Re: [Qemu-block] [PATCH v3 5/5] migration: Use strnlen() for fixed-size string

[Eric Blake]

On 12/18/18 11:51 AM, Philippe Mathieu-Daudé wrote:
> GCC 8 introduced the -Wstringop-overflow, which detect buffer overflow
> by string-modifying functions declared in <string.h>, such strncpy(),
> used in global_state_store_running().
>
> Since the global_state.runstate does not necessarily contains a
> terminating NUL character, We had to use the QEMU_NONSTRING attribute.

s/We/we/

> The GCC manual says about the nonstring attribute:
>
>    However, when the array is declared with the attribute the call to
>    strlen is diagnosed because when the array doesn’t contain a
>    NUL-terminated string the call is undefined. [...]
>    In addition, calling strnlen and strndup with such arrays is safe
>    provided a suitable bound is specified, and not diagnosed.
>
> GCC indeed found an incorrect use of strlen(), because this array
> is loaded by VMSTATE_BUFFER(runstate, GlobalState) then parsed
> using qapi_enum_parse which does not get the buffer length.
>
> Use strnlen() which returns sizeof(s->runstate) if the array is not
> NUL-terminated.
>
> This fixes:
>
>      CC      migration/global_state.o
>    qemu/migration/global_state.c: In function 'global_state_pre_save':
>    qemu/migration/global_state.c:109:15: error: 'strlen' argument 1 declared 
> attribute 'nonstring' [-Werror=stringop-overflow=]
>         s->size = strlen((char *)s->runstate) + 1;
>                   ^~~~~~~~~~~~~~~~~~~~~~~~~~~
>    qemu/migration/global_state.c:24:13: note: argument 'runstate' declared here
>         uint8_t runstate[100] QEMU_NONSTRING;
>                 ^~~~~~~~
>    cc1: all warnings being treated as errors
>    make: *** [qemu/rules.mak:69: migration/global_state.o] Error 1
>
> Signed-off-by: Philippe Mathieu-Daudé <address@hidden>
> ---
>   migration/global_state.c | 2 +-
>   1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/migration/global_state.c b/migration/global_state.c
> index 6e19333422..c19030ef62 100644
> --- a/migration/global_state.c
> +++ b/migration/global_state.c
> @@ -106,7 +106,7 @@ static int global_state_pre_save(void *opaque)
>       GlobalState *s = opaque;
>   
>       trace_migrate_global_state_pre_save((char *)s->runstate);
> -    s->size = strlen((char *)s->runstate) + 1;

The old code sets s->size to the string length + space for the NUL byte 
(by assuming that a NUL byte was present), and accidentally sets it 
beyond the s->runstate array if there was no NUL byte (our existing 
runstate names are shorter than 100 bytes, so this could only happen on 
a malicious stream).

> +    s->size = strnlen((char *)s->runstate, sizeof(s->runstate)) + 1;

The new code can still end up setting s->size beyond the array.  Is that 
intended, or would it be better to use strnlen(s->runstate, 
sizeof(s->runstate) - 1) + 1?

Also, as I argued on 4/5, why isn't this squashed in with the patch that 
marks the field NONSTRING?

--
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3266
Virtualization:  qemu.org | libvirt.org