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 = dma_map_single(host->dev->dev->dma_device,
126 iu->buf, size, direction);
127 if (dma_mapping_error(iu->dma))
131 iu->direction = direction;
143 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
148 dma_unmap_single(host->dev->dev->dma_device,
149 iu->dma, iu->size, iu->direction);
154 static void srp_qp_event(struct ib_event *event, void *context)
156 printk(KERN_ERR PFX "QP event %d\n", event->event);
159 static int srp_init_qp(struct srp_target_port *target,
162 struct ib_qp_attr *attr;
165 attr = kmalloc(sizeof *attr, GFP_KERNEL);
169 ret = ib_find_cached_pkey(target->srp_host->dev->dev,
170 target->srp_host->port,
171 be16_to_cpu(target->path.pkey),
176 attr->qp_state = IB_QPS_INIT;
177 attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
178 IB_ACCESS_REMOTE_WRITE);
179 attr->port_num = target->srp_host->port;
181 ret = ib_modify_qp(qp, attr,
192 static int srp_create_target_ib(struct srp_target_port *target)
194 struct ib_qp_init_attr *init_attr;
197 init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
201 target->cq = ib_create_cq(target->srp_host->dev->dev, srp_completion,
202 NULL, target, SRP_CQ_SIZE);
203 if (IS_ERR(target->cq)) {
204 ret = PTR_ERR(target->cq);
208 ib_req_notify_cq(target->cq, IB_CQ_NEXT_COMP);
210 init_attr->event_handler = srp_qp_event;
211 init_attr->cap.max_send_wr = SRP_SQ_SIZE;
212 init_attr->cap.max_recv_wr = SRP_RQ_SIZE;
213 init_attr->cap.max_recv_sge = 1;
214 init_attr->cap.max_send_sge = 1;
215 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
216 init_attr->qp_type = IB_QPT_RC;
217 init_attr->send_cq = target->cq;
218 init_attr->recv_cq = target->cq;
220 target->qp = ib_create_qp(target->srp_host->dev->pd, init_attr);
221 if (IS_ERR(target->qp)) {
222 ret = PTR_ERR(target->qp);
223 ib_destroy_cq(target->cq);
227 ret = srp_init_qp(target, target->qp);
229 ib_destroy_qp(target->qp);
230 ib_destroy_cq(target->cq);
239 static void srp_free_target_ib(struct srp_target_port *target)
243 ib_destroy_qp(target->qp);
244 ib_destroy_cq(target->cq);
246 for (i = 0; i < SRP_RQ_SIZE; ++i)
247 srp_free_iu(target->srp_host, target->rx_ring[i]);
248 for (i = 0; i < SRP_SQ_SIZE + 1; ++i)
249 srp_free_iu(target->srp_host, target->tx_ring[i]);
252 static void srp_path_rec_completion(int status,
253 struct ib_sa_path_rec *pathrec,
256 struct srp_target_port *target = target_ptr;
258 target->status = status;
260 printk(KERN_ERR PFX "Got failed path rec status %d\n", status);
262 target->path = *pathrec;
263 complete(&target->done);
266 static int srp_lookup_path(struct srp_target_port *target)
268 target->path.numb_path = 1;
270 init_completion(&target->done);
272 target->path_query_id = ib_sa_path_rec_get(&srp_sa_client,
273 target->srp_host->dev->dev,
274 target->srp_host->port,
276 IB_SA_PATH_REC_DGID |
277 IB_SA_PATH_REC_SGID |
278 IB_SA_PATH_REC_NUMB_PATH |
280 SRP_PATH_REC_TIMEOUT_MS,
282 srp_path_rec_completion,
283 target, &target->path_query);
284 if (target->path_query_id < 0)
285 return target->path_query_id;
287 wait_for_completion(&target->done);
289 if (target->status < 0)
290 printk(KERN_WARNING PFX "Path record query failed\n");
292 return target->status;
295 static int srp_send_req(struct srp_target_port *target)
298 struct ib_cm_req_param param;
299 struct srp_login_req priv;
303 req = kzalloc(sizeof *req, GFP_KERNEL);
307 req->param.primary_path = &target->path;
308 req->param.alternate_path = NULL;
309 req->param.service_id = target->service_id;
310 req->param.qp_num = target->qp->qp_num;
311 req->param.qp_type = target->qp->qp_type;
312 req->param.private_data = &req->priv;
313 req->param.private_data_len = sizeof req->priv;
314 req->param.flow_control = 1;
316 get_random_bytes(&req->param.starting_psn, 4);
317 req->param.starting_psn &= 0xffffff;
320 * Pick some arbitrary defaults here; we could make these
321 * module parameters if anyone cared about setting them.
323 req->param.responder_resources = 4;
324 req->param.remote_cm_response_timeout = 20;
325 req->param.local_cm_response_timeout = 20;
326 req->param.retry_count = 7;
327 req->param.rnr_retry_count = 7;
328 req->param.max_cm_retries = 15;
330 req->priv.opcode = SRP_LOGIN_REQ;
332 req->priv.req_it_iu_len = cpu_to_be32(srp_max_iu_len);
333 req->priv.req_buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
334 SRP_BUF_FORMAT_INDIRECT);
336 * In the published SRP specification (draft rev. 16a), the
337 * port identifier format is 8 bytes of ID extension followed
338 * by 8 bytes of GUID. Older drafts put the two halves in the
339 * opposite order, so that the GUID comes first.
341 * Targets conforming to these obsolete drafts can be
342 * recognized by the I/O Class they report.
