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, 0);
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 if (!scsi_sglist(scmnd) ||
459 (scmnd->sc_data_direction != DMA_TO_DEVICE &&
460 scmnd->sc_data_direction != DMA_FROM_DEVICE))
464 ib_fmr_pool_unmap(req->fmr);
468 ib_dma_unmap_sg(target->srp_host->dev->dev, scsi_sglist(scmnd),
469 scsi_sg_count(scmnd), scmnd->sc_data_direction);
472 static void srp_remove_req(struct srp_target_port *target, struct srp_request *req)
474 srp_unmap_data(req->scmnd, target, req);
475 list_move_tail(&req->list, &target->free_reqs);
478 static void srp_reset_req(struct srp_target_port *target, struct srp_request *req)
480 req->scmnd->result = DID_RESET << 16;
481 req->scmnd->scsi_done(req->scmnd);
482 srp_remove_req(target, req);
485 static int srp_reconnect_target(struct srp_target_port *target)
487 struct ib_cm_id *new_cm_id;
488 struct ib_qp_attr qp_attr;
489 struct srp_request *req, *tmp;
493 spin_lock_irq(target->scsi_host->host_lock);
494 if (target->state != SRP_TARGET_LIVE) {
495 spin_unlock_irq(target->scsi_host->host_lock);
498 target->state = SRP_TARGET_CONNECTING;
499 spin_unlock_irq(target->scsi_host->host_lock);
501 srp_disconnect_target(target);
503 * Now get a new local CM ID so that we avoid confusing the
504 * target in case things are really fouled up.
506 new_cm_id = ib_create_cm_id(target->srp_host->dev->dev,
507 srp_cm_handler, target);
508 if (IS_ERR(new_cm_id)) {
509 ret = PTR_ERR(new_cm_id);
512 ib_destroy_cm_id(target->cm_id);
513 target->cm_id = new_cm_id;
515 qp_attr.qp_state = IB_QPS_RESET;
516 ret = ib_modify_qp(target->qp, &qp_attr, IB_QP_STATE);
520 ret = srp_init_qp(target, target->qp);
524 while (ib_poll_cq(target->cq, 1, &wc) > 0)
527 spin_lock_irq(target->scsi_host->host_lock);
528 list_for_each_entry_safe(req, tmp, &target->req_queue, list)
529 srp_reset_req(target, req);
530 spin_unlock_irq(target->scsi_host->host_lock);
536 target->qp_in_error = 0;
537 ret = srp_connect_target(target);
541 spin_lock_irq(target->scsi_host->host_lock);
542 if (target->state == SRP_TARGET_CONNECTING) {
544 target->state = SRP_TARGET_LIVE;
547 spin_unlock_irq(target->scsi_host->host_lock);
552 printk(KERN_ERR PFX "reconnect failed (%d), removing target port.\n", ret);
555 * We couldn't reconnect, so kill our target port off.
556 * However, we have to defer the real removal because we might
557 * be in the context of the SCSI error handler now, which
558 * would deadlock if we call scsi_remove_host().
560 spin_lock_irq(target->scsi_host->host_lock);
561 if (target->state == SRP_TARGET_CONNECTING) {
562 target->state = SRP_TARGET_DEAD;
563 INIT_WORK(&target->work, srp_remove_work);
564 schedule_work(&target->work);
566 spin_unlock_irq(target->scsi_host->host_lock);
571 static int srp_map_fmr(struct srp_target_port *target, struct scatterlist *scat,
572 int sg_cnt, struct srp_request *req,
573 struct srp_direct_buf *buf)
581 struct srp_device *dev = target->srp_host->dev;
582 struct ib_device *ibdev = dev->dev;
583 struct scatterlist *sg;
588 if ((ib_sg_dma_address(ibdev, &scat[0]) & ~dev->fmr_page_mask) &&
589 mellanox_workarounds && !memcmp(&target->ioc_guid, mellanox_oui, 3))
593 scsi_for_each_sg(req->scmnd, sg, sg_cnt, i) {
594 unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
596 if (ib_sg_dma_address(ibdev, sg) & ~dev->fmr_page_mask) {
602 if ((ib_sg_dma_address(ibdev, sg) + dma_len) &
603 ~dev->fmr_page_mask) {
613 page_cnt += len >> dev->fmr_page_shift;
614 if (page_cnt > SRP_FMR_SIZE)
617 dma_pages = kmalloc(sizeof (u64) * page_cnt, GFP_ATOMIC);
622 scsi_for_each_sg(req->scmnd, sg, sg_cnt, i) {
623 unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
625 for (j = 0; j < dma_len; j += dev->fmr_page_size)
626 dma_pages[page_cnt++] =
627 (ib_sg_dma_address(ibdev, sg) &
628 dev->fmr_page_mask) + j;
631 req->fmr = ib_fmr_pool_map_phys(dev->fmr_pool,
632 dma_pages, page_cnt, io_addr);
633 if (IS_ERR(req->fmr)) {
634 ret = PTR_ERR(req->fmr);
639 buf->va = cpu_to_be64(ib_sg_dma_address(ibdev, &scat[0]) &
640 ~dev->fmr_page_mask);
641 buf->key = cpu_to_be32(req->fmr->fmr->rkey);
642 buf->len = cpu_to_be32(len);
652 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_target_port *target,
653 struct srp_request *req)
655 struct scatterlist *scat;
656 struct srp_cmd *cmd = req->cmd->buf;
657 int len, nents, count;
658 u8 fmt = SRP_DATA_DESC_DIRECT;
659 struct srp_device *dev;
660 struct ib_device *ibdev;
662 if (!scsi_sglist(scmnd) || scmnd->sc_data_direction == DMA_NONE)
663 return sizeof (struct srp_cmd);
665 if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
666 scmnd->sc_data_direction != DMA_TO_DEVICE) {
667 printk(KERN_WARNING PFX "Unhandled data direction %d\n",
668 scmnd->sc_data_direction);
672 nents = scsi_sg_count(scmnd);
673 scat = scsi_sglist(scmnd);
675 dev = target->srp_host->dev;
678 count = ib_dma_map_sg(ibdev, scat, nents, scmnd->sc_data_direction);
680 fmt = SRP_DATA_DESC_DIRECT;
681 len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
685 * The midlayer only generated a single gather/scatter
686 * entry, or DMA mapping coalesced everything to a
687 * single entry. So a direct descriptor along with
688 * the DMA MR suffices.
