Actually, we had support for LLVM internally for a while (it is actually not confidential as I published the code beginning of this year: https://github.com/ARM-software/edk2/tree/armcc6). And we also had to force the 4K alignment for the PEI binaries (see https://github.com/ARM-software/edk2/commit/55cebbc5bf6fc3770e78b2b3dc706c0a22f2620a). I did not want to push the code into SVM as I have got concerns about the size of the UEFI firmware if all the XIP PEI binaries have to be 4K aligned in the UEFI firmware.
I wanted to do more investigation to see whether: - it really consume more space (by using padding between the 4K aligned PEI modules) in the UEFI Firmware - there was a way (probably through BaseTools) to play with the relocation offset to prevent these huge paddings -----Original Message----- From: Ard Biesheuvel [mailto:ard.biesheu...@linaro.org] Sent: 22 June 2015 20:28 To: edk2-devel@lists.sourceforge.net; Olivier Martin Cc: Cohen, Eugene Subject: Re: [edk2] [PATCH] BaseTools for AArch64 GCC: Ensure that the correlation .text and .data is consistent between ELF and PE-COFF so the debugger sees global variables correctly On 22 June 2015 at 19:26, Olivier Martin <olivier.mar...@arm.com> wrote: > So I am guessing it is fined to keep 0x260 used by the X64 linker script - > and so reuse the existing GCC ld script. Debuger scripts would need to be > updated to refect this change. > > If I compare the results with and without Eugene's patch: > > Without: > -------- > /work/gcc-linaro-4.9-2015.02-3-x86_64_arm-linux-gnueabihf/bin/arm-linu > x-gnueabihf-readelf -S > Build/ArmJuno-default/DEBUG_GCC49/AARCH64/ArmPkg/Drivers/TimerDxe/Time > rDxe/DEBUG/ArmTimerDxe.dll There are 23 section headers, starting at > offset 0x317a0: > > Section Headers: > [Nr] Name Type Address Offset > Size EntSize Flags Link Info Align > [ 0] NULL 0000000000000000 00000000 > 0000000000000000 0000000000000000 0 0 0 > [ 1] .text PROGBITS 0000000000000000 00010000 > 0000000000004468 0000000000000000 AX 0 0 8 > [ 2] .rela.text RELA 0000000000000000 00031d60 > 00000000000024d8 0000000000000018 21 1 8 > [ 3] .rodata PROGBITS 0000000000004468 00014468 > 00000000000018c8 0000000000000000 A 0 0 8 > [ 4] .data PROGBITS 0000000000015d30 00015d30 > 0000000000000190 0000000000000000 WA 0 0 8 > [ 5] .rela.data RELA 0000000000000000 00034238 > 0000000000000408 0000000000000018 21 4 8 > [ 6] .bss NOBITS 0000000000015ec0 00015ec0 > 0000000000000048 0000000000000000 WA 0 0 8 > (...) > > With: > ----- > /work/gcc-linaro-4.9-2015.02-3-x86_64_arm-linux-gnueabihf/bin/arm-linu > x-gnueabihf-readelf -S > Build/ArmJuno/DEBUG_GCC49/AARCH64/ArmPkg/Drivers/TimerDxe/TimerDxe/DEB > UG/ArmTimerDxe.dll There are 20 section headers, starting at offset > 0x317b0: > > Section Headers: > [Nr] Name Type Address Offset > Size EntSize Flags Link Info Align > [ 0] NULL 0000000000000000 00000000 > 0000000000000000 0000000000000000 0 0 0 > [ 1] .text PROGBITS 0000000000000000 00010000 > 0000000000004480 0000000000000000 AX 0 0 8 > [ 2] .rela.text RELA 0000000000000000 00031cb0 > 00000000000024d8 0000000000000018 18 1 8 > [ 3] .data PROGBITS 0000000000004480 00014480 > 0000000000001aa0 0000000000000000 WA 0 0 8 > [ 4] .rela.data RELA 0000000000000000 00034188 > 0000000000000408 0000000000000018 18 3 8 > (...) > > Some regions have disappeared such as .comment (ok!), but also .bss and > .rodata (which I am more concerned). > I have the impression the linker script combine these regions into .data (I > confirm the size of these 3 regions is equal to 0x1aa0). It means it would > change the PE/COFF binary. We can lose in memory protection if we decide > later to make the '.rodata' section a ReadOnly region. > > I also do not understand why we have .rela.text between .text and .data > sections. I would expect this section to be present after .data if your > script was working as expected. > I have been spending some time investigating this myself. I agree that -Ttext=0x0 seems to be