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 WARN_ON_ONCE(!dma_len);
1265 desc->va = cpu_to_be64(dma_addr);
1266 desc->key = cpu_to_be32(rkey);
1267 desc->len = cpu_to_be32(dma_len);
1269 state->total_len += dma_len;
1274 static int srp_map_finish_fmr(struct srp_map_state *state,
1275 struct srp_rdma_ch *ch)
1277 struct srp_target_port *target = ch->target;
1278 struct srp_device *dev = target->srp_host->srp_dev;
1279 struct ib_pool_fmr *fmr;
1282 if (state->fmr.next >= state->fmr.end)
1285 fmr = ib_fmr_pool_map_phys(ch->fmr_pool, state->pages,
1286 state->npages, io_addr);
1288 return PTR_ERR(fmr);
1290 *state->fmr.next++ = fmr;
1293 srp_map_desc(state, state->base_dma_addr & ~dev->mr_page_mask,
1294 state->dma_len, fmr->fmr->rkey);
1299 static int srp_map_finish_fr(struct srp_map_state *state,
1300 struct srp_rdma_ch *ch)
1302 struct srp_target_port *target = ch->target;
1303 struct srp_device *dev = target->srp_host->srp_dev;
1304 struct ib_send_wr *bad_wr;
1305 struct ib_send_wr wr;
1306 struct srp_fr_desc *desc;
1309 if (state->fr.next >= state->fr.end)
1312 desc = srp_fr_pool_get(ch->fr_pool);
1316 rkey = ib_inc_rkey(desc->mr->rkey);
1317 ib_update_fast_reg_key(desc->mr, rkey);
1319 memcpy(desc->frpl->page_list, state->pages,
1320 sizeof(state->pages[0]) * state->npages);
1322 memset(&wr, 0, sizeof(wr));
1323 wr.opcode = IB_WR_FAST_REG_MR;
1324 wr.wr_id = FAST_REG_WR_ID_MASK;
1325 wr.wr.fast_reg.iova_start = state->base_dma_addr;
1326 wr.wr.fast_reg.page_list = desc->frpl;
1327 wr.wr.fast_reg.page_list_len = state->npages;
1328 wr.wr.fast_reg.page_shift = ilog2(dev->mr_page_size);
1329 wr.wr.fast_reg.length = state->dma_len;
1330 wr.wr.fast_reg.access_flags = (IB_ACCESS_LOCAL_WRITE |
1331 IB_ACCESS_REMOTE_READ |
1332 IB_ACCESS_REMOTE_WRITE);
1333 wr.wr.fast_reg.rkey = desc->mr->lkey;
1335 *state->fr.next++ = desc;
1338 srp_map_desc(state, state->base_dma_addr, state->dma_len,
1341 return ib_post_send(ch->qp, &wr, &bad_wr);
1344 static int srp_finish_mapping(struct srp_map_state *state,
1345 struct srp_rdma_ch *ch)
1347 struct srp_target_port *target = ch->target;
1350 if (state->npages == 0)
1353 if (state->npages == 1 && !register_always)
1354 srp_map_desc(state, state->base_dma_addr, state->dma_len,
1357 ret = target->srp_host->srp_dev->use_fast_reg ?
1358 srp_map_finish_fr(state, ch) :
1359 srp_map_finish_fmr(state, ch);
1369 static int srp_map_sg_entry(struct srp_map_state *state,
1370 struct srp_rdma_ch *ch,
1371 struct scatterlist *sg, int sg_index)
1373 struct srp_target_port *target = ch->target;
1374 struct srp_device *dev = target->srp_host->srp_dev;
1375 struct ib_device *ibdev = dev->dev;
1376 dma_addr_t dma_addr = ib_sg_dma_address(ibdev, sg);
1377 unsigned int dma_len = ib_sg_dma_len(ibdev, sg);
1378 unsigned int len = 0;
1381 WARN_ON_ONCE(!dma_len);
1384 unsigned offset = dma_addr & ~dev->mr_page_mask;
1385 if (state->npages == dev->max_pages_per_mr || offset != 0) {
1386 ret = srp_finish_mapping(state, ch);
1391 len = min_t(unsigned int, dma_len, dev->mr_page_size - offset);
1394 state->base_dma_addr = dma_addr;
1395 state->pages[state->npages++] = dma_addr & dev->mr_page_mask;
1396 state->dma_len += len;
1402 * If the last entry of the MR wasn't a full page, then we need to
1403 * close it out and start a new one -- we can only merge at page
1407 if (len != dev->mr_page_size)
1408 ret = srp_finish_mapping(state, ch);
1412 static int srp_map_sg(struct srp_map_state *state, struct srp_rdma_ch *ch,
1413 struct srp_request *req, struct scatterlist *scat,
1416 struct srp_target_port *target = ch->target;
1417 struct srp_device *dev = target->srp_host->srp_dev;
1418 struct scatterlist *sg;
1422 state->desc = req->indirect_desc;
1423 state->pages = req->map_page;
1424 if (dev->use_fast_reg) {
1425 state->fr.next = req->fr_list;
1426 state->fr.end = req->fr_list + target->cmd_sg_cnt;
1427 use_mr = !!ch->fr_pool;
1429 state->fmr.next = req->fmr_list;
1430 state->fmr.end = req->fmr_list + target->cmd_sg_cnt;
1431 use_mr = !!ch->fmr_pool;
1435 for_each_sg(scat, sg, count, i) {
1436 ret = srp_map_sg_entry(state, ch, sg, i);
1440 ret = srp_finish_mapping(state, ch);
1444 for_each_sg(scat, sg, count, i) {
1445 srp_map_desc(state, ib_sg_dma_address(dev->dev, sg),
1446 ib_sg_dma_len(dev->dev, sg), target->rkey);
1450 req->nmdesc = state->nmdesc;
1457 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_rdma_ch *ch,
1458 struct srp_request *req)
1460 struct srp_target_port *target = ch->target;
1461 struct scatterlist *scat;
1462 struct srp_cmd *cmd = req->cmd->buf;
1463 int len, nents, count;
1464 struct srp_device *dev;
1465 struct ib_device *ibdev;
1466 struct srp_map_state state;
1467 struct srp_indirect_buf *indirect_hdr;
1471 if (!scsi_sglist(scmnd) || scmnd->sc_data_direction == DMA_NONE)
1472 return sizeof (struct srp_cmd);
1474 if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
1475 scmnd->sc_data_direction != DMA_TO_DEVICE) {
1476 shost_printk(KERN_WARNING, target->scsi_host,
1477 PFX "Unhandled data direction %d\n",
1478 scmnd->sc_data_direction);
1482 nents = scsi_sg_count(scmnd);
1483 scat = scsi_sglist(scmnd);
1485 dev = target->srp_host->srp_dev;
1488 count = ib_dma_map_sg(ibdev, scat, nents, scmnd->sc_data_direction);
1489 if (unlikely(count == 0))
1492 fmt = SRP_DATA_DESC_DIRECT;
1493 len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
1495 if (count == 1 && !register_always) {
1497 * The midlayer only generated a single gather/scatter
1498 * entry, or DMA mapping coalesced everything to a
1499 * single entry. So a direct descriptor along with
1500 * the DMA MR suffices.
1502 struct srp_direct_buf *buf = (void *) cmd->add_data;
1504 buf->va = cpu_to_be64(ib_sg_dma_address(ibdev, scat));
1505 buf->key = cpu_to_be32(target->rkey);
1506 buf->len = cpu_to_be32(ib_sg_dma_len(ibdev, scat));
1513 * We have more than one scatter/gather entry, so build our indirect
1514 * descriptor table, trying to merge as many entries as we can.
1516 indirect_hdr = (void *) cmd->add_data;
1518 ib_dma_sync_single_for_cpu(ibdev, req->indirect_dma_addr,
1519 target->indirect_size, DMA_TO_DEVICE);
1521 memset(&state, 0, sizeof(state));
1522 srp_map_sg(&state, ch, req, scat, count);
1524 /* We've mapped the request, now pull as much of the indirect
1525 * descriptor table as we can into the command buffer. If this
1526 * target is not using an external indirect table, we are
1527 * guaranteed to fit into the command, as the SCSI layer won't
1528 * give us more S/G entries than we allow.
1530 if (state.ndesc == 1) {
1532 * Memory registration collapsed the sg-list into one entry,
1533 * so use a direct descriptor.
