diff mbox series

[v2,4/5] PCI: endpoint: Add NVMe endpoint function driver

Message ID 20241011121951.90019-5-dlemoal@kernel.org (mailing list archive)
State New
Delegated to: Manivannan Sadhasivam
Headers show
Series NVMe PCI endpoint function driver | expand

Commit Message

Damien Le Moal Oct. 11, 2024, 12:19 p.m. UTC
From: Damien Le Moal <damien.lemoal@opensource.wdc.com>

Add a Linux PCI Endpoint function driver that implements a PCIe NVMe
controller for a local NVMe fabrics host controller. The NVMe endpoint
function driver relies as most as possible on the NVMe fabrics driver
for executing NVMe commands received from the host, to minimize NVMe
command parsing. However, some admin commands must be modified to
satisfy PCI transport specifications constraints (e.g. queue management
commands support and the optional SGL support).

The NVMe endpoint function drivers is created as follows:
1) Upon binding of the endpoint function driver to the endpoint
   controller (pci_epf_nvme_bind()), the function driver sets up BAR 0
   for the NVMe PCI controller with enough doorbell space to accommodate
   up to PCI_EPF_NVME_MAX_NR_QUEUES (16) queue pairs. The DMA channels
   that will be used to exchange data with the host over the PCI link
   are also initialized.
2) The endpoint function driver then creates the NVMe host fabrics
   controller using nvmef_create_ctrl() (called from
   pci_epf_nvme_create_ctrl()), which connects the host controller to
   its target (e.g. a loop target with a file or block device or a TCP
   remote target).
3) Once the PCI link status is detected to be up, the endpoint
   controller initializes IRQ management and BAR 0 content to advertize
   its capabilities. The capabilities of the fabrics controller are
   mostly used unmodified (pci_epf_nvme_init_ctrl_regs()). With that,
   the endpoint controller starts a delayed task to poll the BAR 0
   register bar to detect changes to the CC register.
4) When the PCI host enables the controller, pci_epf_nvme_enable_ctrl()
   is called to create the admin submission and completion queues and
   start the fabrics controller with nvme_start_ctrl(). The endpoint
   controller then starts a delayed work to poll the admin submission
   queue doorbell to detect commands from the PCI host.
5) Admin commands received from the PCI host are retrieved from the
   admin queue by mapping the queue memory to PCI memory space, copying
   the command locally using a struct pci_epf_nvme_cmd, and proccess the
   command using pci_epf_nvme_process_admin_cmd().
6) I/O commands are similarly handled: each I/O submission queue uses a
   delayed work to poll the queue doorbell and upon detection of a
   command being issued by the host, the I/O command is copied locally
   and processed using pci_epf_nvme_process_io_cmd().

I/O and admin commands are processed as follows:
1) A minimal parsing of the command is done to determine the command
   buffer size and data transfer direction. The command processing then
   continues using a command work scheduled using a per queue-pair
   high-priority workqueue (pci_epf_nvme_exec_cmd_work()).
2) The command execution work calls pci_epf_nvme_exec_cmd() which will
   retrieve and parse the command PRPs to determine the PCI address
   location of the command buffer segments, and retrieve the command
   data if the command is a write command. The command is then executed
   using the host fabrics controller by calling
   __nvme_submit_sync_cmd(). Once done, pci_epf_nvme_complete_cmd() is
   called to complete the command, after having transferred the command
   data back to the PCI host in the case of a read command.
3) pci_epf_nvme_complete_cmd() queues the command in a completion list
   for the completion queue of the command and schedules the queue
   completion work which will batch CQ entry transfers to the PCI host
   with the completion queue memory mapped to the host PCI address of
   the completion queue.

With this processing, most of the command parsing and handling is left
to the NVMe fabrics code. The only NVMe specific parsing implemented in
the endpoint driver is the command PRP parsing. Of note is that the
current code does not support SGL (this capability is thus not
advertized).

For data transfers, the endpoint driver relies by default on the DMA RX
and TX channels of the hardware endpoint PCI controller. If no DMA
channels are available, the NVMe endpoint function driver falls back to
using mmio, which degrades performance significantly but keeps the
function working.

The BAR register polling work also monitors for controller-disable
events (e.g. the PCI host reboots or shutdown). Such events trigger
calls to pci_epf_nvme_disable_ctrl() which drains, cleanups and destroys
the local queue pairs.

The configuration and enablement of this NVMe endpoint function driver
can be fully controlled using configfs, once a NVMe fabrics target is
also setup. The available configfs parameters are:
 - ctrl_opts: Fabrics controller connection arguments, as formatted for
   the nvme cli "connect" command.
 - dma_enable: Enable (default) or disable DMA data transfers.
 - mdts_kb: Change the maximum data transfer size (default: 128 KB).

Early versions of this driver code were based on an RFC submission by
Alan Mikhak <alan.mikhak@sifive.com> (https://lwn.net/Articles/804369/).
The code however has since been completely rewritten.

Co-developed-by: Rick Wertenbroek <rick.wertenbroek@gmail.com>
Signed-off-by: Rick Wertenbroek <rick.wertenbroek@gmail.com>
Signed-off-by: Damien Le Moal <damien.lemoal@opensource.wdc.com>
---
 MAINTAINERS                                   |    7 +
 drivers/pci/endpoint/functions/Kconfig        |    9 +
 drivers/pci/endpoint/functions/Makefile       |    1 +
 drivers/pci/endpoint/functions/pci-epf-nvme.c | 2591 +++++++++++++++++
 4 files changed, 2608 insertions(+)
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-nvme.c

Comments

Christoph Hellwig Oct. 14, 2024, 8:44 a.m. UTC | #1
For one please keep nvme target code in drivers/nvme/  PCI endpoint is
just another transport and should not have device class logic.

But I also really fail to understand the architecture of the whole
thing.  It is a target driver and should in no way tie into the NVMe
host code, the host code runs on the other side of the PCIe wire.
Damien Le Moal Oct. 14, 2024, 10:41 a.m. UTC | #2
On 10/14/24 17:44, Christoph Hellwig wrote:
> For one please keep nvme target code in drivers/nvme/  PCI endpoint is
> just another transport and should not have device class logic.
> 
> But I also really fail to understand the architecture of the whole
> thing.  It is a target driver and should in no way tie into the NVMe
> host code, the host code runs on the other side of the PCIe wire.

Nope, it is not a target driver. It is a PCI endpoint driver which turns the
host running it into a PCIe NVMe device. But the NVMe part implementation is
minimal. Instead I use an endpoint local fabrics host controller which is itself
connected to whatever target you want (loop, tcp, ...).

Overall, it looks like this:

         +-----------------------------------+
         | PCIe Host Machine (Root-Complex)  |
         | (BIOS, Grub, Linux, Windows, ...) |
         |                                   |
         |       +------------------+        |
         |       | NVMe PCIe driver |        |
         +-------+------------------+--------+
                           |
                 PCIe bus  |
                           |
        +----+---------------------------+-----+
        |    | PCIe NVMe endpoint driver |     |
	|    | (Handles BAR registers,   |     |
	|    | doorbells, IRQs, SQs, CQs |     |
	|    | and DMA transfers)        |     |
        |    +---------------------------+     |
        |                  |                   |
        |    +---------------------------+     |
        |    |     NVMe fabrics host     |     |
        |    +---------------------------+     |
        |                  |                   |
        |    +---------------------------+     |
        |    |     NVMe fabrics target   |     |
        |    |      (loop, TCP, ...)     |     |
        |    +---------------------------+     |
        |                                      |
        | PCIe Endpoint Machine (e.g. Rock 5B) |
        +--------------------------------------+

The nvme target can be anything that can be supported with the PCI Endpoint
Machine. With a small board like the Rock 5B, it is loop (file and block
device), TCP target or NVMe passtrhough (using the PCIe Gen2 M.2 E-Key slot).

Unless I am mistaken, if I use a PCI transport as the base for the endpoint
driver, I would be able to connect only to a PCIe nvme device as the backend, no
? With the above design, I can use anything support by nvmf as backend and that
is exposed to the root-complex host through the nvme endpoint PCIe driver.
To do that, the PCI endpoint driver mostly need only to create the fabrics host
with nvmf_create_ctrl(), which connects to the target and then the nvme endpoint
driver can execute the nvme commands with __nvme_submit_sync_cmd().
Only some admin commands need special handling (e.g. create sq/cq).
Christoph Hellwig Oct. 14, 2024, 11:38 a.m. UTC | #3
On Mon, Oct 14, 2024 at 07:41:18PM +0900, Damien Le Moal wrote:
>         |    | PCIe NVMe endpoint driver |     |
> 	|    | (Handles BAR registers,   |     |
> 	|    | doorbells, IRQs, SQs, CQs |     |
> 	|    | and DMA transfers)        |     |
>         |    +---------------------------+     |
>         |                  |                   |
>         |    +---------------------------+     |
>         |    |     NVMe fabrics host     |     |
>         |    +---------------------------+     |

That whole fabrics host here is broken.  The PCI frontend needs to
talk directly to the target code instead of doing a completely
pointless roundtrip through the block layer.

> Unless I am mistaken, if I use a PCI transport as the base for the endpoint
> driver, I would be able to connect only to a PCIe nvme device as the backend, no
> ?

No.  Any fabrisc transport (or frontend) can you any of the the
backends (file / bdev and passthrough).
diff mbox series

Patch

diff --git a/MAINTAINERS b/MAINTAINERS
index a097afd76ded..301e0a1b56e8 100644
--- a/MAINTAINERS
+++ b/MAINTAINERS
@@ -16558,6 +16558,13 @@  S:	Supported
 F:	drivers/platform/x86/nvidia-wmi-ec-backlight.c
 F:	include/linux/platform_data/x86/nvidia-wmi-ec-backlight.h
 
+NVME ENDPOINT DRIVER
+M:	Damien Le Moal <dlemoal@kernel.org>
+L:	linux-pci@vger.kernel.org
+L:	linux-nvme@lists.infradead.org
+S:	Supported
+F:	drivers/pci/endpoint/functions/pci-epf-nvme.c
+
 NVM EXPRESS DRIVER
 M:	Keith Busch <kbusch@kernel.org>
 M:	Jens Axboe <axboe@fb.com>
diff --git a/drivers/pci/endpoint/functions/Kconfig b/drivers/pci/endpoint/functions/Kconfig
index 0c9cea0698d7..ea641d558fb8 100644
--- a/drivers/pci/endpoint/functions/Kconfig
+++ b/drivers/pci/endpoint/functions/Kconfig
@@ -47,3 +47,12 @@  config PCI_EPF_MHI
 	   devices such as SDX55.
 
