2015-04-02 19:12:00

by Corentin Labbe

[permalink] [raw]
Subject: Re: [PATCH v6 4/4] crypto: Add Allwinner Security System crypto accelerator

Le 26/03/2015 19:31, Boris Brezillon a ?crit :
> Hi Corentin,
>
> Here is a quick review, there surely are a lot of other things I didn't
> spot.
>
> On Mon, 16 Mar 2015 20:01:22 +0100
> LABBE Corentin <[email protected]> wrote:
>
>> Add support for the Security System included in Allwinner SoC A20.
>> The Security System is a hardware cryptographic accelerator that support:
>> - MD5 and SHA1 hash algorithms
>> - AES block cipher in CBC mode with 128/196/256bits keys.
>> - DES and 3DES block cipher in CBC mode
>>
>> Signed-off-by: LABBE Corentin <[email protected]>
>> ---
>
> [...]
>
>> +static int sunxi_ss_cipher(struct ablkcipher_request *areq, u32 mode)
>> +{
>> + struct crypto_ablkcipher *tfm = crypto_ablkcipher_reqtfm(areq);
>> + struct sunxi_tfm_ctx *op = crypto_ablkcipher_ctx(tfm);
>> + const char *cipher_type;
>> + struct sunxi_ss_ctx *ss = op->ss;
>> +
>> + if (areq->nbytes == 0)
>> + return 0;
>> +
>> + if (areq->info == NULL) {
>> + dev_err(ss->dev, "ERROR: Empty IV\n");
>> + return -EINVAL;
>> + }
>> +
>> + if (areq->src == NULL || areq->dst == NULL) {
>> + dev_err(ss->dev, "ERROR: Some SGs are NULL\n");
>> + return -EINVAL;
>> + }
>> +
>> + cipher_type = crypto_tfm_alg_name(crypto_ablkcipher_tfm(tfm));
>> +
>> + if (strcmp("cbc(aes)", cipher_type) == 0) {
>> + mode |= SS_OP_AES | SS_CBC | SS_ENABLED | op->keymode;
>> + return sunxi_ss_aes_poll(areq, mode);
>> + }
>> +
>> + if (strcmp("cbc(des)", cipher_type) == 0) {
>> + mode |= SS_OP_DES | SS_CBC | SS_ENABLED | op->keymode;
>> + return sunxi_ss_des_poll(areq, mode);
>> + }
>> +
>> + if (strcmp("cbc(des3_ede)", cipher_type) == 0) {
>> + mode |= SS_OP_3DES | SS_CBC | SS_ENABLED | op->keymode;
>> + return sunxi_ss_des_poll(areq, mode);
>> + }
>
> Hm, I'm not sure doing these string comparisons in the crypto operation
> path is a good idea.
> Moreover, you're doing 3 string comparisons, even though only one can
> be valid at a time (using 'else if' would have been a bit better).
>
> Anyway, IMHO this function should be split into 3 functions, and
> referenced by your alg template definitions.
> Something like:
>
> int sunxi_ss_xxx_encrypt(struct ablkcipher_request *areq)
> {
> /* put your cipher specific intialization here */
>
> return sunxi_ss_xxx_poll(areq, SS_ENCRYPTION);
> }
>
> int sunxi_ss_xxx_decrypt(struct ablkcipher_request *areq)
> {
> /* put your cipher specific intialization here */
>
> return sunxi_ss_xxx_poll(areq, SS_DECRYPTION);
> }
>

You are right, I have added more block mode, and that was added too much strcmp.
So splitting in simplier functions is good.

