| Message ID | 20240807134819.8987-3-alejandro.vallejo@cloud.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | Improve support for EFI multiboot loading | expand |
On 07.08.2024 15:48, Alejandro Vallejo wrote: > If code is loaded by EFI the loader will relocate the image > under 4GB. This is the MB2 EFI path you're talking about? Since there are two paths, I think this needs clearly separating in all descriptions. If it is the MB2 path, then "relocate" isn't quite correct, I think: Relocations aren't applied in that case, as none are present in xen.gz. I'd rather call this "put at an address below 4G". However, that isn't any different from the non-EFI MB1/2 paths, is it? I feel like I'm missing something here. > This cause offsets in x86 code generated by > sym_offs(SYMBOL) to be relocated too (basically they won't be > offsets from image base). In order to get real offset the > formulae "sym_offs(SYMBOL) - sym_offs(__image_base__)" is > used instead. The main calculations of %esi are, if I'm not mistaken, /* Store Xen image load base address in place accessible for 32-bit code. */ lea __image_base__(%rip),%esi and /* Calculate the load base address. */ call 1f 1: pop %esi sub $sym_offs(1b), %esi i.e. both deliberately %rip-relative to be position-independent. What's wrong with this? There are many more uses of sym_esi(). Why is it only this single one which poses a problem? > Also, in some case %esi register (that should point to > __image_base__ addresss) is not set so compute in all cases. Which "some case" is this? > Code tested forcing failures in the code. > > Signed-off-by: Frediano Ziglio <frediano.ziglio@cloud.com> No Fixes: tag? Jan
(re-adding xen-devel@) On 08.08.2024 10:33, Frediano Ziglio wrote: > On Thu, Aug 8, 2024 at 8:49 AM Jan Beulich <jbeulich@suse.com> wrote: >> >> On 07.08.2024 15:48, Alejandro Vallejo wrote: >>> If code is loaded by EFI the loader will relocate the image >>> under 4GB. >> >> This is the MB2 EFI path you're talking about? Since there are two paths, >> I think this needs clearly separating in all descriptions. >> >> If it is the MB2 path, then "relocate" isn't quite correct, I think: >> Relocations aren't applied in that case, as none are present in xen.gz. >> I'd rather call this "put at an address below 4G". However, that isn't >> any different from the non-EFI MB1/2 paths, is it? I feel like I'm >> missing something here. >> > > Yes, xen.gz has no relocation, but xen.efi has them, Of course, I know. I was the one to actually add them, after all. But: Are you (silently) adding a 4th way of booting Xen, using xen.efi yet not its PE-header-specified entry point? So far we had three ways only: Non-EFI (xen.gz:start), native EFI (xen.efi:efi_start), and GrUB2+EFI (xen.gz:__efi64_mb2_start). You effectively suggest GrUB2+EFI (xen.efi:__efi64_mb2_start), if I'm not mistaken. Where is such a mode even specified? When using the MB2 entry point, it's not obvious at all whether the boot loader is even supposed to be respecting the .reloc section of the PE binary (imo that's contrary to the original idea of multiboot). > Normally probably you are using xen.gz instead of xen.efi however we > are adding multiboot with PE support and secure boot so we need to use > PE format for signing. Note how here is the first time that you actually mention you're adding a new boot mode. That's quite relevant as context for the entire series, I would say. >>> This cause offsets in x86 code generated by >>> sym_offs(SYMBOL) to be relocated too (basically they won't be >>> offsets from image base). In order to get real offset the >>> formulae "sym_offs(SYMBOL) - sym_offs(__image_base__)" is >>> used instead. >> >> The main calculations of %esi are, if I'm not mistaken, >> >> /* Store Xen image load base address in place accessible for 32-bit code. */ >> lea __image_base__(%rip),%esi >> > > Which is correct > >> and >> >> /* Calculate the load base address. */ >> call 1f >> 1: pop %esi >> sub $sym_offs(1b), %esi >> >> i.e. both deliberately %rip-relative to be position-independent. What's >> wrong with this? >> > > This can be wrong if sym_offs(1b) was relocated and not patched by > efi_arch_relocate_image. Of course, if in the course of GrUB's loading of xen.efi base relocations are applied (unlike when loading an ELF binary, where afaik base relocs would be ignored, even if there were any), then this calculation is of course going to be wrong. Can't we correct it though, to properly resemble PIC code: /* Calculate the load base address. */ call 1f 1: pop %esi sub 1b - start, %esi or (because start is in a different section): /* Calculate the load base address. */ call 1f 1: pop %esi sub $sym_offs(1b), %esi add $sym_offs(start), %esi (or something along these lines)? > Consider .Lnot_multiboot label call, here we didn't set %esi and in my > case (GRUB using PE) the EFI version was used. How does %esi matter at .Lnot_multiboot? It starts mattering from x86_32_switch onwards, and it is being set there. >> There are many more uses of sym_esi(). Why is it only this single one >> which poses a problem? > > Because potentially %esi is not set (see above example, but better to > make sure is always valid) and becase potentially for EFI code is > relocated. > IMO stripping relocation and calling C code (like efi_multiboot2) > supposed to be execute at a different location is a bomb waiting to > explode. Yet that has been working fine for xen.gz? There definitely is a risk, but that risk isn't new afaict. >>> Also, in some case %esi register (that should point to >>> __image_base__ addresss) is not set so compute in all cases. >> >> Which "some case" is this? > > In __efi64_mb2_start path there are some. Note that we use %rsi for > different purposes there. All cases there where %esi would matter branch to x86_32_switch first afaics. As mentioned, %esi is being set one of the first things after the x86_32_switch label. >>> Code tested forcing failures in the code. >>> >>> Signed-off-by: Frediano Ziglio <frediano.ziglio@cloud.com> >> >> No Fixes: tag? > > Not sure here, should I open some ticker and refer to it? Well, now that I learned that you're trying to add a new boot mode, there's no need for a Fixes: tag. Yet the patch title then also shouldn't say "Fix" - that made it look like you're addressing some problem with one of the existing boot modes we have. Jan
On 08.08.2024 14:50, Frediano Ziglio wrote: > On Thu, Aug 8, 2024 at 10:29 AM Jan Beulich <jbeulich@suse.com> wrote: >> >> (re-adding xen-devel@) Did you notice this in my earlier reply? You dropped the list again. >> On 08.08.2024 10:33, Frediano Ziglio wrote: >>> On Thu, Aug 8, 2024 at 8:49 AM Jan Beulich <jbeulich@suse.com> wrote: >>>>> This cause offsets in x86 code generated by >>>>> sym_offs(SYMBOL) to be relocated too (basically they won't be >>>>> offsets from image base). In order to get real offset the >>>>> formulae "sym_offs(SYMBOL) - sym_offs(__image_base__)" is >>>>> used instead. >>>> >>>> The main calculations of %esi are, if I'm not mistaken, >>>> >>>> /* Store Xen image load base address in place accessible for 32-bit code. */ >>>> lea __image_base__(%rip),%esi >>>> >>> >>> Which is correct >>> >>>> and >>>> >>>> /* Calculate the load base address. */ >>>> call 1f >>>> 1: pop %esi >>>> sub $sym_offs(1b), %esi >>>> >>>> i.e. both deliberately %rip-relative to be position-independent. What's >>>> wrong with this? >>>> >>> >>> This can be wrong if sym_offs(1b) was relocated and not patched by >>> efi_arch_relocate_image. >> >> Of course, if in the course of GrUB's loading of xen.efi base relocations >> are applied (unlike when loading an ELF binary, where afaik base relocs >> would be ignored, even if there were any), then this calculation is of >> course going to be wrong. Can't