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 topspin_workarounds = 1;
67 module_param(topspin_workarounds, int, 0444);
68 MODULE_PARM_DESC(topspin_workarounds,
69 "Enable workarounds for Topspin/Cisco SRP target bugs if != 0");
71 static const u8 topspin_oui[3] = { 0x00, 0x05, 0xad };
73 static void srp_add_one(struct ib_device *device);
74 static void srp_remove_one(struct ib_device *device);
75 static void srp_completion(struct ib_cq *cq, void *target_ptr);
76 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event);
78 static struct ib_client srp_client = {
81 .remove = srp_remove_one
84 static inline struct srp_target_port *host_to_target(struct Scsi_Host *host)
86 return (struct srp_target_port *) host->hostdata;
89 static const char *srp_target_info(struct Scsi_Host *host)
91 return host_to_target(host)->target_name;
94 static struct srp_iu *srp_alloc_iu(struct srp_host *host, size_t size,
96 enum dma_data_direction direction)
100 iu = kmalloc(sizeof *iu, gfp_mask);
104 iu->buf = kzalloc(size, gfp_mask);
108 iu->dma = dma_map_single(host->dev->dev->dma_device,
109 iu->buf, size, direction);
110 if (dma_mapping_error(iu->dma))
114 iu->direction = direction;
126 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
131 dma_unmap_single(host->dev->dev->dma_device,
132 iu->dma, iu->size, iu->direction);
137 static void srp_qp_event(struct ib_event *event, void *context)
139 printk(KERN_ERR PFX "QP event %d\n", event->event);
142 static int srp_init_qp(struct srp_target_port *target,
145 struct ib_qp_attr *attr;
148 attr = kmalloc(sizeof *attr, GFP_KERNEL);
152 ret = ib_find_cached_pkey(target->srp_host->dev->dev,
153 target->srp_host->port,
154 be16_to_cpu(target->path.pkey),
159 attr->qp_state = IB_QPS_INIT;
160 attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
161 IB_ACCESS_REMOTE_WRITE);
162 attr->port_num = target->srp_host->port;
164 ret = ib_modify_qp(qp, attr,
175 static int srp_create_target_ib(struct srp_target_port *target)
177 struct ib_qp_init_attr *init_attr;
180 init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
184 target->cq = ib_create_cq(target->srp_host->dev->dev, srp_completion,
185 NULL, target, SRP_CQ_SIZE);
186 if (IS_ERR(target->cq)) {
187 ret = PTR_ERR(target->cq);
191 ib_req_notify_cq(target->cq, IB_CQ_NEXT_COMP);
193 init_attr->event_handler = srp_qp_event;
194 init_attr->cap.max_send_wr = SRP_SQ_SIZE;
195 init_attr->cap.max_recv_wr = SRP_RQ_SIZE;
196 init_attr->cap.max_recv_sge = 1;
197 init_attr->cap.max_send_sge = 1;
198 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
199 init_attr->qp_type = IB_QPT_RC;
200 init_attr->send_cq = target->cq;
201 init_attr->recv_cq = target->cq;
203 target->qp = ib_create_qp(target->srp_host->dev->pd, init_attr);
204 if (IS_ERR(target->qp)) {
205 ret = PTR_ERR(target->qp);
206 ib_destroy_cq(target->cq);
210 ret = srp_init_qp(target, target->qp);
212 ib_destroy_qp(target->qp);
213 ib_destroy_cq(target->cq);
222 static void srp_free_target_ib(struct srp_target_port *target)
226 ib_destroy_qp(target->qp);
227 ib_destroy_cq(target->cq);
229 for (i = 0; i < SRP_RQ_SIZE; ++i)
230 srp_free_iu(target->srp_host, target->rx_ring[i]);
231 for (i = 0; i < SRP_SQ_SIZE + 1; ++i)
232 srp_free_iu(target->srp_host, target->tx_ring[i]);
235 static void srp_path_rec_completion(int status,
236 struct ib_sa_path_rec *pathrec,
239 struct srp_target_port *target = target_ptr;
241 target->status = status;
243 printk(KERN_ERR PFX "Got failed path rec status %d\n", status);
245 target->path = *pathrec;
246 complete(&target->done);
249 static int srp_lookup_path(struct srp_target_port *target)
251 target->path.numb_path = 1;
253 init_completion(&target->done);
255 target->path_query_id = ib_sa_path_rec_get(target->srp_host->dev->dev,
256 target->srp_host->port,
258 IB_SA_PATH_REC_DGID |
259 IB_SA_PATH_REC_SGID |
260 IB_SA_PATH_REC_NUMB_PATH |
262 SRP_PATH_REC_TIMEOUT_MS,
264 srp_path_rec_completion,
265 target, &target->path_query);
266 if (target->path_query_id < 0)
267 return target->path_query_id;
269 wait_for_completion(&target->done);
271 if (target->status < 0)
272 printk(KERN_WARNING PFX "Path record query failed\n");
274 return target->status;
277 static int srp_send_req(struct srp_target_port *target)
280 struct ib_cm_req_param param;
281 struct srp_login_req priv;
285 req = kzalloc(sizeof *req, GFP_KERNEL);
289 req->param.primary_path = &target->path;
290 req->param.alternate_path = NULL;
291 req->param.service_id = target->service_id;
292 req->param.qp_num = target->qp->qp_num;
293 req->param.qp_type = target->qp->qp_type;
294 req->param.private_data = &req->priv;
295 req->param.private_data_len = sizeof req->priv;
296 req->param.flow_control = 1;
298 get_random_bytes(&req->param.starting_psn, 4);
299 req->param.starting_psn &= 0xffffff;
302 * Pick some arbitrary defaults here; we could make these
303 * module parameters if anyone cared about setting them.