>
>> +
>> +int sunxi_ss_cipher_init(struct crypto_tfm *tfm)
>> +{
>> + const char *name = crypto_tfm_alg_name(tfm);
>> + struct sunxi_tfm_ctx *op = crypto_tfm_ctx(tfm);
>> + struct crypto_alg *alg = tfm->__crt_alg;
>> + struct sunxi_ss_alg_template *algt;
>> + struct sunxi_ss_ctx *ss;
>> +
>> + memset(op, 0, sizeof(struct sunxi_tfm_ctx));
>> +
>> + algt = container_of(alg, struct sunxi_ss_alg_template, alg.crypto);
>> + ss = algt->ss;
>> + op->ss = algt->ss;
>> +
>> + /* fallback is needed only for DES/3DES */
>> + if (strcmp("cbc(des)", name) == 0 ||
>> + strcmp("cbc(des3_ede)", name) == 0) {
>> + op->fallback = crypto_alloc_ablkcipher(name, 0,
>> + CRYPTO_ALG_ASYNC | CRYPTO_ALG_NEED_FALLBACK);
>> + if (IS_ERR(op->fallback)) {
>> + dev_err(ss->dev, "ERROR: allocating fallback algo %s\n",
>> + name);
>> + return PTR_ERR(op->fallback);
>> + }
>> + }
>
> Ditto: just create a specific init function for the des case:
>
> int sunxi_ss_cbc_des_init(struct crypto_tfm *tfm)
> {
> sunxi_ss_cipher_init(tfm);
>
> op->fallback = crypto_alloc_ablkcipher(name, 0,
> CRYPTO_ALG_ASYNC |
> CRYPTO_ALG_NEED_FALLBACK);
> if (IS_ERR(op->fallback)) {
> dev_err(ss->dev, "ERROR: allocating fallback algo %s\n",
> name);
> return PTR_ERR(op->fallback);
> }
>
> return 0;
> }
>

