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 "1.0"
59 #define DRV_RELDATE "July 1, 2013"
61 MODULE_AUTHOR("Roland Dreier");
62 MODULE_DESCRIPTION("InfiniBand SCSI RDMA Protocol initiator "
63 "v" DRV_VERSION " (" DRV_RELDATE ")");
64 MODULE_LICENSE("Dual BSD/GPL");
66 static unsigned int srp_sg_tablesize;
67 static unsigned int cmd_sg_entries;
68 static unsigned int indirect_sg_entries;
69 static bool allow_ext_sg;
70 static bool prefer_fr;
71 static bool register_always;
72 static int topspin_workarounds = 1;
74 module_param(srp_sg_tablesize, uint, 0444);
75 MODULE_PARM_DESC(srp_sg_tablesize, "Deprecated name for cmd_sg_entries");
77 module_param(cmd_sg_entries, uint, 0444);
78 MODULE_PARM_DESC(cmd_sg_entries,
79 "Default number of gather/scatter entries in the SRP command (default is 12, max 255)");
81 module_param(indirect_sg_entries, uint, 0444);
82 MODULE_PARM_DESC(indirect_sg_entries,
83 "Default max number of gather/scatter entries (default is 12, max is " __stringify(SCSI_MAX_SG_CHAIN_SEGMENTS) ")");
85 module_param(allow_ext_sg, bool, 0444);
86 MODULE_PARM_DESC(allow_ext_sg,
87 "Default behavior when there are more than cmd_sg_entries S/G entries after mapping; fails the request when false (default false)");
89 module_param(topspin_workarounds, int, 0444);
90 MODULE_PARM_DESC(topspin_workarounds,
91 "Enable workarounds for Topspin/Cisco SRP target bugs if != 0");
93 module_param(prefer_fr, bool, 0444);
94 MODULE_PARM_DESC(prefer_fr,
95 "Whether to use fast registration if both FMR and fast registration are supported");
97 module_param(register_always, bool, 0444);
98 MODULE_PARM_DESC(register_always,
99 "Use memory registration even for contiguous memory regions");
101 static struct kernel_param_ops srp_tmo_ops;
103 static int srp_reconnect_delay = 10;
104 module_param_cb(reconnect_delay, &srp_tmo_ops, &srp_reconnect_delay,
106 MODULE_PARM_DESC(reconnect_delay, "Time between successive reconnect attempts");
108 static int srp_fast_io_fail_tmo = 15;
109 module_param_cb(fast_io_fail_tmo, &srp_tmo_ops, &srp_fast_io_fail_tmo,
111 MODULE_PARM_DESC(fast_io_fail_tmo,
112 "Number of seconds between the observation of a transport"
113 " layer error and failing all I/O. \"off\" means that this"
114 " functionality is disabled.");
116 static int srp_dev_loss_tmo = 600;
117 module_param_cb(dev_loss_tmo, &srp_tmo_ops, &srp_dev_loss_tmo,
119 MODULE_PARM_DESC(dev_loss_tmo,
120 "Maximum number of seconds that the SRP transport should"
121 " insulate transport layer errors. After this time has been"
122 " exceeded the SCSI host is removed. Should be"
123 " between 1 and " __stringify(SCSI_DEVICE_BLOCK_MAX_TIMEOUT)
124 " if fast_io_fail_tmo has not been set. \"off\" means that"
125 " this functionality is disabled.");
127 static unsigned ch_count;
128 module_param(ch_count, uint, 0444);
129 MODULE_PARM_DESC(ch_count,
130 "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.");
132 static void srp_add_one(struct ib_device *device);
133 static void srp_remove_one(struct ib_device *device);
134 static void srp_recv_completion(struct ib_cq *cq, void *ch_ptr);
135 static void srp_send_completion(struct ib_cq *cq, void *ch_ptr);
136 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event);
138 static struct scsi_transport_template *ib_srp_transport_template;
139 static struct workqueue_struct *srp_remove_wq;
141 static struct ib_client srp_client = {
144 .remove = srp_remove_one
147 static struct ib_sa_client srp_sa_client;
149 static int srp_tmo_get(char *buffer, const struct kernel_param *kp)
151 int tmo = *(int *)kp->arg;
154 return sprintf(buffer, "%d", tmo);
156 return sprintf(buffer, "off");
159 static int srp_tmo_set(const char *val, const struct kernel_param *kp)
163 if (strncmp(val, "off", 3) != 0) {
164 res = kstrtoint(val, 0, &tmo);
170 if (kp->arg == &srp_reconnect_delay)
171 res = srp_tmo_valid(tmo, srp_fast_io_fail_tmo,
173 else if (kp->arg == &srp_fast_io_fail_tmo)
174 res = srp_tmo_valid(srp_reconnect_delay, tmo, srp_dev_loss_tmo);
176 res = srp_tmo_valid(srp_reconnect_delay, srp_fast_io_fail_tmo,
180 *(int *)kp->arg = tmo;
186 static struct kernel_param_ops srp_tmo_ops = {
191 static inline struct srp_target_port *host_to_target(struct Scsi_Host *host)
193 return (struct srp_target_port *) host->hostdata;
196 static const char *srp_target_info(struct Scsi_Host *host)
198 return host_to_target(host)->target_name;
201 static int srp_target_is_topspin(struct srp_target_port *target)
203 static const u8 topspin_oui[3] = { 0x00, 0x05, 0xad };
204 static const u8 cisco_oui[3] = { 0x00, 0x1b, 0x0d };
206 return topspin_workarounds &&
207 (!memcmp(&target->ioc_guid, topspin_oui, sizeof topspin_oui) ||
208 !memcmp(&target->ioc_guid, cisco_oui, sizeof cisco_oui));
211 static struct srp_iu *srp_alloc_iu(struct srp_host *host, size_t size,
213 enum dma_data_direction direction)
217 iu = kmalloc(sizeof *iu, gfp_mask);
221 iu->buf = kzalloc(size, gfp_mask);
225 iu->dma = ib_dma_map_single(host->srp_dev->dev, iu->buf, size,
227 if (ib_dma_mapping_error(host->srp_dev->dev, iu->dma))
231 iu->direction = direction;
243 static void srp_free_iu(struct srp_host *host, struct srp_iu *iu)
248 ib_dma_unmap_single(host->srp_dev->dev, iu->dma, iu->size,
254 static void srp_qp_event(struct ib_event *event, void *context)
256 pr_debug("QP event %d\n", event->event);
259 static int srp_init_qp(struct srp_target_port *target,
262 struct ib_qp_attr *attr;
265 attr = kmalloc(sizeof *attr, GFP_KERNEL);
269 ret = ib_find_cached_pkey(target->srp_host->srp_dev->dev,
270 target->srp_host->port,
271 be16_to_cpu(target->pkey),
276 attr->qp_state = IB_QPS_INIT;
277 attr->qp_access_flags = (IB_ACCESS_REMOTE_READ |
278 IB_ACCESS_REMOTE_WRITE);
279 attr->port_num = target->srp_host->port;
281 ret = ib_modify_qp(qp, attr,
292 static int srp_new_cm_id(struct srp_rdma_ch *ch)
294 struct srp_target_port *target = ch->target;
295 struct ib_cm_id *new_cm_id;
297 new_cm_id = ib_create_cm_id(target->srp_host->srp_dev->dev,
299 if (IS_ERR(new_cm_id))
300 return PTR_ERR(new_cm_id);
303 ib_destroy_cm_id(ch->cm_id);
304 ch->cm_id = new_cm_id;
305 ch->path.sgid = target->sgid;
306 ch->path.dgid = target->orig_dgid;
307 ch->path.pkey = target->pkey;
308 ch->path.service_id = target->service_id;
313 static struct ib_fmr_pool *srp_alloc_fmr_pool(struct srp_target_port *target)
315 struct srp_device *dev = target->srp_host->srp_dev;
316 struct ib_fmr_pool_param fmr_param;
318 memset(&fmr_param, 0, sizeof(fmr_param));
319 fmr_param.pool_size = target->scsi_host->can_queue;
320 fmr_param.dirty_watermark = fmr_param.pool_size / 4;
322 fmr_param.max_pages_per_fmr = dev->max_pages_per_mr;
323 fmr_param.page_shift = ilog2(dev->mr_page_size);
324 fmr_param.access = (IB_ACCESS_LOCAL_WRITE |
325 IB_ACCESS_REMOTE_WRITE |
326 IB_ACCESS_REMOTE_READ);
328 return ib_create_fmr_pool(dev->pd, &fmr_param);
332 * srp_destroy_fr_pool() - free the resources owned by a pool
333 * @pool: Fast registration pool to be destroyed.
335 static void srp_destroy_fr_pool(struct srp_fr_pool *pool)
338 struct srp_fr_desc *d;
343 for (i = 0, d = &pool->desc[0]; i < pool->size; i++, d++) {
345 ib_free_fast_reg_page_list(d->frpl);
353 * srp_create_fr_pool() - allocate and initialize a pool for fast registration
354 * @device: IB device to allocate fast registration descriptors for.
355 * @pd: Protection domain associated with the FR descriptors.
356 * @pool_size: Number of descriptors to allocate.
357 * @max_page_list_len: Maximum fast registration work request page list length.
359 static struct srp_fr_pool *srp_create_fr_pool(struct ib_device *device,
360 struct ib_pd *pd, int pool_size,
361 int max_page_list_len)
363 struct srp_fr_pool *pool;
364 struct srp_fr_desc *d;
366 struct ib_fast_reg_page_list *frpl;
367 int i, ret = -EINVAL;
372 pool = kzalloc(sizeof(struct srp_fr_pool) +
373 pool_size * sizeof(struct srp_fr_desc), GFP_KERNEL);
376 pool->size = pool_size;
377 pool->max_page_list_len = max_page_list_len;
378 spin_lock_init(&pool->lock);
379 INIT_LIST_HEAD(&pool->free_list);
381 for (i = 0, d = &pool->desc[0]; i < pool->size; i++, d++) {
382 mr = ib_alloc_fast_reg_mr(pd, max_page_list_len);
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;
503 init_attr = kzalloc(sizeof *init_attr, GFP_KERNEL);
507 /* + 1 for SRP_LAST_WR_ID */
508 recv_cq = ib_create_cq(dev->dev, srp_recv_completion, NULL, ch,
509 target->queue_size + 1, ch->comp_vector);
510 if (IS_ERR(recv_cq)) {
511 ret = PTR_ERR(recv_cq);
515 send_cq = ib_create_cq(dev->dev, srp_send_completion, NULL, ch,
516 m * target->queue_size, ch->comp_vector);
517 if (IS_ERR(send_cq)) {
518 ret = PTR_ERR(send_cq);
522 ib_req_notify_cq(recv_cq, IB_CQ_NEXT_COMP);
524 init_attr->event_handler = srp_qp_event;
525 init_attr->cap.max_send_wr = m * target->queue_size;
526 init_attr->cap.max_recv_wr = target->queue_size + 1;
527 init_attr->cap.max_recv_sge = 1;
528 init_attr->cap.max_send_sge = 1;
529 init_attr->sq_sig_type = IB_SIGNAL_REQ_WR;
530 init_attr->qp_type = IB_QPT_RC;
531 init_attr->send_cq = send_cq;
532 init_attr->recv_cq = recv_cq;
534 qp = ib_create_qp(dev->pd, init_attr);
540 ret = srp_init_qp(target, qp);
544 if (dev->use_fast_reg && dev->has_fr) {
545 fr_pool = srp_alloc_fr_pool(target);
546 if (IS_ERR(fr_pool)) {
547 ret = PTR_ERR(fr_pool);
548 shost_printk(KERN_WARNING, target->scsi_host, PFX
549 "FR pool allocation failed (%d)\n", ret);
553 srp_destroy_fr_pool(ch->fr_pool);
554 ch->fr_pool = fr_pool;
555 } else if (!dev->use_fast_reg && dev->has_fmr) {
556 fmr_pool = srp_alloc_fmr_pool(target);
557 if (IS_ERR(fmr_pool)) {
558 ret = PTR_ERR(fmr_pool);
559 shost_printk(KERN_WARNING, target->scsi_host, PFX
560 "FMR pool allocation failed (%d)\n", ret);
564 ib_destroy_fmr_pool(ch->fmr_pool);
565 ch->fmr_pool = fmr_pool;
571 ib_destroy_cq(ch->recv_cq);
573 ib_destroy_cq(ch->send_cq);
576 ch->recv_cq = recv_cq;
577 ch->send_cq = send_cq;
586 ib_destroy_cq(send_cq);
589 ib_destroy_cq(recv_cq);
597 * Note: this function may be called without srp_alloc_iu_bufs() having been
598 * invoked. Hence the ch->[rt]x_ring checks.