690 struct srp_direct_buf *buf = (void *) cmd->add_data;
692 buf->va = cpu_to_be64(ib_sg_dma_address(ibdev, scat));
693 buf->key = cpu_to_be32(dev->mr->rkey);
694 buf->len = cpu_to_be32(ib_sg_dma_len(ibdev, scat));
695 } else if (srp_map_fmr(target, scat, count, req,
696 (void *) cmd->add_data)) {
698 * FMR mapping failed, and the scatterlist has more
699 * than one entry. Generate an indirect memory
702 struct srp_indirect_buf *buf = (void *) cmd->add_data;
703 struct scatterlist *sg;
707 fmt = SRP_DATA_DESC_INDIRECT;
708 len = sizeof (struct srp_cmd) +
709 sizeof (struct srp_indirect_buf) +
710 count * sizeof (struct srp_direct_buf);
712 scsi_for_each_sg(scmnd, sg, count, i) {
713 unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
715 buf->desc_list[i].va =
716 cpu_to_be64(ib_sg_dma_address(ibdev, sg));
717 buf->desc_list[i].key =
718 cpu_to_be32(dev->mr->rkey);
719 buf->desc_list[i].len = cpu_to_be32(dma_len);
723 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
724 cmd->data_out_desc_cnt = count;
726 cmd->data_in_desc_cnt = count;
729 cpu_to_be64(req->cmd->dma + sizeof *cmd + sizeof *buf);
730 buf->table_desc.key =
731 cpu_to_be32(target->srp_host->dev->mr->rkey);
732 buf->table_desc.len =
733 cpu_to_be32(count * sizeof (struct srp_direct_buf));
735 buf->len = cpu_to_be32(datalen);
738 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
739 cmd->buf_fmt = fmt << 4;
746 static void srp_process_rsp(struct srp_target_port *target, struct srp_rsp *rsp)
748 struct srp_request *req;
749 struct scsi_cmnd *scmnd;
753 delta = (s32) be32_to_cpu(rsp->req_lim_delta);
755 spin_lock_irqsave(target->scsi_host->host_lock, flags);
757 target->req_lim += delta;
759 req = &target->req_ring[rsp->tag & ~SRP_TAG_TSK_MGMT];
761 if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
762 if (be32_to_cpu(rsp->resp_data_len) < 4)
763 req->tsk_status = -1;
765 req->tsk_status = rsp->data[3];
766 complete(&req->done);
770 printk(KERN_ERR "Null scmnd for RSP w/tag %016llx\n",
771 (unsigned long long) rsp->tag);
772 scmnd->result = rsp->status;
774 if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
775 memcpy(scmnd->sense_buffer, rsp->data +
776 be32_to_cpu(rsp->resp_data_len),
777 min_t(int, be32_to_cpu(rsp->sense_data_len),
778 SCSI_SENSE_BUFFERSIZE));
781 if (rsp->flags & (SRP_RSP_FLAG_DOOVER | SRP_RSP_FLAG_DOUNDER))
782 scsi_set_resid(scmnd, be32_to_cpu(rsp->data_out_res_cnt));
783 else if (rsp->flags & (SRP_RSP_FLAG_DIOVER | SRP_RSP_FLAG_DIUNDER))
784 scsi_set_resid(scmnd, be32_to_cpu(rsp->data_in_res_cnt));
786 if (!req->tsk_mgmt) {
787 scmnd->host_scribble = (void *) -1L;
788 scmnd->scsi_done(scmnd);
790 srp_remove_req(target, req);
795 spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
798 static void srp_handle_recv(struct srp_target_port *target, struct ib_wc *wc)
800 struct ib_device *dev;
804 iu = target->rx_ring[wc->wr_id & ~SRP_OP_RECV];
806 dev = target->srp_host->dev->dev;
807 ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_ti_iu_len,
810 opcode = *(u8 *) iu->buf;
815 printk(KERN_ERR PFX "recv completion, opcode 0x%02x\n", opcode);
817 for (i = 0; i < wc->byte_len; ++i) {
819 printk(KERN_ERR " [%02x] ", i);
820 printk(" %02x", ((u8 *) iu->buf)[i]);
821 if ((i + 1) % 8 == 0)
825 if (wc->byte_len % 8)
831 srp_process_rsp(target, iu->buf);
835 /* XXX Handle target logout */
836 printk(KERN_WARNING PFX "Got target logout request\n");
840 printk(KERN_WARNING PFX "Unhandled SRP opcode 0x%02x\n", opcode);
844 ib_dma_sync_single_for_device(dev, iu->dma, target->max_ti_iu_len,
848 static void srp_completion(struct ib_cq *cq, void *target_ptr)
850 struct srp_target_port *target = target_ptr;
853 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
854 while (ib_poll_cq(cq, 1, &wc) > 0) {
856 printk(KERN_ERR PFX "failed %s status %d\n",
857 wc.wr_id & SRP_OP_RECV ? "receive" : "send",
859 target->qp_in_error = 1;
863 if (wc.wr_id & SRP_OP_RECV)
864 srp_handle_recv(target, &wc);
870 static int __srp_post_recv(struct srp_target_port *target)
874 struct ib_recv_wr wr, *bad_wr;
878 next = target->rx_head & (SRP_RQ_SIZE - 1);
879 wr.wr_id = next | SRP_OP_RECV;
880 iu = target->rx_ring[next];
883 list.length = iu->size;
884 list.lkey = target->srp_host->dev->mr->lkey;
890 ret = ib_post_recv(target->qp, &wr, &bad_wr);
897 static int srp_post_recv(struct srp_target_port *target)
902 spin_lock_irqsave(target->scsi_host->host_lock, flags);
903 ret = __srp_post_recv(target);
904 spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
910 * Must be called with target->scsi_host->host_lock held to protect
911 * req_lim and tx_head. Lock cannot be dropped between call here and
912 * call to __srp_post_send().
