Re: [PATCH v3 2/2] ARM: support PHYS_OFFSET minimum aligned at 64KiB boundary
On Thu, 17 Sep 2020 at 05:16, Leizhen (ThunderTown) wrote: > > > > On 2020/9/16 19:15, Ard Biesheuvel wrote: > > (+ Arnd, Nico) > > > > On Wed, 16 Sep 2020 at 05:51, Zhen Lei wrote: > >> > >> Currently, only support the kernels where the base of physical memory is > >> at a 16MiB boundary. Because the add/sub instructions only contains 8bits > >> unrotated value. But we can use one more "add/sub" instructions to handle > >> bits 23-16. The performance will be slightly affected. > >> > >> Since most boards meet 16 MiB alignment, so add a new configuration > >> option ARM_PATCH_PHYS_VIRT_RADICAL (default n) to control it. Say Y if > >> anyone really needs it. > >> > >> All r0-r7 (r1 = machine no, r2 = atags or dtb, in the start-up phase) are > >> used in __fixup_a_pv_table() now, but the callee saved r11 is not used in > >> the whole head.S file. So choose it. > >> > >> Because the calculation of "y = x + __pv_offset[63:24]" have been done, > >> so we only need to calculate "y = y + __pv_offset[23:16]", that's why > >> the parameters "to" and "from" of __pv_stub() and __pv_add_carry_stub() > >> in the scope of CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL are all passed "t" > >> (above y). > >> > >> Signed-off-by: Zhen Lei > >> --- > >> arch/arm/Kconfig | 17 - > >> arch/arm/include/asm/memory.h | 16 +--- > >> arch/arm/kernel/head.S| 25 +++-- > >> 3 files changed, 48 insertions(+), 10 deletions(-) > >> > >> diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig > >> index e00d94b16658765..073dafa428f3c87 100644 > >> --- a/arch/arm/Kconfig > >> +++ b/arch/arm/Kconfig > >> @@ -240,12 +240,27 @@ config ARM_PATCH_PHYS_VIRT > >> kernel in system memory. > >> > >> This can only be used with non-XIP MMU kernels where the base > >> - of physical memory is at a 16MB boundary. > >> + of physical memory is at a 16MiB boundary. > >> > >> Only disable this option if you know that you do not require > >> this feature (eg, building a kernel for a single machine) and > >> you need to shrink the kernel to the minimal size. > >> > >> +config ARM_PATCH_PHYS_VIRT_RADICAL > >> + bool "Support PHYS_OFFSET minimum aligned at 64KiB boundary" > >> + depends on ARM_PATCH_PHYS_VIRT > >> + depends on !THUMB2_KERNEL > > > > Why is this not implemented for Thumb2 too? > > No Thumb2 boards. > Are you saying none of the boards you tested this on have support for the Thumb2 instruction set?? That seems unlikely to me, given that all ARM cores since Cortex-A8 can run this code, but even if that were true, you can easily test this on QEMU as well, which can definitely run the Thumb2 kernel. > > > > Also, as Russell points out as well, this may end up being enabled for > > all multiarch kernels, so it makes sense to explore whether we can > > enable this unconditionally. > > Yes, In fact, I think we can consider enabling this unconditionally after > the THUMB2 branch is implemented. Performance and code size should not be > a problem. > Excellent. > > Do you have any numbers wrt the impact on > > text size? I would assume it is negligible, but numbers help. > > The text size increased a bit more than 2 KB (2164 Bytes), about 0.0146%. > > make ARCH=arm CROSS_COMPILE=arm-linux-gnueabihf- distclean defconfig > > Before: > $ size vmlinux >textdata bss dec hex filename > 14781964 7508366 420080 2271041015a888a vmlinux > > After: > $ size vmlinux >textdata bss dec hex filename > 14784128 7508366 420080 2271257415a90fe vmlinux > > OK, so negligible in terms of code size. > > > > Being able to decompress the image to any 2MiB aligned base address is > > also quite useful for EFI boot, and it may also help to get rid of the > > TEXT_OFFSET hacks we have for some platforms in the future.