600 static void srp_free_ch_ib(struct srp_target_port *target,
601 struct srp_rdma_ch *ch)
603 struct srp_device *dev = target->srp_host->srp_dev;
610 ib_destroy_cm_id(ch->cm_id);
614 /* If srp_new_cm_id() succeeded but srp_create_ch_ib() not, return. */
618 if (dev->use_fast_reg) {
620 srp_destroy_fr_pool(ch->fr_pool);
623 ib_destroy_fmr_pool(ch->fmr_pool);
626 ib_destroy_cq(ch->send_cq);
627 ib_destroy_cq(ch->recv_cq);
630 * Avoid that the SCSI error handler tries to use this channel after
631 * it has been freed. The SCSI error handler can namely continue
632 * trying to perform recovery actions after scsi_remove_host()
638 ch->send_cq = ch->recv_cq = NULL;
641 for (i = 0; i < target->queue_size; ++i)
642 srp_free_iu(target->srp_host, ch->rx_ring[i]);
647 for (i = 0; i < target->queue_size; ++i)
648 srp_free_iu(target->srp_host, ch->tx_ring[i]);
654 static void srp_path_rec_completion(int status,
655 struct ib_sa_path_rec *pathrec,
658 struct srp_rdma_ch *ch = ch_ptr;
659 struct srp_target_port *target = ch->target;
663 shost_printk(KERN_ERR, target->scsi_host,
664 PFX "Got failed path rec status %d\n", status);
670 static int srp_lookup_path(struct srp_rdma_ch *ch)
672 struct srp_target_port *target = ch->target;
675 ch->path.numb_path = 1;
677 init_completion(&ch->done);
679 ch->path_query_id = ib_sa_path_rec_get(&srp_sa_client,
680 target->srp_host->srp_dev->dev,
681 target->srp_host->port,
683 IB_SA_PATH_REC_SERVICE_ID |
684 IB_SA_PATH_REC_DGID |
685 IB_SA_PATH_REC_SGID |
686 IB_SA_PATH_REC_NUMB_PATH |
688 SRP_PATH_REC_TIMEOUT_MS,
690 srp_path_rec_completion,
691 ch, &ch->path_query);
692 if (ch->path_query_id < 0)
693 return ch->path_query_id;
695 ret = wait_for_completion_interruptible(&ch->done);
700 shost_printk(KERN_WARNING, target->scsi_host,
701 PFX "Path record query failed\n");
706 static int srp_send_req(struct srp_rdma_ch *ch, bool multich)
708 struct srp_target_port *target = ch->target;
710 struct ib_cm_req_param param;
711 struct srp_login_req priv;
715 req = kzalloc(sizeof *req, GFP_KERNEL);
719 req->param.primary_path = &ch->path;
720 req->param.alternate_path = NULL;
721 req->param.service_id = target->service_id;
722 req->param.qp_num = ch->qp->qp_num;
723 req->param.qp_type = ch->qp->qp_type;
724 req->param.private_data = &req->priv;
725 req->param.private_data_len = sizeof req->priv;
726 req->param.flow_control = 1;
728 get_random_bytes(&req->param.starting_psn, 4);
729 req->param.starting_psn &= 0xffffff;
732 * Pick some arbitrary defaults here; we could make these
733 * module parameters if anyone cared about setting them.
735 req->param.responder_resources = 4;
736 req->param.remote_cm_response_timeout = 20;
737 req->param.local_cm_response_timeout = 20;
738 req->param.retry_count = target->tl_retry_count;
739 req->param.rnr_retry_count = 7;
740 req->param.max_cm_retries = 15;
742 req->priv.opcode = SRP_LOGIN_REQ;
744 req->priv.req_it_iu_len = cpu_to_be32(target->max_iu_len);
745 req->priv.req_buf_fmt = cpu_to_be16(SRP_BUF_FORMAT_DIRECT |
746 SRP_BUF_FORMAT_INDIRECT);
747 req->priv.req_flags = (multich ? SRP_MULTICHAN_MULTI :
748 SRP_MULTICHAN_SINGLE);
750 * In the published SRP specification (draft rev. 16a), the
751 * port identifier format is 8 bytes of ID extension followed
752 * by 8 bytes of GUID. Older drafts put the two halves in the
753 * opposite order, so that the GUID comes first.
755 * Targets conforming to these obsolete drafts can be
756 * recognized by the I/O Class they report.
758 if (target->io_class == SRP_REV10_IB_IO_CLASS) {
759 memcpy(req->priv.initiator_port_id,
760 &target->sgid.global.interface_id, 8);
761 memcpy(req->priv.initiator_port_id + 8,
762 &target->initiator_ext, 8);
763 memcpy(req->priv.target_port_id, &target->ioc_guid, 8);
764 memcpy(req->priv.target_port_id + 8, &target->id_ext, 8);
766 memcpy(req->priv.initiator_port_id,
767 &target->initiator_ext, 8);
768 memcpy(req->priv.initiator_port_id + 8,
769 &target->sgid.global.interface_id, 8);
770 memcpy(req->priv.target_port_id, &target->id_ext, 8);
771 memcpy(req->priv.target_port_id + 8, &target->ioc_guid, 8);
775 * Topspin/Cisco SRP targets will reject our login unless we
776 * zero out the first 8 bytes of our initiator port ID and set
777 * the second 8 bytes to the local node GUID.
779 if (srp_target_is_topspin(target)) {
780 shost_printk(KERN_DEBUG, target->scsi_host,
781 PFX "Topspin/Cisco initiator port ID workaround "
782 "activated for target GUID %016llx\n",
783 (unsigned long long) be64_to_cpu(target->ioc_guid));
784 memset(req->priv.initiator_port_id, 0, 8);
785 memcpy(req->priv.initiator_port_id + 8,
786 &target->srp_host->srp_dev->dev->node_guid, 8);
789 status = ib_send_cm_req(ch->cm_id, &req->param);
796 static bool srp_queue_remove_work(struct srp_target_port *target)
798 bool changed = false;
800 spin_lock_irq(&target->lock);
801 if (target->state != SRP_TARGET_REMOVED) {
802 target->state = SRP_TARGET_REMOVED;
805 spin_unlock_irq(&target->lock);
808 queue_work(srp_remove_wq, &target->remove_work);
813 static void srp_disconnect_target(struct srp_target_port *target)
815 struct srp_rdma_ch *ch;
818 /* XXX should send SRP_I_LOGOUT request */
820 for (i = 0; i < target->ch_count; i++) {
822 ch->connected = false;
823 if (ch->cm_id && ib_send_cm_dreq(ch->cm_id, NULL, 0)) {
824 shost_printk(KERN_DEBUG, target->scsi_host,
825 PFX "Sending CM DREQ failed\n");
830 static void srp_free_req_data(struct srp_target_port *target,
831 struct srp_rdma_ch *ch)
833 struct srp_device *dev = target->srp_host->srp_dev;
834 struct ib_device *ibdev = dev->dev;
835 struct srp_request *req;
838 if (!ch->target || !ch->req_ring)
841 for (i = 0; i < target->req_ring_size; ++i) {
842 req = &ch->req_ring[i];
843 if (dev->use_fast_reg)
846 kfree(req->fmr_list);
847 kfree(req->map_page);
848 if (req->indirect_dma_addr) {
849 ib_dma_unmap_single(ibdev, req->indirect_dma_addr,
850 target->indirect_size,
853 kfree(req->indirect_desc);
860 static int srp_alloc_req_data(struct srp_rdma_ch *ch)
862 struct srp_target_port *target = ch->target;
863 struct srp_device *srp_dev = target->srp_host->srp_dev;
864 struct ib_device *ibdev = srp_dev->dev;
865 struct srp_request *req;
868 int i, ret = -ENOMEM;
870 ch->req_ring = kcalloc(target->req_ring_size, sizeof(*ch->req_ring),
875 for (i = 0; i < target->req_ring_size; ++i) {
876 req = &ch->req_ring[i];
877 mr_list = kmalloc(target->cmd_sg_cnt * sizeof(void *),
881 if (srp_dev->use_fast_reg)
882 req->fr_list = mr_list;
884 req->fmr_list = mr_list;
885 req->map_page = kmalloc(srp_dev->max_pages_per_mr *
886 sizeof(void *), GFP_KERNEL);
889 req->indirect_desc = kmalloc(target->indirect_size, GFP_KERNEL);
890 if (!req->indirect_desc)
893 dma_addr = ib_dma_map_single(ibdev, req->indirect_desc,
894 target->indirect_size,
896 if (ib_dma_mapping_error(ibdev, dma_addr))
899 req->indirect_dma_addr = dma_addr;
908 * srp_del_scsi_host_attr() - Remove attributes defined in the host template.
909 * @shost: SCSI host whose attributes to remove from sysfs.
911 * Note: Any attributes defined in the host template and that did not exist
912 * before invocation of this function will be ignored.
914 static void srp_del_scsi_host_attr(struct Scsi_Host *shost)
916 struct device_attribute **attr;
918 for (attr = shost->hostt->shost_attrs; attr && *attr; ++attr)
919 device_remove_file(&shost->shost_dev, *attr);
922 static void srp_remove_target(struct srp_target_port *target)
924 struct srp_rdma_ch *ch;
927 WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
929 srp_del_scsi_host_attr(target->scsi_host);
930 srp_rport_get(target->rport);
931 srp_remove_host(target->scsi_host);
932 scsi_remove_host(target->scsi_host);
933 srp_stop_rport_timers(target->rport);
934 srp_disconnect_target(target);
935 for (i = 0; i < target->ch_count; i++) {
937 srp_free_ch_ib(target, ch);
939 cancel_work_sync(&target->tl_err_work);
940 srp_rport_put(target->rport);
941 for (i = 0; i < target->ch_count; i++) {
943 srp_free_req_data(target, ch);
948 spin_lock(&target->srp_host->target_lock);
949 list_del(&target->list);
950 spin_unlock(&target->srp_host->target_lock);
952 scsi_host_put(target->scsi_host);
955 static void srp_remove_work(struct work_struct *work)
957 struct srp_target_port *target =
958 container_of(work, struct srp_target_port, remove_work);
960 WARN_ON_ONCE(target->state != SRP_TARGET_REMOVED);
962 srp_remove_target(target);
965 static void srp_rport_delete(struct srp_rport *rport)
967 struct srp_target_port *target = rport->lld_data;
969 srp_queue_remove_work(target);
973 * srp_connected_ch() - number of connected channels
974 * @target: SRP target port.
976 static int srp_connected_ch(struct srp_target_port *target)
980 for (i = 0; i < target->ch_count; i++)
981 c += target->ch[i].connected;
986 static int srp_connect_ch(struct srp_rdma_ch *ch, bool multich)
988 struct srp_target_port *target = ch->target;
991 WARN_ON_ONCE(!multich && srp_connected_ch(target) > 0);
993 ret = srp_lookup_path(ch);
998 init_completion(&ch->done);
999 ret = srp_send_req(ch, multich);
1002 ret = wait_for_completion_interruptible(&ch->done);
1007 * The CM event handling code will set status to
1008 * SRP_PORT_REDIRECT if we get a port redirect REJ
1009 * back, or SRP_DLID_REDIRECT if we get a lid/qp
1010 * redirect REJ back.
1012 switch (ch->status) {
1014 ch->connected = true;
1017 case SRP_PORT_REDIRECT:
1018 ret = srp_lookup_path(ch);
1023 case SRP_DLID_REDIRECT:
1026 case SRP_STALE_CONN:
1027 shost_printk(KERN_ERR, target->scsi_host, PFX
1028 "giving up on stale connection\n");
1029 ch->status = -ECONNRESET;
1038 static int srp_inv_rkey(struct srp_rdma_ch *ch, u32 rkey)
1040 struct ib_send_wr *bad_wr;
1041 struct ib_send_wr wr = {
1042 .opcode = IB_WR_LOCAL_INV,
1043 .wr_id = LOCAL_INV_WR_ID_MASK,
1047 .ex.invalidate_rkey = rkey,
1050 return ib_post_send(ch->qp, &wr, &bad_wr);
1053 static void srp_unmap_data(struct scsi_cmnd *scmnd,
1054 struct srp_rdma_ch *ch,
1055 struct srp_request *req)
1057 struct srp_target_port *target = ch->target;
1058 struct srp_device *dev = target->srp_host->srp_dev;
1059 struct ib_device *ibdev = dev->dev;
1062 if (!scsi_sglist(scmnd) ||
1063 (scmnd->sc_data_direction != DMA_TO_DEVICE &&
1064 scmnd->sc_data_direction != DMA_FROM_DEVICE))
1067 if (dev->use_fast_reg) {
1068 struct srp_fr_desc **pfr;
1070 for (i = req->nmdesc, pfr = req->fr_list; i > 0; i--, pfr++) {
1071 res = srp_inv_rkey(ch, (*pfr)->mr->rkey);
1073 shost_printk(KERN_ERR, target->scsi_host, PFX
1074 "Queueing INV WR for rkey %#x failed (%d)\n",
1075 (*pfr)->mr->rkey, res);
1076 queue_work(system_long_wq,
1077 &target->tl_err_work);
1081 srp_fr_pool_put(ch->fr_pool, req->fr_list,
1084 struct ib_pool_fmr **pfmr;
1086 for (i = req->nmdesc, pfmr = req->fmr_list; i > 0; i--, pfmr++)
1087 ib_fmr_pool_unmap(*pfmr);
1090 ib_dma_unmap_sg(ibdev, scsi_sglist(scmnd), scsi_sg_count(scmnd),
1091 scmnd->sc_data_direction);
1095 * srp_claim_req - Take ownership of the scmnd associated with a request.
1096 * @ch: SRP RDMA channel.
1097 * @req: SRP request.
