Skip to contentSkip to conversation

PCI: endpoint: Introduce a virtio-net EP function

RFC PATCH In review Shunsuke Mie 4 participants 29 messages Feb 3, 2023 – Apr 11, 2023

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

cover PCI: endpoint: Introduce a virtio-net EP function

This patchset introduce a virtio-net EP device function. It provides a new option to communiate between PCIe host and endpoint over IP. Advantage of this option is that the driver fully uses a PCIe embedded DMA. It is used to transport data between virtio ring directly each other. It can be expected to better throughput.

To realize the function, this patchset has few changes and introduces a new APIs to PCI EP framework related to virtio. Furthermore, it device depends on the some patchtes that is discussing. Those depended patchset are following:
- [PATCH 1/2] dmaengine: dw-edma: Fix to change for continuous transfer

- [RFC PATCH 0/3] Deal with alignment restriction on EP side

- [RFC PATCH v2 0/7] Introduce a vringh accessor for IO memory

About this patchset has 4 patches. The first of two patch is little changes to virtio. The third patch add APIs to easily access virtio data structure on PCIe Host side memory. The last one introduce a virtio-net EP device function. Details are in commit respectively.

Currently those network devices are testd using ping only. I'll add a result of performance evaluation using iperf and etc to the future version of this patchset.

Shunsuke Mie (4):
virtio_pci: add a definition of queue flag in ISR
virtio_ring: remove const from vring getter
PCI: endpoint: Introduce virtio library for EP functions
PCI: endpoint: function: Add EP function driver to provide virtio net
device

 drivers/pci/endpoint/Kconfig                  |   7 +
 drivers/pci/endpoint/Makefile                 |   1 +
 drivers/pci/endpoint/functions/Kconfig        |  12 +
 drivers/pci/endpoint/functions/Makefile       |   1 +
 .../pci/endpoint/functions/pci-epf-vnet-ep.c  | 343 ++++++++++
 .../pci/endpoint/functions/pci-epf-vnet-rc.c  | 635 ++++++++++++++++++
 drivers/pci/endpoint/functions/pci-epf-vnet.c | 387 +++++++++++
 drivers/pci/endpoint/functions/pci-epf-vnet.h |  62 ++
 drivers/pci/endpoint/pci-epf-virtio.c         | 113 ++++
 drivers/virtio/virtio_ring.c                  |   2 +-
 include/linux/pci-epf-virtio.h                |  25 +
 include/linux/virtio.h                        |   2 +-
 include/uapi/linux/virtio_pci.h               |   2 +
 13 files changed, 1590 insertions(+), 2 deletions(-)
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet-ep.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet-rc.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet.h
 create mode 100644 drivers/pci/endpoint/pci-epf-virtio.c
 create mode 100644 include/linux/pci-epf-virtio.h
Signature · 6 lines
2.25.1

_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
#1 lore raw

RFC PATCH 1/4 virtio_pci: add a definition of queue flag in ISR

Already it has beed defined a config changed flag of ISR, but not the queue flag. Add a macro for it.

---
 include/uapi/linux/virtio_pci.h | 2 ++
 1 file changed, 2 insertions(+)
include/uapi/linux/virtio_pci.h +2 -0
index f703afc7ad31..fa82afd6171a 100644
@@ -94,6 +94,8 @@
 
 #endif /* VIRTIO_PCI_NO_LEGACY */
 
+/* Ths bit of the ISR which indicates a queue entry update */
+#define VIRTIO_PCI_ISR_QUEUE		0x1
 /* The bit of the ISR which indicates a device configuration change. */
 #define VIRTIO_PCI_ISR_CONFIG		0x2
 /* Vector value used to disable MSI for queue */
Signature · 6 lines
2.25.1

_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
#2 lore raw

Re: [RFC PATCH 1/4] virtio_pci: add a definition of queue flag in ISR

Quoted 14 lines · Shunsuke Mie

Already it has beed defined a config changed flag of ISR, but not the queue flag. Add a macro for it.

---
 include/uapi/linux/virtio_pci.h | 2 ++
 1 file changed, 2 insertions(+)
include/uapi/linux/virtio_pci.h +0 -0
index f703afc7ad31..fa82afd6171a 100644

include/uapi/linux/virtio_pci.h line 97

Quoted 4 lines
include/uapi/linux/virtio_pci.h +1 -0
@@ -94,6 +94,8 @@
 
 #endif /* VIRTIO_PCI_NO_LEGACY */
 
+/* Ths bit of the ISR which indicates a queue entry update */

typo

Something to add here:
Note: only when MSI-X is disabled

Quoted 6 lines
include/uapi/linux/virtio_pci.h +1 -0
+#define VIRTIO_PCI_ISR_QUEUE		0x1
 /* The bit of the ISR which indicates a device configuration change. */
 #define VIRTIO_PCI_ISR_CONFIG		0x2
 /* Vector value used to disable MSI for queue */
Signature · 1 line
2.25.1

_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

#3 lore raw

Re: [RFC PATCH 1/4] virtio_pci: add a definition of queue flag in ISR

2023年2月3日(金) 19:16 Michael S. Tsirkin <[email protected]>:

Quoted 17 lines · Shunsuke Mie
Quoted 15 lines

Already it has beed defined a config changed flag of ISR, but not the queue flag. Add a macro for it.

---
 include/uapi/linux/virtio_pci.h | 2 ++
 1 file changed, 2 insertions(+)
include/uapi/linux/virtio_pci.h +0 -0
index f703afc7ad31..fa82afd6171a 100644
@@ -94,6 +94,8 @@
 
Quoted 7 lines

include/uapi/linux/virtio_pci.h line 97

Quoted 3 lines
include/uapi/linux/virtio_pci.h +1 -0
 #endif /* VIRTIO_PCI_NO_LEGACY */
 
+/* Ths bit of the ISR which indicates a queue entry update */

typo
Something to add here:
Note: only when MSI-X is disabled

I'll fix both that way.

Quoted 9 lines
Quoted 8 lines
include/uapi/linux/virtio_pci.h +1 -1
+#define VIRTIO_PCI_ISR_QUEUE         0x1
 /* The bit of the ISR which indicates a device configuration change. */
 #define VIRTIO_PCI_ISR_CONFIG                0x2
 /* Vector value used to disable MSI for queue */
--

2.25.1

Best,
Shunsuke
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

#4 lore raw

RFC PATCH 2/4 virtio_ring: remove const from vring getter

In reply to Shunsuke Mie

There are some method to manage the virto ring in Linux kernel. e.g. vhost and vringh. Remove const from the getter in order to control vring with other APIs, such as vringh.

---
 drivers/virtio/virtio_ring.c | 2 +-
 include/linux/virtio.h       | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)
drivers/virtio/virtio_ring.c +1 -1
index 2e7689bb933b..aa0c455d402b 100644
@@ -2857,7 +2857,7 @@ dma_addr_t virtqueue_get_used_addr(struct virtqueue *_vq)
 EXPORT_SYMBOL_GPL(virtqueue_get_used_addr);
 
 /* Only available for split ring */
-const struct vring *virtqueue_get_vring(struct virtqueue *vq)
+struct vring *virtqueue_get_vring(struct virtqueue *vq)
 {
 	return &to_vvq(vq)->split.vring;
 }
include/linux/virtio.h +1 -1
index dcab9c7e8784..83530b7bc2e9 100644
@@ -88,7 +88,7 @@ unsigned int virtqueue_get_vring_size(struct virtqueue *vq);
 
 bool virtqueue_is_broken(struct virtqueue *vq);
 
-const struct vring *virtqueue_get_vring(struct virtqueue *vq);
+struct vring *virtqueue_get_vring(struct virtqueue *vq);
 dma_addr_t virtqueue_get_desc_addr(struct virtqueue *vq);
 dma_addr_t virtqueue_get_avail_addr(struct virtqueue *vq);
 dma_addr_t virtqueue_get_used_addr(struct virtqueue *vq);
Signature · 6 lines
2.25.1

_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
#5 lore raw

RFC PATCH 3/4 PCI: endpoint: Introduce virtio library for EP functions

In reply to Shunsuke Mie

Add a new library to access a virtio ring located on PCIe host memory. The library generates struct pci_epf_vringh that is introduced in this patch. The struct has a vringh member, so vringh APIs can be used to access the virtio ring.

---
 drivers/pci/endpoint/Kconfig          |   7 ++
 drivers/pci/endpoint/Makefile         |   1 +
 drivers/pci/endpoint/pci-epf-virtio.c | 113 ++++++++++++++++++++++++++
 include/linux/pci-epf-virtio.h        |  25 ++++++
 4 files changed, 146 insertions(+)
 create mode 100644 drivers/pci/endpoint/pci-epf-virtio.c
 create mode 100644 include/linux/pci-epf-virtio.h
drivers/pci/endpoint/Kconfig +7 -0
index 17bbdc9bbde0..07276dcc43c8 100644
@@ -28,6 +28,13 @@ config PCI_ENDPOINT_CONFIGFS
 	   configure the endpoint function and used to bind the
 	   function with a endpoint controller.
 
+config PCI_ENDPOINT_VIRTIO
+	tristate
+	depends on PCI_ENDPOINT
+	select VHOST_IOMEM
+	help
+	  TODO update this comment
+
 source "drivers/pci/endpoint/functions/Kconfig"
 
 endmenu
drivers/pci/endpoint/Makefile +1 -0
index 95b2fe47e3b0..95712f0a13d1 100644
@@ -4,5 +4,6 @@
 #
 
 obj-$(CONFIG_PCI_ENDPOINT_CONFIGFS)	+= pci-ep-cfs.o
+obj-$(CONFIG_PCI_ENDPOINT_VIRTIO)	+= pci-epf-virtio.o
 obj-$(CONFIG_PCI_ENDPOINT)		+= pci-epc-core.o pci-epf-core.o\
 					   pci-epc-mem.o functions/
drivers/pci/endpoint/pci-epf-virtio.c new file +113 -0

113 lines not on this page · raw message

include/linux/pci-epf-virtio.h new file +25 -0
new file mode 100644
index 000000000000..ae09087919a9
@@ -0,0 +1,25 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * PCI Endpoint Function (EPF) for virtio definitions
+ */
+#ifndef __LINUX_PCI_EPF_VIRTIO_H
+#define __LINUX_PCI_EPF_VIRTIO_H
+
+#include <linux/types.h>
+#include <linux/vringh.h>
+#include <linux/pci-epf.h>
+
+struct pci_epf_vringh {
+	struct vringh vrh;
+	void __iomem *virt;
+	phys_addr_t phys;
+	size_t size;
+};
+
+struct pci_epf_vringh *pci_epf_virtio_alloc_vringh(struct pci_epf *epf,
+						   u64 features, u32 pfn,
+						   size_t size);
+void pci_epf_virtio_free_vringh(struct pci_epf *epf,
+				struct pci_epf_vringh *evrh);
+
+#endif // __LINUX_PCI_EPF_VIRTIO_H
Signature · 6 lines
2.25.1

_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
#6 lore raw

Re: [RFC PATCH 3/4] PCI: endpoint: Introduce virtio library for EP functions

Quoted 61 lines · Shunsuke Mie

Add a new library to access a virtio ring located on PCIe host memory. The library generates struct pci_epf_vringh that is introduced in this patch. The struct has a vringh member, so vringh APIs can be used to access the virtio ring.

---
 drivers/pci/endpoint/Kconfig          |   7 ++
 drivers/pci/endpoint/Makefile         |   1 +
 drivers/pci/endpoint/pci-epf-virtio.c | 113 ++++++++++++++++++++++++++
 include/linux/pci-epf-virtio.h        |  25 ++++++
 4 files changed, 146 insertions(+)
 create mode 100644 drivers/pci/endpoint/pci-epf-virtio.c
 create mode 100644 include/linux/pci-epf-virtio.h
drivers/pci/endpoint/Kconfig +7 -0
index 17bbdc9bbde0..07276dcc43c8 100644
@@ -28,6 +28,13 @@ config PCI_ENDPOINT_CONFIGFS
 	   configure the endpoint function and used to bind the
 	   function with a endpoint controller.
 
+config PCI_ENDPOINT_VIRTIO
+	tristate
+	depends on PCI_ENDPOINT
+	select VHOST_IOMEM
+	help
+	  TODO update this comment
+
 source "drivers/pci/endpoint/functions/Kconfig"
 
 endmenu
drivers/pci/endpoint/Makefile +1 -0
index 95b2fe47e3b0..95712f0a13d1 100644
@@ -4,5 +4,6 @@
 #
 
 obj-$(CONFIG_PCI_ENDPOINT_CONFIGFS)	+= pci-ep-cfs.o
+obj-$(CONFIG_PCI_ENDPOINT_VIRTIO)	+= pci-epf-virtio.o
 obj-$(CONFIG_PCI_ENDPOINT)		+= pci-epc-core.o pci-epf-core.o\
 					   pci-epc-mem.o functions/
drivers/pci/endpoint/pci-epf-virtio.c new file +10 -0
new file mode 100644
index 000000000000..7134ca407a03
@@ -0,0 +1,113 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Virtio library for PCI Endpoint function
+ */
+#include <linux/kernel.h>
+#include <linux/pci-epf-virtio.h>
+#include <linux/pci-epc.h>
+#include <linux/virtio_pci.h>
+
+static void __iomem *epf_virtio_map_vq(struct pci_epf *epf, u32 pfn,

drivers/pci/endpoint/pci-epf-virtio.c line 20

Quoted 10 lines
drivers/pci/endpoint/pci-epf-virtio.c new file +10 -0
+				       size_t size, phys_addr_t *vq_phys)
+{
+	int err;
+	phys_addr_t vq_addr;
+	size_t vq_size;
+	void __iomem *vq_virt;
+
+	vq_addr = (phys_addr_t)pfn << VIRTIO_PCI_QUEUE_ADDR_SHIFT;
+
+	vq_size = vring_size(size, VIRTIO_PCI_VRING_ALIGN) + 100;

100?

Also ugh, this uses the legacy vring_size.
Did not look closely but is all this limited to legacy virtio then? Pls make sure you code builds with #define VIRTIO_RING_NO_LEGACY.

Quoted 126 lines
drivers/pci/endpoint/pci-epf-virtio.c new file +93 -0
+
+	vq_virt = pci_epc_mem_alloc_addr(epf->epc, vq_phys, vq_size);
+	if (!vq_virt) {
+		pr_err("Failed to allocate epc memory\n");
+		return ERR_PTR(-ENOMEM);
+	}
+
+	err = pci_epc_map_addr(epf->epc, epf->func_no, epf->vfunc_no, *vq_phys,
+			       vq_addr, vq_size);
+	if (err) {
+		pr_err("Failed to map virtuqueue to local");
+		goto err_free;
+	}
+
+	return vq_virt;
+
+err_free:
+	pci_epc_mem_free_addr(epf->epc, *vq_phys, vq_virt, vq_size);
+
+	return ERR_PTR(err);
+}
+
+static void epf_virtio_unmap_vq(struct pci_epf *epf, void __iomem *vq_virt,
+				phys_addr_t vq_phys, size_t size)
+{
+	pci_epc_unmap_addr(epf->epc, epf->func_no, epf->vfunc_no, vq_phys);
+	pci_epc_mem_free_addr(epf->epc, vq_phys, vq_virt,
+			      vring_size(size, VIRTIO_PCI_VRING_ALIGN));
+}
+
+/**
+ * pci_epf_virtio_alloc_vringh() - allocate epf vringh from @pfn
+ * @epf: the EPF device that communicates to host virtio dirver
+ * @features: the virtio features of device
+ * @pfn: page frame number of virtqueue located on host memory. It is
+ *		passed during virtqueue negotiation.
+ * @size: a length of virtqueue
+ */
+struct pci_epf_vringh *pci_epf_virtio_alloc_vringh(struct pci_epf *epf,
+						   u64 features, u32 pfn,
+						   size_t size)
+{
+	int err;
+	struct vring vring;
+	struct pci_epf_vringh *evrh;
+
+	evrh = kmalloc(sizeof(*evrh), GFP_KERNEL);
+	if (!evrh) {
+		err = -ENOMEM;
+		goto err_unmap_vq;
+	}
+
+	evrh->size = size;
+
+	evrh->virt = epf_virtio_map_vq(epf, pfn, size, &evrh->phys);
+	if (IS_ERR(evrh->virt))
+		return evrh->virt;
+
+	vring_init(&vring, size, evrh->virt, VIRTIO_PCI_VRING_ALIGN);
+
+	err = vringh_init_iomem(&evrh->vrh, features, size, false, GFP_KERNEL,
+				vring.desc, vring.avail, vring.used);
+	if (err)
+		goto err_free_epf_vq;
+
+	return evrh;
+
+err_free_epf_vq:
+	kfree(evrh);
+
+err_unmap_vq:
+	epf_virtio_unmap_vq(epf, evrh->virt, evrh->phys, evrh->size);
+
+	return ERR_PTR(err);
+}
+EXPORT_SYMBOL_GPL(pci_epf_virtio_alloc_vringh);
+
+/**
+ * pci_epf_virtio_free_vringh() - release allocated epf vring
+ * @epf: the EPF device that communicates to host virtio dirver
+ * @evrh: epf vringh to free
+ */
+void pci_epf_virtio_free_vringh(struct pci_epf *epf,
+				struct pci_epf_vringh *evrh)
+{
+	epf_virtio_unmap_vq(epf, evrh->virt, evrh->phys, evrh->size);
+	kfree(evrh);
+}
+EXPORT_SYMBOL_GPL(pci_epf_virtio_free_vringh);
+
+MODULE_DESCRIPTION("PCI EP Virtio Library");
+MODULE_AUTHOR("Shunsuke Mie <[email protected]>");
+MODULE_LICENSE("GPL");
include/linux/pci-epf-virtio.h new file +25 -0
new file mode 100644
index 000000000000..ae09087919a9
@@ -0,0 +1,25 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * PCI Endpoint Function (EPF) for virtio definitions
+ */
+#ifndef __LINUX_PCI_EPF_VIRTIO_H
+#define __LINUX_PCI_EPF_VIRTIO_H
+
+#include <linux/types.h>
+#include <linux/vringh.h>
+#include <linux/pci-epf.h>
+
+struct pci_epf_vringh {
+	struct vringh vrh;
+	void __iomem *virt;
+	phys_addr_t phys;
+	size_t size;
+};
+
+struct pci_epf_vringh *pci_epf_virtio_alloc_vringh(struct pci_epf *epf,
+						   u64 features, u32 pfn,
+						   size_t size);
+void pci_epf_virtio_free_vringh(struct pci_epf *epf,
+				struct pci_epf_vringh *evrh);
+
+#endif // __LINUX_PCI_EPF_VIRTIO_H
Signature · 1 line
2.25.1

_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

#7 lore raw

Re: [RFC PATCH 3/4] PCI: endpoint: Introduce virtio library for EP functions

2023年2月3日(金) 19:20 Michael S. Tsirkin <[email protected]>:

Quoted 67 lines · Shunsuke Mie
Quoted 65 lines

Add a new library to access a virtio ring located on PCIe host memory. The library generates struct pci_epf_vringh that is introduced in this patch. The struct has a vringh member, so vringh APIs can be used to access the virtio ring.

---
 drivers/pci/endpoint/Kconfig          |   7 ++
 drivers/pci/endpoint/Makefile         |   1 +
 drivers/pci/endpoint/pci-epf-virtio.c | 113 ++++++++++++++++++++++++++
 include/linux/pci-epf-virtio.h        |  25 ++++++
 4 files changed, 146 insertions(+)
 create mode 100644 drivers/pci/endpoint/pci-epf-virtio.c
 create mode 100644 include/linux/pci-epf-virtio.h
drivers/pci/endpoint/Kconfig +7 -0
index 17bbdc9bbde0..07276dcc43c8 100644
@@ -28,6 +28,13 @@ config PCI_ENDPOINT_CONFIGFS
         configure the endpoint function and used to bind the
         function with a endpoint controller.
 
+config PCI_ENDPOINT_VIRTIO
+     tristate
+     depends on PCI_ENDPOINT
+     select VHOST_IOMEM
+     help
+       TODO update this comment
+
 source "drivers/pci/endpoint/functions/Kconfig"
 
 endmenu
drivers/pci/endpoint/Makefile +1 -0
index 95b2fe47e3b0..95712f0a13d1 100644
@@ -4,5 +4,6 @@
 #
 
 obj-$(CONFIG_PCI_ENDPOINT_CONFIGFS)  += pci-ep-cfs.o
+obj-$(CONFIG_PCI_ENDPOINT_VIRTIO)    += pci-epf-virtio.o
 obj-$(CONFIG_PCI_ENDPOINT)           += pci-epc-core.o pci-epf-core.o\
                                         pci-epc-mem.o functions/
drivers/pci/endpoint/pci-epf-virtio.c new file +14 -0
new file mode 100644
index 000000000000..7134ca407a03
@@ -0,0 +1,113 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Virtio library for PCI Endpoint function
+ */
+#include <linux/kernel.h>
+#include <linux/pci-epf-virtio.h>
+#include <linux/pci-epc.h>
+#include <linux/virtio_pci.h>
+
+static void __iomem *epf_virtio_map_vq(struct pci_epf *epf, u32 pfn,
+                                    size_t size, phys_addr_t *vq_phys)
+{
+     int err;
+     phys_addr_t vq_addr;
Quoted 8 lines

drivers/pci/endpoint/pci-epf-virtio.c line 20

Quoted 6 lines
drivers/pci/endpoint/pci-epf-virtio.c new file +6 -0
+     size_t vq_size;
+     void __iomem *vq_virt;
+
+     vq_addr = (phys_addr_t)pfn << VIRTIO_PCI_QUEUE_ADDR_SHIFT;
+
+     vq_size = vring_size(size, VIRTIO_PCI_VRING_ALIGN) + 100;

100?

It is a mistake and will be removed.

Quoted 3 lines

Also ugh, this uses the legacy vring_size.
Did not look closely but is all this limited to legacy virtio then? Pls make sure you code builds with #define VIRTIO_RING_NO_LEGACY.

Thanks for your suggestion, but this device works as a legacy device. In this case, the NO_LEGACY macro can not applicable I think. Is that right?

Quoted 118 lines
Quoted 118 lines
drivers/pci/endpoint/pci-epf-virtio.c new file +93 -0
+
+     vq_virt = pci_epc_mem_alloc_addr(epf->epc, vq_phys, vq_size);
+     if (!vq_virt) {
+             pr_err("Failed to allocate epc memory\n");
+             return ERR_PTR(-ENOMEM);
+     }
+
+     err = pci_epc_map_addr(epf->epc, epf->func_no, epf->vfunc_no, *vq_phys,
+                            vq_addr, vq_size);
+     if (err) {
+             pr_err("Failed to map virtuqueue to local");
+             goto err_free;
+     }
+
+     return vq_virt;
+
+err_free:
+     pci_epc_mem_free_addr(epf->epc, *vq_phys, vq_virt, vq_size);
+
+     return ERR_PTR(err);
+}
+
+static void epf_virtio_unmap_vq(struct pci_epf *epf, void __iomem *vq_virt,
+                             phys_addr_t vq_phys, size_t size)
+{
+     pci_epc_unmap_addr(epf->epc, epf->func_no, epf->vfunc_no, vq_phys);
+     pci_epc_mem_free_addr(epf->epc, vq_phys, vq_virt,
+                           vring_size(size, VIRTIO_PCI_VRING_ALIGN));
+}
+
+/**
+ * pci_epf_virtio_alloc_vringh() - allocate epf vringh from @pfn
+ * @epf: the EPF device that communicates to host virtio dirver
+ * @features: the virtio features of device
+ * @pfn: page frame number of virtqueue located on host memory. It is
+ *           passed during virtqueue negotiation.
+ * @size: a length of virtqueue
+ */
+struct pci_epf_vringh *pci_epf_virtio_alloc_vringh(struct pci_epf *epf,
+                                                u64 features, u32 pfn,
+                                                size_t size)
+{
+     int err;
+     struct vring vring;
+     struct pci_epf_vringh *evrh;
+
+     evrh = kmalloc(sizeof(*evrh), GFP_KERNEL);
+     if (!evrh) {
+             err = -ENOMEM;
+             goto err_unmap_vq;
+     }
+
+     evrh->size = size;
+
+     evrh->virt = epf_virtio_map_vq(epf, pfn, size, &evrh->phys);
+     if (IS_ERR(evrh->virt))
+             return evrh->virt;
+
+     vring_init(&vring, size, evrh->virt, VIRTIO_PCI_VRING_ALIGN);
+
+     err = vringh_init_iomem(&evrh->vrh, features, size, false, GFP_KERNEL,
+                             vring.desc, vring.avail, vring.used);
+     if (err)
+             goto err_free_epf_vq;
+
+     return evrh;
+
+err_free_epf_vq:
+     kfree(evrh);
+
+err_unmap_vq:
+     epf_virtio_unmap_vq(epf, evrh->virt, evrh->phys, evrh->size);
+
+     return ERR_PTR(err);
+}
+EXPORT_SYMBOL_GPL(pci_epf_virtio_alloc_vringh);
+
+/**
+ * pci_epf_virtio_free_vringh() - release allocated epf vring
+ * @epf: the EPF device that communicates to host virtio dirver
+ * @evrh: epf vringh to free
+ */
+void pci_epf_virtio_free_vringh(struct pci_epf *epf,
+                             struct pci_epf_vringh *evrh)
+{
+     epf_virtio_unmap_vq(epf, evrh->virt, evrh->phys, evrh->size);
+     kfree(evrh);
+}
+EXPORT_SYMBOL_GPL(pci_epf_virtio_free_vringh);
+
+MODULE_DESCRIPTION("PCI EP Virtio Library");
+MODULE_AUTHOR("Shunsuke Mie <[email protected]>");
+MODULE_LICENSE("GPL");
include/linux/pci-epf-virtio.h new file +19 -0
new file mode 100644
index 000000000000..ae09087919a9
@@ -0,0 +1,25 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+/*
+ * PCI Endpoint Function (EPF) for virtio definitions
+ */
+#ifndef __LINUX_PCI_EPF_VIRTIO_H
+#define __LINUX_PCI_EPF_VIRTIO_H
+
+#include <linux/types.h>
+#include <linux/vringh.h>
+#include <linux/pci-epf.h>
+
+struct pci_epf_vringh {
+     struct vringh vrh;
+     void __iomem *virt;
+     phys_addr_t phys;
+     size_t size;
+};
+
+struct pci_epf_vringh *pci_epf_virtio_alloc_vringh(struct pci_epf *epf,
Quoted 9 lines
Quoted 8 lines
include/linux/pci-epf-virtio.h new file +6 -1
+                                                u64 features, u32 pfn,
+                                                size_t size);
+void pci_epf_virtio_free_vringh(struct pci_epf *epf,
+                             struct pci_epf_vringh *evrh);
+
+#endif // __LINUX_PCI_EPF_VIRTIO_H
--

2.25.1

Best,
Shunsuke
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

#8 lore raw

RFC PATCH 4/4 PCI: endpoint: function: Add EP function driver to provide virtio net device

In reply to Shunsuke Mie

Add a new endpoint(EP) function driver to provide virtio-net device. This function not only shows virtio-net device for PCIe host system, but also provides virtio-net device to EP side(local) system. Virtualy those network devices are connected, so we can use to communicate over IP like a simple NIC.

