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) KBUILD_MODNAME ": " 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>
43 #include <rdma/ib_cache.h>
45 #include <linux/atomic.h>
47 #include <scsi/scsi.h>
48 #include <scsi/scsi_device.h>
49 #include <scsi/scsi_dbg.h>
50 #include <scsi/scsi_tcq.h>
52 #include <scsi/scsi_transport_srp.h>
56 #define DRV_NAME "ib_srp"
57 #define PFX DRV_NAME ": "
58 #define DRV_VERSION "2.0"
59 #define DRV_RELDATE "July 26, 2015"
61 MODULE_AUTHOR("Roland Dreier");
62 MODULE_DESCRIPTION("InfiniBand SCSI RDMA Protocol initiator");
63 MODULE_LICENSE("Dual BSD/GPL");
64 MODULE_VERSION(DRV_VERSION);
65 MODULE_INFO(release_date, DRV_RELDATE);
67 static unsigned int srp_sg_tablesize;
68 static unsigned int cmd_sg_entries;
69 static unsigned int indirect_sg_entries;
70 static bool allow_ext_sg;
71 static bool prefer_fr = true;
72 static bool register_always = true;
73 static int topspin_workarounds = 1;
75 module_param(srp_sg_tablesize, uint, 0444);
76 MODULE_PARM_DESC(srp_sg_tablesize, "Deprecated name for cmd_sg_entries");
78 module_param(cmd_sg_entries, uint, 0444);
79 MODULE_PARM_DESC(cmd_sg_entries,
80 "Default number of gather/scatter entries in the SRP command (default is 12, max 255)");
82 module_param(indirect_sg_entries, uint, 0444);
83 MODULE_PARM_DESC(indirect_sg_entries,
84 "Default max number of gather/scatter entries (default is 12, max is " __stringify(SCSI_MAX_SG_CHAIN_SEGMENTS) ")");
86 module_param(allow_ext_sg, bool, 0444);
87 MODULE_PARM_DESC(allow_ext_sg,
88 "Default behavior when there are more than cmd_sg_entries S/G entries after mapping; fails the request when false (default false)");
90 module_param(topspin_workarounds, int, 0444);
91 MODULE_PARM_DESC(topspin_workarounds,
92 "Enable workarounds for Topspin/Cisco SRP target bugs if != 0");
94 module_param(prefer_fr, bool, 0444);
95 MODULE_PARM_DESC(prefer_fr,
96 "Whether to use fast registration if both FMR and fast registration are supported");
98 module_param(register_always, bool, 0444);
99 MODULE_PARM_DESC(register_always,
100 "Use memory registration even for contiguous memory regions");
102 static const struct kernel_param_ops srp_tmo_ops;
104 static int srp_reconnect_delay = 10;
105 module_param_cb(reconnect_delay, &srp_tmo_ops, &srp_reconnect_delay,
107 MODULE_PARM_DESC(reconnect_delay, "Time between successive reconnect attempts");
109 static int srp_fast_io_fail_tmo = 15;
110 module_param_cb(fast_io_fail_tmo, &srp_tmo_ops, &srp_fast_io_fail_tmo,
112 MODULE_PARM_DESC(fast_io_fail_tmo,
113 "Number of seconds between the observation of a transport"
114 " layer error and failing all I/O. \"off\" means that this"
115 " functionality is disabled.");
117 static int srp_dev_loss_tmo = 600;
118 module_param_cb(dev_loss_tmo, &srp_tmo_ops, &srp_dev_loss_tmo,
120 MODULE_PARM_DESC(dev_loss_tmo,
121 "Maximum number of seconds that the SRP transport should"
122 " insulate transport layer errors. After this time has been"
123 " exceeded the SCSI host is removed. Should be"
124 " between 1 and " __stringify(SCSI_DEVICE_BLOCK_MAX_TIMEOUT)
125 " if fast_io_fail_tmo has not been set. \"off\" means that"
126 " this functionality is disabled.");
128 static unsigned ch_count;
129 module_param(ch_count, uint, 0444);
130 MODULE_PARM_DESC(ch_count,
131 "Number of RDMA channels to use for communication with an SRP target. Using more than one channel improves performance if the HCA supports multiple completion vectors. The default value is the minimum of four times the number of online CPU sockets and the number of completion vectors supported by the HCA.");
133 static void srp_add_one(struct ib_device *device);
134 static void srp_remove_one(struct ib_device *device, void *client_data);
135 static void srp_recv_done(struct ib_cq *cq, struct ib_wc *wc);
136 static void srp_handle_qp_err(struct ib_cq *cq, struct ib_wc *wc,
138 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event);
140 static struct scsi_transport_template *ib_srp_transport_template;
141 static struct workqueue_struct *srp_remove_wq;
143 static struct ib_client srp_client = {
146 .remove = srp_remove_one
149 static struct ib_sa_client srp_sa_client;
151 static int srp_tmo_get(char *buffer, const struct kernel_param *kp)
153 int tmo = *(int *)kp->arg;
156 return sprintf(buffer, "%d", tmo);
158 return sprintf(buffer, "off");
161 static int srp_tmo_set(const char *val, const struct kernel_param *kp)
165 res = srp_parse_tmo(&tmo, val);
169 if (kp->arg == &srp_reconnect_delay)
170 res = srp_tmo_valid(tmo, srp_fast_io_fail_tmo,
172 else if (kp->arg == &srp_fast_io_fail_tmo)
173 res = srp_tmo_valid(srp_reconnect_delay, tmo, srp_dev_loss_tmo);
175 res = srp_tmo_valid(srp_reconnect_delay, srp_fast_io_fail_tmo,
179 *(int *)kp->arg = tmo;
185 static const struct kernel_param_ops srp_tmo_ops = {
190 static inline struct srp_target_port *host_to_target(struct Scsi_Host *host)
192 return (struct srp_target_port *) host->hostdata;
195 static const char *srp_target_info(struct Scsi_Host *host)
197 return host_to_target(host)->target_name;
200 static int srp_target_is_topspin(struct srp_target_port *target)
202 static const u8 topspin_oui[3] = { 0x00, 0x05, 0xad };
203 static const u8 cisco_oui[3] = { 0x00, 0x1b, 0x0d };
205 return topspin_workarounds &&
206 (!memcmp(&target->ioc_guid, topspin_oui, sizeof topspin_oui) ||
207 !memcmp(&target->ioc_guid, cisco_oui, sizeof cisco_oui));
210 static struct srp_iu *srp_alloc_iu(struct srp_host *host, size_t size,
212 enum dma_data_direction direction)
216 iu = kmalloc(sizeof *iu, gfp_mask);
220 iu->buf = kzalloc(size, gfp_mask);
224 iu->dma = ib_dma_map_single(host->srp_dev->dev, iu->buf, size,
226 if (ib_dma_mapping_error(host->srp_dev->dev, iu->dma))
230 iu->direction = direction;
242 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
247 ib_dma_unmap_single(host->srp_dev->dev, iu->dma, iu->size,
253 static void srp_qp_event(struct ib_event *event, void *context)
255 pr_debug("QP event %s (%d)\n",
256 ib_event_msg(event->event), event->event);
259 static int srp_init_qp(struct srp_target_port *target,
262 struct ib_qp_attr *attr;
265 attr = kmalloc(sizeof *attr, GFP_KERNEL);
269 ret = ib_find_cached_pkey(target->srp_host->srp_dev->dev,
270 target->srp_host->port,
271 be16_to_cpu(target->pkey),
276 attr->qp_state = IB_QPS_INIT;
277 attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
278 IB_ACCESS_REMOTE_WRITE);
279 attr->port_num = target->srp_host->port;
281 ret = ib_modify_qp(qp, attr,
292 static int srp_new_cm_id(struct srp_rdma_ch *ch)
294 struct srp_target_port *target = ch->target;
295 struct ib_cm_id *new_cm_id;
297 new_cm_id = ib_create_cm_id(target->srp_host->srp_dev->dev,
299 if (IS_ERR(new_cm_id))
300 return PTR_ERR(new_cm_id);
303 ib_destroy_cm_id(ch->cm_id);
304 ch->cm_id = new_cm_id;
305 ch->path.sgid = target->sgid;
306 ch->path.dgid = target->orig_dgid;
307 ch->path.pkey = target->pkey;
308 ch->path.service_id = target->service_id;
313 static struct ib_fmr_pool *srp_alloc_fmr_pool(struct srp_target_port *target)
315 struct srp_device *dev = target->srp_host->srp_dev;
316 struct ib_fmr_pool_param fmr_param;
318 memset(&fmr_param, 0, sizeof(fmr_param));
319 fmr_param.pool_size = target->mr_pool_size;
320 fmr_param.dirty_watermark = fmr_param.pool_size / 4;
322 fmr_param.max_pages_per_fmr = dev->max_pages_per_mr;
323 fmr_param.page_shift = ilog2(dev->mr_page_size);
324 fmr_param.access = (IB_ACCESS_LOCAL_WRITE |
325 IB_ACCESS_REMOTE_WRITE |
326 IB_ACCESS_REMOTE_READ);
328 return ib_create_fmr_pool(dev->pd, &fmr_param);
332 * srp_destroy_fr_pool() - free the resources owned by a pool
333 * @pool: Fast registration pool to be destroyed.
335 static void srp_destroy_fr_pool(struct srp_fr_pool *pool)
338 struct srp_fr_desc *d;
343 for (i = 0, d = &pool->desc[0]; i < pool->size; i++, d++) {
351 * srp_create_fr_pool() - allocate and initialize a pool for fast registration
352 * @device: IB device to allocate fast registration descriptors for.
353 * @pd: Protection domain associated with the FR descriptors.
354 * @pool_size: Number of descriptors to allocate.
355 * @max_page_list_len: Maximum fast registration work request page list length.
357 static struct srp_fr_pool *srp_create_fr_pool(struct ib_device *device,
358 struct ib_pd *pd, int pool_size,
359 int max_page_list_len)
361 struct srp_fr_pool *pool;
362 struct srp_fr_desc *d;
364 int i, ret = -EINVAL;
369 pool = kzalloc(sizeof(struct srp_fr_pool) +
370 pool_size * sizeof(struct srp_fr_desc), GFP_KERNEL);
373 pool->size = pool_size;
374 pool->max_page_list_len = max_page_list_len;
375 spin_lock_init(&pool->lock);
376 INIT_LIST_HEAD(&pool->free_list);
378 for (i = 0, d = &pool->desc[0]; i < pool->size; i++, d++) {
379 mr = ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG,
386 list_add_tail(&d->entry, &pool->free_list);
393 srp_destroy_fr_pool(pool);
401 * srp_fr_pool_get() - obtain a descriptor suitable for fast registration
402 * @pool: Pool to obtain descriptor from.
404 static struct srp_fr_desc *srp_fr_pool_get(struct srp_fr_pool *pool)
406 struct srp_fr_desc *d = NULL;
409 spin_lock_irqsave(&pool->lock, flags);
410 if (!list_empty(&pool->free_list)) {
411 d = list_first_entry(&pool->free_list, typeof(*d), entry);
414 spin_unlock_irqrestore(&pool->lock, flags);
420 * srp_fr_pool_put() - put an FR descriptor back in the free list
421 * @pool: Pool the descriptor was allocated from.
422 * @desc: Pointer to an array of fast registration descriptor pointers.
423 * @n: Number of descriptors to put back.
425 * Note: The caller must already have queued an invalidation request for
426 * desc->mr->rkey before calling this function.
428 static void srp_fr_pool_put(struct srp_fr_pool *pool, struct srp_fr_desc **desc,
434 spin_lock_irqsave(&pool->lock, flags);
435 for (i = 0; i < n; i++)
436 list_add(&desc[i]->entry, &pool->free_list);
437 spin_unlock_irqrestore(&pool->lock, flags);
440 static struct srp_fr_pool *srp_alloc_fr_pool(struct srp_target_port *target)
442 struct srp_device *dev = target->srp_host->srp_dev;
444 return srp_create_fr_pool(dev->dev, dev->pd, target->mr_pool_size,
445 dev->max_pages_per_mr);
449 * srp_destroy_qp() - destroy an RDMA queue pair
450 * @ch: SRP RDMA channel.