1098 * @sdev: If not NULL, only take ownership for this SCSI device.
1099 * @scmnd: If NULL, take ownership of @req->scmnd. If not NULL, only take
1100 * ownership of @req->scmnd if it equals @scmnd.
1103 * Either NULL or a pointer to the SCSI command the caller became owner of.
1105 static struct scsi_cmnd *srp_claim_req(struct srp_rdma_ch *ch,
1106 struct srp_request *req,
1107 struct scsi_device *sdev,
1108 struct scsi_cmnd *scmnd)
1110 unsigned long flags;
1112 spin_lock_irqsave(&ch->lock, flags);
1114 (!sdev || req->scmnd->device == sdev) &&
1115 (!scmnd || req->scmnd == scmnd)) {
1121 spin_unlock_irqrestore(&ch->lock, flags);
1127 * srp_free_req() - Unmap data and add request to the free request list.
1128 * @ch: SRP RDMA channel.
1129 * @req: Request to be freed.
1130 * @scmnd: SCSI command associated with @req.
1131 * @req_lim_delta: Amount to be added to @target->req_lim.
1133 static void srp_free_req(struct srp_rdma_ch *ch, struct srp_request *req,
1134 struct scsi_cmnd *scmnd, s32 req_lim_delta)
1136 unsigned long flags;
1138 srp_unmap_data(scmnd, ch, req);
1140 spin_lock_irqsave(&ch->lock, flags);
1141 ch->req_lim += req_lim_delta;
1142 spin_unlock_irqrestore(&ch->lock, flags);
1145 static void srp_finish_req(struct srp_rdma_ch *ch, struct srp_request *req,
1146 struct scsi_device *sdev, int result)
1148 struct scsi_cmnd *scmnd = srp_claim_req(ch, req, sdev, NULL);
1151 srp_free_req(ch, req, scmnd, 0);
1152 scmnd->result = result;
1153 scmnd->scsi_done(scmnd);
1157 static void srp_terminate_io(struct srp_rport *rport)
1159 struct srp_target_port *target = rport->lld_data;
1160 struct srp_rdma_ch *ch;
1161 struct Scsi_Host *shost = target->scsi_host;
1162 struct scsi_device *sdev;
1166 * Invoking srp_terminate_io() while srp_queuecommand() is running
1167 * is not safe. Hence the warning statement below.
1169 shost_for_each_device(sdev, shost)
1170 WARN_ON_ONCE(sdev->request_queue->request_fn_active);
1172 for (i = 0; i < target->ch_count; i++) {
1173 ch = &target->ch[i];
1175 for (j = 0; j < target->req_ring_size; ++j) {
1176 struct srp_request *req = &ch->req_ring[j];
1178 srp_finish_req(ch, req, NULL,
1179 DID_TRANSPORT_FAILFAST << 16);
1185 * It is up to the caller to ensure that srp_rport_reconnect() calls are
1186 * serialized and that no concurrent srp_queuecommand(), srp_abort(),
1187 * srp_reset_device() or srp_reset_host() calls will occur while this function
1188 * is in progress. One way to realize that is not to call this function
1189 * directly but to call srp_reconnect_rport() instead since that last function
1190 * serializes calls of this function via rport->mutex and also blocks
1191 * srp_queuecommand() calls before invoking this function.
1193 static int srp_rport_reconnect(struct srp_rport *rport)
1195 struct srp_target_port *target = rport->lld_data;
1196 struct srp_rdma_ch *ch;
1198 bool multich = false;
1200 srp_disconnect_target(target);
1202 if (target->state == SRP_TARGET_SCANNING)
1206 * Now get a new local CM ID so that we avoid confusing the target in
1207 * case things are really fouled up. Doing so also ensures that all CM
1208 * callbacks will have finished before a new QP is allocated.
1210 for (i = 0; i < target->ch_count; i++) {
1211 ch = &target->ch[i];
1214 ret += srp_new_cm_id(ch);
1216 for (i = 0; i < target->ch_count; i++) {
1217 ch = &target->ch[i];
1220 for (j = 0; j < target->req_ring_size; ++j) {
1221 struct srp_request *req = &ch->req_ring[j];
1223 srp_finish_req(ch, req, NULL, DID_RESET << 16);
1226 for (i = 0; i < target->ch_count; i++) {
1227 ch = &target->ch[i];
1231 * Whether or not creating a new CM ID succeeded, create a new
1232 * QP. This guarantees that all completion callback function
1233 * invocations have finished before request resetting starts.
1235 ret += srp_create_ch_ib(ch);
1237 INIT_LIST_HEAD(&ch->free_tx);
1238 for (j = 0; j < target->queue_size; ++j)
1239 list_add(&ch->tx_ring[j]->list, &ch->free_tx);
1242 target->qp_in_error = false;
1244 for (i = 0; i < target->ch_count; i++) {
1245 ch = &target->ch[i];
1246 if (ret || !ch->target)
1248 ret = srp_connect_ch(ch, multich);
1253 shost_printk(KERN_INFO, target->scsi_host,
1254 PFX "reconnect succeeded\n");
1259 static void srp_map_desc(struct srp_map_state *state, dma_addr_t dma_addr,
1260 unsigned int dma_len, u32 rkey)
1262 struct srp_direct_buf *desc = state->desc;
1264 desc->va = cpu_to_be64(dma_addr);
1265 desc->key = cpu_to_be32(rkey);
1266 desc->len = cpu_to_be32(dma_len);
1268 state->total_len += dma_len;
1273 static int srp_map_finish_fmr(struct srp_map_state *state,
1274 struct srp_rdma_ch *ch)
1276 struct ib_pool_fmr *fmr;
1279 fmr = ib_fmr_pool_map_phys(ch->fmr_pool, state->pages,
1280 state->npages, io_addr);
1282 return PTR_ERR(fmr);
1284 *state->next_fmr++ = fmr;
1287 srp_map_desc(state, 0, state->dma_len, fmr->fmr->rkey);
1292 static int srp_map_finish_fr(struct srp_map_state *state,
1293 struct srp_rdma_ch *ch)
1295 struct srp_target_port *target = ch->target;
1296 struct srp_device *dev = target->srp_host->srp_dev;
1297 struct ib_send_wr *bad_wr;
1298 struct ib_send_wr wr;
1299 struct srp_fr_desc *desc;
1302 desc = srp_fr_pool_get(ch->fr_pool);
1306 rkey = ib_inc_rkey(desc->mr->rkey);
1307 ib_update_fast_reg_key(desc->mr, rkey);
1309 memcpy(desc->frpl->page_list, state->pages,
1310 sizeof(state->pages[0]) * state->npages);
1312 memset(&wr, 0, sizeof(wr));
1313 wr.opcode = IB_WR_FAST_REG_MR;
1314 wr.wr_id = FAST_REG_WR_ID_MASK;
1315 wr.wr.fast_reg.iova_start = state->base_dma_addr;
1316 wr.wr.fast_reg.page_list = desc->frpl;
1317 wr.wr.fast_reg.page_list_len = state->npages;
1318 wr.wr.fast_reg.page_shift = ilog2(dev->mr_page_size);
1319 wr.wr.fast_reg.length = state->dma_len;
1320 wr.wr.fast_reg.access_flags = (IB_ACCESS_LOCAL_WRITE |
1321 IB_ACCESS_REMOTE_READ |
1322 IB_ACCESS_REMOTE_WRITE);
1323 wr.wr.fast_reg.rkey = desc->mr->lkey;
1325 *state->next_fr++ = desc;
1328 srp_map_desc(state, state->base_dma_addr, state->dma_len,
1331 return ib_post_send(ch->qp, &wr, &bad_wr);
1334 static int srp_finish_mapping(struct srp_map_state *state,
1335 struct srp_rdma_ch *ch)
1337 struct srp_target_port *target = ch->target;
1340 if (state->npages == 0)
1343 if (state->npages == 1 && !register_always)
1344 srp_map_desc(state, state->base_dma_addr, state->dma_len,
1347 ret = target->srp_host->srp_dev->use_fast_reg ?
1348 srp_map_finish_fr(state, ch) :
1349 srp_map_finish_fmr(state, ch);
1359 static void srp_map_update_start(struct srp_map_state *state,
1360 struct scatterlist *sg, int sg_index,
1361 dma_addr_t dma_addr)
1363 state->unmapped_sg = sg;
1364 state->unmapped_index = sg_index;
1365 state->unmapped_addr = dma_addr;
1368 static int srp_map_sg_entry(struct srp_map_state *state,
1369 struct srp_rdma_ch *ch,
1370 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);
1386 * Once we're in direct map mode for a request, we don't
1387 * go back to FMR or FR mode, so no need to update anything
1388 * other than the descriptor.
1390 srp_map_desc(state, dma_addr, dma_len, target->rkey);
1395 * Since not all RDMA HW drivers support non-zero page offsets for
1396 * FMR, if we start at an offset into a page, don't merge into the
1397 * current FMR mapping. Finish it out, and use the kernel's MR for
1400 if ((!dev->use_fast_reg && dma_addr & ~dev->mr_page_mask) ||
1401 dma_len > dev->mr_max_size) {
1402 ret = srp_finish_mapping(state, ch);
1406 srp_map_desc(state, dma_addr, dma_len, target->rkey);
1407 srp_map_update_start(state, NULL, 0, 0);
1412 * If this is the first sg that will be mapped via FMR or via FR, save
1413 * our position. We need to know the first unmapped entry, its index,
1414 * and the first unmapped address within that entry to be able to
1415 * restart mapping after an error.
1417 if (!state->unmapped_sg)
1418 srp_map_update_start(state, sg, sg_index, dma_addr);
1421 unsigned offset = dma_addr & ~dev->mr_page_mask;
1422 if (state->npages == dev->max_pages_per_mr || offset != 0) {
1423 ret = srp_finish_mapping(state, ch);
1427 srp_map_update_start(state, sg, sg_index, dma_addr);
1430 len = min_t(unsigned int, dma_len, dev->mr_page_size - offset);
1433 state->base_dma_addr = dma_addr;
1434 state->pages[state->npages++] = dma_addr & dev->mr_page_mask;
1435 state->dma_len += len;
1441 * If the last entry of the MR wasn't a full page, then we need to
1442 * close it out and start a new one -- we can only merge at page
1446 if (len != dev->mr_page_size) {
1447 ret = srp_finish_mapping(state, ch);
1449 srp_map_update_start(state, NULL, 0, 0);
1454 static int srp_map_sg(struct srp_map_state *state, struct srp_rdma_ch *ch,
1455 struct srp_request *req, struct scatterlist *scat,
1458 struct srp_target_port *target = ch->target;
1459 struct srp_device *dev = target->srp_host->srp_dev;
1460 struct ib_device *ibdev = dev->dev;
1461 struct scatterlist *sg;
1465 state->desc = req->indirect_desc;
1466 state->pages = req->map_page;
1467 if (dev->use_fast_reg) {
1468 state->next_fr = req->fr_list;
1469 use_mr = !!ch->fr_pool;
1471 state->next_fmr = req->fmr_list;
1472 use_mr = !!ch->fmr_pool;
1475 for_each_sg(scat, sg, count, i) {
1476 if (srp_map_sg_entry(state, ch, sg, i, use_mr)) {
1478 * Memory registration failed, so backtrack to the
1479 * first unmapped entry and continue on without using
1480 * memory registration.
1482 dma_addr_t dma_addr;
1483 unsigned int dma_len;
1486 sg = state->unmapped_sg;
1487 i = state->unmapped_index;
1489 dma_addr = ib_sg_dma_address(ibdev, sg);
1490 dma_len = ib_sg_dma_len(ibdev, sg);
1491 dma_len -= (state->unmapped_addr - dma_addr);
1492 dma_addr = state->unmapped_addr;
1494 srp_map_desc(state, dma_addr, dma_len, target->rkey);
1498 if (use_mr && srp_finish_mapping(state, ch))
1501 req->nmdesc = state->nmdesc;
1506 static int srp_map_data(struct scsi_cmnd *scmnd, struct srp_rdma_ch *ch,
1507 struct srp_request *req)
1509 struct srp_target_port *target = ch->target;
1510 struct scatterlist *scat;
1511 struct srp_cmd *cmd = req->cmd->buf;
1512 int len, nents, count;
1513 struct srp_device *dev;
1514 struct ib_device *ibdev;
1515 struct srp_map_state state;
1516 struct srp_indirect_buf *indirect_hdr;
1520 if (!scsi_sglist(scmnd) || scmnd->sc_data_direction == DMA_NONE)
1521 return sizeof (struct srp_cmd);
1523 if (scmnd->sc_data_direction != DMA_FROM_DEVICE &&
1524 scmnd->sc_data_direction != DMA_TO_DEVICE) {
1525 shost_printk(KERN_WARNING, target->scsi_host,
1526 PFX "Unhandled data direction %d\n",
1527 scmnd->sc_data_direction);
1531 nents = scsi_sg_count(scmnd);
1532 scat = scsi_sglist(scmnd);
1534 dev = target->srp_host->srp_dev;
1537 count = ib_dma_map_sg(ibdev, scat, nents, scmnd->sc_data_direction);
1538 if (unlikely(count == 0))
1541 fmt = SRP_DATA_DESC_DIRECT;
1542 len = sizeof (struct srp_cmd) + sizeof (struct srp_direct_buf);
1544 if (count == 1 && !register_always) {
1546 * The midlayer only generated a single gather/scatter
1547 * entry, or DMA mapping coalesced everything to a
1548 * single entry. So a direct descriptor along with
1549 * the DMA MR suffices.
