Hi Mark,
Thanks for the review.
On Thu, 2020-11-19 at 15:15 -0800, mark gross wrote:
> I don't know the crypto domain well but.
>
> my main feedback is:
>
> * can we remove the sha224 config option and just have it enabled?
Having SHA224 optional (and disabled by default) was a request from
Intel internal crypto team, who, due to internal policy, had to review
this driver before it could be made public.
I understand this may look strange to the Kernel community, but I'm not
allowed to upstream SHA224 support without following the
recommendations from our internal crypto team.
Anyway, to isolate this issue, in v4, I've isolated SHA224 support in a
different patch, so that discussion about SHA224 can continue without
blocking the upstreaming of the rest of the driver.
> * do you need spin lock bh or would a plane spin lock do?
> * would a mutex do instead of the spinlock?
Other drivers (atmel-sha.c, omap-sham.c, stm32-hash.c, etc.) are also
using spin_lock_bh()). I think that's because some ahash functions can
be called from softirqs.
>
> Other than that I'm asking you to look at the log buffer output and ask
> yourself in each case : "what can an admin do with the output or is it
> actionable?"
In v4 (soon to be submitted), I've tried to improve error messages.
>
>
> On Mon, Nov 16, 2020 at 06:58:45PM +0000, Daniele Alessandrelli wrote:
> > From: Declan Murphy <declan.murphy@xxxxxxxxx>
> >
> > Add support for the Hashing Control Unit (HCU) included in the Offload
> > Crypto Subsystem (OCS) of the Intel Keem Bay SoC, thus enabling
> > hardware-accelerated hashing on the Keem Bay SoC for the following
> > algorithms:
> > - sha224 and hmac(sha224)
> > - sha256 and hmac(sha256)
> > - sha384 and hmac(sha384)
> > - sha512 and hmac(sha512)
> > - sm3 and hmac(sm3)
> >
> > The driver passes crypto manager self-tests, including the extra tests
> > (CRYPTO_MANAGER_EXTRA_TESTS=y).
> >
> > Signed-off-by: Declan Murphy <declan.murphy@xxxxxxxxx>
> > Co-developed-by: Daniele Alessandrelli <daniele.alessandrelli@xxxxxxxxx>
> > Signed-off-by: Daniele Alessandrelli <daniele.alessandrelli@xxxxxxxxx>
> > Acked-by: Mark Gross <mgross@xxxxxxxxxxxxxxx>
> > ---
> > drivers/crypto/Kconfig | 2 +
> > drivers/crypto/Makefile | 1 +
> > drivers/crypto/keembay/Kconfig | 20 +
> > drivers/crypto/keembay/Makefile | 5 +
> > drivers/crypto/keembay/keembay-ocs-hcu-core.c | 1484 +++++++++++++++++
> > drivers/crypto/keembay/ocs-hcu.c | 593 +++++++
> > drivers/crypto/keembay/ocs-hcu.h | 115 ++
> > 7 files changed, 2220 insertions(+)
> > create mode 100644 drivers/crypto/keembay/Kconfig
> > create mode 100644 drivers/crypto/keembay/Makefile
> > create mode 100644 drivers/crypto/keembay/keembay-ocs-hcu-core.c
> > create mode 100644 drivers/crypto/keembay/ocs-hcu.c
> > create mode 100644 drivers/crypto/keembay/ocs-hcu.h
> >
> > diff --git a/drivers/crypto/Kconfig b/drivers/crypto/Kconfig
> > index c2950127def6..e636be70b18d 100644
> > --- a/drivers/crypto/Kconfig
> > +++ b/drivers/crypto/Kconfig
> > @@ -881,4 +881,6 @@ config CRYPTO_DEV_SA2UL
> > used for crypto offload. Select this if you want to use hardware
> > acceleration for cryptographic algorithms on these devices.
> >
> > +source "drivers/crypto/keembay/Kconfig"
> > +
> > endif # CRYPTO_HW
> > diff --git a/drivers/crypto/Makefile b/drivers/crypto/Makefile
> > index 53fc115cf459..fff9a70348e1 100644
> > --- a/drivers/crypto/Makefile
> > +++ b/drivers/crypto/Makefile
> > @@ -51,3 +51,4 @@ obj-$(CONFIG_CRYPTO_DEV_ARTPEC6) += axis/
> > obj-$(CONFIG_CRYPTO_DEV_ZYNQMP_AES) += xilinx/
> > obj-y += hisilicon/
> > obj-$(CONFIG_CRYPTO_DEV_AMLOGIC_GXL) += amlogic/
> > +obj-y += keembay/
> > diff --git a/drivers/crypto/keembay/Kconfig b/drivers/crypto/keembay/Kconfig
> > new file mode 100644
> > index 000000000000..db8954dcd057
> > --- /dev/null
> > +++ b/drivers/crypto/keembay/Kconfig
> > @@ -0,0 +1,20 @@
> > +config CRYPTO_DEV_KEEMBAY_OCS_HCU
> > + tristate "Support for Intel Keem Bay OCS HCU HW acceleration"
> > + select CRYPTO_ENGINE
> > + depends on OF || COMPILE_TEST
> > + help
> > + Support for Intel Keem Bay Offload and Crypto Subsystem (OCS) Hash
> > + Control Unit (HCU) hardware acceleration for use with Crypto API.
> > +
> > + Provides OCS HCU hardware acceleration of sha256, sha384, sha512, and
> > + sm3, as well as the HMAC variant of these algorithms.
>
> might want to add a "Say M or Y if ...." that can help a non KEEMBAY engineer
> read the help and know if they should enabling. Keep in mind admins are
> reading this help text.
Done in v4.
>
> > +
> > +config CRYPTO_DEV_KEEMBAY_OCS_HCU_HMAC_SHA224
> > + bool "Enable sha224 and hmac(sha224) support in Intel Keem Bay OCS HCU"
> s/in/to
>
> Looking at the code its clear this option enabls sha224 support to the driver
> in this patch. The question is WHY IS this an option at all? what is the
> traid off for just including sha224 support and not deal with the complexity of
> some configs turning it on while others may not?
>
> Is there a downside to just enabling sha224?
Explained above. Basically, it's just due to some company internal
policy. I understand that the Linux crypto maintainers may reject this.
>
>
> > + depends on CRYPTO_DEV_KEEMBAY_OCS_HCU
> > + help
> > + Enables support for sha224 and hmac(sha224) algorithms to the Intel
> > + Keem Bay OCS HCU driver. Intel recommends not to use these
> > + algorithms.