452 * Drain the qp before destroying it. This avoids that the receive
453 * completion handler can access the queue pair while it is
456 static void srp_destroy_qp(struct srp_rdma_ch *ch)
459 ib_destroy_qp(ch->qp);
462 static int srp_create_ch_ib(struct srp_rdma_ch *ch)
464 struct srp_target_port *target = ch->target;
465 struct srp_device *dev = target->srp_host->srp_dev;
466 struct ib_qp_init_attr *init_attr;
467 struct ib_cq *recv_cq, *send_cq;
469 struct ib_fmr_pool *fmr_pool = NULL;
470 struct srp_fr_pool *fr_pool = NULL;
471 const int m = dev->use_fast_reg ? 3 : 1;
474 init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
478 /* queue_size + 1 for ib_drain_rq() */
479 recv_cq = ib_alloc_cq(dev->dev, ch, target->queue_size + 1,
480 ch->comp_vector, IB_POLL_SOFTIRQ);
481 if (IS_ERR(recv_cq)) {
482 ret = PTR_ERR(recv_cq);
486 send_cq = ib_alloc_cq(dev->dev, ch, m * target->queue_size,
487 ch->comp_vector, IB_POLL_DIRECT);
488 if (IS_ERR(send_cq)) {
489 ret = PTR_ERR(send_cq);
493 init_attr->event_handler = srp_qp_event;
494 init_attr->cap.max_send_wr = m * target->queue_size;
495 init_attr->cap.max_recv_wr = target->queue_size + 1;
496 init_attr->cap.max_recv_sge = 1;
497 init_attr->cap.max_send_sge = 1;
498 init_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
499 init_attr->qp_type = IB_QPT_RC;
500 init_attr->send_cq = send_cq;
501 init_attr->recv_cq = recv_cq;
503 qp = ib_create_qp(dev->pd, init_attr);
509 ret = srp_init_qp(target, qp);
513 if (dev->use_fast_reg) {
514 fr_pool = srp_alloc_fr_pool(target);
515 if (IS_ERR(fr_pool)) {
516 ret = PTR_ERR(fr_pool);
517 shost_printk(KERN_WARNING, target->scsi_host, PFX
518 "FR pool allocation failed (%d)\n", ret);
521 } else if (dev->use_fmr) {
522 fmr_pool = srp_alloc_fmr_pool(target);
523 if (IS_ERR(fmr_pool)) {
524 ret = PTR_ERR(fmr_pool);
525 shost_printk(KERN_WARNING, target->scsi_host, PFX
526 "FMR pool allocation failed (%d)\n", ret);
534 ib_free_cq(ch->recv_cq);
536 ib_free_cq(ch->send_cq);
539 ch->recv_cq = recv_cq;
540 ch->send_cq = send_cq;
542 if (dev->use_fast_reg) {
544 srp_destroy_fr_pool(ch->fr_pool);
545 ch->fr_pool = fr_pool;
546 } else if (dev->use_fmr) {
548 ib_destroy_fmr_pool(ch->fmr_pool);
549 ch->fmr_pool = fmr_pool;
570 * Note: this function may be called without srp_alloc_iu_bufs() having been
571 * invoked. Hence the ch->[rt]x_ring checks.
573 static void srp_free_ch_ib(struct srp_target_port *target,
574 struct srp_rdma_ch *ch)
576 struct srp_device *dev = target->srp_host->srp_dev;
583 ib_destroy_cm_id(ch->cm_id);
587 /* If srp_new_cm_id() succeeded but srp_create_ch_ib() not, return. */
591 if (dev->use_fast_reg) {
593 srp_destroy_fr_pool(ch->fr_pool);
594 } else if (dev->use_fmr) {
596 ib_destroy_fmr_pool(ch->fmr_pool);
600 ib_free_cq(ch->send_cq);
601 ib_free_cq(ch->recv_cq);
604 * Avoid that the SCSI error handler tries to use this channel after
605 * it has been freed. The SCSI error handler can namely continue
606 * trying to perform recovery actions after scsi_remove_host()
612 ch->send_cq = ch->recv_cq = NULL;
615 for (i = 0; i < target->queue_size; ++i)
616 srp_free_iu(target->srp_host, ch->rx_ring[i]);
621 for (i = 0; i < target->queue_size; ++i)
622 srp_free_iu(target->srp_host, ch->tx_ring[i]);
628 static void srp_path_rec_completion(int status,
629 struct ib_sa_path_rec *pathrec,
632 struct srp_rdma_ch *ch = ch_ptr;
633 struct srp_target_port *target = ch->target;
637 shost_printk(KERN_ERR, target->scsi_host,
638 PFX "Got failed path rec status %d\n", status);
644 static int srp_lookup_path(struct srp_rdma_ch *ch)
646 struct srp_target_port *target = ch->target;
649 ch->path.numb_path = 1;
651 init_completion(&ch->done);
653 ch->path_query_id = ib_sa_path_rec_get(&srp_sa_client,
654 target->srp_host->srp_dev->dev,
655 target->srp_host->port,
657 IB_SA_PATH_REC_SERVICE_ID |
658 IB_SA_PATH_REC_DGID |
659 IB_SA_PATH_REC_SGID |
660 IB_SA_PATH_REC_NUMB_PATH |
662 SRP_PATH_REC_TIMEOUT_MS,
664 srp_path_rec_completion,
665 ch, &ch->path_query);
666 if (ch->path_query_id < 0)
667 return ch->path_query_id;
669 ret = wait_for_completion_interruptible(&ch->done);
674 shost_printk(KERN_WARNING, target->scsi_host,
675 PFX "Path record query failed\n");
680 static int srp_send_req(struct srp_rdma_ch *ch, bool multich)
682 struct srp_target_port *target = ch->target;
684 struct ib_cm_req_param param;
685 struct srp_login_req priv;
689 req = kzalloc(sizeof *req, GFP_KERNEL);
693 req->param.primary_path = &ch->path;
694 req->param.alternate_path = NULL;
695 req->param.service_id = target->service_id;
696 req->param.qp_num = ch->qp->qp_num;
697 req->param.qp_type = ch->qp->qp_type;
698 req->param.private_data = &req->priv;
699 req->param.private_data_len = sizeof req->priv;
700 req->param.flow_control = 1;
702 get_random_bytes(&req->param.starting_psn, 4);
703 req->param.starting_psn &= 0xffffff;
706 * Pick some arbitrary defaults here; we could make these
707 * module parameters if anyone cared about setting them.
709 req->param.responder_resources = 4;
710 req->param.remote_cm_response_timeout = 20;
711 req->param.local_cm_response_timeout = 20;
712 req->param.retry_count = target->tl_retry_count;
713 req->param.rnr_retry_count = 7;
714 req->param.max_cm_retries = 15;
716 req->priv.opcode = SRP_LOGIN_REQ;
718 req->priv.req_it_iu_len = cpu_to_be32(target->max_iu_len);
719 req->priv.req_buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
720 SRP_BUF_FORMAT_INDIRECT);
721 req->priv.req_flags = (multich ? SRP_MULTICHAN_MULTI :
722 SRP_MULTICHAN_SINGLE);
724 * In the published SRP specification (draft rev. 16a), the
725 * port identifier format is 8 bytes of ID extension followed
726 * by 8 bytes of GUID. Older drafts put the two halves in the
727 * opposite order, so that the GUID comes first.
729 * Targets conforming to these obsolete drafts can be
730 * recognized by the I/O Class they report.
732 if (target->io_class == SRP_REV10_IB_IO_CLASS) {
733 memcpy(req->priv.initiator_port_id,
734 &target->sgid.global.interface_id, 8);
735 memcpy(req->priv.initiator_port_id + 8,
736 &target->initiator_ext, 8);
737 memcpy(req->priv.target_port_id, &target->ioc_guid, 8);
738 memcpy(req->priv.target_port_id + 8, &target->id_ext, 8);
740 memcpy(req->priv.initiator_port_id,
741 &target->initiator_ext, 8);
742 memcpy(req->priv.initiator_port_id + 8,
743 &target->sgid.global.interface_id, 8);
744 memcpy(req->priv.target_port_id, &target->id_ext, 8);
745 memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
749 * Topspin/Cisco SRP targets will reject our login unless we
750 * zero out the first 8 bytes of our initiator port ID and set
751 * the second 8 bytes to the local node GUID.
753 if (srp_target_is_topspin(target)) {
754 shost_printk(KERN_DEBUG, target->scsi_host,
755 PFX "Topspin/Cisco initiator port ID workaround "
756 "activated for target GUID %016llx\n",
757 be64_to_cpu(target->ioc_guid));
758 memset(req->priv.initiator_port_id, 0, 8);
759 memcpy(req->priv.initiator_port_id + 8,
760 &target->srp_host->srp_dev->dev->node_guid, 8);
763 status = ib_send_cm_req(ch->cm_id, &req->param);
770 static bool srp_queue_remove_work(struct srp_target_port *target)
772 bool changed = false;
774 spin_lock_irq(&target->lock);
775 if (target->state != SRP_TARGET_REMOVED) {
776 target->state = SRP_TARGET_REMOVED;
779 spin_unlock_irq(&target->lock);
782 queue_work(srp_remove_wq, &target->remove_work);
787 static void srp_disconnect_target(struct srp_target_port *target)
789 struct srp_rdma_ch *ch;
792 /* XXX should send SRP_I_LOGOUT request */
794 for (i = 0; i < target->ch_count; i++) {
796 ch->connected = false;
797 if (ch->cm_id && ib_send_cm_dreq(ch->cm_id, NULL, 0)) {
798 shost_printk(KERN_DEBUG, target->scsi_host,
799 PFX "Sending CM DREQ failed\n");
804 static void srp_free_req_data(struct srp_target_port *target,
805 struct srp_rdma_ch *ch)
807 struct srp_device *dev = target->srp_host->srp_dev;
808 struct ib_device *ibdev = dev->dev;
809 struct srp_request *req;
815 for (i = 0; i < target->req_ring_size; ++i) {
816 req = &ch->req_ring[i];
817 if (dev->use_fast_reg) {
820 kfree(req->fmr_list);
821 kfree(req->map_page);
823 if (req->indirect_dma_addr) {
824 ib_dma_unmap_single(ibdev, req->indirect_dma_addr,
825 target->indirect_size,
828 kfree(req->indirect_desc);
835 static int srp_alloc_req_data(struct srp_rdma_ch *ch)
837 struct srp_target_port *target = ch->target;
838 struct srp_device *srp_dev = target->srp_host->srp_dev;
839 struct ib_device *ibdev = srp_dev->dev;
840 struct srp_request *req;
843 int i, ret = -ENOMEM;
845 ch->req_ring = kcalloc(target->req_ring_size, sizeof(*ch->req_ring),
850 for (i = 0; i < target->req_ring_size; ++i) {
851 req = &ch->req_ring[i];
852 mr_list = kmalloc(target->cmd_sg_cnt * sizeof(void *),
856 if (srp_dev->use_fast_reg) {
857 req->fr_list = mr_list;
859 req->fmr_list = mr_list;
860 req->map_page = kmalloc(srp_dev->max_pages_per_mr *
861 sizeof(void *), GFP_KERNEL);
865 req->indirect_desc = kmalloc(target->indirect_size, GFP_KERNEL);
866 if (!req->indirect_desc)
869 dma_addr = ib_dma_map_single(ibdev, req->indirect_desc,
870 target->indirect_size,
872 if (ib_dma_mapping_error(ibdev, dma_addr))
875 req->indirect_dma_addr = dma_addr;
884 * srp_del_scsi_host_attr() - Remove attributes defined in the host template.
885 * @shost: SCSI host whose attributes to remove from sysfs.
887 * Note: Any attributes defined in the host template and that did not exist
888 * before invocation of this function will be ignored.
890 static void srp_del_scsi_host_attr(struct Scsi_Host *shost)
892 struct device_attribute **attr;
894 for (attr = shost->hostt->shost_attrs; attr && *attr; ++attr)
895 device_remove_file(&shost->shost_dev, *attr);
898 static void srp_remove_target(struct srp_target_port *target)
900 struct srp_rdma_ch *ch;
903 WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
905 srp_del_scsi_host_attr(target->scsi_host);
906 srp_rport_get(target->rport);
907 srp_remove_host(target->scsi_host);
908 scsi_remove_host(target->scsi_host);
909 srp_stop_rport_timers(target->rport);
910 srp_disconnect_target(target);
911 for (i = 0; i < target->ch_count; i++) {
913 srp_free_ch_ib(target, ch);
915 cancel_work_sync(&target->tl_err_work);
916 srp_rport_put(target->rport);
917 for (i = 0; i < target->ch_count; i++) {
919 srp_free_req_data(target, ch);
924 spin_lock(&target->srp_host->target_lock);
925 list_del(&target->list);
926 spin_unlock(&target->srp_host->target_lock);
928 scsi_host_put(target->scsi_host);
931 static void srp_remove_work(struct work_struct *work)
933 struct srp_target_port *target =
934 container_of(work, struct srp_target_port, remove_work);
936 WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
938 srp_remove_target(target);
941 static void srp_rport_delete(struct srp_rport *rport)
943 struct srp_target_port *target = rport->lld_data;
945 srp_queue_remove_work(target);
949 * srp_connected_ch() - number of connected channels
950 * @target: SRP target port.