we correct it though, to properly resemble >> PIC code: >> >> /* Calculate the load base address. */ >> call 1f >> 1: pop %esi >> sub 1b - start, %esi >> >> or (because start is in a different section): >> >> /* Calculate the load base address. */ >> call 1f >> 1: pop %esi >> sub $sym_offs(1b), %esi >> add $sym_offs(start), %esi >> >> (or something along these lines)? >> > > Yes, that works. But is a bit painfull, I mean, the %esi will point to > the correct address, but still you will use something like > syms_esi(foo) expecting to work but it won't as there will be applied > a relocation offset. I find your reply contradictory in itself. You first say this works, to then say it can't work. The underlying idea has to be to establish %esi such that it works uniformly. > On 32bit PIC code you could use something like > foo@GOTOFF(%esi), assuing %esi is pointing to the global offset table. > I was trying to use that but linker is complaining a bit as generating > a 64bit relocation. The x64 architecture supports such relocation as > 32bit but I didn't find a way to tell assembler to use the 32bit > version instead of the 64bit one. Also I didn't find a way to set > _GLOBAL_OFFSET_TABLE_ where I want it to be, it looks like that if the > linker is not generating it is not picking up the forcedly set symbol. Even if the toolchain permitted this: We don't have and don't want to have any GOT. Note how the linker script actually has an assertion for .got to be empty (plus a few more ones for other sections). Jan
On Thu, Aug 8, 2024 at 1:58 PM Jan Beulich <jbeulich@suse.com> wrote: > > On 08.08.2024 14:50, Frediano Ziglio wrote: > > On Thu, Aug 8, 2024 at 10:29 AM Jan Beulich <jbeulich@suse.com> wrote: > >> > >> (re-adding xen-devel@) > > Did you notice this in my earlier reply? You dropped the list again. > Yes, later, sorry for that. > >> On 08.08.2024 10:33, Frediano Ziglio wrote: > >>> On Thu, Aug 8, 2024 at 8:49 AM Jan Beulich <jbeulich@suse.com> wrote: > >>>>> This cause offsets in x86 code generated by > >>>>> sym_offs(SYMBOL) to be relocated too (basically they won't be > >>>>> offsets from image base). In order to get real offset the > >>>>> formulae "sym_offs(SYMBOL) - sym_offs(__image_base__)" is > >>>>> used instead. > >>>> > >>>> The main calculations of %esi are, if I'm not mistaken, > >>>> > >>>> /* Store Xen image load base address in place accessible for 32-bit code. */ > >>>> lea __image_base__(%rip),%esi > >>>> > >>> > >>> Which is correct > >>> > >>>> and > >>>> > >>>> /* Calculate the load base address. */ > >>>> call 1f > >>>> 1: pop %esi > >>>> sub $sym_offs(1b), %esi > >>>> > >>>> i.e. both deliberately %rip-relative to be position-independent. What's > >>>> wrong with this? > >>>> > >>> > >>> This can be wrong if sym_offs(1b) was relocated and not patched by > >>> efi_arch_relocate_image. > >> > >> Of course, if in the course of GrUB's loading of xen.efi base relocations > >> are applied (unlike when loading an ELF binary, where afaik base relocs > >> would be ignored, even if there were any), then this calculation is of > >> course going to be wrong. Can't we correct it though, to properly resemble > >> PIC code: > >> > >> /* Calculate the load base address. */ > >> call 1f > >> 1: pop %esi > >> sub 1b - start, %esi > >> > >> or (because start is in a different section): > >> > >> /* Calculate the load base address. */ > >> call 1f > >> 1: pop %esi > >> sub $sym_offs(1b), %esi > >> add $sym_offs(start), %esi > >> > >> (or something along these lines)? > >> > > > > Yes, that works. But is a bit painfull, I mean, the %esi will point to > > the correct address, but still you will use something like > > syms_esi(foo) expecting to work but it won't as there will be applied > > a relocation offset. > > I find your reply contradictory in itself. You first say this works, to > then say it can't work. The underlying