1535 struct srp_direct_buf *buf = (void *) cmd->add_data;
1537 *buf = req->indirect_desc[0];
1541 if (unlikely(target->cmd_sg_cnt < state.ndesc &&
1542 !target->allow_ext_sg)) {
1543 shost_printk(KERN_ERR, target->scsi_host,
1544 "Could not fit S/G list into SRP_CMD\n");
1548 count = min(state.ndesc, target->cmd_sg_cnt);
1549 table_len = state.ndesc * sizeof (struct srp_direct_buf);
1551 fmt = SRP_DATA_DESC_INDIRECT;
1552 len = sizeof(struct srp_cmd) + sizeof (struct srp_indirect_buf);
1553 len += count * sizeof (struct srp_direct_buf);
1555 memcpy(indirect_hdr->desc_list, req->indirect_desc,
1556 count * sizeof (struct srp_direct_buf));
1558 indirect_hdr->table_desc.va = cpu_to_be64(req->indirect_dma_addr);
1559 indirect_hdr->table_desc.key = cpu_to_be32(target->rkey);
1560 indirect_hdr->table_desc.len = cpu_to_be32(table_len);
1561 indirect_hdr->len = cpu_to_be32(state.total_len);
1563 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1564 cmd->data_out_desc_cnt = count;
1566 cmd->data_in_desc_cnt = count;
1568 ib_dma_sync_single_for_device(ibdev, req->indirect_dma_addr, table_len,
1572 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1573 cmd->buf_fmt = fmt << 4;
1581 * Return an IU and possible credit to the free pool
1583 static void srp_put_tx_iu(struct srp_rdma_ch *ch, struct srp_iu *iu,
1584 enum srp_iu_type iu_type)
1586 unsigned long flags;
1588 spin_lock_irqsave(&ch->lock, flags);
1589 list_add(&iu->list, &ch->free_tx);
1590 if (iu_type != SRP_IU_RSP)
1592 spin_unlock_irqrestore(&ch->lock, flags);
1596 * Must be called with ch->lock held to protect req_lim and free_tx.
1597 * If IU is not sent, it must be returned using srp_put_tx_iu().
1600 * An upper limit for the number of allocated information units for each
1602 * - SRP_IU_CMD: SRP_CMD_SQ_SIZE, since the SCSI mid-layer never queues
1603 * more than Scsi_Host.can_queue requests.
1604 * - SRP_IU_TSK_MGMT: SRP_TSK_MGMT_SQ_SIZE.
1605 * - SRP_IU_RSP: 1, since a conforming SRP target never sends more than
1606 * one unanswered SRP request to an initiator.
1608 static struct srp_iu *__srp_get_tx_iu(struct srp_rdma_ch *ch,
1609 enum srp_iu_type iu_type)
1611 struct srp_target_port *target = ch->target;
1612 s32 rsv = (iu_type == SRP_IU_TSK_MGMT) ? 0 : SRP_TSK_MGMT_SQ_SIZE;
1615 srp_send_completion(ch->send_cq, ch);
1617 if (list_empty(&ch->free_tx))
1620 /* Initiator responses to target requests do not consume credits */
1621 if (iu_type != SRP_IU_RSP) {
1622 if (ch->req_lim <= rsv) {
1623 ++target->zero_req_lim;
1630 iu = list_first_entry(&ch->free_tx, struct srp_iu, list);
1631 list_del(&iu->list);
1635 static int srp_post_send(struct srp_rdma_ch *ch, struct srp_iu *iu, int len)
1637 struct srp_target_port *target = ch->target;
1639 struct ib_send_wr wr, *bad_wr;
1641 list.addr = iu->dma;
1643 list.lkey = target->lkey;
1646 wr.wr_id = (uintptr_t) iu;
1649 wr.opcode = IB_WR_SEND;
1650 wr.send_flags = IB_SEND_SIGNALED;
1652 return ib_post_send(ch->qp, &wr, &bad_wr);
1655 static int srp_post_recv(struct srp_rdma_ch *ch, struct srp_iu *iu)
1657 struct srp_target_port *target = ch->target;
1658 struct ib_recv_wr wr, *bad_wr;
1661 list.addr = iu->dma;
1662 list.length = iu->size;
1663 list.lkey = target->lkey;
1666 wr.wr_id = (uintptr_t) iu;
1670 return ib_post_recv(ch->qp, &wr, &bad_wr);
1673 static void srp_process_rsp(struct srp_rdma_ch *ch, struct srp_rsp *rsp)
1675 struct srp_target_port *target = ch->target;
1676 struct srp_request *req;
1677 struct scsi_cmnd *scmnd;
1678 unsigned long flags;
1680 if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
1681 spin_lock_irqsave(&ch->lock, flags);
1682 ch->req_lim += be32_to_cpu(rsp->req_lim_delta);
1683 spin_unlock_irqrestore(&ch->lock, flags);
1685 ch->tsk_mgmt_status = -1;
1686 if (be32_to_cpu(rsp->resp_data_len) >= 4)
1687 ch->tsk_mgmt_status = rsp->data[3];
1688 complete(&ch->tsk_mgmt_done);
1690 scmnd = scsi_host_find_tag(target->scsi_host, rsp->tag);
1692 req = (void *)scmnd->host_scribble;
1693 scmnd = srp_claim_req(ch, req, NULL, scmnd);
1696 shost_printk(KERN_ERR, target->scsi_host,
1697 "Null scmnd for RSP w/tag %#016llx received on ch %td / QP %#x\n",
1698 rsp->tag, ch - target->ch, ch->qp->qp_num);
1700 spin_lock_irqsave(&ch->lock, flags);
1701 ch->req_lim += be32_to_cpu(rsp->req_lim_delta);
1702 spin_unlock_irqrestore(&ch->lock, flags);
1706 scmnd->result = rsp->status;
1708 if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
1709 memcpy(scmnd->sense_buffer, rsp->data +
1710 be32_to_cpu(rsp->resp_data_len),
1711 min_t(int, be32_to_cpu(rsp->sense_data_len),
1712 SCSI_SENSE_BUFFERSIZE));
1715 if (unlikely(rsp->flags & SRP_RSP_FLAG_DIUNDER))
1716 scsi_set_resid(scmnd, be32_to_cpu(rsp->data_in_res_cnt));
1717 else if (unlikely(rsp->flags & SRP_RSP_FLAG_DIOVER))
1718 scsi_set_resid(scmnd, -be32_to_cpu(rsp->data_in_res_cnt));
1719 else if (unlikely(rsp->flags & SRP_RSP_FLAG_DOUNDER))
1720 scsi_set_resid(scmnd, be32_to_cpu(rsp->data_out_res_cnt));
1721 else if (unlikely(rsp->flags & SRP_RSP_FLAG_DOOVER))
1722 scsi_set_resid(scmnd, -be32_to_cpu(rsp->data_out_res_cnt));
1724 srp_free_req(ch, req, scmnd,
1725 be32_to_cpu(rsp->req_lim_delta));
1727 scmnd->host_scribble = NULL;
1728 scmnd->scsi_done(scmnd);
1732 static int srp_response_common(struct srp_rdma_ch *ch, s32 req_delta,
1735 struct srp_target_port *target = ch->target;
1736 struct ib_device *dev = target->srp_host->srp_dev->dev;
1737 unsigned long flags;
1741 spin_lock_irqsave(&ch->lock, flags);
1742 ch->req_lim += req_delta;
1743 iu = __srp_get_tx_iu(ch, SRP_IU_RSP);
1744 spin_unlock_irqrestore(&ch->lock, flags);
1747 shost_printk(KERN_ERR, target->scsi_host, PFX
1748 "no IU available to send response\n");
1752 ib_dma_sync_single_for_cpu(dev, iu->dma, len, DMA_TO_DEVICE);
1753 memcpy(iu->buf, rsp, len);
1754 ib_dma_sync_single_for_device(dev, iu->dma, len, DMA_TO_DEVICE);
1756 err = srp_post_send(ch, iu, len);
1758 shost_printk(KERN_ERR, target->scsi_host, PFX
1759 "unable to post response: %d\n", err);
1760 srp_put_tx_iu(ch, iu, SRP_IU_RSP);
1766 static void srp_process_cred_req(struct srp_rdma_ch *ch,
1767 struct srp_cred_req *req)
1769 struct srp_cred_rsp rsp = {
1770 .opcode = SRP_CRED_RSP,
1773 s32 delta = be32_to_cpu(req->req_lim_delta);
1775 if (srp_response_common(ch, delta, &rsp, sizeof(rsp)))
1776 shost_printk(KERN_ERR, ch->target->scsi_host, PFX
1777 "problems processing SRP_CRED_REQ\n");
1780 static void srp_process_aer_req(struct srp_rdma_ch *ch,
1781 struct srp_aer_req *req)
1783 struct srp_target_port *target = ch->target;
1784 struct srp_aer_rsp rsp = {
1785 .opcode = SRP_AER_RSP,
1788 s32 delta = be32_to_cpu(req->req_lim_delta);
1790 shost_printk(KERN_ERR, target->scsi_host, PFX
1791 "ignoring AER for LUN %llu\n", scsilun_to_int(&req->lun));
1793 if (srp_response_common(ch, delta, &rsp, sizeof(rsp)))
1794 shost_printk(KERN_ERR, target->scsi_host, PFX
1795 "problems processing SRP_AER_REQ\n");
1798 static void srp_handle_recv(struct srp_rdma_ch *ch, struct ib_wc *wc)
1800 struct srp_target_port *target = ch->target;
1801 struct ib_device *dev = target->srp_host->srp_dev->dev;
1802 struct srp_iu *iu = (struct srp_iu *) (uintptr_t) wc->wr_id;
1806 ib_dma_sync_single_for_cpu(dev, iu->dma, ch->max_ti_iu_len,
1809 opcode = *(u8 *) iu->buf;
1812 shost_printk(KERN_ERR, target->scsi_host,
1813 PFX "recv completion, opcode 0x%02x\n", opcode);
1814 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 8, 1,
1815 iu->buf, wc->byte_len, true);
1820 srp_process_rsp(ch, iu->buf);
1824 srp_process_cred_req(ch, iu->buf);
1828 srp_process_aer_req(ch, iu->buf);
1832 /* XXX Handle target logout */
1833 shost_printk(KERN_WARNING, target->scsi_host,
1834 PFX "Got target logout request\n");
1838 shost_printk(KERN_WARNING, target->scsi_host,
1839 PFX "Unhandled SRP opcode 0x%02x\n", opcode);
1843 ib_dma_sync_single_for_device(dev, iu->dma, ch->max_ti_iu_len,
1846 res = srp_post_recv(ch, iu);
1848 shost_printk(KERN_ERR, target->scsi_host,
1849 PFX "Recv failed with error code %d\n", res);
1853 * srp_tl_err_work() - handle a transport layer error
1854 * @work: Work structure embedded in an SRP target port.