Ok

>
> [..]
>
>> +/*
>> + * Optimized function for the case where we have only one SG,
>> + * so we can use kmap_atomic
>> + */
>> +static int sunxi_ss_aes_poll_atomic(struct ablkcipher_request *areq)
>> +{
>> + u32 spaces;
>> + struct scatterlist *in_sg = areq->src;
>> + struct scatterlist *out_sg = areq->dst;
>> + void *src_addr;
>> + void *dst_addr;
>> + unsigned int ileft = areq->nbytes;
>> + unsigned int oleft = areq->nbytes;
>> + unsigned int todo;
>> + u32 *src32;
>> + u32 *dst32;
>> + u32 rx_cnt = 32;
>> + u32 tx_cnt = 0;
>> + int i;
>> + struct crypto_ablkcipher *tfm = crypto_ablkcipher_reqtfm(areq);
>> + struct sunxi_tfm_ctx *op = crypto_ablkcipher_ctx(tfm);
>> + struct sunxi_ss_ctx *ss = op->ss;
>> +
>> + src_addr = kmap_atomic(sg_page(in_sg)) + in_sg->offset;
>> + if (src_addr == NULL) {
>> + dev_err(ss->dev, "kmap_atomic error for src SG\n");
>> + return -EINVAL;
>> + }
>> +
>> + dst_addr = kmap_atomic(sg_page(out_sg)) + out_sg->offset;
>> + if (dst_addr == NULL) {
>> + dev_err(ss->dev, "kmap_atomic error for dst SG\n");
>> + kunmap_atomic(src_addr);
>> + return -EINVAL;
>> + }
>> +
>> + src32 = (u32 *)src_addr;
>> + dst32 = (u32 *)dst_addr;
>> + ileft = areq->nbytes / 4;
>> + oleft = areq->nbytes / 4;
>> + i = 0;
>> + do {
>> + if (ileft > 0 && rx_cnt > 0) {
>> + todo = min(rx_cnt, ileft);
>> + ileft -= todo;
>> + writesl(ss->base + SS_RXFIFO, src32, todo);
>> + src32 += todo;
>> + }
>> + if (tx_cnt > 0) {
>> + todo = min(tx_cnt, oleft);
>> + oleft -= todo;
>> + readsl(ss->base + SS_TXFIFO, dst32, todo);
>> + dst32 += todo;
>> + }
>> + spaces = readl(ss->base + SS_FCSR);
>> + rx_cnt = SS_RXFIFO_SPACES(spaces);
>> + tx_cnt = SS_TXFIFO_SPACES(spaces);
>> + } while (oleft > 0);
>> + kunmap_atomic(dst_addr);
>> + kunmap_atomic(src_addr);
>> + return 0;
>> +}
>> +
>> +int sunxi_ss_aes_poll(struct ablkcipher_request *areq, u32 mode)
>> +{
>> + u32 spaces;
>> + struct crypto_ablkcipher *tfm = crypto_ablkcipher_reqtfm(areq);
>> + struct sunxi_tfm_ctx *op = crypto_ablkcipher_ctx(tfm);
>> + struct sunxi_ss_ctx *ss = op->ss;
>> + unsigned int ivsize = crypto_ablkcipher_ivsize(tfm);
>> + /* when activating SS, the default FIFO space is 32 */
>> + u32 rx_cnt = 32;
>> + u32 tx_cnt = 0;
>> + u32 v;
>> + int i, err = 0;
>> + struct scatterlist *in_sg = areq->src;
>> + struct scatterlist *out_sg = areq->dst;
>> + void *src_addr;
>> + void *dst_addr;
>> + unsigned int ileft = areq->nbytes;
>> + unsigned int oleft = areq->nbytes;
>> + unsigned int sgileft = areq->src->length;
>> + unsigned int sgoleft = areq->dst->length;
>> + unsigned int todo;
>> + u32 *src32;
>> + u32 *dst32;
>> + unsigned long flags;
>> +
>> + spin_lock_irqsave(&ss->slock, flags);
>> +
>> + for (i = 0; i < op->keylen; i += 4)
>> + writel(*(op->key + i/4), ss->base + SS_KEY0 + i);
>> +
>> + if (areq->info != NULL) {
>> + for (i = 0; i < 4 && i < ivsize / 4; i++) {
>> + v = *(u32 *)(areq->info + i * 4);
>> + writel(v, ss->base + SS_IV0 + i * 4);
>> + }
>> + }
>> + writel(mode, ss->base + SS_CTL);
>> +
>> + /* If we have only one SG, we can use kmap_atomic */
>> + if (sg_next(in_sg) == NULL && sg_next(out_sg) == NULL) {
>> + err = sunxi_ss_aes_poll_atomic(areq);
>> + goto release_ss;
>> + }
>> +
>> + /*
>> + * If we have more than one SG, we cannot use kmap_atomic since
>> + * we hold the mapping too long
>> + */
>> + src_addr = kmap(sg_page(in_sg)) + in_sg->offset;
>> + if (src_addr == NULL) {
>> + dev_err(ss->dev, "KMAP error for src SG\n");
>> + err = -EINVAL;
>> + goto release_ss;
>> + }
>> + dst_addr = kmap(sg_page(out_sg)) + out_sg->offset;
>> + if (dst_addr == NULL) {
>> + kunmap(sg_page(in_sg));
>> + dev_err(ss->dev, "KMAP error for dst SG\n");
>> + err = -EINVAL;
>> + goto release_ss;
>> + }
>> + src32 = (u32 *)src_addr;
>> + dst32 = (u32 *)dst_addr;
>> + ileft = areq->nbytes / 4;
>> + oleft = areq->nbytes / 4;
>> + sgileft = in_sg->length / 4;
>> + sgoleft = out_sg->length / 4;
>> + do {
>> + spaces = readl_relaxed(ss->base + SS_FCSR);
>> + rx_cnt = SS_RXFIFO_SPACES(spaces);
>> + tx_cnt = SS_TXFIFO_SPACES(spaces);
>> + todo = min3(rx_cnt, ileft, sgileft);
>> + if (todo > 0) {
>> + ileft -= todo;
>> + sgileft -= todo;
>> + writesl(ss->base + SS_RXFIFO, src32, todo);
>> + src32 += todo;
>> + }
>> + if (in_sg != NULL && sgileft == 0 && ileft > 0) {
>> + kunmap(sg_page(in_sg));
>> + in_sg = sg_next(in_sg);
>> + while (in_sg != NULL && in_sg->length == 0)
>> + in_sg = sg_next(in_sg);
>> + if (in_sg != NULL && ileft > 0) {
>> + src_addr = kmap(sg_page(in_sg)) + in_sg->offset;
>> + if (src_addr == NULL) {
>> + dev_err(ss->dev, "ERROR: KMAP for src SG\n");
>> + err = -EINVAL;
>> + goto release_ss;
>> + }
>> + src32 = src_addr;
>> + sgileft = in_sg->length / 4;
>> + }
>> + }
>> + /* do not test oleft since when oleft == 0 we have finished */
>> + todo = min3(tx_cnt, oleft, sgoleft);
>> + if (todo > 0) {
>> + oleft -= todo;
>> + sgoleft -= todo;
>> + readsl(ss->base + SS_TXFIFO, dst32, todo);
>> + dst32 += todo;
>> + }
>> + if (out_sg != NULL && sgoleft == 0 && oleft >= 0) {
>> + kunmap(sg_page(out_sg));
>> + out_sg = sg_next(out_sg);
>> + while (out_sg != NULL && out_sg->length == 0)
>> + out_sg = sg_next(out_sg);
>> + if (out_sg != NULL && oleft > 0) {
>> + dst_addr = kmap(sg_page(out_sg)) +
>> + out_sg->offset;
>> + if (dst_addr == NULL) {
>> + dev_err(ss->dev, "KMAP error\n");
>> + err = -EINVAL;
>> + goto release_ss;
>> + }
>> + dst32 = dst_addr;
>> + sgoleft = out_sg->length / 4;
>> + }
>> + }
>> + } while (oleft > 0);
>> +
>> +release_ss:
>> + writel_relaxed(0, ss->base + SS_CTL);
>> + spin_unlock_irqrestore(&ss->slock, flags);
>> + return err;
>> +}
>
> Do you really need to have both an "optimized" and "non-optimized"
> function ?
>
> BTW, you should take a look at the sg_copy_buffer function [1], which
> is doing pretty much what you're doing here.
> If you don't want to directly use sg_copy_buffer, you should at least
> use the sg_miter_xxx function to iterate over you're sg list (it's
> taking care of calling kmap or kmap_atomic depending on the
> SG_MITER_ATOMIC flag).
>