344 if (target->io_class == SRP_REV10_IB_IO_CLASS) {
345 memcpy(req->priv.initiator_port_id,
346 target->srp_host->initiator_port_id + 8, 8);
347 memcpy(req->priv.initiator_port_id + 8,
348 target->srp_host->initiator_port_id, 8);
349 memcpy(req->priv.target_port_id, &target->ioc_guid, 8);
350 memcpy(req->priv.target_port_id + 8, &target->id_ext, 8);
352 memcpy(req->priv.initiator_port_id,
353 target->srp_host->initiator_port_id, 16);
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. The
361 * second 8 bytes must be our local node GUID, but we always
364 if (topspin_workarounds && !memcmp(&target->ioc_guid, topspin_oui, 3)) {
365 printk(KERN_DEBUG PFX "Topspin/Cisco initiator port ID workaround "
366 "activated for target GUID %016llx\n",
367 (unsigned long long) be64_to_cpu(target->ioc_guid));
368 memset(req->priv.initiator_port_id, 0, 8);
371 status = ib_send_cm_req(target->cm_id, &req->param);
378 static void srp_disconnect_target(struct srp_target_port *target)
380 /* XXX should send SRP_I_LOGOUT request */
382 init_completion(&target->done);
383 if (ib_send_cm_dreq(target->cm_id, NULL, 0)) {
384 printk(KERN_DEBUG PFX "Sending CM DREQ failed\n");
387 wait_for_completion(&target->done);
390 static void srp_remove_work(void *target_ptr)
392 struct srp_target_port *target = target_ptr;
394 spin_lock_irq(target->scsi_host->host_lock);
395 if (target->state != SRP_TARGET_DEAD) {
396 spin_unlock_irq(target->scsi_host->host_lock);
399 target->state = SRP_TARGET_REMOVED;
400 spin_unlock_irq(target->scsi_host->host_lock);
402 spin_lock(&target->srp_host->target_lock);
403 list_del(&target->list);
404 spin_unlock(&target->srp_host->target_lock);
406 scsi_remove_host(target->scsi_host);
407 ib_destroy_cm_id(target->cm_id);
408 srp_free_target_ib(target);
409 scsi_host_put(target->scsi_host);
412 static int srp_connect_target(struct srp_target_port *target)
416 ret = srp_lookup_path(target);
421 init_completion(&target->done);
422 ret = srp_send_req(target);
425 wait_for_completion(&target->done);
428 * The CM event handling code will set status to
429 * SRP_PORT_REDIRECT if we get a port redirect REJ
430 * back, or SRP_DLID_REDIRECT if we get a lid/qp
433 switch (target->status) {
437 case SRP_PORT_REDIRECT:
438 ret = srp_lookup_path(target);
443 case SRP_DLID_REDIRECT:
447 return target->status;
452 static void srp_unmap_data(struct scsi_cmnd *scmnd,
453 struct srp_target_port *target,
454 struct srp_request *req)
456 struct scatterlist *scat;
459 if (!scmnd->request_buffer ||
460 (scmnd->sc_data_direction != DMA_TO_DEVICE &&
461 scmnd->sc_data_direction != DMA_FROM_DEVICE))
465 ib_fmr_pool_unmap(req->fmr);
470 * This handling of non-SG commands can be killed when the
471 * SCSI midlayer no longer generates non-SG commands.
473 if (likely(scmnd->use_sg)) {
474 nents = scmnd->use_sg;
475 scat = scmnd->request_buffer;
478 scat = &req->fake_sg;
481 dma_unmap_sg(target->srp_host->dev->dev->dma_device, scat, nents,
482 scmnd->sc_data_direction);
485 static void srp_remove_req(struct srp_target_port *target, struct srp_request *req)
487 srp_unmap_data(req->scmnd, target, req);
488 list_move_tail(&req->list, &target->free_reqs);
491 static void srp_reset_req(struct srp_target_port *target, struct srp_request *req)
493 req->scmnd->result = DID_RESET << 16;
494 req->scmnd->scsi_done(req->scmnd);
495 srp_remove_req(target, req);
498 static int srp_reconnect_target(struct srp_target_port *target)
500 struct ib_cm_id *new_cm_id;
501 struct ib_qp_attr qp_attr;
502 struct srp_request *req, *tmp;
506 spin_lock_irq(target->scsi_host->host_lock);
507 if (target->state != SRP_TARGET_LIVE) {
508 spin_unlock_irq(target->scsi_host->host_lock);
511 target->state = SRP_TARGET_CONNECTING;
512 spin_unlock_irq(target->scsi_host->host_lock);
514 srp_disconnect_target(target);
516 * Now get a new local CM ID so that we avoid confusing the
517 * target in case things are really fouled up.
519 new_cm_id = ib_create_cm_id(target->srp_host->dev->dev,
520 srp_cm_handler, target);
521 if (IS_ERR(new_cm_id)) {
522 ret = PTR_ERR(new_cm_id);
525 ib_destroy_cm_id(target->cm_id);
526 target->cm_id = new_cm_id;
528 qp_attr.qp_state = IB_QPS_RESET;
529 ret = ib_modify_qp(target->qp, &qp_attr, IB_QP_STATE);
533 ret = srp_init_qp(target, target->qp);
537 while (ib_poll_cq(target->cq, 1, &wc) > 0)
540 spin_lock_irq(target->scsi_host->host_lock);
541 list_for_each_entry_safe(req, tmp, &target->req_queue, list)
542 srp_reset_req(target, req);
543 spin_unlock_irq(target->scsi_host->host_lock);
549 ret = srp_connect_target(target);
553 spin_lock_irq(target->scsi_host->host_lock);
554 if (target->state == SRP_TARGET_CONNECTING) {
556 target->state = SRP_TARGET_LIVE;
559 spin_unlock_irq(target->scsi_host->host_lock);
564 printk(KERN_ERR PFX "reconnect failed (%d), removing target port.\n", ret);
567 * We couldn't reconnect, so kill our target port off.
568 * However, we have to defer the real removal because we might
569 * be in the context of the SCSI error handler now, which
570 * would deadlock if we call scsi_remove_host().
