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->dma_device, iu->buf, size, direction);
109 if (dma_mapping_error(iu->dma))
113 iu->direction = direction;
125 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
130 dma_unmap_single(host->dev->dma_device, iu->dma, iu->size, iu->direction);
135 static void srp_qp_event(struct ib_event *event, void *context)
137 printk(KERN_ERR PFX "QP event %d\n", event->event);
140 static int srp_init_qp(struct srp_target_port *target,
143 struct ib_qp_attr *attr;
146 attr = kmalloc(sizeof *attr, GFP_KERNEL);
150 ret = ib_find_cached_pkey(target->srp_host->dev,
151 target->srp_host->port,
152 be16_to_cpu(target->path.pkey),
157 attr->qp_state = IB_QPS_INIT;
158 attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
159 IB_ACCESS_REMOTE_WRITE);
160 attr->port_num = target->srp_host->port;
162 ret = ib_modify_qp(qp, attr,
173 static int srp_create_target_ib(struct srp_target_port *target)
175 struct ib_qp_init_attr *init_attr;
178 init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
182 target->cq = ib_create_cq(target->srp_host->dev, srp_completion,
183 NULL, target, SRP_CQ_SIZE);
184 if (IS_ERR(target->cq)) {
185 ret = PTR_ERR(target->cq);
189 ib_req_notify_cq(target->cq, IB_CQ_NEXT_COMP);
191 init_attr->event_handler = srp_qp_event;
192 init_attr->cap.max_send_wr = SRP_SQ_SIZE;
193 init_attr->cap.max_recv_wr = SRP_RQ_SIZE;
194 init_attr->cap.max_recv_sge = 1;
195 init_attr->cap.max_send_sge = 1;
196 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
197 init_attr->qp_type = IB_QPT_RC;
198 init_attr->send_cq = target->cq;
199 init_attr->recv_cq = target->cq;
201 target->qp = ib_create_qp(target->srp_host->pd, init_attr);
202 if (IS_ERR(target->qp)) {
203 ret = PTR_ERR(target->qp);
204 ib_destroy_cq(target->cq);
208 ret = srp_init_qp(target, target->qp);
210 ib_destroy_qp(target->qp);
211 ib_destroy_cq(target->cq);
220 static void srp_free_target_ib(struct srp_target_port *target)
224 ib_destroy_qp(target->qp);
225 ib_destroy_cq(target->cq);
227 for (i = 0; i < SRP_RQ_SIZE; ++i)
228 srp_free_iu(target->srp_host, target->rx_ring[i]);
229 for (i = 0; i < SRP_SQ_SIZE + 1; ++i)
230 srp_free_iu(target->srp_host, target->tx_ring[i]);
233 static void srp_path_rec_completion(int status,
234 struct ib_sa_path_rec *pathrec,
237 struct srp_target_port *target = target_ptr;
239 target->status = status;
241 printk(KERN_ERR PFX "Got failed path rec status %d\n", status);
243 target->path = *pathrec;
244 complete(&target->done);
247 static int srp_lookup_path(struct srp_target_port *target)
249 target->path.numb_path = 1;
251 init_completion(&target->done);
253 target->path_query_id = ib_sa_path_rec_get(target->srp_host->dev,
254 target->srp_host->port,
256 IB_SA_PATH_REC_DGID |
257 IB_SA_PATH_REC_SGID |
258 IB_SA_PATH_REC_NUMB_PATH |
260 SRP_PATH_REC_TIMEOUT_MS,
262 srp_path_rec_completion,
263 target, &target->path_query);
264 if (target->path_query_id < 0)
265 return target->path_query_id;
267 wait_for_completion(&target->done);
269 if (target->status < 0)
270 printk(KERN_WARNING PFX "Path record query failed\n");
272 return target->status;
275 static int srp_send_req(struct srp_target_port *target)
278 struct ib_cm_req_param param;
279 struct srp_login_req priv;
283 req = kzalloc(sizeof *req, GFP_KERNEL);
287 req->param.primary_path = &target->path;
288 req->param.alternate_path = NULL;
289 req->param.service_id = target->service_id;
290 req->param.qp_num = target->qp->qp_num;
291 req->param.qp_type = target->qp->qp_type;
292 req->param.private_data = &req->priv;
293 req->param.private_data_len = sizeof req->priv;
294 req->param.flow_control = 1;
296 get_random_bytes(&req->param.starting_psn, 4);
297 req->param.starting_psn &= 0xffffff;
300 * Pick some arbitrary defaults here; we could make these
301 * module parameters if anyone cared about setting them.
