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
33 #define pr_fmt(fmt) PFX fmt
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 <linux/atomic.h>
46 #include <scsi/scsi.h>
47 #include <scsi/scsi_device.h>
48 #include <scsi/scsi_dbg.h>
50 #include <scsi/scsi_transport_srp.h>
54 #define DRV_NAME "ib_srp"
55 #define PFX DRV_NAME ": "
56 #define DRV_VERSION "0.2"
57 #define DRV_RELDATE "November 1, 2005"
59 MODULE_AUTHOR("Roland Dreier");
60 MODULE_DESCRIPTION("InfiniBand SCSI RDMA Protocol initiator "
61 "v" DRV_VERSION " (" DRV_RELDATE ")");
62 MODULE_LICENSE("Dual BSD/GPL");
64 static unsigned int srp_sg_tablesize;
65 static unsigned int cmd_sg_entries;
66 static unsigned int indirect_sg_entries;
67 static bool allow_ext_sg;
68 static int topspin_workarounds = 1;
70 module_param(srp_sg_tablesize, uint, 0444);
71 MODULE_PARM_DESC(srp_sg_tablesize, "Deprecated name for cmd_sg_entries");
73 module_param(cmd_sg_entries, uint, 0444);
74 MODULE_PARM_DESC(cmd_sg_entries,
75 "Default number of gather/scatter entries in the SRP command (default is 12, max 255)");
77 module_param(indirect_sg_entries, uint, 0444);
78 MODULE_PARM_DESC(indirect_sg_entries,
79 "Default max number of gather/scatter entries (default is 12, max is " __stringify(SCSI_MAX_SG_CHAIN_SEGMENTS) ")");
81 module_param(allow_ext_sg, bool, 0444);
82 MODULE_PARM_DESC(allow_ext_sg,
83 "Default behavior when there are more than cmd_sg_entries S/G entries after mapping; fails the request when false (default false)");
85 module_param(topspin_workarounds, int, 0444);
86 MODULE_PARM_DESC(topspin_workarounds,
87 "Enable workarounds for Topspin/Cisco SRP target bugs if != 0");
89 static void srp_add_one(struct ib_device *device);
90 static void srp_remove_one(struct ib_device *device);
91 static void srp_recv_completion(struct ib_cq *cq, void *target_ptr);
92 static void srp_send_completion(struct ib_cq *cq, void *target_ptr);
93 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event);
95 static struct scsi_transport_template *ib_srp_transport_template;
97 static struct ib_client srp_client = {
100 .remove = srp_remove_one
103 static struct ib_sa_client srp_sa_client;
105 static inline struct srp_target_port *host_to_target(struct Scsi_Host *host)
107 return (struct srp_target_port *) host->hostdata;
110 static const char *srp_target_info(struct Scsi_Host *host)
112 return host_to_target(host)->target_name;
115 static int srp_target_is_topspin(struct srp_target_port *target)
117 static const u8 topspin_oui[3] = { 0x00, 0x05, 0xad };
118 static const u8 cisco_oui[3] = { 0x00, 0x1b, 0x0d };
120 return topspin_workarounds &&
121 (!memcmp(&target->ioc_guid, topspin_oui, sizeof topspin_oui) ||
122 !memcmp(&target->ioc_guid, cisco_oui, sizeof cisco_oui));
125 static struct srp_iu *srp_alloc_iu(struct srp_host *host, size_t size,
127 enum dma_data_direction direction)
131 iu = kmalloc(sizeof *iu, gfp_mask);
135 iu->buf = kzalloc(size, gfp_mask);
139 iu->dma = ib_dma_map_single(host->srp_dev->dev, iu->buf, size,
141 if (ib_dma_mapping_error(host->srp_dev->dev, iu->dma))
145 iu->direction = direction;
157 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
162 ib_dma_unmap_single(host->srp_dev->dev, iu->dma, iu->size,
168 static void srp_qp_event(struct ib_event *event, void *context)
170 pr_debug("QP event %d\n", event->event);
173 static int srp_init_qp(struct srp_target_port *target,
176 struct ib_qp_attr *attr;
179 attr = kmalloc(sizeof *attr, GFP_KERNEL);
183 ret = ib_find_pkey(target->srp_host->srp_dev->dev,
184 target->srp_host->port,
185 be16_to_cpu(target->path.pkey),
190 attr->qp_state = IB_QPS_INIT;
191 attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
192 IB_ACCESS_REMOTE_WRITE);
193 attr->port_num = target->srp_host->port;
195 ret = ib_modify_qp(qp, attr,
206 static int srp_new_cm_id(struct srp_target_port *target)
208 struct ib_cm_id *new_cm_id;
210 new_cm_id = ib_create_cm_id(target->srp_host->srp_dev->dev,
211 srp_cm_handler, target);
212 if (IS_ERR(new_cm_id))
213 return PTR_ERR(new_cm_id);
216 ib_destroy_cm_id(target->cm_id);
217 target->cm_id = new_cm_id;
222 static int srp_create_target_ib(struct srp_target_port *target)
224 struct ib_qp_init_attr *init_attr;
227 init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
231 target->recv_cq = ib_create_cq(target->srp_host->srp_dev->dev,
232 srp_recv_completion, NULL, target, SRP_RQ_SIZE, 0);
233 if (IS_ERR(target->recv_cq)) {
234 ret = PTR_ERR(target->recv_cq);
238 target->send_cq = ib_create_cq(target->srp_host->srp_dev->dev,
239 srp_send_completion, NULL, target, SRP_SQ_SIZE, 0);
240 if (IS_ERR(target->send_cq)) {
241 ret = PTR_ERR(target->send_cq);
245 ib_req_notify_cq(target->recv_cq, IB_CQ_NEXT_COMP);
247 init_attr->event_handler = srp_qp_event;
248 init_attr->cap.max_send_wr = SRP_SQ_SIZE;
249 init_attr->cap.max_recv_wr = SRP_RQ_SIZE;
250 init_attr->cap.max_recv_sge = 1;
251 init_attr->cap.max_send_sge = 1;
252 init_attr->sq_sig_type = IB_SIGNAL_ALL_WR;
253 init_attr->qp_type = IB_QPT_RC;
254 init_attr->send_cq = target->send_cq;
255 init_attr->recv_cq = target->recv_cq;
257 target->qp = ib_create_qp(target->srp_host->srp_dev->pd, init_attr);
258 if (IS_ERR(target->qp)) {
259 ret = PTR_ERR(target->qp);
263 ret = srp_init_qp(target, target->qp);
271 ib_destroy_qp(target->qp);
274 ib_destroy_cq(target->send_cq);
277 ib_destroy_cq(target->recv_cq);
284 static void srp_free_target_ib(struct srp_target_port *target)
288 ib_destroy_qp(target->qp);
289 ib_destroy_cq(target->send_cq);
290 ib_destroy_cq(target->recv_cq);
292 for (i = 0; i < SRP_RQ_SIZE; ++i)
293 srp_free_iu(target->srp_host, target->rx_ring[i]);
294 for (i = 0; i < SRP_SQ_SIZE; ++i)
295 srp_free_iu(target->srp_host, target->tx_ring[i]);
298 static void srp_path_rec_completion(int status,
299 struct ib_sa_path_rec *pathrec,
302 struct srp_target_port *target = target_ptr;
304 target->status = status;
306 shost_printk(KERN_ERR, target->scsi_host,
307 PFX "Got failed path rec status %d\n", status);
309 target->path = *pathrec;
310 complete(&target->done);
313 static int srp_lookup_path(struct srp_target_port *target)
315 target->path.numb_path = 1;
317 init_completion(&target->done);
319 target->path_query_id = ib_sa_path_rec_get(&srp_sa_client,
320 target->srp_host->srp_dev->dev,
321 target->srp_host->port,
323 IB_SA_PATH_REC_SERVICE_ID |
324 IB_SA_PATH_REC_DGID |
325 IB_SA_PATH_REC_SGID |
326 IB_SA_PATH_REC_NUMB_PATH |
328 SRP_PATH_REC_TIMEOUT_MS,
330 srp_path_rec_completion,
331 target, &target->path_query);
332 if (target->path_query_id < 0)
333 return target->path_query_id;
335 wait_for_completion(&target->done);
337 if (target->status < 0)
338 shost_printk(KERN_WARNING, target->scsi_host,
339 PFX "Path record query failed\n");
341 return target->status;
344 static int srp_send_req(struct srp_target_port *target)
347 struct ib_cm_req_param param;
348 struct srp_login_req priv;
352 req = kzalloc(sizeof *req, GFP_KERNEL);
356 req->param.primary_path = &target->path;
357 req->param.alternate_path = NULL;
358 req->param.service_id = target->service_id;
359 req->param.qp_num = target->qp->qp_num;
360 req->param.qp_type = target->qp->qp_type;
361 req->param.private_data = &req->priv;
362 req->param.private_data_len = sizeof req->priv;
363 req->param.flow_control = 1;
365 get_random_bytes(&req->param.starting_psn, 4);
366 req->param.starting_psn &= 0xffffff;
369 * Pick some arbitrary defaults here; we could make these
370 * module parameters if anyone cared about setting them.
