[RFC] ARM: bcm2835: register dmabounce on devices hooked to main interconnect

[RFC] ARM: bcm2835: register dmabounce on devices hooked to main interconnect

[Nicolas Saenz Julienne]

NOTE: This patch builds upon Stefan's series providing basic support for
RPi4[1]. I'm mostly interested in verifying if this is the correct approach
to the issue stated below. If so I assume this will be added to Stefan's
v2 series.

The new Raspberry Pi 4 happens to have weird DMA constraints. Even
though it might contain up to 4 GB of ram, most devices can only access
the first lower GB of memory.

This breaks the overall assumption DMA API makes whereas 32-bit DMA
masks are always supported[2], and potentially breaks DMA addressing for
all streaming DMA users. This has already been observed with
'sdhci-iproc' but might as well happen elsewhere. Note that contiguous
allocations are safe as 'dma_zone_size' is set accordingly.

To get around that limitation we register arm's dmabounce dma-ops on all
devices hooked to the SoC's main interconnect.

[1] https://www.spinics.net/lists/arm-kernel/msg742120.html
[2] https://www.spinics.net/lists/arm-kernel/msg742736.html

Signed-off-by: Nicolas Saenz Julienne <nsaenzjulienne@suse.de>
---
 arch/arm/mach-bcm/Kconfig         |  1 +
 arch/arm/mach-bcm/board_bcm2835.c | 29 +++++++++++++++++++++++++++++
 2 files changed, 30 insertions(+)

diff --git a/arch/arm/mach-bcm/Kconfig b/arch/arm/mach-bcm/Kconfig
index 5e5f1fabc3d4..588326f7e269 100644
--- a/arch/arm/mach-bcm/Kconfig
+++ b/arch/arm/mach-bcm/Kconfig
@@ -168,6 +168,7 @@ config ARCH_BCM2835
 	select PINCTRL
 	select PINCTRL_BCM2835
 	select MFD_CORE
+	select DMABOUNCE if ARCH_MULTI_V7
 	help
 	  This enables support for the Broadcom BCM2835 and BCM2836 SoCs.
 	  This SoC is used in the Raspberry Pi and Roku 2 devices.
diff --git a/arch/arm/mach-bcm/board_bcm2835.c b/arch/arm/mach-bcm/board_bcm2835.c
index c09cf25596af..7aff29f77ca7 100644
--- a/arch/arm/mach-bcm/board_bcm2835.c
+++ b/arch/arm/mach-bcm/board_bcm2835.c
@@ -3,6 +3,8 @@
  * Copyright (C) 2010 Broadcom
  */
 
+#include <linux/device.h>
+#include <linux/dma-mapping.h>
 #include <linux/init.h>
 #include <linux/irqchip.h>
 #include <linux/of_address.h>
@@ -24,8 +26,35 @@ static const char * const bcm2835_compat[] = {
 	NULL
 };
 
+static int bcm2835_needs_bounce(struct device *dev, dma_addr_t dma_addr, size_t size)
+{
+	/*
+	 * The accepted dma addresses are [0xc0000000, 0xffffffff] which map to
+	 * ram's [0x00000000, 0x3fffffff].
+	 */
+	return dma_addr < 3ULL * SZ_1G;
+}
+
+static int bcm2835_platform_notify(struct device *dev)
+{
+	if (dev->parent && !strcmp("soc", dev_name(dev->parent))) {
+		dev->dma_mask = &dev->coherent_dma_mask;
+		dev->coherent_dma_mask = DMA_BIT_MASK(30);
+		dmabounce_register_dev(dev, 2048, 4096, bcm2835_needs_bounce);
+	}
+
+	return 0;
+}
+
+void __init bcm2835_init_early(void)
+{
+	if(of_machine_is_compatible("brcm,bcm2711"))
+		platform_notify = bcm2835_platform_notify;
+}
+
 DT_MACHINE_START(BCM2835, "BCM2835")
 	.dma_zone_size	= SZ_1G,
 	.dt_compat = bcm2835_compat,
 	.smp = smp_ops(bcm2836_smp_ops),
+	.init_early = bcm2835_init_early,
 MACHINE_END
-- 
2.22.0

Re: [RFC] ARM: bcm2835: register dmabounce on devices hooked to main interconnect

[Christoph Hellwig]

> +static int bcm2835_needs_bounce(struct device *dev, dma_addr_t dma_addr, size_t size)

Too long line..

> +void __init bcm2835_init_early(void)
> +{
> +	if(of_machine_is_compatible("brcm,bcm2711"))

Odd formatting.

Otherwise this looks good to me.

Re: [RFC] ARM: bcm2835: register dmabounce on devices hooked to main interconnect

[Florian Fainelli]

On 7/23/2019 6:34 PM, Christoph Hellwig wrote:
>> +static int bcm2835_needs_bounce(struct device *dev, dma_addr_t dma_addr, size_t size)
> 
> Too long line..
> 
>> +void __init bcm2835_init_early(void)
>> +{
>> +	if(of_machine_is_compatible("brcm,bcm2711"))
> 
> Odd formatting.
> 
> Otherwise this looks good to me.

Is this really the right way to solve this problem? First this is ARM
32-bit specific, and second, should not we have a way to indicate via
device tree that all peripherals behind the "soc" simple-bus parent node
are limited to 32-bit of DMA masks, but the specific memory map of the
BCM283x/BCM2711 makes it that only the last 1GB (0xC000_0000 -
0xffff_ffff) (which dma-ranges conveys already) is suitable for DMA into
the VPU uncached alias?
-- 
Florian