303 req->param.responder_resources = 4;
304 req->param.remote_cm_response_timeout = 20;
305 req->param.local_cm_response_timeout = 20;
306 req->param.retry_count = 7;
307 req->param.rnr_retry_count = 7;
308 req->param.max_cm_retries = 15;
310 req->priv.opcode = SRP_LOGIN_REQ;
312 req->priv.req_it_iu_len = cpu_to_be32(SRP_MAX_IU_LEN);
313 req->priv.req_buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
314 SRP_BUF_FORMAT_INDIRECT);
315 memcpy(req->priv.initiator_port_id, target->srp_host->initiator_port_id, 16);
317 * Topspin/Cisco SRP targets will reject our login unless we
318 * zero out the first 8 bytes of our initiator port ID. The
319 * second 8 bytes must be our local node GUID, but we always
322 if (topspin_workarounds && !memcmp(&target->ioc_guid, topspin_oui, 3)) {
323 printk(KERN_DEBUG PFX "Topspin/Cisco initiator port ID workaround "
324 "activated for target GUID %016llx\n",
325 (unsigned long long) be64_to_cpu(target->ioc_guid));
326 memset(req->priv.initiator_port_id, 0, 8);
328 memcpy(req->priv.target_port_id, &target->id_ext, 8);
329 memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
331 status = ib_send_cm_req(target->cm_id, &req->param);
338 static void srp_disconnect_target(struct srp_target_port *target)
340 /* XXX should send SRP_I_LOGOUT request */
342 init_completion(&target->done);
343 ib_send_cm_dreq(target->cm_id, NULL, 0);
344 wait_for_completion(&target->done);
347 static void srp_remove_work(void *target_ptr)
349 struct srp_target_port *target = target_ptr;
351 spin_lock_irq(target->scsi_host->host_lock);
352 if (target->state != SRP_TARGET_DEAD) {
353 spin_unlock_irq(target->scsi_host->host_lock);
354 scsi_host_put(target->scsi_host);
357 target->state = SRP_TARGET_REMOVED;
358 spin_unlock_irq(target->scsi_host->host_lock);
360 mutex_lock(&target->srp_host->target_mutex);
361 list_del(&target->list);
362 mutex_unlock(&target->srp_host->target_mutex);
364 scsi_remove_host(target->scsi_host);
365 ib_destroy_cm_id(target->cm_id);
366 srp_free_target_ib(target);
367 scsi_host_put(target->scsi_host);
368 /* And another put to really free the target port... */
369 scsi_host_put(target->scsi_host);
372 static int srp_connect_target(struct srp_target_port *target)
376 ret = srp_lookup_path(target);
381 init_completion(&target->done);
382 ret = srp_send_req(target);
385 wait_for_completion(&target->done);
388 * The CM event handling code will set status to
389 * SRP_PORT_REDIRECT if we get a port redirect REJ
390 * back, or SRP_DLID_REDIRECT if we get a lid/qp
393 switch (target->status) {
397 case SRP_PORT_REDIRECT:
398 ret = srp_lookup_path(target);
403 case SRP_DLID_REDIRECT:
407 return target->status;
412 static int srp_reconnect_target(struct srp_target_port *target)
414 struct ib_cm_id *new_cm_id;
415 struct ib_qp_attr qp_attr;
416 struct srp_request *req;
421 spin_lock_irq(target->scsi_host->host_lock);
422 if (target->state != SRP_TARGET_LIVE) {
423 spin_unlock_irq(target->scsi_host->host_lock);
426 target->state = SRP_TARGET_CONNECTING;
427 spin_unlock_irq(target->scsi_host->host_lock);
429 srp_disconnect_target(target);
431 * Now get a new local CM ID so that we avoid confusing the
432 * target in case things are really fouled up.
434 new_cm_id = ib_create_cm_id(target->srp_host->dev,
435 srp_cm_handler, target);
436 if (IS_ERR(new_cm_id)) {
437 ret = PTR_ERR(new_cm_id);
440 ib_destroy_cm_id(target->cm_id);
441 target->cm_id = new_cm_id;
443 qp_attr.qp_state = IB_QPS_RESET;
444 ret = ib_modify_qp(target->qp, &qp_attr, IB_QP_STATE);
448 ret = srp_init_qp(target, target->qp);
452 while (ib_poll_cq(target->cq, 1, &wc) > 0)
455 list_for_each_entry(req, &target->req_queue, list) {
456 req->scmnd->result = DID_RESET << 16;
457 req->scmnd->scsi_done(req->scmnd);
463 target->req_head = 0;
464 for (i = 0; i < SRP_SQ_SIZE - 1; ++i)
465 target->req_ring[i].next = i + 1;
466 target->req_ring[SRP_SQ_SIZE - 1].next = -1;
467 INIT_LIST_HEAD(&target->req_queue);
469 ret = srp_connect_target(target);
473 spin_lock_irq(target->scsi_host->host_lock);
474 if (target->state == SRP_TARGET_CONNECTING) {
476 target->state = SRP_TARGET_LIVE;
479 spin_unlock_irq(target->scsi_host->host_lock);
484 printk(KERN_ERR PFX "reconnect failed (%d), removing target port.\n", ret);
487 * We couldn't reconnect, so kill our target port off.
488 * However, we have to defer the real removal because we might
489 * be in the context of the SCSI error handler now, which
490 * would deadlock if we call scsi_remove_host().
