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;
72 static bool register_always;
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_completion(struct ib_cq *cq, void *ch_ptr);
136 static void srp_send_completion(struct ib_cq *cq, void *ch_ptr);
137 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event);
139 static struct scsi_transport_template *ib_srp_transport_template;
140 static struct workqueue_struct *srp_remove_wq;
142 static struct ib_client srp_client = {
145 .remove = srp_remove_one
148 static struct ib_sa_client srp_sa_client;
150 static int srp_tmo_get(char *buffer, const struct kernel_param *kp)
152 int tmo = *(int *)kp->arg;
155 return sprintf(buffer, "%d", tmo);
157 return sprintf(buffer, "off");
160 static int srp_tmo_set(const char *val, const struct kernel_param *kp)
164 res = srp_parse_tmo(&tmo, val);
168 if (kp->arg == &srp_reconnect_delay)
169 res = srp_tmo_valid(tmo, srp_fast_io_fail_tmo,
171 else if (kp->arg == &srp_fast_io_fail_tmo)
172 res = srp_tmo_valid(srp_reconnect_delay, tmo, srp_dev_loss_tmo);
174 res = srp_tmo_valid(srp_reconnect_delay, srp_fast_io_fail_tmo,
178 *(int *)kp->arg = tmo;
184 static const struct kernel_param_ops srp_tmo_ops = {
189 static inline struct srp_target_port *host_to_target(struct Scsi_Host *host)
191 return (struct srp_target_port *) host->hostdata;
194 static const char *srp_target_info(struct Scsi_Host *host)
196 return host_to_target(host)->target_name;
199 static int srp_target_is_topspin(struct srp_target_port *target)
201 static const u8 topspin_oui[3] = { 0x00, 0x05, 0xad };
202 static const u8 cisco_oui[3] = { 0x00, 0x1b, 0x0d };
204 return topspin_workarounds &&
205 (!memcmp(&target->ioc_guid, topspin_oui, sizeof topspin_oui) ||
206 !memcmp(&target->ioc_guid, cisco_oui, sizeof cisco_oui));
209 static struct srp_iu *srp_alloc_iu(struct srp_host *host, size_t size,
211 enum dma_data_direction direction)
215 iu = kmalloc(sizeof *iu, gfp_mask);
219 iu->buf = kzalloc(size, gfp_mask);
223 iu->dma = ib_dma_map_single(host->srp_dev->dev, iu->buf, size,
225 if (ib_dma_mapping_error(host->srp_dev->dev, iu->dma))
229 iu->direction = direction;
241 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
246 ib_dma_unmap_single(host->srp_dev->dev, iu->dma, iu->size,
252 static void srp_qp_event(struct ib_event *event, void *context)
254 pr_debug("QP event %s (%d)\n",
255 ib_event_msg(event->event), event->event);
258 static int srp_init_qp(struct srp_target_port *target,
261 struct ib_qp_attr *attr;
264 attr = kmalloc(sizeof *attr, GFP_KERNEL);
268 ret = ib_find_cached_pkey(target->srp_host->srp_dev->dev,
269 target->srp_host->port,
270 be16_to_cpu(target->pkey),
275 attr->qp_state = IB_QPS_INIT;
276 attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
277 IB_ACCESS_REMOTE_WRITE);
278 attr->port_num = target->srp_host->port;
280 ret = ib_modify_qp(qp, attr,
291 static int srp_new_cm_id(struct srp_rdma_ch *ch)
293 struct srp_target_port *target = ch->target;
294 struct ib_cm_id *new_cm_id;
296 new_cm_id = ib_create_cm_id(target->srp_host->srp_dev->dev,
298 if (IS_ERR(new_cm_id))
299 return PTR_ERR(new_cm_id);
302 ib_destroy_cm_id(ch->cm_id);
303 ch->cm_id = new_cm_id;
304 ch->path.sgid = target->sgid;
305 ch->path.dgid = target->orig_dgid;
306 ch->path.pkey = target->pkey;
307 ch->path.service_id = target->service_id;
312 static struct ib_fmr_pool *srp_alloc_fmr_pool(struct srp_target_port *target)
314 struct srp_device *dev = target->srp_host->srp_dev;
315 struct ib_fmr_pool_param fmr_param;
317 memset(&fmr_param, 0, sizeof(fmr_param));
318 fmr_param.pool_size = target->scsi_host->can_queue;
319 fmr_param.dirty_watermark = fmr_param.pool_size / 4;
321 fmr_param.max_pages_per_fmr = dev->max_pages_per_mr;
322 fmr_param.page_shift = ilog2(dev->mr_page_size);
323 fmr_param.access = (IB_ACCESS_LOCAL_WRITE |
324 IB_ACCESS_REMOTE_WRITE |
325 IB_ACCESS_REMOTE_READ);
327 return ib_create_fmr_pool(dev->pd, &fmr_param);
331 * srp_destroy_fr_pool() - free the resources owned by a pool
332 * @pool: Fast registration pool to be destroyed.
334 static void srp_destroy_fr_pool(struct srp_fr_pool *pool)
337 struct srp_fr_desc *d;
342 for (i = 0, d = &pool->desc[0]; i < pool->size; i++, d++) {
344 ib_free_fast_reg_page_list(d->frpl);
352 * srp_create_fr_pool() - allocate and initialize a pool for fast registration
353 * @device: IB device to allocate fast registration descriptors for.
354 * @pd: Protection domain associated with the FR descriptors.
355 * @pool_size: Number of descriptors to allocate.
356 * @max_page_list_len: Maximum fast registration work request page list length.
358 static struct srp_fr_pool *srp_create_fr_pool(struct ib_device *device,
359 struct ib_pd *pd, int pool_size,
360 int max_page_list_len)
362 struct srp_fr_pool *pool;
363 struct srp_fr_desc *d;
365 struct ib_fast_reg_page_list *frpl;
366 int i, ret = -EINVAL;
371 pool = kzalloc(sizeof(struct srp_fr_pool) +
372 pool_size * sizeof(struct srp_fr_desc), GFP_KERNEL);
375 pool->size = pool_size;
376 pool->max_page_list_len = max_page_list_len;
377 spin_lock_init(&pool->lock);
378 INIT_LIST_HEAD(&pool->free_list);
380 for (i = 0, d = &pool->desc[0]; i < pool->size; i++, d++) {
381 mr = ib_alloc_mr(pd, IB_MR_TYPE_MEM_REG,
388 frpl = ib_alloc_fast_reg_page_list(device, max_page_list_len);
394 list_add_tail(&d->entry, &pool->free_list);
401 srp_destroy_fr_pool(pool);
409 * srp_fr_pool_get() - obtain a descriptor suitable for fast registration
410 * @pool: Pool to obtain descriptor from.
412 static struct srp_fr_desc *srp_fr_pool_get(struct srp_fr_pool *pool)
414 struct srp_fr_desc *d = NULL;
417 spin_lock_irqsave(&pool->lock, flags);
418 if (!list_empty(&pool->free_list)) {
419 d = list_first_entry(&pool->free_list, typeof(*d), entry);
422 spin_unlock_irqrestore(&pool->lock, flags);
428 * srp_fr_pool_put() - put an FR descriptor back in the free list
429 * @pool: Pool the descriptor was allocated from.
430 * @desc: Pointer to an array of fast registration descriptor pointers.
431 * @n: Number of descriptors to put back.
433 * Note: The caller must already have queued an invalidation request for
434 * desc->mr->rkey before calling this function.
436 static void srp_fr_pool_put(struct srp_fr_pool *pool, struct srp_fr_desc **desc,
442 spin_lock_irqsave(&pool->lock, flags);
443 for (i = 0; i < n; i++)
444 list_add(&desc[i]->entry, &pool->free_list);
445 spin_unlock_irqrestore(&pool->lock, flags);
448 static struct srp_fr_pool *srp_alloc_fr_pool(struct srp_target_port *target)
450 struct srp_device *dev = target->srp_host->srp_dev;
452 return srp_create_fr_pool(dev->dev, dev->pd,
453 target->scsi_host->can_queue,
454 dev->max_pages_per_mr);
458 * srp_destroy_qp() - destroy an RDMA queue pair
459 * @ch: SRP RDMA channel.
461 * Change a queue pair into the error state and wait until all receive
462 * completions have been processed before destroying it. This avoids that
463 * the receive completion handler can access the queue pair while it is
466 static void srp_destroy_qp(struct srp_rdma_ch *ch)
468 static struct ib_qp_attr attr = { .qp_state = IB_QPS_ERR };
469 static struct ib_recv_wr wr = { .wr_id = SRP_LAST_WR_ID };
470 struct ib_recv_wr *bad_wr;
473 /* Destroying a QP and reusing ch->done is only safe if not connected */
474 WARN_ON_ONCE(ch->connected);
476 ret = ib_modify_qp(ch->qp, &attr, IB_QP_STATE);
477 WARN_ONCE(ret, "ib_cm_init_qp_attr() returned %d\n", ret);
481 init_completion(&ch->done);
482 ret = ib_post_recv(ch->qp, &wr, &bad_wr);
483 WARN_ONCE(ret, "ib_post_recv() returned %d\n", ret);
485 wait_for_completion(&ch->done);
488 ib_destroy_qp(ch->qp);
491 static int srp_create_ch_ib(struct srp_rdma_ch *ch)
493 struct srp_target_port *target = ch->target;
494 struct srp_device *dev = target->srp_host->srp_dev;
495 struct ib_qp_init_attr *init_attr;
496 struct ib_cq *recv_cq, *send_cq;
498 struct ib_fmr_pool *fmr_pool = NULL;
499 struct srp_fr_pool *fr_pool = NULL;
500 const int m = 1 + dev->use_fast_reg;
501 struct ib_cq_init_attr cq_attr = {};
504 init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
508 /* + 1 for SRP_LAST_WR_ID */
509 cq_attr.cqe = target->queue_size + 1;
510 cq_attr.comp_vector = ch->comp_vector;
511 recv_cq = ib_create_cq(dev->dev, srp_recv_completion, NULL, ch,
513 if (IS_ERR(recv_cq)) {
514 ret = PTR_ERR(recv_cq);
518 cq_attr.cqe = m * target->queue_size;
519 cq_attr.comp_vector = ch->comp_vector;
520 send_cq = ib_create_cq(dev->dev, srp_send_completion, NULL, ch,
522 if (IS_ERR(send_cq)) {
523 ret = PTR_ERR(send_cq);
527 ib_req_notify_cq(recv_cq, IB_CQ_NEXT_COMP);
529 init_attr->event_handler = srp_qp_event;
530 init_attr->cap.max_send_wr = m * target->queue_size;
531 init_attr->cap.max_recv_wr = target->queue_size + 1;
532 init_attr->cap.max_recv_sge = 1;
533 init_attr->cap.max_send_sge = 1;
534 init_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
535 init_attr->qp_type = IB_QPT_RC;
536 init_attr->send_cq = send_cq;
537 init_attr->recv_cq = recv_cq;
539 qp = ib_create_qp(dev->pd, init_attr);
545 ret = srp_init_qp(target, qp);
549 if (dev->use_fast_reg && dev->has_fr) {
550 fr_pool = srp_alloc_fr_pool(target);
551 if (IS_ERR(fr_pool)) {
552 ret = PTR_ERR(fr_pool);
553 shost_printk(KERN_WARNING, target->scsi_host, PFX
554 "FR pool allocation failed (%d)\n", ret);
558 srp_destroy_fr_pool(ch->fr_pool);
559 ch->fr_pool = fr_pool;
560 } else if (!dev->use_fast_reg && dev->has_fmr) {
561 fmr_pool = srp_alloc_fmr_pool(target);
562 if (IS_ERR(fmr_pool)) {
563 ret = PTR_ERR(fmr_pool);
564 shost_printk(KERN_WARNING, target->scsi_host, PFX
565 "FMR pool allocation failed (%d)\n", ret);
569 ib_destroy_fmr_pool(ch->fmr_pool);
570 ch->fmr_pool = fmr_pool;
576 ib_destroy_cq(ch->recv_cq);
578 ib_destroy_cq(ch->send_cq);
581 ch->recv_cq = recv_cq;
582 ch->send_cq = send_cq;
591 ib_destroy_cq(send_cq);
594 ib_destroy_cq(recv_cq);
602 * Note: this function may be called without srp_alloc_iu_bufs() having been
603 * invoked. Hence the ch->[rt]x_ring checks.