1551 struct srp_direct_buf *buf = (void *) cmd->add_data;
1553 buf->va = cpu_to_be64(ib_sg_dma_address(ibdev, scat));
1554 buf->key = cpu_to_be32(target->rkey);
1555 buf->len = cpu_to_be32(ib_sg_dma_len(ibdev, scat));
1562 * We have more than one scatter/gather entry, so build our indirect
1563 * descriptor table, trying to merge as many entries as we can.
1565 indirect_hdr = (void *) cmd->add_data;
1567 ib_dma_sync_single_for_cpu(ibdev, req->indirect_dma_addr,
1568 target->indirect_size, DMA_TO_DEVICE);
1570 memset(&state, 0, sizeof(state));
1571 srp_map_sg(&state, ch, req, scat, count);
1573 /* We've mapped the request, now pull as much of the indirect
1574 * descriptor table as we can into the command buffer. If this
1575 * target is not using an external indirect table, we are
1576 * guaranteed to fit into the command, as the SCSI layer won't
1577 * give us more S/G entries than we allow.
1579 if (state.ndesc == 1) {
1581 * Memory registration collapsed the sg-list into one entry,
1582 * so use a direct descriptor.
1584 struct srp_direct_buf *buf = (void *) cmd->add_data;
1586 *buf = req->indirect_desc[0];
1590 if (unlikely(target->cmd_sg_cnt < state.ndesc &&
1591 !target->allow_ext_sg)) {
1592 shost_printk(KERN_ERR, target->scsi_host,
1593 "Could not fit S/G list into SRP_CMD\n");
1597 count = min(state.ndesc, target->cmd_sg_cnt);
1598 table_len = state.ndesc * sizeof (struct srp_direct_buf);
1600 fmt = SRP_DATA_DESC_INDIRECT;
1601 len = sizeof(struct srp_cmd) + sizeof (struct srp_indirect_buf);
1602 len += count * sizeof (struct srp_direct_buf);
1604 memcpy(indirect_hdr->desc_list, req->indirect_desc,
1605 count * sizeof (struct srp_direct_buf));
1607 indirect_hdr->table_desc.va = cpu_to_be64(req->indirect_dma_addr);
1608 indirect_hdr->table_desc.key = cpu_to_be32(target->rkey);
1609 indirect_hdr->table_desc.len = cpu_to_be32(table_len);
1610 indirect_hdr->len = cpu_to_be32(state.total_len);
1612 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1613 cmd->data_out_desc_cnt = count;
1615 cmd->data_in_desc_cnt = count;
1617 ib_dma_sync_single_for_device(ibdev, req->indirect_dma_addr, table_len,
1621 if (scmnd->sc_data_direction == DMA_TO_DEVICE)
1622 cmd->buf_fmt = fmt << 4;
1630 * Return an IU and possible credit to the free pool
1632 static void srp_put_tx_iu(struct srp_rdma_ch *ch, struct srp_iu *iu,
1633 enum srp_iu_type iu_type)
1635 unsigned long flags;
1637 spin_lock_irqsave(&ch->lock, flags);
1638 list_add(&iu->list, &ch->free_tx);
1639 if (iu_type != SRP_IU_RSP)
1641 spin_unlock_irqrestore(&ch->lock, flags);
1645 * Must be called with ch->lock held to protect req_lim and free_tx.
1646 * If IU is not sent, it must be returned using srp_put_tx_iu().
1649 * An upper limit for the number of allocated information units for each
1651 * - SRP_IU_CMD: SRP_CMD_SQ_SIZE, since the SCSI mid-layer never queues
1652 * more than Scsi_Host.can_queue requests.
1653 * - SRP_IU_TSK_MGMT: SRP_TSK_MGMT_SQ_SIZE.
1654 * - SRP_IU_RSP: 1, since a conforming SRP target never sends more than
1655 * one unanswered SRP request to an initiator.
1657 static struct srp_iu *__srp_get_tx_iu(struct srp_rdma_ch *ch,
1658 enum srp_iu_type iu_type)
1660 struct srp_target_port *target = ch->target;
1661 s32 rsv = (iu_type == SRP_IU_TSK_MGMT) ? 0 : SRP_TSK_MGMT_SQ_SIZE;
1664 srp_send_completion(ch->send_cq, ch);
1666 if (list_empty(&ch->free_tx))
1669 /* Initiator responses to target requests do not consume credits */
1670 if (iu_type != SRP_IU_RSP) {
1671 if (ch->req_lim <= rsv) {
1672 ++target->zero_req_lim;
1679 iu = list_first_entry(&ch->free_tx, struct srp_iu, list);
1680 list_del(&iu->list);
1684 static int srp_post_send(struct srp_rdma_ch *ch, struct srp_iu *iu, int len)
1686 struct srp_target_port *target = ch->target;
1688 struct ib_send_wr wr, *bad_wr;
1690 list.addr = iu->dma;
1692 list.lkey = target->lkey;
1695 wr.wr_id = (uintptr_t) iu;
1698 wr.opcode = IB_WR_SEND;
1699 wr.send_flags = IB_SEND_SIGNALED;
1701 return ib_post_send(ch->qp, &wr, &bad_wr);
1704 static int srp_post_recv(struct srp_rdma_ch *ch, struct srp_iu *iu)
1706 struct srp_target_port *target = ch->target;
1707 struct ib_recv_wr wr, *bad_wr;
1710 list.addr = iu->dma;
1711 list.length = iu->size;
1712 list.lkey = target->lkey;
1715 wr.wr_id = (uintptr_t) iu;
1719 return ib_post_recv(ch->qp, &wr, &bad_wr);
1722 static void srp_process_rsp(struct srp_rdma_ch *ch, struct srp_rsp *rsp)
1724 struct srp_target_port *target = ch->target;
1725 struct srp_request *req;
1726 struct scsi_cmnd *scmnd;
1727 unsigned long flags;
1729 if (unlikely(rsp->tag & SRP_TAG_TSK_MGMT)) {
1730 spin_lock_irqsave(&ch->lock, flags);
1731 ch->req_lim += be32_to_cpu(rsp->req_lim_delta);
1732 spin_unlock_irqrestore(&ch->lock, flags);
1734 ch->tsk_mgmt_status = -1;
1735 if (be32_to_cpu(rsp->resp_data_len) >= 4)
1736 ch->tsk_mgmt_status = rsp->data[3];
1737 complete(&ch->tsk_mgmt_done);
1739 scmnd = scsi_host_find_tag(target->scsi_host, rsp->tag);
1741 req = (void *)scmnd->host_scribble;
1742 scmnd = srp_claim_req(ch, req, NULL, scmnd);
1745 shost_printk(KERN_ERR, target->scsi_host,
1746 "Null scmnd for RSP w/tag %#016llx received on ch %td / QP %#x\n",
1747 rsp->tag, ch - target->ch, ch->qp->qp_num);
1749 spin_lock_irqsave(&ch->lock, flags);
1750 ch->req_lim += be32_to_cpu(rsp->req_lim_delta);
1751 spin_unlock_irqrestore(&ch->lock, flags);
1755 scmnd->result = rsp->status;
1757 if (rsp->flags & SRP_RSP_FLAG_SNSVALID) {
1758 memcpy(scmnd->sense_buffer, rsp->data +
1759 be32_to_cpu(rsp->resp_data_len),
1760 min_t(int, be32_to_cpu(rsp->sense_data_len),
1761 SCSI_SENSE_BUFFERSIZE));
1764 if (unlikely(rsp->flags & SRP_RSP_FLAG_DIUNDER))
1765 scsi_set_resid(scmnd, be32_to_cpu(rsp->data_in_res_cnt));
1766 else if (unlikely(rsp->flags & SRP_RSP_FLAG_DIOVER))
1767 scsi_set_resid(scmnd, -be32_to_cpu(rsp->data_in_res_cnt));
1768 else if (unlikely(rsp->flags & SRP_RSP_FLAG_DOUNDER))
1769 scsi_set_resid(scmnd, be32_to_cpu(rsp->data_out_res_cnt));
1770 else if (unlikely(rsp->flags & SRP_RSP_FLAG_DOOVER))
1771 scsi_set_resid(scmnd, -be32_to_cpu(rsp->data_out_res_cnt));
1773 srp_free_req(ch, req, scmnd,
1774 be32_to_cpu(rsp->req_lim_delta));
1776 scmnd->host_scribble = NULL;
1777 scmnd->scsi_done(scmnd);
1781 static int srp_response_common(struct srp_rdma_ch *ch, s32 req_delta,
1784 struct srp_target_port *target = ch->target;
1785 struct ib_device *dev = target->srp_host->srp_dev->dev;
1786 unsigned long flags;
1790 spin_lock_irqsave(&ch->lock, flags);
1791 ch->req_lim += req_delta;
1792 iu = __srp_get_tx_iu(ch, SRP_IU_RSP);
1793 spin_unlock_irqrestore(&ch->lock, flags);
1796 shost_printk(KERN_ERR, target->scsi_host, PFX
1797 "no IU available to send response\n");
1801 ib_dma_sync_single_for_cpu(dev, iu->dma, len, DMA_TO_DEVICE);
1802 memcpy(iu->buf, rsp, len);
1803 ib_dma_sync_single_for_device(dev, iu->dma, len, DMA_TO_DEVICE);
1805 err = srp_post_send(ch, iu, len);
1807 shost_printk(KERN_ERR, target->scsi_host, PFX
1808 "unable to post response: %d\n", err);
1809 srp_put_tx_iu(ch, iu, SRP_IU_RSP);
1815 static void srp_process_cred_req(struct srp_rdma_ch *ch,
1816 struct srp_cred_req *req)
1818 struct srp_cred_rsp rsp = {
1819 .opcode = SRP_CRED_RSP,
1822 s32 delta = be32_to_cpu(req->req_lim_delta);
1824 if (srp_response_common(ch, delta, &rsp, sizeof(rsp)))
1825 shost_printk(KERN_ERR, ch->target->scsi_host, PFX
1826 "problems processing SRP_CRED_REQ\n");
1829 static void srp_process_aer_req(struct srp_rdma_ch *ch,
1830 struct srp_aer_req *req)
1832 struct srp_target_port *target = ch->target;
1833 struct srp_aer_rsp rsp = {
1834 .opcode = SRP_AER_RSP,
1837 s32 delta = be32_to_cpu(req->req_lim_delta);
1839 shost_printk(KERN_ERR, target->scsi_host, PFX
1840 "ignoring AER for LUN %llu\n", be64_to_cpu(req->lun));
1842 if (srp_response_common(ch, delta, &rsp, sizeof(rsp)))
1843 shost_printk(KERN_ERR, target->scsi_host, PFX
1844 "problems processing SRP_AER_REQ\n");
1847 static void srp_handle_recv(struct srp_rdma_ch *ch, struct ib_wc *wc)
1849 struct srp_target_port *target = ch->target;
1850 struct ib_device *dev = target->srp_host->srp_dev->dev;
1851 struct srp_iu *iu = (struct srp_iu *) (uintptr_t) wc->wr_id;
1855 ib_dma_sync_single_for_cpu(dev, iu->dma, ch->max_ti_iu_len,
1858 opcode = *(u8 *) iu->buf;
1861 shost_printk(KERN_ERR, target->scsi_host,
1862 PFX "recv completion, opcode 0x%02x\n", opcode);
1863 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 8, 1,
1864 iu->buf, wc->byte_len, true);
1869 srp_process_rsp(ch, iu->buf);
1873 srp_process_cred_req(ch, iu->buf);
1877 srp_process_aer_req(ch, iu->buf);
1881 /* XXX Handle target logout */
1882 shost_printk(KERN_WARNING, target->scsi_host,
1883 PFX "Got target logout request\n");
1887 shost_printk(KERN_WARNING, target->scsi_host,
1888 PFX "Unhandled SRP opcode 0x%02x\n", opcode);
1892 ib_dma_sync_single_for_device(dev, iu->dma, ch->max_ti_iu_len,
1895 res = srp_post_recv(ch, iu);
1897 shost_printk(KERN_ERR, target->scsi_host,
1898 PFX "Recv failed with error code %d\n", res);
1902 * srp_tl_err_work() - handle a transport layer error
1903 * @work: Work structure embedded in an SRP target port.
1905 * Note: This function may get invoked before the rport has been created,
1906 * hence the target->rport test.