Architecture overview is following:

to Host       |	                to Endpoint
network stack |                 network stack
      |       |                       |
+-----------+ |	+-----------+   +-----------+
|virtio-net | |	|virtio-net |   |virtio-net |
|driver     | |	|EP function|---|driver     |
+-----------+ |	+-----------+   +-----------+
      |       |	      |
+-----------+ | +-----------+
|PCIeC      | | |PCIeC      |
|Rootcomplex|-|-|Endpoint   |
+-----------+ | +-----------+
  Host side   |          Endpoint side

This driver uses PCIe EP framework to show virtio-net (pci) device Host side, and generate virtual virtio-net device and register to EP side. A communication date is diractly transported between virtqueue level with each other using PCIe embedded DMA controller.

by a limitation of the hardware and Linux EP framework, this function follows a virtio legacy specification.

This function driver has beed tested on S4 Rcar (r8a779fa-spider) board but just use the PCIe EP framework and depends on the PCIe EDMA.

---
 drivers/pci/endpoint/functions/Kconfig        |  12 +
 drivers/pci/endpoint/functions/Makefile       |   1 +
 .../pci/endpoint/functions/pci-epf-vnet-ep.c  | 343 ++++++++++
 .../pci/endpoint/functions/pci-epf-vnet-rc.c  | 635 ++++++++++++++++++
 drivers/pci/endpoint/functions/pci-epf-vnet.c | 387 +++++++++++
 drivers/pci/endpoint/functions/pci-epf-vnet.h |  62 ++
 6 files changed, 1440 insertions(+)
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet-ep.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet-rc.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet.h
drivers/pci/endpoint/functions/Kconfig +12 -0
index 9fd560886871..f88d8baaf689 100644
@@ -37,3 +37,15 @@ config PCI_EPF_VNTB
 	  between PCI Root Port and PCIe Endpoint.
 
 	  If in doubt, say "N" to disable Endpoint NTB driver.
+
+config PCI_EPF_VNET
+	tristate "PCI Endpoint virtio-net driver"
+	depends on PCI_ENDPOINT
+	select PCI_ENDPOINT_VIRTIO
+	select VHOST_RING
+	select VHOST_IOMEM
+	help
+	  PCIe Endpoint virtio-net function implementation. This module enables to
+	  show the virtio-net as pci device to PCIe Host side, and, another
+	  virtio-net device show to local machine. Those devices can communicate
+	  each other.
drivers/pci/endpoint/functions/Makefile +1 -0
index 5c13001deaba..74cc4c330c62 100644
@@ -6,3 +6,4 @@
 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_VNET)		+= pci-epf-vnet.o pci-epf-vnet-rc.o pci-epf-vnet-ep.o
drivers/pci/endpoint/functions/pci-epf-vnet-ep.c new file +343 -0

343 lines not on this page · raw message

drivers/pci/endpoint/functions/pci-epf-vnet-rc.c new file +635 -0

635 lines not on this page · raw message

drivers/pci/endpoint/functions/pci-epf-vnet.c new file +387 -0

387 lines not on this page · raw message

drivers/pci/endpoint/functions/pci-epf-vnet.h new file +62 -0

62 lines not on this page · raw message

Signature · 7 lines
2.25.1


_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization
#9 lore raw

Re: [RFC PATCH 4/4] PCI: endpoint: function: Add EP function driver to provide virtio net device

Quoted 22 lines · Shunsuke Mie

Add a new endpoint(EP) function driver to provide virtio-net device. This function not only shows virtio-net device for PCIe host system, but also provides virtio-net device to EP side(local) system. Virtualy those network devices are connected, so we can use to communicate over IP like a simple NIC.

Architecture overview is following:

to Host       |	                to Endpoint
network stack |                 network stack
      |       |                       |
+-----------+ |	+-----------+   +-----------+
|virtio-net | |	|virtio-net |   |virtio-net |
|driver     | |	|EP function|---|driver     |
+-----------+ |	+-----------+   +-----------+
      |       |	      |
+-----------+ | +-----------+
|PCIeC      | | |PCIeC      |
|Rootcomplex|-|-|Endpoint   |
+-----------+ | +-----------+
  Host side   |          Endpoint side
Quoted 3 lines

This driver uses PCIe EP framework to show virtio-net (pci) device Host side, and generate virtual virtio-net device and register to EP side. A communication date

data?

Quoted 1 line

is diractly

directly?

Quoted 5 lines

transported between virtqueue level
with each other using PCIe embedded DMA controller.

by a limitation of the hardware and Linux EP framework, this function follows a virtio legacy specification.

what exactly is the limitation and why does it force legacy?

Quoted 1,500 lines

This function driver has beed tested on S4 Rcar (r8a779fa-spider) board but just use the PCIe EP framework and depends on the PCIe EDMA.

---
 drivers/pci/endpoint/functions/Kconfig        |  12 +
 drivers/pci/endpoint/functions/Makefile       |   1 +
 .../pci/endpoint/functions/pci-epf-vnet-ep.c  | 343 ++++++++++
 .../pci/endpoint/functions/pci-epf-vnet-rc.c  | 635 ++++++++++++++++++
 drivers/pci/endpoint/functions/pci-epf-vnet.c | 387 +++++++++++
 drivers/pci/endpoint/functions/pci-epf-vnet.h |  62 ++
 6 files changed, 1440 insertions(+)
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet-ep.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet-rc.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet.h
drivers/pci/endpoint/functions/Kconfig +12 -0
index 9fd560886871..f88d8baaf689 100644
@@ -37,3 +37,15 @@ config PCI_EPF_VNTB
 	  between PCI Root Port and PCIe Endpoint.
 