605 static void srp_free_ch_ib(struct srp_target_port *target,
606 struct srp_rdma_ch *ch)
608 struct srp_device *dev = target->srp_host->srp_dev;
615 ib_destroy_cm_id(ch->cm_id);
619 /* If srp_new_cm_id() succeeded but srp_create_ch_ib() not, return. */
623 if (dev->use_fast_reg) {
625 srp_destroy_fr_pool(ch->fr_pool);
628 ib_destroy_fmr_pool(ch->fmr_pool);
631 ib_destroy_cq(ch->send_cq);
632 ib_destroy_cq(ch->recv_cq);
635 * Avoid that the SCSI error handler tries to use this channel after
636 * it has been freed. The SCSI error handler can namely continue
637 * trying to perform recovery actions after scsi_remove_host()
643 ch->send_cq = ch->recv_cq = NULL;
646 for (i = 0; i < target->queue_size; ++i)
647 srp_free_iu(target->srp_host, ch->rx_ring[i]);
652 for (i = 0; i < target->queue_size; ++i)
653 srp_free_iu(target->srp_host, ch->tx_ring[i]);
659 static void srp_path_rec_completion(int status,
660 struct ib_sa_path_rec *pathrec,
663 struct srp_rdma_ch *ch = ch_ptr;
664 struct srp_target_port *target = ch->target;
668 shost_printk(KERN_ERR, target->scsi_host,
669 PFX "Got failed path rec status %d\n", status);
675 static int srp_lookup_path(struct srp_rdma_ch *ch)
677 struct srp_target_port *target = ch->target;
680 ch->path.numb_path = 1;
682 init_completion(&ch->done);
684 ch->path_query_id = ib_sa_path_rec_get(&srp_sa_client,
685 target->srp_host->srp_dev->dev,
686 target->srp_host->port,
688 IB_SA_PATH_REC_SERVICE_ID |
689 IB_SA_PATH_REC_DGID |
690 IB_SA_PATH_REC_SGID |
691 IB_SA_PATH_REC_NUMB_PATH |
693 SRP_PATH_REC_TIMEOUT_MS,
695 srp_path_rec_completion,
696 ch, &ch->path_query);
697 if (ch->path_query_id < 0)
698 return ch->path_query_id;
700 ret = wait_for_completion_interruptible(&ch->done);
705 shost_printk(KERN_WARNING, target->scsi_host,
706 PFX "Path record query failed\n");
711 static int srp_send_req(struct srp_rdma_ch *ch, bool multich)
713 struct srp_target_port *target = ch->target;
715 struct ib_cm_req_param param;
716 struct srp_login_req priv;
720 req = kzalloc(sizeof *req, GFP_KERNEL);
724 req->param.primary_path = &ch->path;
725 req->param.alternate_path = NULL;
726 req->param.service_id = target->service_id;
727 req->param.qp_num = ch->qp->qp_num;
728 req->param.qp_type = ch->qp->qp_type;
729 req->param.private_data = &req->priv;
730 req->param.private_data_len = sizeof req->priv;
731 req->param.flow_control = 1;
733 get_random_bytes(&req->param.starting_psn, 4);
734 req->param.starting_psn &= 0xffffff;
737 * Pick some arbitrary defaults here; we could make these
738 * module parameters if anyone cared about setting them.
740 req->param.responder_resources = 4;
741 req->param.remote_cm_response_timeout = 20;
742 req->param.local_cm_response_timeout = 20;
743 req->param.retry_count = target->tl_retry_count;
744 req->param.rnr_retry_count = 7;
745 req->param.max_cm_retries = 15;
747 req->priv.opcode = SRP_LOGIN_REQ;
749 req->priv.req_it_iu_len = cpu_to_be32(target->max_iu_len);
750 req->priv.req_buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
751 SRP_BUF_FORMAT_INDIRECT);
752 req->priv.req_flags = (multich ? SRP_MULTICHAN_MULTI :
753 SRP_MULTICHAN_SINGLE);
755 * In the published SRP specification (draft rev. 16a), the
756 * port identifier format is 8 bytes of ID extension followed
757 * by 8 bytes of GUID. Older drafts put the two halves in the
758 * opposite order, so that the GUID comes first.
760 * Targets conforming to these obsolete drafts can be
761 * recognized by the I/O Class they report.
763 if (target->io_class == SRP_REV10_IB_IO_CLASS) {
764 memcpy(req->priv.initiator_port_id,
765 &target->sgid.global.interface_id, 8);
766 memcpy(req->priv.initiator_port_id + 8,
767 &target->initiator_ext, 8);
768 memcpy(req->priv.target_port_id, &target->ioc_guid, 8);
769 memcpy(req->priv.target_port_id + 8, &target->id_ext, 8);
771 memcpy(req->priv.initiator_port_id,
772 &target->initiator_ext, 8);
773 memcpy(req->priv.initiator_port_id + 8,
774 &target->sgid.global.interface_id, 8);
775 memcpy(req->priv.target_port_id, &target->id_ext, 8);
776 memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
780 * Topspin/Cisco SRP targets will reject our login unless we
781 * zero out the first 8 bytes of our initiator port ID and set
782 * the second 8 bytes to the local node GUID.
784 if (srp_target_is_topspin(target)) {
785 shost_printk(KERN_DEBUG, target->scsi_host,
786 PFX "Topspin/Cisco initiator port ID workaround "
787 "activated for target GUID %016llx\n",
788 be64_to_cpu(target->ioc_guid));
789 memset(req->priv.initiator_port_id, 0, 8);
790 memcpy(req->priv.initiator_port_id + 8,
791 &target->srp_host->srp_dev->dev->node_guid, 8);
794 status = ib_send_cm_req(ch->cm_id, &req->param);
801 static bool srp_queue_remove_work(struct srp_target_port *target)
803 bool changed = false;
805 spin_lock_irq(&target->lock);
806 if (target->state != SRP_TARGET_REMOVED) {
807 target->state = SRP_TARGET_REMOVED;
810 spin_unlock_irq(&target->lock);
813 queue_work(srp_remove_wq, &target->remove_work);
818 static void srp_disconnect_target(struct srp_target_port *target)
820 struct srp_rdma_ch *ch;
823 /* XXX should send SRP_I_LOGOUT request */
825 for (i = 0; i < target->ch_count; i++) {
827 ch->connected = false;
828 if (ch->cm_id && ib_send_cm_dreq(ch->cm_id, NULL, 0)) {
829 shost_printk(KERN_DEBUG, target->scsi_host,
830 PFX "Sending CM DREQ failed\n");
835 static void srp_free_req_data(struct srp_target_port *target,
836 struct srp_rdma_ch *ch)
838 struct srp_device *dev = target->srp_host->srp_dev;
839 struct ib_device *ibdev = dev->dev;
840 struct srp_request *req;
846 for (i = 0; i < target->req_ring_size; ++i) {
847 req = &ch->req_ring[i];
848 if (dev->use_fast_reg)
851 kfree(req->fmr_list);
852 kfree(req->map_page);
853 if (req->indirect_dma_addr) {
854 ib_dma_unmap_single(ibdev, req->indirect_dma_addr,
855 target->indirect_size,
858 kfree(req->indirect_desc);
865 static int srp_alloc_req_data(struct srp_rdma_ch *ch)
867 struct srp_target_port *target = ch->target;
868 struct srp_device *srp_dev = target->srp_host->srp_dev;
869 struct ib_device *ibdev = srp_dev->dev;
870 struct srp_request *req;
873 int i, ret = -ENOMEM;
875 ch->req_ring = kcalloc(target->req_ring_size, sizeof(*ch->req_ring),
880 for (i = 0; i < target->req_ring_size; ++i) {
881 req = &ch->req_ring[i];
882 mr_list = kmalloc(target->cmd_sg_cnt * sizeof(void *),
886 if (srp_dev->use_fast_reg)
887 req->fr_list = mr_list;
889 req->fmr_list = mr_list;
890 req->map_page = kmalloc(srp_dev->max_pages_per_mr *
891 sizeof(void *), GFP_KERNEL);
894 req->indirect_desc = kmalloc(target->indirect_size, GFP_KERNEL);
895 if (!req->indirect_desc)
898 dma_addr = ib_dma_map_single(ibdev, req->indirect_desc,
899 target->indirect_size,
901 if (ib_dma_mapping_error(ibdev, dma_addr))
904 req->indirect_dma_addr = dma_addr;
913 * srp_del_scsi_host_attr() - Remove attributes defined in the host template.
914 * @shost: SCSI host whose attributes to remove from sysfs.
916 * Note: Any attributes defined in the host template and that did not exist
917 * before invocation of this function will be ignored.
919 static void srp_del_scsi_host_attr(struct Scsi_Host *shost)
921 struct device_attribute **attr;
923 for (attr = shost->hostt->shost_attrs; attr && *attr; ++attr)
924 device_remove_file(&shost->shost_dev, *attr);
927 static void srp_remove_target(struct srp_target_port *target)
929 struct srp_rdma_ch *ch;
932 WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
934 srp_del_scsi_host_attr(target->scsi_host);
935 srp_rport_get(target->rport);
936 srp_remove_host(target->scsi_host);
937 scsi_remove_host(target->scsi_host);
938 srp_stop_rport_timers(target->rport);
939 srp_disconnect_target(target);
940 for (i = 0; i < target->ch_count; i++) {
942 srp_free_ch_ib(target, ch);
944 cancel_work_sync(&target->tl_err_work);
945 srp_rport_put(target->rport);
946 for (i = 0; i < target->ch_count; i++) {
948 srp_free_req_data(target, ch);
953 spin_lock(&target->srp_host->target_lock);
954 list_del(&target->list);
955 spin_unlock(&target->srp_host->target_lock);
957 scsi_host_put(target->scsi_host);
960 static void srp_remove_work(struct work_struct *work)
962 struct srp_target_port *target =
963 container_of(work, struct srp_target_port, remove_work);
965 WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
967 srp_remove_target(target);
970 static void srp_rport_delete(struct srp_rport *rport)
972 struct srp_target_port *target = rport->lld_data;
974 srp_queue_remove_work(target);
978 * srp_connected_ch() - number of connected channels
979 * @target: SRP target port.
981 static int srp_connected_ch(struct srp_target_port *target)
985 for (i = 0; i < target->ch_count; i++)
986 c += target->ch[i].connected;
991 static int srp_connect_ch(struct srp_rdma_ch *ch, bool multich)
993 struct srp_target_port *target = ch->target;
996 WARN_ON_ONCE(!multich && srp_connected_ch(target) > 0);
998 ret = srp_lookup_path(ch);
1003 init_completion(&ch->done);
1004 ret = srp_send_req(ch, multich);
1007 ret = wait_for_completion_interruptible(&ch->done);
1012 * The CM event handling code will set status to
1013 * SRP_PORT_REDIRECT if we get a port redirect REJ
1014 * back, or SRP_DLID_REDIRECT if we get a lid/qp
1015 * redirect REJ back.
1017 switch (ch->status) {
1019 ch->connected = true;
1022 case SRP_PORT_REDIRECT:
1023 ret = srp_lookup_path(ch);
1028 case SRP_DLID_REDIRECT:
1031 case SRP_STALE_CONN:
1032 shost_printk(KERN_ERR, target->scsi_host, PFX
1033 "giving up on stale connection\n");
1034 ch->status = -ECONNRESET;
1043 static int srp_inv_rkey(struct srp_rdma_ch *ch, u32 rkey)
1045 struct ib_send_wr *bad_wr;
1046 struct ib_send_wr wr = {
1047 .opcode = IB_WR_LOCAL_INV,
1048 .wr_id = LOCAL_INV_WR_ID_MASK,
1052 .ex.invalidate_rkey = rkey,
1055 return ib_post_send(ch->qp, &wr, &bad_wr);
1058 static void srp_unmap_data(struct scsi_cmnd *scmnd,
1059 struct srp_rdma_ch *ch,
1060 struct srp_request *req)
1062 struct srp_target_port *target = ch->target;
1063 struct srp_device *dev = target->srp_host->srp_dev;
1064 struct ib_device *ibdev = dev->dev;
1067 if (!scsi_sglist(scmnd) ||
1068 (scmnd->sc_data_direction != DMA_TO_DEVICE &&
1069 scmnd->sc_data_direction != DMA_FROM_DEVICE))
1072 if (dev->use_fast_reg) {
1073 struct srp_fr_desc **pfr;
1075 for (i = req->nmdesc, pfr = req->fr_list; i > 0; i--, pfr++) {
1076 res = srp_inv_rkey(ch, (*pfr)->mr->rkey);
1078 shost_printk(KERN_ERR, target->scsi_host, PFX
1079 "Queueing INV WR for rkey %#x failed (%d)\n",
1080 (*pfr)->mr->rkey, res);
1081 queue_work(system_long_wq,
1082 &target->tl_err_work);
1086 srp_fr_pool_put(ch->fr_pool, req->fr_list,
1089 struct ib_pool_fmr **pfmr;
1091 for (i = req->nmdesc, pfmr = req->fmr_list; i > 0; i--, pfmr++)
1092 ib_fmr_pool_unmap(*pfmr);
1095 ib_dma_unmap_sg(ibdev, scsi_sglist(scmnd), scsi_sg_count(scmnd),
1096 scmnd->sc_data_direction);
1100 * srp_claim_req - Take ownership of the scmnd associated with a request.