372 req->param.responder_resources = 4;
373 req->param.remote_cm_response_timeout = 20;
374 req->param.local_cm_response_timeout = 20;
375 req->param.retry_count = 7;
376 req->param.rnr_retry_count = 7;
377 req->param.max_cm_retries = 15;
379 req->priv.opcode = SRP_LOGIN_REQ;
381 req->priv.req_it_iu_len = cpu_to_be32(target->max_iu_len);
382 req->priv.req_buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
383 SRP_BUF_FORMAT_INDIRECT);
385 * In the published SRP specification (draft rev. 16a), the
386 * port identifier format is 8 bytes of ID extension followed
387 * by 8 bytes of GUID. Older drafts put the two halves in the
388 * opposite order, so that the GUID comes first.
390 * Targets conforming to these obsolete drafts can be
391 * recognized by the I/O Class they report.
393 if (target->io_class == SRP_REV10_IB_IO_CLASS) {
394 memcpy(req->priv.initiator_port_id,
395 &target->path.sgid.global.interface_id, 8);
396 memcpy(req->priv.initiator_port_id + 8,
397 &target->initiator_ext, 8);
398 memcpy(req->priv.target_port_id, &target->ioc_guid, 8);
399 memcpy(req->priv.target_port_id + 8, &target->id_ext, 8);
401 memcpy(req->priv.initiator_port_id,
402 &target->initiator_ext, 8);
403 memcpy(req->priv.initiator_port_id + 8,
404 &target->path.sgid.global.interface_id, 8);
405 memcpy(req->priv.target_port_id, &target->id_ext, 8);
406 memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
410 * Topspin/Cisco SRP targets will reject our login unless we
411 * zero out the first 8 bytes of our initiator port ID and set
412 * the second 8 bytes to the local node GUID.
414 if (srp_target_is_topspin(target)) {
415 shost_printk(KERN_DEBUG, target->scsi_host,
416 PFX "Topspin/Cisco initiator port ID workaround "
417 "activated for target GUID %016llx\n",
418 (unsigned long long) be64_to_cpu(target->ioc_guid));
419 memset(req->priv.initiator_port_id, 0, 8);
420 memcpy(req->priv.initiator_port_id + 8,
421 &target->srp_host->srp_dev->dev->node_guid, 8);
424 status = ib_send_cm_req(target->cm_id, &req->param);
431 static void srp_disconnect_target(struct srp_target_port *target)
433 /* XXX should send SRP_I_LOGOUT request */
435 init_completion(&target->done);
436 if (ib_send_cm_dreq(target->cm_id, NULL, 0)) {
437 shost_printk(KERN_DEBUG, target->scsi_host,
438 PFX "Sending CM DREQ failed\n");
441 wait_for_completion(&target->done);
444 static bool srp_change_state(struct srp_target_port *target,
445 enum srp_target_state old,
446 enum srp_target_state new)
448 bool changed = false;
450 spin_lock_irq(&target->lock);
451 if (target->state == old) {
455 spin_unlock_irq(&target->lock);
459 static void srp_free_req_data(struct srp_target_port *target)
461 struct ib_device *ibdev = target->srp_host->srp_dev->dev;
462 struct srp_request *req;
465 for (i = 0, req = target->req_ring; i < SRP_CMD_SQ_SIZE; ++i, ++req) {
466 kfree(req->fmr_list);
467 kfree(req->map_page);
468 if (req->indirect_dma_addr) {
469 ib_dma_unmap_single(ibdev, req->indirect_dma_addr,
470 target->indirect_size,
473 kfree(req->indirect_desc);
478 * srp_del_scsi_host_attr() - Remove attributes defined in the host template.
479 * @shost: SCSI host whose attributes to remove from sysfs.
481 * Note: Any attributes defined in the host template and that did not exist
482 * before invocation of this function will be ignored.
484 static void srp_del_scsi_host_attr(struct Scsi_Host *shost)
486 struct device_attribute **attr;
488 for (attr = shost->hostt->shost_attrs; attr && *attr; ++attr)
489 device_remove_file(&shost->shost_dev, *attr);
492 static void srp_remove_work(struct work_struct *work)
494 struct srp_target_port *target =
495 container_of(work, struct srp_target_port, work);
497 if (!srp_change_state(target, SRP_TARGET_DEAD, SRP_TARGET_REMOVED))
500 spin_lock(&target->srp_host->target_lock);
501 list_del(&target->list);
502 spin_unlock(&target->srp_host->target_lock);
504 srp_del_scsi_host_attr(target->scsi_host);
505 srp_remove_host(target->scsi_host);
506 scsi_remove_host(target->scsi_host);
507 ib_destroy_cm_id(target->cm_id);
508 srp_free_target_ib(target);
509 srp_free_req_data(target);
510 scsi_host_put(target->scsi_host);
513 static int srp_connect_target(struct srp_target_port *target)
518 ret = srp_lookup_path(target);
523 init_completion(&target->done);
524 ret = srp_send_req(target);
527 wait_for_completion(&target->done);
530 * The CM event handling code will set status to
531 * SRP_PORT_REDIRECT if we get a port redirect REJ
532 * back, or SRP_DLID_REDIRECT if we get a lid/qp
535 switch (target->status) {
539 case SRP_PORT_REDIRECT:
540 ret = srp_lookup_path(target);
545 case SRP_DLID_REDIRECT:
549 /* Our current CM id was stale, and is now in timewait.
550 * Try to reconnect with a new one.
552 if (!retries-- || srp_new_cm_id(target)) {
553 shost_printk(KERN_ERR, target->scsi_host, PFX
554 "giving up on stale connection\n");
555 target->status = -ECONNRESET;
556 return target->status;
559 shost_printk(KERN_ERR, target->scsi_host, PFX
560 "retrying stale connection\n");
564 return target->status;
569 static void srp_unmap_data(struct scsi_cmnd *scmnd,
570 struct srp_target_port *target,
571 struct srp_request *req)
573 struct ib_device *ibdev = target->srp_host->srp_dev->dev;
574 struct ib_pool_fmr **pfmr;
576 if (!scsi_sglist(scmnd) ||
577 (scmnd->sc_data_direction != DMA_TO_DEVICE &&
578 scmnd->sc_data_direction != DMA_FROM_DEVICE))
581 pfmr = req->fmr_list;
583 ib_fmr_pool_unmap(*pfmr++);
585 ib_dma_unmap_sg(ibdev, scsi_sglist(scmnd), scsi_sg_count(scmnd),
586 scmnd->sc_data_direction);
590 * srp_claim_req - Take ownership of the scmnd associated with a request.
591 * @target: SRP target port.
593 * @scmnd: If NULL, take ownership of @req->scmnd. If not NULL, only take
594 * ownership of @req->scmnd if it equals @scmnd.
597 * Either NULL or a pointer to the SCSI command the caller became owner of.
599 static struct scsi_cmnd *srp_claim_req(struct srp_target_port *target,
600 struct srp_request *req,
601 struct scsi_cmnd *scmnd)
605 spin_lock_irqsave(&target->lock, flags);
609 } else if (req->scmnd == scmnd) {
614 spin_unlock_irqrestore(&target->lock, flags);
620 * srp_free_req() - Unmap data and add request to the free request list.
622 static void srp_free_req(struct srp_target_port *target,
623 struct srp_request *req, struct scsi_cmnd *scmnd,
628 srp_unmap_data(scmnd, target, req);
630 spin_lock_irqsave(&target->lock, flags);
631 target->req_lim += req_lim_delta;
632 list_add_tail(&req->list, &target->free_reqs);
633 spin_unlock_irqrestore(&target->lock, flags);
636 static void srp_reset_req(struct srp_target_port *target, struct srp_request *req)
638 struct scsi_cmnd *scmnd = srp_claim_req(target, req, NULL);
641 srp_free_req(target, req, scmnd, 0);
642 scmnd->result = DID_RESET << 16;
643 scmnd->scsi_done(scmnd);
647 static int srp_reconnect_target(struct srp_target_port *target)
649 struct Scsi_Host *shost = target->scsi_host;
650 struct ib_qp_attr qp_attr;
654 if (target->state != SRP_TARGET_LIVE)
657 scsi_target_block(&shost->shost_gendev);
659 srp_disconnect_target(target);
661 * Now get a new local CM ID so that we avoid confusing the
662 * target in case things are really fouled up.
664 ret = srp_new_cm_id(target);
668 qp_attr.qp_state = IB_QPS_RESET;
669 ret = ib_modify_qp(target->qp, &qp_attr, IB_QP_STATE);
673 ret = srp_init_qp(target, target->qp);
677 while (ib_poll_cq(target->recv_cq, 1, &wc) > 0)
679 while (ib_poll_cq(target->send_cq, 1, &wc) > 0)
682 for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
683 struct srp_request *req = &target->req_ring[i];
685 srp_reset_req(target, req);
688 INIT_LIST_HEAD(&target->free_tx);
689 for (i = 0; i < SRP_SQ_SIZE; ++i)
690 list_add(&target->tx_ring[i]->list, &target->free_tx);
692 target->qp_in_error = 0;
693 ret = srp_connect_target(target);
696 scsi_target_unblock(&shost->shost_gendev, ret == 0 ? SDEV_RUNNING :
697 SDEV_TRANSPORT_OFFLINE);
702 shost_printk(KERN_INFO, target->scsi_host, PFX "reconnect succeeded\n");
707 shost_printk(KERN_ERR, target->scsi_host,
708 PFX "reconnect failed (%d), removing target port.\n", ret);
711 * We couldn't reconnect, so kill our target port off.
712 * However, we have to defer the real removal because we
713 * are in the context of the SCSI error handler now, which
714 * will deadlock if we call scsi_remove_host().
716 * Schedule our work inside the lock to avoid a race with
717 * the flush_scheduled_work() in srp_remove_one().
719 spin_lock_irq(&target->lock);
720 if (target->state == SRP_TARGET_LIVE) {
721 target->state = SRP_TARGET_DEAD;
722 INIT_WORK(&target->work, srp_remove_work);
723 queue_work(ib_wq, &target->work);
725 spin_unlock_irq(&target->lock);
730 static void srp_map_desc(struct srp_map_state *state, dma_addr_t dma_addr,
731 unsigned int dma_len, u32 rkey)
733 struct srp_direct_buf *desc = state->desc;
735 desc->va = cpu_to_be64(dma_addr);
736 desc->key = cpu_to_be32(rkey);
737 desc->len = cpu_to_be32(dma_len);
739 state->total_len += dma_len;
744 static int srp_map_finish_fmr(struct srp_map_state *state,
745 struct srp_target_port *target)
747 struct srp_device *dev = target->srp_host->srp_dev;
748 struct ib_pool_fmr *fmr;
754 if (state->npages == 1) {
755 srp_map_desc(state, state->base_dma_addr, state->fmr_len,
757 state->npages = state->fmr_len = 0;
761 fmr = ib_fmr_pool_map_phys(dev->fmr_pool, state->pages,
762 state->npages, io_addr);
766 *state->next_fmr++ = fmr;
769 srp_map_desc(state, 0, state->fmr_len, fmr->fmr->rkey);
770 state->npages = state->fmr_len = 0;
774 static void srp_map_update_start(struct srp_map_state *state,
775 struct scatterlist *sg, int sg_index,
778 state->unmapped_sg = sg;
779 state->unmapped_index = sg_index;
780 state->unmapped_addr = dma_addr;
783 static int srp_map_sg_entry(struct srp_map_state *state,
784 struct srp_target_port *target,
785 struct scatterlist *sg, int sg_index,
788 struct srp_device *dev = target->srp_host->srp_dev;
789 struct ib_device *ibdev = dev->dev;
790 dma_addr_t dma_addr = ib_sg_dma_address(ibdev, sg);
791 unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
798 if (use_fmr == SRP_MAP_NO_FMR) {
799 /* Once we're in direct map mode for a request, we don't
800 * go back to FMR mode, so no need to update anything
801 * other than the descriptor.
803 srp_map_desc(state, dma_addr, dma_len, target->rkey);
807 /* If we start at an offset into the FMR page, don't merge into
808 * the current FMR. Finish it out, and use the kernel's MR for this
809 * sg entry. This is to avoid potential bugs on some SRP targets
810 * that were never quite defined, but went away when the initiator
811 * avoided using FMR on such page fragments.
813 if (dma_addr & ~dev->fmr_page_mask || dma_len > dev->fmr_max_size) {
814 ret = srp_map_finish_fmr(state, target);
818 srp_map_desc(state, dma_addr, dma_len, target->rkey);
819 srp_map_update_start(state, NULL, 0, 0);
823 /* If this is the first sg to go into the FMR, save our position.
824 * We need to know the first unmapped entry, its index, and the
825 * first unmapped address within that entry to be able to restart
826 * mapping after an error.
828 if (!state->unmapped_sg)
829 srp_map_update_start(state, sg, sg_index, dma_addr);
832 if (state->npages == SRP_FMR_SIZE) {
833 ret = srp_map_finish_fmr(state, target);
837 srp_map_update_start(state, sg, sg_index, dma_addr);
840 len = min_t(unsigned int, dma_len, dev->fmr_page_size);
843 state->base_dma_addr = dma_addr;
844 state->pages[state->npages++] = dma_addr;
845 state->fmr_len += len;
850 /* If the last entry of the FMR wasn't a full page, then we need to
851 * close it out and start a new one -- we can only merge at page
855 if (len != dev->fmr_page_size) {
856 ret = srp_map_finish_fmr(state, target);
858 srp_map_update_start(state, NULL, 0, 0);
863 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_target_port *target,
864 struct srp_request *req)
866 struct scatterlist *scat, *sg;
867 struct srp_cmd *cmd = req->cmd->buf;
868 int i, len, nents, count, use_fmr;
869 struct srp_device *dev;
870 struct ib_device *ibdev;
871 struct srp_map_state state;
872 struct srp_indirect_buf *indirect_hdr;
876 if (!scsi_sglist(scmnd) || scmnd->sc_data_direction == DMA_NONE)
877 return sizeof (struct srp_cmd);
879 if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
880 scmnd->sc_data_direction != DMA_TO_DEVICE) {
881 shost_printk(KERN_WARNING, target->scsi_host,
882 PFX "Unhandled data direction %d\n",
883 scmnd->sc_data_direction);
887 nents = scsi_sg_count(scmnd);
888 scat = scsi_sglist(scmnd);
890 dev = target->srp_host->srp_dev;
893 count = ib_dma_map_sg(ibdev, scat, nents, scmnd->sc_data_direction);
894 if (unlikely(count == 0))
897 fmt = SRP_DATA_DESC_DIRECT;
898 len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
902 * The midlayer only generated a single gather/scatter
903 * entry, or DMA mapping coalesced everything to a
904 * single entry. So a direct descriptor along with
905 * the DMA MR suffices.
907 struct srp_direct_buf *buf = (void *) cmd->add_data;
909 buf->va = cpu_to_be64(ib_sg_dma_address(ibdev, scat));
910 buf->key = cpu_to_be32(target->rkey);
911 buf->len = cpu_to_be32(ib_sg_dma_len(ibdev, scat));
917 /* We have more than one scatter/gather entry, so build our indirect
918 * descriptor table, trying to merge as many entries with FMR as we
921 indirect_hdr = (void *) cmd->add_data;
923 ib_dma_sync_single_for_cpu(ibdev, req->indirect_dma_addr,
924 target->indirect_size, DMA_TO_DEVICE);
926 memset(&state, 0, sizeof(state));
927 state.desc = req->indirect_desc;
928 state.pages = req->map_page;
929 state.next_fmr = req->fmr_list;
931 use_fmr = dev->fmr_pool ? SRP_MAP_ALLOW_FMR : SRP_MAP_NO_FMR;
933 for_each_sg(scat, sg, count, i) {
934 if (srp_map_sg_entry(&state, target, sg, i, use_fmr)) {
935 /* FMR mapping failed, so backtrack to the first
936 * unmapped entry and continue on without using FMR.
939 unsigned int dma_len;
942 sg = state.unmapped_sg;
943 i = state.unmapped_index;
945 dma_addr = ib_sg_dma_address(ibdev, sg);
946 dma_len = ib_sg_dma_len(ibdev, sg);
947 dma_len -= (state.unmapped_addr - dma_addr);
948 dma_addr = state.unmapped_addr;
949 use_fmr = SRP_MAP_NO_FMR;
950 srp_map_desc(&state, dma_addr, dma_len, target->rkey);
954 if (use_fmr == SRP_MAP_ALLOW_FMR && srp_map_finish_fmr(&state, target))
957 /* We've mapped the request, now pull as much of the indirect
958 * descriptor table as we can into the command buffer. If this
959 * target is not using an external indirect table, we are
960 * guaranteed to fit into the command, as the SCSI layer won't
961 * give us more S/G entries than we allow.
963 req->nfmr = state.nfmr;
964 if (state.ndesc == 1) {
965 /* FMR mapping was able to collapse this to one entry,
966 * so use a direct descriptor.
968 struct srp_direct_buf *buf = (void *) cmd->add_data;
970 *buf = req->indirect_desc[0];
974 if (unlikely(target->cmd_sg_cnt < state.ndesc &&
975 !target->allow_ext_sg)) {
976 shost_printk(KERN_ERR, target->scsi_host,
977 "Could not fit S/G list into SRP_CMD\n");
981 count = min(state.ndesc, target->cmd_sg_cnt);
982 table_len = state.ndesc * sizeof (struct srp_direct_buf);
984 fmt = SRP_DATA_DESC_INDIRECT;
985 len = sizeof(struct srp_cmd) + sizeof (struct srp_indirect_buf);
986 len += count * sizeof (struct srp_direct_buf);
988 memcpy(indirect_hdr->desc_list, req->indirect_desc,
989 count * sizeof (struct srp_direct_buf));
991 indirect_hdr->table_desc.va = cpu_to_be64(req->indirect_dma_addr);
992 indirect_hdr->table_desc.key = cpu_to_be32(target->rkey);
993 indirect_hdr->table_desc.len = cpu_to_be32(table_len);
994 indirect_hdr->len = cpu_to_be32(state.total_len);
996 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
997 cmd->data_out_desc_cnt = count;
999 cmd->data_in_desc_cnt = count;
1001 ib_dma_sync_single_for_device(ibdev, req->indirect_dma_addr, table_len,
1005 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1006 cmd->buf_fmt = fmt << 4;
1014 * Return an IU and possible credit to the free pool
1016 static void srp_put_tx_iu(struct srp_target_port *target, struct srp_iu *iu,
1017 enum srp_iu_type iu_type)
1019 unsigned long flags;
1021 spin_lock_irqsave(&target->lock, flags);
1022 list_add(&iu->list, &target->free_tx);
1023 if (iu_type != SRP_IU_RSP)
1025 spin_unlock_irqrestore(&target->lock, flags);
1029 * Must be called with target->lock held to protect req_lim and free_tx.