 	  If in doubt, say "N" to disable Endpoint NTB driver.
+
+config PCI_EPF_VNET
+	tristate "PCI Endpoint virtio-net driver"
+	depends on PCI_ENDPOINT
+	select PCI_ENDPOINT_VIRTIO
+	select VHOST_RING
+	select VHOST_IOMEM
+	help
+	  PCIe Endpoint virtio-net function implementation. This module enables to
+	  show the virtio-net as pci device to PCIe Host side, and, another
+	  virtio-net device show to local machine. Those devices can communicate
+	  each other.
drivers/pci/endpoint/functions/Makefile +1 -0
index 5c13001deaba..74cc4c330c62 100644
@@ -6,3 +6,4 @@
 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_VNET)		+= pci-epf-vnet.o pci-epf-vnet-rc.o pci-epf-vnet-ep.o
drivers/pci/endpoint/functions/pci-epf-vnet-ep.c new file +343 -0
new file mode 100644
index 000000000000..93b7e00e8d06
@@ -0,0 +1,343 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Functions work for Endpoint side(local) using EPF framework
+ */
+#include <linux/pci-epc.h>
+#include <linux/virtio_pci.h>
+#include <linux/virtio_net.h>
+#include <linux/virtio_ring.h>
+
+#include "pci-epf-vnet.h"
+
+static inline struct epf_vnet *vdev_to_vnet(struct virtio_device *vdev)
+{
+	return container_of(vdev, struct epf_vnet, ep.vdev);
+}
+
+static void epf_vnet_ep_set_status(struct epf_vnet *vnet, u16 status)
+{
+	vnet->ep.net_config_status |= status;
+}
+
+static void epf_vnet_ep_clear_status(struct epf_vnet *vnet, u16 status)
+{
+	vnet->ep.net_config_status &= ~status;
+}
+
+static void epf_vnet_ep_raise_config_irq(struct epf_vnet *vnet)
+{
+	virtio_config_changed(&vnet->ep.vdev);
+}
+
+void epf_vnet_ep_announce_linkup(struct epf_vnet *vnet)
+{
+	epf_vnet_ep_set_status(vnet,
+			       VIRTIO_NET_S_LINK_UP | VIRTIO_NET_S_ANNOUNCE);
+	epf_vnet_ep_raise_config_irq(vnet);
+}
+
+void epf_vnet_ep_notify(struct epf_vnet *vnet, struct virtqueue *vq)
+{
+	vring_interrupt(0, vq);
+}
+
+static int epf_vnet_ep_process_ctrlq_entry(struct epf_vnet *vnet)
+{
+	struct vringh *vrh = &vnet->ep.ctlvrh;
+	struct vringh_kiov *wiov = &vnet->ep.ctl_riov;
+	struct vringh_kiov *riov = &vnet->ep.ctl_wiov;
+	struct virtio_net_ctrl_hdr *hdr;
+	virtio_net_ctrl_ack *ack;
+	int err;
+	u16 head;
+	size_t len;
+
+	err = vringh_getdesc(vrh, riov, wiov, &head);
+	if (err <= 0)
+		goto done;
+
+	len = vringh_kiov_length(riov);
+	if (len < sizeof(*hdr)) {
+		pr_debug("Command is too short: %ld\n", len);
+		err = -EIO;
+		goto done;
+	}
+
+	if (vringh_kiov_length(wiov) < sizeof(*ack)) {
+		pr_debug("Space for ack is not enough\n");
+		err = -EIO;
+		goto done;
+	}
+
+	hdr = phys_to_virt((unsigned long)riov->iov[riov->i].iov_base);
+	ack = phys_to_virt((unsigned long)wiov->iov[wiov->i].iov_base);
+
+	switch (hdr->class) {
+	case VIRTIO_NET_CTRL_ANNOUNCE:
+		if (hdr->cmd != VIRTIO_NET_CTRL_ANNOUNCE_ACK) {
+			pr_debug("Invalid command: announce: %d\n", hdr->cmd);
+			goto done;
+		}
+
+		epf_vnet_ep_clear_status(vnet, VIRTIO_NET_S_ANNOUNCE);
+		*ack = VIRTIO_NET_OK;
+		break;
+	default:
+		pr_debug("Found not supported class: %d\n", hdr->class);
+		err = -EIO;
+	}
+
+done:
+	vringh_complete(vrh, head, len);
+	return err;
+}
+
+static u64 epf_vnet_ep_vdev_get_features(struct virtio_device *vdev)
+{
+	struct epf_vnet *vnet = vdev_to_vnet(vdev);
+
+	return vnet->virtio_features;
+}
+
+static int epf_vnet_ep_vdev_finalize_features(struct virtio_device *vdev)
+{
+	struct epf_vnet *vnet = vdev_to_vnet(vdev);
+
+	if (vdev->features != vnet->virtio_features)
+		return -EINVAL;
+
+	return 0;
+}
+
+static void epf_vnet_ep_vdev_get_config(struct virtio_device *vdev,
+					unsigned int offset, void *buf,
+					unsigned int len)
+{
+	struct epf_vnet *vnet = vdev_to_vnet(vdev);
+	const unsigned int mac_len = sizeof(vnet->vnet_cfg.mac);
+	const unsigned int status_len = sizeof(vnet->vnet_cfg.status);
+	unsigned int copy_len;
+
+	switch (offset) {
+	case offsetof(struct virtio_net_config, mac):
+		/* This PCIe EP function doesn't provide a VIRTIO_NET_F_MAC feature, so just
+		 * clear the buffer.
+		 */
+		copy_len = len >= mac_len ? mac_len : len;
+		memset(buf, 0x00, copy_len);
+		len -= copy_len;
+		buf += copy_len;
+		fallthrough;
+	case offsetof(struct virtio_net_config, status):
+		copy_len = len >= status_len ? status_len : len;
+		memcpy(buf, &vnet->ep.net_config_status, copy_len);
+		len -= copy_len;
+		buf += copy_len;
+		fallthrough;
+	default:
+		if (offset > sizeof(vnet->vnet_cfg)) {
+			memset(buf, 0x00, len);
+			break;
+		}
+		memcpy(buf, (void *)&vnet->vnet_cfg + offset, len);
+	}
+}
+
+static void epf_vnet_ep_vdev_set_config(struct virtio_device *vdev,
+					unsigned int offset, const void *buf,
+					unsigned int len)
+{
+	/* Do nothing, because all of virtio net config space is readonly. */
+}
+
+static u8 epf_vnet_ep_vdev_get_status(struct virtio_device *vdev)
+{
+	return 0;
+}
+
+static void epf_vnet_ep_vdev_set_status(struct virtio_device *vdev, u8 status)
+{
+	struct epf_vnet *vnet = vdev_to_vnet(vdev);
+
+	if (status & VIRTIO_CONFIG_S_DRIVER_OK)
+		epf_vnet_init_complete(vnet, EPF_VNET_INIT_COMPLETE_EP);
+}
+
+static void epf_vnet_ep_vdev_reset(struct virtio_device *vdev)
+{
+	pr_debug("doesn't support yet");
+}
+
+static bool epf_vnet_ep_vdev_vq_notify(struct virtqueue *vq)
+{
+	struct epf_vnet *vnet = vdev_to_vnet(vq->vdev);
+	struct vringh *tx_vrh = &vnet->ep.txvrh;
+	struct vringh *rx_vrh = &vnet->rc.rxvrh->vrh;
+	struct vringh_kiov *tx_iov = &vnet->ep.tx_iov;
+	struct vringh_kiov *rx_iov = &vnet->rc.rx_iov;
+	int err;
+
+	/* Support only one queue pair */
+	switch (vq->index) {
+	case 0: // rx queue
+		break;
+	case 1: // tx queue
+		while ((err = epf_vnet_transfer(vnet, tx_vrh, rx_vrh, tx_iov,
+						rx_iov, DMA_MEM_TO_DEV)) > 0)
+			;
+		if (err < 0)
+			pr_debug("Failed to transmit: EP -> Host: %d\n", err);
+		break;
+	case 2: // control queue
+		epf_vnet_ep_process_ctrlq_entry(vnet);
+		break;
+	default:
+		return false;
+	}
+
+	return true;
+}
+
+static int epf_vnet_ep_vdev_find_vqs(struct virtio_device *vdev,
+				     unsigned int nvqs, struct virtqueue *vqs[],
+				     vq_callback_t *callback[],
+				     const char *const names[], const bool *ctx,
+				     struct irq_affinity *desc)
+{
+	struct epf_vnet *vnet = vdev_to_vnet(vdev);
+	const size_t vq_size = epf_vnet_get_vq_size();
+	int i;
+	int err;
+	int qidx;
+
+	for (qidx = 0, i = 0; i < nvqs; i++) {
+		struct virtqueue *vq;
+		struct vring *vring;
+		struct vringh *vrh;
+
+		if (!names[i]) {
+			vqs[i] = NULL;
+			continue;
+		}
+
+		vq = vring_create_virtqueue(qidx++, vq_size,
+					    VIRTIO_PCI_VRING_ALIGN, vdev, true,
+					    false, ctx ? ctx[i] : false,
+					    epf_vnet_ep_vdev_vq_notify,
+					    callback[i], names[i]);
+		if (!vq) {
+			err = -ENOMEM;
+			goto err_del_vqs;
+		}
+
+		vqs[i] = vq;
+		vring = virtqueue_get_vring(vq);
+
+		switch (i) {
+		case 0: // rx
+			vrh = &vnet->ep.rxvrh;
+			vnet->ep.rxvq = vq;
+			break;
+		case 1: // tx
+			vrh = &vnet->ep.txvrh;
+			vnet->ep.txvq = vq;
+			break;
+		case 2: // control
+			vrh = &vnet->ep.ctlvrh;
+			vnet->ep.ctlvq = vq;
+			break;
+		default:
+			err = -EIO;
+			goto err_del_vqs;
+		}
+
+		err = vringh_init_kern(vrh, vnet->virtio_features, vq_size,
+				       true, GFP_KERNEL, vring->desc,
+				       vring->avail, vring->used);
+		if (err) {
+			pr_err("failed to init vringh for vring %d\n", i);
+			goto err_del_vqs;
+		}
+	}
+
+	err = epf_vnet_init_kiov(&vnet->ep.tx_iov, vq_size);
+	if (err)
+		goto err_free_kiov;
+	err = epf_vnet_init_kiov(&vnet->ep.rx_iov, vq_size);
+	if (err)
+		goto err_free_kiov;
+	err = epf_vnet_init_kiov(&vnet->ep.ctl_riov, vq_size);
+	if (err)
+		goto err_free_kiov;
+	err = epf_vnet_init_kiov(&vnet->ep.ctl_wiov, vq_size);
+	if (err)
+		goto err_free_kiov;
+
+	return 0;
+
+err_free_kiov:
+	epf_vnet_deinit_kiov(&vnet->ep.tx_iov);
+	epf_vnet_deinit_kiov(&vnet->ep.rx_iov);
+	epf_vnet_deinit_kiov(&vnet->ep.ctl_riov);
+	epf_vnet_deinit_kiov(&vnet->ep.ctl_wiov);
+
+err_del_vqs:
+	for (; i >= 0; i--) {
+		if (!names[i])
+			continue;
+
+		if (!vqs[i])
+			continue;
+
+		vring_del_virtqueue(vqs[i]);
+	}
+	return err;
+}
+
+static void epf_vnet_ep_vdev_del_vqs(struct virtio_device *vdev)
+{
+	struct virtqueue *vq, *n;
+	struct epf_vnet *vnet = vdev_to_vnet(vdev);
+
+	list_for_each_entry_safe(vq, n, &vdev->vqs, list)
+		vring_del_virtqueue(vq);
+
+	epf_vnet_deinit_kiov(&vnet->ep.tx_iov);
+	epf_vnet_deinit_kiov(&vnet->ep.rx_iov);
+	epf_vnet_deinit_kiov(&vnet->ep.ctl_riov);
+	epf_vnet_deinit_kiov(&vnet->ep.ctl_wiov);
+}
+
+static const struct virtio_config_ops epf_vnet_ep_vdev_config_ops = {
+	.get_features = epf_vnet_ep_vdev_get_features,
+	.finalize_features = epf_vnet_ep_vdev_finalize_features,
+	.get = epf_vnet_ep_vdev_get_config,
+	.set = epf_vnet_ep_vdev_set_config,
+	.get_status = epf_vnet_ep_vdev_get_status,
+	.set_status = epf_vnet_ep_vdev_set_status,
+	.reset = epf_vnet_ep_vdev_reset,
+	.find_vqs = epf_vnet_ep_vdev_find_vqs,
+	.del_vqs = epf_vnet_ep_vdev_del_vqs,
+};
+
+void epf_vnet_ep_cleanup(struct epf_vnet *vnet)
+{
+	unregister_virtio_device(&vnet->ep.vdev);
+}
+
+int epf_vnet_ep_setup(struct epf_vnet *vnet)
+{
+	int err;
+	struct virtio_device *vdev = &vnet->ep.vdev;
+
+	vdev->dev.parent = vnet->epf->epc->dev.parent;
+	vdev->config = &epf_vnet_ep_vdev_config_ops;
+	vdev->id.vendor = PCI_VENDOR_ID_REDHAT_QUMRANET;
+	vdev->id.device = VIRTIO_ID_NET;
+
+	err = register_virtio_device(vdev);
+	if (err)
+		return err;
+
+	return 0;
+}
drivers/pci/endpoint/functions/pci-epf-vnet-rc.c new file +635 -0
new file mode 100644
index 000000000000..2ca0245a9134
@@ -0,0 +1,635 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Functions work for PCie Host side(remote) using EPF framework.
+ */
+#include <linux/pci-epf.h>
+#include <linux/pci-epc.h>
+#include <linux/pci_ids.h>
+#include <linux/sched.h>
+#include <linux/virtio_pci.h>
+
+#include "pci-epf-vnet.h"
+
+#define VIRTIO_NET_LEGACY_CFG_BAR BAR_0
+
+/* Returns an out side of the valid queue index. */
+static inline u16 epf_vnet_rc_get_number_of_queues(struct epf_vnet *vnet)
+
+{
+	/* number of queue pairs and control queue */
+	return vnet->vnet_cfg.max_virtqueue_pairs * 2 + 1;
+}
+
+static void epf_vnet_rc_memcpy_config(struct epf_vnet *vnet, size_t offset,
+				      void *buf, size_t len)
+{
+	void __iomem *base = vnet->rc.cfg_base + offset;
+
+	memcpy_toio(base, buf, len);
+}
+
+static void epf_vnet_rc_set_config8(struct epf_vnet *vnet, size_t offset,
+				    u8 config)
+{
+	void __iomem *base = vnet->rc.cfg_base + offset;
+
+	iowrite8(ioread8(base) | config, base);
+}
+
+static void epf_vnet_rc_set_config16(struct epf_vnet *vnet, size_t offset,
+				     u16 config)
+{
+	void __iomem *base = vnet->rc.cfg_base + offset;
+
+	iowrite16(ioread16(base) | config, base);
+}
+
+static void epf_vnet_rc_clear_config16(struct epf_vnet *vnet, size_t offset,
+				       u16 config)
+{
+	void __iomem *base = vnet->rc.cfg_base + offset;
+
+	iowrite16(ioread16(base) & ~config, base);
+}
+
+static void epf_vnet_rc_set_config32(struct epf_vnet *vnet, size_t offset,
+				     u32 config)
+{
+	void __iomem *base = vnet->rc.cfg_base + offset;
+
+	iowrite32(ioread32(base) | config, base);
+}
+
+static void epf_vnet_rc_raise_config_irq(struct epf_vnet *vnet)
+{
+	epf_vnet_rc_set_config16(vnet, VIRTIO_PCI_ISR, VIRTIO_PCI_ISR_CONFIG);
+	queue_work(vnet->rc.irq_wq, &vnet->rc.raise_irq_work);
+}
+
+void epf_vnet_rc_announce_linkup(struct epf_vnet *vnet)
+{
+	epf_vnet_rc_set_config16(vnet,
+				 VIRTIO_PCI_CONFIG_OFF(false) +
+					 offsetof(struct virtio_net_config,
+						  status),
+				 VIRTIO_NET_S_LINK_UP | VIRTIO_NET_S_ANNOUNCE);
+	epf_vnet_rc_raise_config_irq(vnet);
+}
+
+/*
+ * For the PCIe host, this driver shows legacy virtio-net device. Because,
+ * virtio structure pci capabilities is mandatory for modern virtio device,
+ * but there is no PCIe EP hardware that can be configured with any pci
+ * capabilities and Linux PCIe EP framework doesn't support it.
+ */
+static struct pci_epf_header epf_vnet_pci_header = {
+	.vendorid = PCI_VENDOR_ID_REDHAT_QUMRANET,
+	.deviceid = VIRTIO_TRANS_ID_NET,
+	.subsys_vendor_id = PCI_VENDOR_ID_REDHAT_QUMRANET,
+	.subsys_id = VIRTIO_ID_NET,
+	.revid = 0,
+	.baseclass_code = PCI_BASE_CLASS_NETWORK,
+	.interrupt_pin = PCI_INTERRUPT_PIN,
+};
+
+static void epf_vnet_rc_setup_configs(struct epf_vnet *vnet,
+				      void __iomem *cfg_base)
+{
+	u16 default_qindex = epf_vnet_rc_get_number_of_queues(vnet);
+
+	epf_vnet_rc_set_config32(vnet, VIRTIO_PCI_HOST_FEATURES,
+				 vnet->virtio_features);
+
+	epf_vnet_rc_set_config16(vnet, VIRTIO_PCI_ISR, VIRTIO_PCI_ISR_QUEUE);
+	/*
+	 * Initialize the queue notify and selector to outside of the appropriate
+	 * virtqueue index. It is used to detect change with polling. There is no
+	 * other ways to detect host side driver updateing those values
+	 */
+	epf_vnet_rc_set_config16(vnet, VIRTIO_PCI_QUEUE_NOTIFY, default_qindex);
+	epf_vnet_rc_set_config16(vnet, VIRTIO_PCI_QUEUE_SEL, default_qindex);
+	/* This pfn is also set to 0 for the polling as well */
+	epf_vnet_rc_set_config16(vnet, VIRTIO_PCI_QUEUE_PFN, 0);
+
+	epf_vnet_rc_set_config16(vnet, VIRTIO_PCI_QUEUE_NUM,
+				 epf_vnet_get_vq_size());
+	epf_vnet_rc_set_config8(vnet, VIRTIO_PCI_STATUS, 0);
+	epf_vnet_rc_memcpy_config(vnet, VIRTIO_PCI_CONFIG_OFF(false),
+				  &vnet->vnet_cfg, sizeof(vnet->vnet_cfg));
+}
+
+static void epf_vnet_cleanup_bar(struct epf_vnet *vnet)
+{
+	struct pci_epf *epf = vnet->epf;
+
+	pci_epc_clear_bar(epf->epc, epf->func_no, epf->vfunc_no,
+			  &epf->bar[VIRTIO_NET_LEGACY_CFG_BAR]);
+	pci_epf_free_space(epf, vnet->rc.cfg_base, VIRTIO_NET_LEGACY_CFG_BAR,
+			   PRIMARY_INTERFACE);
+}
+
+static int epf_vnet_setup_bar(struct epf_vnet *vnet)
+{
+	int err;
+	size_t cfg_bar_size =
+		VIRTIO_PCI_CONFIG_OFF(false) + sizeof(struct virtio_net_config);
+	struct pci_epf *epf = vnet->epf;
+	const struct pci_epc_features *features;
+	struct pci_epf_bar *config_bar = &epf->bar[VIRTIO_NET_LEGACY_CFG_BAR];
+
+	features = pci_epc_get_features(epf->epc, epf->func_no, epf->vfunc_no);
+	if (!features) {
+		pr_debug("Failed to get PCI EPC features\n");
+		return -EOPNOTSUPP;
+	}
+
+	if (features->reserved_bar & BIT(VIRTIO_NET_LEGACY_CFG_BAR)) {
+		pr_debug("Cannot use the PCI BAR for legacy virtio pci\n");
+		return -EOPNOTSUPP;
+	}
+
+	if (features->bar_fixed_size[VIRTIO_NET_LEGACY_CFG_BAR]) {
+		if (cfg_bar_size >
+		    features->bar_fixed_size[VIRTIO_NET_LEGACY_CFG_BAR]) {
+			pr_debug("PCI BAR size is not enough\n");
+			return -ENOMEM;
+		}
+	}
+
+	config_bar->flags |= PCI_BASE_ADDRESS_MEM_TYPE_64;
+
+	vnet->rc.cfg_base = pci_epf_alloc_space(epf, cfg_bar_size,
+						VIRTIO_NET_LEGACY_CFG_BAR,
+						features->align,
+						PRIMARY_INTERFACE);
+	if (!vnet->rc.cfg_base) {
+		pr_debug("Failed to allocate virtio-net config memory\n");
+		return -ENOMEM;
+	}
+
+	epf_vnet_rc_setup_configs(vnet, vnet->rc.cfg_base);
+
+	err = pci_epc_set_bar(epf->epc, epf->func_no, epf->vfunc_no,
+			      config_bar);
+	if (err) {
+		pr_debug("Failed to set PCI BAR");
+		goto err_free_space;
+	}
+
+	return 0;
+
+err_free_space:
+	pci_epf_free_space(epf, vnet->rc.cfg_base, VIRTIO_NET_LEGACY_CFG_BAR,
+			   PRIMARY_INTERFACE);
+	return err;
+}
+
+static int epf_vnet_rc_negotiate_configs(struct epf_vnet *vnet, u32 *txpfn,
+					 u32 *rxpfn, u32 *ctlpfn)
+{
+	const u16 nqueues = epf_vnet_rc_get_number_of_queues(vnet);
+	const u16 default_sel = nqueues;
+	u32 __iomem *queue_pfn = vnet->rc.cfg_base + VIRTIO_PCI_QUEUE_PFN;
+	u16 __iomem *queue_sel = vnet->rc.cfg_base + VIRTIO_PCI_QUEUE_SEL;
+	u8 __iomem *pci_status = vnet->rc.cfg_base + VIRTIO_PCI_STATUS;
+	u32 pfn;
+	u16 sel;
+	struct {
+		u32 pfn;
+		u16 sel;
+	} tmp[3] = {};
+	int tmp_index = 0;
+
+	*rxpfn = *txpfn = *ctlpfn = 0;
+
+	/* To avoid to miss a getting the pfn and selector for virtqueue wrote by
+	 * host driver, we need to implement fast polling with saving.
+	 *
+	 * This implementation suspects that the host driver writes pfn only once
+	 * for each queues
+	 */
+	while (tmp_index < nqueues) {
+		pfn = ioread32(queue_pfn);
+		if (pfn == 0)
+			continue;
+
+		iowrite32(0, queue_pfn);
+
+		sel = ioread16(queue_sel);
+		if (sel == default_sel)
+			continue;
+
+		tmp[tmp_index].pfn = pfn;
+		tmp[tmp_index].sel = sel;
+		tmp_index++;
+	}
+
+	while (!((ioread8(pci_status) & VIRTIO_CONFIG_S_DRIVER_OK)))
+		;
+
+	for (int i = 0; i < nqueues; ++i) {
+		switch (tmp[i].sel) {
+		case 0:
+			*rxpfn = tmp[i].pfn;
+			break;
+		case 1:
+			*txpfn = tmp[i].pfn;
+			break;
+		case 2:
+			*ctlpfn = tmp[i].pfn;
+			break;
+		}
+	}
+
+	if (!*rxpfn || !*txpfn || !*ctlpfn)
+		return -EIO;
+
+	return 0;
+}
+
+static int epf_vnet_rc_monitor_notify(void *data)
+{
+	struct epf_vnet *vnet = data;
+	u16 __iomem *queue_notify = vnet->rc.cfg_base + VIRTIO_PCI_QUEUE_NOTIFY;
+	const u16 notify_default = epf_vnet_rc_get_number_of_queues(vnet);
+
+	epf_vnet_init_complete(vnet, EPF_VNET_INIT_COMPLETE_RC);
+
+	/* Poll to detect a change of the queue_notify register. Sometimes this
+	 * polling misses the change, so try to check each virtqueues
+	 * everytime.
+	 */
+	while (true) {
+		while (ioread16(queue_notify) == notify_default)
+			;
+		iowrite16(notify_default, queue_notify);
+
+		queue_work(vnet->rc.tx_wq, &vnet->rc.tx_work);
+		queue_work(vnet->rc.ctl_wq, &vnet->rc.ctl_work);
+	}
+
+	return 0;
+}
+
+static int epf_vnet_rc_spawn_notify_monitor(struct epf_vnet *vnet)
+{
+	vnet->rc.notify_monitor_task =
+		kthread_create(epf_vnet_rc_monitor_notify, vnet,
+			       "pci-epf-vnet/cfg_negotiator");
+	if (IS_ERR(vnet->rc.notify_monitor_task))
+		return PTR_ERR(vnet->rc.notify_monitor_task);
+
+	/* Change the thread priority to high for polling. */
+	sched_set_fifo(vnet->rc.notify_monitor_task);
+	wake_up_process(vnet->rc.notify_monitor_task);
+
+	return 0;
+}
+
+static int epf_vnet_rc_device_setup(void *data)
+{
+	struct epf_vnet *vnet = data;
+	struct pci_epf *epf = vnet->epf;
+	u32 txpfn, rxpfn, ctlpfn;
+	const size_t vq_size = epf_vnet_get_vq_size();
+	int err;
+
+	err = epf_vnet_rc_negotiate_configs(vnet, &txpfn, &rxpfn, &ctlpfn);
+	if (err) {
+		pr_debug("Failed to negatiate configs with driver\n");
+		return err;
+	}
+
+	/* Polling phase is finished. This thread backs to normal priority. */
+	sched_set_normal(vnet->rc.device_setup_task, 19);
+
+	vnet->rc.txvrh = pci_epf_virtio_alloc_vringh(epf, vnet->virtio_features,
+						     txpfn, vq_size);
+	if (IS_ERR(vnet->rc.txvrh)) {
+		pr_debug("Failed to setup virtqueue for tx\n");
+		return PTR_ERR(vnet->rc.txvrh);
+	}
+
+	err = epf_vnet_init_kiov(&vnet->rc.tx_iov, vq_size);
+	if (err)
+		goto err_free_epf_tx_vringh;
+
+	vnet->rc.rxvrh = pci_epf_virtio_alloc_vringh(epf, vnet->virtio_features,
+						     rxpfn, vq_size);
+	if (IS_ERR(vnet->rc.rxvrh)) {
+		pr_debug("Failed to setup virtqueue for rx\n");
+		err = PTR_ERR(vnet->rc.rxvrh);
+		goto err_deinit_tx_kiov;
+	}
+
+	err = epf_vnet_init_kiov(&vnet->rc.rx_iov, vq_size);
+	if (err)
+		goto err_free_epf_rx_vringh;
+
+	vnet->rc.ctlvrh = pci_epf_virtio_alloc_vringh(
+		epf, vnet->virtio_features, ctlpfn, vq_size);
+	if (IS_ERR(vnet->rc.ctlvrh)) {
+		pr_err("failed to setup virtqueue\n");
+		err = PTR_ERR(vnet->rc.ctlvrh);
+		goto err_deinit_rx_kiov;
+	}
+
+	err = epf_vnet_init_kiov(&vnet->rc.ctl_riov, vq_size);
+	if (err)
+		goto err_free_epf_ctl_vringh;
+
+	err = epf_vnet_init_kiov(&vnet->rc.ctl_wiov, vq_size);
+	if (err)
+		goto err_deinit_ctl_riov;
+
+	err = epf_vnet_rc_spawn_notify_monitor(vnet);
+	if (err) {
+		pr_debug("Failed to create notify monitor thread\n");
+		goto err_deinit_ctl_wiov;
+	}
+
+	return 0;
+
+err_deinit_ctl_wiov:
+	epf_vnet_deinit_kiov(&vnet->rc.ctl_wiov);
+err_deinit_ctl_riov:
+	epf_vnet_deinit_kiov(&vnet->rc.ctl_riov);
+err_free_epf_ctl_vringh:
+	pci_epf_virtio_free_vringh(epf, vnet->rc.ctlvrh);
+err_deinit_rx_kiov:
+	epf_vnet_deinit_kiov(&vnet->rc.rx_iov);
+err_free_epf_rx_vringh:
+	pci_epf_virtio_free_vringh(epf, vnet->rc.rxvrh);
+err_deinit_tx_kiov:
+	epf_vnet_deinit_kiov(&vnet->rc.tx_iov);
+err_free_epf_tx_vringh:
+	pci_epf_virtio_free_vringh(epf, vnet->rc.txvrh);
+
+	return err;
+}
+
+static int epf_vnet_rc_spawn_device_setup_task(struct epf_vnet *vnet)
+{
+	vnet->rc.device_setup_task = kthread_create(
+		epf_vnet_rc_device_setup, vnet, "pci-epf-vnet/cfg_negotiator");
+	if (IS_ERR(vnet->rc.device_setup_task))
+		return PTR_ERR(vnet->rc.device_setup_task);
+
+	/* Change the thread priority to high for the polling. */
+	sched_set_fifo(vnet->rc.device_setup_task);
+	wake_up_process(vnet->rc.device_setup_task);
+
+	return 0;
+}
+
+static void epf_vnet_rc_tx_handler(struct work_struct *work)
+{
+	struct epf_vnet *vnet = container_of(work, struct epf_vnet, rc.tx_work);
+	struct vringh *tx_vrh = &vnet->rc.txvrh->vrh;
+	struct vringh *rx_vrh = &vnet->ep.rxvrh;
+	struct vringh_kiov *tx_iov = &vnet->rc.tx_iov;
+	struct vringh_kiov *rx_iov = &vnet->ep.rx_iov;
+
+	while (epf_vnet_transfer(vnet, tx_vrh, rx_vrh, tx_iov, rx_iov,
+				 DMA_DEV_TO_MEM) > 0)
+		;
+}
+
+static void epf_vnet_rc_raise_irq_handler(struct work_struct *work)
+{
+	struct epf_vnet *vnet =
+		container_of(work, struct epf_vnet, rc.raise_irq_work);
+	struct pci_epf *epf = vnet->epf;
+
+	pci_epc_raise_irq(epf->epc, epf->func_no, epf->vfunc_no,
+			  PCI_EPC_IRQ_LEGACY, 0);
+}
+
+struct epf_vnet_rc_meminfo {
+	void __iomem *addr, *virt;
+	phys_addr_t phys;
+	size_t len;
+};
+
+/* Util function to access PCIe host side memory from local CPU.  */
+static struct epf_vnet_rc_meminfo *
+epf_vnet_rc_epc_mmap(struct pci_epf *epf, phys_addr_t pci_addr, size_t len)
+{
+	int err;
+	phys_addr_t aaddr, phys_addr;
+	size_t asize, offset;
+	void __iomem *virt_addr;
+	struct epf_vnet_rc_meminfo *meminfo;
+
+	err = pci_epc_mem_align(epf->epc, pci_addr, len, &aaddr, &asize);
+	if (err) {
+		pr_debug("Failed to get EPC align: %d\n", err);
+		return NULL;
+	}
+
+	offset = pci_addr - aaddr;
+
+	virt_addr = pci_epc_mem_alloc_addr(epf->epc, &phys_addr, asize);
+	if (!virt_addr) {
+		pr_debug("Failed to allocate epc memory\n");
+		return NULL;
+	}
+
+	err = pci_epc_map_addr(epf->epc, epf->func_no, epf->vfunc_no, phys_addr,
+			       aaddr, asize);
+	if (err) {
+		pr_debug("Failed to map epc memory\n");
+		goto err_epc_free_addr;
+	}
+
+	meminfo = kmalloc(sizeof(*meminfo), GFP_KERNEL);
+	if (!meminfo)
+		goto err_epc_unmap_addr;
+
+	meminfo->virt = virt_addr;
+	meminfo->phys = phys_addr;
+	meminfo->len = len;
+	meminfo->addr = virt_addr + offset;
+
+	return meminfo;
+
+err_epc_unmap_addr:
+	pci_epc_unmap_addr(epf->epc, epf->func_no, epf->vfunc_no,
+			   meminfo->phys);
+err_epc_free_addr:
+	pci_epc_mem_free_addr(epf->epc, meminfo->phys, meminfo->virt,
+			      meminfo->len);
+
+	return NULL;
+}
+
+static void epf_vnet_rc_epc_munmap(struct pci_epf *epf,
+				   struct epf_vnet_rc_meminfo *meminfo)
+{
+	pci_epc_unmap_addr(epf->epc, epf->func_no, epf->vfunc_no,
+			   meminfo->phys);
+	pci_epc_mem_free_addr(epf->epc, meminfo->phys, meminfo->virt,
+			      meminfo->len);
+	kfree(meminfo);
+}
+
+static int epf_vnet_rc_process_ctrlq_entry(struct epf_vnet *vnet)
+{
+	struct vringh_kiov *riov = &vnet->rc.ctl_riov;
+	struct vringh_kiov *wiov = &vnet->rc.ctl_wiov;
+	struct vringh *vrh = &vnet->rc.ctlvrh->vrh;
+	struct pci_epf *epf = vnet->epf;
+	struct epf_vnet_rc_meminfo *rmem, *wmem;
+	struct virtio_net_ctrl_hdr *hdr;
+	int err;
+	u16 head;
+	size_t total_len;
+	u8 class, cmd;
+
+	err = vringh_getdesc(vrh, riov, wiov, &head);
+	if (err <= 0)
+		return err;
+
+	total_len = vringh_kiov_length(riov);
+
+	rmem = epf_vnet_rc_epc_mmap(epf, (u64)riov->iov[riov->i].iov_base,
+				    riov->iov[riov->i].iov_len);
+	if (!rmem) {
+		err = -ENOMEM;
+		goto err_abandon_descs;
+	}
+
+	wmem = epf_vnet_rc_epc_mmap(epf, (u64)wiov->iov[wiov->i].iov_base,
+				    wiov->iov[wiov->i].iov_len);
+	if (!wmem) {
+		err = -ENOMEM;
+		goto err_epc_unmap_rmem;
+	}
+
+	hdr = rmem->addr;
+	class = ioread8(&hdr->class);
+	cmd = ioread8(&hdr->cmd);
+	switch (ioread8(&hdr->class)) {
+	case VIRTIO_NET_CTRL_ANNOUNCE:
+		if (cmd != VIRTIO_NET_CTRL_ANNOUNCE_ACK) {
+			pr_err("Found invalid command: announce: %d\n", cmd);
+			break;
+		}
+		epf_vnet_rc_clear_config16(
+			vnet,
+			VIRTIO_PCI_CONFIG_OFF(false) +
+				offsetof(struct virtio_net_config, status),
+			VIRTIO_NET_S_ANNOUNCE);
+		epf_vnet_rc_clear_config16(vnet, VIRTIO_PCI_ISR,
+					   VIRTIO_PCI_ISR_CONFIG);
+
+		iowrite8(VIRTIO_NET_OK, wmem->addr);
+		break;
+	default:
+		pr_err("Found unsupported class in control queue: %d\n", class);
+		break;
+	}
+
+	epf_vnet_rc_epc_munmap(epf, rmem);
+	epf_vnet_rc_epc_munmap(epf, wmem);
+	vringh_complete(vrh, head, total_len);
+
+	return 1;
+
+err_epc_unmap_rmem:
+	epf_vnet_rc_epc_munmap(epf, rmem);
+err_abandon_descs:
+	vringh_abandon(vrh, head);
+
+	return err;
+}
+
+static void epf_vnet_rc_process_ctrlq_entries(struct work_struct *work)
+{
+	struct epf_vnet *vnet =
+		container_of(work, struct epf_vnet, rc.ctl_work);
+
+	while (epf_vnet_rc_process_ctrlq_entry(vnet) > 0)
+		;
+}
+
+void epf_vnet_rc_notify(struct epf_vnet *vnet)
+{
+	queue_work(vnet->rc.irq_wq, &vnet->rc.raise_irq_work);
+}
+
+void epf_vnet_rc_cleanup(struct epf_vnet *vnet)
+{
+	epf_vnet_cleanup_bar(vnet);
+	destroy_workqueue(vnet->rc.tx_wq);
+	destroy_workqueue(vnet->rc.irq_wq);
+	destroy_workqueue(vnet->rc.ctl_wq);
+
+	kthread_stop(vnet->rc.device_setup_task);
+}
+
+int epf_vnet_rc_setup(struct epf_vnet *vnet)
+{
+	int err;
+	struct pci_epf *epf = vnet->epf;
+
+	err = pci_epc_write_header(epf->epc, epf->func_no, epf->vfunc_no,
+				   &epf_vnet_pci_header);
+	if (err)
+		return err;
+
+	err = epf_vnet_setup_bar(vnet);
+	if (err)
+		return err;
+
+	vnet->rc.tx_wq =
+		alloc_workqueue("pci-epf-vnet/tx-wq",
+				WQ_MEM_RECLAIM | WQ_HIGHPRI | WQ_UNBOUND, 0);
+	if (!vnet->rc.tx_wq) {
+		pr_debug(
+			"Failed to allocate workqueue for rc -> ep transmission\n");
+		err = -ENOMEM;
+		goto err_cleanup_bar;
+	}
+
+	INIT_WORK(&vnet->rc.tx_work, epf_vnet_rc_tx_handler);
+
+	vnet->rc.irq_wq =
+		alloc_workqueue("pci-epf-vnet/irq-wq",
+				WQ_MEM_RECLAIM | WQ_HIGHPRI | WQ_UNBOUND, 0);
+	if (!vnet->rc.irq_wq) {
+		pr_debug("Failed to allocate workqueue for irq\n");
+		err = -ENOMEM;
+		goto err_destory_tx_wq;
+	}
+
+	INIT_WORK(&vnet->rc.raise_irq_work, epf_vnet_rc_raise_irq_handler);
+
+	vnet->rc.ctl_wq =
+		alloc_workqueue("pci-epf-vnet/ctl-wq",
+				WQ_MEM_RECLAIM | WQ_HIGHPRI | WQ_UNBOUND, 0);
+	if (!vnet->rc.ctl_wq) {
+		pr_err("Failed to allocate work queue for control queue processing\n");
+		err = -ENOMEM;
+		goto err_destory_irq_wq;
+	}
+
+	INIT_WORK(&vnet->rc.ctl_work, epf_vnet_rc_process_ctrlq_entries);
+
+	err = epf_vnet_rc_spawn_device_setup_task(vnet);
+	if (err)
+		goto err_destory_ctl_wq;
+
+	return 0;
+
+err_cleanup_bar:
+	epf_vnet_cleanup_bar(vnet);
+err_destory_tx_wq:
+	destroy_workqueue(vnet->rc.tx_wq);
+err_destory_irq_wq:
+	destroy_workqueue(vnet->rc.irq_wq);
+err_destory_ctl_wq:
+	destroy_workqueue(vnet->rc.ctl_wq);
+
+	return err;
+}
drivers/pci/endpoint/functions/pci-epf-vnet.c new file +387 -0
new file mode 100644
index 000000000000..e48ad8067796
@@ -0,0 +1,387 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * PCI Endpoint function driver to impliment virtio-net device.
+ */
+#include <linux/module.h>
+#include <linux/pci-epf.h>
+#include <linux/pci-epc.h>
+#include <linux/vringh.h>
+#include <linux/dmaengine.h>
+
+#include "pci-epf-vnet.h"
+
+static int virtio_queue_size = 0x100;
+module_param(virtio_queue_size, int, 0444);
+MODULE_PARM_DESC(virtio_queue_size, "A length of virtqueue");
+
+int epf_vnet_get_vq_size(void)
+{
+	return virtio_queue_size;
+}
+
+int epf_vnet_init_kiov(struct vringh_kiov *kiov, const size_t vq_size)
+{
+	struct kvec *kvec;
+
+	kvec = kmalloc_array(vq_size, sizeof(*kvec), GFP_KERNEL);
+	if (!kvec)
+		return -ENOMEM;
+
+	vringh_kiov_init(kiov, kvec, vq_size);
+
+	return 0;
+}
+
+void epf_vnet_deinit_kiov(struct vringh_kiov *kiov)
+{
+	kfree(kiov->iov);
+}
+
+void epf_vnet_init_complete(struct epf_vnet *vnet, u8 from)
+{
+	vnet->init_complete |= from;
+
+	if (!(vnet->init_complete & EPF_VNET_INIT_COMPLETE_EP))
+		return;
+
+	if (!(vnet->init_complete & EPF_VNET_INIT_COMPLETE_RC))
+		return;
+
+	epf_vnet_ep_announce_linkup(vnet);
+	epf_vnet_rc_announce_linkup(vnet);
+}
+
+struct epf_dma_filter_param {
+	struct device *dev;
+	u32 dma_mask;
+};
+
+static bool epf_virtnet_dma_filter(struct dma_chan *chan, void *param)
+{
+	struct epf_dma_filter_param *fparam = param;
+	struct dma_slave_caps caps;
+
+	memset(&caps, 0, sizeof(caps));
+	dma_get_slave_caps(chan, &caps);
+
+	return chan->device->dev == fparam->dev &&
+	       (fparam->dma_mask & caps.directions);
+}
+
+static int epf_vnet_init_edma(struct epf_vnet *vnet, struct device *dma_dev)
+{
+	struct epf_dma_filter_param param;
+	dma_cap_mask_t mask;
+	int err;
+
+	dma_cap_zero(mask);
+	dma_cap_set(DMA_SLAVE, mask);
+
+	param.dev = dma_dev;
+	param.dma_mask = BIT(DMA_MEM_TO_DEV);
+	vnet->lr_dma_chan =
+		dma_request_channel(mask, epf_virtnet_dma_filter, &param);
+	if (!vnet->lr_dma_chan)
+		return -EOPNOTSUPP;
+
+	param.dma_mask = BIT(DMA_DEV_TO_MEM);
+	vnet->rl_dma_chan =
+		dma_request_channel(mask, epf_virtnet_dma_filter, &param);
+	if (!vnet->rl_dma_chan) {
+		err = -EOPNOTSUPP;
+		goto err_release_channel;
+	}
+
+	return 0;
+
+err_release_channel:
+	dma_release_channel(vnet->lr_dma_chan);
+
+	return err;
+}
+
+static void epf_vnet_deinit_edma(struct epf_vnet *vnet)
+{
+	dma_release_channel(vnet->lr_dma_chan);
+	dma_release_channel(vnet->rl_dma_chan);
+}
+
+static int epf_vnet_dma_single(struct epf_vnet *vnet, phys_addr_t pci,
+			       dma_addr_t dma, size_t len,
+			       void (*callback)(void *), void *param,
+			       enum dma_transfer_direction dir)
+{
+	struct dma_async_tx_descriptor *desc;
+	int err;
+	struct dma_chan *chan;
+	struct dma_slave_config sconf;
+	dma_cookie_t cookie;
+	unsigned long flags = 0;
+
+	if (dir == DMA_MEM_TO_DEV) {
+		sconf.dst_addr = pci;
+		chan = vnet->lr_dma_chan;
+	} else {
+		sconf.src_addr = pci;
+		chan = vnet->rl_dma_chan;
+	}
+
+	err = dmaengine_slave_config(chan, &sconf);
+	if (unlikely(err))
+		return err;
+
+	if (callback)
+		flags = DMA_PREP_INTERRUPT | DMA_PREP_FENCE;
+
+	desc = dmaengine_prep_slave_single(chan, dma, len, dir, flags);
+	if (unlikely(!desc))
+		return -EIO;
+
+	desc->callback = callback;
+	desc->callback_param = param;
+
+	cookie = dmaengine_submit(desc);
+	err = dma_submit_error(cookie);
+	if (unlikely(err))
+		return err;
+
+	dma_async_issue_pending(chan);
+
+	return 0;
+}
+
+struct epf_vnet_dma_callback_param {
+	struct epf_vnet *vnet;
+	struct vringh *tx_vrh, *rx_vrh;
+	struct virtqueue *vq;
+	size_t total_len;
+	u16 tx_head, rx_head;
+};
+
+static void epf_vnet_dma_callback(void *p)
+{
+	struct epf_vnet_dma_callback_param *param = p;
+	struct epf_vnet *vnet = param->vnet;
+
+	vringh_complete(param->tx_vrh, param->tx_head, param->total_len);
+	vringh_complete(param->rx_vrh, param->rx_head, param->total_len);
+
+	epf_vnet_rc_notify(vnet);
+	epf_vnet_ep_notify(vnet, param->vq);
+
+	kfree(param);
+}
+
+/**
+ * epf_vnet_transfer() - transfer data between tx vring to rx vring using edma
+ * @vnet: epf virtio net device to do dma
+ * @tx_vrh: vringh related to source tx vring
+ * @rx_vrh: vringh related to target rx vring
+ * @tx_iov: buffer to use tx
+ * @rx_iov: buffer to use rx
+ * @dir: a direction of DMA. local to remote or local from remote
+ *
+ * This function returns 0, 1 or error number. The 0 indicates there is not
+ * data to send. The 1 indicates a request to DMA is succeeded. Other error
+ * numbers shows error, however, ENOSPC means there is no buffer on target
+ * vring, so should retry to call later.
+ */
+int epf_vnet_transfer(struct epf_vnet *vnet, struct vringh *tx_vrh,
+		      struct vringh *rx_vrh, struct vringh_kiov *tx_iov,
+		      struct vringh_kiov *rx_iov,
+		      enum dma_transfer_direction dir)
+{
+	int err;
+	u16 tx_head, rx_head;
+	size_t total_tx_len;
+	struct epf_vnet_dma_callback_param *cb_param;
+	struct vringh_kiov *liov, *riov;
+
+	err = vringh_getdesc(tx_vrh, tx_iov, NULL, &tx_head);
+	if (err <= 0)
+		return err;
+
+	total_tx_len = vringh_kiov_length(tx_iov);
+
+	err = vringh_getdesc(rx_vrh, NULL, rx_iov, &rx_head);
+	if (err < 0) {
+		goto err_tx_complete;
+	} else if (!err) {
+		/* There is not space on a vring of destination to transmit data, so
+		 * rollback tx vringh
+		 */
+		vringh_abandon(tx_vrh, tx_head);
+		return -ENOSPC;
+	}
+
+	cb_param = kmalloc(sizeof(*cb_param), GFP_KERNEL);
+	if (!cb_param) {
+		err = -ENOMEM;
+		goto err_rx_complete;
+	}
+
+	cb_param->tx_vrh = tx_vrh;
+	cb_param->rx_vrh = rx_vrh;
+	cb_param->tx_head = tx_head;
+	cb_param->rx_head = rx_head;
+	cb_param->total_len = total_tx_len;
+	cb_param->vnet = vnet;
+
+	switch (dir) {
+	case DMA_MEM_TO_DEV:
+		liov = tx_iov;
+		riov = rx_iov;
+		cb_param->vq = vnet->ep.txvq;
+		break;
+	case DMA_DEV_TO_MEM:
+		liov = rx_iov;
+		riov = tx_iov;
+		cb_param->vq = vnet->ep.rxvq;
+		break;
+	default:
+		err = -EINVAL;
+		goto err_free_param;
+	}
+
+	for (; tx_iov->i < tx_iov->used; tx_iov->i++, rx_iov->i++) {
+		size_t len;
+		u64 lbase, rbase;
+		void (*callback)(void *) = NULL;
+
+		lbase = (u64)liov->iov[liov->i].iov_base;
+		rbase = (u64)riov->iov[riov->i].iov_base;
+		len = tx_iov->iov[tx_iov->i].iov_len;
+
+		if (tx_iov->i + 1 == tx_iov->used)
+			callback = epf_vnet_dma_callback;
+
+		err = epf_vnet_dma_single(vnet, rbase, lbase, len, callback,
+					  cb_param, dir);
+		if (err)
+			goto err_free_param;
+	}
+
+	return 1;
+
+err_free_param:
+	kfree(cb_param);
+err_rx_complete:
+	vringh_complete(rx_vrh, rx_head, vringh_kiov_length(rx_iov));
+err_tx_complete:
+	vringh_complete(tx_vrh, tx_head, total_tx_len);
+
+	return err;
+}
+
+static int epf_vnet_bind(struct pci_epf *epf)
+{
+	int err;
+	struct epf_vnet *vnet = epf_get_drvdata(epf);
+
+	err = epf_vnet_init_edma(vnet, epf->epc->dev.parent);
+	if (err)
+		return err;
+
+	err = epf_vnet_rc_setup(vnet);
+	if (err)
+		goto err_free_edma;
+
+	err = epf_vnet_ep_setup(vnet);
+	if (err)
+		goto err_cleanup_rc;
+
+	return 0;
+
+err_free_edma:
+	epf_vnet_deinit_edma(vnet);
+err_cleanup_rc:
+	epf_vnet_rc_cleanup(vnet);
+
+	return err;
+}
+
+static void epf_vnet_unbind(struct pci_epf *epf)
+{
+	struct epf_vnet *vnet = epf_get_drvdata(epf);
+
+	epf_vnet_deinit_edma(vnet);
+	epf_vnet_rc_cleanup(vnet);
+	epf_vnet_ep_cleanup(vnet);
+}
+
+static struct pci_epf_ops epf_vnet_ops = {
+	.bind = epf_vnet_bind,
+	.unbind = epf_vnet_unbind,
+};
+
+static const struct pci_epf_device_id epf_vnet_ids[] = {
+	{ .name = "pci_epf_vnet" },
+	{}
+};
+
+static void epf_vnet_virtio_init(struct epf_vnet *vnet)
+{
+	vnet->virtio_features =
+		BIT(VIRTIO_NET_F_MTU) | BIT(VIRTIO_NET_F_STATUS) |
+		/* Following features are to skip any of checking and offloading, Like a
+		 * transmission between virtual machines on same system. Details are on
+		 * section 5.1.5 in virtio specification.
+		 */
+		BIT(VIRTIO_NET_F_GUEST_CSUM) | BIT(VIRTIO_NET_F_GUEST_TSO4) |
+		BIT(VIRTIO_NET_F_GUEST_TSO6) | BIT(VIRTIO_NET_F_GUEST_ECN) |
+		BIT(VIRTIO_NET_F_GUEST_UFO) |
+		// The control queue is just used for linkup announcement.
+		BIT(VIRTIO_NET_F_CTRL_VQ);
+
+	vnet->vnet_cfg.max_virtqueue_pairs = 1;
+	vnet->vnet_cfg.status = 0;
+	vnet->vnet_cfg.mtu = PAGE_SIZE;
+}
+
+static int epf_vnet_probe(struct pci_epf *epf)
+{
+	struct epf_vnet *vnet;
+
+	vnet = devm_kzalloc(&epf->dev, sizeof(*vnet), GFP_KERNEL);
+	if (!vnet)
+		return -ENOMEM;
+
+	epf_set_drvdata(epf, vnet);
+	vnet->epf = epf;
+
+	epf_vnet_virtio_init(vnet);
+
+	return 0;
+}
+
+static struct pci_epf_driver epf_vnet_drv = {
+	.driver.name = "pci_epf_vnet",
+	.ops = &epf_vnet_ops,
+	.id_table = epf_vnet_ids,
+	.probe = epf_vnet_probe,
+	.owner = THIS_MODULE,
+};
+
+static int __init epf_vnet_init(void)
+{
+	int err;
+
+	err = pci_epf_register_driver(&epf_vnet_drv);
+	if (err) {
+		pr_err("Failed to register epf vnet driver\n");
+		return err;
+	}
+
+	return 0;
+}
+module_init(epf_vnet_init);
+
+static void epf_vnet_exit(void)
+{
+	pci_epf_unregister_driver(&epf_vnet_drv);
+}
+module_exit(epf_vnet_exit);
+
+MODULE_LICENSE("GPL");
+MODULE_AUTHOR("Shunsuke Mie <[email protected]>");
+MODULE_DESCRIPTION("PCI endpoint function acts as virtio net device");
drivers/pci/endpoint/functions/pci-epf-vnet.h new file +62 -0
new file mode 100644
index 000000000000..1e0f90c95578
@@ -0,0 +1,62 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _PCI_EPF_VNET_H
+#define _PCI_EPF_VNET_H
+
+#include <linux/pci-epf.h>
+#include <linux/pci-epf-virtio.h>
+#include <linux/virtio_net.h>
+#include <linux/dmaengine.h>
+#include <linux/virtio.h>
+
+struct epf_vnet {
+	//TODO Should this variable be placed here?
+	struct pci_epf *epf;
+	struct virtio_net_config vnet_cfg;
+	u64 virtio_features;
+
+	// dma channels for local to remote(lr) and remote to local(rl)
+	struct dma_chan *lr_dma_chan, *rl_dma_chan;
+
+	struct {
+		void __iomem *cfg_base;
+		struct task_struct *device_setup_task;
+		struct task_struct *notify_monitor_task;
+		struct workqueue_struct *tx_wq, *irq_wq, *ctl_wq;
+		struct work_struct tx_work, raise_irq_work, ctl_work;
+		struct pci_epf_vringh *txvrh, *rxvrh, *ctlvrh;
+		struct vringh_kiov tx_iov, rx_iov, ctl_riov, ctl_wiov;
+	} rc;
+
+	struct {
+		struct virtqueue *rxvq, *txvq, *ctlvq;
+		struct vringh txvrh, rxvrh, ctlvrh;
+		struct vringh_kiov tx_iov, rx_iov, ctl_riov, ctl_wiov;
+		struct virtio_device vdev;
+		u16 net_config_status;
+	} ep;
+
+#define EPF_VNET_INIT_COMPLETE_EP BIT(0)
+#define EPF_VNET_INIT_COMPLETE_RC BIT(1)
+	u8 init_complete;
+};
+
+int epf_vnet_rc_setup(struct epf_vnet *vnet);
+void epf_vnet_rc_cleanup(struct epf_vnet *vnet);
+int epf_vnet_ep_setup(struct epf_vnet *vnet);
+void epf_vnet_ep_cleanup(struct epf_vnet *vnet);
+
+int epf_vnet_get_vq_size(void);
+int epf_vnet_init_kiov(struct vringh_kiov *kiov, const size_t vq_size);
+void epf_vnet_deinit_kiov(struct vringh_kiov *kiov);
+int epf_vnet_transfer(struct epf_vnet *vnet, struct vringh *tx_vrh,
+		      struct vringh *rx_vrh, struct vringh_kiov *tx_iov,
+		      struct vringh_kiov *rx_iov,
+		      enum dma_transfer_direction dir);
+void epf_vnet_rc_notify(struct epf_vnet *vnet);
+void epf_vnet_ep_notify(struct epf_vnet *vnet, struct virtqueue *vq);
+
+void epf_vnet_init_complete(struct epf_vnet *vnet, u8 from);
+void epf_vnet_ep_announce_linkup(struct epf_vnet *vnet);
+void epf_vnet_rc_announce_linkup(struct epf_vnet *vnet);
+
+#endif // _PCI_EPF_VNET_H
Signature · 1 line
2.25.1