492 spin_lock_irq(target->scsi_host->host_lock);
493 if (target->state == SRP_TARGET_CONNECTING) {
494 target->state = SRP_TARGET_DEAD;
495 INIT_WORK(&target->work, srp_remove_work, target);
496 schedule_work(&target->work);
498 spin_unlock_irq(target->scsi_host->host_lock);
503 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_target_port *target,
504 struct srp_request *req)
506 struct srp_cmd *cmd = req->cmd->buf;
510 if (!scmnd->request_buffer || scmnd->sc_data_direction == DMA_NONE)
511 return sizeof (struct srp_cmd);
513 if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
514 scmnd->sc_data_direction != DMA_TO_DEVICE) {
515 printk(KERN_WARNING PFX "Unhandled data direction %d\n",
516 scmnd->sc_data_direction);
521 struct scatterlist *scat = scmnd->request_buffer;
525 n = dma_map_sg(target->srp_host->dev->dma_device,
526 scat, scmnd->use_sg, scmnd->sc_data_direction);
529 struct srp_direct_buf *buf = (void *) cmd->add_data;
531 fmt = SRP_DATA_DESC_DIRECT;
533 buf->va = cpu_to_be64(sg_dma_address(scat));
534 buf->key = cpu_to_be32(target->srp_host->mr->rkey);
535 buf->len = cpu_to_be32(sg_dma_len(scat));
537 len = sizeof (struct srp_cmd) +
538 sizeof (struct srp_direct_buf);
540 struct srp_indirect_buf *buf = (void *) cmd->add_data;
543 fmt = SRP_DATA_DESC_INDIRECT;
545 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
546 cmd->data_out_desc_cnt = n;
548 cmd->data_in_desc_cnt = n;
550 buf->table_desc.va = cpu_to_be64(req->cmd->dma +
553 buf->table_desc.key =
554 cpu_to_be32(target->srp_host->mr->rkey);
555 buf->table_desc.len =
556 cpu_to_be32(n * sizeof (struct srp_direct_buf));
558 for (i = 0; i < n; ++i) {
559 buf->desc_list[i].va = cpu_to_be64(sg_dma_address(&scat[i]));
560 buf->desc_list[i].key =
561 cpu_to_be32(target->srp_host->mr->rkey);
562 buf->desc_list[i].len = cpu_to_be32(sg_dma_len(&scat[i]));
564 datalen += sg_dma_len(&scat[i]);
567 buf->len = cpu_to_be32(datalen);
569 len = sizeof (struct srp_cmd) +
570 sizeof (struct srp_indirect_buf) +
571 n * sizeof (struct srp_direct_buf);
574 struct srp_direct_buf *buf = (void *) cmd->add_data;
577 dma = dma_map_single(target->srp_host->dev->dma_device,
578 scmnd->request_buffer, scmnd->request_bufflen,
579 scmnd->sc_data_direction);
580 if (dma_mapping_error(dma)) {
581 printk(KERN_WARNING PFX "unable to map %p/%d (dir %d)\n",
582 scmnd->request_buffer, (int) scmnd->request_bufflen,
583 scmnd->sc_data_direction);
587 pci_unmap_addr_set(req, direct_mapping, dma);
589 buf->va = cpu_to_be64(dma);
590 buf->key = cpu_to_be32(target->srp_host->mr->rkey);
591 buf->len = cpu_to_be32(scmnd->request_bufflen);
593 fmt = SRP_DATA_DESC_DIRECT;
595 len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
598 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
599 cmd->buf_fmt = fmt << 4;
607 static void srp_unmap_data(struct scsi_cmnd *scmnd,
608 struct srp_target_port *target,
609 struct srp_request *req)
611 if (!scmnd->request_buffer ||
612 (scmnd->sc_data_direction != DMA_TO_DEVICE &&
613 scmnd->sc_data_direction != DMA_FROM_DEVICE))
617 dma_unmap_sg(target->srp_host->dev->dma_device,
618 (struct scatterlist *) scmnd->request_buffer,
619 scmnd->use_sg, scmnd->sc_data_direction);
621 dma_unmap_single(target->srp_host->dev->dma_device,
622 pci_unmap_addr(req, direct_mapping),
623 scmnd->request_bufflen,
624 scmnd->sc_data_direction);
627 static void srp_process_rsp(struct srp_target_port *target, struct srp_rsp *rsp)
629 struct srp_request *req;
630 struct scsi_cmnd *scmnd;
634 delta = (s32) be32_to_cpu(rsp->req_lim_delta);
636 spin_lock_irqsave(target->scsi_host->host_lock, flags);
638 target->req_lim += delta;
640 req = &target->req_ring[rsp->tag & ~SRP_TAG_TSK_MGMT];
642 if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
643 if (be32_to_cpu(rsp->resp_data_len) < 4)
644 req->tsk_status = -1;
646 req->tsk_status = rsp->data[3];
647 complete(&req->done);
651 printk(KERN_ERR "Null scmnd for RSP w/tag %016llx\n",
652 (unsigned long long) rsp->tag);
653 scmnd->result = rsp->status;
655 if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
656 memcpy(scmnd->sense_buffer, rsp->data +
657 be32_to_cpu(rsp->resp_data_len),
658 min_t(int, be32_to_cpu(rsp->sense_data_len),
659 SCSI_SENSE_BUFFERSIZE));
662 if (rsp->flags & (SRP_RSP_FLAG_DOOVER | SRP_RSP_FLAG_DOUNDER))
663 scmnd->resid = be32_to_cpu(rsp->data_out_res_cnt);