redundant, so we should be able to reuse the X86 script. However, there are some other considerations: - the x86 script puts .rodata in .data, which is something I would prefer to avoid since we are moving towards stricter permissions, and putting .rodata in .data strips it from its constness (I don't think we're quite there yet in UEFI 2.5 but it would be nice if EfiRuntimeCode would map cleanly to R-X and EfiRuntimeData to RW-) - if we put .text at 0x800 and align it to 0x800, we could potentially drop the -mcmodel=large configuration, at least for PE32 modules, i.e., something like """ SECTIONS { .text 0x800 : ALIGN(0x800) { *(.text .text.* .rodata .rodata.*) } .data : ALIGN(0x40) { *(.data .data.*) *(.bss .bss.* *COM*) } .rela ALIGN(0x20) : { *(.rela .rela.*) } /DISCARD/ : { *(.note.GNU-stack) *(.interp) *(.dynsym) *(.dynstr) *(.dynamic) *(.hash) *(.comment) } } """ This should also solve Eugene's problem, since it results in the ELF and PE/COFF binary to be laid out identically. It also results in all code sections to be at the same alignment relative to 4K, which is required for the small C model that uses ADRP/ADD and ADRP/LDR pairs instead of 64-bit indirect addressing for globals. I managed to build and run ArmVirtQemu.dsc with the above script, -mcmodel=large removed and ADRP and related relocations ignored by the BaseTools, and this patch applied on top of that diff --git a/ArmVirtPkg/ArmVirtQemu.fdf b/ArmVirtPkg/ArmVirtQemu.fdf index 73d088a3bbdb..de81379105d1 100644 --- a/ArmVirtPkg/ArmVirtQemu.fdf +++ b/ArmVirtPkg/ArmVirtQemu.fdf @@ -313,14 +313,14 @@ [Rule.Common.SEC] [Rule.Common.PEI_CORE] FILE PEI_CORE = $(NAMED_GUID) { - TE TE Align = 8 $(INF_OUTPUT)/$(MODULE_NAME).efi + TE TE Align = 4K $(INF_OUTPUT)/$(MODULE_NAME).efi UI STRING ="$(MODULE_NAME)" Optional } [Rule.Common.PEIM] FILE PEIM = $(NAMED_GUID) { PEI_DEPEX PEI_DEPEX Optional $(INF_OUTPUT)/$(MODULE_NAME).depex - TE TE Align = 8 $(INF_OUTPUT)/$(MODULE_NAME).efi + TE TE Align = 4K $(INF_OUTPUT)/$(MODULE_NAME).efi UI STRING="$(MODULE_NAME)" Optional } > -----Original Message----- > From: Olivier Martin [mailto:olivier.mar...@arm.com] > Sent: 22 June 2015 17:43 > To: Cohen, Eugene; edk2-devel@lists.sourceforge.net > Subject: Re: [edk2] [PATCH] BaseTools for AArch64 GCC: Ensure that the > correlation .text and .data is consistent between ELF and PE-COFF so > the debugger sees global variables correctly > > The text section at zero is there because GNU toolchain used to make this > section at 0x8000(? - I am not sure of the value but it was not zero). > So the debugging script would not even work for debugging the code section. > > I have just checked with the latest Linaro GCC toolchain and I have not > managed to duplicate the issue. Both AArch32 and AArch64 GNU toolchains > generate text section at 0x0 by default. > > > -----Original Message----- > From: Cohen, Eugene [mailto:eug...@hp.com] > Sent: 18 June 2015 14:57 > To: edk2-devel@lists.sourceforge.net; Olivier Martin > Subject: RE: [edk2] [PATCH] BaseTools for AArch64 GCC: Ensure that the > correlation .text and .data is consistent between ELF and PE-COFF so > the debugger sees global variables correctly > >> I would still prefer to use the existing GCC ld script if possible >> instead of rolling our own. Since the existing one does not have the >> 64 KB gap between the RX and RW segments, and it probably also dodges >> the 'Multiple sections' error. > > Sure, I'm fine with that so long as it works. > >> So perhaps Olivier can explain where the -Ttext=0x0 comes from? It >> looks like everything works fine without it ... > > Olivier, do you know if the text at zero is required or can we adopt the X86 > placement in BaseTools/Scripts/gcc-arm-ld-script? > > Eugene > > -----Original Message----- > From: Ard Biesheuvel [mailto:ard.biesheu...