Re: [PATCH] zsmalloc: introduce SG-list based object read API

["Yosry Ahmed" <yosry.ahmed@linux.dev>]

January 16, 2026 at 6:39 PM, "Sergey Senozhatsky" <senozhatsky@chromium.org> wrote:


> 
> On (26/01/16 19:53), Yosry Ahmed wrote:
> [..]
> 
> > 
> > struct zs_pool *zs_create_pool(const char *name);
> >  void zs_destroy_pool(struct zs_pool *pool);
> >  @@ -43,6 +44,9 @@ void *zs_obj_read_begin(struct zs_pool *pool, unsigned long handle,
> >  size_t mem_len, void *local_copy);
> >  void zs_obj_read_end(struct zs_pool *pool, unsigned long handle,
> >  size_t mem_len, void *handle_mem);
> >  +int zs_obj_read_sg_begin(struct zs_pool *pool, unsigned long handle,
> >  
> >  void? The return value is always 0.
> > 
> I thought about returning sg_nents(). Probably can be just void after all.

We don't need it for zswap but it could be useful for zram.

> 
> > 
> > There is a lot of duplication between this and zs_obj_read_begin(). I
> >  wanted to create a common helper for them both that returns the zpdesc
> >  and offset, but we cannot do the same on the read end side as the unlock
> >  needs to happen after kunmap() in zs_obj_read_end().
> >  
> >  Putting parts of this code in helpers makes it a bit obscure due to the
> >  locking rules :/
> >  
> >  I wonder if we can drop zs_obj_read_*() and move the spanning logic into
> >  zram. Looking at zram code, seems like read_from_zspool_raw() and
> >  read_incompressible_page() just copy the return address, so I think they
> >  can trivially move to using the SG list helpers and
> >  memcpy_from_sglist().
> >  
> >  The only non-trivial caller is read_compressed_page(), because it passes
> >  the compressed object to zcomp. So I think we only need to handle the
> >  linearization there, something like this (completely untested):
> > 
> So I was thinking about leaving things as they currently are for this
> dev cycle, because both zram and zsmalloc have enough of new code queued
> up. If you don't mind let's remove memcpy() API and convert zram during
> next cycle (after the upcoming merge window).

Sure. I think we can do all of it in a single series for the next cycle. Add SG interfaces, convert zswap and zram, and remove the old interfaces.