1908 static void srp_tl_err_work(struct work_struct *work)
1910 struct srp_target_port *target;
1912 target = container_of(work, struct srp_target_port, tl_err_work);
1914 srp_start_tl_fail_timers(target->rport);
1917 static void srp_handle_qp_err(u64 wr_id, enum ib_wc_status wc_status,
1918 bool send_err, struct srp_rdma_ch *ch)
1920 struct srp_target_port *target = ch->target;
1922 if (wr_id == SRP_LAST_WR_ID) {
1923 complete(&ch->done);
1927 if (ch->connected && !target->qp_in_error) {
1928 if (wr_id & LOCAL_INV_WR_ID_MASK) {
1929 shost_printk(KERN_ERR, target->scsi_host, PFX
1930 "LOCAL_INV failed with status %d\n",
1932 } else if (wr_id & FAST_REG_WR_ID_MASK) {
1933 shost_printk(KERN_ERR, target->scsi_host, PFX
1934 "FAST_REG_MR failed status %d\n",
1937 shost_printk(KERN_ERR, target->scsi_host,
1938 PFX "failed %s status %d for iu %p\n",
1939 send_err ? "send" : "receive",
1940 wc_status, (void *)(uintptr_t)wr_id);
1942 queue_work(system_long_wq, &target->tl_err_work);
1944 target->qp_in_error = true;
1947 static void srp_recv_completion(struct ib_cq *cq, void *ch_ptr)
1949 struct srp_rdma_ch *ch = ch_ptr;
1952 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1953 while (ib_poll_cq(cq, 1, &wc) > 0) {
1954 if (likely(wc.status == IB_WC_SUCCESS)) {
1955 srp_handle_recv(ch, &wc);
1957 srp_handle_qp_err(wc.wr_id, wc.status, false, ch);
1962 static void srp_send_completion(struct ib_cq *cq, void *ch_ptr)
1964 struct srp_rdma_ch *ch = ch_ptr;
1968 while (ib_poll_cq(cq, 1, &wc) > 0) {
1969 if (likely(wc.status == IB_WC_SUCCESS)) {
1970 iu = (struct srp_iu *) (uintptr_t) wc.wr_id;
1971 list_add(&iu->list, &ch->free_tx);
1973 srp_handle_qp_err(wc.wr_id, wc.status, true, ch);
1978 static int srp_queuecommand(struct Scsi_Host *shost, struct scsi_cmnd *scmnd)
1980 struct srp_target_port *target = host_to_target(shost);
1981 struct srp_rport *rport = target->rport;
1982 struct srp_rdma_ch *ch;
1983 struct srp_request *req;
1985 struct srp_cmd *cmd;
1986 struct ib_device *dev;
1987 unsigned long flags;
1991 const bool in_scsi_eh = !in_interrupt() && current == shost->ehandler;
1994 * The SCSI EH thread is the only context from which srp_queuecommand()
1995 * can get invoked for blocked devices (SDEV_BLOCK /
1996 * SDEV_CREATED_BLOCK). Avoid racing with srp_reconnect_rport() by
1997 * locking the rport mutex if invoked from inside the SCSI EH.
2000 mutex_lock(&rport->mutex);
2002 scmnd->result = srp_chkready(target->rport);
2003 if (unlikely(scmnd->result))
2006 WARN_ON_ONCE(scmnd->request->tag < 0);
2007 tag = blk_mq_unique_tag(scmnd->request);
2008 ch = &target->ch[blk_mq_unique_tag_to_hwq(tag)];
2009 idx = blk_mq_unique_tag_to_tag(tag);
2010 WARN_ONCE(idx >= target->req_ring_size, "%s: tag %#x: idx %d >= %d\n",
2011 dev_name(&shost->shost_gendev), tag, idx,
2012 target->req_ring_size);
2014 spin_lock_irqsave(&ch->lock, flags);
2015 iu = __srp_get_tx_iu(ch, SRP_IU_CMD);
2016 spin_unlock_irqrestore(&ch->lock, flags);
2021 req = &ch->req_ring[idx];
2022 dev = target->srp_host->srp_dev->dev;
2023 ib_dma_sync_single_for_cpu(dev, iu->dma, target->max_iu_len,
2026 scmnd->host_scribble = (void *) req;
2029 memset(cmd, 0, sizeof *cmd);
2031 cmd->opcode = SRP_CMD;
2032 cmd->lun = cpu_to_be64((u64) scmnd->device->lun << 48);
2034 memcpy(cmd->cdb, scmnd->cmnd, scmnd->cmd_len);
2039 len = srp_map_data(scmnd, ch, req);
2041 shost_printk(KERN_ERR, target->scsi_host,
2042 PFX "Failed to map data (%d)\n", len);
2044 * If we ran out of memory descriptors (-ENOMEM) because an
2045 * application is queuing many requests with more than
2046 * max_pages_per_mr sg-list elements, tell the SCSI mid-layer
2047 * to reduce queue depth temporarily.
2049 scmnd->result = len == -ENOMEM ?
2050 DID_OK << 16 | QUEUE_FULL << 1 : DID_ERROR << 16;
2054 ib_dma_sync_single_for_device(dev, iu->dma, target->max_iu_len,
2057 if (srp_post_send(ch, iu, len)) {
2058 shost_printk(KERN_ERR, target->scsi_host, PFX "Send failed\n");
2066 mutex_unlock(&rport->mutex);
2071 srp_unmap_data(scmnd, ch, req);
2074 srp_put_tx_iu(ch, iu, SRP_IU_CMD);
2077 * Avoid that the loops that iterate over the request ring can
2078 * encounter a dangling SCSI command pointer.
2083 if (scmnd->result) {
2084 scmnd->scsi_done(scmnd);
2087 ret = SCSI_MLQUEUE_HOST_BUSY;
2094 * Note: the resources allocated in this function are freed in
2097 static int srp_alloc_iu_bufs(struct srp_rdma_ch *ch)
2099 struct srp_target_port *target = ch->target;
2102 ch->rx_ring = kcalloc(target->queue_size, sizeof(*ch->rx_ring),
2106 ch->tx_ring = kcalloc(target->queue_size, sizeof(*ch->tx_ring),
2111 for (i = 0; i < target->queue_size; ++i) {
2112 ch->rx_ring[i] = srp_alloc_iu(target->srp_host,
2114 GFP_KERNEL, DMA_FROM_DEVICE);
2115 if (!ch->rx_ring[i])
2119 for (i = 0; i < target->queue_size; ++i) {
2120 ch->tx_ring[i] = srp_alloc_iu(target->srp_host,
2122 GFP_KERNEL, DMA_TO_DEVICE);
2123 if (!ch->tx_ring[i])
2126 list_add(&ch->tx_ring[i]->list, &ch->free_tx);
2132 for (i = 0; i < target->queue_size; ++i) {
2133 srp_free_iu(target->srp_host, ch->rx_ring[i]);
2134 srp_free_iu(target->srp_host, ch->tx_ring[i]);
2147 static uint32_t srp_compute_rq_tmo(struct ib_qp_attr *qp_attr, int attr_mask)
2149 uint64_t T_tr_ns, max_compl_time_ms;
2150 uint32_t rq_tmo_jiffies;
2153 * According to section 11.2.4.2 in the IBTA spec (Modify Queue Pair,
2154 * table 91), both the QP timeout and the retry count have to be set
2155 * for RC QP's during the RTR to RTS transition.
2157 WARN_ON_ONCE((attr_mask & (IB_QP_TIMEOUT | IB_QP_RETRY_CNT)) !=
2158 (IB_QP_TIMEOUT | IB_QP_RETRY_CNT));
2161 * Set target->rq_tmo_jiffies to one second more than the largest time
2162 * it can take before an error completion is generated. See also
2163 * C9-140..142 in the IBTA spec for more information about how to
2164 * convert the QP Local ACK Timeout value to nanoseconds.
2166 T_tr_ns = 4096 * (1ULL << qp_attr->timeout);
2167 max_compl_time_ms = qp_attr->retry_cnt * 4 * T_tr_ns;
2168 do_div(max_compl_time_ms, NSEC_PER_MSEC);
2169 rq_tmo_jiffies = msecs_to_jiffies(max_compl_time_ms + 1000);
2171 return rq_tmo_jiffies;
2174 static void srp_cm_rep_handler(struct ib_cm_id *cm_id,
2175 struct srp_login_rsp *lrsp,
2176 struct srp_rdma_ch *ch)
2178 struct srp_target_port *target = ch->target;
2179 struct ib_qp_attr *qp_attr = NULL;
2184 if (lrsp->opcode == SRP_LOGIN_RSP) {
2185 ch->max_ti_iu_len = be32_to_cpu(lrsp->max_ti_iu_len);
2186 ch->req_lim = be32_to_cpu(lrsp->req_lim_delta);
2189 * Reserve credits for task management so we don't
2190 * bounce requests back to the SCSI mid-layer.
2192 target->scsi_host->can_queue
2193 = min(ch->req_lim - SRP_TSK_MGMT_SQ_SIZE,
2194 target->scsi_host->can_queue);
2195 target->scsi_host->cmd_per_lun
2196 = min_t(int, target->scsi_host->can_queue,
2197 target->scsi_host->cmd_per_lun);
2199 shost_printk(KERN_WARNING, target->scsi_host,
2200 PFX "Unhandled RSP opcode %#x\n", lrsp->opcode);
2206 ret = srp_alloc_iu_bufs(ch);
2212 qp_attr = kmalloc(sizeof *qp_attr, GFP_KERNEL);
2216 qp_attr->qp_state = IB_QPS_RTR;
2217 ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
2221 ret = ib_modify_qp(ch->qp, qp_attr, attr_mask);
2225 for (i = 0; i < target->queue_size; i++) {
2226 struct srp_iu *iu = ch->rx_ring[i];
2228 ret = srp_post_recv(ch, iu);
2233 qp_attr->qp_state = IB_QPS_RTS;
2234 ret = ib_cm_init_qp_attr(cm_id, qp_attr, &attr_mask);
2238 target->rq_tmo_jiffies = srp_compute_rq_tmo(qp_attr, attr_mask);
2240 ret = ib_modify_qp(ch->qp, qp_attr, attr_mask);
2244 ret = ib_send_cm_rtu(cm_id, NULL, 0);
2253 static void srp_cm_rej_handler(struct ib_cm_id *cm_id,
2254 struct ib_cm_event *event,
2255 struct srp_rdma_ch *ch)
2257 struct srp_target_port *target = ch->target;
2258 struct Scsi_Host *shost = target->scsi_host;
2259 struct ib_class_port_info *cpi;
2262 switch (event->param.rej_rcvd.reason) {
2263 case IB_CM_REJ_PORT_CM_REDIRECT:
2264 cpi = event->param.rej_rcvd.ari;
2265 ch->path.dlid = cpi->redirect_lid;
2266 ch->path.pkey = cpi->redirect_pkey;
2267 cm_id->remote_cm_qpn = be32_to_cpu(cpi->redirect_qp) & 0x00ffffff;
2268 memcpy(ch->path.dgid.raw, cpi->redirect_gid, 16);
2270 ch->status = ch->path.dlid ?
2271 SRP_DLID_REDIRECT : SRP_PORT_REDIRECT;
2274 case IB_CM_REJ_PORT_REDIRECT:
2275 if (srp_target_is_topspin(target)) {
2277 * Topspin/Cisco SRP gateways incorrectly send
2278 * reject reason code 25 when they mean 24
2281 memcpy(ch->path.dgid.raw,
2282 event->param.rej_rcvd.ari, 16);
2284 shost_printk(KERN_DEBUG, shost,
2285 PFX "Topspin/Cisco redirect to target port GID %016llx%016llx\n",
2286 be64_to_cpu(ch->path.dgid.global.subnet_prefix),
2287 be64_to_cpu(ch->path.dgid.global.interface_id));
2289 ch->status = SRP_PORT_REDIRECT;
2291 shost_printk(KERN_WARNING, shost,
2292 " REJ reason: IB_CM_REJ_PORT_REDIRECT\n");
2293 ch->status = -ECONNRESET;
2297 case IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID:
2298 shost_printk(KERN_WARNING, shost,
2299 " REJ reason: IB_CM_REJ_DUPLICATE_LOCAL_COMM_ID\n");
2300 ch->status = -ECONNRESET;
2303 case IB_CM_REJ_CONSUMER_DEFINED:
2304 opcode = *(u8 *) event->private_data;
2305 if (opcode == SRP_LOGIN_REJ) {
2306 struct srp_login_rej *rej = event->private_data;
2307 u32 reason = be32_to_cpu(rej->reason);
2309 if (reason == SRP_LOGIN_REJ_REQ_IT_IU_LENGTH_TOO_LARGE)
2310 shost_printk(KERN_WARNING, shost,
2311 PFX "SRP_LOGIN_REJ: requested max_it_iu_len too large\n");
2313 shost_printk(KERN_WARNING, shost, PFX
2314 "SRP LOGIN from %pI6 to %pI6 REJECTED, reason 0x%08x\n",
2316 target->orig_dgid.raw, reason);
2318 shost_printk(KERN_WARNING, shost,
2319 " REJ reason: IB_CM_REJ_CONSUMER_DEFINED,"
2320 " opcode 0x%02x\n", opcode);
2321 ch->status = -ECONNRESET;
2324 case IB_CM_REJ_STALE_CONN:
2325 shost_printk(KERN_WARNING, shost, " REJ reason: stale connection\n");
2326 ch->status = SRP_STALE_CONN;
2330 shost_printk(KERN_WARNING, shost, " REJ reason 0x%x\n",
2331 event->param.rej_rcvd.reason);
2332 ch->status = -ECONNRESET;
2336 static int srp_cm_handler(struct ib_cm_id *cm_id, struct ib_cm_event *event)
2338 struct srp_rdma_ch *ch = cm_id->context;
2339 struct srp_target_port *target = ch->target;
2342 switch (event->event) {
2343 case IB_CM_REQ_ERROR:
2344 shost_printk(KERN_DEBUG, target->scsi_host,
2345 PFX "Sending CM REQ failed\n");
2347 ch->status = -ECONNRESET;
2350 case IB_CM_REP_RECEIVED:
2352 srp_cm_rep_handler(cm_id, event->private_data, ch);
2355 case IB_CM_REJ_RECEIVED:
2356 shost_printk(KERN_DEBUG, target->scsi_host, PFX "REJ received\n");
2359 srp_cm_rej_handler(cm_id, event, ch);
2362 case IB_CM_DREQ_RECEIVED:
2363 shost_printk(KERN_WARNING, target->scsi_host,
2364 PFX "DREQ received - connection closed\n");
2365 ch->connected = false;
2366 if (ib_send_cm_drep(cm_id, NULL, 0))
2367 shost_printk(KERN_ERR, target->scsi_host,
2368 PFX "Sending CM DREP failed\n");
2369 queue_work(system_long_wq, &target->tl_err_work);
2372 case IB_CM_TIMEWAIT_EXIT:
2373 shost_printk(KERN_ERR, target->scsi_host,
2374 PFX "connection closed\n");
2380 case IB_CM_MRA_RECEIVED:
2381 case IB_CM_DREQ_ERROR:
2382 case IB_CM_DREP_RECEIVED:
2386 shost_printk(KERN_WARNING, target->scsi_host,
2387 PFX "Unhandled CM event %d\n", event->event);
2392 complete(&ch->done);
2398 * srp_change_queue_depth - setting device queue depth
2399 * @sdev: scsi device struct
2400 * @qdepth: requested queue depth
2402 * Returns queue depth.