> > +
> > + Provides OCS HCU hardware acceleration of sha224 and hmac(224).
> > diff --git a/drivers/crypto/keembay/Makefile b/drivers/crypto/keembay/Makefile
> > new file mode 100644
> > index 000000000000..2aa2647fab1b
> > --- /dev/null
> > +++ b/drivers/crypto/keembay/Makefile
> > @@ -0,0 +1,5 @@
> > +#
> > +# Makefile for Keem Bay OCS Crypto API Linux drivers
> > +#
> > +obj-$(CONFIG_CRYPTO_DEV_KEEMBAY_OCS_HCU) += keembay-ocs-hcu.o
> > +keembay-ocs-hcu-objs := keembay-ocs-hcu-core.o ocs-hcu.o
> > diff --git a/drivers/crypto/keembay/keembay-ocs-hcu-core.c b/drivers/crypto/keembay/keembay-ocs-hcu-core.c
> > new file mode 100644
> > index 000000000000..2c8f8f331501
> > --- /dev/null
> > +++ b/drivers/crypto/keembay/keembay-ocs-hcu-core.c
> > @@ -0,0 +1,1484 @@
> > +// SPDX-License-Identifier: GPL-2.0-only
> > +/*
> > + * Intel Keem Bay OCS HCU Crypto Driver.
> > + *
> > + * Copyright (C) 2018-2020 Intel Corporation
> > + */
> > +
> > +#include <linux/delay.h>
> > +#include <linux/dma-mapping.h>
> > +#include <linux/interrupt.h>
> > +#include <linux/module.h>
> > +#include <linux/of_device.h>
> > +
> > +#include <crypto/engine.h>
> > +#include <crypto/scatterwalk.h>
> > +#include <crypto/sha.h>
> > +#include <crypto/sm3.h>
> > +#include <crypto/hmac.h>
> > +#include <crypto/internal/hash.h>
> > +
> > +#include "ocs-hcu.h"
> > +
> > +#define DRV_NAME "keembay-ocs-hcu"
> > +
> > +/* Flag marking a final request. */
> > +#define REQ_FINAL BIT(0)
> > +/* Flag set when there is data to be processed by the HW. */
> > +#define REQ_FLAGS_DO_DATA BIT(1)
> > +/* Flag set when there is intermediate data from a previous request. */
> > +#define REQ_FLAGS_INTERMEDIATE_DATA BIT(2)
> > +/* Flag marking a HMAC request. */
> > +#define REQ_FLAGS_HMAC BIT(3)
> > +/* Flag set when HW HMAC is being used. */
> > +#define REQ_FLAGS_HMAC_HW BIT(4)
> > +/* Flag set when SW HMAC is being used. */
> > +#define REQ_FLAGS_HMAC_SW BIT(5)
> > +/* Flag set when HW HMAC key has been already set (only used in HW HMAC). */
> > +#define REQ_FLAGS_HMAC_HW_KEY_SET BIT(6)
> > +/* Flag set when OPAD computing is pending (only used in SW HMAC). */
> > +#define REQ_FLAGS_HMAC_SW_DO_OPAD BIT(7)
> > +
> > +#define KMB_OCS_HCU_ALIGNMENT 8
> > +#define KMB_OCS_HCU_ALIGN_MASK (KMB_OCS_HCU_ALIGNMENT - 1)
> > +
> > +/**
> > + * struct ocs_hcu_ctx: OCS HCU Transform context.
> > + * @engine_ctx: Crypto Engine context.
> > + * @hcu_dev: The OCS HCU device used by the transformation.
> > + * @key: The key (used only for HMAC transformations).
> > + * @key_len: The length of the key.
> > + * @is_sm3_tfm: Whether or not this is an SM3 transformation.
> > + * @is_hmac_tfm: Whether or not this is a HMAC transformation.
> > + */
> > +struct ocs_hcu_ctx {
> > + struct crypto_engine_ctx engine_ctx;
> > + struct ocs_hcu_dev *hcu_dev;
> > + u8 key[SHA512_BLOCK_SIZE];
> > + size_t key_len;
> > + bool is_sm3_tfm;
> > + bool is_hmac_tfm;
> > +};
> > +
> > +/**
> > + * struct dma_buf - DMA-able buffer.
> > + * @dma_addr: The DMA address of the buffer.
> > + * @vaddr: The Kernel virtual address of the buffer.
> > + * @size: The size of the buffer.
> > + */
> > +struct dma_buf {
> > + dma_addr_t dma_addr;
> > + u8 *vaddr;
> > + size_t size;
> > +};
> > +
> > +/**
> > + * struct ocs_dma_list - OCS-specific DMA linked list.
> > + * @head: The head of the list (points to the array backing the list).
> > + * @tail: The current tail of the list; NULL if the list is empty.
> > + * @dma_addr: The DMA address of @head (i.e., the DMA address of the backing
> > + * array).
> > + * @nents: Maximum number of entries in the list (i.e., number of elements
> > + * in the backing array).
> > + *
> > + * The OCS DMA list is an array-backed list of OCS DMA descriptors. The array
> > + * backing the list is allocated with dma_alloc_coherent() and pointed by
> > + * @head.
> > + */
> > +struct ocs_dma_list {
> > + struct ocs_dma_desc *head;
> > + struct ocs_dma_desc *tail;
> > + dma_addr_t dma_addr;
> > + size_t nents;
> > +};
> > +
> > +/**
> > + * struct ocs_hcu_rctx - Context for the request.
> > + * @flags: Flags tracking request status.
> > + * @algo: Algorithm to use for the request.
> > + * @blk_sz: Block size of the transformation / request.
> > + * @dig_sz: Digest size of the transformation / request.
> > + * @dma_list: OCS DMA linked list.
> > + * @hcu_dev: OCS HCU device to be used to service the request.
> > + * @idata: Intermediate data (used to store intermediate results).
> > + * @buffer: Buffer to store: partial block of data and SW HMAC
> > + * artifacts (ipad, opad, etc.).
> > + * @buf_cnt: Number of bytes currently stored in the buffer.
> > + * @dma_buf: DMA-able copy of buffer.
> > + * @sg: Head of the scatterlist entries containing data.
> > + * @sg_data_total: Total data in the SG list at any time.
> > + * @sg_data_offset: Offset into the data of the current individual SG node.
> > + * @sg_dma_nents: Number of sg entries mapped in dma_list.
