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 int mellanox_workarounds = 1;
80 module_param(mellanox_workarounds, int, 0444);
81 MODULE_PARM_DESC(mellanox_workarounds,
82 "Enable workarounds for Mellanox SRP target bugs if != 0");
84 static void srp_add_one(struct ib_device *device);
85 static void srp_remove_one(struct ib_device *device);
86 static void srp_completion(struct ib_cq *cq, void *target_ptr);
87 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event);
89 static struct ib_client srp_client = {
92 .remove = srp_remove_one
95 static struct ib_sa_client srp_sa_client;
97 static inline struct srp_target_port *host_to_target(struct Scsi_Host *host)
99 return (struct srp_target_port *) host->hostdata;
102 static const char *srp_target_info(struct Scsi_Host *host)
104 return host_to_target(host)->target_name;
107 static int srp_target_is_topspin(struct srp_target_port *target)
109 static const u8 topspin_oui[3] = { 0x00, 0x05, 0xad };
110 static const u8 cisco_oui[3] = { 0x00, 0x1b, 0x0d };
112 return topspin_workarounds &&
113 (!memcmp(&target->ioc_guid, topspin_oui, sizeof topspin_oui) ||
114 !memcmp(&target->ioc_guid, cisco_oui, sizeof cisco_oui));
117 static int srp_target_is_mellanox(struct srp_target_port *target)
119 static const u8 mellanox_oui[3] = { 0x00, 0x02, 0xc9 };
121 return mellanox_workarounds &&
122 !memcmp(&target->ioc_guid, mellanox_oui, sizeof mellanox_oui);
125 static struct srp_iu *srp_alloc_iu(struct srp_host *host, size_t size,
127 enum dma_data_direction direction)
131 iu = kmalloc(sizeof *iu, gfp_mask);
135 iu->buf = kzalloc(size, gfp_mask);
139 iu->dma = ib_dma_map_single(host->dev->dev, iu->buf, size, direction);
140 if (ib_dma_mapping_error(host->dev->dev, iu->dma))
144 iu->direction = direction;
156 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
161 ib_dma_unmap_single(host->dev->dev, iu->dma, iu->size, iu->direction);
166 static void srp_qp_event(struct ib_event *event, void *context)
168 printk(KERN_ERR PFX "QP event %d\n", event->event);
171 static int srp_init_qp(struct srp_target_port *target,
174 struct ib_qp_attr *attr;
177 attr = kmalloc(sizeof *attr, GFP_KERNEL);
181 ret = ib_find_cached_pkey(target->srp_host->dev->dev,
182 target->srp_host->port,
183 be16_to_cpu(target->path.pkey),
188 attr->qp_state = IB_QPS_INIT;
189 attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
190 IB_ACCESS_REMOTE_WRITE);
191 attr->port_num = target->srp_host->port;
193 ret = ib_modify_qp(qp, attr,
204 static int srp_create_target_ib(struct srp_target_port *target)
206 struct ib_qp_init_attr *init_attr;
209 init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
213 target->cq = ib_create_cq(target->srp_host->dev->dev, srp_completion,
214 NULL, target, SRP_CQ_SIZE, 0);
215 if (IS_ERR(target->cq)) {
216 ret = PTR_ERR(target->cq);
220 ib_req_notify_cq(target->cq, IB_CQ_NEXT_COMP);
222 init_attr->event_handler = srp_qp_event;
223 init_attr->cap.max_send_wr = SRP_SQ_SIZE;
224 init_attr->cap.max_recv_wr = SRP_RQ_SIZE;
225 init_attr->cap.max_recv_sge = 1;
226 init_attr->cap.max_send_sge = 1;
227 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
228 init_attr->qp_type = IB_QPT_RC;
229 init_attr->send_cq = target->cq;
230 init_attr->recv_cq = target->cq;
232 target->qp = ib_create_qp(target->srp_host->dev->pd, init_attr);
233 if (IS_ERR(target->qp)) {
234 ret = PTR_ERR(target->qp);
235 ib_destroy_cq(target->cq);
239 ret = srp_init_qp(target, target->qp);
241 ib_destroy_qp(target->qp);
242 ib_destroy_cq(target->cq);
251 static void srp_free_target_ib(struct srp_target_port *target)
255 ib_destroy_qp(target->qp);
256 ib_destroy_cq(target->cq);
258 for (i = 0; i < SRP_RQ_SIZE; ++i)
259 srp_free_iu(target->srp_host, target->rx_ring[i]);
260 for (i = 0; i < SRP_SQ_SIZE + 1; ++i)
261 srp_free_iu(target->srp_host, target->tx_ring[i]);
264 static void srp_path_rec_completion(int status,
265 struct ib_sa_path_rec *pathrec,
268 struct srp_target_port *target = target_ptr;
270 target->status = status;
272 printk(KERN_ERR PFX "Got failed path rec status %d\n", status);
274 target->path = *pathrec;
275 complete(&target->done);
278 static int srp_lookup_path(struct srp_target_port *target)
280 target->path.numb_path = 1;
282 init_completion(&target->done);
284 target->path_query_id = ib_sa_path_rec_get(&srp_sa_client,
285 target->srp_host->dev->dev,
286 target->srp_host->port,
288 IB_SA_PATH_REC_DGID |
289 IB_SA_PATH_REC_SGID |
290 IB_SA_PATH_REC_NUMB_PATH |
292 SRP_PATH_REC_TIMEOUT_MS,
294 srp_path_rec_completion,
295 target, &target->path_query);
296 if (target->path_query_id < 0)
297 return target->path_query_id;
299 wait_for_completion(&target->done);
301 if (target->status < 0)
302 printk(KERN_WARNING PFX "Path record query failed\n");
304 return target->status;
307 static int srp_send_req(struct srp_target_port *target)
310 struct ib_cm_req_param param;
311 struct srp_login_req priv;
315 req = kzalloc(sizeof *req, GFP_KERNEL);
319 req->param.primary_path = &target->path;
320 req->param.alternate_path = NULL;
321 req->param.service_id = target->service_id;
322 req->param.qp_num = target->qp->qp_num;
323 req->param.qp_type = target->qp->qp_type;
324 req->param.private_data = &req->priv;
325 req->param.private_data_len = sizeof req->priv;
326 req->param.flow_control = 1;
328 get_random_bytes(&req->param.starting_psn, 4);
329 req->param.starting_psn &= 0xffffff;
332 * Pick some arbitrary defaults here; we could make these
333 * module parameters if anyone cared about setting them.
