nvme-loop: add a NVMe loopback host driver
authorChristoph Hellwig <hch@lst.de>
Tue, 21 Jun 2016 16:04:21 +0000 (18:04 +0200)
committerJens Axboe <axboe@fb.com>
Tue, 5 Jul 2016 17:30:36 +0000 (11:30 -0600)
This patch implements adds nvme-loop which allows to access local devices
exported as NVMe over Fabrics namespaces. This module can be useful for
easy evaluation, testing and also feature experimentation.

To createa nvme-loop device you need to configure the NVMe target to
export a loop port (see the nvmetcli documentaton for that) and then
connect to it using

nvme connect-all -t loop

which requires the very latest nvme-cli version with Fabrics support.

Signed-off-by: Jay Freyensee <james.p.freyensee@intel.com>
Signed-off-by: Ming Lin <ming.l@ssi.samsung.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
drivers/nvme/target/Kconfig
drivers/nvme/target/Makefile
drivers/nvme/target/loop.c [new file with mode: 0644]

index acf0c070e50d973151a4509f3c413fe3fd5c1976..b77ce4350695ecfa560d6d569448168db672a353 100644 (file)
@@ -13,4 +13,14 @@ config NVME_TARGET
          To configure the NVMe target you probably want to use the nvmetcli
          tool from http://git.infradead.org/users/hch/nvmetcli.git.
 
+config NVME_TARGET_LOOP
+       tristate "NVMe loopback device support"
+       depends on BLK_DEV_NVME
+       select NVME_TARGET
+       select NVME_FABRICS
+       select SG_POOL
+       help
+         This enables the NVMe loopback device support, which can be useful
+         to test NVMe host and target side features.
+
          If unsure, say N.
index b4600b6f5724ed59b3e08e4156e4c15d62530835..e49ba60756d47e68a40f2946505934aee6eff6b8 100644 (file)
@@ -1,5 +1,7 @@
 
 obj-$(CONFIG_NVME_TARGET)              += nvmet.o
+obj-$(CONFIG_NVME_TARGET_LOOP)         += nvme-loop.o
 
 nvmet-y                += core.o configfs.o admin-cmd.o io-cmd.o fabrics-cmd.o \
                        discovery.o
+nvme-loop-y    += loop.o
diff --git a/drivers/nvme/target/loop.c b/drivers/nvme/target/loop.c
new file mode 100644 (file)
index 0000000..94e7829
--- /dev/null
@@ -0,0 +1,754 @@
+/*
+ * NVMe over Fabrics loopback device.
+ * Copyright (c) 2015-2016 HGST, a Western Digital Company.
+ *
+ * This program is free software; you can redistribute it and/or modify it
+ * under the terms and conditions of the GNU General Public License,
+ * version 2, as published by the Free Software Foundation.
+ *
+ * This program is distributed in the hope it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
+ * more details.
+ */
+#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
+#include <linux/scatterlist.h>
+#include <linux/delay.h>
+#include <linux/blk-mq.h>
+#include <linux/nvme.h>
+#include <linux/module.h>
+#include <linux/parser.h>
+#include <linux/t10-pi.h>
+#include "nvmet.h"
+#include "../host/nvme.h"
+#include "../host/fabrics.h"
+
+#define NVME_LOOP_AQ_DEPTH             256
+
+#define NVME_LOOP_MAX_SEGMENTS         256
+
+/*
+ * We handle AEN commands ourselves and don't even let the
+ * block layer know about them.
