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 inline struct srp_target_port *host_to_target(struct Scsi_Host *host)
101 return (struct srp_target_port *) host->hostdata;
104 static const char *srp_target_info(struct Scsi_Host *host)
106 return host_to_target(host)->target_name;
109 static struct srp_iu *srp_alloc_iu(struct srp_host *host, size_t size,
111 enum dma_data_direction direction)
115 iu = kmalloc(sizeof *iu, gfp_mask);
119 iu->buf = kzalloc(size, gfp_mask);
123 iu->dma = dma_map_single(host->dev->dev->dma_device,
124 iu->buf, size, direction);
125 if (dma_mapping_error(iu->dma))
129 iu->direction = direction;
141 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
146 dma_unmap_single(host->dev->dev->dma_device,
147 iu->dma, iu->size, iu->direction);
152 static void srp_qp_event(struct ib_event *event, void *context)
154 printk(KERN_ERR PFX "QP event %d\n", event->event);
157 static int srp_init_qp(struct srp_target_port *target,
160 struct ib_qp_attr *attr;
163 attr = kmalloc(sizeof *attr, GFP_KERNEL);
167 ret = ib_find_cached_pkey(target->srp_host->dev->dev,
168 target->srp_host->port,
169 be16_to_cpu(target->path.pkey),
174 attr->qp_state = IB_QPS_INIT;
175 attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
176 IB_ACCESS_REMOTE_WRITE);
177 attr->port_num = target->srp_host->port;
179 ret = ib_modify_qp(qp, attr,
190 static int srp_create_target_ib(struct srp_target_port *target)
192 struct ib_qp_init_attr *init_attr;
195 init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
199 target->cq = ib_create_cq(target->srp_host->dev->dev, srp_completion,
200 NULL, target, SRP_CQ_SIZE);
201 if (IS_ERR(target->cq)) {
202 ret = PTR_ERR(target->cq);
206 ib_req_notify_cq(target->cq, IB_CQ_NEXT_COMP);
208 init_attr->event_handler = srp_qp_event;
209 init_attr->cap.max_send_wr = SRP_SQ_SIZE;
210 init_attr->cap.max_recv_wr = SRP_RQ_SIZE;
211 init_attr->cap.max_recv_sge = 1;
212 init_attr->cap.max_send_sge = 1;
213 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
214 init_attr->qp_type = IB_QPT_RC;
215 init_attr->send_cq = target->cq;
216 init_attr->recv_cq = target->cq;
218 target->qp = ib_create_qp(target->srp_host->dev->pd, init_attr);
219 if (IS_ERR(target->qp)) {
220 ret = PTR_ERR(target->qp);
221 ib_destroy_cq(target->cq);
225 ret = srp_init_qp(target, target->qp);
227 ib_destroy_qp(target->qp);
228 ib_destroy_cq(target->cq);
237 static void srp_free_target_ib(struct srp_target_port *target)
241 ib_destroy_qp(target->qp);
242 ib_destroy_cq(target->cq);
244 for (i = 0; i < SRP_RQ_SIZE; ++i)
245 srp_free_iu(target->srp_host, target->rx_ring[i]);
246 for (i = 0; i < SRP_SQ_SIZE + 1; ++i)
247 srp_free_iu(target->srp_host, target->tx_ring[i]);
250 static void srp_path_rec_completion(int status,
251 struct ib_sa_path_rec *pathrec,
254 struct srp_target_port *target = target_ptr;
256 target->status = status;
258 printk(KERN_ERR PFX "Got failed path rec status %d\n", status);
260 target->path = *pathrec;
261 complete(&target->done);
264 static int srp_lookup_path(struct srp_target_port *target)
266 target->path.numb_path = 1;
268 init_completion(&target->done);
270 target->path_query_id = ib_sa_path_rec_get(target->srp_host->dev->dev,
271 target->srp_host->port,
273 IB_SA_PATH_REC_DGID |
274 IB_SA_PATH_REC_SGID |
275 IB_SA_PATH_REC_NUMB_PATH |
277 SRP_PATH_REC_TIMEOUT_MS,
279 srp_path_rec_completion,
280 target, &target->path_query);
281 if (target->path_query_id < 0)
282 return target->path_query_id;
284 wait_for_completion(&target->done);
286 if (target->status < 0)
287 printk(KERN_WARNING PFX "Path record query failed\n");
289 return target->status;
292 static int srp_send_req(struct srp_target_port *target)
295 struct ib_cm_req_param param;
296 struct srp_login_req priv;
300 req = kzalloc(sizeof *req, GFP_KERNEL);
304 req->param.primary_path = &target->path;
305 req->param.alternate_path = NULL;
306 req->param.service_id = target->service_id;
307 req->param.qp_num = target->qp->qp_num;
308 req->param.qp_type = target->qp->qp_type;
309 req->param.private_data = &req->priv;
310 req->param.private_data_len = sizeof req->priv;
311 req->param.flow_control = 1;
313 get_random_bytes(&req->param.starting_psn, 4);
314 req->param.starting_psn &= 0xffffff;
317 * Pick some arbitrary defaults here; we could make these
318 * module parameters if anyone cared about setting them.
320 req->param.responder_resources = 4;
321 req->param.remote_cm_response_timeout = 20;
322 req->param.local_cm_response_timeout = 20;
323 req->param.retry_count = 7;
324 req->param.rnr_retry_count = 7;
325 req->param.max_cm_retries = 15;
327 req->priv.opcode = SRP_LOGIN_REQ;
329 req->priv.req_it_iu_len = cpu_to_be32(srp_max_iu_len);
330 req->priv.req_buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
331 SRP_BUF_FORMAT_INDIRECT);
333 * In the published SRP specification (draft rev. 16a), the
334 * port identifier format is 8 bytes of ID extension followed
335 * by 8 bytes of GUID. Older drafts put the two halves in the
336 * opposite order, so that the GUID comes first.
338 * Targets conforming to these obsolete drafts can be
339 * recognized by the I/O Class they report.