572 spin_lock_irq(target->scsi_host->host_lock);
573 if (target->state == SRP_TARGET_CONNECTING) {
574 target->state = SRP_TARGET_DEAD;
575 INIT_WORK(&target->work, srp_remove_work, target);
576 schedule_work(&target->work);
578 spin_unlock_irq(target->scsi_host->host_lock);
583 static int srp_map_fmr(struct srp_target_port *target, struct scatterlist *scat,
584 int sg_cnt, struct srp_request *req,
585 struct srp_direct_buf *buf)
593 struct srp_device *dev = target->srp_host->dev;
598 if ((sg_dma_address(&scat[0]) & ~dev->fmr_page_mask) &&
599 mellanox_workarounds && !memcmp(&target->ioc_guid, mellanox_oui, 3))
603 for (i = 0; i < sg_cnt; ++i) {
604 if (sg_dma_address(&scat[i]) & ~dev->fmr_page_mask) {
610 if ((sg_dma_address(&scat[i]) + sg_dma_len(&scat[i])) &
611 ~dev->fmr_page_mask) {
618 len += sg_dma_len(&scat[i]);
621 page_cnt += len >> dev->fmr_page_shift;
622 if (page_cnt > SRP_FMR_SIZE)
625 dma_pages = kmalloc(sizeof (u64) * page_cnt, GFP_ATOMIC);
630 for (i = 0; i < sg_cnt; ++i)
631 for (j = 0; j < sg_dma_len(&scat[i]); j += dev->fmr_page_size)
632 dma_pages[page_cnt++] =
633 (sg_dma_address(&scat[i]) & dev->fmr_page_mask) + j;
635 req->fmr = ib_fmr_pool_map_phys(dev->fmr_pool,
636 dma_pages, page_cnt, io_addr);
637 if (IS_ERR(req->fmr)) {
638 ret = PTR_ERR(req->fmr);
643 buf->va = cpu_to_be64(sg_dma_address(&scat[0]) & ~dev->fmr_page_mask);
644 buf->key = cpu_to_be32(req->fmr->fmr->rkey);
645 buf->len = cpu_to_be32(len);
655 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_target_port *target,
656 struct srp_request *req)
658 struct scatterlist *scat;
659 struct srp_cmd *cmd = req->cmd->buf;
660 int len, nents, count;
661 u8 fmt = SRP_DATA_DESC_DIRECT;
663 if (!scmnd->request_buffer || scmnd->sc_data_direction == DMA_NONE)
664 return sizeof (struct srp_cmd);
666 if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
667 scmnd->sc_data_direction != DMA_TO_DEVICE) {
668 printk(KERN_WARNING PFX "Unhandled data direction %d\n",
669 scmnd->sc_data_direction);
674 * This handling of non-SG commands can be killed when the
675 * SCSI midlayer no longer generates non-SG commands.
677 if (likely(scmnd->use_sg)) {
678 nents = scmnd->use_sg;
679 scat = scmnd->request_buffer;
682 scat = &req->fake_sg;
683 sg_init_one(scat, scmnd->request_buffer, scmnd->request_bufflen);
686 count = dma_map_sg(target->srp_host->dev->dev->dma_device,
687 scat, nents, scmnd->sc_data_direction);
689 fmt = SRP_DATA_DESC_DIRECT;
690 len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
694 * The midlayer only generated a single gather/scatter
695 * entry, or DMA mapping coalesced everything to a
696 * single entry. So a direct descriptor along with
697 * the DMA MR suffices.
699 struct srp_direct_buf *buf = (void *) cmd->add_data;
701 buf->va = cpu_to_be64(sg_dma_address(scat));
702 buf->key = cpu_to_be32(target->srp_host->dev->mr->rkey);
703 buf->len = cpu_to_be32(sg_dma_len(scat));
704 } else if (srp_map_fmr(target, scat, count, req,
705 (void *) cmd->add_data)) {
707 * FMR mapping failed, and the scatterlist has more
708 * than one entry. Generate an indirect memory
711 struct srp_indirect_buf *buf = (void *) cmd->add_data;
715 fmt = SRP_DATA_DESC_INDIRECT;
716 len = sizeof (struct srp_cmd) +
717 sizeof (struct srp_indirect_buf) +
718 count * sizeof (struct srp_direct_buf);
720 for (i = 0; i < count; ++i) {
721 buf->desc_list[i].va =
722 cpu_to_be64(sg_dma_address(&scat[i]));
723 buf->desc_list[i].key =
724 cpu_to_be32(target->srp_host->dev->mr->rkey);
725 buf->desc_list[i].len =
726 cpu_to_be32(sg_dma_len(&scat[i]));
727 datalen += sg_dma_len(&scat[i]);
730 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
731 cmd->data_out_desc_cnt = count;
733 cmd->data_in_desc_cnt = count;
736 cpu_to_be64(req->cmd->dma + sizeof *cmd + sizeof *buf);
737 buf->table_desc.key =
738 cpu_to_be32(target->srp_host->dev->mr->rkey);
739 buf->table_desc.len =
740 cpu_to_be32(count * sizeof (struct srp_direct_buf));
742 buf->len = cpu_to_be32(datalen);
745 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
746 cmd->buf_fmt = fmt << 4;
753 static void srp_process_rsp(struct srp_target_port *target, struct srp_rsp *rsp)
755 struct srp_request *req;
756 struct scsi_cmnd *scmnd;
760 delta = (s32) be32_to_cpu(rsp->req_lim_delta);
762 spin_lock_irqsave(target->scsi_host->host_lock, flags);
764 target->req_lim += delta;
766 req = &target->req_ring[rsp->tag & ~SRP_TAG_TSK_MGMT];
768 if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
769 if (be32_to_cpu(rsp->resp_data_len) < 4)
770 req->tsk_status = -1;
772 req->tsk_status = rsp->data[3];
773 complete(&req->done);
777 printk(KERN_ERR "Null scmnd for RSP w/tag %016llx\n",
778 (unsigned long long) rsp->tag);
779 scmnd->result = rsp->status;
781 if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
782 memcpy(scmnd->sense_buffer, rsp->data +
783 be32_to_cpu(rsp->resp_data_len),
784 min_t(int, be32_to_cpu(rsp->sense_data_len),
785 SCSI_SENSE_BUFFERSIZE));
788 if (rsp->flags & (SRP_RSP_FLAG_DOOVER | SRP_RSP_FLAG_DOUNDER))
789 scmnd->resid = be32_to_cpu(rsp->data_out_res_cnt);
790 else if (rsp->flags & (SRP_RSP_FLAG_DIOVER | SRP_RSP_FLAG_DIUNDER))
791 scmnd->resid = be32_to_cpu(rsp->data_in_res_cnt);
793 if (!req->tsk_mgmt) {
794 scmnd->host_scribble = (void *) -1L;
795 scmnd->scsi_done(scmnd);
797 srp_remove_req(target, req);