+ */
+#define NVME_LOOP_NR_AEN_COMMANDS      1
+#define NVME_LOOP_AQ_BLKMQ_DEPTH       \
+       (NVME_LOOP_AQ_DEPTH - NVME_LOOP_NR_AEN_COMMANDS)
+
+struct nvme_loop_iod {
+       struct nvme_command     cmd;
+       struct nvme_completion  rsp;
+       struct nvmet_req        req;
+       struct nvme_loop_queue  *queue;
+       struct work_struct      work;
+       struct sg_table         sg_table;
+       struct scatterlist      first_sgl[];
+};
+
+struct nvme_loop_ctrl {
+       spinlock_t              lock;
+       struct nvme_loop_queue  *queues;
+       u32                     queue_count;
+
+       struct blk_mq_tag_set   admin_tag_set;
+
+       struct list_head        list;
+       u64                     cap;
+       struct blk_mq_tag_set   tag_set;
+       struct nvme_loop_iod    async_event_iod;
+       struct nvme_ctrl        ctrl;
+
+       struct nvmet_ctrl       *target_ctrl;
+       struct work_struct      delete_work;
+       struct work_struct      reset_work;
+};
+
+static inline struct nvme_loop_ctrl *to_loop_ctrl(struct nvme_ctrl *ctrl)
+{
+       return container_of(ctrl, struct nvme_loop_ctrl, ctrl);
+}
+
+struct nvme_loop_queue {
+       struct nvmet_cq         nvme_cq;
+       struct nvmet_sq         nvme_sq;
+       struct nvme_loop_ctrl   *ctrl;
+};
+
+static struct nvmet_port *nvmet_loop_port;
+
+static LIST_HEAD(nvme_loop_ctrl_list);
+static DEFINE_MUTEX(nvme_loop_ctrl_mutex);
+
+static void nvme_loop_queue_response(struct nvmet_req *nvme_req);
+static void nvme_loop_delete_ctrl(struct nvmet_ctrl *ctrl);
+
+static struct nvmet_fabrics_ops nvme_loop_ops;
+
+static inline int nvme_loop_queue_idx(struct nvme_loop_queue *queue)
+{
+       return queue - queue->ctrl->queues;
+}
+
+static void nvme_loop_complete_rq(struct request *req)
+{
+       struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
+       int error = 0;
+
+       nvme_cleanup_cmd(req);
+       sg_free_table_chained(&iod->sg_table, true);
+
+       if (unlikely(req->errors)) {
+               if (nvme_req_needs_retry(req, req->errors)) {
+                       nvme_requeue_req(req);
+                       return;
+               }
+
+               if (req->cmd_type == REQ_TYPE_DRV_PRIV)
+                       error = req->errors;
+               else
+                       error = nvme_error_status(req->errors);
+       }
+
+       blk_mq_end_request(req, error);
+}
+
+static void nvme_loop_queue_response(struct nvmet_req *nvme_req)
+{
+       struct nvme_loop_iod *iod =
+               container_of(nvme_req, struct nvme_loop_iod, req);
+       struct nvme_completion *cqe = &iod->rsp;
+
+       /*
+        * AEN requests are special as they don't time out and can
+        * survive any kind of queue freeze and often don't respond to
+        * aborts.  We don't even bother to allocate a struct request
+        * for them but rather special case them here.
+        */
+       if (unlikely(nvme_loop_queue_idx(iod->queue) == 0 &&
+                       cqe->command_id >= NVME_LOOP_AQ_BLKMQ_DEPTH)) {
+               nvme_complete_async_event(&iod->queue->ctrl->ctrl, cqe);
+       } else {
+               struct request *req = blk_mq_rq_from_pdu(iod);
+
+               if (req->cmd_type == REQ_TYPE_DRV_PRIV && req->special)
+                       memcpy(req->special, cqe, sizeof(*cqe));
+               blk_mq_complete_request(req, le16_to_cpu(cqe->status) >> 1);
+       }
+}
+
+static void nvme_loop_execute_work(struct work_struct *work)
+{
+       struct nvme_loop_iod *iod =
+               container_of(work, struct nvme_loop_iod, work);
+
+       iod->req.execute(&iod->req);
+}
+
+static enum blk_eh_timer_return
+nvme_loop_timeout(struct request *rq, bool reserved)
+{
+       struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(rq);
+
+       /* queue error recovery */
+       schedule_work(&iod->queue->ctrl->reset_work);