> > > > >> + help > >> + This can only be used with non-XIP MMU kernels where the base > >> + of physical memory is at a 64KiB boundary. > >> + > >> + Compared with ARM_PATCH_PHYS_VIRT, one or two more instructions > >> + need to be added to implement the conversion of bits 23-16 of > >> + the VA/PA in phys-to-virt and virt-to-phys. The performance is > >> + slightly affected. > >> + > > > > Does it affect performance in other ways beyond code size/Icache density? > > I just want to say it will slightly slower than !ARM_PATCH_PHYS_VIRT_RADICAL, > because one or two more instructions. It certainly cannot affect system > performance. > > Because of your doubts, I think I should remove the statement: "The > performance is > slightly affected." > Yeah, I think you can drop that statement. > > > >> + If unsure say N here. > >> + > >> config NEED_MACH_IO_H > >> bool > >> help > >> diff --git a/arch/arm/include/asm/memory.h b/arch/arm/include/asm/memory.h > >> index
Re: [PATCH v3 2/2] ARM: support PHYS_OFFSET minimum aligned at 64KiB boundary
On 2020/9/16 19:15, Ard Biesheuvel wrote: > (+ Arnd, Nico) > > On Wed, 16 Sep 2020 at 05:51, Zhen Lei wrote: >> >> Currently, only support the kernels where the base of physical memory is >> at a 16MiB boundary. Because the add/sub instructions only contains 8bits >> unrotated value. But we can use one more "add/sub" instructions to handle >> bits 23-16. The performance will be slightly affected. >> >> Since most boards meet 16 MiB alignment, so add a new configuration >> option ARM_PATCH_PHYS_VIRT_RADICAL (default n) to control it. Say Y if >> anyone really needs it. >> >> All r0-r7 (r1 = machine no, r2 = atags or dtb, in the start-up phase) are >> used in __fixup_a_pv_table() now, but the callee saved r11 is not used in >> the whole head.S file. So choose it. >> >> Because the calculation of "y = x + __pv_offset[63:24]" have been done, >> so we only need to calculate "y = y + __pv_offset[23:16]", that's why >> the parameters "to" and "from" of __pv_stub() and __pv_add_carry_stub() >> in the scope of CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL are all passed "t" >> (above y). >> >> Signed-off-by: Zhen Lei >> --- >> arch/arm/Kconfig | 17 - >> arch/arm/include/asm/memory.h | 16 +--- >> arch/arm/kernel/head.S| 25 +++-- >> 3 files changed, 48 insertions(+), 10 deletions(-) >> >> diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig >> index e00d94b16658765..073dafa428f3c87 100644 >> --- a/arch/arm/Kconfig >> +++ b/arch/arm/Kconfig >> @@ -240,12 +240,27 @@ config ARM_PATCH_PHYS_VIRT >> kernel in system memory. >> >> This can only be used with non-XIP MMU kernels where the base >> - of physical memory is at a 16MB boundary. >> + of physical memory is at a 16MiB boundary. >> >> Only disable this option if you know that you do not require >> this feature (eg, building a kernel for a single machine) and >> you need to shrink the kernel to the minimal size. >> >> +config ARM_PATCH_PHYS_VIRT_RADICAL >> + bool "Support PHYS_OFFSET minimum aligned at 64KiB boundary" >> + depends on ARM_PATCH_PHYS_VIRT >> + depends on !THUMB2_KERNEL > > Why is this not implemented for Thumb2 too? No Thumb2 boards. > > Also, as Russell points out as well, this may end up being enabled for > all multiarch kernels, so it makes sense to explore whether we can > enable this unconditionally. Yes, In fact, I think we can consider enabling this unconditionally