| Message ID | 20230424112313.3408363-1-glider@google.com (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | string: use __builtin_memcpy() in strlcpy/strlcat | expand |
On Mon, 24 Apr 2023 at 13:23, Alexander Potapenko <glider@google.com> wrote: > > lib/string.c is built with -ffreestanding, which prevents the compiler > from replacing certain functions with calls to their library versions. > > On the other hand, this also prevents Clang and GCC from instrumenting > calls to memcpy() when building with KASAN, KCSAN or KMSAN: > - KASAN normally replaces memcpy() with __asan_memcpy() with the > additional cc-param,asan-kernel-mem-intrinsic-prefix=1; > - KCSAN and KMSAN replace memcpy() with __tsan_memcpy() and > __msan_memcpy() by default. > > To let the tools catch memory accesses from strlcpy/strlcat, replace > the calls to memcpy() with __builtin_memcpy(), which KASAN, KCSAN and > KMSAN are able to replace even in -ffreestanding mode. > > This preserves the behavior in normal builds (__builtin_memcpy() ends up > being replaced with memcpy()), and does not introduce new instrumentation > in unwanted places, as strlcpy/strlcat are already instrumented. > > Suggested-by: Marco Elver <elver@google.com> > Signed-off-by: Alexander Potapenko <glider@google.com> > Link: https://lore.kernel.org/all/20230224085942.1791837-1-elver@google.com/ Reviewed-by: Marco Elver <elver@google.com> Looks reasonable. > --- > lib/string.c | 4 ++-- > 1 file changed, 2 insertions(+), 2 deletions(-) > > diff --git a/lib/string.c b/lib/string.c > index 3d55ef8901068..be26623953d2e 100644 > --- a/lib/string.c > +++ b/lib/string.c > @@ -110,7 +110,7 @@ size_t strlcpy(char *dest, const char *src, size_t size) > > if (size) { > size_t len = (ret >= size) ? size - 1 : ret; > - memcpy(dest, src, len); > + __builtin_memcpy(dest, src, len); > dest[len] = '\0'; > } > return ret; > @@ -260,7 +260,7 @@ size_t strlcat(char *dest, const char *src, size_t count) > count -= dsize; > if (len >= count) > len = count-1; > - memcpy(dest, src, len); > + __builtin_memcpy(dest, src, len); > dest[len] = 0; > return res; > } > -- > 2.40.0.634.g4ca3ef3211-goog >
On Mon, Apr 24, 2023 at 01:23:13PM +0200, Alexander Potapenko wrote: > lib/string.c is built with -ffreestanding, which prevents the compiler > from replacing certain functions with calls to their library versions. > > On the other hand, this also prevents Clang and GCC from instrumenting > calls to memcpy() when building with KASAN, KCSAN or KMSAN: > - KASAN normally replaces memcpy() with __asan_memcpy() with the > additional cc-param,asan-kernel-mem-intrinsic-prefix=1; > - KCSAN and KMSAN replace memcpy() with __tsan_memcpy() and > __msan_memcpy() by default. > > To let the tools catch memory accesses from strlcpy/strlcat, replace > the calls to memcpy() with __builtin_memcpy(), which KASAN, KCSAN and > KMSAN are able to replace even in -ffreestanding mode. > > This preserves the behavior in normal builds (__builtin_memcpy() ends up > being replaced with memcpy()), and does not introduce new instrumentation > in unwanted places, as strlcpy/strlcat are already instrumented. > > Suggested-by: Marco Elver <elver@google.com> > Signed-off-by: Alexander Potapenko <glider@google.com> > Link: https://lore.kernel.org/all/20230224085942.1791837-1-elver@google.com/ > --- > lib/string.c | 4 ++-- > 1 file changed, 2 insertions(+), 2 deletions(-) > > diff --git a/lib/string.c b/lib/string.c > index 3d55ef8901068..be26623953d2e 100644 > --- a/lib/string.c > +++ b/lib/string.c > @@ -110,7 +110,7 @@ size_t strlcpy(char *dest, const char *src, size_t size) > > if (size) { > size_t len = (ret >= size) ? size - 1 : ret; > - memcpy(dest, src, len); > + __builtin_memcpy(dest, src, len); > dest[len] = '\0'; > } > return ret; > @@ -260,7 +260,7 @@ size_t strlcat(char *dest, const char *src, size_t count) > count -= dsize; > if (len >= count) > len = count-1; > - memcpy(dest, src, len); > + __builtin_memcpy(dest, src, len); > dest[len] = 0; > return res; I *think* this isn't a problem for CONFIG_FORTIFY, since these will be replaced and checked separately -- but it still seems strange that you need to explicitly use __builtin_memcpy. Does this end up changing fortify coverage?