335 req->param.responder_resources = 4;
336 req->param.remote_cm_response_timeout = 20;
337 req->param.local_cm_response_timeout = 20;
338 req->param.retry_count = 7;
339 req->param.rnr_retry_count = 7;
340 req->param.max_cm_retries = 15;
342 req->priv.opcode = SRP_LOGIN_REQ;
344 req->priv.req_it_iu_len = cpu_to_be32(srp_max_iu_len);
345 req->priv.req_buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
346 SRP_BUF_FORMAT_INDIRECT);
348 * In the published SRP specification (draft rev. 16a), the
349 * port identifier format is 8 bytes of ID extension followed
350 * by 8 bytes of GUID. Older drafts put the two halves in the
351 * opposite order, so that the GUID comes first.
353 * Targets conforming to these obsolete drafts can be
354 * recognized by the I/O Class they report.
356 if (target->io_class == SRP_REV10_IB_IO_CLASS) {
357 memcpy(req->priv.initiator_port_id,
358 &target->path.sgid.global.interface_id, 8);
359 memcpy(req->priv.initiator_port_id + 8,
360 &target->initiator_ext, 8);
361 memcpy(req->priv.target_port_id, &target->ioc_guid, 8);
362 memcpy(req->priv.target_port_id + 8, &target->id_ext, 8);
364 memcpy(req->priv.initiator_port_id,
365 &target->initiator_ext, 8);
366 memcpy(req->priv.initiator_port_id + 8,
367 &target->path.sgid.global.interface_id, 8);
368 memcpy(req->priv.target_port_id, &target->id_ext, 8);
369 memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
373 * Topspin/Cisco SRP targets will reject our login unless we
374 * zero out the first 8 bytes of our initiator port ID and set
375 * the second 8 bytes to the local node GUID.
377 if (srp_target_is_topspin(target)) {
378 printk(KERN_DEBUG PFX "Topspin/Cisco initiator port ID workaround "
379 "activated for target GUID %016llx\n",
380 (unsigned long long) be64_to_cpu(target->ioc_guid));
381 memset(req->priv.initiator_port_id, 0, 8);
382 memcpy(req->priv.initiator_port_id + 8,
383 &target->srp_host->dev->dev->node_guid, 8);
386 status = ib_send_cm_req(target->cm_id, &req->param);
393 static void srp_disconnect_target(struct srp_target_port *target)
395 /* XXX should send SRP_I_LOGOUT request */
397 init_completion(&target->done);
398 if (ib_send_cm_dreq(target->cm_id, NULL, 0)) {
399 printk(KERN_DEBUG PFX "Sending CM DREQ failed\n");
402 wait_for_completion(&target->done);
405 static void srp_remove_work(struct work_struct *work)
407 struct srp_target_port *target =
408 container_of(work, struct srp_target_port, work);
410 spin_lock_irq(target->scsi_host->host_lock);
411 if (target->state != SRP_TARGET_DEAD) {
412 spin_unlock_irq(target->scsi_host->host_lock);
415 target->state = SRP_TARGET_REMOVED;
416 spin_unlock_irq(target->scsi_host->host_lock);
418 spin_lock(&target->srp_host->target_lock);
419 list_del(&target->list);
420 spin_unlock(&target->srp_host->target_lock);
422 scsi_remove_host(target->scsi_host);
423 ib_destroy_cm_id(target->cm_id);
424 srp_free_target_ib(target);
425 scsi_host_put(target->scsi_host);
428 static int srp_connect_target(struct srp_target_port *target)
432 ret = srp_lookup_path(target);
437 init_completion(&target->done);
438 ret = srp_send_req(target);
441 wait_for_completion(&target->done);
444 * The CM event handling code will set status to
445 * SRP_PORT_REDIRECT if we get a port redirect REJ
446 * back, or SRP_DLID_REDIRECT if we get a lid/qp
449 switch (target->status) {
453 case SRP_PORT_REDIRECT:
454 ret = srp_lookup_path(target);
459 case SRP_DLID_REDIRECT:
463 return target->status;
468 static void srp_unmap_data(struct scsi_cmnd *scmnd,
469 struct srp_target_port *target,
470 struct srp_request *req)
472 if (!scsi_sglist(scmnd) ||
473 (scmnd->sc_data_direction != DMA_TO_DEVICE &&
474 scmnd->sc_data_direction != DMA_FROM_DEVICE))
478 ib_fmr_pool_unmap(req->fmr);
482 ib_dma_unmap_sg(target->srp_host->dev->dev, scsi_sglist(scmnd),
483 scsi_sg_count(scmnd), scmnd->sc_data_direction);
486 static void srp_remove_req(struct srp_target_port *target, struct srp_request *req)
488 srp_unmap_data(req->scmnd, target, req);
489 list_move_tail(&req->list, &target->free_reqs);
492 static void srp_reset_req(struct srp_target_port *target, struct srp_request *req)
494 req->scmnd->result = DID_RESET << 16;
495 req->scmnd->scsi_done(req->scmnd);
496 srp_remove_req(target, req);
499 static int srp_reconnect_target(struct srp_target_port *target)
501 struct ib_cm_id *new_cm_id;
502 struct ib_qp_attr qp_attr;
503 struct srp_request *req, *tmp;
507 spin_lock_irq(target->scsi_host->host_lock);
508 if (target->state != SRP_TARGET_LIVE) {
509 spin_unlock_irq(target->scsi_host->host_lock);
512 target->state = SRP_TARGET_CONNECTING;
513 spin_unlock_irq(target->scsi_host->host_lock);
515 srp_disconnect_target(target);
517 * Now get a new local CM ID so that we avoid confusing the
518 * target in case things are really fouled up.