914 static struct srp_iu *__srp_get_tx_iu(struct srp_target_port *target)
916 if (target->tx_head - target->tx_tail >= SRP_SQ_SIZE)
919 if (unlikely(target->req_lim < 1))
920 ++target->zero_req_lim;
922 return target->tx_ring[target->tx_head & SRP_SQ_SIZE];
926 * Must be called with target->scsi_host->host_lock held to protect
927 * req_lim and tx_head.
929 static int __srp_post_send(struct srp_target_port *target,
930 struct srp_iu *iu, int len)
933 struct ib_send_wr wr, *bad_wr;
938 list.lkey = target->srp_host->dev->mr->lkey;
941 wr.wr_id = target->tx_head & SRP_SQ_SIZE;
944 wr.opcode = IB_WR_SEND;
945 wr.send_flags = IB_SEND_SIGNALED;
947 ret = ib_post_send(target->qp, &wr, &bad_wr);
957 static int srp_queuecommand(struct scsi_cmnd *scmnd,
958 void (*done)(struct scsi_cmnd *))
960 struct srp_target_port *target = host_to_target(scmnd->device->host);
961 struct srp_request *req;
964 struct ib_device *dev;
967 if (target->state == SRP_TARGET_CONNECTING)
970 if (target->state == SRP_TARGET_DEAD ||
971 target->state == SRP_TARGET_REMOVED) {
972 scmnd->result = DID_BAD_TARGET << 16;
977 iu = __srp_get_tx_iu(target);
981 dev = target->srp_host->dev->dev;
982 ib_dma_sync_single_for_cpu(dev, iu->dma, srp_max_iu_len,
985 req = list_entry(target->free_reqs.next, struct srp_request, list);
987 scmnd->scsi_done = done;
989 scmnd->host_scribble = (void *) (long) req->index;
992 memset(cmd, 0, sizeof *cmd);
994 cmd->opcode = SRP_CMD;
995 cmd->lun = cpu_to_be64((u64) scmnd->device->lun << 48);
996 cmd->tag = req->index;
997 memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
1002 req->tsk_mgmt = NULL;
1004 len = srp_map_data(scmnd, target, req);
1006 printk(KERN_ERR PFX "Failed to map data\n");
1010 if (__srp_post_recv(target)) {
1011 printk(KERN_ERR PFX "Recv failed\n");
1015 ib_dma_sync_single_for_device(dev, iu->dma, srp_max_iu_len,
1018 if (__srp_post_send(target, iu, len)) {
1019 printk(KERN_ERR PFX "Send failed\n");
1023 list_move_tail(&req->list, &target->req_queue);
1028 srp_unmap_data(scmnd, target, req);
1031 return SCSI_MLQUEUE_HOST_BUSY;
1034 static int srp_alloc_iu_bufs(struct srp_target_port *target)
1038 for (i = 0; i < SRP_RQ_SIZE; ++i) {
1039 target->rx_ring[i] = srp_alloc_iu(target->srp_host,
1040 target->max_ti_iu_len,
1041 GFP_KERNEL, DMA_FROM_DEVICE);
1042 if (!target->rx_ring[i])
1046 for (i = 0; i < SRP_SQ_SIZE + 1; ++i) {
1047 target->tx_ring[i] = srp_alloc_iu(target->srp_host,
1049 GFP_KERNEL, DMA_TO_DEVICE);
1050 if (!target->tx_ring[i])
1057 for (i = 0; i < SRP_RQ_SIZE; ++i) {
1058 srp_free_iu(target->srp_host, target->rx_ring[i]);
1059 target->rx_ring[i] = NULL;
1062 for (i = 0; i < SRP_SQ_SIZE + 1; ++i) {
1063 srp_free_iu(target->srp_host, target->tx_ring[i]);
1064 target->tx_ring[i] = NULL;
1070 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
1071 struct ib_cm_event *event,
1072 struct srp_target_port *target)
1074 struct ib_class_port_info *cpi;
1077 switch (event->param.rej_rcvd.reason) {
1078 case IB_CM_REJ_PORT_CM_REDIRECT:
1079 cpi = event->param.rej_rcvd.ari;
1080 target->path.dlid = cpi->redirect_lid;
1081 target->path.pkey = cpi->redirect_pkey;
1082 cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
1083 memcpy(target->path.dgid.raw, cpi->redirect_gid, 16);
1085 target->status = target->path.dlid ?