341 if (target->io_class == SRP_REV10_IB_IO_CLASS) {
342 memcpy(req->priv.initiator_port_id,
343 target->srp_host->initiator_port_id + 8, 8);
344 memcpy(req->priv.initiator_port_id + 8,
345 target->srp_host->initiator_port_id, 8);
346 memcpy(req->priv.target_port_id, &target->ioc_guid, 8);
347 memcpy(req->priv.target_port_id + 8, &target->id_ext, 8);
349 memcpy(req->priv.initiator_port_id,
350 target->srp_host->initiator_port_id, 16);
351 memcpy(req->priv.target_port_id, &target->id_ext, 8);
352 memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
356 * Topspin/Cisco SRP targets will reject our login unless we
357 * zero out the first 8 bytes of our initiator port ID. The
358 * second 8 bytes must be our local node GUID, but we always
361 if (topspin_workarounds && !memcmp(&target->ioc_guid, topspin_oui, 3)) {
362 printk(KERN_DEBUG PFX "Topspin/Cisco initiator port ID workaround "
363 "activated for target GUID %016llx\n",
364 (unsigned long long) be64_to_cpu(target->ioc_guid));
365 memset(req->priv.initiator_port_id, 0, 8);
368 status = ib_send_cm_req(target->cm_id, &req->param);
375 static void srp_disconnect_target(struct srp_target_port *target)
377 /* XXX should send SRP_I_LOGOUT request */
379 init_completion(&target->done);
380 if (ib_send_cm_dreq(target->cm_id, NULL, 0)) {
381 printk(KERN_DEBUG PFX "Sending CM DREQ failed\n");
384 wait_for_completion(&target->done);
387 static void srp_remove_work(void *target_ptr)
389 struct srp_target_port *target = target_ptr;
391 spin_lock_irq(target->scsi_host->host_lock);
392 if (target->state != SRP_TARGET_DEAD) {
393 spin_unlock_irq(target->scsi_host->host_lock);
396 target->state = SRP_TARGET_REMOVED;
397 spin_unlock_irq(target->scsi_host->host_lock);
399 spin_lock(&target->srp_host->target_lock);
400 list_del(&target->list);
401 spin_unlock(&target->srp_host->target_lock);
403 scsi_remove_host(target->scsi_host);
404 ib_destroy_cm_id(target->cm_id);
405 srp_free_target_ib(target);
406 scsi_host_put(target->scsi_host);
409 static int srp_connect_target(struct srp_target_port *target)
413 ret = srp_lookup_path(target);
418 init_completion(&target->done);
419 ret = srp_send_req(target);
422 wait_for_completion(&target->done);
425 * The CM event handling code will set status to
426 * SRP_PORT_REDIRECT if we get a port redirect REJ
427 * back, or SRP_DLID_REDIRECT if we get a lid/qp
430 switch (target->status) {
434 case SRP_PORT_REDIRECT:
435 ret = srp_lookup_path(target);
440 case SRP_DLID_REDIRECT:
444 return target->status;
449 static void srp_unmap_data(struct scsi_cmnd *scmnd,
450 struct srp_target_port *target,
451 struct srp_request *req)
453 struct scatterlist *scat;
456 if (!scmnd->request_buffer ||
457 (scmnd->sc_data_direction != DMA_TO_DEVICE &&
458 scmnd->sc_data_direction != DMA_FROM_DEVICE))
462 ib_fmr_pool_unmap(req->fmr);
467 * This handling of non-SG commands can be killed when the
468 * SCSI midlayer no longer generates non-SG commands.
470 if (likely(scmnd->use_sg)) {
471 nents = scmnd->use_sg;
472 scat = scmnd->request_buffer;
475 scat = &req->fake_sg;
478 dma_unmap_sg(target->srp_host->dev->dev->dma_device, scat, nents,
479 scmnd->sc_data_direction);
482 static void srp_remove_req(struct srp_target_port *target, struct srp_request *req)
484 srp_unmap_data(req->scmnd, target, req);
485 list_move_tail(&req->list, &target->free_reqs);
488 static void srp_reset_req(struct srp_target_port *target, struct srp_request *req)
490 req->scmnd->result = DID_RESET << 16;
491 req->scmnd->scsi_done(req->scmnd);
492 srp_remove_req(target, req);
495 static int srp_reconnect_target(struct srp_target_port *target)
497 struct ib_cm_id *new_cm_id;
498 struct ib_qp_attr qp_attr;
499 struct srp_request *req, *tmp;
503 spin_lock_irq(target->scsi_host->host_lock);
504 if (target->state != SRP_TARGET_LIVE) {
505 spin_unlock_irq(target->scsi_host->host_lock);
508 target->state = SRP_TARGET_CONNECTING;
509 spin_unlock_irq(target->scsi_host->host_lock);
511 srp_disconnect_target(target);
513 * Now get a new local CM ID so that we avoid confusing the
514 * target in case things are really fouled up.
516 new_cm_id = ib_create_cm_id(target->srp_host->dev->dev,
517 srp_cm_handler, target);
518 if (IS_ERR(new_cm_id)) {
519 ret = PTR_ERR(new_cm_id);
522 ib_destroy_cm_id(target->cm_id);
523 target->cm_id = new_cm_id;
525 qp_attr.qp_state = IB_QPS_RESET;
526 ret = ib_modify_qp(target->qp, &qp_attr, IB_QP_STATE);
530 ret = srp_init_qp(target, target->qp);
534 while (ib_poll_cq(target->cq, 1, &wc) > 0)
537 spin_lock_irq(target->scsi_host->host_lock);
538 list_for_each_entry_safe(req, tmp, &target->req_queue, list)
539 srp_reset_req(target, req);
540 spin_unlock_irq(target->scsi_host->host_lock);
546 ret = srp_connect_target(target);
550 spin_lock_irq(target->scsi_host->host_lock);
551 if (target->state == SRP_TARGET_CONNECTING) {
553 target->state = SRP_TARGET_LIVE;
556 spin_unlock_irq(target->scsi_host->host_lock);
561 printk(KERN_ERR PFX "reconnect failed (%d), removing target port.\n", ret);
564 * We couldn't reconnect, so kill our target port off.
565 * However, we have to defer the real removal because we might
566 * be in the context of the SCSI error handler now, which
567 * would deadlock if we call scsi_remove_host().