664 else if (rsp->flags & (SRP_RSP_FLAG_DIOVER | SRP_RSP_FLAG_DIUNDER))
665 scmnd->resid = be32_to_cpu(rsp->data_in_res_cnt);
667 srp_unmap_data(scmnd, target, req);
669 if (!req->tsk_mgmt) {
671 scmnd->host_scribble = (void *) -1L;
672 scmnd->scsi_done(scmnd);
674 list_del(&req->list);
675 req->next = target->req_head;
676 target->req_head = rsp->tag & ~SRP_TAG_TSK_MGMT;
681 spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
684 static void srp_reconnect_work(void *target_ptr)
686 struct srp_target_port *target = target_ptr;
688 srp_reconnect_target(target);
691 static void srp_handle_recv(struct srp_target_port *target, struct ib_wc *wc)
696 iu = target->rx_ring[wc->wr_id & ~SRP_OP_RECV];
698 dma_sync_single_for_cpu(target->srp_host->dev->dma_device, iu->dma,
699 target->max_ti_iu_len, DMA_FROM_DEVICE);
701 opcode = *(u8 *) iu->buf;
706 printk(KERN_ERR PFX "recv completion, opcode 0x%02x\n", opcode);
708 for (i = 0; i < wc->byte_len; ++i) {
710 printk(KERN_ERR " [%02x] ", i);
711 printk(" %02x", ((u8 *) iu->buf)[i]);
712 if ((i + 1) % 8 == 0)
716 if (wc->byte_len % 8)
722 srp_process_rsp(target, iu->buf);
726 /* XXX Handle target logout */
727 printk(KERN_WARNING PFX "Got target logout request\n");
731 printk(KERN_WARNING PFX "Unhandled SRP opcode 0x%02x\n", opcode);
735 dma_sync_single_for_device(target->srp_host->dev->dma_device, iu->dma,
736 target->max_ti_iu_len, DMA_FROM_DEVICE);
739 static void srp_completion(struct ib_cq *cq, void *target_ptr)
741 struct srp_target_port *target = target_ptr;
745 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
746 while (ib_poll_cq(cq, 1, &wc) > 0) {
748 printk(KERN_ERR PFX "failed %s status %d\n",
749 wc.wr_id & SRP_OP_RECV ? "receive" : "send",
751 spin_lock_irqsave(target->scsi_host->host_lock, flags);
752 if (target->state == SRP_TARGET_LIVE)
753 schedule_work(&target->work);
754 spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
758 if (wc.wr_id & SRP_OP_RECV)
759 srp_handle_recv(target, &wc);
765 static int __srp_post_recv(struct srp_target_port *target)
769 struct ib_recv_wr wr, *bad_wr;
773 next = target->rx_head & (SRP_RQ_SIZE - 1);
774 wr.wr_id = next | SRP_OP_RECV;
775 iu = target->rx_ring[next];
778 list.length = iu->size;
779 list.lkey = target->srp_host->mr->lkey;
785 ret = ib_post_recv(target->qp, &wr, &bad_wr);
792 static int srp_post_recv(struct srp_target_port *target)
797 spin_lock_irqsave(target->scsi_host->host_lock, flags);
798 ret = __srp_post_recv(target);
799 spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
805 * Must be called with target->scsi_host->host_lock held to protect
806 * req_lim and tx_head. Lock cannot be dropped between call here and
807 * call to __srp_post_send().
809 static struct srp_iu *__srp_get_tx_iu(struct srp_target_port *target)
811 if (target->tx_head - target->tx_tail >= SRP_SQ_SIZE)
814 if (unlikely(target->req_lim < 1)) {
815 if (printk_ratelimit())
816 printk(KERN_DEBUG PFX "Target has req_lim %d\n",
821 return target->tx_ring[target->tx_head & SRP_SQ_SIZE];
825 * Must be called with target->scsi_host->host_lock held to protect
826 * req_lim and tx_head.
828 static int __srp_post_send(struct srp_target_port *target,
829 struct srp_iu *iu, int len)
832 struct ib_send_wr wr, *bad_wr;
837 list.lkey = target->srp_host->mr->lkey;
840 wr.wr_id = target->tx_head & SRP_SQ_SIZE;
843 wr.opcode = IB_WR_SEND;
844 wr.send_flags = IB_SEND_SIGNALED;
846 ret = ib_post_send(target->qp, &wr, &bad_wr);
856 static int srp_queuecommand(struct scsi_cmnd *scmnd,
857 void (*done)(struct scsi_cmnd *))
859 struct srp_target_port *target = host_to_target(scmnd->device->host);
860 struct srp_request *req;
866 if (target->state == SRP_TARGET_CONNECTING)
869 if (target->state == SRP_TARGET_DEAD ||
870 target->state == SRP_TARGET_REMOVED) {
871 scmnd->result = DID_BAD_TARGET << 16;
876 iu = __srp_get_tx_iu(target);
880 dma_sync_single_for_cpu(target->srp_host->dev->dma_device, iu->dma,
881 SRP_MAX_IU_LEN, DMA_TO_DEVICE);
883 req_index = target->req_head;
885 scmnd->scsi_done = done;
887 scmnd->host_scribble = (void *) req_index;
890 memset(cmd, 0, sizeof *cmd);
892 cmd->opcode = SRP_CMD;
893 cmd->lun = cpu_to_be64((u64) scmnd->device->lun << 48);
894 cmd->tag = req_index;
895 memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
897 req = &target->req_ring[req_index];
