| Message ID | 20170615223543.22867-3-david.daney@cavium.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On 06/16/2017 12:35 AM, David Daney wrote: > ... this allows gating of inline assembly code that causes llvm to > fail when emitting BPF. > > Signed-off-by: David Daney <david.daney@cavium.com> I don't have a better idea at the moment, perhaps there could be a clang rewrite plugin that would ignore all inline assembly code since this is never used from BPF progs. Hmm. Really ugly that we have to add this __EMITTING_BPF__ into arch asm files, but I don't have a better idea for an immediate workaround right now ... I would really prefer if we could avoid just for the sake of the kernel samples going down the road of adding a !defined(__EMITTING_BPF__) into a uapi asm header for mips, though. Is this coming from networking sample code or rather tracing? > --- > samples/bpf/Makefile | 8 ++++---- > 1 file changed, 4 insertions(+), 4 deletions(-) > > diff --git a/samples/bpf/Makefile b/samples/bpf/Makefile > index a0561dc762fe..4979e6b56662 100644 > --- a/samples/bpf/Makefile > +++ b/samples/bpf/Makefile > @@ -193,12 +193,12 @@ $(src)/*.c: verify_target_bpf > > $(obj)/tracex5_kern.o: $(obj)/syscall_nrs.h > > -# asm/sysreg.h - inline assembly used by it is incompatible with llvm. > -# But, there is no easy way to fix it, so just exclude it since it is > -# useless for BPF samples. > +# __EMITTING_BPF__ used to exclude inline assembly, which cannot be > +# emitted in BPF code. > $(obj)/%.o: $(src)/%.c > $(CLANG) $(NOSTDINC_FLAGS) $(LINUXINCLUDE) $(EXTRA_CFLAGS) \ > - -D__KERNEL__ -D__ASM_SYSREG_H -Wno-unused-value -Wno-pointer-sign \ > + -D__KERNEL__ -D__EMITTING_BPF__ \ > + -Wno-unused-value -Wno-pointer-sign \ > -Wno-compare-distinct-pointer-types \ > -Wno-gnu-variable-sized-type-not-at-end \ > -Wno-address-of-packed-member -Wno-tautological-compare \ >
From: Daniel Borkmann <daniel@iogearbox.net> Date: Fri, 16 Jun 2017 12:24:06 +0200 > On 06/16/2017 12:35 AM, David Daney wrote: >> ... this allows gating of inline assembly code that causes llvm to >> fail when emitting BPF. >> >> Signed-off-by: David Daney <david.daney@cavium.com> > > I don't have a better idea at the moment, perhaps there could be > a clang rewrite plugin that would ignore all inline assembly code > since this is never used from BPF progs. Hmm. Really ugly that > we have to add this __EMITTING_BPF__ into arch asm files, but I > don't have a better idea for an immediate workaround right now ... > I would really prefer if we could avoid just for the sake of the > kernel samples going down the road of adding a > !defined(__EMITTING_BPF__) > into a uapi asm header for mips, though. Is this coming from > networking sample code or rather tracing? The problem is that we include the arch include/asm files at all. When we build bpf stuff, we should have a set of asm/ files specifically for builds targetting BPF. Let's just fix this right and stop putting all of these hacks in. Thank you.
On Thu, Jun 15, 2017 at 03:35:42PM -0700, David Daney wrote: > ... this allows gating of inline assembly code that causes llvm to > fail when emitting BPF. I floated essentially the same patch in Feb without much luck: http://lists.infradead.org/pipermail/linux-arm-kernel/2017-March/492758.html Many of the folks on the cc-list here had objections then and I suspect they unfortunately still object. I like the fact that this fix allowed architectures that are currently problematic to move forward before there is a possibly mythical "fix in llvm" to address this problem. When talking about this in one of the IOVisor calls it was also discussed that this needs to be fixed for tracing, so there are more than just the BPF use-case where this is important. I wasn't sure there was buy-in from the ARM developers, but my thought had been that a cleaner solution to this would be to reorganize sysreg.h into multiple files. The inline assembly would be the only thing in sysreg-asm.h (that was actually the only thing originally in sysreg.h) and the rest of the code would be in sysreg.h. That is not what Dave suggested, but it would be a good starting point for a custom asm/ layer for BPF/tracing. I'm with Dave and think a specialized set of asm/ files for tracing/BPF to avoid these issues all-together and let arch developers to do whatever they want in their code. > Signed-off-by: David Daney <david.daney@cavium.com> > --- > samples/bpf/Makefile | 8 ++++---- > 1 file changed, 4 insertions(+), 4 deletions(-) > > diff --git a/samples/bpf/Makefile b/samples/bpf/Makefile > index a0561dc762fe..4979e6b56662 100644 > --- a/samples/bpf/Makefile > +++ b/samples/bpf/Makefile > @@ -193,12 +193,12 @@ $(src)/*.c: verify_target_bpf > > $(obj)/tracex5_kern.o: $(obj)/syscall_nrs.h > > -# asm/sysreg.h - inline assembly used by it is incompatible with llvm. > -# But, there is no easy way to fix it, so just exclude it since it is > -# useless for BPF samples. > +# __EMITTING_BPF__ used to exclude inline assembly, which cannot be > +# emitted in BPF code. > $(obj)/%.o: $(src)/%.c > $(CLANG) $(NOSTDINC_FLAGS) $(LINUXINCLUDE) $(EXTRA_CFLAGS) \ > - -D__KERNEL__ -D__ASM_SYSREG_H -Wno-unused-value -Wno-pointer-sign \ > + -D__KERNEL__ -D__EMITTING_BPF__ \ > + -Wno-unused-value -Wno-pointer-sign \ > -Wno-compare-distinct-pointer-types \ > -Wno-gnu-variable-sized-type-not-at-end \ > -Wno-address-of-packed-member -Wno-tautological-compare \
diff --git a/samples/bpf/Makefile b/samples/bpf/Makefile index a0561dc762fe..4979e6b56662 100644 --- a/samples/bpf/Makefile +++ b/samples/bpf/Makefile @@ -193,12 +193,12 @@ $(src)/*.c: verify_target_bpf $(obj)/tracex5_kern.o: $(obj)/syscall_nrs.h -# asm/sysreg.h - inline assembly used by it is incompatible with llvm. -# But, there is no easy way to fix it, so just exclude it since it is -# useless for BPF samples. +# __EMITTING_BPF__ used to exclude inline assembly, which cannot be +# emitted in BPF code. $(obj)/%.o: $(src)/%.c $(CLANG) $(NOSTDINC_FLAGS) $(LINUXINCLUDE) $(EXTRA_CFLAGS) \ - -D__KERNEL__ -D__ASM_SYSREG_H -Wno-unused-value -Wno-pointer-sign \ + -D__KERNEL__ -D__EMITTING_BPF__ \ + -Wno-unused-value -Wno-pointer-sign \ -Wno-compare-distinct-pointer-types \ -Wno-gnu-variable-sized-type-not-at-end \ -Wno-address-of-packed-member -Wno-tautological-compare \
... this allows gating of inline assembly code that causes llvm to fail when emitting BPF. Signed-off-by: David Daney <david.daney@cavium.com> --- samples/bpf/Makefile | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-)