_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

#10 lore raw

RE: [EXT] Re: [RFC PATCH 4/4] PCI: endpoint: function: Add EP function driver to provide virtio net device

Quoted 36 lines

Caution: EXT Email

Quoted 22 lines · Shunsuke Mie

Add a new endpoint(EP) function driver to provide virtio-net device. This function not only shows virtio-net device for PCIe host system, but also provides virtio-net device to EP side(local) system. Virtualy those network devices are connected, so we can use to communicate over IP like a simple NIC.

Architecture overview is following:

to Host       |                       to Endpoint
network stack |                 network stack
      |       |                       |
+-----------+ |       +-----------+   +-----------+
|virtio-net | |       |virtio-net |   |virtio-net |
|driver     | |       |EP function|---|driver     |
+-----------+ |       +-----------+   +-----------+
      |       |             |
+-----------+ | +-----------+
|PCIeC      | | |PCIeC      |
|Rootcomplex|-|-|Endpoint   |
+-----------+ | +-----------+
  Host side   |          Endpoint side
Quoted 3 lines

This driver uses PCIe EP framework to show virtio-net (pci) device Host side, and generate virtual virtio-net device and register to EP side. A communication date

data?

Quoted 1 line

is diractly

directly?

Quoted 17 lines
Quoted 5 lines

transported between virtqueue level
with each other using PCIe embedded DMA controller.

by a limitation of the hardware and Linux EP framework, this function follows a virtio legacy specification.

what exactly is the limitation and why does it force legacy?

Quoted 9 lines

This function driver has beed tested on S4 Rcar (r8a779fa-spider) board but just use the PCIe EP framework and depends on the PCIe EDMA.

---
 drivers/pci/endpoint/functions/Kconfig        |  12 +
 drivers/pci/endpoint/functions/Makefile       |   1 +
 .../pci/endpoint/functions/pci-epf-vnet-ep.c  | 343 ++++++++++

It is actually that not related vnet. Just virtio. I think pci-epf-virtio.c is better.

Quoted 1 line
Quoted 1 line
 .../pci/endpoint/functions/pci-epf-vnet-rc.c  | 635 ++++++++++++++++++

It is epf driver. rc is quite confused.
Maybe you can combine pci-epf-vnet-ep.c and pci-epf-vnet-rc.c to one file.

Quoted 1 line
Quoted 1 line
 drivers/pci/endpoint/functions/pci-epf-vnet.c | 387 +++++++++++

This file setup dma transfer according virtio-ring. How about pci-epf-virtio-dma.c ?

Quoted 8 lines

Review

Quoted 2 lines
(quoted diff) +2 -0
+
+     epf_vnet_rc_set_config16(vnet, VIRTIO_PCI_ISR,

VIRTIO_PCI_ISR_QUEUE);

Quoted 5 lines
drivers/pci/endpoint/functions/pci-epf-vnet-rc.c new file +5 -0
+     /*
+      * Initialize the queue notify and selector to outside of the appropriate
+      * virtqueue index. It is used to detect change with polling. There is no
+      * other ways to detect host side driver updateing those values
+      */

I am try to use gic-its or other msi controller as doorbell.
https://lore.kernel.org/imx/[email protected]/T/#u

but it may need update host side pci virtio driver.

Quoted 9 lines
Quoted 1 line

+ epf_vnet_rc_set_config16(vnet, VIRTIO_PCI_QUEUE_NOTIFY,

default_qindex);

Quoted 1 line

+ epf_vnet_rc_set_config16(vnet, VIRTIO_PCI_QUEUE_SEL,

default_qindex);

drivers/pci/endpoint/functions/pci-epf-vnet-rc.c line 113

Quoted 3 lines
drivers/pci/endpoint/functions/pci-epf-vnet-rc.c new file +3 -0
+     /* This pfn is also set to 0 for the polling as well */
+     epf_vnet_rc_set_config16(vnet, VIRTIO_PCI_QUEUE_PFN, 0);
+

--

Quoted 1 line

2.25.1

#11 lore raw

Re: [EXT] Re: [RFC PATCH 4/4] PCI: endpoint: function: Add EP function driver to provide virtio net device

2023年2月4日(土) 7:15 Frank Li <[email protected]>:

Quoted 55 lines
Quoted 37 lines

Caution: EXT Email

Quoted 22 lines · Shunsuke Mie

Add a new endpoint(EP) function driver to provide virtio-net device. This function not only shows virtio-net device for PCIe host system, but also provides virtio-net device to EP side(local) system. Virtualy those network devices are connected, so we can use to communicate over IP like a simple NIC.

Architecture overview is following:

to Host       |                       to Endpoint
network stack |                 network stack
      |       |                       |
+-----------+ |       +-----------+   +-----------+
|virtio-net | |       |virtio-net |   |virtio-net |
|driver     | |       |EP function|---|driver     |
+-----------+ |       +-----------+   +-----------+
      |       |             |
+-----------+ | +-----------+
|PCIeC      | | |PCIeC      |
|Rootcomplex|-|-|Endpoint   |
+-----------+ | +-----------+
  Host side   |          Endpoint side
Quoted 3 lines

This driver uses PCIe EP framework to show virtio-net (pci) device Host side, and generate virtual virtio-net device and register to EP side. A communication date

data?

Quoted 1 line

is diractly

directly?

Quoted 17 lines
Quoted 5 lines

transported between virtqueue level
with each other using PCIe embedded DMA controller.

by a limitation of the hardware and Linux EP framework, this function follows a virtio legacy specification.

what exactly is the limitation and why does it force legacy?

Quoted 9 lines

This function driver has beed tested on S4 Rcar (r8a779fa-spider) board but just use the PCIe EP framework and depends on the PCIe EDMA.

---
 drivers/pci/endpoint/functions/Kconfig        |  12 +
 drivers/pci/endpoint/functions/Makefile       |   1 +
 .../pci/endpoint/functions/pci-epf-vnet-ep.c  | 343 ++++++++++
Quoted 2 lines

It is actually that not related vnet. Just virtio. I think pci-epf-virtio.c is better.

Yes, it have to be.

Quoted 4 lines
Quoted 1 line
Quoted 1 line
 .../pci/endpoint/functions/pci-epf-vnet-rc.c  | 635 ++++++++++++++++++

It is epf driver. rc is quite confused.
Maybe you can combine pci-epf-vnet-ep.c and pci-epf-vnet-rc.c to one file.

I agree. Try to combine them

Quoted 4 lines
Quoted 1 line
Quoted 1 line
 drivers/pci/endpoint/functions/pci-epf-vnet.c | 387 +++++++++++

This file setup dma transfer according virtio-ring. How about pci-epf-virtio-dma.c ?

I attempt to rearrange the location of code and filenames.

Quoted 9 lines
Quoted 8 lines

Review

Quoted 2 lines
(quoted diff) +2 -0
+
+     epf_vnet_rc_set_config16(vnet, VIRTIO_PCI_ISR,

VIRTIO_PCI_ISR_QUEUE);

Quoted 5 lines
drivers/pci/endpoint/functions/pci-epf-vnet-rc.c new file +5 -0
+     /*
+      * Initialize the queue notify and selector to outside of the appropriate
+      * virtqueue index. It is used to detect change with polling. There is no
+      * other ways to detect host side driver updateing those values
+      */
Quoted 4 lines

I am try to use gic-its or other msi controller as doorbell.
https://lore.kernel.org/imx/[email protected]/T/#u

but it may need update host side pci virtio driver.

Thanks, is it possible to use MSI-X as well? The virtio spec indicates to use legacy irq or
MSI-X only.

Quoted 10 lines
Quoted 9 lines
Quoted 1 line

+ epf_vnet_rc_set_config16(vnet, VIRTIO_PCI_QUEUE_NOTIFY,

default_qindex);

Quoted 1 line

+ epf_vnet_rc_set_config16(vnet, VIRTIO_PCI_QUEUE_SEL,

default_qindex);

drivers/pci/endpoint/functions/pci-epf-vnet-rc.c line 113

Quoted 3 lines
drivers/pci/endpoint/functions/pci-epf-vnet-rc.c new file +3 -0
+     /* This pfn is also set to 0 for the polling as well */
+     epf_vnet_rc_set_config16(vnet, VIRTIO_PCI_QUEUE_PFN, 0);
+

--

Quoted 1 line

2.25.1

Best,
Shunsuke.

#12 lore raw

RE: [EXT] Re: [RFC PATCH 4/4] PCI: endpoint: function: Add EP function driver to provide virtio net device

Quoted 4 lines
Quoted 1 line

but it may need update host side pci virtio driver.

Thanks, is it possible to use MSI-X as well? The virtio spec indicates to use legacy irq or
MSI-X only.

I supposed yes. It is depend MSI controller type in EP side. But not like standard PCI MSI-X, it is platform MSI-X irq.

If use GIC-its, it should support MSI-X.

Thomas Gleixner is working on pre-device msi irq domain.
https://lore.kernel.org/all/[email protected]

I hope Thomas can finish their work soon.
so I can continue my patch upstream work.
https://lore.kernel.org/imx/87wn7evql7.ffs@tglx/T/#u

Quoted 3 lines
Quoted 1 line

Best,
Shunsuke.

#13 lore raw

Re: [EXT] Re: [RFC PATCH 4/4] PCI: endpoint: function: Add EP function driver to provide virtio net device

Quoted 7 lines · Frank Li
Quoted 4 lines
Quoted 1 line

but it may need update host side pci virtio driver.

Thanks, is it possible to use MSI-X as well? The virtio spec indicates to use legacy irq or
MSI-X only.

I supposed yes. It is depend MSI controller type in EP side. But not like standard PCI MSI-X, it is platform MSI-X irq.

Quoted 8 lines

If use GIC-its, it should support MSI-X.

Thomas Gleixner is working on pre-device msi irq domain.
https://lore.kernel.org/all/[email protected]

I hope Thomas can finish their work soon.
so I can continue my patch upstream work.
https://lore.kernel.org/imx/87wn7evql7.ffs@tglx/T/#u

Thank for sharing this those information. I'll see the details to embed.

Quoted 2 lines
Quoted 2 lines

Best,
Shunsuke.

Best,

Shunsuke.

#14 lore raw

Re: [RFC PATCH 4/4] PCI: endpoint: function: Add EP function driver to provide virtio net device

In reply to Michael S. Tsirkin

2023年2月3日(金) 19:22 Michael S. Tsirkin <[email protected]>:

Quoted 28 lines · Shunsuke Mie
Quoted 22 lines

Add a new endpoint(EP) function driver to provide virtio-net device. This function not only shows virtio-net device for PCIe host system, but also provides virtio-net device to EP side(local) system. Virtualy those network devices are connected, so we can use to communicate over IP like a simple NIC.

Architecture overview is following:

to Host       |                       to Endpoint
network stack |                 network stack
      |       |                       |
+-----------+ |       +-----------+   +-----------+
|virtio-net | |       |virtio-net |   |virtio-net |
|driver     | |       |EP function|---|driver     |
+-----------+ |       +-----------+   +-----------+
      |       |             |
+-----------+ | +-----------+
|PCIeC      | | |PCIeC      |
|Rootcomplex|-|-|Endpoint   |
+-----------+ | +-----------+
  Host side   |          Endpoint side
Quoted 3 lines

This driver uses PCIe EP framework to show virtio-net (pci) device Host side, and generate virtual virtio-net device and register to EP side. A communication date

Quoted 5 lines

data?

Quoted 1 line

is diractly

directly?

Sorry, I have to revise this comment.

Quoted 7 lines
Quoted 5 lines

transported between virtqueue level
with each other using PCIe embedded DMA controller.

by a limitation of the hardware and Linux EP framework, this function follows a virtio legacy specification.

what exactly is the limitation and why does it force legacy?

Modern virtio pci device have to provide a virtio pci capability, Designware's PCIe controller is equipped to several boards. There is
no functionality in the
controller to implement custom pci capability at
least. And the PCI EP framework is not supported either.

Those explanations have to be located on the cover letter. I'll add these.

Quoted 1,492 lines
Quoted 1,492 lines

This function driver has beed tested on S4 Rcar (r8a779fa-spider) board but just use the PCIe EP framework and depends on the PCIe EDMA.

---
 drivers/pci/endpoint/functions/Kconfig        |  12 +
 drivers/pci/endpoint/functions/Makefile       |   1 +
 .../pci/endpoint/functions/pci-epf-vnet-ep.c  | 343 ++++++++++
 .../pci/endpoint/functions/pci-epf-vnet-rc.c  | 635 ++++++++++++++++++
 drivers/pci/endpoint/functions/pci-epf-vnet.c | 387 +++++++++++
 drivers/pci/endpoint/functions/pci-epf-vnet.h |  62 ++
 6 files changed, 1440 insertions(+)
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet-ep.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet-rc.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet.h
drivers/pci/endpoint/functions/Kconfig +12 -0
index 9fd560886871..f88d8baaf689 100644
@@ -37,3 +37,15 @@ config PCI_EPF_VNTB
        between PCI Root Port and PCIe Endpoint.
 
