On 2021/05/18 6:21, Romain Naour wrote: Note that I am seeing similar problems on RISC-V NOMMU builds with the latest buildroot/busybox 1.33. I had no time to explore the reason for it yet though. It sounds like it may be a similar problem as yours.

Same early segfault I am seeing on init shell startup (used as init process in my case).

>>>> "Damien" == Damien Le Moal <Damien.LeMoal(a)wdc.com> writes:

> The upstream Buildroot project only support RISC-V with MMU > (BR2_ARCH_HAS_MMU_MANDATORY).

I have a series of patches for building bootable images for boards based on the Canaan Kendryte K210 SoC. These boards use NOMMU builds that the patch series enables. I also have elf2flt fixed and the kernel flatbin loader is already fixed in 5.13-rc. The crash on startup I am seeing with the latest buildroot/busybox is blocking me from posting the patch series though.

>>>> "Damien" == Damien Le Moal <Damien.LeMoal(a)wdc.com> writes:

> Yes, I remember. What is the status of the elf2flt patch? It would be > good if you could send an updated patch series to the Buildroot list.

The elf2flt patch is ready. All seems good on this front now that the kernel flatbin loader is fixed to not break riscv global pointer.

Also, with all the K210 improvments that went into kernel 5.12, the buildroot patch series was expanded to add SD card image build and 4 different K210 boards.

However, when I rebased my series on the latest upstream tree, I ran into this crash on startup and have not had time to debug yet. So I have not sent anything yet, nor did I send the elf2flt patch as I want to check that the problem cause is not that one. Really not sure what is going on. I only had a quick look after my last rebase and added "need debug" to my to-do list.

Will try to get to it next week. I may just wait for 5.13 to come out though as the needed flatbin loader patch will be included. Since we are at rc2, still have some time :)

Hi Waldemar, Le 19/05/2021 à 00:09, Waldemar Brodkorb a écrit : Good question. I guess it should work with pie (see PIEFLAG_NAME:=-fpie) https://cgit.uclibc-ng.org/cgi/cgit/uclibc-ng.git/tree/Rules.mak?h=v1.0.38#… I did a try with Glibc without any problem with BR2_PIC_PIE enabled. Best regards, Romain

Hello, I've stayed silent so far but I've done some investigation on this one. It's not yet clear to me why there is this code path: |#ifdef __PIC__ # ifdef HAVE_ASM_PPC_REL16 bcl 20,31,1f 1: mflr r31 addis r31,r31,_GLOBAL_OFFSET_TABLE_-1b@ha addi r31,r31,_GLOBAL_OFFSET_TABLE_-1b@l # else bl _GLOBAL_OFFSET_TABLE_-4@local mflr r31 # endif #endif ||| If I understood correctly, the "new" (so I guess "better ?") way of doing is by using the PPC_REL16 reloc. Code seems bigger, and uses 2 relocs, compared to the other code path which uses just 1 reloc. So, why is the 1st better ? I am betting here that : it's because by using 2 16-bit relocs, you can indeed address 32 bit and reach the .got even if it is very far away from the current code. If you are stuck with only 16 bits ... it can fail. So, I am guessing (please if someone can confirm) that your fix will work in most cases ... but can fail when got is too far away to fit in the 16 bit reloc targeting the "bl" insn immediate. I think it would be better to understand why the first code path does not work well. Here is the current state of my investigations: I compiled both versions (with and without ppc_rel16) to compare what's happening in both cases. busybox_unstripped in the no_rel16 (#else) case: https://salamandar.fr/lufi/r/4TrBm3U7q5#z3J3Z3G24BTPJnQKD/EzGhAk/n3YHZSZF3h… busybox_unstripped in the ppc_rel16 (#if) case: https://salamandar.fr/lufi/r/pQ79KBb5wU#HFlcBZIVngZFVzvObe/afqeNgMEbMtwDk9a… In the working case (#else): ################# the _start symbol contains:    b022c:       48 03 fd f9     bl      f0024 <__TMC_END__+0x10>    b0230:       7f e8 02 a6     mflr    r31    b0234:       81 bf ff f0     lwz     r13,-16(r31) to load the r13 reg with SDA (Small Data Area) pointer. 0xf0024 is in the got which contains: 000f0014 <.got>:         ...    f0024:       4e 80 00 21     blrl 000f0028 <_GLOBAL_OFFSET_TABLE_>:    f0028:       00 0e ff 34     .long 0xeff34         ... so basically we jump to the got, we execute "blrl" which IIUC jumps back to our _start but with LR (link register) == 0xf0028 (nextpc) then mflr (move from link register) makes it so that r31 == 0xf0028 Then we lwz r13 = * (0xf0028 - 16) == * 0xf0018 0xf0018 is in the .got too and seems to be the target of a R_PPC_RELATIVE relocation which I *think* will end up with the 0xf0028 value which is the address of "_GLOBAL_OFFSET_TABLE_" symbol. In http://refspecs.linux-foundation.org/elf/elfspec_ppc.pdf we can read p.76: "In a shared object,_SDA_BASE_is defined to have the same value as_GLOBAL_OFFSET_TABLE_" So it seems all is indeed OK. |##############################################| |Now, in the crashing case (HAVE_ASM_PPC_REL16 = 1):| |||| the _start symbol contains:    b022c:       42 9f 00 05     bcl     20,4*cr7+so,b0230 <_start+0xc>    b0230:       7f e8 02 a6     mflr    r31    b0234:       3f ff 00 05     addis   r31,r31,5    b0238:       3b ff f8 6c     addi    r31,r31,-1940    b023c:       81 bf ff f4     lwz     r13,-12(r31) So we branch to 0xb0230 (nextpc) to move link register to r31 which ends up with the value 0xb0230. we do r31 = r31 + 0x10000 * 5 = 0x100230 we do r31 = r31 - 1940 = 0xffa9c Then we load r13 = *(0xffa9c - 12) = *(0xffa90) 0xffa90 seems to be the target of a R_PPC_RELATIVE relocation that targets 0xffa9c which is the address of the _GLOBAL_OFFSET_TABLE_ symbol. The code seems correct, I am guessing the relocation handling might be the cause of the problem? It might be interesting to compare glibc and uclibc-ng for PowerPC for the handling of R_PPC_RELATIVE reloc. Sorry I don't have any quick fix, I'm just sharing my thoughts on the issue. Cheers! Yann On 21/05/2021 10:34, Waldemar Brodkorb wrote: _______________________________________________ devel mailing list devel(a)uclibc-ng.org https://mailman.uclibc-ng.org/cgi-bin/mailman/listinfo/devel