+
+       /* fail with DNR on admin cmd timeout */
+       rq->errors = NVME_SC_ABORT_REQ | NVME_SC_DNR;
+
+       return BLK_EH_HANDLED;
+}
+
+static int nvme_loop_queue_rq(struct blk_mq_hw_ctx *hctx,
+               const struct blk_mq_queue_data *bd)
+{
+       struct nvme_ns *ns = hctx->queue->queuedata;
+       struct nvme_loop_queue *queue = hctx->driver_data;
+       struct request *req = bd->rq;
+       struct nvme_loop_iod *iod = blk_mq_rq_to_pdu(req);
+       int ret;
+
+       ret = nvme_setup_cmd(ns, req, &iod->cmd);
+       if (ret)
+               return ret;
+
+       iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
+       iod->req.port = nvmet_loop_port;
+       if (!nvmet_req_init(&iod->req, &queue->nvme_cq,
+                       &queue->nvme_sq, &nvme_loop_ops)) {
+               nvme_cleanup_cmd(req);
+               blk_mq_start_request(req);
+               nvme_loop_queue_response(&iod->req);
+               return 0;
+       }
+
+       if (blk_rq_bytes(req)) {
+               iod->sg_table.sgl = iod->first_sgl;
+               ret = sg_alloc_table_chained(&iod->sg_table,
+                       req->nr_phys_segments, iod->sg_table.sgl);
+               if (ret)
+                       return BLK_MQ_RQ_QUEUE_BUSY;
+
+               iod->req.sg = iod->sg_table.sgl;
+               iod->req.sg_cnt = blk_rq_map_sg(req->q, req, iod->sg_table.sgl);
+               BUG_ON(iod->req.sg_cnt > req->nr_phys_segments);
+       }
+
+       iod->cmd.common.command_id = req->tag;
+       blk_mq_start_request(req);
+
+       schedule_work(&iod->work);
+       return 0;
+}
+
+static void nvme_loop_submit_async_event(struct nvme_ctrl *arg, int aer_idx)
+{
+       struct nvme_loop_ctrl *ctrl = to_loop_ctrl(arg);
+       struct nvme_loop_queue *queue = &ctrl->queues[0];
+       struct nvme_loop_iod *iod = &ctrl->async_event_iod;
+
+       memset(&iod->cmd, 0, sizeof(iod->cmd));
+       iod->cmd.common.opcode = nvme_admin_async_event;
+       iod->cmd.common.command_id = NVME_LOOP_AQ_BLKMQ_DEPTH;
+       iod->cmd.common.flags |= NVME_CMD_SGL_METABUF;
+
+       if (!nvmet_req_init(&iod->req, &queue->nvme_cq, &queue->nvme_sq,
+                       &nvme_loop_ops)) {
+               dev_err(ctrl->ctrl.device, "failed async event work\n");
+               return;
+       }
+
+       schedule_work(&iod->work);
+}
+
+static int nvme_loop_init_iod(struct nvme_loop_ctrl *ctrl,
+               struct nvme_loop_iod *iod, unsigned int queue_idx)
+{
+       BUG_ON(queue_idx >= ctrl->queue_count);
+
+       iod->req.cmd = &iod->cmd;
+       iod->req.rsp = &iod->rsp;
+       iod->queue = &ctrl->queues[queue_idx];
+       INIT_WORK(&iod->work, nvme_loop_execute_work);
+       return 0;
+}
+
+static int nvme_loop_init_request(void *data, struct request *req,
+                               unsigned int hctx_idx, unsigned int rq_idx,
+                               unsigned int numa_node)
+{
+       return nvme_loop_init_iod(data, blk_mq_rq_to_pdu(req), hctx_idx + 1);
+}
+
+static int nvme_loop_init_admin_request(void *data, struct request *req,
+                               unsigned int hctx_idx, unsigned int rq_idx,
+                               unsigned int numa_node)
+{
+       return nvme_loop_init_iod(data, blk_mq_rq_to_pdu(req), 0);
+}
+
+static int nvme_loop_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
+               unsigned int hctx_idx)
+{
+       struct nvme_loop_ctrl *ctrl = data;
+       struct nvme_loop_queue *queue = &ctrl->queues[hctx_idx + 1];
+
+       BUG_ON(hctx_idx >= ctrl->queue_count);
+
+       hctx->driver_data = queue;
+       return 0;
+}
+
+static int nvme_loop_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
+               unsigned int hctx_idx)
+{
+       struct nvme_loop_ctrl *ctrl = data;
+       struct nvme_loop_queue *queue = &ctrl->queues[0];