902 req->tsk_mgmt = NULL;
904 len = srp_map_data(scmnd, target, req);
906 printk(KERN_ERR PFX "Failed to map data\n");
910 if (__srp_post_recv(target)) {
911 printk(KERN_ERR PFX "Recv failed\n");
915 dma_sync_single_for_device(target->srp_host->dev->dma_device, iu->dma,
916 SRP_MAX_IU_LEN, DMA_TO_DEVICE);
918 if (__srp_post_send(target, iu, len)) {
919 printk(KERN_ERR PFX "Send failed\n");
923 target->req_head = req->next;
924 list_add_tail(&req->list, &target->req_queue);
929 srp_unmap_data(scmnd, target, req);
932 return SCSI_MLQUEUE_HOST_BUSY;
935 static int srp_alloc_iu_bufs(struct srp_target_port *target)
939 for (i = 0; i < SRP_RQ_SIZE; ++i) {
940 target->rx_ring[i] = srp_alloc_iu(target->srp_host,
941 target->max_ti_iu_len,
942 GFP_KERNEL, DMA_FROM_DEVICE);
943 if (!target->rx_ring[i])
947 for (i = 0; i < SRP_SQ_SIZE + 1; ++i) {
948 target->tx_ring[i] = srp_alloc_iu(target->srp_host,
950 GFP_KERNEL, DMA_TO_DEVICE);
951 if (!target->tx_ring[i])
958 for (i = 0; i < SRP_RQ_SIZE; ++i) {
959 srp_free_iu(target->srp_host, target->rx_ring[i]);
960 target->rx_ring[i] = NULL;
963 for (i = 0; i < SRP_SQ_SIZE + 1; ++i) {
964 srp_free_iu(target->srp_host, target->tx_ring[i]);
965 target->tx_ring[i] = NULL;
971 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
972 struct ib_cm_event *event,
973 struct srp_target_port *target)
975 struct ib_class_port_info *cpi;
978 switch (event->param.rej_rcvd.reason) {
979 case IB_CM_REJ_PORT_CM_REDIRECT:
980 cpi = event->param.rej_rcvd.ari;
981 target->path.dlid = cpi->redirect_lid;
982 target->path.pkey = cpi->redirect_pkey;
983 cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
984 memcpy(target->path.dgid.raw, cpi->redirect_gid, 16);
986 target->status = target->path.dlid ?
987 SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
990 case IB_CM_REJ_PORT_REDIRECT:
991 if (topspin_workarounds &&
992 !memcmp(&target->ioc_guid, topspin_oui, 3)) {
994 * Topspin/Cisco SRP gateways incorrectly send
995 * reject reason code 25 when they mean 24
998 memcpy(target->path.dgid.raw,
999 event->param.rej_rcvd.ari, 16);
1001 printk(KERN_DEBUG PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
1002 (unsigned long long) be64_to_cpu(target->path.dgid.global.subnet_prefix),
1003 (unsigned long long) be64_to_cpu(target->path.dgid.global.interface_id));
1005 target->status = SRP_PORT_REDIRECT;
1007 printk(KERN_WARNING " REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
1008 target->status = -ECONNRESET;
1012 case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
1013 printk(KERN_WARNING " REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
1014 target->status = -ECONNRESET;
1017 case IB_CM_REJ_CONSUMER_DEFINED:
1018 opcode = *(u8 *) event->private_data;
1019 if (opcode == SRP_LOGIN_REJ) {
1020 struct srp_login_rej *rej = event->private_data;
1021 u32 reason = be32_to_cpu(rej->reason);
1023 if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
1024 printk(KERN_WARNING PFX
1025 "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
1027 printk(KERN_WARNING PFX
1028 "SRP LOGIN REJECTED, reason 0x%08x\n", reason);
1030 printk(KERN_WARNING " REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
1031 " opcode 0x%02x\n", opcode);
1032 target->status = -ECONNRESET;
1036 printk(KERN_WARNING " REJ reason 0x%x\n",
1037 event->param.rej_rcvd.reason);
1038 target->status = -ECONNRESET;
1042 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
1044 struct srp_target_port *target = cm_id->context;
1045 struct ib_qp_attr *qp_attr = NULL;
1050 switch (event->event) {
1051 case IB_CM_REQ_ERROR:
1052 printk(KERN_DEBUG PFX "Sending CM REQ failed\n");
1054 target->status = -ECONNRESET;
1057 case IB_CM_REP_RECEIVED:
1059 opcode = *(u8 *) event->private_data;
1061 if (opcode == SRP_LOGIN_RSP) {
1062 struct srp_login_rsp *rsp = event->private_data;
1064 target->max_ti_iu_len = be32_to_cpu(rsp->max_ti_iu_len);
1065 target->req_lim = be32_to_cpu(rsp->req_lim_delta);
1067 target->scsi_host->can_queue = min(target->req_lim,
1068 target->scsi_host->can_queue);
1070 printk(KERN_WARNING PFX "Unhandled RSP opcode %#x\n", opcode);
1071 target->status = -ECONNRESET;
1075 target->status = srp_alloc_iu_bufs(target);
1079 qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
1081 target->status = -ENOMEM;
1085 qp_attr->qp_state = IB_QPS_RTR;