520 new_cm_id = ib_create_cm_id(target->srp_host->dev->dev,
521 srp_cm_handler, target);
522 if (IS_ERR(new_cm_id)) {
523 ret = PTR_ERR(new_cm_id);
526 ib_destroy_cm_id(target->cm_id);
527 target->cm_id = new_cm_id;
529 qp_attr.qp_state = IB_QPS_RESET;
530 ret = ib_modify_qp(target->qp, &qp_attr, IB_QP_STATE);
534 ret = srp_init_qp(target, target->qp);
538 while (ib_poll_cq(target->cq, 1, &wc) > 0)
541 spin_lock_irq(target->scsi_host->host_lock);
542 list_for_each_entry_safe(req, tmp, &target->req_queue, list)
543 srp_reset_req(target, req);
544 spin_unlock_irq(target->scsi_host->host_lock);
550 target->qp_in_error = 0;
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;
597 struct scatterlist *sg;
602 if (srp_target_is_mellanox(target) &&
603 (ib_sg_dma_address(ibdev, &scat[0]) & ~dev->fmr_page_mask))
607 scsi_for_each_sg(req->scmnd, sg, sg_cnt, i) {
608 unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
610 if (ib_sg_dma_address(ibdev, sg) & ~dev->fmr_page_mask) {
616 if ((ib_sg_dma_address(ibdev, sg) + dma_len) &
617 ~dev->fmr_page_mask) {
627 page_cnt += len >> dev->fmr_page_shift;
628 if (page_cnt > SRP_FMR_SIZE)
631 dma_pages = kmalloc(sizeof (u64) * page_cnt, GFP_ATOMIC);
636 scsi_for_each_sg(req->scmnd, sg, sg_cnt, i) {
637 unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
639 for (j = 0; j < dma_len; j += dev->fmr_page_size)
640 dma_pages[page_cnt++] =
641 (ib_sg_dma_address(ibdev, sg) &
642 dev->fmr_page_mask) + j;
645 req->fmr = ib_fmr_pool_map_phys(dev->fmr_pool,
646 dma_pages, page_cnt, io_addr);
647 if (IS_ERR(req->fmr)) {
648 ret = PTR_ERR(req->fmr);
653 buf->va = cpu_to_be64(ib_sg_dma_address(ibdev, &scat[0]) &
654 ~dev->fmr_page_mask);
655 buf->key = cpu_to_be32(req->fmr->fmr->rkey);
656 buf->len = cpu_to_be32(len);
666 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_target_port *target,
667 struct srp_request *req)
669 struct scatterlist *scat;
670 struct srp_cmd *cmd = req->cmd->buf;
671 int len, nents, count;
672 u8 fmt = SRP_DATA_DESC_DIRECT;
673 struct srp_device *dev;
674 struct ib_device *ibdev;
676 if (!scsi_sglist(scmnd) || scmnd->sc_data_direction == DMA_NONE)
677 return sizeof (struct srp_cmd);
679 if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
680 scmnd->sc_data_direction != DMA_TO_DEVICE) {
681 printk(KERN_WARNING PFX "Unhandled data direction %d\n",
682 scmnd->sc_data_direction);
686 nents = scsi_sg_count(scmnd);
687 scat = scsi_sglist(scmnd);
689 dev = target->srp_host->dev;
692 count = ib_dma_map_sg(ibdev, scat, nents, scmnd->sc_data_direction);
694 fmt = SRP_DATA_DESC_DIRECT;
695 len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
699 * The midlayer only generated a single gather/scatter
700 * entry, or DMA mapping coalesced everything to a
701 * single entry. So a direct descriptor along with
702 * the DMA MR suffices.
704 struct srp_direct_buf *buf = (void *) cmd->add_data;
706 buf->va = cpu_to_be64(ib_sg_dma_address(ibdev, scat));
707 buf->key = cpu_to_be32(dev->mr->rkey);
708 buf->len = cpu_to_be32(ib_sg_dma_len(ibdev, scat));
709 } else if (srp_map_fmr(target, scat, count, req,
710 (void *) cmd->add_data)) {
712 * FMR mapping failed, and the scatterlist has more
713 * than one entry. Generate an indirect memory
716 struct srp_indirect_buf *buf = (void *) cmd->add_data;
717 struct scatterlist *sg;
721 fmt = SRP_DATA_DESC_INDIRECT;
722 len = sizeof (struct srp_cmd) +
723 sizeof (struct srp_indirect_buf) +
724 count * sizeof (struct srp_direct_buf);
726 scsi_for_each_sg(scmnd, sg, count, i) {
727 unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
729 buf->desc_list[i].va =
730 cpu_to_be64(ib_sg_dma_address(ibdev, sg));
731 buf->desc_list[i].key =
732 cpu_to_be32(dev->mr->rkey);
733 buf->desc_list[i].len = cpu_to_be32(dma_len);
737 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
738 cmd->data_out_desc_cnt = count;
740 cmd->data_in_desc_cnt = count;
743 cpu_to_be64(req->cmd->dma + sizeof *cmd + sizeof *buf);
744 buf->table_desc.key =
745 cpu_to_be32(target->srp_host->dev->mr->rkey);
746 buf->table_desc.len =
747 cpu_to_be32(count * sizeof (struct srp_direct_buf));
749 buf->len = cpu_to_be32(datalen);
752 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
753 cmd->buf_fmt = fmt << 4;
760 static void srp_process_rsp(struct srp_target_port *target, struct srp_rsp *rsp)
762 struct srp_request *req;
763 struct scsi_cmnd *scmnd;
767 delta = (s32) be32_to_cpu(rsp->req_lim_delta);
769 spin_lock_irqsave(target->scsi_host->host_lock, flags);
771 target->req_lim += delta;
773 req = &target->req_ring[rsp->tag & ~SRP_TAG_TSK_MGMT];
775 if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
776 if (be32_to_cpu(rsp->resp_data_len) < 4)
777 req->tsk_status = -1;
779 req->tsk_status = rsp->data[3];
780 complete(&req->done);
784 printk(KERN_ERR "Null scmnd for RSP w/tag %016llx\n",
785 (unsigned long long) rsp->tag);
786 scmnd->result = rsp->status;
788 if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
789 memcpy(scmnd->sense_buffer, rsp->data +
790 be32_to_cpu(rsp->resp_data_len),
791 min_t(int, be32_to_cpu(rsp->sense_data_len),
792 SCSI_SENSE_BUFFERSIZE));
795 if (rsp->flags & (SRP_RSP_FLAG_DOOVER | SRP_RSP_FLAG_DOUNDER))
796 scsi_set_resid(scmnd, be32_to_cpu(rsp->data_out_res_cnt));
797 else if (rsp->flags & (SRP_RSP_FLAG_DIOVER | SRP_RSP_FLAG_DIUNDER))
798 scsi_set_resid(scmnd, be32_to_cpu(rsp->data_in_res_cnt));
800 if (!req->tsk_mgmt) {
801 scmnd->host_scribble = (void *) -1L;
802 scmnd->scsi_done(scmnd);
804 srp_remove_req(target, req);
809 spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
812 static void srp_handle_recv(struct srp_target_port *target, struct ib_wc *wc)
814 struct ib_device *dev;
818 iu = target->rx_ring[wc->wr_id & ~SRP_OP_RECV];
820 dev = target->srp_host->dev->dev;
821 ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_ti_iu_len,
824 opcode = *(u8 *) iu->buf;
829 printk(KERN_ERR PFX "recv completion, opcode 0x%02x\n", opcode);
831 for (i = 0; i < wc->byte_len; ++i) {
833 printk(KERN_ERR " [%02x] ", i);
834 printk(" %02x", ((u8 *) iu->buf)[i]);
835 if ((i + 1) % 8 == 0)
839 if (wc->byte_len % 8)
845 srp_process_rsp(target, iu->buf);
849 /* XXX Handle target logout */
850 printk(KERN_WARNING PFX "Got target logout request\n");
854 printk(KERN_WARNING PFX "Unhandled SRP opcode 0x%02x\n", opcode);
858 ib_dma_sync_single_for_device(dev, iu->dma, target->max_ti_iu_len,
862 static void srp_completion(struct ib_cq *cq, void *target_ptr)
864 struct srp_target_port *target = target_ptr;
867 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
868 while (ib_poll_cq(cq, 1, &wc) > 0) {
870 printk(KERN_ERR PFX "failed %s status %d\n",
871 wc.wr_id & SRP_OP_RECV ? "receive" : "send",
873 target->qp_in_error = 1;
877 if (wc.wr_id & SRP_OP_RECV)
878 srp_handle_recv(target, &wc);
884 static int __srp_post_recv(struct srp_target_port *target)
888 struct ib_recv_wr wr, *bad_wr;
892 next = target->rx_head & (SRP_RQ_SIZE - 1);
893 wr.wr_id = next | SRP_OP_RECV;
894 iu = target->rx_ring[next];
897 list.length = iu->size;
898 list.lkey = target->srp_host->dev->mr->lkey;
904 ret = ib_post_recv(target->qp, &wr, &bad_wr);
911 static int srp_post_recv(struct srp_target_port *target)
916 spin_lock_irqsave(target->scsi_host->host_lock, flags);
917 ret = __srp_post_recv(target);
918 spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
924 * Must be called with target->scsi_host->host_lock held to protect
925 * req_lim and tx_head. Lock cannot be dropped between call here and
926 * call to __srp_post_send().