1101 * @ch: SRP RDMA channel.
1102 * @req: SRP request.
1103 * @sdev: If not NULL, only take ownership for this SCSI device.
1104 * @scmnd: If NULL, take ownership of @req->scmnd. If not NULL, only take
1105 * ownership of @req->scmnd if it equals @scmnd.
1108 * Either NULL or a pointer to the SCSI command the caller became owner of.
1110 static struct scsi_cmnd *srp_claim_req(struct srp_rdma_ch *ch,
1111 struct srp_request *req,
1112 struct scsi_device *sdev,
1113 struct scsi_cmnd *scmnd)
1115 unsigned long flags;
1117 spin_lock_irqsave(&ch->lock, flags);
1119 (!sdev || req->scmnd->device == sdev) &&
1120 (!scmnd || req->scmnd == scmnd)) {
1126 spin_unlock_irqrestore(&ch->lock, flags);
1132 * srp_free_req() - Unmap data and add request to the free request list.
1133 * @ch: SRP RDMA channel.
1134 * @req: Request to be freed.
1135 * @scmnd: SCSI command associated with @req.
1136 * @req_lim_delta: Amount to be added to @target->req_lim.
1138 static void srp_free_req(struct srp_rdma_ch *ch, struct srp_request *req,
1139 struct scsi_cmnd *scmnd, s32 req_lim_delta)
1141 unsigned long flags;
1143 srp_unmap_data(scmnd, ch, req);
1145 spin_lock_irqsave(&ch->lock, flags);
1146 ch->req_lim += req_lim_delta;
1147 spin_unlock_irqrestore(&ch->lock, flags);
1150 static void srp_finish_req(struct srp_rdma_ch *ch, struct srp_request *req,
1151 struct scsi_device *sdev, int result)
1153 struct scsi_cmnd *scmnd = srp_claim_req(ch, req, sdev, NULL);
1156 srp_free_req(ch, req, scmnd, 0);
1157 scmnd->result = result;
1158 scmnd->scsi_done(scmnd);
1162 static void srp_terminate_io(struct srp_rport *rport)
1164 struct srp_target_port *target = rport->lld_data;
1165 struct srp_rdma_ch *ch;
1166 struct Scsi_Host *shost = target->scsi_host;
1167 struct scsi_device *sdev;
1171 * Invoking srp_terminate_io() while srp_queuecommand() is running
1172 * is not safe. Hence the warning statement below.
1174 shost_for_each_device(sdev, shost)
1175 WARN_ON_ONCE(sdev->request_queue->request_fn_active);
1177 for (i = 0; i < target->ch_count; i++) {
1178 ch = &target->ch[i];
1180 for (j = 0; j < target->req_ring_size; ++j) {
1181 struct srp_request *req = &ch->req_ring[j];
1183 srp_finish_req(ch, req, NULL,
1184 DID_TRANSPORT_FAILFAST << 16);
1190 * It is up to the caller to ensure that srp_rport_reconnect() calls are
1191 * serialized and that no concurrent srp_queuecommand(), srp_abort(),
1192 * srp_reset_device() or srp_reset_host() calls will occur while this function
1193 * is in progress. One way to realize that is not to call this function
1194 * directly but to call srp_reconnect_rport() instead since that last function
1195 * serializes calls of this function via rport->mutex and also blocks
1196 * srp_queuecommand() calls before invoking this function.
1198 static int srp_rport_reconnect(struct srp_rport *rport)
1200 struct srp_target_port *target = rport->lld_data;
1201 struct srp_rdma_ch *ch;
1203 bool multich = false;
1205 srp_disconnect_target(target);
1207 if (target->state == SRP_TARGET_SCANNING)
1211 * Now get a new local CM ID so that we avoid confusing the target in
1212 * case things are really fouled up. Doing so also ensures that all CM
1213 * callbacks will have finished before a new QP is allocated.
1215 for (i = 0; i < target->ch_count; i++) {
1216 ch = &target->ch[i];
1217 ret += srp_new_cm_id(ch);
1219 for (i = 0; i < target->ch_count; i++) {
1220 ch = &target->ch[i];
1221 for (j = 0; j < target->req_ring_size; ++j) {
1222 struct srp_request *req = &ch->req_ring[j];
1224 srp_finish_req(ch, req, NULL, DID_RESET << 16);
1227 for (i = 0; i < target->ch_count; i++) {
1228 ch = &target->ch[i];
1230 * Whether or not creating a new CM ID succeeded, create a new
1231 * QP. This guarantees that all completion callback function
1232 * invocations have finished before request resetting starts.
1234 ret += srp_create_ch_ib(ch);
1236 INIT_LIST_HEAD(&ch->free_tx);
1237 for (j = 0; j < target->queue_size; ++j)
1238 list_add(&ch->tx_ring[j]->list, &ch->free_tx);
1241 target->qp_in_error = false;
1243 for (i = 0; i < target->ch_count; i++) {
1244 ch = &target->ch[i];
1247 ret = srp_connect_ch(ch, multich);
1252 shost_printk(KERN_INFO, target->scsi_host,
1253 PFX "reconnect succeeded\n");
1258 static void srp_map_desc(struct srp_map_state *state, dma_addr_t dma_addr,
1259 unsigned int dma_len, u32 rkey)
1261 struct srp_direct_buf *desc = state->desc;
1263 desc->va = cpu_to_be64(dma_addr);
1264 desc->key = cpu_to_be32(rkey);
1265 desc->len = cpu_to_be32(dma_len);
1267 state->total_len += dma_len;
1272 static int srp_map_finish_fmr(struct srp_map_state *state,
1273 struct srp_rdma_ch *ch)
1275 struct srp_target_port *target = ch->target;
1276 struct srp_device *dev = target->srp_host->srp_dev;
1277 struct ib_pool_fmr *fmr;
1280 if (state->fmr.next >= state->fmr.end)
1283 fmr = ib_fmr_pool_map_phys(ch->fmr_pool, state->pages,
1284 state->npages, io_addr);
1286 return PTR_ERR(fmr);
1288 *state->fmr.next++ = fmr;
1291 srp_map_desc(state, state->base_dma_addr & ~dev->mr_page_mask,
1292 state->dma_len, fmr->fmr->rkey);
1297 static int srp_map_finish_fr(struct srp_map_state *state,
1298 struct srp_rdma_ch *ch)
1300 struct srp_target_port *target = ch->target;
1301 struct srp_device *dev = target->srp_host->srp_dev;
1302 struct ib_send_wr *bad_wr;
1303 struct ib_send_wr wr;
1304 struct srp_fr_desc *desc;
1307 if (state->fr.next >= state->fr.end)
1310 desc = srp_fr_pool_get(ch->fr_pool);
1314 rkey = ib_inc_rkey(desc->mr->rkey);
1315 ib_update_fast_reg_key(desc->mr, rkey);
1317 memcpy(desc->frpl->page_list, state->pages,
1318 sizeof(state->pages[0]) * state->npages);
1320 memset(&wr, 0, sizeof(wr));
1321 wr.opcode = IB_WR_FAST_REG_MR;
1322 wr.wr_id = FAST_REG_WR_ID_MASK;
1323 wr.wr.fast_reg.iova_start = state->base_dma_addr;
1324 wr.wr.fast_reg.page_list = desc->frpl;
1325 wr.wr.fast_reg.page_list_len = state->npages;
1326 wr.wr.fast_reg.page_shift = ilog2(dev->mr_page_size);
1327 wr.wr.fast_reg.length = state->dma_len;
1328 wr.wr.fast_reg.access_flags = (IB_ACCESS_LOCAL_WRITE |
1329 IB_ACCESS_REMOTE_READ |
1330 IB_ACCESS_REMOTE_WRITE);
1331 wr.wr.fast_reg.rkey = desc->mr->lkey;
1333 *state->fr.next++ = desc;
1336 srp_map_desc(state, state->base_dma_addr, state->dma_len,
1339 return ib_post_send(ch->qp, &wr, &bad_wr);
1342 static int srp_finish_mapping(struct srp_map_state *state,
1343 struct srp_rdma_ch *ch)
1345 struct srp_target_port *target = ch->target;
1348 if (state->npages == 0)
1351 if (state->npages == 1 && !register_always)
1352 srp_map_desc(state, state->base_dma_addr, state->dma_len,
1355 ret = target->srp_host->srp_dev->use_fast_reg ?
1356 srp_map_finish_fr(state, ch) :
1357 srp_map_finish_fmr(state, ch);
1367 static void srp_map_update_start(struct srp_map_state *state,
1368 struct scatterlist *sg, int sg_index,
1369 dma_addr_t dma_addr)
1371 state->unmapped_sg = sg;
1372 state->unmapped_index = sg_index;
1373 state->unmapped_addr = dma_addr;
1376 static int srp_map_sg_entry(struct srp_map_state *state,
1377 struct srp_rdma_ch *ch,
1378 struct scatterlist *sg, int sg_index,
1381 struct srp_target_port *target = ch->target;
1382 struct srp_device *dev = target->srp_host->srp_dev;
1383 struct ib_device *ibdev = dev->dev;
1384 dma_addr_t dma_addr = ib_sg_dma_address(ibdev, sg);
1385 unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
1394 * Once we're in direct map mode for a request, we don't
1395 * go back to FMR or FR mode, so no need to update anything
1396 * other than the descriptor.
1398 srp_map_desc(state, dma_addr, dma_len, target->rkey);
1403 * If this is the first sg that will be mapped via FMR or via FR, save
1404 * our position. We need to know the first unmapped entry, its index,
1405 * and the first unmapped address within that entry to be able to
1406 * restart mapping after an error.
1408 if (!state->unmapped_sg)
1409 srp_map_update_start(state, sg, sg_index, dma_addr);
1412 unsigned offset = dma_addr & ~dev->mr_page_mask;
1413 if (state->npages == dev->max_pages_per_mr || offset != 0) {
1414 ret = srp_finish_mapping(state, ch);
1418 srp_map_update_start(state, sg, sg_index, dma_addr);
1421 len = min_t(unsigned int, dma_len, dev->mr_page_size - offset);
1424 state->base_dma_addr = dma_addr;
1425 state->pages[state->npages++] = dma_addr & dev->mr_page_mask;
1426 state->dma_len += len;
1432 * If the last entry of the MR wasn't a full page, then we need to
1433 * close it out and start a new one -- we can only merge at page
1437 if (len != dev->mr_page_size) {
1438 ret = srp_finish_mapping(state, ch);
1440 srp_map_update_start(state, NULL, 0, 0);
1445 static int srp_map_sg(struct srp_map_state *state, struct srp_rdma_ch *ch,
1446 struct srp_request *req, struct scatterlist *scat,
1449 struct srp_target_port *target = ch->target;
1450 struct srp_device *dev = target->srp_host->srp_dev;
1451 struct ib_device *ibdev = dev->dev;
1452 struct scatterlist *sg;
1456 state->desc = req->indirect_desc;
1457 state->pages = req->map_page;
1458 if (dev->use_fast_reg) {
1459 state->fr.next = req->fr_list;
1460 state->fr.end = req->fr_list + target->cmd_sg_cnt;
1461 use_mr = !!ch->fr_pool;
1463 state->fmr.next = req->fmr_list;
1464 state->fmr.end = req->fmr_list + target->cmd_sg_cnt;
1465 use_mr = !!ch->fmr_pool;
1468 for_each_sg(scat, sg, count, i) {
1469 if (srp_map_sg_entry(state, ch, sg, i, use_mr)) {
1471 * Memory registration failed, so backtrack to the
1472 * first unmapped entry and continue on without using
1473 * memory registration.
1475 dma_addr_t dma_addr;
1476 unsigned int dma_len;
1479 sg = state->unmapped_sg;
1480 i = state->unmapped_index;
1482 dma_addr = ib_sg_dma_address(ibdev, sg);
1483 dma_len = ib_sg_dma_len(ibdev, sg);
1484 dma_len -= (state->unmapped_addr - dma_addr);
1485 dma_addr = state->unmapped_addr;
1487 srp_map_desc(state, dma_addr, dma_len, target->rkey);
1491 if (use_mr && srp_finish_mapping(state, ch))
1494 req->nmdesc = state->nmdesc;
1499 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_rdma_ch *ch,
1500 struct srp_request *req)
1502 struct srp_target_port *target = ch->target;
1503 struct scatterlist *scat;
1504 struct srp_cmd *cmd = req->cmd->buf;
1505 int len, nents, count;
1506 struct srp_device *dev;
1507 struct ib_device *ibdev;
1508 struct srp_map_state state;
1509 struct srp_indirect_buf *indirect_hdr;
1513 if (!scsi_sglist(scmnd) || scmnd->sc_data_direction == DMA_NONE)
1514 return sizeof (struct srp_cmd);
1516 if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
1517 scmnd->sc_data_direction != DMA_TO_DEVICE) {
1518 shost_printk(KERN_WARNING, target->scsi_host,
1519 PFX "Unhandled data direction %d\n",
1520 scmnd->sc_data_direction);
1524 nents = scsi_sg_count(scmnd);
1525 scat = scsi_sglist(scmnd);
1527 dev = target->srp_host->srp_dev;
1530 count = ib_dma_map_sg(ibdev, scat, nents, scmnd->sc_data_direction);
1531 if (unlikely(count == 0))
1534 fmt = SRP_DATA_DESC_DIRECT;
1535 len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
1537 if (count == 1 && !register_always) {
1539 * The midlayer only generated a single gather/scatter
1540 * entry, or DMA mapping coalesced everything to a
1541 * single entry. So a direct descriptor along with
1542 * the DMA MR suffices.