1030 * If IU is not sent, it must be returned using srp_put_tx_iu().
1033 * An upper limit for the number of allocated information units for each
1035 * - SRP_IU_CMD: SRP_CMD_SQ_SIZE, since the SCSI mid-layer never queues
1036 * more than Scsi_Host.can_queue requests.
1037 * - SRP_IU_TSK_MGMT: SRP_TSK_MGMT_SQ_SIZE.
1038 * - SRP_IU_RSP: 1, since a conforming SRP target never sends more than
1039 * one unanswered SRP request to an initiator.
1041 static struct srp_iu *__srp_get_tx_iu(struct srp_target_port *target,
1042 enum srp_iu_type iu_type)
1044 s32 rsv = (iu_type == SRP_IU_TSK_MGMT) ? 0 : SRP_TSK_MGMT_SQ_SIZE;
1047 srp_send_completion(target->send_cq, target);
1049 if (list_empty(&target->free_tx))
1052 /* Initiator responses to target requests do not consume credits */
1053 if (iu_type != SRP_IU_RSP) {
1054 if (target->req_lim <= rsv) {
1055 ++target->zero_req_lim;
1062 iu = list_first_entry(&target->free_tx, struct srp_iu, list);
1063 list_del(&iu->list);
1067 static int srp_post_send(struct srp_target_port *target,
1068 struct srp_iu *iu, int len)
1071 struct ib_send_wr wr, *bad_wr;
1073 list.addr = iu->dma;
1075 list.lkey = target->lkey;
1078 wr.wr_id = (uintptr_t) iu;
1081 wr.opcode = IB_WR_SEND;
1082 wr.send_flags = IB_SEND_SIGNALED;
1084 return ib_post_send(target->qp, &wr, &bad_wr);
1087 static int srp_post_recv(struct srp_target_port *target, struct srp_iu *iu)
1089 struct ib_recv_wr wr, *bad_wr;
1092 list.addr = iu->dma;
1093 list.length = iu->size;
1094 list.lkey = target->lkey;
1097 wr.wr_id = (uintptr_t) iu;
1101 return ib_post_recv(target->qp, &wr, &bad_wr);
1104 static void srp_process_rsp(struct srp_target_port *target, struct srp_rsp *rsp)
1106 struct srp_request *req;
1107 struct scsi_cmnd *scmnd;
1108 unsigned long flags;
1110 if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
1111 spin_lock_irqsave(&target->lock, flags);
1112 target->req_lim += be32_to_cpu(rsp->req_lim_delta);
1113 spin_unlock_irqrestore(&target->lock, flags);
1115 target->tsk_mgmt_status = -1;
1116 if (be32_to_cpu(rsp->resp_data_len) >= 4)
1117 target->tsk_mgmt_status = rsp->data[3];
1118 complete(&target->tsk_mgmt_done);
1120 req = &target->req_ring[rsp->tag];
1121 scmnd = srp_claim_req(target, req, NULL);
1123 shost_printk(KERN_ERR, target->scsi_host,
1124 "Null scmnd for RSP w/tag %016llx\n",
1125 (unsigned long long) rsp->tag);
1127 spin_lock_irqsave(&target->lock, flags);
1128 target->req_lim += be32_to_cpu(rsp->req_lim_delta);
1129 spin_unlock_irqrestore(&target->lock, flags);
1133 scmnd->result = rsp->status;
1135 if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
1136 memcpy(scmnd->sense_buffer, rsp->data +
1137 be32_to_cpu(rsp->resp_data_len),
1138 min_t(int, be32_to_cpu(rsp->sense_data_len),
1139 SCSI_SENSE_BUFFERSIZE));
1142 if (rsp->flags & (SRP_RSP_FLAG_DOOVER | SRP_RSP_FLAG_DOUNDER))
1143 scsi_set_resid(scmnd, be32_to_cpu(rsp->data_out_res_cnt));
1144 else if (rsp->flags & (SRP_RSP_FLAG_DIOVER | SRP_RSP_FLAG_DIUNDER))
1145 scsi_set_resid(scmnd, be32_to_cpu(rsp->data_in_res_cnt));
1147 srp_free_req(target, req, scmnd,
1148 be32_to_cpu(rsp->req_lim_delta));
1150 scmnd->host_scribble = NULL;
1151 scmnd->scsi_done(scmnd);
1155 static int srp_response_common(struct srp_target_port *target, s32 req_delta,
1158 struct ib_device *dev = target->srp_host->srp_dev->dev;
1159 unsigned long flags;
1163 spin_lock_irqsave(&target->lock, flags);
1164 target->req_lim += req_delta;
1165 iu = __srp_get_tx_iu(target, SRP_IU_RSP);
1166 spin_unlock_irqrestore(&target->lock, flags);
1169 shost_printk(KERN_ERR, target->scsi_host, PFX
1170 "no IU available to send response\n");
1174 ib_dma_sync_single_for_cpu(dev, iu->dma, len, DMA_TO_DEVICE);
1175 memcpy(iu->buf, rsp, len);
1176 ib_dma_sync_single_for_device(dev, iu->dma, len, DMA_TO_DEVICE);
1178 err = srp_post_send(target, iu, len);
1180 shost_printk(KERN_ERR, target->scsi_host, PFX
1181 "unable to post response: %d\n", err);
1182 srp_put_tx_iu(target, iu, SRP_IU_RSP);
1188 static void srp_process_cred_req(struct srp_target_port *target,
1189 struct srp_cred_req *req)
1191 struct srp_cred_rsp rsp = {
1192 .opcode = SRP_CRED_RSP,
1195 s32 delta = be32_to_cpu(req->req_lim_delta);
1197 if (srp_response_common(target, delta, &rsp, sizeof rsp))
1198 shost_printk(KERN_ERR, target->scsi_host, PFX
1199 "problems processing SRP_CRED_REQ\n");
1202 static void srp_process_aer_req(struct srp_target_port *target,
1203 struct srp_aer_req *req)
1205 struct srp_aer_rsp rsp = {
1206 .opcode = SRP_AER_RSP,
1209 s32 delta = be32_to_cpu(req->req_lim_delta);
1211 shost_printk(KERN_ERR, target->scsi_host, PFX
1212 "ignoring AER for LUN %llu\n", be64_to_cpu(req->lun));
1214 if (srp_response_common(target, delta, &rsp, sizeof rsp))
1215 shost_printk(KERN_ERR, target->scsi_host, PFX
1216 "problems processing SRP_AER_REQ\n");
1219 static void srp_handle_recv(struct srp_target_port *target, struct ib_wc *wc)
1221 struct ib_device *dev = target->srp_host->srp_dev->dev;
1222 struct srp_iu *iu = (struct srp_iu *) (uintptr_t) wc->wr_id;
1226 ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_ti_iu_len,
1229 opcode = *(u8 *) iu->buf;
1232 shost_printk(KERN_ERR, target->scsi_host,
1233 PFX "recv completion, opcode 0x%02x\n", opcode);
1234 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 8, 1,
1235 iu->buf, wc->byte_len, true);
1240 srp_process_rsp(target, iu->buf);
1244 srp_process_cred_req(target, iu->buf);
1248 srp_process_aer_req(target, iu->buf);
1252 /* XXX Handle target logout */
1253 shost_printk(KERN_WARNING, target->scsi_host,
1254 PFX "Got target logout request\n");
1258 shost_printk(KERN_WARNING, target->scsi_host,
1259 PFX "Unhandled SRP opcode 0x%02x\n", opcode);
1263 ib_dma_sync_single_for_device(dev, iu->dma, target->max_ti_iu_len,
1266 res = srp_post_recv(target, iu);
1268 shost_printk(KERN_ERR, target->scsi_host,
1269 PFX "Recv failed with error code %d\n", res);
1272 static void srp_recv_completion(struct ib_cq *cq, void *target_ptr)
1274 struct srp_target_port *target = target_ptr;
1277 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1278 while (ib_poll_cq(cq, 1, &wc) > 0) {
1280 shost_printk(KERN_ERR, target->scsi_host,
1281 PFX "failed receive status %d\n",
1283 target->qp_in_error = 1;
1287 srp_handle_recv(target, &wc);
1291 static void srp_send_completion(struct ib_cq *cq, void *target_ptr)
1293 struct srp_target_port *target = target_ptr;
1297 while (ib_poll_cq(cq, 1, &wc) > 0) {
1299 shost_printk(KERN_ERR, target->scsi_host,
1300 PFX "failed send status %d\n",
1302 target->qp_in_error = 1;
1306 iu = (struct srp_iu *) (uintptr_t) wc.wr_id;
1307 list_add(&iu->list, &target->free_tx);
1311 static int srp_queuecommand(struct Scsi_Host *shost, struct scsi_cmnd *scmnd)
1313 struct srp_target_port *target = host_to_target(shost);
1314 struct srp_request *req;
1316 struct srp_cmd *cmd;
1317 struct ib_device *dev;
1318 unsigned long flags;
1321 spin_lock_irqsave(&target->lock, flags);
1322 iu = __srp_get_tx_iu(target, SRP_IU_CMD);
1326 req = list_first_entry(&target->free_reqs, struct srp_request, list);
1327 list_del(&req->list);
1328 spin_unlock_irqrestore(&target->lock, flags);
1330 dev = target->srp_host->srp_dev->dev;
1331 ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_iu_len,
1335 scmnd->host_scribble = (void *) req;
1338 memset(cmd, 0, sizeof *cmd);
1340 cmd->opcode = SRP_CMD;
1341 cmd->lun = cpu_to_be64((u64) scmnd->device->lun << 48);
1342 cmd->tag = req->index;
1343 memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
1348 len = srp_map_data(scmnd, target, req);
1350 shost_printk(KERN_ERR, target->scsi_host,
1351 PFX "Failed to map data\n");
1355 ib_dma_sync_single_for_device(dev, iu->dma, target->max_iu_len,
1358 if (srp_post_send(target, iu, len)) {
1359 shost_printk(KERN_ERR, target->scsi_host, PFX "Send failed\n");
1366 srp_unmap_data(scmnd, target, req);
1369 srp_put_tx_iu(target, iu, SRP_IU_CMD);
1371 spin_lock_irqsave(&target->lock, flags);
1372 list_add(&req->list, &target->free_reqs);
1375 spin_unlock_irqrestore(&target->lock, flags);
1377 return SCSI_MLQUEUE_HOST_BUSY;
1380 static int srp_alloc_iu_bufs(struct srp_target_port *target)
1384 for (i = 0; i < SRP_RQ_SIZE; ++i) {
1385 target->rx_ring[i] = srp_alloc_iu(target->srp_host,
1386 target->max_ti_iu_len,
1387 GFP_KERNEL, DMA_FROM_DEVICE);
1388 if (!target->rx_ring[i])
1392 for (i = 0; i < SRP_SQ_SIZE; ++i) {
1393 target->tx_ring[i] = srp_alloc_iu(target->srp_host,
1395 GFP_KERNEL, DMA_TO_DEVICE);
1396 if (!target->tx_ring[i])
1399 list_add(&target->tx_ring[i]->list, &target->free_tx);
1405 for (i = 0; i < SRP_RQ_SIZE; ++i) {
1406 srp_free_iu(target->srp_host, target->rx_ring[i]);
1407 target->rx_ring[i] = NULL;
1410 for (i = 0; i < SRP_SQ_SIZE; ++i) {
1411 srp_free_iu(target->srp_host, target->tx_ring[i]);
1412 target->tx_ring[i] = NULL;
1418 static uint32_t srp_compute_rq_tmo(struct ib_qp_attr *qp_attr, int attr_mask)
1420 uint64_t T_tr_ns, max_compl_time_ms;
1421 uint32_t rq_tmo_jiffies;
1424 * According to section 11.2.4.2 in the IBTA spec (Modify Queue Pair,
1425 * table 91), both the QP timeout and the retry count have to be set
1426 * for RC QP's during the RTR to RTS transition.