> > + */
> > +struct ocs_hcu_rctx {
> > + u32 flags;
> > + u32 algo;
> > + size_t blk_sz;
> > + size_t dig_sz;
> > + struct ocs_dma_list dma_list;
> > + struct ocs_hcu_dev *hcu_dev;
> > + struct ocs_hcu_idata idata;
> > + /*
> > + * Buffer is double the block size because we need space for SW HMAC
> > + * artifacts, i.e:
> > + * - ipad (1 block) + a possible partial block of data.
> > + * - opad (1 block) + digest of H(k ^ ipad || m)
> > + */
> > + u8 buffer[2 * SHA512_BLOCK_SIZE];
> > + size_t buf_cnt;
> > + struct dma_buf dma_buf;
> > + struct scatterlist *sg;
> > + unsigned int sg_data_total;
> > + unsigned int sg_data_offset;
> > + unsigned int sg_dma_nents;
> > +};
> > +
> > +/**
> > + * struct ocs_hcu_drv - Driver data
> > + * @dev_list: The list of HCU devices.
> > + * @lock: The lock protecting dev_list.
> > + */
> > +struct ocs_hcu_drv {
> > + struct list_head dev_list;
> > + spinlock_t lock; /* Protects dev_list. */
> > +};
> > +
> > +static struct ocs_hcu_drv ocs_hcu = {
> > + .dev_list = LIST_HEAD_INIT(ocs_hcu.dev_list),
> > + .lock = __SPIN_LOCK_UNLOCKED(ocs_hcu.lock),
> > +};
> > +
> > +/*
> > + * Return the total amount of data in the request; that is: the data in the
> > + * request buffer + the data in the sg list.
> > + */
> > +static inline unsigned int kmb_get_total_data(struct ocs_hcu_rctx *rctx)
> > +{
> > + return rctx->sg_data_total + rctx->buf_cnt;
> > +}
> > +
> > +/* Move remaining content of scatter-gather list to context buffer. */
> > +static int flush_sg_to_ocs_buffer(struct ocs_hcu_rctx *rctx)
> > +{
> > + size_t count;
> > +
> > + if (rctx->sg_data_total > (sizeof(rctx->buffer) - rctx->buf_cnt)) {
> > + WARN(1, "sg data does not fit in buffer\n");
> > + return -EINVAL;
> > + }
> > +
> > + while (rctx->sg_data_total) {
> > + if (!rctx->sg) {
> > + WARN(1, "sg == NULL, but sg_data_total != 0\n");
> > + return -EINVAL;
> > + }
> > + /*
> > + * If current sg has been fully processed, skip to the next
> > + * one.
> > + */
> > + if (rctx->sg_data_offset == rctx->sg->length) {
> > + rctx->sg = sg_next(rctx->sg);
> > + rctx->sg_data_offset = 0;
> > + continue;
> > + }
> > + /*
> > + * Determine the maximum data available to copy from the node.
> > + * Minimum of the length left in the sg node, or the total data
> > + * in the request.
> > + */
> > + count = min(rctx->sg->length - rctx->sg_data_offset,
> > + rctx->sg_data_total);
> > + /* Copy from scatter-list entry to context buffer. */
> > + scatterwalk_map_and_copy(&rctx->buffer[rctx->buf_cnt],
> > + rctx->sg, rctx->sg_data_offset,
> > + count, 0);
> > +
> > + rctx->sg_data_offset += count;
> > + rctx->sg_data_total -= count;
> > + rctx->buf_cnt += count;
> > + }
> > +
> > + return 0;
> > +}
> > +
> > +static struct ocs_hcu_dev *kmb_ocs_hcu_find_dev(struct ahash_request *req)
> > +{
> > + struct crypto_ahash *tfm = crypto_ahash_reqtfm(req);
> > + struct ocs_hcu_ctx *tctx = crypto_ahash_ctx(tfm);
> > +
> > + /* If the HCU device for the request was previously set, return it. */
> > + if (tctx->hcu_dev)
> > + return tctx->hcu_dev;
> > +
> > + /*
> > + * Otherwise, get the first HCU device available (there should be one
> > + * and only one device).
> > + */
> > + spin_lock_bh(&ocs_hcu.lock);
> _bh spinlocks are kindof big hammers. they disable softIRQ's is it justified?
> wouldn't a normal spinlock work just as well?
> do you have list processing in a interrupt handler?
>
> heck would a mutex work just as well?
My understanding is that some crypto ahash functions can be called from
softirqs, so spin_lock_bh() is required.
Regarding the list processing, we just get the first element, so the
operation is time bound.
>
> > + tctx->hcu_dev = list_first_entry_or_null(&ocs_hcu.dev_list,
> > + struct ocs_hcu_dev,
> > + list);
> > + spin_unlock_bh(&ocs_hcu.lock);
> > +
> > + WARN(!tctx->hcu_dev, "No HCU device available\n");
> > +
> > + return tctx->hcu_dev;
> > +}
> > +
> > +/* Free OCS DMA linked list and DMA-able context buffer. */
> > +static void kmb_ocs_free_dma_list(struct ocs_hcu_rctx *rctx)
> > +{
> > + struct ocs_hcu_dev *hcu_dev = rctx->hcu_dev;
> > + struct device *dev = hcu_dev->dev;
> > +
> > + /* Free DMA-able context buffer, if it was allocated. */
> > + if (rctx->dma_buf.vaddr) {
> > + memzero_explicit(rctx->dma_buf.vaddr, rctx->dma_buf.size);
> > + dma_free_coherent(dev, rctx->dma_buf.size, rctx->dma_buf.vaddr,
> > + rctx->dma_buf.dma_addr);
> > + memset(&rctx->dma_buf, 0, sizeof(rctx->dma_buf));
> > + }
> > +
> > + if (!rctx->dma_list.head)
> > + return;
> > +
> > + if (rctx->sg_dma_nents > 0)
> > + dma_unmap_sg(dev, hcu_dev->req->src, rctx->sg_dma_nents,
> > + DMA_TO_DEVICE);
> > +
> > + dma_free_coherent(dev,
> > + sizeof(*rctx->dma_list.head) * rctx->dma_list.nents,
> > + rctx->dma_list.head, rctx->dma_list.dma_addr);
> > +
> > + memset(&rctx->dma_list, 0, sizeof(rctx->dma_list));
> > +}
> > +
> > +/* Add a new DMA entry at the end of the OCS DMA list. */
> > +static int kmb_ocs_add_dma_tail(struct ocs_hcu_rctx *rctx,
> > + dma_addr_t addr, size_t len)
> > +{
> > + struct ocs_hcu_dev *hcu_dev = rctx->hcu_dev;
> > + struct device *dev = hcu_dev->dev;
> > + struct ocs_dma_list *dma_list = &rctx->dma_list;
> > + struct ocs_dma_desc *old_tail = dma_list->tail;
> > + struct ocs_dma_desc *new_tail = old_tail ?