1544 struct srp_direct_buf *buf = (void *) cmd->add_data;
1546 buf->va = cpu_to_be64(ib_sg_dma_address(ibdev, scat));
1547 buf->key = cpu_to_be32(target->rkey);
1548 buf->len = cpu_to_be32(ib_sg_dma_len(ibdev, scat));
1555 * We have more than one scatter/gather entry, so build our indirect
1556 * descriptor table, trying to merge as many entries as we can.
1558 indirect_hdr = (void *) cmd->add_data;
1560 ib_dma_sync_single_for_cpu(ibdev, req->indirect_dma_addr,
1561 target->indirect_size, DMA_TO_DEVICE);
1563 memset(&state, 0, sizeof(state));
1564 srp_map_sg(&state, ch, req, scat, count);
1566 /* We've mapped the request, now pull as much of the indirect
1567 * descriptor table as we can into the command buffer. If this
1568 * target is not using an external indirect table, we are
1569 * guaranteed to fit into the command, as the SCSI layer won't
1570 * give us more S/G entries than we allow.
1572 if (state.ndesc == 1) {
1574 * Memory registration collapsed the sg-list into one entry,
1575 * so use a direct descriptor.
1577 struct srp_direct_buf *buf = (void *) cmd->add_data;
1579 *buf = req->indirect_desc[0];
1583 if (unlikely(target->cmd_sg_cnt < state.ndesc &&
1584 !target->allow_ext_sg)) {
1585 shost_printk(KERN_ERR, target->scsi_host,
1586 "Could not fit S/G list into SRP_CMD\n");
1590 count = min(state.ndesc, target->cmd_sg_cnt);
1591 table_len = state.ndesc * sizeof (struct srp_direct_buf);
1593 fmt = SRP_DATA_DESC_INDIRECT;
1594 len = sizeof(struct srp_cmd) + sizeof (struct srp_indirect_buf);
1595 len += count * sizeof (struct srp_direct_buf);
1597 memcpy(indirect_hdr->desc_list, req->indirect_desc,
1598 count * sizeof (struct srp_direct_buf));
1600 indirect_hdr->table_desc.va = cpu_to_be64(req->indirect_dma_addr);
1601 indirect_hdr->table_desc.key = cpu_to_be32(target->rkey);
1602 indirect_hdr->table_desc.len = cpu_to_be32(table_len);
1603 indirect_hdr->len = cpu_to_be32(state.total_len);
1605 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1606 cmd->data_out_desc_cnt = count;
1608 cmd->data_in_desc_cnt = count;
1610 ib_dma_sync_single_for_device(ibdev, req->indirect_dma_addr, table_len,
1614 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1615 cmd->buf_fmt = fmt << 4;
1623 * Return an IU and possible credit to the free pool
1625 static void srp_put_tx_iu(struct srp_rdma_ch *ch, struct srp_iu *iu,
1626 enum srp_iu_type iu_type)
1628 unsigned long flags;
1630 spin_lock_irqsave(&ch->lock, flags);
1631 list_add(&iu->list, &ch->free_tx);
1632 if (iu_type != SRP_IU_RSP)
1634 spin_unlock_irqrestore(&ch->lock, flags);
1638 * Must be called with ch->lock held to protect req_lim and free_tx.
1639 * If IU is not sent, it must be returned using srp_put_tx_iu().
1642 * An upper limit for the number of allocated information units for each
1644 * - SRP_IU_CMD: SRP_CMD_SQ_SIZE, since the SCSI mid-layer never queues
1645 * more than Scsi_Host.can_queue requests.
1646 * - SRP_IU_TSK_MGMT: SRP_TSK_MGMT_SQ_SIZE.
1647 * - SRP_IU_RSP: 1, since a conforming SRP target never sends more than
1648 * one unanswered SRP request to an initiator.
1650 static struct srp_iu *__srp_get_tx_iu(struct srp_rdma_ch *ch,
1651 enum srp_iu_type iu_type)
1653 struct srp_target_port *target = ch->target;
1654 s32 rsv = (iu_type == SRP_IU_TSK_MGMT) ? 0 : SRP_TSK_MGMT_SQ_SIZE;
1657 srp_send_completion(ch->send_cq, ch);
1659 if (list_empty(&ch->free_tx))
1662 /* Initiator responses to target requests do not consume credits */
1663 if (iu_type != SRP_IU_RSP) {
1664 if (ch->req_lim <= rsv) {
1665 ++target->zero_req_lim;
1672 iu = list_first_entry(&ch->free_tx, struct srp_iu, list);
1673 list_del(&iu->list);
1677 static int srp_post_send(struct srp_rdma_ch *ch, struct srp_iu *iu, int len)
1679 struct srp_target_port *target = ch->target;
1681 struct ib_send_wr wr, *bad_wr;
1683 list.addr = iu->dma;
1685 list.lkey = target->lkey;
1688 wr.wr_id = (uintptr_t) iu;
1691 wr.opcode = IB_WR_SEND;
1692 wr.send_flags = IB_SEND_SIGNALED;
1694 return ib_post_send(ch->qp, &wr, &bad_wr);
1697 static int srp_post_recv(struct srp_rdma_ch *ch, struct srp_iu *iu)
1699 struct srp_target_port *target = ch->target;
1700 struct ib_recv_wr wr, *bad_wr;
1703 list.addr = iu->dma;
1704 list.length = iu->size;
1705 list.lkey = target->lkey;
1708 wr.wr_id = (uintptr_t) iu;
1712 return ib_post_recv(ch->qp, &wr, &bad_wr);
1715 static void srp_process_rsp(struct srp_rdma_ch *ch, struct srp_rsp *rsp)
1717 struct srp_target_port *target = ch->target;
1718 struct srp_request *req;
1719 struct scsi_cmnd *scmnd;
1720 unsigned long flags;
1722 if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
1723 spin_lock_irqsave(&ch->lock, flags);
1724 ch->req_lim += be32_to_cpu(rsp->req_lim_delta);
1725 spin_unlock_irqrestore(&ch->lock, flags);
1727 ch->tsk_mgmt_status = -1;
1728 if (be32_to_cpu(rsp->resp_data_len) >= 4)
1729 ch->tsk_mgmt_status = rsp->data[3];
1730 complete(&ch->tsk_mgmt_done);
1732 scmnd = scsi_host_find_tag(target->scsi_host, rsp->tag);
1734 req = (void *)scmnd->host_scribble;
1735 scmnd = srp_claim_req(ch, req, NULL, scmnd);
1738 shost_printk(KERN_ERR, target->scsi_host,
1739 "Null scmnd for RSP w/tag %#016llx received on ch %td / QP %#x\n",
1740 rsp->tag, ch - target->ch, ch->qp->qp_num);
1742 spin_lock_irqsave(&ch->lock, flags);
1743 ch->req_lim += be32_to_cpu(rsp->req_lim_delta);
1744 spin_unlock_irqrestore(&ch->lock, flags);
1748 scmnd->result = rsp->status;
1750 if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
1751 memcpy(scmnd->sense_buffer, rsp->data +
1752 be32_to_cpu(rsp->resp_data_len),
1753 min_t(int, be32_to_cpu(rsp->sense_data_len),
1754 SCSI_SENSE_BUFFERSIZE));
1757 if (unlikely(rsp->flags & SRP_RSP_FLAG_DIUNDER))
1758 scsi_set_resid(scmnd, be32_to_cpu(rsp->data_in_res_cnt));
1759 else if (unlikely(rsp->flags & SRP_RSP_FLAG_DIOVER))
1760 scsi_set_resid(scmnd, -be32_to_cpu(rsp->data_in_res_cnt));
1761 else if (unlikely(rsp->flags & SRP_RSP_FLAG_DOUNDER))
1762 scsi_set_resid(scmnd, be32_to_cpu(rsp->data_out_res_cnt));
1763 else if (unlikely(rsp->flags & SRP_RSP_FLAG_DOOVER))
1764 scsi_set_resid(scmnd, -be32_to_cpu(rsp->data_out_res_cnt));
1766 srp_free_req(ch, req, scmnd,
1767 be32_to_cpu(rsp->req_lim_delta));
1769 scmnd->host_scribble = NULL;
1770 scmnd->scsi_done(scmnd);
1774 static int srp_response_common(struct srp_rdma_ch *ch, s32 req_delta,
1777 struct srp_target_port *target = ch->target;
1778 struct ib_device *dev = target->srp_host->srp_dev->dev;
1779 unsigned long flags;
1783 spin_lock_irqsave(&ch->lock, flags);
1784 ch->req_lim += req_delta;
1785 iu = __srp_get_tx_iu(ch, SRP_IU_RSP);
1786 spin_unlock_irqrestore(&ch->lock, flags);
1789 shost_printk(KERN_ERR, target->scsi_host, PFX
1790 "no IU available to send response\n");
1794 ib_dma_sync_single_for_cpu(dev, iu->dma, len, DMA_TO_DEVICE);
1795 memcpy(iu->buf, rsp, len);
1796 ib_dma_sync_single_for_device(dev, iu->dma, len, DMA_TO_DEVICE);
1798 err = srp_post_send(ch, iu, len);
1800 shost_printk(KERN_ERR, target->scsi_host, PFX
1801 "unable to post response: %d\n", err);
1802 srp_put_tx_iu(ch, iu, SRP_IU_RSP);
1808 static void srp_process_cred_req(struct srp_rdma_ch *ch,
1809 struct srp_cred_req *req)
1811 struct srp_cred_rsp rsp = {
1812 .opcode = SRP_CRED_RSP,
1815 s32 delta = be32_to_cpu(req->req_lim_delta);
1817 if (srp_response_common(ch, delta, &rsp, sizeof(rsp)))
1818 shost_printk(KERN_ERR, ch->target->scsi_host, PFX
1819 "problems processing SRP_CRED_REQ\n");
1822 static void srp_process_aer_req(struct srp_rdma_ch *ch,
1823 struct srp_aer_req *req)
1825 struct srp_target_port *target = ch->target;
1826 struct srp_aer_rsp rsp = {
1827 .opcode = SRP_AER_RSP,
1830 s32 delta = be32_to_cpu(req->req_lim_delta);
1832 shost_printk(KERN_ERR, target->scsi_host, PFX
1833 "ignoring AER for LUN %llu\n", scsilun_to_int(&req->lun));
1835 if (srp_response_common(ch, delta, &rsp, sizeof(rsp)))
1836 shost_printk(KERN_ERR, target->scsi_host, PFX
1837 "problems processing SRP_AER_REQ\n");
1840 static void srp_handle_recv(struct srp_rdma_ch *ch, struct ib_wc *wc)
1842 struct srp_target_port *target = ch->target;
1843 struct ib_device *dev = target->srp_host->srp_dev->dev;
1844 struct srp_iu *iu = (struct srp_iu *) (uintptr_t) wc->wr_id;
1848 ib_dma_sync_single_for_cpu(dev, iu->dma, ch->max_ti_iu_len,
1851 opcode = *(u8 *) iu->buf;
1854 shost_printk(KERN_ERR, target->scsi_host,
1855 PFX "recv completion, opcode 0x%02x\n", opcode);
1856 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 8, 1,
1857 iu->buf, wc->byte_len, true);
1862 srp_process_rsp(ch, iu->buf);
1866 srp_process_cred_req(ch, iu->buf);
1870 srp_process_aer_req(ch, iu->buf);
1874 /* XXX Handle target logout */
1875 shost_printk(KERN_WARNING, target->scsi_host,
1876 PFX "Got target logout request\n");
1880 shost_printk(KERN_WARNING, target->scsi_host,
1881 PFX "Unhandled SRP opcode 0x%02x\n", opcode);
1885 ib_dma_sync_single_for_device(dev, iu->dma, ch->max_ti_iu_len,
1888 res = srp_post_recv(ch, iu);
1890 shost_printk(KERN_ERR, target->scsi_host,
1891 PFX "Recv failed with error code %d\n", res);
1895 * srp_tl_err_work() - handle a transport layer error
1896 * @work: Work structure embedded in an SRP target port.
