Re: [PATCH] soc/fsl: qbman: fix conflicting alignment attributes

On Fri, Mar 26, 2021 at 3:17 AM Li Yang [off-list ref] wrote:

On Tue, Mar 23, 2021 at 8:17 AM Arnd Bergmann [off-list ref] wrote:

quoted

From: Arnd Bergmann <arnd@arndb.de>

When building with W=1, gcc points out that the __packed attribute
on struct qm_eqcr_entry conflicts with the 8-byte alignment
attribute on struct qm_fd inside it:

drivers/soc/fsl/qbman/qman.c:189:1: error: alignment 1 of 'struct qm_eqcr_entry' is less than 8 [-Werror=packed-not-aligned]

I assume that the alignment attribute is the correct one, and
that qm_eqcr_entry cannot actually be unaligned in memory,
so add the same alignment on the outer struct.

Fixes: c535e923bb97 ("soc/fsl: Introduce DPAA 1.x QMan device driver")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
---
 drivers/soc/fsl/qbman/qman.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/drivers/soc/fsl/qbman/qman.c b/drivers/soc/fsl/qbman/qman.c
index a1b9be1d105a..fde4edd83c14 100644
--- a/drivers/soc/fsl/qbman/qman.c
+++ b/drivers/soc/fsl/qbman/qman.c

@@ -186,7 +186,7 @@ struct qm_eqcr_entry {
        __be32 tag;
        struct qm_fd fd;
        u8 __reserved3[32];
-} __packed;
+} __packed __aligned(8);

The EQCR structure is actually aligned on 64-byte from the manual.
But probably 8 is enough to let the compiler not complain.

The important bit is that all members inside are at most 8-byte long
and naturally aligned, so the compiler can now optimize the accesses
because it can assume that each member can be updated atomically.

Marking a structure as unaligned with the __packed attribute can lead
to rather unoptimized code, which is particularly important when the
structure is mapped as uncached -- accessing it one byte at a time
means you have eight times the latency for a simple read. I think on
arm64, the compiler doesn't actually have to do that, but at least if you
run a 32-bit kernel, the packing would prevent the use of ldrd or ldm
instructions that need an aligned word.

       Arnd