1086 SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
1089 case IB_CM_REJ_PORT_REDIRECT:
1090 if (topspin_workarounds &&
1091 !memcmp(&target->ioc_guid, topspin_oui, 3)) {
1093 * Topspin/Cisco SRP gateways incorrectly send
1094 * reject reason code 25 when they mean 24
1097 memcpy(target->path.dgid.raw,
1098 event->param.rej_rcvd.ari, 16);
1100 printk(KERN_DEBUG PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
1101 (unsigned long long) be64_to_cpu(target->path.dgid.global.subnet_prefix),
1102 (unsigned long long) be64_to_cpu(target->path.dgid.global.interface_id));
1104 target->status = SRP_PORT_REDIRECT;
1106 printk(KERN_WARNING " REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
1107 target->status = -ECONNRESET;
1111 case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
1112 printk(KERN_WARNING " REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
1113 target->status = -ECONNRESET;
1116 case IB_CM_REJ_CONSUMER_DEFINED:
1117 opcode = *(u8 *) event->private_data;
1118 if (opcode == SRP_LOGIN_REJ) {
1119 struct srp_login_rej *rej = event->private_data;
1120 u32 reason = be32_to_cpu(rej->reason);
1122 if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
1123 printk(KERN_WARNING PFX
1124 "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
1126 printk(KERN_WARNING PFX
1127 "SRP LOGIN REJECTED, reason 0x%08x\n", reason);
1129 printk(KERN_WARNING " REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
1130 " opcode 0x%02x\n", opcode);
1131 target->status = -ECONNRESET;
1135 printk(KERN_WARNING " REJ reason 0x%x\n",
1136 event->param.rej_rcvd.reason);
1137 target->status = -ECONNRESET;
1141 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
1143 struct srp_target_port *target = cm_id->context;
1144 struct ib_qp_attr *qp_attr = NULL;
1149 switch (event->event) {
1150 case IB_CM_REQ_ERROR:
1151 printk(KERN_DEBUG PFX "Sending CM REQ failed\n");
1153 target->status = -ECONNRESET;
1156 case IB_CM_REP_RECEIVED:
1158 opcode = *(u8 *) event->private_data;
1160 if (opcode == SRP_LOGIN_RSP) {
1161 struct srp_login_rsp *rsp = event->private_data;
1163 target->max_ti_iu_len = be32_to_cpu(rsp->max_ti_iu_len);
1164 target->req_lim = be32_to_cpu(rsp->req_lim_delta);
1166 target->scsi_host->can_queue = min(target->req_lim,
1167 target->scsi_host->can_queue);
1169 printk(KERN_WARNING PFX "Unhandled RSP opcode %#x\n", opcode);
1170 target->status = -ECONNRESET;
1174 if (!target->rx_ring[0]) {
1175 target->status = srp_alloc_iu_bufs(target);
1180 qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
1182 target->status = -ENOMEM;
1186 qp_attr->qp_state = IB_QPS_RTR;
1187 target->status = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1191 target->status = ib_modify_qp(target->qp, qp_attr, attr_mask);
1195 target->status = srp_post_recv(target);
1199 qp_attr->qp_state = IB_QPS_RTS;
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 = ib_send_cm_rtu(cm_id, NULL, 0);
1214 case IB_CM_REJ_RECEIVED:
1215 printk(KERN_DEBUG PFX "REJ received\n");
1218 srp_cm_rej_handler(cm_id, event, target);
1221 case IB_CM_DREQ_RECEIVED:
1222 printk(KERN_WARNING PFX "DREQ received - connection closed\n");
1223 if (ib_send_cm_drep(cm_id, NULL, 0))
1224 printk(KERN_ERR PFX "Sending CM DREP failed\n");
1227 case IB_CM_TIMEWAIT_EXIT:
1228 printk(KERN_ERR PFX "connection closed\n");
1234 case IB_CM_MRA_RECEIVED:
1235 case IB_CM_DREQ_ERROR:
1236 case IB_CM_DREP_RECEIVED:
1240 printk(KERN_WARNING PFX "Unhandled CM event %d\n", event->event);
1245 complete(&target->done);
1252 static int srp_send_tsk_mgmt(struct srp_target_port *target,
1253 struct srp_request *req, u8 func)
1256 struct srp_tsk_mgmt *tsk_mgmt;
1258 spin_lock_irq(target->scsi_host->host_lock);
1260 if (target->state == SRP_TARGET_DEAD ||
1261 target->state == SRP_TARGET_REMOVED) {
1262 req->scmnd->result = DID_BAD_TARGET << 16;
1266 init_completion(&req->done);
1268 iu = __srp_get_tx_iu(target);
1273 memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
1275 tsk_mgmt->opcode = SRP_TSK_MGMT;
1276 tsk_mgmt->lun = cpu_to_be64((u64) req->scmnd->device->lun << 48);
1277 tsk_mgmt->tag = req->index | SRP_TAG_TSK_MGMT;
1278 tsk_mgmt->tsk_mgmt_func = func;
1279 tsk_mgmt->task_tag = req->index;
1281 if (__srp_post_send(target, iu, sizeof *tsk_mgmt))
1286 spin_unlock_irq(target->scsi_host->host_lock);
1288 if (!wait_for_completion_timeout(&req->done,
1289 msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
1295 spin_unlock_irq(target->scsi_host->host_lock);
1299 static int srp_find_req(struct srp_target_port *target,
1300 struct scsi_cmnd *scmnd,
1301 struct srp_request **req)
1303 if (scmnd->host_scribble == (void *) -1L)
1306 *req = &target->req_ring[(long) scmnd->host_scribble];
1311 static int srp_abort(struct scsi_cmnd *scmnd)
1313 struct srp_target_port *target = host_to_target(scmnd->device->host);
1314 struct srp_request *req;
1317 printk(KERN_ERR "SRP abort called\n");
1319 if (target->qp_in_error)