2405 srp_change_queue_depth(struct scsi_device *sdev, int qdepth)
2407 if (!sdev->tagged_supported)
2409 return scsi_change_queue_depth(sdev, qdepth);
2412 static int srp_send_tsk_mgmt(struct srp_rdma_ch *ch, u64 req_tag,
2413 unsigned int lun, u8 func)
2415 struct srp_target_port *target = ch->target;
2416 struct srp_rport *rport = target->rport;
2417 struct ib_device *dev = target->srp_host->srp_dev->dev;
2419 struct srp_tsk_mgmt *tsk_mgmt;
2421 if (!ch->connected || target->qp_in_error)
2424 init_completion(&ch->tsk_mgmt_done);
2427 * Lock the rport mutex to avoid that srp_create_ch_ib() is
2428 * invoked while a task management function is being sent.
2430 mutex_lock(&rport->mutex);
2431 spin_lock_irq(&ch->lock);
2432 iu = __srp_get_tx_iu(ch, SRP_IU_TSK_MGMT);
2433 spin_unlock_irq(&ch->lock);
2436 mutex_unlock(&rport->mutex);
2441 ib_dma_sync_single_for_cpu(dev, iu->dma, sizeof *tsk_mgmt,
2444 memset(tsk_mgmt, 0, sizeof *tsk_mgmt);
2446 tsk_mgmt->opcode = SRP_TSK_MGMT;
2447 tsk_mgmt->lun = cpu_to_be64((u64) lun << 48);
2448 tsk_mgmt->tag = req_tag | SRP_TAG_TSK_MGMT;
2449 tsk_mgmt->tsk_mgmt_func = func;
2450 tsk_mgmt->task_tag = req_tag;
2452 ib_dma_sync_single_for_device(dev, iu->dma, sizeof *tsk_mgmt,
2454 if (srp_post_send(ch, iu, sizeof(*tsk_mgmt))) {
2455 srp_put_tx_iu(ch, iu, SRP_IU_TSK_MGMT);
2456 mutex_unlock(&rport->mutex);
2460 mutex_unlock(&rport->mutex);
2462 if (!wait_for_completion_timeout(&ch->tsk_mgmt_done,
2463 msecs_to_jiffies(SRP_ABORT_TIMEOUT_MS)))
2469 static int srp_abort(struct scsi_cmnd *scmnd)
2471 struct srp_target_port *target = host_to_target(scmnd->device->host);
2472 struct srp_request *req = (struct srp_request *) scmnd->host_scribble;
2475 struct srp_rdma_ch *ch;
2478 shost_printk(KERN_ERR, target->scsi_host, "SRP abort called\n");
2482 tag = blk_mq_unique_tag(scmnd->request);
2483 ch_idx = blk_mq_unique_tag_to_hwq(tag);
2484 if (WARN_ON_ONCE(ch_idx >= target->ch_count))
2486 ch = &target->ch[ch_idx];
2487 if (!srp_claim_req(ch, req, NULL, scmnd))
2489 shost_printk(KERN_ERR, target->scsi_host,
2490 "Sending SRP abort for tag %#x\n", tag);
2491 if (srp_send_tsk_mgmt(ch, tag, scmnd->device->lun,
2492 SRP_TSK_ABORT_TASK) == 0)
2494 else if (target->rport->state == SRP_RPORT_LOST)
2498 srp_free_req(ch, req, scmnd, 0);
2499 scmnd->result = DID_ABORT << 16;
2500 scmnd->scsi_done(scmnd);
2505 static int srp_reset_device(struct scsi_cmnd *scmnd)
2507 struct srp_target_port *target = host_to_target(scmnd->device->host);
2508 struct srp_rdma_ch *ch;
2511 shost_printk(KERN_ERR, target->scsi_host, "SRP reset_device called\n");
2513 ch = &target->ch[0];
2514 if (srp_send_tsk_mgmt(ch, SRP_TAG_NO_REQ, scmnd->device->lun,
2517 if (ch->tsk_mgmt_status)
2520 for (i = 0; i < target->ch_count; i++) {
2521 ch = &target->ch[i];
2522 for (i = 0; i < target->req_ring_size; ++i) {
2523 struct srp_request *req = &ch->req_ring[i];
2525 srp_finish_req(ch, req, scmnd->device, DID_RESET << 16);
2532 static int srp_reset_host(struct scsi_cmnd *scmnd)
2534 struct srp_target_port *target = host_to_target(scmnd->device->host);
2536 shost_printk(KERN_ERR, target->scsi_host, PFX "SRP reset_host called\n");
2538 return srp_reconnect_rport(target->rport) == 0 ? SUCCESS : FAILED;
2541 static int srp_slave_configure(struct scsi_device *sdev)
2543 struct Scsi_Host *shost = sdev->host;
2544 struct srp_target_port *target = host_to_target(shost);
2545 struct request_queue *q = sdev->request_queue;
2546 unsigned long timeout;
2548 if (sdev->type == TYPE_DISK) {
2549 timeout = max_t(unsigned, 30 * HZ, target->rq_tmo_jiffies);
2550 blk_queue_rq_timeout(q, timeout);
2556 static ssize_t show_id_ext(struct device *dev, struct device_attribute *attr,
2559 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2561 return sprintf(buf, "0x%016llx\n",
2562 (unsigned long long) be64_to_cpu(target->id_ext));
2565 static ssize_t show_ioc_guid(struct device *dev, struct device_attribute *attr,
2568 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2570 return sprintf(buf, "0x%016llx\n",
2571 (unsigned long long) be64_to_cpu(target->ioc_guid));
2574 static ssize_t show_service_id(struct device *dev,
2575 struct device_attribute *attr, char *buf)
2577 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2579 return sprintf(buf, "0x%016llx\n",
2580 (unsigned long long) be64_to_cpu(target->service_id));
2583 static ssize_t show_pkey(struct device *dev, struct device_attribute *attr,
2586 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2588 return sprintf(buf, "0x%04x\n", be16_to_cpu(target->pkey));
2591 static ssize_t show_sgid(struct device *dev, struct device_attribute *attr,
2594 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2596 return sprintf(buf, "%pI6\n", target->sgid.raw);
2599 static ssize_t show_dgid(struct device *dev, struct device_attribute *attr,
2602 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2603 struct srp_rdma_ch *ch = &target->ch[0];
2605 return sprintf(buf, "%pI6\n", ch->path.dgid.raw);
2608 static ssize_t show_orig_dgid(struct device *dev,
2609 struct device_attribute *attr, char *buf)
2611 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2613 return sprintf(buf, "%pI6\n", target->orig_dgid.raw);
2616 static ssize_t show_req_lim(struct device *dev,
2617 struct device_attribute *attr, char *buf)
2619 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2620 struct srp_rdma_ch *ch;
2621 int i, req_lim = INT_MAX;
2623 for (i = 0; i < target->ch_count; i++) {
2624 ch = &target->ch[i];
2625 req_lim = min(req_lim, ch->req_lim);
2627 return sprintf(buf, "%d\n", req_lim);
2630 static ssize_t show_zero_req_lim(struct device *dev,
2631 struct device_attribute *attr, char *buf)
2633 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2635 return sprintf(buf, "%d\n", target->zero_req_lim);
2638 static ssize_t show_local_ib_port(struct device *dev,
2639 struct device_attribute *attr, char *buf)
2641 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2643 return sprintf(buf, "%d\n", target->srp_host->port);
2646 static ssize_t show_local_ib_device(struct device *dev,
2647 struct device_attribute *attr, char *buf)
2649 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2651 return sprintf(buf, "%s\n", target->srp_host->srp_dev->dev->name);
2654 static ssize_t show_ch_count(struct device *dev, struct device_attribute *attr,
2657 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2659 return sprintf(buf, "%d\n", target->ch_count);
2662 static ssize_t show_comp_vector(struct device *dev,
2663 struct device_attribute *attr, char *buf)
2665 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2667 return sprintf(buf, "%d\n", target->comp_vector);
2670 static ssize_t show_tl_retry_count(struct device *dev,
2671 struct device_attribute *attr, char *buf)
2673 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2675 return sprintf(buf, "%d\n", target->tl_retry_count);
2678 static ssize_t show_cmd_sg_entries(struct device *dev,
2679 struct device_attribute *attr, char *buf)
2681 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2683 return sprintf(buf, "%u\n", target->cmd_sg_cnt);
2686 static ssize_t show_allow_ext_sg(struct device *dev,
2687 struct device_attribute *attr, char *buf)
2689 struct srp_target_port *target = host_to_target(class_to_shost(dev));
2691 return sprintf(buf, "%s\n", target->allow_ext_sg ? "true" : "false");
2694 static DEVICE_ATTR(id_ext, S_IRUGO, show_id_ext, NULL);
2695 static DEVICE_ATTR(ioc_guid, S_IRUGO, show_ioc_guid, NULL);
2696 static DEVICE_ATTR(service_id, S_IRUGO, show_service_id, NULL);
2697 static DEVICE_ATTR(pkey, S_IRUGO, show_pkey, NULL);
2698 static DEVICE_ATTR(sgid, S_IRUGO, show_sgid, NULL);
2699 static DEVICE_ATTR(dgid, S_IRUGO, show_dgid, NULL);
2700 static DEVICE_ATTR(orig_dgid, S_IRUGO, show_orig_dgid, NULL);
2701 static DEVICE_ATTR(req_lim, S_IRUGO, show_req_lim, NULL);
2702 static DEVICE_ATTR(zero_req_lim, S_IRUGO, show_zero_req_lim, NULL);
2703 static DEVICE_ATTR(local_ib_port, S_IRUGO, show_local_ib_port, NULL);
2704 static DEVICE_ATTR(local_ib_device, S_IRUGO, show_local_ib_device, NULL);
2705 static DEVICE_ATTR(ch_count, S_IRUGO, show_ch_count, NULL);
2706 static DEVICE_ATTR(comp_vector, S_IRUGO, show_comp_vector, NULL);
2707 static DEVICE_ATTR(tl_retry_count, S_IRUGO, show_tl_retry_count, NULL);
2708 static DEVICE_ATTR(cmd_sg_entries, S_IRUGO, show_cmd_sg_entries, NULL);
2709 static DEVICE_ATTR(allow_ext_sg, S_IRUGO, show_allow_ext_sg, NULL);
2711 static struct device_attribute *srp_host_attrs[] = {
2714 &dev_attr_service_id,
2718 &dev_attr_orig_dgid,
2720 &dev_attr_zero_req_lim,
2721 &dev_attr_local_ib_port,
2722 &dev_attr_local_ib_device,
2724 &dev_attr_comp_vector,
2725 &dev_attr_tl_retry_count,
2726 &dev_attr_cmd_sg_entries,
2727 &dev_attr_allow_ext_sg,
2731 static struct scsi_host_template srp_template = {
2732 .module = THIS_MODULE,
2733 .name = "InfiniBand SRP initiator",
2734 .proc_name = DRV_NAME,
2735 .slave_configure = srp_slave_configure,
2736 .info = srp_target_info,
2737 .queuecommand = srp_queuecommand,
2738 .change_queue_depth = srp_change_queue_depth,
2739 .eh_abort_handler = srp_abort,
2740 .eh_device_reset_handler = srp_reset_device,
2741 .eh_host_reset_handler = srp_reset_host,
2742 .skip_settle_delay = true,
2743 .sg_tablesize = SRP_DEF_SG_TABLESIZE,
2744 .can_queue = SRP_DEFAULT_CMD_SQ_SIZE,
2746 .cmd_per_lun = SRP_DEFAULT_CMD_SQ_SIZE,
2747 .use_clustering = ENABLE_CLUSTERING,
2748 .shost_attrs = srp_host_attrs,
2750 .track_queue_depth = 1,
2753 static int srp_sdev_count(struct Scsi_Host *host)
2755 struct scsi_device *sdev;
2758 shost_for_each_device(sdev, host)
2764 static int srp_add_target(struct srp_host *host, struct srp_target_port *target)
2766 struct srp_rport_identifiers ids;
2767 struct srp_rport *rport;
2769 target->state = SRP_TARGET_SCANNING;
2770 sprintf(target->target_name, "SRP.T10:%016llX",
2771 (unsigned long long) be64_to_cpu(target->id_ext));
2773 if (scsi_add_host(target->scsi_host, host->srp_dev->dev->dma_device))
2776 memcpy(ids.port_id, &target->id_ext, 8);
2777 memcpy(ids.port_id + 8, &target->ioc_guid, 8);
2778 ids.roles = SRP_RPORT_ROLE_TARGET;
2779 rport = srp_rport_add(target->scsi_host, &ids);
2780 if (IS_ERR(rport)) {
2781 scsi_remove_host(target->scsi_host);
2782 return PTR_ERR(rport);
2785 rport->lld_data = target;
2786 target->rport = rport;
2788 spin_lock(&host->target_lock);
2789 list_add_tail(&target->list, &host->target_list);
2790 spin_unlock(&host->target_lock);
2792 scsi_scan_target(&target->scsi_host->shost_gendev,
2793 0, target->scsi_id, SCAN_WILD_CARD, 0);
2795 if (srp_connected_ch(target) < target->ch_count ||
2796 target->qp_in_error) {
2797 shost_printk(KERN_INFO, target->scsi_host,
2798 PFX "SCSI scan failed - removing SCSI host\n");
2799 srp_queue_remove_work(target);
2803 pr_debug(PFX "%s: SCSI scan succeeded - detected %d LUNs\n",
2804 dev_name(&target->scsi_host->shost_gendev),
2805 srp_sdev_count(target->scsi_host));
2807 spin_lock_irq(&target->lock);
2808 if (target->state == SRP_TARGET_SCANNING)
2809 target->state = SRP_TARGET_LIVE;
2810 spin_unlock_irq(&target->lock);
2816 static void srp_release_dev(struct device *dev)
2818 struct srp_host *host =
2819 container_of(dev, struct srp_host, dev);
2821 complete(&host->released);
2824 static struct class srp_class = {
2825 .name = "infiniband_srp",
2826 .dev_release = srp_release_dev
2830 * srp_conn_unique() - check whether the connection to a target is unique
2832 * @target: SRP target port.
