[V4] LoongArch: Make WriteCombine configurable for ioremap()
Commit Message
LoongArch maintains cache coherency in hardware, but when paired with
LS7A chipsets the WUC attribute (Weak-ordered UnCached, which is similar
to WriteCombine) is out of the scope of cache coherency machanism for
PCIe devices (this is a PCIe protocol violation, which may be fixed in
newer chipsets).
This means WUC can only used for write-only memory regions now, so this
option is disabled by default, making WUC silently fallback to SUC for
ioremap(). You can enable this option if the kernel is ensured to run on
hardware without this bug.
Kernel parameter writecombine=on/off can be used to override the Kconfig
option.
Suggested-by: WANG Xuerui <kernel@xen0n.name>
Reviewed-by: WANG Xuerui <kernel@xen0n.name>
Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
---
V2: Add kernel prameter and update commit messages.
V3: Add a warning message for invalid kernel parameters.
V4: Document the writecombine=on/off parameter.
.../admin-guide/kernel-parameters.rst | 1 +
.../admin-guide/kernel-parameters.txt | 6 ++++++
arch/loongarch/Kconfig | 16 ++++++++++++++
arch/loongarch/include/asm/io.h | 4 +++-
arch/loongarch/kernel/setup.c | 21 +++++++++++++++++++
5 files changed, 47 insertions(+), 1 deletion(-)
Comments
On Thu, 2023-03-16 at 14:41 +0800, Huacai Chen wrote:
> LoongArch maintains cache coherency in hardware, but when paired with
> LS7A chipsets the WUC attribute (Weak-ordered UnCached, which is similar
> to WriteCombine) is out of the scope of cache coherency machanism for
> PCIe devices (this is a PCIe protocol violation, which may be fixed in
> newer chipsets).
>
> This means WUC can only used for write-only memory regions now, so this
> option is disabled by default, making WUC silently fallback to SUC for
> ioremap(). You can enable this option if the kernel is ensured to run on
> hardware without this bug.
>
> Kernel parameter writecombine=on/off can be used to override the Kconfig
> option.
>
> Suggested-by: WANG Xuerui <kernel@xen0n.name>
> Reviewed-by: WANG Xuerui <kernel@xen0n.name>
> Signed-off-by: Huacai Chen <chenhuacai@loongson.cn>
LGTM.
I still prefer an automatic way, but anyway we can implement it later
after a bug-free LS7A successor is launched.
Should we Cc: stable@vger.kernel.org and make a PR for 6.3 as well? To
me it's a "bug fix" and needed for stable releases, but I'm not sure.
@@ -128,6 +128,7 @@ parameter is applicable::
KVM Kernel Virtual Machine support is enabled.
LIBATA Libata driver is enabled
LP Printer support is enabled.
+ LOONGARCH LoongArch architecture is enabled.
LOOP Loopback device support is enabled.
M68k M68k architecture is enabled.
These options have more detailed description inside of
@@ -6933,6 +6933,12 @@
When enabled, memory and cache locality will be
impacted.
+ writecombine= [LOONGARCH] Control the MAT (Memory Access Type) of
+ ioremap_wc().
+
+ on - Enable writecombine, use WUC for ioremap_wc()
+ off - Disable writecombine, use SUC for ioremap_wc()
+
x2apic_phys [X86-64,APIC] Use x2apic physical mode instead of
default x2apic cluster mode on platforms
supporting x2apic.
@@ -447,6 +447,22 @@ config ARCH_IOREMAP
protection support. However, you can enable LoongArch DMW-based
ioremap() for better performance.
+config ARCH_WRITECOMBINE
+ bool "Enable WriteCombine (WUC) for ioremap()"
+ help
+ LoongArch maintains cache coherency in hardware, but when paired
+ with LS7A chipsets the WUC attribute (Weak-ordered UnCached, which
+ is similar to WriteCombine) is out of the scope of cache coherency
+ machanism for PCIe devices (this is a PCIe protocol violation, which
+ may be fixed in newer chipsets).
+
+ This means WUC can only used for write-only memory regions now, so
+ this option is disabled by default, making WUC silently fallback to
+ SUC for ioremap(). You can enable this option if the kernel is ensured
+ to run on hardware without this bug.
+
+ You can override this setting via writecombine=on/off boot parameter.
+
config ARCH_STRICT_ALIGN
bool "Enable -mstrict-align to prevent unaligned accesses" if EXPERT
default y
@@ -54,8 +54,10 @@ static inline void __iomem *ioremap_prot(phys_addr_t offset, unsigned long size,
* @offset: bus address of the memory
* @size: size of the resource to map
*/
+extern pgprot_t pgprot_wc;
+
#define ioremap_wc(offset, size) \
- ioremap_prot((offset), (size), pgprot_val(PAGE_KERNEL_WUC))
+ ioremap_prot((offset), (size), pgprot_val(pgprot_wc))
#define ioremap_cache(offset, size) \
ioremap_prot((offset), (size), pgprot_val(PAGE_KERNEL))
@@ -160,6 +160,27 @@ static void __init smbios_parse(void)
dmi_walk(find_tokens, NULL);
}
+#ifdef CONFIG_ARCH_WRITECOMBINE
+pgprot_t pgprot_wc = PAGE_KERNEL_WUC;
+#else
+pgprot_t pgprot_wc = PAGE_KERNEL_SUC;
+#endif
+
+EXPORT_SYMBOL(pgprot_wc);
+
+static int __init setup_writecombine(char *p)
+{
+ if (!strcmp(p, "on"))
+ pgprot_wc = PAGE_KERNEL_WUC;
+ else if (!strcmp(p, "off"))
+ pgprot_wc = PAGE_KERNEL_SUC;
+ else
+ pr_warn("Unknown writecombine setting \"%s\".\n", p);
+
+ return 0;
+}
+early_param("writecombine", setup_writecombine);
+
static int usermem __initdata;
static int __init early_parse_mem(char *p)