On Thu, May 15, 2025 at 11:06:56PM -0700, Christoph Hellwig wrote:
> On Thu, May 15, 2025 at 10:31:00PM -0700, Eric Biggers wrote:
> > +static inline __le32 nvme_tcp_hdgst(const void *pdu, size_t len)
> > +{
> > + return cpu_to_le32(~crc32c(NVME_TCP_CRC_SEED, pdu, len));
> > }
>
> This drops the unaligned handling. Now in the NVMe protocol it will
> always be properly aligned, but my TCP-foo is not good enough to
> remember if the networking code will also guarantee 32-bit alignment
> for the start of the packet?
>
> Otherwise this looks great:
>
> Reviewed-by: Christoph Hellwig <hch@xxxxxx>
The nvme-tcp driver already assumes that the header is at least 4-byte aligned,
considering that it accesses hdr->plen and the struct doesn't use __packed:
struct nvme_tcp_hdr {
__u8 type;
__u8 flags;
__u8 hlen;
__u8 pdo;
__le32 plen;
};
On the send size, the header size is always sizeof(struct nvme_tcp_cmd_pdu) or
sizeof(struct nvme_tcp_data_pdu) which are multiples of 4. On the receive side,
nvme-tcp validates hdr->hlen == sizeof(struct nvme_tcp_rsp_pdu) and then does:
recv_digest = *(__le32 *)(pdu + hdr->hlen);
So, using put_unaligned_le32() is unnecessary. I just had it there in v1
because I had directly translated crypto_ahash_digest(), which does ultimately
do a put_unaligned_le32() once you unravel all the API layers.
- Eric
