| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-Id: <20210130022124.65083-26-mgross@linux.intel.com>
Date: Fri, 29 Jan 2021 18:20:40 -0800
From: mgross@...ux.intel.com
To: markgross@...nel.org, mgross@...ux.intel.com, arnd@...db.de,
bp@...e.de, damien.lemoal@....com, dragan.cvetic@...inx.com,
gregkh@...uxfoundation.org, corbet@....net,
palmerdabbelt@...gle.com, paul.walmsley@...ive.com,
peng.fan@....com, robh+dt@...nel.org, shawnguo@...nel.org,
jassisinghbrar@...il.com
Cc: linux-kernel@...r.kernel.org,
"Li, Tingqian" <tingqian.li@...el.com>, Li@...ux.intel.com,
Zhou@...ux.intel.com, Luwei <luwei.zhou@...el.com>,
Wang@...ux.intel.com, jue <wang.jue@...el.com>
Subject: [PATCH v3 25/34] misc: Add Keem Bay VPU manager
From: "Li, Tingqian" <tingqian.li@...el.com>
VPU IP on Keem Bay SOC is a vision acceleration IP complex
under the control of a RTOS-based firmware (running on RISC
MCU inside the VPU IP) serving user-space application
running on CPU side for HW accelerated computer vision tasks.
This module is kernel counterpart of the VPUAL(VPU abstraction
layer) which bridges firmware on VPU side and applications on
CPU user-space, it assists firmware on VPU side serving multiple
user space application processes on CPU side concurrently while
also performing necessary data buffer management on behave of
VPU IP.
objmgr provides basic infrastructure for create/destroy VPU side
software object concurrently on demand of user-space application
and also automatically release leaked objects during handling of
application termination. Note this module only cares about the
life-cycle of such objects, it's up to the application and firmware
to define the behavior/operations of each object.
objmgr does it's job by communicating with firmware through a fixed
reserved xlink channel, using a very simple message protocol.
smm provides DMABuf allocation/import facilities to allow user-space
app pass data to/from VPU in zero-copy fashion. it also provided a
convenient ioctl function for converting virtual pointer of a mem-mapped
and imported DMABuf into it's corresponding dma address, to allow
user-space app to specify the sub-regions of a bigger DMABuf to be
processed by VPU.
Signed-off-by: Li, Tingqian <tingqian.li@...el.com>
Signed-off-by: Zhou, Luwei <luwei.zhou@...el.com>
Signed-off-by: Wang, jue <wang.jue@...el.com>
---
drivers/misc/Kconfig | 1 +
drivers/misc/Makefile | 1 +
drivers/misc/vpumgr/Kconfig | 9 +
drivers/misc/vpumgr/Makefile | 3 +
drivers/misc/vpumgr/vpu_common.h | 31 ++
drivers/misc/vpumgr/vpu_mgr.c | 370 ++++++++++++++++++++
drivers/misc/vpumgr/vpu_smm.c | 554 +++++++++++++++++++++++++++++
drivers/misc/vpumgr/vpu_smm.h | 30 ++
drivers/misc/vpumgr/vpu_vcm.c | 584 +++++++++++++++++++++++++++++++
drivers/misc/vpumgr/vpu_vcm.h | 84 +++++
include/uapi/misc/vpumgr.h | 64 ++++
11 files changed, 1731 insertions(+)
create mode 100644 drivers/misc/vpumgr/Kconfig
create mode 100644 drivers/misc/vpumgr/Makefile
create mode 100644 drivers/misc/vpumgr/vpu_common.h
create mode 100644 drivers/misc/vpumgr/vpu_mgr.c
create mode 100644 drivers/misc/vpumgr/vpu_smm.c
create mode 100644 drivers/misc/vpumgr/vpu_smm.h
create mode 100644 drivers/misc/vpumgr/vpu_vcm.c
create mode 100644 drivers/misc/vpumgr/vpu_vcm.h
create mode 100644 include/uapi/misc/vpumgr.h
diff --git a/drivers/misc/Kconfig b/drivers/misc/Kconfig
index 09ae65e98681..2d1f7b165cc8 100644
--- a/drivers/misc/Kconfig
+++ b/drivers/misc/Kconfig
@@ -484,4 +484,5 @@ source "drivers/misc/uacce/Kconfig"
source "drivers/misc/xlink-pcie/Kconfig"
source "drivers/misc/xlink-ipc/Kconfig"
source "drivers/misc/xlink-core/Kconfig"
+source "drivers/misc/vpumgr/Kconfig"
endmenu
diff --git a/drivers/misc/Makefile b/drivers/misc/Makefile
index f3a6eb03bae9..2936930f3edc 100644
--- a/drivers/misc/Makefile
+++ b/drivers/misc/Makefile
@@ -60,3 +60,4 @@ obj-$(CONFIG_HISI_HIKEY_USB) += hisi_hikey_usb.o
obj-y += xlink-pcie/
obj-$(CONFIG_XLINK_IPC) += xlink-ipc/
obj-$(CONFIG_XLINK_CORE) += xlink-core/
+obj-$(CONFIG_VPUMGR) += vpumgr/
diff --git a/drivers/misc/vpumgr/Kconfig b/drivers/misc/vpumgr/Kconfig
new file mode 100644
index 000000000000..bb82ff83afd3
--- /dev/null
+++ b/drivers/misc/vpumgr/Kconfig
@@ -0,0 +1,9 @@
+config VPUMGR
+ tristate "VPU Manager"
+ depends on ARM64 && XLINK_CORE
+ help
+ VPUMGR manages life-cycle of VPU related resources which were
+ dynamically allocated on demands of user-space application
+
+ Select y or m if you have a processor including the Intel
+ Vision Processor (VPU) on it.
diff --git a/drivers/misc/vpumgr/Makefile b/drivers/misc/vpumgr/Makefile
new file mode 100644
index 000000000000..51441dc8a930
--- /dev/null
+++ b/drivers/misc/vpumgr/Makefile
@@ -0,0 +1,3 @@
+# SPDX-License-Identifier: GPL-2.0-only
+obj-$(CONFIG_VPUMGR) += vpumgr.o
+vpumgr-objs := vpu_mgr.o vpu_smm.o vpu_vcm.o
diff --git a/drivers/misc/vpumgr/vpu_common.h b/drivers/misc/vpumgr/vpu_common.h
new file mode 100644
index 000000000000..cd474ffc05f3
--- /dev/null
+++ b/drivers/misc/vpumgr/vpu_common.h
@@ -0,0 +1,31 @@
+/* SPDX-License-Identifier: GPL-2.0-only
+ * VPUMGR Kernel module - common definition
+ * Copyright (C) 2020-2021 Intel Corporation
+ */
+#ifndef _VPU_COMMON_H
+#define _VPU_COMMON_H
+#include <linux/cdev.h>
+#include <linux/platform_device.h>
+
+#include <uapi/misc/vpumgr.h>
+
+#include "vpu_vcm.h"
+
+/* there will be one such device for each HW instance */
+struct vpumgr_device {
+ struct device *sdev;
+ struct device *dev;
+ dev_t devnum;
+ struct cdev cdev;
+ struct platform_device *pdev;
+
+ struct vcm_dev vcm;
+ struct dentry *debugfs_root;
+
+ struct mutex client_mutex; /* protect client_list */
+ struct list_head client_list;
+};
+
+#define XLINK_INVALID_SW_DEVID 0xDEADBEEF
+
+#endif
diff --git a/drivers/misc/vpumgr/vpu_mgr.c b/drivers/misc/vpumgr/vpu_mgr.c
new file mode 100644
index 000000000000..75be64ebc3b0
--- /dev/null
+++ b/drivers/misc/vpumgr/vpu_mgr.c
@@ -0,0 +1,370 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * VPU Manager Kernel module.
+ *
+ * Copyright (C) 2020-2021 Intel Corporation
+ *
+ */
+#include <linux/debugfs.h>
+#include <linux/kernel.h>
+#include <linux/module.h>
+#include <linux/of_device.h>
+#include <linux/slab.h>
+#include <linux/uaccess.h>
+
+#include "vpu_common.h"
+#include "vpu_smm.h"
+#include "vpu_vcm.h"
+
+#define DRIVER_NAME "vpumgr"
+#define MAX_DEV_CNT 32
+
+/* Define the max xlink device number */
+#define MAX_SW_DEV_CNT 20
+
+/* Define the SW_DEVICE_ID bit mask and offset */
+#define IPC_INTERFACE 0x00
+#define PCIE_INTERFACE 0x01
+#define MASK_INTERFACE 0x7000000
+#define BIT_OFFSET_INTERFACE 24
+#define MASK_VPU_IPC_ID 0x000E
+#define BIT_OFFSET_VPU_ID 1
+
+#define SWDEVID_INTERFACE(sw_dev_id) (((sw_dev_id) & MASK_INTERFACE) >> BIT_OFFSET_INTERFACE)
+#define SWDEVID_VPU_IPC_ID(sw_dev_id) (((sw_dev_id) & MASK_VPU_IPC_ID) >> BIT_OFFSET_VPU_ID)
+
+static dev_t vpumgr_devnum;
+static struct class *vpumgr_class;
+
+/**
+ * struct vpumgr_fpriv - per-process context stored in FD private data.
