openwrt/target/linux/bcm27xx/patches-6.6/950-0674-vc_mem-Add-the-DMA-memcpy-support-from-bcm2708_fb.patch
Álvaro Fernández Rojas 8c405cdccc bcm27xx: add 6.6 kernel patches
The patches were generated from the RPi repo with the following command:
git format-patch v6.6.34..rpi-6.1.y

Some patches needed rebasing and, as usual, the applied and reverted, wireless
drivers, Github workflows, READMEs and defconfigs patches were removed.

Signed-off-by: Álvaro Fernández Rojas <noltari@gmail.com>
2024-06-18 18:52:49 +02:00

346 lines
8.2 KiB
Diff

From 76e509fd83156e96156e089eccfd49e2ebda5ec5 Mon Sep 17 00:00:00 2001
From: Phil Elwell <phil@raspberrypi.com>
Date: Sat, 14 Oct 2023 14:57:49 +0100
Subject: [PATCH 0674/1085] vc_mem: Add the DMA memcpy support from bcm2708_fb
bcm2708_fb is disabled by the vc4-kms-v3d overlay, which means that the
DMA memcpy support it provides is not available to allow vclog to read
the VC logs from the top 16MB on Pi 2 and Pi 3. Add the code to the
vc_mem driver, which will still be enabled.
It ought to be possible to do a proper DMA_MEM_TO_MEM copy via the
generic DMA customer API, but that can be a later step.
Signed-off-by: Phil Elwell <phil@raspberrypi.com>
---
drivers/char/broadcom/vc_mem.c | 259 +++++++++++++++++++++++++++++++++
1 file changed, 259 insertions(+)
--- a/drivers/char/broadcom/vc_mem.c
+++ b/drivers/char/broadcom/vc_mem.c
@@ -23,9 +23,21 @@
#include <linux/uaccess.h>
#include <linux/dma-mapping.h>
#include <linux/broadcom/vc_mem.h>
+#include <linux/compat.h>
+#include <linux/platform_data/dma-bcm2708.h>
+#include <soc/bcm2835/raspberrypi-firmware.h>
#define DRIVER_NAME "vc-mem"
+/* N.B. These use a different magic value for compatibility with bmc7208_fb */
+#define VC_MEM_IOC_DMACOPY _IOW('z', 0x22, struct vc_mem_dmacopy)
+#define VC_MEM_IOC_DMACOPY32 _IOW('z', 0x22, struct vc_mem_dmacopy32)
+
+/* address with no aliases */
+#define INTALIAS_NORMAL(x) ((x) & ~0xc0000000)
+/* cache coherent but non-allocating in L1 and L2 */
+#define INTALIAS_L1L2_NONALLOCATING(x) (((x) & ~0xc0000000) | 0x80000000)
+
/* Device (/dev) related variables */
static dev_t vc_mem_devnum;
static struct class *vc_mem_class;
@@ -36,6 +48,20 @@ static int vc_mem_inited;
static struct dentry *vc_mem_debugfs_entry;
#endif
+struct vc_mem_dmacopy {
+ void *dst;
+ __u32 src;
+ __u32 length;
+};
+
+#ifdef CONFIG_COMPAT
+struct vc_mem_dmacopy32 {
+ compat_uptr_t dst;
+ __u32 src;
+ __u32 length;
+};
+#endif
+
/*
* Videocore memory addresses and size
*
@@ -62,6 +88,20 @@ static uint phys_addr;
static uint mem_size;
static uint mem_base;
+struct vc_mem_dma {
+ struct device *dev;
+ int dma_chan;
+ int dma_irq;
+ void __iomem *dma_chan_base;
+ wait_queue_head_t dma_waitq;
+ void *cb_base; /* DMA control blocks */
+ dma_addr_t cb_handle;
+};
+
+struct { u32 base, length; } gpu_mem;
+static struct mutex dma_mutex;
+static struct vc_mem_dma vc_mem_dma;
+
static int
vc_mem_open(struct inode *inode, struct file *file)
{
@@ -99,6 +139,189 @@ vc_mem_get_current_size(void)
}
EXPORT_SYMBOL_GPL(vc_mem_get_current_size);
+static int
+vc_mem_dma_init(void)
+{
+ struct vc_mem_dma *vcdma = &vc_mem_dma;
+ struct platform_device *pdev;
+ struct device_node *fwnode;
+ struct rpi_firmware *fw;
+ struct device *dev;
+ u32 revision;
+ int rc;
+
+ if (vcdma->dev)
+ return 0;
+
+ fwnode = of_find_node_by_path("/system");
+ rc = of_property_read_u32(fwnode, "linux,revision", &revision);
+ revision = (revision >> 12) & 0xf;
+ if (revision != 1 && revision != 2) {
+ /* Only BCM2709 and BCM2710 may have logs where the ARMs
+ * can't see them.
