| Message ID | 20200928092641.2070-3-thunder.leizhen@huawei.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | ARM: decompressor: relax the loading restriction of the decompressed kernel | expand |
On Mon, 28 Sep 2020 at 11:27, Zhen Lei <thunder.leizhen@huawei.com> wrote: > > mov r4, pc > and r4, r4, #0xf8000000 //truncated to 128MiB boundary > add r4, r4, #TEXT_OFFSET //PA(_start) > > Currently, the decompressed kernel must be placed at the position: 128MiB > boundary + TEXT_OFFSET. This limitation is just because we masked PC with > 0xf80000000. Actually, we can directly obtain PA(_start) by using formula > : VA(_start) + (PHYS_OFFSET - PAGE_OFFSET). > > So the "PA(_start) - TEXT_OFFSET" can be 2MiB boundary, 1MiB boundary, > and so on. > > Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com> No, this won't work. The whole reason for rounding to a multiple of 128 MB is that we cannot infer the start of DRAM from the placement of the zImage (which provides _start). There are patches on the list by Geert [0] to obtain the start of DRAM from the device tree directly, to work around this masking. And when booting via EFI (which is supported by u-boot too these days), the zImage could be anywhere in [32-bit addressable] memory, and the start of DRAM is obtained from the EFI memory map. Note to Geert: this is a followup to an effort by Zhen Lei and myself [1] to lower the minimum alignment of PHYS_OFFSET from 16 MiB to 2 MiB, and this affects your patch as well. [0] https://lore.kernel.org/linux-arm-kernel/20200902153606.13652-1-geert+renesas@glider.be/ [1] https://lore.kernel.org/linux-arm-kernel/20200921154117.757-1-ardb@kernel.org/ > --- > arch/arm/boot/compressed/head.S | 33 ++++++++++++++------------------- > 1 file changed, 14 insertions(+), 19 deletions(-) > > diff --git a/arch/arm/boot/compressed/head.S b/arch/arm/boot/compressed/head.S > index 434a16982e344fe..e5ba2ad2ea4700f 100644 > --- a/arch/arm/boot/compressed/head.S > +++ b/arch/arm/boot/compressed/head.S > @@ -255,26 +255,16 @@ not_angel: > > #ifdef CONFIG_AUTO_ZRELADDR > /* > - * Find the start of physical memory. As we are executing > - * without the MMU on, we are in the physical address space. > - * We just need to get rid of any offset by aligning the > - * address. > - * > - * This alignment is a balance between the requirements of > - * different platforms - we have chosen 128MB to allow > - * platforms which align the start of their physical memory > - * to 128MB to use this feature, while allowing the zImage > - * to be placed within the first 128MB of memory on other > - * platforms. Increasing the alignment means we place > - * stricter alignment requirements on the start of physical > - * memory, but relaxing it means that we break people who > - * are already placing their zImage in (eg) the top 64MB > - * of this range. > + * Find ZRELADDR (Address where the decompressed kernel was > + * placed, usually == PHYS_OFFSET + TEXT_OFFSET). That's the > + * start physical address of the text section, PA(_start). > + * As we are executing without the MMU on, we are in the > + * physical address space. > */ > - mov r4, pc > - and r4, r4, #0xf8000000 > - /* Determine final kernel image address. */ > - add r4, r4, #TEXT_OFFSET > + adr r0, LC2 > + ldmia r0, {r3-r4} > + sub r3, r0, r3 @ PHYS_OFFSET - PAGE_OFFSET > + add r4, r4, r3 @ PA(_start) > #else > ldr r4, =zreladdr > #endif > @@ -660,6 +650,11 @@ LC1: .word .L_user_stack_end - LC1 @ sp > .word _edata - LC1 @ r6 > .size LC1, . - LC1 > > + .align 2 > + .type LC2, #object > +LC2: .word LC2 > + .word _start @ start VA of text section > + > .Lheadroom: > .word _end - restart + 16384 + 1024*1024 > > -- > 1.8.3 > >