        If in doubt, say "N" to disable Endpoint NTB driver.
+
+config PCI_EPF_VNET
+     tristate "PCI Endpoint virtio-net driver"
+     depends on PCI_ENDPOINT
+     select PCI_ENDPOINT_VIRTIO
+     select VHOST_RING
+     select VHOST_IOMEM
+     help
+       PCIe Endpoint virtio-net function implementation. This module enables to
+       show the virtio-net as pci device to PCIe Host side, and, another
+       virtio-net device show to local machine. Those devices can communicate
+       each other.
drivers/pci/endpoint/functions/Makefile +1 -0
index 5c13001deaba..74cc4c330c62 100644
@@ -6,3 +6,4 @@
 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_VNET)           += pci-epf-vnet.o pci-epf-vnet-rc.o pci-epf-vnet-ep.o
drivers/pci/endpoint/functions/pci-epf-vnet-ep.c new file +343 -0
new file mode 100644
index 000000000000..93b7e00e8d06
@@ -0,0 +1,343 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Functions work for Endpoint side(local) using EPF framework
+ */
+#include <linux/pci-epc.h>
+#include <linux/virtio_pci.h>
+#include <linux/virtio_net.h>
+#include <linux/virtio_ring.h>
+
+#include "pci-epf-vnet.h"
+
+static inline struct epf_vnet *vdev_to_vnet(struct virtio_device *vdev)
+{
+     return container_of(vdev, struct epf_vnet, ep.vdev);
+}
+
+static void epf_vnet_ep_set_status(struct epf_vnet *vnet, u16 status)
+{
+     vnet->ep.net_config_status |= status;
+}
+
+static void epf_vnet_ep_clear_status(struct epf_vnet *vnet, u16 status)
+{
+     vnet->ep.net_config_status &= ~status;
+}
+
+static void epf_vnet_ep_raise_config_irq(struct epf_vnet *vnet)
+{
+     virtio_config_changed(&vnet->ep.vdev);
+}
+
+void epf_vnet_ep_announce_linkup(struct epf_vnet *vnet)
+{
+     epf_vnet_ep_set_status(vnet,
+                            VIRTIO_NET_S_LINK_UP | VIRTIO_NET_S_ANNOUNCE);
+     epf_vnet_ep_raise_config_irq(vnet);
+}
+
+void epf_vnet_ep_notify(struct epf_vnet *vnet, struct virtqueue *vq)
+{
+     vring_interrupt(0, vq);
+}
+
+static int epf_vnet_ep_process_ctrlq_entry(struct epf_vnet *vnet)
+{
+     struct vringh *vrh = &vnet->ep.ctlvrh;
+     struct vringh_kiov *wiov = &vnet->ep.ctl_riov;
+     struct vringh_kiov *riov = &vnet->ep.ctl_wiov;
+     struct virtio_net_ctrl_hdr *hdr;
+     virtio_net_ctrl_ack *ack;
+     int err;
+     u16 head;
+     size_t len;
+
+     err = vringh_getdesc(vrh, riov, wiov, &head);
+     if (err <= 0)
+             goto done;
+
+     len = vringh_kiov_length(riov);
+     if (len < sizeof(*hdr)) {
+             pr_debug("Command is too short: %ld\n", len);
+             err = -EIO;
+             goto done;
+     }
+
+     if (vringh_kiov_length(wiov) < sizeof(*ack)) {
+             pr_debug("Space for ack is not enough\n");
+             err = -EIO;
+             goto done;
+     }
+
+     hdr = phys_to_virt((unsigned long)riov->iov[riov->i].iov_base);
+     ack = phys_to_virt((unsigned long)wiov->iov[wiov->i].iov_base);
+
+     switch (hdr->class) {
+     case VIRTIO_NET_CTRL_ANNOUNCE:
+             if (hdr->cmd != VIRTIO_NET_CTRL_ANNOUNCE_ACK) {
+                     pr_debug("Invalid command: announce: %d\n", hdr->cmd);
+                     goto done;
+             }
+
+             epf_vnet_ep_clear_status(vnet, VIRTIO_NET_S_ANNOUNCE);
+             *ack = VIRTIO_NET_OK;
+             break;
+     default:
+             pr_debug("Found not supported class: %d\n", hdr->class);
+             err = -EIO;
+     }
+
+done:
+     vringh_complete(vrh, head, len);
+     return err;
+}
+
+static u64 epf_vnet_ep_vdev_get_features(struct virtio_device *vdev)
+{
+     struct epf_vnet *vnet = vdev_to_vnet(vdev);
+
+     return vnet->virtio_features;
+}
+
+static int epf_vnet_ep_vdev_finalize_features(struct virtio_device *vdev)
+{
+     struct epf_vnet *vnet = vdev_to_vnet(vdev);
+
+     if (vdev->features != vnet->virtio_features)
+             return -EINVAL;
+
+     return 0;
+}
+
+static void epf_vnet_ep_vdev_get_config(struct virtio_device *vdev,
+                                     unsigned int offset, void *buf,
+                                     unsigned int len)
+{
+     struct epf_vnet *vnet = vdev_to_vnet(vdev);
+     const unsigned int mac_len = sizeof(vnet->vnet_cfg.mac);
+     const unsigned int status_len = sizeof(vnet->vnet_cfg.status);
+     unsigned int copy_len;
+
+     switch (offset) {
+     case offsetof(struct virtio_net_config, mac):
+             /* This PCIe EP function doesn't provide a VIRTIO_NET_F_MAC feature, so just
+              * clear the buffer.
+              */
+             copy_len = len >= mac_len ? mac_len : len;
+             memset(buf, 0x00, copy_len);
+             len -= copy_len;
+             buf += copy_len;
+             fallthrough;
+     case offsetof(struct virtio_net_config, status):
+             copy_len = len >= status_len ? status_len : len;
+             memcpy(buf, &vnet->ep.net_config_status, copy_len);
+             len -= copy_len;
+             buf += copy_len;
+             fallthrough;
+     default:
+             if (offset > sizeof(vnet->vnet_cfg)) {
+                     memset(buf, 0x00, len);
+                     break;
+             }
+             memcpy(buf, (void *)&vnet->vnet_cfg + offset, len);
+     }
+}
+
+static void epf_vnet_ep_vdev_set_config(struct virtio_device *vdev,
+                                     unsigned int offset, const void *buf,
+                                     unsigned int len)
+{
+     /* Do nothing, because all of virtio net config space is readonly. */
+}
+
+static u8 epf_vnet_ep_vdev_get_status(struct virtio_device *vdev)
+{
+     return 0;
+}
+
+static void epf_vnet_ep_vdev_set_status(struct virtio_device *vdev, u8 status)
+{
+     struct epf_vnet *vnet = vdev_to_vnet(vdev);
+
+     if (status & VIRTIO_CONFIG_S_DRIVER_OK)
+             epf_vnet_init_complete(vnet, EPF_VNET_INIT_COMPLETE_EP);
+}
+
+static void epf_vnet_ep_vdev_reset(struct virtio_device *vdev)
+{
+     pr_debug("doesn't support yet");
+}
+
+static bool epf_vnet_ep_vdev_vq_notify(struct virtqueue *vq)
+{
+     struct epf_vnet *vnet = vdev_to_vnet(vq->vdev);
+     struct vringh *tx_vrh = &vnet->ep.txvrh;
+     struct vringh *rx_vrh = &vnet->rc.rxvrh->vrh;
+     struct vringh_kiov *tx_iov = &vnet->ep.tx_iov;
+     struct vringh_kiov *rx_iov = &vnet->rc.rx_iov;
+     int err;
+
+     /* Support only one queue pair */
+     switch (vq->index) {
+     case 0: // rx queue
+             break;
+     case 1: // tx queue
+             while ((err = epf_vnet_transfer(vnet, tx_vrh, rx_vrh, tx_iov,
+                                             rx_iov, DMA_MEM_TO_DEV)) > 0)
+                     ;
+             if (err < 0)
+                     pr_debug("Failed to transmit: EP -> Host: %d\n", err);
+             break;
+     case 2: // control queue
+             epf_vnet_ep_process_ctrlq_entry(vnet);
+             break;
+     default:
+             return false;
+     }
+
+     return true;
+}
+
+static int epf_vnet_ep_vdev_find_vqs(struct virtio_device *vdev,
+                                  unsigned int nvqs, struct virtqueue *vqs[],
+                                  vq_callback_t *callback[],
+                                  const char *const names[], const bool *ctx,
+                                  struct irq_affinity *desc)
+{
+     struct epf_vnet *vnet = vdev_to_vnet(vdev);
+     const size_t vq_size = epf_vnet_get_vq_size();
+     int i;
+     int err;
+     int qidx;
+
+     for (qidx = 0, i = 0; i < nvqs; i++) {
+             struct virtqueue *vq;
+             struct vring *vring;
+             struct vringh *vrh;
+
+             if (!names[i]) {
+                     vqs[i] = NULL;
+                     continue;
+             }
+
+             vq = vring_create_virtqueue(qidx++, vq_size,
+                                         VIRTIO_PCI_VRING_ALIGN, vdev, true,
+                                         false, ctx ? ctx[i] : false,
+                                         epf_vnet_ep_vdev_vq_notify,
+                                         callback[i], names[i]);
+             if (!vq) {
+                     err = -ENOMEM;
+                     goto err_del_vqs;
+             }
+
+             vqs[i] = vq;
+             vring = virtqueue_get_vring(vq);
+
+             switch (i) {
+             case 0: // rx
+                     vrh = &vnet->ep.rxvrh;
+                     vnet->ep.rxvq = vq;
+                     break;
+             case 1: // tx
+                     vrh = &vnet->ep.txvrh;
+                     vnet->ep.txvq = vq;
+                     break;
+             case 2: // control
+                     vrh = &vnet->ep.ctlvrh;
+                     vnet->ep.ctlvq = vq;
+                     break;
+             default:
+                     err = -EIO;
+                     goto err_del_vqs;
+             }
+
+             err = vringh_init_kern(vrh, vnet->virtio_features, vq_size,
+                                    true, GFP_KERNEL, vring->desc,
+                                    vring->avail, vring->used);
+             if (err) {
+                     pr_err("failed to init vringh for vring %d\n", i);
+                     goto err_del_vqs;
+             }
+     }
+
+     err = epf_vnet_init_kiov(&vnet->ep.tx_iov, vq_size);
+     if (err)
+             goto err_free_kiov;
+     err = epf_vnet_init_kiov(&vnet->ep.rx_iov, vq_size);
+     if (err)
+             goto err_free_kiov;
+     err = epf_vnet_init_kiov(&vnet->ep.ctl_riov, vq_size);
+     if (err)
+             goto err_free_kiov;
+     err = epf_vnet_init_kiov(&vnet->ep.ctl_wiov, vq_size);
+     if (err)
+             goto err_free_kiov;
+
+     return 0;
+
+err_free_kiov:
+     epf_vnet_deinit_kiov(&vnet->ep.tx_iov);
+     epf_vnet_deinit_kiov(&vnet->ep.rx_iov);
+     epf_vnet_deinit_kiov(&vnet->ep.ctl_riov);
+     epf_vnet_deinit_kiov(&vnet->ep.ctl_wiov);
+
+err_del_vqs:
+     for (; i >= 0; i--) {
+             if (!names[i])
+                     continue;
+
+             if (!vqs[i])
+                     continue;
+
+             vring_del_virtqueue(vqs[i]);
+     }
+     return err;
+}
+
+static void epf_vnet_ep_vdev_del_vqs(struct virtio_device *vdev)
+{
+     struct virtqueue *vq, *n;
+     struct epf_vnet *vnet = vdev_to_vnet(vdev);
+
+     list_for_each_entry_safe(vq, n, &vdev->vqs, list)
+             vring_del_virtqueue(vq);
+
+     epf_vnet_deinit_kiov(&vnet->ep.tx_iov);
+     epf_vnet_deinit_kiov(&vnet->ep.rx_iov);
+     epf_vnet_deinit_kiov(&vnet->ep.ctl_riov);
+     epf_vnet_deinit_kiov(&vnet->ep.ctl_wiov);
+}
+
+static const struct virtio_config_ops epf_vnet_ep_vdev_config_ops = {
+     .get_features = epf_vnet_ep_vdev_get_features,
+     .finalize_features = epf_vnet_ep_vdev_finalize_features,
+     .get = epf_vnet_ep_vdev_get_config,
+     .set = epf_vnet_ep_vdev_set_config,
+     .get_status = epf_vnet_ep_vdev_get_status,
+     .set_status = epf_vnet_ep_vdev_set_status,
+     .reset = epf_vnet_ep_vdev_reset,
+     .find_vqs = epf_vnet_ep_vdev_find_vqs,
+     .del_vqs = epf_vnet_ep_vdev_del_vqs,
+};
+
+void epf_vnet_ep_cleanup(struct epf_vnet *vnet)
+{
+     unregister_virtio_device(&vnet->ep.vdev);
+}
+
+int epf_vnet_ep_setup(struct epf_vnet *vnet)
+{
+     int err;
+     struct virtio_device *vdev = &vnet->ep.vdev;
+
+     vdev->dev.parent = vnet->epf->epc->dev.parent;
+     vdev->config = &epf_vnet_ep_vdev_config_ops;
+     vdev->id.vendor = PCI_VENDOR_ID_REDHAT_QUMRANET;
+     vdev->id.device = VIRTIO_ID_NET;
+
+     err = register_virtio_device(vdev);
+     if (err)
+             return err;
+
+     return 0;
+}
drivers/pci/endpoint/functions/pci-epf-vnet-rc.c new file +635 -0
new file mode 100644
index 000000000000..2ca0245a9134
@@ -0,0 +1,635 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Functions work for PCie Host side(remote) using EPF framework.
+ */
+#include <linux/pci-epf.h>
+#include <linux/pci-epc.h>
+#include <linux/pci_ids.h>
+#include <linux/sched.h>
+#include <linux/virtio_pci.h>
+
+#include "pci-epf-vnet.h"
+
+#define VIRTIO_NET_LEGACY_CFG_BAR BAR_0
+
+/* Returns an out side of the valid queue index. */
+static inline u16 epf_vnet_rc_get_number_of_queues(struct epf_vnet *vnet)
+
+{
+     /* number of queue pairs and control queue */
+     return vnet->vnet_cfg.max_virtqueue_pairs * 2 + 1;
+}
+
+static void epf_vnet_rc_memcpy_config(struct epf_vnet *vnet, size_t offset,
+                                   void *buf, size_t len)
+{
+     void __iomem *base = vnet->rc.cfg_base + offset;
+
+     memcpy_toio(base, buf, len);
+}
+
+static void epf_vnet_rc_set_config8(struct epf_vnet *vnet, size_t offset,
+                                 u8 config)
+{
+     void __iomem *base = vnet->rc.cfg_base + offset;
+
+     iowrite8(ioread8(base) | config, base);
+}
+
+static void epf_vnet_rc_set_config16(struct epf_vnet *vnet, size_t offset,
+                                  u16 config)
+{
+     void __iomem *base = vnet->rc.cfg_base + offset;
+
+     iowrite16(ioread16(base) | config, base);
+}
+
+static void epf_vnet_rc_clear_config16(struct epf_vnet *vnet, size_t offset,
+                                    u16 config)
+{
+     void __iomem *base = vnet->rc.cfg_base + offset;
+
+     iowrite16(ioread16(base) & ~config, base);
+}
+
+static void epf_vnet_rc_set_config32(struct epf_vnet *vnet, size_t offset,
+                                  u32 config)
+{
+     void __iomem *base = vnet->rc.cfg_base + offset;
+
+     iowrite32(ioread32(base) | config, base);
+}
+
+static void epf_vnet_rc_raise_config_irq(struct epf_vnet *vnet)
+{
+     epf_vnet_rc_set_config16(vnet, VIRTIO_PCI_ISR, VIRTIO_PCI_ISR_CONFIG);
+     queue_work(vnet->rc.irq_wq, &vnet->rc.raise_irq_work);
+}
+
+void epf_vnet_rc_announce_linkup(struct epf_vnet *vnet)
+{
+     epf_vnet_rc_set_config16(vnet,
+                              VIRTIO_PCI_CONFIG_OFF(false) +
+                                      offsetof(struct virtio_net_config,
+                                               status),
+                              VIRTIO_NET_S_LINK_UP | VIRTIO_NET_S_ANNOUNCE);
+     epf_vnet_rc_raise_config_irq(vnet);
+}
+
+/*
+ * For the PCIe host, this driver shows legacy virtio-net device. Because,
+ * virtio structure pci capabilities is mandatory for modern virtio device,
+ * but there is no PCIe EP hardware that can be configured with any pci
+ * capabilities and Linux PCIe EP framework doesn't support it.
+ */
+static struct pci_epf_header epf_vnet_pci_header = {
+     .vendorid = PCI_VENDOR_ID_REDHAT_QUMRANET,
+     .deviceid = VIRTIO_TRANS_ID_NET,
+     .subsys_vendor_id = PCI_VENDOR_ID_REDHAT_QUMRANET,
+     .subsys_id = VIRTIO_ID_NET,
+     .revid = 0,
+     .baseclass_code = PCI_BASE_CLASS_NETWORK,
+     .interrupt_pin = PCI_INTERRUPT_PIN,
+};
+
+static void epf_vnet_rc_setup_configs(struct epf_vnet *vnet,
+                                   void __iomem *cfg_base)
+{
+     u16 default_qindex = epf_vnet_rc_get_number_of_queues(vnet);
+
+     epf_vnet_rc_set_config32(vnet, VIRTIO_PCI_HOST_FEATURES,
+                              vnet->virtio_features);
+
+     epf_vnet_rc_set_config16(vnet, VIRTIO_PCI_ISR, VIRTIO_PCI_ISR_QUEUE);
+     /*
+      * Initialize the queue notify and selector to outside of the appropriate
+      * virtqueue index. It is used to detect change with polling. There is no
+      * other ways to detect host side driver updateing those values
+      */
+     epf_vnet_rc_set_config16(vnet, VIRTIO_PCI_QUEUE_NOTIFY, default_qindex);
+     epf_vnet_rc_set_config16(vnet, VIRTIO_PCI_QUEUE_SEL, default_qindex);
+     /* This pfn is also set to 0 for the polling as well */
+     epf_vnet_rc_set_config16(vnet, VIRTIO_PCI_QUEUE_PFN, 0);
+
+     epf_vnet_rc_set_config16(vnet, VIRTIO_PCI_QUEUE_NUM,
+                              epf_vnet_get_vq_size());
+     epf_vnet_rc_set_config8(vnet, VIRTIO_PCI_STATUS, 0);
+     epf_vnet_rc_memcpy_config(vnet, VIRTIO_PCI_CONFIG_OFF(false),
+                               &vnet->vnet_cfg, sizeof(vnet->vnet_cfg));
+}
+
+static void epf_vnet_cleanup_bar(struct epf_vnet *vnet)
+{
+     struct pci_epf *epf = vnet->epf;
+
+     pci_epc_clear_bar(epf->epc, epf->func_no, epf->vfunc_no,
+                       &epf->bar[VIRTIO_NET_LEGACY_CFG_BAR]);
+     pci_epf_free_space(epf, vnet->rc.cfg_base, VIRTIO_NET_LEGACY_CFG_BAR,
+                        PRIMARY_INTERFACE);
+}
+
+static int epf_vnet_setup_bar(struct epf_vnet *vnet)
+{
+     int err;
+     size_t cfg_bar_size =
+             VIRTIO_PCI_CONFIG_OFF(false) + sizeof(struct virtio_net_config);
+     struct pci_epf *epf = vnet->epf;
+     const struct pci_epc_features *features;
+     struct pci_epf_bar *config_bar = &epf->bar[VIRTIO_NET_LEGACY_CFG_BAR];
+
+     features = pci_epc_get_features(epf->epc, epf->func_no, epf->vfunc_no);
+     if (!features) {
+             pr_debug("Failed to get PCI EPC features\n");
+             return -EOPNOTSUPP;
+     }
+
+     if (features->reserved_bar & BIT(VIRTIO_NET_LEGACY_CFG_BAR)) {
+             pr_debug("Cannot use the PCI BAR for legacy virtio pci\n");
+             return -EOPNOTSUPP;
+     }
+
+     if (features->bar_fixed_size[VIRTIO_NET_LEGACY_CFG_BAR]) {
+             if (cfg_bar_size >
+                 features->bar_fixed_size[VIRTIO_NET_LEGACY_CFG_BAR]) {
+                     pr_debug("PCI BAR size is not enough\n");
+                     return -ENOMEM;
+             }
+     }
+
+     config_bar->flags |= PCI_BASE_ADDRESS_MEM_TYPE_64;
+
+     vnet->rc.cfg_base = pci_epf_alloc_space(epf, cfg_bar_size,
+                                             VIRTIO_NET_LEGACY_CFG_BAR,
+                                             features->align,
+                                             PRIMARY_INTERFACE);
+     if (!vnet->rc.cfg_base) {
+             pr_debug("Failed to allocate virtio-net config memory\n");
+             return -ENOMEM;
+     }
+
+     epf_vnet_rc_setup_configs(vnet, vnet->rc.cfg_base);
+
+     err = pci_epc_set_bar(epf->epc, epf->func_no, epf->vfunc_no,
+                           config_bar);
+     if (err) {
+             pr_debug("Failed to set PCI BAR");
+             goto err_free_space;
+     }
+
+     return 0;
+
+err_free_space:
+     pci_epf_free_space(epf, vnet->rc.cfg_base, VIRTIO_NET_LEGACY_CFG_BAR,
+                        PRIMARY_INTERFACE);
+     return err;
+}
+
+static int epf_vnet_rc_negotiate_configs(struct epf_vnet *vnet, u32 *txpfn,
+                                      u32 *rxpfn, u32 *ctlpfn)
+{
+     const u16 nqueues = epf_vnet_rc_get_number_of_queues(vnet);
+     const u16 default_sel = nqueues;
+     u32 __iomem *queue_pfn = vnet->rc.cfg_base + VIRTIO_PCI_QUEUE_PFN;
+     u16 __iomem *queue_sel = vnet->rc.cfg_base + VIRTIO_PCI_QUEUE_SEL;
+     u8 __iomem *pci_status = vnet->rc.cfg_base + VIRTIO_PCI_STATUS;
+     u32 pfn;
+     u16 sel;
+     struct {
+             u32 pfn;
+             u16 sel;
+     } tmp[3] = {};
+     int tmp_index = 0;
+
+     *rxpfn = *txpfn = *ctlpfn = 0;
+
+     /* To avoid to miss a getting the pfn and selector for virtqueue wrote by
+      * host driver, we need to implement fast polling with saving.
+      *
+      * This implementation suspects that the host driver writes pfn only once
+      * for each queues
+      */
+     while (tmp_index < nqueues) {
+             pfn = ioread32(queue_pfn);
+             if (pfn == 0)
+                     continue;
+
+             iowrite32(0, queue_pfn);
+
+             sel = ioread16(queue_sel);
+             if (sel == default_sel)
+                     continue;
+
+             tmp[tmp_index].pfn = pfn;
+             tmp[tmp_index].sel = sel;
+             tmp_index++;
+     }
+
+     while (!((ioread8(pci_status) & VIRTIO_CONFIG_S_DRIVER_OK)))
+             ;
+
+     for (int i = 0; i < nqueues; ++i) {
+             switch (tmp[i].sel) {
+             case 0:
+                     *rxpfn = tmp[i].pfn;
+                     break;
+             case 1:
+                     *txpfn = tmp[i].pfn;
+                     break;
+             case 2:
+                     *ctlpfn = tmp[i].pfn;
+                     break;
+             }
+     }
+
+     if (!*rxpfn || !*txpfn || !*ctlpfn)
+             return -EIO;
+
+     return 0;
+}
+
+static int epf_vnet_rc_monitor_notify(void *data)
+{
+     struct epf_vnet *vnet = data;
+     u16 __iomem *queue_notify = vnet->rc.cfg_base + VIRTIO_PCI_QUEUE_NOTIFY;
+     const u16 notify_default = epf_vnet_rc_get_number_of_queues(vnet);
+
+     epf_vnet_init_complete(vnet, EPF_VNET_INIT_COMPLETE_RC);
+
+     /* Poll to detect a change of the queue_notify register. Sometimes this
+      * polling misses the change, so try to check each virtqueues
+      * everytime.
+      */
+     while (true) {
+             while (ioread16(queue_notify) == notify_default)
+                     ;
+             iowrite16(notify_default, queue_notify);
+
+             queue_work(vnet->rc.tx_wq, &vnet->rc.tx_work);
+             queue_work(vnet->rc.ctl_wq, &vnet->rc.ctl_work);
+     }
+
+     return 0;
+}
+
+static int epf_vnet_rc_spawn_notify_monitor(struct epf_vnet *vnet)
+{
+     vnet->rc.notify_monitor_task =
+             kthread_create(epf_vnet_rc_monitor_notify, vnet,
+                            "pci-epf-vnet/cfg_negotiator");
+     if (IS_ERR(vnet->rc.notify_monitor_task))
+             return PTR_ERR(vnet->rc.notify_monitor_task);
+
+     /* Change the thread priority to high for polling. */
+     sched_set_fifo(vnet->rc.notify_monitor_task);
+     wake_up_process(vnet->rc.notify_monitor_task);
+
+     return 0;
+}
+
+static int epf_vnet_rc_device_setup(void *data)
+{
+     struct epf_vnet *vnet = data;
+     struct pci_epf *epf = vnet->epf;
+     u32 txpfn, rxpfn, ctlpfn;
+     const size_t vq_size = epf_vnet_get_vq_size();
+     int err;
+
+     err = epf_vnet_rc_negotiate_configs(vnet, &txpfn, &rxpfn, &ctlpfn);
+     if (err) {
+             pr_debug("Failed to negatiate configs with driver\n");
+             return err;
+     }
+
+     /* Polling phase is finished. This thread backs to normal priority. */
+     sched_set_normal(vnet->rc.device_setup_task, 19);
+
+     vnet->rc.txvrh = pci_epf_virtio_alloc_vringh(epf, vnet->virtio_features,
+                                                  txpfn, vq_size);
+     if (IS_ERR(vnet->rc.txvrh)) {
+             pr_debug("Failed to setup virtqueue for tx\n");
+             return PTR_ERR(vnet->rc.txvrh);
+     }
+
+     err = epf_vnet_init_kiov(&vnet->rc.tx_iov, vq_size);
+     if (err)
+             goto err_free_epf_tx_vringh;
+
+     vnet->rc.rxvrh = pci_epf_virtio_alloc_vringh(epf, vnet->virtio_features,
+                                                  rxpfn, vq_size);
+     if (IS_ERR(vnet->rc.rxvrh)) {
+             pr_debug("Failed to setup virtqueue for rx\n");
+             err = PTR_ERR(vnet->rc.rxvrh);
+             goto err_deinit_tx_kiov;
+     }
+
+     err = epf_vnet_init_kiov(&vnet->rc.rx_iov, vq_size);
+     if (err)
+             goto err_free_epf_rx_vringh;
+
+     vnet->rc.ctlvrh = pci_epf_virtio_alloc_vringh(
+             epf, vnet->virtio_features, ctlpfn, vq_size);
+     if (IS_ERR(vnet->rc.ctlvrh)) {
+             pr_err("failed to setup virtqueue\n");
+             err = PTR_ERR(vnet->rc.ctlvrh);
+             goto err_deinit_rx_kiov;
+     }
+
+     err = epf_vnet_init_kiov(&vnet->rc.ctl_riov, vq_size);
+     if (err)
+             goto err_free_epf_ctl_vringh;
+
+     err = epf_vnet_init_kiov(&vnet->rc.ctl_wiov, vq_size);
+     if (err)
+             goto err_deinit_ctl_riov;
+
+     err = epf_vnet_rc_spawn_notify_monitor(vnet);
+     if (err) {
+             pr_debug("Failed to create notify monitor thread\n");
+             goto err_deinit_ctl_wiov;
+     }
+
+     return 0;
+
+err_deinit_ctl_wiov:
+     epf_vnet_deinit_kiov(&vnet->rc.ctl_wiov);
+err_deinit_ctl_riov:
+     epf_vnet_deinit_kiov(&vnet->rc.ctl_riov);
+err_free_epf_ctl_vringh:
+     pci_epf_virtio_free_vringh(epf, vnet->rc.ctlvrh);
+err_deinit_rx_kiov:
+     epf_vnet_deinit_kiov(&vnet->rc.rx_iov);
+err_free_epf_rx_vringh:
+     pci_epf_virtio_free_vringh(epf, vnet->rc.rxvrh);
+err_deinit_tx_kiov:
+     epf_vnet_deinit_kiov(&vnet->rc.tx_iov);
+err_free_epf_tx_vringh:
+     pci_epf_virtio_free_vringh(epf, vnet->rc.txvrh);
+
+     return err;
+}
+
+static int epf_vnet_rc_spawn_device_setup_task(struct epf_vnet *vnet)
+{
+     vnet->rc.device_setup_task = kthread_create(
+             epf_vnet_rc_device_setup, vnet, "pci-epf-vnet/cfg_negotiator");
+     if (IS_ERR(vnet->rc.device_setup_task))
+             return PTR_ERR(vnet->rc.device_setup_task);
+
+     /* Change the thread priority to high for the polling. */
+     sched_set_fifo(vnet->rc.device_setup_task);
+     wake_up_process(vnet->rc.device_setup_task);
+
+     return 0;
+}
+
+static void epf_vnet_rc_tx_handler(struct work_struct *work)
+{
+     struct epf_vnet *vnet = container_of(work, struct epf_vnet, rc.tx_work);
+     struct vringh *tx_vrh = &vnet->rc.txvrh->vrh;
+     struct vringh *rx_vrh = &vnet->ep.rxvrh;
+     struct vringh_kiov *tx_iov = &vnet->rc.tx_iov;
+     struct vringh_kiov *rx_iov = &vnet->ep.rx_iov;
+
+     while (epf_vnet_transfer(vnet, tx_vrh, rx_vrh, tx_iov, rx_iov,
+                              DMA_DEV_TO_MEM) > 0)
+             ;
+}
+
+static void epf_vnet_rc_raise_irq_handler(struct work_struct *work)
+{
+     struct epf_vnet *vnet =
+             container_of(work, struct epf_vnet, rc.raise_irq_work);
+     struct pci_epf *epf = vnet->epf;
+
+     pci_epc_raise_irq(epf->epc, epf->func_no, epf->vfunc_no,
+                       PCI_EPC_IRQ_LEGACY, 0);
+}
+
+struct epf_vnet_rc_meminfo {
+     void __iomem *addr, *virt;
+     phys_addr_t phys;
+     size_t len;
+};
+
+/* Util function to access PCIe host side memory from local CPU.  */
+static struct epf_vnet_rc_meminfo *
+epf_vnet_rc_epc_mmap(struct pci_epf *epf, phys_addr_t pci_addr, size_t len)
+{
+     int err;
+     phys_addr_t aaddr, phys_addr;
+     size_t asize, offset;
+     void __iomem *virt_addr;
+     struct epf_vnet_rc_meminfo *meminfo;
+
+     err = pci_epc_mem_align(epf->epc, pci_addr, len, &aaddr, &asize);
+     if (err) {
+             pr_debug("Failed to get EPC align: %d\n", err);
+             return NULL;
+     }
+
+     offset = pci_addr - aaddr;
+
+     virt_addr = pci_epc_mem_alloc_addr(epf->epc, &phys_addr, asize);
+     if (!virt_addr) {
+             pr_debug("Failed to allocate epc memory\n");
+             return NULL;
+     }
+
+     err = pci_epc_map_addr(epf->epc, epf->func_no, epf->vfunc_no, phys_addr,
+                            aaddr, asize);
+     if (err) {
+             pr_debug("Failed to map epc memory\n");
+             goto err_epc_free_addr;
+     }
+
+     meminfo = kmalloc(sizeof(*meminfo), GFP_KERNEL);
+     if (!meminfo)
+             goto err_epc_unmap_addr;
+
+     meminfo->virt = virt_addr;
+     meminfo->phys = phys_addr;
+     meminfo->len = len;
+     meminfo->addr = virt_addr + offset;
+
+     return meminfo;
+
+err_epc_unmap_addr:
+     pci_epc_unmap_addr(epf->epc, epf->func_no, epf->vfunc_no,
+                        meminfo->phys);
+err_epc_free_addr:
+     pci_epc_mem_free_addr(epf->epc, meminfo->phys, meminfo->virt,
+                           meminfo->len);
+
+     return NULL;
+}
+
+static void epf_vnet_rc_epc_munmap(struct pci_epf *epf,
+                                struct epf_vnet_rc_meminfo *meminfo)
+{
+     pci_epc_unmap_addr(epf->epc, epf->func_no, epf->vfunc_no,
+                        meminfo->phys);
+     pci_epc_mem_free_addr(epf->epc, meminfo->phys, meminfo->virt,
+                           meminfo->len);
+     kfree(meminfo);
+}
+
+static int epf_vnet_rc_process_ctrlq_entry(struct epf_vnet *vnet)
+{
+     struct vringh_kiov *riov = &vnet->rc.ctl_riov;
+     struct vringh_kiov *wiov = &vnet->rc.ctl_wiov;
+     struct vringh *vrh = &vnet->rc.ctlvrh->vrh;
+     struct pci_epf *epf = vnet->epf;
+     struct epf_vnet_rc_meminfo *rmem, *wmem;
+     struct virtio_net_ctrl_hdr *hdr;
+     int err;
+     u16 head;
+     size_t total_len;
+     u8 class, cmd;
+
+     err = vringh_getdesc(vrh, riov, wiov, &head);
+     if (err <= 0)
+             return err;
+
+     total_len = vringh_kiov_length(riov);
+
+     rmem = epf_vnet_rc_epc_mmap(epf, (u64)riov->iov[riov->i].iov_base,
+                                 riov->iov[riov->i].iov_len);
+     if (!rmem) {
+             err = -ENOMEM;
+             goto err_abandon_descs;
+     }
+
+     wmem = epf_vnet_rc_epc_mmap(epf, (u64)wiov->iov[wiov->i].iov_base,
+                                 wiov->iov[wiov->i].iov_len);
+     if (!wmem) {
+             err = -ENOMEM;
+             goto err_epc_unmap_rmem;
+     }
+
+     hdr = rmem->addr;
+     class = ioread8(&hdr->class);
+     cmd = ioread8(&hdr->cmd);
+     switch (ioread8(&hdr->class)) {
+     case VIRTIO_NET_CTRL_ANNOUNCE:
+             if (cmd != VIRTIO_NET_CTRL_ANNOUNCE_ACK) {
+                     pr_err("Found invalid command: announce: %d\n", cmd);
+                     break;
+             }
+             epf_vnet_rc_clear_config16(
+                     vnet,
+                     VIRTIO_PCI_CONFIG_OFF(false) +
+                             offsetof(struct virtio_net_config, status),
+                     VIRTIO_NET_S_ANNOUNCE);
+             epf_vnet_rc_clear_config16(vnet, VIRTIO_PCI_ISR,
+                                        VIRTIO_PCI_ISR_CONFIG);
+
+             iowrite8(VIRTIO_NET_OK, wmem->addr);
+             break;
+     default:
+             pr_err("Found unsupported class in control queue: %d\n", class);
+             break;
+     }
+
+     epf_vnet_rc_epc_munmap(epf, rmem);
+     epf_vnet_rc_epc_munmap(epf, wmem);
+     vringh_complete(vrh, head, total_len);
+
+     return 1;
+
+err_epc_unmap_rmem:
+     epf_vnet_rc_epc_munmap(epf, rmem);
+err_abandon_descs:
+     vringh_abandon(vrh, head);
+
+     return err;
+}
+
+static void epf_vnet_rc_process_ctrlq_entries(struct work_struct *work)
+{
+     struct epf_vnet *vnet =
+             container_of(work, struct epf_vnet, rc.ctl_work);
+
+     while (epf_vnet_rc_process_ctrlq_entry(vnet) > 0)
+             ;
+}
+
+void epf_vnet_rc_notify(struct epf_vnet *vnet)
+{
+     queue_work(vnet->rc.irq_wq, &vnet->rc.raise_irq_work);
+}
+
+void epf_vnet_rc_cleanup(struct epf_vnet *vnet)
+{
+     epf_vnet_cleanup_bar(vnet);
+     destroy_workqueue(vnet->rc.tx_wq);
+     destroy_workqueue(vnet->rc.irq_wq);
+     destroy_workqueue(vnet->rc.ctl_wq);
+
+     kthread_stop(vnet->rc.device_setup_task);
+}
+
+int epf_vnet_rc_setup(struct epf_vnet *vnet)
+{
+     int err;
+     struct pci_epf *epf = vnet->epf;
+
+     err = pci_epc_write_header(epf->epc, epf->func_no, epf->vfunc_no,
+                                &epf_vnet_pci_header);
+     if (err)
+             return err;
+
+     err = epf_vnet_setup_bar(vnet);
+     if (err)
+             return err;
+
+     vnet->rc.tx_wq =
+             alloc_workqueue("pci-epf-vnet/tx-wq",
+                             WQ_MEM_RECLAIM | WQ_HIGHPRI | WQ_UNBOUND, 0);
+     if (!vnet->rc.tx_wq) {
+             pr_debug(
+                     "Failed to allocate workqueue for rc -> ep transmission\n");
+             err = -ENOMEM;
+             goto err_cleanup_bar;
+     }
+
+     INIT_WORK(&vnet->rc.tx_work, epf_vnet_rc_tx_handler);
+
+     vnet->rc.irq_wq =
+             alloc_workqueue("pci-epf-vnet/irq-wq",
+                             WQ_MEM_RECLAIM | WQ_HIGHPRI | WQ_UNBOUND, 0);
+     if (!vnet->rc.irq_wq) {
+             pr_debug("Failed to allocate workqueue for irq\n");
+             err = -ENOMEM;
+             goto err_destory_tx_wq;
+     }
+
+     INIT_WORK(&vnet->rc.raise_irq_work, epf_vnet_rc_raise_irq_handler);
+
+     vnet->rc.ctl_wq =
+             alloc_workqueue("pci-epf-vnet/ctl-wq",
+                             WQ_MEM_RECLAIM | WQ_HIGHPRI | WQ_UNBOUND, 0);
+     if (!vnet->rc.ctl_wq) {
+             pr_err("Failed to allocate work queue for control queue processing\n");
+             err = -ENOMEM;
+             goto err_destory_irq_wq;
+     }
+
+     INIT_WORK(&vnet->rc.ctl_work, epf_vnet_rc_process_ctrlq_entries);
+
+     err = epf_vnet_rc_spawn_device_setup_task(vnet);
+     if (err)
+             goto err_destory_ctl_wq;
+
+     return 0;
+
+err_cleanup_bar:
+     epf_vnet_cleanup_bar(vnet);
+err_destory_tx_wq:
+     destroy_workqueue(vnet->rc.tx_wq);
+err_destory_irq_wq:
+     destroy_workqueue(vnet->rc.irq_wq);
+err_destory_ctl_wq:
+     destroy_workqueue(vnet->rc.ctl_wq);
+
+     return err;
+}
drivers/pci/endpoint/functions/pci-epf-vnet.c new file +387 -0
new file mode 100644
index 000000000000..e48ad8067796
@@ -0,0 +1,387 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * PCI Endpoint function driver to impliment virtio-net device.
+ */
+#include <linux/module.h>
+#include <linux/pci-epf.h>
+#include <linux/pci-epc.h>
+#include <linux/vringh.h>
+#include <linux/dmaengine.h>
+
+#include "pci-epf-vnet.h"
+
+static int virtio_queue_size = 0x100;
+module_param(virtio_queue_size, int, 0444);
+MODULE_PARM_DESC(virtio_queue_size, "A length of virtqueue");
+
+int epf_vnet_get_vq_size(void)
+{
+     return virtio_queue_size;
+}
+
+int epf_vnet_init_kiov(struct vringh_kiov *kiov, const size_t vq_size)
+{
+     struct kvec *kvec;
+
+     kvec = kmalloc_array(vq_size, sizeof(*kvec), GFP_KERNEL);
+     if (!kvec)
+             return -ENOMEM;
+
+     vringh_kiov_init(kiov, kvec, vq_size);
+
+     return 0;
+}
+
+void epf_vnet_deinit_kiov(struct vringh_kiov *kiov)
+{
+     kfree(kiov->iov);
+}
+
+void epf_vnet_init_complete(struct epf_vnet *vnet, u8 from)
+{
+     vnet->init_complete |= from;
+
+     if (!(vnet->init_complete & EPF_VNET_INIT_COMPLETE_EP))
+             return;
+
+     if (!(vnet->init_complete & EPF_VNET_INIT_COMPLETE_RC))
+             return;
+
+     epf_vnet_ep_announce_linkup(vnet);
+     epf_vnet_rc_announce_linkup(vnet);
+}
+
+struct epf_dma_filter_param {
+     struct device *dev;
+     u32 dma_mask;
+};
+
+static bool epf_virtnet_dma_filter(struct dma_chan *chan, void *param)
+{
+     struct epf_dma_filter_param *fparam = param;
+     struct dma_slave_caps caps;
+
+     memset(&caps, 0, sizeof(caps));
+     dma_get_slave_caps(chan, &caps);
+
+     return chan->device->dev == fparam->dev &&
+            (fparam->dma_mask & caps.directions);
+}
+
+static int epf_vnet_init_edma(struct epf_vnet *vnet, struct device *dma_dev)
+{
+     struct epf_dma_filter_param param;
+     dma_cap_mask_t mask;
+     int err;
+
+     dma_cap_zero(mask);
+     dma_cap_set(DMA_SLAVE, mask);
+
+     param.dev = dma_dev;
+     param.dma_mask = BIT(DMA_MEM_TO_DEV);
+     vnet->lr_dma_chan =
+             dma_request_channel(mask, epf_virtnet_dma_filter, &param);
+     if (!vnet->lr_dma_chan)
+             return -EOPNOTSUPP;
+
+     param.dma_mask = BIT(DMA_DEV_TO_MEM);
+     vnet->rl_dma_chan =
+             dma_request_channel(mask, epf_virtnet_dma_filter, &param);
+     if (!vnet->rl_dma_chan) {
+             err = -EOPNOTSUPP;
+             goto err_release_channel;
+     }
+
+     return 0;
+
+err_release_channel:
+     dma_release_channel(vnet->lr_dma_chan);
+
+     return err;
+}
+
+static void epf_vnet_deinit_edma(struct epf_vnet *vnet)
+{
+     dma_release_channel(vnet->lr_dma_chan);
+     dma_release_channel(vnet->rl_dma_chan);
+}
+
+static int epf_vnet_dma_single(struct epf_vnet *vnet, phys_addr_t pci,
+                            dma_addr_t dma, size_t len,
+                            void (*callback)(void *), void *param,
+                            enum dma_transfer_direction dir)
+{
+     struct dma_async_tx_descriptor *desc;
+     int err;
+     struct dma_chan *chan;
+     struct dma_slave_config sconf;
+     dma_cookie_t cookie;
+     unsigned long flags = 0;
+
+     if (dir == DMA_MEM_TO_DEV) {
+             sconf.dst_addr = pci;
+             chan = vnet->lr_dma_chan;
+     } else {
+             sconf.src_addr = pci;
+             chan = vnet->rl_dma_chan;
+     }
+
+     err = dmaengine_slave_config(chan, &sconf);
+     if (unlikely(err))
+             return err;
+
+     if (callback)
+             flags = DMA_PREP_INTERRUPT | DMA_PREP_FENCE;
+
+     desc = dmaengine_prep_slave_single(chan, dma, len, dir, flags);
+     if (unlikely(!desc))
+             return -EIO;
+
+     desc->callback = callback;
+     desc->callback_param = param;
+
+     cookie = dmaengine_submit(desc);
+     err = dma_submit_error(cookie);
+     if (unlikely(err))
+             return err;
+
+     dma_async_issue_pending(chan);
+
+     return 0;
+}
+
+struct epf_vnet_dma_callback_param {
+     struct epf_vnet *vnet;
+     struct vringh *tx_vrh, *rx_vrh;
+     struct virtqueue *vq;
+     size_t total_len;
+     u16 tx_head, rx_head;
+};
+
+static void epf_vnet_dma_callback(void *p)
+{
+     struct epf_vnet_dma_callback_param *param = p;
+     struct epf_vnet *vnet = param->vnet;
+
+     vringh_complete(param->tx_vrh, param->tx_head, param->total_len);
+     vringh_complete(param->rx_vrh, param->rx_head, param->total_len);
+
+     epf_vnet_rc_notify(vnet);
+     epf_vnet_ep_notify(vnet, param->vq);
+
+     kfree(param);
+}
+
+/**
+ * epf_vnet_transfer() - transfer data between tx vring to rx vring using edma
+ * @vnet: epf virtio net device to do dma
+ * @tx_vrh: vringh related to source tx vring
+ * @rx_vrh: vringh related to target rx vring
+ * @tx_iov: buffer to use tx
+ * @rx_iov: buffer to use rx
+ * @dir: a direction of DMA. local to remote or local from remote
+ *
+ * This function returns 0, 1 or error number. The 0 indicates there is not
+ * data to send. The 1 indicates a request to DMA is succeeded. Other error
+ * numbers shows error, however, ENOSPC means there is no buffer on target
+ * vring, so should retry to call later.
+ */
+int epf_vnet_transfer(struct epf_vnet *vnet, struct vringh *tx_vrh,
+                   struct vringh *rx_vrh, struct vringh_kiov *tx_iov,
+                   struct vringh_kiov *rx_iov,
+                   enum dma_transfer_direction dir)
+{
+     int err;
+     u16 tx_head, rx_head;
+     size_t total_tx_len;
+     struct epf_vnet_dma_callback_param *cb_param;
+     struct vringh_kiov *liov, *riov;
+
+     err = vringh_getdesc(tx_vrh, tx_iov, NULL, &tx_head);
+     if (err <= 0)
+             return err;
+
+     total_tx_len = vringh_kiov_length(tx_iov);
+
+     err = vringh_getdesc(rx_vrh, NULL, rx_iov, &rx_head);
+     if (err < 0) {
+             goto err_tx_complete;
+     } else if (!err) {
+             /* There is not space on a vring of destination to transmit data, so
+              * rollback tx vringh
+              */
+             vringh_abandon(tx_vrh, tx_head);
+             return -ENOSPC;
+     }
+
+     cb_param = kmalloc(sizeof(*cb_param), GFP_KERNEL);
+     if (!cb_param) {
+             err = -ENOMEM;
+             goto err_rx_complete;
+     }
+
+     cb_param->tx_vrh = tx_vrh;
+     cb_param->rx_vrh = rx_vrh;
+     cb_param->tx_head = tx_head;
+     cb_param->rx_head = rx_head;
+     cb_param->total_len = total_tx_len;
+     cb_param->vnet = vnet;
+
+     switch (dir) {
+     case DMA_MEM_TO_DEV:
+             liov = tx_iov;
+             riov = rx_iov;
+             cb_param->vq = vnet->ep.txvq;
+             break;
+     case DMA_DEV_TO_MEM:
+             liov = rx_iov;
+             riov = tx_iov;
+             cb_param->vq = vnet->ep.rxvq;
+             break;
+     default:
+             err = -EINVAL;
+             goto err_free_param;
+     }
+
+     for (; tx_iov->i < tx_iov->used; tx_iov->i++, rx_iov->i++) {
+             size_t len;
+             u64 lbase, rbase;
+             void (*callback)(void *) = NULL;
+
+             lbase = (u64)liov->iov[liov->i].iov_base;
+             rbase = (u64)riov->iov[riov->i].iov_base;
+             len = tx_iov->iov[tx_iov->i].iov_len;
+
+             if (tx_iov->i + 1 == tx_iov->used)
+                     callback = epf_vnet_dma_callback;
+
+             err = epf_vnet_dma_single(vnet, rbase, lbase, len, callback,
+                                       cb_param, dir);
+             if (err)
+                     goto err_free_param;
+     }
+
+     return 1;
+
+err_free_param:
+     kfree(cb_param);
+err_rx_complete:
+     vringh_complete(rx_vrh, rx_head, vringh_kiov_length(rx_iov));
+err_tx_complete:
+     vringh_complete(tx_vrh, tx_head, total_tx_len);
+
+     return err;
+}
+
+static int epf_vnet_bind(struct pci_epf *epf)
+{
+     int err;
+     struct epf_vnet *vnet = epf_get_drvdata(epf);
+
+     err = epf_vnet_init_edma(vnet, epf->epc->dev.parent);
+     if (err)
+             return err;
+
+     err = epf_vnet_rc_setup(vnet);
+     if (err)
+             goto err_free_edma;
+
+     err = epf_vnet_ep_setup(vnet);
+     if (err)
+             goto err_cleanup_rc;
+
+     return 0;
+
+err_free_edma:
+     epf_vnet_deinit_edma(vnet);
+err_cleanup_rc:
+     epf_vnet_rc_cleanup(vnet);
+
+     return err;
+}
+
+static void epf_vnet_unbind(struct pci_epf *epf)
+{
+     struct epf_vnet *vnet = epf_get_drvdata(epf);
+
+     epf_vnet_deinit_edma(vnet);
+     epf_vnet_rc_cleanup(vnet);
+     epf_vnet_ep_cleanup(vnet);
+}
+
+static struct pci_epf_ops epf_vnet_ops = {
+     .bind = epf_vnet_bind,
+     .unbind = epf_vnet_unbind,
+};
+
+static const struct pci_epf_device_id epf_vnet_ids[] = {
+     { .name = "pci_epf_vnet" },
+     {}
+};
+
+static void epf_vnet_virtio_init(struct epf_vnet *vnet)
+{
+     vnet->virtio_features =
+             BIT(VIRTIO_NET_F_MTU) | BIT(VIRTIO_NET_F_STATUS) |
+             /* Following features are to skip any of checking and offloading, Like a
+              * transmission between virtual machines on same system. Details are on
+              * section 5.1.5 in virtio specification.
+              */
+             BIT(VIRTIO_NET_F_GUEST_CSUM) | BIT(VIRTIO_NET_F_GUEST_TSO4) |
+             BIT(VIRTIO_NET_F_GUEST_TSO6) | BIT(VIRTIO_NET_F_GUEST_ECN) |
+             BIT(VIRTIO_NET_F_GUEST_UFO) |
+             // The control queue is just used for linkup announcement.
+             BIT(VIRTIO_NET_F_CTRL_VQ);
+
+     vnet->vnet_cfg.max_virtqueue_pairs = 1;
+     vnet->vnet_cfg.status = 0;
+     vnet->vnet_cfg.mtu = PAGE_SIZE;
+}
+
+static int epf_vnet_probe(struct pci_epf *epf)
+{
+     struct epf_vnet *vnet;
+
+     vnet = devm_kzalloc(&epf->dev, sizeof(*vnet), GFP_KERNEL);
+     if (!vnet)
+             return -ENOMEM;
+
+     epf_set_drvdata(epf, vnet);
+     vnet->epf = epf;
+
+     epf_vnet_virtio_init(vnet);
+
+     return 0;
+}
+
+static struct pci_epf_driver epf_vnet_drv = {
+     .driver.name = "pci_epf_vnet",
+     .ops = &epf_vnet_ops,
+     .id_table = epf_vnet_ids,
+     .probe = epf_vnet_probe,
+     .owner = THIS_MODULE,
+};
+
+static int __init epf_vnet_init(void)
+{
+     int err;
+
+     err = pci_epf_register_driver(&epf_vnet_drv);
+     if (err) {
+             pr_err("Failed to register epf vnet driver\n");
+             return err;
+     }
+
+     return 0;
+}
+module_init(epf_vnet_init);
+
+static void epf_vnet_exit(void)
+{
+     pci_epf_unregister_driver(&epf_vnet_drv);
+}
+module_exit(epf_vnet_exit);
+
+MODULE_LICENSE("GPL");
+MODULE_AUTHOR("Shunsuke Mie <[email protected]>");
+MODULE_DESCRIPTION("PCI endpoint function acts as virtio net device");
drivers/pci/endpoint/functions/pci-epf-vnet.h new file +56 -0
new file mode 100644
index 000000000000..1e0f90c95578
@@ -0,0 +1,62 @@
+/* SPDX-License-Identifier: GPL-2.0 */
+#ifndef _PCI_EPF_VNET_H
+#define _PCI_EPF_VNET_H
+
+#include <linux/pci-epf.h>
+#include <linux/pci-epf-virtio.h>
+#include <linux/virtio_net.h>
+#include <linux/dmaengine.h>
+#include <linux/virtio.h>
+
+struct epf_vnet {
+     //TODO Should this variable be placed here?
+     struct pci_epf *epf;
+     struct virtio_net_config vnet_cfg;
+     u64 virtio_features;
+
+     // dma channels for local to remote(lr) and remote to local(rl)
+     struct dma_chan *lr_dma_chan, *rl_dma_chan;
+
+     struct {
+             void __iomem *cfg_base;
+             struct task_struct *device_setup_task;
+             struct task_struct *notify_monitor_task;
+             struct workqueue_struct *tx_wq, *irq_wq, *ctl_wq;
+             struct work_struct tx_work, raise_irq_work, ctl_work;
+             struct pci_epf_vringh *txvrh, *rxvrh, *ctlvrh;
+             struct vringh_kiov tx_iov, rx_iov, ctl_riov, ctl_wiov;
+     } rc;
+
+     struct {
+             struct virtqueue *rxvq, *txvq, *ctlvq;
+             struct vringh txvrh, rxvrh, ctlvrh;
+             struct vringh_kiov tx_iov, rx_iov, ctl_riov, ctl_wiov;
+             struct virtio_device vdev;
+             u16 net_config_status;
+     } ep;
+
+#define EPF_VNET_INIT_COMPLETE_EP BIT(0)
+#define EPF_VNET_INIT_COMPLETE_RC BIT(1)
+     u8 init_complete;
+};
+
+int epf_vnet_rc_setup(struct epf_vnet *vnet);
+void epf_vnet_rc_cleanup(struct epf_vnet *vnet);
+int epf_vnet_ep_setup(struct epf_vnet *vnet);
+void epf_vnet_ep_cleanup(struct epf_vnet *vnet);
+
+int epf_vnet_get_vq_size(void);
+int epf_vnet_init_kiov(struct vringh_kiov *kiov, const size_t vq_size);
+void epf_vnet_deinit_kiov(struct vringh_kiov *kiov);
+int epf_vnet_transfer(struct epf_vnet *vnet, struct vringh *tx_vrh,
+                   struct vringh *rx_vrh, struct vringh_kiov *tx_iov,
+                   struct vringh_kiov *rx_iov,
+                   enum dma_transfer_direction dir);
+void epf_vnet_rc_notify(struct epf_vnet *vnet);
+void epf_vnet_ep_notify(struct epf_vnet *vnet, struct virtqueue *vq);
Quoted 9 lines
Quoted 8 lines
drivers/pci/endpoint/functions/pci-epf-vnet.h new file +6 -1
+
+void epf_vnet_init_complete(struct epf_vnet *vnet, u8 from);
+void epf_vnet_ep_announce_linkup(struct epf_vnet *vnet);
+void epf_vnet_rc_announce_linkup(struct epf_vnet *vnet);
+
+#endif // _PCI_EPF_VNET_H
--