952 static int srp_connected_ch(struct srp_target_port *target)
956 for (i = 0; i < target->ch_count; i++)
957 c += target->ch[i].connected;
962 static int srp_connect_ch(struct srp_rdma_ch *ch, bool multich)
964 struct srp_target_port *target = ch->target;
967 WARN_ON_ONCE(!multich && srp_connected_ch(target) > 0);
969 ret = srp_lookup_path(ch);
974 init_completion(&ch->done);
975 ret = srp_send_req(ch, multich);
978 ret = wait_for_completion_interruptible(&ch->done);
983 * The CM event handling code will set status to
984 * SRP_PORT_REDIRECT if we get a port redirect REJ
985 * back, or SRP_DLID_REDIRECT if we get a lid/qp
991 ch->connected = true;
994 case SRP_PORT_REDIRECT:
995 ret = srp_lookup_path(ch);
1000 case SRP_DLID_REDIRECT:
1003 case SRP_STALE_CONN:
1004 shost_printk(KERN_ERR, target->scsi_host, PFX
1005 "giving up on stale connection\n");
1015 return ret <= 0 ? ret : -ENODEV;
1018 static void srp_inv_rkey_err_done(struct ib_cq *cq, struct ib_wc *wc)
1020 srp_handle_qp_err(cq, wc, "INV RKEY");
1023 static int srp_inv_rkey(struct srp_request *req, struct srp_rdma_ch *ch,
1026 struct ib_send_wr *bad_wr;
1027 struct ib_send_wr wr = {
1028 .opcode = IB_WR_LOCAL_INV,
1032 .ex.invalidate_rkey = rkey,
1035 wr.wr_cqe = &req->reg_cqe;
1036 req->reg_cqe.done = srp_inv_rkey_err_done;
1037 return ib_post_send(ch->qp, &wr, &bad_wr);
1040 static void srp_unmap_data(struct scsi_cmnd *scmnd,
1041 struct srp_rdma_ch *ch,
1042 struct srp_request *req)
1044 struct srp_target_port *target = ch->target;
1045 struct srp_device *dev = target->srp_host->srp_dev;
1046 struct ib_device *ibdev = dev->dev;
1049 if (!scsi_sglist(scmnd) ||
1050 (scmnd->sc_data_direction != DMA_TO_DEVICE &&
1051 scmnd->sc_data_direction != DMA_FROM_DEVICE))
1054 if (dev->use_fast_reg) {
1055 struct srp_fr_desc **pfr;
1057 for (i = req->nmdesc, pfr = req->fr_list; i > 0; i--, pfr++) {
1058 res = srp_inv_rkey(req, ch, (*pfr)->mr->rkey);
1060 shost_printk(KERN_ERR, target->scsi_host, PFX
1061 "Queueing INV WR for rkey %#x failed (%d)\n",
1062 (*pfr)->mr->rkey, res);
1063 queue_work(system_long_wq,
1064 &target->tl_err_work);
1068 srp_fr_pool_put(ch->fr_pool, req->fr_list,
1070 } else if (dev->use_fmr) {
1071 struct ib_pool_fmr **pfmr;
1073 for (i = req->nmdesc, pfmr = req->fmr_list; i > 0; i--, pfmr++)
1074 ib_fmr_pool_unmap(*pfmr);
1077 ib_dma_unmap_sg(ibdev, scsi_sglist(scmnd), scsi_sg_count(scmnd),
1078 scmnd->sc_data_direction);
1082 * srp_claim_req - Take ownership of the scmnd associated with a request.
1083 * @ch: SRP RDMA channel.
1084 * @req: SRP request.
1085 * @sdev: If not NULL, only take ownership for this SCSI device.
1086 * @scmnd: If NULL, take ownership of @req->scmnd. If not NULL, only take
1087 * ownership of @req->scmnd if it equals @scmnd.
1090 * Either NULL or a pointer to the SCSI command the caller became owner of.
1092 static struct scsi_cmnd *srp_claim_req(struct srp_rdma_ch *ch,
1093 struct srp_request *req,
1094 struct scsi_device *sdev,
1095 struct scsi_cmnd *scmnd)
1097 unsigned long flags;
1099 spin_lock_irqsave(&ch->lock, flags);
1101 (!sdev || req->scmnd->device == sdev) &&
1102 (!scmnd || req->scmnd == scmnd)) {
1108 spin_unlock_irqrestore(&ch->lock, flags);
1114 * srp_free_req() - Unmap data and adjust ch->req_lim.
1115 * @ch: SRP RDMA channel.
1116 * @req: Request to be freed.
1117 * @scmnd: SCSI command associated with @req.
1118 * @req_lim_delta: Amount to be added to @target->req_lim.
1120 static void srp_free_req(struct srp_rdma_ch *ch, struct srp_request *req,
1121 struct scsi_cmnd *scmnd, s32 req_lim_delta)
1123 unsigned long flags;
1125 srp_unmap_data(scmnd, ch, req);
1127 spin_lock_irqsave(&ch->lock, flags);
1128 ch->req_lim += req_lim_delta;
1129 spin_unlock_irqrestore(&ch->lock, flags);
1132 static void srp_finish_req(struct srp_rdma_ch *ch, struct srp_request *req,
1133 struct scsi_device *sdev, int result)
1135 struct scsi_cmnd *scmnd = srp_claim_req(ch, req, sdev, NULL);
1138 srp_free_req(ch, req, scmnd, 0);
1139 scmnd->result = result;
1140 scmnd->scsi_done(scmnd);
1144 static void srp_terminate_io(struct srp_rport *rport)
1146 struct srp_target_port *target = rport->lld_data;
1147 struct srp_rdma_ch *ch;
1148 struct Scsi_Host *shost = target->scsi_host;
1149 struct scsi_device *sdev;
1153 * Invoking srp_terminate_io() while srp_queuecommand() is running
1154 * is not safe. Hence the warning statement below.
1156 shost_for_each_device(sdev, shost)
1157 WARN_ON_ONCE(sdev->request_queue->request_fn_active);
1159 for (i = 0; i < target->ch_count; i++) {
1160 ch = &target->ch[i];
1162 for (j = 0; j < target->req_ring_size; ++j) {
1163 struct srp_request *req = &ch->req_ring[j];
1165 srp_finish_req(ch, req, NULL,
1166 DID_TRANSPORT_FAILFAST << 16);
1172 * It is up to the caller to ensure that srp_rport_reconnect() calls are
1173 * serialized and that no concurrent srp_queuecommand(), srp_abort(),
1174 * srp_reset_device() or srp_reset_host() calls will occur while this function
1175 * is in progress. One way to realize that is not to call this function
1176 * directly but to call srp_reconnect_rport() instead since that last function
1177 * serializes calls of this function via rport->mutex and also blocks
1178 * srp_queuecommand() calls before invoking this function.
1180 static int srp_rport_reconnect(struct srp_rport *rport)
1182 struct srp_target_port *target = rport->lld_data;
1183 struct srp_rdma_ch *ch;
1185 bool multich = false;
1187 srp_disconnect_target(target);
1189 if (target->state == SRP_TARGET_SCANNING)
1193 * Now get a new local CM ID so that we avoid confusing the target in
1194 * case things are really fouled up. Doing so also ensures that all CM
1195 * callbacks will have finished before a new QP is allocated.
1197 for (i = 0; i < target->ch_count; i++) {
1198 ch = &target->ch[i];
1199 ret += srp_new_cm_id(ch);
1201 for (i = 0; i < target->ch_count; i++) {
1202 ch = &target->ch[i];
1203 for (j = 0; j < target->req_ring_size; ++j) {
1204 struct srp_request *req = &ch->req_ring[j];
1206 srp_finish_req(ch, req, NULL, DID_RESET << 16);
1209 for (i = 0; i < target->ch_count; i++) {
1210 ch = &target->ch[i];
1212 * Whether or not creating a new CM ID succeeded, create a new
1213 * QP. This guarantees that all completion callback function
1214 * invocations have finished before request resetting starts.
1216 ret += srp_create_ch_ib(ch);
1218 INIT_LIST_HEAD(&ch->free_tx);
1219 for (j = 0; j < target->queue_size; ++j)
1220 list_add(&ch->tx_ring[j]->list, &ch->free_tx);
1223 target->qp_in_error = false;
1225 for (i = 0; i < target->ch_count; i++) {
1226 ch = &target->ch[i];
1229 ret = srp_connect_ch(ch, multich);
1234 shost_printk(KERN_INFO, target->scsi_host,
1235 PFX "reconnect succeeded\n");
1240 static void srp_map_desc(struct srp_map_state *state, dma_addr_t dma_addr,
1241 unsigned int dma_len, u32 rkey)
1243 struct srp_direct_buf *desc = state->desc;
1245 WARN_ON_ONCE(!dma_len);
1247 desc->va = cpu_to_be64(dma_addr);
1248 desc->key = cpu_to_be32(rkey);
1249 desc->len = cpu_to_be32(dma_len);
1251 state->total_len += dma_len;
1256 static int srp_map_finish_fmr(struct srp_map_state *state,
1257 struct srp_rdma_ch *ch)
1259 struct srp_target_port *target = ch->target;
1260 struct srp_device *dev = target->srp_host->srp_dev;
1261 struct ib_pool_fmr *fmr;
1264 if (state->fmr.next >= state->fmr.end)
1267 WARN_ON_ONCE(!dev->use_fmr);
1269 if (state->npages == 0)
1272 if (state->npages == 1 && target->global_mr) {
1273 srp_map_desc(state, state->base_dma_addr, state->dma_len,
1274 target->global_mr->rkey);
1278 fmr = ib_fmr_pool_map_phys(ch->fmr_pool, state->pages,
1279 state->npages, io_addr);
1281 return PTR_ERR(fmr);
1283 *state->fmr.next++ = fmr;
1286 srp_map_desc(state, state->base_dma_addr & ~dev->mr_page_mask,
1287 state->dma_len, fmr->fmr->rkey);
1296 static void srp_reg_mr_err_done(struct ib_cq *cq, struct ib_wc *wc)
1298 srp_handle_qp_err(cq, wc, "FAST REG");
1301 static int srp_map_finish_fr(struct srp_map_state *state,
1302 struct srp_request *req,
1303 struct srp_rdma_ch *ch, int sg_nents)
1305 struct srp_target_port *target = ch->target;
1306 struct srp_device *dev = target->srp_host->srp_dev;
1307 struct ib_send_wr *bad_wr;
1308 struct ib_reg_wr wr;
1309 struct srp_fr_desc *desc;
1313 if (state->fr.next >= state->fr.end)
1316 WARN_ON_ONCE(!dev->use_fast_reg);
1318 if (sg_nents == 1 && target->global_mr) {
1319 srp_map_desc(state, sg_dma_address(state->sg),
1320 sg_dma_len(state->sg),
1321 target->global_mr->rkey);
1325 desc = srp_fr_pool_get(ch->fr_pool);
1329 rkey = ib_inc_rkey(desc->mr->rkey);
1330 ib_update_fast_reg_key(desc->mr, rkey);
1332 n = ib_map_mr_sg(desc->mr, state->sg, sg_nents, 0, dev->mr_page_size);
1333 if (unlikely(n < 0))
1336 req->reg_cqe.done = srp_reg_mr_err_done;
1339 wr.wr.opcode = IB_WR_REG_MR;
1340 wr.wr.wr_cqe = &req->reg_cqe;
1342 wr.wr.send_flags = 0;
1344 wr.key = desc->mr->rkey;
1345 wr.access = (IB_ACCESS_LOCAL_WRITE |
1346 IB_ACCESS_REMOTE_READ |
1347 IB_ACCESS_REMOTE_WRITE);
1349 *state->fr.next++ = desc;
1352 srp_map_desc(state, desc->mr->iova,
1353 desc->mr->length, desc->mr->rkey);
1355 err = ib_post_send(ch->qp, &wr.wr, &bad_wr);
1362 static int srp_map_sg_entry(struct srp_map_state *state,
1363 struct srp_rdma_ch *ch,
1364 struct scatterlist *sg, int sg_index)
1366 struct srp_target_port *target = ch->target;
1367 struct srp_device *dev = target->srp_host->srp_dev;
1368 struct ib_device *ibdev = dev->dev;
1369 dma_addr_t dma_addr = ib_sg_dma_address(ibdev, sg);
1370 unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
1371 unsigned int len = 0;
1374 WARN_ON_ONCE(!dma_len);
1377 unsigned offset = dma_addr & ~dev->mr_page_mask;
1378 if (state->npages == dev->max_pages_per_mr || offset != 0) {
1379 ret = srp_map_finish_fmr(state, ch);
1384 len = min_t(unsigned int, dma_len, dev->mr_page_size - offset);
1387 state->base_dma_addr = dma_addr;
1388 state->pages[state->npages++] = dma_addr & dev->mr_page_mask;
1389 state->dma_len += len;
1395 * If the last entry of the MR wasn't a full page, then we need to
1396 * close it out and start a new one -- we can only merge at page
1400 if (len != dev->mr_page_size)
1401 ret = srp_map_finish_fmr(state, ch);
1405 static int srp_map_sg_fmr(struct srp_map_state *state, struct srp_rdma_ch *ch,
1406 struct srp_request *req, struct scatterlist *scat,
1409 struct scatterlist *sg;
1412 state->pages = req->map_page;
1413 state->fmr.next = req->fmr_list;
1414 state->fmr.end = req->fmr_list + ch->target->cmd_sg_cnt;
1416 for_each_sg(scat, sg, count, i) {
1417 ret = srp_map_sg_entry(state, ch, sg, i);
1422 ret = srp_map_finish_fmr(state, ch);
1429 static int srp_map_sg_fr(struct srp_map_state *state, struct srp_rdma_ch *ch,
1430 struct srp_request *req, struct scatterlist *scat,
1433 state->desc = req->indirect_desc;
1434 state->fr.next = req->fr_list;
1435 state->fr.end = req->fr_list + ch->target->cmd_sg_cnt;
1444 n = srp_map_finish_fr(state, req, ch, count);
1445 if (unlikely(n < 0))
1449 for (i = 0; i < n; i++)
1450 state->sg = sg_next(state->sg);
1456 static int srp_map_sg_dma(struct srp_map_state *state, struct srp_rdma_ch *ch,
1457 struct srp_request *req, struct scatterlist *scat,
1460 struct srp_target_port *target = ch->target;
1461 struct srp_device *dev = target->srp_host->srp_dev;
1462 struct scatterlist *sg;
1465 state->desc = req->indirect_desc;
1466 for_each_sg(scat, sg, count, i) {
1467 srp_map_desc(state, ib_sg_dma_address(dev->dev, sg),
1468 ib_sg_dma_len(dev->dev, sg),
1469 target->global_mr->rkey);
1476 * Register the indirect data buffer descriptor with the HCA.
1478 * Note: since the indirect data buffer descriptor has been allocated with
1479 * kmalloc() it is guaranteed that this buffer is a physically contiguous
1482 static int srp_map_idb(struct srp_rdma_ch *ch, struct srp_request *req,
1483 void **next_mr, void **end_mr, u32 idb_len,
1486 struct srp_target_port *target = ch->target;
1487 struct srp_device *dev = target->srp_host->srp_dev;
1488 struct srp_map_state state;
1489 struct srp_direct_buf idb_desc;
1491 struct scatterlist idb_sg[1];
1494 memset(&state, 0, sizeof(state));
1495 memset(&idb_desc, 0, sizeof(idb_desc));
1496 state.gen.next = next_mr;
1497 state.gen.end = end_mr;
1498 state.desc = &idb_desc;
1499 state.base_dma_addr = req->indirect_dma_addr;
1500 state.dma_len = idb_len;
1502 if (dev->use_fast_reg) {
1504 sg_set_buf(idb_sg, req->indirect_desc, idb_len);
1505 idb_sg->dma_address = req->indirect_dma_addr; /* hack! */
1506 #ifdef CONFIG_NEED_SG_DMA_LENGTH
1507 idb_sg->dma_length = idb_sg->length; /* hack^2 */
1509 ret = srp_map_finish_fr(&state, req, ch, 1);
1512 } else if (dev->use_fmr) {
1513 state.pages = idb_pages;
1514 state.pages[0] = (req->indirect_dma_addr &
1517 ret = srp_map_finish_fmr(&state, ch);
1524 *idb_rkey = idb_desc.key;
1530 * srp_map_data() - map SCSI data buffer onto an SRP request
1531 * @scmnd: SCSI command to map
1532 * @ch: SRP RDMA channel
1535 * Returns the length in bytes of the SRP_CMD IU or a negative value if
1538 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_rdma_ch *ch,
1539 struct srp_request *req)
1541 struct srp_target_port *target = ch->target;
1542 struct scatterlist *scat;
1543 struct srp_cmd *cmd = req->cmd->buf;
1544 int len, nents, count, ret;
1545 struct srp_device *dev;
1546 struct ib_device *ibdev;
1547 struct srp_map_state state;
1548 struct srp_indirect_buf *indirect_hdr;
1549 u32 idb_len, table_len;
1553 if (!scsi_sglist(scmnd) || scmnd->sc_data_direction == DMA_NONE)
1554 return sizeof (struct srp_cmd);
1556 if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
1557 scmnd->sc_data_direction != DMA_TO_DEVICE) {
1558 shost_printk(KERN_WARNING, target->scsi_host,
1559 PFX "Unhandled data direction %d\n",
1560 scmnd->sc_data_direction);
1564 nents = scsi_sg_count(scmnd);
1565 scat = scsi_sglist(scmnd);
1567 dev = target->srp_host->srp_dev;
1570 count = ib_dma_map_sg(ibdev, scat, nents, scmnd->sc_data_direction);
1571 if (unlikely(count == 0))
1574 fmt = SRP_DATA_DESC_DIRECT;
1575 len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
1577 if (count == 1 && target->global_mr) {
1579 * The midlayer only generated a single gather/scatter
1580 * entry, or DMA mapping coalesced everything to a
1581 * single entry. So a direct descriptor along with
1582 * the DMA MR suffices.
1584 struct srp_direct_buf *buf = (void *) cmd->add_data;
1586 buf->va = cpu_to_be64(ib_sg_dma_address(ibdev, scat));
1587 buf->key = cpu_to_be32(target->global_mr->rkey);
1588 buf->len = cpu_to_be32(ib_sg_dma_len(ibdev, scat));
1595 * We have more than one scatter/gather entry, so build our indirect
1596 * descriptor table, trying to merge as many entries as we can.
1598 indirect_hdr = (void *) cmd->add_data;
1600 ib_dma_sync_single_for_cpu(ibdev, req->indirect_dma_addr,
1601 target->indirect_size, DMA_TO_DEVICE);
1603 memset(&state, 0, sizeof(state));
1604 if (dev->use_fast_reg)
1605 ret = srp_map_sg_fr(&state, ch, req, scat, count);
1606 else if (dev->use_fmr)
1607 ret = srp_map_sg_fmr(&state, ch, req, scat, count);
1609 ret = srp_map_sg_dma(&state, ch, req, scat, count);
1610 req->nmdesc = state.nmdesc;
1614 /* We've mapped the request, now pull as much of the indirect
1615 * descriptor table as we can into the command buffer. If this
1616 * target is not using an external indirect table, we are
1617 * guaranteed to fit into the command, as the SCSI layer won't
1618 * give us more S/G entries than we allow.
1620 if (state.ndesc == 1) {
1622 * Memory registration collapsed the sg-list into one entry,
1623 * so use a direct descriptor.
1625 struct srp_direct_buf *buf = (void *) cmd->add_data;
1627 *buf = req->indirect_desc[0];
1631 if (unlikely(target->cmd_sg_cnt < state.ndesc &&
1632 !target->allow_ext_sg)) {
1633 shost_printk(KERN_ERR, target->scsi_host,
1634 "Could not fit S/G list into SRP_CMD\n");
1639 count = min(state.ndesc, target->cmd_sg_cnt);
1640 table_len = state.ndesc * sizeof (struct srp_direct_buf);
1641 idb_len = sizeof(struct srp_indirect_buf) + table_len;
1643 fmt = SRP_DATA_DESC_INDIRECT;
1644 len = sizeof(struct srp_cmd) + sizeof (struct srp_indirect_buf);
1645 len += count * sizeof (struct srp_direct_buf);
1647 memcpy(indirect_hdr->desc_list, req->indirect_desc,
1648 count * sizeof (struct srp_direct_buf));
1650 if (!target->global_mr) {
1651 ret = srp_map_idb(ch, req, state.gen.next, state.gen.end,
1652 idb_len, &idb_rkey);
1657 idb_rkey = cpu_to_be32(target->global_mr->rkey);
1660 indirect_hdr->table_desc.va = cpu_to_be64(req->indirect_dma_addr);
1661 indirect_hdr->table_desc.key = idb_rkey;
1662 indirect_hdr->table_desc.len = cpu_to_be32(table_len);
1663 indirect_hdr->len = cpu_to_be32(state.total_len);
1665 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1666 cmd->data_out_desc_cnt = count;
1668 cmd->data_in_desc_cnt = count;
1670 ib_dma_sync_single_for_device(ibdev, req->indirect_dma_addr, table_len,
1674 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1675 cmd->buf_fmt = fmt << 4;
1682 srp_unmap_data(scmnd, ch, req);
1683 if (ret == -ENOMEM && req->nmdesc >= target->mr_pool_size)
1689 * Return an IU and possible credit to the free pool
1691 static void srp_put_tx_iu(struct srp_rdma_ch *ch, struct srp_iu *iu,
1692 enum srp_iu_type iu_type)
1694 unsigned long flags;
1696 spin_lock_irqsave(&ch->lock, flags);
1697 list_add(&iu->list, &ch->free_tx);
1698 if (iu_type != SRP_IU_RSP)
1700 spin_unlock_irqrestore(&ch->lock, flags);
1704 * Must be called with ch->lock held to protect req_lim and free_tx.
1705 * If IU is not sent, it must be returned using srp_put_tx_iu().
1708 * An upper limit for the number of allocated information units for each
1710 * - SRP_IU_CMD: SRP_CMD_SQ_SIZE, since the SCSI mid-layer never queues
1711 * more than Scsi_Host.can_queue requests.
1712 * - SRP_IU_TSK_MGMT: SRP_TSK_MGMT_SQ_SIZE.
1713 * - SRP_IU_RSP: 1, since a conforming SRP target never sends more than
1714 * one unanswered SRP request to an initiator.
1716 static struct srp_iu *__srp_get_tx_iu(struct srp_rdma_ch *ch,
1717 enum srp_iu_type iu_type)
1719 struct srp_target_port *target = ch->target;
1720 s32 rsv = (iu_type == SRP_IU_TSK_MGMT) ? 0 : SRP_TSK_MGMT_SQ_SIZE;
1723 ib_process_cq_direct(ch->send_cq, -1);
1725 if (list_empty(&ch->free_tx))
1728 /* Initiator responses to target requests do not consume credits */
1729 if (iu_type != SRP_IU_RSP) {
1730 if (ch->req_lim <= rsv) {
1731 ++target->zero_req_lim;
1738 iu = list_first_entry(&ch->free_tx, struct srp_iu, list);
1739 list_del(&iu->list);
1743 static void srp_send_done(struct ib_cq *cq, struct ib_wc *wc)
1745 struct srp_iu *iu = container_of(wc->wr_cqe, struct srp_iu, cqe);
1746 struct srp_rdma_ch *ch = cq->cq_context;
1748 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1749 srp_handle_qp_err(cq, wc, "SEND");
1753 list_add(&iu->list, &ch->free_tx);
1756 static int srp_post_send(struct srp_rdma_ch *ch, struct srp_iu *iu, int len)
1758 struct srp_target_port *target = ch->target;
1760 struct ib_send_wr wr, *bad_wr;
1762 list.addr = iu->dma;
1764 list.lkey = target->lkey;
1766 iu->cqe.done = srp_send_done;
1769 wr.wr_cqe = &iu->cqe;
1772 wr.opcode = IB_WR_SEND;
1773 wr.send_flags = IB_SEND_SIGNALED;
1775 return ib_post_send(ch->qp, &wr, &bad_wr);
1778 static int srp_post_recv(struct srp_rdma_ch *ch, struct srp_iu *iu)
1780 struct srp_target_port *target = ch->target;
1781 struct ib_recv_wr wr, *bad_wr;
1784 list.addr = iu->dma;
1785 list.length = iu->size;
1786 list.lkey = target->lkey;
1788 iu->cqe.done = srp_recv_done;
1791 wr.wr_cqe = &iu->cqe;
1795 return ib_post_recv(ch->qp, &wr, &bad_wr);
1798 static void srp_process_rsp(struct srp_rdma_ch *ch, struct srp_rsp *rsp)
1800 struct srp_target_port *target = ch->target;
1801 struct srp_request *req;
1802 struct scsi_cmnd *scmnd;
1803 unsigned long flags;
1805 if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
1806 spin_lock_irqsave(&ch->lock, flags);
1807 ch->req_lim += be32_to_cpu(rsp->req_lim_delta);
1808 spin_unlock_irqrestore(&ch->lock, flags);
1810 ch->tsk_mgmt_status = -1;
1811 if (be32_to_cpu(rsp->resp_data_len) >= 4)
1812 ch->tsk_mgmt_status = rsp->data[3];
1813 complete(&ch->tsk_mgmt_done);
1815 scmnd = scsi_host_find_tag(target->scsi_host, rsp->tag);
1817 req = (void *)scmnd->host_scribble;
1818 scmnd = srp_claim_req(ch, req, NULL, scmnd);
1821 shost_printk(KERN_ERR, target->scsi_host,
1822 "Null scmnd for RSP w/tag %#016llx received on ch %td / QP %#x\n",
1823 rsp->tag, ch - target->ch, ch->qp->qp_num);
1825 spin_lock_irqsave(&ch->lock, flags);
1826 ch->req_lim += be32_to_cpu(rsp->req_lim_delta);
1827 spin_unlock_irqrestore(&ch->lock, flags);
1831 scmnd->result = rsp->status;
1833 if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
1834 memcpy(scmnd->sense_buffer, rsp->data +
1835 be32_to_cpu(rsp->resp_data_len),
1836 min_t(int, be32_to_cpu(rsp->sense_data_len),
1837 SCSI_SENSE_BUFFERSIZE));
1840 if (unlikely(rsp->flags & SRP_RSP_FLAG_DIUNDER))
1841 scsi_set_resid(scmnd, be32_to_cpu(rsp->data_in_res_cnt));
1842 else if (unlikely(rsp->flags & SRP_RSP_FLAG_DIOVER))
1843 scsi_set_resid(scmnd, -be32_to_cpu(rsp->data_in_res_cnt));
1844 else if (unlikely(rsp->flags & SRP_RSP_FLAG_DOUNDER))
1845 scsi_set_resid(scmnd, be32_to_cpu(rsp->data_out_res_cnt));
1846 else if (unlikely(rsp->flags & SRP_RSP_FLAG_DOOVER))
1847 scsi_set_resid(scmnd, -be32_to_cpu(rsp->data_out_res_cnt));
1849 srp_free_req(ch, req, scmnd,
1850 be32_to_cpu(rsp->req_lim_delta));
1852 scmnd->host_scribble = NULL;
1853 scmnd->scsi_done(scmnd);
1857 static int srp_response_common(struct srp_rdma_ch *ch, s32 req_delta,
1860 struct srp_target_port *target = ch->target;
1861 struct ib_device *dev = target->srp_host->srp_dev->dev;
1862 unsigned long flags;
1866 spin_lock_irqsave(&ch->lock, flags);
1867 ch->req_lim += req_delta;
1868 iu = __srp_get_tx_iu(ch, SRP_IU_RSP);
1869 spin_unlock_irqrestore(&ch->lock, flags);
1872 shost_printk(KERN_ERR, target->scsi_host, PFX
1873 "no IU available to send response\n");
1877 ib_dma_sync_single_for_cpu(dev, iu->dma, len, DMA_TO_DEVICE);
1878 memcpy(iu->buf, rsp, len);
1879 ib_dma_sync_single_for_device(dev, iu->dma, len, DMA_TO_DEVICE);
1881 err = srp_post_send(ch, iu, len);
1883 shost_printk(KERN_ERR, target->scsi_host, PFX
1884 "unable to post response: %d\n", err);
1885 srp_put_tx_iu(ch, iu, SRP_IU_RSP);
1891 static void srp_process_cred_req(struct srp_rdma_ch *ch,
1892 struct srp_cred_req *req)
1894 struct srp_cred_rsp rsp = {
1895 .opcode = SRP_CRED_RSP,
1898 s32 delta = be32_to_cpu(req->req_lim_delta);
1900 if (srp_response_common(ch, delta, &rsp, sizeof(rsp)))
1901 shost_printk(KERN_ERR, ch->target->scsi_host, PFX
1902 "problems processing SRP_CRED_REQ\n");
1905 static void srp_process_aer_req(struct srp_rdma_ch *ch,
1906 struct srp_aer_req *req)
1908 struct srp_target_port *target = ch->target;
1909 struct srp_aer_rsp rsp = {
1910 .opcode = SRP_AER_RSP,
1913 s32 delta = be32_to_cpu(req->req_lim_delta);
1915 shost_printk(KERN_ERR, target->scsi_host, PFX
1916 "ignoring AER for LUN %llu\n", scsilun_to_int(&req->lun));
1918 if (srp_response_common(ch, delta, &rsp, sizeof(rsp)))
1919 shost_printk(KERN_ERR, target->scsi_host, PFX
1920 "problems processing SRP_AER_REQ\n");
1923 static void srp_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1925 struct srp_iu *iu = container_of(wc->wr_cqe, struct srp_iu, cqe);
1926 struct srp_rdma_ch *ch = cq->cq_context;
1927 struct srp_target_port *target = ch->target;
1928 struct ib_device *dev = target->srp_host->srp_dev->dev;
1932 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1933 srp_handle_qp_err(cq, wc, "RECV");
1937 ib_dma_sync_single_for_cpu(dev, iu->dma, ch->max_ti_iu_len,
1940 opcode = *(u8 *) iu->buf;
1943 shost_printk(KERN_ERR, target->scsi_host,
1944 PFX "recv completion, opcode 0x%02x\n", opcode);
1945 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 8, 1,
1946 iu->buf, wc->byte_len, true);
1951 srp_process_rsp(ch, iu->buf);
1955 srp_process_cred_req(ch, iu->buf);
1959 srp_process_aer_req(ch, iu->buf);
1963 /* XXX Handle target logout */
1964 shost_printk(KERN_WARNING, target->scsi_host,
1965 PFX "Got target logout request\n");
1969 shost_printk(KERN_WARNING, target->scsi_host,
1970 PFX "Unhandled SRP opcode 0x%02x\n", opcode);
1974 ib_dma_sync_single_for_device(dev, iu->dma, ch->max_ti_iu_len,
1977 res = srp_post_recv(ch, iu);
1979 shost_printk(KERN_ERR, target->scsi_host,
1980 PFX "Recv failed with error code %d\n", res);
1984 * srp_tl_err_work() - handle a transport layer error
1985 * @work: Work structure embedded in an SRP target port.