1856 * Note: This function may get invoked before the rport has been created,
1857 * hence the target->rport test.
1859 static void srp_tl_err_work(struct work_struct *work)
1861 struct srp_target_port *target;
1863 target = container_of(work, struct srp_target_port, tl_err_work);
1865 srp_start_tl_fail_timers(target->rport);
1868 static void srp_handle_qp_err(u64 wr_id, enum ib_wc_status wc_status,
1869 bool send_err, struct srp_rdma_ch *ch)
1871 struct srp_target_port *target = ch->target;
1873 if (wr_id == SRP_LAST_WR_ID) {
1874 complete(&ch->done);
1878 if (ch->connected && !target->qp_in_error) {
1879 if (wr_id & LOCAL_INV_WR_ID_MASK) {
1880 shost_printk(KERN_ERR, target->scsi_host, PFX
1881 "LOCAL_INV failed with status %s (%d)\n",
1882 ib_wc_status_msg(wc_status), wc_status);
1883 } else if (wr_id & FAST_REG_WR_ID_MASK) {
1884 shost_printk(KERN_ERR, target->scsi_host, PFX
1885 "FAST_REG_MR failed status %s (%d)\n",
1886 ib_wc_status_msg(wc_status), wc_status);
1888 shost_printk(KERN_ERR, target->scsi_host,
1889 PFX "failed %s status %s (%d) for iu %p\n",
1890 send_err ? "send" : "receive",
1891 ib_wc_status_msg(wc_status), wc_status,
1892 (void *)(uintptr_t)wr_id);
1894 queue_work(system_long_wq, &target->tl_err_work);
1896 target->qp_in_error = true;
1899 static void srp_recv_completion(struct ib_cq *cq, void *ch_ptr)
1901 struct srp_rdma_ch *ch = ch_ptr;
1904 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1905 while (ib_poll_cq(cq, 1, &wc) > 0) {
1906 if (likely(wc.status == IB_WC_SUCCESS)) {
1907 srp_handle_recv(ch, &wc);
1909 srp_handle_qp_err(wc.wr_id, wc.status, false, ch);
1914 static void srp_send_completion(struct ib_cq *cq, void *ch_ptr)
1916 struct srp_rdma_ch *ch = ch_ptr;
1920 while (ib_poll_cq(cq, 1, &wc) > 0) {
1921 if (likely(wc.status == IB_WC_SUCCESS)) {
1922 iu = (struct srp_iu *) (uintptr_t) wc.wr_id;
1923 list_add(&iu->list, &ch->free_tx);
1925 srp_handle_qp_err(wc.wr_id, wc.status, true, ch);
1930 static int srp_queuecommand(struct Scsi_Host *shost, struct scsi_cmnd *scmnd)
1932 struct srp_target_port *target = host_to_target(shost);
1933 struct srp_rport *rport = target->rport;
1934 struct srp_rdma_ch *ch;
1935 struct srp_request *req;
1937 struct srp_cmd *cmd;
1938 struct ib_device *dev;
1939 unsigned long flags;
1943 const bool in_scsi_eh = !in_interrupt() && current == shost->ehandler;
1946 * The SCSI EH thread is the only context from which srp_queuecommand()
1947 * can get invoked for blocked devices (SDEV_BLOCK /
1948 * SDEV_CREATED_BLOCK). Avoid racing with srp_reconnect_rport() by
1949 * locking the rport mutex if invoked from inside the SCSI EH.
1952 mutex_lock(&rport->mutex);
1954 scmnd->result = srp_chkready(target->rport);
1955 if (unlikely(scmnd->result))
1958 WARN_ON_ONCE(scmnd->request->tag < 0);
1959 tag = blk_mq_unique_tag(scmnd->request);
1960 ch = &target->ch[blk_mq_unique_tag_to_hwq(tag)];
1961 idx = blk_mq_unique_tag_to_tag(tag);
1962 WARN_ONCE(idx >= target->req_ring_size, "%s: tag %#x: idx %d >= %d\n",
1963 dev_name(&shost->shost_gendev), tag, idx,
1964 target->req_ring_size);
1966 spin_lock_irqsave(&ch->lock, flags);
1967 iu = __srp_get_tx_iu(ch, SRP_IU_CMD);
1968 spin_unlock_irqrestore(&ch->lock, flags);
1973 req = &ch->req_ring[idx];
1974 dev = target->srp_host->srp_dev->dev;
1975 ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_iu_len,
1978 scmnd->host_scribble = (void *) req;
1981 memset(cmd, 0, sizeof *cmd);
1983 cmd->opcode = SRP_CMD;
1984 int_to_scsilun(scmnd->device->lun, &cmd->lun);
1986 memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
1991 len = srp_map_data(scmnd, ch, req);
1993 shost_printk(KERN_ERR, target->scsi_host,
1994 PFX "Failed to map data (%d)\n", len);
1996 * If we ran out of memory descriptors (-ENOMEM) because an
1997 * application is queuing many requests with more than
1998 * max_pages_per_mr sg-list elements, tell the SCSI mid-layer
1999 * to reduce queue depth temporarily.
2001 scmnd->result = len == -ENOMEM ?
2002 DID_OK << 16 | QUEUE_FULL << 1 : DID_ERROR << 16;
2006 ib_dma_sync_single_for_device(dev, iu->dma, target->max_iu_len,
2009 if (srp_post_send(ch, iu, len)) {
2010 shost_printk(KERN_ERR, target->scsi_host, PFX "Send failed\n");
2018 mutex_unlock(&rport->mutex);
2023 srp_unmap_data(scmnd, ch, req);
2026 srp_put_tx_iu(ch, iu, SRP_IU_CMD);
2029 * Avoid that the loops that iterate over the request ring can
2030 * encounter a dangling SCSI command pointer.
2035 if (scmnd->result) {
2036 scmnd->scsi_done(scmnd);
2039 ret = SCSI_MLQUEUE_HOST_BUSY;
2046 * Note: the resources allocated in this function are freed in
2049 static int srp_alloc_iu_bufs(struct srp_rdma_ch *ch)
2051 struct srp_target_port *target = ch->target;
2054 ch->rx_ring = kcalloc(target->queue_size, sizeof(*ch->rx_ring),
2058 ch->tx_ring = kcalloc(target->queue_size, sizeof(*ch->tx_ring),
2063 for (i = 0; i < target->queue_size; ++i) {
2064 ch->rx_ring[i] = srp_alloc_iu(target->srp_host,
2066 GFP_KERNEL, DMA_FROM_DEVICE);
2067 if (!ch->rx_ring[i])
2071 for (i = 0; i < target->queue_size; ++i) {
2072 ch->tx_ring[i] = srp_alloc_iu(target->srp_host,
2074 GFP_KERNEL, DMA_TO_DEVICE);
2075 if (!ch->tx_ring[i])
2078 list_add(&ch->tx_ring[i]->list, &ch->free_tx);
2084 for (i = 0; i < target->queue_size; ++i) {
2085 srp_free_iu(target->srp_host, ch->rx_ring[i]);
2086 srp_free_iu(target->srp_host, ch->tx_ring[i]);
2099 static uint32_t srp_compute_rq_tmo(struct ib_qp_attr *qp_attr, int attr_mask)
2101 uint64_t T_tr_ns, max_compl_time_ms;
2102 uint32_t rq_tmo_jiffies;
2105 * According to section 11.2.4.2 in the IBTA spec (Modify Queue Pair,
2106 * table 91), both the QP timeout and the retry count have to be set
2107 * for RC QP's during the RTR to RTS transition.