+ * @vdev: vpumgr device corresponding to the file
+ * @smm: memory manager
+ * @ctx: vpu context manager
+ * @list: for global list of all opened file
+ * @pid: process which opens the device file
+ */
+struct vpumgr_fpriv {
+ struct vpumgr_device *vdev;
+ struct vpumgr_smm smm;
+ struct vpumgr_ctx ctx;
+ struct list_head list;
+ pid_t pid;
+};
+
+static u32 get_sw_device_id(int vpu_ipc_id)
+{
+ u32 sw_id_list[MAX_SW_DEV_CNT];
+ enum xlink_error rc;
+ u32 num = 0;
+ u32 swid;
+ int i;
+
+ rc = xlink_get_device_list(sw_id_list, &num);
+ if (rc) {
+ pr_err("XLINK get device list error %d in %s\n", rc, __func__);
+ return XLINK_INVALID_SW_DEVID;
+ }
+
+ for (i = 0; i < num; i++) {
+ swid = sw_id_list[i];
+ if (SWDEVID_INTERFACE(swid) == IPC_INTERFACE &&
+ SWDEVID_VPU_IPC_ID(swid) == vpu_ipc_id)
+ return swid;
+ }
+ return XLINK_INVALID_SW_DEVID;
+}
+
+static int vpumgr_open(struct inode *inode, struct file *filp)
+{
+ struct vpumgr_fpriv *vpriv;
+ struct vpumgr_device *vdev;
+ int rc;
+
+ vpriv = kzalloc(sizeof(*vpriv), GFP_KERNEL);
+ if (!vpriv)
+ return -ENOMEM;
+
+ vdev = container_of(inode->i_cdev, struct vpumgr_device, cdev);
+ rc = smm_open(&vpriv->smm, vdev);
+ if (rc)
+ goto free_priv;
+
+ rc = vcm_open(&vpriv->ctx, vdev);
+ if (rc)
+ goto close_smm;
+
+ vpriv->vdev = vdev;
+ vpriv->pid = task_pid_nr(current);
+ INIT_LIST_HEAD(&vpriv->list);
+
+ mutex_lock(&vdev->client_mutex);
+ list_add_tail(&vpriv->list, &vdev->client_list);
+ mutex_unlock(&vdev->client_mutex);
+
+ filp->private_data = vpriv;
+ return 0;
+
+close_smm:
+ smm_close(&vpriv->smm);
+free_priv:
+ kfree(vpriv);
+ return rc;
+}
+
+static int vpumgr_release(struct inode *inode, struct file *filp)
+{
+ struct vpumgr_fpriv *vpriv = filp->private_data;
+ struct vpumgr_device *vdev = container_of(inode->i_cdev, struct vpumgr_device, cdev);
+
+ vcm_close(&vpriv->ctx);
+ smm_close(&vpriv->smm);
+
+ mutex_lock(&vdev->client_mutex);
+ list_del(&vpriv->list);
+ mutex_unlock(&vdev->client_mutex);
+
+ kfree(vpriv);
+ filp->private_data = NULL;
+ return 0;
+}
+
+static long vpumgr_ioctl(struct file *filp, unsigned int cmd, unsigned long arg)
+{
+ struct vpumgr_fpriv *vpriv = filp->private_data;
+ const unsigned int io_dir = _IOC_DIR(cmd);
+ const unsigned int io_size = _IOC_SIZE(cmd);
+ struct vpumgr_vcm_submit *vs;
+ struct vpumgr_vcm_wait *vw;
+ char tmp[128];
+ int rc = 0;
+
+ if (_IOC_TYPE(cmd) != VPUMGR_MAGIC || _IOC_NR(cmd) >= _IOC_NR(VPUMGR_IOCTL_END))
+ return -EINVAL;
+
+ if (io_size > sizeof(tmp))
+ return -EFAULT;
+
+ if (io_dir & _IOC_READ) {
+ if (copy_from_user(tmp, (void __user *)arg, io_size) != 0)
+ return -EFAULT;
+ }
+
+ switch (cmd) {
+ case VPUMGR_IOCTL_DMABUF_ALLOC:
+ rc = smm_alloc(&vpriv->smm, (void *)tmp);
+ break;
+ case VPUMGR_IOCTL_DMABUF_IMPORT:
+ rc = smm_import(&vpriv->smm, (void *)tmp);
+ break;
+ case VPUMGR_IOCTL_DMABUF_UNIMPORT:
+ rc = smm_unimport(&vpriv->smm, (void *)tmp);
+ break;
+ case VPUMGR_IOCTL_DMABUF_PTR2VPU:
+ rc = smm_ptr2vpu(&vpriv->smm, (void *)tmp);
+ break;
+ case VPUMGR_IOCTL_VCM_SUBMIT:
+ vs = (struct vpumgr_vcm_submit *)tmp;
+ if (vs->cmd <= VCTX_KMD_RESERVED_CMD_LAST) {
+ /*
+ * user-space can talk to vpu context lives in firmware
+ * with any commands other than those reserved for kernel
+ * mode.
+ */
+ rc = -EACCES;
+ break;
+ }
+ rc = vcm_submit(&vpriv->ctx, vs->cmd,
+ (const void *)vs->in, vs->in_len, &vs->submit_id);
+ break;
+ case VPUMGR_IOCTL_VCM_WAIT:
+ vw = (struct vpumgr_vcm_wait *)tmp;
+ rc = vcm_wait(&vpriv->ctx, vw->submit_id, &vw->vpu_rc,
+ (void *)vw->out, &vw->out_len, vw->timeout_ms);
+ break;
+ }
+
+ if (!rc) {
+ if (io_dir & _IOC_WRITE) {
+ if (copy_to_user((void __user *)arg, tmp, io_size) != 0)
+ return -EFAULT;
+ }
+ }
+ return rc;
+}
+
+static const struct file_operations vpumgr_devfile_fops = {
+ .owner = THIS_MODULE,
+ .open = vpumgr_open,
+ .release = vpumgr_release,
+ .unlocked_ioctl = vpumgr_ioctl,
+};
+
+static int vpumgr_debugfs_stats_show(struct seq_file *file, void *offset)
+{
+ struct vpumgr_device *vdev = dev_get_drvdata(file->private);
+ struct vpumgr_fpriv *fpriv;
+ int i = 0;
+
+ mutex_lock(&vdev->client_mutex);
+ list_for_each_entry(fpriv, &vdev->client_list, list) {
+ seq_printf(file, "client #%d pid:%d\n", i++, fpriv->pid);
+ vcm_debugfs_stats_show(file, &fpriv->ctx);
+ smm_debugfs_stats_show(file, &fpriv->smm);
+ }
+ mutex_unlock(&vdev->client_mutex);
+ return 0;
+}
+
+static const struct of_device_id keembay_vpumgr_of_match[] = {
+ { .compatible = "intel,keembay-vpu-mgr"},
+ { .compatible = "intel,keembay-vpusmm"},
+ {}
+};
+MODULE_DEVICE_TABLE(of, keembay_vpumgr_of_match);
+
+static int vpumgr_driver_probe(struct platform_device *pdev)
+{
+ struct device *dev = &pdev->dev;
+ struct vpumgr_device *vdev;
+ u32 ipc_sw_device_id;
+ u32 vpu_ipc_id = 0;
+ int rc;
+
+ /* get device id */
+ rc = of_property_read_u32(dev->of_node, "intel,keembay-vpu-ipc-id",
+ &vpu_ipc_id);
+ if (rc && rc != -EINVAL) {
+ dev_err(dev, "%s: vpu-ipc-id read failed with rc %d\n", __func__, rc);
+ return -EINVAL;
+ }
+
+ ipc_sw_device_id = get_sw_device_id(vpu_ipc_id);
+ if (ipc_sw_device_id == XLINK_INVALID_SW_DEVID)
+ dev_warn(dev, "%s: no xlink sw device for vpu_ipc_id %d\n",
+ __func__, vpu_ipc_id);
+
+ vdev = devm_kzalloc(dev, sizeof(struct vpumgr_device), GFP_KERNEL);
+ if (!vdev)
+ return -ENOMEM;
+
+ vdev->devnum = MKDEV(MAJOR(vpumgr_devnum), vpu_ipc_id);
+ vdev->pdev = pdev;
+ vdev->dev = dev;
+
+ dev_dbg(dev, "dev->devnum %u, id %u, major %u\n",
+ vdev->devnum, vpu_ipc_id, MAJOR(vdev->devnum));
+
+ vdev->sdev = device_create(vpumgr_class, dev, vdev->devnum,
+ NULL, DRIVER_NAME "%d", vpu_ipc_id);
+ if (IS_ERR(vdev->sdev)) {
+ dev_err(dev, "%s: device_create failed\n", __func__);
+ return PTR_ERR(vdev->sdev);
+ }
+
+ cdev_init(&vdev->cdev, &vpumgr_devfile_fops);
+ vdev->cdev.owner = THIS_MODULE;
+ rc = cdev_add(&vdev->cdev, vdev->devnum, 1);
+ if (rc) {
+ dev_err(dev, "%s: cdev_add failed.\n", __func__);
+ goto detroy_device;
+ }
+
+ vdev->debugfs_root = debugfs_create_dir(dev_name(vdev->sdev), NULL);
+
+ debugfs_create_devm_seqfile(dev, "stats", vdev->debugfs_root,
+ vpumgr_debugfs_stats_show);
+
+ rc = smm_init(vdev);
+ if (rc)
+ goto remove_debugfs;
+