> > + old_tail + 1 : dma_list->head;
> > +
> > + if (!len)
> > + return 0;
> > +
> > + if (addr & ~OCS_HCU_DMA_BIT_MASK) {
> > + dev_err(dev,
> > + "Unexpected error: Invalid DMA address for OCS HCU\n");
> > + return -EINVAL;
> > + }
> > +
> > + if (new_tail - dma_list->head >= dma_list->nents) {
> > + dev_err(dev, "Unexpected error: OCS DMA list is full\n");
> > + return -ENOMEM;
> > + }
> > +
> > + /*
> > + * If there was an old tail (i.e., this is not the first element we are
> > + * adding), un-terminate the old tail and make it point to the new one.
> > + */
> > + if (old_tail) {
> > + old_tail->ll_flags &= ~OCS_LL_DMA_FLAG_TERMINATE;
> > + /*
> > + * The old tail 'nxt_desc' must point to the DMA address of the
> > + * new tail.
> > + */
> > + old_tail->nxt_desc = dma_list->dma_addr +
> > + sizeof(*dma_list->tail) * (new_tail -
> > + dma_list->head);
> > + }
> > +
> > + new_tail->src_adr = (u32)addr;
> > + new_tail->src_len = (u32)len;
> > + new_tail->ll_flags = OCS_LL_DMA_FLAG_TERMINATE;
> > + new_tail->nxt_desc = 0;
> > +
> > + /* Update list tail with new tail. */
> > + dma_list->tail = new_tail;
> > +
> > + return 0;
> > +}
> > +
> > +/*
> > + * Initialize OCS DMA list:
> > + * - Allocate DMA linked list (number of elements: SG entries to process +
> > + * context buffer, if not empty).
> > + * - Allocate DMA-able context buffer (if needed) and add it to the DMA list.
> > + * - Add SG entries to the DMA list.
> > + *
> > + * Note: if this is a final request, we process all the data in the SG list,
> > + * otherwise we can only process up to the maximum amount of block-aligned data
> > + * (the remainder will be put into the context buffer and processed in the next
> > + * request).
> > + */
> > +static int kmb_ocs_init_dma_list(struct ocs_hcu_dev *hcu_dev)
> > +{
> > + struct ahash_request *req = hcu_dev->req;
> > + struct ocs_hcu_rctx *rctx = ahash_request_ctx(req);
> > + struct device *dev = hcu_dev->dev;
> > + struct scatterlist *sg = req->src;
> > + unsigned int remainder = 0;
> > + unsigned int total;
> > + size_t nents;
> what is a nent?
Number of entries (I think), it's a variable name commonly used by the
scatterlist API.
>
> > + size_t count;
> > + int rc;
> > + int i;
> > +
> > + total = kmb_get_total_data(rctx);
> > + if (!total)
> > + return 0;
> > +
> > + /*
> > + * If this is not a final DMA (terminated DMA), the data passed to the
> > + * HCU must be aligned to the block size; compute the remainder data to
> > + * be processed in the next request.
> > + */
> > + if (!(rctx->flags & REQ_FINAL))
> > + remainder = total % rctx->blk_sz;
> > +
> > + /* Determine the number of scatter gather list nodes to process. */
> > + nents = sg_nents_for_len(sg, rctx->sg_data_total - remainder);
> > + rctx->sg_dma_nents = nents;
> > +
> > + /* Add extra DMA entry for data in context buffer, if any. */
> > + if (rctx->buf_cnt)
> > + nents++;
> > +
> > + /* Total size of the DMA list to allocate. */
> > + rctx->dma_list.head =
> > + dma_alloc_coherent(dev, sizeof(*rctx->dma_list.head) * nents,
> > + &rctx->dma_list.dma_addr, GFP_KERNEL);
> > + if (!rctx->dma_list.head) {
> > + dev_err(dev, "Failed to allocated DMA list\n");
> > + return -ENOMEM;
> > + }
> > + rctx->dma_list.nents = nents;
> > + rctx->dma_list.tail = NULL;
> > +
> > + rc = dma_map_sg(hcu_dev->dev, req->src, rctx->sg_dma_nents,
> > + DMA_TO_DEVICE);
> > + if (rc != rctx->sg_dma_nents) {
> > + dev_err(dev, "Failed to MAP SG\n");
> > + rc = -ENOMEM;
> > + goto cleanup;
> > + }
> > +
> > + /*
> > + * If context buffer is not empty, create DMA-able buffer for it, copy
> > + * context to new buffer and add it to the DMA list.
> > + */
> > + if (rctx->buf_cnt) {
> > + rctx->dma_buf.size = rctx->buf_cnt;
> > + rctx->dma_buf.vaddr =
> > + dma_alloc_coherent(dev, rctx->dma_buf.size,
> > + &rctx->dma_buf.dma_addr,
> > + GFP_KERNEL);
> > + if (!rctx->dma_buf.vaddr) {
> > + dev_err(dev, "Failed to allocate DMA Buf\n");
> > + goto cleanup;
> > + }
> > + memcpy(rctx->dma_buf.vaddr, rctx->buffer, rctx->buf_cnt);
> > + rc = kmb_ocs_add_dma_tail(rctx, rctx->dma_buf.dma_addr,
> > + rctx->dma_buf.size);
> > + if (rc)
> > + goto cleanup;
> > + rctx->buf_cnt = 0;
> > + }
> > +
> > + /* Add the SG nodes to be processed to the DMA linked list. */
> > + for_each_sg(req->src, rctx->sg, rctx->sg_dma_nents, i) {
> > + count = min(rctx->sg_data_total - remainder,
> > + rctx->sg->length - rctx->sg_data_offset);
> > + /*
> > + * Do not create a zero length DMA descriptor. Check in case of
> > + * zero length SG node.
> > + */
> > + if (count == 0)
> > + continue;
> > + /* Add sg to HCU DMA list. */
> > + rc = kmb_ocs_add_dma_tail(rctx, rctx->sg->dma_address,
> > + count);
> > + if (rc)
> > + goto cleanup;
> > +
> > + /* Update amount of data remaining in SG list. */
> > + rctx->sg_data_total -= count;
> > +
> > + /*
> > + * If remaining data is equal to remainder (note: 'less than'
> > + * case will never happen in practice), we are done: update
> never is a strong word. so why the <=? shouldn't it be == then?