1321 if (srp_find_req(target, scmnd, &req))
1323 if (srp_send_tsk_mgmt(target, req, SRP_TSK_ABORT_TASK))
1326 spin_lock_irq(target->scsi_host->host_lock);
1328 if (req->cmd_done) {
1329 srp_remove_req(target, req);
1330 scmnd->scsi_done(scmnd);
1331 } else if (!req->tsk_status) {
1332 srp_remove_req(target, req);
1333 scmnd->result = DID_ABORT << 16;
1337 spin_unlock_irq(target->scsi_host->host_lock);
1342 static int srp_reset_device(struct scsi_cmnd *scmnd)
1344 struct srp_target_port *target = host_to_target(scmnd->device->host);
1345 struct srp_request *req, *tmp;
1347 printk(KERN_ERR "SRP reset_device called\n");
1349 if (target->qp_in_error)
1351 if (srp_find_req(target, scmnd, &req))
1353 if (srp_send_tsk_mgmt(target, req, SRP_TSK_LUN_RESET))
1355 if (req->tsk_status)
1358 spin_lock_irq(target->scsi_host->host_lock);
1360 list_for_each_entry_safe(req, tmp, &target->req_queue, list)
1361 if (req->scmnd->device == scmnd->device)
1362 srp_reset_req(target, req);
1364 spin_unlock_irq(target->scsi_host->host_lock);
1369 static int srp_reset_host(struct scsi_cmnd *scmnd)
1371 struct srp_target_port *target = host_to_target(scmnd->device->host);
1374 printk(KERN_ERR PFX "SRP reset_host called\n");
1376 if (!srp_reconnect_target(target))
1382 static ssize_t show_id_ext(struct class_device *cdev, char *buf)
1384 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1386 if (target->state == SRP_TARGET_DEAD ||
1387 target->state == SRP_TARGET_REMOVED)
1390 return sprintf(buf, "0x%016llx\n",
1391 (unsigned long long) be64_to_cpu(target->id_ext));
1394 static ssize_t show_ioc_guid(struct class_device *cdev, char *buf)
1396 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1398 if (target->state == SRP_TARGET_DEAD ||
1399 target->state == SRP_TARGET_REMOVED)
1402 return sprintf(buf, "0x%016llx\n",
1403 (unsigned long long) be64_to_cpu(target->ioc_guid));
1406 static ssize_t show_service_id(struct class_device *cdev, char *buf)
1408 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1410 if (target->state == SRP_TARGET_DEAD ||
1411 target->state == SRP_TARGET_REMOVED)
1414 return sprintf(buf, "0x%016llx\n",
1415 (unsigned long long) be64_to_cpu(target->service_id));
1418 static ssize_t show_pkey(struct class_device *cdev, char *buf)
1420 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1422 if (target->state == SRP_TARGET_DEAD ||
1423 target->state == SRP_TARGET_REMOVED)
1426 return sprintf(buf, "0x%04x\n", be16_to_cpu(target->path.pkey));
1429 static ssize_t show_dgid(struct class_device *cdev, char *buf)
1431 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1433 if (target->state == SRP_TARGET_DEAD ||
1434 target->state == SRP_TARGET_REMOVED)
1437 return sprintf(buf, "%04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x\n",
1438 be16_to_cpu(((__be16 *) target->path.dgid.raw)[0]),
1439 be16_to_cpu(((__be16 *) target->path.dgid.raw)[1]),
1440 be16_to_cpu(((__be16 *) target->path.dgid.raw)[2]),
1441 be16_to_cpu(((__be16 *) target->path.dgid.raw)[3]),
1442 be16_to_cpu(((__be16 *) target->path.dgid.raw)[4]),
1443 be16_to_cpu(((__be16 *) target->path.dgid.raw)[5]),
1444 be16_to_cpu(((__be16 *) target->path.dgid.raw)[6]),
1445 be16_to_cpu(((__be16 *) target->path.dgid.raw)[7]));
1448 static ssize_t show_orig_dgid(struct class_device *cdev, char *buf)
1450 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1452 if (target->state == SRP_TARGET_DEAD ||
1453 target->state == SRP_TARGET_REMOVED)
1456 return sprintf(buf, "%04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x\n",
1457 be16_to_cpu(target->orig_dgid[0]),
1458 be16_to_cpu(target->orig_dgid[1]),
1459 be16_to_cpu(target->orig_dgid[2]),
1460 be16_to_cpu(target->orig_dgid[3]),
1461 be16_to_cpu(target->orig_dgid[4]),
1462 be16_to_cpu(target->orig_dgid[5]),
1463 be16_to_cpu(target->orig_dgid[6]),
1464 be16_to_cpu(target->orig_dgid[7]));
1467 static ssize_t show_zero_req_lim(struct class_device *cdev, char *buf)
1469 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1471 if (target->state == SRP_TARGET_DEAD ||
1472 target->state == SRP_TARGET_REMOVED)
1475 return sprintf(buf, "%d\n", target->zero_req_lim);
1478 static ssize_t show_local_ib_port(struct class_device *cdev, char *buf)
1480 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1482 return sprintf(buf, "%d\n", target->srp_host->port);
1485 static ssize_t show_local_ib_device(struct class_device *cdev, char *buf)
1487 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1489 return sprintf(buf, "%s\n", target->srp_host->dev->dev->name);
1492 static CLASS_DEVICE_ATTR(id_ext, S_IRUGO, show_id_ext, NULL);
1493 static CLASS_DEVICE_ATTR(ioc_guid, S_IRUGO, show_ioc_guid, NULL);
1494 static CLASS_DEVICE_ATTR(service_id, S_IRUGO, show_service_id, NULL);
1495 static CLASS_DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1496 static CLASS_DEVICE_ATTR(dgid, S_IRUGO, show_dgid, NULL);
1497 static CLASS_DEVICE_ATTR(orig_dgid, S_IRUGO, show_orig_dgid, NULL);