1898 * Note: This function may get invoked before the rport has been created,
1899 * hence the target->rport test.
1901 static void srp_tl_err_work(struct work_struct *work)
1903 struct srp_target_port *target;
1905 target = container_of(work, struct srp_target_port, tl_err_work);
1907 srp_start_tl_fail_timers(target->rport);
1910 static void srp_handle_qp_err(u64 wr_id, enum ib_wc_status wc_status,
1911 bool send_err, struct srp_rdma_ch *ch)
1913 struct srp_target_port *target = ch->target;
1915 if (wr_id == SRP_LAST_WR_ID) {
1916 complete(&ch->done);
1920 if (ch->connected && !target->qp_in_error) {
1921 if (wr_id & LOCAL_INV_WR_ID_MASK) {
1922 shost_printk(KERN_ERR, target->scsi_host, PFX
1923 "LOCAL_INV failed with status %s (%d)\n",
1924 ib_wc_status_msg(wc_status), wc_status);
1925 } else if (wr_id & FAST_REG_WR_ID_MASK) {
1926 shost_printk(KERN_ERR, target->scsi_host, PFX
1927 "FAST_REG_MR failed status %s (%d)\n",
1928 ib_wc_status_msg(wc_status), wc_status);
1930 shost_printk(KERN_ERR, target->scsi_host,
1931 PFX "failed %s status %s (%d) for iu %p\n",
1932 send_err ? "send" : "receive",
1933 ib_wc_status_msg(wc_status), wc_status,
1934 (void *)(uintptr_t)wr_id);
1936 queue_work(system_long_wq, &target->tl_err_work);
1938 target->qp_in_error = true;
1941 static void srp_recv_completion(struct ib_cq *cq, void *ch_ptr)
1943 struct srp_rdma_ch *ch = ch_ptr;
1946 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1947 while (ib_poll_cq(cq, 1, &wc) > 0) {
1948 if (likely(wc.status == IB_WC_SUCCESS)) {
1949 srp_handle_recv(ch, &wc);
1951 srp_handle_qp_err(wc.wr_id, wc.status, false, ch);
1956 static void srp_send_completion(struct ib_cq *cq, void *ch_ptr)
1958 struct srp_rdma_ch *ch = ch_ptr;
1962 while (ib_poll_cq(cq, 1, &wc) > 0) {
1963 if (likely(wc.status == IB_WC_SUCCESS)) {
1964 iu = (struct srp_iu *) (uintptr_t) wc.wr_id;
1965 list_add(&iu->list, &ch->free_tx);
1967 srp_handle_qp_err(wc.wr_id, wc.status, true, ch);
1972 static int srp_queuecommand(struct Scsi_Host *shost, struct scsi_cmnd *scmnd)
1974 struct srp_target_port *target = host_to_target(shost);
1975 struct srp_rport *rport = target->rport;
1976 struct srp_rdma_ch *ch;
1977 struct srp_request *req;
1979 struct srp_cmd *cmd;
1980 struct ib_device *dev;
1981 unsigned long flags;
1985 const bool in_scsi_eh = !in_interrupt() && current == shost->ehandler;
1988 * The SCSI EH thread is the only context from which srp_queuecommand()
1989 * can get invoked for blocked devices (SDEV_BLOCK /
1990 * SDEV_CREATED_BLOCK). Avoid racing with srp_reconnect_rport() by
1991 * locking the rport mutex if invoked from inside the SCSI EH.
1994 mutex_lock(&rport->mutex);
1996 scmnd->result = srp_chkready(target->rport);
1997 if (unlikely(scmnd->result))
2000 WARN_ON_ONCE(scmnd->request->tag < 0);
2001 tag = blk_mq_unique_tag(scmnd->request);
2002 ch = &target->ch[blk_mq_unique_tag_to_hwq(tag)];
2003 idx = blk_mq_unique_tag_to_tag(tag);
2004 WARN_ONCE(idx >= target->req_ring_size, "%s: tag %#x: idx %d >= %d\n",
2005 dev_name(&shost->shost_gendev), tag, idx,
2006 target->req_ring_size);
2008 spin_lock_irqsave(&ch->lock, flags);
2009 iu = __srp_get_tx_iu(ch, SRP_IU_CMD);
2010 spin_unlock_irqrestore(&ch->lock, flags);
2015 req = &ch->req_ring[idx];
2016 dev = target->srp_host->srp_dev->dev;
2017 ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_iu_len,
2020 scmnd->host_scribble = (void *) req;
2023 memset(cmd, 0, sizeof *cmd);
2025 cmd->opcode = SRP_CMD;
2026 int_to_scsilun(scmnd->device->lun, &cmd->lun);
2028 memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
2033 len = srp_map_data(scmnd, ch, req);
2035 shost_printk(KERN_ERR, target->scsi_host,
2036 PFX "Failed to map data (%d)\n", len);
2038 * If we ran out of memory descriptors (-ENOMEM) because an
2039 * application is queuing many requests with more than
2040 * max_pages_per_mr sg-list elements, tell the SCSI mid-layer
2041 * to reduce queue depth temporarily.
2043 scmnd->result = len == -ENOMEM ?
2044 DID_OK << 16 | QUEUE_FULL << 1 : DID_ERROR << 16;
2048 ib_dma_sync_single_for_device(dev, iu->dma, target->max_iu_len,
2051 if (srp_post_send(ch, iu, len)) {
2052 shost_printk(KERN_ERR, target->scsi_host, PFX "Send failed\n");
2060 mutex_unlock(&rport->mutex);
2065 srp_unmap_data(scmnd, ch, req);
2068 srp_put_tx_iu(ch, iu, SRP_IU_CMD);
2071 * Avoid that the loops that iterate over the request ring can
2072 * encounter a dangling SCSI command pointer.
2077 if (scmnd->result) {
2078 scmnd->scsi_done(scmnd);
2081 ret = SCSI_MLQUEUE_HOST_BUSY;
2088 * Note: the resources allocated in this function are freed in
2091 static int srp_alloc_iu_bufs(struct srp_rdma_ch *ch)
2093 struct srp_target_port *target = ch->target;
2096 ch->rx_ring = kcalloc(target->queue_size, sizeof(*ch->rx_ring),
2100 ch->tx_ring = kcalloc(target->queue_size, sizeof(*ch->tx_ring),
2105 for (i = 0; i < target->queue_size; ++i) {
2106 ch->rx_ring[i] = srp_alloc_iu(target->srp_host,
2108 GFP_KERNEL, DMA_FROM_DEVICE);
2109 if (!ch->rx_ring[i])
2113 for (i = 0; i < target->queue_size; ++i) {
2114 ch->tx_ring[i] = srp_alloc_iu(target->srp_host,
2116 GFP_KERNEL, DMA_TO_DEVICE);
2117 if (!ch->tx_ring[i])
2120 list_add(&ch->tx_ring[i]->list, &ch->free_tx);
2126 for (i = 0; i < target->queue_size; ++i) {
2127 srp_free_iu(target->srp_host, ch->rx_ring[i]);
2128 srp_free_iu(target->srp_host, ch->tx_ring[i]);
2141 static uint32_t srp_compute_rq_tmo(struct ib_qp_attr *qp_attr, int attr_mask)
2143 uint64_t T_tr_ns, max_compl_time_ms;
2144 uint32_t rq_tmo_jiffies;
2147 * According to section 11.2.4.2 in the IBTA spec (Modify Queue Pair,
2148 * table 91), both the QP timeout and the retry count have to be set
2149 * for RC QP's during the RTR to RTS transition.
2151 WARN_ON_ONCE((attr_mask & (IB_QP_TIMEOUT | IB_QP_RETRY_CNT)) !=
2152 (IB_QP_TIMEOUT | IB_QP_RETRY_CNT));
2155 * Set target->rq_tmo_jiffies to one second more than the largest time
2156 * it can take before an error completion is generated. See also
2157 * C9-140..142 in the IBTA spec for more information about how to
2158 * convert the QP Local ACK Timeout value to nanoseconds.
2160 T_tr_ns = 4096 * (1ULL << qp_attr->timeout);
2161 max_compl_time_ms = qp_attr->retry_cnt * 4 * T_tr_ns;
2162 do_div(max_compl_time_ms, NSEC_PER_MSEC);
2163 rq_tmo_jiffies = msecs_to_jiffies(max_compl_time_ms + 1000);
2165 return rq_tmo_jiffies;
2168 static void srp_cm_rep_handler(struct ib_cm_id *cm_id,
2169 const struct srp_login_rsp *lrsp,
2170 struct srp_rdma_ch *ch)
2172 struct srp_target_port *target = ch->target;
2173 struct ib_qp_attr *qp_attr = NULL;
2178 if (lrsp->opcode == SRP_LOGIN_RSP) {
2179 ch->max_ti_iu_len = be32_to_cpu(lrsp->max_ti_iu_len);
2180 ch->req_lim = be32_to_cpu(lrsp->req_lim_delta);
2183 * Reserve credits for task management so we don't
2184 * bounce requests back to the SCSI mid-layer.
2186 target->scsi_host->can_queue
2187 = min(ch->req_lim - SRP_TSK_MGMT_SQ_SIZE,
2188 target->scsi_host->can_queue);
2189 target->scsi_host->cmd_per_lun
2190 = min_t(int, target->scsi_host->can_queue,
2191 target->scsi_host->cmd_per_lun);
2193 shost_printk(KERN_WARNING, target->scsi_host,
2194 PFX "Unhandled RSP opcode %#x\n", lrsp->opcode);
2200 ret = srp_alloc_iu_bufs(ch);
2206 qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
2210 qp_attr->qp_state = IB_QPS_RTR;
2211 ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
2215 ret = ib_modify_qp(ch->qp, qp_attr, attr_mask);
2219 for (i = 0; i < target->queue_size; i++) {
2220 struct srp_iu *iu = ch->rx_ring[i];
2222 ret = srp_post_recv(ch, iu);
2227 qp_attr->qp_state = IB_QPS_RTS;
2228 ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
2232 target->rq_tmo_jiffies = srp_compute_rq_tmo(qp_attr, attr_mask);
2234 ret = ib_modify_qp(ch->qp, qp_attr, attr_mask);
2238 ret = ib_send_cm_rtu(cm_id, NULL, 0);
2247 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
2248 struct ib_cm_event *event,
2249 struct srp_rdma_ch *ch)
2251 struct srp_target_port *target = ch->target;
2252 struct Scsi_Host *shost = target->scsi_host;
2253 struct ib_class_port_info *cpi;
2256 switch (event->param.rej_rcvd.reason) {
2257 case IB_CM_REJ_PORT_CM_REDIRECT:
2258 cpi = event->param.rej_rcvd.ari;
2259 ch->path.dlid = cpi->redirect_lid;
2260 ch->path.pkey = cpi->redirect_pkey;
2261 cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
2262 memcpy(ch->path.dgid.raw, cpi->redirect_gid, 16);
2264 ch->status = ch->path.dlid ?
2265 SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
2268 case IB_CM_REJ_PORT_REDIRECT:
2269 if (srp_target_is_topspin(target)) {
2271 * Topspin/Cisco SRP gateways incorrectly send
2272 * reject reason code 25 when they mean 24
2275 memcpy(ch->path.dgid.raw,
2276 event->param.rej_rcvd.ari, 16);
2278 shost_printk(KERN_DEBUG, shost,
2279 PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
2280 be64_to_cpu(ch->path.dgid.global.subnet_prefix),
2281 be64_to_cpu(ch->path.dgid.global.interface_id));
2283 ch->status = SRP_PORT_REDIRECT;
2285 shost_printk(KERN_WARNING, shost,
2286 " REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
2287 ch->status = -ECONNRESET;
2291 case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
2292 shost_printk(KERN_WARNING, shost,
2293 " REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
2294 ch->status = -ECONNRESET;
2297 case IB_CM_REJ_CONSUMER_DEFINED:
2298 opcode = *(u8 *) event->private_data;
2299 if (opcode == SRP_LOGIN_REJ) {
2300 struct srp_login_rej *rej = event->private_data;
2301 u32 reason = be32_to_cpu(rej->reason);
2303 if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
2304 shost_printk(KERN_WARNING, shost,
2305 PFX "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
2307 shost_printk(KERN_WARNING, shost, PFX
2308 "SRP LOGIN from %pI6 to %pI6 REJECTED, reason 0x%08x\n",
2310 target->orig_dgid.raw, reason);
2312 shost_printk(KERN_WARNING, shost,
2313 " REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
2314 " opcode 0x%02x\n", opcode);
2315 ch->status = -ECONNRESET;
2318 case IB_CM_REJ_STALE_CONN:
2319 shost_printk(KERN_WARNING, shost, " REJ reason: stale connection\n");
2320 ch->status = SRP_STALE_CONN;
2324 shost_printk(KERN_WARNING, shost, " REJ reason 0x%x\n",
2325 event->param.rej_rcvd.reason);
2326 ch->status = -ECONNRESET;
2330 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
2332 struct srp_rdma_ch *ch = cm_id->context;
2333 struct srp_target_port *target = ch->target;
2336 switch (event->event) {
2337 case IB_CM_REQ_ERROR:
2338 shost_printk(KERN_DEBUG, target->scsi_host,
2339 PFX "Sending CM REQ failed\n");
2341 ch->status = -ECONNRESET;
2344 case IB_CM_REP_RECEIVED:
2346 srp_cm_rep_handler(cm_id, event->private_data, ch);
2349 case IB_CM_REJ_RECEIVED:
2350 shost_printk(KERN_DEBUG, target->scsi_host, PFX "REJ received\n");
2353 srp_cm_rej_handler(cm_id, event, ch);
2356 case IB_CM_DREQ_RECEIVED:
2357 shost_printk(KERN_WARNING, target->scsi_host,
2358 PFX "DREQ received - connection closed\n");
2359 ch->connected = false;
2360 if (ib_send_cm_drep(cm_id, NULL, 0))
2361 shost_printk(KERN_ERR, target->scsi_host,
2362 PFX "Sending CM DREP failed\n");
2363 queue_work(system_long_wq, &target->tl_err_work);
2366 case IB_CM_TIMEWAIT_EXIT:
2367 shost_printk(KERN_ERR, target->scsi_host,
2368 PFX "connection closed\n");
2374 case IB_CM_MRA_RECEIVED:
2375 case IB_CM_DREQ_ERROR:
2376 case IB_CM_DREP_RECEIVED:
2380 shost_printk(KERN_WARNING, target->scsi_host,
2381 PFX "Unhandled CM event %d\n", event->event);
2386 complete(&ch->done);
2392 * srp_change_queue_depth - setting device queue depth
2393 * @sdev: scsi device struct
2394 * @qdepth: requested queue depth
2396 * Returns queue depth.
