On Mon, Jun 19, 2023 at 8:17 PM Yonghong Song yhs@meta.com wrote:
On 6/19/23 7:03 AM, Florent Revest wrote:
On Fri, Jun 16, 2023 at 6:57 PM Andrii Nakryiko andrii.nakryiko@gmail.com wrote:
On Thu, Jun 15, 2023 at 7:56 AM Florent Revest revest@chromium.org wrote:
When building a kernel with LLVM=1, LLVM_IAS=0 and CONFIG_KASAN=y, LLVM leaves DWARF tags for the "asan.module_ctor" & co symbols. In turn, pahole creates BTF_KIND_FUNC entries for these and this makes the BTF metadata validation fail because they contain a dot.
In a dramatic turn of event, this BTF verification failure can cause the netfilter_bpf initialization to fail, causing netfilter_core to free the netfilter_helper hashmap and netfilter_ftp to trigger a use-after-free. The risk of u-a-f in netfilter will be addressed separately but the existence of "asan.module_ctor" debug info under some build conditions sounds like a good enough reason to accept functions that contain dots in BTF.
I don't see much harm in allowing dots. There are also all those .isra and other modifications to functions that we currently don't have in BTF, but with the discussions about recording function addrs we might eventually have those as well. So:
Acked-by: Andrii Nakryiko andrii@kernel.org
Thanks Andrii! :)
Cc: stable@vger.kernel.org Fixes: 1dc92851849c ("bpf: kernel side support for BTF Var and DataSec")
So do you think these trailers should be kept ? I suppose we can either see this as a "new feature" to accommodate .isra that should go through bpf-next or as a bug fix that goes through bpf and gets backported to stable (without this, BTF wouldn't work on old kernels built under a new clang and with LLVM_IAS=0 and CONFIG_KASAN=y so this sounds like a legitimate bug fix to me, I just wanted to double check)
How many people really build the kernel with LLVM=1 LLVM_IAS=0 which uses clang compiler ans gcc 'as'. I think distro most likely won't do this if they intend to build the kernel with clang.
I tend to agree with you
Note that LLVM=1 implies to use both clang compiler and clang assembler.
However, this is only true since: f12b034afeb3 ("scripts/Makefile.clang: default to LLVM_IAS=1")
5.10 stable for example does not have that commit and LLVM_IAS=0 is still the default there. (actually that's how I stumbled upon this: by building a 5.10 LTS and then finding a way to reproduce it upstream by disabling LLVM_IAS)
Using clang17 and 'LLVM=1 LLVM_IAS=0', with latest bpf-next, I actually hit some build errors like:
/tmp/video-bios-59fa52.s: Assembler messages: /tmp/video-bios-59fa52.s:4: Error: junk at end of line, first unrecognized character is `"' /tmp/video-bios-59fa52.s:4: Error: file number less than one /tmp/video-bios-59fa52.s:5: Error: junk at end of line, first unrecognized character is `"' /tmp/video-bios-59fa52.s:6: Error: junk at end of line, first unrecognized character is `"' /tmp/video-bios-59fa52.s:7: Error: junk at end of line, first unrecognized character is `"' /tmp/video-bios-59fa52.s:8: Error: junk at end of line, first unrecognized character is `"' /tmp/video-bios-59fa52.s:9: Error: junk at end of line, first unrecognized character is `"' /tmp/video-bios-59fa52.s:10: Error: junk at end of line, first unrecognized character is `"' /tmp/video-bios-59fa52.s:68: Error: junk at end of line, first unrecognized character is `"' clang: error: assembler command failed with exit code 1 (use -v to see invocation) make[4]: *** [/home/yhs/work/bpf-next/scripts/Makefile.build:252: arch/x86/realmode/rm/video-bios.o] Error 1 make[4]: *** Waiting for unfinished jobs.... /tmp/wakemain-88777c.s: Assembler messages: /tmp/wakemain-88777c.s:4: Error: junk at end of line, first unrecognized character is `"' /tmp/wakemain-88777c.s:4: Error: file number less than one /tmp/wakemain-88777c.s:5: Error: junk at end of line, first unrecognized character is `"' /tmp/wakemain-88777c.s:6: Error: junk at end of line, first unrecognized character is `"' /tmp/wakemain-88777c.s:7: Error: junk at end of line, first unrecognized character is `"' /tmp/wakemain-88777c.s:8: Error: junk at end of line, first unrecognized character is `"' /tmp/wakemain-88777c.s:81: Error: junk at end of line, first unrecognized character is `"' /tmp/wakemain-88777c.s:312: Error: junk at end of line, first unrecognized character is `"' clang: error: assembler command failed with exit code 1 (use -v to see invocation)
Potentially because of my local gnu assembler 2.30-120.el8 won't work with some syntax generated by clang. Mixing clang compiler and arbitrary gnu assembler are not a good idea (see the above example). It might work with close-to-latest gnu assembler.
I did not hit this bug with clang 17 and bpf-next so it's probably an incompatibility with that gnu assembler indeed.
To support function name like '<fname>.isra', some llvm work will be needed, and it may take some time.
So in my opinion, this patch is NOT a bug fix. It won't affect distro. Whether we should backport to the old kernel, I am not sure whether it is absolutely necessary as casual build can always remove LLVM_IAS=0 or hack the kernel source itself.
If you think it's not worth backporting to 5.10 (where LLVM_IAS=0 is the default) then I could drop these trailers and send it to bpf-next with a different justification. Either way is fine by me.