2834 static bool srp_conn_unique(struct srp_host *host,
2835 struct srp_target_port *target)
2837 struct srp_target_port *t;
2840 if (target->state == SRP_TARGET_REMOVED)
2845 spin_lock(&host->target_lock);
2846 list_for_each_entry(t, &host->target_list, list) {
2848 target->id_ext == t->id_ext &&
2849 target->ioc_guid == t->ioc_guid &&
2850 target->initiator_ext == t->initiator_ext) {
2855 spin_unlock(&host->target_lock);
2862 * Target ports are added by writing
2864 * id_ext=<SRP ID ext>,ioc_guid=<SRP IOC GUID>,dgid=<dest GID>,
2865 * pkey=<P_Key>,service_id=<service ID>
2867 * to the add_target sysfs attribute.
2871 SRP_OPT_ID_EXT = 1 << 0,
2872 SRP_OPT_IOC_GUID = 1 << 1,
2873 SRP_OPT_DGID = 1 << 2,
2874 SRP_OPT_PKEY = 1 << 3,
2875 SRP_OPT_SERVICE_ID = 1 << 4,
2876 SRP_OPT_MAX_SECT = 1 << 5,
2877 SRP_OPT_MAX_CMD_PER_LUN = 1 << 6,
2878 SRP_OPT_IO_CLASS = 1 << 7,
2879 SRP_OPT_INITIATOR_EXT = 1 << 8,
2880 SRP_OPT_CMD_SG_ENTRIES = 1 << 9,
2881 SRP_OPT_ALLOW_EXT_SG = 1 << 10,
2882 SRP_OPT_SG_TABLESIZE = 1 << 11,
2883 SRP_OPT_COMP_VECTOR = 1 << 12,
2884 SRP_OPT_TL_RETRY_COUNT = 1 << 13,
2885 SRP_OPT_QUEUE_SIZE = 1 << 14,
2886 SRP_OPT_ALL = (SRP_OPT_ID_EXT |
2890 SRP_OPT_SERVICE_ID),
2893 static const match_table_t srp_opt_tokens = {
2894 { SRP_OPT_ID_EXT, "id_ext=%s" },
2895 { SRP_OPT_IOC_GUID, "ioc_guid=%s" },
2896 { SRP_OPT_DGID, "dgid=%s" },
2897 { SRP_OPT_PKEY, "pkey=%x" },
2898 { SRP_OPT_SERVICE_ID, "service_id=%s" },
2899 { SRP_OPT_MAX_SECT, "max_sect=%d" },
2900 { SRP_OPT_MAX_CMD_PER_LUN, "max_cmd_per_lun=%d" },
2901 { SRP_OPT_IO_CLASS, "io_class=%x" },
2902 { SRP_OPT_INITIATOR_EXT, "initiator_ext=%s" },
2903 { SRP_OPT_CMD_SG_ENTRIES, "cmd_sg_entries=%u" },
2904 { SRP_OPT_ALLOW_EXT_SG, "allow_ext_sg=%u" },
2905 { SRP_OPT_SG_TABLESIZE, "sg_tablesize=%u" },
2906 { SRP_OPT_COMP_VECTOR, "comp_vector=%u" },
2907 { SRP_OPT_TL_RETRY_COUNT, "tl_retry_count=%u" },
2908 { SRP_OPT_QUEUE_SIZE, "queue_size=%d" },
2909 { SRP_OPT_ERR, NULL }
2912 static int srp_parse_options(const char *buf, struct srp_target_port *target)
2914 char *options, *sep_opt;
2917 substring_t args[MAX_OPT_ARGS];
2923 options = kstrdup(buf, GFP_KERNEL);
2928 while ((p = strsep(&sep_opt, ",\n")) != NULL) {
2932 token = match_token(p, srp_opt_tokens, args);
2936 case SRP_OPT_ID_EXT:
2937 p = match_strdup(args);
2942 target->id_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
2946 case SRP_OPT_IOC_GUID:
2947 p = match_strdup(args);
2952 target->ioc_guid = cpu_to_be64(simple_strtoull(p, NULL, 16));
2957 p = match_strdup(args);
2962 if (strlen(p) != 32) {
2963 pr_warn("bad dest GID parameter '%s'\n", p);
2968 for (i = 0; i < 16; ++i) {
2969 strlcpy(dgid, p + i * 2, sizeof(dgid));
2970 if (sscanf(dgid, "%hhx",
2971 &target->orig_dgid.raw[i]) < 1) {
2981 if (match_hex(args, &token)) {
2982 pr_warn("bad P_Key parameter '%s'\n", p);
2985 target->pkey = cpu_to_be16(token);
2988 case SRP_OPT_SERVICE_ID:
2989 p = match_strdup(args);
2994 target->service_id = cpu_to_be64(simple_strtoull(p, NULL, 16));
2998 case SRP_OPT_MAX_SECT:
2999 if (match_int(args, &token)) {
3000 pr_warn("bad max sect parameter '%s'\n", p);
3003 target->scsi_host->max_sectors = token;
3006 case SRP_OPT_QUEUE_SIZE:
3007 if (match_int(args, &token) || token < 1) {
3008 pr_warn("bad queue_size parameter '%s'\n", p);
3011 target->scsi_host->can_queue = token;
3012 target->queue_size = token + SRP_RSP_SQ_SIZE +
3013 SRP_TSK_MGMT_SQ_SIZE;
3014 if (!(opt_mask & SRP_OPT_MAX_CMD_PER_LUN))
3015 target->scsi_host->cmd_per_lun = token;
3018 case SRP_OPT_MAX_CMD_PER_LUN:
3019 if (match_int(args, &token) || token < 1) {
3020 pr_warn("bad max cmd_per_lun parameter '%s'\n",
3024 target->scsi_host->cmd_per_lun = token;
3027 case SRP_OPT_IO_CLASS:
3028 if (match_hex(args, &token)) {
3029 pr_warn("bad IO class parameter '%s'\n", p);
3032 if (token != SRP_REV10_IB_IO_CLASS &&
3033 token != SRP_REV16A_IB_IO_CLASS) {
3034 pr_warn("unknown IO class parameter value %x specified (use %x or %x).\n",
3035 token, SRP_REV10_IB_IO_CLASS,
3036 SRP_REV16A_IB_IO_CLASS);
3039 target->io_class = token;
3042 case SRP_OPT_INITIATOR_EXT:
3043 p = match_strdup(args);
3048 target->initiator_ext = cpu_to_be64(simple_strtoull(p, NULL, 16));
3052 case SRP_OPT_CMD_SG_ENTRIES:
3053 if (match_int(args, &token) || token < 1 || token > 255) {
3054 pr_warn("bad max cmd_sg_entries parameter '%s'\n",
3058 target->cmd_sg_cnt = token;
3061 case SRP_OPT_ALLOW_EXT_SG:
3062 if (match_int(args, &token)) {
3063 pr_warn("bad allow_ext_sg parameter '%s'\n", p);
3066 target->allow_ext_sg = !!token;
3069 case SRP_OPT_SG_TABLESIZE:
3070 if (match_int(args, &token) || token < 1 ||
3071 token > SCSI_MAX_SG_CHAIN_SEGMENTS) {
3072 pr_warn("bad max sg_tablesize parameter '%s'\n",
3076 target->sg_tablesize = token;
3079 case SRP_OPT_COMP_VECTOR:
3080 if (match_int(args, &token) || token < 0) {
3081 pr_warn("bad comp_vector parameter '%s'\n", p);
3084 target->comp_vector = token;
3087 case SRP_OPT_TL_RETRY_COUNT:
3088 if (match_int(args, &token) || token < 2 || token > 7) {
3089 pr_warn("bad tl_retry_count parameter '%s' (must be a number between 2 and 7)\n",
3093 target->tl_retry_count = token;
3097 pr_warn("unknown parameter or missing value '%s' in target creation request\n",
3103 if ((opt_mask & SRP_OPT_ALL) == SRP_OPT_ALL)
3106 for (i = 0; i < ARRAY_SIZE(srp_opt_tokens); ++i)
3107 if ((srp_opt_tokens[i].token & SRP_OPT_ALL) &&
3108 !(srp_opt_tokens[i].token & opt_mask))
3109 pr_warn("target creation request is missing parameter '%s'\n",
3110 srp_opt_tokens[i].pattern);
3112 if (target->scsi_host->cmd_per_lun > target->scsi_host->can_queue
3113 && (opt_mask & SRP_OPT_MAX_CMD_PER_LUN))
3114 pr_warn("cmd_per_lun = %d > queue_size = %d\n",
3115 target->scsi_host->cmd_per_lun,
3116 target->scsi_host->can_queue);
3123 static ssize_t srp_create_target(struct device *dev,
3124 struct device_attribute *attr,
3125 const char *buf, size_t count)
3127 struct srp_host *host =
3128 container_of(dev, struct srp_host, dev);
3129 struct Scsi_Host *target_host;
3130 struct srp_target_port *target;
3131 struct srp_rdma_ch *ch;
3132 struct srp_device *srp_dev = host->srp_dev;
3133 struct ib_device *ibdev = srp_dev->dev;
3134 int ret, node_idx, node, cpu, i;
3135 bool multich = false;
3137 target_host = scsi_host_alloc(&srp_template,
3138 sizeof (struct srp_target_port));
3142 target_host->transportt = ib_srp_transport_template;
3143 target_host->max_channel = 0;
3144 target_host->max_id = 1;
3145 target_host->max_lun = SRP_MAX_LUN;
3146 target_host->max_cmd_len = sizeof ((struct srp_cmd *) (void *) 0L)->cdb;
3148 target = host_to_target(target_host);
3150 target->io_class = SRP_REV16A_IB_IO_CLASS;
3151 target->scsi_host = target_host;
3152 target->srp_host = host;
3153 target->lkey = host->srp_dev->mr->lkey;
3154 target->rkey = host->srp_dev->mr->rkey;
3155 target->cmd_sg_cnt = cmd_sg_entries;
3156 target->sg_tablesize = indirect_sg_entries ? : cmd_sg_entries;
3157 target->allow_ext_sg = allow_ext_sg;
3158 target->tl_retry_count = 7;
3159 target->queue_size = SRP_DEFAULT_QUEUE_SIZE;
3162 * Avoid that the SCSI host can be removed by srp_remove_target()
3163 * before this function returns.