569 spin_lock_irq(target->scsi_host->host_lock);
570 if (target->state == SRP_TARGET_CONNECTING) {
571 target->state = SRP_TARGET_DEAD;
572 INIT_WORK(&target->work, srp_remove_work, target);
573 schedule_work(&target->work);
575 spin_unlock_irq(target->scsi_host->host_lock);
580 static int srp_map_fmr(struct srp_target_port *target, struct scatterlist *scat,
581 int sg_cnt, struct srp_request *req,
582 struct srp_direct_buf *buf)
590 struct srp_device *dev = target->srp_host->dev;
595 if ((sg_dma_address(&scat[0]) & ~dev->fmr_page_mask) &&
596 mellanox_workarounds && !memcmp(&target->ioc_guid, mellanox_oui, 3))
600 for (i = 0; i < sg_cnt; ++i) {
601 if (sg_dma_address(&scat[i]) & ~dev->fmr_page_mask) {
607 if ((sg_dma_address(&scat[i]) + sg_dma_len(&scat[i])) &
608 ~dev->fmr_page_mask) {
615 len += sg_dma_len(&scat[i]);
618 page_cnt += len >> dev->fmr_page_shift;
619 if (page_cnt > SRP_FMR_SIZE)
622 dma_pages = kmalloc(sizeof (u64) * page_cnt, GFP_ATOMIC);
627 for (i = 0; i < sg_cnt; ++i)
628 for (j = 0; j < sg_dma_len(&scat[i]); j += dev->fmr_page_size)
629 dma_pages[page_cnt++] =
630 (sg_dma_address(&scat[i]) & dev->fmr_page_mask) + j;
632 req->fmr = ib_fmr_pool_map_phys(dev->fmr_pool,
633 dma_pages, page_cnt, io_addr);
634 if (IS_ERR(req->fmr)) {
635 ret = PTR_ERR(req->fmr);
640 buf->va = cpu_to_be64(sg_dma_address(&scat[0]) & ~dev->fmr_page_mask);
641 buf->key = cpu_to_be32(req->fmr->fmr->rkey);
642 buf->len = cpu_to_be32(len);
652 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_target_port *target,
653 struct srp_request *req)
655 struct scatterlist *scat;
656 struct srp_cmd *cmd = req->cmd->buf;
657 int len, nents, count;
658 u8 fmt = SRP_DATA_DESC_DIRECT;
660 if (!scmnd->request_buffer || scmnd->sc_data_direction == DMA_NONE)
661 return sizeof (struct srp_cmd);
663 if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
664 scmnd->sc_data_direction != DMA_TO_DEVICE) {
665 printk(KERN_WARNING PFX "Unhandled data direction %d\n",
666 scmnd->sc_data_direction);
671 * This handling of non-SG commands can be killed when the
672 * SCSI midlayer no longer generates non-SG commands.
674 if (likely(scmnd->use_sg)) {
675 nents = scmnd->use_sg;
676 scat = scmnd->request_buffer;
679 scat = &req->fake_sg;
680 sg_init_one(scat, scmnd->request_buffer, scmnd->request_bufflen);
683 count = dma_map_sg(target->srp_host->dev->dev->dma_device,
684 scat, nents, scmnd->sc_data_direction);
686 fmt = SRP_DATA_DESC_DIRECT;
687 len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
691 * The midlayer only generated a single gather/scatter
692 * entry, or DMA mapping coalesced everything to a
693 * single entry. So a direct descriptor along with
694 * the DMA MR suffices.
696 struct srp_direct_buf *buf = (void *) cmd->add_data;
698 buf->va = cpu_to_be64(sg_dma_address(scat));
699 buf->key = cpu_to_be32(target->srp_host->dev->mr->rkey);
700 buf->len = cpu_to_be32(sg_dma_len(scat));
701 } else if (srp_map_fmr(target, scat, count, req,
702 (void *) cmd->add_data)) {
704 * FMR mapping failed, and the scatterlist has more
705 * than one entry. Generate an indirect memory
708 struct srp_indirect_buf *buf = (void *) cmd->add_data;
712 fmt = SRP_DATA_DESC_INDIRECT;
713 len = sizeof (struct srp_cmd) +
714 sizeof (struct srp_indirect_buf) +
715 count * sizeof (struct srp_direct_buf);
717 for (i = 0; i < count; ++i) {
718 buf->desc_list[i].va =
719 cpu_to_be64(sg_dma_address(&scat[i]));
720 buf->desc_list[i].key =
721 cpu_to_be32(target->srp_host->dev->mr->rkey);
722 buf->desc_list[i].len =
723 cpu_to_be32(sg_dma_len(&scat[i]));
724 datalen += sg_dma_len(&scat[i]);
727 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
728 cmd->data_out_desc_cnt = count;
730 cmd->data_in_desc_cnt = count;
733 cpu_to_be64(req->cmd->dma + sizeof *cmd + sizeof *buf);
734 buf->table_desc.key =
735 cpu_to_be32(target->srp_host->dev->mr->rkey);
736 buf->table_desc.len =
737 cpu_to_be32(count * sizeof (struct srp_direct_buf));
739 buf->len = cpu_to_be32(datalen);
742 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
743 cmd->buf_fmt = fmt << 4;
750 static void srp_process_rsp(struct srp_target_port *target, struct srp_rsp *rsp)
752 struct srp_request *req;
753 struct scsi_cmnd *scmnd;
757 delta = (s32) be32_to_cpu(rsp->req_lim_delta);
759 spin_lock_irqsave(target->scsi_host->host_lock, flags);
761 target->req_lim += delta;
763 req = &target->req_ring[rsp->tag & ~SRP_TAG_TSK_MGMT];
765 if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
766 if (be32_to_cpu(rsp->resp_data_len) < 4)
767 req->tsk_status = -1;
769 req->tsk_status = rsp->data[3];
770 complete(&req->done);
774 printk(KERN_ERR "Null scmnd for RSP w/tag %016llx\n",
775 (unsigned long long) rsp->tag);
776 scmnd->result = rsp->status;
778 if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
779 memcpy(scmnd->sense_buffer, rsp->data +
780 be32_to_cpu(rsp->resp_data_len),
781 min_t(int, be32_to_cpu(rsp->sense_data_len),
782 SCSI_SENSE_BUFFERSIZE));
785 if (rsp->flags & (SRP_RSP_FLAG_DOOVER | SRP_RSP_FLAG_DOUNDER))
786 scmnd->resid = be32_to_cpu(rsp->data_out_res_cnt);
787 else if (rsp->flags & (SRP_RSP_FLAG_DIOVER | SRP_RSP_FLAG_DIUNDER))
788 scmnd->resid = be32_to_cpu(rsp->data_in_res_cnt);
790 if (!req->tsk_mgmt) {
791 scmnd->host_scribble = (void *) -1L;
792 scmnd->scsi_done(scmnd);
794 srp_remove_req(target, req);