1428 WARN_ON_ONCE((attr_mask & (IB_QP_TIMEOUT | IB_QP_RETRY_CNT)) !=
1429 (IB_QP_TIMEOUT | IB_QP_RETRY_CNT));
1432 * Set target->rq_tmo_jiffies to one second more than the largest time
1433 * it can take before an error completion is generated. See also
1434 * C9-140..142 in the IBTA spec for more information about how to
1435 * convert the QP Local ACK Timeout value to nanoseconds.
1437 T_tr_ns = 4096 * (1ULL << qp_attr->timeout);
1438 max_compl_time_ms = qp_attr->retry_cnt * 4 * T_tr_ns;
1439 do_div(max_compl_time_ms, NSEC_PER_MSEC);
1440 rq_tmo_jiffies = msecs_to_jiffies(max_compl_time_ms + 1000);
1442 return rq_tmo_jiffies;
1445 static void srp_cm_rep_handler(struct ib_cm_id *cm_id,
1446 struct srp_login_rsp *lrsp,
1447 struct srp_target_port *target)
1449 struct ib_qp_attr *qp_attr = NULL;
1454 if (lrsp->opcode == SRP_LOGIN_RSP) {
1455 target->max_ti_iu_len = be32_to_cpu(lrsp->max_ti_iu_len);
1456 target->req_lim = be32_to_cpu(lrsp->req_lim_delta);
1459 * Reserve credits for task management so we don't
1460 * bounce requests back to the SCSI mid-layer.
1462 target->scsi_host->can_queue
1463 = min(target->req_lim - SRP_TSK_MGMT_SQ_SIZE,
1464 target->scsi_host->can_queue);
1466 shost_printk(KERN_WARNING, target->scsi_host,
1467 PFX "Unhandled RSP opcode %#x\n", lrsp->opcode);
1472 if (!target->rx_ring[0]) {
1473 ret = srp_alloc_iu_bufs(target);
1479 qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
1483 qp_attr->qp_state = IB_QPS_RTR;
1484 ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1488 ret = ib_modify_qp(target->qp, qp_attr, attr_mask);
1492 for (i = 0; i < SRP_RQ_SIZE; i++) {
1493 struct srp_iu *iu = target->rx_ring[i];
1494 ret = srp_post_recv(target, iu);
1499 qp_attr->qp_state = IB_QPS_RTS;
1500 ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
1504 target->rq_tmo_jiffies = srp_compute_rq_tmo(qp_attr, attr_mask);
1506 ret = ib_modify_qp(target->qp, qp_attr, attr_mask);
1510 ret = ib_send_cm_rtu(cm_id, NULL, 0);
1516 target->status = ret;
1519 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
1520 struct ib_cm_event *event,
1521 struct srp_target_port *target)
1523 struct Scsi_Host *shost = target->scsi_host;
1524 struct ib_class_port_info *cpi;
1527 switch (event->param.rej_rcvd.reason) {
1528 case IB_CM_REJ_PORT_CM_REDIRECT:
1529 cpi = event->param.rej_rcvd.ari;
1530 target->path.dlid = cpi->redirect_lid;
1531 target->path.pkey = cpi->redirect_pkey;
1532 cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
1533 memcpy(target->path.dgid.raw, cpi->redirect_gid, 16);
1535 target->status = target->path.dlid ?
1536 SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
1539 case IB_CM_REJ_PORT_REDIRECT:
1540 if (srp_target_is_topspin(target)) {
1542 * Topspin/Cisco SRP gateways incorrectly send
1543 * reject reason code 25 when they mean 24
1546 memcpy(target->path.dgid.raw,
1547 event->param.rej_rcvd.ari, 16);
1549 shost_printk(KERN_DEBUG, shost,
1550 PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
1551 (unsigned long long) be64_to_cpu(target->path.dgid.global.subnet_prefix),
1552 (unsigned long long) be64_to_cpu(target->path.dgid.global.interface_id));
1554 target->status = SRP_PORT_REDIRECT;
1556 shost_printk(KERN_WARNING, shost,
1557 " REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
1558 target->status = -ECONNRESET;
1562 case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
1563 shost_printk(KERN_WARNING, shost,
1564 " REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
1565 target->status = -ECONNRESET;
1568 case IB_CM_REJ_CONSUMER_DEFINED:
1569 opcode = *(u8 *) event->private_data;
1570 if (opcode == SRP_LOGIN_REJ) {
1571 struct srp_login_rej *rej = event->private_data;
1572 u32 reason = be32_to_cpu(rej->reason);
1574 if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
1575 shost_printk(KERN_WARNING, shost,
1576 PFX "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
1578 shost_printk(KERN_WARNING, shost,
1579 PFX "SRP LOGIN REJECTED, reason 0x%08x\n", reason);
1581 shost_printk(KERN_WARNING, shost,
1582 " REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
1583 " opcode 0x%02x\n", opcode);
1584 target->status = -ECONNRESET;
1587 case IB_CM_REJ_STALE_CONN:
1588 shost_printk(KERN_WARNING, shost, " REJ reason: stale connection\n");
1589 target->status = SRP_STALE_CONN;
1593 shost_printk(KERN_WARNING, shost, " REJ reason 0x%x\n",
1594 event->param.rej_rcvd.reason);
1595 target->status = -ECONNRESET;
1599 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
1601 struct srp_target_port *target = cm_id->context;
1604 switch (event->event) {
1605 case IB_CM_REQ_ERROR:
1606 shost_printk(KERN_DEBUG, target->scsi_host,
1607 PFX "Sending CM REQ failed\n");
1609 target->status = -ECONNRESET;
1612 case IB_CM_REP_RECEIVED:
1614 srp_cm_rep_handler(cm_id, event->private_data, target);
1617 case IB_CM_REJ_RECEIVED:
1618 shost_printk(KERN_DEBUG, target->scsi_host, PFX "REJ received\n");
1621 srp_cm_rej_handler(cm_id, event, target);
1624 case IB_CM_DREQ_RECEIVED:
1625 shost_printk(KERN_WARNING, target->scsi_host,
1626 PFX "DREQ received - connection closed\n");
1627 if (ib_send_cm_drep(cm_id, NULL, 0))
1628 shost_printk(KERN_ERR, target->scsi_host,
1629 PFX "Sending CM DREP failed\n");
1632 case IB_CM_TIMEWAIT_EXIT:
1633 shost_printk(KERN_ERR, target->scsi_host,
1634 PFX "connection closed\n");
1640 case IB_CM_MRA_RECEIVED:
1641 case IB_CM_DREQ_ERROR:
1642 case IB_CM_DREP_RECEIVED:
1646 shost_printk(KERN_WARNING, target->scsi_host,
1647 PFX "Unhandled CM event %d\n", event->event);
1652 complete(&target->done);
1657 static int srp_send_tsk_mgmt(struct srp_target_port *target,
1658 u64 req_tag, unsigned int lun, u8 func)
1660 struct ib_device *dev = target->srp_host->srp_dev->dev;
1662 struct srp_tsk_mgmt *tsk_mgmt;
1664 init_completion(&target->tsk_mgmt_done);
1666 spin_lock_irq(&target->lock);
1667 iu = __srp_get_tx_iu(target, SRP_IU_TSK_MGMT);
1668 spin_unlock_irq(&target->lock);
1673 ib_dma_sync_single_for_cpu(dev, iu->dma, sizeof *tsk_mgmt,
1676 memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
1678 tsk_mgmt->opcode = SRP_TSK_MGMT;
1679 tsk_mgmt->lun = cpu_to_be64((u64) lun << 48);
1680 tsk_mgmt->tag = req_tag | SRP_TAG_TSK_MGMT;
1681 tsk_mgmt->tsk_mgmt_func = func;
1682 tsk_mgmt->task_tag = req_tag;
1684 ib_dma_sync_single_for_device(dev, iu->dma, sizeof *tsk_mgmt,
1686 if (srp_post_send(target, iu, sizeof *tsk_mgmt)) {
1687 srp_put_tx_iu(target, iu, SRP_IU_TSK_MGMT);
1691 if (!wait_for_completion_timeout(&target->tsk_mgmt_done,
1692 msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
1698 static int srp_abort(struct scsi_cmnd *scmnd)
1700 struct srp_target_port *target = host_to_target(scmnd->device->host);
1701 struct srp_request *req = (struct srp_request *) scmnd->host_scribble;
1703 shost_printk(KERN_ERR, target->scsi_host, "SRP abort called\n");
1705 if (!req || target->qp_in_error || !srp_claim_req(target, req, scmnd))
1707 srp_send_tsk_mgmt(target, req->index, scmnd->device->lun,
1708 SRP_TSK_ABORT_TASK);