@linaro.org] > Sent: Wednesday, June 17, 2015 10:48 PM > To: edk2-devel@lists.sourceforge.net > Subject: Re: [edk2] [PATCH] BaseTools for AArch64 GCC: Ensure that the > correlation .text and .data is consistent between ELF and PE-COFF so > the debugger sees global variables correctly > > On 18 June 2015 at 00:01, Cohen, Eugene <eug...@hp.com> wrote: >> Thanks for the great feedback -- you made me do more homework than I was >> expecting which is not a bad thing. >> >> Since we understand the problem and the fix, I think it's time to get >> Olivier's review and approval to move forward. >> > > I would still prefer to use the existing GCC ld script if possible > instead of rolling our own. Since the existing one does not have the > 64 KB gap between the RX and RW segments, and it probably also dodges the > 'Multiple sections' error. > > So perhaps Olivier can explain where the -Ttext=0x0 comes from? It looks like > everything works fine without it ... > > -- > Ard. > > >> -----Original Message----- >> From: Ard Biesheuvel [mailto:ard.biesheu...@linaro.org] >> Sent: Tuesday, June 16, 2015 2:51 PM >> To: edk2-devel@lists.sourceforge.net; Olivier Martin >> Subject: Re: [edk2] [PATCH] BaseTools for AArch64 GCC: Ensure that >> the correlation .text and .data is consistent between ELF and PE-COFF >> so the debugger sees global variables correctly >> >> On 16 June 2015 at 22:32, Cohen, Eugene <eug...@hp.com> wrote: >>>> OK. so does that mean we are using a builtin linker script for >>>> AArch64? That sounds fragile to me ... >>> >>> Apparently we are using the builtin script. I learned that this can be >>> queried with the -verbose switch for ld and have attached it. It includes >>> this interesting alignment mechanism: >>> >>> /* Adjust the address for the data segment. We want to adjust up to >>> the same address within the page on the next page up. */ >>> . = ALIGN(CONSTANT (MAXPAGESIZE)) + (. & (CONSTANT (MAXPAGESIZE) >>> - 1)); >>> >>> MAXPAGESIZE doesn't appear to be well documented but I can see patches to >>> ld that show this being increased to 64KB which is consistent with my >>> objdump output before the change. So I think this is the "offending" line >>> in the original script since it shifted data out in ELF even though the >>> PE-COFF converter packed it tightly, accounting for the section alignment >>> requirements from ELF. >>> >>>> Could you please look at the X64 approach, and compare it to yours? >>> >>> It's really a question for the developer originating -Ttext=0x0 - maybe >>> Olivier? I don't have the history on how the GCC linker configuration for >>> edk2 came to be. >>> >>>> Perhaps you could share some numbers or other details to get a feel >>>> for what exactly goes on here. >>> >>> I think I described in this in my email titled " AArch64 Debugger Global >>> Variable Correlation Issues" - I gave the ELF dump showing data shifted out >>> by 64KB. Strangely, this was shifted out not to the next 64KB aligned >>> boundary but to 0x10000 beyond the last .test/.rodata section - I'm not >>> sure why but perhaps somebody with more GNU ld experience can figure out >>> why. >>> >> >> That is *precisely* what the expression above aims to achieve. It >> wants to preserve the relative alignment of all the sections, but >> make sure that there is a 64 KB aligned boundary in between so that >> the two regions can always be mapped with different permissions even >> on a 64K pagesize OS. I.e., the expression aligns to the next >> boundary, and then adds the unaligned fraction of ".", which >> effectively just adds MAXPAGESIZE unless "." is already 64K aligned (if I am >> not mistaken). >> This matches your observation, right? >> >> I assume MAXPAGESIZE is a build time constant for ld, so there is not >> a lot of wiggle room here unless we switch to a custom linker script. >> >> -- >> Ard. >> >> >> >> >>> -----Original Message----- >>> From: Ard Biesheuvel [mailto:ard.biesheu...@linaro.org] >>> Sent: Tuesday, June 16, 2015 9:21 AM >>> To: edk2-devel@lists.sourceforge.net >>> Subject: Re: [edk2] [PATCH] BaseTools for AArch64 GCC: Ensure that >>> the correlation .text and .data is consistent between ELF and >>> PE-COFF so the debugger sees global variables correctly >>> >>> On 16 June 2015 at 16:50, Cohen, Eugene <eug...