799 spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
802 static void srp_handle_recv(struct srp_target_port *target, struct ib_wc *wc)
807 iu = target->rx_ring[wc->wr_id & ~SRP_OP_RECV];
809 dma_sync_single_for_cpu(target->srp_host->dev->dev->dma_device, iu->dma,
810 target->max_ti_iu_len, DMA_FROM_DEVICE);
812 opcode = *(u8 *) iu->buf;
817 printk(KERN_ERR PFX "recv completion, opcode 0x%02x\n", opcode);
819 for (i = 0; i < wc->byte_len; ++i) {
821 printk(KERN_ERR " [%02x] ", i);
822 printk(" %02x", ((u8 *) iu->buf)[i]);
823 if ((i + 1) % 8 == 0)
827 if (wc->byte_len % 8)
833 srp_process_rsp(target, iu->buf);
837 /* XXX Handle target logout */
838 printk(KERN_WARNING PFX "Got target logout request\n");
842 printk(KERN_WARNING PFX "Unhandled SRP opcode 0x%02x\n", opcode);
846 dma_sync_single_for_device(target->srp_host->dev->dev->dma_device, iu->dma,
847 target->max_ti_iu_len, DMA_FROM_DEVICE);
850 static void srp_completion(struct ib_cq *cq, void *target_ptr)
852 struct srp_target_port *target = target_ptr;
855 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
856 while (ib_poll_cq(cq, 1, &wc) > 0) {
858 printk(KERN_ERR PFX "failed %s status %d\n",
859 wc.wr_id & SRP_OP_RECV ? "receive" : "send",
864 if (wc.wr_id & SRP_OP_RECV)
865 srp_handle_recv(target, &wc);
871 static int __srp_post_recv(struct srp_target_port *target)
875 struct ib_recv_wr wr, *bad_wr;
879 next = target->rx_head & (SRP_RQ_SIZE - 1);
880 wr.wr_id = next | SRP_OP_RECV;
881 iu = target->rx_ring[next];
884 list.length = iu->size;
885 list.lkey = target->srp_host->dev->mr->lkey;
891 ret = ib_post_recv(target->qp, &wr, &bad_wr);
898 static int srp_post_recv(struct srp_target_port *target)
903 spin_lock_irqsave(target->scsi_host->host_lock, flags);
904 ret = __srp_post_recv(target);
905 spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
911 * Must be called with target->scsi_host->host_lock held to protect
912 * req_lim and tx_head. Lock cannot be dropped between call here and
913 * call to __srp_post_send().
915 static struct srp_iu *__srp_get_tx_iu(struct srp_target_port *target)
917 if (target->tx_head - target->tx_tail >= SRP_SQ_SIZE)
920 if (unlikely(target->req_lim < 1))
921 ++target->zero_req_lim;
923 return target->tx_ring[target->tx_head & SRP_SQ_SIZE];
927 * Must be called with target->scsi_host->host_lock held to protect
928 * req_lim and tx_head.
930 static int __srp_post_send(struct srp_target_port *target,
931 struct srp_iu *iu, int len)
934 struct ib_send_wr wr, *bad_wr;
939 list.lkey = target->srp_host->dev->mr->lkey;
942 wr.wr_id = target->tx_head & SRP_SQ_SIZE;
945 wr.opcode = IB_WR_SEND;
946 wr.send_flags = IB_SEND_SIGNALED;
948 ret = ib_post_send(target->qp, &wr, &bad_wr);
958 static int srp_queuecommand(struct scsi_cmnd *scmnd,
959 void (*done)(struct scsi_cmnd *))
961 struct srp_target_port *target = host_to_target(scmnd->device->host);
962 struct srp_request *req;
967 if (target->state == SRP_TARGET_CONNECTING)
970 if (target->state == SRP_TARGET_DEAD ||
971 target->state == SRP_TARGET_REMOVED) {
972 scmnd->result = DID_BAD_TARGET << 16;
977 iu = __srp_get_tx_iu(target);
981 dma_sync_single_for_cpu(target->srp_host->dev->dev->dma_device, iu->dma,
982 srp_max_iu_len, DMA_TO_DEVICE);
984 req = list_entry(target->free_reqs.next, struct srp_request, list);
986 scmnd->scsi_done = done;
988 scmnd->host_scribble = (void *) (long) req->index;
991 memset(cmd, 0, sizeof *cmd);
993 cmd->opcode = SRP_CMD;
994 cmd->lun = cpu_to_be64((u64) scmnd->device->lun << 48);
995 cmd->tag = req->index;
996 memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
1001 req->tsk_mgmt = NULL;
1003 len = srp_map_data(scmnd, target, req);
1005 printk(KERN_ERR PFX "Failed to map data\n");
1009 if (__srp_post_recv(target)) {
1010 printk(KERN_ERR PFX "Recv failed\n");
1014 dma_sync_single_for_device(target->srp_host->dev->dev->dma_device, iu->dma,
1015 srp_max_iu_len, DMA_TO_DEVICE);
1017 if (__srp_post_send(target, iu, len)) {
1018 printk(KERN_ERR PFX "Send failed\n");
1022 list_move_tail(&req->list, &target->req_queue);
1027 srp_unmap_data(scmnd, target, req);
1030 return SCSI_MLQUEUE_HOST_BUSY;
1033 static int srp_alloc_iu_bufs(struct srp_target_port *target)
1037 for (i = 0; i < SRP_RQ_SIZE; ++i) {
1038 target->rx_ring[i] = srp_alloc_iu(target->srp_host,
1039 target->max_ti_iu_len,
1040 GFP_KERNEL, DMA_FROM_DEVICE);
1041 if (!target->rx_ring[i])
1045 for (i = 0; i < SRP_SQ_SIZE + 1; ++i) {
1046 target->tx_ring[i] = srp_alloc_iu(target->srp_host,
1048 GFP_KERNEL, DMA_TO_DEVICE);
1049 if (!target->tx_ring[i])
1056 for (i = 0; i < SRP_RQ_SIZE; ++i) {
1057 srp_free_iu(target->srp_host, target->rx_ring[i]);
1058 target->rx_ring[i] = NULL;
1061 for (i = 0; i < SRP_SQ_SIZE + 1; ++i) {
1062 srp_free_iu(target->srp_host, target->tx_ring[i]);
1063 target->tx_ring[i] = NULL;
1069 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
1070 struct ib_cm_event *event,
1071 struct srp_target_port *target)
1073 struct ib_class_port_info *cpi;
1076 switch (event->param.rej_rcvd.reason) {
1077 case IB_CM_REJ_PORT_CM_REDIRECT:
1078 cpi = event->param.rej_rcvd.ari;
1079 target->path.dlid = cpi->redirect_lid;
1080 target->path.pkey = cpi->redirect_pkey;
1081 cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
1082 memcpy(target->path.dgid.raw, cpi->redirect_gid, 16);
1084 target->status = target->path.dlid ?