>
> maybe explain why you handle the < case. This comment makes me want to do an :
> if ==
> else if <
> where I'd log the < case hitting because its not expected in practice....
Good point. The '<' is just to handle possible bugs gracefully.
In v4, I've added a WARN() to notify of the unexpected event.
>
> > + * offset and exit the loop.
> > + */
> > + if (rctx->sg_data_total <= remainder) {
> > + rctx->sg_data_offset += count;
> > + break;
> > + }
> > +
> > + /*
> > + * If we get here is because we need to process the next sg in
> > + * the list; set offset within the sg to 0.
> > + */
> > + rctx->sg_data_offset = 0;
> > + }
> > +
> > + return 0;
> > +cleanup:
> > + dev_err(dev, "Failed to map DMA buffer.\n");
> > + kmb_ocs_free_dma_list(rctx);
> > +
> > + return rc;
> > +}
> > +
> > +static void kmb_ocs_hcu_secure_cleanup(struct ocs_hcu_dev *hcu_dev)
> > +{
> > + struct ocs_hcu_rctx *rctx = ahash_request_ctx(hcu_dev->req);
> > +
> > + /* Clear buffer of any data. */
> > + memzero_explicit(rctx->buffer, sizeof(rctx->buffer));
> > + /*
> > + * Clear the key in HW.
> > + *
> > + * NOTE: the key in the tfm context (ctx->key) is cleared in
> > + * kmb_ocs_hcu_hmac_cra_exit(), i.e., when tfm is de-initialized.
> > + */
> > + ocs_hcu_clear_key(hcu_dev);
> > +}
> > +
> > +static void kmb_ocs_hcu_finish_request(struct ocs_hcu_dev *hcu_dev, int error)
> > +{
> > + struct ocs_hcu_rctx *rctx = ahash_request_ctx(hcu_dev->req);
> > +
> > + if (error)
> > + goto exit;
> > +
> > + /* Get the intermediate data (including digest). */
> > + error = ocs_hcu_get_intermediate_data(hcu_dev, &rctx->idata,
> > + rctx->algo);
> > + if (error)
> > + goto exit;
> > + /*
> > + * If this is not the final request
> > + * - Set the intermediate data flag so that intermediate data will be
> > + * written back to HW during the next request processing.
> > + * - Move remaining data in the SG list to the request buffer, so that
> > + * it will be processed during the next request.
> > + */
> > + if (!(rctx->flags & REQ_FINAL)) {
> > + rctx->flags |= REQ_FLAGS_INTERMEDIATE_DATA;
> > + error = flush_sg_to_ocs_buffer(rctx);
> > + goto exit;
> > + }
> > +
> > + /*
> > + * If we get here, this was the final request and the digest in the
> > + * intermediate date is actually the final digest: copy it to the
> > + * request result field.
> > + */
> > + memcpy(hcu_dev->req->result, rctx->idata.digest, rctx->dig_sz);
> > +
> > +exit:
> > + if (error || rctx->flags & REQ_FINAL)
> > + kmb_ocs_hcu_secure_cleanup(hcu_dev);
> > + ocs_hcu_hw_disable(hcu_dev);
> > + kmb_ocs_free_dma_list(rctx);
> > + crypto_finalize_hash_request(hcu_dev->engine, hcu_dev->req, error);
> > +}
> > +
> > +static int kmb_ocs_hcu_set_hw(struct ocs_hcu_dev *hcu_dev, int algo)
> > +{
> > + struct ocs_hcu_rctx *rctx = ahash_request_ctx(hcu_dev->req);
> > + struct crypto_ahash *tfm = crypto_ahash_reqtfm(hcu_dev->req);
> > + struct ocs_hcu_ctx *ctx = crypto_ahash_ctx(tfm);
> > + struct device *dev = hcu_dev->dev;
> > + int rc;
> > +
> > + rc = ocs_hcu_wait_busy(hcu_dev);
> > + if (rc) {
> > + dev_err(hcu_dev->dev, "%s HW Busy, cannot continue.\n",
> > + __func__);
> > + return -EIO;
> > + }
> > +
> > + /* Configure the hardware for the current request. */
> > + rc = ocs_hcu_hw_cfg(hcu_dev, algo);
> > + if (rc)
> > + return rc;
> > +
> > + if (rctx->flags & REQ_FLAGS_INTERMEDIATE_DATA) {
> > + ocs_hcu_set_intermediate_data(hcu_dev, &rctx->idata, algo);
> > + rctx->flags &= ~REQ_FLAGS_INTERMEDIATE_DATA;
> > + }
> > +
> > + /*
> > + * Set key if all the following conditions are met:
> > + * 1. We are using HW HMAC
> > + * 2. Key has not been set already
> > + */
> > + if (rctx->flags & REQ_FLAGS_HMAC_HW &&
> > + !(rctx->flags & REQ_FLAGS_HMAC_HW_KEY_SET)) {
> > + if (ctx->key_len > HCU_HW_KEY_LEN) {
> > + dev_err(dev,
> > + "Unexpected error: key len > HCU_HW_KEY_LEN\n");
> > + return -EINVAL;
> > + }
> > + /*
> > + * Hardware requires all the bytes of the HW Key vector to be
> > + * written. So pad with zero until we reach HCU_HW_KEY_LEN.
> > + */
> > + memzero_explicit(&ctx->key[ctx->key_len],
> > + HCU_HW_KEY_LEN - ctx->key_len);
> > + rc = ocs_hcu_write_key(hcu_dev, ctx->key, HCU_HW_KEY_LEN);
> > + if (rc)
> > + return rc;
> > + rctx->flags |= REQ_FLAGS_HMAC_HW_KEY_SET;
> > + }
> > +
> > + return 0;
> > +}
> > +
> > +static int kmb_ocs_hcu_hash_pio(struct ocs_hcu_dev *hcu_dev,
> > + u8 *buf, u32 sz, u32 algo, bool term)
> > +{
> > + int rc = 0;
> > +
> > + /* Configure the hardware for the current request. */
> > + rc = kmb_ocs_hcu_set_hw(hcu_dev, algo);
> > +
> > + if (rc)
> > + return rc;
> > +
> > + if (sz != ocs_hcu_hash_cpu(hcu_dev, buf, sz, algo, term))
> > + return -EIO;
> > +
> > + return 0;
> > +}
> > +
> > +static int kmb_ocs_hcu_tx_dma(struct ocs_hcu_dev *hcu_dev)
> > +{
> > + struct ocs_hcu_rctx *rctx = ahash_request_ctx(hcu_dev->req);
> > + bool term = !!(rctx->flags & REQ_FINAL);
> > + int rc;
> > +
> > + if (!rctx->dma_list.head)
> > + return 0;
> > +
> > + /* Configure the hardware for the current request. */
> > + rc = kmb_ocs_hcu_set_hw(hcu_dev, rctx->algo);
> > +
> > + if (rc)
> > + return rc;
> > +
> > + /* Start the DMA engine with the descriptor address stored. */
> > + rc = ocs_hcu_ll_dma_start(hcu_dev, rctx->dma_list.dma_addr, term);
> > + if (rc)
> > + return rc;
> > +
> > + return 0;
> > +}
> > +
> > +static int kmb_ocs_hcu_handle_queue(struct ahash_request *req)
> > +{
> > + struct ocs_hcu_dev *hcu_dev = kmb_ocs_hcu_find_dev(req);
> > +
> > + if (!hcu_dev)
> > + return -ENOENT;
> > +
> > + return crypto_transfer_hash_request_to_engine(hcu_dev->engine,
> > + req);
> > +}
> > +
> > +static int kmb_ocs_hcu_prepare_request(struct crypto_engine *engine, void *areq)
> > +{
> > + struct ahash_request *req = container_of(areq, struct ahash_request,
> > + base);
> > + struct ocs_hcu_dev *hcu_dev = kmb_ocs_hcu_find_dev(req);
> > + struct ocs_hcu_rctx *rctx = ahash_request_ctx(req);
> > + int rc;
> > +
> > + if (!hcu_dev)
> > + return -ENOENT;
> > +
> > + hcu_dev->req = req;
> > +
> > + /*
> > + * If the request has no data (i.e., it's a final request without data),
> > + * we are done.
> > + */
> > + if (kmb_get_total_data(rctx) == 0)
> > + return 0;
> > + /* Otherwise, move the data into the HCU DMA linked list. */
> > + rc = kmb_ocs_init_dma_list(hcu_dev);
> > + if (rc)
> > + return rc;
> > + /* Set the flag to notify of pending data. */
> > + rctx->flags |= REQ_FLAGS_DO_DATA;
> > +
> > + return 0;
> > +}
> > +
> > +static int kmb_ocs_hcu_do_final(struct ocs_hcu_dev *hcu_dev)
> > +{
> > + struct ocs_hcu_rctx *rctx = ahash_request_ctx(hcu_dev->req);
> > + int rc;
> > +
> > + rc = kmb_ocs_hcu_set_hw(hcu_dev, rctx->algo);
> > + if (rc)
> > + return rc;
> > +
> > + ocs_hcu_tx_data_done(hcu_dev);
> > +
> > + return 0;
> > +}
> > +
> > +static int prepare_ipad(struct ahash_request *req)
> > +{
> > + struct ocs_hcu_rctx *rctx = ahash_request_ctx(req);
> > + struct crypto_ahash *tfm = crypto_ahash_reqtfm(req);
> > + struct ocs_hcu_ctx *ctx = crypto_ahash_ctx(tfm);
> > + int i;
> > +
> > + WARN(rctx->buf_cnt != 0, "rctx->buf_cnt = %zu\n", rctx->buf_cnt);
> is this enough information for an admin to do something useful or report an
> issue to us?
> looks like tribal knowledge is needed to take action if seen.
>
> FWIW If I saw rctx->buf_cnt = <some non-zero number> in a log buff I'd shrug.
I have tried to improve the error message in v4.
>
>
> > + WARN(!(rctx->flags & REQ_FLAGS_HMAC_SW), "Not HMAC SW");
> Same here is this an actionalbe output to the log buffer?
Fixed in v4.
>
> > + /*
> > + * Key length must be equal to block size. If key is shorter,
> > + * we pad it with zero (note: key cannot be longer, since
> > + * longer keys are hashed by kmb_ocs_hcu_setkey()).
> > + */
> > + WARN_ON(ctx->key_len > rctx->blk_sz);
> maybe if this warn hits we should return -EINVAL?
> (why not return an error?)
Good point, I've fixed it in v4.
>
> > + memzero_explicit(&ctx->key[ctx->key_len],
> > + rctx->blk_sz - ctx->key_len);
> > + ctx->key_len = rctx->blk_sz;
> > + /*
> > + * Prepare IPAD for HMAC. Only done for first block.
> > + * HMAC(k,m) = H(k ^ opad || H(k ^ ipad || m))
> > + * k ^ ipad will be first hashed block.
> > + * k ^ opad will be calculated in the final request.
> > + * Only needed if not using HW HMAC.
> > + */
> > + for (i = 0; i < rctx->blk_sz; i++)
> > + rctx->buffer[i] = ctx->key[i] ^ HMAC_IPAD_VALUE;
> > + rctx->buf_cnt = rctx->blk_sz;
> > +
> > + return 0;
> > +}
> > +
> > +static int kmb_ocs_hcu_do_one_request(struct crypto_engine *engine, void *areq)
> > +{
> > + struct ahash_request *req = container_of(areq, struct ahash_request,
> > + base);
> > + struct ocs_hcu_dev *hcu_dev = kmb_ocs_hcu_find_dev(req);
> > + struct ocs_hcu_rctx *rctx = ahash_request_ctx(req);
> > +
> > + if (!hcu_dev)
> > + return -ENOENT;
> > +
> > + hcu_dev->req = req;
> > +
> > + /* Initialize the hardware. */
> > + ocs_hcu_hw_init(hcu_dev);
> > +
> > + /*
> > + * If there is data to process (i.e., this is an update request or a
> > + * final request with data), process the data.