On 2020/9/28 18:14, Ard Biesheuvel wrote: > On Mon, 28 Sep 2020 at 11:27, Zhen Lei <thunder.leizhen@huawei.com> wrote: >> >> mov r4, pc >> and r4, r4, #0xf8000000 //truncated to 128MiB boundary >> add r4, r4, #TEXT_OFFSET //PA(_start) >> >> Currently, the decompressed kernel must be placed at the position: 128MiB >> boundary + TEXT_OFFSET. This limitation is just because we masked PC with >> 0xf80000000. Actually, we can directly obtain PA(_start) by using formula >> : VA(_start) + (PHYS_OFFSET - PAGE_OFFSET). >> >> So the "PA(_start) - TEXT_OFFSET" can be 2MiB boundary, 1MiB boundary, >> and so on. >> >> Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com> > > No, this won't work. But it works well on my board. > > The whole reason for rounding to a multiple of 128 MB is that we > cannot infer the start of DRAM from the placement of the zImage (which > provides _start). Maybe this is further guaranteed by the following code: /* * Set up a page table only if it won't overwrite ourself. * That means r4 < pc || r4 - 16k page directory > &_end. * Given that r4 > &_end is most unfrequent, we add a rough * additional 1MB of room for a possible appended DTB. */ In addition, the Image has already been loaded when this position is reached. ----------- <--128MiB boundary | | ----------- <--TEXT_OFFSET <-- | (1)Image | | ------------ | | | | ----------- (2)--copyto----- | (2)Image | ----------- I don't think it's the case of (2), but (1). Because no code modification is required for the case (2). By the way, I'm not familiar with the arm32 code, so I'm just speculating. > > There are patches on the list by Geert [0] to obtain the start of DRAM > from the device tree directly, to work around this masking. And when > booting via EFI (which is supported by u-boot too these days), the > zImage could be anywhere in [32-bit addressable] memory, and the start > of DRAM is obtained from the EFI memory map. > > Note to Geert: this is a followup to an effort by Zhen Lei and myself > [1] to lower the minimum alignment of PHYS_OFFSET from 16 MiB to 2 > MiB, and this affects your patch as well. > > > [0] https://lore.kernel.org/linux-arm-kernel/20200902153606.13652-1-geert+renesas@glider.be/ > [1] https://lore.kernel.org/linux-arm-kernel/20200921154117.757-1-ardb@kernel.org/ > > >> --- >> arch/arm/boot/compressed/head.S | 33 ++++++++++++++------------------- >> 1 file changed, 14 insertions(+), 19 deletions(-) >> >> diff --git a/arch/arm/boot/compressed/head.S b/arch/arm/boot/compressed/head.S >> index 434a16982e344fe..e5ba2ad2ea4700f 100644 >> --- a/arch/arm/boot/compressed/head.S >> +++ b/arch/arm/boot/compressed/head.S >> @@ -255,26 +255,16 @@ not_angel: >> >> #ifdef CONFIG_AUTO_ZRELADDR >> /* >> - * Find the start of physical memory. As we are executing >> - * without the MMU on, we are in the physical address space. >> - * We just need to get rid of any offset by aligning the >> - * address. >> - * >> - * This alignment is a balance between the requirements of >> - * different platforms - we have chosen 128MB to allow >> - * platforms which align the start of their physical memory >> - * to 128MB to use this feature, while allowing the zImage >> - * to be placed within the first 128MB of memory on other >> - * platforms. Increasing the alignment means we