2399 srp_change_queue_depth(struct scsi_device *sdev, int qdepth)
2401 if (!sdev->tagged_supported)
2403 return scsi_change_queue_depth(sdev, qdepth);
2406 static int srp_send_tsk_mgmt(struct srp_rdma_ch *ch, u64 req_tag, u64 lun,
2409 struct srp_target_port *target = ch->target;
2410 struct srp_rport *rport = target->rport;
2411 struct ib_device *dev = target->srp_host->srp_dev->dev;
2413 struct srp_tsk_mgmt *tsk_mgmt;
2415 if (!ch->connected || target->qp_in_error)
2418 init_completion(&ch->tsk_mgmt_done);
2421 * Lock the rport mutex to avoid that srp_create_ch_ib() is
2422 * invoked while a task management function is being sent.
2424 mutex_lock(&rport->mutex);
2425 spin_lock_irq(&ch->lock);
2426 iu = __srp_get_tx_iu(ch, SRP_IU_TSK_MGMT);
2427 spin_unlock_irq(&ch->lock);
2430 mutex_unlock(&rport->mutex);
2435 ib_dma_sync_single_for_cpu(dev, iu->dma, sizeof *tsk_mgmt,
2438 memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
2440 tsk_mgmt->opcode = SRP_TSK_MGMT;
2441 int_to_scsilun(lun, &tsk_mgmt->lun);
2442 tsk_mgmt->tag = req_tag | SRP_TAG_TSK_MGMT;
2443 tsk_mgmt->tsk_mgmt_func = func;
2444 tsk_mgmt->task_tag = req_tag;
2446 ib_dma_sync_single_for_device(dev, iu->dma, sizeof *tsk_mgmt,
2448 if (srp_post_send(ch, iu, sizeof(*tsk_mgmt))) {
2449 srp_put_tx_iu(ch, iu, SRP_IU_TSK_MGMT);
2450 mutex_unlock(&rport->mutex);
2454 mutex_unlock(&rport->mutex);
2456 if (!wait_for_completion_timeout(&ch->tsk_mgmt_done,
2457 msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
2463 static int srp_abort(struct scsi_cmnd *scmnd)
2465 struct srp_target_port *target = host_to_target(scmnd->device->host);
2466 struct srp_request *req = (struct srp_request *) scmnd->host_scribble;
2469 struct srp_rdma_ch *ch;
2472 shost_printk(KERN_ERR, target->scsi_host, "SRP abort called\n");
2476 tag = blk_mq_unique_tag(scmnd->request);
2477 ch_idx = blk_mq_unique_tag_to_hwq(tag);
2478 if (WARN_ON_ONCE(ch_idx >= target->ch_count))
2480 ch = &target->ch[ch_idx];
2481 if (!srp_claim_req(ch, req, NULL, scmnd))
2483 shost_printk(KERN_ERR, target->scsi_host,
2484 "Sending SRP abort for tag %#x\n", tag);
2485 if (srp_send_tsk_mgmt(ch, tag, scmnd->device->lun,
2486 SRP_TSK_ABORT_TASK) == 0)
2488 else if (target->rport->state == SRP_RPORT_LOST)
2492 srp_free_req(ch, req, scmnd, 0);
2493 scmnd->result = DID_ABORT << 16;
2494 scmnd->scsi_done(scmnd);
2499 static int srp_reset_device(struct scsi_cmnd *scmnd)
2501 struct srp_target_port *target = host_to_target(scmnd->device->host);
2502 struct srp_rdma_ch *ch;
2505 shost_printk(KERN_ERR, target->scsi_host, "SRP reset_device called\n");
2507 ch = &target->ch[0];
2508 if (srp_send_tsk_mgmt(ch, SRP_TAG_NO_REQ, scmnd->device->lun,
2511 if (ch->tsk_mgmt_status)
2514 for (i = 0; i < target->ch_count; i++) {
2515 ch = &target->ch[i];
2516 for (i = 0; i < target->req_ring_size; ++i) {
2517 struct srp_request *req = &ch->req_ring[i];
2519 srp_finish_req(ch, req, scmnd->device, DID_RESET << 16);
2526 static int srp_reset_host(struct scsi_cmnd *scmnd)
2528 struct srp_target_port *target = host_to_target(scmnd->device->host);
2530 shost_printk(KERN_ERR, target->scsi_host, PFX "SRP reset_host called\n");
2532 return srp_reconnect_rport(target->rport) == 0 ? SUCCESS : FAILED;
2535 static int srp_slave_configure(struct scsi_device *sdev)
2537 struct Scsi_Host *shost = sdev->host;
2538 struct srp_target_port *target = host_to_target(shost);
2539 struct request_queue *q = sdev->request_queue;
2540 unsigned long timeout;
2542 if (sdev->type == TYPE_DISK) {
2543 timeout = max_t(unsigned, 30 * HZ, target->rq_tmo_jiffies);
2544 blk_queue_rq_timeout(q, timeout);
2550 static ssize_t show_id_ext(struct device *dev, struct device_attribute *attr,
2553 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2555 return sprintf(buf, "0x%016llx\n", be64_to_cpu(target->id_ext));
2558 static ssize_t show_ioc_guid(struct device *dev, struct device_attribute *attr,
2561 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2563 return sprintf(buf, "0x%016llx\n", be64_to_cpu(target->ioc_guid));
2566 static ssize_t show_service_id(struct device *dev,
2567 struct device_attribute *attr, char *buf)
2569 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2571 return sprintf(buf, "0x%016llx\n", be64_to_cpu(target->service_id));
2574 static ssize_t show_pkey(struct device *dev, struct device_attribute *attr,
2577 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2579 return sprintf(buf, "0x%04x\n", be16_to_cpu(target->pkey));
2582 static ssize_t show_sgid(struct device *dev, struct device_attribute *attr,
2585 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2587 return sprintf(buf, "%pI6\n", target->sgid.raw);
2590 static ssize_t show_dgid(struct device *dev, struct device_attribute *attr,
2593 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2594 struct srp_rdma_ch *ch = &target->ch[0];
2596 return sprintf(buf, "%pI6\n", ch->path.dgid.raw);
2599 static ssize_t show_orig_dgid(struct device *dev,
2600 struct device_attribute *attr, char *buf)
2602 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2604 return sprintf(buf, "%pI6\n", target->orig_dgid.raw);
2607 static ssize_t show_req_lim(struct device *dev,
2608 struct device_attribute *attr, char *buf)
2610 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2611 struct srp_rdma_ch *ch;
2612 int i, req_lim = INT_MAX;
2614 for (i = 0; i < target->ch_count; i++) {
2615 ch = &target->ch[i];
2616 req_lim = min(req_lim, ch->req_lim);
2618 return sprintf(buf, "%d\n", req_lim);
2621 static ssize_t show_zero_req_lim(struct device *dev,
2622 struct device_attribute *attr, char *buf)
2624 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2626 return sprintf(buf, "%d\n", target->zero_req_lim);
2629 static ssize_t show_local_ib_port(struct device *dev,
2630 struct device_attribute *attr, char *buf)
2632 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2634 return sprintf(buf, "%d\n", target->srp_host->port);
2637 static ssize_t show_local_ib_device(struct device *dev,
2638 struct device_attribute *attr, char *buf)
2640 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2642 return sprintf(buf, "%s\n", target->srp_host->srp_dev->dev->name);
2645 static ssize_t show_ch_count(struct device *dev, struct device_attribute *attr,
2648 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2650 return sprintf(buf, "%d\n", target->ch_count);
2653 static ssize_t show_comp_vector(struct device *dev,
2654 struct device_attribute *attr, char *buf)
2656 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2658 return sprintf(buf, "%d\n", target->comp_vector);
2661 static ssize_t show_tl_retry_count(struct device *dev,
2662 struct device_attribute *attr, char *buf)
2664 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2666 return sprintf(buf, "%d\n", target->tl_retry_count);
2669 static ssize_t show_cmd_sg_entries(struct device *dev,
2670 struct device_attribute *attr, char *buf)
2672 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2674 return sprintf(buf, "%u\n", target->cmd_sg_cnt);
2677 static ssize_t show_allow_ext_sg(struct device *dev,
2678 struct device_attribute *attr, char *buf)
2680 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2682 return sprintf(buf, "%s\n", target->allow_ext_sg ? "true" : "false");
2685 static DEVICE_ATTR(id_ext, S_IRUGO, show_id_ext, NULL);
2686 static DEVICE_ATTR(ioc_guid, S_IRUGO, show_ioc_guid, NULL);
2687 static DEVICE_ATTR(service_id, S_IRUGO, show_service_id, NULL);
2688 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
2689 static DEVICE_ATTR(sgid, S_IRUGO, show_sgid, NULL);
2690 static DEVICE_ATTR(dgid, S_IRUGO, show_dgid, NULL);
2691 static DEVICE_ATTR(orig_dgid, S_IRUGO, show_orig_dgid, NULL);
2692 static DEVICE_ATTR(req_lim, S_IRUGO, show_req_lim, NULL);
2693 static DEVICE_ATTR(zero_req_lim, S_IRUGO, show_zero_req_lim, NULL);
2694 static DEVICE_ATTR(local_ib_port, S_IRUGO, show_local_ib_port, NULL);
2695 static DEVICE_ATTR(local_ib_device, S_IRUGO, show_local_ib_device, NULL);
2696 static DEVICE_ATTR(ch_count, S_IRUGO, show_ch_count, NULL);
2697 static DEVICE_ATTR(comp_vector, S_IRUGO, show_comp_vector, NULL);
2698 static DEVICE_ATTR(tl_retry_count, S_IRUGO, show_tl_retry_count, NULL);
2699 static DEVICE_ATTR(cmd_sg_entries, S_IRUGO, show_cmd_sg_entries, NULL);
2700 static DEVICE_ATTR(allow_ext_sg, S_IRUGO, show_allow_ext_sg, NULL);
2702 static struct device_attribute *srp_host_attrs[] = {
2705 &dev_attr_service_id,
2709 &dev_attr_orig_dgid,
2711 &dev_attr_zero_req_lim,
2712 &dev_attr_local_ib_port,
2713 &dev_attr_local_ib_device,
2715 &dev_attr_comp_vector,
2716 &dev_attr_tl_retry_count,
2717 &dev_attr_cmd_sg_entries,
2718 &dev_attr_allow_ext_sg,
2722 static struct scsi_host_template srp_template = {
2723 .module = THIS_MODULE,
2724 .name = "InfiniBand SRP initiator",
2725 .proc_name = DRV_NAME,
2726 .slave_configure = srp_slave_configure,
2727 .info = srp_target_info,
2728 .queuecommand = srp_queuecommand,
2729 .change_queue_depth = srp_change_queue_depth,
2730 .eh_abort_handler = srp_abort,
2731 .eh_device_reset_handler = srp_reset_device,
2732 .eh_host_reset_handler = srp_reset_host,
2733 .skip_settle_delay = true,
2734 .sg_tablesize = SRP_DEF_SG_TABLESIZE,
2735 .can_queue = SRP_DEFAULT_CMD_SQ_SIZE,
2737 .cmd_per_lun = SRP_DEFAULT_CMD_SQ_SIZE,
2738 .use_clustering = ENABLE_CLUSTERING,
2739 .shost_attrs = srp_host_attrs,
2741 .track_queue_depth = 1,
2744 static int srp_sdev_count(struct Scsi_Host *host)
2746 struct scsi_device *sdev;
2749 shost_for_each_device(sdev, host)
2757 * < 0 upon failure. Caller is responsible for SRP target port cleanup.