+ */
+ return -ENXIO;
+ }
+
+ fwnode = rpi_firmware_find_node();
+ if (!fwnode)
+ return -ENXIO;
+
+ pdev = of_find_device_by_node(fwnode);
+ dev = &pdev->dev;
+
+ rc = dma_coerce_mask_and_coherent(dev, DMA_BIT_MASK(32));
+ if (rc)
+ return rc;
+
+ fw = rpi_firmware_get(fwnode);
+ if (!fw)
+ return -ENXIO;
+ rc = rpi_firmware_property(fw, RPI_FIRMWARE_GET_VC_MEMORY,
+ &gpu_mem, sizeof(gpu_mem));
+ if (rc)
+ return rc;
+
+ gpu_mem.base = INTALIAS_NORMAL(gpu_mem.base);
+
+ if (!gpu_mem.base || !gpu_mem.length) {
+ dev_err(dev, "%s: unable to determine gpu memory (%x,%x)\n",
+ __func__, gpu_mem.base, gpu_mem.length);
+ return -EFAULT;
+ }
+
+ vcdma->cb_base = dma_alloc_wc(dev, SZ_4K, &vcdma->cb_handle, GFP_KERNEL);
+ if (!vcdma->cb_base) {
+ dev_err(dev, "failed to allocate DMA CBs\n");
+ return -ENOMEM;
+ }
+
+ rc = bcm_dma_chan_alloc(BCM_DMA_FEATURE_BULK,
+ &vcdma->dma_chan_base,
+ &vcdma->dma_irq);
+ if (rc < 0) {
+ dev_err(dev, "failed to allocate a DMA channel\n");
+ goto free_cb;
+ }
+
+ vcdma->dma_chan = rc;
+
+ init_waitqueue_head(&vcdma->dma_waitq);
+
+ vcdma->dev = dev;
+
+ return 0;
+
+free_cb:
+ dma_free_wc(dev, SZ_4K, vcdma->cb_base, vcdma->cb_handle);
+
+ return rc;
+}
+
+static void
+vc_mem_dma_uninit(void)
+{
+ struct vc_mem_dma *vcdma = &vc_mem_dma;
+
+ if (vcdma->dev) {
+ bcm_dma_chan_free(vcdma->dma_chan);
+ dma_free_wc(vcdma->dev, SZ_4K, vcdma->cb_base, vcdma->cb_handle);
+ vcdma->dev = NULL;
+ }
+}
+
+static int dma_memcpy(struct vc_mem_dma *vcdma, dma_addr_t dst, dma_addr_t src,
+ int size)
+{
+ struct bcm2708_dma_cb *cb = vcdma->cb_base;
+ int burst_size = (vcdma->dma_chan == 0) ? 8 : 2;
+
+ cb->info = BCM2708_DMA_BURST(burst_size) | BCM2708_DMA_S_WIDTH |
+ BCM2708_DMA_S_INC | BCM2708_DMA_D_WIDTH |
+ BCM2708_DMA_D_INC;
+ cb->dst = dst;
+ cb->src = src;
+ cb->length = size;
+ cb->stride = 0;
+ cb->pad[0] = 0;
+ cb->pad[1] = 0;
+ cb->next = 0;
+
+ bcm_dma_start(vcdma->dma_chan_base, vcdma->cb_handle);
+ bcm_dma_wait_idle(vcdma->dma_chan_base);
+
+ return 0;
+}
+
+static long vc_mem_copy(struct vc_mem_dmacopy *ioparam)
+{
+ struct vc_mem_dma *vcdma = &vc_mem_dma;
+ size_t size = PAGE_SIZE;
+ const u32 dma_xfer_chunk = 256;
+ u32 *buf = NULL;
+ dma_addr_t bus_addr;
+ long rc = 0;
+ size_t offset;
+
+ /* restrict this to root user */
+ if (!uid_eq(current_euid(), GLOBAL_ROOT_UID))
+ return -EFAULT;
+
+ if (mutex_lock_interruptible(&dma_mutex))
+ return -EINTR;
+
+ rc = vc_mem_dma_init();
+ if (rc)
+ goto out;
+
+ vcdma = &vc_mem_dma;
+
+ if (INTALIAS_NORMAL(ioparam->src) < gpu_mem.base ||
