2021-10-10 16:01:27

by Cai,Huoqing

Subject: [PATCH] crypto: octeontx2 - Use dma_alloc_coherent() instead of kzalloc/dma_map_single()

Replacing kzalloc/kfree/dma_map_single/dma_unmap_single()
with dma_alloc_coherent/dma_free_coherent() helps to reduce
code size, and simplify the code, and coherent DMA will not
clear the cache every time.

Signed-off-by: Cai Huoqing <[email protected]>
---
drivers/crypto/marvell/octeontx2/otx2_cptpf_ucode.c | 13 ++-----------
1 file changed, 2 insertions(+), 11 deletions(-)

diff --git a/drivers/crypto/marvell/octeontx2/otx2_cptpf_ucode.c b/drivers/crypto/marvell/octeontx2/otx2_cptpf_ucode.c
index dff34b3ec09e..60d62ce049a6 100644
--- a/drivers/crypto/marvell/octeontx2/otx2_cptpf_ucode.c
+++ b/drivers/crypto/marvell/octeontx2/otx2_cptpf_ucode.c
@@ -1449,18 +1449,11 @@ int otx2_cpt_discover_eng_capabilities(struct otx2_cptpf_dev *cptpf)
compl_rlen = ALIGN(sizeof(union otx2_cpt_res_s), OTX2_CPT_DMA_MINALIGN);
len = compl_rlen + LOADFVC_RLEN;

- result = kzalloc(len, GFP_KERNEL);
+ result = dma_alloc_coherent(&pdev->dev, len, &rptr_baddr, GFP_KERNEL);
if (!result) {
ret = -ENOMEM;
goto lf_cleanup;
}
- rptr_baddr = dma_map_single(&pdev->dev, (void *)result, len,
- DMA_BIDIRECTIONAL);
- if (dma_mapping_error(&pdev->dev, rptr_baddr)) {
- dev_err(&pdev->dev, "DMA mapping failed\n");
- ret = -EFAULT;
- goto free_result;
- }
rptr = (u8 *)result + compl_rlen;

/* Fill in the command */
@@ -1489,11 +1482,9 @@ int otx2_cpt_discover_eng_capabilities(struct otx2_cptpf_dev *cptpf)

cptpf->eng_caps[etype].u = be64_to_cpup(rptr);
}
- dma_unmap_single(&pdev->dev, rptr_baddr, len, DMA_BIDIRECTIONAL);
+ dma_free_coherent(&pdev->dev, len, (void *)result, rptr_baddr);
cptpf->is_eng_caps_discovered = true;

-free_result:
- kfree(result);
lf_cleanup:
otx2_cptlf_shutdown(&cptpf->lfs);
delete_grps:
--
2.25.1

2021-10-11 08:43:18

by Christoph Hellwig

Subject: Re: [PATCH] crypto: octeontx2 - Use dma_alloc_coherent() instead of kzalloc/dma_map_single()

On Mon, Oct 11, 2021 at 12:00:04AM +0800, Cai Huoqing wrote:
> Replacing kzalloc/kfree/dma_map_single/dma_unmap_single()
> with dma_alloc_coherent/dma_free_coherent() helps to reduce
> code size, and simplify the code, and coherent DMA will not
> clear the cache every time.

As just explained in another patch like this please qualify the
savings as we have actual results that shows the reverse of your claims.