928 static struct srp_iu *__srp_get_tx_iu(struct srp_target_port *target)
930 if (target->tx_head - target->tx_tail >= SRP_SQ_SIZE)
933 if (unlikely(target->req_lim < 1))
934 ++target->zero_req_lim;
936 return target->tx_ring[target->tx_head & SRP_SQ_SIZE];
940 * Must be called with target->scsi_host->host_lock held to protect
941 * req_lim and tx_head.
943 static int __srp_post_send(struct srp_target_port *target,
944 struct srp_iu *iu, int len)
947 struct ib_send_wr wr, *bad_wr;
952 list.lkey = target->srp_host->dev->mr->lkey;
955 wr.wr_id = target->tx_head & SRP_SQ_SIZE;
958 wr.opcode = IB_WR_SEND;
959 wr.send_flags = IB_SEND_SIGNALED;
961 ret = ib_post_send(target->qp, &wr, &bad_wr);
971 static int srp_queuecommand(struct scsi_cmnd *scmnd,
972 void (*done)(struct scsi_cmnd *))
974 struct srp_target_port *target = host_to_target(scmnd->device->host);
975 struct srp_request *req;
978 struct ib_device *dev;
981 if (target->state == SRP_TARGET_CONNECTING)
984 if (target->state == SRP_TARGET_DEAD ||
985 target->state == SRP_TARGET_REMOVED) {
986 scmnd->result = DID_BAD_TARGET << 16;
991 iu = __srp_get_tx_iu(target);
995 dev = target->srp_host->dev->dev;
996 ib_dma_sync_single_for_cpu(dev, iu->dma, srp_max_iu_len,
999 req = list_entry(target->free_reqs.next, struct srp_request, list);
1001 scmnd->scsi_done = done;
1003 scmnd->host_scribble = (void *) (long) req->index;
1006 memset(cmd, 0, sizeof *cmd);
1008 cmd->opcode = SRP_CMD;
1009 cmd->lun = cpu_to_be64((u64) scmnd->device->lun << 48);
1010 cmd->tag = req->index;
1011 memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
1016 req->tsk_mgmt = NULL;
1018 len = srp_map_data(scmnd, target, req);
1020 printk(KERN_ERR PFX "Failed to map data\n");
1024 if (__srp_post_recv(target)) {
1025 printk(KERN_ERR PFX "Recv failed\n");
1029 ib_dma_sync_single_for_device(dev, iu->dma, srp_max_iu_len,
1032 if (__srp_post_send(target, iu, len)) {
1033 printk(KERN_ERR PFX "Send failed\n");
1037 list_move_tail(&req->list, &target->req_queue);
1042 srp_unmap_data(scmnd, target, req);
1045 return SCSI_MLQUEUE_HOST_BUSY;
1048 static int srp_alloc_iu_bufs(struct srp_target_port *target)
1052 for (i = 0; i < SRP_RQ_SIZE; ++i) {
1053 target->rx_ring[i] = srp_alloc_iu(target->srp_host,
1054 target->max_ti_iu_len,
1055 GFP_KERNEL, DMA_FROM_DEVICE);
1056 if (!target->rx_ring[i])
1060 for (i = 0; i < SRP_SQ_SIZE + 1; ++i) {
1061 target->tx_ring[i] = srp_alloc_iu(target->srp_host,
1063 GFP_KERNEL, DMA_TO_DEVICE);
1064 if (!target->tx_ring[i])
1071 for (i = 0; i < SRP_RQ_SIZE; ++i) {
1072 srp_free_iu(target->srp_host, target->rx_ring[i]);
1073 target->rx_ring[i] = NULL;
1076 for (i = 0; i < SRP_SQ_SIZE + 1; ++i) {
1077 srp_free_iu(target->srp_host, target->tx_ring[i]);
1078 target->tx_ring[i] = NULL;
1084 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
1085 struct ib_cm_event *event,
1086 struct srp_target_port *target)
1088 struct ib_class_port_info *cpi;
1091 switch (event->param.rej_rcvd.reason) {
1092 case IB_CM_REJ_PORT_CM_REDIRECT:
1093 cpi = event->param.rej_rcvd.ari;
1094 target->path.dlid = cpi->redirect_lid;
1095 target->path.pkey = cpi->redirect_pkey;
1096 cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
1097 memcpy(target->path.dgid.raw, cpi->redirect_gid, 16);
1099 target->status = target->path.dlid ?
1100 SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
1103 case IB_CM_REJ_PORT_REDIRECT:
1104 if (srp_target_is_topspin(target)) {
1106 * Topspin/Cisco SRP gateways incorrectly send
1107 * reject reason code 25 when they mean 24
1110 memcpy(target->path.dgid.raw,
1111 event->param.rej_rcvd.ari, 16);
1113 printk(KERN_DEBUG PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
1114 (unsigned long long) be64_to_cpu(target->path.dgid.global.subnet_prefix),
1115 (unsigned long long) be64_to_cpu(target->path.dgid.global.interface_id));
1117 target->status = SRP_PORT_REDIRECT;
1119 printk(KERN_WARNING " REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
1120 target->status = -ECONNRESET;
1124 case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
1125 printk(KERN_WARNING " REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
1126 target->status = -ECONNRESET;
1129 case IB_CM_REJ_CONSUMER_DEFINED:
1130 opcode = *(u8 *) event->private_data;
1131 if (opcode == SRP_LOGIN_REJ) {
1132 struct srp_login_rej *rej = event->private_data;
1133 u32 reason = be32_to_cpu(rej->reason);
1135 if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
1136 printk(KERN_WARNING PFX
1137 "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
1139 printk(KERN_WARNING PFX
1140 "SRP LOGIN REJECTED, reason 0x%08x\n", reason);
1142 printk(KERN_WARNING " REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
1143 " opcode 0x%02x\n", opcode);
1144 target->status = -ECONNRESET;
1148 printk(KERN_WARNING " REJ reason 0x%x\n",
1149 event->param.rej_rcvd.reason);
1150 target->status = -ECONNRESET;
1154 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
1156 struct srp_target_port *target = cm_id->context;
1157 struct ib_qp_attr *qp_attr = NULL;
1162 switch (event->event) {
1163 case IB_CM_REQ_ERROR:
1164 printk(KERN_DEBUG PFX "Sending CM REQ failed\n");
1166 target->status = -ECONNRESET;
1169 case IB_CM_REP_RECEIVED:
1171 opcode = *(u8 *) event->private_data;
1173 if (opcode == SRP_LOGIN_RSP) {
1174 struct srp_login_rsp *rsp = event->private_data;
1176 target->max_ti_iu_len = be32_to_cpu(rsp->max_ti_iu_len);
1177 target->req_lim = be32_to_cpu(rsp->req_lim_delta);
1179 target->scsi_host->can_queue = min(target->req_lim,
1180 target->scsi_host->can_queue);
1182 printk(KERN_WARNING PFX "Unhandled RSP opcode %#x\n", opcode);
1183 target->status = -ECONNRESET;
1187 if (!target->rx_ring[0]) {
1188 target->status = srp_alloc_iu_bufs(target);
1193 qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
1195 target->status = -ENOMEM;
1199 qp_attr->qp_state = IB_QPS_RTR;
1200 target->status = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1204 target->status = ib_modify_qp(target->qp, qp_attr, attr_mask);
1208 target->status = srp_post_recv(target);
1212 qp_attr->qp_state = IB_QPS_RTS;
1213 target->status = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1217 target->status = ib_modify_qp(target->qp, qp_attr, attr_mask);
1221 target->status = ib_send_cm_rtu(cm_id, NULL, 0);
1227 case IB_CM_REJ_RECEIVED:
1228 printk(KERN_DEBUG PFX "REJ received\n");
1231 srp_cm_rej_handler(cm_id, event, target);
1234 case IB_CM_DREQ_RECEIVED:
1235 printk(KERN_WARNING PFX "DREQ received - connection closed\n");
1236 if (ib_send_cm_drep(cm_id, NULL, 0))
1237 printk(KERN_ERR PFX "Sending CM DREP failed\n");
1240 case IB_CM_TIMEWAIT_EXIT:
1241 printk(KERN_ERR PFX "connection closed\n");
1247 case IB_CM_MRA_RECEIVED:
1248 case IB_CM_DREQ_ERROR:
1249 case IB_CM_DREP_RECEIVED:
1253 printk(KERN_WARNING PFX "Unhandled CM event %d\n", event->event);
1258 complete(&target->done);
1265 static int srp_send_tsk_mgmt(struct srp_target_port *target,
1266 struct srp_request *req, u8 func)
1269 struct srp_tsk_mgmt *tsk_mgmt;
1271 spin_lock_irq(target->scsi_host->host_lock);
1273 if (target->state == SRP_TARGET_DEAD ||
1274 target->state == SRP_TARGET_REMOVED) {
1275 req->scmnd->result = DID_BAD_TARGET << 16;
1279 init_completion(&req->done);
1281 iu = __srp_get_tx_iu(target);
1286 memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
1288 tsk_mgmt->opcode = SRP_TSK_MGMT;
1289 tsk_mgmt->lun = cpu_to_be64((u64) req->scmnd->device->lun << 48);
1290 tsk_mgmt->tag = req->index | SRP_TAG_TSK_MGMT;
1291 tsk_mgmt->tsk_mgmt_func = func;
1292 tsk_mgmt->task_tag = req->index;
1294 if (__srp_post_send(target, iu, sizeof *tsk_mgmt))
1299 spin_unlock_irq(target->scsi_host->host_lock);
1301 if (!wait_for_completion_timeout(&req->done,
1302 msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
1308 spin_unlock_irq(target->scsi_host->host_lock);
1312 static int srp_find_req(struct srp_target_port *target,
1313 struct scsi_cmnd *scmnd,
1314 struct srp_request **req)
1316 if (scmnd->host_scribble == (void *) -1L)
1319 *req = &target->req_ring[(long) scmnd->host_scribble];
1324 static int srp_abort(struct scsi_cmnd *scmnd)
1326 struct srp_target_port *target = host_to_target(scmnd->device->host);
1327 struct srp_request *req;
1330 printk(KERN_ERR "SRP abort called\n");
1332 if (target->qp_in_error)
1334 if (srp_find_req(target, scmnd, &req))
1336 if (srp_send_tsk_mgmt(target, req, SRP_TSK_ABORT_TASK))
1339 spin_lock_irq(target->scsi_host->host_lock);
1341 if (req->cmd_done) {
1342 srp_remove_req(target, req);
1343 scmnd->scsi_done(scmnd);
1344 } else if (!req->tsk_status) {
1345 srp_remove_req(target, req);
1346 scmnd->result = DID_ABORT << 16;
1350 spin_unlock_irq(target->scsi_host->host_lock);
1355 static int srp_reset_device(struct scsi_cmnd *scmnd)
1357 struct srp_target_port *target = host_to_target(scmnd->device->host);
1358 struct srp_request *req, *tmp;
1360 printk(KERN_ERR "SRP reset_device called\n");
1362 if (target->qp_in_error)
1364 if (srp_find_req(target, scmnd, &req))
1366 if (srp_send_tsk_mgmt(target, req, SRP_TSK_LUN_RESET))
1368 if (req->tsk_status)
1371 spin_lock_irq(target->scsi_host->host_lock);
1373 list_for_each_entry_safe(req, tmp, &target->req_queue, list)
1374 if (req->scmnd->device == scmnd->device)
1375 srp_reset_req(target, req);
1377 spin_unlock_irq(target->scsi_host->host_lock);
1382 static int srp_reset_host(struct scsi_cmnd *scmnd)
1384 struct srp_target_port *target = host_to_target(scmnd->device->host);
1387 printk(KERN_ERR PFX "SRP reset_host called\n");
1389 if (!srp_reconnect_target(target))
1395 static ssize_t show_id_ext(struct class_device *cdev, char *buf)
1397 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1399 if (target->state == SRP_TARGET_DEAD ||
1400 target->state == SRP_TARGET_REMOVED)
1403 return sprintf(buf, "0x%016llx\n",
1404 (unsigned long long) be64_to_cpu(target->id_ext));
1407 static ssize_t show_ioc_guid(struct class_device *cdev, char *buf)
1409 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1411 if (target->state == SRP_TARGET_DEAD ||
1412 target->state == SRP_TARGET_REMOVED)
1415 return sprintf(buf, "0x%016llx\n",
1416 (unsigned long long) be64_to_cpu(target->ioc_guid));
1419 static ssize_t show_service_id(struct class_device *cdev, char *buf)
1421 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1423 if (target->state == SRP_TARGET_DEAD ||
1424 target->state == SRP_TARGET_REMOVED)
1427 return sprintf(buf, "0x%016llx\n",
1428 (unsigned long long) be64_to_cpu(target->service_id));
1431 static ssize_t show_pkey(struct class_device *cdev, char *buf)
1433 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1435 if (target->state == SRP_TARGET_DEAD ||
1436 target->state == SRP_TARGET_REMOVED)
1439 return sprintf(buf, "0x%04x\n", be16_to_cpu(target->path.pkey));
1442 static ssize_t show_dgid(struct class_device *cdev, char *buf)
1444 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1446 if (target->state == SRP_TARGET_DEAD ||
1447 target->state == SRP_TARGET_REMOVED)
1450 return sprintf(buf, "%04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x\n",
1451 be16_to_cpu(((__be16 *) target->path.dgid.raw)[0]),
1452 be16_to_cpu(((__be16 *) target->path.dgid.raw)[1]),
1453 be16_to_cpu(((__be16 *) target->path.dgid.raw)[2]),
1454 be16_to_cpu(((__be16 *) target->path.dgid.raw)[3]),
1455 be16_to_cpu(((__be16 *) target->path.dgid.raw)[4]),
1456 be16_to_cpu(((__be16 *) target->path.dgid.raw)[5]),
1457 be16_to_cpu(((__be16 *) target->path.dgid.raw)[6]),
1458 be16_to_cpu(((__be16 *) target->path.dgid.raw)[7]));
1461 static ssize_t show_orig_dgid(struct class_device *cdev, char *buf)
1463 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1465 if (target->state == SRP_TARGET_DEAD ||
1466 target->state == SRP_TARGET_REMOVED)
1469 return sprintf(buf, "%04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x\n",
1470 be16_to_cpu(target->orig_dgid[0]),
1471 be16_to_cpu(target->orig_dgid[1]),
1472 be16_to_cpu(target->orig_dgid[2]),
1473 be16_to_cpu(target->orig_dgid[3]),
1474 be16_to_cpu(target->orig_dgid[4]),
1475 be16_to_cpu(target->orig_dgid[5]),
1476 be16_to_cpu(target->orig_dgid[6]),
1477 be16_to_cpu(target->orig_dgid[7]));
1480 static ssize_t show_zero_req_lim(struct class_device *cdev, char *buf)
1482 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1484 if (target->state == SRP_TARGET_DEAD ||
1485 target->state == SRP_TARGET_REMOVED)
1488 return sprintf(buf, "%d\n", target->zero_req_lim);
1491 static ssize_t show_local_ib_port(struct class_device *cdev, char *buf)
1493 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1495 return sprintf(buf, "%d\n", target->srp_host->port);
1498 static ssize_t show_local_ib_device(struct class_device *cdev, char *buf)
1500 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1502 return sprintf(buf, "%s\n", target->srp_host->dev->dev->name);
1505 static CLASS_DEVICE_ATTR(id_ext, S_IRUGO, show_id_ext, NULL);
1506 static CLASS_DEVICE_ATTR(ioc_guid, S_IRUGO, show_ioc_guid, NULL);
1507 static CLASS_DEVICE_ATTR(service_id, S_IRUGO, show_service_id, NULL);
1508 static CLASS_DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1509 static CLASS_DEVICE_ATTR(dgid, S_IRUGO, show_dgid, NULL);
1510 static CLASS_DEVICE_ATTR(orig_dgid, S_IRUGO, show_orig_dgid, NULL);
1511 static CLASS_DEVICE_ATTR(zero_req_lim, S_IRUGO, show_zero_req_lim, NULL);
1512 static CLASS_DEVICE_ATTR(local_ib_port, S_IRUGO, show_local_ib_port, NULL);
1513 static CLASS_DEVICE_ATTR(local_ib_device, S_IRUGO, show_local_ib_device, NULL);
1515 static struct class_device_attribute *srp_host_attrs[] = {
1516 &class_device_attr_id_ext,
1517 &class_device_attr_ioc_guid,
1518 &class_device_attr_service_id,
1519 &class_device_attr_pkey,
1520 &class_device_attr_dgid,
1521 &class_device_attr_orig_dgid,
1522 &class_device_attr_zero_req_lim,
1523 &class_device_attr_local_ib_port,
1524 &class_device_attr_local_ib_device,
1528 static struct scsi_host_template srp_template = {
1529 .module = THIS_MODULE,
1530 .name = "InfiniBand SRP initiator",
1531 .proc_name = DRV_NAME,
1532 .info = srp_target_info,
1533 .queuecommand = srp_queuecommand,
1534 .eh_abort_handler = srp_abort,
1535 .eh_device_reset_handler = srp_reset_device,
1536 .eh_host_reset_handler = srp_reset_host,
1537 .can_queue = SRP_SQ_SIZE,
1539 .cmd_per_lun = SRP_SQ_SIZE,
1540 .use_clustering = ENABLE_CLUSTERING,
1541 .shost_attrs = srp_host_attrs
1544 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
1546 sprintf(target->target_name, "SRP.T10:%016llX",
1547 (unsigned long long) be64_to_cpu(target->id_ext));
1549 if (scsi_add_host(target->scsi_host, host->dev->dev->dma_device))
1552 spin_lock(&host->target_lock);
1553 list_add_tail(&target->list, &host->target_list);
1554 spin_unlock(&host->target_lock);
1556 target->state = SRP_TARGET_LIVE;
1558 scsi_scan_target(&target->scsi_host->shost_gendev,
1559 0, target->scsi_id, SCAN_WILD_CARD, 0);
1564 static void srp_release_class_dev(struct class_device *class_dev)
1566 struct srp_host *host =
1567 container_of(class_dev, struct srp_host, class_dev);
1569 complete(&host->released);
1572 static struct class srp_class = {
1573 .name = "infiniband_srp",
1574 .release = srp_release_class_dev
1578 * Target ports are added by writing
1580 * id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
1581 * pkey=<P_Key>,service_id=<service ID>
1583 * to the add_target sysfs attribute.