1709 srp_free_req(target, req, scmnd, 0);
1710 scmnd->result = DID_ABORT << 16;
1711 scmnd->scsi_done(scmnd);
1716 static int srp_reset_device(struct scsi_cmnd *scmnd)
1718 struct srp_target_port *target = host_to_target(scmnd->device->host);
1721 shost_printk(KERN_ERR, target->scsi_host, "SRP reset_device called\n");
1723 if (target->qp_in_error)
1725 if (srp_send_tsk_mgmt(target, SRP_TAG_NO_REQ, scmnd->device->lun,
1728 if (target->tsk_mgmt_status)
1731 for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
1732 struct srp_request *req = &target->req_ring[i];
1733 if (req->scmnd && req->scmnd->device == scmnd->device)
1734 srp_reset_req(target, req);
1740 static int srp_reset_host(struct scsi_cmnd *scmnd)
1742 struct srp_target_port *target = host_to_target(scmnd->device->host);
1745 shost_printk(KERN_ERR, target->scsi_host, PFX "SRP reset_host called\n");
1747 if (!srp_reconnect_target(target))
1753 static int srp_slave_configure(struct scsi_device *sdev)
1755 struct Scsi_Host *shost = sdev->host;
1756 struct srp_target_port *target = host_to_target(shost);
1757 struct request_queue *q = sdev->request_queue;
1758 unsigned long timeout;
1760 if (sdev->type == TYPE_DISK) {
1761 timeout = max_t(unsigned, 30 * HZ, target->rq_tmo_jiffies);
1762 blk_queue_rq_timeout(q, timeout);
1768 static ssize_t show_id_ext(struct device *dev, struct device_attribute *attr,
1771 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1773 return sprintf(buf, "0x%016llx\n",
1774 (unsigned long long) be64_to_cpu(target->id_ext));
1777 static ssize_t show_ioc_guid(struct device *dev, struct device_attribute *attr,
1780 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1782 return sprintf(buf, "0x%016llx\n",
1783 (unsigned long long) be64_to_cpu(target->ioc_guid));
1786 static ssize_t show_service_id(struct device *dev,
1787 struct device_attribute *attr, char *buf)
1789 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1791 return sprintf(buf, "0x%016llx\n",
1792 (unsigned long long) be64_to_cpu(target->service_id));
1795 static ssize_t show_pkey(struct device *dev, struct device_attribute *attr,
1798 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1800 return sprintf(buf, "0x%04x\n", be16_to_cpu(target->path.pkey));
1803 static ssize_t show_dgid(struct device *dev, struct device_attribute *attr,
1806 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1808 return sprintf(buf, "%pI6\n", target->path.dgid.raw);
1811 static ssize_t show_orig_dgid(struct device *dev,
1812 struct device_attribute *attr, char *buf)
1814 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1816 return sprintf(buf, "%pI6\n", target->orig_dgid);
1819 static ssize_t show_req_lim(struct device *dev,
1820 struct device_attribute *attr, char *buf)
1822 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1824 return sprintf(buf, "%d\n", target->req_lim);
1827 static ssize_t show_zero_req_lim(struct device *dev,
1828 struct device_attribute *attr, char *buf)
1830 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1832 return sprintf(buf, "%d\n", target->zero_req_lim);
1835 static ssize_t show_local_ib_port(struct device *dev,
1836 struct device_attribute *attr, char *buf)
1838 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1840 return sprintf(buf, "%d\n", target->srp_host->port);
1843 static ssize_t show_local_ib_device(struct device *dev,
1844 struct device_attribute *attr, char *buf)
1846 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1848 return sprintf(buf, "%s\n", target->srp_host->srp_dev->dev->name);
1851 static ssize_t show_cmd_sg_entries(struct device *dev,
1852 struct device_attribute *attr, char *buf)
1854 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1856 return sprintf(buf, "%u\n", target->cmd_sg_cnt);
1859 static ssize_t show_allow_ext_sg(struct device *dev,
1860 struct device_attribute *attr, char *buf)
1862 struct srp_target_port *target = host_to_target(class_to_shost(dev));
1864 return sprintf(buf, "%s\n", target->allow_ext_sg ? "true" : "false");
1867 static DEVICE_ATTR(id_ext, S_IRUGO, show_id_ext, NULL);
1868 static DEVICE_ATTR(ioc_guid, S_IRUGO, show_ioc_guid, NULL);
1869 static DEVICE_ATTR(service_id, S_IRUGO, show_service_id, NULL);
1870 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
1871 static DEVICE_ATTR(dgid, S_IRUGO, show_dgid, NULL);
1872 static DEVICE_ATTR(orig_dgid, S_IRUGO, show_orig_dgid, NULL);
1873 static DEVICE_ATTR(req_lim, S_IRUGO, show_req_lim, NULL);
1874 static DEVICE_ATTR(zero_req_lim, S_IRUGO, show_zero_req_lim, NULL);
1875 static DEVICE_ATTR(local_ib_port, S_IRUGO, show_local_ib_port, NULL);
1876 static DEVICE_ATTR(local_ib_device, S_IRUGO, show_local_ib_device, NULL);
1877 static DEVICE_ATTR(cmd_sg_entries, S_IRUGO, show_cmd_sg_entries, NULL);
1878 static DEVICE_ATTR(allow_ext_sg, S_IRUGO, show_allow_ext_sg, NULL);
1880 static struct device_attribute *srp_host_attrs[] = {
1883 &dev_attr_service_id,
1886 &dev_attr_orig_dgid,
1888 &dev_attr_zero_req_lim,
1889 &dev_attr_local_ib_port,
1890 &dev_attr_local_ib_device,
1891 &dev_attr_cmd_sg_entries,
1892 &dev_attr_allow_ext_sg,
1896 static struct scsi_host_template srp_template = {
1897 .module = THIS_MODULE,
1898 .name = "InfiniBand SRP initiator",
1899 .proc_name = DRV_NAME,
1900 .slave_configure = srp_slave_configure,
1901 .info = srp_target_info,
1902 .queuecommand = srp_queuecommand,
1903 .eh_abort_handler = srp_abort,
1904 .eh_device_reset_handler = srp_reset_device,
1905 .eh_host_reset_handler = srp_reset_host,
1906 .sg_tablesize = SRP_DEF_SG_TABLESIZE,
1907 .can_queue = SRP_CMD_SQ_SIZE,
1909 .cmd_per_lun = SRP_CMD_SQ_SIZE,
1910 .use_clustering = ENABLE_CLUSTERING,
1911 .shost_attrs = srp_host_attrs
1914 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
1916 struct srp_rport_identifiers ids;
1917 struct srp_rport *rport;
1919 sprintf(target->target_name, "SRP.T10:%016llX",
1920 (unsigned long long) be64_to_cpu(target->id_ext));
1922 if (scsi_add_host(target->scsi_host, host->srp_dev->dev->dma_device))
1925 memcpy(ids.port_id, &target->id_ext, 8);
1926 memcpy(ids.port_id + 8, &target->ioc_guid, 8);
1927 ids.roles = SRP_RPORT_ROLE_TARGET;
1928 rport = srp_rport_add(target->scsi_host, &ids);
1929 if (IS_ERR(rport)) {
1930 scsi_remove_host(target->scsi_host);
1931 return PTR_ERR(rport);
1934 spin_lock(&host->target_lock);
1935 list_add_tail(&target->list, &host->target_list);
1936 spin_unlock(&host->target_lock);
1938 target->state = SRP_TARGET_LIVE;
1940 scsi_scan_target(&target->scsi_host->shost_gendev,
1941 0, target->scsi_id, SCAN_WILD_CARD, 0);
1946 static void srp_release_dev(struct device *dev)
1948 struct srp_host *host =
1949 container_of(dev, struct srp_host, dev);
1951 complete(&host->released);
1954 static struct class srp_class = {
1955 .name = "infiniband_srp",
1956 .dev_release = srp_release_dev
1960 * Target ports are added by writing
1962 * id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
1963 * pkey=<P_Key>,service_id=<service ID>
1965 * to the add_target sysfs attribute.