2758 * 0 and target->state == SRP_TARGET_REMOVED if asynchronous target port
2759 * removal has been scheduled.
2760 * 0 and target->state != SRP_TARGET_REMOVED upon success.
2762 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
2764 struct srp_rport_identifiers ids;
2765 struct srp_rport *rport;
2767 target->state = SRP_TARGET_SCANNING;
2768 sprintf(target->target_name, "SRP.T10:%016llX",
2769 be64_to_cpu(target->id_ext));
2771 if (scsi_add_host(target->scsi_host, host->srp_dev->dev->dma_device))
2774 memcpy(ids.port_id, &target->id_ext, 8);
2775 memcpy(ids.port_id + 8, &target->ioc_guid, 8);
2776 ids.roles = SRP_RPORT_ROLE_TARGET;
2777 rport = srp_rport_add(target->scsi_host, &ids);
2778 if (IS_ERR(rport)) {
2779 scsi_remove_host(target->scsi_host);
2780 return PTR_ERR(rport);
2783 rport->lld_data = target;
2784 target->rport = rport;
2786 spin_lock(&host->target_lock);
2787 list_add_tail(&target->list, &host->target_list);
2788 spin_unlock(&host->target_lock);
2790 scsi_scan_target(&target->scsi_host->shost_gendev,
2791 0, target->scsi_id, SCAN_WILD_CARD, 0);
2793 if (srp_connected_ch(target) < target->ch_count ||
2794 target->qp_in_error) {
2795 shost_printk(KERN_INFO, target->scsi_host,
2796 PFX "SCSI scan failed - removing SCSI host\n");
2797 srp_queue_remove_work(target);
2801 pr_debug(PFX "%s: SCSI scan succeeded - detected %d LUNs\n",
2802 dev_name(&target->scsi_host->shost_gendev),
2803 srp_sdev_count(target->scsi_host));
2805 spin_lock_irq(&target->lock);
2806 if (target->state == SRP_TARGET_SCANNING)
2807 target->state = SRP_TARGET_LIVE;
2808 spin_unlock_irq(&target->lock);
2814 static void srp_release_dev(struct device *dev)
2816 struct srp_host *host =
2817 container_of(dev, struct srp_host, dev);
2819 complete(&host->released);
2822 static struct class srp_class = {
2823 .name = "infiniband_srp",
2824 .dev_release = srp_release_dev
2828 * srp_conn_unique() - check whether the connection to a target is unique
2830 * @target: SRP target port.
2832 static bool srp_conn_unique(struct srp_host *host,
2833 struct srp_target_port *target)
2835 struct srp_target_port *t;
2838 if (target->state == SRP_TARGET_REMOVED)
2843 spin_lock(&host->target_lock);
2844 list_for_each_entry(t, &host->target_list, list) {
2846 target->id_ext == t->id_ext &&
2847 target->ioc_guid == t->ioc_guid &&
2848 target->initiator_ext == t->initiator_ext) {
2853 spin_unlock(&host->target_lock);
2860 * Target ports are added by writing
2862 * id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
2863 * pkey=<P_Key>,service_id=<service ID>
2865 * to the add_target sysfs attribute.
2869 SRP_OPT_ID_EXT = 1 << 0,
2870 SRP_OPT_IOC_GUID = 1 << 1,
2871 SRP_OPT_DGID = 1 << 2,
2872 SRP_OPT_PKEY = 1 << 3,
2873 SRP_OPT_SERVICE_ID = 1 << 4,
2874 SRP_OPT_MAX_SECT = 1 << 5,
2875 SRP_OPT_MAX_CMD_PER_LUN = 1 << 6,
2876 SRP_OPT_IO_CLASS = 1 << 7,
2877 SRP_OPT_INITIATOR_EXT = 1 << 8,
2878 SRP_OPT_CMD_SG_ENTRIES = 1 << 9,
2879 SRP_OPT_ALLOW_EXT_SG = 1 << 10,
2880 SRP_OPT_SG_TABLESIZE = 1 << 11,
2881 SRP_OPT_COMP_VECTOR = 1 << 12,
2882 SRP_OPT_TL_RETRY_COUNT = 1 << 13,
2883 SRP_OPT_QUEUE_SIZE = 1 << 14,
2884 SRP_OPT_ALL = (SRP_OPT_ID_EXT |
2888 SRP_OPT_SERVICE_ID),
2891 static const match_table_t srp_opt_tokens = {
2892 { SRP_OPT_ID_EXT, "id_ext=%s" },
2893 { SRP_OPT_IOC_GUID, "ioc_guid=%s" },
2894 { SRP_OPT_DGID, "dgid=%s" },
2895 { SRP_OPT_PKEY, "pkey=%x" },
2896 { SRP_OPT_SERVICE_ID, "service_id=%s" },
2897 { SRP_OPT_MAX_SECT, "max_sect=%d" },
2898 { SRP_OPT_MAX_CMD_PER_LUN, "max_cmd_per_lun=%d" },
2899 { SRP_OPT_IO_CLASS, "io_class=%x" },
2900 { SRP_OPT_INITIATOR_EXT, "initiator_ext=%s" },
2901 { SRP_OPT_CMD_SG_ENTRIES, "cmd_sg_entries=%u" },
2902 { SRP_OPT_ALLOW_EXT_SG, "allow_ext_sg=%u" },
2903 { SRP_OPT_SG_TABLESIZE, "sg_tablesize=%u" },
2904 { SRP_OPT_COMP_VECTOR, "comp_vector=%u" },
2905 { SRP_OPT_TL_RETRY_COUNT, "tl_retry_count=%u" },
2906 { SRP_OPT_QUEUE_SIZE, "queue_size=%d" },
2907 { SRP_OPT_ERR, NULL }
2910 static int srp_parse_options(const char *buf, struct srp_target_port *target)
2912 char *options, *sep_opt;
2915 substring_t args[MAX_OPT_ARGS];
2921 options = kstrdup(buf, GFP_KERNEL);
2926 while ((p = strsep(&sep_opt, ",\n")) != NULL) {
2930 token = match_token(p, srp_opt_tokens, args);
2934 case SRP_OPT_ID_EXT:
2935 p = match_strdup(args);
2940 target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2944 case SRP_OPT_IOC_GUID:
2945 p = match_strdup(args);
2950 target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
2955 p = match_strdup(args);
2960 if (strlen(p) != 32) {
2961 pr_warn("bad dest GID parameter '%s'\n", p);
2966 for (i = 0; i < 16; ++i) {
2967 strlcpy(dgid, p + i * 2, sizeof(dgid));
2968 if (sscanf(dgid, "%hhx",
2969 &target->orig_dgid.raw[i]) < 1) {
2979 if (match_hex(args, &token)) {
2980 pr_warn("bad P_Key parameter '%s'\n", p);
2983 target->pkey = cpu_to_be16(token);
2986 case SRP_OPT_SERVICE_ID:
2987 p = match_strdup(args);
2992 target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
2996 case SRP_OPT_MAX_SECT:
2997 if (match_int(args, &token)) {
2998 pr_warn("bad max sect parameter '%s'\n", p);
3001 target->scsi_host->max_sectors = token;
3004 case SRP_OPT_QUEUE_SIZE:
3005 if (match_int(args, &token) || token < 1) {
3006 pr_warn("bad queue_size parameter '%s'\n", p);
3009 target->scsi_host->can_queue = token;
3010 target->queue_size = token + SRP_RSP_SQ_SIZE +
3011 SRP_TSK_MGMT_SQ_SIZE;
3012 if (!(opt_mask & SRP_OPT_MAX_CMD_PER_LUN))
3013 target->scsi_host->cmd_per_lun = token;
3016 case SRP_OPT_MAX_CMD_PER_LUN:
3017 if (match_int(args, &token) || token < 1) {
3018 pr_warn("bad max cmd_per_lun parameter '%s'\n",
3022 target->scsi_host->cmd_per_lun = token;
3025 case SRP_OPT_IO_CLASS:
3026 if (match_hex(args, &token)) {
3027 pr_warn("bad IO class parameter '%s'\n", p);
3030 if (token != SRP_REV10_IB_IO_CLASS &&
3031 token != SRP_REV16A_IB_IO_CLASS) {
3032 pr_warn("unknown IO class parameter value %x specified (use %x or %x).\n",
3033 token, SRP_REV10_IB_IO_CLASS,
3034 SRP_REV16A_IB_IO_CLASS);
3037 target->io_class = token;
3040 case SRP_OPT_INITIATOR_EXT:
3041 p = match_strdup(args);
3046 target->initiator_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
3050 case SRP_OPT_CMD_SG_ENTRIES:
3051 if (match_int(args, &token) || token < 1 || token > 255) {
3052 pr_warn("bad max cmd_sg_entries parameter '%s'\n",
3056 target->cmd_sg_cnt = token;
3059 case SRP_OPT_ALLOW_EXT_SG:
3060 if (match_int(args, &token)) {
3061 pr_warn("bad allow_ext_sg parameter '%s'\n", p);
3064 target->allow_ext_sg = !!token;
3067 case SRP_OPT_SG_TABLESIZE:
3068 if (match_int(args, &token) || token < 1 ||
3069 token > SCSI_MAX_SG_CHAIN_SEGMENTS) {
3070 pr_warn("bad max sg_tablesize parameter '%s'\n",
3074 target->sg_tablesize = token;
3077 case SRP_OPT_COMP_VECTOR:
3078 if (match_int(args, &token) || token < 0) {
3079 pr_warn("bad comp_vector parameter '%s'\n", p);
3082 target->comp_vector = token;
3085 case SRP_OPT_TL_RETRY_COUNT:
3086 if (match_int(args, &token) || token < 2 || token > 7) {
3087 pr_warn("bad tl_retry_count parameter '%s' (must be a number between 2 and 7)\n",
3091 target->tl_retry_count = token;
3095 pr_warn("unknown parameter or missing value '%s' in target creation request\n",
3101 if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
3104 for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
3105 if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
3106 !(srp_opt_tokens[i].token & opt_mask))
3107 pr_warn("target creation request is missing parameter '%s'\n",
3108 srp_opt_tokens[i].pattern);
3110 if (target->scsi_host->cmd_per_lun > target->scsi_host->can_queue
3111 && (opt_mask & SRP_OPT_MAX_CMD_PER_LUN))
3112 pr_warn("cmd_per_lun = %d > queue_size = %d\n",
3113 target->scsi_host->cmd_per_lun,
3114 target->scsi_host->can_queue);
3121 static ssize_t srp_create_target(struct device *dev,
3122 struct device_attribute *attr,
3123 const char *buf, size_t count)
3125 struct srp_host *host =
3126 container_of(dev, struct srp_host, dev);
3127 struct Scsi_Host *target_host;
3128 struct srp_target_port *target;
3129 struct srp_rdma_ch *ch;
3130 struct srp_device *srp_dev = host->srp_dev;
3131 struct ib_device *ibdev = srp_dev->dev;
3132 int ret, node_idx, node, cpu, i;
3133 bool multich = false;
3135 target_host = scsi_host_alloc(&srp_template,
3136 sizeof (struct srp_target_port));
3140 target_host->transportt = ib_srp_transport_template;
3141 target_host->max_channel = 0;
3142 target_host->max_id = 1;
3143 target_host->max_lun = -1LL;
3144 target_host->max_cmd_len = sizeof ((struct srp_cmd *) (void *) 0L)->cdb;
3146 target = host_to_target(target_host);
3148 target->io_class = SRP_REV16A_IB_IO_CLASS;
3149 target->scsi_host = target_host;
3150 target->srp_host = host;
3151 target->lkey = host->srp_dev->pd->local_dma_lkey;
3152 target->rkey = host->srp_dev->mr->rkey;
3153 target->cmd_sg_cnt = cmd_sg_entries;
3154 target->sg_tablesize = indirect_sg_entries ? : cmd_sg_entries;
3155 target->allow_ext_sg = allow_ext_sg;
3156 target->tl_retry_count = 7;
3157 target->queue_size = SRP_DEFAULT_QUEUE_SIZE;
3160 * Avoid that the SCSI host can be removed by srp_remove_target()
3161 * before this function returns.