1086 target->status = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1090 target->status = ib_modify_qp(target->qp, qp_attr, attr_mask);
1094 target->status = srp_post_recv(target);
1098 qp_attr->qp_state = IB_QPS_RTS;
1099 target->status = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1103 target->status = ib_modify_qp(target->qp, qp_attr, attr_mask);
1107 target->status = ib_send_cm_rtu(cm_id, NULL, 0);
1113 case IB_CM_REJ_RECEIVED:
1114 printk(KERN_DEBUG PFX "REJ received\n");
1117 srp_cm_rej_handler(cm_id, event, target);
1120 case IB_CM_MRA_RECEIVED:
1121 printk(KERN_ERR PFX "MRA received\n");
1124 case IB_CM_DREP_RECEIVED:
1127 case IB_CM_TIMEWAIT_EXIT:
1128 printk(KERN_ERR PFX "connection closed\n");
1135 printk(KERN_WARNING PFX "Unhandled CM event %d\n", event->event);
1140 complete(&target->done);
1147 static int srp_send_tsk_mgmt(struct scsi_cmnd *scmnd, u8 func)
1149 struct srp_target_port *target = host_to_target(scmnd->device->host);
1150 struct srp_request *req;
1152 struct srp_tsk_mgmt *tsk_mgmt;
1156 spin_lock_irq(target->scsi_host->host_lock);
1158 if (target->state == SRP_TARGET_DEAD ||
1159 target->state == SRP_TARGET_REMOVED) {
1160 scmnd->result = DID_BAD_TARGET << 16;
1164 if (scmnd->host_scribble == (void *) -1L)
1167 req_index = (long) scmnd->host_scribble;
1168 printk(KERN_ERR "Abort for req_index %d\n", req_index);
1170 req = &target->req_ring[req_index];
1171 init_completion(&req->done);
1173 iu = __srp_get_tx_iu(target);
1178 memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
1180 tsk_mgmt->opcode = SRP_TSK_MGMT;
1181 tsk_mgmt->lun = cpu_to_be64((u64) scmnd->device->lun << 48);
1182 tsk_mgmt->tag = req_index | SRP_TAG_TSK_MGMT;
1183 tsk_mgmt->tsk_mgmt_func = func;
1184 tsk_mgmt->task_tag = req_index;
1186 if (__srp_post_send(target, iu, sizeof *tsk_mgmt))
1191 spin_unlock_irq(target->scsi_host->host_lock);
1192 if (!wait_for_completion_timeout(&req->done,
1193 msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
1195 spin_lock_irq(target->scsi_host->host_lock);
1197 if (req->cmd_done) {
1198 list_del(&req->list);
1199 req->next = target->req_head;
1200 target->req_head = req_index;
1202 scmnd->scsi_done(scmnd);
1203 } else if (!req->tsk_status) {
1204 scmnd->result = DID_ABORT << 16;
1209 spin_unlock_irq(target->scsi_host->host_lock);
1213 static int srp_abort(struct scsi_cmnd *scmnd)
1215 printk(KERN_ERR "SRP abort called\n");
1217 return srp_send_tsk_mgmt(scmnd, SRP_TSK_ABORT_TASK);
1220 static int srp_reset_device(struct scsi_cmnd *scmnd)
1222 printk(KERN_ERR "SRP reset_device called\n");
1224 return srp_send_tsk_mgmt(scmnd, SRP_TSK_LUN_RESET);
1227 static int srp_reset_host(struct scsi_cmnd *scmnd)
1229 struct srp_target_port *target = host_to_target(scmnd->device->host);
1232 printk(KERN_ERR PFX "SRP reset_host called\n");
1234 if (!srp_reconnect_target(target))
1240 static struct scsi_host_template srp_template = {
1241 .module = THIS_MODULE,
1243 .info = srp_target_info,
1244 .queuecommand = srp_queuecommand,
1245 .eh_abort_handler = srp_abort,
1246 .eh_device_reset_handler = srp_reset_device,
1247 .eh_host_reset_handler = srp_reset_host,
1248 .can_queue = SRP_SQ_SIZE,
1250 .sg_tablesize = SRP_MAX_INDIRECT,
1251 .cmd_per_lun = SRP_SQ_SIZE,
1252 .use_clustering = ENABLE_CLUSTERING
1255 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
1257 sprintf(target->target_name, "SRP.T10:%016llX",
1258 (unsigned long long) be64_to_cpu(target->id_ext));
1260 if (scsi_add_host(target->scsi_host, host->dev->dma_device))
1263 mutex_lock(&host->target_mutex);
1264 list_add_tail(&target->list, &host->target_list);
1265 mutex_unlock(&host->target_mutex);
1267 target->state = SRP_TARGET_LIVE;
1269 /* XXX: are we supposed to have a definition of SCAN_WILD_CARD ?? */
1270 scsi_scan_target(&target->scsi_host->shost_gendev,
1271 0, target->scsi_id, ~0, 0);
1276 static void srp_release_class_dev(struct class_device *class_dev)
1278 struct srp_host *host =
1279 container_of(class_dev, struct srp_host, class_dev);
1281 complete(&host->released);
1284 static struct class srp_class = {
1285 .name = "infiniband_srp",
1286 .release = srp_release_class_dev
1290 * Target ports are added by writing
1292 * id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
1293 * pkey=<P_Key>,service_id=<service ID>
1295 * to the add_target sysfs attribute.
