Re: [PATCH 2/4] mm/slub: allow to set node and align in k[v]realloc

[Uladzislau Rezki <urezki@gmail.com>]

On Fri, Jun 27, 2025 at 02:01:23PM +0200, Vitaly Wool wrote:
> 
> 
> > On Jun 27, 2025, at 1:42 PM, Uladzislau Rezki <urezki@gmail.com> wrote:
> > 
> > Hello, Vitaly, Danilo.
> > 
> >> On Fri, Jun 27, 2025 at 11:37:14AM +0200, Vitaly Wool wrote:
> >>> Reimplement k[v]realloc_node() to be able to set node and
> >>> alignment should a user need to do so. In order to do that while
> >>> retaining the maximal backward compatibility, the following rules
> >>> are honored:
> >>> * kmalloc/kzalloc/krealloc remain unchanged
> >>> * kvmalloc/kvrealloc/kvcalloc remain unchanged
> >>> * kvrealloc remains unchanged
> >>> * kvrealloc_node is implemented as a new function taking align and
> >>>  NUMA id as extra parameters compared to kvrealloc.
> >>> * krealloc_node is implemented as a new function taking NUMA id
> >>>  as an extra parameter compared to krealloc
> >>> * kvmalloc_node/kvzalloc_node/kvcalloc_node get an extra parameter
> >>>  (alignment)
> >> 
> >> I see what you're doing here:
> >> 
> >> You created vrealloc_node_noprof() in the previous patch, taking the following
> >> arguments:
> >> 
> >> vrealloc_node_noprof(const void *p, size_t size,
> >>      unsigned long align,
> >>      gfp_t flags, int nid)
> >> 
> >> And now you're aligning the newly introduced krealloc_node() and
> >> kvrealloc_node() with that.
> >> 
> >> The idea for having an align argument on krealloc_node() simply is that it
> >> fails if the alignment requirement can't be fulfilled by the corresponding
> >> kmalloc bucket, such that we can fall back to vrealloc_node() in
> >> kvrealloc_node().
> >> 
> >> Generally, this makes sense to me.
> >> 
> >> However, now you consequently have to add the align argument to kvmalloc_node(),
> >> kvzalloc_node(), kvcalloc_node() as well.
> >> 
> >> This is what creates this huge diffstat changing all the users.
> >> 
> >> IMHO, the problem here was introduced already with vrealloc_node_noprof() taking
> >> an align argument in your previous patch, since now you have to adjust
> >> everything else to logically follow the same naming scheme.
> >> 
> >> Instead, I think you should introduce vrealloc_node_align(),
> >> 
> > I am probably missing something. Could you please clarify why do you
> > need the vrealloc_node_align() and other friends? Do you have users
> > which require vrealloc() or kvrealloc() to support nid and align from
> > Rust API point of view? 
> > 
> > 
> 
> Alignment for Rust allocators should generally be supported, it’s been listed as TODO for a while.
> Node awareness is very desirable for e.g. KVBox and for the coming zpool mapping.
> 
Thank you for clarification. The comment about TODO i saw. Initially i
thought that there is a real demand in having it for some users. But
this is just for future potential ones. So, maybe the comment should
be removed instead? But i do not have a strong opinion here.

vrealloc_node_align() sounds indeed better for that purpose.

--
Uladzislau Rezki