305 req->param.responder_resources = 4;
306 req->param.remote_cm_response_timeout = 20;
307 req->param.local_cm_response_timeout = 20;
308 req->param.retry_count = 7;
309 req->param.rnr_retry_count = 7;
310 req->param.max_cm_retries = 15;
312 req->priv.opcode = SRP_LOGIN_REQ;
314 req->priv.req_it_iu_len = cpu_to_be32(SRP_MAX_IU_LEN);
315 req->priv.req_buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
316 SRP_BUF_FORMAT_INDIRECT);
317 memcpy(req->priv.initiator_port_id, target->srp_host->initiator_port_id, 16);
319 * Topspin/Cisco SRP targets will reject our login unless we
320 * zero out the first 8 bytes of our initiator port ID. The
321 * second 8 bytes must be our local node GUID, but we always
324 if (topspin_workarounds && !memcmp(&target->ioc_guid, topspin_oui, 3)) {
325 printk(KERN_DEBUG PFX "Topspin/Cisco initiator port ID workaround "
326 "activated for target GUID %016llx\n",
327 (unsigned long long) be64_to_cpu(target->ioc_guid));
328 memset(req->priv.initiator_port_id, 0, 8);
330 memcpy(req->priv.target_port_id, &target->id_ext, 8);
331 memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
333 status = ib_send_cm_req(target->cm_id, &req->param);
340 static void srp_disconnect_target(struct srp_target_port *target)
342 /* XXX should send SRP_I_LOGOUT request */
344 init_completion(&target->done);
345 if (ib_send_cm_dreq(target->cm_id, NULL, 0)) {
346 printk(KERN_DEBUG PFX "Sending CM DREQ failed\n");
349 wait_for_completion(&target->done);
352 static void srp_remove_work(void *target_ptr)
354 struct srp_target_port *target = target_ptr;
356 spin_lock_irq(target->scsi_host->host_lock);
357 if (target->state != SRP_TARGET_DEAD) {
358 spin_unlock_irq(target->scsi_host->host_lock);
361 target->state = SRP_TARGET_REMOVED;
362 spin_unlock_irq(target->scsi_host->host_lock);
364 mutex_lock(&target->srp_host->target_mutex);
365 list_del(&target->list);
366 mutex_unlock(&target->srp_host->target_mutex);
368 scsi_remove_host(target->scsi_host);
369 ib_destroy_cm_id(target->cm_id);
370 srp_free_target_ib(target);
371 scsi_host_put(target->scsi_host);
374 static int srp_connect_target(struct srp_target_port *target)
378 ret = srp_lookup_path(target);
383 init_completion(&target->done);
384 ret = srp_send_req(target);
387 wait_for_completion(&target->done);
390 * The CM event handling code will set status to
391 * SRP_PORT_REDIRECT if we get a port redirect REJ
392 * back, or SRP_DLID_REDIRECT if we get a lid/qp
395 switch (target->status) {
399 case SRP_PORT_REDIRECT:
400 ret = srp_lookup_path(target);
405 case SRP_DLID_REDIRECT:
409 return target->status;
414 static void srp_unmap_data(struct scsi_cmnd *scmnd,
415 struct srp_target_port *target,
416 struct srp_request *req)
418 struct scatterlist *scat;
421 if (!scmnd->request_buffer ||
422 (scmnd->sc_data_direction != DMA_TO_DEVICE &&
423 scmnd->sc_data_direction != DMA_FROM_DEVICE))
427 ib_fmr_pool_unmap(req->fmr);
432 * This handling of non-SG commands can be killed when the
433 * SCSI midlayer no longer generates non-SG commands.
435 if (likely(scmnd->use_sg)) {
436 nents = scmnd->use_sg;
437 scat = scmnd->request_buffer;
440 scat = &req->fake_sg;
443 dma_unmap_sg(target->srp_host->dev->dev->dma_device, scat, nents,
444 scmnd->sc_data_direction);
447 static int srp_reconnect_target(struct srp_target_port *target)
449 struct ib_cm_id *new_cm_id;
450 struct ib_qp_attr qp_attr;
451 struct srp_request *req;
456 spin_lock_irq(target->scsi_host->host_lock);
457 if (target->state != SRP_TARGET_LIVE) {
458 spin_unlock_irq(target->scsi_host->host_lock);
461 target->state = SRP_TARGET_CONNECTING;
462 spin_unlock_irq(target->scsi_host->host_lock);
464 srp_disconnect_target(target);
466 * Now get a new local CM ID so that we avoid confusing the
467 * target in case things are really fouled up.
469 new_cm_id = ib_create_cm_id(target->srp_host->dev->dev,
470 srp_cm_handler, target);
471 if (IS_ERR(new_cm_id)) {
472 ret = PTR_ERR(new_cm_id);
475 ib_destroy_cm_id(target->cm_id);
476 target->cm_id = new_cm_id;
478 qp_attr.qp_state = IB_QPS_RESET;
479 ret = ib_modify_qp(target->qp, &qp_attr, IB_QP_STATE);
483 ret = srp_init_qp(target, target->qp);
487 while (ib_poll_cq(target->cq, 1, &wc) > 0)
490 list_for_each_entry(req, &target->req_queue, list) {
491 req->scmnd->result = DID_RESET << 16;
492 req->scmnd->scsi_done(req->scmnd);
493 srp_unmap_data(req->scmnd, target, req);
499 INIT_LIST_HEAD(&target->free_reqs);
500 INIT_LIST_HEAD(&target->req_queue);
501 for (i = 0; i < SRP_SQ_SIZE; ++i)
502 list_add_tail(&target->req_ring[i].list, &target->free_reqs);
504 ret = srp_connect_target(target);
508 spin_lock_irq(target->scsi_host->host_lock);
509 if (target->state == SRP_TARGET_CONNECTING) {
511 target->state = SRP_TARGET_LIVE;
514 spin_unlock_irq(target->scsi_host->host_lock);
519 printk(KERN_ERR PFX "reconnect failed (%d), removing target port.\n", ret);
522 * We couldn't reconnect, so kill our target port off.
523 * However, we have to defer the real removal because we might
524 * be in the context of the SCSI error handler now, which
525 * would deadlock if we call scsi_remove_host().