3163 scsi_host_get(target->scsi_host);
3165 mutex_lock(&host->add_target_mutex);
3167 ret = srp_parse_options(buf, target);
3171 ret = scsi_init_shared_tag_map(target_host, target_host->can_queue);
3175 target->req_ring_size = target->queue_size - SRP_TSK_MGMT_SQ_SIZE;
3177 if (!srp_conn_unique(target->srp_host, target)) {
3178 shost_printk(KERN_INFO, target->scsi_host,
3179 PFX "Already connected to target port with id_ext=%016llx;ioc_guid=%016llx;initiator_ext=%016llx\n",
3180 be64_to_cpu(target->id_ext),
3181 be64_to_cpu(target->ioc_guid),
3182 be64_to_cpu(target->initiator_ext));
3187 if (!srp_dev->has_fmr && !srp_dev->has_fr && !target->allow_ext_sg &&
3188 target->cmd_sg_cnt < target->sg_tablesize) {
3189 pr_warn("No MR pool and no external indirect descriptors, limiting sg_tablesize to cmd_sg_cnt\n");
3190 target->sg_tablesize = target->cmd_sg_cnt;
3193 target_host->sg_tablesize = target->sg_tablesize;
3194 target->indirect_size = target->sg_tablesize *
3195 sizeof (struct srp_direct_buf);
3196 target->max_iu_len = sizeof (struct srp_cmd) +
3197 sizeof (struct srp_indirect_buf) +
3198 target->cmd_sg_cnt * sizeof (struct srp_direct_buf);
3200 INIT_WORK(&target->tl_err_work, srp_tl_err_work);
3201 INIT_WORK(&target->remove_work, srp_remove_work);
3202 spin_lock_init(&target->lock);
3203 ret = ib_query_gid(ibdev, host->port, 0, &target->sgid);
3208 target->ch_count = max_t(unsigned, num_online_nodes(),
3210 min(4 * num_online_nodes(),
3211 ibdev->num_comp_vectors),
3212 num_online_cpus()));
3213 target->ch = kcalloc(target->ch_count, sizeof(*target->ch),
3219 for_each_online_node(node) {
3220 const int ch_start = (node_idx * target->ch_count /
3221 num_online_nodes());
3222 const int ch_end = ((node_idx + 1) * target->ch_count /
3223 num_online_nodes());
3224 const int cv_start = (node_idx * ibdev->num_comp_vectors /
3225 num_online_nodes() + target->comp_vector)
3226 % ibdev->num_comp_vectors;
3227 const int cv_end = ((node_idx + 1) * ibdev->num_comp_vectors /
3228 num_online_nodes() + target->comp_vector)
3229 % ibdev->num_comp_vectors;
3232 for_each_online_cpu(cpu) {
3233 if (cpu_to_node(cpu) != node)
3235 if (ch_start + cpu_idx >= ch_end)
3237 ch = &target->ch[ch_start + cpu_idx];
3238 ch->target = target;
3239 ch->comp_vector = cv_start == cv_end ? cv_start :
3240 cv_start + cpu_idx % (cv_end - cv_start);
3241 spin_lock_init(&ch->lock);
3242 INIT_LIST_HEAD(&ch->free_tx);
3243 ret = srp_new_cm_id(ch);
3245 goto err_disconnect;
3247 ret = srp_create_ch_ib(ch);
3249 goto err_disconnect;
3251 ret = srp_alloc_req_data(ch);
3253 goto err_disconnect;
3255 ret = srp_connect_ch(ch, multich);
3257 shost_printk(KERN_ERR, target->scsi_host,
3258 PFX "Connection %d/%d failed\n",
3261 if (node_idx == 0 && cpu_idx == 0) {
3262 goto err_disconnect;
3264 srp_free_ch_ib(target, ch);
3265 srp_free_req_data(target, ch);
3266 target->ch_count = ch - target->ch;
3278 target->scsi_host->nr_hw_queues = target->ch_count;
3280 ret = srp_add_target(host, target);
3282 goto err_disconnect;
3284 if (target->state != SRP_TARGET_REMOVED) {
3285 shost_printk(KERN_DEBUG, target->scsi_host, PFX
3286 "new target: id_ext %016llx ioc_guid %016llx pkey %04x service_id %016llx sgid %pI6 dgid %pI6\n",
3287 be64_to_cpu(target->id_ext),
3288 be64_to_cpu(target->ioc_guid),
3289 be16_to_cpu(target->pkey),
3290 be64_to_cpu(target->service_id),
3291 target->sgid.raw, target->orig_dgid.raw);
3297 mutex_unlock(&host->add_target_mutex);
3299 scsi_host_put(target->scsi_host);
3301 scsi_host_put(target->scsi_host);
3306 srp_disconnect_target(target);
3308 for (i = 0; i < target->ch_count; i++) {
3309 ch = &target->ch[i];
3310 srp_free_ch_ib(target, ch);
3311 srp_free_req_data(target, ch);
3318 static DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
3320 static ssize_t show_ibdev(struct device *dev, struct device_attribute *attr,
3323 struct srp_host *host = container_of(dev, struct srp_host, dev);
3325 return sprintf(buf, "%s\n", host->srp_dev->dev->name);
3328 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
3330 static ssize_t show_port(struct device *dev, struct device_attribute *attr,
3333 struct srp_host *host = container_of(dev, struct srp_host, dev);
3335 return sprintf(buf, "%d\n", host->port);
3338 static DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
3340 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
3342 struct srp_host *host;
3344 host = kzalloc(sizeof *host, GFP_KERNEL);
3348 INIT_LIST_HEAD(&host->target_list);
3349 spin_lock_init(&host->target_lock);
3350 init_completion(&host->released);
3351 mutex_init(&host->add_target_mutex);
3352 host->srp_dev = device;
3355 host->dev.class = &srp_class;
3356 host->dev.parent = device->dev->dma_device;
3357 dev_set_name(&host->dev, "srp-%s-%d", device->dev->name, port);
3359 if (device_register(&host->dev))
3361 if (device_create_file(&host->dev, &dev_attr_add_target))
3363 if (device_create_file(&host->dev, &dev_attr_ibdev))
3365 if (device_create_file(&host->dev, &dev_attr_port))
3371 device_unregister(&host->dev);
3379 static void srp_add_one(struct ib_device *device)
3381 struct srp_device *srp_dev;
3382 struct ib_device_attr *dev_attr;
3383 struct srp_host *host;
3384 int mr_page_shift, p;
3385 u64 max_pages_per_mr;
3387 dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
3391 if (ib_query_device(device, dev_attr)) {
3392 pr_warn("Query device failed for %s\n", device->name);
3396 srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
3400 srp_dev->has_fmr = (device->alloc_fmr && device->dealloc_fmr &&
3401 device->map_phys_fmr && device->unmap_fmr);
3402 srp_dev->has_fr = (dev_attr->device_cap_flags &
3403 IB_DEVICE_MEM_MGT_EXTENSIONS);
3404 if (!srp_dev->has_fmr && !srp_dev->has_fr)
3405 dev_warn(&device->dev, "neither FMR nor FR is supported\n");
3407 srp_dev->use_fast_reg = (srp_dev->has_fr &&
3408 (!srp_dev->has_fmr || prefer_fr));
3411 * Use the smallest page size supported by the HCA, down to a
3412 * minimum of 4096 bytes. We're unlikely to build large sglists
3413 * out of smaller entries.
3415 mr_page_shift = max(12, ffs(dev_attr->page_size_cap) - 1);
3416 srp_dev->mr_page_size = 1 << mr_page_shift;
3417 srp_dev->mr_page_mask = ~((u64) srp_dev->mr_page_size - 1);
3418 max_pages_per_mr = dev_attr->max_mr_size;
3419 do_div(max_pages_per_mr, srp_dev->mr_page_size);
3420 srp_dev->max_pages_per_mr = min_t(u64, SRP_MAX_PAGES_PER_MR,
3422 if (srp_dev->use_fast_reg) {
3423 srp_dev->max_pages_per_mr =
3424 min_t(u32, srp_dev->max_pages_per_mr,
3425 dev_attr->max_fast_reg_page_list_len);
3427 srp_dev->mr_max_size = srp_dev->mr_page_size *
3428 srp_dev->max_pages_per_mr;
3429 pr_debug("%s: mr_page_shift = %d, dev_attr->max_mr_size = %#llx, dev_attr->max_fast_reg_page_list_len = %u, max_pages_per_mr = %d, mr_max_size = %#x\n",
3430 device->name, mr_page_shift, dev_attr->max_mr_size,
3431 dev_attr->max_fast_reg_page_list_len,
3432 srp_dev->max_pages_per_mr, srp_dev->mr_max_size);
3434 INIT_LIST_HEAD(&srp_dev->dev_list);
3436 srp_dev->dev = device;
3437 srp_dev->pd = ib_alloc_pd(device);
3438 if (IS_ERR(srp_dev->pd))
3441 srp_dev->mr = ib_get_dma_mr(srp_dev->pd,
3442 IB_ACCESS_LOCAL_WRITE |
3443 IB_ACCESS_REMOTE_READ |
3444 IB_ACCESS_REMOTE_WRITE);
3445 if (IS_ERR(srp_dev->mr))
3448 for (p = rdma_start_port(device); p <= rdma_end_port(device); ++p) {
3449 host = srp_add_port(srp_dev, p);
3451 list_add_tail(&host->list, &srp_dev->dev_list);
3454 ib_set_client_data(device, &srp_client, srp_dev);
3459 ib_dealloc_pd(srp_dev->pd);
3468 static void srp_remove_one(struct ib_device *device, void *client_data)
3470 struct srp_device *srp_dev;
3471 struct srp_host *host, *tmp_host;
3472 struct srp_target_port *target;
3474 srp_dev = client_data;
3478 list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
3479 device_unregister(&host->dev);
3481 * Wait for the sysfs entry to go away, so that no new
3482 * target ports can be created.
3484 wait_for_completion(&host->released);
3487 * Remove all target ports.
3489 spin_lock(&host->target_lock);
3490 list_for_each_entry(target, &host->target_list, list)
3491 srp_queue_remove_work(target);
3492 spin_unlock(&host->target_lock);
3495 * Wait for tl_err and target port removal tasks.
3497 flush_workqueue(system_long_wq);
3498 flush_workqueue(srp_remove_wq);
3503 ib_dereg_mr(srp_dev->mr);
3504 ib_dealloc_pd(srp_dev->pd);
3509 static struct srp_function_template ib_srp_transport_functions = {
3510 .has_rport_state = true,
3511 .reset_timer_if_blocked = true,
3512 .reconnect_delay = &srp_reconnect_delay,
3513 .fast_io_fail_tmo = &srp_fast_io_fail_tmo,
3514 .dev_loss_tmo = &srp_dev_loss_tmo,
3515 .reconnect = srp_rport_reconnect,
3516 .rport_delete = srp_rport_delete,
3517 .terminate_rport_io = srp_terminate_io,
3520 static int __init srp_init_module(void)
3524 BUILD_BUG_ON(FIELD_SIZEOF(struct ib_wc, wr_id) < sizeof(void *));
3526 if (srp_sg_tablesize) {
3527 pr_warn("srp_sg_tablesize is deprecated, please use cmd_sg_entries\n");
3528 if (!cmd_sg_entries)
3529 cmd_sg_entries = srp_sg_tablesize;
3532 if (!cmd_sg_entries)
3533 cmd_sg_entries = SRP_DEF_SG_TABLESIZE;
3535 if (cmd_sg_entries > 255) {
3536 pr_warn("Clamping cmd_sg_entries to 255\n");
3537 cmd_sg_entries = 255;
3540 if (!indirect_sg_entries)
3541 indirect_sg_entries = cmd_sg_entries;
3542 else if (indirect_sg_entries < cmd_sg_entries) {
3543 pr_warn("Bumping up indirect_sg_entries to match cmd_sg_entries (%u)\n",
3545 indirect_sg_entries = cmd_sg_entries;
3548 srp_remove_wq = create_workqueue("srp_remove");
3549 if (!srp_remove_wq) {
3555 ib_srp_transport_template =
3556 srp_attach_transport(&ib_srp_transport_functions);
3557 if (!ib_srp_transport_template)
3560 ret = class_register(&srp_class);
3562 pr_err("couldn't register class infiniband_srp\n");
3566 ib_sa_register_client(&srp_sa_client);
3568 ret = ib_register_client(&srp_client);
3570 pr_err("couldn't register IB client\n");
3578 ib_sa_unregister_client(&srp_sa_client);
3579 class_unregister(&srp_class);
3582 srp_release_transport(ib_srp_transport_template);
3585 destroy_workqueue(srp_remove_wq);
3589 static void __exit srp_cleanup_module(void)
3591 ib_unregister_client(&srp_client);
3592 ib_sa_unregister_client(&srp_sa_client);
3593 class_unregister(&srp_class);
3594 srp_release_transport(ib_srp_transport_template);
3595 destroy_workqueue(srp_remove_wq);
3598 module_init(srp_init_module);
3599 module_exit(srp_cleanup_module);