On Mon, Jul 26, 2021 at 2:48 PM Corentin Labbe
<clabbe.montjoie@xxxxxxxxx> wrote:
>
> Le Mon, Jul 26, 2021 at 02:27:50PM +0800, Dongliang Mu a écrit :
> > This patch fixes some memory leak caused by dma_mmap_sg/single
> > in the error handling code. In addition, it fixes the return value
> > when errors related with dma_mmap_sg/single occur.
> >
> > Reported-by: Dongliang Mu <mudongliangabcd@xxxxxxxxx>
> > Fixes: 732b764099f65 ("crypto: sun8i-ce - fix two error path's memory leak")
> > Signed-off-by: Dongliang Mu <mudongliangabcd@xxxxxxxxx>
> > ---
> > .../crypto/allwinner/sun8i-ss/sun8i-ss-hash.c | 37 ++++++++++---------
> > 1 file changed, 20 insertions(+), 17 deletions(-)
> >
> > diff --git a/drivers/crypto/allwinner/sun8i-ss/sun8i-ss-hash.c b/drivers/crypto/allwinner/sun8i-ss/sun8i-ss-hash.c
> > index 3c073eb3db03..7c4ed19f5466 100644
> > --- a/drivers/crypto/allwinner/sun8i-ss/sun8i-ss-hash.c
> > +++ b/drivers/crypto/allwinner/sun8i-ss/sun8i-ss-hash.c
> > @@ -324,11 +324,11 @@ int sun8i_ss_hash_run(struct crypto_engine *engine, void *breq)
> > struct sun8i_ss_alg_template *algt;
> > struct sun8i_ss_dev *ss;
> > struct scatterlist *sg;
> > - int nr_sgs, err, digestsize;
> > + int j, i, todo, nr_sgs, tmp_err, digestsize;
> > + int err = 0;
> > unsigned int len;
> > u64 fill, min_fill, byte_count;
> > void *pad, *result;
> > - int j, i, todo;
> > __be64 *bebits;
> > __le64 *lebits;
> > dma_addr_t addr_res, addr_pad;
> > @@ -368,14 +368,14 @@ int sun8i_ss_hash_run(struct crypto_engine *engine, void *breq)
> > if (nr_sgs <= 0 || nr_sgs > MAX_SG) {
> > dev_err(ss->dev, "Invalid sg number %d\n", nr_sgs);
> > err = -EINVAL;
> > - goto theend;
> > + goto err_result;
> > }
> >
> > addr_res = dma_map_single(ss->dev, result, digestsize, DMA_FROM_DEVICE);
> > if (dma_mapping_error(ss->dev, addr_res)) {
> > dev_err(ss->dev, "DMA map dest\n");
> > err = -EINVAL;
> > - goto theend;
> > + goto err_unmap_sg;
> > }
> >
> > len = areq->nbytes;
> > @@ -390,7 +390,7 @@ int sun8i_ss_hash_run(struct crypto_engine *engine, void *breq)
> > if (len > 0) {
> > dev_err(ss->dev, "remaining len %d\n", len);
> > err = -EINVAL;
> > - goto theend;
> > + goto err_addr_res;
> > }
> >
> > byte_count = areq->nbytes;
> > @@ -421,27 +421,30 @@ int sun8i_ss_hash_run(struct crypto_engine *engine, void *breq)
> > }
> >
> > addr_pad = dma_map_single(ss->dev, pad, j * 4, DMA_TO_DEVICE);
> > - rctx->t_src[i].addr = addr_pad;
> > - rctx->t_src[i].len = j;
> > - rctx->t_dst[i].addr = addr_res;
> > - rctx->t_dst[i].len = digestsize / 4;
> > if (dma_mapping_error(ss->dev, addr_pad)) {
> > dev_err(ss->dev, "DMA error on padding SG\n");
> > err = -EINVAL;
> > - goto theend;
> > + goto err_addr_res;
> > }
> > + rctx->t_src[i].addr = addr_pad;
> > + rctx->t_src[i].len = j;
> > + rctx->t_dst[i].addr = addr_res;
> > + rctx->t_dst[i].len = digestsize / 4;
> >
> > - err = sun8i_ss_run_hash_task(ss, rctx, crypto_tfm_alg_name(areq->base.tfm));
> > + tmp_err = sun8i_ss_run_hash_task(ss, rctx, crypto_tfm_alg_name(areq->base.tfm));
> > +
> > + memcpy(areq->result, result, algt->alg.hash.halg.digestsize);
> > +
> > + crypto_finalize_hash_request(engine, breq, tmp_err);
> >
> > dma_unmap_single(ss->dev, addr_pad, j * 4, DMA_TO_DEVICE);
> > +err_addr_res:
> > + dma_unmap_single(ss->dev, addr_res, digestsize, DMA_FROM_DEVICE);
> > +err_unmap_sg:
> > dma_unmap_sg(ss->dev, areq->src, sg_nents(areq->src),
> > DMA_TO_DEVICE);
> > - dma_unmap_single(ss->dev, addr_res, digestsize, DMA_FROM_DEVICE);
> > -
> > - memcpy(areq->result, result, algt->alg.hash.halg.digestsize);
> > -theend:
> > +err_result:
> > kfree(pad);
> > kfree(result);
> > - crypto_finalize_hash_request(engine, breq, err);
> > - return 0;
> > + return err;
> > }
>
> Hello
>
> This is wrong, you need to always call crypto_finalize_hash_request()
>
> Do you have tested your changes ? Copying results before dma_unmap() is also very wrong, this will lead to random cache fail.
I am sorry I don't test my changes due to no PoC or input at hand. If
you know any PoC for me to test, please let me know. Thanks in
advance.
I am not familiar with this code. All my intention is to fix the leak
caused by the incorrect error handling code.
Your reply indicated two pieces of information:
1. crypto_finalize_hash_request must be executed and placed at the end
2. memcpy must be executed after dma_unmap_*() functions.
Any other information I missed? I can address those issues and send a v2 patch.
>
> Regards