527 spin_lock_irq(target->scsi_host->host_lock);
528 if (target->state == SRP_TARGET_CONNECTING) {
529 target->state = SRP_TARGET_DEAD;
530 INIT_WORK(&target->work, srp_remove_work, target);
531 schedule_work(&target->work);
533 spin_unlock_irq(target->scsi_host->host_lock);
538 static int srp_map_fmr(struct srp_device *dev, struct scatterlist *scat,
539 int sg_cnt, struct srp_request *req,
540 struct srp_direct_buf *buf)
553 for (i = 0; i < sg_cnt; ++i) {
554 if (sg_dma_address(&scat[i]) & ~dev->fmr_page_mask) {
560 if ((sg_dma_address(&scat[i]) + sg_dma_len(&scat[i])) &
561 ~dev->fmr_page_mask) {
568 len += sg_dma_len(&scat[i]);
571 page_cnt += len >> dev->fmr_page_shift;
572 if (page_cnt > SRP_FMR_SIZE)
575 dma_pages = kmalloc(sizeof (u64) * page_cnt, GFP_ATOMIC);
580 for (i = 0; i < sg_cnt; ++i)
581 for (j = 0; j < sg_dma_len(&scat[i]); j += dev->fmr_page_size)
582 dma_pages[page_cnt++] =
583 (sg_dma_address(&scat[i]) & dev->fmr_page_mask) + j;
585 req->fmr = ib_fmr_pool_map_phys(dev->fmr_pool,
586 dma_pages, page_cnt, &io_addr);
587 if (IS_ERR(req->fmr)) {
588 ret = PTR_ERR(req->fmr);
592 buf->va = cpu_to_be64(sg_dma_address(&scat[0]) & ~dev->fmr_page_mask);
593 buf->key = cpu_to_be32(req->fmr->fmr->rkey);
594 buf->len = cpu_to_be32(len);
604 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_target_port *target,
605 struct srp_request *req)
607 struct scatterlist *scat;
608 struct srp_cmd *cmd = req->cmd->buf;
609 int len, nents, count;
610 u8 fmt = SRP_DATA_DESC_DIRECT;
612 if (!scmnd->request_buffer || scmnd->sc_data_direction == DMA_NONE)
613 return sizeof (struct srp_cmd);
615 if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
616 scmnd->sc_data_direction != DMA_TO_DEVICE) {
617 printk(KERN_WARNING PFX "Unhandled data direction %d\n",
618 scmnd->sc_data_direction);
623 * This handling of non-SG commands can be killed when the
624 * SCSI midlayer no longer generates non-SG commands.
626 if (likely(scmnd->use_sg)) {
627 nents = scmnd->use_sg;
628 scat = scmnd->request_buffer;
631 scat = &req->fake_sg;
632 sg_init_one(scat, scmnd->request_buffer, scmnd->request_bufflen);
635 count = dma_map_sg(target->srp_host->dev->dev->dma_device,
636 scat, nents, scmnd->sc_data_direction);
638 fmt = SRP_DATA_DESC_DIRECT;
639 len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
643 * The midlayer only generated a single gather/scatter
644 * entry, or DMA mapping coalesced everything to a
645 * single entry. So a direct descriptor along with
646 * the DMA MR suffices.
648 struct srp_direct_buf *buf = (void *) cmd->add_data;
650 buf->va = cpu_to_be64(sg_dma_address(scat));
651 buf->key = cpu_to_be32(target->srp_host->dev->mr->rkey);
652 buf->len = cpu_to_be32(sg_dma_len(scat));
653 } else if (srp_map_fmr(target->srp_host->dev, scat, count, req,
654 (void *) cmd->add_data)) {
656 * FMR mapping failed, and the scatterlist has more
657 * than one entry. Generate an indirect memory
660 struct srp_indirect_buf *buf = (void *) cmd->add_data;
664 fmt = SRP_DATA_DESC_INDIRECT;
665 len = sizeof (struct srp_cmd) +
666 sizeof (struct srp_indirect_buf) +
667 count * sizeof (struct srp_direct_buf);
669 for (i = 0; i < count; ++i) {
670 buf->desc_list[i].va =
671 cpu_to_be64(sg_dma_address(&scat[i]));
672 buf->desc_list[i].key =
673 cpu_to_be32(target->srp_host->dev->mr->rkey);
674 buf->desc_list[i].len =
675 cpu_to_be32(sg_dma_len(&scat[i]));
676 datalen += sg_dma_len(&scat[i]);
679 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
680 cmd->data_out_desc_cnt = count;
682 cmd->data_in_desc_cnt = count;
685 cpu_to_be64(req->cmd->dma + sizeof *cmd + sizeof *buf);
686 buf->table_desc.key =
687 cpu_to_be32(target->srp_host->dev->mr->rkey);
688 buf->table_desc.len =
689 cpu_to_be32(count * sizeof (struct srp_direct_buf));
691 buf->len = cpu_to_be32(datalen);
694 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
695 cmd->buf_fmt = fmt << 4;
702 static void srp_remove_req(struct srp_target_port *target, struct srp_request *req)
704 srp_unmap_data(req->scmnd, target, req);
705 list_move_tail(&req->list, &target->free_reqs);
708 static void srp_process_rsp(struct srp_target_port *target, struct srp_rsp *rsp)
710 struct srp_request *req;
711 struct scsi_cmnd *scmnd;
715 delta = (s32) be32_to_cpu(rsp->req_lim_delta);
717 spin_lock_irqsave(target->scsi_host->host_lock, flags);
719 target->req_lim += delta;
721 req = &target->req_ring[rsp->tag & ~SRP_TAG_TSK_MGMT];
723 if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
724 if (be32_to_cpu(rsp->resp_data_len) < 4)
725 req->tsk_status = -1;
727 req->tsk_status = rsp->data[3];
728 complete(&req->done);
732 printk(KERN_ERR "Null scmnd for RSP w/tag %016llx\n",
733 (unsigned long long) rsp->tag);
734 scmnd->result = rsp->status;
736 if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
737 memcpy(scmnd->sense_buffer, rsp->data +
738 be32_to_cpu(rsp->resp_data_len),
739 min_t(int, be32_to_cpu(rsp->sense_data_len),
740 SCSI_SENSE_BUFFERSIZE));
743 if (rsp->flags & (SRP_RSP_FLAG_DOOVER | SRP_RSP_FLAG_DOUNDER))
744 scmnd->resid = be32_to_cpu(rsp->data_out_res_cnt);
745 else if (rsp->flags & (SRP_RSP_FLAG_DIOVER | SRP_RSP_FLAG_DIUNDER))
746 scmnd->resid = be32_to_cpu(rsp->data_in_res_cnt);
748 if (!req->tsk_mgmt) {
749 scmnd->host_scribble = (void *) -1L;
750 scmnd->scsi_done(scmnd);
752 srp_remove_req(target, req);
757 spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
760 static void srp_reconnect_work(void *target_ptr)
762 struct srp_target_port *target = target_ptr;
764 srp_reconnect_target(target);