> > I *think* this isn't a problem for CONFIG_FORTIFY, since these will be > replaced and checked separately -- but it still seems strange that you > need to explicitly use __builtin_memcpy. > > Does this end up changing fortify coverage? Is fortify relevant here? Note that the whole file is compiled with __NO_FORTIFY.
On Fri, Apr 28, 2023 at 3:48 PM Alexander Potapenko <glider@google.com> wrote: > >FORTIFY_SOURCE glidear > > I *think* this isn't a problem for CONFIG_FORTIFY, since these will be > > replaced and checked separately -- but it still seems strange that you > > need to explicitly use __builtin_memcpy. > > Or did you mean we'd better use __underlying_memcpy() here instead? I am a bit puzzled.
On Wed, May 10, 2023 at 9:48 AM Alexander Potapenko <glider@google.com> wrote: > > > > On Fri, Apr 28, 2023 at 3:48 PM Alexander Potapenko <glider@google.com> wrote: >> >> > I *think* this isn't a problem for CONFIG_FORTIFY, since these will be >> > replaced and checked separately -- but it still seems strange that you >> > need to explicitly use __builtin_memcpy. >> > > Or did you mean we'd better use __underlying_memcpy() here instead? I am a bit puzzled. Kees told me offline that the patch in question is fine. @Andrew, would it be possible to queue it for 6.4? -- Alexander Potapenko Software Engineer Google Germany GmbH Erika-Mann-Straße, 33 80636 München Geschäftsführer: Paul Manicle, Liana Sebastian Registergericht und -nummer: Hamburg, HRB 86891 Sitz der Gesellschaft: Hamburg
On Fri, Apr 28, 2023 at 03:48:28PM +0200, Alexander Potapenko wrote: > > > > I *think* this isn't a problem for CONFIG_FORTIFY, since these will be > > replaced and checked separately -- but it still seems strange that you > > need to explicitly use __builtin_memcpy. > > > > Does this end up changing fortify coverage? > > Is fortify relevant here? Note that the whole file is compiled with > __NO_FORTIFY. Yeah, agreed. I think I was just curious if that got verified. I'm good with this. Acked-by: Kees Cook <keescook@chromium.org>
diff --git a/lib/string.c b/lib/string.c index 3d55ef8901068..be26623953d2e 100644 --- a/lib/string.c +++ b/lib/string.c @@ -110,7 +110,7 @@ size_t strlcpy(char *dest, const char *src, size_t size) if (size) { size_t len = (ret >= size) ? size - 1 : ret; - memcpy(dest, src, len); + __builtin_memcpy(dest, src, len); dest[len] = '\0'; } return ret; @@ -260,7 +260,7 @@ size_t strlcat(char *dest, const char *src, size_t count) count -= dsize; if (len >= count) len = count-1; - memcpy(dest, src, len); + __builtin_memcpy(dest, src, len); dest[len] = 0; return res; }
lib/string.c is built with -ffreestanding, which prevents the compiler from replacing certain functions with calls to their library versions. On the other hand, this also prevents Clang and GCC from instrumenting calls to memcpy() when building with KASAN, KCSAN or KMSAN: - KASAN normally replaces memcpy() with __asan_memcpy() with the additional cc-param,asan-kernel-mem-intrinsic-prefix=1; - KCSAN and KMSAN replace memcpy() with __tsan_memcpy() and __msan_memcpy() by default. To let the tools catch memory accesses from strlcpy/strlcat, replace the calls to memcpy() with __builtin_memcpy(), which KASAN, KCSAN and KMSAN are able to replace even in -ffreestanding mode. This preserves the behavior in normal builds (__builtin_memcpy() ends up being replaced with memcpy()), and does not introduce new instrumentation in unwanted places, as strlcpy/strlcat are already instrumented. Suggested-by: Marco Elver <elver@google.com> Signed-off-by: Alexander Potapenko <glider@google.com> Link: https://lore.kernel.org/all/20230224085942.1791837-1-elver@google.com/ --- lib/string.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-)