I have dropped sg_copy_buffer that I used for DES/3DES some times ago for handling SG with length not multiple of 4.
And using sg_miter is not usefull.
It could be usefull for DES/3DEs for non 4 bytes multiple SG, but I use a fallback for that case.

>> +
>> +/* Pure CPU driven way of doing DES/3DES with SS */
>> +int sunxi_ss_des_poll(struct ablkcipher_request *areq, u32 mode)
>> +{
>> + struct crypto_ablkcipher *tfm = crypto_ablkcipher_reqtfm(areq);
>> + struct sunxi_tfm_ctx *op = crypto_ablkcipher_ctx(tfm);
>> + struct sunxi_ss_ctx *ss = op->ss;
>> + int i, err = 0;
>> + int no_chunk = 1;
>> + struct scatterlist *in_sg = areq->src;
>> + struct scatterlist *out_sg = areq->dst;
>> + u8 kkey[256 / 8];
>> +
>> + /*
>> + * if we have only SGs with size multiple of 4,
>> + * we can use the SS AES function
>> + */
>> + while (in_sg != NULL && no_chunk == 1) {
>> + if ((in_sg->length % 4) != 0)
>> + no_chunk = 0;
>> + in_sg = sg_next(in_sg);
>> + }
>> + while (out_sg != NULL && no_chunk == 1) {
>> + if ((out_sg->length % 4) != 0)
>> + no_chunk = 0;
>> + out_sg = sg_next(out_sg);
>> + }
>> +
>> + if (no_chunk == 1)
>> + return sunxi_ss_aes_poll(areq, mode);
>> +
>
> Given your explanations, I'm not sure the names sunxi_ss_aes_poll and
> sunxi_ss_des_poll are appropriate.
>
> What about sunxi_ss_aligned_cipher_op_poll and sunxi_ss_cipher_op_poll.

The difference is not about aligned or not since according to my understanding, the data will always be aligned following cra_alignmask.

Thanks for your review

>
> That's all I got for now.
> I haven't reviewed the hash part yet, but I guess some of my comments
> apply to this code too.
>
> Best Regards,
>
> Boris
>
>
> [1]http://lxr.free-electrons.com/source/lib/scatterlist.c#L621
>