From: Eric Biggers <ebiggers@google.com>

bcm963xx_nvram_checksum() was using crc32_le_combine() to update a CRC
with four zero bytes.  However, this is about 5x slower than just
CRC'ing four zero bytes in the normal way.  Just do that instead.

(We could instead make crc32_le_combine() faster on short lengths.  But
all its callers do just fine without it, so I'd like to just remove it.)

Signed-off-by: Eric Biggers <ebiggers@google.com>
---

Please consider this patch for the mips tree for 6.16.  If it doesn't
get picked up, I'll take it through the crc tree.

 include/linux/bcm963xx_nvram.h | 16 ++++++----------
 1 file changed, 6 insertions(+), 10 deletions(-)

diff --git a/include/linux/bcm963xx_nvram.h b/include/linux/bcm963xx_nvram.h
index c8c7f01159fed..48830bf180427 100644
--- a/include/linux/bcm963xx_nvram.h
+++ b/include/linux/bcm963xx_nvram.h
@@ -79,29 +79,25 @@ static inline u64 __pure bcm963xx_nvram_nand_part_size(
  */
 static int __maybe_unused bcm963xx_nvram_checksum(
 	const struct bcm963xx_nvram *nvram,
 	u32 *expected_out, u32 *actual_out)
 {
+	const u32 zero = 0;
 	u32 expected, actual;
 	size_t len;
 
 	if (nvram->version <= 4) {
 		expected = nvram->checksum_v4;
-		len = BCM963XX_NVRAM_V4_SIZE - sizeof(u32);
+		len = BCM963XX_NVRAM_V4_SIZE;
 	} else {
 		expected = nvram->checksum_v5;
-		len = BCM963XX_NVRAM_V5_SIZE - sizeof(u32);
+		len = BCM963XX_NVRAM_V5_SIZE;
 	}
 
-	/*
-	 * Calculate the CRC32 value for the nvram with a checksum value
-	 * of 0 without modifying or copying the nvram by combining:
-	 * - The CRC32 of the nvram without the checksum value
-	 * - The CRC32 of a zero checksum value (which is also 0)
-	 */
-	actual = crc32_le_combine(
-		crc32_le(~0, (u8 *)nvram, len), 0, sizeof(u32));
+	/* Calculate the CRC32 of the nvram with the checksum field set to 0. */
+	actual = crc32_le(~0, nvram, len - sizeof(u32));
+	actual = crc32_le(actual, &zero, sizeof(u32));
 
 	if (expected_out)
 		*expected_out = expected;
 
 	if (actual_out)

base-commit: 3ce9925823c7d6bb0e6eb951bf2db0e9e182582d
-- 
2.49.0

On Sun, May 11, 2025 at 11:28:36AM -0700, Eric Biggers wrote:
> From: Eric Biggers <ebiggers@google.com>
> 
> bcm963xx_nvram_checksum() was using crc32_le_combine() to update a CRC
> with four zero bytes.  However, this is about 5x slower than just
> CRC'ing four zero bytes in the normal way.  Just do that instead.
> 
> (We could instead make crc32_le_combine() faster on short lengths.  But
> all its callers do just fine without it, so I'd like to just remove it.)
> 
> Signed-off-by: Eric Biggers <ebiggers@google.com>
> ---
> 
> Please consider this patch for the mips tree for 6.16.  If it doesn't
> get picked up, I'll take it through the crc tree.
> 
>  include/linux/bcm963xx_nvram.h | 16 ++++++----------
>  1 file changed, 6 insertions(+), 10 deletions(-)

applied to mips-next

Thomas.

-- 
Crap can work. Given enough thrust pigs will fly, but it's not necessarily a
good idea.                                                [ RFC1925, 2.3 ]