+ rc = vcm_init(vdev, ipc_sw_device_id);
+ if (rc)
+ goto fini_smm;
+
+ INIT_LIST_HEAD(&vdev->client_list);
+ mutex_init(&vdev->client_mutex);
+
+ dev_set_drvdata(dev, vdev);
+ return 0;
+
+fini_smm:
+ smm_fini(vdev);
+remove_debugfs:
+ debugfs_remove_recursive(vdev->debugfs_root);
+ cdev_del(&vdev->cdev);
+detroy_device:
+ device_destroy(vpumgr_class, vdev->devnum);
+ return rc;
+}
+
+static int vpumgr_driver_remove(struct platform_device *pdev)
+{
+ struct vpumgr_device *vdev = dev_get_drvdata(&pdev->dev);
+
+ mutex_destroy(&vdev->client_mutex);
+ vcm_fini(vdev);
+ smm_fini(vdev);
+ debugfs_remove_recursive(vdev->debugfs_root);
+ cdev_del(&vdev->cdev);
+ device_destroy(vpumgr_class, vdev->devnum);
+ return 0;
+}
+
+static struct platform_driver vpumgr_driver = {
+ .probe = vpumgr_driver_probe,
+ .remove = vpumgr_driver_remove,
+ .driver = {
+ .owner = THIS_MODULE,
+ .name = "keembay-vpu-mgr",
+ .of_match_table = keembay_vpumgr_of_match,
+ },
+};
+
+static int __init vpumgr_init(void)
+{
+ int rc;
+
+ rc = alloc_chrdev_region(&vpumgr_devnum, 0, MAX_DEV_CNT, DRIVER_NAME);
+ if (rc < 0) {
+ pr_err("[%s] err: alloc_chrdev_region\n", __func__);
+ return rc;
+ }
+
+ vpumgr_class = class_create(THIS_MODULE, DRIVER_NAME "_class");
+ if (IS_ERR(vpumgr_class)) {
+ rc = PTR_ERR(vpumgr_class);
+ pr_err("[%s] err: class_create\n", __func__);
+ goto unreg_chrdev;
+ }
+
+ rc = platform_driver_register(&vpumgr_driver);
+ if (rc) {
+ pr_err("[%s] err platform_driver_register\n", __func__);
+ goto destroy_class;
+ }
+
+ return 0;
+
+destroy_class:
+ class_destroy(vpumgr_class);
+unreg_chrdev:
+ unregister_chrdev_region(vpumgr_devnum, MAX_DEV_CNT);
+ return rc;
+}
+
+static void vpumgr_exit(void)
+{
+ platform_driver_unregister(&vpumgr_driver);
+ class_destroy(vpumgr_class);
+ unregister_chrdev_region(vpumgr_devnum, MAX_DEV_CNT);
+}
+
+module_init(vpumgr_init)
+module_exit(vpumgr_exit)
+
+MODULE_DESCRIPTION("VPU resource manager driver");
+MODULE_AUTHOR("Tingqian Li <tingqian.li@...el.com>");
+MODULE_AUTHOR("Luwei Zhou <luwie.zhou@...el.com>");
+MODULE_LICENSE("GPL");
diff --git a/drivers/misc/vpumgr/vpu_smm.c b/drivers/misc/vpumgr/vpu_smm.c
new file mode 100644
index 000000000000..a89f62984a48
--- /dev/null
+++ b/drivers/misc/vpumgr/vpu_smm.c
@@ -0,0 +1,554 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * Copyright (C) 2020 Intel Corporation
+ */
+#include <linux/cdev.h>
+#include <linux/debugfs.h>
+#include <linux/dma-mapping.h>
+#include <linux/dma-buf.h>
+#include <linux/dma-direct.h>
+#include <linux/err.h>
+#include <linux/kernel.h>
+#include <linux/module.h>
+#include <linux/mm.h>
+#include <linux/of_device.h>
+#include <linux/of_reserved_mem.h>
+#include <linux/of_address.h>
+#include <linux/platform_device.h>
+#include <linux/rbtree.h>
+#include <linux/slab.h>
+#include <linux/seq_file.h>
+#include <linux/uaccess.h>
+
+#include <linux/sched.h>
+#include <linux/sched/mm.h>
+#include <linux/sched/task.h>
+
+#include <uapi/misc/vpumgr.h>
+
+#include "vpu_common.h"
+#include "vpu_smm.h"
+
+/*
+ * DMABuf exported by VPU device is described by following data structure
+ * the life of the buffer may be longer than session(it may be shared with
+ * other driver),so it contains pointer to device rather than to session
+ */
+struct vpusmm_buffer {
+ struct device *dev;
+ void *cookie;
+ dma_addr_t dma_addr;
+ size_t size;
+ unsigned long dma_attrs;
+};
+
+/*
+ * DMABufs imported to VPU device are maintained in a rb tree with dmabuf's
+ * pointer as key. so user space can unimport it by specifying fd and ptr2vpu
+ * API can work by searching this rb tree.
+ */
+struct vpusmm_impbuf {
+ struct rb_node node;
+ struct dma_buf *dmabuf;
+ struct dma_buf_attachment *attach;
+ struct sg_table *sgt;
+ enum dma_data_direction direction;
+ dma_addr_t vpu_addr;
+ int refcount;
+};
+
+/*
+ * VPU imported dmabuf management
+ */
+static void vpusmm_insert_impbuf(struct vpumgr_smm *sess,
+ struct vpusmm_impbuf *new_item)
+{
+ struct rb_root *root = &sess->imp_rb;
+ struct rb_node **iter = &root->rb_node, *parent = NULL;
+ struct dma_buf *value = new_item->dmabuf;
+ struct vpusmm_impbuf *item;
+
+ while (*iter) {
+ parent = *iter;
+ item = rb_entry(parent, struct vpusmm_impbuf, node);
+
+ if (item->dmabuf > value)
+ iter = &(*iter)->rb_left;
+ else
+ iter = &(*iter)->rb_right;
+ }
+
+ /* Put the new node there */
+ rb_link_node(&new_item->node, parent, iter);
+ rb_insert_color(&new_item->node, root);
+}
+
+static struct vpusmm_impbuf *vpusmm_find_impbuf(struct vpumgr_smm *sess,
+ struct dma_buf *dmabuf)
+{
+ struct rb_node *node = sess->imp_rb.rb_node;
+ struct vpusmm_impbuf *item;
+
+ while (node) {
+ item = rb_entry(node, struct vpusmm_impbuf, node);
+
+ if (item->dmabuf > dmabuf)
+ node = node->rb_left;
+ else if (item->dmabuf < dmabuf)
+ node = node->rb_right;
+ else
+ return item;
+ }
+ return NULL;
+}
+
+static struct vpusmm_impbuf *vpusmm_import_dmabuf(struct vpumgr_smm *sess,
+ int dmabuf_fd,
+ enum dma_data_direction direction, int vpu_id)
+{
+ struct vpusmm_impbuf *item;
+ struct dma_buf_attachment *attach;
+ struct device *dma_dev = sess->vdev->dev;
+ struct dma_buf *dmabuf;
+ struct sg_table *sgt;
+
+ dmabuf = dma_buf_get(dmabuf_fd);
+ if (IS_ERR(dmabuf))
+ return ERR_CAST(dmabuf);
+
+ mutex_lock(&sess->imp_rb_lock);
+ item = vpusmm_find_impbuf(sess, dmabuf);
+ if (item) {
+ item->refcount++;
+ goto found_impbuf;
+ }
+
+ attach = dma_buf_attach(dmabuf, dma_dev);
+ if (IS_ERR(attach)) {
+ item = ERR_CAST(attach);
+ goto fail_attach;
+ }
+
+ sgt = dma_buf_map_attachment(attach, direction);
+ if (IS_ERR(sgt)) {
+ item = ERR_CAST(sgt);
+ goto fail_map;
+ }
+
+ if (sgt->nents > 1) {
+ item = ERR_PTR(-EINVAL);
+ goto fail_import;
+ }
+
+ item = kzalloc(sizeof(*item), GFP_KERNEL);
+ if (!item) {
+ item = ERR_PTR(-ENOMEM);
+ goto fail_import;
+ }
+
+ item->dmabuf = dmabuf;
+ item->attach = attach;
+ item->sgt = sgt;
+ item->direction = direction;
+ item->vpu_addr = sg_dma_address(sgt->sgl);
+ item->refcount = 1;
+
+ vpusmm_insert_impbuf(sess, item);
+
+ mutex_unlock(&sess->imp_rb_lock);
+ return item;
+
+fail_import:
+ dma_buf_unmap_attachment(attach, sgt, direction);
+fail_map:
+ dma_buf_detach(dmabuf, attach);
+fail_attach:
