net: atlantic: fix ring buffer alignment
Subsystems (3)
PATCH 1/1 net: atlantic: fix ring buffer alignment
From: Arnd Bergmann <[email protected]>
clang warns about conflicting packing annotations:
drivers/net/ethernet/aquantia/atlantic/aq_ring.h:72:2: error: field within 'struct aq_ring_buff_s' is less aligned than 'union aq_ring_buff_s::(anonymous at drivers/net/ethernet/aquantia/atlantic/aq_ring.h:72:2)' and is usually due to 'struct aq_ring_buff_s' being packed, which can lead to unaligned accesses [-Werror,-Wunaligned-access]
This was originally intended to ensure the structure fits exactly into 32 bytes on 64-bit architectures, but apparently never did, and instead just produced misaligned pointers as well as forcing byte-wise access on hardware without unaligned load/store instructions.
Update the comment to more closely reflect the layout and remove the broken __packed annotation.
--- .../net/ethernet/aquantia/atlantic/aq_ring.h | 26 +++++++++++-------- 1 file changed, 15 insertions(+), 11 deletions(-)
drivers/net/ethernet/aquantia/atlantic/aq_ring.h +15 -11
@@ -26,19 +26,23 @@ struct aq_rxpage {unsigned int pg_off; }; -/* TxC SOP DX EOP -* +----------+----------+----------+----------- -* 8bytes|len l3,l4 | pa | pa | pa -* +----------+----------+----------+----------- -* 4/8bytes|len pkt |len pkt | | skb -* +----------+----------+----------+----------- -* 4/8bytes|is_gso |len,flags |len |len,is_eop -* +----------+----------+----------+----------- +/* TxC SOP DX EOP RX +* +----------+----------+----------+----------+------- +* 8bytes|len l3,l4 | pa | pa | pa | hash +* +----------+----------+----------+----------+------- +* 4/8bytes|len pkt |len pkt | | skb | page +* +----------+----------+----------+----------+------- +* 4/8bytes|is_gso |len,flags |len |len,is_eop| daddr +* +----------+----------+----------+----------+------- +* 4/8bytes| | | | | order,pgoff +* +----------+----------+----------+----------+------- +* 2bytes | | | | | vlan_rx_tag +* +----------+----------+----------+----------+------- +* 8bytes + flags +* +----------+----------+----------+----------+------- * -* This aq_ring_buff_s doesn't have endianness dependency. -* It is __packed for cache line optimizations. */ -struct __packed aq_ring_buff_s { +struct aq_ring_buff_s {union {/* RX/TX */ dma_addr_t pa;
Signature · 1 line
2.39.2
Quoted 5 lines · Arnd Bergmann
From: Arnd Bergmann <[email protected]>
clang warns about conflicting packing annotations:
drivers/net/ethernet/aquantia/atlantic/aq_ring.h:72:2: error: field within 'structaq_ring_buff_s' is less aligned than 'union aq_ring_buff_s::(anonymous atdrivers/net/ethernet/aquantia/atlantic/aq_ring.h:72:2)' and is usually due to 'structaq_ring_buff_s' being packed, which can lead to unaligned accesses [-Werror,-Wunaligned-access]
FWIIW, I was able to reproduce this (warning) with clang-16 on ARM (32bit). And I agree with the approach you have taken here.
Quoted 33 lines
This was originally intended to ensure the structure fits exactly into 32 bytes on 64-bit architectures, but apparently never did, and instead just produced misaligned pointers as well as forcing byte-wise access on hardware without unaligned load/store instructions.
Update the comment to more closely reflect the layout and remove the broken
__packedannotation.Signed-off-by Arnd Bergmann <[email protected]>--- .../net/ethernet/aquantia/atlantic/aq_ring.h | 26 +++++++++++-------- 1 file changed, 15 insertions(+), 11 deletions(-)drivers/net/ethernet/aquantia/atlantic/aq_ring.h +4
-8@@ -26,19 +26,23 @@ struct aq_rxpage {unsigned int pg_off; }; -/* TxC SOP DX EOP -* +----------+----------+----------+----------- -* 8bytes|len l3,l4 | pa | pa | pa -* +----------+----------+----------+----------- -* 4/8bytes|len pkt |len pkt | | skb -* +----------+----------+----------+----------- -* 4/8bytes|is_gso |len,flags |len |len,is_eop -* +----------+----------+----------+----------- +/* TxC SOP DX EOP RX +* +----------+----------+----------+----------+------- +* 8bytes|len l3,l4 | pa | pa | pa | hash +* +----------+----------+----------+----------+-------
drivers/net/ethernet/aquantia/atlantic/aq_ring.h struct aq_rxpage { · line 42
Quoted 10 lines
drivers/net/ethernet/aquantia/atlantic/aq_ring.h +10
-0+* 4/8bytes|len pkt |len pkt | | skb | page +* +----------+----------+----------+----------+------- +* 4/8bytes|is_gso |len,flags |len |len,is_eop| daddr +* +----------+----------+----------+----------+------- +* 4/8bytes| | | | | order,pgoff +* +----------+----------+----------+----------+------- +* 2bytes | | | | | vlan_rx_tag +* +----------+----------+----------+----------+------- +* 8bytes + flags +* +----------+----------+----------+----------+-------
Perhaps it just me. But I do have trouble reconciling the description above with the structure below. As such, my suggest would be to simply delete it.
Quoted 9 lines
drivers/net/ethernet/aquantia/atlantic/aq_ring.h +1
-3* -* This aq_ring_buff_s doesn't have endianness dependency. -* It is __packed for cache line optimizations. */ -struct __packed aq_ring_buff_s { +struct aq_ring_buff_s {union {/* RX/TX */ dma_addr_t pa;
Jakub Kicinski
maintainerQuoted 3 lines · On Fri, 16 Jun 2023 15:10:03 +0200 Simon Horman wrote:
Perhaps it just me. But I do have trouble reconciling the description above with the structure below. As such, my suggest would be to simply delete it.
Agreed that the comment is confusing seems to be incorrect post-change. Flags for instance are overlapped with len, is_gso etc. so they can't be a separate 8B at the end.
Igor, please advise how you want to proceed.
Igor Russkikh
Acked-byHi Jakub, Arnd,
Quoted 10 lines · On Fri, 16 Jun 2023 15:10:03 +0200 Simon Horman wrote:
Quoted 3 lines
Perhaps it just me. But I do have trouble reconciling the description above with the structure below. As such, my suggest would be to simply delete it.
Agreed that the comment is confusing seems to be incorrect post-change. Flags for instance are overlapped with len,
is_gsoetc. so they can't be a separate 8B at the end.Igor, please advise how you want to proceed.
I do agree better to remove the comment at all - it explains almost nothing.
Thats not a hardware related structure, so explicit pack is for sure not required.
Regards,
Igor
Jakub Kicinski
maintainerQuoted 11 lines · On Thu, 22 Jun 2023 07:49:52 +0200 Igor Russkikh wrote:
Quoted 5 lines
Agreed that the comment is confusing seems to be incorrect post-change. Flags for instance are overlapped with len,
is_gsoetc. so they can't be a separate 8B at the end.Igor, please advise how you want to proceed.
I do agree better to remove the comment at all - it explains almost nothing.
Thats not a hardware related structure, so explicit pack is for sure not required.
Acked-by Igor Russkikh <[email protected]>
Thanks Igor!
Arnd, could you respin with the comment removed?