1498 static CLASS_DEVICE_ATTR(zero_req_lim, S_IRUGO, show_zero_req_lim, NULL);
1499 static CLASS_DEVICE_ATTR(local_ib_port, S_IRUGO, show_local_ib_port, NULL);
1500 static CLASS_DEVICE_ATTR(local_ib_device, S_IRUGO, show_local_ib_device, NULL);
1502 static struct class_device_attribute *srp_host_attrs[] = {
1503 &class_device_attr_id_ext,
1504 &class_device_attr_ioc_guid,
1505 &class_device_attr_service_id,
1506 &class_device_attr_pkey,
1507 &class_device_attr_dgid,
1508 &class_device_attr_orig_dgid,
1509 &class_device_attr_zero_req_lim,
1510 &class_device_attr_local_ib_port,
1511 &class_device_attr_local_ib_device,
1515 static struct scsi_host_template srp_template = {
1516 .module = THIS_MODULE,
1517 .name = "InfiniBand SRP initiator",
1518 .proc_name = DRV_NAME,
1519 .info = srp_target_info,
1520 .queuecommand = srp_queuecommand,
1521 .eh_abort_handler = srp_abort,
1522 .eh_device_reset_handler = srp_reset_device,
1523 .eh_host_reset_handler = srp_reset_host,
1524 .can_queue = SRP_SQ_SIZE,
1526 .cmd_per_lun = SRP_SQ_SIZE,
1527 .use_clustering = ENABLE_CLUSTERING,
1528 .shost_attrs = srp_host_attrs
1531 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
1533 sprintf(target->target_name, "SRP.T10:%016llX",
1534 (unsigned long long) be64_to_cpu(target->id_ext));
1536 if (scsi_add_host(target->scsi_host, host->dev->dev->dma_device))
1539 spin_lock(&host->target_lock);
1540 list_add_tail(&target->list, &host->target_list);
1541 spin_unlock(&host->target_lock);
1543 target->state = SRP_TARGET_LIVE;
1545 scsi_scan_target(&target->scsi_host->shost_gendev,
1546 0, target->scsi_id, SCAN_WILD_CARD, 0);
1551 static void srp_release_class_dev(struct class_device *class_dev)
1553 struct srp_host *host =
1554 container_of(class_dev, struct srp_host, class_dev);
1556 complete(&host->released);
1559 static struct class srp_class = {
1560 .name = "infiniband_srp",
1561 .release = srp_release_class_dev
1565 * Target ports are added by writing
1567 * id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
1568 * pkey=<P_Key>,service_id=<service ID>
1570 * to the add_target sysfs attribute.
1574 SRP_OPT_ID_EXT = 1 << 0,
1575 SRP_OPT_IOC_GUID = 1 << 1,
1576 SRP_OPT_DGID = 1 << 2,
1577 SRP_OPT_PKEY = 1 << 3,
1578 SRP_OPT_SERVICE_ID = 1 << 4,
1579 SRP_OPT_MAX_SECT = 1 << 5,
1580 SRP_OPT_MAX_CMD_PER_LUN = 1 << 6,
1581 SRP_OPT_IO_CLASS = 1 << 7,
1582 SRP_OPT_INITIATOR_EXT = 1 << 8,
1583 SRP_OPT_ALL = (SRP_OPT_ID_EXT |
1587 SRP_OPT_SERVICE_ID),
1590 static match_table_t srp_opt_tokens = {
1591 { SRP_OPT_ID_EXT, "id_ext=%s" },
1592 { SRP_OPT_IOC_GUID, "ioc_guid=%s" },
1593 { SRP_OPT_DGID, "dgid=%s" },
1594 { SRP_OPT_PKEY, "pkey=%x" },
1595 { SRP_OPT_SERVICE_ID, "service_id=%s" },
1596 { SRP_OPT_MAX_SECT, "max_sect=%d" },
1597 { SRP_OPT_MAX_CMD_PER_LUN, "max_cmd_per_lun=%d" },
1598 { SRP_OPT_IO_CLASS, "io_class=%x" },
1599 { SRP_OPT_INITIATOR_EXT, "initiator_ext=%s" },
1600 { SRP_OPT_ERR, NULL }
1603 static int srp_parse_options(const char *buf, struct srp_target_port *target)
1605 char *options, *sep_opt;
1608 substring_t args[MAX_OPT_ARGS];
1614 options = kstrdup(buf, GFP_KERNEL);
1619 while ((p = strsep(&sep_opt, ",")) != NULL) {
1623 token = match_token(p, srp_opt_tokens, args);
1627 case SRP_OPT_ID_EXT:
1628 p = match_strdup(args);
1633 target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
1637 case SRP_OPT_IOC_GUID:
1638 p = match_strdup(args);
1643 target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
1648 p = match_strdup(args);
1653 if (strlen(p) != 32) {
1654 printk(KERN_WARNING PFX "bad dest GID parameter '%s'\n", p);
1659 for (i = 0; i < 16; ++i) {
1660 strlcpy(dgid, p + i * 2, 3);
1661 target->path.dgid.raw[i] = simple_strtoul(dgid, NULL, 16);
1664 memcpy(target->orig_dgid, target->path.dgid.raw, 16);
1668 if (match_hex(args, &token)) {
1669 printk(KERN_WARNING PFX "bad P_Key parameter '%s'\n", p);
1672 target->path.pkey = cpu_to_be16(token);
1675 case SRP_OPT_SERVICE_ID:
1676 p = match_strdup(args);
1681 target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
1685 case SRP_OPT_MAX_SECT:
1686 if (match_int(args, &token)) {
1687 printk(KERN_WARNING PFX "bad max sect parameter '%s'\n", p);
1690 target->scsi_host->max_sectors = token;
1693 case SRP_OPT_MAX_CMD_PER_LUN:
1694 if (match_int(args, &token)) {
1695 printk(KERN_WARNING PFX "bad max cmd_per_lun parameter '%s'\n", p);
1698 target->scsi_host->cmd_per_lun = min(token, SRP_SQ_SIZE);
1701 case SRP_OPT_IO_CLASS:
1702 if (match_hex(args, &token)) {
1703 printk(KERN_WARNING PFX "bad IO class parameter '%s' \n", p);
1706 if (token != SRP_REV10_IB_IO_CLASS &&
1707 token != SRP_REV16A_IB_IO_CLASS) {
1708 printk(KERN_WARNING PFX "unknown IO class parameter value"
1709 " %x specified (use %x or %x).\n",
1710 token, SRP_REV10_IB_IO_CLASS, SRP_REV16A_IB_IO_CLASS);
1713 target->io_class = token;
1716 case SRP_OPT_INITIATOR_EXT:
1717 p = match_strdup(args);