2.25.1

Best,
Shunsuke

#15 lore raw

Re: [RFC PATCH 4/4] PCI: endpoint: function: Add EP function driver to provide virtio net device

In reply to Shunsuke Mie
[If your patch is applied to the wrong git tree, kindly drop us a note. And when submitting patch, we suggest to use '--base' as documented in…

Hi Shunsuke,

[FYI, it's a private test report for your RFC patch.]
[auto build test WARNING on helgaas-pci/next]
[also build test WARNING on helgaas-pci/for-linus linus/master v6.2-rc6 next-20230203] [If your patch is applied to the wrong git tree, kindly drop us a note. And when submitting patch, we suggest to use '--base' as documented in
https://git-scm.com/docs/git-format-patch#_base_tree_information]

url:    https://github.com/intel-lab-lkp/linux/commits/Shunsuke-Mie/virtio_pci-add-a-definition-of-queue-flag-in-ISR/20230203-180631
base:   https://git.kernel.org/pub/scm/linux/kernel/git/helgaas/pci.git next
patch link:    https://lore.kernel.org/r/20230203100418.2981144-5-mie%40igel.co.jp
patch subject: [RFC PATCH 4/4] PCI: endpoint: function: Add EP function driver to provide virtio net device
config: sparc-allyesconfig (https://download.01.org/0day-ci/archive/20230203/[email protected]/config)
compiler: sparc64-linux-gcc (GCC) 12.1.0

reproduce (this is a W=1 build):
wget https://raw.githubusercontent.com/intel/lkp-tests/master/sbin/make.cross -O ~/bin/make.cross chmod +x ~/bin/make.cross
# https://github.com/intel-lab-lkp/linux/commit/a76cd5956970d04bf5e4e72da87cfdaa4da86164
git remote add linux-review https://github.com/intel-lab-lkp/linux
git fetch --no-tags linux-review Shunsuke-Mie/virtio_pci-add-a-definition-of-queue-flag-in-ISR/20230203-180631 git checkout a76cd5956970d04bf5e4e72da87cfdaa4da86164
# save the config file
mkdir build_dir && cp config build_dir/.config
COMPILER_INSTALL_PATH=$HOME/0day COMPILER=gcc-12.1.0 make.cross W=1 O=build_dir ARCH=sparc olddefconfig COMPILER_INSTALL_PATH=$HOME/0day COMPILER=gcc-12.1.0 make.cross W=1 O=build_dir ARCH=sparc SHELL=/bin/bash drivers/pci/

If you fix the issue, kindly add following tag where applicable | Reported-by: kernel test robot <[email protected]>

All warnings (new ones prefixed by >>):

drivers/pci/endpoint/functions/pci-epf-vnet-ep.c: In function 'epf_vnet_ep_process_ctrlq_entry':
drivers/pci/endpoint/functions/pci-epf-vnet-ep.c:55:15: error: implicit declaration of function 'vringh_getdesc'; did you mean 'vringh_getdesc_kern'? [-Werror=implicit-function-declaration]

      55 |         err = vringh_getdesc(vrh, riov, wiov, &head);
         |               ^~~~~~~~~~~~~~
         |               vringh_getdesc_kern
   drivers/pci/endpoint/functions/pci-epf-vnet-ep.c:91:9: error: implicit declaration of function 'vringh_complete'; did you mean 'vringh_complete_kern'? [-Werror=implicit-function-declaration]
      91 |         vringh_complete(vrh, head, len);
         |         ^~~~~~~~~~~~~~~
         |         vringh_complete_kern
   In file included from include/linux/cpumask.h:15,
                    from include/linux/smp.h:13,
                    from include/linux/lockdep.h:14,
                    from include/linux/spinlock.h:63,
                    from include/linux/kref.h:16,
                    from include/linux/configfs.h:25,
                    from include/linux/pci-epf.h:12,
                    from include/linux/pci-epc.h:12,
                    from drivers/pci/endpoint/functions/pci-epf-vnet-ep.c:5:
   drivers/pci/endpoint/functions/pci-epf-vnet-ep.c: In function 'epf_vnet_ep_vdev_find_vqs':
>> include/linux/gfp_types.h:333:25: warning: passing argument 5 of 'vringh_init_kern' makes pointer from integer without a cast [-Wint-conversion]
     333 | #define GFP_KERNEL      (__GFP_RECLAIM | __GFP_IO | __GFP_FS)
         |                         ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
         |                         |
         |                         unsigned int
   drivers/pci/endpoint/functions/pci-epf-vnet-ep.c:255:46: note: in expansion of macro 'GFP_KERNEL'
     255 |                                        true, GFP_KERNEL, vring->desc,
         |                                              ^~~~~~~~~~
   In file included from include/linux/pci-epf-virtio.h:9,
                    from drivers/pci/endpoint/functions/pci-epf-vnet.h:6,
                    from drivers/pci/endpoint/functions/pci-epf-vnet-ep.c:10:
   include/linux/vringh.h:174:41: note: expected 'struct vring_desc *' but argument is of type 'unsigned int'
     174 |                      struct vring_desc *desc,
         |                      ~~~~~~~~~~~~~~~~~~~^~~~
   drivers/pci/endpoint/functions/pci-epf-vnet-ep.c:255:63: error: passing argument 6 of 'vringh_init_kern' from incompatible pointer type [-Werror=incompatible-pointer-types]
     255 |                                        true, GFP_KERNEL, vring->desc,
         |                                                          ~~~~~^~~~~~
         |                                                               |
         |                                                               vring_desc_t * {aka struct vring_desc *}
   include/linux/vringh.h:175:42: note: expected 'struct vring_avail *' but argument is of type 'vring_desc_t *' {aka 'struct vring_desc *'}
     175 |                      struct vring_avail *avail,
         |                      ~~~~~~~~~~~~~~~~~~~~^~~~~
   drivers/pci/endpoint/functions/pci-epf-vnet-ep.c:256:45: error: passing argument 7 of 'vringh_init_kern' from incompatible pointer type [-Werror=incompatible-pointer-types]
     256 |                                        vring->avail, vring->used);
         |                                        ~~~~~^~~~~~~
         |                                             |
         |                                             vring_avail_t * {aka struct vring_avail *}
   include/linux/vringh.h:176:41: note: expected 'struct vring_used *' but argument is of type 'vring_avail_t *' {aka 'struct vring_avail *'}
     176 |                      struct vring_used *used);
         |                      ~~~~~~~~~~~~~~~~~~~^~~~
   drivers/pci/endpoint/functions/pci-epf-vnet-ep.c:254:23: error: too many arguments to function 'vringh_init_kern'
     254 |                 err = vringh_init_kern(vrh, vnet->virtio_features, vq_size,
         |                       ^~~~~~~~~~~~~~~~

include/linux/vringh.h:172:5: note: declared here
172 | int vringh_init_kern(struct vringh *vrh, u64 features,
| ^~~~~~~~~~~~~~~~
cc1: some warnings being treated as errors

vim +/vringh_init_kern +333 include/linux/gfp_types.h

cb5a065b4ea9c0 Ingo Molnar    2022-04-14  261  
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  262  /**
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  263   * DOC: Useful GFP flag combinations
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  264   *
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  265   * Useful GFP flag combinations
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  266   * ----------------------------
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  267   *
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  268   * Useful GFP flag combinations that are commonly used. It is recommended
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  269   * that subsystems start with one of these combinations and then set/clear
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  270   * %__GFP_FOO flags as necessary.
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  271   *
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  272   * %GFP_ATOMIC users can not sleep and need the allocation to succeed. A lower
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  273   * watermark is applied to allow access to "atomic reserves".
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  274   * The current implementation doesn't support NMI and few other strict
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  275   * non-preemptive contexts (e.g. raw_spin_lock). The same applies to %GFP_NOWAIT.
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  276   *
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  277   * %GFP_KERNEL is typical for kernel-internal allocations. The caller requires
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  278   * %ZONE_NORMAL or a lower zone for direct access but can direct reclaim.
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  279   *
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  280   * %GFP_KERNEL_ACCOUNT is the same as GFP_KERNEL, except the allocation is
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  281   * accounted to kmemcg.
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  282   *
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  283   * %GFP_NOWAIT is for kernel allocations that should not stall for direct
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  284   * reclaim, start physical IO or use any filesystem callback.
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  285   *
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  286   * %GFP_NOIO will use direct reclaim to discard clean pages or slab pages
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  287   * that do not require the starting of any physical IO.
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  288   * Please try to avoid using this flag directly and instead use
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  289   * memalloc_noio_{save,restore} to mark the whole scope which cannot
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  290   * perform any IO with a short explanation why. All allocation requests
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  291   * will inherit GFP_NOIO implicitly.
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  292   *
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  293   * %GFP_NOFS will use direct reclaim but will not use any filesystem interfaces.
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  294   * Please try to avoid using this flag directly and instead use
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  295   * memalloc_nofs_{save,restore} to mark the whole scope which cannot/shouldn't
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  296   * recurse into the FS layer with a short explanation why. All allocation
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  297   * requests will inherit GFP_NOFS implicitly.
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  298   *
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  299   * %GFP_USER is for userspace allocations that also need to be directly
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  300   * accessibly by the kernel or hardware. It is typically used by hardware
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  301   * for buffers that are mapped to userspace (e.g. graphics) that hardware
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  302   * still must DMA to. cpuset limits are enforced for these allocations.
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  303   *
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  304   * %GFP_DMA exists for historical reasons and should be avoided where possible.
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  305   * The flags indicates that the caller requires that the lowest zone be
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  306   * used (%ZONE_DMA or 16M on x86-64). Ideally, this would be removed but
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  307   * it would require careful auditing as some users really require it and
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  308   * others use the flag to avoid lowmem reserves in %ZONE_DMA and treat the
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  309   * lowest zone as a type of emergency reserve.
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  310   *
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  311   * %GFP_DMA32 is similar to %GFP_DMA except that the caller requires a 32-bit
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  312   * address. Note that kmalloc(..., GFP_DMA32) does not return DMA32 memory
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  313   * because the DMA32 kmalloc cache array is not implemented.
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  314   * (Reason: there is no such user in kernel).
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  315   *
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  316   * %GFP_HIGHUSER is for userspace allocations that may be mapped to userspace,
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  317   * do not need to be directly accessible by the kernel but that cannot
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  318   * move once in use. An example may be a hardware allocation that maps
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  319   * data directly into userspace but has no addressing limitations.
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  320   *
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  321   * %GFP_HIGHUSER_MOVABLE is for userspace allocations that the kernel does not
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  322   * need direct access to but can use kmap() when access is required. They
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  323   * are expected to be movable via page reclaim or page migration. Typically,
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  324   * pages on the LRU would also be allocated with %GFP_HIGHUSER_MOVABLE.
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  325   *
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  326   * %GFP_TRANSHUGE and %GFP_TRANSHUGE_LIGHT are used for THP allocations. They
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  327   * are compound allocations that will generally fail quickly if memory is not
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  328   * available and will not wake kswapd/kcompactd on failure. The _LIGHT
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  329   * version does not attempt reclaim/compaction at all and is by default used
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  330   * in page fault path, while the non-light is used by khugepaged.
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  331   */
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  332  #define GFP_ATOMIC	(__GFP_HIGH|__GFP_ATOMIC|__GFP_KSWAPD_RECLAIM)
cb5a065b4ea9c0 Ingo Molnar    2022-04-14 @333  #define GFP_KERNEL	(__GFP_RECLAIM | __GFP_IO | __GFP_FS)
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  334  #define GFP_KERNEL_ACCOUNT (GFP_KERNEL | __GFP_ACCOUNT)
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  335  #define GFP_NOWAIT	(__GFP_KSWAPD_RECLAIM)
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  336  #define GFP_NOIO	(__GFP_RECLAIM)
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  337  #define GFP_NOFS	(__GFP_RECLAIM | __GFP_IO)
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  338  #define GFP_USER	(__GFP_RECLAIM | __GFP_IO | __GFP_FS | __GFP_HARDWALL)
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  339  #define GFP_DMA		__GFP_DMA
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  340  #define GFP_DMA32	__GFP_DMA32
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  341  #define GFP_HIGHUSER	(GFP_USER | __GFP_HIGHMEM)
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  342  #define GFP_HIGHUSER_MOVABLE	(GFP_HIGHUSER | __GFP_MOVABLE | \
4e23eeebb2e57f Linus Torvalds 2022-08-07  343  			 __GFP_SKIP_KASAN_POISON | __GFP_SKIP_KASAN_UNPOISON)
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  344  #define GFP_TRANSHUGE_LIGHT	((GFP_HIGHUSER_MOVABLE | __GFP_COMP | \
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  345  			 __GFP_NOMEMALLOC | __GFP_NOWARN) & ~__GFP_RECLAIM)
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  346  #define GFP_TRANSHUGE	(GFP_TRANSHUGE_LIGHT | __GFP_DIRECT_RECLAIM)
cb5a065b4ea9c0 Ingo Molnar    2022-04-14  347  
Signature · 2 lines
0-DAY CI Kernel Test Service
https://github.com/intel/lkp-tests
#16 lore raw

