On Thu, Feb 27, 2025 at 10:38:47AM -0800, Eric Biggers wrote: > > Well, unfortunately this patchset still uses sg_init_one() on the virtual > address, so AFAICS it doesn't work with arbitrary virtual addresses. Right, you can't do sg_init_one on highmem memory. But the problem here is that this pointer (the source for decompression) should never have been linear to start with. Had it been an SG list, things would have just worked. In fact, digging deeper reveals even more reasons why it should be non-linear. The object that we're decompressing is often split over two physical pages. In order to produce a linear pointer, the zsmalloc code is actually forcing a copy (incidentally, that's why an extra copy is being added in zswap because the zsmalloc copy is done with a per-CPU buffer that cannot be used with async). Had this be done with SG lists none of these issues would've existed and things would just work whether you're doing software compression or hardware offload. So I'm withdrawing my acomp patch-set because the premise was wrong. The only user for acomp actually wants SG lists. Cheers, -- Email: Herbert Xu <herbert@xxxxxxxxxxxxxxxxxxx> Home Page: http://gondor.apana.org.au/~herbert/ PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt
