| Message ID | 20210616092521.800788-1-Tony.Ambardar@gmail.com |
|---|---|
| State | New |
| Headers | show |
| Series | [bpf,v1] bpf: fix libelf endian handling in resolv_btfids | expand |
On Wed, 16 Jun 2021 at 09:38, Jiri Olsa <jolsa@redhat.com> wrote: > > On Wed, Jun 16, 2021 at 08:56:42AM -0700, Yonghong Song wrote: > > > > On 6/16/21 2:25 AM, Tony Ambardar wrote: > > > While patching the .BTF_ids section in vmlinux, resolve_btfids writes type > > > ids using host-native endianness, and relies on libelf for any required > > > translation when finally updating vmlinux. However, the default type of the > > > .BTF_ids section content is ELF_T_BYTE (i.e. unsigned char), and undergoes > > > no translation. This results in incorrect patched values if cross-compiling > > > to non-native endianness, and can manifest as kernel Oops and test failures > > > which are difficult to debug. > > nice catch, great libelf can do that ;-) Funny, I'd actually assumed that was your intention, but I just couldn't find where the data type was being set, so resorted to this "kludge". While there's a .BTF_ids section definition in include/linux/btf_ids.h, there's no means I can see to specify the data type either (i.e. in the gcc asm .pushsection() options). That approach would be cleaner. > > > > > > > Explicitly set the type of patched data to ELF_T_WORD, allowing libelf to > > > transparently handle the endian conversions. > > > > > > Fixes: fbbb68de80a4 ("bpf: Add resolve_btfids tool to resolve BTF IDs in ELF object") > > > Cc: stable@vger.kernel.org # v5.10+ > > > Cc: Jiri Olsa <jolsa@kernel.org> > > > Cc: Yonghong Song <yhs@fb.com> > > > Link: https://lore.kernel.org/bpf/CAPGftE_eY-Zdi3wBcgDfkz_iOr1KF10n=9mJHm1_a_PykcsoeA@mail.gmail.com/ > > > Signed-off-by: Tony Ambardar <Tony.Ambardar@gmail.com> > > > --- > > > tools/bpf/resolve_btfids/main.c | 3 +++ > > > 1 file changed, 3 insertions(+) > > > > > > diff --git a/tools/bpf/resolve_btfids/main.c b/tools/bpf/resolve_btfids/main.c > > > index d636643ddd35..f32c059fbfb4 100644 > > > --- a/tools/bpf/resolve_btfids/main.c > > > +++ b/tools/bpf/resolve_btfids/main.c > > > @@ -649,6 +649,9 @@ static int symbols_patch(struct object *obj) > > > if (sets_patch(obj)) > > > return -1; > > > + /* Set type to ensure endian translation occurs. */ > > > + obj->efile.idlist->d_type = ELF_T_WORD; > > > > The change makes sense to me as .BTF_ids contains just a list of > > u32's. > > > > Jiri, could you double check on this? > > the comment in ELF_T_WORD declaration suggests the size depends on > elf's class? > > ELF_T_WORD, /* Elf32_Word, Elf64_Word, ... */ > > data in .BTF_ids section are allways u32 > I believe the Elf32/Elf64 refer to the arch since some data structures vary between the two, but ELF_T_WORD is common to both, and valid as the data type of Elf_Data struct holding the .BTF_ids contents. See elf(5): Basic types The following types are used for N-bit architectures (N=32,64, ElfN stands for Elf32 or Elf64, uintN_t stands for uint32_t or uint64_t): ... ElfN_Word uint32_t Also see the code and comments in "elf.h": /* Types for signed and unsigned 32-bit quantities. */ typedef uint32_t Elf32_Word; typedef uint32_t Elf64_Word; > I have no idea how is this handled in libelf (perhaps it's ok), > but just that comment above suggests it could be also 64 bits, > cc-ing Frank and Mark for more insight > One other area I'd like to confirm is with section compression. Is it safe to ignore this for .BTF_ids? I've done so because include/linux/btf_ids.h appears to define the section with SHF_ALLOC flag set, which is incompatible with compression based on "libelf.h" comments. Thanks for reviewing, Tony > thanks, > jirka > > > > > > + > > > elf_flagdata(obj->efile.idlist, ELF_C_SET, ELF_F_DIRTY); > > > err = elf_update(obj->efile.elf, ELF_C_WRITE); > > > > > >
