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 void srp_add_one(struct ib_device *device);
81 static void srp_remove_one(struct ib_device *device);
82 static void srp_completion(struct ib_cq *cq, void *target_ptr);
83 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event);
85 static struct ib_client srp_client = {
88 .remove = srp_remove_one
91 static inline struct srp_target_port *host_to_target(struct Scsi_Host *host)
93 return (struct srp_target_port *) host->hostdata;
96 static const char *srp_target_info(struct Scsi_Host *host)
98 return host_to_target(host)->target_name;
101 static struct srp_iu *srp_alloc_iu(struct srp_host *host, size_t size,
103 enum dma_data_direction direction)
107 iu = kmalloc(sizeof *iu, gfp_mask);
111 iu->buf = kzalloc(size, gfp_mask);
115 iu->dma = dma_map_single(host->dev->dev->dma_device,
116 iu->buf, size, direction);
117 if (dma_mapping_error(iu->dma))
121 iu->direction = direction;
133 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
138 dma_unmap_single(host->dev->dev->dma_device,
139 iu->dma, iu->size, iu->direction);
144 static void srp_qp_event(struct ib_event *event, void *context)
146 printk(KERN_ERR PFX "QP event %d\n", event->event);
149 static int srp_init_qp(struct srp_target_port *target,
152 struct ib_qp_attr *attr;
155 attr = kmalloc(sizeof *attr, GFP_KERNEL);
159 ret = ib_find_cached_pkey(target->srp_host->dev->dev,
160 target->srp_host->port,
161 be16_to_cpu(target->path.pkey),
166 attr->qp_state = IB_QPS_INIT;
167 attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
168 IB_ACCESS_REMOTE_WRITE);
169 attr->port_num = target->srp_host->port;
171 ret = ib_modify_qp(qp, attr,
182 static int srp_create_target_ib(struct srp_target_port *target)
184 struct ib_qp_init_attr *init_attr;
187 init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
191 target->cq = ib_create_cq(target->srp_host->dev->dev, srp_completion,
192 NULL, target, SRP_CQ_SIZE);
193 if (IS_ERR(target->cq)) {
194 ret = PTR_ERR(target->cq);
198 ib_req_notify_cq(target->cq, IB_CQ_NEXT_COMP);
200 init_attr->event_handler = srp_qp_event;
201 init_attr->cap.max_send_wr = SRP_SQ_SIZE;
202 init_attr->cap.max_recv_wr = SRP_RQ_SIZE;
203 init_attr->cap.max_recv_sge = 1;
204 init_attr->cap.max_send_sge = 1;
205 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
206 init_attr->qp_type = IB_QPT_RC;
207 init_attr->send_cq = target->cq;
208 init_attr->recv_cq = target->cq;
210 target->qp = ib_create_qp(target->srp_host->dev->pd, init_attr);
211 if (IS_ERR(target->qp)) {
212 ret = PTR_ERR(target->qp);
213 ib_destroy_cq(target->cq);
217 ret = srp_init_qp(target, target->qp);
219 ib_destroy_qp(target->qp);
220 ib_destroy_cq(target->cq);
229 static void srp_free_target_ib(struct srp_target_port *target)
233 ib_destroy_qp(target->qp);
234 ib_destroy_cq(target->cq);
236 for (i = 0; i < SRP_RQ_SIZE; ++i)
237 srp_free_iu(target->srp_host, target->rx_ring[i]);
238 for (i = 0; i < SRP_SQ_SIZE + 1; ++i)
239 srp_free_iu(target->srp_host, target->tx_ring[i]);
242 static void srp_path_rec_completion(int status,
243 struct ib_sa_path_rec *pathrec,
246 struct srp_target_port *target = target_ptr;
248 target->status = status;
250 printk(KERN_ERR PFX "Got failed path rec status %d\n", status);
252 target->path = *pathrec;
253 complete(&target->done);
256 static int srp_lookup_path(struct srp_target_port *target)
258 target->path.numb_path = 1;
260 init_completion(&target->done);
262 target->path_query_id = ib_sa_path_rec_get(target->srp_host->dev->dev,
263 target->srp_host->port,
265 IB_SA_PATH_REC_DGID |
266 IB_SA_PATH_REC_SGID |
267 IB_SA_PATH_REC_NUMB_PATH |
269 SRP_PATH_REC_TIMEOUT_MS,
271 srp_path_rec_completion,
272 target, &target->path_query);
273 if (target->path_query_id < 0)
274 return target->path_query_id;
276 wait_for_completion(&target->done);
278 if (target->status < 0)
279 printk(KERN_WARNING PFX "Path record query failed\n");
281 return target->status;
284 static int srp_send_req(struct srp_target_port *target)
287 struct ib_cm_req_param param;
288 struct srp_login_req priv;
292 req = kzalloc(sizeof *req, GFP_KERNEL);
296 req->param.primary_path = &target->path;
297 req->param.alternate_path = NULL;
298 req->param.service_id = target->service_id;
299 req->param.qp_num = target->qp->qp_num;
300 req->param.qp_type = target->qp->qp_type;
301 req->param.private_data = &req->priv;
302 req->param.private_data_len = sizeof req->priv;
303 req->param.flow_control = 1;
305 get_random_bytes(&req->param.starting_psn, 4);
306 req->param.starting_psn &= 0xffffff;
309 * Pick some arbitrary defaults here; we could make these
310 * module parameters if anyone cared about setting them.
312 req->param.responder_resources = 4;
313 req->param.remote_cm_response_timeout = 20;
314 req->param.local_cm_response_timeout = 20;
315 req->param.retry_count = 7;
316 req->param.rnr_retry_count = 7;
317 req->param.max_cm_retries = 15;
319 req->priv.opcode = SRP_LOGIN_REQ;
321 req->priv.req_it_iu_len = cpu_to_be32(srp_max_iu_len);
322 req->priv.req_buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
323 SRP_BUF_FORMAT_INDIRECT);
325 * In the published SRP specification (draft rev. 16a), the
326 * port identifier format is 8 bytes of ID extension followed
327 * by 8 bytes of GUID. Older drafts put the two halves in the
328 * opposite order, so that the GUID comes first.
330 * Targets conforming to these obsolete drafts can be
331 * recognized by the I/O Class they report.