@hp.com> wrote: >>>>> Could you please send patches inline? Attachments are a pain to review. >>>> >>>> Sure, I will do that. Because I use Outlook I prefer the attachment since >>>> I can pull it into a nice patch viewer. >>>> >>> >>> I can see how that would be preferred for just viewing, but for >>> commenting inline it is far from optimal. >>> >>>>> Applying your patch and doing 'git show --find-copies-harder' gives me >>>>> (tools_def.tempate omitted): >>>> >>>> Unfortunately git found a copy that is not appropriate for >>>> comparison. It found the X86/X64 linker script that, while >>>> similar, is different than what we were already doing for AArch64. >>>> (I'm not saying the X86 approach doesn't work but my goal here was >>>> to minimize the impact to what was already being done for AArch64.) >>>> >>> >>> OK. so does that mean we are using a builtin linker script for >>> AArch64? That sounds fragile to me ... >>> Could you please look at the X64 approach, and compare it to yours? >>> Personally, I'd prefer to stay as close as possible to what is being >>> done on Intel, for obvious reasons ... >>> >>> >>>> On AArch64 the .text section was being placed with this linker switch: >>>> >>>> -Ttext=0x0 >>>> >>>> But this was not sufficient to pack the .data section in a manner that is >>>> consistent with the PE-COFF conversion that happens later in the build. >>>> So when I converted this to a linker control file, I maintained the zero >>>> starting address with: >>>> >>>> . = 0; >>>> >>> >>> OK, that parts seems obvious. So how is the packing of the .data >>> section being affected by your version of the linker script? >>> Perhaps you could share some numbers or other details to get a feel >>> for what exactly goes on here. >>> >>>>> So are you saying that the resulting PE/COFF is identical (for all >>>>> intents and purposes), and only the ELF intermediate file deviates? >>>> >>>> Yes, from my testing the PE/COFF looks the same but the ELF is updated to >>>> move .data such that its offset relative to .text is consistent with >>>> PE/COFF. This fixes debugger correlation for stuff in the .data section >>>> like global variables. >>>> >>>>> How does this affect >>>>> ArmPkg/Library/DebugPeCoffExtraActionLib/DebugPeCoffExtraActionLib. >>>>> inf, >>>>> which outputs lines like >>>> >>>> >From my testing it has no effect since this output is only outputting the >>>> >PE-COFF image address offset by the size of the PE-COFF headers (0x260 in >>>> >this case): >>>> >>>> (UINTN)(ImageContext->ImageAddress + >>>> ImageContext->SizeOfHeaders)) >>>> >>>> The reason the SizeOfHeaders is added is because the debugger is unaware >>>> of the PE-COFF conversion and added executable headers. >>>> >>> >>> OK, that makes sense. >>> >>>> -----Original Message----- >>>> From: Ard Biesheuvel [mailto:ard.biesheu...@linaro.org] >>>> Sent: Tuesday, June 16, 2015 5:56 AM >>>> To: edk2-devel@lists.sourceforge.net >>>> Subject: Re: [edk2] [PATCH] BaseTools for AArch64 GCC: Ensure that >>>> the correlation .text and .data is consistent between ELF and >>>> PE-COFF so the debugger sees global variables correctly >>>> >>>> On 4 June 2015 at 21:48, Cohen, Eugene <eug...@hp.com> wrote: >>>>> Oops, left off the contribution agreement line, trying again >>>>> >>>>> Dear ArmPkg maintainers (and later BaseTools maintainer), >>>>> >>>>> This is a fix for debugger correlation of global variables for AArch64 >>>>> built on GCC. >>>>> >>>>> Before this change looking at global variables with a debugger showed >>>>> bogus memory locations. This is because the offset of the .data section >>>>> in the ELF file did not reflect where it was placed in the PE-COFF (.efi) >>>>> output. >>>>> >>>>> This change passes a linker control script so that the data section is >>>>> packed next to .text so the ELF accurately reflects the relationship >>>>> between the sections when converted to PE-COFF by GenFw. >>>>> >>>>> I have tested this with the Lauterbach debugger. I don't know how well >>>>> it will work with other debuggers and debug scripts. >>>>> >>>>> If you would rather view the change as a github pull request: >>>>> https://github.com/tianocore/edk2/pull/12 >>>>> >>>>> Contributed-under: TianoCore Contribution Agreement 1.0 >>>>> Signed-off-by: Eugene Cohen <eug...