1299 SRP_OPT_ID_EXT = 1 << 0,
1300 SRP_OPT_IOC_GUID = 1 << 1,
1301 SRP_OPT_DGID = 1 << 2,
1302 SRP_OPT_PKEY = 1 << 3,
1303 SRP_OPT_SERVICE_ID = 1 << 4,
1304 SRP_OPT_MAX_SECT = 1 << 5,
1305 SRP_OPT_ALL = (SRP_OPT_ID_EXT |
1309 SRP_OPT_SERVICE_ID),
1312 static match_table_t srp_opt_tokens = {
1313 { SRP_OPT_ID_EXT, "id_ext=%s" },
1314 { SRP_OPT_IOC_GUID, "ioc_guid=%s" },
1315 { SRP_OPT_DGID, "dgid=%s" },
1316 { SRP_OPT_PKEY, "pkey=%x" },
1317 { SRP_OPT_SERVICE_ID, "service_id=%s" },
1318 { SRP_OPT_MAX_SECT, "max_sect=%d" },
1319 { SRP_OPT_ERR, NULL }
1322 static int srp_parse_options(const char *buf, struct srp_target_port *target)
1324 char *options, *sep_opt;
1327 substring_t args[MAX_OPT_ARGS];
1333 options = kstrdup(buf, GFP_KERNEL);
1338 while ((p = strsep(&sep_opt, ",")) != NULL) {
1342 token = match_token(p, srp_opt_tokens, args);
1346 case SRP_OPT_ID_EXT:
1347 p = match_strdup(args);
1348 target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
1352 case SRP_OPT_IOC_GUID:
1353 p = match_strdup(args);
1354 target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
1359 p = match_strdup(args);
1360 if (strlen(p) != 32) {
1361 printk(KERN_WARNING PFX "bad dest GID parameter '%s'\n", p);
1365 for (i = 0; i < 16; ++i) {
1366 strlcpy(dgid, p + i * 2, 3);
1367 target->path.dgid.raw[i] = simple_strtoul(dgid, NULL, 16);
1372 if (match_hex(args, &token)) {
1373 printk(KERN_WARNING PFX "bad P_Key parameter '%s'\n", p);
1376 target->path.pkey = cpu_to_be16(token);
1379 case SRP_OPT_SERVICE_ID:
1380 p = match_strdup(args);
1381 target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
1385 case SRP_OPT_MAX_SECT:
1386 if (match_int(args, &token)) {
1387 printk(KERN_WARNING PFX "bad max sect parameter '%s'\n", p);
1390 target->scsi_host->max_sectors = token;
1394 printk(KERN_WARNING PFX "unknown parameter or missing value "
1395 "'%s' in target creation request\n", p);
1400 if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
1403 for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
1404 if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
1405 !(srp_opt_tokens[i].token & opt_mask))
1406 printk(KERN_WARNING PFX "target creation request is "
1407 "missing parameter '%s'\n",
1408 srp_opt_tokens[i].pattern);
1415 static ssize_t srp_create_target(struct class_device *class_dev,
1416 const char *buf, size_t count)
1418 struct srp_host *host =
1419 container_of(class_dev, struct srp_host, class_dev);
1420 struct Scsi_Host *target_host;
1421 struct srp_target_port *target;
1425 target_host = scsi_host_alloc(&srp_template,
1426 sizeof (struct srp_target_port));
1430 target_host->max_lun = SRP_MAX_LUN;
1432 target = host_to_target(target_host);
1433 memset(target, 0, sizeof *target);
1435 target->scsi_host = target_host;
1436 target->srp_host = host;
1438 INIT_WORK(&target->work, srp_reconnect_work, target);
1440 for (i = 0; i < SRP_SQ_SIZE - 1; ++i)
1441 target->req_ring[i].next = i + 1;
1442 target->req_ring[SRP_SQ_SIZE - 1].next = -1;
1443 INIT_LIST_HEAD(&target->req_queue);
1445 ret = srp_parse_options(buf, target);
1449 ib_get_cached_gid(host->dev, host->port, 0, &target->path.sgid);
1451 printk(KERN_DEBUG PFX "new target: id_ext %016llx ioc_guid %016llx pkey %04x "
1452 "service_id %016llx dgid %04x:%04x:%04x:%04x:%04x:%04x:%04x:%04x\n",
1453 (unsigned long long) be64_to_cpu(target->id_ext),
1454 (unsigned long long) be64_to_cpu(target->ioc_guid),
1455 be16_to_cpu(target->path.pkey),
1456 (unsigned long long) be64_to_cpu(target->service_id),
1457 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[0]),
1458 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[2]),
1459 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[4]),
1460 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[6]),
1461 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[8]),
1462 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[10]),
1463 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[12]),
1464 (int) be16_to_cpu(*(__be16 *) &target->path.dgid.raw[14]));
1466 ret = srp_create_target_ib(target);
1470 target->cm_id = ib_create_cm_id(host->dev, srp_cm_handler, target);
1471 if (IS_ERR(target->cm_id)) {
1472 ret = PTR_ERR(target->cm_id);
1476 ret = srp_connect_target(target);
1478 printk(KERN_ERR PFX "Connection failed\n");
1482 ret = srp_add_target(host, target);
1484 goto err_disconnect;
1489 srp_disconnect_target(target);