333 if (target->io_class == SRP_REV10_IB_IO_CLASS) {
334 memcpy(req->priv.initiator_port_id,
335 target->srp_host->initiator_port_id + 8, 8);
336 memcpy(req->priv.initiator_port_id + 8,
337 target->srp_host->initiator_port_id, 8);
338 memcpy(req->priv.target_port_id, &target->ioc_guid, 8);
339 memcpy(req->priv.target_port_id + 8, &target->id_ext, 8);
341 memcpy(req->priv.initiator_port_id,
342 target->srp_host->initiator_port_id, 16);
343 memcpy(req->priv.target_port_id, &target->id_ext, 8);
344 memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
348 * Topspin/Cisco SRP targets will reject our login unless we
349 * zero out the first 8 bytes of our initiator port ID. The
350 * second 8 bytes must be our local node GUID, but we always
353 if (topspin_workarounds && !memcmp(&target->ioc_guid, topspin_oui, 3)) {
354 printk(KERN_DEBUG PFX "Topspin/Cisco initiator port ID workaround "
355 "activated for target GUID %016llx\n",
356 (unsigned long long) be64_to_cpu(target->ioc_guid));
357 memset(req->priv.initiator_port_id, 0, 8);
360 status = ib_send_cm_req(target->cm_id, &req->param);
367 static void srp_disconnect_target(struct srp_target_port *target)
369 /* XXX should send SRP_I_LOGOUT request */
371 init_completion(&target->done);
372 if (ib_send_cm_dreq(target->cm_id, NULL, 0)) {
373 printk(KERN_DEBUG PFX "Sending CM DREQ failed\n");
376 wait_for_completion(&target->done);
379 static void srp_remove_work(void *target_ptr)
381 struct srp_target_port *target = target_ptr;
383 spin_lock_irq(target->scsi_host->host_lock);
384 if (target->state != SRP_TARGET_DEAD) {
385 spin_unlock_irq(target->scsi_host->host_lock);
388 target->state = SRP_TARGET_REMOVED;
389 spin_unlock_irq(target->scsi_host->host_lock);
391 spin_lock(&target->srp_host->target_lock);
392 list_del(&target->list);
393 spin_unlock(&target->srp_host->target_lock);
395 scsi_remove_host(target->scsi_host);
396 ib_destroy_cm_id(target->cm_id);
397 srp_free_target_ib(target);
398 scsi_host_put(target->scsi_host);
401 static int srp_connect_target(struct srp_target_port *target)
405 ret = srp_lookup_path(target);
410 init_completion(&target->done);
411 ret = srp_send_req(target);
414 wait_for_completion(&target->done);
417 * The CM event handling code will set status to
418 * SRP_PORT_REDIRECT if we get a port redirect REJ
419 * back, or SRP_DLID_REDIRECT if we get a lid/qp
422 switch (target->status) {
426 case SRP_PORT_REDIRECT:
427 ret = srp_lookup_path(target);
432 case SRP_DLID_REDIRECT:
436 return target->status;
441 static void srp_unmap_data(struct scsi_cmnd *scmnd,
442 struct srp_target_port *target,
443 struct srp_request *req)
445 struct scatterlist *scat;
448 if (!scmnd->request_buffer ||
449 (scmnd->sc_data_direction != DMA_TO_DEVICE &&
450 scmnd->sc_data_direction != DMA_FROM_DEVICE))
454 ib_fmr_pool_unmap(req->fmr);
459 * This handling of non-SG commands can be killed when the
460 * SCSI midlayer no longer generates non-SG commands.
462 if (likely(scmnd->use_sg)) {
463 nents = scmnd->use_sg;
464 scat = scmnd->request_buffer;
467 scat = &req->fake_sg;
470 dma_unmap_sg(target->srp_host->dev->dev->dma_device, scat, nents,
471 scmnd->sc_data_direction);
474 static void srp_remove_req(struct srp_target_port *target, struct srp_request *req)
476 srp_unmap_data(req->scmnd, target, req);
477 list_move_tail(&req->list, &target->free_reqs);
480 static void srp_reset_req(struct srp_target_port *target, struct srp_request *req)
482 req->scmnd->result = DID_RESET << 16;
483 req->scmnd->scsi_done(req->scmnd);
484 srp_remove_req(target, req);
487 static int srp_reconnect_target(struct srp_target_port *target)
489 struct ib_cm_id *new_cm_id;
490 struct ib_qp_attr qp_attr;
491 struct srp_request *req, *tmp;
495 spin_lock_irq(target->scsi_host->host_lock);
496 if (target->state != SRP_TARGET_LIVE) {
497 spin_unlock_irq(target->scsi_host->host_lock);
500 target->state = SRP_TARGET_CONNECTING;
501 spin_unlock_irq(target->scsi_host->host_lock);
503 srp_disconnect_target(target);
505 * Now get a new local CM ID so that we avoid confusing the
506 * target in case things are really fouled up.
508 new_cm_id = ib_create_cm_id(target->srp_host->dev->dev,
509 srp_cm_handler, target);
510 if (IS_ERR(new_cm_id)) {
511 ret = PTR_ERR(new_cm_id);
514 ib_destroy_cm_id(target->cm_id);
515 target->cm_id = new_cm_id;
517 qp_attr.qp_state = IB_QPS_RESET;
518 ret = ib_modify_qp(target->qp, &qp_attr, IB_QP_STATE);
522 ret = srp_init_qp(target, target->qp);
526 while (ib_poll_cq(target->cq, 1, &wc) > 0)
529 list_for_each_entry_safe(req, tmp, &target->req_queue, list)
530 srp_reset_req(target, req);
536 ret = srp_connect_target(target);
540 spin_lock_irq(target->scsi_host->host_lock);
541 if (target->state == SRP_TARGET_CONNECTING) {
543 target->state = SRP_TARGET_LIVE;
546 spin_unlock_irq(target->scsi_host->host_lock);
551 printk(KERN_ERR PFX "reconnect failed (%d), removing target port.\n", ret);
554 * We couldn't reconnect, so kill our target port off.
555 * However, we have to defer the real removal because we might
556 * be in the context of the SCSI error handler now, which
557 * would deadlock if we call scsi_remove_host().