On Wed, Jun 16, 2021 at 03:09:13PM -0700, Tony Ambardar wrote: > On Wed, 16 Jun 2021 at 09:38, Jiri Olsa <jolsa@redhat.com> wrote: > > > > On Wed, Jun 16, 2021 at 08:56:42AM -0700, Yonghong Song wrote: > > > > > > On 6/16/21 2:25 AM, Tony Ambardar wrote: > > > > While patching the .BTF_ids section in vmlinux, resolve_btfids writes type > > > > ids using host-native endianness, and relies on libelf for any required > > > > translation when finally updating vmlinux. However, the default type of the > > > > .BTF_ids section content is ELF_T_BYTE (i.e. unsigned char), and undergoes > > > > no translation. This results in incorrect patched values if cross-compiling > > > > to non-native endianness, and can manifest as kernel Oops and test failures > > > > which are difficult to debug. > > > > nice catch, great libelf can do that ;-) > > Funny, I'd actually assumed that was your intention, but I just > couldn't find where the > data type was being set, so resorted to this "kludge". While there's a .BTF_ids > section definition in include/linux/btf_ids.h, there's no means I can > see to specify > the data type either (i.e. in the gcc asm .pushsection() options). That approach > would be cleaner. > > > > > > > > > > > Explicitly set the type of patched data to ELF_T_WORD, allowing libelf to > > > > transparently handle the endian conversions. > > > > > > > > Fixes: fbbb68de80a4 ("bpf: Add resolve_btfids tool to resolve BTF IDs in ELF object") > > > > Cc: stable@vger.kernel.org # v5.10+ > > > > Cc: Jiri Olsa <jolsa@kernel.org> > > > > Cc: Yonghong Song <yhs@fb.com> > > > > Link: https://lore.kernel.org/bpf/CAPGftE_eY-Zdi3wBcgDfkz_iOr1KF10n=9mJHm1_a_PykcsoeA@mail.gmail.com/ > > > > Signed-off-by: Tony Ambardar <Tony.Ambardar@gmail.com> > > > > --- > > > > tools/bpf/resolve_btfids/main.c | 3 +++ > > > > 1 file changed, 3 insertions(+) > > > > > > > > diff --git a/tools/bpf/resolve_btfids/main.c b/tools/bpf/resolve_btfids/main.c > > > > index d636643ddd35..f32c059fbfb4 100644 > > > > --- a/tools/bpf/resolve_btfids/main.c > > > > +++ b/tools/bpf/resolve_btfids/main.c > > > > @@ -649,6 +649,9 @@ static int symbols_patch(struct object *obj) > > > > if (sets_patch(obj)) > > > > return -1; > > > > + /* Set type to ensure endian translation occurs. */ > > > > + obj->efile.idlist->d_type = ELF_T_WORD; > > > > > > The change makes sense to me as .BTF_ids contains just a list of > > > u32's. > > > > > > Jiri, could you double check on this? > > > > the comment in ELF_T_WORD declaration suggests the size depends on > > elf's class? > > > > ELF_T_WORD, /* Elf32_Word, Elf64_Word, ... */ > > > > data in .BTF_ids section are allways u32 > > > > I believe the Elf32/Elf64 refer to the arch since some data structures vary > between the two, but ELF_T_WORD is common to both, and valid as the > data type of Elf_Data struct holding the .BTF_ids contents. See elf(5): > > Basic types > The following types are used for N-bit architectures (N=32,64, ElfN > stands for Elf32 or Elf64, uintN_t stands for uint32_t or uint64_t): > ... > ElfN_Word uint32_t > > Also see the code and comments in "elf.h": > /* Types for signed and unsigned 32-bit quantities. */ > typedef uint32_t Elf32_Word; > typedef uint32_t Elf64_Word; ok > > > I have no idea how is this handled in libelf (perhaps it's ok), > > but just that comment above suggests it could be also 64 bits, > > cc-ing Frank and Mark for more insight > > > > One other area I'd like to confirm is with section compression. Is it safe > to ignore this for .BTF_ids? I've done so because include/linux/btf_ids.h > appears to define the section with SHF_ALLOC flag set, which is > incompatible with compression based on "libelf.h" comments. not sure what you mean.. where it wouldn't be safe? what workflow/processing thanks, jirka