767 static void srp_handle_recv(struct srp_target_port *target, struct ib_wc *wc)
772 iu = target->rx_ring[wc->wr_id & ~SRP_OP_RECV];
774 dma_sync_single_for_cpu(target->srp_host->dev->dev->dma_device, iu->dma,
775 target->max_ti_iu_len, DMA_FROM_DEVICE);
777 opcode = *(u8 *) iu->buf;
782 printk(KERN_ERR PFX "recv completion, opcode 0x%02x\n", opcode);
784 for (i = 0; i < wc->byte_len; ++i) {
786 printk(KERN_ERR " [%02x] ", i);
787 printk(" %02x", ((u8 *) iu->buf)[i]);
788 if ((i + 1) % 8 == 0)
792 if (wc->byte_len % 8)
798 srp_process_rsp(target, iu->buf);
802 /* XXX Handle target logout */
803 printk(KERN_WARNING PFX "Got target logout request\n");
807 printk(KERN_WARNING PFX "Unhandled SRP opcode 0x%02x\n", opcode);
811 dma_sync_single_for_device(target->srp_host->dev->dev->dma_device, iu->dma,
812 target->max_ti_iu_len, DMA_FROM_DEVICE);
815 static void srp_completion(struct ib_cq *cq, void *target_ptr)
817 struct srp_target_port *target = target_ptr;
821 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
822 while (ib_poll_cq(cq, 1, &wc) > 0) {
824 printk(KERN_ERR PFX "failed %s status %d\n",
825 wc.wr_id & SRP_OP_RECV ? "receive" : "send",
827 spin_lock_irqsave(target->scsi_host->host_lock, flags);
828 if (target->state == SRP_TARGET_LIVE)
829 schedule_work(&target->work);
830 spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
834 if (wc.wr_id & SRP_OP_RECV)
835 srp_handle_recv(target, &wc);
841 static int __srp_post_recv(struct srp_target_port *target)
845 struct ib_recv_wr wr, *bad_wr;
849 next = target->rx_head & (SRP_RQ_SIZE - 1);
850 wr.wr_id = next | SRP_OP_RECV;
851 iu = target->rx_ring[next];
854 list.length = iu->size;
855 list.lkey = target->srp_host->dev->mr->lkey;
861 ret = ib_post_recv(target->qp, &wr, &bad_wr);
868 static int srp_post_recv(struct srp_target_port *target)
873 spin_lock_irqsave(target->scsi_host->host_lock, flags);
874 ret = __srp_post_recv(target);
875 spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
881 * Must be called with target->scsi_host->host_lock held to protect
882 * req_lim and tx_head. Lock cannot be dropped between call here and
883 * call to __srp_post_send().
885 static struct srp_iu *__srp_get_tx_iu(struct srp_target_port *target)
887 if (target->tx_head - target->tx_tail >= SRP_SQ_SIZE)
890 if (unlikely(target->req_lim < 1)) {
891 if (printk_ratelimit())
892 printk(KERN_DEBUG PFX "Target has req_lim %d\n",
897 return target->tx_ring[target->tx_head & SRP_SQ_SIZE];
901 * Must be called with target->scsi_host->host_lock held to protect
902 * req_lim and tx_head.
904 static int __srp_post_send(struct srp_target_port *target,
905 struct srp_iu *iu, int len)
908 struct ib_send_wr wr, *bad_wr;
913 list.lkey = target->srp_host->dev->mr->lkey;
916 wr.wr_id = target->tx_head & SRP_SQ_SIZE;
919 wr.opcode = IB_WR_SEND;
920 wr.send_flags = IB_SEND_SIGNALED;
922 ret = ib_post_send(target->qp, &wr, &bad_wr);
932 static int srp_queuecommand(struct scsi_cmnd *scmnd,
933 void (*done)(struct scsi_cmnd *))
935 struct srp_target_port *target = host_to_target(scmnd->device->host);
936 struct srp_request *req;
941 if (target->state == SRP_TARGET_CONNECTING)
944 if (target->state == SRP_TARGET_DEAD ||
945 target->state == SRP_TARGET_REMOVED) {
946 scmnd->result = DID_BAD_TARGET << 16;
951 iu = __srp_get_tx_iu(target);
955 dma_sync_single_for_cpu(target->srp_host->dev->dev->dma_device, iu->dma,
956 SRP_MAX_IU_LEN, DMA_TO_DEVICE);
958 req = list_entry(target->free_reqs.next, struct srp_request, list);
960 scmnd->scsi_done = done;
962 scmnd->host_scribble = (void *) (long) req->index;
965 memset(cmd, 0, sizeof *cmd);
967 cmd->opcode = SRP_CMD;
968 cmd->lun = cpu_to_be64((u64) scmnd->device->lun << 48);
969 cmd->tag = req->index;
970 memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
975 req->tsk_mgmt = NULL;
977 len = srp_map_data(scmnd, target, req);
979 printk(KERN_ERR PFX "Failed to map data\n");
983 if (__srp_post_recv(target)) {
984 printk(KERN_ERR PFX "Recv failed\n");
988 dma_sync_single_for_device(target->srp_host->dev->dev->dma_device, iu->dma,
989 SRP_MAX_IU_LEN, DMA_TO_DEVICE);
991 if (__srp_post_send(target, iu, len)) {
992 printk(KERN_ERR PFX "Send failed\n");
996 list_move_tail(&req->list, &target->req_queue);
1001 srp_unmap_data(scmnd, target, req);
1004 return SCSI_MLQUEUE_HOST_BUSY;
1007 static int srp_alloc_iu_bufs(struct srp_target_port *target)
1011 for (i = 0; i < SRP_RQ_SIZE; ++i) {
1012 target->rx_ring[i] = srp_alloc_iu(target->srp_host,
1013 target->max_ti_iu_len,
1014 GFP_KERNEL, DMA_FROM_DEVICE);
1015 if (!target->rx_ring[i])
1019 for (i = 0; i < SRP_SQ_SIZE + 1; ++i) {
1020 target->tx_ring[i] = srp_alloc_iu(target->srp_host,
1022 GFP_KERNEL, DMA_TO_DEVICE);
1023 if (!target->tx_ring[i])
1030 for (i = 0; i < SRP_RQ_SIZE; ++i) {
1031 srp_free_iu(target->srp_host, target->rx_ring[i]);
1032 target->rx_ring[i] = NULL;
1035 for (i = 0; i < SRP_SQ_SIZE + 1; ++i) {
1036 srp_free_iu(target->srp_host, target->tx_ring[i]);
1037 target->tx_ring[i] = NULL;
1043 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
1044 struct ib_cm_event *event,
1045 struct srp_target_port *target)
1047 struct ib_class_port_info *cpi;
1050 switch (event->param.rej_rcvd.reason) {
1051 case IB_CM_REJ_PORT_CM_REDIRECT:
1052 cpi = event->param.rej_rcvd.ari;
1053 target->path.dlid = cpi->redirect_lid;
1054 target->path.pkey = cpi->redirect_pkey;
1055 cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
1056 memcpy(target->path.dgid.raw, cpi->redirect_gid, 16);
1058 target->status = target->path.dlid ?
