> From: Stephen Hemminger <stephen@networkplumber.org>
> Date: Monday, 3 January 2022 at 19:05
> To: Elena Agostini <eagostini@nvidia.com>
> Cc: dev@dpdk.org <dev@dpdk.org>
> Subject: Re: [PATCH v1 3/3] gpu/cuda: mem alloc aligned memory
> External email: Use caution opening links or attachments>
>

> On Tue, 4 Jan 2022 01:47:21 +0000
> <eagostini@nvidia.com> wrote:>

> >  static int
> > -cuda_mem_alloc(struct rte_gpu *dev, size_t size, void **ptr)
> > +cuda_mem_alloc(struct rte_gpu *dev, size_t size, void **ptr, unsigned int align)
> >  {
> >       CUresult res;
> >       const char *err_string;
> > @@ -610,8 +612,10 @@ cuda_mem_alloc(struct rte_gpu *dev, size_t size, void **ptr)
> >
> >       /* Allocate memory */
> >       mem_alloc_list_tail->size = size;
> > -     res = pfn_cuMemAlloc(&(mem_alloc_list_tail->ptr_d),
> > -                     mem_alloc_list_tail->size);
> > +     mem_alloc_list_tail->size_orig = size + align;
> > +
> > +     res = pfn_cuMemAlloc(&(mem_alloc_list_tail->ptr_orig_d),
> > +                     mem_alloc_list_tail->size_orig);
> >       if (res != 0) {
> >               pfn_cuGetErrorString(res, &(err_string));
> >               rte_cuda_log(ERR, "cuCtxSetCurrent current failed with %s",
> > @@ -620,6 +624,13 @@ cuda_mem_alloc(struct rte_gpu *dev, size_t size, void **ptr)
> >               return -rte_errno;
> >       }
> >
> > +
> > +     /* Align memory address */
> > +     mem_alloc_list_tail->ptr_d = mem_alloc_list_tail->ptr_orig_d;
> > +     if (align && ((uintptr_t)mem_alloc_list_tail->ptr_d) % align)
> > +             mem_alloc_list_tail->ptr_d += (align -
> > +                             (((uintptr_t)mem_alloc_list_tail->ptr_d) % align));>
>

> Posix memalign takes size_t for both size and alignment.

I've created this patch based on the rte_malloc function definition for consistency.

void * rte_malloc(const char *type, size_t size, unsigned align)


> Better to put the input parameters first, and then the resulting output parameter last
> for consistency; follows the Rusty API design manifesto.

Got it, will do.

> Alignment only makes sense if power of two. The code should check that and optimize
> for that.
>

The alignment value is checked in the gpudev library before
passing it to the driver.

Adding this kind of checks in the driver has been rejected in the past because it was
considered dead code (the library was already checking input parameters).

Let me know what are the preferred options.