802 spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
805 static void srp_handle_recv(struct srp_target_port *target, struct ib_wc *wc)
810 iu = target->rx_ring[wc->wr_id & ~SRP_OP_RECV];
812 dma_sync_single_for_cpu(target->srp_host->dev->dev->dma_device, iu->dma,
813 target->max_ti_iu_len, DMA_FROM_DEVICE);
815 opcode = *(u8 *) iu->buf;
820 printk(KERN_ERR PFX "recv completion, opcode 0x%02x\n", opcode);
822 for (i = 0; i < wc->byte_len; ++i) {
824 printk(KERN_ERR " [%02x] ", i);
825 printk(" %02x", ((u8 *) iu->buf)[i]);
826 if ((i + 1) % 8 == 0)
830 if (wc->byte_len % 8)
836 srp_process_rsp(target, iu->buf);
840 /* XXX Handle target logout */
841 printk(KERN_WARNING PFX "Got target logout request\n");
845 printk(KERN_WARNING PFX "Unhandled SRP opcode 0x%02x\n", opcode);
849 dma_sync_single_for_device(target->srp_host->dev->dev->dma_device, iu->dma,
850 target->max_ti_iu_len, DMA_FROM_DEVICE);
853 static void srp_completion(struct ib_cq *cq, void *target_ptr)
855 struct srp_target_port *target = target_ptr;
858 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
859 while (ib_poll_cq(cq, 1, &wc) > 0) {
861 printk(KERN_ERR PFX "failed %s status %d\n",
862 wc.wr_id & SRP_OP_RECV ? "receive" : "send",
867 if (wc.wr_id & SRP_OP_RECV)
868 srp_handle_recv(target, &wc);
874 static int __srp_post_recv(struct srp_target_port *target)
878 struct ib_recv_wr wr, *bad_wr;
882 next = target->rx_head & (SRP_RQ_SIZE - 1);
883 wr.wr_id = next | SRP_OP_RECV;
884 iu = target->rx_ring[next];
887 list.length = iu->size;
888 list.lkey = target->srp_host->dev->mr->lkey;
894 ret = ib_post_recv(target->qp, &wr, &bad_wr);
901 static int srp_post_recv(struct srp_target_port *target)
906 spin_lock_irqsave(target->scsi_host->host_lock, flags);
907 ret = __srp_post_recv(target);
908 spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
914 * Must be called with target->scsi_host->host_lock held to protect
915 * req_lim and tx_head. Lock cannot be dropped between call here and
916 * call to __srp_post_send().
918 static struct srp_iu *__srp_get_tx_iu(struct srp_target_port *target)
920 if (target->tx_head - target->tx_tail >= SRP_SQ_SIZE)
923 if (unlikely(target->req_lim < 1))
924 ++target->zero_req_lim;
926 return target->tx_ring[target->tx_head & SRP_SQ_SIZE];
930 * Must be called with target->scsi_host->host_lock held to protect
931 * req_lim and tx_head.
933 static int __srp_post_send(struct srp_target_port *target,
934 struct srp_iu *iu, int len)
937 struct ib_send_wr wr, *bad_wr;
942 list.lkey = target->srp_host->dev->mr->lkey;
945 wr.wr_id = target->tx_head & SRP_SQ_SIZE;
948 wr.opcode = IB_WR_SEND;
949 wr.send_flags = IB_SEND_SIGNALED;
951 ret = ib_post_send(target->qp, &wr, &bad_wr);
961 static int srp_queuecommand(struct scsi_cmnd *scmnd,
962 void (*done)(struct scsi_cmnd *))
964 struct srp_target_port *target = host_to_target(scmnd->device->host);
965 struct srp_request *req;
970 if (target->state == SRP_TARGET_CONNECTING)
973 if (target->state == SRP_TARGET_DEAD ||
974 target->state == SRP_TARGET_REMOVED) {
975 scmnd->result = DID_BAD_TARGET << 16;
980 iu = __srp_get_tx_iu(target);
984 dma_sync_single_for_cpu(target->srp_host->dev->dev->dma_device, iu->dma,
985 srp_max_iu_len, DMA_TO_DEVICE);
987 req = list_entry(target->free_reqs.next, struct srp_request, list);
989 scmnd->scsi_done = done;
991 scmnd->host_scribble = (void *) (long) req->index;
994 memset(cmd, 0, sizeof *cmd);
996 cmd->opcode = SRP_CMD;
997 cmd->lun = cpu_to_be64((u64) scmnd->device->lun << 48);
998 cmd->tag = req->index;
999 memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
1004 req->tsk_mgmt = NULL;
1006 len = srp_map_data(scmnd, target, req);
1008 printk(KERN_ERR PFX "Failed to map data\n");
1012 if (__srp_post_recv(target)) {
1013 printk(KERN_ERR PFX "Recv failed\n");
1017 dma_sync_single_for_device(target->srp_host->dev->dev->dma_device, iu->dma,
1018 srp_max_iu_len, DMA_TO_DEVICE);
1020 if (__srp_post_send(target, iu, len)) {
1021 printk(KERN_ERR PFX "Send failed\n");
1025 list_move_tail(&req->list, &target->req_queue);
1030 srp_unmap_data(scmnd, target, req);
1033 return SCSI_MLQUEUE_HOST_BUSY;
1036 static int srp_alloc_iu_bufs(struct srp_target_port *target)
1040 for (i = 0; i < SRP_RQ_SIZE; ++i) {
1041 target->rx_ring[i] = srp_alloc_iu(target->srp_host,
1042 target->max_ti_iu_len,
1043 GFP_KERNEL, DMA_FROM_DEVICE);
1044 if (!target->rx_ring[i])
1048 for (i = 0; i < SRP_SQ_SIZE + 1; ++i) {
1049 target->tx_ring[i] = srp_alloc_iu(target->srp_host,
1051 GFP_KERNEL, DMA_TO_DEVICE);
1052 if (!target->tx_ring[i])
1059 for (i = 0; i < SRP_RQ_SIZE; ++i) {
1060 srp_free_iu(target->srp_host, target->rx_ring[i]);
1061 target->rx_ring[i] = NULL;
1064 for (i = 0; i < SRP_SQ_SIZE + 1; ++i) {
1065 srp_free_iu(target->srp_host, target->tx_ring[i]);
1066 target->tx_ring[i] = NULL;
1072 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
1073 struct ib_cm_event *event,
1074 struct srp_target_port *target)
1076 struct ib_class_port_info *cpi;
1079 switch (event->param.rej_rcvd.reason) {
1080 case IB_CM_REJ_PORT_CM_REDIRECT:
1081 cpi = event->param.rej_rcvd.ari;
1082 target->path.dlid = cpi->redirect_lid;
1083 target->path.pkey = cpi->redirect_pkey;
1084 cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
1085 memcpy(target->path.dgid.raw, cpi->redirect_gid, 16);
1087 target->status = target->path.dlid ?