1059 SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
1062 case IB_CM_REJ_PORT_REDIRECT:
1063 if (topspin_workarounds &&
1064 !memcmp(&target->ioc_guid, topspin_oui, 3)) {
1066 * Topspin/Cisco SRP gateways incorrectly send
1067 * reject reason code 25 when they mean 24
1070 memcpy(target->path.dgid.raw,
1071 event->param.rej_rcvd.ari, 16);
1073 printk(KERN_DEBUG PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
1074 (unsigned long long) be64_to_cpu(target->path.dgid.global.subnet_prefix),
1075 (unsigned long long) be64_to_cpu(target->path.dgid.global.interface_id));
1077 target->status = SRP_PORT_REDIRECT;
1079 printk(KERN_WARNING " REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
1080 target->status = -ECONNRESET;
1084 case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
1085 printk(KERN_WARNING " REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
1086 target->status = -ECONNRESET;
1089 case IB_CM_REJ_CONSUMER_DEFINED:
1090 opcode = *(u8 *) event->private_data;
1091 if (opcode == SRP_LOGIN_REJ) {
1092 struct srp_login_rej *rej = event->private_data;
1093 u32 reason = be32_to_cpu(rej->reason);
1095 if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
1096 printk(KERN_WARNING PFX
1097 "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
1099 printk(KERN_WARNING PFX
1100 "SRP LOGIN REJECTED, reason 0x%08x\n", reason);
1102 printk(KERN_WARNING " REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
1103 " opcode 0x%02x\n", opcode);
1104 target->status = -ECONNRESET;
1108 printk(KERN_WARNING " REJ reason 0x%x\n",
1109 event->param.rej_rcvd.reason);
1110 target->status = -ECONNRESET;
1114 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
1116 struct srp_target_port *target = cm_id->context;
1117 struct ib_qp_attr *qp_attr = NULL;
1122 switch (event->event) {
1123 case IB_CM_REQ_ERROR:
1124 printk(KERN_DEBUG PFX "Sending CM REQ failed\n");
1126 target->status = -ECONNRESET;
1129 case IB_CM_REP_RECEIVED:
1131 opcode = *(u8 *) event->private_data;
1133 if (opcode == SRP_LOGIN_RSP) {
1134 struct srp_login_rsp *rsp = event->private_data;
1136 target->max_ti_iu_len = be32_to_cpu(rsp->max_ti_iu_len);
1137 target->req_lim = be32_to_cpu(rsp->req_lim_delta);
1139 target->scsi_host->can_queue = min(target->req_lim,
1140 target->scsi_host->can_queue);
1142 printk(KERN_WARNING PFX "Unhandled RSP opcode %#x\n", opcode);
1143 target->status = -ECONNRESET;
1147 target->status = srp_alloc_iu_bufs(target);
1151 qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
1153 target->status = -ENOMEM;
1157 qp_attr->qp_state = IB_QPS_RTR;
1158 target->status = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1162 target->status = ib_modify_qp(target->qp, qp_attr, attr_mask);
1166 target->status = srp_post_recv(target);
1170 qp_attr->qp_state = IB_QPS_RTS;
1171 target->status = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1175 target->status = ib_modify_qp(target->qp, qp_attr, attr_mask);
1179 target->status = ib_send_cm_rtu(cm_id, NULL, 0);
1185 case IB_CM_REJ_RECEIVED:
1186 printk(KERN_DEBUG PFX "REJ received\n");
1189 srp_cm_rej_handler(cm_id, event, target);
1192 case IB_CM_MRA_RECEIVED:
1193 printk(KERN_ERR PFX "MRA received\n");
1196 case IB_CM_DREP_RECEIVED:
1199 case IB_CM_TIMEWAIT_EXIT:
1200 printk(KERN_ERR PFX "connection closed\n");
1207 printk(KERN_WARNING PFX "Unhandled CM event %d\n", event->event);
1212 complete(&target->done);
1219 static int srp_send_tsk_mgmt(struct srp_target_port *target,
1220 struct srp_request *req, u8 func)
1223 struct srp_tsk_mgmt *tsk_mgmt;
1225 spin_lock_irq(target->scsi_host->host_lock);
1227 if (target->state == SRP_TARGET_DEAD ||
1228 target->state == SRP_TARGET_REMOVED) {
1229 req->scmnd->result = DID_BAD_TARGET << 16;
1233 init_completion(&req->done);
1235 iu = __srp_get_tx_iu(target);
1240 memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
1242 tsk_mgmt->opcode = SRP_TSK_MGMT;
1243 tsk_mgmt->lun = cpu_to_be64((u64) req->scmnd->device->lun << 48);
1244 tsk_mgmt->tag = req->index | SRP_TAG_TSK_MGMT;
1245 tsk_mgmt->tsk_mgmt_func = func;
1246 tsk_mgmt->task_tag = req->index;
1248 if (__srp_post_send(target, iu, sizeof *tsk_mgmt))
1253 spin_unlock_irq(target->scsi_host->host_lock);
1255 if (!wait_for_completion_timeout(&req->done,
1256 msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
1262 spin_unlock_irq(target->scsi_host->host_lock);
1266 static int srp_find_req(struct srp_target_port *target,
1267 struct scsi_cmnd *scmnd,
1268 struct srp_request **req)
1270 if (scmnd->host_scribble == (void *) -1L)
1273 *req = &target->req_ring[(long) scmnd->host_scribble];
1278 static int srp_abort(struct scsi_cmnd *scmnd)
1280 struct srp_target_port *target = host_to_target(scmnd->device->host);
1281 struct srp_request *req;
1284 printk(KERN_ERR "SRP abort called\n");
1286 if (srp_find_req(target, scmnd, &req))
1288 if (srp_send_tsk_mgmt(target, req, SRP_TSK_ABORT_TASK))
1291 spin_lock_irq(target->scsi_host->host_lock);
1293 if (req->cmd_done) {
1294 srp_remove_req(target, req);
1295 scmnd->scsi_done(scmnd);