+found_impbuf:
+ mutex_unlock(&sess->imp_rb_lock);
+ dma_buf_put(dmabuf);
+ return item;
+}
+
+int smm_open(struct vpumgr_smm *sess, struct vpumgr_device *vdev)
+{
+ sess->vdev = vdev;
+ sess->imp_rb = RB_ROOT;
+ mutex_init(&sess->imp_rb_lock);
+ return 0;
+}
+
+int smm_close(struct vpumgr_smm *sess)
+{
+ struct device *dev = sess->vdev->sdev;
+ struct rb_node *cur, *next;
+ struct vpusmm_impbuf *item;
+
+ mutex_destroy(&sess->imp_rb_lock);
+
+ cur = rb_first(&sess->imp_rb);
+ while (cur) {
+ item = rb_entry(cur, struct vpusmm_impbuf, node);
+ next = rb_next(cur);
+ if (item) {
+ dev_dbg(dev, "[%s] PID:%d free leaked imported dmabuf %zu bytes, %d refs\n",
+ __func__, current->pid, item->dmabuf->size, item->refcount);
+ dma_buf_unmap_attachment(item->attach, item->sgt, item->direction);
+ dma_buf_detach(item->dmabuf, item->attach);
+ dma_buf_put(item->dmabuf);
+ rb_erase(&item->node, &sess->imp_rb);
+ kfree(item);
+ }
+ cur = next;
+ }
+ return 0;
+}
+
+/*
+ * DMABuf
+ */
+static struct sg_table *map_dma_buf_vpusmm(struct dma_buf_attachment *attach,
+ enum dma_data_direction dir)
+{
+ struct vpusmm_buffer *buff = attach->dmabuf->priv;
+ struct sg_table *sgt;
+ int rc;
+
+ if (WARN_ON(dir == DMA_NONE))
+ return ERR_PTR(-EINVAL);
+
+ sgt = kzalloc(sizeof(*sgt), GFP_KERNEL);
+ if (!sgt)
+ return NULL;
+
+ rc = dma_get_sgtable(buff->dev, sgt, buff->cookie, buff->dma_addr, buff->size);
+ if (rc < 0)
+ goto free_sgt;
+
+ rc = dma_map_sg_attrs(attach->dev, sgt->sgl, sgt->nents, dir, DMA_ATTR_SKIP_CPU_SYNC);
+ if (!rc)
+ goto free_sg_table;
+ return sgt;
+
+free_sg_table:
+ sg_free_table(sgt);
+free_sgt:
+ kfree(sgt);
+ return ERR_PTR(rc);
+}
+
+static void unmap_dma_buf_vpusmm(struct dma_buf_attachment *attach,
+ struct sg_table *sgt, enum dma_data_direction dir)
+{
+ dma_unmap_sg_attrs(attach->dev, sgt->sgl, sgt->nents, dir, DMA_ATTR_SKIP_CPU_SYNC);
+ sg_free_table(sgt);
+ kfree(sgt);
+}
+
+static void release_vpusmm(struct dma_buf *dmabuf)
+{
+ struct vpusmm_buffer *buff = dmabuf->priv;
+
+ dma_free_attrs(buff->dev, buff->size, buff->cookie, buff->dma_addr, buff->dma_attrs);
+ kfree(buff);
+}
+
+static int mmap_vpusmm(struct dma_buf *dmabuf, struct vm_area_struct *vma)
+{
+ struct vpusmm_buffer *buff = dmabuf->priv;
+ unsigned long vm_size;
+
+ vm_size = vma->vm_end - vma->vm_start;
+ if (vm_size > PAGE_ALIGN(buff->size))
+ return -EINVAL;
+
+ vma->vm_flags |= VM_IO | VM_DONTEXPAND | VM_DONTDUMP;
+ vma->vm_pgoff = 0;
+
+ return dma_mmap_attrs(buff->dev, vma, buff->cookie, buff->dma_addr,
+ buff->size, buff->dma_attrs);
+}
+
+static int vmap_vpusmm(struct dma_buf *dmabuf, struct dma_buf_map *map)
+{
+ struct vpusmm_buffer *buff = dmabuf->priv;
+
+ dma_buf_map_set_vaddr(map, buff->cookie);
+
+ return 0;
+}
+
+static const struct dma_buf_ops vpusmm_dmabuf_ops = {
+ .cache_sgt_mapping = true,
+ .map_dma_buf = map_dma_buf_vpusmm,
+ .unmap_dma_buf = unmap_dma_buf_vpusmm,
+ .release = release_vpusmm,
+ .mmap = mmap_vpusmm,
+ .vmap = vmap_vpusmm,
+};
+
+/*
+ * Allocate dma buffer suitable for VPU access.
+ * export as DMABuf fd
+ * sess will hold additional refcount to the dmabuf
+ * on request of passing it to VPU side for processing
+ */
+int smm_alloc(struct vpumgr_smm *sess, struct vpumgr_args_alloc *arg)
+{
+ struct vpumgr_device *vdev = sess->vdev;
+ const int flags = O_RDWR | O_CLOEXEC;
+ size_t buffer_size = arg->size;
+ struct dma_buf *dmabuf = NULL;
+ phys_addr_t phys_addr;
+ struct dma_buf_export_info exp_info = {
+ .exp_name = dev_name(vdev->sdev),
+ .owner = THIS_MODULE,
+ .ops = &vpusmm_dmabuf_ops,
+ .size = buffer_size,
+ .flags = flags
+ };
+ struct vpusmm_buffer *buff;
+ int retval;
+
+ buff = kzalloc(sizeof(*buff), GFP_KERNEL);
+ if (!buff) {
+ retval = -ENOMEM;
+ goto failed;
+ }
+
+ buff->dev = vdev->dev;
+ buff->size = buffer_size;
+ buff->dma_attrs = DMA_ATTR_FORCE_CONTIGUOUS | DMA_ATTR_WRITE_COMBINE;
+ buff->cookie = dma_alloc_attrs(vdev->dev, buff->size, &buff->dma_addr,
+ GFP_KERNEL | GFP_DMA, buff->dma_attrs);
+ if (!buff->cookie) {
+ retval = -ENOMEM;
+ goto failed;
+ }
+
+ phys_addr = dma_to_phys(vdev->dev, buff->dma_addr);
+
+ exp_info.priv = buff;
+ dmabuf = dma_buf_export(&exp_info);
+ if (IS_ERR(dmabuf)) {
+ retval = PTR_ERR(dmabuf);
+ dmabuf = NULL;
+ goto failed;
+ }
+
+ retval = dma_buf_fd(dmabuf, flags);
+ if (retval < 0) {
+ goto failed;
+ } else {
+ arg->fd = retval;
+ retval = 0;
+ }
+
+ dev_dbg(vdev->dev, "%s: dma_addr=%llx, phys_addr=%llx allocated from %s\n",
+ __func__, buff->dma_addr, phys_addr, dev_name(vdev->dev));
+
+ return 0;
+failed:
+ dev_err(vdev->dev, "%s failed with %d\n", __func__, retval);
+
+ if (dmabuf) {
+ /* this will finally release underlying buff */
+ dma_buf_put(dmabuf);
+ } else if (buff) {
+ if (buff->cookie)
+ dma_free_attrs(vdev->dev, buff->size, buff->cookie,
+ buff->dma_addr, buff->dma_attrs);
+ kfree(buff);
+ }
+ return retval;
+}
+
+int smm_import(struct vpumgr_smm *sess, struct vpumgr_args_import *arg)
+{
+ struct device *dev = sess->vdev->sdev;
+ enum dma_data_direction direction;
+ struct vpusmm_impbuf *item;
+
+ switch (arg->vpu_access) {
+ case VPU_ACCESS_READ:
+ direction = DMA_TO_DEVICE;
+ break;
+ case VPU_ACCESS_WRITE:
+ direction = DMA_FROM_DEVICE;
+ break;
+ case VPU_ACCESS_DEFAULT:
+ case VPU_ACCESS_RW:
+ direction = DMA_BIDIRECTIONAL;
+ break;
+ default:
+ dev_err(dev, "Unknown vpu_access:%d\n", arg->vpu_access);
+ return -EINVAL;
+ }
+
+ item = vpusmm_import_dmabuf(sess, arg->fd, direction, 0);
+ if (IS_ERR(item))
+ return PTR_ERR(item);
+
+ arg->vpu_addr = item->vpu_addr;
+ arg->size = item->dmabuf->size;
+ return 0;
+}
+
+int smm_unimport(struct vpumgr_smm *sess, int *p_dmabuf_fd)
+{
+ struct vpusmm_impbuf *item;
+ struct dma_buf *dmabuf;
+ int rc = 0;
+
+ dmabuf = dma_buf_get(*p_dmabuf_fd);
+ if (IS_ERR(dmabuf))
+ return PTR_ERR(dmabuf);
+
+ mutex_lock(&sess->imp_rb_lock);
+ item = vpusmm_find_impbuf(sess, dmabuf);
+ if (!item) {
+ rc = -EINVAL;
+ goto exit;
+ }
+
+ item->refcount--;
+ if (item->refcount <= 0) {
+ rb_erase(&item->node, &sess->imp_rb);
+ dma_buf_unmap_attachment(item->attach, item->sgt, item->direction);
+ dma_buf_detach(item->dmabuf, item->attach);
+ dma_buf_put(item->dmabuf);
+ kfree(item);
+ }
+exit:
+ mutex_unlock(&sess->imp_rb_lock);
+ dma_buf_put(dmabuf);
+ return rc;
+}
+
+int smm_ptr2vpu(struct vpumgr_smm *sess, unsigned long *arg)
+{
+ struct device *dev = sess->vdev->sdev;
+ struct task_struct *task = current;
+ struct dma_buf *dmabuf = NULL;
+ unsigned long vaddr = *arg;
+ struct vm_area_struct *vma;
+ struct vpusmm_impbuf *item;
+ struct mm_struct *mm;
+
+ mm = get_task_mm(task);
+ if (!mm)
+ goto failed;
+
+ mmap_read_lock(mm);
+
+ vma = find_vma(mm, vaddr);
+ if (!vma) {
+ dev_dbg(dev, "cannot find vaddr: %lx\n", vaddr);
+ goto failed;
+ }
+
+ if (vaddr < vma->vm_start) {
+ dev_dbg(dev, "failed at line %d, vaddr=%lx, vma->vm_start=%lx\n",
+ __LINE__, vaddr, vma->vm_start);
+ goto failed;
+ }
+
+ /* make sure the vma is backed by a dmabuf */
+ if (!vma->vm_file) {
+ dev_dbg(dev, "failed at line %d\n", __LINE__);
+ goto failed;
+ }
+
+ dmabuf = vma->vm_file->private_data;
+ if (!dmabuf) {
+ dev_dbg(dev, "failed at line %d\n", __LINE__);
+ goto failed;
+ }
+
+ if (dmabuf->file != vma->vm_file) {
+ dev_dbg(dev, "failed at line %d\n", __LINE__);
+ goto failed;
+ }
+ mmap_read_unlock(mm);
+ mmput(mm);
+
+ mutex_lock(&sess->imp_rb_lock);
+ item = vpusmm_find_impbuf(sess, dmabuf);
+ mutex_unlock(&sess->imp_rb_lock);
+
+ if (!item) {
+ dev_dbg(dev, "failed to find dmabuf in imported list for vaddr=0x%lx (%d)\n",
+ vaddr, __LINE__);
+ return -EFAULT;
+ }
+
+ *arg = (dma_addr_t)(vaddr - vma->vm_start) + item->vpu_addr;
+ return 0;
+
+failed:
+ if (mm) {
+ mmap_read_unlock(mm);
+ mmput(mm);
+ }
+ return -EFAULT;
+}
+
+int smm_debugfs_stats_show(struct seq_file *file, struct vpumgr_smm *sess)
+{
+ struct rb_node *cur, *next;
+ struct vpusmm_impbuf *item;
+ int i;
+
+ seq_puts(file, "\tdmabuf\texpname\tsize(bytes)\tfilecount\trefs\n");
+
+ mutex_lock(&sess->imp_rb_lock);
+ cur = rb_first(&sess->imp_rb);
+ i = 0;
+ while (cur) {
+ item = rb_entry(cur, struct vpusmm_impbuf, node);
+ next = rb_next(cur);
+ if (item)
+ seq_printf(file, "\t%d:%s\t%s\t%zu\t%ld\t%d\n", i++,
+ item->dmabuf->name ? : "",
+ item->dmabuf->exp_name,
+ item->dmabuf->size,
+ file_count(item->dmabuf->file),
+ item->refcount);
+ cur = next;
+ }
+ mutex_unlock(&sess->imp_rb_lock);
+ return 0;
+}
+
+int smm_init(struct vpumgr_device *vdev)
+{
+ int rc = 0;
+
+ if (!vdev->dev->of_node) {
+ /*
+ * no of_node imply:
+ * 1. no IOMMU, VPU device is only 32-bit DMA capable
+ * 2. use default CMA because no device tree node specifying memory-region
+ */
+ dma_set_mask(vdev->dev, DMA_BIT_MASK(32));
+ dma_set_coherent_mask(vdev->dev, DMA_BIT_MASK(32));
+ } else {
+ /* Initialize reserved memory resources */
+ rc = of_reserved_mem_device_init(vdev->dev);
+ if (rc) {
+ if (rc == -ENODEV) {
+ dev_warn(vdev->dev,
+ "No reserved memory specified, use default cma\n");
+ rc = 0;
+ } else {
+ dev_err(vdev->dev,
+ "Failed to init reserved memory, rc=%d\n", rc);
+ }
+ }
+ }
+ return rc;
+}
+
+int smm_fini(struct vpumgr_device *vdev)
+{
+ return 0;
+}
diff --git a/drivers/misc/vpumgr/vpu_smm.h b/drivers/misc/vpumgr/vpu_smm.h
new file mode 100644
index 000000000000..ff547649d95c
--- /dev/null
+++ b/drivers/misc/vpumgr/vpu_smm.h
@@ -0,0 +1,30 @@
+/* SPDX-License-Identifier: GPL-2.0-only
+ * Copyright (C) 2020 Intel Corporation
+ */
+#ifndef _VPU_SMM_H
+#define _VPU_SMM_H
+#include <linux/kernel.h>
+#include <linux/rbtree.h>
+
+#include "vpu_common.h"
+
+struct vpumgr_smm {
+ struct vpumgr_device *vdev;
+ struct rb_root imp_rb;
+ struct mutex imp_rb_lock; /* protects imp_rb */
+};
+
+int smm_init(struct vpumgr_device *vdev);
+int smm_fini(struct vpumgr_device *vdev);
+
+int smm_open(struct vpumgr_smm *sess, struct vpumgr_device *vdev);
+int smm_close(struct vpumgr_smm *sess);
+
+int smm_alloc(struct vpumgr_smm *sess, struct vpumgr_args_alloc *arg);
+int smm_import(struct vpumgr_smm *sess, struct vpumgr_args_import *arg);
+int smm_unimport(struct vpumgr_smm *sess, int *p_dmabuf_fd);
+int smm_ptr2vpu(struct vpumgr_smm *sess, unsigned long *arg);
+
+int smm_debugfs_stats_show(struct seq_file *file, struct vpumgr_smm *sess);
+
+#endif
diff --git a/drivers/misc/vpumgr/vpu_vcm.c b/drivers/misc/vpumgr/vpu_vcm.c
new file mode 100644
index 000000000000..44b9408d478e
--- /dev/null
+++ b/drivers/misc/vpumgr/vpu_vcm.c
@@ -0,0 +1,584 @@
+// SPDX-License-Identifier: GPL-2.0-only
+/*
+ * Copyright (C) 2020-2021 Intel Corporation
+ */
+#include <linux/delay.h>
+#include <linux/dma-buf.h>
+#include <linux/idr.h>
+#include <linux/mutex.h>
+#include <linux/kthread.h>
+#include <linux/sched/task.h>
+#include <linux/seq_file.h>
+#include <linux/uaccess.h>
+#include <linux/workqueue.h>
+#include <linux/xlink.h>
+#include "vpu_common.h"
+#include "vpu_vcm.h"
+
+#define XLINK_IPC_TIMEOUT 1000u
+
+/* Static xlink configuration */
+#define VCM_XLINK_CHANNEL 1
+#define VCM_XLINK_CHAN_SIZE 128
+
+static const int msg_header_size = offsetof(struct vcm_msg, payload.data);
+
+struct vpu_cmd {
+ struct work_struct work;
+ struct kref refcount;
+ struct xlink_handle *handle;
+ struct vpumgr_ctx *vctx; /* the submitting vpu context */
+ struct vcm_msg msg; /* message buffer for send/recv */
+ struct completion complete; /* completion for async submit/reply */
+ int submit_err; /* error code for submittion process */
+};
+
+static int vcm_vpu_link_init(struct vcm_dev *pvcm)
+{
+ struct vpumgr_device *vdev = container_of(pvcm, struct vpumgr_device, vcm);
+ enum xlink_error rc;
+
+ pvcm->ipc_xlink_handle.dev_type = VPUIP_DEVICE;
+ pvcm->ipc_xlink_handle.sw_device_id = pvcm->sw_dev_id;
+
+ rc = xlink_initialize();
+ if (rc != X_LINK_SUCCESS)
+ goto exit;
+
+ rc = xlink_connect(&pvcm->ipc_xlink_handle);
+ if (rc != X_LINK_SUCCESS)
+ goto exit;
+
+ rc = xlink_open_channel(&pvcm->ipc_xlink_handle, VCM_XLINK_CHANNEL,
+ RXB_TXB, VCM_XLINK_CHAN_SIZE, XLINK_IPC_TIMEOUT);
+ if (rc != X_LINK_SUCCESS && rc != X_LINK_ALREADY_OPEN) {
+ xlink_disconnect(&pvcm->ipc_xlink_handle);
+ goto exit;
+ }
+
+ rc = 0;
+exit:
+ dev_info(vdev->dev, "%s: rc = %d\n", __func__, rc);
+ return -(int)rc;
+}
+
+static int vcm_vpu_link_fini(struct vcm_dev *pvcm)
+{
+ xlink_close_channel(&pvcm->ipc_xlink_handle, VCM_XLINK_CHANNEL);
+ xlink_disconnect(&pvcm->ipc_xlink_handle);
+ return 0;
+}
+
+/*
+ * Send a vcm_msg by xlink.