 	   If in doubt, say "N" to disable Endpoint driver for MHI bus.
+
+config PCI_EPF_NVME
+	tristate "PCI Endpoint NVMe function driver"
+	depends on PCI_ENDPOINT && NVME_TARGET
+	help
+	   Enable this configuration option to enable the NVMe PCI endpoint
+	   function driver.
+
+	   If in doubt, say "N".
diff --git a/drivers/pci/endpoint/functions/Makefile b/drivers/pci/endpoint/functions/Makefile
index 696473fce50e..fe2d6cf8c502 100644
--- a/drivers/pci/endpoint/functions/Makefile
+++ b/drivers/pci/endpoint/functions/Makefile
@@ -7,3 +7,4 @@  obj-$(CONFIG_PCI_EPF_TEST)		+= pci-epf-test.o
 obj-$(CONFIG_PCI_EPF_NTB)		+= pci-epf-ntb.o
 obj-$(CONFIG_PCI_EPF_VNTB) 		+= pci-epf-vntb.o
 obj-$(CONFIG_PCI_EPF_MHI)		+= pci-epf-mhi.o
+obj-$(CONFIG_PCI_EPF_NVME)		+= pci-epf-nvme.o
diff --git a/drivers/pci/endpoint/functions/pci-epf-nvme.c b/drivers/pci/endpoint/functions/pci-epf-nvme.c
new file mode 100644
index 000000000000..00314a7d1b20
--- /dev/null
+++ b/drivers/pci/endpoint/functions/pci-epf-nvme.c
@@ -0,0 +1,2591 @@ 
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * NVMe function driver for PCI Endpoint Framework
+ *
+ * Copyright (c) 2024, Western Digital Corporation or its affiliates.
+ * Copyright (c) 2024, Rick Wertenbroek <rick.wertenbroek@gmail.com>
+ *                     REDS Institute, HEIG-VD, HES-SO, Switzerland
+ */
+
+#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
+
+#include <linux/delay.h>
+#include <linux/dmaengine.h>
+#include <linux/io.h>
+#include <linux/module.h>
+#include <linux/mutex.h>
+#include <linux/nvme.h>
+#include <linux/pci_ids.h>
+#include <linux/pci-epc.h>
+#include <linux/pci-epf.h>
+#include <linux/pci_regs.h>
+#include <linux/slab.h>
+
+#include "../../../nvme/host/nvme.h"
+#include "../../../nvme/host/fabrics.h"
+
+/*
+ * Maximum number of queue pairs: A higheer this number, the more mapping
+ * windows of the PCI endpoint controller will be used. To avoid exceeding the
+ * maximum number of mapping windows available (i.e. avoid PCI space mapping
+ * failures) errors, the maximum number of queue pairs should be limited to
+ * the number of mapping windows minus 2 (one window for IRQ issuing and one
+ * window for data transfers) and divided by 2 (one mapping windows for the SQ
+ * and one mapping window for the CQ).
+ */
+#define PCI_EPF_NVME_MAX_NR_QUEUES	16
+
+/*
+ * Default maximum data transfer size: limit to 128 KB to avoid
+ * excessive local memory use for buffers.
+ */
+#define PCI_EPF_NVME_MDTS_KB		128
+#define PCI_EPF_NVME_MAX_MDTS_KB	1024
+
+/*
+ * Queue flags.
+ */
+#define PCI_EPF_NVME_QUEUE_IS_SQ	(1U << 0)
+#define PCI_EPF_NVME_QUEUE_LIVE		(1U << 1)
+
+/* PRP manipulation macros */
+#define pci_epf_nvme_prp_addr(ctrl, prp)	((prp) & ~(ctrl)->mps_mask)
+#define pci_epf_nvme_prp_ofst(ctrl, prp)	((prp) & (ctrl)->mps_mask)
+#define pci_epf_nvme_prp_size(ctrl, prp)	\
+	((size_t)((ctrl)->mps - pci_epf_nvme_prp_ofst(ctrl, prp)))
+
+static struct kmem_cache *epf_nvme_cmd_cache;
+
+struct pci_epf_nvme;
+
+/*
+ * Host PCI memory segment for admin and IO commands.
+ */
+struct pci_epf_nvme_segment {
+	phys_addr_t	pci_addr;
+	size_t		size;
+};
+
+/*
+ * Queue definition and mapping for the local PCI controller.
+ */
+struct pci_epf_nvme_queue {
+	struct pci_epf_nvme	*epf_nvme;
+
+	unsigned int		qflags;
+	int			ref;
+
+	phys_addr_t		pci_addr;
+	size_t			pci_size;
+	struct pci_epc_map	pci_map;
+
+	u16			qid;
+	u16			cqid;
+	u16			size;
+	u16			depth;
+	u16			flags;
+	u16			vector;
+	u16			head;
+	u16			tail;
+	u16			phase;
+	u32			db;
+
+	size_t			qes;
+
+	struct workqueue_struct	*cmd_wq;
+	struct delayed_work	work;
+	spinlock_t		lock;
+	struct list_head	list;
+
+	struct pci_epf_nvme_queue *sq;
+};
+
+/*
+ * Local PCI controller exposed with the endpoint function.
+ */
+struct pci_epf_nvme_ctrl {
+	/* Fabrics host controller */
+	struct nvme_ctrl		*ctrl;
+
+	/* Registers of the local PCI controller */
+	void				*reg;
+	u64				cap;
+	u32				vs;
+	u32				cc;
+	u32				csts;
+	u32				aqa;
+	u64				asq;
+	u64				acq;
+
+	size_t				adm_sqes;
+	size_t				adm_cqes;
+	size_t				io_sqes;
+	size_t				io_cqes;
+
+	size_t				mps_shift;
+	size_t				mps;
+	size_t				mps_mask;
+
+	size_t				mdts;
+
+	unsigned int			nr_queues;
+	struct pci_epf_nvme_queue	*sq;
+	struct pci_epf_nvme_queue	*cq;
+
+	struct workqueue_struct		*wq;
+};
+
+/*
+ * Descriptor of commands sent by the host.
+ */
+struct pci_epf_nvme_cmd {
+	struct list_head		link;
+	struct pci_epf_nvme		*epf_nvme;
+
+	int				sqid;
+	int				cqid;
+	unsigned int			status;
+	struct nvme_ns			*ns;
+	struct nvme_command		cmd;
+	struct nvme_completion		cqe;
+
+	/* Internal buffer that we will transfer over PCI */
+	size_t				buffer_size;
+	void				*buffer;
+	enum dma_data_direction		dma_dir;
+
+	/*
+	 * Host PCI address segments: if nr_segs is 1, we use only "seg",
+	 * otherwise, the segs array is allocated and used to store
+	 * multiple segments.
+	 */
+	unsigned int			nr_segs;
+	struct pci_epf_nvme_segment	seg;
+	struct pci_epf_nvme_segment	*segs;
+
+	struct work_struct		work;
+};
+
+/*
+ * EPF function private data representing our NVMe subsystem.
+ */
+struct pci_epf_nvme {
+	struct pci_epf			*epf;
+	const struct pci_epc_features	*epc_features;
+
+	void				*reg_bar;
+	size_t				msix_table_offset;
+
+	unsigned int			irq_type;
+	unsigned int			nr_vectors;
+
+	unsigned int			queue_count;
+
+	struct pci_epf_nvme_ctrl	ctrl;
+	bool				ctrl_enabled;
+
+	__le64				*prp_list_buf;
+
+	struct dma_chan			*dma_chan_tx;
+	struct dma_chan			*dma_chan_rx;
+	struct mutex			xfer_lock;
+
+	struct mutex			irq_lock;
+
+	struct delayed_work		reg_poll;
+
+	/* Function configfs attributes */
+	struct config_group		group;
+	char				*ctrl_opts_buf;
+	bool				dma_enable;
+	size_t				mdts_kb;
+};
+
+/*
+ * Read a 32-bits BAR register (equivalent to readl()).
+ */
+static inline u32 pci_epf_nvme_reg_read32(struct pci_epf_nvme_ctrl *ctrl,
+					  u32 reg)
+{
+	__le32 *ctrl_reg = ctrl->reg + reg;
+
+	return le32_to_cpu(READ_ONCE(*ctrl_reg));
+}
+
+/*
+ * Write a 32-bits BAR register (equivalent to writel()).
+ */
+static inline void pci_epf_nvme_reg_write32(struct pci_epf_nvme_ctrl *ctrl,
+					    u32 reg, u32 val)
+{
+	__le32 *ctrl_reg = ctrl->reg + reg;
+
+	WRITE_ONCE(*ctrl_reg, cpu_to_le32(val));
+}
+
+/*
+ * Read a 64-bits BAR register (equivalent to lo_hi_readq()).
+ */
+static inline u64 pci_epf_nvme_reg_read64(struct pci_epf_nvme_ctrl *ctrl,
+					  u32 reg)
+{
+	return (u64)pci_epf_nvme_reg_read32(ctrl, reg) |
+		((u64)pci_epf_nvme_reg_read32(ctrl, reg + 4) << 32);
+}
+
+/*
+ * Write a 64-bits BAR register (equivalent to lo_hi_writeq()).
+ */
+static inline void pci_epf_nvme_reg_write64(struct pci_epf_nvme_ctrl *ctrl,
+					    u32 reg, u64 val)
+{
+	pci_epf_nvme_reg_write32(ctrl, reg, val & 0xFFFFFFFF);
+	pci_epf_nvme_reg_write32(ctrl, reg + 4, (val >> 32) & 0xFFFFFFFF);
+}
+
+static inline bool pci_epf_nvme_ctrl_ready(struct pci_epf_nvme *epf_nvme)
+{
+	struct pci_epf_nvme_ctrl *ctrl = &epf_nvme->ctrl;
+
+	if (!epf_nvme->ctrl_enabled)
+		return false;
+	return (ctrl->cc & NVME_CC_ENABLE) && (ctrl->csts & NVME_CSTS_RDY);
+}
+
+struct pci_epf_nvme_dma_filter {
+	struct device *dev;
+	u32 dma_mask;
+};
+
+static bool pci_epf_nvme_dma_filter(struct dma_chan *chan, void *arg)
+{
+	struct pci_epf_nvme_dma_filter *filter = arg;
+	struct dma_slave_caps caps;
+
+	memset(&caps, 0, sizeof(caps));
+	dma_get_slave_caps(chan, &caps);
+
+	return chan->device->dev == filter->dev &&
+		(filter->dma_mask & caps.directions);
+}
+
+static bool pci_epf_nvme_init_dma(struct pci_epf_nvme *epf_nvme)
+{
+	struct pci_epf *epf = epf_nvme->epf;
+	struct device *dev = &epf->dev;
+	struct pci_epf_nvme_dma_filter filter;
+	struct dma_chan *chan;
+	dma_cap_mask_t mask;
+
+	mutex_init(&epf_nvme->xfer_lock);
+	mutex_init(&epf_nvme->irq_lock);
+
+	dma_cap_zero(mask);
+	dma_cap_set(DMA_SLAVE, mask);
+
+	filter.dev = epf->epc->dev.parent;
+	filter.dma_mask = BIT(DMA_DEV_TO_MEM);
+
+	chan = dma_request_channel(mask, pci_epf_nvme_dma_filter, &filter);
+	if (!chan)
+		return false;
+	epf_nvme->dma_chan_rx = chan;
+
+	filter.dma_mask = BIT(DMA_MEM_TO_DEV);
+	chan = dma_request_channel(mask, pci_epf_nvme_dma_filter, &filter);
+	if (!chan) {
+		dma_release_channel(epf_nvme->dma_chan_rx);
+		epf_nvme->dma_chan_rx = NULL;
+		return false;
+	}
+	epf_nvme->dma_chan_tx = chan;
+
+	dev_info(dev, "DMA RX channel %s, maximum segment size %u B\n",
+		 dma_chan_name(epf_nvme->dma_chan_rx),
+		 dma_get_max_seg_size(epf_nvme->dma_chan_rx->device->dev));
+	dev_info(dev, "DMA TX channel %s, maximum segment size %u B\n",
+		 dma_chan_name(epf_nvme->dma_chan_tx),
+		 dma_get_max_seg_size(epf_nvme->dma_chan_tx->device->dev));
+
+	return true;
+}
+
+static void pci_epf_nvme_clean_dma(struct pci_epf *epf)
+{
+	struct pci_epf_nvme *epf_nvme = epf_get_drvdata(epf);
+
+	if (epf_nvme->dma_chan_tx) {
+		dma_release_channel(epf_nvme->dma_chan_tx);
+		epf_nvme->dma_chan_tx = NULL;
+	}
+
+	if (epf_nvme->dma_chan_rx) {
+		dma_release_channel(epf_nvme->dma_chan_rx);
+		epf_nvme->dma_chan_rx = NULL;
+	}
+}
+
+static void pci_epf_nvme_dma_callback(void *param)
+{
+	complete(param);
+}
+
+static ssize_t pci_epf_nvme_dma_transfer(struct pci_epf_nvme *epf_nvme,
+					 struct pci_epf_nvme_segment *seg,
+					 enum dma_data_direction dir, void *buf)
+{
+	struct pci_epf *epf = epf_nvme->epf;
+	struct device *dma_dev = epf->epc->dev.parent;
+	struct dma_async_tx_descriptor *desc;
+	DECLARE_COMPLETION_ONSTACK(complete);
+	struct dma_slave_config sconf = {};
+	struct device *dev = &epf->dev;
+	phys_addr_t dma_addr;
+	struct dma_chan *chan;
+	dma_cookie_t cookie;
+	int ret;
+
+	switch (dir) {
+	case DMA_FROM_DEVICE:
+		chan = epf_nvme->dma_chan_rx;
+		sconf.direction = DMA_DEV_TO_MEM;
+		sconf.src_addr = seg->pci_addr;
+		break;
+	case DMA_TO_DEVICE:
+		chan = epf_nvme->dma_chan_tx;