@hp.com> >>>>> >>>> >>>> Could you please send patches inline? Attachments are a pain to review. >>>> >>>> Also, since the patch introduces a completely new linker script, it >>>> is hard to review how yours deviates from the default. >>>> Git is actually quite helpful since it can figure out if your newly >>>> introduced file resembles an existing file, and only shows the diff >>>> with respect to the original. >>>> >>>> Applying your patch and doing 'git show --find-copies-harder' gives >>>> me (tools_def.tempate omitted): >>>> >>>> """ >>>> diff --git a/BaseTools/Scripts/gcc4.4-ld-script >>>> b/BaseTools/Scripts/gcc-arm-ld-script >>>> similarity index 59% >>>> copy from BaseTools/Scripts/gcc4.4-ld-script >>>> copy to BaseTools/Scripts/gcc-arm-ld-script >>>> index 68b2767590ac..e1589a4d03bf 100644 >>>> --- a/BaseTools/Scripts/gcc4.4-ld-script >>>> +++ b/BaseTools/Scripts/gcc-arm-ld-script >>>> @@ -1,8 +1,10 @@ >>>> -/* OUTPUT_FORMAT(efi-bsdrv-x86_64) */ SECTIONS { >>>> - /* . = 0 + SIZEOF_HEADERS; */ >>>> - . = 0x280; >>>> + /* Start at 0 so we can meet more aggressive alignment requires >>>> after the PE-COFF conversion >>>> + like those for ARM exception vectors. This requires debugger >>>> scripts to offset past >>>> + the PE-COFF header (typically 0x260). When the PE-COFF >>>> conversion occurs we will >>>> + get proper alignment since the ELF section alignment is >>>> + applied >>>> in the conversion process. */ >>>> + . = 0; >>>> .text ALIGN(0x20) : >>>> { >>>> *(.text .stub .text.* .gnu.linkonce.t.*) """ >>>> >>>> So are you saying that the resulting PE/COFF is identical (for all >>>> intents and purposes), and only the ELF intermediate file deviates? >>>> How does this affect >>>> ArmPkg/Library/DebugPeCoffExtraActionLib/DebugPeCoffExtraActionLib. >>>> i >>>> nf, >>>> which outputs lines like >>>> >>>> add-symbol-file >>>> /home/ard/build/uefi-next/Build/ArmVirtQemu-AARCH64/DEBUG_GCC48/AAR >>>> C H64/MdeModulePkg/Core/Dxe/DxeMain/DEBUG/DxeCore.dll >>>> 0x5F2C7260 >>>> >>>> where the .dll is and ELF file, and the line seems to incorporate >>>> the header offset that you are removing. >>>> >>>> Regards, >>>> Ard. >>>> >>>> ------------------------------------------------------------------- >>>> - >>>> ---------- _______________________________________________ >>>> edk2-devel mailing list >>>> edk2-devel@lists.sourceforge.net >>>> https://lists.sourceforge.net/lists/listinfo/edk2-devel >>>> >>>> ------------------------------------------------------------------- >>>> - >>>> ---------- _______________________________________________ >>>> edk2-devel mailing list >>>> edk2-devel@lists.sourceforge.net >>>> https://lists.sourceforge.net/lists/listinfo/edk2-devel >>> >>> -------------------------------------------------------------------- >>> - >>> --------- _______________________________________________ >>> edk2-devel mailing list >>> edk2-devel@lists.sourceforge.net >>> https://lists.sourceforge.net/lists/listinfo/edk2-devel >>> >>> -------------------------------------------------------------------- >>> - >>> --------- >>> >>> _______________________________________________ >>> edk2-devel mailing list >>> edk2-devel@lists.sourceforge.net >>> https://lists.sourceforge.net/lists/listinfo/edk2-devel >>> >> >> --------------------------------------------------------------------- >> - >> -------- _______________________________________________ >> edk2-devel mailing list >> edk2-devel@lists.sourceforge.net >> https://lists.sourceforge.net/lists/listinfo/edk2-devel >> >> --------------------------------------------------------------------- >> - >> -------- _______________________________________________ >> edk2-devel mailing list >> edk2-devel@lists.sourceforge.net >> https://lists.sourceforge.net/lists/listinfo/edk2-devel > > ---------------------------------------------------------------------- > -------- _______________________________________________ > edk2-devel mailing list > edk2-devel@lists.sourceforge.net > https://lists.sourceforge.net/lists/listinfo/edk2-devel > > > -- IMPORTANT NOTICE: The contents of this email and any attachments are > confidential and may also be privileged. If you are not the intended > recipient, please notify the sender immediately and do not disclose the > contents to any other person, use it for any purpose, or store or copy the > information in any medium. Thank you. > > ARM Limited, Registered office 110 Fulbourn Road, Cambridge CB1 9NJ, > Registered in England & Wales, Company No: 2557590 ARM Holdings plc, > Registered office 110 Fulbourn Road, Cambridge CB1 9NJ, Registered in > England & Wales, Company No: 2548782 > > > ---------------------------------------------------------------------- > -------- Monitor 25 network devices or servers for free with > OpManager! > OpManager is web-based network management software that monitors > network devices and physical & virtual servers, alerts via email & sms > for fault. Monitor 25 devices for free with no restriction. Download > now http://ad.doubleclick.net/ddm/clk/292181274;119417398;o > _______________________________________________ > edk2-devel mailing list > edk2-devel@lists.sourceforge.net > https://lists.sourceforge.net/lists/listinfo/edk2-devel > > > -- IMPORTANT NOTICE: The contents of this email and any attachments are > confidential and may also be privileged. If you are not the intended > recipient, please notify the sender immediately and do not disclose the > contents to any other person, use it for any purpose, or store or copy the > information in any medium. Thank you. > > ARM Limited, Registered office 110 Fulbourn Road, Cambridge CB1 9NJ, > Registered in England & Wales, Company No: 2557590 ARM Holdings plc, > Registered office 110 Fulbourn Road, Cambridge CB1 9NJ, Registered in > England & Wales, Company No: 2548782 > > > ---------------------------------------------------------------------- > -------- Monitor 25 network devices or servers for free with > OpManager! > OpManager is web-based network management software that monitors > network devices and physical & virtual servers, alerts via email & sms > for fault. Monitor 25 devices for free with no restriction. Download > now http://ad.doubleclick.net/ddm/clk/292181274;119417398;o > _______________________________________________ > edk2-devel mailing list > edk2-devel@lists.sourceforge.net > https://lists.sourceforge.net/lists/listinfo/edk2-devel -- IMPORTANT NOTICE: The contents of this email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please notify the sender immediately and do not disclose the contents to any other person, use it for any purpose, or store or copy the information in any medium. Thank you. ARM Limited, Registered office 110 Fulbourn Road, Cambridge CB1 9NJ, Registered in England & Wales, Company No: 2557590 ARM Holdings plc, Registered office 110 Fulbourn Road, Cambridge CB1 9NJ, Registered in England & Wales, Company No: 2548782 ------------------------------------------------------------------------------ Monitor 25 network devices or servers for free with OpManager! OpManager is web-based network management software that monitors network devices and physical & virtual servers, alerts via email & sms for fault. Monitor 25 devices for free with no restriction. Download now http://ad.doubleclick.net/ddm/clk/292181274;119417398;o _______________________________________________ edk2-devel mailing list edk2-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/edk2-devel