1088 SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
1091 case IB_CM_REJ_PORT_REDIRECT:
1092 if (topspin_workarounds &&
1093 !memcmp(&target->ioc_guid, topspin_oui, 3)) {
1095 * Topspin/Cisco SRP gateways incorrectly send
1096 * reject reason code 25 when they mean 24
1099 memcpy(target->path.dgid.raw,
1100 event->param.rej_rcvd.ari, 16);
1102 printk(KERN_DEBUG PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
1103 (unsigned long long) be64_to_cpu(target->path.dgid.global.subnet_prefix),
1104 (unsigned long long) be64_to_cpu(target->path.dgid.global.interface_id));
1106 target->status = SRP_PORT_REDIRECT;
1108 printk(KERN_WARNING " REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
1109 target->status = -ECONNRESET;
1113 case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
1114 printk(KERN_WARNING " REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
1115 target->status = -ECONNRESET;
1118 case IB_CM_REJ_CONSUMER_DEFINED:
1119 opcode = *(u8 *) event->private_data;
1120 if (opcode == SRP_LOGIN_REJ) {
1121 struct srp_login_rej *rej = event->private_data;
1122 u32 reason = be32_to_cpu(rej->reason);
1124 if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
1125 printk(KERN_WARNING PFX
1126 "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
1128 printk(KERN_WARNING PFX
1129 "SRP LOGIN REJECTED, reason 0x%08x\n", reason);
1131 printk(KERN_WARNING " REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
1132 " opcode 0x%02x\n", opcode);
1133 target->status = -ECONNRESET;
1137 printk(KERN_WARNING " REJ reason 0x%x\n",
1138 event->param.rej_rcvd.reason);
1139 target->status = -ECONNRESET;
1143 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
1145 struct srp_target_port *target = cm_id->context;
1146 struct ib_qp_attr *qp_attr = NULL;
1151 switch (event->event) {
1152 case IB_CM_REQ_ERROR:
1153 printk(KERN_DEBUG PFX "Sending CM REQ failed\n");
1155 target->status = -ECONNRESET;
1158 case IB_CM_REP_RECEIVED:
1160 opcode = *(u8 *) event->private_data;
1162 if (opcode == SRP_LOGIN_RSP) {
1163 struct srp_login_rsp *rsp = event->private_data;
1165 target->max_ti_iu_len = be32_to_cpu(rsp->max_ti_iu_len);
1166 target->req_lim = be32_to_cpu(rsp->req_lim_delta);
1168 target->scsi_host->can_queue = min(target->req_lim,
1169 target->scsi_host->can_queue);
1171 printk(KERN_WARNING PFX "Unhandled RSP opcode %#x\n", opcode);
1172 target->status = -ECONNRESET;
1176 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 (srp_find_req(target, scmnd, &req))
1321 if (srp_send_tsk_mgmt(target, req, SRP_TSK_ABORT_TASK))
1324 spin_lock_irq(target->scsi_host->host_lock);
1326 if (req->cmd_done) {
1327 srp_remove_req(target, req);
1328 scmnd->scsi_done(scmnd);
1329 } else if (!req->tsk_status) {
1330 srp_remove_req(target, req);
1331 scmnd->result = DID_ABORT << 16;
1335 spin_unlock_irq(target->scsi_host->host_lock);
1340 static int srp_reset_device(struct scsi_cmnd *scmnd)
1342 struct srp_target_port *target = host_to_target(scmnd->device->host);
1343 struct srp_request *req, *tmp;
1345 printk(KERN_ERR "SRP reset_device called\n");
1347 if (srp_find_req(target, scmnd, &req))
1349 if (srp_send_tsk_mgmt(target, req, SRP_TSK_LUN_RESET))
1351 if (req->tsk_status)
1354 spin_lock_irq(target->scsi_host->host_lock);
1356 list_for_each_entry_safe(req, tmp, &target->req_queue, list)
1357 if (req->scmnd->device == scmnd->device)
1358 srp_reset_req(target, req);
1360 spin_unlock_irq(target->scsi_host->host_lock);
1365 static int srp_reset_host(struct scsi_cmnd *scmnd)
1367 struct srp_target_port *target = host_to_target(scmnd->device->host);
1370 printk(KERN_ERR PFX "SRP reset_host called\n");
1372 if (!srp_reconnect_target(target))
1378 static ssize_t show_id_ext(struct class_device *cdev, char *buf)
1380 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1382 if (target->state == SRP_TARGET_DEAD ||
1383 target->state == SRP_TARGET_REMOVED)
1386 return sprintf(buf, "0x%016llx\n",
1387 (unsigned long long) be64_to_cpu(target->id_ext));
1390 static ssize_t show_ioc_guid(struct class_device *cdev, char *buf)
1392 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1394 if (target->state == SRP_TARGET_DEAD ||
1395 target->state == SRP_TARGET_REMOVED)
1398 return sprintf(buf, "0x%016llx\n",
1399 (unsigned long long) be64_to_cpu(target->ioc_guid));