+		sconf.direction = DMA_MEM_TO_DEV;
+		sconf.dst_addr = seg->pci_addr;
+		break;
+	default:
+		return -EINVAL;
+	}
+
+	ret = dmaengine_slave_config(chan, &sconf);
+	if (ret) {
+		dev_err(dev, "Failed to configure DMA channel\n");
+		return ret;
+	}
+
+	dma_addr = dma_map_single(dma_dev, buf, seg->size, dir);
+	ret = dma_mapping_error(dma_dev, dma_addr);
+	if (ret) {
+		dev_err(dev, "Failed to map remote memory\n");
+		return ret;
+	}
+
+	desc = dmaengine_prep_slave_single(chan, dma_addr,
+					   seg->size, sconf.direction,
+					   DMA_CTRL_ACK | DMA_PREP_INTERRUPT);
+	if (!desc) {
+		dev_err(dev, "Failed to prepare DMA\n");
+		ret = -EIO;
+		goto unmap;
+	}
+
+	desc->callback = pci_epf_nvme_dma_callback;
+	desc->callback_param = &complete;
+
+	cookie = dmaengine_submit(desc);
+	ret = dma_submit_error(cookie);
+	if (ret) {
+		dev_err(dev, "DMA submit failed %d\n", ret);
+		goto unmap;
+	}
+
+	dma_async_issue_pending(chan);
+	ret = wait_for_completion_timeout(&complete, msecs_to_jiffies(1000));
+	if (!ret) {
+		dev_err(dev, "DMA transfer timeout\n");
+		dmaengine_terminate_sync(chan);
+		ret = -ETIMEDOUT;
+		goto unmap;
+	}
+
+	ret = seg->size;
+
+unmap:
+	dma_unmap_single(dma_dev, dma_addr, seg->size, dir);
+
+	return ret;
+}
+
+static ssize_t pci_epf_nvme_mmio_transfer(struct pci_epf_nvme *epf_nvme,
+					  struct pci_epf_nvme_segment *seg,
+					  enum dma_data_direction dir,
+					  void *buf)
+{
+	struct pci_epf *epf = epf_nvme->epf;
+	struct pci_epc_map map;
+	int ret;
+
+	/* Map segment */
+	ret = pci_epc_mem_map(epf->epc, epf->func_no, epf->vfunc_no,
+			      seg->pci_addr, seg->size, &map);
+	if (ret)
+		return ret;
+
+	switch (dir) {
+	case DMA_FROM_DEVICE:
+		memcpy_fromio(buf, map.virt_addr, map.pci_size);
+		ret = map.pci_size;
+		break;
+	case DMA_TO_DEVICE:
+		memcpy_toio(map.virt_addr, buf, map.pci_size);
+		ret = map.pci_size;
+		break;
+	default:
+		ret = -EINVAL;
+		break;
+	}
+
+	pci_epc_mem_unmap(epf->epc, epf->func_no, epf->vfunc_no, &map);
+
+	return ret;
+}
+
+static int pci_epf_nvme_transfer(struct pci_epf_nvme *epf_nvme,
+				 struct pci_epf_nvme_segment *seg,
+				 enum dma_data_direction dir, void *buf)
+{
+	size_t size = seg->size;
+	int ret;
+
+	while (size) {
+		/*
+		 * Note: mmio transfers do not need serialization, but
+		 * this is an easy way to prevent using too many mapped
+		 * memory areauiswhich would lead to errors.
+		 */
+		mutex_lock(&epf_nvme->xfer_lock);
+		if (!epf_nvme->dma_enable)
+			ret = pci_epf_nvme_mmio_transfer(epf_nvme, seg,
+							 dir, buf);
+		else
+			ret = pci_epf_nvme_dma_transfer(epf_nvme, seg,
+							dir, buf);
+		mutex_unlock(&epf_nvme->xfer_lock);
+
+		if (ret < 0)
+			return ret;
+
+		size -= ret;
+		buf += ret;
+	}
+
+	return 0;
+}
+
+static const char *pci_epf_nvme_cmd_name(struct pci_epf_nvme_cmd *epcmd)
+{
+	u8 opcode = epcmd->cmd.common.opcode;
+
+	if (epcmd->sqid)
+		return nvme_get_opcode_str(opcode);
+	return nvme_get_admin_opcode_str(opcode);
+}
+
+static inline struct pci_epf_nvme_cmd *
+pci_epf_nvme_alloc_cmd(struct pci_epf_nvme *nvme)
+{
+	return kmem_cache_alloc(epf_nvme_cmd_cache, GFP_KERNEL);
+}
+
+static void pci_epf_nvme_exec_cmd_work(struct work_struct *work);
+
+static void pci_epf_nvme_init_cmd(struct pci_epf_nvme *epf_nvme,
+				  struct pci_epf_nvme_cmd *epcmd,
+				  int sqid, int cqid)
+{
+	memset(epcmd, 0, sizeof(*epcmd));
+	INIT_LIST_HEAD(&epcmd->link);
+	INIT_WORK(&epcmd->work, pci_epf_nvme_exec_cmd_work);
+	epcmd->epf_nvme = epf_nvme;
+	epcmd->sqid = sqid;
+	epcmd->cqid = cqid;
+	epcmd->status = NVME_SC_SUCCESS;
+	epcmd->dma_dir = DMA_NONE;
+}
+
+static int pci_epf_nvme_alloc_cmd_buffer(struct pci_epf_nvme_cmd *epcmd)
+{
+	void *buffer;
+
+	buffer = kmalloc(epcmd->buffer_size, GFP_KERNEL);
+	if (!buffer) {
+		epcmd->buffer_size = 0;
+		return -ENOMEM;
+	}
+
+	if (!epcmd->sqid)
+		memset(buffer, 0, epcmd->buffer_size);
+	epcmd->buffer = buffer;
+
+	return 0;
+}
+
+static int pci_epf_nvme_alloc_cmd_segs(struct pci_epf_nvme_cmd *epcmd,
+				       int nr_segs)
+{
+	struct pci_epf_nvme_segment *segs;
+
+	/* Single segment case: use the command embedded structure */
+	if (nr_segs == 1) {
+		epcmd->segs = &epcmd->seg;
+		epcmd->nr_segs = 1;
+		return 0;
+	}
+
+	/* More than one segment needed: allocate an array */
+	segs = kcalloc(nr_segs, sizeof(struct pci_epf_nvme_segment), GFP_KERNEL);
+	if (!segs)
+		return -ENOMEM;
+
+	epcmd->nr_segs = nr_segs;
+	epcmd->segs = segs;
+
+	return 0;
+}
+
+static void pci_epf_nvme_free_cmd(struct pci_epf_nvme_cmd *epcmd)
+{
+	if (epcmd->ns)
+		nvme_put_ns(epcmd->ns);
+
+	kfree(epcmd->buffer);
+
+	if (epcmd->segs && epcmd->segs != &epcmd->seg)
+		kfree(epcmd->segs);
+
+	kmem_cache_free(epf_nvme_cmd_cache, epcmd);
+}
+
+static void pci_epf_nvme_complete_cmd(struct pci_epf_nvme_cmd *epcmd)
+{
+	struct pci_epf_nvme *epf_nvme = epcmd->epf_nvme;
+	struct pci_epf_nvme_queue *cq;
+	unsigned long flags;
+
+	if (!pci_epf_nvme_ctrl_ready(epf_nvme)) {
+		pci_epf_nvme_free_cmd(epcmd);
+		return;
+	}
+
+	/*
+	 * Add the command to the list of completed commands for the
+	 * target cq and schedule the list processing.
+	 */
+	cq = &epf_nvme->ctrl.cq[epcmd->cqid];
+	spin_lock_irqsave(&cq->lock, flags);
+	list_add_tail(&epcmd->link, &cq->list);
+	queue_delayed_work(epf_nvme->ctrl.wq, &cq->work, 0);
+	spin_unlock_irqrestore(&cq->lock, flags);
+}
+
+static int pci_epf_nvme_transfer_cmd_data(struct pci_epf_nvme_cmd *epcmd)
+{
+	struct pci_epf_nvme *epf_nvme = epcmd->epf_nvme;
+	struct pci_epf_nvme_segment *seg;
+	void *buf = epcmd->buffer;
+	size_t size = 0;
+	int i, ret;
+
+	/* Transfer each segment of the command */
+	for (i = 0; i < epcmd->nr_segs; i++) {
+		seg = &epcmd->segs[i];
+
+		if (size >= epcmd->buffer_size) {
+			dev_err(&epf_nvme->epf->dev, "Invalid transfer size\n");
+			goto xfer_err;
+		}
+
+		ret = pci_epf_nvme_transfer(epf_nvme, seg, epcmd->dma_dir, buf);
+		if (ret)
+			goto xfer_err;
+
+		buf += seg->size;
+		size += seg->size;
+	}
+
+	return 0;
+
+xfer_err:
+	epcmd->status = NVME_SC_DATA_XFER_ERROR | NVME_STATUS_DNR;
+	return -EIO;
+}
+
+static void pci_epf_nvme_raise_irq(struct pci_epf_nvme *epf_nvme,
+				   struct pci_epf_nvme_queue *cq)
+{
+	struct pci_epf *epf = epf_nvme->epf;
+	int ret;
+
+	if (!(cq->qflags & NVME_CQ_IRQ_ENABLED))
+		return;
+
+	mutex_lock(&epf_nvme->irq_lock);
+
+	switch (epf_nvme->irq_type) {
+	case PCI_IRQ_MSIX:
+	case PCI_IRQ_MSI:
+		ret = pci_epc_raise_irq(epf->epc, epf->func_no, epf->vfunc_no,
+					epf_nvme->irq_type, cq->vector + 1);
+		if (!ret)
+			break;
+		/*
+		 * If we got an error, it is likely because the host is using
+		 * legacy IRQs (e.g. BIOS, grub).
+		 */
+		fallthrough;
+	case PCI_IRQ_INTX:
+		ret = pci_epc_raise_irq(epf->epc, epf->func_no, epf->vfunc_no,
+					PCI_IRQ_INTX, 0);
+		break;
+	default:
+		WARN_ON_ONCE(1);
+		ret = -EINVAL;
+		break;
+	}
+
+	if (ret)
+		dev_err(&epf->dev, "Raise IRQ failed %d\n", ret);
+
+	mutex_unlock(&epf_nvme->irq_lock);
+}
+
+/*
+ * Transfer a prp list from the host and return the number of prps.
+ */
+static int pci_epf_nvme_get_prp_list(struct pci_epf_nvme *epf_nvme, u64 prp,
+				     size_t xfer_len)
+{
+	struct pci_epf_nvme_ctrl *ctrl = &epf_nvme->ctrl;
+	size_t nr_prps = (xfer_len + ctrl->mps_mask) >> ctrl->mps_shift;
+	struct pci_epf_nvme_segment seg;
+	int ret;
+
+	/*
+	 * Compute the number of PRPs required for the number of bytes to
+	 * transfer (xfer_len). If this number overflows the memory page size
+	 * with the PRP list pointer specified, only return the space available
+	 * in the memory page, the last PRP in there will be a PRP list pointer
+	 * to the remaining PRPs.
+	 */
+	seg.pci_addr = prp;
+	seg.size = min(pci_epf_nvme_prp_size(ctrl, prp), nr_prps << 3);
+	ret = pci_epf_nvme_transfer(epf_nvme, &seg, DMA_FROM_DEVICE,
+				    epf_nvme->prp_list_buf);
+	if (ret)
+		return ret;
+
+	return seg.size >> 3;
+}
+
+static int pci_epf_nvme_cmd_parse_prp_list(struct pci_epf_nvme *epf_nvme,
+					   struct pci_epf_nvme_cmd *epcmd)
+{
+	struct pci_epf_nvme_ctrl *ctrl = &epf_nvme->ctrl;
+	struct nvme_command *cmd = &epcmd->cmd;
+	__le64 *prps = epf_nvme->prp_list_buf;
+	struct pci_epf_nvme_segment *seg;
+	size_t size = 0, ofst, prp_size, xfer_len;
+	size_t transfer_len = epcmd->buffer_size;
+	int nr_segs, nr_prps = 0;
+	phys_addr_t pci_addr;
+	int i = 0, ret;
+	u64 prp;
+
+	/*
+	 * Allocate segments for the command: this considers the worst case
+	 * scenario where all prps are discontiguous, so get as many segments
+	 * as we can have prps. In practice, most of the time, we will have
+	 * far less segments than prps.
+	 */
+	prp = le64_to_cpu(cmd->common.dptr.prp1);
+	if (!prp)
+		goto invalid_field;
+
+	ofst = pci_epf_nvme_prp_ofst(ctrl, prp);
+	nr_segs = (transfer_len + ofst + NVME_CTRL_PAGE_SIZE - 1)
+		>> NVME_CTRL_PAGE_SHIFT;
+
+	ret = pci_epf_nvme_alloc_cmd_segs(epcmd, nr_segs);
+	if (ret)
+		goto internal;
+
+	/* Set the first segment using prp1 */
+	seg = &epcmd->segs[0];
+	seg->pci_addr = prp;
+	seg->size = pci_epf_nvme_prp_size(ctrl, prp);
+
+	size = seg->size;
+	pci_addr = prp + size;
+	nr_segs = 1;
+
+	/*
+	 * Now build the PCI address segments using the prp lists, starting
+	 * from prp2.
+	 */
+	prp = le64_to_cpu(cmd->common.dptr.prp2);
+	if (!prp)
+		goto invalid_field;
+
+	while (size < transfer_len) {
+		xfer_len = transfer_len - size;
+
+		if (!nr_prps) {
+			/* Get the prp list */
+			nr_prps = pci_epf_nvme_get_prp_list(epf_nvme, prp,
+							    xfer_len);
+			if (nr_prps < 0)
+				goto internal;
+
+			i = 0;
+			ofst = 0;
+		}
+
+		/* Current entry */
+		prp = le64_to_cpu(prps[i]);
+		if (!prp)
+			goto invalid_field;
+
+		/* Did we reach the last prp entry of the list ? */
+		if (xfer_len > ctrl->mps && i == nr_prps - 1) {
+			/* We need more PRPs: prp is a list pointer */
+			nr_prps = 0;
+			continue;
+		}
+
+		/* Only the first prp is allowed to have an offset */