place >> - * stricter alignment requirements on the start of physical >> - * memory, but relaxing it means that we break people who >> - * are already placing their zImage in (eg) the top 64MB >> - * of this range. >> + * Find ZRELADDR (Address where the decompressed kernel was >> + * placed, usually == PHYS_OFFSET + TEXT_OFFSET). That's the >> + * start physical address of the text section, PA(_start). >> + * As we are executing without the MMU on, we are in the >> + * physical address space. >> */ >> - mov r4, pc >> - and r4, r4, #0xf8000000 >> - /* Determine final kernel image address. */ >> - add r4, r4, #TEXT_OFFSET >> + adr r0, LC2 >> + ldmia r0, {r3-r4} >> + sub r3, r0, r3 @ PHYS_OFFSET - PAGE_OFFSET >> + add r4, r4, r3 @ PA(_start) >> #else >> ldr r4, =zreladdr >> #endif >> @@ -660,6 +650,11 @@ LC1: .word .L_user_stack_end - LC1 @ sp >> .word _edata - LC1 @ r6 >> .size LC1, . - LC1 >> >> + .align 2 >> + .type LC2, #object >> +LC2: .word LC2 >> + .word _start @ start VA of text section >> + >> .Lheadroom: >> .word _end - restart + 16384 + 1024*1024 >> >> -- >> 1.8.3 >> >> > > . >
On Mon, 28 Sep 2020 at 13:57, Leizhen (ThunderTown) <thunder.leizhen@huawei.com> wrote: > > > > On 2020/9/28 18:14, Ard Biesheuvel wrote: > > On Mon, 28 Sep 2020 at 11:27, Zhen Lei <thunder.leizhen@huawei.com> wrote: > >> > >> mov r4, pc > >> and r4, r4, #0xf8000000 //truncated to 128MiB boundary > >> add r4, r4, #TEXT_OFFSET //PA(_start) > >> > >> Currently, the decompressed kernel must be placed at the position: 128MiB > >> boundary + TEXT_OFFSET. This limitation is just because we masked PC with > >> 0xf80000000. Actually, we can directly obtain PA(_start) by using formula > >> : VA(_start) + (PHYS_OFFSET - PAGE_OFFSET). > >> > >> So the "PA(_start) - TEXT_OFFSET" can be 2MiB boundary, 1MiB boundary, > >> and so on. > >> > >> Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com> > > > > No, this won't work. > > But it works well on my board. > That is because you load zImage at the base of DRAM. > > > > The whole reason for rounding to a multiple of 128 MB is that we > > cannot infer the start of DRAM from the placement of the zImage (which > > provides _start). > > Maybe this is further guaranteed by the following code: > /* > * Set up a page table only if it won't overwrite ourself. > * That means r4 < pc || r4 - 16k page directory > &_end. > * Given that r4 > &_end is most unfrequent, we add a rough > * additional 1MB of room for a possible appended DTB. > */ > > In addition, the Image has already been loaded when this position is reached. > > ----------- <--128MiB boundary > | | > ----------- <--TEXT_OFFSET <-- > | (1)Image | | > ------------ | > | | | > ----------- (2)--copyto----- > | (2)Image | > ----------- > > I don't think it's the case of (2), but (1). Because no code modification is > required for the case (2). > > By the way, I'm not familiar with the arm32 code, so I'm just speculating. > The zImage code that runs has not received *any* information from the platform on where DRAM starts, so the only info it has is the current placement of zImage. So when zImage is loaded at the intended base of DRAM, things work fine. If the zImage is loaded close to the end of a 128 MB region, the rounding would pick the start of that 128 MB region. However, the _start symbol you are using will point to an address that is close to the end of the 128 MB [given that it is inside zImage] so your logic will pick an address that is much higher in memory.