Re: [RFC PATCH 4/4] PCI: endpoint: function: Add EP function driver to provide virtio net device

In reply to Shunsuke Mie
[If your patch is applied to the wrong git tree, kindly drop us a note. And when submitting patch, we suggest to use '--base' as documented in…

Hi Shunsuke,

[FYI, it's a private test report for your RFC patch.]
[auto build test WARNING on helgaas-pci/next]
[also build test WARNING on helgaas-pci/for-linus linus/master v6.2-rc6 next-20230203] [If your patch is applied to the wrong git tree, kindly drop us a note. And when submitting patch, we suggest to use '--base' as documented in
https://git-scm.com/docs/git-format-patch#_base_tree_information]

url:    https://github.com/intel-lab-lkp/linux/commits/Shunsuke-Mie/virtio_pci-add-a-definition-of-queue-flag-in-ISR/20230203-180631
base:   https://git.kernel.org/pub/scm/linux/kernel/git/helgaas/pci.git next
patch link:    https://lore.kernel.org/r/20230203100418.2981144-5-mie%40igel.co.jp
patch subject: [RFC PATCH 4/4] PCI: endpoint: function: Add EP function driver to provide virtio net device
config: arc-allyesconfig (https://download.01.org/0day-ci/archive/20230204/[email protected]/config)
compiler: arceb-elf-gcc (GCC) 12.1.0

reproduce (this is a W=1 build):
wget https://raw.githubusercontent.com/intel/lkp-tests/master/sbin/make.cross -O ~/bin/make.cross
chmod +x ~/bin/make.cross
# https://github.com/intel-lab-lkp/linux/commit/a76cd5956970d04bf5e4e72da87cfdaa4da86164
git remote add linux-review https://github.com/intel-lab-lkp/linux
git fetch --no-tags linux-review Shunsuke-Mie/virtio_pci-add-a-definition-of-queue-flag-in-ISR/20230203-180631 git checkout a76cd5956970d04bf5e4e72da87cfdaa4da86164
# save the config file
mkdir build_dir && cp config build_dir/.config
COMPILER_INSTALL_PATH=$HOME/0day COMPILER=gcc-12.1.0 make.cross W=1 O=build_dir ARCH=arc olddefconfig COMPILER_INSTALL_PATH=$HOME/0day COMPILER=gcc-12.1.0 make.cross W=1 O=build_dir ARCH=arc SHELL=/bin/bash drivers/pci/endpoint/functions/

If you fix the issue, kindly add following tag where applicable | Reported-by: kernel test robot <[email protected]>

All warnings (new ones prefixed by >>):

drivers/pci/endpoint/functions/pci-epf-vnet.c: In function 'epf_vnet_dma_callback':
drivers/pci/endpoint/functions/pci-epf-vnet.c:166:9: error: implicit declaration of function 'vringh_complete'; did you mean 'vringh_complete_kern'? [-Werror=implicit-function-declaration]

     166 |         vringh_complete(param->tx_vrh, param->tx_head, param->total_len);
         |         ^~~~~~~~~~~~~~~
         |         vringh_complete_kern
   drivers/pci/endpoint/functions/pci-epf-vnet.c: In function 'epf_vnet_transfer':
   drivers/pci/endpoint/functions/pci-epf-vnet.c:200:15: error: implicit declaration of function 'vringh_getdesc'; did you mean 'vringh_getdesc_kern'? [-Werror=implicit-function-declaration]
     200 |         err = vringh_getdesc(tx_vrh, tx_iov, NULL, &tx_head);
         |               ^~~~~~~~~~~~~~
         |               vringh_getdesc_kern
   drivers/pci/endpoint/functions/pci-epf-vnet.c:213:17: error: implicit declaration of function 'vringh_abandon'; did you mean 'vringh_abandon_kern'? [-Werror=implicit-function-declaration]
     213 |                 vringh_abandon(tx_vrh, tx_head);
         |                 ^~~~~~~~~~~~~~
         |                 vringh_abandon_kern
>> drivers/pci/endpoint/functions/pci-epf-vnet.c:251:25: warning: cast from pointer to integer of different size [-Wpointer-to-int-cast]
     251 |                 lbase = (u64)liov->iov[liov->i].iov_base;
         |                         ^
   drivers/pci/endpoint/functions/pci-epf-vnet.c:252:25: warning: cast from pointer to integer of different size [-Wpointer-to-int-cast]
     252 |                 rbase = (u64)riov->iov[riov->i].iov_base;
         |                         ^

cc1: some warnings being treated as errors
--
drivers/pci/endpoint/functions/pci-epf-vnet-rc.c: In function 'epf_vnet_rc_epc_mmap':
drivers/pci/endpoint/functions/pci-epf-vnet-rc.c:424:15: error: implicit declaration of function 'pci_epc_mem_align'; did you mean 'pci_epc_mem_exit'? [-Werror=implicit-function-declaration]

     424 |         err = pci_epc_mem_align(epf->epc, pci_addr, len, &aaddr, &asize);
         |               ^~~~~~~~~~~~~~~~~
         |               pci_epc_mem_exit
   drivers/pci/endpoint/functions/pci-epf-vnet-rc.c: In function 'epf_vnet_rc_process_ctrlq_entry':
   drivers/pci/endpoint/functions/pci-epf-vnet-rc.c:489:15: error: implicit declaration of function 'vringh_getdesc'; did you mean 'vringh_getdesc_kern'? [-Werror=implicit-function-declaration]
     489 |         err = vringh_getdesc(vrh, riov, wiov, &head);
         |               ^~~~~~~~~~~~~~
         |               vringh_getdesc_kern
>> drivers/pci/endpoint/functions/pci-epf-vnet-rc.c:495:42: warning: cast from pointer to integer of different size [-Wpointer-to-int-cast]
     495 |         rmem = epf_vnet_rc_epc_mmap(epf, (u64)riov->iov[riov->i].iov_base,
         |                                          ^
   drivers/pci/endpoint/functions/pci-epf-vnet-rc.c:502:42: warning: cast from pointer to integer of different size [-Wpointer-to-int-cast]
     502 |         wmem = epf_vnet_rc_epc_mmap(epf, (u64)wiov->iov[wiov->i].iov_base,
         |                                          ^
   drivers/pci/endpoint/functions/pci-epf-vnet-rc.c:535:9: error: implicit declaration of function 'vringh_complete'; did you mean 'vringh_complete_kern'? [-Werror=implicit-function-declaration]
     535 |         vringh_complete(vrh, head, total_len);
         |         ^~~~~~~~~~~~~~~
         |         vringh_complete_kern
   drivers/pci/endpoint/functions/pci-epf-vnet-rc.c:542:9: error: implicit declaration of function 'vringh_abandon'; did you mean 'vringh_abandon_kern'? [-Werror=implicit-function-declaration]
     542 |         vringh_abandon(vrh, head);
         |         ^~~~~~~~~~~~~~
         |         vringh_abandon_kern

cc1: some warnings being treated as errors
--
drivers/pci/endpoint/functions/pci-epf-vnet-ep.c: In function 'epf_vnet_ep_process_ctrlq_entry':
drivers/pci/endpoint/functions/pci-epf-vnet-ep.c:55:15: error: implicit declaration of function 'vringh_getdesc'; did you mean 'vringh_getdesc_kern'? [-Werror=implicit-function-declaration]

      55 |         err = vringh_getdesc(vrh, riov, wiov, &head);
         |               ^~~~~~~~~~~~~~
         |               vringh_getdesc_kern
   In file included from include/asm-generic/bug.h:22,
                    from arch/arc/include/asm/bug.h:30,
                    from include/linux/bug.h:5,
                    from include/linux/thread_info.h:13,
                    from include/asm-generic/preempt.h:5,
                    from ./arch/arc/include/generated/asm/preempt.h:1,
                    from include/linux/preempt.h:78,
                    from include/linux/spinlock.h:56,
                    from include/linux/kref.h:16,
                    from include/linux/configfs.h:25,
                    from include/linux/pci-epf.h:12,
                    from include/linux/pci-epc.h:12,
                    from drivers/pci/endpoint/functions/pci-epf-vnet-ep.c:5:
>> drivers/pci/endpoint/functions/pci-epf-vnet-ep.c:61:26: warning: format '%ld' expects argument of type 'long int', but argument 3 has type 'size_t' {aka 'unsigned int'} [-Wformat=]
      61 |                 pr_debug("Command is too short: %ld\n", len);
         |                          ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~
   include/linux/printk.h:347:21: note: in definition of macro 'pr_fmt'
     347 | #define pr_fmt(fmt) fmt
         |                     ^~~
   include/linux/dynamic_debug.h:247:9: note: in expansion of macro '__dynamic_func_call_cls'
     247 |         __dynamic_func_call_cls(__UNIQUE_ID(ddebug), cls, fmt, func, ##__VA_ARGS__)
         |         ^~~~~~~~~~~~~~~~~~~~~~~
   include/linux/dynamic_debug.h:249:9: note: in expansion of macro '_dynamic_func_call_cls'
     249 |         _dynamic_func_call_cls(_DPRINTK_CLASS_DFLT, fmt, func, ##__VA_ARGS__)
         |         ^~~~~~~~~~~~~~~~~~~~~~
   include/linux/dynamic_debug.h:268:9: note: in expansion of macro '_dynamic_func_call'
     268 |         _dynamic_func_call(fmt, __dynamic_pr_debug,             \
         |         ^~~~~~~~~~~~~~~~~~
   include/linux/printk.h:581:9: note: in expansion of macro 'dynamic_pr_debug'
     581 |         dynamic_pr_debug(fmt, ##__VA_ARGS__)
         |         ^~~~~~~~~~~~~~~~
   drivers/pci/endpoint/functions/pci-epf-vnet-ep.c:61:17: note: in expansion of macro 'pr_debug'
      61 |                 pr_debug("Command is too short: %ld\n", len);
         |                 ^~~~~~~~
   drivers/pci/endpoint/functions/pci-epf-vnet-ep.c:61:51: note: format string is defined here
      61 |                 pr_debug("Command is too short: %ld\n", len);
         |                                                 ~~^
         |                                                   |
         |                                                   long int
         |                                                 %d
   drivers/pci/endpoint/functions/pci-epf-vnet-ep.c:91:9: error: implicit declaration of function 'vringh_complete'; did you mean 'vringh_complete_kern'? [-Werror=implicit-function-declaration]
      91 |         vringh_complete(vrh, head, len);
         |         ^~~~~~~~~~~~~~~
         |         vringh_complete_kern
   In file included from include/linux/cpumask.h:15,
                    from include/linux/smp.h:13,
                    from include/linux/lockdep.h:14,
                    from include/linux/spinlock.h:63:
   drivers/pci/endpoint/functions/pci-epf-vnet-ep.c: In function 'epf_vnet_ep_vdev_find_vqs':
   include/linux/gfp_types.h:333:25: warning: passing argument 5 of 'vringh_init_kern' makes pointer from integer without a cast [-Wint-conversion]
     333 | #define GFP_KERNEL      (__GFP_RECLAIM | __GFP_IO | __GFP_FS)
         |                         ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
         |                         |
         |                         unsigned int
   drivers/pci/endpoint/functions/pci-epf-vnet-ep.c:255:46: note: in expansion of macro 'GFP_KERNEL'
     255 |                                        true, GFP_KERNEL, vring->desc,
         |                                              ^~~~~~~~~~
   In file included from include/linux/pci-epf-virtio.h:9,
                    from drivers/pci/endpoint/functions/pci-epf-vnet.h:6,
                    from drivers/pci/endpoint/functions/pci-epf-vnet-ep.c:10:
   include/linux/vringh.h:174:41: note: expected 'struct vring_desc *' but argument is of type 'unsigned int'
     174 |                      struct vring_desc *desc,
         |                      ~~~~~~~~~~~~~~~~~~~^~~~
   drivers/pci/endpoint/functions/pci-epf-vnet-ep.c:255:63: error: passing argument 6 of 'vringh_init_kern' from incompatible pointer type [-Werror=incompatible-pointer-types]
     255 |                                        true, GFP_KERNEL, vring->desc,
         |                                                          ~~~~~^~~~~~
         |                                                               |
         |                                                               vring_desc_t * {aka struct vring_desc *}
   include/linux/vringh.h:175:42: note: expected 'struct vring_avail *' but argument is of type 'vring_desc_t *' {aka 'struct vring_desc *'}
     175 |                      struct vring_avail *avail,
         |                      ~~~~~~~~~~~~~~~~~~~~^~~~~
   drivers/pci/endpoint/functions/pci-epf-vnet-ep.c:256:45: error: passing argument 7 of 'vringh_init_kern' from incompatible pointer type [-Werror=incompatible-pointer-types]
     256 |                                        vring->avail, vring->used);
         |                                        ~~~~~^~~~~~~
         |                                             |
         |                                             vring_avail_t * {aka struct vring_avail *}
   include/linux/vringh.h:176:41: note: expected 'struct vring_used *' but argument is of type 'vring_avail_t *' {aka 'struct vring_avail *'}
     176 |                      struct vring_used *used);
         |                      ~~~~~~~~~~~~~~~~~~~^~~~
   drivers/pci/endpoint/functions/pci-epf-vnet-ep.c:254:23: error: too many arguments to function 'vringh_init_kern'
     254 |                 err = vringh_init_kern(vrh, vnet->virtio_features, vq_size,
         |                       ^~~~~~~~~~~~~~~~

include/linux/vringh.h:172:5: note: declared here
172 | int vringh_init_kern(struct vringh *vrh, u64 features,
| ^~~~~~~~~~~~~~~~
cc1: some warnings being treated as errors

vim +251 drivers/pci/endpoint/functions/pci-epf-vnet.c

174
175 /**
176 * epf_vnet_transfer() - transfer data between tx vring to rx vring using edma
177 * @vnet: epf virtio net device to do dma
178 * @tx_vrh: vringh related to source tx vring
179 * @rx_vrh: vringh related to target rx vring
180 * @tx_iov: buffer to use tx
181 * @rx_iov: buffer to use rx
182 * @dir: a direction of DMA. local to remote or local from remote
183 *
184 * This function returns 0, 1 or error number. The 0 indicates there is not
185 * data to send. The 1 indicates a request to DMA is succeeded. Other error
186 * numbers shows error, however, ENOSPC means there is no buffer on target
187 * vring, so should retry to call later.
188 */
189 int epf_vnet_transfer(struct epf_vnet *vnet, struct vringh *tx_vrh,

   190			      struct vringh *rx_vrh, struct vringh_kiov *tx_iov,
   191			      struct vringh_kiov *rx_iov,
   192			      enum dma_transfer_direction dir)
   193	{
   194		int err;
   195		u16 tx_head, rx_head;
   196		size_t total_tx_len;
   197		struct epf_vnet_dma_callback_param *cb_param;
   198		struct vringh_kiov *liov, *riov;
   199	
   200		err = vringh_getdesc(tx_vrh, tx_iov, NULL, &tx_head);
   201		if (err <= 0)
   202			return err;
   203	
   204		total_tx_len = vringh_kiov_length(tx_iov);
   205	
   206		err = vringh_getdesc(rx_vrh, NULL, rx_iov, &rx_head);
   207		if (err < 0) {
   208			goto err_tx_complete;
   209		} else if (!err) {
   210			/* There is not space on a vring of destination to transmit data, so
   211			 * rollback tx vringh
   212			 */
   213			vringh_abandon(tx_vrh, tx_head);
   214			return -ENOSPC;
   215		}
   216	
   217		cb_param = kmalloc(sizeof(*cb_param), GFP_KERNEL);
   218		if (!cb_param) {
   219			err = -ENOMEM;
   220			goto err_rx_complete;
   221		}
   222	
   223		cb_param->tx_vrh = tx_vrh;
   224		cb_param->rx_vrh = rx_vrh;
   225		cb_param->tx_head = tx_head;
   226		cb_param->rx_head = rx_head;
   227		cb_param->total_len = total_tx_len;
   228		cb_param->vnet = vnet;
   229	
   230		switch (dir) {
   231		case DMA_MEM_TO_DEV:
   232			liov = tx_iov;
   233			riov = rx_iov;
   234			cb_param->vq = vnet->ep.txvq;
   235			break;
   236		case DMA_DEV_TO_MEM:
   237			liov = rx_iov;
   238			riov = tx_iov;
   239			cb_param->vq = vnet->ep.rxvq;
   240			break;
   241		default:
   242			err = -EINVAL;
   243			goto err_free_param;
   244		}
   245	
   246		for (; tx_iov->i < tx_iov->used; tx_iov->i++, rx_iov->i++) {
   247			size_t len;
   248			u64 lbase, rbase;
   249			void (*callback)(void *) = NULL;
   250	
 > 251			lbase = (u64)liov->iov[liov->i].iov_base;

252 rbase = (u64)riov->iov[riov->i].iov_base;
253 len = tx_iov->iov[tx_iov->i].iov_len;
254
255 if (tx_iov->i + 1 == tx_iov->used)
256 callback = epf_vnet_dma_callback;
257
258 err = epf_vnet_dma_single(vnet, rbase, lbase, len, callback,
259 cb_param, dir);
260 if (err)
261 goto err_free_param;
262 }
263
264 return 1;
265
266 err_free_param:
267 kfree(cb_param);
268 err_rx_complete:
269 vringh_complete(rx_vrh, rx_head, vringh_kiov_length(rx_iov));
270 err_tx_complete:
271 vringh_complete(tx_vrh, tx_head, total_tx_len);
272
273 return err;
274 }
275

Signature · 2 lines
0-DAY CI Kernel Test Service
https://github.com/intel/lkp-tests
#17 lore raw
In reply to Shunsuke Mie
Quoted 17 lines

-----Original Message-----
From: Shunsuke Mie <[email protected]>
Sent: Friday, February 3, 2023 4:04 AM
To: Lorenzo Pieralisi <[email protected]>
Cc: Krzysztof Wilczyński <[email protected]>; Manivannan Sadhasivam <[email protected]>; Kishon Vijay Abraham I <[email protected]>; Bjorn Helgaas <[email protected]>; Michael S. Tsirkin <[email protected]>; Jason Wang <[email protected]>; Shunsuke Mie <[email protected]>; Frank Li <[email protected]>; Jon Mason <[email protected]>; Ren Zhijie <[email protected]>; Takanari Hayama <[email protected]>; linux- [email protected]; [email protected]; [email protected] foundation.org
Subject: [EXT] [RFC PATCH 0/4] PCI: endpoint: Introduce a virtio-net EP function

Quoted 7 lines

Caution: EXT Email

This patchset introduce a virtio-net EP device function. It provides a new option to communiate between PCIe host and endpoint over IP. Advantage of this option is that the driver fully uses a PCIe embedded DMA. It is used to transport data between virtio ring directly each other. It can be expected to better throughput.

Thanks, basic that's what I want. I am trying use RDMA. But I think virtio-net still be good solution.

Frank Li

Quoted 75 lines

To realize the function, this patchset has few changes and introduces a new APIs to PCI EP framework related to virtio. Furthermore, it device depends on the some patchtes that is discussing. Those depended patchset are following:
- [PATCH 1/2] dmaengine: dw-edma: Fix to change for continuous transfer link:
https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.k
ernel.org%2Fdmaengine%2F20221223022608.550697-1- mie%40igel.co.jp%2F&data=05%7C01%7CFrank.Li%40nxp.com%7Cac57a62d4 10b458a5ba408db05ce0a4e%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C0% 7C0%7C638110154722945380%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4 wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7 C%7C%7C&sdata=tIn0MHzEvrdxaC4KKTvTRvYXBzQ6MyrFa2GXpa3ePv0%3D& reserved=0
- [RFC PATCH 0/3] Deal with alignment restriction on EP side link:
https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.k
ernel.org%2Flinux-pci%2F20230113090350.1103494-1- mie%40igel.co.jp%2F&data=05%7C01%7CFrank.Li%40nxp.com%7Cac57a62d4 10b458a5ba408db05ce0a4e%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C0% 7C0%7C638110154722945380%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4 wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7 C%7C%7C&sdata=RLpnDiLwfqQd5QMXdiQyPVCkfOj8q2AyVeZOwWHvlsM%3 D&reserved=0
- [RFC PATCH v2 0/7] Introduce a vringh accessor for IO memory link:
https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.k
ernel.org%2Fvirtualization%2F20230202090934.549556-1- mie%40igel.co.jp%2F&data=05%7C01%7CFrank.Li%40nxp.com%7Cac57a62d4 10b458a5ba408db05ce0a4e%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C0% 7C0%7C638110154722945380%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4 wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7 C%7C%7C&sdata=6jgY76BMSbvamb%2Fl3Urjt4Gcizeqon%2BZE5nPssc2kDA% 3D&reserved=0

About this patchset has 4 patches. The first of two patch is little changes to virtio. The third patch add APIs to easily access virtio data structure on PCIe Host side memory. The last one introduce a virtio-net EP device function. Details are in commit respectively.

Currently those network devices are testd using ping only. I'll add a result of performance evaluation using iperf and etc to the future version of this patchset.

Shunsuke Mie (4):
virtio_pci: add a definition of queue flag in ISR
virtio_ring: remove const from vring getter
PCI: endpoint: Introduce virtio library for EP functions
PCI: endpoint: function: Add EP function driver to provide virtio net
device

 drivers/pci/endpoint/Kconfig                  |   7 +
 drivers/pci/endpoint/Makefile                 |   1 +
 drivers/pci/endpoint/functions/Kconfig        |  12 +
 drivers/pci/endpoint/functions/Makefile       |   1 +
 .../pci/endpoint/functions/pci-epf-vnet-ep.c  | 343 ++++++++++
 .../pci/endpoint/functions/pci-epf-vnet-rc.c  | 635 ++++++++++++++++++
 drivers/pci/endpoint/functions/pci-epf-vnet.c | 387 +++++++++++
 drivers/pci/endpoint/functions/pci-epf-vnet.h |  62 ++
 drivers/pci/endpoint/pci-epf-virtio.c         | 113 ++++
 drivers/virtio/virtio_ring.c                  |   2 +-
 include/linux/pci-epf-virtio.h                |  25 +
 include/linux/virtio.h                        |   2 +-
 include/uapi/linux/virtio_pci.h               |   2 +
 13 files changed, 1590 insertions(+), 2 deletions(-)
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet-ep.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet-rc.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet.h
 create mode 100644 drivers/pci/endpoint/pci-epf-virtio.c
 create mode 100644 include/linux/pci-epf-virtio.h

--
2.25.1

#18 lore raw

2023年2月4日(土) 1:45 Frank Li <[email protected]>:

Quoted 20 lines
Quoted 20 lines

-----Original Message-----
From: Shunsuke Mie <[email protected]>
Sent: Friday, February 3, 2023 4:04 AM
To: Lorenzo Pieralisi <[email protected]>
Cc: Krzysztof Wilczyński <[email protected]>; Manivannan Sadhasivam <[email protected]>; Kishon Vijay Abraham I <[email protected]>; Bjorn Helgaas <[email protected]>; Michael S. Tsirkin <[email protected]>; Jason Wang <[email protected]>; Shunsuke Mie <[email protected]>; Frank Li <[email protected]>; Jon Mason <[email protected]>; Ren Zhijie <[email protected]>; Takanari Hayama <[email protected]>; linux- [email protected]; [email protected]; [email protected] foundation.org
Subject: [EXT] [RFC PATCH 0/4] PCI: endpoint: Introduce a virtio-net EP function

Caution: EXT Email

Quoted 8 lines
Quoted 5 lines

This patchset introduce a virtio-net EP device function. It provides a new option to communiate between PCIe host and endpoint over IP. Advantage of this option is that the driver fully uses a PCIe embedded DMA. It is used to transport data between virtio ring directly each other. It can be expected to better throughput.

Thanks, basic that's what I want. I am trying use RDMA. But I think virtio-net still be good solution.

We project extending this module to support RDMA. The plan is based on virtio-rdma[1].
It extends the virtio-net and we are plan to implement the proposed spec based on this patch.
[1] virtio-rdma
- proposal:
https://lore.kernel.org/all/[email protected]/T/
- presentation on kvm forum:
https://youtu.be/Qrhv6hC_YK4

Please feel free to comment and suggest.