+		if (pci_epf_nvme_prp_ofst(ctrl, prp))
+			goto invalid_offset;
+
+		if (prp != pci_addr) {
+			/* Discontiguous prp: new segment */
+			nr_segs++;
+			if (WARN_ON_ONCE(nr_segs > epcmd->nr_segs))
+				goto internal;
+
+			seg++;
+			seg->pci_addr = prp;
+			seg->size = 0;
+			pci_addr = prp;
+		}
+
+		prp_size = min_t(size_t, ctrl->mps, xfer_len);
+		seg->size += prp_size;
+		pci_addr += prp_size;
+		size += prp_size;
+
+		i++;
+	}
+
+	epcmd->nr_segs = nr_segs;
+	ret = 0;
+
+	if (size != transfer_len) {
+		dev_err(&epf_nvme->epf->dev,
+			"PRPs transfer length mismatch %zu / %zu\n",
+			size, transfer_len);
+		goto internal;
+	}
+
+	return 0;
+
+internal:
+	epcmd->status = NVME_SC_INTERNAL | NVME_STATUS_DNR;
+	return -EINVAL;
+
+invalid_offset:
+	epcmd->status = NVME_SC_PRP_INVALID_OFFSET | NVME_STATUS_DNR;
+	return -EINVAL;
+
+invalid_field:
+	epcmd->status = NVME_SC_INVALID_FIELD | NVME_STATUS_DNR;
+	return -EINVAL;
+}
+
+static int pci_epf_nvme_cmd_parse_prp_simple(struct pci_epf_nvme *epf_nvme,
+					     struct pci_epf_nvme_cmd *epcmd)
+{
+	struct pci_epf_nvme_ctrl *ctrl = &epf_nvme->ctrl;
+	struct nvme_command *cmd = &epcmd->cmd;
+	size_t transfer_len = epcmd->buffer_size;
+	int ret, nr_segs = 1;
+	u64 prp1, prp2 = 0;
+	size_t prp1_size;
+
+	/* prp1 */
+	prp1 = le64_to_cpu(cmd->common.dptr.prp1);
+	prp1_size = pci_epf_nvme_prp_size(ctrl, prp1);
+
+	/* For commands crossing a page boundary, we should have a valid prp2 */
+	if (transfer_len > prp1_size) {
+		prp2 = le64_to_cpu(cmd->common.dptr.prp2);
+		if (!prp2)
+			goto invalid_field;
+		if (pci_epf_nvme_prp_ofst(ctrl, prp2))
+			goto invalid_offset;
+		if (prp2 != prp1 + prp1_size)
+			nr_segs = 2;
+	}
+
+	/* Create segments using the prps */
+	ret = pci_epf_nvme_alloc_cmd_segs(epcmd, nr_segs);
+	if (ret)
+		goto internal;
+
+	epcmd->segs[0].pci_addr = prp1;
+	if (nr_segs == 1) {
+		epcmd->segs[0].size = transfer_len;
+	} else {
+		epcmd->segs[0].size = prp1_size;
+		epcmd->segs[1].pci_addr = prp2;
+		epcmd->segs[1].size = transfer_len - prp1_size;
+	}
+
+	return 0;
+
+invalid_offset:
+	epcmd->status = NVME_SC_PRP_INVALID_OFFSET | NVME_STATUS_DNR;
+	return -EINVAL;
+
+invalid_field:
+	epcmd->status = NVME_SC_INVALID_FIELD | NVME_STATUS_DNR;
+	return -EINVAL;
+
+internal:
+	epcmd->status = NVME_SC_INTERNAL | NVME_STATUS_DNR;
+	return ret;
+}
+
+static int pci_epf_nvme_cmd_parse_dptr(struct pci_epf_nvme_cmd *epcmd)
+{
+	struct pci_epf_nvme *epf_nvme = epcmd->epf_nvme;
+	struct pci_epf_nvme_ctrl *ctrl = &epf_nvme->ctrl;
+	struct nvme_command *cmd = &epcmd->cmd;
+	u64 prp1 = le64_to_cpu(cmd->common.dptr.prp1);
+	size_t ofst;
+	int ret;
+
+	if (epcmd->buffer_size > ctrl->mdts)
+		goto invalid_field;
+
+	/* We do not support SGL for now */
+	if (epcmd->cmd.common.flags & NVME_CMD_SGL_ALL)
+		goto invalid_field;
+
+	/* Get PCI address segments for the command using its prps */
+	ofst = pci_epf_nvme_prp_ofst(ctrl, prp1);
+	if (ofst & 0x3)
+		goto invalid_offset;
+
+	if (epcmd->buffer_size + ofst <= NVME_CTRL_PAGE_SIZE * 2)
+		ret = pci_epf_nvme_cmd_parse_prp_simple(epf_nvme, epcmd);
+	else
+		ret = pci_epf_nvme_cmd_parse_prp_list(epf_nvme, epcmd);
+	if (ret)
+		return ret;
+
+	/* Get an internal buffer for the command */
+	ret = pci_epf_nvme_alloc_cmd_buffer(epcmd);
+	if (ret) {
+		epcmd->status = NVME_SC_INTERNAL | NVME_STATUS_DNR;
+		return ret;
+	}
+
+	return 0;
+
+invalid_field:
+	epcmd->status = NVME_SC_INVALID_FIELD | NVME_STATUS_DNR;
+	return -EINVAL;
+
+invalid_offset:
+	epcmd->status = NVME_SC_PRP_INVALID_OFFSET | NVME_STATUS_DNR;
+	return -EINVAL;
+}
+
+static void pci_epf_nvme_exec_cmd(struct pci_epf_nvme_cmd *epcmd,
+			void (*post_exec_hook)(struct pci_epf_nvme_cmd *))
+{
+	struct pci_epf_nvme *epf_nvme = epcmd->epf_nvme;
+	struct nvme_command *cmd = &epcmd->cmd;
+	struct request_queue *q;
+	int ret;
+
+	if (epcmd->ns)
+		q = epcmd->ns->queue;
+	else
+		q = epf_nvme->ctrl.ctrl->admin_q;
+
+	if (epcmd->buffer_size) {
+		/* Setup the command buffer */
+		ret = pci_epf_nvme_cmd_parse_dptr(epcmd);
+		if (ret)
+			return;
+
+		/* Get data from the host if needed */
+		if (epcmd->dma_dir == DMA_FROM_DEVICE) {
+			ret = pci_epf_nvme_transfer_cmd_data(epcmd);
+			if (ret)
+				return;
+		}
+	}
+
+	/* Synchronously execute the command */
+	ret = __nvme_submit_sync_cmd(q, cmd, &epcmd->cqe.result,
+				     epcmd->buffer, epcmd->buffer_size,
+				     NVME_QID_ANY, 0);
+	if (ret < 0)
+		epcmd->status = NVME_SC_INTERNAL | NVME_STATUS_DNR;
+	else if (ret > 0)
+		epcmd->status = ret;
+
+	if (epcmd->status != NVME_SC_SUCCESS) {
+		dev_err(&epf_nvme->epf->dev,
+			"QID %d: submit command %s (0x%x) failed, status 0x%0x\n",
+			epcmd->sqid, pci_epf_nvme_cmd_name(epcmd),
+			epcmd->cmd.common.opcode, epcmd->status);
+		return;
+	}
+
+	if (post_exec_hook)
+		post_exec_hook(epcmd);
+
+	if (epcmd->buffer_size && epcmd->dma_dir == DMA_TO_DEVICE)
+		pci_epf_nvme_transfer_cmd_data(epcmd);
+}
+
+static void pci_epf_nvme_exec_cmd_work(struct work_struct *work)
+{
+	struct pci_epf_nvme_cmd *epcmd =
+		container_of(work, struct pci_epf_nvme_cmd, work);
+
+	pci_epf_nvme_exec_cmd(epcmd, NULL);
+
+	pci_epf_nvme_complete_cmd(epcmd);
+}
+
+static bool pci_epf_nvme_queue_response(struct pci_epf_nvme_cmd *epcmd)
+{
+	struct pci_epf_nvme *epf_nvme = epcmd->epf_nvme;
+	struct pci_epf *epf = epf_nvme->epf;
+	struct pci_epf_nvme_ctrl *ctrl = &epf_nvme->ctrl;
+	struct pci_epf_nvme_queue *sq = &ctrl->sq[epcmd->sqid];
+	struct pci_epf_nvme_queue *cq = &ctrl->cq[epcmd->cqid];
+	struct nvme_completion *cqe = &epcmd->cqe;
+
+	/*
+	 * Do not try to complete commands if the controller is not ready
+	 * anymore, e.g. after the host cleared CC.EN.
+	 */
+	if (!pci_epf_nvme_ctrl_ready(epf_nvme) ||
+	    !(cq->qflags & PCI_EPF_NVME_QUEUE_LIVE))
+		goto free_cmd;
+
+	/* Check completion queue full state */
+	cq->head = pci_epf_nvme_reg_read32(ctrl, cq->db);
+	if (cq->head == cq->tail + 1)
+		return false;
+
+	/* Setup the completion entry */
+	cqe->sq_id = cpu_to_le16(epcmd->sqid);
+	cqe->sq_head = cpu_to_le16(sq->head);
+	cqe->command_id = epcmd->cmd.common.command_id;
+	cqe->status = cpu_to_le16((epcmd->status << 1) | cq->phase);
+
+	/* Post the completion entry */
+	dev_dbg(&epf->dev,
+		"cq[%d]: %s status 0x%x, head %d, tail %d, phase %d\n",
+		epcmd->cqid, pci_epf_nvme_cmd_name(epcmd),
+		epcmd->status, cq->head, cq->tail, cq->phase);
+
+	memcpy_toio(cq->pci_map.virt_addr + cq->tail * cq->qes, cqe,
+		    sizeof(struct nvme_completion));
+
+	/* Advance the tail */
+	cq->tail++;
+	if (cq->tail >= cq->depth) {
+		cq->tail = 0;
+		cq->phase ^= 1;
+	}
+
+free_cmd:
+	pci_epf_nvme_free_cmd(epcmd);
+
+	return true;
+}
+
+static int pci_epf_nvme_map_queue(struct pci_epf_nvme *epf_nvme,
+				  struct pci_epf_nvme_queue *q)
+{
+	struct pci_epf *epf = epf_nvme->epf;
+	int ret;
+
+	ret = pci_epc_mem_map(epf->epc, epf->func_no, epf->vfunc_no,
+			      q->pci_addr, q->pci_size, &q->pci_map);
+	if (ret) {
+		dev_err(&epf->dev, "Map %cQ %d failed %d\n",
+			q->qflags & PCI_EPF_NVME_QUEUE_IS_SQ ? 'S' : 'C',
+			q->qid, ret);
+		return ret;
+	}
+
+	if (q->pci_map.pci_size < q->pci_size) {
+		dev_err(&epf->dev, "Partial %cQ %d mapping\n",
+			q->qflags & PCI_EPF_NVME_QUEUE_IS_SQ ? 'S' : 'C',
+			q->qid);
+		pci_epc_mem_unmap(epf->epc, epf->func_no, epf->vfunc_no,
+				  &q->pci_map);
+		return -ENOMEM;
+	}
+
+	return 0;
+}
+
+static inline void pci_epf_nvme_unmap_queue(struct pci_epf_nvme *epf_nvme,
+					    struct pci_epf_nvme_queue *q)
+{
+	struct pci_epf *epf = epf_nvme->epf;
+
+	pci_epc_mem_unmap(epf->epc, epf->func_no, epf->vfunc_no,
+			  &q->pci_map);
+}
+
+static void pci_epf_nvme_delete_queue(struct pci_epf_nvme *epf_nvme,
+				      struct pci_epf_nvme_queue *q)
+{
+	struct pci_epf_nvme_cmd *epcmd;
+
+	q->qflags &= ~PCI_EPF_NVME_QUEUE_LIVE;
+
+	if (q->cmd_wq) {
+		flush_workqueue(q->cmd_wq);
+		destroy_workqueue(q->cmd_wq);
+		q->cmd_wq = NULL;
+	}
+
+	flush_delayed_work(&q->work);
+	cancel_delayed_work_sync(&q->work);
+
+	while (!list_empty(&q->list)) {
+		epcmd = list_first_entry(&q->list,
+					 struct pci_epf_nvme_cmd, link);
+		list_del_init(&epcmd->link);
+		pci_epf_nvme_free_cmd(epcmd);
+	}
+}
+
+static void pci_epf_nvme_cq_work(struct work_struct *work);
+
+static int pci_epf_nvme_create_cq(struct pci_epf_nvme *epf_nvme, int qid,
+				  int flags, int size, int vector,
+				  phys_addr_t pci_addr)
+{
+	struct pci_epf_nvme_ctrl *ctrl = &epf_nvme->ctrl;
+	struct pci_epf_nvme_queue *cq = &ctrl->cq[qid];
+	struct pci_epf *epf = epf_nvme->epf;
+
+	/*
+	 * Increment the queue reference count: if the queue is already being
+	 * used, we have nothing to do.
+	 */
+	cq->ref++;
+	if (cq->ref > 1)
+		return 0;
+
+	/* Setup the completion queue */
+	cq->pci_addr = pci_addr;
+	cq->qid = qid;
+	cq->cqid = qid;
+	cq->size = size;
+	cq->flags = flags;
+	cq->depth = size + 1;
+	cq->vector = vector;
+	cq->head = 0;
+	cq->tail = 0;
+	cq->phase = 1;
+	cq->db = NVME_REG_DBS + (((qid * 2) + 1) * sizeof(u32));
+	pci_epf_nvme_reg_write32(ctrl, cq->db, 0);
+	INIT_DELAYED_WORK(&cq->work, pci_epf_nvme_cq_work);
+	if (!qid)
+		cq->qes = ctrl->adm_cqes;
+	else
+		cq->qes = ctrl->io_cqes;
+	cq->pci_size = cq->qes * cq->depth;
+
+	dev_dbg(&epf->dev,
+		"CQ %d: %d entries of %zu B, vector IRQ %d\n",
+		qid, cq->size, cq->qes, (int)cq->vector + 1);
+
+	cq->qflags = PCI_EPF_NVME_QUEUE_LIVE;
+
+	return 0;
+}
+
+static void pci_epf_nvme_delete_cq(struct pci_epf_nvme *epf_nvme, int qid)
+{
+	struct pci_epf_nvme_queue *cq = &epf_nvme->ctrl.cq[qid];
+
+	if (cq->ref < 1)
+		return;
+
+	cq->ref--;
+	if (cq->ref)
+		return;
+
+	pci_epf_nvme_delete_queue(epf_nvme, cq);
+}
+
+static void pci_epf_nvme_sq_work(struct work_struct *work);
+
+static int pci_epf_nvme_create_sq(struct pci_epf_nvme *epf_nvme, int qid,
+				  int cqid, int flags, int size,
+				  phys_addr_t pci_addr)
+{
