Re: [PATCH 3/4 RESEND] crypto: octeontx2: Fix address alignment on CN10K A0/A1 and OcteonTX2

On Mon, May 19, 2025 at 11:47:18AM +0530, Bharat Bhushan wrote:

...

...

• /* Allocate extra memory for SG and response address alignment */
  

• total_mem_len = ALIGN(info_len, OTX2_CPT_DPTR_RPTR_ALIGN) + dlen;
  

This add extra memory for 8-byte (OTX2_CPT_DPTR_RPTR_ALIGN) alignment

...

...

• total_mem_len = ALIGN(total_mem_len, OTX2_CPT_RES_ADDR_ALIGN) +
  

•                  sizeof(union otx2_cpt_res_s);
  

This add extra memory for 32-byte (OTX2_CPT_RES_ADDR_ALIGN)) In case not observed, OTX2_CPT_RES_ADDR_ALIGN is not the same as OTX2_CPT_DPTR_RPTR_ALIGN.

But it doesn't do that. Look, assume that total_mem_len is 64, then ALIGN(64, 32) will still be 64. You're not adding any extra space for the alignment padding.

OTOH, kmalloc can return something that has a page offset of 8, and you will need 24 extra bytes in your structure to make it align at 32.

Now of course if you're very lucky, and total_mem_len starts out at 8, then it would work but that's purely by chance.

Cheers,