after the THUMB2 branch is implemented. Performance and code size should not be a problem. > Do you have any numbers wrt the impact on > text size? I would assume it is negligible, but numbers help. The text size increased a bit more than 2 KB (2164 Bytes), about 0.0146%. make ARCH=arm CROSS_COMPILE=arm-linux-gnueabihf- distclean defconfig Before: $ size vmlinux textdata bss dec hex filename 14781964 7508366 420080 2271041015a888a vmlinux After: $ size vmlinux textdata bss dec hex filename 14784128 7508366 420080 2271257415a90fe vmlinux > > Being able to decompress the image to any 2MiB aligned base address is > also quite useful for EFI boot, and it may also help to get rid of the > TEXT_OFFSET hacks we have for some platforms in the future.> > >> + help >> + This can only be used with non-XIP MMU kernels where the base >> + of physical memory is at a 64KiB boundary. >> + >> + Compared with ARM_PATCH_PHYS_VIRT, one or two more instructions >> + need to be added to implement the conversion of bits 23-16 of >> + the VA/PA in phys-to-virt and virt-to-phys. The performance is >> + slightly affected. >> + > > Does it affect performance in other ways beyond code size/Icache density? I just want to say it will slightly slower than !ARM_PATCH_PHYS_VIRT_RADICAL, because one or two more instructions. It certainly cannot affect system performance. Because of your doubts, I think I should remove the statement: "The performance is slightly affected." > >> + If unsure say N here. >> + >> config NEED_MACH_IO_H >> bool >> help >> diff --git a/arch/arm/include/asm/memory.h b/arch/arm/include/asm/memory.h >> index 99035b5891ef442..f97b37303a00f60 100644 >> --- a/arch/arm/include/asm/memory.h >> +++ b/arch/arm/include/asm/memory.h >> @@ -173,6 +173,7 @@ >> * so that all we need to do is modify the 8-bit constant field. >> */ >> #define __PV_BITS_31_240x8100 >> +#define __PV_BITS_23_160x0081 >> #define __PV_BITS_7_0 0x81 >> >> extern unsigned long __pv_phys_pfn_offset; >> @@ -201,7 +202,7 @@ >> : "=r" (t) \ >> : "I" (__PV_BITS_7_0)) >> >> -#define __pv_add_carry_stub(x, y) \ >> +#define __pv_add_carry_stub(x, y, type)\ >> __asm__ volatile("@ __pv_add_carry_stub\n"
Re: [PATCH v3 2/2] ARM: support PHYS_OFFSET minimum aligned at 64KiB boundary
(+ Arnd, Nico) On Wed, 16 Sep 2020 at 05:51, Zhen Lei wrote: > > Currently, only support the kernels where the base of physical memory is > at a 16MiB boundary. Because the add/sub instructions only contains 8bits > unrotated value. But we can use one more "add/sub" instructions to handle > bits 23-16. The performance will be slightly affected. > > Since most boards meet 16 MiB alignment, so add a new configuration > option ARM_PATCH_PHYS_VIRT_RADICAL (default n) to control it. Say Y if > anyone really needs it. > > All r0-r7 (r1 = machine no, r2 = atags or dtb, in the start-up phase) are > used in __fixup_a_pv_table() now, but the callee saved r11 is not used in > the whole head.S file. So choose it. > > Because the calculation of "y = x + __pv_offset[63:24]" have been done, > so we only need to calculate "y = y + __pv_offset[23:16]", that's why > the parameters "to" and "from" of __pv_stub() and __pv_add_carry_stub() > in the scope of CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL are all passed "t" > (above y). > > Signed-off-by: Zhen Lei > --- > arch/arm/Kconfig | 17 - > arch/arm/include/asm/memory.h | 16 +--- > arch/arm/kernel/head.S| 25 +++-- > 3 files changed, 48 insertions(+), 10 deletions(-) > > diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig > index e00d94b16658765..073dafa428f3c87 100644 > --- a/arch/arm/Kconfig > +++ b/arch/arm/Kconfig > @@ -240,12 +240,27 @@ config ARM_PATCH_PHYS_VIRT > kernel in system memory. > > This can only be used with non-XIP MMU kernels where the base > - of physical memory is at a 16MB boundary. > + of physical memory is at a 16MiB boundary. > > Only disable this option if you know that you do not require > this feature (eg, building a kernel for a single machine) and > you need to shrink the kernel to the minimal size. > > +config ARM_PATCH_PHYS_VIRT_RADICAL > + bool "Support PHYS_OFFSET minimum aligned at 64KiB boundary" > + depends on ARM_PATCH_PHYS_VIRT > + depends on !THUMB2_KERNEL Why is this not implemented for Thumb2 too? Also, as Russell points out as well, this may end up being enabled for all multiarch kernels, so it makes sense to explore whether we can enable this unconditionally. Do you have any numbers wrt the impact on text size? I would assume it is negligible, but numbers help. Being able to decompress the image to any 2MiB aligned base address is also quite useful for EFI boot, and it may also help to get rid of the TEXT_OFFSET hacks we have for some platforms in the future. > + help > + This can only be used with non-XIP MMU kernels where the base > + of physical memory is at a 64KiB boundary. > + > + Compared with ARM_PATCH_PHYS_VIRT, one or two more instructions > + need to be added to implement the conversion of bits 23-16 of > + the VA/PA in phys-to-virt and virt-to-phys. The performance is > + slightly affected. > + Does it affect performance in other ways beyond code size/Icache density? > + If unsure say N here. > + > config NEED_MACH_IO_H > bool > help > diff --git a/arch/arm/include/asm/memory.h b/arch/arm/include/asm/memory.h > index 99035b5891ef442..f97b37303a00f60 100644 > --- a/arch/arm/include/asm/memory.h > +++ b/arch/arm/include/asm/memory.h > @@ -173,6 +173,7 @@ > * so that all we need to do is modify the 8-bit constant field. > */ > #define __PV_BITS_31_240x8100 > +#define __PV_BITS_23_160x0081 > #define __PV_BITS_7_0 0x81 > > extern unsigned long __pv_phys_pfn_offset; > @@ -201,7 +202,7 @@ > : "=r" (t) \ > : "I" (__PV_BITS_7_0)) > > -#define __pv_add_carry_stub(x, y) \ > +#define __pv_add_carry_stub(x, y, type)\ > __asm__ volatile("@ __pv_add_carry_stub\n" \ > "1: adds%Q0, %1, %2\n" \ > " adc %R0, %R0, #0\n" \ > @@ -209,7 +210,7 @@ > " .long 1b\n" \ > " .popsection\n" \ > : "+r" (y) \ > - : "r" (x), "I" (__PV_BITS_31_24)\ > + : "r" (x), "I" (type) \ > : "cc") > > static inline phys_addr_t __virt_to_phys_nodebug(unsigned long x) > @@ -218,9 +219,15 @@ static inline phys_addr_t > __virt_to_phys_nodebug(unsigned long x) > > if (sizeof(phys_addr_t) == 4) { > __pv_stub(x, t, "add", __PV_BITS_31_24); > +#ifdef CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL > + __pv_stub(t, t, "add", __PV_BITS_23_16); > +#endif > } else { > __pv_stub_mov_hi(t); > - __pv_add_carry_stub(x, t); > + __pv_add_carry_stub(x, t, __PV_BITS_31_24); > +#ifdef
[PATCH v3 2/2] ARM: support PHYS_OFFSET minimum aligned at 64KiB boundary
Currently, only support the kernels where the base of physical memory is at a 16MiB boundary. Because the add/sub instructions only contains 8bits unrotated value. But we can use one more "add/sub" instructions to handle bits 23-16. The performance will be slightly affected. Since most boards meet 16 MiB alignment, so add a new configuration option ARM_PATCH_PHYS_VIRT_RADICAL (default n) to control it. Say Y if anyone really needs it. All r0-r7 (r1 = machine no, r2 = atags or dtb, in the start-up phase) are used in __fixup_a_pv_table() now, but the callee saved r11 is