1987 * Note: This function may get invoked before the rport has been created,
1988 * hence the target->rport test.
1990 static void srp_tl_err_work(struct work_struct *work)
1992 struct srp_target_port *target;
1994 target = container_of(work, struct srp_target_port, tl_err_work);
1996 srp_start_tl_fail_timers(target->rport);
1999 static void srp_handle_qp_err(struct ib_cq *cq, struct ib_wc *wc,
2002 struct srp_rdma_ch *ch = cq->cq_context;
2003 struct srp_target_port *target = ch->target;
2005 if (ch->connected && !target->qp_in_error) {
2006 shost_printk(KERN_ERR, target->scsi_host,
2007 PFX "failed %s status %s (%d) for CQE %p\n",
2008 opname, ib_wc_status_msg(wc->status), wc->status,
2010 queue_work(system_long_wq, &target->tl_err_work);
2012 target->qp_in_error = true;
2015 static int srp_queuecommand(struct Scsi_Host *shost, struct scsi_cmnd *scmnd)
2017 struct srp_target_port *target = host_to_target(shost);
2018 struct srp_rport *rport = target->rport;
2019 struct srp_rdma_ch *ch;
2020 struct srp_request *req;
2022 struct srp_cmd *cmd;
2023 struct ib_device *dev;
2024 unsigned long flags;
2028 const bool in_scsi_eh = !in_interrupt() && current == shost->ehandler;
2031 * The SCSI EH thread is the only context from which srp_queuecommand()
2032 * can get invoked for blocked devices (SDEV_BLOCK /
2033 * SDEV_CREATED_BLOCK). Avoid racing with srp_reconnect_rport() by
2034 * locking the rport mutex if invoked from inside the SCSI EH.
2037 mutex_lock(&rport->mutex);
2039 scmnd->result = srp_chkready(target->rport);
2040 if (unlikely(scmnd->result))
2043 WARN_ON_ONCE(scmnd->request->tag < 0);
2044 tag = blk_mq_unique_tag(scmnd->request);
2045 ch = &target->ch[blk_mq_unique_tag_to_hwq(tag)];
2046 idx = blk_mq_unique_tag_to_tag(tag);
2047 WARN_ONCE(idx >= target->req_ring_size, "%s: tag %#x: idx %d >= %d\n",
2048 dev_name(&shost->shost_gendev), tag, idx,
2049 target->req_ring_size);
2051 spin_lock_irqsave(&ch->lock, flags);
2052 iu = __srp_get_tx_iu(ch, SRP_IU_CMD);
2053 spin_unlock_irqrestore(&ch->lock, flags);
2058 req = &ch->req_ring[idx];
2059 dev = target->srp_host->srp_dev->dev;
2060 ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_iu_len,
2063 scmnd->host_scribble = (void *) req;
2066 memset(cmd, 0, sizeof *cmd);
2068 cmd->opcode = SRP_CMD;
2069 int_to_scsilun(scmnd->device->lun, &cmd->lun);
2071 memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
2076 len = srp_map_data(scmnd, ch, req);
2078 shost_printk(KERN_ERR, target->scsi_host,
2079 PFX "Failed to map data (%d)\n", len);
2081 * If we ran out of memory descriptors (-ENOMEM) because an
2082 * application is queuing many requests with more than
2083 * max_pages_per_mr sg-list elements, tell the SCSI mid-layer
2084 * to reduce queue depth temporarily.
2086 scmnd->result = len == -ENOMEM ?
2087 DID_OK << 16 | QUEUE_FULL << 1 : DID_ERROR << 16;
2091 ib_dma_sync_single_for_device(dev, iu->dma, target->max_iu_len,
2094 if (srp_post_send(ch, iu, len)) {
2095 shost_printk(KERN_ERR, target->scsi_host, PFX "Send failed\n");
2103 mutex_unlock(&rport->mutex);
2108 srp_unmap_data(scmnd, ch, req);
2111 srp_put_tx_iu(ch, iu, SRP_IU_CMD);
2114 * Avoid that the loops that iterate over the request ring can
2115 * encounter a dangling SCSI command pointer.
2120 if (scmnd->result) {
2121 scmnd->scsi_done(scmnd);
2124 ret = SCSI_MLQUEUE_HOST_BUSY;
2131 * Note: the resources allocated in this function are freed in
2134 static int srp_alloc_iu_bufs(struct srp_rdma_ch *ch)
2136 struct srp_target_port *target = ch->target;
2139 ch->rx_ring = kcalloc(target->queue_size, sizeof(*ch->rx_ring),
2143 ch->tx_ring = kcalloc(target->queue_size, sizeof(*ch->tx_ring),
2148 for (i = 0; i < target->queue_size; ++i) {
2149 ch->rx_ring[i] = srp_alloc_iu(target->srp_host,
2151 GFP_KERNEL, DMA_FROM_DEVICE);
2152 if (!ch->rx_ring[i])
2156 for (i = 0; i < target->queue_size; ++i) {
2157 ch->tx_ring[i] = srp_alloc_iu(target->srp_host,
2159 GFP_KERNEL, DMA_TO_DEVICE);
2160 if (!ch->tx_ring[i])
2163 list_add(&ch->tx_ring[i]->list, &ch->free_tx);
2169 for (i = 0; i < target->queue_size; ++i) {
2170 srp_free_iu(target->srp_host, ch->rx_ring[i]);
2171 srp_free_iu(target->srp_host, ch->tx_ring[i]);
2184 static uint32_t srp_compute_rq_tmo(struct ib_qp_attr *qp_attr, int attr_mask)
2186 uint64_t T_tr_ns, max_compl_time_ms;
2187 uint32_t rq_tmo_jiffies;
2190 * According to section 11.2.4.2 in the IBTA spec (Modify Queue Pair,
2191 * table 91), both the QP timeout and the retry count have to be set
2192 * for RC QP's during the RTR to RTS transition.
2194 WARN_ON_ONCE((attr_mask & (IB_QP_TIMEOUT | IB_QP_RETRY_CNT)) !=
2195 (IB_QP_TIMEOUT | IB_QP_RETRY_CNT));
2198 * Set target->rq_tmo_jiffies to one second more than the largest time
2199 * it can take before an error completion is generated. See also
2200 * C9-140..142 in the IBTA spec for more information about how to
2201 * convert the QP Local ACK Timeout value to nanoseconds.
2203 T_tr_ns = 4096 * (1ULL << qp_attr->timeout);
2204 max_compl_time_ms = qp_attr->retry_cnt * 4 * T_tr_ns;
2205 do_div(max_compl_time_ms, NSEC_PER_MSEC);
2206 rq_tmo_jiffies = msecs_to_jiffies(max_compl_time_ms + 1000);
2208 return rq_tmo_jiffies;
2211 static void srp_cm_rep_handler(struct ib_cm_id *cm_id,
2212 const struct srp_login_rsp *lrsp,
2213 struct srp_rdma_ch *ch)
2215 struct srp_target_port *target = ch->target;
2216 struct ib_qp_attr *qp_attr = NULL;
2221 if (lrsp->opcode == SRP_LOGIN_RSP) {
2222 ch->max_ti_iu_len = be32_to_cpu(lrsp->max_ti_iu_len);
2223 ch->req_lim = be32_to_cpu(lrsp->req_lim_delta);
2226 * Reserve credits for task management so we don't
2227 * bounce requests back to the SCSI mid-layer.
2229 target->scsi_host->can_queue
2230 = min(ch->req_lim - SRP_TSK_MGMT_SQ_SIZE,
2231 target->scsi_host->can_queue);
2232 target->scsi_host->cmd_per_lun
2233 = min_t(int, target->scsi_host->can_queue,
2234 target->scsi_host->cmd_per_lun);
2236 shost_printk(KERN_WARNING, target->scsi_host,
2237 PFX "Unhandled RSP opcode %#x\n", lrsp->opcode);
2243 ret = srp_alloc_iu_bufs(ch);
2249 qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
2253 qp_attr->qp_state = IB_QPS_RTR;
2254 ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
2258 ret = ib_modify_qp(ch->qp, qp_attr, attr_mask);
2262 for (i = 0; i < target->queue_size; i++) {
2263 struct srp_iu *iu = ch->rx_ring[i];
2265 ret = srp_post_recv(ch, iu);
2270 qp_attr->qp_state = IB_QPS_RTS;
2271 ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
2275 target->rq_tmo_jiffies = srp_compute_rq_tmo(qp_attr, attr_mask);
2277 ret = ib_modify_qp(ch->qp, qp_attr, attr_mask);
2281 ret = ib_send_cm_rtu(cm_id, NULL, 0);
2290 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
2291 struct ib_cm_event *event,
2292 struct srp_rdma_ch *ch)
2294 struct srp_target_port *target = ch->target;
2295 struct Scsi_Host *shost = target->scsi_host;
2296 struct ib_class_port_info *cpi;
2299 switch (event->param.rej_rcvd.reason) {
2300 case IB_CM_REJ_PORT_CM_REDIRECT:
2301 cpi = event->param.rej_rcvd.ari;
2302 ch->path.dlid = cpi->redirect_lid;
2303 ch->path.pkey = cpi->redirect_pkey;
2304 cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
2305 memcpy(ch->path.dgid.raw, cpi->redirect_gid, 16);
2307 ch->status = ch->path.dlid ?