2109 WARN_ON_ONCE((attr_mask & (IB_QP_TIMEOUT | IB_QP_RETRY_CNT)) !=
2110 (IB_QP_TIMEOUT | IB_QP_RETRY_CNT));
2113 * Set target->rq_tmo_jiffies to one second more than the largest time
2114 * it can take before an error completion is generated. See also
2115 * C9-140..142 in the IBTA spec for more information about how to
2116 * convert the QP Local ACK Timeout value to nanoseconds.
2118 T_tr_ns = 4096 * (1ULL << qp_attr->timeout);
2119 max_compl_time_ms = qp_attr->retry_cnt * 4 * T_tr_ns;
2120 do_div(max_compl_time_ms, NSEC_PER_MSEC);
2121 rq_tmo_jiffies = msecs_to_jiffies(max_compl_time_ms + 1000);
2123 return rq_tmo_jiffies;
2126 static void srp_cm_rep_handler(struct ib_cm_id *cm_id,
2127 const struct srp_login_rsp *lrsp,
2128 struct srp_rdma_ch *ch)
2130 struct srp_target_port *target = ch->target;
2131 struct ib_qp_attr *qp_attr = NULL;
2136 if (lrsp->opcode == SRP_LOGIN_RSP) {
2137 ch->max_ti_iu_len = be32_to_cpu(lrsp->max_ti_iu_len);
2138 ch->req_lim = be32_to_cpu(lrsp->req_lim_delta);
2141 * Reserve credits for task management so we don't
2142 * bounce requests back to the SCSI mid-layer.
2144 target->scsi_host->can_queue
2145 = min(ch->req_lim - SRP_TSK_MGMT_SQ_SIZE,
2146 target->scsi_host->can_queue);
2147 target->scsi_host->cmd_per_lun
2148 = min_t(int, target->scsi_host->can_queue,
2149 target->scsi_host->cmd_per_lun);
2151 shost_printk(KERN_WARNING, target->scsi_host,
2152 PFX "Unhandled RSP opcode %#x\n", lrsp->opcode);
2158 ret = srp_alloc_iu_bufs(ch);
2164 qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
2168 qp_attr->qp_state = IB_QPS_RTR;
2169 ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
2173 ret = ib_modify_qp(ch->qp, qp_attr, attr_mask);
2177 for (i = 0; i < target->queue_size; i++) {
2178 struct srp_iu *iu = ch->rx_ring[i];
2180 ret = srp_post_recv(ch, iu);
2185 qp_attr->qp_state = IB_QPS_RTS;
2186 ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
2190 target->rq_tmo_jiffies = srp_compute_rq_tmo(qp_attr, attr_mask);
2192 ret = ib_modify_qp(ch->qp, qp_attr, attr_mask);
2196 ret = ib_send_cm_rtu(cm_id, NULL, 0);
2205 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
2206 struct ib_cm_event *event,
2207 struct srp_rdma_ch *ch)
2209 struct srp_target_port *target = ch->target;
2210 struct Scsi_Host *shost = target->scsi_host;
2211 struct ib_class_port_info *cpi;
2214 switch (event->param.rej_rcvd.reason) {
2215 case IB_CM_REJ_PORT_CM_REDIRECT:
2216 cpi = event->param.rej_rcvd.ari;
2217 ch->path.dlid = cpi->redirect_lid;
2218 ch->path.pkey = cpi->redirect_pkey;
2219 cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
2220 memcpy(ch->path.dgid.raw, cpi->redirect_gid, 16);
2222 ch->status = ch->path.dlid ?
2223 SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
2226 case IB_CM_REJ_PORT_REDIRECT:
2227 if (srp_target_is_topspin(target)) {
2229 * Topspin/Cisco SRP gateways incorrectly send
2230 * reject reason code 25 when they mean 24
2233 memcpy(ch->path.dgid.raw,
2234 event->param.rej_rcvd.ari, 16);
2236 shost_printk(KERN_DEBUG, shost,
2237 PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
2238 be64_to_cpu(ch->path.dgid.global.subnet_prefix),
2239 be64_to_cpu(ch->path.dgid.global.interface_id));
2241 ch->status = SRP_PORT_REDIRECT;
2243 shost_printk(KERN_WARNING, shost,
2244 " REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
2245 ch->status = -ECONNRESET;
2249 case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
2250 shost_printk(KERN_WARNING, shost,
2251 " REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
2252 ch->status = -ECONNRESET;
2255 case IB_CM_REJ_CONSUMER_DEFINED:
2256 opcode = *(u8 *) event->private_data;
2257 if (opcode == SRP_LOGIN_REJ) {
2258 struct srp_login_rej *rej = event->private_data;
2259 u32 reason = be32_to_cpu(rej->reason);
2261 if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
2262 shost_printk(KERN_WARNING, shost,
2263 PFX "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
2265 shost_printk(KERN_WARNING, shost, PFX
2266 "SRP LOGIN from %pI6 to %pI6 REJECTED, reason 0x%08x\n",
2268 target->orig_dgid.raw, reason);
2270 shost_printk(KERN_WARNING, shost,
2271 " REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
2272 " opcode 0x%02x\n", opcode);
2273 ch->status = -ECONNRESET;
2276 case IB_CM_REJ_STALE_CONN:
2277 shost_printk(KERN_WARNING, shost, " REJ reason: stale connection\n");
2278 ch->status = SRP_STALE_CONN;
2282 shost_printk(KERN_WARNING, shost, " REJ reason 0x%x\n",
2283 event->param.rej_rcvd.reason);
2284 ch->status = -ECONNRESET;
2288 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
2290 struct srp_rdma_ch *ch = cm_id->context;
2291 struct srp_target_port *target = ch->target;
2294 switch (event->event) {
2295 case IB_CM_REQ_ERROR:
2296 shost_printk(KERN_DEBUG, target->scsi_host,
2297 PFX "Sending CM REQ failed\n");
2299 ch->status = -ECONNRESET;
2302 case IB_CM_REP_RECEIVED:
2304 srp_cm_rep_handler(cm_id, event->private_data, ch);
2307 case IB_CM_REJ_RECEIVED:
2308 shost_printk(KERN_DEBUG, target->scsi_host, PFX "REJ received\n");
2311 srp_cm_rej_handler(cm_id, event, ch);
2314 case IB_CM_DREQ_RECEIVED:
2315 shost_printk(KERN_WARNING, target->scsi_host,
2316 PFX "DREQ received - connection closed\n");
2317 ch->connected = false;
2318 if (ib_send_cm_drep(cm_id, NULL, 0))
2319 shost_printk(KERN_ERR, target->scsi_host,
2320 PFX "Sending CM DREP failed\n");
2321 queue_work(system_long_wq, &target->tl_err_work);
2324 case IB_CM_TIMEWAIT_EXIT:
2325 shost_printk(KERN_ERR, target->scsi_host,
2326 PFX "connection closed\n");
2332 case IB_CM_MRA_RECEIVED:
2333 case IB_CM_DREQ_ERROR:
2334 case IB_CM_DREP_RECEIVED:
2338 shost_printk(KERN_WARNING, target->scsi_host,
2339 PFX "Unhandled CM event %d\n", event->event);
2344 complete(&ch->done);
2350 * srp_change_queue_depth - setting device queue depth
2351 * @sdev: scsi device struct
2352 * @qdepth: requested queue depth
2354 * Returns queue depth.
2357 srp_change_queue_depth(struct scsi_device *sdev, int qdepth)
2359 if (!sdev->tagged_supported)
2361 return scsi_change_queue_depth(sdev, qdepth);
2364 static int srp_send_tsk_mgmt(struct srp_rdma_ch *ch, u64 req_tag, u64 lun,
2367 struct srp_target_port *target = ch->target;
2368 struct srp_rport *rport = target->rport;
2369 struct ib_device *dev = target->srp_host->srp_dev->dev;
2371 struct srp_tsk_mgmt *tsk_mgmt;
2373 if (!ch->connected || target->qp_in_error)
2376 init_completion(&ch->tsk_mgmt_done);
2379 * Lock the rport mutex to avoid that srp_create_ch_ib() is
2380 * invoked while a task management function is being sent.