1296 } else if (!req->tsk_status) {
1297 srp_remove_req(target, req);
1298 scmnd->result = DID_ABORT << 16;
1302 spin_unlock_irq(target->scsi_host->host_lock);
1307 static int srp_reset_device(struct scsi_cmnd *scmnd)
1309 struct srp_target_port *target = host_to_target(scmnd->device->host);
1310 struct srp_request *req, *tmp;
1312 printk(KERN_ERR "SRP reset_device called\n");
1314 if (srp_find_req(target, scmnd, &req))
1316 if (srp_send_tsk_mgmt(target, req, SRP_TSK_LUN_RESET))
1318 if (req->tsk_status)
1321 spin_lock_irq(target->scsi_host->host_lock);
1323 list_for_each_entry_safe(req, tmp, &target->req_queue, list)
1324 if (req->scmnd->device == scmnd->device) {
1325 req->scmnd->result = DID_RESET << 16;
1326 req->scmnd->scsi_done(req->scmnd);
1327 srp_remove_req(target, req);
1330 spin_unlock_irq(target->scsi_host->host_lock);
1335 static int srp_reset_host(struct scsi_cmnd *scmnd)
1337 struct srp_target_port *target = host_to_target(scmnd->device->host);
1340 printk(KERN_ERR PFX "SRP reset_host called\n");
1342 if (!srp_reconnect_target(target))
1348 static ssize_t show_id_ext(struct class_device *cdev, char *buf)
1350 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1352 if (target->state == SRP_TARGET_DEAD ||
1353 target->state == SRP_TARGET_REMOVED)
1356 return sprintf(buf, "0x%016llx\n",
1357 (unsigned long long) be64_to_cpu(target->id_ext));
1360 static ssize_t show_ioc_guid(struct class_device *cdev, char *buf)
1362 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1364 if (target->state == SRP_TARGET_DEAD ||
1365 target->state == SRP_TARGET_REMOVED)
1368 return sprintf(buf, "0x%016llx\n",
1369 (unsigned long long) be64_to_cpu(target->ioc_guid));
1372 static ssize_t show_service_id(struct class_device *cdev, char *buf)
1374 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1376 if (target->state == SRP_TARGET_DEAD ||
1377 target->state == SRP_TARGET_REMOVED)
1380 return sprintf(buf, "0x%016llx\n",
1381 (unsigned long long) be64_to_cpu(target->service_id));
1384 static ssize_t show_pkey(struct class_device *cdev, char *buf)
1386 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1388 if (target->state == SRP_TARGET_DEAD ||
1389 target->state == SRP_TARGET_REMOVED)
1392 return sprintf(buf, "0x%04x\n", be16_to_cpu(target->path.pkey));
1395 static ssize_t show_dgid(struct class_device *cdev, char *buf)
1397 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1399 if (target->state == SRP_TARGET_DEAD ||
1400 target->state == SRP_TARGET_REMOVED)
1403 return sprintf(buf, "%04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x\n",
1404 be16_to_cpu(((__be16 *) target->path.dgid.raw)[0]),
1405 be16_to_cpu(((__be16 *) target->path.dgid.raw)[1]),
1406 be16_to_cpu(((__be16 *) target->path.dgid.raw)[2]),
1407 be16_to_cpu(((__be16 *) target->path.dgid.raw)[3]),
1408 be16_to_cpu(((__be16 *) target->path.dgid.raw)[4]),
1409 be16_to_cpu(((__be16 *) target->path.dgid.raw)[5]),
1410 be16_to_cpu(((__be16 *) target->path.dgid.raw)[6]),
1411 be16_to_cpu(((__be16 *) target->path.dgid.raw)[7]));
1414 static CLASS_DEVICE_ATTR(id_ext, S_IRUGO, show_id_ext, NULL);
1415 static CLASS_DEVICE_ATTR(ioc_guid, S_IRUGO, show_ioc_guid, NULL);
1416 static CLASS_DEVICE_ATTR(service_id, S_IRUGO, show_service_id, NULL);
1417 static CLASS_DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1418 static CLASS_DEVICE_ATTR(dgid, S_IRUGO, show_dgid, NULL);
1420 static struct class_device_attribute *srp_host_attrs[] = {
1421 &class_device_attr_id_ext,
1422 &class_device_attr_ioc_guid,
1423 &class_device_attr_service_id,
1424 &class_device_attr_pkey,
1425 &class_device_attr_dgid,
1429 static struct scsi_host_template srp_template = {
1430 .module = THIS_MODULE,
1432 .info = srp_target_info,
1433 .queuecommand = srp_queuecommand,
1434 .eh_abort_handler = srp_abort,
1435 .eh_device_reset_handler = srp_reset_device,
1436 .eh_host_reset_handler = srp_reset_host,
1437 .can_queue = SRP_SQ_SIZE,
1439 .sg_tablesize = SRP_MAX_INDIRECT,
1440 .cmd_per_lun = SRP_SQ_SIZE,
1441 .use_clustering = ENABLE_CLUSTERING,
1442 .shost_attrs = srp_host_attrs
1445 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
1447 sprintf(target->target_name, "SRP.T10:%016llX",
1448 (unsigned long long) be64_to_cpu(target->id_ext));
1450 if (scsi_add_host(target->scsi_host, host->dev->dev->dma_device))
1453 mutex_lock(&host->target_mutex);
1454 list_add_tail(&target->list, &host->target_list);
1455 mutex_unlock(&host->target_mutex);
1457 target->state = SRP_TARGET_LIVE;
1459 scsi_scan_target(&target->scsi_host->shost_gendev,
1460 0, target->scsi_id, SCAN_WILD_CARD, 0);
1465 static void srp_release_class_dev(struct class_device *class_dev)
1467 struct srp_host *host =
1468 container_of(class_dev, struct srp_host, class_dev);
1470 complete(&host->released);
1473 static struct class srp_class = {
1474 .name = "infiniband_srp",
1475 .release = srp_release_class_dev
1479 * Target ports are added by writing
1481 * id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
1482 * pkey=<P_Key>,service_id=<service ID>
1484 * to the add_target sysfs attribute.