2308 SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
2311 case IB_CM_REJ_PORT_REDIRECT:
2312 if (srp_target_is_topspin(target)) {
2314 * Topspin/Cisco SRP gateways incorrectly send
2315 * reject reason code 25 when they mean 24
2318 memcpy(ch->path.dgid.raw,
2319 event->param.rej_rcvd.ari, 16);
2321 shost_printk(KERN_DEBUG, shost,
2322 PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
2323 be64_to_cpu(ch->path.dgid.global.subnet_prefix),
2324 be64_to_cpu(ch->path.dgid.global.interface_id));
2326 ch->status = SRP_PORT_REDIRECT;
2328 shost_printk(KERN_WARNING, shost,
2329 " REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
2330 ch->status = -ECONNRESET;
2334 case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
2335 shost_printk(KERN_WARNING, shost,
2336 " REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
2337 ch->status = -ECONNRESET;
2340 case IB_CM_REJ_CONSUMER_DEFINED:
2341 opcode = *(u8 *) event->private_data;
2342 if (opcode == SRP_LOGIN_REJ) {
2343 struct srp_login_rej *rej = event->private_data;
2344 u32 reason = be32_to_cpu(rej->reason);
2346 if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
2347 shost_printk(KERN_WARNING, shost,
2348 PFX "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
2350 shost_printk(KERN_WARNING, shost, PFX
2351 "SRP LOGIN from %pI6 to %pI6 REJECTED, reason 0x%08x\n",
2353 target->orig_dgid.raw, reason);
2355 shost_printk(KERN_WARNING, shost,
2356 " REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
2357 " opcode 0x%02x\n", opcode);
2358 ch->status = -ECONNRESET;
2361 case IB_CM_REJ_STALE_CONN:
2362 shost_printk(KERN_WARNING, shost, " REJ reason: stale connection\n");
2363 ch->status = SRP_STALE_CONN;
2367 shost_printk(KERN_WARNING, shost, " REJ reason 0x%x\n",
2368 event->param.rej_rcvd.reason);
2369 ch->status = -ECONNRESET;
2373 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
2375 struct srp_rdma_ch *ch = cm_id->context;
2376 struct srp_target_port *target = ch->target;
2379 switch (event->event) {
2380 case IB_CM_REQ_ERROR:
2381 shost_printk(KERN_DEBUG, target->scsi_host,
2382 PFX "Sending CM REQ failed\n");
2384 ch->status = -ECONNRESET;
2387 case IB_CM_REP_RECEIVED:
2389 srp_cm_rep_handler(cm_id, event->private_data, ch);
2392 case IB_CM_REJ_RECEIVED:
2393 shost_printk(KERN_DEBUG, target->scsi_host, PFX "REJ received\n");
2396 srp_cm_rej_handler(cm_id, event, ch);
2399 case IB_CM_DREQ_RECEIVED:
2400 shost_printk(KERN_WARNING, target->scsi_host,
2401 PFX "DREQ received - connection closed\n");
2402 ch->connected = false;
2403 if (ib_send_cm_drep(cm_id, NULL, 0))
2404 shost_printk(KERN_ERR, target->scsi_host,
2405 PFX "Sending CM DREP failed\n");
2406 queue_work(system_long_wq, &target->tl_err_work);
2409 case IB_CM_TIMEWAIT_EXIT:
2410 shost_printk(KERN_ERR, target->scsi_host,
2411 PFX "connection closed\n");
2417 case IB_CM_MRA_RECEIVED:
2418 case IB_CM_DREQ_ERROR:
2419 case IB_CM_DREP_RECEIVED:
2423 shost_printk(KERN_WARNING, target->scsi_host,
2424 PFX "Unhandled CM event %d\n", event->event);
2429 complete(&ch->done);
2435 * srp_change_queue_depth - setting device queue depth
2436 * @sdev: scsi device struct
2437 * @qdepth: requested queue depth
2439 * Returns queue depth.
2442 srp_change_queue_depth(struct scsi_device *sdev, int qdepth)
2444 if (!sdev->tagged_supported)
2446 return scsi_change_queue_depth(sdev, qdepth);
2449 static int srp_send_tsk_mgmt(struct srp_rdma_ch *ch, u64 req_tag, u64 lun,
2452 struct srp_target_port *target = ch->target;
2453 struct srp_rport *rport = target->rport;
2454 struct ib_device *dev = target->srp_host->srp_dev->dev;
2456 struct srp_tsk_mgmt *tsk_mgmt;
2458 if (!ch->connected || target->qp_in_error)
2461 init_completion(&ch->tsk_mgmt_done);
2464 * Lock the rport mutex to avoid that srp_create_ch_ib() is
2465 * invoked while a task management function is being sent.
2467 mutex_lock(&rport->mutex);
2468 spin_lock_irq(&ch->lock);
2469 iu = __srp_get_tx_iu(ch, SRP_IU_TSK_MGMT);
2470 spin_unlock_irq(&ch->lock);
2473 mutex_unlock(&rport->mutex);
2478 ib_dma_sync_single_for_cpu(dev, iu->dma, sizeof *tsk_mgmt,
2481 memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
2483 tsk_mgmt->opcode = SRP_TSK_MGMT;
2484 int_to_scsilun(lun, &tsk_mgmt->lun);
2485 tsk_mgmt->tag = req_tag | SRP_TAG_TSK_MGMT;
2486 tsk_mgmt->tsk_mgmt_func = func;
2487 tsk_mgmt->task_tag = req_tag;
2489 ib_dma_sync_single_for_device(dev, iu->dma, sizeof *tsk_mgmt,
2491 if (srp_post_send(ch, iu, sizeof(*tsk_mgmt))) {
2492 srp_put_tx_iu(ch, iu, SRP_IU_TSK_MGMT);
2493 mutex_unlock(&rport->mutex);
2497 mutex_unlock(&rport->mutex);
2499 if (!wait_for_completion_timeout(&ch->tsk_mgmt_done,
2500 msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
2506 static int srp_abort(struct scsi_cmnd *scmnd)
2508 struct srp_target_port *target = host_to_target(scmnd->device->host);
2509 struct srp_request *req = (struct srp_request *) scmnd->host_scribble;
2512 struct srp_rdma_ch *ch;
2515 shost_printk(KERN_ERR, target->scsi_host, "SRP abort called\n");
2519 tag = blk_mq_unique_tag(scmnd->request);
2520 ch_idx = blk_mq_unique_tag_to_hwq(tag);
2521 if (WARN_ON_ONCE(ch_idx >= target->ch_count))
2523 ch = &target->ch[ch_idx];
2524 if (!srp_claim_req(ch, req, NULL, scmnd))
2526 shost_printk(KERN_ERR, target->scsi_host,
2527 "Sending SRP abort for tag %#x\n", tag);
2528 if (srp_send_tsk_mgmt(ch, tag, scmnd->device->lun,
2529 SRP_TSK_ABORT_TASK) == 0)
2531 else if (target->rport->state == SRP_RPORT_LOST)
2535 srp_free_req(ch, req, scmnd, 0);
2536 scmnd->result = DID_ABORT << 16;
2537 scmnd->scsi_done(scmnd);
2542 static int srp_reset_device(struct scsi_cmnd *scmnd)
2544 struct srp_target_port *target = host_to_target(scmnd->device->host);
2545 struct srp_rdma_ch *ch;
2548 shost_printk(KERN_ERR, target->scsi_host, "SRP reset_device called\n");
2550 ch = &target->ch[0];
2551 if (srp_send_tsk_mgmt(ch, SRP_TAG_NO_REQ, scmnd->device->lun,
2554 if (ch->tsk_mgmt_status)
2557 for (i = 0; i < target->ch_count; i++) {
2558 ch = &target->ch[i];
2559 for (i = 0; i < target->req_ring_size; ++i) {
2560 struct srp_request *req = &ch->req_ring[i];
2562 srp_finish_req(ch, req, scmnd->device, DID_RESET << 16);
2569 static int srp_reset_host(struct scsi_cmnd *scmnd)
2571 struct srp_target_port *target = host_to_target(scmnd->device->host);
2573 shost_printk(KERN_ERR, target->scsi_host, PFX "SRP reset_host called\n");
2575 return srp_reconnect_rport(target->rport) == 0 ? SUCCESS : FAILED;
2578 static int srp_slave_configure(struct scsi_device *sdev)
2580 struct Scsi_Host *shost = sdev->host;
2581 struct srp_target_port *target = host_to_target(shost);
2582 struct request_queue *q = sdev->request_queue;
2583 unsigned long timeout;
2585 if (sdev->type == TYPE_DISK) {
2586 timeout = max_t(unsigned, 30 * HZ, target->rq_tmo_jiffies);
2587 blk_queue_rq_timeout(q, timeout);
2593 static ssize_t show_id_ext(struct device *dev, struct device_attribute *attr,
2596 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2598 return sprintf(buf, "0x%016llx\n", be64_to_cpu(target->id_ext));
2601 static ssize_t show_ioc_guid(struct device *dev, struct device_attribute *attr,
2604 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2606 return sprintf(buf, "0x%016llx\n", be64_to_cpu(target->ioc_guid));
2609 static ssize_t show_service_id(struct device *dev,
2610 struct device_attribute *attr, char *buf)
2612 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2614 return sprintf(buf, "0x%016llx\n", be64_to_cpu(target->service_id));
2617 static ssize_t show_pkey(struct device *dev, struct device_attribute *attr,
2620 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2622 return sprintf(buf, "0x%04x\n", be16_to_cpu(target->pkey));
2625 static ssize_t show_sgid(struct device *dev, struct device_attribute *attr,
2628 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2630 return sprintf(buf, "%pI6\n", target->sgid.raw);
2633 static ssize_t show_dgid(struct device *dev, struct device_attribute *attr,
2636 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2637 struct srp_rdma_ch *ch = &target->ch[0];
2639 return sprintf(buf, "%pI6\n", ch->path.dgid.raw);
2642 static ssize_t show_orig_dgid(struct device *dev,
2643 struct device_attribute *attr, char *buf)
2645 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2647 return sprintf(buf, "%pI6\n", target->orig_dgid.raw);
2650 static ssize_t show_req_lim(struct device *dev,
2651 struct device_attribute *attr, char *buf)
2653 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2654 struct srp_rdma_ch *ch;
2655 int i, req_lim = INT_MAX;
2657 for (i = 0; i < target->ch_count; i++) {
2658 ch = &target->ch[i];
2659 req_lim = min(req_lim, ch->req_lim);
2661 return sprintf(buf, "%d\n", req_lim);
2664 static ssize_t show_zero_req_lim(struct device *dev,
2665 struct device_attribute *attr, char *buf)
2667 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2669 return sprintf(buf, "%d\n", target->zero_req_lim);
2672 static ssize_t show_local_ib_port(struct device *dev,
2673 struct device_attribute *attr, char *buf)
2675 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2677 return sprintf(buf, "%d\n", target->srp_host->port);
2680 static ssize_t show_local_ib_device(struct device *dev,
2681 struct device_attribute *attr, char *buf)
2683 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2685 return sprintf(buf, "%s\n", target->srp_host->srp_dev->dev->name);
2688 static ssize_t show_ch_count(struct device *dev, struct device_attribute *attr,
2691 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2693 return sprintf(buf, "%d\n", target->ch_count);
2696 static ssize_t show_comp_vector(struct device *dev,
2697 struct device_attribute *attr, char *buf)
2699 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2701 return sprintf(buf, "%d\n", target->comp_vector);
2704 static ssize_t show_tl_retry_count(struct device *dev,
2705 struct device_attribute *attr, char *buf)
2707 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2709 return sprintf(buf, "%d\n", target->tl_retry_count);
2712 static ssize_t show_cmd_sg_entries(struct device *dev,
2713 struct device_attribute *attr, char *buf)
2715 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2717 return sprintf(buf, "%u\n", target->cmd_sg_cnt);
2720 static ssize_t show_allow_ext_sg(struct device *dev,
2721 struct device_attribute *attr, char *buf)
2723 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2725 return sprintf(buf, "%s\n", target->allow_ext_sg ? "true" : "false");
2728 static DEVICE_ATTR(id_ext, S_IRUGO, show_id_ext, NULL);
2729 static DEVICE_ATTR(ioc_guid, S_IRUGO, show_ioc_guid, NULL);
2730 static DEVICE_ATTR(service_id, S_IRUGO, show_service_id, NULL);
2731 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
2732 static DEVICE_ATTR(sgid, S_IRUGO, show_sgid, NULL);
2733 static DEVICE_ATTR(dgid, S_IRUGO, show_dgid, NULL);
2734 static DEVICE_ATTR(orig_dgid, S_IRUGO, show_orig_dgid, NULL);
2735 static DEVICE_ATTR(req_lim, S_IRUGO, show_req_lim, NULL);
2736 static DEVICE_ATTR(zero_req_lim, S_IRUGO, show_zero_req_lim, NULL);
2737 static DEVICE_ATTR(local_ib_port, S_IRUGO, show_local_ib_port, NULL);
2738 static DEVICE_ATTR(local_ib_device, S_IRUGO, show_local_ib_device, NULL);
2739 static DEVICE_ATTR(ch_count, S_IRUGO, show_ch_count, NULL);
2740 static DEVICE_ATTR(comp_vector, S_IRUGO, show_comp_vector, NULL);
2741 static DEVICE_ATTR(tl_retry_count, S_IRUGO, show_tl_retry_count, NULL);
2742 static DEVICE_ATTR(cmd_sg_entries, S_IRUGO, show_cmd_sg_entries, NULL);
2743 static DEVICE_ATTR(allow_ext_sg, S_IRUGO, show_allow_ext_sg, NULL);
2745 static struct device_attribute *srp_host_attrs[] = {
2748 &dev_attr_service_id,
2752 &dev_attr_orig_dgid,
2754 &dev_attr_zero_req_lim,
2755 &dev_attr_local_ib_port,
2756 &dev_attr_local_ib_device,
2758 &dev_attr_comp_vector,
2759 &dev_attr_tl_retry_count,
2760 &dev_attr_cmd_sg_entries,
2761 &dev_attr_allow_ext_sg,
2765 static struct scsi_host_template srp_template = {
2766 .module = THIS_MODULE,
2767 .name = "InfiniBand SRP initiator",
2768 .proc_name = DRV_NAME,
2769 .slave_configure = srp_slave_configure,
2770 .info = srp_target_info,
2771 .queuecommand = srp_queuecommand,
2772 .change_queue_depth = srp_change_queue_depth,
2773 .eh_abort_handler = srp_abort,
2774 .eh_device_reset_handler = srp_reset_device,
2775 .eh_host_reset_handler = srp_reset_host,
2776 .skip_settle_delay = true,
2777 .sg_tablesize = SRP_DEF_SG_TABLESIZE,
2778 .can_queue = SRP_DEFAULT_CMD_SQ_SIZE,
2780 .cmd_per_lun = SRP_DEFAULT_CMD_SQ_SIZE,
2781 .use_clustering = ENABLE_CLUSTERING,
2782 .shost_attrs = srp_host_attrs,
2783 .track_queue_depth = 1,
2786 static int srp_sdev_count(struct Scsi_Host *host)
2788 struct scsi_device *sdev;
2791 shost_for_each_device(sdev, host)
2799 * < 0 upon failure. Caller is responsible for SRP target port cleanup.