1085 SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
1088 case IB_CM_REJ_PORT_REDIRECT:
1089 if (topspin_workarounds &&
1090 !memcmp(&target->ioc_guid, topspin_oui, 3)) {
1092 * Topspin/Cisco SRP gateways incorrectly send
1093 * reject reason code 25 when they mean 24
1096 memcpy(target->path.dgid.raw,
1097 event->param.rej_rcvd.ari, 16);
1099 printk(KERN_DEBUG PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
1100 (unsigned long long) be64_to_cpu(target->path.dgid.global.subnet_prefix),
1101 (unsigned long long) be64_to_cpu(target->path.dgid.global.interface_id));
1103 target->status = SRP_PORT_REDIRECT;
1105 printk(KERN_WARNING " REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
1106 target->status = -ECONNRESET;
1110 case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
1111 printk(KERN_WARNING " REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
1112 target->status = -ECONNRESET;
1115 case IB_CM_REJ_CONSUMER_DEFINED:
1116 opcode = *(u8 *) event->private_data;
1117 if (opcode == SRP_LOGIN_REJ) {
1118 struct srp_login_rej *rej = event->private_data;
1119 u32 reason = be32_to_cpu(rej->reason);
1121 if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
1122 printk(KERN_WARNING PFX
1123 "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
1125 printk(KERN_WARNING PFX
1126 "SRP LOGIN REJECTED, reason 0x%08x\n", reason);
1128 printk(KERN_WARNING " REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
1129 " opcode 0x%02x\n", opcode);
1130 target->status = -ECONNRESET;
1134 printk(KERN_WARNING " REJ reason 0x%x\n",
1135 event->param.rej_rcvd.reason);
1136 target->status = -ECONNRESET;
1140 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
1142 struct srp_target_port *target = cm_id->context;
1143 struct ib_qp_attr *qp_attr = NULL;
1148 switch (event->event) {
1149 case IB_CM_REQ_ERROR:
1150 printk(KERN_DEBUG PFX "Sending CM REQ failed\n");
1152 target->status = -ECONNRESET;
1155 case IB_CM_REP_RECEIVED:
1157 opcode = *(u8 *) event->private_data;
1159 if (opcode == SRP_LOGIN_RSP) {
1160 struct srp_login_rsp *rsp = event->private_data;
1162 target->max_ti_iu_len = be32_to_cpu(rsp->max_ti_iu_len);
1163 target->req_lim = be32_to_cpu(rsp->req_lim_delta);
1165 target->scsi_host->can_queue = min(target->req_lim,
1166 target->scsi_host->can_queue);
1168 printk(KERN_WARNING PFX "Unhandled RSP opcode %#x\n", opcode);
1169 target->status = -ECONNRESET;
1173 target->status = srp_alloc_iu_bufs(target);
1177 qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
1179 target->status = -ENOMEM;
1183 qp_attr->qp_state = IB_QPS_RTR;
1184 target->status = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1188 target->status = ib_modify_qp(target->qp, qp_attr, attr_mask);
1192 target->status = srp_post_recv(target);
1196 qp_attr->qp_state = IB_QPS_RTS;
1197 target->status = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1201 target->status = ib_modify_qp(target->qp, qp_attr, attr_mask);
1205 target->status = ib_send_cm_rtu(cm_id, NULL, 0);
1211 case IB_CM_REJ_RECEIVED:
1212 printk(KERN_DEBUG PFX "REJ received\n");
1215 srp_cm_rej_handler(cm_id, event, target);
1218 case IB_CM_DREQ_RECEIVED:
1219 printk(KERN_WARNING PFX "DREQ received - connection closed\n");
1220 if (ib_send_cm_drep(cm_id, NULL, 0))
1221 printk(KERN_ERR PFX "Sending CM DREP failed\n");
1224 case IB_CM_TIMEWAIT_EXIT:
1225 printk(KERN_ERR PFX "connection closed\n");
1231 case IB_CM_MRA_RECEIVED:
1232 case IB_CM_DREQ_ERROR:
1233 case IB_CM_DREP_RECEIVED:
1237 printk(KERN_WARNING PFX "Unhandled CM event %d\n", event->event);
1242 complete(&target->done);
1249 static int srp_send_tsk_mgmt(struct srp_target_port *target,
1250 struct srp_request *req, u8 func)
1253 struct srp_tsk_mgmt *tsk_mgmt;
1255 spin_lock_irq(target->scsi_host->host_lock);
1257 if (target->state == SRP_TARGET_DEAD ||
1258 target->state == SRP_TARGET_REMOVED) {
1259 req->scmnd->result = DID_BAD_TARGET << 16;
1263 init_completion(&req->done);
1265 iu = __srp_get_tx_iu(target);
1270 memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
1272 tsk_mgmt->opcode = SRP_TSK_MGMT;
1273 tsk_mgmt->lun = cpu_to_be64((u64) req->scmnd->device->lun << 48);
1274 tsk_mgmt->tag = req->index | SRP_TAG_TSK_MGMT;
1275 tsk_mgmt->tsk_mgmt_func = func;
1276 tsk_mgmt->task_tag = req->index;
1278 if (__srp_post_send(target, iu, sizeof *tsk_mgmt))
1283 spin_unlock_irq(target->scsi_host->host_lock);
1285 if (!wait_for_completion_timeout(&req->done,
1286 msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
1292 spin_unlock_irq(target->scsi_host->host_lock);
1296 static int srp_find_req(struct srp_target_port *target,
1297 struct scsi_cmnd *scmnd,
1298 struct srp_request **req)
1300 if (scmnd->host_scribble == (void *) -1L)
1303 *req = &target->req_ring[(long) scmnd->host_scribble];
1308 static int srp_abort(struct scsi_cmnd *scmnd)
1310 struct srp_target_port *target = host_to_target(scmnd->device->host);
1311 struct srp_request *req;
1314 printk(KERN_ERR "SRP abort called\n");
1316 if (srp_find_req(target, scmnd, &req))
1318 if (srp_send_tsk_mgmt(target, req, SRP_TSK_ABORT_TASK))
1321 spin_lock_irq(target->scsi_host->host_lock);
1323 if (req->cmd_done) {
1324 srp_remove_req(target, req);
1325 scmnd->scsi_done(scmnd);
1326 } else if (!req->tsk_status) {
1327 srp_remove_req(target, req);
