Re: [PATCH net 1/4] virtchnl: Fix layout of RSS structures

From: Samudrala, Sridhar
Date: Sun Mar 28 2021 - 23:20:08 EST


On 3/27/2021 2:53 AM, Geert Uytterhoeven wrote:
Hi Samudrala,

On Fri, Mar 26, 2021 at 11:45 PM Samudrala, Sridhar
<sridhar.samudrala@xxxxxxxxx> wrote:
On 3/26/2021 1:06 AM, Geert Uytterhoeven wrote:
On Thu, Mar 25, 2021 at 11:29 PM Tony Nguyen <anthony.l.nguyen@xxxxxxxxx> wrote:
From: Norbert Ciosek <norbertx.ciosek@xxxxxxxxx>

Remove padding from RSS structures. Previous layout
could lead to unwanted compiler optimizations
in loops when iterating over key and lut arrays.

From an earlier private conversation with Mateusz, I understand the real
explanation is that key[] and lut[] must be at the end of the
structures, because they are used as flexible array members?

Fixes: 65ece6de0114 ("virtchnl: Add missing explicit padding to structures")
Signed-off-by: Norbert Ciosek <norbertx.ciosek@xxxxxxxxx>
Tested-by: Konrad Jankowski <konrad0.jankowski@xxxxxxxxx>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@xxxxxxxxx>

--- a/include/linux/avf/virtchnl.h
+++ b/include/linux/avf/virtchnl.h
@@ -476,7 +476,6 @@ struct virtchnl_rss_key {
u16 vsi_id;
u16 key_len;
u8 key[1]; /* RSS hash key, packed bytes */
- u8 pad[1];
};

VIRTCHNL_CHECK_STRUCT_LEN(6, virtchnl_rss_key);
@@ -485,7 +484,6 @@ struct virtchnl_rss_lut {
u16 vsi_id;
u16 lut_entries;
u8 lut[1]; /* RSS lookup table */
- u8 pad[1];
};

If you use a flexible array member, it should be declared without a size,
i.e.

u8 key[];

Everything else is (trying to) fool the compiler, and leading to undefined
behavior, and people (re)adding explicit padding.
This header file is shared across other OSes that use C++ that doesn't support
flexible arrays. So the structures in this file use an array of size 1 as a last
element to enable variable sized arrays.
I don't think it is accepted practice to have non-Linux-isms in
include/*linux*/avf/virtchnl.h header files. Moreover, using a size
of 1 is counter-intuitive for people used to Linux kernel development,
and may lead to off-by-one errors in calculation of sizes.

If you insist on ignoring the above, this definitely deserves a
comment next to the member's declaration.
Sure. We can add a comment indicating that these fields are used variable sized arrays.

Thanks
Sridhar