Quoted 78 lines

Frank Li

Quoted 75 lines

To realize the function, this patchset has few changes and introduces a new APIs to PCI EP framework related to virtio. Furthermore, it device depends on the some patchtes that is discussing. Those depended patchset are following:
- [PATCH 1/2] dmaengine: dw-edma: Fix to change for continuous transfer link:
https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.k
ernel.org%2Fdmaengine%2F20221223022608.550697-1- mie%40igel.co.jp%2F&data=05%7C01%7CFrank.Li%40nxp.com%7Cac57a62d4 10b458a5ba408db05ce0a4e%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C0% 7C0%7C638110154722945380%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4 wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7 C%7C%7C&sdata=tIn0MHzEvrdxaC4KKTvTRvYXBzQ6MyrFa2GXpa3ePv0%3D& reserved=0
- [RFC PATCH 0/3] Deal with alignment restriction on EP side link:
https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.k
ernel.org%2Flinux-pci%2F20230113090350.1103494-1- mie%40igel.co.jp%2F&data=05%7C01%7CFrank.Li%40nxp.com%7Cac57a62d4 10b458a5ba408db05ce0a4e%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C0% 7C0%7C638110154722945380%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4 wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7 C%7C%7C&sdata=RLpnDiLwfqQd5QMXdiQyPVCkfOj8q2AyVeZOwWHvlsM%3 D&reserved=0
- [RFC PATCH v2 0/7] Introduce a vringh accessor for IO memory link:
https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.k
ernel.org%2Fvirtualization%2F20230202090934.549556-1- mie%40igel.co.jp%2F&data=05%7C01%7CFrank.Li%40nxp.com%7Cac57a62d4 10b458a5ba408db05ce0a4e%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C0% 7C0%7C638110154722945380%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4 wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7 C%7C%7C&sdata=6jgY76BMSbvamb%2Fl3Urjt4Gcizeqon%2BZE5nPssc2kDA% 3D&reserved=0

About this patchset has 4 patches. The first of two patch is little changes to virtio. The third patch add APIs to easily access virtio data structure on PCIe Host side memory. The last one introduce a virtio-net EP device function. Details are in commit respectively.

Currently those network devices are testd using ping only. I'll add a result of performance evaluation using iperf and etc to the future version of this patchset.

Shunsuke Mie (4):
virtio_pci: add a definition of queue flag in ISR
virtio_ring: remove const from vring getter
PCI: endpoint: Introduce virtio library for EP functions
PCI: endpoint: function: Add EP function driver to provide virtio net
device

 drivers/pci/endpoint/Kconfig                  |   7 +
 drivers/pci/endpoint/Makefile                 |   1 +
 drivers/pci/endpoint/functions/Kconfig        |  12 +
 drivers/pci/endpoint/functions/Makefile       |   1 +
 .../pci/endpoint/functions/pci-epf-vnet-ep.c  | 343 ++++++++++
 .../pci/endpoint/functions/pci-epf-vnet-rc.c  | 635 ++++++++++++++++++
 drivers/pci/endpoint/functions/pci-epf-vnet.c | 387 +++++++++++
 drivers/pci/endpoint/functions/pci-epf-vnet.h |  62 ++
 drivers/pci/endpoint/pci-epf-virtio.c         | 113 ++++
 drivers/virtio/virtio_ring.c                  |   2 +-
 include/linux/pci-epf-virtio.h                |  25 +
 include/linux/virtio.h                        |   2 +-
 include/uapi/linux/virtio_pci.h               |   2 +
 13 files changed, 1590 insertions(+), 2 deletions(-)
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet-ep.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet-rc.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet.h
 create mode 100644 drivers/pci/endpoint/pci-epf-virtio.c
 create mode 100644 include/linux/pci-epf-virtio.h

--
2.25.1

Best,
Shunsuke
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

#19 lore raw
Quoted 15 lines

We project extending this module to support RDMA. The plan is based on virtio-rdma[1].
It extends the virtio-net and we are plan to implement the proposed spec based on this patch.
[1] virtio-rdma
- proposal:
https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.k
ernel.org%2Fall%2F20220511095900.343-1- xieyongji%40bytedance.com%2FT%2F&data=05%7C01%7Cfrank.li%40nxp.co m%7C0ef2bd62eda945c413be08db08f62ba3%7C686ea1d3bc2b4c6fa92cd99c5 c301635%7C0%7C0%7C638113625610341574%7CUnknown%7CTWFpbGZsb3d 8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0% 3D%7C3000%7C%7C%7C&sdata=HyhpRTG8MNx%2BtfmWn6x3srmdBjHcZAo 2qbxL9USph9o%3D&reserved=0

Quoted 9 lines

- presentation on kvm forum:
https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fyout
u.be%2FQrhv6hC_YK4&data=05%7C01%7Cfrank.li%40nxp.com%7C0ef2bd62 eda945c413be08db08f62ba3%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C0% 7C0%7C638113625610341574%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4 wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7 C%7C%7C&sdata=ucOsGR1letTjxf0gKN6uls5y951CXaIspZtLGnASEC8%3D&res erved=0

Sorry for our outlook client always change link. This previous discussion.
https://lore.kernel.org/imx/[email protected]/T/#t

Look like Endpoint maintainer Kishon like endpoint side work as vhost. Previous Haotian Wang submit similar patches, which just not use eDMA, just use memcpy. But overall idea is the same.

I think your and haotian's method is more reasonable for PCI-RC EP connection.

Kishon is not active recently. Maybe need Lorenzo Pieralisi and Bjorn helgass's comments
for overall directions.

Frank Li

Quoted 114 lines

Please feel free to comment and suggest.

Quoted 6 lines

Frank Li

Quoted 4 lines

To realize the function, this patchset has few changes and introduces a new APIs to PCI EP framework related to virtio. Furthermore, it device depends on the some patchtes that is discussing. Those depended

patchset

Quoted 2 lines
Quoted 2 lines

are following:
- [PATCH 1/2] dmaengine: dw-edma: Fix to change for continuous

transfer

Quoted 2 lines
Quoted 2 lines

link:

https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.k
%2F&data=05%7C01%7Cfrank.li%40nxp.com%7C0ef2bd62eda945c413be08db 08f62ba3%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C0%7C0%7C6381136256 10341574%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoi V2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=d VZMaheX3eR1xA2wQtecmT857h2%2BFtUbhDSHXwgvsEY%3D&reserved=0

Quoted 2 lines
Quoted 2 lines

ernel.org%2Fdmaengine%2F20221223022608.550697-1-

mie%40igel.co.jp%2F&data=05%7C01%7CFrank.Li%40nxp.com%7Cac57a62d4

Quoted 2 lines
Quoted 2 lines

10b458a5ba408db05ce0a4e%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C0%

7C0%7C638110154722945380%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4

Quoted 1 line
Quoted 1 line

wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7

Quoted 1 line
Quoted 1 line

C%7C%7C&sdata=tIn0MHzEvrdxaC4KKTvTRvYXBzQ6MyrFa2GXpa3ePv0%3D&

Quoted 4 lines
Quoted 4 lines

reserved=0
- [RFC PATCH 0/3] Deal with alignment restriction on EP side link:

https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.k
%2F&data=05%7C01%7Cfrank.li%40nxp.com%7C0ef2bd62eda945c413be08db 08f62ba3%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C0%7C0%7C6381136256 10341574%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoi V2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=d VZMaheX3eR1xA2wQtecmT857h2%2BFtUbhDSHXwgvsEY%3D&reserved=0

Quoted 2 lines
Quoted 2 lines

ernel.org%2Flinux-pci%2F20230113090350.1103494-1-

mie%40igel.co.jp%2F&data=05%7C01%7CFrank.Li%40nxp.com%7Cac57a62d4

Quoted 2 lines
Quoted 2 lines

10b458a5ba408db05ce0a4e%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C0%

7C0%7C638110154722945380%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4

Quoted 1 line
Quoted 1 line

wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7

Quoted 1 line
Quoted 1 line

C%7C%7C&sdata=RLpnDiLwfqQd5QMXdiQyPVCkfOj8q2AyVeZOwWHvlsM%3

Quoted 4 lines
Quoted 4 lines

D&reserved=0
- [RFC PATCH v2 0/7] Introduce a vringh accessor for IO memory link:

https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.k
%2F&data=05%7C01%7Cfrank.li%40nxp.com%7C0ef2bd62eda945c413be08db 08f62ba3%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C0%7C0%7C6381136256 10341574%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoi V2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=d VZMaheX3eR1xA2wQtecmT857h2%2BFtUbhDSHXwgvsEY%3D&reserved=0

Quoted 2 lines
Quoted 2 lines

ernel.org%2Fvirtualization%2F20230202090934.549556-1-

mie%40igel.co.jp%2F&data=05%7C01%7CFrank.Li%40nxp.com%7Cac57a62d4

Quoted 2 lines
Quoted 2 lines

10b458a5ba408db05ce0a4e%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C0%

7C0%7C638110154722945380%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4

Quoted 1 line
Quoted 1 line

wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7

Quoted 1 line
Quoted 1 line

C%7C%7C&sdata=6jgY76BMSbvamb%2Fl3Urjt4Gcizeqon%2BZE5nPssc2kDA%

Quoted 18 lines
Quoted 18 lines

3D&reserved=0

About this patchset has 4 patches. The first of two patch is little changes to virtio. The third patch add APIs to easily access virtio data structure on PCIe Host side memory. The last one introduce a virtio-net EP device function. Details are in commit respectively.

Currently those network devices are testd using ping only. I'll add a result of performance evaluation using iperf and etc to the future version of this patchset.

Shunsuke Mie (4):
virtio_pci: add a definition of queue flag in ISR
virtio_ring: remove const from vring getter
PCI: endpoint: Introduce virtio library for EP functions
PCI: endpoint: function: Add EP function driver to provide virtio net
device

Quoted 6 lines
Quoted 6 lines
 drivers/pci/endpoint/Kconfig                  |   7 +
 drivers/pci/endpoint/Makefile                 |   1 +
 drivers/pci/endpoint/functions/Kconfig        |  12 +
 drivers/pci/endpoint/functions/Makefile       |   1 +
 .../pci/endpoint/functions/pci-epf-vnet-ep.c  | 343 ++++++++++
 .../pci/endpoint/functions/pci-epf-vnet-rc.c  | 635

++++++++++++++++++

Quoted 15 lines
Quoted 15 lines
 drivers/pci/endpoint/functions/pci-epf-vnet.c | 387 +++++++++++
 drivers/pci/endpoint/functions/pci-epf-vnet.h |  62 ++
 drivers/pci/endpoint/pci-epf-virtio.c         | 113 ++++
 drivers/virtio/virtio_ring.c                  |   2 +-
 include/linux/pci-epf-virtio.h                |  25 +
 include/linux/virtio.h                        |   2 +-
 include/uapi/linux/virtio_pci.h               |   2 +
 13 files changed, 1590 insertions(+), 2 deletions(-)
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet-ep.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet-rc.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet.h
 create mode 100644 drivers/pci/endpoint/pci-epf-virtio.c
 create mode 100644 include/linux/pci-epf-virtio.h
Quoted 3 lines
Quoted 2 lines

--
2.25.1

Best,
Shunsuke

#20 lore raw
Quoted 26 lines · Frank Li
Quoted 14 lines

We project extending this module to support RDMA. The plan is based on virtio-rdma[1].
It extends the virtio-net and we are plan to implement the proposed spec based on this patch.
[1] virtio-rdma
- proposal:
https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.k
ernel.org%2Fall%2F20220511095900.343-1- xieyongji%40bytedance.com%2FT%2F&data=05%7C01%7Cfrank.li%40nxp.co m%7C0ef2bd62eda945c413be08db08f62ba3%7C686ea1d3bc2b4c6fa92cd99c5 c301635%7C0%7C0%7C638113625610341574%7CUnknown%7CTWFpbGZsb3d 8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0% 3D%7C3000%7C%7C%7C&sdata=HyhpRTG8MNx%2BtfmWn6x3srmdBjHcZAo 2qbxL9USph9o%3D&reserved=0

Quoted 9 lines

- presentation on kvm forum:
https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fyout
u.be%2FQrhv6hC_YK4&data=05%7C01%7Cfrank.li%40nxp.com%7C0ef2bd62 eda945c413be08db08f62ba3%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C0% 7C0%7C638113625610341574%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4 wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7 C%7C%7C&sdata=ucOsGR1letTjxf0gKN6uls5y951CXaIspZtLGnASEC8%3D&res erved=0

Sorry for our outlook client always change link. This previous discussion.
https://lore.kernel.org/imx/[email protected]/T/#t

Quoted 8 lines

Look like Endpoint maintainer Kishon like endpoint side work as vhost. Previous Haotian Wang submit similar patches, which just not use eDMA, just use memcpy. But overall idea is the same.

I think your and haotian's method is more reasonable for PCI-RC EP connection.

Kishon is not active recently. Maybe need Lorenzo Pieralisi and Bjorn helgass's comments
for overall directions.

I think so too. Thank you for your summarization. I've commented on the e-mail.

Quoted 108 lines

Frank Li

Quoted 106 lines

Please feel free to comment and suggest.

Quoted 5 lines

Frank Li

Quoted 3 lines

To realize the function, this patchset has few changes and introduces a new APIs to PCI EP framework related to virtio. Furthermore, it device depends on the some patchtes that is discussing. Those depended

patchset

Quoted 2 lines
Quoted 2 lines

are following:
- [PATCH 1/2] dmaengine: dw-edma: Fix to change for continuous

transfer

Quoted 2 lines
Quoted 2 lines

link:

https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.k
%2F&data=05%7C01%7Cfrank.li%40nxp.com%7C0ef2bd62eda945c413be08db 08f62ba3%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C0%7C0%7C6381136256 10341574%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoi V2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=d VZMaheX3eR1xA2wQtecmT857h2%2BFtUbhDSHXwgvsEY%3D&reserved=0

Quoted 2 lines
Quoted 2 lines

ernel.org%2Fdmaengine%2F20221223022608.550697-1-

mie%40igel.co.jp%2F&data=05%7C01%7CFrank.Li%40nxp.com%7Cac57a62d4

Quoted 2 lines
Quoted 2 lines

10b458a5ba408db05ce0a4e%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C0%

7C0%7C638110154722945380%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4 wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7 C%7C%7C&sdata=tIn0MHzEvrdxaC4KKTvTRvYXBzQ6MyrFa2GXpa3ePv0%3D&

Quoted 4 lines
Quoted 4 lines

reserved=0
- [RFC PATCH 0/3] Deal with alignment restriction on EP side link:

https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.k
%2F&data=05%7C01%7Cfrank.li%40nxp.com%7C0ef2bd62eda945c413be08db 08f62ba3%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C0%7C0%7C6381136256 10341574%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoi V2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=d VZMaheX3eR1xA2wQtecmT857h2%2BFtUbhDSHXwgvsEY%3D&reserved=0

Quoted 2 lines
Quoted 2 lines

ernel.org%2Flinux-pci%2F20230113090350.1103494-1-

mie%40igel.co.jp%2F&data=05%7C01%7CFrank.Li%40nxp.com%7Cac57a62d4

Quoted 2 lines
Quoted 2 lines

10b458a5ba408db05ce0a4e%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C0%

7C0%7C638110154722945380%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4 wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7 C%7C%7C&sdata=RLpnDiLwfqQd5QMXdiQyPVCkfOj8q2AyVeZOwWHvlsM%3

Quoted 4 lines
Quoted 4 lines

D&reserved=0
- [RFC PATCH v2 0/7] Introduce a vringh accessor for IO memory link:

https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Flore.k
%2F&data=05%7C01%7Cfrank.li%40nxp.com%7C0ef2bd62eda945c413be08db 08f62ba3%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C0%7C0%7C6381136256 10341574%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoi V2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=d VZMaheX3eR1xA2wQtecmT857h2%2BFtUbhDSHXwgvsEY%3D&reserved=0

Quoted 2 lines
Quoted 2 lines

ernel.org%2Fvirtualization%2F20230202090934.549556-1-

mie%40igel.co.jp%2F&data=05%7C01%7CFrank.Li%40nxp.com%7Cac57a62d4

Quoted 2 lines
Quoted 2 lines

10b458a5ba408db05ce0a4e%7C686ea1d3bc2b4c6fa92cd99c5c301635%7C0%

7C0%7C638110154722945380%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4 wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7 C%7C%7C&sdata=6jgY76BMSbvamb%2Fl3Urjt4Gcizeqon%2BZE5nPssc2kDA%

Quoted 18 lines
Quoted 18 lines

3D&reserved=0

About this patchset has 4 patches. The first of two patch is little changes to virtio. The third patch add APIs to easily access virtio data structure on PCIe Host side memory. The last one introduce a virtio-net EP device function. Details are in commit respectively.

Currently those network devices are testd using ping only. I'll add a result of performance evaluation using iperf and etc to the future version of this patchset.

Shunsuke Mie (4):
virtio_pci: add a definition of queue flag in ISR
virtio_ring: remove const from vring getter
PCI: endpoint: Introduce virtio library for EP functions
PCI: endpoint: function: Add EP function driver to provide virtio net
device

Quoted 6 lines
Quoted 6 lines
  drivers/pci/endpoint/Kconfig                  |   7 +
  drivers/pci/endpoint/Makefile                 |   1 +
  drivers/pci/endpoint/functions/Kconfig        |  12 +
  drivers/pci/endpoint/functions/Makefile       |   1 +
  .../pci/endpoint/functions/pci-epf-vnet-ep.c  | 343 ++++++++++
  .../pci/endpoint/functions/pci-epf-vnet-rc.c  | 635

++++++++++++++++++

Quoted 15 lines
Quoted 15 lines
  drivers/pci/endpoint/functions/pci-epf-vnet.c | 387 +++++++++++
  drivers/pci/endpoint/functions/pci-epf-vnet.h |  62 ++
  drivers/pci/endpoint/pci-epf-virtio.c         | 113 ++++
  drivers/virtio/virtio_ring.c                  |   2 +-
  include/linux/pci-epf-virtio.h                |  25 +
  include/linux/virtio.h                        |   2 +-
  include/uapi/linux/virtio_pci.h               |   2 +
  13 files changed, 1590 insertions(+), 2 deletions(-)

create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet-ep.c create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet-rc.c create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet.c create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet.h create mode 100644 drivers/pci/endpoint/pci-epf-virtio.c create mode 100644 include/linux/pci-epf-virtio.h

Quoted 2 lines
Quoted 2 lines

--
2.25.1

Best,
Shunsuke

Best,

Shunsuke.

#21 lore raw
Quoted 3 lines · Frank Li

Did you have chance to improve this?

Best regards
Frank Li

#22 lore raw
Quoted 2 lines · Frank Li
Quoted 1 line · Frank Li

Did you have chance to improve this?

Yes. I'm working on it.I'd like to submit new one in this week.

Quoted 2 lines

Best regards
Frank Li

Best,

Shunsuke,

#24 lore raw
In reply to Shunsuke Mie
Quoted 4 lines

foundation.org
Subject: [EXT] [RFC PATCH 0/4] PCI: endpoint: Introduce a virtio-net EP function

The dependent EDMA patch can't be applied at last linux-next. Can you provide a git link? So I can try directly.

Frank

Quoted 6 lines

About this patchset has 4 patches. The first of two patch is little changes to virtio. The third patch add APIs to easily access virtio data structure on PCIe Host side memory. The last one introduce a virtio-net EP device function. Details are in commit respectively.

#25 lore raw

2023年2月4日(土) 6:48 Frank Li <[email protected]>:

Quoted 8 lines
Quoted 5 lines

foundation.org
Subject: [EXT] [RFC PATCH 0/4] PCI: endpoint: Introduce a virtio-net EP function

The dependent EDMA patch can't be applied at last linux-next. Can you provide a git link? So I can try directly.

Sorry, I've missed it. The embedded DMA's patchset is
https://lore.kernel.org/linux-pci/[email protected]/
and, merged to a pci/dwc branch on kernel/git/lpieralisi/pci.git . The link is here:
https://git.kernel.org/pub/scm/linux/kernel/git/lpieralisi/pci.git/commit/?h=pci/dwc

I'll add the information to a cover letter from the next submission.

Quoted 9 lines

Frank

Quoted 6 lines

About this patchset has 4 patches. The first of two patch is little changes to virtio. The third patch add APIs to easily access virtio data structure on PCIe Host side memory. The last one introduce a virtio-net EP device function. Details are in commit respectively.

Best,
Shunsuke
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

#26 lore raw

2023年2月7日(火) 10:43 Shunsuke Mie <[email protected]>:

Quoted 8 lines

2023年2月4日(土) 6:48 Frank Li <[email protected]>:

Quoted 6 lines
Quoted 5 lines

foundation.org
Subject: [EXT] [RFC PATCH 0/4] PCI: endpoint: Introduce a virtio-net EP function

Quoted 7 lines
Quoted 2 lines

The dependent EDMA patch can't be applied at last linux-next. Can you provide a git link? So I can try directly.

Sorry, I've missed it. The embedded DMA's patchset is
https://lore.kernel.org/linux-pci/[email protected]/
and, merged to a pci/dwc branch on kernel/git/lpieralisi/pci.git . The link is here:
https://git.kernel.org/pub/scm/linux/kernel/git/lpieralisi/pci.git/commit/?h=pci/dwc

In addition, the patches are merged into next-20230131 .
https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git/tag/?h=next-20230131

Quoted 12 lines

I'll add the information to a cover letter from the next submission.

Quoted 9 lines

Frank

Quoted 6 lines

About this patchset has 4 patches. The first of two patch is little changes to virtio. The third patch add APIs to easily access virtio data structure on PCIe Host side memory. The last one introduce a virtio-net EP device function. Details are in commit respectively.

Best,
Shunsuke

#27 lore raw
In reply to Shunsuke Mie
Quoted 17 lines · Shunsuke Mie

This patchset introduce a virtio-net EP device function. It provides a new option to communiate between PCIe host and endpoint over IP. Advantage of this option is that the driver fully uses a PCIe embedded DMA. It is used to transport data between virtio ring directly each other. It can be expected to better throughput.

To realize the function, this patchset has few changes and introduces a new APIs to PCI EP framework related to virtio. Furthermore, it device depends on the some patchtes that is discussing. Those depended patchset are following:
- [PATCH 1/2] dmaengine: dw-edma: Fix to change for continuous transfer

- [RFC PATCH 0/3] Deal with alignment restriction on EP side

- [RFC PATCH v2 0/7] Introduce a vringh accessor for IO memory

Quoted 8 lines

About this patchset has 4 patches. The first of two patch is little changes to virtio. The third patch add APIs to easily access virtio data structure on PCIe Host side memory. The last one introduce a virtio-net EP device function. Details are in commit respectively.

Currently those network devices are testd using ping only. I'll add a result of performance evaluation using iperf and etc to the future version of this patchset.

All this feels like it'd need a virtio spec extension but I'm not 100% sure without spending much more time understanding this. what do you say?

Quoted 19 lines

Shunsuke Mie (4):
virtio_pci: add a definition of queue flag in ISR
virtio_ring: remove const from vring getter
PCI: endpoint: Introduce virtio library for EP functions
PCI: endpoint: function: Add EP function driver to provide virtio net
device

 drivers/pci/endpoint/Kconfig                  |   7 +
 drivers/pci/endpoint/Makefile                 |   1 +
 drivers/pci/endpoint/functions/Kconfig        |  12 +
 drivers/pci/endpoint/functions/Makefile       |   1 +
 .../pci/endpoint/functions/pci-epf-vnet-ep.c  | 343 ++++++++++
 .../pci/endpoint/functions/pci-epf-vnet-rc.c  | 635 ++++++++++++++++++
 drivers/pci/endpoint/functions/pci-epf-vnet.c | 387 +++++++++++
 drivers/pci/endpoint/functions/pci-epf-vnet.h |  62 ++
 drivers/pci/endpoint/pci-epf-virtio.c         | 113 ++++
 drivers/virtio/virtio_ring.c                  |   2 +-
 include/linux/pci-epf-virtio.h                |  25 +
 include/linux/virtio.h                        |   2 +-
Quoted 11 lines
 include/uapi/linux/virtio_pci.h               |   2 +
 13 files changed, 1590 insertions(+), 2 deletions(-)
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet-ep.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet-rc.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet.h
 create mode 100644 drivers/pci/endpoint/pci-epf-virtio.c
 create mode 100644 include/linux/pci-epf-virtio.h
Signature · 1 line
2.25.1

_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

#28 lore raw

2023年2月5日(日) 19:02 Michael S. Tsirkin <[email protected]>:

Quoted 24 lines · Shunsuke Mie
Quoted 22 lines

This patchset introduce a virtio-net EP device function. It provides a new option to communiate between PCIe host and endpoint over IP. Advantage of this option is that the driver fully uses a PCIe embedded DMA. It is used to transport data between virtio ring directly each other. It can be expected to better throughput.

To realize the function, this patchset has few changes and introduces a new APIs to PCI EP framework related to virtio. Furthermore, it device depends on the some patchtes that is discussing. Those depended patchset are following:
- [PATCH 1/2] dmaengine: dw-edma: Fix to change for continuous transfer

- [RFC PATCH 0/3] Deal with alignment restriction on EP side

- [RFC PATCH v2 0/7] Introduce a vringh accessor for IO memory

About this patchset has 4 patches. The first of two patch is little changes to virtio. The third patch add APIs to easily access virtio data structure on PCIe Host side memory. The last one introduce a virtio-net EP device function. Details are in commit respectively.

Quoted 8 lines
Quoted 3 lines

Currently those network devices are testd using ping only. I'll add a result of performance evaluation using iperf and etc to the future version of this patchset.

All this feels like it'd need a virtio spec extension but I'm not 100% sure without spending much more time understanding this. what do you say?

This patch shows the virtio-net device as pcie device. Could you tell me what part
of the spec are you concerned about?

Quoted 28 lines
Quoted 28 lines

Shunsuke Mie (4):
virtio_pci: add a definition of queue flag in ISR
virtio_ring: remove const from vring getter
PCI: endpoint: Introduce virtio library for EP functions
PCI: endpoint: function: Add EP function driver to provide virtio net
device

 drivers/pci/endpoint/Kconfig                  |   7 +
 drivers/pci/endpoint/Makefile                 |   1 +
 drivers/pci/endpoint/functions/Kconfig        |  12 +
 drivers/pci/endpoint/functions/Makefile       |   1 +
 .../pci/endpoint/functions/pci-epf-vnet-ep.c  | 343 ++++++++++
 .../pci/endpoint/functions/pci-epf-vnet-rc.c  | 635 ++++++++++++++++++
 drivers/pci/endpoint/functions/pci-epf-vnet.c | 387 +++++++++++
 drivers/pci/endpoint/functions/pci-epf-vnet.h |  62 ++
 drivers/pci/endpoint/pci-epf-virtio.c         | 113 ++++
 drivers/virtio/virtio_ring.c                  |   2 +-
 include/linux/pci-epf-virtio.h                |  25 +
 include/linux/virtio.h                        |   2 +-
 include/uapi/linux/virtio_pci.h               |   2 +
 13 files changed, 1590 insertions(+), 2 deletions(-)
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet-ep.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet-rc.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet.c
 create mode 100644 drivers/pci/endpoint/functions/pci-epf-vnet.h
 create mode 100644 drivers/pci/endpoint/pci-epf-virtio.c
 create mode 100644 include/linux/pci-epf-virtio.h
Quoted 3 lines
Quoted 2 lines

--
2.25.1

Best,
Shunsuke
_______________________________________________
Virtualization mailing list
[email protected]
https://lists.linuxfoundation.org/mailman/listinfo/virtualization