Re: [PATCH 4/4] btrfs: introduce btrfs_bio::async_csum

[Eric Biggers <ebiggers@kernel.org>]

On Sat, Oct 25, 2025 at 08:45:09AM +1030, Qu Wenruo wrote:
> Even if the new API can cause black magic to make CRC32C to be faster than
> DRAM bandwidth, it will not remove the latency.

For what it's worth, crc32c checksumming is often faster than DRAM
already.  Though, being faster than DRAM is still useful when the data
is already in cache.

For example, on AMD Ryzen 9 9950X (Zen 5), I get 89 GB/s crc32c with the
kernel's current crc32c code for data that's already in L1 cache.

However, Zen 5 tripled the number of ALUs that can execute the crc32
instruction, resulting in new code being optimal.  I've tested that
130 GB/s crc32c is theoretically possible with Zen 5 optimized code.

The indirect calls and other overhead in the traditional crypto API
makes a notable difference in checksumming throughput when the actual
calculation is this fast...

But absolutely, the real bottleneck for I/O is almost always going to be
elsewhere in the stack.

- Eric