+	struct pci_epf_nvme_ctrl *ctrl = &epf_nvme->ctrl;
+	struct pci_epf_nvme_queue *sq = &ctrl->sq[qid];
+	struct pci_epf_nvme_queue *cq = &ctrl->cq[cqid];
+	struct pci_epf *epf = epf_nvme->epf;
+
+	/* Setup the submission queue */
+	sq->qflags = PCI_EPF_NVME_QUEUE_IS_SQ;
+	sq->pci_addr = pci_addr;
+	sq->ref = 1;
+	sq->qid = qid;
+	sq->cqid = cqid;
+	sq->size = size;
+	sq->flags = flags;
+	sq->depth = size + 1;
+	sq->head = 0;
+	sq->tail = 0;
+	sq->phase = 0;
+	sq->db = NVME_REG_DBS + (qid * 2 * sizeof(u32));
+	pci_epf_nvme_reg_write32(ctrl, sq->db, 0);
+	INIT_DELAYED_WORK(&sq->work, pci_epf_nvme_sq_work);
+	if (!qid)
+		sq->qes = ctrl->adm_sqes;
+	else
+		sq->qes = ctrl->io_sqes;
+	sq->pci_size = sq->qes * sq->depth;
+
+	sq->cmd_wq = alloc_workqueue("sq%d_wq", WQ_HIGHPRI | WQ_UNBOUND,
+				     min_t(int, sq->depth, WQ_MAX_ACTIVE), qid);
+	if (!sq->cmd_wq) {
+		dev_err(&epf->dev, "Create SQ %d cmd wq failed\n", qid);
+		memset(sq, 0, sizeof(*sq));
+		return -ENOMEM;
+	}
+
+	/* Get a reference on the completion queue */
+	cq->ref++;
+	cq->sq = sq;
+
+	dev_dbg(&epf->dev,
+		"SQ %d: %d queue entries of %zu B, CQ %d\n",
+		qid, size, sq->qes, cqid);
+
+	sq->qflags |= PCI_EPF_NVME_QUEUE_LIVE;
+
+	return 0;
+}
+
+static void pci_epf_nvme_delete_sq(struct pci_epf_nvme *epf_nvme, int qid)
+{
+	struct pci_epf_nvme_queue *sq = &epf_nvme->ctrl.sq[qid];
+
+	if (!sq->ref)
+		return;
+
+	sq->ref--;
+	if (WARN_ON_ONCE(sq->ref != 0))
+		return;
+
+	pci_epf_nvme_delete_queue(epf_nvme, sq);
+
+	if (epf_nvme->ctrl.cq[sq->cqid].ref)
+		epf_nvme->ctrl.cq[sq->cqid].ref--;
+}
+
+static void pci_epf_nvme_disable_ctrl(struct pci_epf_nvme *epf_nvme)
+{
+	struct pci_epf_nvme_ctrl *ctrl = &epf_nvme->ctrl;
+	struct pci_epf *epf = epf_nvme->epf;
+	int qid;
+
+	if (!epf_nvme->ctrl_enabled)
+		return;
+
+	dev_info(&epf->dev, "Disabling controller\n");
+
+	/*
+	 * Delete the submission queues first to release all references
+	 * to the completion queues. This also stops polling for submissions
+	 * and drains any pending command from the queue.
+	 */
+	for (qid = 1; qid < ctrl->nr_queues; qid++)
+		pci_epf_nvme_delete_sq(epf_nvme, qid);
+
+	for (qid = 1; qid < ctrl->nr_queues; qid++)
+		pci_epf_nvme_delete_cq(epf_nvme, qid);
+
+	/* Unmap the admin queue last */
+	pci_epf_nvme_delete_sq(epf_nvme, 0);
+	pci_epf_nvme_delete_cq(epf_nvme, 0);
+
+	/* Tell the host we are done */
+	ctrl->csts &= ~NVME_CSTS_RDY;
+	if (ctrl->cc & NVME_CC_SHN_NORMAL) {
+		ctrl->csts |= NVME_CSTS_SHST_CMPLT;
+		ctrl->cc &= ~NVME_CC_SHN_NORMAL;
+	}
+	ctrl->cc &= ~NVME_CC_ENABLE;
+	pci_epf_nvme_reg_write32(ctrl, NVME_REG_CSTS, ctrl->csts);
+	pci_epf_nvme_reg_write32(ctrl, NVME_REG_CC, ctrl->cc);
+
+	epf_nvme->ctrl_enabled = false;
+}
+
+static void pci_epf_nvme_delete_ctrl(struct pci_epf *epf)
+{
+	struct pci_epf_nvme *epf_nvme = epf_get_drvdata(epf);
+	struct pci_epf_nvme_ctrl *ctrl = &epf_nvme->ctrl;
+
+	dev_info(&epf->dev, "Deleting controller\n");
+
+	if (ctrl->ctrl) {
+		nvme_put_ctrl(ctrl->ctrl);
+		ctrl->ctrl = NULL;
+
+		ctrl->cc &= ~NVME_CC_SHN_NORMAL;
+		ctrl->csts |= NVME_CSTS_SHST_CMPLT;
+	}
+
+	pci_epf_nvme_disable_ctrl(epf_nvme);
+
+	if (ctrl->wq) {
+		flush_workqueue(ctrl->wq);
+		destroy_workqueue(ctrl->wq);
+		ctrl->wq = NULL;
+	}
+
+	ctrl->nr_queues = 0;
+	kfree(ctrl->cq);
+	ctrl->cq = NULL;
+	kfree(ctrl->sq);
+	ctrl->sq = NULL;
+}
+
+static struct pci_epf_nvme_queue *
+pci_epf_nvme_alloc_queues(struct pci_epf_nvme *epf_nvme, int nr_queues)
+{
+	struct pci_epf_nvme_queue *q;
+	int i;
+
+	q = kcalloc(nr_queues, sizeof(struct pci_epf_nvme_queue), GFP_KERNEL);
+	if (!q)
+		return NULL;
+
+	for (i = 0; i < nr_queues; i++) {
+		q[i].epf_nvme = epf_nvme;
+		spin_lock_init(&q[i].lock);
+		INIT_LIST_HEAD(&q[i].list);
+	}
+
+	return q;
+}
+
+static int pci_epf_nvme_create_ctrl(struct pci_epf *epf)
+{
+	struct pci_epf_nvme *epf_nvme = epf_get_drvdata(epf);
+	const struct pci_epc_features *features = epf_nvme->epc_features;
+	struct pci_epf_nvme_ctrl *ctrl = &epf_nvme->ctrl;
+	struct nvme_ctrl *fctrl;
+	int ret;
+
+	/* We must have nvme fabrics options. */
+	if (!epf_nvme->ctrl_opts_buf) {
+		dev_err(&epf->dev, "No nvme fabrics options specified\n");
+		return -EINVAL;
+	}
+
+	/* Create the fabrics controller */
+	fctrl = nvmf_create_ctrl(&epf->dev, epf_nvme->ctrl_opts_buf);
+	if (IS_ERR(fctrl)) {
+		dev_err(&epf->dev, "Create nvme fabrics controller failed\n");
+		return PTR_ERR(fctrl);
+	}
+
+	/* We only support IO controllers */
+	if (fctrl->cntrltype != NVME_CTRL_IO) {
+		dev_err(&epf->dev, "Unsupported controller type\n");
+		ret = -EINVAL;
+		goto out_delete_ctrl;
+	}
+
+	dev_info(&epf->dev, "NVMe fabrics controller created, %u I/O queues\n",
+		 fctrl->queue_count - 1);
+
+	epf_nvme->queue_count =
+		min(fctrl->queue_count, PCI_EPF_NVME_MAX_NR_QUEUES);
+	if (features->msix_capable && epf->msix_interrupts) {
+		dev_info(&epf->dev,
+			 "NVMe PCI controller supports MSI-X, %u vectors\n",
+			 epf->msix_interrupts);
+		epf_nvme->queue_count =
+			min(epf_nvme->queue_count, epf->msix_interrupts);
+	} else if (features->msi_capable && epf->msi_interrupts) {
+		dev_info(&epf->dev,
+			 "NVMe PCI controller supports MSI, %u vectors\n",
+			 epf->msi_interrupts);
+		epf_nvme->queue_count =
+			min(epf_nvme->queue_count, epf->msi_interrupts);
+	}
+
+	if (epf_nvme->queue_count < 2) {
+		dev_info(&epf->dev, "Invalid number of queues %u\n",
+			 epf_nvme->queue_count);
+		ret = -EINVAL;
+		goto out_delete_ctrl;
+	}
+
+	if (epf_nvme->queue_count != fctrl->queue_count)
+		dev_info(&epf->dev, "Limiting number of queues to %u\n",
+			 epf_nvme->queue_count);
+
+	dev_info(&epf->dev, "NVMe PCI controller: %u I/O queues\n",
+		 epf_nvme->queue_count - 1);
+
+	ret = -ENOMEM;
+
+	/* Create the workqueue for processing our SQs and CQs */
+	ctrl->wq = alloc_workqueue("ctrl_wq", WQ_HIGHPRI | WQ_UNBOUND,
+				min_t(int, ctrl->nr_queues * 2, WQ_MAX_ACTIVE));
+	if (!ctrl->wq) {
+		dev_err(&epf->dev, "Create controller wq failed\n");
+		goto out_delete_ctrl;
+	}
+
+	/* Allocate queues */
+	ctrl->nr_queues = epf_nvme->queue_count;
+	ctrl->sq = pci_epf_nvme_alloc_queues(epf_nvme, ctrl->nr_queues);
+	if (!ctrl->sq)
+		goto out_delete_ctrl;
+
+	ctrl->cq = pci_epf_nvme_alloc_queues(epf_nvme, ctrl->nr_queues);
+	if (!ctrl->cq)
+		goto out_delete_ctrl;
+
+	epf_nvme->ctrl.ctrl = fctrl;
+
+	return 0;
+
+out_delete_ctrl:
+	pci_epf_nvme_delete_ctrl(epf);
+
+	return ret;
+}
+
+static void pci_epf_nvme_init_ctrl_regs(struct pci_epf *epf)
+{
+	struct pci_epf_nvme *epf_nvme = epf_get_drvdata(epf);
+	struct pci_epf_nvme_ctrl *ctrl = &epf_nvme->ctrl;
+
+	ctrl->reg = epf_nvme->reg_bar;
+
+	/* Copy the fabrics controller capabilities as a base */
+	ctrl->cap = ctrl->ctrl->cap;
+
+	/* Contiguous Queues Required (CQR) */
+	ctrl->cap |= 0x1ULL << 16;
+
+	/* Set Doorbell stride to 4B (DSTRB) */
+	ctrl->cap &= ~GENMASK(35, 32);
+
+	/* Clear NVM Subsystem Reset Supported (NSSRS) */
+	ctrl->cap &= ~(0x1ULL << 36);
+
+	/* Clear Boot Partition Support (BPS) */
+	ctrl->cap &= ~(0x1ULL << 45);
+
+	/* Memory Page Size minimum (MPSMIN) = 4K */
+	ctrl->cap |= (NVME_CTRL_PAGE_SHIFT - 12) << NVME_CC_MPS_SHIFT;
+
+	/* Memory Page Size maximum (MPSMAX) = 4K */
+	ctrl->cap |= (NVME_CTRL_PAGE_SHIFT - 12) << NVME_CC_MPS_SHIFT;
+
+	/* Clear Persistent Memory Region Supported (PMRS) */
+	ctrl->cap &= ~(0x1ULL << 56);
+
+	/* Clear Controller Memory Buffer Supported (CMBS) */
+	ctrl->cap &= ~(0x1ULL << 57);
+
+	/* NVMe version supported */
+	ctrl->vs = ctrl->ctrl->vs;
+
+	/* Controller configuration */
+	ctrl->cc = ctrl->ctrl->ctrl_config & (~NVME_CC_ENABLE);
+
+	/* Controller Status (not ready) */
+	ctrl->csts = 0;
+
+	pci_epf_nvme_reg_write64(ctrl, NVME_REG_CAP, ctrl->cap);
+	pci_epf_nvme_reg_write32(ctrl, NVME_REG_VS, ctrl->vs);
+	pci_epf_nvme_reg_write32(ctrl, NVME_REG_CSTS, ctrl->csts);
+	pci_epf_nvme_reg_write32(ctrl, NVME_REG_CC, ctrl->cc);
+}
+
+static void pci_epf_nvme_enable_ctrl(struct pci_epf_nvme *epf_nvme)
+{
+	struct pci_epf_nvme_ctrl *ctrl = &epf_nvme->ctrl;
+	struct pci_epf *epf = epf_nvme->epf;
+	int ret;
+
+	dev_info(&epf->dev, "Enabling controller\n");
+
+	ctrl->mdts = epf_nvme->mdts_kb * SZ_1K;
+
+	ctrl->mps_shift = ((ctrl->cc >> NVME_CC_MPS_SHIFT) & 0xf) + 12;
+	ctrl->mps = 1UL << ctrl->mps_shift;
+	ctrl->mps_mask = ctrl->mps - 1;
+
+	ctrl->adm_sqes = 1UL << NVME_ADM_SQES;
+	ctrl->adm_cqes = sizeof(struct nvme_completion);
+	ctrl->io_sqes = 1UL << ((ctrl->cc >> NVME_CC_IOSQES_SHIFT) & 0xf);
+	ctrl->io_cqes = 1UL << ((ctrl->cc >> NVME_CC_IOCQES_SHIFT) & 0xf);
+
+	if (ctrl->io_sqes < sizeof(struct nvme_command)) {
+		dev_err(&epf->dev, "Unsupported IO sqes %zu (need %zu)\n",
+			ctrl->io_sqes, sizeof(struct nvme_command));
+		return;
+	}
+
+	if (ctrl->io_cqes < sizeof(struct nvme_completion)) {
+		dev_err(&epf->dev, "Unsupported IO cqes %zu (need %zu)\n",
+			ctrl->io_sqes, sizeof(struct nvme_completion));
+		return;
+	}
+
+	ctrl->aqa = pci_epf_nvme_reg_read32(ctrl, NVME_REG_AQA);
+	ctrl->asq = pci_epf_nvme_reg_read64(ctrl, NVME_REG_ASQ);
+	ctrl->acq = pci_epf_nvme_reg_read64(ctrl, NVME_REG_ACQ);
+
+	/*
+	 * Create the PCI controller admin completion and submission queues.
+	 */
+	ret = pci_epf_nvme_create_cq(epf_nvme, 0,
+				NVME_QUEUE_PHYS_CONTIG | NVME_CQ_IRQ_ENABLED,
+				(ctrl->aqa & 0x0fff0000) >> 16, 0,
+				ctrl->acq & GENMASK(63, 12));
+	if (ret)
+		return;
+
+	ret = pci_epf_nvme_create_sq(epf_nvme, 0, 0, NVME_QUEUE_PHYS_CONTIG,
+				     ctrl->aqa & 0x0fff,
+				     ctrl->asq & GENMASK(63, 12));
+	if (ret) {
+		pci_epf_nvme_delete_cq(epf_nvme, 0);
+		return;
+	}
+
+	nvme_start_ctrl(ctrl->ctrl);
+
+	/* Tell the host we are now ready */
+	ctrl->csts |= NVME_CSTS_RDY;
+	pci_epf_nvme_reg_write32(ctrl, NVME_REG_CSTS, ctrl->csts);
+
+	/* Start polling the admin submission queue */
+	queue_delayed_work(ctrl->wq, &ctrl->sq[0].work, msecs_to_jiffies(5));
+
+	epf_nvme->ctrl_enabled = true;