2382 mutex_lock(&rport->mutex);
2383 spin_lock_irq(&ch->lock);
2384 iu = __srp_get_tx_iu(ch, SRP_IU_TSK_MGMT);
2385 spin_unlock_irq(&ch->lock);
2388 mutex_unlock(&rport->mutex);
2393 ib_dma_sync_single_for_cpu(dev, iu->dma, sizeof *tsk_mgmt,
2396 memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
2398 tsk_mgmt->opcode = SRP_TSK_MGMT;
2399 int_to_scsilun(lun, &tsk_mgmt->lun);
2400 tsk_mgmt->tag = req_tag | SRP_TAG_TSK_MGMT;
2401 tsk_mgmt->tsk_mgmt_func = func;
2402 tsk_mgmt->task_tag = req_tag;
2404 ib_dma_sync_single_for_device(dev, iu->dma, sizeof *tsk_mgmt,
2406 if (srp_post_send(ch, iu, sizeof(*tsk_mgmt))) {
2407 srp_put_tx_iu(ch, iu, SRP_IU_TSK_MGMT);
2408 mutex_unlock(&rport->mutex);
2412 mutex_unlock(&rport->mutex);
2414 if (!wait_for_completion_timeout(&ch->tsk_mgmt_done,
2415 msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
2421 static int srp_abort(struct scsi_cmnd *scmnd)
2423 struct srp_target_port *target = host_to_target(scmnd->device->host);
2424 struct srp_request *req = (struct srp_request *) scmnd->host_scribble;
2427 struct srp_rdma_ch *ch;
2430 shost_printk(KERN_ERR, target->scsi_host, "SRP abort called\n");
2434 tag = blk_mq_unique_tag(scmnd->request);
2435 ch_idx = blk_mq_unique_tag_to_hwq(tag);
2436 if (WARN_ON_ONCE(ch_idx >= target->ch_count))
2438 ch = &target->ch[ch_idx];
2439 if (!srp_claim_req(ch, req, NULL, scmnd))
2441 shost_printk(KERN_ERR, target->scsi_host,
2442 "Sending SRP abort for tag %#x\n", tag);
2443 if (srp_send_tsk_mgmt(ch, tag, scmnd->device->lun,
2444 SRP_TSK_ABORT_TASK) == 0)
2446 else if (target->rport->state == SRP_RPORT_LOST)
2450 srp_free_req(ch, req, scmnd, 0);
2451 scmnd->result = DID_ABORT << 16;
2452 scmnd->scsi_done(scmnd);
2457 static int srp_reset_device(struct scsi_cmnd *scmnd)
2459 struct srp_target_port *target = host_to_target(scmnd->device->host);
2460 struct srp_rdma_ch *ch;
2463 shost_printk(KERN_ERR, target->scsi_host, "SRP reset_device called\n");
2465 ch = &target->ch[0];
2466 if (srp_send_tsk_mgmt(ch, SRP_TAG_NO_REQ, scmnd->device->lun,
2469 if (ch->tsk_mgmt_status)
2472 for (i = 0; i < target->ch_count; i++) {
2473 ch = &target->ch[i];
2474 for (i = 0; i < target->req_ring_size; ++i) {
2475 struct srp_request *req = &ch->req_ring[i];
2477 srp_finish_req(ch, req, scmnd->device, DID_RESET << 16);
2484 static int srp_reset_host(struct scsi_cmnd *scmnd)
2486 struct srp_target_port *target = host_to_target(scmnd->device->host);
2488 shost_printk(KERN_ERR, target->scsi_host, PFX "SRP reset_host called\n");
2490 return srp_reconnect_rport(target->rport) == 0 ? SUCCESS : FAILED;
2493 static int srp_slave_configure(struct scsi_device *sdev)
2495 struct Scsi_Host *shost = sdev->host;
2496 struct srp_target_port *target = host_to_target(shost);
2497 struct request_queue *q = sdev->request_queue;
2498 unsigned long timeout;
2500 if (sdev->type == TYPE_DISK) {
2501 timeout = max_t(unsigned, 30 * HZ, target->rq_tmo_jiffies);
2502 blk_queue_rq_timeout(q, timeout);
2508 static ssize_t show_id_ext(struct device *dev, struct device_attribute *attr,
2511 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2513 return sprintf(buf, "0x%016llx\n", be64_to_cpu(target->id_ext));
2516 static ssize_t show_ioc_guid(struct device *dev, struct device_attribute *attr,
2519 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2521 return sprintf(buf, "0x%016llx\n", be64_to_cpu(target->ioc_guid));
2524 static ssize_t show_service_id(struct device *dev,
2525 struct device_attribute *attr, char *buf)
2527 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2529 return sprintf(buf, "0x%016llx\n", be64_to_cpu(target->service_id));
2532 static ssize_t show_pkey(struct device *dev, struct device_attribute *attr,
2535 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2537 return sprintf(buf, "0x%04x\n", be16_to_cpu(target->pkey));
2540 static ssize_t show_sgid(struct device *dev, struct device_attribute *attr,
2543 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2545 return sprintf(buf, "%pI6\n", target->sgid.raw);
2548 static ssize_t show_dgid(struct device *dev, struct device_attribute *attr,
2551 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2552 struct srp_rdma_ch *ch = &target->ch[0];
2554 return sprintf(buf, "%pI6\n", ch->path.dgid.raw);
2557 static ssize_t show_orig_dgid(struct device *dev,
2558 struct device_attribute *attr, char *buf)
2560 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2562 return sprintf(buf, "%pI6\n", target->orig_dgid.raw);
2565 static ssize_t show_req_lim(struct device *dev,
2566 struct device_attribute *attr, char *buf)
2568 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2569 struct srp_rdma_ch *ch;
2570 int i, req_lim = INT_MAX;
2572 for (i = 0; i < target->ch_count; i++) {
2573 ch = &target->ch[i];
2574 req_lim = min(req_lim, ch->req_lim);
2576 return sprintf(buf, "%d\n", req_lim);
2579 static ssize_t show_zero_req_lim(struct device *dev,
2580 struct device_attribute *attr, char *buf)
2582 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2584 return sprintf(buf, "%d\n", target->zero_req_lim);
2587 static ssize_t show_local_ib_port(struct device *dev,
2588 struct device_attribute *attr, char *buf)
2590 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2592 return sprintf(buf, "%d\n", target->srp_host->port);
2595 static ssize_t show_local_ib_device(struct device *dev,
2596 struct device_attribute *attr, char *buf)
2598 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2600 return sprintf(buf, "%s\n", target->srp_host->srp_dev->dev->name);
2603 static ssize_t show_ch_count(struct device *dev, struct device_attribute *attr,
2606 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2608 return sprintf(buf, "%d\n", target->ch_count);
2611 static ssize_t show_comp_vector(struct device *dev,
2612 struct device_attribute *attr, char *buf)
2614 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2616 return sprintf(buf, "%d\n", target->comp_vector);
2619 static ssize_t show_tl_retry_count(struct device *dev,
2620 struct device_attribute *attr, char *buf)
2622 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2624 return sprintf(buf, "%d\n", target->tl_retry_count);
2627 static ssize_t show_cmd_sg_entries(struct device *dev,
2628 struct device_attribute *attr, char *buf)
2630 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2632 return sprintf(buf, "%u\n", target->cmd_sg_cnt);
2635 static ssize_t show_allow_ext_sg(struct device *dev,
2636 struct device_attribute *attr, char *buf)
2638 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2640 return sprintf(buf, "%s\n", target->allow_ext_sg ? "true" : "false");
2643 static DEVICE_ATTR(id_ext, S_IRUGO, show_id_ext, NULL);
2644 static DEVICE_ATTR(ioc_guid, S_IRUGO, show_ioc_guid, NULL);
2645 static DEVICE_ATTR(service_id, S_IRUGO, show_service_id, NULL);
2646 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
2647 static DEVICE_ATTR(sgid, S_IRUGO, show_sgid, NULL);
2648 static DEVICE_ATTR(dgid, S_IRUGO, show_dgid, NULL);
2649 static DEVICE_ATTR(orig_dgid, S_IRUGO, show_orig_dgid, NULL);
2650 static DEVICE_ATTR(req_lim, S_IRUGO, show_req_lim, NULL);
2651 static DEVICE_ATTR(zero_req_lim, S_IRUGO, show_zero_req_lim, NULL);
2652 static DEVICE_ATTR(local_ib_port, S_IRUGO, show_local_ib_port, NULL);
2653 static DEVICE_ATTR(local_ib_device, S_IRUGO, show_local_ib_device, NULL);
2654 static DEVICE_ATTR(ch_count, S_IRUGO, show_ch_count, NULL);
2655 static DEVICE_ATTR(comp_vector, S_IRUGO, show_comp_vector, NULL);
2656 static DEVICE_ATTR(tl_retry_count, S_IRUGO, show_tl_retry_count, NULL);
2657 static DEVICE_ATTR(cmd_sg_entries, S_IRUGO, show_cmd_sg_entries, NULL);
2658 static DEVICE_ATTR(allow_ext_sg, S_IRUGO, show_allow_ext_sg, NULL);
2660 static struct device_attribute *srp_host_attrs[] = {
2663 &dev_attr_service_id,
2667 &dev_attr_orig_dgid,
2669 &dev_attr_zero_req_lim,
2670 &dev_attr_local_ib_port,
2671 &dev_attr_local_ib_device,
2673 &dev_attr_comp_vector,
2674 &dev_attr_tl_retry_count,
2675 &dev_attr_cmd_sg_entries,
2676 &dev_attr_allow_ext_sg,
2680 static struct scsi_host_template srp_template = {
2681 .module = THIS_MODULE,
2682 .name = "InfiniBand SRP initiator",
2683 .proc_name = DRV_NAME,
2684 .slave_configure = srp_slave_configure,
2685 .info = srp_target_info,
2686 .queuecommand = srp_queuecommand,
2687 .change_queue_depth = srp_change_queue_depth,
2688 .eh_abort_handler = srp_abort,
2689 .eh_device_reset_handler = srp_reset_device,
2690 .eh_host_reset_handler = srp_reset_host,
2691 .skip_settle_delay = true,
2692 .sg_tablesize = SRP_DEF_SG_TABLESIZE,
2693 .can_queue = SRP_DEFAULT_CMD_SQ_SIZE,
2695 .cmd_per_lun = SRP_DEFAULT_CMD_SQ_SIZE,
2696 .use_clustering = ENABLE_CLUSTERING,
2697 .shost_attrs = srp_host_attrs,
2699 .track_queue_depth = 1,
2702 static int srp_sdev_count(struct Scsi_Host *host)
2704 struct scsi_device *sdev;
2707 shost_for_each_device(sdev, host)
2715 * < 0 upon failure. Caller is responsible for SRP target port cleanup.