1722 target->initiator_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
1727 printk(KERN_WARNING PFX "unknown parameter or missing value "
1728 "'%s' in target creation request\n", p);
1733 if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
1736 for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
1737 if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
1738 !(srp_opt_tokens[i].token & opt_mask))
1739 printk(KERN_WARNING PFX "target creation request is "
1740 "missing parameter '%s'\n",
1741 srp_opt_tokens[i].pattern);
1748 static ssize_t srp_create_target(struct class_device *class_dev,
1749 const char *buf, size_t count)
1751 struct srp_host *host =
1752 container_of(class_dev, struct srp_host, class_dev);
1753 struct Scsi_Host *target_host;
1754 struct srp_target_port *target;
1758 target_host = scsi_host_alloc(&srp_template,
1759 sizeof (struct srp_target_port));
1763 target_host->max_lun = SRP_MAX_LUN;
1764 target_host->max_cmd_len = sizeof ((struct srp_cmd *) (void *) 0L)->cdb;
1766 target = host_to_target(target_host);
1768 target->io_class = SRP_REV16A_IB_IO_CLASS;
1769 target->scsi_host = target_host;
1770 target->srp_host = host;
1772 INIT_LIST_HEAD(&target->free_reqs);
1773 INIT_LIST_HEAD(&target->req_queue);
1774 for (i = 0; i < SRP_SQ_SIZE; ++i) {
1775 target->req_ring[i].index = i;
1776 list_add_tail(&target->req_ring[i].list, &target->free_reqs);
1779 ret = srp_parse_options(buf, target);
1783 ib_get_cached_gid(host->dev->dev, host->port, 0, &target->path.sgid);
1785 printk(KERN_DEBUG PFX "new target: id_ext %016llx ioc_guid %016llx pkey %04x "
1786 "service_id %016llx dgid %04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x\n",
1787 (unsigned long long) be64_to_cpu(target->id_ext),
1788 (unsigned long long) be64_to_cpu(target->ioc_guid),
1789 be16_to_cpu(target->path.pkey),
1790 (unsigned long long) be64_to_cpu(target->service_id),
1791 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[0]),
1792 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[2]),
1793 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[4]),
1794 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[6]),
1795 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[8]),
1796 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[10]),
1797 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[12]),
1798 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[14]));
1800 ret = srp_create_target_ib(target);
1804 target->cm_id = ib_create_cm_id(host->dev->dev, srp_cm_handler, target);
1805 if (IS_ERR(target->cm_id)) {
1806 ret = PTR_ERR(target->cm_id);
1810 target->qp_in_error = 0;
1811 ret = srp_connect_target(target);
1813 printk(KERN_ERR PFX "Connection failed\n");
1817 ret = srp_add_target(host, target);
1819 goto err_disconnect;
1824 srp_disconnect_target(target);
1827 ib_destroy_cm_id(target->cm_id);
1830 srp_free_target_ib(target);
1833 scsi_host_put(target_host);
1838 static CLASS_DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
1840 static ssize_t show_ibdev(struct class_device *class_dev, char *buf)
1842 struct srp_host *host =
1843 container_of(class_dev, struct srp_host, class_dev);
1845 return sprintf(buf, "%s\n", host->dev->dev->name);
1848 static CLASS_DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
1850 static ssize_t show_port(struct class_device *class_dev, char *buf)
1852 struct srp_host *host =
1853 container_of(class_dev, struct srp_host, class_dev);
1855 return sprintf(buf, "%d\n", host->port);
1858 static CLASS_DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
1860 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
1862 struct srp_host *host;
1864 host = kzalloc(sizeof *host, GFP_KERNEL);
1868 INIT_LIST_HEAD(&host->target_list);
1869 spin_lock_init(&host->target_lock);
1870 init_completion(&host->released);
1874 host->class_dev.class = &srp_class;
1875 host->class_dev.dev = device->dev->dma_device;
1876 snprintf(host->class_dev.class_id, BUS_ID_SIZE, "srp-%s-%d",
1877 device->dev->name, port);
1879 if (class_device_register(&host->class_dev))
1881 if (class_device_create_file(&host->class_dev, &class_device_attr_add_target))
1883 if (class_device_create_file(&host->class_dev, &class_device_attr_ibdev))
1885 if (class_device_create_file(&host->class_dev, &class_device_attr_port))
1891 class_device_unregister(&host->class_dev);
1899 static void srp_add_one(struct ib_device *device)
1901 struct srp_device *srp_dev;
1902 struct ib_device_attr *dev_attr;
1903 struct ib_fmr_pool_param fmr_param;
1904 struct srp_host *host;
1907 dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
1911 if (ib_query_device(device, dev_attr)) {
1912 printk(KERN_WARNING PFX "Query device failed for %s\n",
1917 srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
1922 * Use the smallest page size supported by the HCA, down to a
1923 * minimum of 512 bytes (which is the smallest sector that a
1924 * SCSI command will ever carry).