1587 SRP_OPT_ID_EXT = 1 << 0,
1588 SRP_OPT_IOC_GUID = 1 << 1,
1589 SRP_OPT_DGID = 1 << 2,
1590 SRP_OPT_PKEY = 1 << 3,
1591 SRP_OPT_SERVICE_ID = 1 << 4,
1592 SRP_OPT_MAX_SECT = 1 << 5,
1593 SRP_OPT_MAX_CMD_PER_LUN = 1 << 6,
1594 SRP_OPT_IO_CLASS = 1 << 7,
1595 SRP_OPT_INITIATOR_EXT = 1 << 8,
1596 SRP_OPT_ALL = (SRP_OPT_ID_EXT |
1600 SRP_OPT_SERVICE_ID),
1603 static match_table_t srp_opt_tokens = {
1604 { SRP_OPT_ID_EXT, "id_ext=%s" },
1605 { SRP_OPT_IOC_GUID, "ioc_guid=%s" },
1606 { SRP_OPT_DGID, "dgid=%s" },
1607 { SRP_OPT_PKEY, "pkey=%x" },
1608 { SRP_OPT_SERVICE_ID, "service_id=%s" },
1609 { SRP_OPT_MAX_SECT, "max_sect=%d" },
1610 { SRP_OPT_MAX_CMD_PER_LUN, "max_cmd_per_lun=%d" },
1611 { SRP_OPT_IO_CLASS, "io_class=%x" },
1612 { SRP_OPT_INITIATOR_EXT, "initiator_ext=%s" },
1613 { SRP_OPT_ERR, NULL }
1616 static int srp_parse_options(const char *buf, struct srp_target_port *target)
1618 char *options, *sep_opt;
1621 substring_t args[MAX_OPT_ARGS];
1627 options = kstrdup(buf, GFP_KERNEL);
1632 while ((p = strsep(&sep_opt, ",")) != NULL) {
1636 token = match_token(p, srp_opt_tokens, args);
1640 case SRP_OPT_ID_EXT:
1641 p = match_strdup(args);
1646 target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
1650 case SRP_OPT_IOC_GUID:
1651 p = match_strdup(args);
1656 target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
1661 p = match_strdup(args);
1666 if (strlen(p) != 32) {
1667 printk(KERN_WARNING PFX "bad dest GID parameter '%s'\n", p);
1672 for (i = 0; i < 16; ++i) {
1673 strlcpy(dgid, p + i * 2, 3);
1674 target->path.dgid.raw[i] = simple_strtoul(dgid, NULL, 16);
1677 memcpy(target->orig_dgid, target->path.dgid.raw, 16);
1681 if (match_hex(args, &token)) {
1682 printk(KERN_WARNING PFX "bad P_Key parameter '%s'\n", p);
1685 target->path.pkey = cpu_to_be16(token);
1688 case SRP_OPT_SERVICE_ID:
1689 p = match_strdup(args);
1694 target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
1698 case SRP_OPT_MAX_SECT:
1699 if (match_int(args, &token)) {
1700 printk(KERN_WARNING PFX "bad max sect parameter '%s'\n", p);
1703 target->scsi_host->max_sectors = token;
1706 case SRP_OPT_MAX_CMD_PER_LUN:
1707 if (match_int(args, &token)) {
1708 printk(KERN_WARNING PFX "bad max cmd_per_lun parameter '%s'\n", p);
1711 target->scsi_host->cmd_per_lun = min(token, SRP_SQ_SIZE);
1714 case SRP_OPT_IO_CLASS:
1715 if (match_hex(args, &token)) {
1716 printk(KERN_WARNING PFX "bad IO class parameter '%s' \n", p);
1719 if (token != SRP_REV10_IB_IO_CLASS &&
1720 token != SRP_REV16A_IB_IO_CLASS) {
1721 printk(KERN_WARNING PFX "unknown IO class parameter value"
1722 " %x specified (use %x or %x).\n",
1723 token, SRP_REV10_IB_IO_CLASS, SRP_REV16A_IB_IO_CLASS);
1726 target->io_class = token;
1729 case SRP_OPT_INITIATOR_EXT:
1730 p = match_strdup(args);
1735 target->initiator_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
1740 printk(KERN_WARNING PFX "unknown parameter or missing value "
1741 "'%s' in target creation request\n", p);
1746 if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
1749 for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
1750 if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
1751 !(srp_opt_tokens[i].token & opt_mask))
1752 printk(KERN_WARNING PFX "target creation request is "
1753 "missing parameter '%s'\n",
1754 srp_opt_tokens[i].pattern);
1761 static ssize_t srp_create_target(struct class_device *class_dev,
1762 const char *buf, size_t count)
1764 struct srp_host *host =
1765 container_of(class_dev, struct srp_host, class_dev);
1766 struct Scsi_Host *target_host;
1767 struct srp_target_port *target;
1771 target_host = scsi_host_alloc(&srp_template,
1772 sizeof (struct srp_target_port));
1776 target_host->max_lun = SRP_MAX_LUN;
1777 target_host->max_cmd_len = sizeof ((struct srp_cmd *) (void *) 0L)->cdb;
1779 target = host_to_target(target_host);
1781 target->io_class = SRP_REV16A_IB_IO_CLASS;
1782 target->scsi_host = target_host;
1783 target->srp_host = host;
1785 INIT_LIST_HEAD(&target->free_reqs);
1786 INIT_LIST_HEAD(&target->req_queue);
1787 for (i = 0; i < SRP_SQ_SIZE; ++i) {
1788 target->req_ring[i].index = i;
1789 list_add_tail(&target->req_ring[i].list, &target->free_reqs);
1792 ret = srp_parse_options(buf, target);
1796 ib_get_cached_gid(host->dev->dev, host->port, 0, &target->path.sgid);
1798 printk(KERN_DEBUG PFX "new target: id_ext %016llx ioc_guid %016llx pkey %04x "
1799 "service_id %016llx dgid %04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x\n",
1800 (unsigned long long) be64_to_cpu(target->id_ext),
1801 (unsigned long long) be64_to_cpu(target->ioc_guid),
1802 be16_to_cpu(target->path.pkey),
1803 (unsigned long long) be64_to_cpu(target->service_id),
1804 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[0]),
1805 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[2]),
1806 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[4]),
1807 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[6]),
1808 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[8]),
1809 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[10]),
1810 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[12]),
1811 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[14]));
1813 ret = srp_create_target_ib(target);
1817 target->cm_id = ib_create_cm_id(host->dev->dev, srp_cm_handler, target);
1818 if (IS_ERR(target->cm_id)) {
1819 ret = PTR_ERR(target->cm_id);
1823 target->qp_in_error = 0;
1824 ret = srp_connect_target(target);
1826 printk(KERN_ERR PFX "Connection failed\n");
1830 ret = srp_add_target(host, target);
1832 goto err_disconnect;
1837 srp_disconnect_target(target);
1840 ib_destroy_cm_id(target->cm_id);
1843 srp_free_target_ib(target);
1846 scsi_host_put(target_host);
1851 static CLASS_DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
1853 static ssize_t show_ibdev(struct class_device *class_dev, char *buf)
1855 struct srp_host *host =
1856 container_of(class_dev, struct srp_host, class_dev);
1858 return sprintf(buf, "%s\n", host->dev->dev->name);
1861 static CLASS_DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
1863 static ssize_t show_port(struct class_device *class_dev, char *buf)
1865 struct srp_host *host =
1866 container_of(class_dev, struct srp_host, class_dev);
1868 return sprintf(buf, "%d\n", host->port);
1871 static CLASS_DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
1873 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
1875 struct srp_host *host;
1877 host = kzalloc(sizeof *host, GFP_KERNEL);
1881 INIT_LIST_HEAD(&host->target_list);
1882 spin_lock_init(&host->target_lock);
1883 init_completion(&host->released);
1887 host->class_dev.class = &srp_class;
1888 host->class_dev.dev = device->dev->dma_device;
1889 snprintf(host->class_dev.class_id, BUS_ID_SIZE, "srp-%s-%d",
1890 device->dev->name, port);
1892 if (class_device_register(&host->class_dev))
1894 if (class_device_create_file(&host->class_dev, &class_device_attr_add_target))
1896 if (class_device_create_file(&host->class_dev, &class_device_attr_ibdev))
1898 if (class_device_create_file(&host->class_dev, &class_device_attr_port))
1904 class_device_unregister(&host->class_dev);
1912 static void srp_add_one(struct ib_device *device)
1914 struct srp_device *srp_dev;
1915 struct ib_device_attr *dev_attr;
1916 struct ib_fmr_pool_param fmr_param;
1917 struct srp_host *host;
1920 dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
1924 if (ib_query_device(device, dev_attr)) {
1925 printk(KERN_WARNING PFX "Query device failed for %s\n",
1930 srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
1935 * Use the smallest page size supported by the HCA, down to a
1936 * minimum of 512 bytes (which is the smallest sector that a
1937 * SCSI command will ever carry).