2716 * 0 and target->state == SRP_TARGET_REMOVED if asynchronous target port
2717 * removal has been scheduled.
2718 * 0 and target->state != SRP_TARGET_REMOVED upon success.
2720 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
2722 struct srp_rport_identifiers ids;
2723 struct srp_rport *rport;
2725 target->state = SRP_TARGET_SCANNING;
2726 sprintf(target->target_name, "SRP.T10:%016llX",
2727 be64_to_cpu(target->id_ext));
2729 if (scsi_add_host(target->scsi_host, host->srp_dev->dev->dma_device))
2732 memcpy(ids.port_id, &target->id_ext, 8);
2733 memcpy(ids.port_id + 8, &target->ioc_guid, 8);
2734 ids.roles = SRP_RPORT_ROLE_TARGET;
2735 rport = srp_rport_add(target->scsi_host, &ids);
2736 if (IS_ERR(rport)) {
2737 scsi_remove_host(target->scsi_host);
2738 return PTR_ERR(rport);
2741 rport->lld_data = target;
2742 target->rport = rport;
2744 spin_lock(&host->target_lock);
2745 list_add_tail(&target->list, &host->target_list);
2746 spin_unlock(&host->target_lock);
2748 scsi_scan_target(&target->scsi_host->shost_gendev,
2749 0, target->scsi_id, SCAN_WILD_CARD, 0);
2751 if (srp_connected_ch(target) < target->ch_count ||
2752 target->qp_in_error) {
2753 shost_printk(KERN_INFO, target->scsi_host,
2754 PFX "SCSI scan failed - removing SCSI host\n");
2755 srp_queue_remove_work(target);
2759 pr_debug(PFX "%s: SCSI scan succeeded - detected %d LUNs\n",
2760 dev_name(&target->scsi_host->shost_gendev),
2761 srp_sdev_count(target->scsi_host));
2763 spin_lock_irq(&target->lock);
2764 if (target->state == SRP_TARGET_SCANNING)
2765 target->state = SRP_TARGET_LIVE;
2766 spin_unlock_irq(&target->lock);
2772 static void srp_release_dev(struct device *dev)
2774 struct srp_host *host =
2775 container_of(dev, struct srp_host, dev);
2777 complete(&host->released);
2780 static struct class srp_class = {
2781 .name = "infiniband_srp",
2782 .dev_release = srp_release_dev
2786 * srp_conn_unique() - check whether the connection to a target is unique
2788 * @target: SRP target port.
2790 static bool srp_conn_unique(struct srp_host *host,
2791 struct srp_target_port *target)
2793 struct srp_target_port *t;
2796 if (target->state == SRP_TARGET_REMOVED)
2801 spin_lock(&host->target_lock);
2802 list_for_each_entry(t, &host->target_list, list) {
2804 target->id_ext == t->id_ext &&
2805 target->ioc_guid == t->ioc_guid &&
2806 target->initiator_ext == t->initiator_ext) {
2811 spin_unlock(&host->target_lock);
2818 * Target ports are added by writing
2820 * id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
2821 * pkey=<P_Key>,service_id=<service ID>
2823 * to the add_target sysfs attribute.
2827 SRP_OPT_ID_EXT = 1 << 0,
2828 SRP_OPT_IOC_GUID = 1 << 1,
2829 SRP_OPT_DGID = 1 << 2,
2830 SRP_OPT_PKEY = 1 << 3,
2831 SRP_OPT_SERVICE_ID = 1 << 4,
2832 SRP_OPT_MAX_SECT = 1 << 5,
2833 SRP_OPT_MAX_CMD_PER_LUN = 1 << 6,
2834 SRP_OPT_IO_CLASS = 1 << 7,
2835 SRP_OPT_INITIATOR_EXT = 1 << 8,
2836 SRP_OPT_CMD_SG_ENTRIES = 1 << 9,
2837 SRP_OPT_ALLOW_EXT_SG = 1 << 10,
2838 SRP_OPT_SG_TABLESIZE = 1 << 11,
2839 SRP_OPT_COMP_VECTOR = 1 << 12,
2840 SRP_OPT_TL_RETRY_COUNT = 1 << 13,
2841 SRP_OPT_QUEUE_SIZE = 1 << 14,
2842 SRP_OPT_ALL = (SRP_OPT_ID_EXT |
2846 SRP_OPT_SERVICE_ID),
2849 static const match_table_t srp_opt_tokens = {
2850 { SRP_OPT_ID_EXT, "id_ext=%s" },
2851 { SRP_OPT_IOC_GUID, "ioc_guid=%s" },
2852 { SRP_OPT_DGID, "dgid=%s" },
2853 { SRP_OPT_PKEY, "pkey=%x" },
2854 { SRP_OPT_SERVICE_ID, "service_id=%s" },
2855 { SRP_OPT_MAX_SECT, "max_sect=%d" },
2856 { SRP_OPT_MAX_CMD_PER_LUN, "max_cmd_per_lun=%d" },
2857 { SRP_OPT_IO_CLASS, "io_class=%x" },
2858 { SRP_OPT_INITIATOR_EXT, "initiator_ext=%s" },
2859 { SRP_OPT_CMD_SG_ENTRIES, "cmd_sg_entries=%u" },
2860 { SRP_OPT_ALLOW_EXT_SG, "allow_ext_sg=%u" },
2861 { SRP_OPT_SG_TABLESIZE, "sg_tablesize=%u" },
2862 { SRP_OPT_COMP_VECTOR, "comp_vector=%u" },
2863 { SRP_OPT_TL_RETRY_COUNT, "tl_retry_count=%u" },
2864 { SRP_OPT_QUEUE_SIZE, "queue_size=%d" },
2865 { SRP_OPT_ERR, NULL }
2868 static int srp_parse_options(const char *buf, struct srp_target_port *target)
2870 char *options, *sep_opt;
2873 substring_t args[MAX_OPT_ARGS];
2879 options = kstrdup(buf, GFP_KERNEL);
2884 while ((p = strsep(&sep_opt, ",\n")) != NULL) {
2888 token = match_token(p, srp_opt_tokens, args);
2892 case SRP_OPT_ID_EXT:
2893 p = match_strdup(args);
2898 target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2902 case SRP_OPT_IOC_GUID:
2903 p = match_strdup(args);
2908 target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
2913 p = match_strdup(args);
2918 if (strlen(p) != 32) {
2919 pr_warn("bad dest GID parameter '%s'\n", p);
2924 for (i = 0; i < 16; ++i) {
2925 strlcpy(dgid, p + i * 2, sizeof(dgid));
2926 if (sscanf(dgid, "%hhx",
2927 &target->orig_dgid.raw[i]) < 1) {
2937 if (match_hex(args, &token)) {
2938 pr_warn("bad P_Key parameter '%s'\n", p);
2941 target->pkey = cpu_to_be16(token);
2944 case SRP_OPT_SERVICE_ID:
2945 p = match_strdup(args);
2950 target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
2954 case SRP_OPT_MAX_SECT:
2955 if (match_int(args, &token)) {
2956 pr_warn("bad max sect parameter '%s'\n", p);
2959 target->scsi_host->max_sectors = token;
2962 case SRP_OPT_QUEUE_SIZE:
2963 if (match_int(args, &token) || token < 1) {
2964 pr_warn("bad queue_size parameter '%s'\n", p);
2967 target->scsi_host->can_queue = token;
2968 target->queue_size = token + SRP_RSP_SQ_SIZE +
2969 SRP_TSK_MGMT_SQ_SIZE;
2970 if (!(opt_mask & SRP_OPT_MAX_CMD_PER_LUN))
2971 target->scsi_host->cmd_per_lun = token;
2974 case SRP_OPT_MAX_CMD_PER_LUN:
2975 if (match_int(args, &token) || token < 1) {
2976 pr_warn("bad max cmd_per_lun parameter '%s'\n",
2980 target->scsi_host->cmd_per_lun = token;
2983 case SRP_OPT_IO_CLASS:
2984 if (match_hex(args, &token)) {
2985 pr_warn("bad IO class parameter '%s'\n", p);
2988 if (token != SRP_REV10_IB_IO_CLASS &&
2989 token != SRP_REV16A_IB_IO_CLASS) {
2990 pr_warn("unknown IO class parameter value %x specified (use %x or %x).\n",
2991 token, SRP_REV10_IB_IO_CLASS,
2992 SRP_REV16A_IB_IO_CLASS);
2995 target->io_class = token;
2998 case SRP_OPT_INITIATOR_EXT:
2999 p = match_strdup(args);
3004 target->initiator_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
3008 case SRP_OPT_CMD_SG_ENTRIES:
3009 if (match_int(args, &token) || token < 1 || token > 255) {
3010 pr_warn("bad max cmd_sg_entries parameter '%s'\n",
3014 target->cmd_sg_cnt = token;
3017 case SRP_OPT_ALLOW_EXT_SG:
3018 if (match_int(args, &token)) {
3019 pr_warn("bad allow_ext_sg parameter '%s'\n", p);