idea has to be to establish %esi > such that it works uniformly. > The computation of %esi is correct after the additional "add" command, in the sense it will point to the current base (under 4GB) however then you will use syms_esi(foo) thinking "if %esi is correct then also syms_esi is correct" and it isn't. So either you need to add another offset to make syms_esi(foo) correct having %esi not pointing to the base or assuming that syms_esi(foo) would need fixing. Potentially the first option would be better, you just need to remember to correct %esi after rolling back relocations. > > On 32bit PIC code you could use something like > > foo@GOTOFF(%esi), assuing %esi is pointing to the global offset table. > > I was trying to use that but linker is complaining a bit as generating > > a 64bit relocation. The x64 architecture supports such relocation as > > 32bit but I didn't find a way to tell assembler to use the 32bit > > version instead of the 64bit one. Also I didn't find a way to set > > _GLOBAL_OFFSET_TABLE_ where I want it to be, it looks like that if the > > linker is not generating it is not picking up the forcedly set symbol. > > Even if the toolchain permitted this: We don't have and don't want to > have any GOT. Note how the linker script actually has an assertion for > .got to be empty (plus a few more ones for other sections). > I know, @GOTOFF does not generate .got entries, that's the reason I tried to use it and tell linker my idea of _GLOBAL_OFFSET_TABLE_. It's just that we need real offsets without relocations and I was looking at a way to get it. That would simply solve the relocation issue without having to offset %esi to some weird value. > Jan Frediano
On 08.08.2024 15:17, Frediano Ziglio wrote: > The computation of %esi is correct after the additional "add" command, > in the sense it will point to the current base (under 4GB) however > then you will use syms_esi(foo) thinking "if %esi is correct then also > syms_esi is correct" and it isn't. > So either you need to add another offset to make syms_esi(foo) correct > having %esi not pointing to the base or assuming that syms_esi(foo) > would need fixing. > Potentially the first option would be better, you just need to > remember to correct %esi after rolling back relocations. Right, the preferred goal is to have sym_esi() working right, so you wouldn't need to touch all of them. The number of direct uses of %esi is, I think, far smaller. Jan
diff --git a/xen/arch/x86/boot/head.S b/xen/arch/x86/boot/head.S index f027ff45fd..296f76146a 100644 --- a/xen/arch/x86/boot/head.S +++ b/xen/arch/x86/boot/head.S @@ -188,8 +188,27 @@ early_error: /* Here to improve the disassembly. */ xor %edi,%edi # No VGA text buffer jmp .Lprint_err .Lget_vtb: - mov sym_esi(vga_text_buffer), %edi + mov $sym_offs(vga_text_buffer), %edi .Lprint_err: + mov $sym_offs(__image_base__), %ebx + + /* compute base, relocation or not */ + call 1f +1: + pop %esi + subl $sym_offs(1b), %esi + addl %ebx, %esi + + /* adjust offset and load */ + test %edi, %edi + jz 1f + subl %ebx, %edi + movl (%edi,%esi,1), %edi +1: + + /* adjust message offset */ + subl %ebx, %ecx + add %ecx, %esi # Add string offset to relocation base. # NOTE: No further use of sym_esi() till the end of the "function"! 1:
If code is loaded by EFI the loader will relocate the image under 4GB. This cause offsets in x86 code generated by sym_offs(SYMBOL) to be relocated too (basically they won't be offsets from image base). In order to get real offset the formulae "sym_offs(SYMBOL) - sym_offs(__image_base__)" is used instead. Also, in some case %esi register (that should point to __image_base__ addresss) is not set so compute in all cases. Code tested forcing failures in the code. Signed-off-by: Frediano Ziglio <frediano.ziglio@cloud.com> --- xen/arch/x86/boot/head.S | 21 ++++++++++++++++++++- 1 file changed, 20 insertions(+), 1 deletion(-)