1969 SRP_OPT_ID_EXT = 1 << 0,
1970 SRP_OPT_IOC_GUID = 1 << 1,
1971 SRP_OPT_DGID = 1 << 2,
1972 SRP_OPT_PKEY = 1 << 3,
1973 SRP_OPT_SERVICE_ID = 1 << 4,
1974 SRP_OPT_MAX_SECT = 1 << 5,
1975 SRP_OPT_MAX_CMD_PER_LUN = 1 << 6,
1976 SRP_OPT_IO_CLASS = 1 << 7,
1977 SRP_OPT_INITIATOR_EXT = 1 << 8,
1978 SRP_OPT_CMD_SG_ENTRIES = 1 << 9,
1979 SRP_OPT_ALLOW_EXT_SG = 1 << 10,
1980 SRP_OPT_SG_TABLESIZE = 1 << 11,
1981 SRP_OPT_ALL = (SRP_OPT_ID_EXT |
1985 SRP_OPT_SERVICE_ID),
1988 static const match_table_t srp_opt_tokens = {
1989 { SRP_OPT_ID_EXT, "id_ext=%s" },
1990 { SRP_OPT_IOC_GUID, "ioc_guid=%s" },
1991 { SRP_OPT_DGID, "dgid=%s" },
1992 { SRP_OPT_PKEY, "pkey=%x" },
1993 { SRP_OPT_SERVICE_ID, "service_id=%s" },
1994 { SRP_OPT_MAX_SECT, "max_sect=%d" },
1995 { SRP_OPT_MAX_CMD_PER_LUN, "max_cmd_per_lun=%d" },
1996 { SRP_OPT_IO_CLASS, "io_class=%x" },
1997 { SRP_OPT_INITIATOR_EXT, "initiator_ext=%s" },
1998 { SRP_OPT_CMD_SG_ENTRIES, "cmd_sg_entries=%u" },
1999 { SRP_OPT_ALLOW_EXT_SG, "allow_ext_sg=%u" },
2000 { SRP_OPT_SG_TABLESIZE, "sg_tablesize=%u" },
2001 { SRP_OPT_ERR, NULL }
2004 static int srp_parse_options(const char *buf, struct srp_target_port *target)
2006 char *options, *sep_opt;
2009 substring_t args[MAX_OPT_ARGS];
2015 options = kstrdup(buf, GFP_KERNEL);
2020 while ((p = strsep(&sep_opt, ",")) != NULL) {
2024 token = match_token(p, srp_opt_tokens, args);
2028 case SRP_OPT_ID_EXT:
2029 p = match_strdup(args);
2034 target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2038 case SRP_OPT_IOC_GUID:
2039 p = match_strdup(args);
2044 target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
2049 p = match_strdup(args);
2054 if (strlen(p) != 32) {
2055 pr_warn("bad dest GID parameter '%s'\n", p);
2060 for (i = 0; i < 16; ++i) {
2061 strlcpy(dgid, p + i * 2, 3);
2062 target->path.dgid.raw[i] = simple_strtoul(dgid, NULL, 16);
2065 memcpy(target->orig_dgid, target->path.dgid.raw, 16);
2069 if (match_hex(args, &token)) {
2070 pr_warn("bad P_Key parameter '%s'\n", p);
2073 target->path.pkey = cpu_to_be16(token);
2076 case SRP_OPT_SERVICE_ID:
2077 p = match_strdup(args);
2082 target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
2083 target->path.service_id = target->service_id;
2087 case SRP_OPT_MAX_SECT:
2088 if (match_int(args, &token)) {
2089 pr_warn("bad max sect parameter '%s'\n", p);
2092 target->scsi_host->max_sectors = token;
2095 case SRP_OPT_MAX_CMD_PER_LUN:
2096 if (match_int(args, &token)) {
2097 pr_warn("bad max cmd_per_lun parameter '%s'\n",
2101 target->scsi_host->cmd_per_lun = min(token, SRP_CMD_SQ_SIZE);
2104 case SRP_OPT_IO_CLASS:
2105 if (match_hex(args, &token)) {
2106 pr_warn("bad IO class parameter '%s'\n", p);
2109 if (token != SRP_REV10_IB_IO_CLASS &&
2110 token != SRP_REV16A_IB_IO_CLASS) {
2111 pr_warn("unknown IO class parameter value %x specified (use %x or %x).\n",
2112 token, SRP_REV10_IB_IO_CLASS,
2113 SRP_REV16A_IB_IO_CLASS);
2116 target->io_class = token;
2119 case SRP_OPT_INITIATOR_EXT:
2120 p = match_strdup(args);
2125 target->initiator_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2129 case SRP_OPT_CMD_SG_ENTRIES:
2130 if (match_int(args, &token) || token < 1 || token > 255) {
2131 pr_warn("bad max cmd_sg_entries parameter '%s'\n",
2135 target->cmd_sg_cnt = token;
2138 case SRP_OPT_ALLOW_EXT_SG:
2139 if (match_int(args, &token)) {
2140 pr_warn("bad allow_ext_sg parameter '%s'\n", p);
2143 target->allow_ext_sg = !!token;
2146 case SRP_OPT_SG_TABLESIZE:
2147 if (match_int(args, &token) || token < 1 ||
2148 token > SCSI_MAX_SG_CHAIN_SEGMENTS) {
2149 pr_warn("bad max sg_tablesize parameter '%s'\n",
2153 target->sg_tablesize = token;
2157 pr_warn("unknown parameter or missing value '%s' in target creation request\n",
2163 if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
2166 for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
2167 if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
2168 !(srp_opt_tokens[i].token & opt_mask))
2169 pr_warn("target creation request is missing parameter '%s'\n",
2170 srp_opt_tokens[i].pattern);
2177 static ssize_t srp_create_target(struct device *dev,
2178 struct device_attribute *attr,
2179 const char *buf, size_t count)
2181 struct srp_host *host =
2182 container_of(dev, struct srp_host, dev);
2183 struct Scsi_Host *target_host;
2184 struct srp_target_port *target;
2185 struct ib_device *ibdev = host->srp_dev->dev;
2186 dma_addr_t dma_addr;
2189 target_host = scsi_host_alloc(&srp_template,
2190 sizeof (struct srp_target_port));
2194 target_host->transportt = ib_srp_transport_template;
2195 target_host->max_channel = 0;
2196 target_host->max_id = 1;
2197 target_host->max_lun = SRP_MAX_LUN;
2198 target_host->max_cmd_len = sizeof ((struct srp_cmd *) (void *) 0L)->cdb;
2200 target = host_to_target(target_host);
2202 target->io_class = SRP_REV16A_IB_IO_CLASS;
2203 target->scsi_host = target_host;
2204 target->srp_host = host;
2205 target->lkey = host->srp_dev->mr->lkey;
2206 target->rkey = host->srp_dev->mr->rkey;
2207 target->cmd_sg_cnt = cmd_sg_entries;
2208 target->sg_tablesize = indirect_sg_entries ? : cmd_sg_entries;
2209 target->allow_ext_sg = allow_ext_sg;
2211 ret = srp_parse_options(buf, target);
2215 if (!host->srp_dev->fmr_pool && !target->allow_ext_sg &&
2216 target->cmd_sg_cnt < target->sg_tablesize) {
2217 pr_warn("No FMR pool and no external indirect descriptors, limiting sg_tablesize to cmd_sg_cnt\n");
2218 target->sg_tablesize = target->cmd_sg_cnt;
2221 target_host->sg_tablesize = target->sg_tablesize;
2222 target->indirect_size = target->sg_tablesize *
2223 sizeof (struct srp_direct_buf);
2224 target->max_iu_len = sizeof (struct srp_cmd) +
2225 sizeof (struct srp_indirect_buf) +
2226 target->cmd_sg_cnt * sizeof (struct srp_direct_buf);
2228 spin_lock_init(&target->lock);
2229 INIT_LIST_HEAD(&target->free_tx);
2230 INIT_LIST_HEAD(&target->free_reqs);
2231 for (i = 0; i < SRP_CMD_SQ_SIZE; ++i) {
2232 struct srp_request *req = &target->req_ring[i];
2234 req->fmr_list = kmalloc(target->cmd_sg_cnt * sizeof (void *),
2236 req->map_page = kmalloc(SRP_FMR_SIZE * sizeof (void *),
2238 req->indirect_desc = kmalloc(target->indirect_size, GFP_KERNEL);
2239 if (!req->fmr_list || !req->map_page || !req->indirect_desc)
2242 dma_addr = ib_dma_map_single(ibdev, req->indirect_desc,
2243 target->indirect_size,
2245 if (ib_dma_mapping_error(ibdev, dma_addr))
2248 req->indirect_dma_addr = dma_addr;
2250 list_add_tail(&req->list, &target->free_reqs);
2253 ib_query_gid(ibdev, host->port, 0, &target->path.sgid);
2255 shost_printk(KERN_DEBUG, target->scsi_host, PFX
2256 "new target: id_ext %016llx ioc_guid %016llx pkey %04x "
2257 "service_id %016llx dgid %pI6\n",
2258 (unsigned long long) be64_to_cpu(target->id_ext),
2259 (unsigned long long) be64_to_cpu(target->ioc_guid),
2260 be16_to_cpu(target->path.pkey),
2261 (unsigned long long) be64_to_cpu(target->service_id),
2262 target->path.dgid.raw);
2264 ret = srp_create_target_ib(target);
2268 ret = srp_new_cm_id(target);
2272 target->qp_in_error = 0;