Hi Zhen, On Mon, Sep 28, 2020 at 2:15 PM Ard Biesheuvel <ardb@kernel.org> wrote: > On Mon, 28 Sep 2020 at 13:57, Leizhen (ThunderTown) > <thunder.leizhen@huawei.com> wrote: > > On 2020/9/28 18:14, Ard Biesheuvel wrote: > > > On Mon, 28 Sep 2020 at 11:27, Zhen Lei <thunder.leizhen@huawei.com> wrote: > > >> > > >> mov r4, pc > > >> and r4, r4, #0xf8000000 //truncated to 128MiB boundary > > >> add r4, r4, #TEXT_OFFSET //PA(_start) > > >> > > >> Currently, the decompressed kernel must be placed at the position: 128MiB > > >> boundary + TEXT_OFFSET. This limitation is just because we masked PC with > > >> 0xf80000000. Actually, we can directly obtain PA(_start) by using formula > > >> : VA(_start) + (PHYS_OFFSET - PAGE_OFFSET). > > >> > > >> So the "PA(_start) - TEXT_OFFSET" can be 2MiB boundary, 1MiB boundary, > > >> and so on. > > >> > > >> Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com> > > > > > > No, this won't work. > > > > But it works well on my board. > > > > That is because you load zImage at the base of DRAM. > > > > > > > The whole reason for rounding to a multiple of 128 MB is that we > > > cannot infer the start of DRAM from the placement of the zImage (which > > > provides _start). > > > > Maybe this is further guaranteed by the following code: > > /* > > * Set up a page table only if it won't overwrite ourself. > > * That means r4 < pc || r4 - 16k page directory > &_end. > > * Given that r4 > &_end is most unfrequent, we add a rough > > * additional 1MB of room for a possible appended DTB. > > */ > > > > In addition, the Image has already been loaded when this position is reached. > > > > ----------- <--128MiB boundary > > | | > > ----------- <--TEXT_OFFSET <-- > > | (1)Image | | > > ------------ | > > | | | > > ----------- (2)--copyto----- > > | (2)Image | > > ----------- > > > > I don't think it's the case of (2), but (1). Because no code modification is > > required for the case (2). > > > > By the way, I'm not familiar with the arm32 code, so I'm just speculating. > > > > The zImage code that runs has not received *any* information from the > platform on where DRAM starts, so the only info it has is the current > placement of zImage. > > So when zImage is loaded at the intended base of DRAM, things work fine. > > If the zImage is loaded close to the end of a 128 MB region, the > rounding would pick the start of that 128 MB region. However, the > _start symbol you are using will point to an address that is close to > the end of the 128 MB [given that it is inside zImage] so your logic > will pick an address that is much higher in memory. https://people.kernel.org/linusw/how-the-arm32-linux-kernel-decompresses https://people.kernel.org/linusw/how-the-arm32-kernel-starts are good reads. Gr{oetje,eeting}s, Geert -- Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org In personal conversations with technical people, I call myself a hacker. But when I'm talking to journalists I just say "programmer" or something like that. -- Linus Torvalds
On 2020/9/28 20:57, Geert Uytterhoeven wrote: > Hi Zhen, > > On Mon, Sep 28, 2020 at 2:15 PM Ard Biesheuvel <ardb@kernel.org> wrote: >> On Mon, 28 Sep 2020 at 13:57, Leizhen (ThunderTown) >> <thunder.leizhen@huawei.com> wrote: >>> On 2020/9/28 18:14, Ard Biesheuvel wrote: >>>> On Mon, 28 Sep 2020 at 11:27, Zhen Lei <thunder.leizhen@huawei.com> wrote: >>>>> >>>>> mov r4, pc >>>>> and r4, r4, #0xf8000000 //truncated to 128MiB boundary >>>>> add r4, r4, #TEXT_OFFSET //PA(_start) >>>>> >>>>> Currently, the decompressed kernel must be placed at the position: 128MiB >>>>> boundary + TEXT_OFFSET. This limitation is just because we masked PC with >>>>> 0xf80000000. Actually, we can directly obtain PA(_start) by using formula >>>>> : VA(_start) + (PHYS_OFFSET - PAGE_OFFSET). >>>>> >>>>> So the "PA(_start) - TEXT_OFFSET" can be 2MiB boundary, 1MiB boundary, >>>>> and so on. >>>>> >>>>> Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com> >>>> >>>> No, this won't work. >>> >>> But it works well on my board. >>> >> >> That is because you load zImage at the base of DRAM. >> >>>> >>>> The whole reason for rounding to a multiple of 128 MB is that we >>>> cannot infer the start of DRAM from the placement of the zImage (which >>>> provides _start). >>> >>> Maybe