+ * Given limited xlink payload size, packing is also performed.
+ */
+static int vcm_send(struct xlink_handle *xlnk_handle, struct vcm_msg *req)
+{
+ struct vpumgr_device *vdev;
+ enum xlink_error rc;
+ u8 *ptr = (u8 *)req;
+ u32 size = 0;
+ u32 len = req->size;
+
+ vdev = container_of(xlnk_handle, struct vpumgr_device, vcm.ipc_xlink_handle);
+ if (len > sizeof(*req))
+ return -EINVAL;
+ do {
+ size = len > VCM_XLINK_CHAN_SIZE ? VCM_XLINK_CHAN_SIZE : len;
+ rc = xlink_write_volatile(xlnk_handle, VCM_XLINK_CHANNEL, ptr, size);
+ if (rc != X_LINK_SUCCESS) {
+ dev_warn(vdev->dev, "%s xlink_write_volatile error %d\n", __func__, rc);
+ return -EINVAL;
+ }
+ ptr += size;
+ len -= size;
+ } while (len > 0);
+
+ return 0;
+}
+
+/*
+ * Receives a vcm_msg by xlink.
+ * Given limited xlink payload size, unpacking is also performed.
+ */
+static int vcm_recv(struct xlink_handle *xlnk_handle, struct vcm_msg *rep)
+{
+ struct vpumgr_device *vdev;
+ enum xlink_error rc;
+ u64 size;
+ u32 total_size = 0;
+ u32 rx_size = 0;
+ u8 *ptr = (u8 *)rep;
+
+ vdev = container_of(xlnk_handle, struct vpumgr_device, vcm.ipc_xlink_handle);
+ do {
+ /* workaround for a bug in xlink_read_data_to_buffer()
+ * although it's last argument is declared to be of type (u32 *), the
+ * function actually writes 64-bit value into that address.
+ */
+ rc = xlink_read_data_to_buffer(xlnk_handle, VCM_XLINK_CHANNEL, ptr, (u32 *)&size);
+ if (rc != X_LINK_SUCCESS) {
+ dev_warn(vdev->dev, "%s: xlink_read_data_to_buffer failed, rc:%d\n",
+ __func__, rc);
+ return -EPIPE;
+ }
+
+ if (total_size == 0) {
+ if (size < msg_header_size) {
+ dev_warn(vdev->dev, "%s: first packet is too small (%llu)\n",
+ __func__, size);
+ return -EINVAL;
+ }
+
+ total_size = rep->size;
+ if (total_size > sizeof(*rep)) {
+ dev_warn(vdev->dev, "%s: packet size (%u) is too big\n",
+ __func__, total_size);
+ return -EINVAL;
+ }
+ if (total_size < size) {
+ dev_warn(vdev->dev,
+ "%s: first packet is smaller than claimed (%llu)\n",
+ __func__, size);
+ return -EINVAL;
+ }
+ }
+
+ ptr += size;
+ rx_size += size;
+ } while (rx_size < total_size);
+
+ if (rx_size != total_size) {
+ dev_warn(vdev->dev, "%s: actuall size %u exceeds claimed size %ud\n",
+ __func__, rx_size, total_size);
+ return -EINVAL;
+ }
+
+ return 0;
+}
+
+static void vcmd_free(struct kref *kref)
+{
+ struct vpu_cmd *vcmd = container_of(kref, struct vpu_cmd, refcount);
+
+ kvfree(vcmd);
+}
+
+static struct vpu_cmd *vcmd_get(struct vcm_dev *pvcm, int msgid)
+{
+ struct vpu_cmd *vcmd;
+
+ mutex_lock(&pvcm->msg_idr_lock);
+ vcmd = idr_find(&pvcm->msg_idr, msgid);
+ if (vcmd)
+ kref_get(&vcmd->refcount);
+ mutex_unlock(&pvcm->msg_idr_lock);
+
+ return vcmd;
+}
+
+static void vcmd_put(struct vpu_cmd *vcmd)
+{
+ kref_put(&vcmd->refcount, vcmd_free);
+}
+
+static int vcmd_alloc_msgid(struct vcm_dev *pvcm, struct vpu_cmd *vcmd)
+{
+ int msgid;
+
+ mutex_lock(&pvcm->msg_idr_lock);
+ msgid = idr_alloc_cyclic(&pvcm->msg_idr, vcmd, 1, 0, GFP_KERNEL);
+ if (msgid >= 0)
+ kref_init(&vcmd->refcount);
+ mutex_unlock(&pvcm->msg_idr_lock);
+ return msgid;
+}
+
+static void vcmd_remove_msgid(struct vcm_dev *pvcm, int msgid)
+{
+ struct vpu_cmd *vcmd;
+
+ mutex_lock(&pvcm->msg_idr_lock);
+ vcmd = idr_remove(&pvcm->msg_idr, msgid);
+ if (vcmd)
+ kref_put(&vcmd->refcount, vcmd_free);
+ mutex_unlock(&pvcm->msg_idr_lock);
+}
+
+static void vcmd_clean_ctx(struct vcm_dev *pvcm, struct vpumgr_ctx *v)
+{
+ struct vpu_cmd *vcmd;
+ int i;
+
+ mutex_lock(&pvcm->msg_idr_lock);
+ idr_for_each_entry(&pvcm->msg_idr, vcmd, i)
+ if (vcmd->vctx == v) {
+ idr_remove(&pvcm->msg_idr, i);
+ kref_put(&vcmd->refcount, vcmd_free);
+ }
+ mutex_unlock(&pvcm->msg_idr_lock);
+}
+
+static int vcmd_count_ctx(struct vcm_dev *pvcm, struct vpumgr_ctx *v)
+{
+ struct vpu_cmd *vcmd;
+ int i;
+ int count = 0;
+
+ mutex_lock(&pvcm->msg_idr_lock);
+ idr_for_each_entry(&pvcm->msg_idr, vcmd, i)
+ if (vcmd->vctx == v)
+ count++;
+ mutex_unlock(&pvcm->msg_idr_lock);
+
+ return count;
+}
+
+static void vpu_cmd_submit(struct work_struct *work)
+{
+ struct vpu_cmd *p = container_of(work, struct vpu_cmd, work);
+
+ p->submit_err = vcm_send(p->handle, &p->msg);
+}
+
+/*
+ * vcm_submit() - Submit a command to VPU
+ * @v: Pointer to local vpu context data structure.