559 spin_lock_irq(target->scsi_host->host_lock);
560 if (target->state == SRP_TARGET_CONNECTING) {
561 target->state = SRP_TARGET_DEAD;
562 INIT_WORK(&target->work, srp_remove_work, target);
563 schedule_work(&target->work);
565 spin_unlock_irq(target->scsi_host->host_lock);
570 static int srp_map_fmr(struct srp_device *dev, struct scatterlist *scat,
571 int sg_cnt, struct srp_request *req,
572 struct srp_direct_buf *buf)
585 for (i = 0; i < sg_cnt; ++i) {
586 if (sg_dma_address(&scat[i]) & ~dev->fmr_page_mask) {
592 if ((sg_dma_address(&scat[i]) + sg_dma_len(&scat[i])) &
593 ~dev->fmr_page_mask) {
600 len += sg_dma_len(&scat[i]);
603 page_cnt += len >> dev->fmr_page_shift;
604 if (page_cnt > SRP_FMR_SIZE)
607 dma_pages = kmalloc(sizeof (u64) * page_cnt, GFP_ATOMIC);
612 for (i = 0; i < sg_cnt; ++i)
613 for (j = 0; j < sg_dma_len(&scat[i]); j += dev->fmr_page_size)
614 dma_pages[page_cnt++] =
615 (sg_dma_address(&scat[i]) & dev->fmr_page_mask) + j;
617 req->fmr = ib_fmr_pool_map_phys(dev->fmr_pool,
618 dma_pages, page_cnt, &io_addr);
619 if (IS_ERR(req->fmr)) {
620 ret = PTR_ERR(req->fmr);
624 buf->va = cpu_to_be64(sg_dma_address(&scat[0]) & ~dev->fmr_page_mask);
625 buf->key = cpu_to_be32(req->fmr->fmr->rkey);
626 buf->len = cpu_to_be32(len);
636 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_target_port *target,
637 struct srp_request *req)
639 struct scatterlist *scat;
640 struct srp_cmd *cmd = req->cmd->buf;
641 int len, nents, count;
642 u8 fmt = SRP_DATA_DESC_DIRECT;
644 if (!scmnd->request_buffer || scmnd->sc_data_direction == DMA_NONE)
645 return sizeof (struct srp_cmd);
647 if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
648 scmnd->sc_data_direction != DMA_TO_DEVICE) {
649 printk(KERN_WARNING PFX "Unhandled data direction %d\n",
650 scmnd->sc_data_direction);
655 * This handling of non-SG commands can be killed when the
656 * SCSI midlayer no longer generates non-SG commands.
658 if (likely(scmnd->use_sg)) {
659 nents = scmnd->use_sg;
660 scat = scmnd->request_buffer;
663 scat = &req->fake_sg;
664 sg_init_one(scat, scmnd->request_buffer, scmnd->request_bufflen);
667 count = dma_map_sg(target->srp_host->dev->dev->dma_device,
668 scat, nents, scmnd->sc_data_direction);
670 fmt = SRP_DATA_DESC_DIRECT;
671 len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
675 * The midlayer only generated a single gather/scatter
676 * entry, or DMA mapping coalesced everything to a
677 * single entry. So a direct descriptor along with
678 * the DMA MR suffices.
680 struct srp_direct_buf *buf = (void *) cmd->add_data;
682 buf->va = cpu_to_be64(sg_dma_address(scat));
683 buf->key = cpu_to_be32(target->srp_host->dev->mr->rkey);
684 buf->len = cpu_to_be32(sg_dma_len(scat));
685 } else if (srp_map_fmr(target->srp_host->dev, scat, count, req,
686 (void *) cmd->add_data)) {
688 * FMR mapping failed, and the scatterlist has more
689 * than one entry. Generate an indirect memory
692 struct srp_indirect_buf *buf = (void *) cmd->add_data;
696 fmt = SRP_DATA_DESC_INDIRECT;
697 len = sizeof (struct srp_cmd) +
698 sizeof (struct srp_indirect_buf) +
699 count * sizeof (struct srp_direct_buf);
701 for (i = 0; i < count; ++i) {
702 buf->desc_list[i].va =
703 cpu_to_be64(sg_dma_address(&scat[i]));
704 buf->desc_list[i].key =
705 cpu_to_be32(target->srp_host->dev->mr->rkey);
706 buf->desc_list[i].len =
707 cpu_to_be32(sg_dma_len(&scat[i]));
708 datalen += sg_dma_len(&scat[i]);
711 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
712 cmd->data_out_desc_cnt = count;
714 cmd->data_in_desc_cnt = count;
717 cpu_to_be64(req->cmd->dma + sizeof *cmd + sizeof *buf);
718 buf->table_desc.key =
719 cpu_to_be32(target->srp_host->dev->mr->rkey);
720 buf->table_desc.len =
721 cpu_to_be32(count * sizeof (struct srp_direct_buf));
723 buf->len = cpu_to_be32(datalen);
726 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
727 cmd->buf_fmt = fmt << 4;
734 static void srp_process_rsp(struct srp_target_port *target, struct srp_rsp *rsp)
736 struct srp_request *req;
737 struct scsi_cmnd *scmnd;
741 delta = (s32) be32_to_cpu(rsp->req_lim_delta);
743 spin_lock_irqsave(target->scsi_host->host_lock, flags);
745 target->req_lim += delta;
747 req = &target->req_ring[rsp->tag & ~SRP_TAG_TSK_MGMT];
749 if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
750 if (be32_to_cpu(rsp->resp_data_len) < 4)
751 req->tsk_status = -1;
753 req->tsk_status = rsp->data[3];
754 complete(&req->done);
758 printk(KERN_ERR "Null scmnd for RSP w/tag %016llx\n",
759 (unsigned long long) rsp->tag);
760 scmnd->result = rsp->status;
762 if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
763 memcpy(scmnd->sense_buffer, rsp->data +
764 be32_to_cpu(rsp->resp_data_len),
765 min_t(int, be32_to_cpu(rsp->sense_data_len),
766 SCSI_SENSE_BUFFERSIZE));
769 if (rsp->flags & (SRP_RSP_FLAG_DOOVER | SRP_RSP_FLAG_DOUNDER))
770 scmnd->resid = be32_to_cpu(rsp->data_out_res_cnt);
771 else if (rsp->flags & (SRP_RSP_FLAG_DIOVER | SRP_RSP_FLAG_DIUNDER))
772 scmnd->resid = be32_to_cpu(rsp->data_in_res_cnt);
774 if (!req->tsk_mgmt) {
775 scmnd->host_scribble = (void *) -1L;
776 scmnd->scsi_done(scmnd);
778 srp_remove_req(target, req);
783 spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
786 static void srp_reconnect_work(void *target_ptr)
788 struct srp_target_port *target = target_ptr;
790 srp_reconnect_target(target);
793 static void srp_handle_recv(struct srp_target_port *target, struct ib_wc *wc)