> > + */
> > + if (rctx->flags & REQ_FLAGS_DO_DATA) {
> > + rctx->flags &= ~REQ_FLAGS_DO_DATA;
> > + return kmb_ocs_hcu_tx_dma(hcu_dev);
> > + }
> > + /* Otherwise, if this is a final request, finalize the hash. */
> > + if (rctx->flags & REQ_FINAL)
> > + return kmb_ocs_hcu_do_final(hcu_dev);
> > + /* If we end up here, something was wrong with the request. */
> > + return -EINVAL;
> > +}
> > +
> > +static int kmb_ocs_hcu_init(struct ahash_request *req)
> > +{
> > + struct ocs_hcu_dev *hcu_dev = kmb_ocs_hcu_find_dev(req);
> > + struct ocs_hcu_rctx *rctx = ahash_request_ctx(req);
> > + struct crypto_ahash *tfm = crypto_ahash_reqtfm(req);
> > + struct ocs_hcu_ctx *ctx = crypto_ahash_ctx(tfm);
> > +
> > + if (!hcu_dev)
> > + return -ENOENT;
> > +
> > + /* Initialize entire request context to zero. */
> > + memset(rctx, 0, sizeof(*rctx));
> > +
> > + rctx->hcu_dev = hcu_dev;
> > + rctx->dig_sz = crypto_ahash_digestsize(tfm);
> > +
> > + switch (rctx->dig_sz) {
> > +#ifdef CONFIG_CRYPTO_DEV_KEEMBAY_OCS_HCU_HMAC_SHA224
> > + case SHA224_DIGEST_SIZE:
> > + rctx->blk_sz = SHA224_BLOCK_SIZE;
> > + rctx->algo = OCS_HCU_ALGO_SHA224;
> > + break;
> > +#endif /* CONFIG_CRYPTO_DEV_KEEMBAY_OCS_HCU_HMAC_SHA224 */
> why not always include sha224 support and remove this config option???
Discussed above.
>
> > + case SHA256_DIGEST_SIZE:
> > + rctx->blk_sz = SHA256_BLOCK_SIZE;
> > + /*
> > + * SHA256 and SM3 have the same digest size: use info from tfm
> > + * context to find out which one we should use.
> > + */
> > + rctx->algo = ctx->is_sm3_tfm ? OCS_HCU_ALGO_SM3 :
> > + OCS_HCU_ALGO_SHA256;
> > + break;
> > + case SHA384_DIGEST_SIZE:
> > + rctx->blk_sz = SHA384_BLOCK_SIZE;
> > + rctx->algo = OCS_HCU_ALGO_SHA384;
> > + break;
> > + case SHA512_DIGEST_SIZE:
> > + rctx->blk_sz = SHA512_BLOCK_SIZE;
> > + rctx->algo = OCS_HCU_ALGO_SHA512;
> > + break;
> > + default:
> > + return -EINVAL;
> > + }
> > +
> > + /* Clear the intermediate data. */
> > + memzero_explicit(&rctx->idata, sizeof(rctx->idata));
> > + /* If this is not a HMAC request, we are done. */
> > + if (!ctx->is_hmac_tfm)
> > + return 0;
> > +
> > + rctx->flags |= REQ_FLAGS_HMAC;
> > +
> > + return 0;
> > +}
> > +
> > +static int kmb_ocs_hcu_update(struct ahash_request *req)
> > +{
> > + struct ocs_hcu_rctx *rctx = ahash_request_ctx(req);
> > + int rc;
> > +
> > + if (!req->nbytes)
> > + return 0;
> > +
> > + /* Check for overflow */
> > + if (check_add_overflow(rctx->sg_data_total, req->nbytes,
> > + &rctx->sg_data_total))
> > + return -EINVAL;
> > +
> > + rctx->sg_data_offset = 0;
> > + rctx->sg = req->src;
> > +
> > + /*
> > + * If we are doing HMAC, then we must use SW-assisted HMAC, since HW
> > + * HMAC does not support context switching (so can only be used with
> > + * finup() or digest().
> > + */
> > + if (rctx->flags & REQ_FLAGS_HMAC &&
> > + !(rctx->flags & REQ_FLAGS_HMAC_SW)) {
> > + rctx->flags |= REQ_FLAGS_HMAC_SW;
> > + rc = prepare_ipad(req);
> > + if (rc)
> > + return rc;
> > + }
> > +
> > + /*
> > + * If remaining sg_data fits into ctx buffer, just copy it there; we'll
> > + * process it at the next update() or final().
> > + */
> > + if (rctx->sg_data_total <= (sizeof(rctx->buffer) - rctx->buf_cnt))
> > + return flush_sg_to_ocs_buffer(rctx);
> > +
> > + return kmb_ocs_hcu_handle_queue(req);
> > +}
> > +
> > +static int kmb_ocs_hcu_final(struct ahash_request *req)
> > +{
> > + struct ocs_hcu_rctx *rctx = ahash_request_ctx(req);
> > + struct crypto_ahash *tfm = crypto_ahash_reqtfm(req);
> > + struct ocs_hcu_ctx *ctx = crypto_ahash_ctx(tfm);
> > + int rc;
> > +
> > + rctx->sg = req->src;
> > + rctx->sg_data_offset = 0;
> > + rctx->flags |= REQ_FINAL;
> > +
> > + /*
> > + * If this is a HMAC request and, so far, we didn't have to switch to
> > + * SW HMAC, check if we can use HW HMAC.
> > + */
> > + if (rctx->flags & REQ_FLAGS_HMAC &&
> > + !(rctx->flags & REQ_FLAGS_HMAC_SW)) {
> > + /*
> > + * If we are here, it means we never processed any data so far,
> > + * so we can use HW HMAC, but only if there is some data to
> > + * process (since OCS HW MAC does not support zero-length
> > + * messages) and the key length is supported by the hardware
> > + * (OCS HCU HW only supports length <= 64); if HW HMAC cannot
> > + * be used, fall back to SW-assisted HMAC.
> > + */
> > + if (kmb_get_total_data(rctx) &&
> > + ctx->key_len <= HCU_HW_KEY_LEN) {
> > + rctx->algo |= OCS_HCU_ALGO_HMAC_MASK;
> > + rctx->flags |= REQ_FLAGS_HMAC_HW;
> > + } else {
> > + rctx->flags |= REQ_FLAGS_HMAC_SW;
> > + rc = prepare_ipad(req);
> > + if (rc)
> > + return rc;
> > + }
> > + }
> > + /*
> > + * If SW HMAC is being used, request OPAD computation (do in IRQ
> > + * thread, after the full message has been hashed).