3165 scsi_host_get(target->scsi_host);
3167 mutex_lock(&host->add_target_mutex);
3169 ret = srp_parse_options(buf, target);
3173 ret = scsi_init_shared_tag_map(target_host, target_host->can_queue);
3177 target->req_ring_size = target->queue_size - SRP_TSK_MGMT_SQ_SIZE;
3179 if (!srp_conn_unique(target->srp_host, target)) {
3180 shost_printk(KERN_INFO, target->scsi_host,
3181 PFX "Already connected to target port with id_ext=%016llx;ioc_guid=%016llx;initiator_ext=%016llx\n",
3182 be64_to_cpu(target->id_ext),
3183 be64_to_cpu(target->ioc_guid),
3184 be64_to_cpu(target->initiator_ext));
3189 if (!srp_dev->has_fmr && !srp_dev->has_fr && !target->allow_ext_sg &&
3190 target->cmd_sg_cnt < target->sg_tablesize) {
3191 pr_warn("No MR pool and no external indirect descriptors, limiting sg_tablesize to cmd_sg_cnt\n");
3192 target->sg_tablesize = target->cmd_sg_cnt;
3195 target_host->sg_tablesize = target->sg_tablesize;
3196 target->indirect_size = target->sg_tablesize *
3197 sizeof (struct srp_direct_buf);
3198 target->max_iu_len = sizeof (struct srp_cmd) +
3199 sizeof (struct srp_indirect_buf) +
3200 target->cmd_sg_cnt * sizeof (struct srp_direct_buf);
3202 INIT_WORK(&target->tl_err_work, srp_tl_err_work);
3203 INIT_WORK(&target->remove_work, srp_remove_work);
3204 spin_lock_init(&target->lock);
3205 ret = ib_query_gid(ibdev, host->port, 0, &target->sgid);
3210 target->ch_count = max_t(unsigned, num_online_nodes(),
3212 min(4 * num_online_nodes(),
3213 ibdev->num_comp_vectors),
3214 num_online_cpus()));
3215 target->ch = kcalloc(target->ch_count, sizeof(*target->ch),
3221 for_each_online_node(node) {
3222 const int ch_start = (node_idx * target->ch_count /
3223 num_online_nodes());
3224 const int ch_end = ((node_idx + 1) * target->ch_count /
3225 num_online_nodes());
3226 const int cv_start = (node_idx * ibdev->num_comp_vectors /
3227 num_online_nodes() + target->comp_vector)
3228 % ibdev->num_comp_vectors;
3229 const int cv_end = ((node_idx + 1) * ibdev->num_comp_vectors /
3230 num_online_nodes() + target->comp_vector)
3231 % ibdev->num_comp_vectors;
3234 for_each_online_cpu(cpu) {
3235 if (cpu_to_node(cpu) != node)
3237 if (ch_start + cpu_idx >= ch_end)
3239 ch = &target->ch[ch_start + cpu_idx];
3240 ch->target = target;
3241 ch->comp_vector = cv_start == cv_end ? cv_start :
3242 cv_start + cpu_idx % (cv_end - cv_start);
3243 spin_lock_init(&ch->lock);
3244 INIT_LIST_HEAD(&ch->free_tx);
3245 ret = srp_new_cm_id(ch);
3247 goto err_disconnect;
3249 ret = srp_create_ch_ib(ch);
3251 goto err_disconnect;
3253 ret = srp_alloc_req_data(ch);
3255 goto err_disconnect;
3257 ret = srp_connect_ch(ch, multich);
3259 shost_printk(KERN_ERR, target->scsi_host,
3260 PFX "Connection %d/%d failed\n",
3263 if (node_idx == 0 && cpu_idx == 0) {
3264 goto err_disconnect;
3266 srp_free_ch_ib(target, ch);
3267 srp_free_req_data(target, ch);
3268 target->ch_count = ch - target->ch;
3279 target->scsi_host->nr_hw_queues = target->ch_count;
3281 ret = srp_add_target(host, target);
3283 goto err_disconnect;
3285 if (target->state != SRP_TARGET_REMOVED) {
3286 shost_printk(KERN_DEBUG, target->scsi_host, PFX
3287 "new target: id_ext %016llx ioc_guid %016llx pkey %04x service_id %016llx sgid %pI6 dgid %pI6\n",
3288 be64_to_cpu(target->id_ext),
3289 be64_to_cpu(target->ioc_guid),
3290 be16_to_cpu(target->pkey),
3291 be64_to_cpu(target->service_id),
3292 target->sgid.raw, target->orig_dgid.raw);
3298 mutex_unlock(&host->add_target_mutex);
3300 scsi_host_put(target->scsi_host);
3305 srp_disconnect_target(target);
3307 for (i = 0; i < target->ch_count; i++) {
3308 ch = &target->ch[i];
3309 srp_free_ch_ib(target, ch);
3310 srp_free_req_data(target, ch);
3317 static DEVICE_ATTR(add_target, S_IWUSR, NULL, srp_create_target);
3319 static ssize_t show_ibdev(struct device *dev, struct device_attribute *attr,
3322 struct srp_host *host = container_of(dev, struct srp_host, dev);
3324 return sprintf(buf, "%s\n", host->srp_dev->dev->name);
3327 static DEVICE_ATTR(ibdev, S_IRUGO, show_ibdev, NULL);
3329 static ssize_t show_port(struct device *dev, struct device_attribute *attr,
3332 struct srp_host *host = container_of(dev, struct srp_host, dev);
3334 return sprintf(buf, "%d\n", host->port);
3337 static DEVICE_ATTR(port, S_IRUGO, show_port, NULL);
3339 static struct srp_host *srp_add_port(struct srp_device *device, u8 port)
3341 struct srp_host *host;
3343 host = kzalloc(sizeof *host, GFP_KERNEL);
3347 INIT_LIST_HEAD(&host->target_list);
3348 spin_lock_init(&host->target_lock);
3349 init_completion(&host->released);
3350 mutex_init(&host->add_target_mutex);
3351 host->srp_dev = device;
3354 host->dev.class = &srp_class;
3355 host->dev.parent = device->dev->dma_device;
3356 dev_set_name(&host->dev, "srp-%s-%d", device->dev->name, port);
3358 if (device_register(&host->dev))
3360 if (device_create_file(&host->dev, &dev_attr_add_target))
3362 if (device_create_file(&host->dev, &dev_attr_ibdev))
3364 if (device_create_file(&host->dev, &dev_attr_port))
3370 device_unregister(&host->dev);
3378 static void srp_add_one(struct ib_device *device)
3380 struct srp_device *srp_dev;
3381 struct ib_device_attr *dev_attr;
3382 struct srp_host *host;
3383 int mr_page_shift, s, e, p;
3384 u64 max_pages_per_mr;
3386 dev_attr = kmalloc(sizeof *dev_attr, GFP_KERNEL);
3390 if (ib_query_device(device, dev_attr)) {
3391 pr_warn("Query device failed for %s\n", device->name);
3395 srp_dev = kmalloc(sizeof *srp_dev, GFP_KERNEL);
3399 srp_dev->has_fmr = (device->alloc_fmr && device->dealloc_fmr &&
3400 device->map_phys_fmr && device->unmap_fmr);
3401 srp_dev->has_fr = (dev_attr->device_cap_flags &
3402 IB_DEVICE_MEM_MGT_EXTENSIONS);
3403 if (!srp_dev->has_fmr && !srp_dev->has_fr)
3404 dev_warn(&device->dev, "neither FMR nor FR is supported\n");
3406 srp_dev->use_fast_reg = (srp_dev->has_fr &&
3407 (!srp_dev->has_fmr || prefer_fr));
3410 * Use the smallest page size supported by the HCA, down to a
3411 * minimum of 4096 bytes. We're unlikely to build large sglists
3412 * out of smaller entries.
3414 mr_page_shift = max(12, ffs(dev_attr->page_size_cap) - 1);
3415 srp_dev->mr_page_size = 1 << mr_page_shift;
3416 srp_dev->mr_page_mask = ~((u64) srp_dev->mr_page_size - 1);
3417 max_pages_per_mr = dev_attr->max_mr_size;
3418 do_div(max_pages_per_mr, srp_dev->mr_page_size);
3419 srp_dev->max_pages_per_mr = min_t(u64, SRP_MAX_PAGES_PER_MR,
3421 if (srp_dev->use_fast_reg) {
3422 srp_dev->max_pages_per_mr =
3423 min_t(u32, srp_dev->max_pages_per_mr,
3424 dev_attr->max_fast_reg_page_list_len);
3426 srp_dev->mr_max_size = srp_dev->mr_page_size *
3427 srp_dev->max_pages_per_mr;
3428 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",
3429 device->name, mr_page_shift, dev_attr->max_mr_size,
3430 dev_attr->max_fast_reg_page_list_len,
3431 srp_dev->max_pages_per_mr, srp_dev->mr_max_size);
3433 INIT_LIST_HEAD(&srp_dev->dev_list);
3435 srp_dev->dev = device;
3436 srp_dev->pd = ib_alloc_pd(device);
3437 if (IS_ERR(srp_dev->pd))
3440 srp_dev->mr = ib_get_dma_mr(srp_dev->pd,
3441 IB_ACCESS_LOCAL_WRITE |
3442 IB_ACCESS_REMOTE_READ |
3443 IB_ACCESS_REMOTE_WRITE);
3444 if (IS_ERR(srp_dev->mr))
3447 if (device->node_type == RDMA_NODE_IB_SWITCH) {
3452 e = device->phys_port_cnt;
3455 for (p = s; p <= e; ++p) {
3456 host = srp_add_port(srp_dev, p);
3458 list_add_tail(&host->list, &srp_dev->dev_list);
3461 ib_set_client_data(device, &srp_client, srp_dev);
3466 ib_dealloc_pd(srp_dev->pd);
3475 static void srp_remove_one(struct ib_device *device)
3477 struct srp_device *srp_dev;
3478 struct srp_host *host, *tmp_host;
3479 struct srp_target_port *target;
3481 srp_dev = ib_get_client_data(device, &srp_client);
3485 list_for_each_entry_safe(host, tmp_host, &srp_dev->dev_list, list) {
3486 device_unregister(&host->dev);
3488 * Wait for the sysfs entry to go away, so that no new
3489 * target ports can be created.
3491 wait_for_completion(&host->released);
3494 * Remove all target ports.
3496 spin_lock(&host->target_lock);
3497 list_for_each_entry(target, &host->target_list, list)
3498 srp_queue_remove_work(target);
3499 spin_unlock(&host->target_lock);
3502 * Wait for tl_err and target port removal tasks.
3504 flush_workqueue(system_long_wq);
3505 flush_workqueue(srp_remove_wq);
3510 ib_dereg_mr(srp_dev->mr);
3511 ib_dealloc_pd(srp_dev->pd);
3516 static struct srp_function_template ib_srp_transport_functions = {
3517 .has_rport_state = true,
3518 .reset_timer_if_blocked = true,
3519 .reconnect_delay = &srp_reconnect_delay,
3520 .fast_io_fail_tmo = &srp_fast_io_fail_tmo,
3521 .dev_loss_tmo = &srp_dev_loss_tmo,
3522 .reconnect = srp_rport_reconnect,
3523 .rport_delete = srp_rport_delete,
3524 .terminate_rport_io = srp_terminate_io,
3527 static int __init srp_init_module(void)
3531 BUILD_BUG_ON(FIELD_SIZEOF(struct ib_wc, wr_id) < sizeof(void *));
3533 if (srp_sg_tablesize) {
3534 pr_warn("srp_sg_tablesize is deprecated, please use cmd_sg_entries\n");
3535 if (!cmd_sg_entries)
3536 cmd_sg_entries = srp_sg_tablesize;
3539 if (!cmd_sg_entries)
3540 cmd_sg_entries = SRP_DEF_SG_TABLESIZE;
3542 if (cmd_sg_entries > 255) {
3543 pr_warn("Clamping cmd_sg_entries to 255\n");
3544 cmd_sg_entries = 255;
3547 if (!indirect_sg_entries)
3548 indirect_sg_entries = cmd_sg_entries;
3549 else if (indirect_sg_entries < cmd_sg_entries) {
3550 pr_warn("Bumping up indirect_sg_entries to match cmd_sg_entries (%u)\n",
3552 indirect_sg_entries = cmd_sg_entries;
3555 srp_remove_wq = create_workqueue("srp_remove");
3556 if (!srp_remove_wq) {
3562 ib_srp_transport_template =
3563 srp_attach_transport(&ib_srp_transport_functions);
3564 if (!ib_srp_transport_template)
3567 ret = class_register(&srp_class);
3569 pr_err("couldn't register class infiniband_srp\n");
3573 ib_sa_register_client(&srp_sa_client);
3575 ret = ib_register_client(&srp_client);
3577 pr_err("couldn't register IB client\n");
3585 ib_sa_unregister_client(&srp_sa_client);
3586 class_unregister(&srp_class);
3589 srp_release_transport(ib_srp_transport_template);
3592 destroy_workqueue(srp_remove_wq);
3596 static void __exit srp_cleanup_module(void)
3598 ib_unregister_client(&srp_client);
3599 ib_sa_unregister_client(&srp_sa_client);
3600 class_unregister(&srp_class);
3601 srp_release_transport(ib_srp_transport_template);
3602 destroy_workqueue(srp_remove_wq);
3605 module_init(srp_init_module);
3606 module_exit(srp_cleanup_module);