Hi Romain, can you confirm that attached patch works for you? I tested with non-PIE and PIE and both seems to work.

best regards Waldemar Romain Naour wrote, > Hi Waldemar, > > Le 19/05/2021 à 00:09, Waldemar Brodkorb a écrit : >> Hi Romain, >> Romain Naour wrote, >> >>> Hello, >>> >>> Recently in Buildroot the option BR2_PIC_PIE has been enabled by default along >>> with other hardening features [1]. Since then some ppc defconfig such >>> qemu_ppc_e500mc_defconfig are failing to boot due to a segfault in init program. >>> >>> The segfault appear very early in __uClibc_main while starting any binaries, >>> an issue located in crt1.S (powerpc)[2]. >>> >>> After some trial and error, removing HAVE_ASM_PPC_REL16 from CFLAGS [3] >>> allow to generate a working system again. But this is actually wrong since >>> instead we should consider HAVE_ASM_PPC_REL16 always true nowadays. >> >> What if the assembly inside HAVE_ASM_PPC_REL16 isn't pie safe? > > Good question. > > I guess it should work with pie (see PIEFLAG_NAME:=-fpie) > https://cgit.uclibc-ng.org/cgi/cgit/uclibc-ng.git/tree/Rules.mak?h=v1.0.38#… > > I did a try with Glibc without any problem with BR2_PIC_PIE enabled. > > Best regards, > Romain > >> >>> Indeed, Glibc removed HAVE_ASM_PPC_REL16 since version 2.22 [4] since "the >>> minimum binutils supports rel16 relocs". Binutils 2.22 supports R_PPC_REL16 as >>> default. >>> >>> uClibc-ng should remove HAVE_ASM_PPC_REL16 but keep the code as it was defined. >>> But this doesn't fix the initial issue. >>> >>> Any idea ? >>> >>> [1] >>> https://git.buildroot.net/buildroot/commit/?id=810ba387bec3c5b6904e8893fb4c… >>> [2] >>> https://cgit.uclibc-ng.org/cgi/cgit/uclibc-ng.git/tree/libc/sysdeps/linux/p… >>> [3] >>> https://cgit.uclibc-ng.org/cgi/cgit/uclibc-ng.git/tree/Rules.mak?id=2bf4991… >>> [4] >>> https://sourceware.org/git/?p=glibc.git;a=commit;h=59261ad3eb345e0d7b9f5c73… >> >> best regards >> Waldemar >> >