1402 static ssize_t show_service_id(struct class_device *cdev, char *buf)
1404 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1406 if (target->state == SRP_TARGET_DEAD ||
1407 target->state == SRP_TARGET_REMOVED)
1410 return sprintf(buf, "0x%016llx\n",
1411 (unsigned long long) be64_to_cpu(target->service_id));
1414 static ssize_t show_pkey(struct class_device *cdev, char *buf)
1416 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1418 if (target->state == SRP_TARGET_DEAD ||
1419 target->state == SRP_TARGET_REMOVED)
1422 return sprintf(buf, "0x%04x\n", be16_to_cpu(target->path.pkey));
1425 static ssize_t show_dgid(struct class_device *cdev, char *buf)
1427 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1429 if (target->state == SRP_TARGET_DEAD ||
1430 target->state == SRP_TARGET_REMOVED)
1433 return sprintf(buf, "%04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x\n",
1434 be16_to_cpu(((__be16 *) target->path.dgid.raw)[0]),
1435 be16_to_cpu(((__be16 *) target->path.dgid.raw)[1]),
1436 be16_to_cpu(((__be16 *) target->path.dgid.raw)[2]),
1437 be16_to_cpu(((__be16 *) target->path.dgid.raw)[3]),
1438 be16_to_cpu(((__be16 *) target->path.dgid.raw)[4]),
1439 be16_to_cpu(((__be16 *) target->path.dgid.raw)[5]),
1440 be16_to_cpu(((__be16 *) target->path.dgid.raw)[6]),
1441 be16_to_cpu(((__be16 *) target->path.dgid.raw)[7]));
1444 static ssize_t show_zero_req_lim(struct class_device *cdev, char *buf)
1446 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1448 if (target->state == SRP_TARGET_DEAD ||
1449 target->state == SRP_TARGET_REMOVED)
1452 return sprintf(buf, "%d\n", target->zero_req_lim);
1455 static ssize_t show_local_ib_port(struct class_device *cdev, char *buf)
1457 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1459 return sprintf(buf, "%d\n", target->srp_host->port);
1462 static ssize_t show_local_ib_device(struct class_device *cdev, char *buf)
1464 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1466 return sprintf(buf, "%s\n", target->srp_host->dev->dev->name);
1469 static CLASS_DEVICE_ATTR(id_ext, S_IRUGO, show_id_ext, NULL);
1470 static CLASS_DEVICE_ATTR(ioc_guid, S_IRUGO, show_ioc_guid, NULL);
1471 static CLASS_DEVICE_ATTR(service_id, S_IRUGO, show_service_id, NULL);
1472 static CLASS_DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1473 static CLASS_DEVICE_ATTR(dgid, S_IRUGO, show_dgid, NULL);
1474 static CLASS_DEVICE_ATTR(zero_req_lim, S_IRUGO, show_zero_req_lim, NULL);
1475 static CLASS_DEVICE_ATTR(local_ib_port, S_IRUGO, show_local_ib_port, NULL);
1476 static CLASS_DEVICE_ATTR(local_ib_device, S_IRUGO, show_local_ib_device, NULL);
1478 static struct class_device_attribute *srp_host_attrs[] = {
1479 &class_device_attr_id_ext,
1480 &class_device_attr_ioc_guid,
1481 &class_device_attr_service_id,
1482 &class_device_attr_pkey,
1483 &class_device_attr_dgid,
1484 &class_device_attr_zero_req_lim,
1485 &class_device_attr_local_ib_port,
1486 &class_device_attr_local_ib_device,
1490 static struct scsi_host_template srp_template = {
1491 .module = THIS_MODULE,
1493 .info = srp_target_info,
1494 .queuecommand = srp_queuecommand,
1495 .eh_abort_handler = srp_abort,
1496 .eh_device_reset_handler = srp_reset_device,
1497 .eh_host_reset_handler = srp_reset_host,
1498 .can_queue = SRP_SQ_SIZE,
1500 .cmd_per_lun = SRP_SQ_SIZE,
1501 .use_clustering = ENABLE_CLUSTERING,
1502 .shost_attrs = srp_host_attrs
1505 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
1507 sprintf(target->target_name, "SRP.T10:%016llX",
1508 (unsigned long long) be64_to_cpu(target->id_ext));
1510 if (scsi_add_host(target->scsi_host, host->dev->dev->dma_device))
1513 spin_lock(&host->target_lock);
1514 list_add_tail(&target->list, &host->target_list);
1515 spin_unlock(&host->target_lock);
1517 target->state = SRP_TARGET_LIVE;
1519 scsi_scan_target(&target->scsi_host->shost_gendev,
1520 0, target->scsi_id, SCAN_WILD_CARD, 0);
1525 static void srp_release_class_dev(struct class_device *class_dev)
1527 struct srp_host *host =
1528 container_of(class_dev, struct srp_host, class_dev);
1530 complete(&host->released);
1533 static struct class srp_class = {
1534 .name = "infiniband_srp",
1535 .release = srp_release_class_dev
1539 * Target ports are added by writing
1541 * id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
1542 * pkey=<P_Key>,service_id=<service ID>
1544 * to the add_target sysfs attribute.