798 iu = target->rx_ring[wc->wr_id & ~SRP_OP_RECV];
800 dma_sync_single_for_cpu(target->srp_host->dev->dev->dma_device, iu->dma,
801 target->max_ti_iu_len, DMA_FROM_DEVICE);
803 opcode = *(u8 *) iu->buf;
808 printk(KERN_ERR PFX "recv completion, opcode 0x%02x\n", opcode);
810 for (i = 0; i < wc->byte_len; ++i) {
812 printk(KERN_ERR " [%02x] ", i);
813 printk(" %02x", ((u8 *) iu->buf)[i]);
814 if ((i + 1) % 8 == 0)
818 if (wc->byte_len % 8)
824 srp_process_rsp(target, iu->buf);
828 /* XXX Handle target logout */
829 printk(KERN_WARNING PFX "Got target logout request\n");
833 printk(KERN_WARNING PFX "Unhandled SRP opcode 0x%02x\n", opcode);
837 dma_sync_single_for_device(target->srp_host->dev->dev->dma_device, iu->dma,
838 target->max_ti_iu_len, DMA_FROM_DEVICE);
841 static void srp_completion(struct ib_cq *cq, void *target_ptr)
843 struct srp_target_port *target = target_ptr;
847 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
848 while (ib_poll_cq(cq, 1, &wc) > 0) {
850 printk(KERN_ERR PFX "failed %s status %d\n",
851 wc.wr_id & SRP_OP_RECV ? "receive" : "send",
853 spin_lock_irqsave(target->scsi_host->host_lock, flags);
854 if (target->state == SRP_TARGET_LIVE)
855 schedule_work(&target->work);
856 spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
860 if (wc.wr_id & SRP_OP_RECV)
861 srp_handle_recv(target, &wc);
867 static int __srp_post_recv(struct srp_target_port *target)
871 struct ib_recv_wr wr, *bad_wr;
875 next = target->rx_head & (SRP_RQ_SIZE - 1);
876 wr.wr_id = next | SRP_OP_RECV;
877 iu = target->rx_ring[next];
880 list.length = iu->size;
881 list.lkey = target->srp_host->dev->mr->lkey;
887 ret = ib_post_recv(target->qp, &wr, &bad_wr);
894 static int srp_post_recv(struct srp_target_port *target)
899 spin_lock_irqsave(target->scsi_host->host_lock, flags);
900 ret = __srp_post_recv(target);
901 spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
907 * Must be called with target->scsi_host->host_lock held to protect
908 * req_lim and tx_head. Lock cannot be dropped between call here and
909 * call to __srp_post_send().
911 static struct srp_iu *__srp_get_tx_iu(struct srp_target_port *target)
913 if (target->tx_head - target->tx_tail >= SRP_SQ_SIZE)
916 if (unlikely(target->req_lim < 1))
917 ++target->zero_req_lim;
919 return target->tx_ring[target->tx_head & SRP_SQ_SIZE];
923 * Must be called with target->scsi_host->host_lock held to protect
924 * req_lim and tx_head.
926 static int __srp_post_send(struct srp_target_port *target,
927 struct srp_iu *iu, int len)
930 struct ib_send_wr wr, *bad_wr;
935 list.lkey = target->srp_host->dev->mr->lkey;
938 wr.wr_id = target->tx_head & SRP_SQ_SIZE;
941 wr.opcode = IB_WR_SEND;
942 wr.send_flags = IB_SEND_SIGNALED;
944 ret = ib_post_send(target->qp, &wr, &bad_wr);
954 static int srp_queuecommand(struct scsi_cmnd *scmnd,
955 void (*done)(struct scsi_cmnd *))
957 struct srp_target_port *target = host_to_target(scmnd->device->host);
958 struct srp_request *req;
963 if (target->state == SRP_TARGET_CONNECTING)
966 if (target->state == SRP_TARGET_DEAD ||
967 target->state == SRP_TARGET_REMOVED) {
968 scmnd->result = DID_BAD_TARGET << 16;
973 iu = __srp_get_tx_iu(target);
977 dma_sync_single_for_cpu(target->srp_host->dev->dev->dma_device, iu->dma,
978 srp_max_iu_len, DMA_TO_DEVICE);
980 req = list_entry(target->free_reqs.next, struct srp_request, list);
982 scmnd->scsi_done = done;
984 scmnd->host_scribble = (void *) (long) req->index;
987 memset(cmd, 0, sizeof *cmd);
989 cmd->opcode = SRP_CMD;
990 cmd->lun = cpu_to_be64((u64) scmnd->device->lun << 48);
991 cmd->tag = req->index;
992 memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
997 req->tsk_mgmt = NULL;
999 len = srp_map_data(scmnd, target, req);
1001 printk(KERN_ERR PFX "Failed to map data\n");
1005 if (__srp_post_recv(target)) {
1006 printk(KERN_ERR PFX "Recv failed\n");
1010 dma_sync_single_for_device(target->srp_host->dev->dev->dma_device, iu->dma,
1011 srp_max_iu_len, DMA_TO_DEVICE);
1013 if (__srp_post_send(target, iu, len)) {
1014 printk(KERN_ERR PFX "Send failed\n");
1018 list_move_tail(&req->list, &target->req_queue);
1023 srp_unmap_data(scmnd, target, req);
1026 return SCSI_MLQUEUE_HOST_BUSY;
1029 static int srp_alloc_iu_bufs(struct srp_target_port *target)
1033 for (i = 0; i < SRP_RQ_SIZE; ++i) {
1034 target->rx_ring[i] = srp_alloc_iu(target->srp_host,
1035 target->max_ti_iu_len,
1036 GFP_KERNEL, DMA_FROM_DEVICE);
1037 if (!target->rx_ring[i])
1041 for (i = 0; i < SRP_SQ_SIZE + 1; ++i) {
1042 target->tx_ring[i] = srp_alloc_iu(target->srp_host,
1044 GFP_KERNEL, DMA_TO_DEVICE);
1045 if (!target->tx_ring[i])
1052 for (i = 0; i < SRP_RQ_SIZE; ++i) {
1053 srp_free_iu(target->srp_host, target->rx_ring[i]);
1054 target->rx_ring[i] = NULL;
1057 for (i = 0; i < SRP_SQ_SIZE + 1; ++i) {
1058 srp_free_iu(target->srp_host, target->tx_ring[i]);
1059 target->tx_ring[i] = NULL;
1065 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
1066 struct ib_cm_event *event,
1067 struct srp_target_port *target)
1069 struct ib_class_port_info *cpi;
1072 switch (event->param.rej_rcvd.reason) {
1073 case IB_CM_REJ_PORT_CM_REDIRECT:
1074 cpi = event->param.rej_rcvd.ari;
1075 target->path.dlid = cpi->redirect_lid;
1076 target->path.pkey = cpi->redirect_pkey;
1077 cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
1078 memcpy(target->path.dgid.raw, cpi->redirect_gid, 16);
1080 target->status = target->path.dlid ?
