| Message ID | 714d7ab8a48172c67ddc027c85b2a0dad0312a74.1726806756.git.tony.ambardar@gmail.com |
|---|---|
| State | New |
| Headers | show |
| Series | Improve .BTF_ids patching and alignment | expand |
On Fri, 2024-09-20 at 00:49 -0700, Tony Ambardar wrote: > While building of vmlinux employs a linker script to align the .BTF_ids > section to 4 bytes, other usage leaves .BTF_ids unaligned and may lead to > problems (e.g. [1]). Post-processing and libelf-based endian translation by > resolve_btfids may also potentially suffer from misalignment. > > Update encoding macros in btf_ids.h to always align BTF ID data to 4 bytes. > > [1]: 3effc06a4dde ("selftests/bpf: Fix alignment of .BTF_ids") > > Signed-off-by: Tony Ambardar <tony.ambardar@gmail.com> > --- > include/linux/btf_ids.h | 1 + > 1 file changed, 1 insertion(+) > > diff --git a/include/linux/btf_ids.h b/include/linux/btf_ids.h > index c0e3e1426a82..c10b163dc340 100644 > --- a/include/linux/btf_ids.h > +++ b/include/linux/btf_ids.h > @@ -89,6 +89,7 @@ word \ > #define __BTF_ID_LIST(name, scope) \ > asm( \ > ".pushsection " BTF_IDS_SECTION ",\"a\"; \n" \ > +".balign 4, 0; \n" \ > "." #scope " " #name "; \n" \ > #name ":; \n" \ > ".popsection; \n"); This forces all id list symbols to be aligned on 4 bytes. Should the same be done for __BTF_SET_START? Also, is it guaranteed that all btf ids are organized in lists and sets? Grepping through the code it seems they are, but it looks like resolve_btfids does not really enforce this, simply looking for symbols matching a special name __BTF_ID__<type>__<symbol>[__<id>] .
On Sat, Sep 21, 2024 at 02:46:17AM -0700, Eduard Zingerman wrote: > On Fri, 2024-09-20 at 00:49 -0700, Tony Ambardar wrote: > > While building of vmlinux employs a linker script to align the .BTF_ids > > section to 4 bytes, other usage leaves .BTF_ids unaligned and may lead to > > problems (e.g. [1]). Post-processing and libelf-based endian translation by > > resolve_btfids may also potentially suffer from misalignment. > > > > Update encoding macros in btf_ids.h to always align BTF ID data to 4 bytes. > > > > [1]: 3effc06a4dde ("selftests/bpf: Fix alignment of .BTF_ids") > > > > Signed-off-by: Tony Ambardar <tony.ambardar@gmail.com> > > --- > > include/linux/btf_ids.h | 1 + > > 1 file changed, 1 insertion(+) > > > > diff --git a/include/linux/btf_ids.h b/include/linux/btf_ids.h > > index c0e3e1426a82..c10b163dc340 100644 > > --- a/include/linux/btf_ids.h > > +++ b/include/linux/btf_ids.h > > @@ -89,6 +89,7 @@ word \ > > #define __BTF_ID_LIST(name, scope) \ > > asm( \ > > ".pushsection " BTF_IDS_SECTION ",\"a\"; \n" \ > > +".balign 4, 0; \n" \ > > "." #scope " " #name "; \n" \ > > #name ":; \n" \ > > ".popsection; \n"); > > This forces all id list symbols to be aligned on 4 bytes. > Should the same be done for __BTF_SET_START? it seems all the set macros use __BTF_ID_LIST, so it should be taken care of by that > > Also, is it guaranteed that all btf ids are organized in lists and sets? > Grepping through the code it seems they are, but it looks like resolve_btfids > does not really enforce this, simply looking for symbols matching a special name > __BTF_ID__<type>__<symbol>[__<id>] . yes, you need the BTF_ID to be part of list or set to be able to access it resolve_btfids does not enforce some loose BTF_ID definition without list/set, but that does not seem to be a problem thanks, jirka
On Mon, 2024-09-23 at 12:54 +0200, Jiri Olsa wrote: [...] > > > diff --git a/include/linux/btf_ids.h b/include/linux/btf_ids.h > > > index c0e3e1426a82..c10b163dc340 100644 > > > --- a/include/linux/btf_ids.h > > > +++ b/include/linux/btf_ids.h > > > @@ -89,6 +89,7 @@ word \ > > > #define __BTF_ID_LIST(name, scope) \ > > > asm( \ > > > ".pushsection " BTF_IDS_SECTION ",\"a\"; \n" \ > > > +".balign 4, 0; \n" \ > > > "." #scope " " #name "; \n" \ > > > #name ":; \n" \ > > > ".popsection; \n"); > > > > This forces all id list symbols to be aligned on 4 bytes. > > Should the same be done for __BTF_SET_START? > > it seems all the set macros use __BTF_ID_LIST, so it should be taken > care of by that Apologies, I don't know how I missed __BTF_ID_LIST invocation in the __BTF_SET_START definition :( > > Also, is it guaranteed that all btf ids are organized in lists and sets? > > Grepping through the code it seems they are, but it looks like resolve_btfids > > does not really enforce this, simply looking for symbols matching a special name > > __BTF_ID__<type>__<symbol>[__<id>] . > > yes, you need the BTF_ID to be part of list or set to be able to access it > > resolve_btfids does not enforce some loose BTF_ID definition without list/set, > but that does not seem to be a problem Understood, thank you.
diff --git a/include/linux/btf_ids.h b/include/linux/btf_ids.h index c0e3e1426a82..c10b163dc340 100644 --- a/include/linux/btf_ids.h +++ b/include/linux/btf_ids.h @@ -89,6 +89,7 @@ word \ #define __BTF_ID_LIST(name, scope) \ asm( \ ".pushsection " BTF_IDS_SECTION ",\"a\"; \n" \ +".balign 4, 0; \n" \ "." #scope " " #name "; \n" \ #name ":; \n" \ ".popsection; \n");
While building of vmlinux employs a linker script to align the .BTF_ids section to 4 bytes, other usage leaves .BTF_ids unaligned and may lead to problems (e.g. [1]). Post-processing and libelf-based endian translation by resolve_btfids may also potentially suffer from misalignment. Update encoding macros in btf_ids.h to always align BTF ID data to 4 bytes. [1]: 3effc06a4dde ("selftests/bpf: Fix alignment of .BTF_ids") Signed-off-by: Tony Ambardar <tony.ambardar@gmail.com> --- include/linux/btf_ids.h | 1 + 1 file changed, 1 insertion(+)