+ * @cmd: Command code
+ * @data_in: Data arguments
+ * @in_len: Length of the data arguments
+ * @submit_id: On return, this will containe a newly allocated
+ * vpu-device-wise unique ID for the submitted command
+ *
+ * Submit a command to corresponding vpu context running on firmware to execute
+ */
+int vcm_submit(struct vpumgr_ctx *v,
+ u32 cmd, const void *data_in, u32 in_len, s32 *submit_id)
+{
+ struct vcm_dev *pvcm = &v->vdev->vcm;
+ int ctx = v->vpu_ctx_id;
+ int rc = 0;
+ struct vpu_cmd *vcmd;
+
+ if (!v->vdev->vcm.enabled)
+ return -ENOENT;
+
+ if (in_len > VCM_PAYLOAD_SIZE)
+ return -EINVAL;
+
+ vcmd = kvmalloc(sizeof(*vcmd), GFP_KERNEL);
+ if (!vcmd)
+ return -ENOMEM;
+
+ vcmd->vctx = v;
+
+ rc = vcmd_alloc_msgid(pvcm, vcmd);
+ if (rc < 0) {
+ rc = -EEXIST;
+ kvfree(vcmd);
+ return rc;
+ }
+
+ /* from now on, vcmd can is refcount managed */
+ vcmd->msg.id = rc;
+ *submit_id = vcmd->msg.id;
+
+ if (data_in && in_len > 0) {
+ if (access_ok((void __user *)data_in, in_len)) {
+ rc = copy_from_user(vcmd->msg.payload.data,
+ (const void __user *)data_in, in_len);
+ if (rc)
+ goto remove_msgid;
+ } else {
+ memcpy(vcmd->msg.payload.data, data_in, in_len);
+ }
+ }
+
+ init_completion(&vcmd->complete);
+ vcmd->handle = &pvcm->ipc_xlink_handle;
+ vcmd->msg.size = msg_header_size + in_len;
+ vcmd->msg.ctx = ctx;
+ vcmd->msg.cmd = cmd;
+ vcmd->msg.rc = 0;
+ INIT_WORK(&vcmd->work, vpu_cmd_submit);
+
+ if (!queue_work(pvcm->wq, &vcmd->work)) {
+ rc = -EEXIST;
+ goto remove_msgid;
+ }
+
+ return 0;
+
+remove_msgid:
+ vcmd_remove_msgid(pvcm, vcmd->msg.id);
+ return rc;
+}
+
+/*
+ * vcm_wait() - Wait a submitted command to finish
+ * @v: Pointer to local vpu context data structure.
+ * @submit_id: Unique ID of the submitted command to wait for
+ * @vpu_rc: Return code of the submitted commands
+ * @data_out: Return data payload of the submitted command
+ * @p_out_len: Length of the returned paylaod
+ * @timeout_ms: Time in milliseconds before the wait expires
+ *
+ * Wait for a submitted command to finish and retrieves the
+ * return code and outputs on success with timeout.
+ */
+int vcm_wait(struct vpumgr_ctx *v, s32 submit_id,
+ s32 *vpu_rc, void *data_out, u32 *p_out_len, u32 timeout_ms)
+{
+ struct vcm_dev *pvcm = &v->vdev->vcm;
+ struct device *dev = v->vdev->sdev;
+ unsigned long timeout = msecs_to_jiffies(timeout_ms);
+ struct vpu_cmd *vcmd;
+ int rc, len;
+
+ if (!v->vdev->vcm.enabled)
+ return -ENOENT;
+
+ vcmd = vcmd_get(pvcm, submit_id);
+ if (!vcmd) {
+ dev_err(dev, "%s:cannot find submit_id %d\n", __func__, submit_id);
+ return -EINVAL;
+ }
+
+ if (v != vcmd->vctx) {
+ dev_err(dev, "%s:trying to wait on submit %d doesn't belong to vpu context %d\n",
+ __func__, submit_id, v->vpu_ctx_id);
+ return -EINVAL;
+ }
+
+ /* wait for submission work to be done */
+ flush_work(&vcmd->work);
+ rc = vcmd->submit_err;
+ if (rc)
+ goto exit;
+
+ /* wait for reply */
+ rc = wait_for_completion_interruptible_timeout(&vcmd->complete, timeout);
+ if (rc < 0) {
+ goto exit;
+ } else if (rc == 0) {
+ rc = -ETIMEDOUT;
+ goto exit;
+ } else {
+ /* wait_for_completion_interruptible_timeout return positive
+ * rc on success, but we return zero as success.
+ */
+ rc = 0;
+ }
+
+ if (vpu_rc)
+ *vpu_rc = vcmd->msg.rc;
+
+ if (data_out && p_out_len) {
+ /* truncate payload to fit output buffer size provided */
+ len = vcmd->msg.size - msg_header_size;
+ if (len > (*p_out_len)) {
+ dev_err(dev, "%s: output is truncated from %d to %d to fit buffer size.\n",
+ __func__, len, (*p_out_len));
+ len = (*p_out_len);
+ }
+
+ if (len > 0) {
+ if (access_ok((void __user *)data_out, len)) {
+ rc = copy_to_user((void __user *)data_out,
+ vcmd->msg.payload.data, len);
+ if (rc)
+ goto exit;
+ } else {
+ memcpy(data_out, vcmd->msg.payload.data, len);
+ }
+ }
+
+ /* tell the caller the exact length received, even if
+ * we have truncated due to output buffer size limitation
+ */
+ *p_out_len = len;
+ }
+exit:
+ v->total_vcmds++;
+ vcmd_put(vcmd);
+ vcmd_remove_msgid(pvcm, submit_id);
+ return rc;
+}
+
+static int vcm_call(struct vpumgr_ctx *v,
+ s32 cmd, const void *data_in, u32 in_len,
+ s32 *res_rc, void *data_out, u32 *p_out_len)
+{
+ int submit_id, rc;
+
+ if (!v->vdev->vcm.enabled)
+ return -ENOENT;
+
+ rc = vcm_submit(v, cmd, data_in, in_len, &submit_id);
+ if (rc)
+ return rc;
+
+ return vcm_wait(v, submit_id, res_rc, data_out, p_out_len, 1000);
+}
+
+int vcm_open(struct vpumgr_ctx *v, struct vpumgr_device *vdev)
+{
+ struct device *dev = vdev->sdev;
+ int rep_rc, rc;
+
+ v->vdev = vdev;
+
+ if (!vdev->vcm.enabled)
+ return 0;
+
+ rc = vcm_call(v, VCTX_MSG_CREATE, NULL, 0, &rep_rc, NULL, NULL);
+
+ if (rc != 0 || rep_rc < 0)
+ dev_err(dev, "%s: Vpu context create with rc:%d and vpu reply rc:%d\n",
+ __func__, rc, rep_rc);
+ if (rc)
+ return rc;
+ if (rep_rc < 0)
+ return -ENXIO;
+
+ v->vpu_ctx_id = rep_rc;
+ v->total_vcmds = 0;
+ return 0;
+}
+
+int vcm_close(struct vpumgr_ctx *v)
+{
+ struct device *dev = v->vdev->sdev;
+ int rep_rc, rc;
+
+ if (!v->vdev->vcm.enabled)
+ return 0;
+
+ rc = vcm_call(v, VCTX_MSG_DESTROY, NULL, 0, &rep_rc, NULL, NULL);
+ dev_dbg(dev, "vpu context %d is destroyed with rc:%d and vpu reply rc:%d\n",
+ v->vpu_ctx_id, rc, rep_rc);
+
+ /* remove submit belongs to this context */
+ vcmd_clean_ctx(&v->vdev->vcm, v);
+ return 0;
+}
+
+int vcm_debugfs_stats_show(struct seq_file *file, struct vpumgr_ctx *v)
+{
+ if (!v->vdev->vcm.enabled)
+ return 0;
+ seq_printf(file, "\tvpu context: #%d\n", v->vpu_ctx_id);
+ seq_printf(file, "\t\tNum of completed cmds: %llu\n", v->total_vcmds);
+ seq_printf(file, "\t\tNum of on-going cmds: %d\n", vcmd_count_ctx(&v->vdev->vcm, v));
+ return 0;
+}
+
+static int vcm_rxthread(void *param)
+{
+ struct vpumgr_device *vdev = param;
+ struct device *dev = vdev->sdev;
+ struct vcm_dev *pvcm = &vdev->vcm;
+ struct vcm_msg *msg = &pvcm->rxmsg;
+ struct vpu_cmd *vcmd;
+ int rc;
+
+ while (!kthread_should_stop()) {
+ rc = vcm_recv(&pvcm->ipc_xlink_handle, msg);
+ if (rc == -EPIPE)
+ break;
+ if (rc)
+ continue;
+
+ switch (msg->cmd) {
+ case VCTX_MSG_REPLY:
+ /* find local data associated with that msg id */
+ vcmd = vcmd_get(pvcm, (unsigned long)msg->id);
+ if (!vcmd)
+ break;
+
+ if (msg->ctx != vcmd->msg.ctx)
