Re: [PATCH RFC 1/9] lib/lz4: import upstream LZ4 sources verbatim

[David Laight]

On Sun, 4 Oct 2026 01:50:39 +0200
Michal Wilczynski <m.wilczynski@xxxxxxxxxxx> wrote:

> On 9/28/26 07:19, Sergey Senozhatsky wrote:
> > On (26/09/25 13:27), Michal Wilczynski wrote:
> > [..]  
> >> +int LZ4_compress_fast_extState(void* state, const char* source, char* dest, int inputSize, int maxOutputSize, int acceleration)
> >> +{
> >> +    LZ4_stream_t_internal* const ctx = & LZ4_initStream(state, sizeof(LZ4_stream_t)) -> internal_donotuse;
> >> +    assert(ctx != NULL);  
> > 
> > Probably in some other places as well:
> > 
> > LZ4_initStream() can return NULL, but the code immediately derefs
> > the pointer.  The assert() check seems to be misplaced, unless I'm
> > missing something.
> > 
> > I suppose those need to be decoupled.  First LZ4_initStream() call and
> > NULL check, then stream->internal_donotuse.  
> 
> Good catch, this assert() is a noop in the kernel (as any build without
> LZ4_DEBUG), so a NULL, undersized or misaligned wrkmem oopses in
> LZ4_compress_fast(), and LZ4_compress_destSize() and LZ4_resetStreamHC()
> carry on with an uninitialised state.

The assert() is also just a waste of code.
Debugging the NULL pointer dereference is just as easy.

David 
 
> 
> No in-tree caller hits it (all use kmalloc/vmalloc), and I listed the
> alignment requirement as a behaviour change in the cover letter. Still,
> the vendored files need to stay verbatim.
> 
> So in v2 our own entry points can check alignment with a WARN_ON_ONCE()
> and bail out early. It might be something that we can fix upstream as
> well.
> 
> 
> >   
> 
> Best regards,