2273 ret = srp_connect_target(target);
2275 shost_printk(KERN_ERR, target->scsi_host,
2276 PFX "Connection failed\n");
2280 ret = srp_add_target(host, target);
2282 goto err_disconnect;
2287 srp_disconnect_target(target);
2290 ib_destroy_cm_id(target->cm_id);
2293 srp_free_target_ib(target);
2296 srp_free_req_data(target);
2299 scsi_host_put(target_host);
2304 static DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
2306 static ssize_t show_ibdev(struct device *dev, struct device_attribute *attr,
2309 struct srp_host *host = container_of(dev, struct srp_host, dev);
2311 return sprintf(buf, "%s\n", host->srp_dev->dev->name);
2314 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
2316 static ssize_t show_port(struct device *dev, struct device_attribute *attr,
2319 struct srp_host *host = container_of(dev, struct srp_host, dev);
2321 return sprintf(buf, "%d\n", host->port);
2324 static DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
2326 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
2328 struct srp_host *host;
2330 host = kzalloc(sizeof *host, GFP_KERNEL);
2334 INIT_LIST_HEAD(&host->target_list);
2335 spin_lock_init(&host->target_lock);
2336 init_completion(&host->released);
2337 host->srp_dev = device;
2340 host->dev.class = &srp_class;
2341 host->dev.parent = device->dev->dma_device;
2342 dev_set_name(&host->dev, "srp-%s-%d", device->dev->name, port);
2344 if (device_register(&host->dev))
2346 if (device_create_file(&host->dev, &dev_attr_add_target))
2348 if (device_create_file(&host->dev, &dev_attr_ibdev))
2350 if (device_create_file(&host->dev, &dev_attr_port))
2356 device_unregister(&host->dev);
2364 static void srp_add_one(struct ib_device *device)
2366 struct srp_device *srp_dev;
2367 struct ib_device_attr *dev_attr;
2368 struct ib_fmr_pool_param fmr_param;
2369 struct srp_host *host;
2370 int max_pages_per_fmr, fmr_page_shift, s, e, p;
2372 dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
2376 if (ib_query_device(device, dev_attr)) {
2377 pr_warn("Query device failed for %s\n", device->name);
2381 srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
2386 * Use the smallest page size supported by the HCA, down to a
2387 * minimum of 4096 bytes. We're unlikely to build large sglists
2388 * out of smaller entries.
2390 fmr_page_shift = max(12, ffs(dev_attr->page_size_cap) - 1);
2391 srp_dev->fmr_page_size = 1 << fmr_page_shift;
2392 srp_dev->fmr_page_mask = ~((u64) srp_dev->fmr_page_size - 1);
2393 srp_dev->fmr_max_size = srp_dev->fmr_page_size * SRP_FMR_SIZE;
2395 INIT_LIST_HEAD(&srp_dev->dev_list);
2397 srp_dev->dev = device;
2398 srp_dev->pd = ib_alloc_pd(device);
2399 if (IS_ERR(srp_dev->pd))
2402 srp_dev->mr = ib_get_dma_mr(srp_dev->pd,
2403 IB_ACCESS_LOCAL_WRITE |
2404 IB_ACCESS_REMOTE_READ |
2405 IB_ACCESS_REMOTE_WRITE);
2406 if (IS_ERR(srp_dev->mr))
2409 for (max_pages_per_fmr = SRP_FMR_SIZE;
2410 max_pages_per_fmr >= SRP_FMR_MIN_SIZE;
2411 max_pages_per_fmr /= 2, srp_dev->fmr_max_size /= 2) {
2412 memset(&fmr_param, 0, sizeof fmr_param);
2413 fmr_param.pool_size = SRP_FMR_POOL_SIZE;
2414 fmr_param.dirty_watermark = SRP_FMR_DIRTY_SIZE;
2415 fmr_param.cache = 1;
2416 fmr_param.max_pages_per_fmr = max_pages_per_fmr;
2417 fmr_param.page_shift = fmr_page_shift;
2418 fmr_param.access = (IB_ACCESS_LOCAL_WRITE |
2419 IB_ACCESS_REMOTE_WRITE |
2420 IB_ACCESS_REMOTE_READ);
2422 srp_dev->fmr_pool = ib_create_fmr_pool(srp_dev->pd, &fmr_param);
2423 if (!IS_ERR(srp_dev->fmr_pool))
2427 if (IS_ERR(srp_dev->fmr_pool))
2428 srp_dev->fmr_pool = NULL;
2430 if (device->node_type == RDMA_NODE_IB_SWITCH) {
2435 e = device->phys_port_cnt;
2438 for (p = s; p <= e; ++p) {
2439 host = srp_add_port(srp_dev, p);
2441 list_add_tail(&host->list, &srp_dev->dev_list);
2444 ib_set_client_data(device, &srp_client, srp_dev);
2449 ib_dealloc_pd(srp_dev->pd);
2458 static void srp_remove_one(struct ib_device *device)
2460 struct srp_device *srp_dev;
2461 struct srp_host *host, *tmp_host;
2462 LIST_HEAD(target_list);
2463 struct srp_target_port *target, *tmp_target;
2465 srp_dev = ib_get_client_data(device, &srp_client);
2467 list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
2468 device_unregister(&host->dev);
2470 * Wait for the sysfs entry to go away, so that no new
2471 * target ports can be created.
2473 wait_for_completion(&host->released);
2476 * Mark all target ports as removed, so we stop queueing
2477 * commands and don't try to reconnect.
2479 spin_lock(&host->target_lock);
2480 list_for_each_entry(target, &host->target_list, list) {
2481 spin_lock_irq(&target->lock);
2482 target->state = SRP_TARGET_REMOVED;
2483 spin_unlock_irq(&target->lock);
2485 spin_unlock(&host->target_lock);
2488 * Wait for any reconnection tasks that may have
2489 * started before we marked our target ports as
2490 * removed, and any target port removal tasks.
2492 flush_workqueue(ib_wq);
2494 list_for_each_entry_safe(target, tmp_target,
2495 &host->target_list, list) {
2496 srp_del_scsi_host_attr(target->scsi_host);
2497 srp_remove_host(target->scsi_host);
2498 scsi_remove_host(target->scsi_host);
2499 srp_disconnect_target(target);
2500 ib_destroy_cm_id(target->cm_id);
2501 srp_free_target_ib(target);
2502 srp_free_req_data(target);
2503 scsi_host_put(target->scsi_host);
2509 if (srp_dev->fmr_pool)
2510 ib_destroy_fmr_pool(srp_dev->fmr_pool);
2511 ib_dereg_mr(srp_dev->mr);
2512 ib_dealloc_pd(srp_dev->pd);
2517 static struct srp_function_template ib_srp_transport_functions = {
2520 static int __init srp_init_module(void)
2524 BUILD_BUG_ON(FIELD_SIZEOF(struct ib_wc, wr_id) < sizeof(void *));
2526 if (srp_sg_tablesize) {
2527 pr_warn("srp_sg_tablesize is deprecated, please use cmd_sg_entries\n");
2528 if (!cmd_sg_entries)
2529 cmd_sg_entries = srp_sg_tablesize;
2532 if (!cmd_sg_entries)
2533 cmd_sg_entries = SRP_DEF_SG_TABLESIZE;
2535 if (cmd_sg_entries > 255) {
2536 pr_warn("Clamping cmd_sg_entries to 255\n");
2537 cmd_sg_entries = 255;
2540 if (!indirect_sg_entries)
2541 indirect_sg_entries = cmd_sg_entries;
2542 else if (indirect_sg_entries < cmd_sg_entries) {
2543 pr_warn("Bumping up indirect_sg_entries to match cmd_sg_entries (%u)\n",
2545 indirect_sg_entries = cmd_sg_entries;
2548 ib_srp_transport_template =
2549 srp_attach_transport(&ib_srp_transport_functions);
2550 if (!ib_srp_transport_template)
2553 ret = class_register(&srp_class);
2555 pr_err("couldn't register class infiniband_srp\n");
2556 srp_release_transport(ib_srp_transport_template);
2560 ib_sa_register_client(&srp_sa_client);
2562 ret = ib_register_client(&srp_client);
2564 pr_err("couldn't register IB client\n");
2565 srp_release_transport(ib_srp_transport_template);
2566 ib_sa_unregister_client(&srp_sa_client);
2567 class_unregister(&srp_class);
2574 static void __exit srp_cleanup_module(void)
2576 ib_unregister_client(&srp_client);
2577 ib_sa_unregister_client(&srp_sa_client);
2578 class_unregister(&srp_class);
2579 srp_release_transport(ib_srp_transport_template);
2582 module_init(srp_init_module);
2583 module_exit(srp_cleanup_module);