1488 SRP_OPT_ID_EXT = 1 << 0,
1489 SRP_OPT_IOC_GUID = 1 << 1,
1490 SRP_OPT_DGID = 1 << 2,
1491 SRP_OPT_PKEY = 1 << 3,
1492 SRP_OPT_SERVICE_ID = 1 << 4,
1493 SRP_OPT_MAX_SECT = 1 << 5,
1494 SRP_OPT_ALL = (SRP_OPT_ID_EXT |
1498 SRP_OPT_SERVICE_ID),
1501 static match_table_t srp_opt_tokens = {
1502 { SRP_OPT_ID_EXT, "id_ext=%s" },
1503 { SRP_OPT_IOC_GUID, "ioc_guid=%s" },
1504 { SRP_OPT_DGID, "dgid=%s" },
1505 { SRP_OPT_PKEY, "pkey=%x" },
1506 { SRP_OPT_SERVICE_ID, "service_id=%s" },
1507 { SRP_OPT_MAX_SECT, "max_sect=%d" },
1508 { SRP_OPT_ERR, NULL }
1511 static int srp_parse_options(const char *buf, struct srp_target_port *target)
1513 char *options, *sep_opt;
1516 substring_t args[MAX_OPT_ARGS];
1522 options = kstrdup(buf, GFP_KERNEL);
1527 while ((p = strsep(&sep_opt, ",")) != NULL) {
1531 token = match_token(p, srp_opt_tokens, args);
1535 case SRP_OPT_ID_EXT:
1536 p = match_strdup(args);
1537 target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
1541 case SRP_OPT_IOC_GUID:
1542 p = match_strdup(args);
1543 target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
1548 p = match_strdup(args);
1549 if (strlen(p) != 32) {
1550 printk(KERN_WARNING PFX "bad dest GID parameter '%s'\n", p);
1555 for (i = 0; i < 16; ++i) {
1556 strlcpy(dgid, p + i * 2, 3);
1557 target->path.dgid.raw[i] = simple_strtoul(dgid, NULL, 16);
1563 if (match_hex(args, &token)) {
1564 printk(KERN_WARNING PFX "bad P_Key parameter '%s'\n", p);
1567 target->path.pkey = cpu_to_be16(token);
1570 case SRP_OPT_SERVICE_ID:
1571 p = match_strdup(args);
1572 target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
1576 case SRP_OPT_MAX_SECT:
1577 if (match_int(args, &token)) {
1578 printk(KERN_WARNING PFX "bad max sect parameter '%s'\n", p);
1581 target->scsi_host->max_sectors = token;
1585 printk(KERN_WARNING PFX "unknown parameter or missing value "
1586 "'%s' in target creation request\n", p);
1591 if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
1594 for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
1595 if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
1596 !(srp_opt_tokens[i].token & opt_mask))
1597 printk(KERN_WARNING PFX "target creation request is "
1598 "missing parameter '%s'\n",
1599 srp_opt_tokens[i].pattern);
1606 static ssize_t srp_create_target(struct class_device *class_dev,
1607 const char *buf, size_t count)
1609 struct srp_host *host =
1610 container_of(class_dev, struct srp_host, class_dev);
1611 struct Scsi_Host *target_host;
1612 struct srp_target_port *target;
1616 target_host = scsi_host_alloc(&srp_template,
1617 sizeof (struct srp_target_port));
1621 target_host->max_lun = SRP_MAX_LUN;
1623 target = host_to_target(target_host);
1624 memset(target, 0, sizeof *target);
1626 target->scsi_host = target_host;
1627 target->srp_host = host;
1629 INIT_WORK(&target->work, srp_reconnect_work, target);
1631 INIT_LIST_HEAD(&target->free_reqs);
1632 INIT_LIST_HEAD(&target->req_queue);
1633 for (i = 0; i < SRP_SQ_SIZE; ++i) {
1634 target->req_ring[i].index = i;
1635 list_add_tail(&target->req_ring[i].list, &target->free_reqs);
1638 ret = srp_parse_options(buf, target);
1642 ib_get_cached_gid(host->dev->dev, host->port, 0, &target->path.sgid);
1644 printk(KERN_DEBUG PFX "new target: id_ext %016llx ioc_guid %016llx pkey %04x "
1645 "service_id %016llx dgid %04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x\n",
1646 (unsigned long long) be64_to_cpu(target->id_ext),
1647 (unsigned long long) be64_to_cpu(target->ioc_guid),
1648 be16_to_cpu(target->path.pkey),
1649 (unsigned long long) be64_to_cpu(target->service_id),
1650 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[0]),
1651 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[2]),
1652 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[4]),
1653 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[6]),
1654 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[8]),
1655 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[10]),
1656 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[12]),
1657 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[14]));
1659 ret = srp_create_target_ib(target);
1663 target->cm_id = ib_create_cm_id(host->dev->dev, srp_cm_handler, target);
1664 if (IS_ERR(target->cm_id)) {
1665 ret = PTR_ERR(target->cm_id);
1669 ret = srp_connect_target(target);
1671 printk(KERN_ERR PFX "Connection failed\n");
1675 ret = srp_add_target(host, target);
1677 goto err_disconnect;
1682 srp_disconnect_target(target);
1685 ib_destroy_cm_id(target->cm_id);
1688 srp_free_target_ib(target);
1691 scsi_host_put(target_host);
1696 static CLASS_DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
1698 static ssize_t show_ibdev(struct class_device *class_dev, char *buf)
1700 struct srp_host *host =
1701 container_of(class_dev, struct srp_host, class_dev);
1703 return sprintf(buf, "%s\n", host->dev->dev->name);
1706 static CLASS_DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
1708 static ssize_t show_port(struct class_device *class_dev, char *buf)
1710 struct srp_host *host =
1711 container_of(class_dev, struct srp_host, class_dev);
1713 return sprintf(buf, "%d\n", host->port);
1716 static CLASS_DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
1718 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
1720 struct srp_host *host;
1722 host = kzalloc(sizeof *host, GFP_KERNEL);
1726 INIT_LIST_HEAD(&host->target_list);
1727 mutex_init(&host->target_mutex);
1728 init_completion(&host->released);
1732 host->initiator_port_id[7] = port;
1733 memcpy(host->initiator_port_id + 8, &device->dev->node_guid, 8);
1735 host->class_dev.class = &srp_class;
1736 host->class_dev.dev = device->dev->dma_device;
1737 snprintf(host->class_dev.class_id, BUS_ID_SIZE, "srp-%s-%d",
1738 device->dev->name, port);
1740 if (class_device_register(&host->class_dev))
1742 if (class_device_create_file(&host->class_dev, &class_device_attr_add_target))
1744 if (class_device_create_file(&host->class_dev, &class_device_attr_ibdev))
1746 if (class_device_create_file(&host->class_dev, &class_device_attr_port))
1752 class_device_unregister(&host->class_dev);
1760 static void srp_add_one(struct ib_device *device)
1762 struct srp_device *srp_dev;
1763 struct ib_device_attr *dev_attr;
1764 struct ib_fmr_pool_param fmr_param;
1765 struct srp_host *host;
1768 dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
1772 if (ib_query_device(device, dev_attr)) {
1773 printk(KERN_WARNING PFX "Query device failed for %s\n",
1778 srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
1783 * Use the smallest page size supported by the HCA, down to a
1784 * minimum of 512 bytes (which is the smallest sector that a
1785 * SCSI command will ever carry).