On Thu, 2021-06-17 at 11:10 +0200, Jiri Olsa wrote: > On Wed, Jun 16, 2021 at 03:09:13PM -0700, Tony Ambardar wrote: > > On Wed, 16 Jun 2021 at 09:38, Jiri Olsa <jolsa@redhat.com> wrote: > > > I have no idea how is this handled in libelf (perhaps it's ok), > > > but just that comment above suggests it could be also 64 bits, > > > cc-ing Frank and Mark for more insight > > > > > > > One other area I'd like to confirm is with section compression. Is > > it safe > > to ignore this for .BTF_ids? I've done so because > > include/linux/btf_ids.h > > appears to define the section with SHF_ALLOC flag set, which is > > incompatible with compression based on "libelf.h" comments. > > not sure what you mean.. where it wouldn't be safe? > what workflow/processing I haven't looked at the code/patch, but Tony is correct that if a section has SHF_ALLOC set it cannot be a compressed section. SHF_COMPRESSED is incompatbile with SHF_ALLOC (or SHF_NOBITS) sections, because it would be unclear what a loader would need to do with them (uncompress the data first, then map it, or map the compressed data as is into memory). So ignoring whether or not a section is compressed for SHF_ALLOC sections is fine. Cheers, Mark
On 6/17/21 11:02 AM, Jiri Olsa wrote: > On Thu, Jun 17, 2021 at 12:28:00AM +0200, Mark Wielaard wrote: >> On Wed, Jun 16, 2021 at 06:38:33PM +0200, Jiri Olsa wrote: >>>>> diff --git a/tools/bpf/resolve_btfids/main.c b/tools/bpf/resolve_btfids/main.c >>>>> index d636643ddd35..f32c059fbfb4 100644 >>>>> --- a/tools/bpf/resolve_btfids/main.c >>>>> +++ b/tools/bpf/resolve_btfids/main.c >>>>> @@ -649,6 +649,9 @@ static int symbols_patch(struct object *obj) >>>>> if (sets_patch(obj)) >>>>> return -1; >>>>> + /* Set type to ensure endian translation occurs. */ >>>>> + obj->efile.idlist->d_type = ELF_T_WORD; >>>> >>>> The change makes sense to me as .BTF_ids contains just a list of >>>> u32's. >>>> >>>> Jiri, could you double check on this? >>> >>> the comment in ELF_T_WORD declaration suggests the size depends on >>> elf's class? >>> >>> ELF_T_WORD, /* Elf32_Word, Elf64_Word, ... */ >>> >>> data in .BTF_ids section are allways u32 >>> >>> I have no idea how is this handled in libelf (perhaps it's ok), >>> but just that comment above suggests it could be also 64 bits, >>> cc-ing Frank and Mark for more insight >> >> It is correct to use ELF_T_WORD, which means a 32bit unsigned word. >> >> The comment is meant to explain that, but is really confusing if you >> don't know that Elf32_Word and Elf64_Word are the same thing (a 32bit >> unsigned word). This comes from being "too consistent" in defining all >> data types for both 32bit and 64bit ELF, even if those types are the >> same in both formats... >> >> Only Elf32_Addr/Elf64_Addr and Elf32_Off/Elf64_Off are different >> sizes. But Elf32/Elf_64_Half (16 bit), Elf32/Elf64_Word (32 bit), >> Elf32/Elf64_Xword (64 bit) and their Sword/Sxword (signed) variants >> are all identical data types in both the Elf32 and Elf64 formats. >> >> I don't really know why. It seems the original ELF spec was 32bit only >> and when introducing the ELF64 format "they" simply duplicated all >> data types whether or not those data type were actually different >> between the 32 and 64 bit format. > > nice, thanks for details > > Acked-by: Jiri Olsa <jolsa@redhat.com> Tony, could you do a v2 and summarize the remainder of the discussion in here for the commit message? Would be good to explicitly document the assumptions made and why they work. Thanks everyone, Daniel
