Hi Baruch,
On Fri, May 3, 2019 at 1:52 PM Baruch Siach <[email protected]> wrote:
> On Fri, May 03 2019, Geert Uytterhoeven wrote:
> > On Fri, May 3, 2019 at 6:06 AM Baruch Siach <[email protected]> wrote:
> >> strace 5.0 fails to build for m86k/5208 with the Buildroot generated
> >> toolchain:
> >>
> >> In file included from bpf_attr_check.c:6:0:
> >> static_assert.h:20:25: error: static assertion failed: "bpf_prog_info_struct.nr_jited_ksyms offset mismatch"
> >> # define static_assert _Static_assert
> >> ^
> >> bpf_attr_check.c:913:2: note: in expansion of macro ‘static_assert’
> >> static_assert(offsetof(struct bpf_prog_info_struct, nr_jited_ksyms) == offsetof(struct bpf_prog_info, nr_jited_ksyms),
> >> ^~~~~~~~~~~~~
> >>
> >> The direct cause is a difference in the hole after the gpl_compatible
> >> field. Here is pahole output for the kernel struct (from v4.19):
> >>
> >> struct bpf_prog_info {
> >> ...
> >> __u32 ifindex; /* 80 4 */
> >> __u32 gpl_compatible:1; /* 84: 0 4 */
> >>
> >> /* XXX 15 bits hole, try to pack */
> >> /* Bitfield combined with next fields */
> >>
> >> __u64 netns_dev; /* 86 8 */
> >
> > I guess that should be "__aligned_u64 netns_dev;", to not rely on
> > implicit alignment.
>
> Thanks. I can confirm that this minimal change fixes strace build:
>
> diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
> index 929c8e537a14..709d4dddc229 100644
> --- a/include/uapi/linux/bpf.h
> +++ b/include/uapi/linux/bpf.h
> @@ -2869,7 +2869,7 @@ struct bpf_prog_info {
> char name[BPF_OBJ_NAME_LEN];
> __u32 ifindex;
> __u32 gpl_compatible:1;
> - __u64 netns_dev;
> + __aligned_u64 netns_dev;
> __u64 netns_ino;
> __u32 nr_jited_ksyms;
> __u32 nr_jited_func_lens;
>
> Won't that break ABI compatibility for affected architectures?
Yes it will. Or it may have been unusable without the fix. I don't know
for sure.
> >> And this is for the strace struct:
> >>
> >> struct bpf_prog_info_struct {
> >> ...
> >> uint32_t ifindex; /* 80 4 */
> >> uint32_t gpl_compatible:1; /* 84: 0 4 */
> >>
> >> /* XXX 31 bits hole, try to pack */
> >
> > How come the uint64_t below is 8-byte aligned, not 2-byte aligned?
> > Does strace use a special definition of uint64_t?
>
> I guess this is because of the netns_dev field definition in struct
> bpf_prog_info_struct at bpf_attr.h:
>
> struct bpf_prog_info_struct {
> ...
> uint32_t gpl_compatible:1;
> /*
> * The kernel UAPI is broken by Linux commit
> * v4.16-rc1~123^2~227^2~5^2~2 .
> */
> uint64_t ATTRIBUTE_ALIGNED(8) netns_dev; /* skip check */
Oh, the bug was even documented, with its cause ;-)
That's commit 675fc275a3a2d905 ("bpf: offload: report device information
for offloaded programs").
Partially fixed by commit 36f9814a494a874d ("bpf: fix uapi hole for 32 bit
compat applications"), which left architectures with 16-bit alignment
broken...
Gr{oetje,eeting}s,
Geert
--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- [email protected]
In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
-- Linus Torvalds
Hi Baruch, Geert,
Could you share these findings with bpf and netdev people, please?