+}
+
+static void pci_epf_nvme_process_create_cq(struct pci_epf_nvme *epf_nvme,
+					   struct pci_epf_nvme_cmd *epcmd)
+{
+	struct nvme_command *cmd = &epcmd->cmd;
+	int mqes = NVME_CAP_MQES(epf_nvme->ctrl.cap);
+	u16 cqid, cq_flags, qsize, vector;
+	int ret;
+
+	cqid = le16_to_cpu(cmd->create_cq.cqid);
+	if (cqid >= epf_nvme->ctrl.nr_queues || epf_nvme->ctrl.cq[cqid].ref) {
+		epcmd->status = NVME_SC_QID_INVALID | NVME_STATUS_DNR;
+		return;
+	}
+
+	cq_flags = le16_to_cpu(cmd->create_cq.cq_flags);
+	if (!(cq_flags & NVME_QUEUE_PHYS_CONTIG)) {
+		epcmd->status = NVME_SC_INVALID_QUEUE | NVME_STATUS_DNR;
+		return;
+	}
+
+	qsize = le16_to_cpu(cmd->create_cq.qsize);
+	if (!qsize || qsize > NVME_CAP_MQES(epf_nvme->ctrl.cap)) {
+		if (qsize > mqes)
+			dev_warn(&epf_nvme->epf->dev,
+				 "Create CQ %d, qsize %d > mqes %d: buggy driver?\n",
+				 cqid, (int)qsize, mqes);
+		epcmd->status = NVME_SC_QUEUE_SIZE | NVME_STATUS_DNR;
+		return;
+	}
+
+	vector = le16_to_cpu(cmd->create_cq.irq_vector);
+	if (vector >= epf_nvme->nr_vectors) {
+		epcmd->status = NVME_SC_INVALID_VECTOR | NVME_STATUS_DNR;
+		return;
+	}
+
+	ret = pci_epf_nvme_create_cq(epf_nvme, cqid, cq_flags, qsize, vector,
+				     le64_to_cpu(cmd->create_cq.prp1));
+	if (ret)
+		epcmd->status = NVME_SC_INTERNAL | NVME_STATUS_DNR;
+}
+
+static void pci_epf_nvme_process_delete_cq(struct pci_epf_nvme *epf_nvme,
+					   struct pci_epf_nvme_cmd *epcmd)
+{
+	struct nvme_command *cmd = &epcmd->cmd;
+	u16 cqid;
+
+	cqid = le16_to_cpu(cmd->delete_queue.qid);
+	if (!cqid ||
+	    cqid >= epf_nvme->ctrl.nr_queues ||
+	    !epf_nvme->ctrl.cq[cqid].ref) {
+		epcmd->status = NVME_SC_QID_INVALID | NVME_STATUS_DNR;
+		return;
+	}
+
+	pci_epf_nvme_delete_cq(epf_nvme, cqid);
+}
+
+static void pci_epf_nvme_process_create_sq(struct pci_epf_nvme *epf_nvme,
+					   struct pci_epf_nvme_cmd *epcmd)
+{
+	struct nvme_command *cmd = &epcmd->cmd;
+	int mqes = NVME_CAP_MQES(epf_nvme->ctrl.cap);
+	u16 sqid, cqid, sq_flags, qsize;
+	int ret;
+
+	sqid = le16_to_cpu(cmd->create_sq.sqid);
+	if (!sqid || sqid > epf_nvme->ctrl.nr_queues ||
+	    epf_nvme->ctrl.sq[sqid].ref) {
+		epcmd->status = NVME_SC_QID_INVALID | NVME_STATUS_DNR;
+		return;
+	}
+
+	cqid = le16_to_cpu(cmd->create_sq.cqid);
+	if (!cqid || !epf_nvme->ctrl.cq[cqid].ref) {
+		epcmd->status = NVME_SC_CQ_INVALID | NVME_STATUS_DNR;
+		return;
+	}
+
+	sq_flags = le16_to_cpu(cmd->create_sq.sq_flags);
+	if (!(sq_flags & NVME_QUEUE_PHYS_CONTIG)) {
+		epcmd->status = NVME_SC_INVALID_QUEUE | NVME_STATUS_DNR;
+		return;
+	}
+
+	qsize = le16_to_cpu(cmd->create_sq.qsize);
+	if (!qsize || qsize > mqes) {
+		if (qsize > mqes)
+			dev_warn(&epf_nvme->epf->dev,
+				 "Create SQ %d, qsize %d > mqes %d: buggy driver?\n",
+				 sqid, (int)qsize, mqes);
+		epcmd->status = NVME_SC_QUEUE_SIZE | NVME_STATUS_DNR;
+		return;
+	}
+
+	ret = pci_epf_nvme_create_sq(epf_nvme, sqid, cqid, sq_flags, qsize,
+				     le64_to_cpu(cmd->create_sq.prp1));
+	if (ret) {
+		epcmd->status = NVME_SC_INTERNAL | NVME_STATUS_DNR;
+		return;
+	}
+
+	/* Start polling the submission queue */
+	queue_delayed_work(epf_nvme->ctrl.wq, &epf_nvme->ctrl.sq[sqid].work, 1);
+}
+
+static void pci_epf_nvme_process_delete_sq(struct pci_epf_nvme *epf_nvme,
+					   struct pci_epf_nvme_cmd *epcmd)
+{
+	struct nvme_command *cmd = &epcmd->cmd;
+	u16 sqid;
+
+	sqid = le16_to_cpu(cmd->delete_queue.qid);
+	if (!sqid ||
+	    sqid >= epf_nvme->ctrl.nr_queues ||
+	    !epf_nvme->ctrl.sq[sqid].ref) {
+		epcmd->status = NVME_SC_QID_INVALID | NVME_STATUS_DNR;
+		return;
+	}
+
+	pci_epf_nvme_delete_sq(epf_nvme, sqid);
+}
+
+static void pci_epf_nvme_identify_hook(struct pci_epf_nvme_cmd *epcmd)
+{
+	struct pci_epf_nvme *epf_nvme = epcmd->epf_nvme;
+	struct nvme_command *cmd = &epcmd->cmd;
+	struct nvme_id_ctrl *id = epcmd->buffer;
+	unsigned int page_shift;
+
+	if (cmd->identify.cns != NVME_ID_CNS_CTRL)
+		return;
+
+	/* Set device vendor IDs */
+	id->vid = cpu_to_le16(epf_nvme->epf->header->vendorid);
+	id->ssvid = id->vid;
+
+	/* Set Maximum Data Transfer Size (MDTS) */
+	page_shift = NVME_CAP_MPSMIN(epf_nvme->ctrl.ctrl->cap) + 12;
+	id->mdts = ilog2(epf_nvme->ctrl.mdts) - page_shift;
+
+	/* Clear Controller Multi-Path I/O and Namespace Sharing Capabilities */
+	id->cmic = 0;
+
+	/* Do not report support for Autonomous Power State Transitions */
+	id->apsta = 0;
+
+	/* Indicate no support for SGLs */
+	id->sgls = 0;
+}
+
+static void pci_epf_nvme_get_log_hook(struct pci_epf_nvme_cmd *epcmd)
+{
+	struct nvme_command *cmd = &epcmd->cmd;
+	struct nvme_effects_log *log = epcmd->buffer;
+
+	if (cmd->get_log_page.lid != NVME_LOG_CMD_EFFECTS)
+		return;
+
+	/*
+	 * ACS0     [Delete I/O Submission Queue     ] 00000001
+	 * CSUPP+  LBCC-  NCC-  NIC-  CCC-  USS-  No command restriction
+	 */
+	log->acs[0] |= cpu_to_le32(NVME_CMD_EFFECTS_CSUPP);
+
+	/*
+	 * ACS1     [Create I/O Submission Queue     ] 00000001
+	 * CSUPP+  LBCC-  NCC-  NIC-  CCC-  USS-  No command restriction
+	 */
+	log->acs[1] |= cpu_to_le32(NVME_CMD_EFFECTS_CSUPP);
+
+	/*
+	 * ACS4     [Delete I/O Completion Queue     ] 00000001
+	 * CSUPP+  LBCC-  NCC-  NIC-  CCC-  USS-  No command restriction
+	 */
+	log->acs[4] |= cpu_to_le32(NVME_CMD_EFFECTS_CSUPP);
+
+	/*
+	 * ACS5     [Create I/O Completion Queue     ] 00000001
+	 * CSUPP+  LBCC-  NCC-  NIC-  CCC-  USS-  No command restriction
+	 */
+	log->acs[5] |= cpu_to_le32(NVME_CMD_EFFECTS_CSUPP);
+}
+
+/*
+ * Returns true if the command has been handled
+ */
+static bool pci_epf_nvme_process_set_features(struct pci_epf_nvme_cmd *epcmd)
+{
+	struct pci_epf_nvme *epf_nvme = epcmd->epf_nvme;
+	struct pci_epf_nvme_ctrl *ctrl = &epf_nvme->ctrl;
+	u32 cdw10 = le32_to_cpu(epcmd->cmd.common.cdw10);
+	u32 cdw11 = le32_to_cpu(epcmd->cmd.common.cdw11);
+	u8 feat = cdw10 & 0xff;
+	u16 nr_ioq, nsqr, ncqr;
+	int qid;
+
+	switch (feat) {
+	case NVME_FEAT_NUM_QUEUES:
+		ncqr = (cdw11 >> 16) & 0xffff;
+		nsqr = cdw11 & 0xffff;
+		if (ncqr == 0xffff || nsqr == 0xffff) {
+			epcmd->status = NVME_SC_INVALID_FIELD | NVME_STATUS_DNR;
+			return true;
+		}
+
+		/* We cannot accept this command if we already have IO queues */
+		for (qid = 1; qid < ctrl->nr_queues; qid++) {
+			if (epf_nvme->ctrl.sq[qid].ref ||
+			    epf_nvme->ctrl.cq[qid].ref) {
+				epcmd->status =
+					NVME_SC_CMD_SEQ_ERROR | NVME_STATUS_DNR;
+				return true;
+			}
+		}
+
+		/*
+		 * Number of I/O queues to report must not include the admin
+		 * queue and is a 0-based value, so it is the total number of
+		 * queues minus two.
+		 */
+		nr_ioq = ctrl->nr_queues - 2;
+		epcmd->cqe.result.u32 = cpu_to_le32(nr_ioq | (nr_ioq << 16));
+		return true;
+	case NVME_FEAT_IRQ_COALESCE:
+	case NVME_FEAT_ARBITRATION:
+		/* We do not need to do anything special here. */
+		epcmd->status = NVME_SC_SUCCESS;
+		return true;
+	default:
+		return false;
+	}
+}
+
+/*
+ * Returns true if the command has been handled
+ */
+static bool pci_epf_nvme_process_get_features(struct pci_epf_nvme_cmd *epcmd)
+{
+	struct pci_epf_nvme *epf_nvme = epcmd->epf_nvme;
+	struct pci_epf_nvme_ctrl *ctrl = &epf_nvme->ctrl;
+	u32 cdw10 = le32_to_cpu(epcmd->cmd.common.cdw10);
+	u8 feat = cdw10 & 0xff;
+	u16 nr_ioq;
+
+	switch (feat) {
+	case NVME_FEAT_NUM_QUEUES:
+		/*
+		 * Number of I/O queues to report must not include the admin
+		 * queue and is a 0-based value, so it is the total number of
+		 * queues minus two.
+		 */
+		nr_ioq = ctrl->nr_queues - 2;
+		epcmd->cqe.result.u32 = cpu_to_le32(nr_ioq | (nr_ioq << 16));
+		return true;
+	case NVME_FEAT_IRQ_COALESCE:
+	case NVME_FEAT_ARBITRATION:
+		/* We do not need to do anything special here. */
+		epcmd->status = NVME_SC_SUCCESS;
+		return true;
+	default:
+		return false;
+	}
+}
+
+static void pci_epf_nvme_process_admin_cmd(struct pci_epf_nvme_cmd *epcmd)
+{
+	struct pci_epf_nvme *epf_nvme = epcmd->epf_nvme;
+	void (*post_exec_hook)(struct pci_epf_nvme_cmd *) = NULL;
+	struct nvme_command *cmd = &epcmd->cmd;
+
+	switch (cmd->common.opcode) {
+	case nvme_admin_identify:
+		post_exec_hook = pci_epf_nvme_identify_hook;
+		epcmd->buffer_size = NVME_IDENTIFY_DATA_SIZE;
+		epcmd->dma_dir = DMA_TO_DEVICE;
+		break;
+
+	case nvme_admin_get_log_page:
+		post_exec_hook = pci_epf_nvme_get_log_hook;
+		epcmd->buffer_size = nvme_get_log_page_len(cmd);
+		epcmd->dma_dir = DMA_TO_DEVICE;
+		break;
+
+	case nvme_admin_async_event:
+		/*
+		 * Async events are a pain to deal with as they get canceled
+		 * only once we delete the fabrics controller, which happens
+		 * after the epf function is deleted, thus causing access to
+		 * freed memory or leaking of epcmd. So ignore these commands
+		 * for now, which is fine. The host will simply never see any
+		 * event.
+		 */
+		pci_epf_nvme_free_cmd(epcmd);
+		return;
+
+	case nvme_admin_set_features:
+		/*
+		 * Several NVMe features do not apply to the NVMe fabrics
+		 * host controller, so handle them directly here.
+		 */
+		if (pci_epf_nvme_process_set_features(epcmd))
+			goto complete;
+		break;
+
+	case nvme_admin_get_features:
+		/*
+		 * Several NVMe features do not apply to the NVMe fabrics
+		 * host controller, so handle them directly here.
+		 */
+		if (pci_epf_nvme_process_get_features(epcmd))
+			goto complete;
+
+	case nvme_admin_abort_cmd:
+		break;
+
+	case nvme_admin_create_cq:
+		pci_epf_nvme_process_create_cq(epf_nvme, epcmd);
+		goto complete;
+
+	case nvme_admin_create_sq:
+		pci_epf_nvme_process_create_sq(epf_nvme, epcmd);
+		goto complete;
+
+	case nvme_admin_delete_cq:
+		pci_epf_nvme_process_delete_cq(epf_nvme, epcmd);
+		goto complete;
+
+	case nvme_admin_delete_sq:
+		pci_epf_nvme_process_delete_sq(epf_nvme, epcmd);
+		goto complete;
+
+	default:
+		dev_err(&epf_nvme->epf->dev,
+			"Unhandled admin command %s (0x%02x)\n",
+			pci_epf_nvme_cmd_name(epcmd), cmd->common.opcode);
+		epcmd->status = NVME_SC_INVALID_OPCODE | NVME_STATUS_DNR;