1081 SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
1084 case IB_CM_REJ_PORT_REDIRECT:
1085 if (topspin_workarounds &&
1086 !memcmp(&target->ioc_guid, topspin_oui, 3)) {
1088 * Topspin/Cisco SRP gateways incorrectly send
1089 * reject reason code 25 when they mean 24
1092 memcpy(target->path.dgid.raw,
1093 event->param.rej_rcvd.ari, 16);
1095 printk(KERN_DEBUG PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
1096 (unsigned long long) be64_to_cpu(target->path.dgid.global.subnet_prefix),
1097 (unsigned long long) be64_to_cpu(target->path.dgid.global.interface_id));
1099 target->status = SRP_PORT_REDIRECT;
1101 printk(KERN_WARNING " REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
1102 target->status = -ECONNRESET;
1106 case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
1107 printk(KERN_WARNING " REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
1108 target->status = -ECONNRESET;
1111 case IB_CM_REJ_CONSUMER_DEFINED:
1112 opcode = *(u8 *) event->private_data;
1113 if (opcode == SRP_LOGIN_REJ) {
1114 struct srp_login_rej *rej = event->private_data;
1115 u32 reason = be32_to_cpu(rej->reason);
1117 if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
1118 printk(KERN_WARNING PFX
1119 "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
1121 printk(KERN_WARNING PFX
1122 "SRP LOGIN REJECTED, reason 0x%08x\n", reason);
1124 printk(KERN_WARNING " REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
1125 " opcode 0x%02x\n", opcode);
1126 target->status = -ECONNRESET;
1130 printk(KERN_WARNING " REJ reason 0x%x\n",
1131 event->param.rej_rcvd.reason);
1132 target->status = -ECONNRESET;
1136 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
1138 struct srp_target_port *target = cm_id->context;
1139 struct ib_qp_attr *qp_attr = NULL;
1144 switch (event->event) {
1145 case IB_CM_REQ_ERROR:
1146 printk(KERN_DEBUG PFX "Sending CM REQ failed\n");
1148 target->status = -ECONNRESET;
1151 case IB_CM_REP_RECEIVED:
1153 opcode = *(u8 *) event->private_data;
1155 if (opcode == SRP_LOGIN_RSP) {
1156 struct srp_login_rsp *rsp = event->private_data;
1158 target->max_ti_iu_len = be32_to_cpu(rsp->max_ti_iu_len);
1159 target->req_lim = be32_to_cpu(rsp->req_lim_delta);
1161 target->scsi_host->can_queue = min(target->req_lim,
1162 target->scsi_host->can_queue);
1164 printk(KERN_WARNING PFX "Unhandled RSP opcode %#x\n", opcode);
1165 target->status = -ECONNRESET;
1169 target->status = srp_alloc_iu_bufs(target);
1173 qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
1175 target->status = -ENOMEM;
1179 qp_attr->qp_state = IB_QPS_RTR;
1180 target->status = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1184 target->status = ib_modify_qp(target->qp, qp_attr, attr_mask);
1188 target->status = srp_post_recv(target);
1192 qp_attr->qp_state = IB_QPS_RTS;
1193 target->status = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1197 target->status = ib_modify_qp(target->qp, qp_attr, attr_mask);
1201 target->status = ib_send_cm_rtu(cm_id, NULL, 0);
1207 case IB_CM_REJ_RECEIVED:
1208 printk(KERN_DEBUG PFX "REJ received\n");
1211 srp_cm_rej_handler(cm_id, event, target);
1214 case IB_CM_DREQ_RECEIVED:
1215 printk(KERN_WARNING PFX "DREQ received - connection closed\n");
1216 if (ib_send_cm_drep(cm_id, NULL, 0))
1217 printk(KERN_ERR PFX "Sending CM DREP failed\n");
1220 case IB_CM_TIMEWAIT_EXIT:
1221 printk(KERN_ERR PFX "connection closed\n");
1227 case IB_CM_MRA_RECEIVED:
1228 case IB_CM_DREQ_ERROR:
1229 case IB_CM_DREP_RECEIVED:
1233 printk(KERN_WARNING PFX "Unhandled CM event %d\n", event->event);
1238 complete(&target->done);
1245 static int srp_send_tsk_mgmt(struct srp_target_port *target,
1246 struct srp_request *req, u8 func)
1249 struct srp_tsk_mgmt *tsk_mgmt;
1251 spin_lock_irq(target->scsi_host->host_lock);
1253 if (target->state == SRP_TARGET_DEAD ||
1254 target->state == SRP_TARGET_REMOVED) {
1255 req->scmnd->result = DID_BAD_TARGET << 16;
1259 init_completion(&req->done);
1261 iu = __srp_get_tx_iu(target);
1266 memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
1268 tsk_mgmt->opcode = SRP_TSK_MGMT;
1269 tsk_mgmt->lun = cpu_to_be64((u64) req->scmnd->device->lun << 48);
1270 tsk_mgmt->tag = req->index | SRP_TAG_TSK_MGMT;
1271 tsk_mgmt->tsk_mgmt_func = func;
1272 tsk_mgmt->task_tag = req->index;
1274 if (__srp_post_send(target, iu, sizeof *tsk_mgmt))
1279 spin_unlock_irq(target->scsi_host->host_lock);
1281 if (!wait_for_completion_timeout(&req->done,
1282 msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
1288 spin_unlock_irq(target->scsi_host->host_lock);
1292 static int srp_find_req(struct srp_target_port *target,
1293 struct scsi_cmnd *scmnd,
1294 struct srp_request **req)
1296 if (scmnd->host_scribble == (void *) -1L)
1299 *req = &target->req_ring[(long) scmnd->host_scribble];
1304 static int srp_abort(struct scsi_cmnd *scmnd)
1306 struct srp_target_port *target = host_to_target(scmnd->device->host);
1307 struct srp_request *req;
1310 printk(KERN_ERR "SRP abort called\n");
1312 if (srp_find_req(target, scmnd, &req))
1314 if (srp_send_tsk_mgmt(target, req, SRP_TSK_ABORT_TASK))
1317 spin_lock_irq(target->scsi_host->host_lock);
1319 if (req->cmd_done) {
1320 srp_remove_req(target, req);