3022 target->allow_ext_sg = !!token;
3025 case SRP_OPT_SG_TABLESIZE:
3026 if (match_int(args, &token) || token < 1 ||
3027 token > SCSI_MAX_SG_CHAIN_SEGMENTS) {
3028 pr_warn("bad max sg_tablesize parameter '%s'\n",
3032 target->sg_tablesize = token;
3035 case SRP_OPT_COMP_VECTOR:
3036 if (match_int(args, &token) || token < 0) {
3037 pr_warn("bad comp_vector parameter '%s'\n", p);
3040 target->comp_vector = token;
3043 case SRP_OPT_TL_RETRY_COUNT:
3044 if (match_int(args, &token) || token < 2 || token > 7) {
3045 pr_warn("bad tl_retry_count parameter '%s' (must be a number between 2 and 7)\n",
3049 target->tl_retry_count = token;
3053 pr_warn("unknown parameter or missing value '%s' in target creation request\n",
3059 if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
3062 for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
3063 if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
3064 !(srp_opt_tokens[i].token & opt_mask))
3065 pr_warn("target creation request is missing parameter '%s'\n",
3066 srp_opt_tokens[i].pattern);
3068 if (target->scsi_host->cmd_per_lun > target->scsi_host->can_queue
3069 && (opt_mask & SRP_OPT_MAX_CMD_PER_LUN))
3070 pr_warn("cmd_per_lun = %d > queue_size = %d\n",
3071 target->scsi_host->cmd_per_lun,
3072 target->scsi_host->can_queue);
3079 static ssize_t srp_create_target(struct device *dev,
3080 struct device_attribute *attr,
3081 const char *buf, size_t count)
3083 struct srp_host *host =
3084 container_of(dev, struct srp_host, dev);
3085 struct Scsi_Host *target_host;
3086 struct srp_target_port *target;
3087 struct srp_rdma_ch *ch;
3088 struct srp_device *srp_dev = host->srp_dev;
3089 struct ib_device *ibdev = srp_dev->dev;
3090 int ret, node_idx, node, cpu, i;
3091 bool multich = false;
3093 target_host = scsi_host_alloc(&srp_template,
3094 sizeof (struct srp_target_port));
3098 target_host->transportt = ib_srp_transport_template;
3099 target_host->max_channel = 0;
3100 target_host->max_id = 1;
3101 target_host->max_lun = -1LL;
3102 target_host->max_cmd_len = sizeof ((struct srp_cmd *) (void *) 0L)->cdb;
3104 target = host_to_target(target_host);
3106 target->io_class = SRP_REV16A_IB_IO_CLASS;
3107 target->scsi_host = target_host;
3108 target->srp_host = host;
3109 target->lkey = host->srp_dev->pd->local_dma_lkey;
3110 target->rkey = host->srp_dev->mr->rkey;
3111 target->cmd_sg_cnt = cmd_sg_entries;
3112 target->sg_tablesize = indirect_sg_entries ? : cmd_sg_entries;
3113 target->allow_ext_sg = allow_ext_sg;
3114 target->tl_retry_count = 7;
3115 target->queue_size = SRP_DEFAULT_QUEUE_SIZE;
3118 * Avoid that the SCSI host can be removed by srp_remove_target()
3119 * before this function returns.
3121 scsi_host_get(target->scsi_host);
3123 mutex_lock(&host->add_target_mutex);
3125 ret = srp_parse_options(buf, target);
3129 ret = scsi_init_shared_tag_map(target_host, target_host->can_queue);
3133 target->req_ring_size = target->queue_size - SRP_TSK_MGMT_SQ_SIZE;
3135 if (!srp_conn_unique(target->srp_host, target)) {
3136 shost_printk(KERN_INFO, target->scsi_host,
3137 PFX "Already connected to target port with id_ext=%016llx;ioc_guid=%016llx;initiator_ext=%016llx\n",
3138 be64_to_cpu(target->id_ext),
3139 be64_to_cpu(target->ioc_guid),
3140 be64_to_cpu(target->initiator_ext));
3145 if (!srp_dev->has_fmr && !srp_dev->has_fr && !target->allow_ext_sg &&
3146 target->cmd_sg_cnt < target->sg_tablesize) {
3147 pr_warn("No MR pool and no external indirect descriptors, limiting sg_tablesize to cmd_sg_cnt\n");
3148 target->sg_tablesize = target->cmd_sg_cnt;
3151 target_host->sg_tablesize = target->sg_tablesize;
3152 target->indirect_size = target->sg_tablesize *
3153 sizeof (struct srp_direct_buf);
3154 target->max_iu_len = sizeof (struct srp_cmd) +
3155 sizeof (struct srp_indirect_buf) +
3156 target->cmd_sg_cnt * sizeof (struct srp_direct_buf);
3158 INIT_WORK(&target->tl_err_work, srp_tl_err_work);
3159 INIT_WORK(&target->remove_work, srp_remove_work);
3160 spin_lock_init(&target->lock);
3161 ret = ib_query_gid(ibdev, host->port, 0, &target->sgid);
3166 target->ch_count = max_t(unsigned, num_online_nodes(),
3168 min(4 * num_online_nodes(),
3169 ibdev->num_comp_vectors),
3170 num_online_cpus()));
3171 target->ch = kcalloc(target->ch_count, sizeof(*target->ch),
3177 for_each_online_node(node) {
3178 const int ch_start = (node_idx * target->ch_count /
3179 num_online_nodes());
3180 const int ch_end = ((node_idx + 1) * target->ch_count /
3181 num_online_nodes());
3182 const int cv_start = (node_idx * ibdev->num_comp_vectors /
3183 num_online_nodes() + target->comp_vector)
3184 % ibdev->num_comp_vectors;
3185 const int cv_end = ((node_idx + 1) * ibdev->num_comp_vectors /
3186 num_online_nodes() + target->comp_vector)
3187 % ibdev->num_comp_vectors;
3190 for_each_online_cpu(cpu) {
3191 if (cpu_to_node(cpu) != node)
3193 if (ch_start + cpu_idx >= ch_end)
3195 ch = &target->ch[ch_start + cpu_idx];
3196 ch->target = target;
3197 ch->comp_vector = cv_start == cv_end ? cv_start :
3198 cv_start + cpu_idx % (cv_end - cv_start);
3199 spin_lock_init(&ch->lock);
3200 INIT_LIST_HEAD(&ch->free_tx);
3201 ret = srp_new_cm_id(ch);
3203 goto err_disconnect;
3205 ret = srp_create_ch_ib(ch);
3207 goto err_disconnect;
3209 ret = srp_alloc_req_data(ch);
3211 goto err_disconnect;
3213 ret = srp_connect_ch(ch, multich);
3215 shost_printk(KERN_ERR, target->scsi_host,
3216 PFX "Connection %d/%d failed\n",
3219 if (node_idx == 0 && cpu_idx == 0) {
3220 goto err_disconnect;
3222 srp_free_ch_ib(target, ch);
3223 srp_free_req_data(target, ch);
3224 target->ch_count = ch - target->ch;
3236 target->scsi_host->nr_hw_queues = target->ch_count;
3238 ret = srp_add_target(host, target);
3240 goto err_disconnect;
3242 if (target->state != SRP_TARGET_REMOVED) {
3243 shost_printk(KERN_DEBUG, target->scsi_host, PFX
3244 "new target: id_ext %016llx ioc_guid %016llx pkey %04x service_id %016llx sgid %pI6 dgid %pI6\n",
3245 be64_to_cpu(target->id_ext),
3246 be64_to_cpu(target->ioc_guid),
3247 be16_to_cpu(target->pkey),
3248 be64_to_cpu(target->service_id),
3249 target->sgid.raw, target->orig_dgid.raw);
3255 mutex_unlock(&host->add_target_mutex);
3257 scsi_host_put(target->scsi_host);
3259 scsi_host_put(target->scsi_host);
3264 srp_disconnect_target(target);
3266 for (i = 0; i < target->ch_count; i++) {
3267 ch = &target->ch[i];
3268 srp_free_ch_ib(target, ch);
3269 srp_free_req_data(target, ch);
3276 static DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