On Thu, 17 Jun 2021 at 04:22, Daniel Borkmann <daniel@iogearbox.net> wrote: > > On 6/17/21 11:02 AM, Jiri Olsa wrote: > > On Thu, Jun 17, 2021 at 12:28:00AM +0200, Mark Wielaard wrote: > >> On Wed, Jun 16, 2021 at 06:38:33PM +0200, Jiri Olsa wrote: > >>>>> diff --git a/tools/bpf/resolve_btfids/main.c b/tools/bpf/resolve_btfids/main.c > >>>>> index d636643ddd35..f32c059fbfb4 100644 > >>>>> --- a/tools/bpf/resolve_btfids/main.c > >>>>> +++ b/tools/bpf/resolve_btfids/main.c > >>>>> @@ -649,6 +649,9 @@ static int symbols_patch(struct object *obj) > >>>>> if (sets_patch(obj)) > >>>>> return -1; > >>>>> + /* Set type to ensure endian translation occurs. */ > >>>>> + obj->efile.idlist->d_type = ELF_T_WORD; > >>>> > >>>> The change makes sense to me as .BTF_ids contains just a list of > >>>> u32's. > >>>> > >>>> Jiri, could you double check on this? > >>> > >>> the comment in ELF_T_WORD declaration suggests the size depends on > >>> elf's class? > >>> > >>> ELF_T_WORD, /* Elf32_Word, Elf64_Word, ... */ > >>> > >>> data in .BTF_ids section are allways u32 > >>> > >>> I have no idea how is this handled in libelf (perhaps it's ok), > >>> but just that comment above suggests it could be also 64 bits, > >>> cc-ing Frank and Mark for more insight > >> > >> It is correct to use ELF_T_WORD, which means a 32bit unsigned word. > >> > >> The comment is meant to explain that, but is really confusing if you > >> don't know that Elf32_Word and Elf64_Word are the same thing (a 32bit > >> unsigned word). This comes from being "too consistent" in defining all > >> data types for both 32bit and 64bit ELF, even if those types are the > >> same in both formats... > >> > >> Only Elf32_Addr/Elf64_Addr and Elf32_Off/Elf64_Off are different > >> sizes. But Elf32/Elf_64_Half (16 bit), Elf32/Elf64_Word (32 bit), > >> Elf32/Elf64_Xword (64 bit) and their Sword/Sxword (signed) variants > >> are all identical data types in both the Elf32 and Elf64 formats. > >> > >> I don't really know why. It seems the original ELF spec was 32bit only > >> and when introducing the ELF64 format "they" simply duplicated all > >> data types whether or not those data type were actually different > >> between the 32 and 64 bit format. > > > > nice, thanks for details > > > > Acked-by: Jiri Olsa <jolsa@redhat.com> > > Tony, could you do a v2 and summarize the remainder of the discussion in > here for the commit message? Would be good to explicitly document the > assumptions made and why they work. Sure, Daniel, I'll update the commit details and resend. Thanks, Tony > Thanks everyone, > Daniel
diff --git a/tools/bpf/resolve_btfids/main.c b/tools/bpf/resolve_btfids/main.c index d636643ddd35..f32c059fbfb4 100644 --- a/tools/bpf/resolve_btfids/main.c +++ b/tools/bpf/resolve_btfids/main.c @@ -649,6 +649,9 @@ static int symbols_patch(struct object *obj) if (sets_patch(obj)) return -1; + /* Set type to ensure endian translation occurs. */ + obj->efile.idlist->d_type = ELF_T_WORD; + elf_flagdata(obj->efile.idlist, ELF_C_SET, ELF_F_DIRTY); err = elf_update(obj->efile.elf, ELF_C_WRITE);
While patching the .BTF_ids section in vmlinux, resolve_btfids writes type ids using host-native endianness, and relies on libelf for any required translation when finally updating vmlinux. However, the default type of the .BTF_ids section content is ELF_T_BYTE (i.e. unsigned char), and undergoes no translation. This results in incorrect patched values if cross-compiling to non-native endianness, and can manifest as kernel Oops and test failures which are difficult to debug. Explicitly set the type of patched data to ELF_T_WORD, allowing libelf to transparently handle the endian conversions. Fixes: fbbb68de80a4 ("bpf: Add resolve_btfids tool to resolve BTF IDs in ELF object") Cc: stable@vger.kernel.org # v5.10+ Cc: Jiri Olsa <jolsa@kernel.org> Cc: Yonghong Song <yhs@fb.com> Link: https://lore.kernel.org/bpf/CAPGftE_eY-Zdi3wBcgDfkz_iOr1KF10n=9mJHm1_a_PykcsoeA@mail.gmail.com/ Signed-off-by: Tony Ambardar <Tony.Ambardar@gmail.com> --- tools/bpf/resolve_btfids/main.c | 3 +++ 1 file changed, 3 insertions(+)