1321 scmnd->scsi_done(scmnd);
1322 } else if (!req->tsk_status) {
1323 srp_remove_req(target, req);
1324 scmnd->result = DID_ABORT << 16;
1328 spin_unlock_irq(target->scsi_host->host_lock);
1333 static int srp_reset_device(struct scsi_cmnd *scmnd)
1335 struct srp_target_port *target = host_to_target(scmnd->device->host);
1336 struct srp_request *req, *tmp;
1338 printk(KERN_ERR "SRP reset_device called\n");
1340 if (srp_find_req(target, scmnd, &req))
1342 if (srp_send_tsk_mgmt(target, req, SRP_TSK_LUN_RESET))
1344 if (req->tsk_status)
1347 spin_lock_irq(target->scsi_host->host_lock);
1349 list_for_each_entry_safe(req, tmp, &target->req_queue, list)
1350 if (req->scmnd->device == scmnd->device)
1351 srp_reset_req(target, req);
1353 spin_unlock_irq(target->scsi_host->host_lock);
1358 static int srp_reset_host(struct scsi_cmnd *scmnd)
1360 struct srp_target_port *target = host_to_target(scmnd->device->host);
1363 printk(KERN_ERR PFX "SRP reset_host called\n");
1365 if (!srp_reconnect_target(target))
1371 static ssize_t show_id_ext(struct class_device *cdev, char *buf)
1373 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1375 if (target->state == SRP_TARGET_DEAD ||
1376 target->state == SRP_TARGET_REMOVED)
1379 return sprintf(buf, "0x%016llx\n",
1380 (unsigned long long) be64_to_cpu(target->id_ext));
1383 static ssize_t show_ioc_guid(struct class_device *cdev, char *buf)
1385 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1387 if (target->state == SRP_TARGET_DEAD ||
1388 target->state == SRP_TARGET_REMOVED)
1391 return sprintf(buf, "0x%016llx\n",
1392 (unsigned long long) be64_to_cpu(target->ioc_guid));
1395 static ssize_t show_service_id(struct class_device *cdev, char *buf)
1397 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1399 if (target->state == SRP_TARGET_DEAD ||
1400 target->state == SRP_TARGET_REMOVED)
1403 return sprintf(buf, "0x%016llx\n",
1404 (unsigned long long) be64_to_cpu(target->service_id));
1407 static ssize_t show_pkey(struct class_device *cdev, char *buf)
1409 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1411 if (target->state == SRP_TARGET_DEAD ||
1412 target->state == SRP_TARGET_REMOVED)
1415 return sprintf(buf, "0x%04x\n", be16_to_cpu(target->path.pkey));
1418 static ssize_t show_dgid(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, "%04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x\n",
1427 be16_to_cpu(((__be16 *) target->path.dgid.raw)[0]),
1428 be16_to_cpu(((__be16 *) target->path.dgid.raw)[1]),
1429 be16_to_cpu(((__be16 *) target->path.dgid.raw)[2]),
1430 be16_to_cpu(((__be16 *) target->path.dgid.raw)[3]),
1431 be16_to_cpu(((__be16 *) target->path.dgid.raw)[4]),
1432 be16_to_cpu(((__be16 *) target->path.dgid.raw)[5]),
1433 be16_to_cpu(((__be16 *) target->path.dgid.raw)[6]),
1434 be16_to_cpu(((__be16 *) target->path.dgid.raw)[7]));
1437 static ssize_t show_zero_req_lim(struct class_device *cdev, char *buf)
1439 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1441 if (target->state == SRP_TARGET_DEAD ||
1442 target->state == SRP_TARGET_REMOVED)
1445 return sprintf(buf, "%d\n", target->zero_req_lim);
1448 static CLASS_DEVICE_ATTR(id_ext, S_IRUGO, show_id_ext, NULL);
1449 static CLASS_DEVICE_ATTR(ioc_guid, S_IRUGO, show_ioc_guid, NULL);
1450 static CLASS_DEVICE_ATTR(service_id, S_IRUGO, show_service_id, NULL);
1451 static CLASS_DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1452 static CLASS_DEVICE_ATTR(dgid, S_IRUGO, show_dgid, NULL);
1453 static CLASS_DEVICE_ATTR(zero_req_lim, S_IRUGO, show_zero_req_lim, NULL);
1455 static struct class_device_attribute *srp_host_attrs[] = {
1456 &class_device_attr_id_ext,
1457 &class_device_attr_ioc_guid,
1458 &class_device_attr_service_id,
1459 &class_device_attr_pkey,
1460 &class_device_attr_dgid,
1461 &class_device_attr_zero_req_lim,
1465 static struct scsi_host_template srp_template = {
1466 .module = THIS_MODULE,
1468 .info = srp_target_info,
1469 .queuecommand = srp_queuecommand,
1470 .eh_abort_handler = srp_abort,
1471 .eh_device_reset_handler = srp_reset_device,
1472 .eh_host_reset_handler = srp_reset_host,
1473 .can_queue = SRP_SQ_SIZE,
1475 .cmd_per_lun = SRP_SQ_SIZE,
1476 .use_clustering = ENABLE_CLUSTERING,
1477 .shost_attrs = srp_host_attrs
1480 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
1482 sprintf(target->target_name, "SRP.T10:%016llX",
1483 (unsigned long long) be64_to_cpu(target->id_ext));
1485 if (scsi_add_host(target->scsi_host, host->dev->dev->dma_device))
1488 spin_lock(&host->target_lock);
1489 list_add_tail(&target->list, &host->target_list);
1490 spin_unlock(&host->target_lock);
1492 target->state = SRP_TARGET_LIVE;
1494 scsi_scan_target(&target->scsi_host->shost_gendev,
1495 0, target->scsi_id, SCAN_WILD_CARD, 0);
1500 static void srp_release_class_dev(struct class_device *class_dev)
1502 struct srp_host *host =
1503 container_of(class_dev, struct srp_host, class_dev);
1505 complete(&host->released);
1508 static struct class srp_class = {
1509 .name = "infiniband_srp",
1510 .release = srp_release_class_dev
1514 * Target ports are added by writing
1516 * id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
1517 * pkey=<P_Key>,service_id=<service ID>
1519 * to the add_target sysfs attribute.