+
+       BUG_ON(hctx_idx != 0);
+
+       hctx->driver_data = queue;
+       return 0;
+}
+
+static struct blk_mq_ops nvme_loop_mq_ops = {
+       .queue_rq       = nvme_loop_queue_rq,
+       .complete       = nvme_loop_complete_rq,
+       .map_queue      = blk_mq_map_queue,
+       .init_request   = nvme_loop_init_request,
+       .init_hctx      = nvme_loop_init_hctx,
+       .timeout        = nvme_loop_timeout,
+};
+
+static struct blk_mq_ops nvme_loop_admin_mq_ops = {
+       .queue_rq       = nvme_loop_queue_rq,
+       .complete       = nvme_loop_complete_rq,
+       .map_queue      = blk_mq_map_queue,
+       .init_request   = nvme_loop_init_admin_request,
+       .init_hctx      = nvme_loop_init_admin_hctx,
+       .timeout        = nvme_loop_timeout,
+};
+
+static void nvme_loop_destroy_admin_queue(struct nvme_loop_ctrl *ctrl)
+{
+       blk_cleanup_queue(ctrl->ctrl.admin_q);
+       blk_mq_free_tag_set(&ctrl->admin_tag_set);
+       nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
+}
+
+static void nvme_loop_free_ctrl(struct nvme_ctrl *nctrl)
+{
+       struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
+
+       if (list_empty(&ctrl->list))
+               goto free_ctrl;
+
+       mutex_lock(&nvme_loop_ctrl_mutex);
+       list_del(&ctrl->list);
+       mutex_unlock(&nvme_loop_ctrl_mutex);
+
+       if (nctrl->tagset) {
+               blk_cleanup_queue(ctrl->ctrl.connect_q);
+               blk_mq_free_tag_set(&ctrl->tag_set);
+       }
+       kfree(ctrl->queues);
+       nvmf_free_options(nctrl->opts);
+free_ctrl:
+       kfree(ctrl);
+}
+
+static int nvme_loop_configure_admin_queue(struct nvme_loop_ctrl *ctrl)
+{
+       int error;
+
+       memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
+       ctrl->admin_tag_set.ops = &nvme_loop_admin_mq_ops;
+       ctrl->admin_tag_set.queue_depth = NVME_LOOP_AQ_BLKMQ_DEPTH;
+       ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
+       ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
+       ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
+               SG_CHUNK_SIZE * sizeof(struct scatterlist);
+       ctrl->admin_tag_set.driver_data = ctrl;
+       ctrl->admin_tag_set.nr_hw_queues = 1;
+       ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
+
+       ctrl->queues[0].ctrl = ctrl;
+       error = nvmet_sq_init(&ctrl->queues[0].nvme_sq);
+       if (error)
+               return error;
+       ctrl->queue_count = 1;
+
+       error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
+       if (error)
+               goto out_free_sq;
+
+       ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
+       if (IS_ERR(ctrl->ctrl.admin_q)) {
+               error = PTR_ERR(ctrl->ctrl.admin_q);
+               goto out_free_tagset;
+       }
+
+       error = nvmf_connect_admin_queue(&ctrl->ctrl);
+       if (error)
+               goto out_cleanup_queue;
+
+       error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->cap);
+       if (error) {
+               dev_err(ctrl->ctrl.device,
+                       "prop_get NVME_REG_CAP failed\n");
+               goto out_cleanup_queue;
+       }
+
+       ctrl->ctrl.sqsize =
+               min_t(int, NVME_CAP_MQES(ctrl->cap) + 1, ctrl->ctrl.sqsize);
+
+       error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
+       if (error)
+               goto out_cleanup_queue;
+
+       ctrl->ctrl.max_hw_sectors =
+               (NVME_LOOP_MAX_SEGMENTS - 1) << (PAGE_SHIFT - 9);
+
+       error = nvme_init_identify(&ctrl->ctrl);
+       if (error)
+               goto out_cleanup_queue;
+
+       nvme_start_keep_alive(&ctrl->ctrl);
+
+       return 0;
+
+out_cleanup_queue:
+       blk_cleanup_queue(ctrl->ctrl.admin_q);
+out_free_tagset:
+       blk_mq_free_tag_set(&ctrl->admin_tag_set);
+out_free_sq:
+       nvmet_sq_destroy(&ctrl->queues[0].nvme_sq);