1548 SRP_OPT_ID_EXT = 1 << 0,
1549 SRP_OPT_IOC_GUID = 1 << 1,
1550 SRP_OPT_DGID = 1 << 2,
1551 SRP_OPT_PKEY = 1 << 3,
1552 SRP_OPT_SERVICE_ID = 1 << 4,
1553 SRP_OPT_MAX_SECT = 1 << 5,
1554 SRP_OPT_MAX_CMD_PER_LUN = 1 << 6,
1555 SRP_OPT_IO_CLASS = 1 << 7,
1556 SRP_OPT_ALL = (SRP_OPT_ID_EXT |
1560 SRP_OPT_SERVICE_ID),
1563 static match_table_t srp_opt_tokens = {
1564 { SRP_OPT_ID_EXT, "id_ext=%s" },
1565 { SRP_OPT_IOC_GUID, "ioc_guid=%s" },
1566 { SRP_OPT_DGID, "dgid=%s" },
1567 { SRP_OPT_PKEY, "pkey=%x" },
1568 { SRP_OPT_SERVICE_ID, "service_id=%s" },
1569 { SRP_OPT_MAX_SECT, "max_sect=%d" },
1570 { SRP_OPT_MAX_CMD_PER_LUN, "max_cmd_per_lun=%d" },
1571 { SRP_OPT_IO_CLASS, "io_class=%x" },
1572 { SRP_OPT_ERR, NULL }
1575 static int srp_parse_options(const char *buf, struct srp_target_port *target)
1577 char *options, *sep_opt;
1580 substring_t args[MAX_OPT_ARGS];
1586 options = kstrdup(buf, GFP_KERNEL);
1591 while ((p = strsep(&sep_opt, ",")) != NULL) {
1595 token = match_token(p, srp_opt_tokens, args);
1599 case SRP_OPT_ID_EXT:
1600 p = match_strdup(args);
1601 target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
1605 case SRP_OPT_IOC_GUID:
1606 p = match_strdup(args);
1607 target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
1612 p = match_strdup(args);
1613 if (strlen(p) != 32) {
1614 printk(KERN_WARNING PFX "bad dest GID parameter '%s'\n", p);
1619 for (i = 0; i < 16; ++i) {
1620 strlcpy(dgid, p + i * 2, 3);
1621 target->path.dgid.raw[i] = simple_strtoul(dgid, NULL, 16);
1627 if (match_hex(args, &token)) {
1628 printk(KERN_WARNING PFX "bad P_Key parameter '%s'\n", p);
1631 target->path.pkey = cpu_to_be16(token);
1634 case SRP_OPT_SERVICE_ID:
1635 p = match_strdup(args);
1636 target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
1640 case SRP_OPT_MAX_SECT:
1641 if (match_int(args, &token)) {
1642 printk(KERN_WARNING PFX "bad max sect parameter '%s'\n", p);
1645 target->scsi_host->max_sectors = token;
1648 case SRP_OPT_MAX_CMD_PER_LUN:
1649 if (match_int(args, &token)) {
1650 printk(KERN_WARNING PFX "bad max cmd_per_lun parameter '%s'\n", p);
1653 target->scsi_host->cmd_per_lun = min(token, SRP_SQ_SIZE);
1656 case SRP_OPT_IO_CLASS:
1657 if (match_hex(args, &token)) {
1658 printk(KERN_WARNING PFX "bad IO class parameter '%s' \n", p);
1661 if (token != SRP_REV10_IB_IO_CLASS &&
1662 token != SRP_REV16A_IB_IO_CLASS) {
1663 printk(KERN_WARNING PFX "unknown IO class parameter value"
1664 " %x specified (use %x or %x).\n",
1665 token, SRP_REV10_IB_IO_CLASS, SRP_REV16A_IB_IO_CLASS);
1668 target->io_class = token;
1672 printk(KERN_WARNING PFX "unknown parameter or missing value "
1673 "'%s' in target creation request\n", p);
1678 if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
1681 for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
1682 if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
1683 !(srp_opt_tokens[i].token & opt_mask))
1684 printk(KERN_WARNING PFX "target creation request is "
1685 "missing parameter '%s'\n",
1686 srp_opt_tokens[i].pattern);
1693 static ssize_t srp_create_target(struct class_device *class_dev,
1694 const char *buf, size_t count)
1696 struct srp_host *host =
1697 container_of(class_dev, struct srp_host, class_dev);
1698 struct Scsi_Host *target_host;
1699 struct srp_target_port *target;
1703 target_host = scsi_host_alloc(&srp_template,
1704 sizeof (struct srp_target_port));
1708 target_host->max_lun = SRP_MAX_LUN;
1710 target = host_to_target(target_host);
1711 memset(target, 0, sizeof *target);
1713 target->io_class = SRP_REV16A_IB_IO_CLASS;
1714 target->scsi_host = target_host;
1715 target->srp_host = host;
1717 INIT_LIST_HEAD(&target->free_reqs);
1718 INIT_LIST_HEAD(&target->req_queue);
1719 for (i = 0; i < SRP_SQ_SIZE; ++i) {
1720 target->req_ring[i].index = i;
1721 list_add_tail(&target->req_ring[i].list, &target->free_reqs);
1724 ret = srp_parse_options(buf, target);
1728 ib_get_cached_gid(host->dev->dev, host->port, 0, &target->path.sgid);
1730 printk(KERN_DEBUG PFX "new target: id_ext %016llx ioc_guid %016llx pkey %04x "
1731 "service_id %016llx dgid %04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x\n",
1732 (unsigned long long) be64_to_cpu(target->id_ext),
1733 (unsigned long long) be64_to_cpu(target->ioc_guid),
1734 be16_to_cpu(target->path.pkey),
1735 (unsigned long long) be64_to_cpu(target->service_id),
1736 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[0]),
1737 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[2]),
1738 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[4]),
1739 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[6]),
1740 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[8]),
1741 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[10]),
1742 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[12]),
1743 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[14]));
1745 ret = srp_create_target_ib(target);
1749 target->cm_id = ib_create_cm_id(host->dev->dev, srp_cm_handler, target);
1750 if (IS_ERR(target->cm_id)) {
1751 ret = PTR_ERR(target->cm_id);
1755 ret = srp_connect_target(target);
1757 printk(KERN_ERR PFX "Connection failed\n");
1761 ret = srp_add_target(host, target);
1763 goto err_disconnect;
1768 srp_disconnect_target(target);
1771 ib_destroy_cm_id(target->cm_id);
1774 srp_free_target_ib(target);
1777 scsi_host_put(target_host);
1782 static CLASS_DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
1784 static ssize_t show_ibdev(struct class_device *class_dev, char *buf)
1786 struct srp_host *host =
1787 container_of(class_dev, struct srp_host, class_dev);
1789 return sprintf(buf, "%s\n", host->dev->dev->name);
1792 static CLASS_DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
1794 static ssize_t show_port(struct class_device *class_dev, char *buf)
1796 struct srp_host *host =
1797 container_of(class_dev, struct srp_host, class_dev);
1799 return sprintf(buf, "%d\n", host->port);
1802 static CLASS_DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
1804 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
1806 struct srp_host *host;
1808 host = kzalloc(sizeof *host, GFP_KERNEL);
1812 INIT_LIST_HEAD(&host->target_list);
1813 spin_lock_init(&host->target_lock);
1814 init_completion(&host->released);
1818 host->initiator_port_id[7] = port;
1819 memcpy(host->initiator_port_id + 8, &device->dev->node_guid, 8);
1821 host->class_dev.class = &srp_class;
1822 host->class_dev.dev = device->dev->dma_device;
1823 snprintf(host->class_dev.class_id, BUS_ID_SIZE, "srp-%s-%d",
1824 device->dev->name, port);
1826 if (class_device_register(&host->class_dev))
1828 if (class_device_create_file(&host->class_dev, &class_device_attr_add_target))
1830 if (class_device_create_file(&host->class_dev, &class_device_attr_ibdev))
1832 if (class_device_create_file(&host->class_dev, &class_device_attr_port))
1838 class_device_unregister(&host->class_dev);
1846 static void srp_add_one(struct ib_device *device)
1848 struct srp_device *srp_dev;
1849 struct ib_device_attr *dev_attr;
1850 struct ib_fmr_pool_param fmr_param;
1851 struct srp_host *host;
1854 dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
1858 if (ib_query_device(device, dev_attr)) {
1859 printk(KERN_WARNING PFX "Query device failed for %s\n",
1864 srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
1869 * Use the smallest page size supported by the HCA, down to a
1870 * minimum of 512 bytes (which is the smallest sector that a
1871 * SCSI command will ever carry).