1523 SRP_OPT_ID_EXT = 1 << 0,
1524 SRP_OPT_IOC_GUID = 1 << 1,
1525 SRP_OPT_DGID = 1 << 2,
1526 SRP_OPT_PKEY = 1 << 3,
1527 SRP_OPT_SERVICE_ID = 1 << 4,
1528 SRP_OPT_MAX_SECT = 1 << 5,
1529 SRP_OPT_MAX_CMD_PER_LUN = 1 << 6,
1530 SRP_OPT_IO_CLASS = 1 << 7,
1531 SRP_OPT_ALL = (SRP_OPT_ID_EXT |
1535 SRP_OPT_SERVICE_ID),
1538 static match_table_t srp_opt_tokens = {
1539 { SRP_OPT_ID_EXT, "id_ext=%s" },
1540 { SRP_OPT_IOC_GUID, "ioc_guid=%s" },
1541 { SRP_OPT_DGID, "dgid=%s" },
1542 { SRP_OPT_PKEY, "pkey=%x" },
1543 { SRP_OPT_SERVICE_ID, "service_id=%s" },
1544 { SRP_OPT_MAX_SECT, "max_sect=%d" },
1545 { SRP_OPT_MAX_CMD_PER_LUN, "max_cmd_per_lun=%d" },
1546 { SRP_OPT_IO_CLASS, "io_class=%x" },
1547 { SRP_OPT_ERR, NULL }
1550 static int srp_parse_options(const char *buf, struct srp_target_port *target)
1552 char *options, *sep_opt;
1555 substring_t args[MAX_OPT_ARGS];
1561 options = kstrdup(buf, GFP_KERNEL);
1566 while ((p = strsep(&sep_opt, ",")) != NULL) {
1570 token = match_token(p, srp_opt_tokens, args);
1574 case SRP_OPT_ID_EXT:
1575 p = match_strdup(args);
1576 target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
1580 case SRP_OPT_IOC_GUID:
1581 p = match_strdup(args);
1582 target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
1587 p = match_strdup(args);
1588 if (strlen(p) != 32) {
1589 printk(KERN_WARNING PFX "bad dest GID parameter '%s'\n", p);
1594 for (i = 0; i < 16; ++i) {
1595 strlcpy(dgid, p + i * 2, 3);
1596 target->path.dgid.raw[i] = simple_strtoul(dgid, NULL, 16);
1602 if (match_hex(args, &token)) {
1603 printk(KERN_WARNING PFX "bad P_Key parameter '%s'\n", p);
1606 target->path.pkey = cpu_to_be16(token);
1609 case SRP_OPT_SERVICE_ID:
1610 p = match_strdup(args);
1611 target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
1615 case SRP_OPT_MAX_SECT:
1616 if (match_int(args, &token)) {
1617 printk(KERN_WARNING PFX "bad max sect parameter '%s'\n", p);
1620 target->scsi_host->max_sectors = token;
1623 case SRP_OPT_MAX_CMD_PER_LUN:
1624 if (match_int(args, &token)) {
1625 printk(KERN_WARNING PFX "bad max cmd_per_lun parameter '%s'\n", p);
1628 target->scsi_host->cmd_per_lun = min(token, SRP_SQ_SIZE);
1631 case SRP_OPT_IO_CLASS:
1632 if (match_hex(args, &token)) {
1633 printk(KERN_WARNING PFX "bad IO class parameter '%s' \n", p);
1636 if (token != SRP_REV10_IB_IO_CLASS &&
1637 token != SRP_REV16A_IB_IO_CLASS) {
1638 printk(KERN_WARNING PFX "unknown IO class parameter value"
1639 " %x specified (use %x or %x).\n",
1640 token, SRP_REV10_IB_IO_CLASS, SRP_REV16A_IB_IO_CLASS);
1643 target->io_class = token;
1647 printk(KERN_WARNING PFX "unknown parameter or missing value "
1648 "'%s' in target creation request\n", p);
1653 if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
1656 for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
1657 if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
1658 !(srp_opt_tokens[i].token & opt_mask))
1659 printk(KERN_WARNING PFX "target creation request is "
1660 "missing parameter '%s'\n",
1661 srp_opt_tokens[i].pattern);
1668 static ssize_t srp_create_target(struct class_device *class_dev,
1669 const char *buf, size_t count)
1671 struct srp_host *host =
1672 container_of(class_dev, struct srp_host, class_dev);
1673 struct Scsi_Host *target_host;
1674 struct srp_target_port *target;
1678 target_host = scsi_host_alloc(&srp_template,
1679 sizeof (struct srp_target_port));
1683 target_host->max_lun = SRP_MAX_LUN;
1685 target = host_to_target(target_host);
1686 memset(target, 0, sizeof *target);
1688 target->io_class = SRP_REV16A_IB_IO_CLASS;
1689 target->scsi_host = target_host;
1690 target->srp_host = host;
1692 INIT_WORK(&target->work, srp_reconnect_work, target);
1694 INIT_LIST_HEAD(&target->free_reqs);
1695 INIT_LIST_HEAD(&target->req_queue);
1696 for (i = 0; i < SRP_SQ_SIZE; ++i) {
1697 target->req_ring[i].index = i;
1698 list_add_tail(&target->req_ring[i].list, &target->free_reqs);
1701 ret = srp_parse_options(buf, target);
1705 ib_get_cached_gid(host->dev->dev, host->port, 0, &target->path.sgid);
1707 printk(KERN_DEBUG PFX "new target: id_ext %016llx ioc_guid %016llx pkey %04x "
1708 "service_id %016llx dgid %04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x\n",
1709 (unsigned long long) be64_to_cpu(target->id_ext),
1710 (unsigned long long) be64_to_cpu(target->ioc_guid),
1711 be16_to_cpu(target->path.pkey),
1712 (unsigned long long) be64_to_cpu(target->service_id),
1713 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[0]),
1714 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[2]),
1715 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[4]),
1716 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[6]),
1717 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[8]),
1718 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[10]),
1719 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[12]),
1720 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[14]));
1722 ret = srp_create_target_ib(target);
1726 target->cm_id = ib_create_cm_id(host->dev->dev, srp_cm_handler, target);
1727 if (IS_ERR(target->cm_id)) {
1728 ret = PTR_ERR(target->cm_id);
1732 ret = srp_connect_target(target);
1734 printk(KERN_ERR PFX "Connection failed\n");
1738 ret = srp_add_target(host, target);
1740 goto err_disconnect;
1745 srp_disconnect_target(target);
1748 ib_destroy_cm_id(target->cm_id);
1751 srp_free_target_ib(target);
1754 scsi_host_put(target_host);
1759 static CLASS_DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
1761 static ssize_t show_ibdev(struct class_device *class_dev, char *buf)
1763 struct srp_host *host =
1764 container_of(class_dev, struct srp_host, class_dev);
1766 return sprintf(buf, "%s\n", host->dev->dev->name);
1769 static CLASS_DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
1771 static ssize_t show_port(struct class_device *class_dev, char *buf)
1773 struct srp_host *host =
1774 container_of(class_dev, struct srp_host, class_dev);
1776 return sprintf(buf, "%d\n", host->port);
1779 static CLASS_DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
1781 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
1783 struct srp_host *host;
1785 host = kzalloc(sizeof *host, GFP_KERNEL);
1789 INIT_LIST_HEAD(&host->target_list);
1790 spin_lock_init(&host->target_lock);
1791 init_completion(&host->released);
1795 host->initiator_port_id[7] = port;
1796 memcpy(host->initiator_port_id + 8, &device->dev->node_guid, 8);
1798 host->class_dev.class = &srp_class;
1799 host->class_dev.dev = device->dev->dma_device;
1800 snprintf(host->class_dev.class_id, BUS_ID_SIZE, "srp-%s-%d",
1801 device->dev->name, port);
1803 if (class_device_register(&host->class_dev))
1805 if (class_device_create_file(&host->class_dev, &class_device_attr_add_target))
1807 if (class_device_create_file(&host->class_dev, &class_device_attr_ibdev))
1809 if (class_device_create_file(&host->class_dev, &class_device_attr_port))
1815 class_device_unregister(&host->class_dev);
1823 static void srp_add_one(struct ib_device *device)
1825 struct srp_device *srp_dev;
1826 struct ib_device_attr *dev_attr;
1827 struct ib_fmr_pool_param fmr_param;
1828 struct srp_host *host;
1831 dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
1835 if (ib_query_device(device, dev_attr)) {
1836 printk(KERN_WARNING PFX "Query device failed for %s\n",
1841 srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
1846 * Use the smallest page size supported by the HCA, down to a
1847 * minimum of 512 bytes (which is the smallest sector that a
1848 * SCSI command will ever carry).