1926 srp_dev->fmr_page_shift = max(9, ffs(dev_attr->page_size_cap) - 1);
1927 srp_dev->fmr_page_size = 1 << srp_dev->fmr_page_shift;
1928 srp_dev->fmr_page_mask = ~((u64) srp_dev->fmr_page_size - 1);
1930 INIT_LIST_HEAD(&srp_dev->dev_list);
1932 srp_dev->dev = device;
1933 srp_dev->pd = ib_alloc_pd(device);
1934 if (IS_ERR(srp_dev->pd))
1937 srp_dev->mr = ib_get_dma_mr(srp_dev->pd,
1938 IB_ACCESS_LOCAL_WRITE |
1939 IB_ACCESS_REMOTE_READ |
1940 IB_ACCESS_REMOTE_WRITE);
1941 if (IS_ERR(srp_dev->mr))
1944 memset(&fmr_param, 0, sizeof fmr_param);
1945 fmr_param.pool_size = SRP_FMR_POOL_SIZE;
1946 fmr_param.dirty_watermark = SRP_FMR_DIRTY_SIZE;
1947 fmr_param.cache = 1;
1948 fmr_param.max_pages_per_fmr = SRP_FMR_SIZE;
1949 fmr_param.page_shift = srp_dev->fmr_page_shift;
1950 fmr_param.access = (IB_ACCESS_LOCAL_WRITE |
1951 IB_ACCESS_REMOTE_WRITE |
1952 IB_ACCESS_REMOTE_READ);
1954 srp_dev->fmr_pool = ib_create_fmr_pool(srp_dev->pd, &fmr_param);
1955 if (IS_ERR(srp_dev->fmr_pool))
1956 srp_dev->fmr_pool = NULL;
1958 if (device->node_type == RDMA_NODE_IB_SWITCH) {
1963 e = device->phys_port_cnt;
1966 for (p = s; p <= e; ++p) {
1967 host = srp_add_port(srp_dev, p);
1969 list_add_tail(&host->list, &srp_dev->dev_list);
1972 ib_set_client_data(device, &srp_client, srp_dev);
1977 ib_dealloc_pd(srp_dev->pd);
1986 static void srp_remove_one(struct ib_device *device)
1988 struct srp_device *srp_dev;
1989 struct srp_host *host, *tmp_host;
1990 LIST_HEAD(target_list);
1991 struct srp_target_port *target, *tmp_target;
1993 srp_dev = ib_get_client_data(device, &srp_client);
1995 list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
1996 class_device_unregister(&host->class_dev);
1998 * Wait for the sysfs entry to go away, so that no new
1999 * target ports can be created.
2001 wait_for_completion(&host->released);
2004 * Mark all target ports as removed, so we stop queueing
2005 * commands and don't try to reconnect.
2007 spin_lock(&host->target_lock);
2008 list_for_each_entry(target, &host->target_list, list) {
2009 spin_lock_irq(target->scsi_host->host_lock);
2010 target->state = SRP_TARGET_REMOVED;
2011 spin_unlock_irq(target->scsi_host->host_lock);
2013 spin_unlock(&host->target_lock);
2016 * Wait for any reconnection tasks that may have
2017 * started before we marked our target ports as
2018 * removed, and any target port removal tasks.
2020 flush_scheduled_work();
2022 list_for_each_entry_safe(target, tmp_target,
2023 &host->target_list, list) {
2024 scsi_remove_host(target->scsi_host);
2025 srp_disconnect_target(target);
2026 ib_destroy_cm_id(target->cm_id);
2027 srp_free_target_ib(target);
2028 scsi_host_put(target->scsi_host);
2034 if (srp_dev->fmr_pool)
2035 ib_destroy_fmr_pool(srp_dev->fmr_pool);
2036 ib_dereg_mr(srp_dev->mr);
2037 ib_dealloc_pd(srp_dev->pd);
2042 static int __init srp_init_module(void)
2046 srp_template.sg_tablesize = srp_sg_tablesize;
2047 srp_max_iu_len = (sizeof (struct srp_cmd) +
2048 sizeof (struct srp_indirect_buf) +
2049 srp_sg_tablesize * 16);
2051 ret = class_register(&srp_class);
2053 printk(KERN_ERR PFX "couldn't register class infiniband_srp\n");
2057 ib_sa_register_client(&srp_sa_client);
2059 ret = ib_register_client(&srp_client);
2061 printk(KERN_ERR PFX "couldn't register IB client\n");
2062 ib_sa_unregister_client(&srp_sa_client);
2063 class_unregister(&srp_class);
2070 static void __exit srp_cleanup_module(void)
2072 ib_unregister_client(&srp_client);
2073 ib_sa_unregister_client(&srp_sa_client);
2074 class_unregister(&srp_class);
2077 module_init(srp_init_module);
2078 module_exit(srp_cleanup_module);