1492 ib_destroy_cm_id(target->cm_id);
1495 srp_free_target_ib(target);
1498 scsi_host_put(target_host);
1503 static CLASS_DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
1505 static ssize_t show_ibdev(struct class_device *class_dev, char *buf)
1507 struct srp_host *host =
1508 container_of(class_dev, struct srp_host, class_dev);
1510 return sprintf(buf, "%s\n", host->dev->name);
1513 static CLASS_DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
1515 static ssize_t show_port(struct class_device *class_dev, char *buf)
1517 struct srp_host *host =
1518 container_of(class_dev, struct srp_host, class_dev);
1520 return sprintf(buf, "%d\n", host->port);
1523 static CLASS_DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
1525 static struct srp_host *srp_add_port(struct ib_device *device, u8 port)
1527 struct srp_host *host;
1529 host = kzalloc(sizeof *host, GFP_KERNEL);
1533 INIT_LIST_HEAD(&host->target_list);
1534 mutex_init(&host->target_mutex);
1535 init_completion(&host->released);
1539 host->initiator_port_id[7] = port;
1540 memcpy(host->initiator_port_id + 8, &device->node_guid, 8);
1542 host->pd = ib_alloc_pd(device);
1543 if (IS_ERR(host->pd))
1546 host->mr = ib_get_dma_mr(host->pd,
1547 IB_ACCESS_LOCAL_WRITE |
1548 IB_ACCESS_REMOTE_READ |
1549 IB_ACCESS_REMOTE_WRITE);
1550 if (IS_ERR(host->mr))
1553 host->class_dev.class = &srp_class;
1554 host->class_dev.dev = device->dma_device;
1555 snprintf(host->class_dev.class_id, BUS_ID_SIZE, "srp-%s-%d",
1556 device->name, port);
1558 if (class_device_register(&host->class_dev))
1560 if (class_device_create_file(&host->class_dev, &class_device_attr_add_target))
1562 if (class_device_create_file(&host->class_dev, &class_device_attr_ibdev))
1564 if (class_device_create_file(&host->class_dev, &class_device_attr_port))
1570 class_device_unregister(&host->class_dev);
1573 ib_dereg_mr(host->mr);
1576 ib_dealloc_pd(host->pd);
1584 static void srp_add_one(struct ib_device *device)
1586 struct list_head *dev_list;
1587 struct srp_host *host;
1590 dev_list = kmalloc(sizeof *dev_list, GFP_KERNEL);
1594 INIT_LIST_HEAD(dev_list);
1596 if (device->node_type == IB_NODE_SWITCH) {
1601 e = device->phys_port_cnt;
1604 for (p = s; p <= e; ++p) {
1605 host = srp_add_port(device, p);
1607 list_add_tail(&host->list, dev_list);
1610 ib_set_client_data(device, &srp_client, dev_list);
1613 static void srp_remove_one(struct ib_device *device)
1615 struct list_head *dev_list;
1616 struct srp_host *host, *tmp_host;
1617 LIST_HEAD(target_list);
1618 struct srp_target_port *target, *tmp_target;
1619 unsigned long flags;
1621 dev_list = ib_get_client_data(device, &srp_client);
1623 list_for_each_entry_safe(host, tmp_host, dev_list, list) {
1624 class_device_unregister(&host->class_dev);
1626 * Wait for the sysfs entry to go away, so that no new
1627 * target ports can be created.
1629 wait_for_completion(&host->released);
1632 * Mark all target ports as removed, so we stop queueing
1633 * commands and don't try to reconnect.
1635 mutex_lock(&host->target_mutex);
1636 list_for_each_entry_safe(target, tmp_target,
1637 &host->target_list, list) {
1638 spin_lock_irqsave(target->scsi_host->host_lock, flags);
1639 if (target->state != SRP_TARGET_REMOVED)
1640 target->state = SRP_TARGET_REMOVED;
1641 spin_unlock_irqrestore(target->scsi_host->host_lock, flags);
1643 mutex_unlock(&host->target_mutex);
1646 * Wait for any reconnection tasks that may have
1647 * started before we marked our target ports as
1648 * removed, and any target port removal tasks.
1650 flush_scheduled_work();
1652 list_for_each_entry_safe(target, tmp_target,
1653 &host->target_list, list) {
1654 scsi_remove_host(target->scsi_host);
1655 srp_disconnect_target(target);
1656 ib_destroy_cm_id(target->cm_id);
1657 srp_free_target_ib(target);
1658 scsi_host_put(target->scsi_host);
1661 ib_dereg_mr(host->mr);
1662 ib_dealloc_pd(host->pd);
1669 static int __init srp_init_module(void)
1673 ret = class_register(&srp_class);
1675 printk(KERN_ERR PFX "couldn't register class infiniband_srp\n");
1679 ret = ib_register_client(&srp_client);
1681 printk(KERN_ERR PFX "couldn't register IB client\n");
1682 class_unregister(&srp_class);
1689 static void __exit srp_cleanup_module(void)
1691 ib_unregister_client(&srp_client);
1692 class_unregister(&srp_class);
1695 module_init(srp_init_module);
1696 module_exit(srp_cleanup_module);