+       return error;
+}
+
+static void nvme_loop_shutdown_ctrl(struct nvme_loop_ctrl *ctrl)
+{
+       int i;
+
+       nvme_stop_keep_alive(&ctrl->ctrl);
+
+       if (ctrl->queue_count > 1) {
+               nvme_stop_queues(&ctrl->ctrl);
+               blk_mq_tagset_busy_iter(&ctrl->tag_set,
+                                       nvme_cancel_request, &ctrl->ctrl);
+
+               for (i = 1; i < ctrl->queue_count; i++)
+                       nvmet_sq_destroy(&ctrl->queues[i].nvme_sq);
+       }
+
+       if (ctrl->ctrl.state == NVME_CTRL_LIVE)
+               nvme_shutdown_ctrl(&ctrl->ctrl);
+
+       blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
+       blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
+                               nvme_cancel_request, &ctrl->ctrl);
+       nvme_loop_destroy_admin_queue(ctrl);
+}
+
+static void nvme_loop_del_ctrl_work(struct work_struct *work)
+{
+       struct nvme_loop_ctrl *ctrl = container_of(work,
+                               struct nvme_loop_ctrl, delete_work);
+
+       nvme_remove_namespaces(&ctrl->ctrl);
+       nvme_loop_shutdown_ctrl(ctrl);
+       nvme_uninit_ctrl(&ctrl->ctrl);
+       nvme_put_ctrl(&ctrl->ctrl);
+}
+
+static int __nvme_loop_del_ctrl(struct nvme_loop_ctrl *ctrl)
+{
+       if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING))
+               return -EBUSY;
+
+       if (!schedule_work(&ctrl->delete_work))
+               return -EBUSY;
+
+       return 0;
+}
+
+static int nvme_loop_del_ctrl(struct nvme_ctrl *nctrl)
+{
+       struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
+       int ret;
+
+       ret = __nvme_loop_del_ctrl(ctrl);
+       if (ret)
+               return ret;
+
+       flush_work(&ctrl->delete_work);
+
+       return 0;
+}
+
+static void nvme_loop_delete_ctrl(struct nvmet_ctrl *nctrl)
+{
+       struct nvme_loop_ctrl *ctrl;
+
+       mutex_lock(&nvme_loop_ctrl_mutex);
+       list_for_each_entry(ctrl, &nvme_loop_ctrl_list, list) {
+               if (ctrl->ctrl.cntlid == nctrl->cntlid)
+                       __nvme_loop_del_ctrl(ctrl);
+       }
+       mutex_unlock(&nvme_loop_ctrl_mutex);
+}
+
+static void nvme_loop_reset_ctrl_work(struct work_struct *work)
+{
+       struct nvme_loop_ctrl *ctrl = container_of(work,
+                                       struct nvme_loop_ctrl, reset_work);
+       bool changed;
+       int i, ret;
+
+       nvme_loop_shutdown_ctrl(ctrl);
+
+       ret = nvme_loop_configure_admin_queue(ctrl);
+       if (ret)
+               goto out_disable;
+
+       for (i = 1; i <= ctrl->ctrl.opts->nr_io_queues; i++) {
+               ctrl->queues[i].ctrl = ctrl;
+               ret = nvmet_sq_init(&ctrl->queues[i].nvme_sq);
+               if (ret)
+                       goto out_free_queues;
+
+               ctrl->queue_count++;
+       }
+
+       for (i = 1; i <= ctrl->ctrl.opts->nr_io_queues; i++) {
+               ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
+               if (ret)
+                       goto out_free_queues;
+       }
+
+       changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
+       WARN_ON_ONCE(!changed);
+
+       nvme_queue_scan(&ctrl->ctrl);
+       nvme_queue_async_events(&ctrl->ctrl);
+
+       nvme_start_queues(&ctrl->ctrl);
+
+       return;
+
+out_free_queues:
+       for (i = 1; i < ctrl->queue_count; i++)
+               nvmet_sq_destroy(&ctrl->queues[i].nvme_sq);
+       nvme_loop_destroy_admin_queue(ctrl);
+out_disable:
+       dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
+       nvme_remove_namespaces(&ctrl->ctrl);
+       nvme_uninit_ctrl(&ctrl->ctrl);
+       nvme_put_ctrl(&ctrl->ctrl);
+}
+
+static int nvme_loop_reset_ctrl(struct nvme_ctrl *nctrl)
+{
+       struct nvme_loop_ctrl *ctrl = to_loop_ctrl(nctrl);
+
+       if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