On Fri, May 03, 2019 at 02:16:04PM +0200, Geert Uytterhoeven wrote:
> Hi Baruch,
>
> On Fri, May 3, 2019 at 1:52 PM Baruch Siach <[email protected]> wrote:
> > On Fri, May 03 2019, Geert Uytterhoeven wrote:
> > > On Fri, May 3, 2019 at 6:06 AM Baruch Siach <[email protected]> wrote:
> > >> strace 5.0 fails to build for m86k/5208 with the Buildroot generated
> > >> toolchain:
> > >>
> > >> In file included from bpf_attr_check.c:6:0:
> > >> static_assert.h:20:25: error: static assertion failed: "bpf_prog_info_struct.nr_jited_ksyms offset mismatch"
> > >> # define static_assert _Static_assert
> > >> ^
> > >> bpf_attr_check.c:913:2: note: in expansion of macro ‘static_assert’
> > >> static_assert(offsetof(struct bpf_prog_info_struct, nr_jited_ksyms) == offsetof(struct bpf_prog_info, nr_jited_ksyms),
> > >> ^~~~~~~~~~~~~
> > >>
> > >> The direct cause is a difference in the hole after the gpl_compatible
> > >> field. Here is pahole output for the kernel struct (from v4.19):
> > >>
> > >> struct bpf_prog_info {
> > >> ...
> > >> __u32 ifindex; /* 80 4 */
> > >> __u32 gpl_compatible:1; /* 84: 0 4 */
> > >>
> > >> /* XXX 15 bits hole, try to pack */
> > >> /* Bitfield combined with next fields */
> > >>
> > >> __u64 netns_dev; /* 86 8 */
> > >
> > > I guess that should be "__aligned_u64 netns_dev;", to not rely on
> > > implicit alignment.
> >
> > Thanks. I can confirm that this minimal change fixes strace build:
> >
> > diff --git a/include/uapi/linux/bpf.h b/include/uapi/linux/bpf.h
> > index 929c8e537a14..709d4dddc229 100644
> > --- a/include/uapi/linux/bpf.h
> > +++ b/include/uapi/linux/bpf.h
> > @@ -2869,7 +2869,7 @@ struct bpf_prog_info {
> > char name[BPF_OBJ_NAME_LEN];
> > __u32 ifindex;
> > __u32 gpl_compatible:1;
> > - __u64 netns_dev;
> > + __aligned_u64 netns_dev;
> > __u64 netns_ino;
> > __u32 nr_jited_ksyms;
> > __u32 nr_jited_func_lens;
> >
> > Won't that break ABI compatibility for affected architectures?
>
> Yes it will. Or it may have been unusable without the fix. I don't know
> for sure.
>
> > >> And this is for the strace struct:
> > >>
> > >> struct bpf_prog_info_struct {
> > >> ...
> > >> uint32_t ifindex; /* 80 4 */
> > >> uint32_t gpl_compatible:1; /* 84: 0 4 */
> > >>
> > >> /* XXX 31 bits hole, try to pack */
> > >
> > > How come the uint64_t below is 8-byte aligned, not 2-byte aligned?
> > > Does strace use a special definition of uint64_t?
> >
> > I guess this is because of the netns_dev field definition in struct
> > bpf_prog_info_struct at bpf_attr.h:
> >
> > struct bpf_prog_info_struct {
> > ...
> > uint32_t gpl_compatible:1;
> > /*
> > * The kernel UAPI is broken by Linux commit
> > * v4.16-rc1~123^2~227^2~5^2~2 .
> > */
> > uint64_t ATTRIBUTE_ALIGNED(8) netns_dev; /* skip check */
>
> Oh, the bug was even documented, with its cause ;-)
> That's commit 675fc275a3a2d905 ("bpf: offload: report device information
> for offloaded programs").
>
> Partially fixed by commit 36f9814a494a874d ("bpf: fix uapi hole for 32 bit
> compat applications"), which left architectures with 16-bit alignment
> broken...
The offending commit seems to be the merge commit v4.18-rc1~114
that replaced "__u32 :32;" from the fix commit v4.17~4^2^2 with
"__u32 gpl_compatible:1;" from earlier commit v4.18-rc1~114^2~376^2~6.
> Gr{oetje,eeting}s,
>
> Geert
>
> --
> Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- [email protected]
>
> In personal conversations with technical people, I call myself a hacker. But
> when I'm talking to journalists I just say "programmer" or something like that.
> -- Linus Torvalds
> --
> Strace-devel mailing list
> [email protected]
> https://lists.strace.io/mailman/listinfo/strace-devel
--
ldv