1850 srp_dev->fmr_page_shift = max(9, ffs(dev_attr->page_size_cap) - 1);
1851 srp_dev->fmr_page_size = 1 << srp_dev->fmr_page_shift;
1852 srp_dev->fmr_page_mask = ~((unsigned long) srp_dev->fmr_page_size - 1);
1854 INIT_LIST_HEAD(&srp_dev->dev_list);
1856 srp_dev->dev = device;
1857 srp_dev->pd = ib_alloc_pd(device);
1858 if (IS_ERR(srp_dev->pd))
1861 srp_dev->mr = ib_get_dma_mr(srp_dev->pd,
1862 IB_ACCESS_LOCAL_WRITE |
1863 IB_ACCESS_REMOTE_READ |
1864 IB_ACCESS_REMOTE_WRITE);
1865 if (IS_ERR(srp_dev->mr))
1868 memset(&fmr_param, 0, sizeof fmr_param);
1869 fmr_param.pool_size = SRP_FMR_POOL_SIZE;
1870 fmr_param.dirty_watermark = SRP_FMR_DIRTY_SIZE;
1871 fmr_param.cache = 1;
1872 fmr_param.max_pages_per_fmr = SRP_FMR_SIZE;
1873 fmr_param.page_shift = srp_dev->fmr_page_shift;
1874 fmr_param.access = (IB_ACCESS_LOCAL_WRITE |
1875 IB_ACCESS_REMOTE_WRITE |
1876 IB_ACCESS_REMOTE_READ);
1878 srp_dev->fmr_pool = ib_create_fmr_pool(srp_dev->pd, &fmr_param);
1879 if (IS_ERR(srp_dev->fmr_pool))
1880 srp_dev->fmr_pool = NULL;
1882 if (device->node_type == IB_NODE_SWITCH) {
1887 e = device->phys_port_cnt;
1890 for (p = s; p <= e; ++p) {
1891 host = srp_add_port(srp_dev, p);
1893 list_add_tail(&host->list, &srp_dev->dev_list);
1896 ib_set_client_data(device, &srp_client, srp_dev);
1901 ib_dealloc_pd(srp_dev->pd);
1910 static void srp_remove_one(struct ib_device *device)
1912 struct srp_device *srp_dev;
1913 struct srp_host *host, *tmp_host;
1914 LIST_HEAD(target_list);
1915 struct srp_target_port *target, *tmp_target;
1917 srp_dev = ib_get_client_data(device, &srp_client);
1919 list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
1920 class_device_unregister(&host->class_dev);
1922 * Wait for the sysfs entry to go away, so that no new
1923 * target ports can be created.
1925 wait_for_completion(&host->released);
1928 * Mark all target ports as removed, so we stop queueing
1929 * commands and don't try to reconnect.
1931 spin_lock(&host->target_lock);
1932 list_for_each_entry(target, &host->target_list, list) {
1933 spin_lock_irq(target->scsi_host->host_lock);
1934 target->state = SRP_TARGET_REMOVED;
1935 spin_unlock_irq(target->scsi_host->host_lock);
1937 spin_unlock(&host->target_lock);
1940 * Wait for any reconnection tasks that may have
1941 * started before we marked our target ports as
1942 * removed, and any target port removal tasks.
1944 flush_scheduled_work();
1946 list_for_each_entry_safe(target, tmp_target,
1947 &host->target_list, list) {
1948 scsi_remove_host(target->scsi_host);
1949 srp_disconnect_target(target);
1950 ib_destroy_cm_id(target->cm_id);
1951 srp_free_target_ib(target);
1952 scsi_host_put(target->scsi_host);
1958 if (srp_dev->fmr_pool)
1959 ib_destroy_fmr_pool(srp_dev->fmr_pool);
1960 ib_dereg_mr(srp_dev->mr);
1961 ib_dealloc_pd(srp_dev->pd);
1966 static int __init srp_init_module(void)
1970 srp_template.sg_tablesize = srp_sg_tablesize;
1971 srp_max_iu_len = (sizeof (struct srp_cmd) +
1972 sizeof (struct srp_indirect_buf) +
1973 srp_sg_tablesize * 16);
1975 ret = class_register(&srp_class);
1977 printk(KERN_ERR PFX "couldn't register class infiniband_srp\n");
1981 ret = ib_register_client(&srp_client);
1983 printk(KERN_ERR PFX "couldn't register IB client\n");
1984 class_unregister(&srp_class);
1991 static void __exit srp_cleanup_module(void)
1993 ib_unregister_client(&srp_client);
1994 class_unregister(&srp_class);
1997 module_init(srp_init_module);
1998 module_exit(srp_cleanup_module);