not used in the whole head.S file. So choose it. Because the calculation of "y = x + __pv_offset[63:24]" have been done, so we only need to calculate "y = y + __pv_offset[23:16]", that's why the parameters "to" and "from" of __pv_stub() and __pv_add_carry_stub() in the scope of CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL are all passed "t" (above y). Signed-off-by: Zhen Lei --- arch/arm/Kconfig | 17 - arch/arm/include/asm/memory.h | 16 +--- arch/arm/kernel/head.S| 25 +++-- 3 files changed, 48 insertions(+), 10 deletions(-) diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig index e00d94b16658765..073dafa428f3c87 100644 --- a/arch/arm/Kconfig +++ b/arch/arm/Kconfig @@ -240,12 +240,27 @@ config ARM_PATCH_PHYS_VIRT kernel in system memory. This can only be used with non-XIP MMU kernels where the base - of physical memory is at a 16MB boundary. + of physical memory is at a 16MiB boundary. Only disable this option if you know that you do not require this feature (eg, building a kernel for a single machine) and you need to shrink the kernel to the minimal size. +config ARM_PATCH_PHYS_VIRT_RADICAL + bool "Support PHYS_OFFSET minimum aligned at 64KiB boundary" + depends on ARM_PATCH_PHYS_VIRT + depends on !THUMB2_KERNEL + help + This can only be used with non-XIP MMU kernels where the base + of physical memory is at a 64KiB boundary. + + Compared with ARM_PATCH_PHYS_VIRT, one or two more instructions + need to be added to implement the conversion of bits 23-16 of + the VA/PA in phys-to-virt and virt-to-phys. The performance is + slightly affected. + + If unsure say N here. + config NEED_MACH_IO_H bool help diff --git a/arch/arm/include/asm/memory.h b/arch/arm/include/asm/memory.h index 99035b5891ef442..f97b37303a00f60 100644 --- a/arch/arm/include/asm/memory.h +++ b/arch/arm/include/asm/memory.h @@ -173,6 +173,7 @@ * so that all we need to do is modify the 8-bit constant field. */ #define __PV_BITS_31_240x8100 +#define __PV_BITS_23_160x0081 #define __PV_BITS_7_0 0x81 extern unsigned long __pv_phys_pfn_offset; @@ -201,7 +202,7 @@ : "=r" (t) \ : "I" (__PV_BITS_7_0)) -#define __pv_add_carry_stub(x, y) \ +#define __pv_add_carry_stub(x, y, type)\ __asm__ volatile("@ __pv_add_carry_stub\n" \ "1: adds%Q0, %1, %2\n" \ " adc %R0, %R0, #0\n" \ @@ -209,7 +210,7 @@ " .long 1b\n" \ " .popsection\n" \ : "+r" (y) \ - : "r" (x), "I" (__PV_BITS_31_24)\ + : "r" (x), "I" (type) \ : "cc") static inline phys_addr_t __virt_to_phys_nodebug(unsigned long x) @@ -218,9 +219,15 @@ static inline phys_addr_t __virt_to_phys_nodebug(unsigned long x) if (sizeof(phys_addr_t) == 4) { __pv_stub(x, t, "add", __PV_BITS_31_24); +#ifdef CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL + __pv_stub(t, t, "add", __PV_BITS_23_16); +#endif } else { __pv_stub_mov_hi(t); - __pv_add_carry_stub(x, t); + __pv_add_carry_stub(x, t, __PV_BITS_31_24); +#ifdef CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL + __pv_add_carry_stub(t, t, __PV_BITS_23_16); +#endif } return t; } @@ -236,6 +243,9 @@ static inline unsigned long __phys_to_virt(phys_addr_t x) * in place where 'r' 32 bit operand is expected. */ __pv_stub((unsigned long) x, t, "sub", __PV_BITS_31_24); +#ifdef CONFIG_ARM_PATCH_PHYS_VIRT_RADICAL + __pv_stub(t, t, "sub", __PV_BITS_23_16); +#endif return t; } diff --git a/arch/arm/kernel/head.S b/arch/arm/kernel/head.S index 02d78c9198d0e8d..d9fb226a24d43ae 100644 --- a/arch/arm/kernel/head.S +++ b/arch/arm/kernel/head.S @@ -120,7 +120,7 @@ ENTRY(stext) bl __fixup_smp #endif #ifdef CONFIG_ARM_PATCH_PHYS_VIRT - bl __fixup_pv_table + bl __fixup_pv_table@r11 will be used #endif