3278 static ssize_t show_ibdev(struct device *dev, struct device_attribute *attr,
3281 struct srp_host *host = container_of(dev, struct srp_host, dev);
3283 return sprintf(buf, "%s\n", host->srp_dev->dev->name);
3286 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
3288 static ssize_t show_port(struct device *dev, struct device_attribute *attr,
3291 struct srp_host *host = container_of(dev, struct srp_host, dev);
3293 return sprintf(buf, "%d\n", host->port);
3296 static DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
3298 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
3300 struct srp_host *host;
3302 host = kzalloc(sizeof *host, GFP_KERNEL);
3306 INIT_LIST_HEAD(&host->target_list);
3307 spin_lock_init(&host->target_lock);
3308 init_completion(&host->released);
3309 mutex_init(&host->add_target_mutex);
3310 host->srp_dev = device;
3313 host->dev.class = &srp_class;
3314 host->dev.parent = device->dev->dma_device;
3315 dev_set_name(&host->dev, "srp-%s-%d", device->dev->name, port);
3317 if (device_register(&host->dev))
3319 if (device_create_file(&host->dev, &dev_attr_add_target))
3321 if (device_create_file(&host->dev, &dev_attr_ibdev))
3323 if (device_create_file(&host->dev, &dev_attr_port))
3329 device_unregister(&host->dev);
3337 static void srp_add_one(struct ib_device *device)
3339 struct srp_device *srp_dev;
3340 struct ib_device_attr *dev_attr;
3341 struct srp_host *host;
3342 int mr_page_shift, p;
3343 u64 max_pages_per_mr;
3345 dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
3349 if (ib_query_device(device, dev_attr)) {
3350 pr_warn("Query device failed for %s\n", device->name);
3354 srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
3358 srp_dev->has_fmr = (device->alloc_fmr && device->dealloc_fmr &&
3359 device->map_phys_fmr && device->unmap_fmr);
3360 srp_dev->has_fr = (dev_attr->device_cap_flags &
3361 IB_DEVICE_MEM_MGT_EXTENSIONS);
3362 if (!srp_dev->has_fmr && !srp_dev->has_fr)
3363 dev_warn(&device->dev, "neither FMR nor FR is supported\n");
3365 srp_dev->use_fast_reg = (srp_dev->has_fr &&
3366 (!srp_dev->has_fmr || prefer_fr));
3369 * Use the smallest page size supported by the HCA, down to a
3370 * minimum of 4096 bytes. We're unlikely to build large sglists
3371 * out of smaller entries.
3373 mr_page_shift = max(12, ffs(dev_attr->page_size_cap) - 1);
3374 srp_dev->mr_page_size = 1 << mr_page_shift;
3375 srp_dev->mr_page_mask = ~((u64) srp_dev->mr_page_size - 1);
3376 max_pages_per_mr = dev_attr->max_mr_size;
3377 do_div(max_pages_per_mr, srp_dev->mr_page_size);
3378 srp_dev->max_pages_per_mr = min_t(u64, SRP_MAX_PAGES_PER_MR,
3380 if (srp_dev->use_fast_reg) {
3381 srp_dev->max_pages_per_mr =
3382 min_t(u32, srp_dev->max_pages_per_mr,
3383 dev_attr->max_fast_reg_page_list_len);
3385 srp_dev->mr_max_size = srp_dev->mr_page_size *
3386 srp_dev->max_pages_per_mr;
3387 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",
3388 device->name, mr_page_shift, dev_attr->max_mr_size,
3389 dev_attr->max_fast_reg_page_list_len,
3390 srp_dev->max_pages_per_mr, srp_dev->mr_max_size);
3392 INIT_LIST_HEAD(&srp_dev->dev_list);
3394 srp_dev->dev = device;
3395 srp_dev->pd = ib_alloc_pd(device);
3396 if (IS_ERR(srp_dev->pd))
3399 srp_dev->mr = ib_get_dma_mr(srp_dev->pd,
3400 IB_ACCESS_LOCAL_WRITE |
3401 IB_ACCESS_REMOTE_READ |
3402 IB_ACCESS_REMOTE_WRITE);
3403 if (IS_ERR(srp_dev->mr))
3406 for (p = rdma_start_port(device); p <= rdma_end_port(device); ++p) {
3407 host = srp_add_port(srp_dev, p);
3409 list_add_tail(&host->list, &srp_dev->dev_list);
3412 ib_set_client_data(device, &srp_client, srp_dev);
3417 ib_dealloc_pd(srp_dev->pd);
3426 static void srp_remove_one(struct ib_device *device, void *client_data)
3428 struct srp_device *srp_dev;
3429 struct srp_host *host, *tmp_host;
3430 struct srp_target_port *target;
3432 srp_dev = client_data;
3436 list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
3437 device_unregister(&host->dev);
3439 * Wait for the sysfs entry to go away, so that no new
3440 * target ports can be created.
3442 wait_for_completion(&host->released);
3445 * Remove all target ports.
3447 spin_lock(&host->target_lock);
3448 list_for_each_entry(target, &host->target_list, list)
3449 srp_queue_remove_work(target);
3450 spin_unlock(&host->target_lock);
3453 * Wait for tl_err and target port removal tasks.
3455 flush_workqueue(system_long_wq);
3456 flush_workqueue(srp_remove_wq);
3461 ib_dereg_mr(srp_dev->mr);
3462 ib_dealloc_pd(srp_dev->pd);
3467 static struct srp_function_template ib_srp_transport_functions = {
3468 .has_rport_state = true,
3469 .reset_timer_if_blocked = true,
3470 .reconnect_delay = &srp_reconnect_delay,
3471 .fast_io_fail_tmo = &srp_fast_io_fail_tmo,
3472 .dev_loss_tmo = &srp_dev_loss_tmo,
3473 .reconnect = srp_rport_reconnect,
3474 .rport_delete = srp_rport_delete,
3475 .terminate_rport_io = srp_terminate_io,
3478 static int __init srp_init_module(void)
3482 BUILD_BUG_ON(FIELD_SIZEOF(struct ib_wc, wr_id) < sizeof(void *));
3484 if (srp_sg_tablesize) {
3485 pr_warn("srp_sg_tablesize is deprecated, please use cmd_sg_entries\n");
3486 if (!cmd_sg_entries)
3487 cmd_sg_entries = srp_sg_tablesize;
3490 if (!cmd_sg_entries)
3491 cmd_sg_entries = SRP_DEF_SG_TABLESIZE;
3493 if (cmd_sg_entries > 255) {
3494 pr_warn("Clamping cmd_sg_entries to 255\n");
3495 cmd_sg_entries = 255;
3498 if (!indirect_sg_entries)
3499 indirect_sg_entries = cmd_sg_entries;
3500 else if (indirect_sg_entries < cmd_sg_entries) {
3501 pr_warn("Bumping up indirect_sg_entries to match cmd_sg_entries (%u)\n",
3503 indirect_sg_entries = cmd_sg_entries;
3506 srp_remove_wq = create_workqueue("srp_remove");
3507 if (!srp_remove_wq) {
3513 ib_srp_transport_template =
3514 srp_attach_transport(&ib_srp_transport_functions);
3515 if (!ib_srp_transport_template)
3518 ret = class_register(&srp_class);
3520 pr_err("couldn't register class infiniband_srp\n");
3524 ib_sa_register_client(&srp_sa_client);
3526 ret = ib_register_client(&srp_client);
3528 pr_err("couldn't register IB client\n");
3536 ib_sa_unregister_client(&srp_sa_client);
3537 class_unregister(&srp_class);
3540 srp_release_transport(ib_srp_transport_template);
3543 destroy_workqueue(srp_remove_wq);
3547 static void __exit srp_cleanup_module(void)
3549 ib_unregister_client(&srp_client);
3550 ib_sa_unregister_client(&srp_sa_client);
3551 class_unregister(&srp_class);
3552 srp_release_transport(ib_srp_transport_template);
3553 destroy_workqueue(srp_remove_wq);
3556 module_init(srp_init_module);
3557 module_exit(srp_cleanup_module);