+               return -EBUSY;
+
+       if (!schedule_work(&ctrl->reset_work))
+               return -EBUSY;
+
+       flush_work(&ctrl->reset_work);
+
+       return 0;
+}
+
+static const struct nvme_ctrl_ops nvme_loop_ctrl_ops = {
+       .name                   = "loop",
+       .module                 = THIS_MODULE,
+       .is_fabrics             = true,
+       .reg_read32             = nvmf_reg_read32,
+       .reg_read64             = nvmf_reg_read64,
+       .reg_write32            = nvmf_reg_write32,
+       .reset_ctrl             = nvme_loop_reset_ctrl,
+       .free_ctrl              = nvme_loop_free_ctrl,
+       .submit_async_event     = nvme_loop_submit_async_event,
+       .delete_ctrl            = nvme_loop_del_ctrl,
+       .get_subsysnqn          = nvmf_get_subsysnqn,
+};
+
+static int nvme_loop_create_io_queues(struct nvme_loop_ctrl *ctrl)
+{
+       struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
+       int ret, i;
+
+       ret = nvme_set_queue_count(&ctrl->ctrl, &opts->nr_io_queues);
+       if (ret || !opts->nr_io_queues)
+               return ret;
+
+       dev_info(ctrl->ctrl.device, "creating %d I/O queues.\n",
+               opts->nr_io_queues);
+
+       for (i = 1; i <= opts->nr_io_queues; i++) {
+               ctrl->queues[i].ctrl = ctrl;
+               ret = nvmet_sq_init(&ctrl->queues[i].nvme_sq);
+               if (ret)
+                       goto out_destroy_queues;
+
+               ctrl->queue_count++;
+       }
+
+       memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
+       ctrl->tag_set.ops = &nvme_loop_mq_ops;
+       ctrl->tag_set.queue_depth = ctrl->ctrl.sqsize;
+       ctrl->tag_set.reserved_tags = 1; /* fabric connect */
+       ctrl->tag_set.numa_node = NUMA_NO_NODE;
+       ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
+       ctrl->tag_set.cmd_size = sizeof(struct nvme_loop_iod) +
+               SG_CHUNK_SIZE * sizeof(struct scatterlist);
+       ctrl->tag_set.driver_data = ctrl;
+       ctrl->tag_set.nr_hw_queues = ctrl->queue_count - 1;
+       ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
+       ctrl->ctrl.tagset = &ctrl->tag_set;
+
+       ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
+       if (ret)
+               goto out_destroy_queues;
+
+       ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
+       if (IS_ERR(ctrl->ctrl.connect_q)) {
+               ret = PTR_ERR(ctrl->ctrl.connect_q);
+               goto out_free_tagset;
+       }
+
+       for (i = 1; i <= opts->nr_io_queues; i++) {
+               ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
+               if (ret)
+                       goto out_cleanup_connect_q;
+       }
+
+       return 0;
+
+out_cleanup_connect_q:
+       blk_cleanup_queue(ctrl->ctrl.connect_q);
+out_free_tagset:
+       blk_mq_free_tag_set(&ctrl->tag_set);
+out_destroy_queues:
+       for (i = 1; i < ctrl->queue_count; i++)
+               nvmet_sq_destroy(&ctrl->queues[i].nvme_sq);
+       return ret;
+}
+
+static struct nvme_ctrl *nvme_loop_create_ctrl(struct device *dev,
+               struct nvmf_ctrl_options *opts)
+{
+       struct nvme_loop_ctrl *ctrl;
+       bool changed;
+       int ret;
+
+       ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
+       if (!ctrl)
+               return ERR_PTR(-ENOMEM);
+       ctrl->ctrl.opts = opts;
+       INIT_LIST_HEAD(&ctrl->list);
+
+       INIT_WORK(&ctrl->delete_work, nvme_loop_del_ctrl_work);
+       INIT_WORK(&ctrl->reset_work, nvme_loop_reset_ctrl_work);
+
+       ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_loop_ctrl_ops,
+                               0 /* no quirks, we're perfect! */);
+       if (ret)
+               goto out_put_ctrl;
+
+       spin_lock_init(&ctrl->lock);
+
+       ret = -ENOMEM;
+
+       ctrl->ctrl.sqsize = opts->queue_size;
+       ctrl->ctrl.kato = opts->kato;
