Re: [regression] oops on heavy compilations ("kernel BUG at mm/zswap.c:1005!" and "Oops: invalid opcode: 0000")

[Nhat Pham <nphamcs@gmail.com>]

On Fri, Aug 23, 2024 at 9:13 AM Matthew Wilcox <willy@infradead.org> wrote:
>
>
> That said, zswap could handle this better.  There's no need to panic the
> entire machine over being unable to read a page from swap.  Killing just
> the process that needed this page is sufficient.

Agree 100%. It is silly to kill the entire host for a swap read error,
and extra silly to kill the process because we fail to writeback - for
all we know that page might never be needed by the process again!!!

>
> Suggested patch at end after the oops.
>
> @@ -1601,6 +1613,7 @@ bool zswap_load(struct folio *folio)
>         bool swapcache = folio_test_swapcache(folio);
>         struct xarray *tree = swap_zswap_tree(swp);
>         struct zswap_entry *entry;
> +       int err;
>
>         VM_WARN_ON_ONCE(!folio_test_locked(folio));
>
> @@ -1638,10 +1651,13 @@ bool zswap_load(struct folio *folio)
>         if (!entry)
>                 return false;
>
> -       if (entry->length)
> -               zswap_decompress(entry, folio);
> -       else
> +       if (entry->length) {
> +               err = zswap_decompress(entry, folio);
> +               if (err)
> +                       return false;

Here, if zswap decompression fails and zswap load returns false, the
page_io logic will proceed as if zswap does not have the page and
reads garbage from the backing device instead. This could potentially
lead to silent data/memory corruption right? Or am I missing something
:) Maybe we could be extra careful here and treat it as if there is a
bio read error in the case zswap owns the page, but cannot decompress
it?

The rest seems solid to me :)