2800 * 0 and target->state == SRP_TARGET_REMOVED if asynchronous target port
2801 * removal has been scheduled.
2802 * 0 and target->state != SRP_TARGET_REMOVED upon success.
2804 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
2806 struct srp_rport_identifiers ids;
2807 struct srp_rport *rport;
2809 target->state = SRP_TARGET_SCANNING;
2810 sprintf(target->target_name, "SRP.T10:%016llX",
2811 be64_to_cpu(target->id_ext));
2813 if (scsi_add_host(target->scsi_host, host->srp_dev->dev->dma_device))
2816 memcpy(ids.port_id, &target->id_ext, 8);
2817 memcpy(ids.port_id + 8, &target->ioc_guid, 8);
2818 ids.roles = SRP_RPORT_ROLE_TARGET;
2819 rport = srp_rport_add(target->scsi_host, &ids);
2820 if (IS_ERR(rport)) {
2821 scsi_remove_host(target->scsi_host);
2822 return PTR_ERR(rport);
2825 rport->lld_data = target;
2826 target->rport = rport;
2828 spin_lock(&host->target_lock);
2829 list_add_tail(&target->list, &host->target_list);
2830 spin_unlock(&host->target_lock);
2832 scsi_scan_target(&target->scsi_host->shost_gendev,
2833 0, target->scsi_id, SCAN_WILD_CARD, 0);
2835 if (srp_connected_ch(target) < target->ch_count ||
2836 target->qp_in_error) {
2837 shost_printk(KERN_INFO, target->scsi_host,
2838 PFX "SCSI scan failed - removing SCSI host\n");
2839 srp_queue_remove_work(target);
2843 pr_debug(PFX "%s: SCSI scan succeeded - detected %d LUNs\n",
2844 dev_name(&target->scsi_host->shost_gendev),
2845 srp_sdev_count(target->scsi_host));
2847 spin_lock_irq(&target->lock);
2848 if (target->state == SRP_TARGET_SCANNING)
2849 target->state = SRP_TARGET_LIVE;
2850 spin_unlock_irq(&target->lock);
2856 static void srp_release_dev(struct device *dev)
2858 struct srp_host *host =
2859 container_of(dev, struct srp_host, dev);
2861 complete(&host->released);
2864 static struct class srp_class = {
2865 .name = "infiniband_srp",
2866 .dev_release = srp_release_dev
2870 * srp_conn_unique() - check whether the connection to a target is unique
2872 * @target: SRP target port.
2874 static bool srp_conn_unique(struct srp_host *host,
2875 struct srp_target_port *target)
2877 struct srp_target_port *t;
2880 if (target->state == SRP_TARGET_REMOVED)
2885 spin_lock(&host->target_lock);
2886 list_for_each_entry(t, &host->target_list, list) {
2888 target->id_ext == t->id_ext &&
2889 target->ioc_guid == t->ioc_guid &&
2890 target->initiator_ext == t->initiator_ext) {
2895 spin_unlock(&host->target_lock);
2902 * Target ports are added by writing
2904 * id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
2905 * pkey=<P_Key>,service_id=<service ID>
2907 * to the add_target sysfs attribute.
2911 SRP_OPT_ID_EXT = 1 << 0,
2912 SRP_OPT_IOC_GUID = 1 << 1,
2913 SRP_OPT_DGID = 1 << 2,
2914 SRP_OPT_PKEY = 1 << 3,
2915 SRP_OPT_SERVICE_ID = 1 << 4,
2916 SRP_OPT_MAX_SECT = 1 << 5,
2917 SRP_OPT_MAX_CMD_PER_LUN = 1 << 6,
2918 SRP_OPT_IO_CLASS = 1 << 7,
2919 SRP_OPT_INITIATOR_EXT = 1 << 8,
2920 SRP_OPT_CMD_SG_ENTRIES = 1 << 9,
2921 SRP_OPT_ALLOW_EXT_SG = 1 << 10,
2922 SRP_OPT_SG_TABLESIZE = 1 << 11,
2923 SRP_OPT_COMP_VECTOR = 1 << 12,
2924 SRP_OPT_TL_RETRY_COUNT = 1 << 13,
2925 SRP_OPT_QUEUE_SIZE = 1 << 14,
2926 SRP_OPT_ALL = (SRP_OPT_ID_EXT |
2930 SRP_OPT_SERVICE_ID),
2933 static const match_table_t srp_opt_tokens = {
2934 { SRP_OPT_ID_EXT, "id_ext=%s" },
2935 { SRP_OPT_IOC_GUID, "ioc_guid=%s" },
2936 { SRP_OPT_DGID, "dgid=%s" },
2937 { SRP_OPT_PKEY, "pkey=%x" },
2938 { SRP_OPT_SERVICE_ID, "service_id=%s" },
2939 { SRP_OPT_MAX_SECT, "max_sect=%d" },
2940 { SRP_OPT_MAX_CMD_PER_LUN, "max_cmd_per_lun=%d" },
2941 { SRP_OPT_IO_CLASS, "io_class=%x" },
2942 { SRP_OPT_INITIATOR_EXT, "initiator_ext=%s" },
2943 { SRP_OPT_CMD_SG_ENTRIES, "cmd_sg_entries=%u" },
2944 { SRP_OPT_ALLOW_EXT_SG, "allow_ext_sg=%u" },
2945 { SRP_OPT_SG_TABLESIZE, "sg_tablesize=%u" },
2946 { SRP_OPT_COMP_VECTOR, "comp_vector=%u" },
2947 { SRP_OPT_TL_RETRY_COUNT, "tl_retry_count=%u" },
2948 { SRP_OPT_QUEUE_SIZE, "queue_size=%d" },
2949 { SRP_OPT_ERR, NULL }
2952 static int srp_parse_options(const char *buf, struct srp_target_port *target)
2954 char *options, *sep_opt;
2957 substring_t args[MAX_OPT_ARGS];
2963 options = kstrdup(buf, GFP_KERNEL);
2968 while ((p = strsep(&sep_opt, ",\n")) != NULL) {
2972 token = match_token(p, srp_opt_tokens, args);
2976 case SRP_OPT_ID_EXT:
2977 p = match_strdup(args);
2982 target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2986 case SRP_OPT_IOC_GUID:
2987 p = match_strdup(args);
2992 target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
2997 p = match_strdup(args);
3002 if (strlen(p) != 32) {
3003 pr_warn("bad dest GID parameter '%s'\n", p);
3008 for (i = 0; i < 16; ++i) {
3009 strlcpy(dgid, p + i * 2, sizeof(dgid));
3010 if (sscanf(dgid, "%hhx",
3011 &target->orig_dgid.raw[i]) < 1) {
3021 if (match_hex(args, &token)) {
3022 pr_warn("bad P_Key parameter '%s'\n", p);
3025 target->pkey = cpu_to_be16(token);
3028 case SRP_OPT_SERVICE_ID:
3029 p = match_strdup(args);
3034 target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
3038 case SRP_OPT_MAX_SECT:
3039 if (match_int(args, &token)) {
3040 pr_warn("bad max sect parameter '%s'\n", p);
3043 target->scsi_host->max_sectors = token;
3046 case SRP_OPT_QUEUE_SIZE:
3047 if (match_int(args, &token) || token < 1) {
3048 pr_warn("bad queue_size parameter '%s'\n", p);
3051 target->scsi_host->can_queue = token;
3052 target->queue_size = token + SRP_RSP_SQ_SIZE +
3053 SRP_TSK_MGMT_SQ_SIZE;
3054 if (!(opt_mask & SRP_OPT_MAX_CMD_PER_LUN))
3055 target->scsi_host->cmd_per_lun = token;
3058 case SRP_OPT_MAX_CMD_PER_LUN:
3059 if (match_int(args, &token) || token < 1) {
3060 pr_warn("bad max cmd_per_lun parameter '%s'\n",
3064 target->scsi_host->cmd_per_lun = token;
3067 case SRP_OPT_IO_CLASS:
3068 if (match_hex(args, &token)) {
3069 pr_warn("bad IO class parameter '%s'\n", p);
3072 if (token != SRP_REV10_IB_IO_CLASS &&
3073 token != SRP_REV16A_IB_IO_CLASS) {
3074 pr_warn("unknown IO class parameter value %x specified (use %x or %x).\n",
3075 token, SRP_REV10_IB_IO_CLASS,
3076 SRP_REV16A_IB_IO_CLASS);
3079 target->io_class = token;
3082 case SRP_OPT_INITIATOR_EXT:
3083 p = match_strdup(args);
3088 target->initiator_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
3092 case SRP_OPT_CMD_SG_ENTRIES:
3093 if (match_int(args, &token) || token < 1 || token > 255) {
3094 pr_warn("bad max cmd_sg_entries parameter '%s'\n",
3098 target->cmd_sg_cnt = token;
3101 case SRP_OPT_ALLOW_EXT_SG:
3102 if (match_int(args, &token)) {
3103 pr_warn("bad allow_ext_sg parameter '%s'\n", p);
3106 target->allow_ext_sg = !!token;
3109 case SRP_OPT_SG_TABLESIZE:
3110 if (match_int(args, &token) || token < 1 ||
3111 token > SCSI_MAX_SG_CHAIN_SEGMENTS) {
3112 pr_warn("bad max sg_tablesize parameter '%s'\n",
3116 target->sg_tablesize = token;
3119 case SRP_OPT_COMP_VECTOR:
3120 if (match_int(args, &token) || token < 0) {
3121 pr_warn("bad comp_vector parameter '%s'\n", p);
3124 target->comp_vector = token;
3127 case SRP_OPT_TL_RETRY_COUNT:
3128 if (match_int(args, &token) || token < 2 || token > 7) {
3129 pr_warn("bad tl_retry_count parameter '%s' (must be a number between 2 and 7)\n",
3133 target->tl_retry_count = token;
3137 pr_warn("unknown parameter or missing value '%s' in target creation request\n",
3143 if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
3146 for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
3147 if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
3148 !(srp_opt_tokens[i].token & opt_mask))
3149 pr_warn("target creation request is missing parameter '%s'\n",
3150 srp_opt_tokens[i].pattern);
3152 if (target->scsi_host->cmd_per_lun > target->scsi_host->can_queue
3153 && (opt_mask & SRP_OPT_MAX_CMD_PER_LUN))
3154 pr_warn("cmd_per_lun = %d > queue_size = %d\n",
3155 target->scsi_host->cmd_per_lun,
3156 target->scsi_host->can_queue);
3163 static ssize_t srp_create_target(struct device *dev,
3164 struct device_attribute *attr,
3165 const char *buf, size_t count)
3167 struct srp_host *host =
3168 container_of(dev, struct srp_host, dev);
3169 struct Scsi_Host *target_host;
3170 struct srp_target_port *target;
3171 struct srp_rdma_ch *ch;
3172 struct srp_device *srp_dev = host->srp_dev;
3173 struct ib_device *ibdev = srp_dev->dev;
3174 int ret, node_idx, node, cpu, i;
3175 bool multich = false;
3177 target_host = scsi_host_alloc(&srp_template,
3178 sizeof (struct srp_target_port));
3182 target_host->transportt = ib_srp_transport_template;
3183 target_host->max_channel = 0;
3184 target_host->max_id = 1;
3185 target_host->max_lun = -1LL;
3186 target_host->max_cmd_len = sizeof ((struct srp_cmd *) (void *) 0L)->cdb;
3188 target = host_to_target(target_host);
3190 target->io_class = SRP_REV16A_IB_IO_CLASS;
3191 target->scsi_host = target_host;
3192 target->srp_host = host;
3193 target->lkey = host->srp_dev->pd->local_dma_lkey;
3194 target->global_mr = host->srp_dev->global_mr;
3195 target->cmd_sg_cnt = cmd_sg_entries;
3196 target->sg_tablesize = indirect_sg_entries ? : cmd_sg_entries;
3197 target->allow_ext_sg = allow_ext_sg;
3198 target->tl_retry_count = 7;
3199 target->queue_size = SRP_DEFAULT_QUEUE_SIZE;
3202 * Avoid that the SCSI host can be removed by srp_remove_target()
3203 * before this function returns.
3205 scsi_host_get(target->scsi_host);
3207 mutex_lock(&host->add_target_mutex);
3209 ret = srp_parse_options(buf, target);
3213 target->req_ring_size = target->queue_size - SRP_TSK_MGMT_SQ_SIZE;
3215 if (!srp_conn_unique(target->srp_host, target)) {
3216 shost_printk(KERN_INFO, target->scsi_host,
3217 PFX "Already connected to target port with id_ext=%016llx;ioc_guid=%016llx;initiator_ext=%016llx\n",
3218 be64_to_cpu(target->id_ext),
3219 be64_to_cpu(target->ioc_guid),
3220 be64_to_cpu(target->initiator_ext));
3225 if (!srp_dev->has_fmr && !srp_dev->has_fr && !target->allow_ext_sg &&
3226 target->cmd_sg_cnt < target->sg_tablesize) {
3227 pr_warn("No MR pool and no external indirect descriptors, limiting sg_tablesize to cmd_sg_cnt\n");
3228 target->sg_tablesize = target->cmd_sg_cnt;
3231 target_host->sg_tablesize = target->sg_tablesize;
3232 target->mr_pool_size = target->scsi_host->can_queue;
3233 target->indirect_size = target->sg_tablesize *
3234 sizeof (struct srp_direct_buf);
3235 target->max_iu_len = sizeof (struct srp_cmd) +
3236 sizeof (struct srp_indirect_buf) +
3237 target->cmd_sg_cnt * sizeof (struct srp_direct_buf);
3239 INIT_WORK(&target->tl_err_work, srp_tl_err_work);
3240 INIT_WORK(&target->remove_work, srp_remove_work);
3241 spin_lock_init(&target->lock);
3242 ret = ib_query_gid(ibdev, host->port, 0, &target->sgid, NULL);
3247 target->ch_count = max_t(unsigned, num_online_nodes(),
3249 min(4 * num_online_nodes(),
3250 ibdev->num_comp_vectors),
3251 num_online_cpus()));
3252 target->ch = kcalloc(target->ch_count, sizeof(*target->ch),
3258 for_each_online_node(node) {
3259 const int ch_start = (node_idx * target->ch_count /
3260 num_online_nodes());
3261 const int ch_end = ((node_idx + 1) * target->ch_count /
3262 num_online_nodes());
3263 const int cv_start = (node_idx * ibdev->num_comp_vectors /
3264 num_online_nodes() + target->comp_vector)
3265 % ibdev->num_comp_vectors;
3266 const int cv_end = ((node_idx + 1) * ibdev->num_comp_vectors /
3267 num_online_nodes() + target->comp_vector)
3268 % ibdev->num_comp_vectors;
3271 for_each_online_cpu(cpu) {
3272 if (cpu_to_node(cpu) != node)
3274 if (ch_start + cpu_idx >= ch_end)
3276 ch = &target->ch[ch_start + cpu_idx];
3277 ch->target = target;
3278 ch->comp_vector = cv_start == cv_end ? cv_start :
3279 cv_start + cpu_idx % (cv_end - cv_start);
3280 spin_lock_init(&ch->lock);
3281 INIT_LIST_HEAD(&ch->free_tx);
3282 ret = srp_new_cm_id(ch);
3284 goto err_disconnect;
3286 ret = srp_create_ch_ib(ch);
3288 goto err_disconnect;
3290 ret = srp_alloc_req_data(ch);
3292 goto err_disconnect;
3294 ret = srp_connect_ch(ch, multich);
3296 shost_printk(KERN_ERR, target->scsi_host,
3297 PFX "Connection %d/%d failed\n",
3300 if (node_idx == 0 && cpu_idx == 0) {
3301 goto err_disconnect;
3303 srp_free_ch_ib(target, ch);
3304 srp_free_req_data(target, ch);
3305 target->ch_count = ch - target->ch;
3317 target->scsi_host->nr_hw_queues = target->ch_count;
3319 ret = srp_add_target(host, target);
3321 goto err_disconnect;
3323 if (target->state != SRP_TARGET_REMOVED) {
3324 shost_printk(KERN_DEBUG, target->scsi_host, PFX
3325 "new target: id_ext %016llx ioc_guid %016llx pkey %04x service_id %016llx sgid %pI6 dgid %pI6\n",
3326 be64_to_cpu(target->id_ext),
3327 be64_to_cpu(target->ioc_guid),
3328 be16_to_cpu(target->pkey),
3329 be64_to_cpu(target->service_id),
3330 target->sgid.raw, target->orig_dgid.raw);
3336 mutex_unlock(&host->add_target_mutex);
3338 scsi_host_put(target->scsi_host);
3340 scsi_host_put(target->scsi_host);
3345 srp_disconnect_target(target);
3347 for (i = 0; i < target->ch_count; i++) {
3348 ch = &target->ch[i];
3349 srp_free_ch_ib(target, ch);
3350 srp_free_req_data(target, ch);
3357 static DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
3359 static ssize_t show_ibdev(struct device *dev, struct device_attribute *attr,
3362 struct srp_host *host = container_of(dev, struct srp_host, dev);
3364 return sprintf(buf, "%s\n", host->srp_dev->dev->name);
3367 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
3369 static ssize_t show_port(struct device *dev, struct device_attribute *attr,
3372 struct srp_host *host = container_of(dev, struct srp_host, dev);
3374 return sprintf(buf, "%d\n", host->port);
3377 static DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
3379 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
3381 struct srp_host *host;
3383 host = kzalloc(sizeof *host, GFP_KERNEL);
3387 INIT_LIST_HEAD(&host->target_list);
3388 spin_lock_init(&host->target_lock);
3389 init_completion(&host->released);
3390 mutex_init(&host->add_target_mutex);
3391 host->srp_dev = device;
3394 host->dev.class = &srp_class;
3395 host->dev.parent = device->dev->dma_device;
3396 dev_set_name(&host->dev, "srp-%s-%d", device->dev->name, port);
3398 if (device_register(&host->dev))
3400 if (device_create_file(&host->dev, &dev_attr_add_target))
3402 if (device_create_file(&host->dev, &dev_attr_ibdev))
3404 if (device_create_file(&host->dev, &dev_attr_port))
3410 device_unregister(&host->dev);
3418 static void srp_add_one(struct ib_device *device)
3420 struct srp_device *srp_dev;
3421 struct srp_host *host;
3422 int mr_page_shift, p;
3423 u64 max_pages_per_mr;
3425 srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
3429 srp_dev->has_fmr = (device->alloc_fmr && device->dealloc_fmr &&
3430 device->map_phys_fmr && device->unmap_fmr);
3431 srp_dev->has_fr = (device->attrs.device_cap_flags &
3432 IB_DEVICE_MEM_MGT_EXTENSIONS);
3433 if (!srp_dev->has_fmr && !srp_dev->has_fr)
3434 dev_warn(&device->dev, "neither FMR nor FR is supported\n");
3436 srp_dev->use_fast_reg = (srp_dev->has_fr &&
3437 (!srp_dev->has_fmr || prefer_fr));
3438 srp_dev->use_fmr = !srp_dev->use_fast_reg && srp_dev->has_fmr;
3441 * Use the smallest page size supported by the HCA, down to a
3442 * minimum of 4096 bytes. We're unlikely to build large sglists
3443 * out of smaller entries.
3445 mr_page_shift = max(12, ffs(device->attrs.page_size_cap) - 1);
3446 srp_dev->mr_page_size = 1 << mr_page_shift;
3447 srp_dev->mr_page_mask = ~((u64) srp_dev->mr_page_size - 1);
3448 max_pages_per_mr = device->attrs.max_mr_size;
3449 do_div(max_pages_per_mr, srp_dev->mr_page_size);
3450 srp_dev->max_pages_per_mr = min_t(u64, SRP_MAX_PAGES_PER_MR,
3452 if (srp_dev->use_fast_reg) {
3453 srp_dev->max_pages_per_mr =
3454 min_t(u32, srp_dev->max_pages_per_mr,
3455 device->attrs.max_fast_reg_page_list_len);
3457 srp_dev->mr_max_size = srp_dev->mr_page_size *
3458 srp_dev->max_pages_per_mr;
3459 pr_debug("%s: mr_page_shift = %d, device->max_mr_size = %#llx, device->max_fast_reg_page_list_len = %u, max_pages_per_mr = %d, mr_max_size = %#x\n",
3460 device->name, mr_page_shift, device->attrs.max_mr_size,
3461 device->attrs.max_fast_reg_page_list_len,
3462 srp_dev->max_pages_per_mr, srp_dev->mr_max_size);
3464 INIT_LIST_HEAD(&srp_dev->dev_list);
3466 srp_dev->dev = device;
3467 srp_dev->pd = ib_alloc_pd(device);
3468 if (IS_ERR(srp_dev->pd))
3471 if (!register_always || (!srp_dev->has_fmr && !srp_dev->has_fr)) {
3472 srp_dev->global_mr = ib_get_dma_mr(srp_dev->pd,
3473 IB_ACCESS_LOCAL_WRITE |
3474 IB_ACCESS_REMOTE_READ |
3475 IB_ACCESS_REMOTE_WRITE);
3476 if (IS_ERR(srp_dev->global_mr))
3479 srp_dev->global_mr = NULL;
3482 for (p = rdma_start_port(device); p <= rdma_end_port(device); ++p) {
3483 host = srp_add_port(srp_dev, p);
3485 list_add_tail(&host->list, &srp_dev->dev_list);
3488 ib_set_client_data(device, &srp_client, srp_dev);
3492 ib_dealloc_pd(srp_dev->pd);
3498 static void srp_remove_one(struct ib_device *device, void *client_data)
3500 struct srp_device *srp_dev;
3501 struct srp_host *host, *tmp_host;
3502 struct srp_target_port *target;
3504 srp_dev = client_data;
3508 list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
3509 device_unregister(&host->dev);
3511 * Wait for the sysfs entry to go away, so that no new
3512 * target ports can be created.
3514 wait_for_completion(&host->released);
3517 * Remove all target ports.
3519 spin_lock(&host->target_lock);
3520 list_for_each_entry(target, &host->target_list, list)
3521 srp_queue_remove_work(target);
3522 spin_unlock(&host->target_lock);
3525 * Wait for tl_err and target port removal tasks.
3527 flush_workqueue(system_long_wq);
3528 flush_workqueue(srp_remove_wq);
3533 if (srp_dev->global_mr)
3534 ib_dereg_mr(srp_dev->global_mr);
3535 ib_dealloc_pd(srp_dev->pd);
3540 static struct srp_function_template ib_srp_transport_functions = {
3541 .has_rport_state = true,
3542 .reset_timer_if_blocked = true,
3543 .reconnect_delay = &srp_reconnect_delay,
3544 .fast_io_fail_tmo = &srp_fast_io_fail_tmo,
3545 .dev_loss_tmo = &srp_dev_loss_tmo,
3546 .reconnect = srp_rport_reconnect,
3547 .rport_delete = srp_rport_delete,
3548 .terminate_rport_io = srp_terminate_io,
3551 static int __init srp_init_module(void)
3555 if (srp_sg_tablesize) {
3556 pr_warn("srp_sg_tablesize is deprecated, please use cmd_sg_entries\n");
3557 if (!cmd_sg_entries)
3558 cmd_sg_entries = srp_sg_tablesize;
3561 if (!cmd_sg_entries)
3562 cmd_sg_entries = SRP_DEF_SG_TABLESIZE;
3564 if (cmd_sg_entries > 255) {
3565 pr_warn("Clamping cmd_sg_entries to 255\n");
3566 cmd_sg_entries = 255;
3569 if (!indirect_sg_entries)
3570 indirect_sg_entries = cmd_sg_entries;
3571 else if (indirect_sg_entries < cmd_sg_entries) {
3572 pr_warn("Bumping up indirect_sg_entries to match cmd_sg_entries (%u)\n",
3574 indirect_sg_entries = cmd_sg_entries;
3577 srp_remove_wq = create_workqueue("srp_remove");
3578 if (!srp_remove_wq) {
3584 ib_srp_transport_template =
3585 srp_attach_transport(&ib_srp_transport_functions);
3586 if (!ib_srp_transport_template)
3589 ret = class_register(&srp_class);
3591 pr_err("couldn't register class infiniband_srp\n");
3595 ib_sa_register_client(&srp_sa_client);
3597 ret = ib_register_client(&srp_client);
3599 pr_err("couldn't register IB client\n");
3607 ib_sa_unregister_client(&srp_sa_client);
3608 class_unregister(&srp_class);
3611 srp_release_transport(ib_srp_transport_template);
3614 destroy_workqueue(srp_remove_wq);
3618 static void __exit srp_cleanup_module(void)
3620 ib_unregister_client(&srp_client);
3621 ib_sa_unregister_client(&srp_sa_client);
3622 class_unregister(&srp_class);
3623 srp_release_transport(ib_srp_transport_template);
3624 destroy_workqueue(srp_remove_wq);
3627 module_init(srp_init_module);
3628 module_exit(srp_cleanup_module);