+		goto complete;
+	}
+
+	/* Synchronously execute the command */
+	pci_epf_nvme_exec_cmd(epcmd, post_exec_hook);
+
+complete:
+	pci_epf_nvme_complete_cmd(epcmd);
+}
+
+static inline size_t pci_epf_nvme_rw_data_len(struct pci_epf_nvme_cmd *epcmd)
+{
+	return ((u32)le16_to_cpu(epcmd->cmd.rw.length) + 1) <<
+		epcmd->ns->head->lba_shift;
+}
+
+static void pci_epf_nvme_process_io_cmd(struct pci_epf_nvme_cmd *epcmd,
+					struct pci_epf_nvme_queue *sq)
+{
+	struct pci_epf_nvme *epf_nvme = epcmd->epf_nvme;
+
+	/* Get the command target namespace */
+	epcmd->ns = nvme_find_get_ns(epf_nvme->ctrl.ctrl,
+				     le32_to_cpu(epcmd->cmd.common.nsid));
+	if (!epcmd->ns) {
+		epcmd->status = NVME_SC_INVALID_NS | NVME_STATUS_DNR;
+		goto complete;
+	}
+
+	switch (epcmd->cmd.common.opcode) {
+	case nvme_cmd_read:
+		epcmd->buffer_size = pci_epf_nvme_rw_data_len(epcmd);
+		epcmd->dma_dir = DMA_TO_DEVICE;
+		break;
+
+	case nvme_cmd_write:
+		epcmd->buffer_size = pci_epf_nvme_rw_data_len(epcmd);
+		epcmd->dma_dir = DMA_FROM_DEVICE;
+		break;
+
+	case nvme_cmd_dsm:
+		epcmd->buffer_size = (le32_to_cpu(epcmd->cmd.dsm.nr) + 1) *
+			sizeof(struct nvme_dsm_range);
+		epcmd->dma_dir = DMA_FROM_DEVICE;
+		goto complete;
+
+	case nvme_cmd_flush:
+	case nvme_cmd_write_zeroes:
+		break;
+
+	default:
+		dev_err(&epf_nvme->epf->dev,
+			"Unhandled IO command %s (0x%02x)\n",
+			pci_epf_nvme_cmd_name(epcmd),
+			epcmd->cmd.common.opcode);
+		epcmd->status = NVME_SC_INVALID_OPCODE | NVME_STATUS_DNR;
+		goto complete;
+	}
+
+	queue_work(sq->cmd_wq, &epcmd->work);
+
+	return;
+
+complete:
+	pci_epf_nvme_complete_cmd(epcmd);
+}
+
+static bool pci_epf_nvme_fetch_cmd(struct pci_epf_nvme *epf_nvme,
+				   struct pci_epf_nvme_queue *sq)
+{
+	struct pci_epf_nvme_ctrl *ctrl = &epf_nvme->ctrl;
+	struct pci_epf_nvme_cmd *epcmd;
+	int ret;
+
+	if (!(sq->qflags & PCI_EPF_NVME_QUEUE_LIVE))
+		return false;
+
+	sq->tail = pci_epf_nvme_reg_read32(ctrl, sq->db);
+	if (sq->head == sq->tail)
+		return false;
+
+	ret = pci_epf_nvme_map_queue(epf_nvme, sq);
+	if (ret)
+		return false;
+
+	while (sq->head != sq->tail) {
+		epcmd = pci_epf_nvme_alloc_cmd(epf_nvme);
+		if (!epcmd)
+			break;
+
+		/* Get the NVMe command submitted by the host */
+		pci_epf_nvme_init_cmd(epf_nvme, epcmd, sq->qid, sq->cqid);
+		memcpy_fromio(&epcmd->cmd,
+			      sq->pci_map.virt_addr + sq->head * sq->qes,
+			      sizeof(struct nvme_command));
+
+		dev_dbg(&epf_nvme->epf->dev,
+			"sq[%d]: head %d/%d, tail %d, command %s\n",
+			sq->qid, (int)sq->head, (int)sq->depth,
+			(int)sq->tail, pci_epf_nvme_cmd_name(epcmd));
+
+		sq->head++;
+		if (sq->head == sq->depth)
+			sq->head = 0;
+
+		list_add_tail(&epcmd->link, &sq->list);
+	}
+
+	pci_epf_nvme_unmap_queue(epf_nvme, sq);
+
+	return !list_empty(&sq->list);
+}
+
+static void pci_epf_nvme_sq_work(struct work_struct *work)
+{
+	struct pci_epf_nvme_queue *sq =
+		container_of(work, struct pci_epf_nvme_queue, work.work);
+	struct pci_epf_nvme *epf_nvme = sq->epf_nvme;
+	struct pci_epf_nvme_cmd *epcmd;
+	unsigned long poll_interval = 1;
+	unsigned long j = jiffies;
+
+	while (pci_epf_nvme_ctrl_ready(epf_nvme) &&
+	       (sq->qflags & PCI_EPF_NVME_QUEUE_LIVE)) {
+		/*
+		 * Try to get commands from the host. If We do not yet have any
+		 * command, aggressively keep polling the SQ of IO queues for at
+		 * most one tick and fall back to rescheduling the SQ work if we
+		 * have not received any command after that. This hybrid
+		 * spin-polling method significantly increases the IOPS for
+		 * shallow queue depth operation (e.g. QD=1).
+		 */
+		if (!pci_epf_nvme_fetch_cmd(epf_nvme, sq)) {
+			if (!sq->qid || j != jiffies)
+				break;
+			usleep_range(1, 2);
+			continue;
+		}
+
+		while (!list_empty(&sq->list)) {
+			epcmd = list_first_entry(&sq->list,
+						 struct pci_epf_nvme_cmd, link);
+			list_del_init(&epcmd->link);
+			if (sq->qid)
+				pci_epf_nvme_process_io_cmd(epcmd, sq);
+			else
+				pci_epf_nvme_process_admin_cmd(epcmd);
+		}
+	}
+
+	if (!pci_epf_nvme_ctrl_ready(epf_nvme))
+		return;
+
+	/* No need to aggressively poll the admin queue. */
+	if (!sq->qid)
+		poll_interval = msecs_to_jiffies(5);
+	queue_delayed_work(epf_nvme->ctrl.wq, &sq->work, poll_interval);
+}
+
+static void pci_epf_nvme_cq_work(struct work_struct *work)
+{
+	struct pci_epf_nvme_queue *cq =
+		container_of(work, struct pci_epf_nvme_queue, work.work);
+	struct pci_epf_nvme *epf_nvme = cq->epf_nvme;
+	struct pci_epf_nvme_cmd *epcmd;
+	unsigned long flags;
+	LIST_HEAD(list);
+	int ret;
+
+	spin_lock_irqsave(&cq->lock, flags);
+
+	while (!list_empty(&cq->list)) {
+
+		list_splice_tail_init(&cq->list, &list);
+		spin_unlock_irqrestore(&cq->lock, flags);
+
+		ret = pci_epf_nvme_map_queue(epf_nvme, cq);
+		if (ret) {
+			queue_delayed_work(epf_nvme->ctrl.wq, &cq->work, 1);
+			return;
+		}
+
+		while (!list_empty(&list)) {
+			epcmd = list_first_entry(&list,
+						 struct pci_epf_nvme_cmd, link);
+			list_del_init(&epcmd->link);
+			if (!pci_epf_nvme_queue_response(epcmd))
+				break;
+		}
+
+		pci_epf_nvme_unmap_queue(epf_nvme, cq);
+
+		if (pci_epf_nvme_ctrl_ready(cq->epf_nvme))
+			pci_epf_nvme_raise_irq(cq->epf_nvme, cq);
+
+		spin_lock_irqsave(&cq->lock, flags);
+	}
+
+	/*
+	 * Completions on the host may trigger issuing of new commands. Try to
+	 * get these early to improve IOPS and reduce latency.
+	 */
+	if (cq->qid)
+		queue_delayed_work(epf_nvme->ctrl.wq, &cq->sq->work, 0);
+
+	spin_unlock_irqrestore(&cq->lock, flags);
+}
+
+static void pci_epf_nvme_reg_poll(struct work_struct *work)
+{
+	struct pci_epf_nvme *epf_nvme =
+		container_of(work, struct pci_epf_nvme, reg_poll.work);
+	struct pci_epf_nvme_ctrl *ctrl = &epf_nvme->ctrl;
+	u32 old_cc;
+
+	/* Set the controller register bar */
+	ctrl->reg = epf_nvme->reg_bar;
+	if (!ctrl->reg) {
+		dev_err(&epf_nvme->epf->dev, "No register BAR set\n");
+		goto again;
+	}
+
+	/* Check CC.EN to determine what we need to do */
+	old_cc = ctrl->cc;
+	ctrl->cc = pci_epf_nvme_reg_read32(ctrl, NVME_REG_CC);
+
+	/* If not enabled yet, wait */
+	if (!(old_cc & NVME_CC_ENABLE) && !(ctrl->cc & NVME_CC_ENABLE))
+		goto again;
+
+	/* If CC.EN was set by the host, enable the controller */
+	if (!(old_cc & NVME_CC_ENABLE) && (ctrl->cc & NVME_CC_ENABLE)) {
+		pci_epf_nvme_enable_ctrl(epf_nvme);
+		goto again;
+	}
+
+	/* If CC.EN was cleared by the host, disable the controller */
+	if (((old_cc & NVME_CC_ENABLE) && !(ctrl->cc & NVME_CC_ENABLE)) ||
+	    ctrl->cc & NVME_CC_SHN_NORMAL)
+		pci_epf_nvme_disable_ctrl(epf_nvme);
+
+again:
+	schedule_delayed_work(&epf_nvme->reg_poll, msecs_to_jiffies(5));
+}
+
+static int pci_epf_nvme_configure_bar(struct pci_epf *epf)
+{
+	struct pci_epf_nvme *epf_nvme = epf_get_drvdata(epf);
+	const struct pci_epc_features *features = epf_nvme->epc_features;
+	size_t reg_size, reg_bar_size;
+	size_t msix_table_size = 0;
+
+	/*
+	 * The first free BAR will be our register BAR and per NVMe
+	 * specifications, it must be BAR 0.
+	 */
+	if (pci_epc_get_first_free_bar(features) != BAR_0) {
+		dev_err(&epf->dev, "BAR 0 is not free\n");
+		return -EINVAL;
+	}
+
+	/* Initialize BAR flags */
+	if (features->bar[BAR_0].only_64bit)
+		epf->bar[BAR_0].flags |= PCI_BASE_ADDRESS_MEM_TYPE_64;
+
+	/*
+	 * Calculate the size of the register bar: NVMe registers first with
+	 * enough space for the doorbells, followed by the MSI-X table
+	 * if supported.
+	 */
+	reg_size = NVME_REG_DBS + (PCI_EPF_NVME_MAX_NR_QUEUES * 2 * sizeof(u32));
+	reg_size = ALIGN(reg_size, 8);
+
+	if (features->msix_capable) {
+		size_t pba_size;
+
+		msix_table_size = PCI_MSIX_ENTRY_SIZE * epf->msix_interrupts;
+		epf_nvme->msix_table_offset = reg_size;
+		pba_size = ALIGN(DIV_ROUND_UP(epf->msix_interrupts, 8), 8);
+
+		reg_size += msix_table_size + pba_size;
+	}
+
+	reg_bar_size = ALIGN(reg_size, 4096);
+
+	if (features->bar[BAR_0].type == BAR_FIXED) {
+		if (reg_bar_size > features->bar[BAR_0].fixed_size) {
+			dev_err(&epf->dev,
+				"Reg BAR 0 size %llu B too small, need %zu B\n",
+				features->bar[BAR_0].fixed_size,
+				reg_bar_size);
+			return -ENOMEM;
+		}
+		reg_bar_size = features->bar[BAR_0].fixed_size;
+	}
+
+	epf_nvme->reg_bar = pci_epf_alloc_space(epf, reg_bar_size, BAR_0,
+						features, PRIMARY_INTERFACE);
+	if (!epf_nvme->reg_bar) {
+		dev_err(&epf->dev, "Allocate register BAR failed\n");
+		return -ENOMEM;
+	}
+	memset(epf_nvme->reg_bar, 0, reg_bar_size);
+
+	return 0;
+}
+
+static void pci_epf_nvme_clear_bar(struct pci_epf *epf)
+{
+	struct pci_epf_nvme *epf_nvme = epf_get_drvdata(epf);
+
+	pci_epc_clear_bar(epf->epc, epf->func_no, epf->vfunc_no,
+			  &epf->bar[BAR_0]);
+	pci_epf_free_space(epf, epf_nvme->reg_bar, BAR_0, PRIMARY_INTERFACE);
+	epf_nvme->reg_bar = NULL;
+}
+
+static int pci_epf_nvme_init_irq(struct pci_epf *epf)
+{
+	struct pci_epf_nvme *epf_nvme = epf_get_drvdata(epf);
+	int ret;
+
+	/* Enable MSI-X if supported, otherwise, use MSI */
+	if (epf_nvme->epc_features->msix_capable && epf->msix_interrupts) {
+		ret = pci_epc_set_msix(epf->epc, epf->func_no, epf->vfunc_no,
+				       epf->msix_interrupts, BAR_0,
+				       epf_nvme->msix_table_offset);
+		if (ret) {
+			dev_err(&epf->dev, "MSI-X configuration failed\n");
+			return ret;
+		}
+
+		epf_nvme->nr_vectors = epf->msix_interrupts;
+		epf_nvme->irq_type = PCI_IRQ_MSIX;
+
+		return 0;
+	}
+
+	if (epf_nvme->epc_features->msi_capable && epf->msi_interrupts) {
+		ret = pci_epc_set_msi(epf->epc, epf->func_no, epf->vfunc_no,
+				      epf->msi_interrupts);
+		if (ret) {
+			dev_err(&epf->dev, "MSI configuration failed\n");
+			return ret;
+		}
+
+		epf_nvme->nr_vectors = epf->msi_interrupts;
+		epf_nvme->irq_type = PCI_IRQ_MSI;
+
+		return 0;
+	}
+
+	/* MSI and MSI-X are not supported: fall back to INTX */
+	epf_nvme->nr_vectors = 1;
+	epf_nvme->irq_type = PCI_IRQ_INTX;
+
+	return 0;
+}
+