+
+       ctrl->queues = kcalloc(opts->nr_io_queues + 1, sizeof(*ctrl->queues),
+                       GFP_KERNEL);
+       if (!ctrl->queues)
+               goto out_uninit_ctrl;
+
+       ret = nvme_loop_configure_admin_queue(ctrl);
+       if (ret)
+               goto out_free_queues;
+
+       if (opts->queue_size > ctrl->ctrl.maxcmd) {
+               /* warn if maxcmd is lower than queue_size */
+               dev_warn(ctrl->ctrl.device,
+                       "queue_size %zu > ctrl maxcmd %u, clamping down\n",
+                       opts->queue_size, ctrl->ctrl.maxcmd);
+               opts->queue_size = ctrl->ctrl.maxcmd;
+       }
+
+       if (opts->nr_io_queues) {
+               ret = nvme_loop_create_io_queues(ctrl);
+               if (ret)
+                       goto out_remove_admin_queue;
+       }
+
+       nvme_loop_init_iod(ctrl, &ctrl->async_event_iod, 0);
+
+       dev_info(ctrl->ctrl.device,
+                "new ctrl: \"%s\"\n", ctrl->ctrl.opts->subsysnqn);
+
+       kref_get(&ctrl->ctrl.kref);
+
+       changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
+       WARN_ON_ONCE(!changed);
+
+       mutex_lock(&nvme_loop_ctrl_mutex);
+       list_add_tail(&ctrl->list, &nvme_loop_ctrl_list);
+       mutex_unlock(&nvme_loop_ctrl_mutex);
+
+       if (opts->nr_io_queues) {
+               nvme_queue_scan(&ctrl->ctrl);
+               nvme_queue_async_events(&ctrl->ctrl);
+       }
+
+       return &ctrl->ctrl;
+
+out_remove_admin_queue:
+       nvme_loop_destroy_admin_queue(ctrl);
+out_free_queues:
+       kfree(ctrl->queues);
+out_uninit_ctrl:
+       nvme_uninit_ctrl(&ctrl->ctrl);
+out_put_ctrl:
+       nvme_put_ctrl(&ctrl->ctrl);
+       if (ret > 0)
+               ret = -EIO;
+       return ERR_PTR(ret);
+}
+
+static int nvme_loop_add_port(struct nvmet_port *port)
+{
+       /*
+        * XXX: disalow adding more than one port so
+        * there is no connection rejections when a
+        * a subsystem is assigned to a port for which
+        * loop doesn't have a pointer.
+        * This scenario would be possible if we allowed
+        * more than one port to be added and a subsystem
+        * was assigned to a port other than nvmet_loop_port.
+        */
+
+       if (nvmet_loop_port)
+               return -EPERM;
+
+       nvmet_loop_port = port;
+       return 0;
+}
+
+static void nvme_loop_remove_port(struct nvmet_port *port)
+{
+       if (port == nvmet_loop_port)
+               nvmet_loop_port = NULL;
+}
+
+static struct nvmet_fabrics_ops nvme_loop_ops = {
+       .owner          = THIS_MODULE,
+       .type           = NVMF_TRTYPE_LOOP,
+       .add_port       = nvme_loop_add_port,
+       .remove_port    = nvme_loop_remove_port,
+       .queue_response = nvme_loop_queue_response,
+       .delete_ctrl    = nvme_loop_delete_ctrl,
+};
+
+static struct nvmf_transport_ops nvme_loop_transport = {
+       .name           = "loop",
+       .create_ctrl    = nvme_loop_create_ctrl,
+};
+
+static int __init nvme_loop_init_module(void)
+{
+       int ret;
+
+       ret = nvmet_register_transport(&nvme_loop_ops);
+       if (ret)
+               return ret;
+       nvmf_register_transport(&nvme_loop_transport);
+       return 0;
+}
+
+static void __exit nvme_loop_cleanup_module(void)
+{
+       struct nvme_loop_ctrl *ctrl, *next;
+
+       nvmf_unregister_transport(&nvme_loop_transport);
+       nvmet_unregister_transport(&nvme_loop_ops);
+
+       mutex_lock(&nvme_loop_ctrl_mutex);
+       list_for_each_entry_safe(ctrl, next, &nvme_loop_ctrl_list, list)
+               __nvme_loop_del_ctrl(ctrl);
+       mutex_unlock(&nvme_loop_ctrl_mutex);
+
+       flush_scheduled_work();
+}
+
+module_init(nvme_loop_init_module);
+module_exit(nvme_loop_cleanup_module);
+
+MODULE_LICENSE("GPL v2");
+MODULE_ALIAS("nvmet-transport-254"); /* 254 == NVMF_TRTYPE_LOOP */