1873 srp_dev->fmr_page_shift = max(9, ffs(dev_attr->page_size_cap) - 1);
1874 srp_dev->fmr_page_size = 1 << srp_dev->fmr_page_shift;
1875 srp_dev->fmr_page_mask = ~((unsigned long) srp_dev->fmr_page_size - 1);
1877 INIT_LIST_HEAD(&srp_dev->dev_list);
1879 srp_dev->dev = device;
1880 srp_dev->pd = ib_alloc_pd(device);
1881 if (IS_ERR(srp_dev->pd))
1884 srp_dev->mr = ib_get_dma_mr(srp_dev->pd,
1885 IB_ACCESS_LOCAL_WRITE |
1886 IB_ACCESS_REMOTE_READ |
1887 IB_ACCESS_REMOTE_WRITE);
1888 if (IS_ERR(srp_dev->mr))
1891 memset(&fmr_param, 0, sizeof fmr_param);
1892 fmr_param.pool_size = SRP_FMR_POOL_SIZE;
1893 fmr_param.dirty_watermark = SRP_FMR_DIRTY_SIZE;
1894 fmr_param.cache = 1;
1895 fmr_param.max_pages_per_fmr = SRP_FMR_SIZE;
1896 fmr_param.page_shift = srp_dev->fmr_page_shift;
1897 fmr_param.access = (IB_ACCESS_LOCAL_WRITE |
1898 IB_ACCESS_REMOTE_WRITE |
1899 IB_ACCESS_REMOTE_READ);
1901 srp_dev->fmr_pool = ib_create_fmr_pool(srp_dev->pd, &fmr_param);
1902 if (IS_ERR(srp_dev->fmr_pool))
1903 srp_dev->fmr_pool = NULL;
1905 if (device->node_type == RDMA_NODE_IB_SWITCH) {
1910 e = device->phys_port_cnt;
1913 for (p = s; p <= e; ++p) {
1914 host = srp_add_port(srp_dev, p);
1916 list_add_tail(&host->list, &srp_dev->dev_list);
1919 ib_set_client_data(device, &srp_client, srp_dev);
1924 ib_dealloc_pd(srp_dev->pd);
1933 static void srp_remove_one(struct ib_device *device)
1935 struct srp_device *srp_dev;
1936 struct srp_host *host, *tmp_host;
1937 LIST_HEAD(target_list);
1938 struct srp_target_port *target, *tmp_target;
1940 srp_dev = ib_get_client_data(device, &srp_client);
1942 list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
1943 class_device_unregister(&host->class_dev);
1945 * Wait for the sysfs entry to go away, so that no new
1946 * target ports can be created.
1948 wait_for_completion(&host->released);
1951 * Mark all target ports as removed, so we stop queueing
1952 * commands and don't try to reconnect.
1954 spin_lock(&host->target_lock);
1955 list_for_each_entry(target, &host->target_list, list) {
1956 spin_lock_irq(target->scsi_host->host_lock);
1957 target->state = SRP_TARGET_REMOVED;
1958 spin_unlock_irq(target->scsi_host->host_lock);
1960 spin_unlock(&host->target_lock);
1963 * Wait for any reconnection tasks that may have
1964 * started before we marked our target ports as
1965 * removed, and any target port removal tasks.
1967 flush_scheduled_work();
1969 list_for_each_entry_safe(target, tmp_target,
1970 &host->target_list, list) {
1971 scsi_remove_host(target->scsi_host);
1972 srp_disconnect_target(target);
1973 ib_destroy_cm_id(target->cm_id);
1974 srp_free_target_ib(target);
1975 scsi_host_put(target->scsi_host);
1981 if (srp_dev->fmr_pool)
1982 ib_destroy_fmr_pool(srp_dev->fmr_pool);
1983 ib_dereg_mr(srp_dev->mr);
1984 ib_dealloc_pd(srp_dev->pd);
1989 static int __init srp_init_module(void)
1993 srp_template.sg_tablesize = srp_sg_tablesize;
1994 srp_max_iu_len = (sizeof (struct srp_cmd) +
1995 sizeof (struct srp_indirect_buf) +
1996 srp_sg_tablesize * 16);
1998 ret = class_register(&srp_class);
2000 printk(KERN_ERR PFX "couldn't register class infiniband_srp\n");
2004 ib_sa_register_client(&srp_sa_client);
2006 ret = ib_register_client(&srp_client);
2008 printk(KERN_ERR PFX "couldn't register IB client\n");
2009 ib_sa_unregister_client(&srp_sa_client);
2010 class_unregister(&srp_class);
2017 static void __exit srp_cleanup_module(void)
2019 ib_unregister_client(&srp_client);
2020 ib_sa_unregister_client(&srp_sa_client);
2021 class_unregister(&srp_class);
2024 module_init(srp_init_module);
2025 module_exit(srp_cleanup_module);