+static int pci_epf_nvme_epc_init(struct pci_epf *epf)
+{
+	struct pci_epf_nvme *epf_nvme = epf_get_drvdata(epf);
+	int ret;
+
+	if (epf->vfunc_no <= 1) {
+		/* Set device ID, class, etc */
+		ret = pci_epc_write_header(epf->epc, epf->func_no, epf->vfunc_no,
+					   epf->header);
+		if (ret) {
+			dev_err(&epf->dev,
+				"Write configuration header failed %d\n", ret);
+			return ret;
+		}
+	}
+
+	/* Setup the PCIe BAR and enable interrupts */
+	ret = pci_epc_set_bar(epf->epc, epf->func_no, epf->vfunc_no,
+			      &epf->bar[BAR_0]);
+	if (ret) {
+		dev_err(&epf->dev, "Set BAR 0 failed\n");
+		pci_epf_free_space(epf, epf_nvme->reg_bar, BAR_0,
+				   PRIMARY_INTERFACE);
+		return ret;
+	}
+
+	ret = pci_epf_nvme_init_irq(epf);
+	if (ret)
+		return ret;
+
+	pci_epf_nvme_init_ctrl_regs(epf);
+
+	if (!epf_nvme->epc_features->linkup_notifier)
+		schedule_delayed_work(&epf_nvme->reg_poll, msecs_to_jiffies(5));
+
+	return 0;
+}
+
+static void pci_epf_nvme_epc_deinit(struct pci_epf *epf)
+{
+	struct pci_epf_nvme *epf_nvme = epf_get_drvdata(epf);
+
+	/* Stop polling BAR registers and disable the controller */
+	cancel_delayed_work_sync(&epf_nvme->reg_poll);
+
+	pci_epf_nvme_delete_ctrl(epf);
+	pci_epf_nvme_clean_dma(epf);
+	pci_epf_nvme_clear_bar(epf);
+}
+
+static int pci_epf_nvme_link_up(struct pci_epf *epf)
+{
+	struct pci_epf_nvme *epf_nvme = epf_get_drvdata(epf);
+
+	dev_info(&epf->dev, "Link UP\n");
+
+	pci_epf_nvme_init_ctrl_regs(epf);
+
+	/* Start polling the BAR registers to detect controller enable */
+	schedule_delayed_work(&epf_nvme->reg_poll, 0);
+
+	return 0;
+}
+
+static int pci_epf_nvme_link_down(struct pci_epf *epf)
+{
+	struct pci_epf_nvme *epf_nvme = epf_get_drvdata(epf);
+
+	dev_info(&epf->dev, "Link DOWN\n");
+
+	/* Stop polling BAR registers and disable the controller */
+	cancel_delayed_work_sync(&epf_nvme->reg_poll);
+	pci_epf_nvme_disable_ctrl(epf_nvme);
+
+	return 0;
+}
+
+static const struct pci_epc_event_ops pci_epf_nvme_event_ops = {
+	.epc_init = pci_epf_nvme_epc_init,
+	.epc_deinit = pci_epf_nvme_epc_deinit,
+	.link_up = pci_epf_nvme_link_up,
+	.link_down = pci_epf_nvme_link_down,
+};
+
+static int pci_epf_nvme_bind(struct pci_epf *epf)
+{
+	struct pci_epf_nvme *epf_nvme = epf_get_drvdata(epf);
+	const struct pci_epc_features *epc_features;
+	struct pci_epc *epc = epf->epc;
+	bool dma_supported;
+	int ret;
+
+	if (!epc) {
+		dev_err(&epf->dev, "No endpoint controller\n");
+		return -EINVAL;
+	}
+
+	epc_features = pci_epc_get_features(epc, epf->func_no, epf->vfunc_no);
+	if (!epc_features) {
+		dev_err(&epf->dev, "epc_features not implemented\n");
+		return -EOPNOTSUPP;
+	}
+	epf_nvme->epc_features = epc_features;
+
+	ret = pci_epf_nvme_configure_bar(epf);
+	if (ret)
+		return ret;
+
+	if (epf_nvme->dma_enable) {
+		dma_supported = pci_epf_nvme_init_dma(epf_nvme);
+		if (dma_supported) {
+			dev_info(&epf->dev, "DMA supported\n");
+		} else {
+			dev_info(&epf->dev,
+				 "DMA not supported, falling back to mmio\n");
+			epf_nvme->dma_enable = false;
+		}
+	} else {
+		dev_info(&epf->dev, "DMA disabled\n");
+	}
+
+	/* Create the fabrics host controller */
+	ret = pci_epf_nvme_create_ctrl(epf);
+	if (ret)
+		goto clean_dma;
+
+	return 0;
+
+clean_dma:
+	pci_epf_nvme_clean_dma(epf);
+	pci_epf_nvme_clear_bar(epf);
+
+	return ret;
+}
+
+static void pci_epf_nvme_unbind(struct pci_epf *epf)
+{
+	struct pci_epf_nvme *epf_nvme = epf_get_drvdata(epf);
+	struct pci_epc *epc = epf->epc;
+
+	cancel_delayed_work_sync(&epf_nvme->reg_poll);
+
+	pci_epf_nvme_delete_ctrl(epf);
+
+	if (epc->init_complete) {
+		pci_epf_nvme_clean_dma(epf);
+		pci_epf_nvme_clear_bar(epf);
+	}
+}
+
+static struct pci_epf_header epf_nvme_pci_header = {
+	.vendorid	= PCI_ANY_ID,
+	.deviceid	= PCI_ANY_ID,
+	.progif_code	= 0x02, /* NVM Express */
+	.baseclass_code = PCI_BASE_CLASS_STORAGE,
+	.subclass_code	= 0x08, /* Non-Volatile Memory controller */
+	.interrupt_pin	= PCI_INTERRUPT_INTA,
+};
+
+static int pci_epf_nvme_probe(struct pci_epf *epf,
+			      const struct pci_epf_device_id *id)
+{
+	struct pci_epf_nvme *epf_nvme;
+
+	epf_nvme = devm_kzalloc(&epf->dev, sizeof(*epf_nvme), GFP_KERNEL);
+	if (!epf_nvme)
+		return -ENOMEM;
+
+	epf_nvme->epf = epf;
+	INIT_DELAYED_WORK(&epf_nvme->reg_poll, pci_epf_nvme_reg_poll);
+
+	epf_nvme->prp_list_buf = devm_kzalloc(&epf->dev, NVME_CTRL_PAGE_SIZE,
+					      GFP_KERNEL);
+	if (!epf_nvme->prp_list_buf)
+		return -ENOMEM;
+
+	/* Set default attribute values */
+	epf_nvme->dma_enable = true;
+	epf_nvme->mdts_kb = PCI_EPF_NVME_MDTS_KB;
+
+	epf->event_ops = &pci_epf_nvme_event_ops;
+	epf->header = &epf_nvme_pci_header;
+	epf_set_drvdata(epf, epf_nvme);
+
+	return 0;
+}
+
+#define to_epf_nvme(epf_group)	\
+	container_of((epf_group), struct pci_epf_nvme, group)
+
+static ssize_t pci_epf_nvme_ctrl_opts_show(struct config_item *item,
+					   char *page)
+{
+	struct config_group *group = to_config_group(item);
+	struct pci_epf_nvme *epf_nvme = to_epf_nvme(group);
+
+	if (!epf_nvme->ctrl_opts_buf)
+		return 0;
+
+	return sysfs_emit(page, "%s\n", epf_nvme->ctrl_opts_buf);
+}
+
+#define PCI_EPF_NVME_OPT_HIDDEN_NS	"hidden_ns"
+
+static ssize_t pci_epf_nvme_ctrl_opts_store(struct config_item *item,
+					    const char *page, size_t len)
+{
+	struct config_group *group = to_config_group(item);
+	struct pci_epf_nvme *epf_nvme = to_epf_nvme(group);
+	size_t opt_buf_size;
+
+	/* Do not allow setting options when the function is already started */
+	if (epf_nvme->ctrl.ctrl)
+		return -EBUSY;
+
+	if (!len)
+		return -EINVAL;
+
+	kfree(epf_nvme->ctrl_opts_buf);
+
+	/*
+	 * Make sure we have enough room to add the hidden_ns option
+	 * if it is missing.
+	 */
+	opt_buf_size = len + strlen(PCI_EPF_NVME_OPT_HIDDEN_NS) + 2;
+	epf_nvme->ctrl_opts_buf = kzalloc(opt_buf_size, GFP_KERNEL);
+	if (!epf_nvme->ctrl_opts_buf)
+		return -ENOMEM;
+
+	strscpy(epf_nvme->ctrl_opts_buf, page, opt_buf_size);
+	if (!strnstr(page, PCI_EPF_NVME_OPT_HIDDEN_NS, len))
+		strncat(epf_nvme->ctrl_opts_buf,
+			"," PCI_EPF_NVME_OPT_HIDDEN_NS, opt_buf_size);
+
+	dev_dbg(&epf_nvme->epf->dev,
+		"NVMe fabrics controller options: %s\n",
+		epf_nvme->ctrl_opts_buf);
+
+	return len;
+}
+
+CONFIGFS_ATTR(pci_epf_nvme_, ctrl_opts);
+
+static ssize_t pci_epf_nvme_dma_enable_show(struct config_item *item,
+					    char *page)
+{
+	struct config_group *group = to_config_group(item);
+	struct pci_epf_nvme *epf_nvme = to_epf_nvme(group);
+
+	return sysfs_emit(page, "%d\n", epf_nvme->dma_enable);
+}
+
+static ssize_t pci_epf_nvme_dma_enable_store(struct config_item *item,
+					     const char *page, size_t len)
+{
+	struct config_group *group = to_config_group(item);
+	struct pci_epf_nvme *epf_nvme = to_epf_nvme(group);
+	int ret;
+
+	if (epf_nvme->ctrl_enabled)
+		return -EBUSY;
+
+	ret = kstrtobool(page, &epf_nvme->dma_enable);
+	if (ret)
+		return ret;
+
+	return len;
+}
+
+CONFIGFS_ATTR(pci_epf_nvme_, dma_enable);
+
+static ssize_t pci_epf_nvme_mdts_kb_show(struct config_item *item, char *page)
+{
+	struct config_group *group = to_config_group(item);
+	struct pci_epf_nvme *epf_nvme = to_epf_nvme(group);
+
+	return sysfs_emit(page, "%zu\n", epf_nvme->mdts_kb);
+}
+
+static ssize_t pci_epf_nvme_mdts_kb_store(struct config_item *item,
+					  const char *page, size_t len)
+{
+	struct config_group *group = to_config_group(item);
+	struct pci_epf_nvme *epf_nvme = to_epf_nvme(group);
+	unsigned long mdts_kb;
+	int ret;
+
+	if (epf_nvme->ctrl_enabled)
+		return -EBUSY;
+
+	ret = kstrtoul(page, 0, &mdts_kb);
+	if (ret)
+		return ret;
+	if (!mdts_kb)
+		mdts_kb = PCI_EPF_NVME_MDTS_KB;
+	else if (mdts_kb > PCI_EPF_NVME_MAX_MDTS_KB)
+		mdts_kb = PCI_EPF_NVME_MAX_MDTS_KB;
+
+	if (!is_power_of_2(mdts_kb))
+		return -EINVAL;
+
+	epf_nvme->mdts_kb = mdts_kb;
+
+	return len;
+}
+
+CONFIGFS_ATTR(pci_epf_nvme_, mdts_kb);
+
+static struct configfs_attribute *pci_epf_nvme_attrs[] = {
+	&pci_epf_nvme_attr_ctrl_opts,
+	&pci_epf_nvme_attr_dma_enable,
+	&pci_epf_nvme_attr_mdts_kb,
+	NULL,
+};
+
+static const struct config_item_type pci_epf_nvme_group_type = {
+	.ct_attrs	= pci_epf_nvme_attrs,
+	.ct_owner	= THIS_MODULE,
+};
+
+static struct config_group *pci_epf_nvme_add_cfs(struct pci_epf *epf,
+						 struct config_group *group)
+{
+	struct pci_epf_nvme *epf_nvme = epf_get_drvdata(epf);
+
+	/* Add the NVMe target attributes */
+	config_group_init_type_name(&epf_nvme->group, "nvme",
+				    &pci_epf_nvme_group_type);
+
+	return &epf_nvme->group;
+}
+
+static const struct pci_epf_device_id pci_epf_nvme_ids[] = {
+	{ .name = "pci_epf_nvme" },
+	{},
+};
+
+static struct pci_epf_ops pci_epf_nvme_ops = {
+	.bind	= pci_epf_nvme_bind,
+	.unbind	= pci_epf_nvme_unbind,
+	.add_cfs = pci_epf_nvme_add_cfs,
+};
+
+static struct pci_epf_driver epf_nvme_driver = {
+	.driver.name	= "pci_epf_nvme",
+	.probe		= pci_epf_nvme_probe,
+	.id_table	= pci_epf_nvme_ids,
+	.ops		= &pci_epf_nvme_ops,
+	.owner		= THIS_MODULE,
+};
+
+static int __init pci_epf_nvme_init(void)
+{
+	int ret;
+
+	epf_nvme_cmd_cache = kmem_cache_create("epf_nvme_cmd",
+					sizeof(struct pci_epf_nvme_cmd),
+					0, SLAB_HWCACHE_ALIGN, NULL);
+	if (!epf_nvme_cmd_cache)
+		return -ENOMEM;
+
+	ret = pci_epf_register_driver(&epf_nvme_driver);
+	if (ret)
+		goto out_cache;
+
+	pr_info("Registered nvme EPF driver\n");
+
+	return 0;
+
+out_cache:
+	kmem_cache_destroy(epf_nvme_cmd_cache);
+
+	pr_err("Register nvme EPF driver failed\n");
+
+	return ret;
+}
+module_init(pci_epf_nvme_init);
+
+static void __exit pci_epf_nvme_exit(void)
+{
+	pci_epf_unregister_driver(&epf_nvme_driver);
+
+	kmem_cache_destroy(epf_nvme_cmd_cache);
+
+	pr_info("Unregistered nvme EPF driver\n");
+}
+module_exit(pci_epf_nvme_exit);
+
+MODULE_DESCRIPTION("PCI endpoint NVMe function driver");
+MODULE_AUTHOR("Damien Le Moal <dlemoal@kernel.org>");
+MODULE_IMPORT_NS(NVME_TARGET_PASSTHRU);
+MODULE_IMPORT_NS(NVME_FABRICS);
+MODULE_LICENSE("GPL");