this is further guaranteed by the following code: >>> /* >>> * Set up a page table only if it won't overwrite ourself. >>> * That means r4 < pc || r4 - 16k page directory > &_end. >>> * Given that r4 > &_end is most unfrequent, we add a rough >>> * additional 1MB of room for a possible appended DTB. >>> */ >>> >>> In addition, the Image has already been loaded when this position is reached. >>> >>> ----------- <--128MiB boundary >>> | | >>> ----------- <--TEXT_OFFSET <-- >>> | (1)Image | | >>> ------------ | >>> | | | >>> ----------- (2)--copyto----- >>> | (2)Image | >>> ----------- >>> >>> I don't think it's the case of (2), but (1). Because no code modification is >>> required for the case (2). >>> >>> By the way, I'm not familiar with the arm32 code, so I'm just speculating. >>> >> >> The zImage code that runs has not received *any* information from the >> platform on where DRAM starts, so the only info it has is the current >> placement of zImage. >> >> So when zImage is loaded at the intended base of DRAM, things work fine. >> >> If the zImage is loaded close to the end of a 128 MB region, the >> rounding would pick the start of that 128 MB region. However, the >> _start symbol you are using will point to an address that is close to >> the end of the 128 MB [given that it is inside zImage] so your logic >> will pick an address that is much higher in memory. > > https://people.kernel.org/linusw/how-the-arm32-linux-kernel-decompresses > https://people.kernel.org/linusw/how-the-arm32-kernel-starts > are good reads. Thanks for your information. > > Gr{oetje,eeting}s, > > Geert > > > -- > Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org > > In personal conversations with technical people, I call myself a hacker. But > when I'm talking to journalists I just say "programmer" or something like that. > -- Linus Torvalds > > . >
diff --git a/arch/arm/boot/compressed/head.S b/arch/arm/boot/compressed/head.S index 434a16982e344fe..e5ba2ad2ea4700f 100644 --- a/arch/arm/boot/compressed/head.S +++ b/arch/arm/boot/compressed/head.S @@ -255,26 +255,16 @@ not_angel: #ifdef CONFIG_AUTO_ZRELADDR /* - * Find the start of physical memory. As we are executing - * without the MMU on, we are in the physical address space. - * We just need to get rid of any offset by aligning the - * address. - * - * This alignment is a balance between the requirements of - * different platforms - we have chosen 128MB to allow - * platforms which align the start of their physical memory - * to 128MB to use this feature, while allowing the zImage - * to be placed within the first 128MB of memory on other - * platforms. Increasing the alignment means we place - * stricter alignment requirements on the start of physical - * memory, but relaxing it means that we break people who - * are already placing their zImage in (eg) the top 64MB - * of this range. + * Find ZRELADDR (Address where the decompressed kernel was + * placed, usually == PHYS_OFFSET + TEXT_OFFSET). That's the + * start physical address of the text section, PA(_start). + * As we are executing without the MMU on, we are in the + * physical address space. */ - mov r4, pc - and r4, r4, #0xf8000000 - /* Determine final kernel image address. */ - add r4, r4, #TEXT_OFFSET + adr r0, LC2 + ldmia r0, {r3-r4} + sub r3, r0, r3 @ PHYS_OFFSET - PAGE_OFFSET + add r4, r4, r3 @ PA(_start) #else ldr r4, =zreladdr #endif @@ -660,6 +650,11 @@ LC1: .word .L_user_stack_end - LC1 @ sp .word _edata - LC1 @ r6 .size LC1, . - LC1 + .align 2 + .type LC2, #object +LC2: .word LC2 + .word _start @ start VA of text section + .Lheadroom: .word _end - restart + 16384 + 1024*1024
mov r4, pc and r4, r4, #0xf8000000 //truncated to 128MiB boundary add r4, r4, #TEXT_OFFSET //PA(_start) Currently, the decompressed kernel must be placed at the position: 128MiB boundary + TEXT_OFFSET. This limitation is just because we masked PC with 0xf80000000. Actually, we can directly obtain PA(_start) by using formula : VA(_start) + (PHYS_OFFSET - PAGE_OFFSET). So the "PA(_start) - TEXT_OFFSET" can be 2MiB boundary, 1MiB boundary, and so on. Signed-off-by: Zhen Lei <thunder.leizhen@huawei.com> --- arch/arm/boot/compressed/head.S | 33 ++++++++++++++------------------- 1 file changed, 14 insertions(+), 19 deletions(-)