On 2023-09-14 14:49:44-0500, Eric W. Biederman wrote:
Thomas Weißschuh linux@weissschuh.net writes:
When allocating the pages for bss the start address needs to be rounded down instead of up. Otherwise the start of the bss segment may be unmapped.
The was reported to happen on Aarch64:
Those program headers you quote look corrupt.
To reproduce:
$ cat test.c char foo[1];
void __attribute__((weak, noreturn, optimize("Os", "omit-frame-pointer"))) _start(void) { __asm__ volatile ( "mov x0, 123\n" "mov x8, 93\n" /* NR_exit == 93 */ "svc #0\n" ); __builtin_unreachable(); }
$ aarch64-linux-gnu-gcc -fno-stack-protector -o nolibc-test -nostdlib -static test.c
Note: it works in qemu-user, newer versions need the workaround from https://gitlab.com/qemu-project/qemu/-/issues/1854 The issue in qemu-user seems to be related to the question at hand.
The address 0x41ffe8 is not 0x10000 aligned.
I don't think anything in the elf specification allows that.
The most common way to have bss is for a elf segment to have a larger memsize than filesize. In which case rounding up is the correct way to handle things.
We definitely need to verify the appended bss case works, before taking this patch, or we will get random application failures because parts of the data segment are being zeroed, or the binaries won't load because the bss won't be able to map over the initialized data.
My hope in posting this patch was also for the bots to uncover any obvious breakage. So far there were no reports.
The note segment living at a conflicting virtual address also looks suspicious. It is probably harmless, as note segments are not loaded.
Are you by any chance using an experimental linker?
I'm using GNU ld 2.41 as supplied by my distro. (ArchLinux, aarch64-linux-gnu-binutils 2.41-2)
In general every segment in an elf executable needs to be aligned to the SYSVABI's architecture page size. I think that is 64k on ARM. Which it looks like the linker tried to implement by setting the alignment to 0x10000, and then ignored by putting a byte offset beginning to the page.
Looking at Figure A-5 of [0] this seems not to be the case. It shows p_vaddr=0x8048100 and p_align=0x1000. (On x86_64 with PAGE_SIZE=0x1000)
At a minimum someone needs to sort through what the elf specification says needs to happen is a weird case like this where the start address of a load segment does not match the alignment of the segment.
I'll take a look.
To see how common this is I looked at a binary known to be working, and my /usr/bin/ls binary has one segment that has one of these unaligned starts as well.
Same for my /usr/bin/busybox, also the .data and .bss segment.
So it must be defined to work somewhere but I need to see the definition to even have a good opinion on the nonsense of saying an unaligned value should be aligned.
Figure 2-1 from [0]:
p_align:
Loadable process segments must have congruent values for p_vaddr and p_offset, modulo the page size.This member gives the value to which the segments are aligned in memory and in the file. Values 0 and 1 mean that no alignment is required. Otherwise, p_align should be a positive, integral power of 2, and p_addr should equal p_offset, modulo p_align.
0x41ffe8 (p_vaddr) % 0x1000 = 0xfe8 0x00ffe8 (p_offset) % 0x1000 = 0xfe8
0x41ffe8 (p_addr) % 0x10000 = 0xffe8 0x00ffe8 (p_offset) % 0x10000 = 0xffe8
So this seems to be satisfied.
All I know is that we need to limit our support to what memory mapping pieces from the elf executable can support. Which at a minimum requires: virt_addr % ELF_MIN_ALIGN == file_offset % ELF_MIN_ALIGN
Aarch64 can also handle 4k pages so this invariant should be satisfied. 4k pages seems to be the default for the kernel, too.
[0] https://refspecs.linuxfoundation.org/elf/elf.pdf
Memory allocated by set_brk(): Before: start=0x420000 end=0x420000 After: start=0x41f000 end=0x420000
The triggering binary looks like this:
Elf file type is EXEC (Executable file) Entry point 0x400144 There are 4 program headers, starting at offset 64 Program Headers: Type Offset VirtAddr PhysAddr FileSiz MemSiz Flags Align LOAD 0x0000000000000000 0x0000000000400000 0x0000000000400000 0x0000000000000178 0x0000000000000178 R E 0x10000 LOAD 0x000000000000ffe8 0x000000000041ffe8 0x000000000041ffe8 0x0000000000000000 0x0000000000000008 RW 0x10000 NOTE 0x0000000000000120 0x0000000000400120 0x0000000000400120 0x0000000000000024 0x0000000000000024 R 0x4 GNU_STACK 0x0000000000000000 0x0000000000000000 0x0000000000000000 0x0000000000000000 0x0000000000000000 RW 0x10 Section to Segment mapping: Segment Sections... 00 .note.gnu.build-id .text .eh_frame 01 .bss 02 .note.gnu.build-id 03Reported-by: Sebastian Ott sebott@redhat.com Closes: https://lore.kernel.org/lkml/5d49767a-fbdc-fbe7-5fb2-d99ece3168cb@redhat.com... Fixes: 1da177e4c3f4 ("Linux-2.6.12-rc2") Cc: stable@vger.kernel.org Signed-off-by: Thomas Weißschuh linux@weissschuh.net
I'm not really familiar with the ELF loading process, so putting this out as RFC.
A example binary compiled with aarch64-linux-gnu-gcc 13.2.0 is available at https://test.t-8ch.de/binfmt-bss-repro.bin
fs/binfmt_elf.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/fs/binfmt_elf.c b/fs/binfmt_elf.c index 7b3d2d491407..4008a57d388b 100644 --- a/fs/binfmt_elf.c +++ b/fs/binfmt_elf.c @@ -112,7 +112,7 @@ static struct linux_binfmt elf_format = { static int set_brk(unsigned long start, unsigned long end, int prot) {
- start = ELF_PAGEALIGN(start);
- start = ELF_PAGESTART(start); end = ELF_PAGEALIGN(end); if (end > start) { /*
base-commit: aed8aee11130a954356200afa3f1b8753e8a9482 change-id: 20230914-bss-alloc-f523fa61718c
Best regards,