1787 srp_dev->fmr_page_shift = max(9, ffs(dev_attr->page_size_cap) - 1);
1788 srp_dev->fmr_page_size = 1 << srp_dev->fmr_page_shift;
1789 srp_dev->fmr_page_mask = ~((unsigned long) srp_dev->fmr_page_size - 1);
1791 INIT_LIST_HEAD(&srp_dev->dev_list);
1793 srp_dev->dev = device;
1794 srp_dev->pd = ib_alloc_pd(device);
1795 if (IS_ERR(srp_dev->pd))
1798 srp_dev->mr = ib_get_dma_mr(srp_dev->pd,
1799 IB_ACCESS_LOCAL_WRITE |
1800 IB_ACCESS_REMOTE_READ |
1801 IB_ACCESS_REMOTE_WRITE);
1802 if (IS_ERR(srp_dev->mr))
1805 memset(&fmr_param, 0, sizeof fmr_param);
1806 fmr_param.pool_size = SRP_FMR_POOL_SIZE;
1807 fmr_param.dirty_watermark = SRP_FMR_DIRTY_SIZE;
1808 fmr_param.cache = 1;
1809 fmr_param.max_pages_per_fmr = SRP_FMR_SIZE;
1810 fmr_param.page_shift = srp_dev->fmr_page_shift;
1811 fmr_param.access = (IB_ACCESS_LOCAL_WRITE |
1812 IB_ACCESS_REMOTE_WRITE |
1813 IB_ACCESS_REMOTE_READ);
1815 srp_dev->fmr_pool = ib_create_fmr_pool(srp_dev->pd, &fmr_param);
1816 if (IS_ERR(srp_dev->fmr_pool))
1817 srp_dev->fmr_pool = NULL;
1819 if (device->node_type == IB_NODE_SWITCH) {
1824 e = device->phys_port_cnt;
1827 for (p = s; p <= e; ++p) {
1828 host = srp_add_port(srp_dev, p);
1830 list_add_tail(&host->list, &srp_dev->dev_list);
1833 ib_set_client_data(device, &srp_client, srp_dev);
1838 ib_dealloc_pd(srp_dev->pd);
1847 static void srp_remove_one(struct ib_device *device)
1849 struct srp_device *srp_dev;
1850 struct srp_host *host, *tmp_host;
1851 LIST_HEAD(target_list);
1852 struct srp_target_port *target, *tmp_target;
1853 unsigned long flags;
1855 srp_dev = ib_get_client_data(device, &srp_client);
1857 list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
1858 class_device_unregister(&host->class_dev);
1860 * Wait for the sysfs entry to go away, so that no new
1861 * target ports can be created.
1863 wait_for_completion(&host->released);
1866 * Mark all target ports as removed, so we stop queueing
1867 * commands and don't try to reconnect.
1869 mutex_lock(&host->target_mutex);
1870 list_for_each_entry(target, &host->target_list, list) {
1871 spin_lock_irqsave(target->scsi_host->host_lock, flags);
1872 if (target->state != SRP_TARGET_REMOVED)
1873 target->state = SRP_TARGET_REMOVED;
1874 spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
1876 mutex_unlock(&host->target_mutex);
1879 * Wait for any reconnection tasks that may have
1880 * started before we marked our target ports as
1881 * removed, and any target port removal tasks.
1883 flush_scheduled_work();
1885 list_for_each_entry_safe(target, tmp_target,
1886 &host->target_list, list) {
1887 scsi_remove_host(target->scsi_host);
1888 srp_disconnect_target(target);
1889 ib_destroy_cm_id(target->cm_id);
1890 srp_free_target_ib(target);
1891 scsi_host_put(target->scsi_host);
1897 if (srp_dev->fmr_pool)
1898 ib_destroy_fmr_pool(srp_dev->fmr_pool);
1899 ib_dereg_mr(srp_dev->mr);
1900 ib_dealloc_pd(srp_dev->pd);
1905 static int __init srp_init_module(void)
1909 ret = class_register(&srp_class);
1911 printk(KERN_ERR PFX "couldn't register class infiniband_srp\n");
1915 ret = ib_register_client(&srp_client);
1917 printk(KERN_ERR PFX "couldn't register IB client\n");
1918 class_unregister(&srp_class);
1925 static void __exit srp_cleanup_module(void)
1927 ib_unregister_client(&srp_client);
1928 class_unregister(&srp_class);
1931 module_init(srp_init_module);
1932 module_exit(srp_cleanup_module);