr = SRP_FMR_SIZE;
1942 fmr_param.page_shift = srp_dev->fmr_page_shift;
1943 fmr_param.access = (IB_ACCESS_LOCAL_WRITE |
1944 IB_ACCESS_REMOTE_WRITE |
1945 IB_ACCESS_REMOTE_READ);
1947 srp_dev->fmr_pool = ib_create_fmr_pool(srp_dev->pd, &fmr_param);
1948 if (IS_ERR(srp_dev->fmr_pool))
1949 srp_dev->fmr_pool = NULL;
1951 if (device->node_type == RDMA_NODE_IB_SWITCH) {
1956 e = device->phys_port_cnt;
1959 for (p = s; p <= e; ++p) {
1960 host = srp_add_port(srp_dev, p);
1962 list_add_tail(&host->list, &srp_dev->dev_list);
1965 ib_set_client_data(device, &srp_client, srp_dev);
1970 ib_dealloc_pd(srp_dev->pd);
1979 static void srp_remove_one(struct ib_device *device)
1981 struct srp_device *srp_dev;
1982 struct srp_host *host, *tmp_host;
1983 LIST_HEAD(target_list);
1984 struct srp_target_port *target, *tmp_target;
1986 srp_dev = ib_get_client_data(device, &srp_client);
1988 list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
1989 class_device_unregister(&host->class_dev);
1991 * Wait for the sysfs entry to go away, so that no new
1992 * target ports can be created.
1994 wait_for_completion(&host->released);
1997 * Mark all target ports as removed, so we stop queueing
1998 * commands and don't try to reconnect.
2000 spin_lock(&host->target_lock);
2001 list_for_each_entry(target, &host->target_list, list) {
2002 spin_lock_irq(target->scsi_host->host_lock);
2003 target->state = SRP_TARGET_REMOVED;
2004 spin_unlock_irq(target->scsi_host->host_lock);
2006 spin_unlock(&host->target_lock);
2009 * Wait for any reconnection tasks that may have
2010 * started before we marked our target ports as
2011 * removed, and any target port removal tasks.
2013 flush_scheduled_work();
2015 list_for_each_entry_safe(target, tmp_target,
2016 &host->target_list, list) {
2017 scsi_remove_host(target->scsi_host);
2018 srp_disconnect_target(target);
2019 ib_destroy_cm_id(target->cm_id);
2020 srp_free_target_ib(target);
2021 scsi_host_put(target->scsi_host);
2027 if (srp_dev->fmr_pool)
2028 ib_destroy_fmr_pool(srp_dev->fmr_pool);
2029 ib_dereg_mr(srp_dev->mr);
2030 ib_dealloc_pd(srp_dev->pd);
2035 static int __init srp_init_module(void)
2039 srp_template.sg_tablesize = srp_sg_tablesize;
2040 srp_max_iu_len = (sizeof (struct srp_cmd) +
2041 sizeof (struct srp_indirect_buf) +
2042 srp_sg_tablesize * 16);
2044 ret = class_register(&srp_class);
2046 printk(KERN_ERR PFX "couldn't register class infiniband_srp\n");
2050 ib_sa_register_client(&srp_sa_client);
2052 ret = ib_register_client(&srp_client);
2054 printk(KERN_ERR PFX "couldn't register IB client\n");
2055 ib_sa_unregister_client(&srp_sa_client);
2056 class_unregister(&srp_class);
2063 static void __exit srp_cleanup_module(void)
2065 ib_unregister_client(&srp_client);
2066 ib_sa_unregister_client(&srp_sa_client);
2067 class_unregister(&srp_class);
2070 module_init(srp_init_module);
2071 module_exit(srp_cleanup_module);