1328 scmnd->result = DID_ABORT << 16;
1332 spin_unlock_irq(target->scsi_host->host_lock);
1337 static int srp_reset_device(struct scsi_cmnd *scmnd)
1339 struct srp_target_port *target = host_to_target(scmnd->device->host);
1340 struct srp_request *req, *tmp;
1342 printk(KERN_ERR "SRP reset_device called\n");
1344 if (srp_find_req(target, scmnd, &req))
1346 if (srp_send_tsk_mgmt(target, req, SRP_TSK_LUN_RESET))
1348 if (req->tsk_status)
1351 spin_lock_irq(target->scsi_host->host_lock);
1353 list_for_each_entry_safe(req, tmp, &target->req_queue, list)
1354 if (req->scmnd->device == scmnd->device)
1355 srp_reset_req(target, req);
1357 spin_unlock_irq(target->scsi_host->host_lock);
1362 static int srp_reset_host(struct scsi_cmnd *scmnd)
1364 struct srp_target_port *target = host_to_target(scmnd->device->host);
1367 printk(KERN_ERR PFX "SRP reset_host called\n");
1369 if (!srp_reconnect_target(target))
1375 static ssize_t show_id_ext(struct class_device *cdev, char *buf)
1377 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1379 if (target->state == SRP_TARGET_DEAD ||
1380 target->state == SRP_TARGET_REMOVED)
1383 return sprintf(buf, "0x%016llx\n",
1384 (unsigned long long) be64_to_cpu(target->id_ext));
1387 static ssize_t show_ioc_guid(struct class_device *cdev, char *buf)
1389 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1391 if (target->state == SRP_TARGET_DEAD ||
1392 target->state == SRP_TARGET_REMOVED)
1395 return sprintf(buf, "0x%016llx\n",
1396 (unsigned long long) be64_to_cpu(target->ioc_guid));
1399 static ssize_t show_service_id(struct class_device *cdev, char *buf)
1401 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1403 if (target->state == SRP_TARGET_DEAD ||
1404 target->state == SRP_TARGET_REMOVED)
1407 return sprintf(buf, "0x%016llx\n",
1408 (unsigned long long) be64_to_cpu(target->service_id));
1411 static ssize_t show_pkey(struct class_device *cdev, char *buf)
1413 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1415 if (target->state == SRP_TARGET_DEAD ||
1416 target->state == SRP_TARGET_REMOVED)
1419 return sprintf(buf, "0x%04x\n", be16_to_cpu(target->path.pkey));
1422 static ssize_t show_dgid(struct class_device *cdev, char *buf)
1424 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1426 if (target->state == SRP_TARGET_DEAD ||
1427 target->state == SRP_TARGET_REMOVED)
1430 return sprintf(buf, "%04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x\n",
1431 be16_to_cpu(((__be16 *) target->path.dgid.raw)[0]),
1432 be16_to_cpu(((__be16 *) target->path.dgid.raw)[1]),
1433 be16_to_cpu(((__be16 *) target->path.dgid.raw)[2]),
1434 be16_to_cpu(((__be16 *) target->path.dgid.raw)[3]),
1435 be16_to_cpu(((__be16 *) target->path.dgid.raw)[4]),
1436 be16_to_cpu(((__be16 *) target->path.dgid.raw)[5]),
1437 be16_to_cpu(((__be16 *) target->path.dgid.raw)[6]),
1438 be16_to_cpu(((__be16 *) target->path.dgid.raw)[7]));
1441 static ssize_t show_zero_req_lim(struct class_device *cdev, char *buf)
1443 struct srp_target_port *target = host_to_target(class_to_shost(cdev));
1445 if (target->state == SRP_TARGET_DEAD ||
1446 target->state == SRP_TARGET_REMOVED)
1449 return sprintf(buf, "%d\n", target->zero_req_lim);
1452 static CLASS_DEVICE_ATTR(id_ext, S_IRUGO, show_id_ext, NULL);
1453 static CLASS_DEVICE_ATTR(ioc_guid, S_IRUGO, show_ioc_guid, NULL);
1454 static CLASS_DEVICE_ATTR(service_id, S_IRUGO, show_service_id, NULL);
1455 static CLASS_DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1456 static CLASS_DEVICE_ATTR(dgid, S_IRUGO, show_dgid, NULL);
1457 static CLASS_DEVICE_ATTR(zero_req_lim, S_IRUGO, show_zero_req_lim, NULL);
1459 static struct class_device_attribute *srp_host_attrs[] = {
1460 &class_device_attr_id_ext,
1461 &class_device_attr_ioc_guid,
1462 &class_device_attr_service_id,
1463 &class_device_attr_pkey,
1464 &class_device_attr_dgid,
1465 &class_device_attr_zero_req_lim,
1469 static struct scsi_host_template srp_template = {
1470 .module = THIS_MODULE,
1472 .info = srp_target_info,
1473 .queuecommand = srp_queuecommand,
1474 .eh_abort_handler = srp_abort,
1475 .eh_device_reset_handler = srp_reset_device,
1476 .eh_host_reset_handler = srp_reset_host,
1477 .can_queue = SRP_SQ_SIZE,
1479 .cmd_per_lun = SRP_SQ_SIZE,
1480 .use_clustering = ENABLE_CLUSTERING,
1481 .shost_attrs = srp_host_attrs
1484 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
1486 sprintf(target->target_name, "SRP.T10:%016llX",
1487 (unsigned long long) be64_to_cpu(target->id_ext));
1489 if (scsi_add_host(target->scsi_host, host->dev->dev->dma_device))
1492 spin_lock(&host->target_lock);
1493 list_add_tail(&target->list, &host->target_list);
1494 spin_unlock(&host->target_lock);
1496 target->state = SRP_TARGET_LIVE;
1498 scsi_scan_target(&target->scsi_host->shost_gendev,
1499 0, target->scsi_id, SCAN_WILD_CARD, 0);
1504 static void srp_release_class_dev(struct class_device *class_dev)
1506 struct srp_host *host =
1507 container_of(class_dev, struct srp_host, class_dev);
1509 complete(&host->released);
1512 static struct class srp_class = {
1513 .name = "infiniband_srp",
1514 .release = srp_release_class_dev
1518 * Target ports are added by writing
1520 * id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
1521 * pkey=<P_Key>,service_id=<service ID>
1523 * to the add_target sysfs attribute.
