On Wed, Mar 6, 2019 at 2:57 PM Peter Zijlstra <pet...@infradead.org> wrote:
>
> On Wed, Mar 06, 2019 at 02:39:33PM +0100, Dmitry Vyukov wrote:
> > On Wed, Mar 6, 2019 at 2:13 PM Peter Zijlstra <pet...@infradead.org> wrote:
>
> > > annotated:
> > >
> > > 0000 0000000000000150 <__asan_load1>:
> > > 0000 150: 48 b8 ff ff ff ff ff movabs $0xffff7fffffffffff,%rax
> > > 0007 157: 7f ff ff
> > > 000a 15a: 53 push %rbx
> >
> > /\/\/\/\/\/\
> >
> > This push is unpleasant on hot fast path. I think we need to move
> > whole report cold path into a separate noinline function as it is now,
> > and that function will do the magic with smap. Then this won't prevent
> > tail calling and won't affect fast-path codegen.
>
> It's a bit daft of GCC to do that anyway; since it only uses that rbx
> thing in the cold path at __asan_load1+0x30.
>
> But yes, that wants fixing or something. Then again; a kernel with KASAN
> on is unbearable slow anyway.
>
> > > 000b 15b: 48 8b 4c 24 08 mov 0x8(%rsp),%rcx
> > > 0010 160: 48 39 c7 cmp %rax,%rdi
> > > 0013 163: 76 24 jbe 189 <__asan_load1+0x39>
> > > 0015 165: 48 b8 00 00 00 00 00 movabs $0xdffffc0000000000,%rax
> > > 001c 16c: fc ff df
> > > 001f 16f: 48 89 fa mov %rdi,%rdx
> > > 0022 172: 48 c1 ea 03 shr $0x3,%rdx
> > > 0026 176: 0f b6 04 02 movzbl (%rdx,%rax,1),%eax
> > > 002a 17a: 84 c0 test %al,%al
> > > 002c 17c: 75 02 jne 180 <__asan_load1+0x30>
> > > 002e 17e: 5b pop %rbx
> > > 002f 17f: c3 retq
>
> ^^^ hot path, vvv cold path
>
> > > 0030 180: 89 fa mov %edi,%edx
> > > 0032 182: 83 e2 07 and $0x7,%edx
> > > 0035 185: 38 d0 cmp %dl,%al
> > > 0037 187: 7f f5 jg 17e <__asan_load1+0x2e>
> > > 0039 189: 9c pushfq
> > > 003a 18a: 5b pop %rbx
> > > 003b 18b: 90 nop
> > > 003c 18c: 90 nop
> > > 003d 18d: 90 nop
> > > 003e 18e: 31 d2 xor %edx,%edx
> > > 0040 190: be 01 00 00 00 mov $0x1,%esi
> > > 0045 195: e8 00 00 00 00 callq 19a <__asan_load1+0x4a>
> > > 0046 196: R_X86_64_PLT32 __kasan_report-0x4
> > > 004a 19a: 53 push %rbx
> > > 004b 19b: 9d popfq
> > > 004c 19c: 5b pop %rbx
> > > 004d 19d: c3 retq
>
> > > +static __always_inline
> > > +void kasan_report(unsigned long addr, size_t size, bool is_write,
> > > unsigned long ip)
> > > +{
> > > + unsigned long flags;
> > > +
> > > + flags = smap_save();
> >
> > Previously you said that messing with smap here causes boot errors.
> > Shouldn't we do smap_save iff kasan_report_enabled? Otherwise we just
> > bail out, so no need to enable/disable smap.
> >
> > > + __kasan_report(addr, size, is_write, ip);
> > > + smap_restore(flags);
> > > +
> > > +}
>
> Ah, you think I booted this :-) Still, this is only PUSHF;CLAC, which I
> think should actually work really early. It was that #UD thing that
> didn't work early, simply because we'd not set up the exception vector
> yet when first this happens.
>
> > > --- a/mm/kasan/generic_report.c
> > > +++ b/mm/kasan/generic_report.c
> > > @@ -118,14 +118,14 @@ const char *get_bug_type(struct kasan_ac
> > > #define DEFINE_ASAN_REPORT_LOAD(size) \
> > > void __asan_report_load##size##_noabort(unsigned long addr) \
> > > { \
> > > - kasan_report(addr, size, false, _RET_IP_); \
> > > + __kasan_report(addr, size, false, _RET_IP_); \
> >
> > Unless I am missing something, this seems to make this patch no-op. We
> > fixed kasan_report for smap, but here we now use __kasan_report which
> > is not fixed. So this won't work with smap again?..
>
> I've not found callers of __asan_report_load* with AC=1 in the kernel
> yet. Under what condtions does GCC emit calls to these functions?
CONFIG_KASAN_INLINE=y
Then compiler inlines fast path into generated code and only calls
into runtime to report errors (also, faster, this should be a default
for anything other than tiny ROM controllers).
