On 09/13, Mina Almasry wrote:
> Building net-next with powerpc with GCC 14 compiler results in this
> build error:
>
> /home/sfr/next/tmp/ccuSzwiR.s: Assembler messages:
> /home/sfr/next/tmp/ccuSzwiR.s:2579: Error: operand out of domain (39 is
> not a multiple of 4)
> make[5]: *** [/home/sfr/next/next/scripts/Makefile.build:229:
> net/core/page_pool.o] Error 1
>
> Root caused in this thread:
> https://lore.kernel.org/netdev/913e2fbd-d318-4c9b-aed2-4d333a1d5...@cs-soprasteria.com/
>
> We try to access offset 40 in the pointer returned by this function:
>
> static inline unsigned long _compound_head(const struct page *page)
> {
> unsigned long head = READ_ONCE(page->compound_head);
>
> if (unlikely(head & 1))
> return head - 1;
> return (unsigned long)page_fixed_fake_head(page);
> }
>
> The GCC 14 (but not 11) compiler optimizes this by doing:
>
> ld page + 39
>
> Rather than:
>
> ld (page - 1) + 40
>
> And causing an unaligned load. Get around this by issuing a READ_ONCE as
> we convert the page to netmem. That disables the compiler optimizing the
> load in this way.
>
> Cc: Simon Horman <ho...@kernel.org>
> Cc: Stephen Rothwell <s...@canb.auug.org.au>
> Cc: Jakub Kicinski <k...@kernel.org>
> Cc: David Miller <da...@davemloft.net>
> Cc: Paolo Abeni <pab...@redhat.com>
> Cc: Networking <net...@vger.kernel.org>
> Cc: Linux Kernel Mailing List <linux-ker...@vger.kernel.org>
> Cc: Linux Next Mailing List <linux-n...@vger.kernel.org>
> Cc: Arnd Bergmann <a...@arndb.de>
> Cc: "linuxppc-dev@lists.ozlabs.org" <linuxppc-dev@lists.ozlabs.org>
> Cc: Matthew Wilcox <wi...@infradead.org>
>
> Suggested-by: Jakub Kicinski <k...@kernel.org>
> Signed-off-by: Mina Almasry <almasrym...@google.com>
>
> ---
>
> v2:
> https://lore.kernel.org/netdev/20240913192036.3289003-1-almasrym...@google.com/
>
> - Work around this issue as we convert the page to netmem, instead of
> a generic change that affects compound_head().
> ---
> net/core/page_pool.c | 15 ++++++++++++++-
> 1 file changed, 14 insertions(+), 1 deletion(-)
>
> diff --git a/net/core/page_pool.c b/net/core/page_pool.c
> index a813d30d2135..74ea491d0ab2 100644
> --- a/net/core/page_pool.c
> +++ b/net/core/page_pool.c
> @@ -859,12 +859,25 @@ void page_pool_put_page_bulk(struct page_pool *pool,
> void **data,
> {
> int i, bulk_len = 0;
> bool allow_direct;
> + netmem_ref netmem;
> + struct page *page;
> bool in_softirq;
>
> allow_direct = page_pool_napi_local(pool);
>
> for (i = 0; i < count; i++) {
> - netmem_ref netmem = page_to_netmem(virt_to_head_page(data[i]));
> + page = virt_to_head_page(data[i]);
> +
> + /* GCC 14 powerpc compiler will optimize reads into the
> + * resulting netmem_ref into unaligned reads as it sees address
> + * arithmetic in _compound_head() call that the page has come
> + * from.
> + *
> + * The READ_ONCE here gets around that by breaking the
> + * optimization chain between the address arithmetic and later
> + * indexing.
> + */
> + netmem = page_to_netmem(READ_ONCE(page));
>
> /* It is not the last user for the page frag case */
> if (!page_pool_is_last_ref(netmem))
Are we sure this is the only place where we can hit by this?
Any reason not to hide this inside page_to_netmem?
diff --git a/include/net/netmem.h b/include/net/netmem.h
index 8a6e20be4b9d..46bc362acec4 100644
--- a/include/net/netmem.h
+++ b/include/net/netmem.h
@@ -100,7 +100,7 @@ static inline netmem_ref net_iov_to_netmem(struct net_iov
*niov)
static inline netmem_ref page_to_netmem(struct page *page)
{
- return (__force netmem_ref)page;
+ return (__force netmem_ref)READ_ONCE(page);
}
static inline int netmem_ref_count(netmem_ref netmem)
Is it gonna generate slower code elsewhere?