1527 SRP_OPT_ID_EXT = 1 << 0,
1528 SRP_OPT_IOC_GUID = 1 << 1,
1529 SRP_OPT_DGID = 1 << 2,
1530 SRP_OPT_PKEY = 1 << 3,
1531 SRP_OPT_SERVICE_ID = 1 << 4,
1532 SRP_OPT_MAX_SECT = 1 << 5,
1533 SRP_OPT_MAX_CMD_PER_LUN = 1 << 6,
1534 SRP_OPT_IO_CLASS = 1 << 7,
1535 SRP_OPT_ALL = (SRP_OPT_ID_EXT |
1539 SRP_OPT_SERVICE_ID),
1542 static match_table_t srp_opt_tokens = {
1543 { SRP_OPT_ID_EXT, "id_ext=%s" },
1544 { SRP_OPT_IOC_GUID, "ioc_guid=%s" },
1545 { SRP_OPT_DGID, "dgid=%s" },
1546 { SRP_OPT_PKEY, "pkey=%x" },
1547 { SRP_OPT_SERVICE_ID, "service_id=%s" },
1548 { SRP_OPT_MAX_SECT, "max_sect=%d" },
1549 { SRP_OPT_MAX_CMD_PER_LUN, "max_cmd_per_lun=%d" },
1550 { SRP_OPT_IO_CLASS, "io_class=%x" },
1551 { SRP_OPT_ERR, NULL }
1554 static int srp_parse_options(const char *buf, struct srp_target_port *target)
1556 char *options, *sep_opt;
1559 substring_t args[MAX_OPT_ARGS];
1565 options = kstrdup(buf, GFP_KERNEL);
1570 while ((p = strsep(&sep_opt, ",")) != NULL) {
1574 token = match_token(p, srp_opt_tokens, args);
1578 case SRP_OPT_ID_EXT:
1579 p = match_strdup(args);
1580 target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
1584 case SRP_OPT_IOC_GUID:
1585 p = match_strdup(args);
1586 target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
1591 p = match_strdup(args);
1592 if (strlen(p) != 32) {
1593 printk(KERN_WARNING PFX "bad dest GID parameter '%s'\n", p);
1598 for (i = 0; i < 16; ++i) {
1599 strlcpy(dgid, p + i * 2, 3);
1600 target->path.dgid.raw[i] = simple_strtoul(dgid, NULL, 16);
1606 if (match_hex(args, &token)) {
1607 printk(KERN_WARNING PFX "bad P_Key parameter '%s'\n", p);
1610 target->path.pkey = cpu_to_be16(token);
1613 case SRP_OPT_SERVICE_ID:
1614 p = match_strdup(args);
1615 target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
1619 case SRP_OPT_MAX_SECT:
1620 if (match_int(args, &token)) {
1621 printk(KERN_WARNING PFX "bad max sect parameter '%s'\n", p);
1624 target->scsi_host->max_sectors = token;
1627 case SRP_OPT_MAX_CMD_PER_LUN:
1628 if (match_int(args, &token)) {
1629 printk(KERN_WARNING PFX "bad max cmd_per_lun parameter '%s'\n", p);
1632 target->scsi_host->cmd_per_lun = min(token, SRP_SQ_SIZE);
1635 case SRP_OPT_IO_CLASS:
1636 if (match_hex(args, &token)) {
1637 printk(KERN_WARNING PFX "bad IO class parameter '%s' \n", p);
1640 if (token != SRP_REV10_IB_IO_CLASS &&
1641 token != SRP_REV16A_IB_IO_CLASS) {
1642 printk(KERN_WARNING PFX "unknown IO class parameter value"
1643 " %x specified (use %x or %x).\n",
1644 token, SRP_REV10_IB_IO_CLASS, SRP_REV16A_IB_IO_CLASS);
1647 target->io_class = token;
1651 printk(KERN_WARNING PFX "unknown parameter or missing value "
1652 "'%s' in target creation request\n", p);
1657 if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
1660 for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
1661 if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
1662 !(srp_opt_tokens[i].token & opt_mask))
1663 printk(KERN_WARNING PFX "target creation request is "
1664 "missing parameter '%s'\n",
1665 srp_opt_tokens[i].pattern);
1672 static ssize_t srp_create_target(struct class_device *class_dev,
1673 const char *buf, size_t count)
1675 struct srp_host *host =
1676 container_of(class_dev, struct srp_host, class_dev);
1677 struct Scsi_Host *target_host;
1678 struct srp_target_port *target;
1682 target_host = scsi_host_alloc(&srp_template,
1683 sizeof (struct srp_target_port));
1687 target_host->max_lun = SRP_MAX_LUN;
1689 target = host_to_target(target_host);
1690 memset(target, 0, sizeof *target);
1692 target->io_class = SRP_REV16A_IB_IO_CLASS;
1693 target->scsi_host = target_host;
1694 target->srp_host = host;
1696 INIT_LIST_HEAD(&target->free_reqs);
1697 INIT_LIST_HEAD(&target->req_queue);
1698 for (i = 0; i < SRP_SQ_SIZE; ++i) {
1699 target->req_ring[i].index = i;
1700 list_add_tail(&target->req_ring[i].list, &target->free_reqs);
1703 ret = srp_parse_options(buf, target);
1707 ib_get_cached_gid(host->dev->dev, host->port, 0, &target->path.sgid);
1709 printk(KERN_DEBUG PFX "new target: id_ext %016llx ioc_guid %016llx pkey %04x "
1710 "service_id %016llx dgid %04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x\n",
1711 (unsigned long long) be64_to_cpu(target->id_ext),
1712 (unsigned long long) be64_to_cpu(target->ioc_guid),
1713 be16_to_cpu(target->path.pkey),
1714 (unsigned long long) be64_to_cpu(target->service_id),
1715 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[0]),
1716 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[2]),
1717 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[4]),
1718 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[6]),
1719 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[8]),
1720 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[10]),
1721 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[12]),
1722 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[14]));
1724 ret = srp_create_target_ib(target);
1728 target->cm_id = ib_create_cm_id(host->dev->dev, srp_cm_handler, target);
1729 if (IS_ERR(target->cm_id)) {
1730 ret = PTR_ERR(target->cm_id);
1734 ret = srp_connect_target(target);
1736 printk(KERN_ERR PFX "Connection failed\n");
1740 ret = srp_add_target(host, target);
1742 goto err_disconnect;
1747 srp_disconnect_target(target);
1750 ib_destroy_cm_id(target->cm_id);
1753 srp_free_target_ib(target);
1756 scsi_host_put(target_host);
1761 static CLASS_DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
1763 static ssize_t show_ibdev(struct class_device *class_dev, char *buf)
1765 struct srp_host *host =
1766 container_of(class_dev, struct srp_host, class_dev);
1768 return sprintf(buf, "%s\n", host->dev->dev->name);
1771 static CLASS_DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
1773 static ssize_t show_port(struct class_device *class_dev, char *buf)
1775 struct srp_host *host =
1776 container_of(class_dev, struct srp_host, class_dev);
1778 return sprintf(buf, "%d\n", host->port);
1781 static CLASS_DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
1783 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
1785 struct srp_host *host;
1787 host = kzalloc(sizeof *host, GFP_KERNEL);
1791 INIT_LIST_HEAD(&host->target_list);
1792 spin_lock_init(&host->target_lock);
1793 init_completion(&host->released);
1797 host->initiator_port_id[7] = port;
1798 memcpy(host->initiator_port_id + 8, &device->dev->node_guid, 8);
1800 host->class_dev.class = &srp_class;
1801 host->class_dev.dev = device->dev->dma_device;
1802 snprintf(host->class_dev.class_id, BUS_ID_SIZE, "srp-%s-%d",
1803 device->dev->name, port);
1805 if (class_device_register(&host->class_dev))
1807 if (class_device_create_file(&host->class_dev, &class_device_attr_add_target))
1809 if (class_device_create_file(&host->class_dev, &class_device_attr_ibdev))
1811 if (class_device_create_file(&host->class_dev, &class_device_attr_port))
1817 class_device_unregister(&host->class_dev);
1825 static void srp_add_one(struct ib_device *device)
1827 struct srp_device *srp_dev;
1828 struct ib_device_attr *dev_attr;
1829 struct ib_fmr_pool_param fmr_param;
1830 struct srp_host *host;
1833 dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
1837 if (ib_query_device(device, dev_attr)) {
1838 printk(KERN_WARNING PFX "Query device failed for %s\n",
1843 srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
1848 * Use the smallest page size supported by the HCA, down to a
1849 * minimum of 512 bytes (which is the smallest sector that a
1850 * SCSI command will ever carry).
1852 srp_dev->fmr_page_shift = max(9, ffs(dev_attr->page_size_cap) - 1);
1853 srp_dev->fmr_page_size = 1 << srp_dev->fmr_page_shift;
1854 srp_dev->fmr_page_mask = ~((unsigned long) srp_dev->fmr_page_size - 1);
1856 INIT_LIST_HEAD(&srp_dev->dev_list);
1858 srp_dev->dev = device;
1859 srp_dev->pd = ib_alloc_pd(device);
1860 if (IS_ERR(srp_dev->pd))
1863 srp_dev->mr = ib_get_dma_mr(srp_dev->pd,
1864 IB_ACCESS_LOCAL_WRITE |
1865 IB_ACCESS_REMOTE_READ |
1866 IB_ACCESS_REMOTE_WRITE);
1867 if (IS_ERR(srp_dev->mr))
1870 memset(&fmr_param, 0, sizeof fmr_param);
1871 fmr_param.pool_size = SRP_FMR_POOL_SIZE;
1872 fmr_param.dirty_watermark = SRP_FMR_DIRTY_SIZE;
1873 fmr_param.cache = 1;
1874 fmr_param.max_pages_per_fmr = SRP_FMR_SIZE;
1875 fmr_param.page_shift = srp_dev->fmr_page_shift;
1876 fmr_param.access = (IB_ACCESS_LOCAL_WRITE |
1877 IB_ACCESS_REMOTE_WRITE |
1878 IB_ACCESS_REMOTE_READ);
1880 srp_dev->fmr_pool = ib_create_fmr_pool(srp_dev->pd, &fmr_param);
1881 if (IS_ERR(srp_dev->fmr_pool))
1882 srp_dev->fmr_pool = NULL;
1884 if (device->node_type == IB_NODE_SWITCH) {
1889 e = device->phys_port_cnt;
1892 for (p = s; p <= e; ++p) {
1893 host = srp_add_port(srp_dev, p);
1895 list_add_tail(&host->list, &srp_dev->dev_list);
1898 ib_set_client_data(device, &srp_client, srp_dev);
1903 ib_dealloc_pd(srp_dev->pd);
1912 static void srp_remove_one(struct ib_device *device)
1914 struct srp_device *srp_dev;
1915 struct srp_host *host, *tmp_host;
1916 LIST_HEAD(target_list);
1917 struct srp_target_port *target, *tmp_target;
1919 srp_dev = ib_get_client_data(device, &srp_client);
1921 list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
1922 class_device_unregister(&host->class_dev);
1924 * Wait for the sysfs entry to go away, so that no new
1925 * target ports can be created.
1927 wait_for_completion(&host->released);
1930 * Mark all target ports as removed, so we stop queueing
1931 * commands and don't try to reconnect.
1933 spin_lock(&host->target_lock);
1934 list_for_each_entry(target, &host->target_list, list) {
1935 spin_lock_irq(target->scsi_host->host_lock);
1936 target->state = SRP_TARGET_REMOVED;
1937 spin_unlock_irq(target->scsi_host->host_lock);
1939 spin_unlock(&host->target_lock);
1942 * Wait for any reconnection tasks that may have
1943 * started before we marked our target ports as
1944 * removed, and any target port removal tasks.
1946 flush_scheduled_work();
1948 list_for_each_entry_safe(target, tmp_target,
1949 &host->target_list, list) {
1950 scsi_remove_host(target->scsi_host);
1951 srp_disconnect_target(target);
1952 ib_destroy_cm_id(target->cm_id);
1953 srp_free_target_ib(target);
1954 scsi_host_put(target->scsi_host);
1960 if (srp_dev->fmr_pool)
1961 ib_destroy_fmr_pool(srp_dev->fmr_pool);
1962 ib_dereg_mr(srp_dev->mr);
1963 ib_dealloc_pd(srp_dev->pd);
1968 static int __init srp_init_module(void)
1972 srp_template.sg_tablesize = srp_sg_tablesize;
1973 srp_max_iu_len = (sizeof (struct srp_cmd) +
1974 sizeof (struct srp_indirect_buf) +
1975 srp_sg_tablesize * 16);
1977 ret = class_register(&srp_class);
1979 printk(KERN_ERR PFX "couldn't register class infiniband_srp\n");
1983 ret = ib_register_client(&srp_client);
1985 printk(KERN_ERR PFX "couldn't register IB client\n");
1986 class_unregister(&srp_class);
1993 static void __exit srp_cleanup_module(void)
1995 ib_unregister_client(&srp_client);
1996 class_unregister(&srp_class);
1999 module_init(srp_init_module);
2000 module_exit(srp_cleanup_module);