+ INTALIAS_NORMAL(ioparam->src) >= gpu_mem.base + gpu_mem.length) {
+ pr_err("%s: invalid memory access %x (%x-%x)", __func__,
+ INTALIAS_NORMAL(ioparam->src), gpu_mem.base,
+ gpu_mem.base + gpu_mem.length);
+ rc = -EFAULT;
+ goto out;
+ }
+
+ buf = dma_alloc_coherent(vcdma->dev, PAGE_ALIGN(size), &bus_addr,
+ GFP_ATOMIC);
+ if (!buf) {
+ rc = -ENOMEM;
+ goto out;
+ }
+
+ for (offset = 0; offset < ioparam->length; offset += size) {
+ size_t remaining = ioparam->length - offset;
+ size_t s = min(size, remaining);
+ u8 *p = (u8 *)((uintptr_t)ioparam->src + offset);
+ u8 *q = (u8 *)ioparam->dst + offset;
+
+ rc = dma_memcpy(vcdma, bus_addr,
+ INTALIAS_L1L2_NONALLOCATING((u32)(uintptr_t)p),
+ (s + dma_xfer_chunk - 1) & ~(dma_xfer_chunk - 1));
+ if (rc) {
+ dev_err(vcdma->dev, "dma_memcpy failed\n");
+ break;
+ }
+ if (copy_to_user(q, buf, s) != 0) {
+ pr_err("%s: copy_to_user failed\n", __func__);
+ rc = -EFAULT;
+ break;
+ }
+ }
+
+out:
+ if (buf)
+ dma_free_coherent(vcdma->dev, PAGE_ALIGN(size), buf,
+ bus_addr);
+
+ mutex_unlock(&dma_mutex);
+
+ return rc;
+}
+
static long
vc_mem_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
{
@@ -163,6 +386,21 @@ vc_mem_ioctl(struct file *file, unsigned
}
break;
}
+ case VC_MEM_IOC_DMACOPY:
+ {
+ struct vc_mem_dmacopy ioparam;
+ /* Get the parameter data.
+ */
+ if (copy_from_user
+ (&ioparam, (void *)arg, sizeof(ioparam))) {
+ pr_err("%s: copy_from_user failed\n", __func__);
+ rc = -EFAULT;
+ break;
+ }
+
+ rc = vc_mem_copy(&ioparam);
+ break;
+ }
default:
{
return -ENOTTY;
@@ -193,6 +431,24 @@ vc_mem_compat_ioctl(struct file *file, u
break;
+ case VC_MEM_IOC_DMACOPY32:
+ {
+ struct vc_mem_dmacopy32 param32;
+ struct vc_mem_dmacopy param;
+ /* Get the parameter data.
+ */
+ if (copy_from_user(&param32, (void *)arg, sizeof(param32))) {
+ pr_err("%s: copy_from_user failed\n", __func__);
+ rc = -EFAULT;
+ break;
+ }
+ param.dst = compat_ptr(param32.dst);
+ param.src = param32.src;
+ param.length = param32.length;
+ rc = vc_mem_copy(&param);
+ break;
+ }
+
default:
rc = vc_mem_ioctl(file, cmd, arg);
break;
@@ -330,6 +586,7 @@ vc_mem_init(void)
vc_mem_debugfs_init(dev);
#endif
+ mutex_init(&dma_mutex);
vc_mem_inited = 1;
return 0;
@@ -352,6 +609,7 @@ vc_mem_exit(void)
{
pr_debug("%s: called\n", __func__);
+ vc_mem_dma_uninit();
if (vc_mem_inited) {
#ifdef CONFIG_DEBUG_FS
vc_mem_debugfs_deinit();
@@ -360,6 +618,7 @@ vc_mem_exit(void)
class_destroy(vc_mem_class);
cdev_del(&vc_mem_cdev);
unregister_chrdev_region(vc_mem_devnum, 1);
+ vc_mem_inited = 0;
}
}