+ dev_warn(dev, "reply msg #%u's ctx (%u) mismatches vcmd ctx (%u)\n",
+ msg->id, msg->ctx, vcmd->msg.ctx);
+
+ vcmd->submit_err = 0;
+
+ /* submit corresponding to msg->id is done, do post process */
+ memcpy(&vcmd->msg, msg, msg->size);
+ complete(&vcmd->complete);
+
+ vcmd_put(vcmd);
+ break;
+ default:
+ break;
+ }
+ }
+ return rc;
+}
+
+int vcm_init(struct vpumgr_device *vdev, u32 sw_dev_id)
+{
+ struct vcm_dev *pvcm = &vdev->vcm;
+ struct task_struct *rxthread;
+ int rc = 0;
+
+ if (sw_dev_id == XLINK_INVALID_SW_DEVID) {
+ dev_warn(vdev->dev, "%s: vcm is not enabled!\n",
+ __func__);
+ rc = 0;
+ goto exit;
+ }
+
+ pvcm->sw_dev_id = sw_dev_id;
+ rc = vcm_vpu_link_init(pvcm);
+ if (rc)
+ goto exit;
+
+ pvcm->wq = alloc_ordered_workqueue("vcm workqueue", WQ_MEM_RECLAIM | WQ_HIGHPRI);
+ if (!pvcm->wq) {
+ rc = -ENOMEM;
+ goto vpu_link_fini;
+ }
+
+ mutex_init(&pvcm->msg_idr_lock);
+ idr_init(&pvcm->msg_idr);
+
+ rxthread = kthread_run(vcm_rxthread,
+ (void *)vdev, "vcmrx");
+ if (IS_ERR(rxthread)) {
+ rc = PTR_ERR(rxthread);
+ goto destroy_idr;
+ }
+
+ pvcm->rxthread = get_task_struct(rxthread);
+
+ pvcm->enabled = true;
+ return 0;
+
+destroy_idr:
+ idr_destroy(&pvcm->msg_idr);
+ destroy_workqueue(pvcm->wq);
+vpu_link_fini:
+ vcm_vpu_link_fini(pvcm);
+exit:
+ pvcm->enabled = false;
+ return rc;
+}
+
+int vcm_fini(struct vpumgr_device *vdev)
+{
+ struct vcm_dev *pvcm = &vdev->vcm;
+
+ if (!pvcm->enabled)
+ return 0;
+
+ vcm_vpu_link_fini(pvcm);
+
+ kthread_stop(pvcm->rxthread);
+ put_task_struct(pvcm->rxthread);
+
+ idr_destroy(&pvcm->msg_idr);
+ destroy_workqueue(pvcm->wq);
+ mutex_destroy(&pvcm->msg_idr_lock);
+ return 0;
+}
diff --git a/drivers/misc/vpumgr/vpu_vcm.h b/drivers/misc/vpumgr/vpu_vcm.h
new file mode 100644
index 000000000000..9e89c281092b
--- /dev/null
+++ b/drivers/misc/vpumgr/vpu_vcm.h
@@ -0,0 +1,84 @@
+/* SPDX-License-Identifier: GPL-2.0-only
+ * RESMGR driver - VPU Context Manager
+ * Copyright (C) 2020-2021 Intel Corporation
+ */
+#ifndef __VPU_VCM_H
+#define __VPU_VCM_H
+
+#include <linux/xlink.h>
+
+struct vpumgr_device;
+
+/* Command code for message to/from VPU context manager on firmware */
+#define VCTX_MSG_CREATE 1
+#define VCTX_MSG_DESTROY 2
+#define VCTX_MSG_REPLY 3
+
+/* Maximal payload size supported for request or reply */
+#define VCM_PAYLOAD_SIZE (8192 - 5 * sizeof(u32))
+
+/**
+ * struct vcm_msg: VPU context manager message
+ * @size: size tag must be at the begin
+ * @ctx: to/from which context the msg is
+ * @cmd: the type of this message
+ * @id: index to identify this message in context ctx
+ * @rc: return code or misc args
+ * @payload: the payload of message
+ */
+struct vcm_msg {
+ u32 size;
+ u32 ctx;
+ u32 cmd;
+ u32 id;
+ s32 rc;
+ union {
+ char data[VCM_PAYLOAD_SIZE];
+ } payload;
+} __packed;
+
+struct vcm_dev {
+ bool enabled;
+ /*
+ * XLINK IPC related.
+ */
+ struct xlink_handle ipc_xlink_handle;
+ s32 sw_dev_id;
+
+ /*
+ * dispatch work queue.
+ * Xlink transcations would handled in the work queue;
+ * Decouple the xlink API invoking with user space systemcall
+ * because SIGINT would cause xlink_read* erro.
+ */
+ struct workqueue_struct *wq;
+
+ /* kthread for rx */
+ struct task_struct *rxthread;
+
+ /* message buffer for receiving thread */
+ struct vcm_msg rxmsg;
+
+ struct mutex msg_idr_lock; /* protects msg_idr */
+ struct idr msg_idr;
+};
+
+struct vpumgr_ctx {
+ struct vpumgr_device *vdev;
+ u32 vpu_ctx_id;
+ u64 total_vcmds;
+};
+
+int vcm_init(struct vpumgr_device *vdev, u32 sw_dev_id);
+int vcm_fini(struct vpumgr_device *vdev);
+
+int vcm_open(struct vpumgr_ctx *v, struct vpumgr_device *vdev);
+int vcm_close(struct vpumgr_ctx *v);
+int vcm_debugfs_stats_show(struct seq_file *file, struct vpumgr_ctx *v);
+
+int vcm_submit(struct vpumgr_ctx *v,
+ u32 cmd, const void *data_in, u32 in_len, s32 *submit_id);
+int vcm_wait(struct vpumgr_ctx *v, s32 submit_id,
+ s32 *vpu_rc, void *data_out, u32 *p_out_len, u32 timeout_ms);
+
+#endif /* __VPU_VCM_H */
diff --git a/include/uapi/misc/vpumgr.h b/include/uapi/misc/vpumgr.h
new file mode 100644
index 000000000000..910b26e60097
--- /dev/null
+++ b/include/uapi/misc/vpumgr.h
@@ -0,0 +1,64 @@
+/* SPDX-License-Identifier: GPL-2.0+ WITH Linux-syscall-note
+ * VPU manager Linux Kernel API
+ * Copyright (C) 2020-2021 Intel Corporation
+ *
+ */
+#ifndef __VPUMGR_UAPI_H
+#define __VPUMGR_UAPI_H
+
+#include <linux/types.h>
+
+/* ioctl numbers */
+#define VPUMGR_MAGIC 'V'
+/* VPU manager IOCTLs */
+#define VPUMGR_IOCTL_DMABUF_ALLOC _IOWR(VPUMGR_MAGIC, 2, struct vpumgr_args_alloc)
+#define VPUMGR_IOCTL_DMABUF_IMPORT _IOWR(VPUMGR_MAGIC, 3, struct vpumgr_args_import)
+#define VPUMGR_IOCTL_DMABUF_UNIMPORT _IOWR(VPUMGR_MAGIC, 4, __s32)
+#define VPUMGR_IOCTL_DMABUF_PTR2VPU _IOWR(VPUMGR_MAGIC, 5, __u64)
+#define VPUMGR_IOCTL_VCM_SUBMIT _IOWR(VPUMGR_MAGIC, 6, struct vpumgr_vcm_submit)
+#define VPUMGR_IOCTL_VCM_WAIT _IOWR(VPUMGR_MAGIC, 7, struct vpumgr_vcm_wait)
+#define VPUMGR_IOCTL_END _IO(VPUMGR_MAGIC, 8)
+
+struct vpumgr_args_alloc {
+ __s32 fd; /* out: DMABuf fd */
+ __s32 reserved[2]; /* in: reserved */
+ __u64 size; /* in: required buffer size */
+};
+
+/* vpu_access flags */
+enum vpu_access_type {
+ VPU_ACCESS_DEFAULT = 0,
+ VPU_ACCESS_READ = 1,
+ VPU_ACCESS_WRITE = 2,
+ VPU_ACCESS_RW = 3
+};
+
+struct vpumgr_args_import {
+ __s32 fd; /* in: input DMABuf fd */
+ __s32 vpu_access; /* in: how vpu is going to access the buffer */
+ __u64 vpu_addr; /* out: vpu dma address of the DMABuf */
+ __u64 size; /* out: the size of the DMABuf */
+};
+
+/* Command code reserved for kernel mode driver,
+ * user-space should not use commmand code smaller
+ * than or equal to this micro
+ */
+#define VCTX_KMD_RESERVED_CMD_LAST 31
+
+struct vpumgr_vcm_submit {
+ __u32 cmd; /* in: command code */
+ __u64 in; /* in: input paramer buffer address */
+ __u32 in_len; /* in: input paramer buffer length */
+ __s32 submit_id; /* out: submit id */
+};
+
+struct vpumgr_vcm_wait {
+ __s32 submit_id; /* in: submit id */
+ __s32 vpu_rc; /* out: vpu return code */
+ __u64 out; /* in: address of the buffer for receiving result */
+ __u32 out_len; /* in: length of the result buffer */
+ __u32 timeout_ms; /* in: timeout in milliseconds */
+};
+
+#endif /* __VPUMGR_UAPI_H */
--
2.17.1
Powered by blists - more mailing lists