1939 srp_dev->fmr_page_shift = max(9, ffs(dev_attr->page_size_cap) - 1);
1940 srp_dev->fmr_page_size = 1 << srp_dev->fmr_page_shift;
1941 srp_dev->fmr_page_mask = ~((u64) srp_dev->fmr_page_size - 1);
1943 INIT_LIST_HEAD(&srp_dev->dev_list);
1945 srp_dev->dev = device;
1946 srp_dev->pd = ib_alloc_pd(device);
1947 if (IS_ERR(srp_dev->pd))
1950 srp_dev->mr = ib_get_dma_mr(srp_dev->pd,
1951 IB_ACCESS_LOCAL_WRITE |
1952 IB_ACCESS_REMOTE_READ |
1953 IB_ACCESS_REMOTE_WRITE);
1954 if (IS_ERR(srp_dev->mr))
1957 memset(&fmr_param, 0, sizeof fmr_param);
1958 fmr_param.pool_size = SRP_FMR_POOL_SIZE;
1959 fmr_param.dirty_watermark = SRP_FMR_DIRTY_SIZE;
1960 fmr_param.cache = 1;
1961 fmr_param.max_pages_per_fmr = SRP_FMR_SIZE;
1962 fmr_param.page_shift = srp_dev->fmr_page_shift;
1963 fmr_param.access = (IB_ACCESS_LOCAL_WRITE |
1964 IB_ACCESS_REMOTE_WRITE |
1965 IB_ACCESS_REMOTE_READ);
1967 srp_dev->fmr_pool = ib_create_fmr_pool(srp_dev->pd, &fmr_param);
1968 if (IS_ERR(srp_dev->fmr_pool))
1969 srp_dev->fmr_pool = NULL;
1971 if (device->node_type == RDMA_NODE_IB_SWITCH) {
1976 e = device->phys_port_cnt;
1979 for (p = s; p <= e; ++p) {
1980 host = srp_add_port(srp_dev, p);
1982 list_add_tail(&host->list, &srp_dev->dev_list);
1985 ib_set_client_data(device, &srp_client, srp_dev);
1990 ib_dealloc_pd(srp_dev->pd);
1999 static void srp_remove_one(struct ib_device *device)
2001 struct srp_device *srp_dev;
2002 struct srp_host *host, *tmp_host;
2003 LIST_HEAD(target_list);
2004 struct srp_target_port *target, *tmp_target;
2006 srp_dev = ib_get_client_data(device, &srp_client);
2008 list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
2009 class_device_unregister(&host->class_dev);
2011 * Wait for the sysfs entry to go away, so that no new
2012 * target ports can be created.
2014 wait_for_completion(&host->released);
2017 * Mark all target ports as removed, so we stop queueing
2018 * commands and don't try to reconnect.
2020 spin_lock(&host->target_lock);
2021 list_for_each_entry(target, &host->target_list, list) {
2022 spin_lock_irq(target->scsi_host->host_lock);
2023 target->state = SRP_TARGET_REMOVED;
2024 spin_unlock_irq(target->scsi_host->host_lock);
2026 spin_unlock(&host->target_lock);
2029 * Wait for any reconnection tasks that may have
2030 * started before we marked our target ports as
2031 * removed, and any target port removal tasks.
2033 flush_scheduled_work();
2035 list_for_each_entry_safe(target, tmp_target,
2036 &host->target_list, list) {
2037 scsi_remove_host(target->scsi_host);
2038 srp_disconnect_target(target);
2039 ib_destroy_cm_id(target->cm_id);
2040 srp_free_target_ib(target);
2041 scsi_host_put(target->scsi_host);
2047 if (srp_dev->fmr_pool)
2048 ib_destroy_fmr_pool(srp_dev->fmr_pool);
2049 ib_dereg_mr(srp_dev->mr);
2050 ib_dealloc_pd(srp_dev->pd);
2055 static int __init srp_init_module(void)
2059 srp_template.sg_tablesize = srp_sg_tablesize;
2060 srp_max_iu_len = (sizeof (struct srp_cmd) +
2061 sizeof (struct srp_indirect_buf) +
2062 srp_sg_tablesize * 16);
2064 ret = class_register(&srp_class);
2066 printk(KERN_ERR PFX "couldn't register class infiniband_srp\n");
2070 ib_sa_register_client(&srp_sa_client);
2072 ret = ib_register_client(&srp_client);
2074 printk(KERN_ERR PFX "couldn't register IB client\n");
2075 ib_sa_unregister_client(&srp_sa_client);
2076 class_unregister(&srp_class);
2083 static void __exit srp_cleanup_module(void)
2085 ib_unregister_client(&srp_client);
2086 ib_sa_unregister_client(&srp_sa_client);
2087 class_unregister(&srp_class);
2090 module_init(srp_init_module);
2091 module_exit(srp_cleanup_module);