> > + */
> > + if (rctx->flags & REQ_FLAGS_HMAC_SW)
> > + rctx->flags |= REQ_FLAGS_HMAC_SW_DO_OPAD;
> > +
> > + return kmb_ocs_hcu_handle_queue(req);
> > +}
> > +
> > +static int kmb_ocs_hcu_finup(struct ahash_request *req)
> > +{
> > + struct ocs_hcu_rctx *rctx = ahash_request_ctx(req);
> > +
> > + /* Check for overflow */
> > + if (check_add_overflow(rctx->sg_data_total, req->nbytes,
> > + &rctx->sg_data_total))
> > + return -EINVAL;
> > +
> > + return kmb_ocs_hcu_final(req);
> > +}
> > +
> > +static int kmb_ocs_hcu_digest(struct ahash_request *req)
> > +{
> > + int rc = 0;
> > + struct ocs_hcu_dev *hcu_dev = kmb_ocs_hcu_find_dev(req);
> > +
> > + if (!hcu_dev)
> > + return -ENOENT;
> > +
> > + rc = kmb_ocs_hcu_init(req);
> > + if (rc)
> > + return rc;
> > +
> > + rc = kmb_ocs_hcu_finup(req);
> > +
> > + return rc;
> > +}
> > +
> > +static int kmb_ocs_hcu_export(struct ahash_request *req, void *out)
> > +{
> > + struct ocs_hcu_rctx *rctx = ahash_request_ctx(req);
> > +
> > + /* Intermediate data is always stored and applied per request. */
> > + memcpy(out, rctx, sizeof(*rctx));
> > +
> > + return 0;
> > +}
> > +
> > +static int kmb_ocs_hcu_import(struct ahash_request *req, const void *in)
> > +{
> > + struct ocs_hcu_rctx *rctx = ahash_request_ctx(req);
> > +
> > + /* Intermediate data is always stored and applied per request. */
> > + memcpy(rctx, in, sizeof(*rctx));
> > +
> > + return 0;
> > +}
> > +
> > +static int kmb_ocs_hcu_setkey(struct crypto_ahash *tfm, const u8 *key,
> > + unsigned int keylen)
> > +{
> > + struct ocs_hcu_ctx *ctx = crypto_ahash_ctx(tfm);
> > + unsigned int digestsize = crypto_ahash_digestsize(tfm);
> > + size_t blk_sz = crypto_ahash_blocksize(tfm);
> > + struct crypto_ahash *ahash_tfm;
> > + struct ahash_request *req;
> > + struct crypto_wait wait;
> > + struct scatterlist sg;
> > + const char *alg_name;
> > + u8 *buf;
> > + int rc;
> > +
> > + /*
> > + * Key length must be equal to block size:
> > + * - If key is shorter, we are done for now (the key will be padded
> > + * later on); this is to maximize the use of HW HMAC (which works
> > + * only for keys <= 64 bytes).
> > + * - If key is longer, we hash it.
> > + */
> > + if (keylen <= blk_sz) {
> > + memcpy(ctx->key, key, keylen);
> > + ctx->key_len = keylen;
> > + return 0;
> > + }
> > +
> > + switch (digestsize) {
> > +#ifdef CONFIG_CRYPTO_DEV_KEEMBAY_OCS_HCU_HMAC_SHA224
> > + case SHA224_DIGEST_SIZE:
> > + alg_name = "sha224-keembay-ocs";
> > + break;
> > +#endif /* CONFIG_CRYPTO_DEV_KEEMBAY_OCS_HCU_HMAC_SHA224 */
> > + case SHA256_DIGEST_SIZE:
> > + alg_name = ctx->is_sm3_tfm ? "sm3-keembay-ocs" :
> > + "sha256-keembay-ocs";
> > + break;
> > + case SHA384_DIGEST_SIZE:
> > + alg_name = "sha384-keembay-ocs";
> > + break;
> > + case SHA512_DIGEST_SIZE:
> > + alg_name = "sha512-keembay-ocs";
> > + break;
> > + default:
> > + return -EINVAL;
> > + }
> > +
> > + ahash_tfm = crypto_alloc_ahash(alg_name, 0, 0);
> > + if (IS_ERR(ahash_tfm))
> > + return PTR_ERR(ahash_tfm);
> > +
> > + req = ahash_request_alloc(ahash_tfm, GFP_KERNEL);
> > + if (!req) {
> > + rc = -ENOMEM;
> > + goto err_free_ahash;
> > + }
> > +
> > + crypto_init_wait(&wait);
> > + ahash_request_set_callback(req, CRYPTO_TFM_REQ_MAY_BACKLOG,
> > + crypto_req_done, &wait);
> > + crypto_ahash_clear_flags(ahash_tfm, ~0);
> > +
> > + buf = kzalloc(keylen, GFP_KERNEL);
> > + if (!buf) {
> > + rc = -ENOMEM;
> > + goto err_free_req;
> > + }
> > +
> > + memcpy(buf, key, keylen);
> > + sg_init_one(&sg, buf, keylen);
> > + ahash_request_set_crypt(req, &sg, ctx->key, keylen);
> > +
> > + rc = crypto_wait_req(crypto_ahash_digest(req), &wait);
> > + if (rc == 0)
> > + ctx->key_len = digestsize;
> > +
> > + kfree(buf);
> > +err_free_req:
> > + ahash_request_free(req);
> > +err_free_ahash:
> > + crypto_free_ahash(ahash_tfm);
> > +
> > + return rc;
> > +}
> > +
> > +/* Set request size and initialize tfm context. */
> > +static void __cra_init(struct crypto_tfm *tfm, struct ocs_hcu_ctx *ctx)
> > +{
> > + crypto_ahash_set_reqsize(__crypto_ahash_cast(tfm),
> > + sizeof(struct ocs_hcu_rctx));
> > +
> > + /* Init context to 0. */
> > + memzero_explicit(ctx, sizeof(*ctx));
> > + /* Set engine ops. */
> > + ctx->engine_ctx.op.do_one_request = kmb_ocs_hcu_do_one_request;
> > + ctx->engine_ctx.op.prepare_request = kmb_ocs_hcu_prepare_request;
> > +}
> > +
> > +static int kmb_ocs_hcu_sha_cra_init(struct crypto_tfm *tfm)
> > +{
> > + struct ocs_hcu_ctx *ctx = crypto_tfm_ctx(tfm);
> > +
> > + __cra_init(tfm, ctx);
> > +
> > + return 0;
> > +}
> > +
> > +static int kmb_ocs_hcu_sm3_cra_init(struct crypto_tfm *tfm)
> > +{
> > + struct ocs_hcu_ctx *ctx = crypto_tfm_ctx(tfm);
> > +
> > + __cra_init(tfm, ctx);
> > +
> > + ctx->is_sm3_tfm = true;
> > +
> > + return 0;
> > +}
> > +
> > +static int kmb_ocs_hcu_hmac_sm3_cra_init(struct crypto_tfm *tfm)
> > +{
> > + struct ocs_hcu_ctx *ctx = crypto_tfm_ctx(tfm);
> > +
> > + __cra_init(tfm, ctx);
> > +
> > + ctx->is_sm3_tfm = true;
> > + ctx->is_hmac_tfm = true;
> > +
> > + return 0;
> > +}
> > +
> > +static int kmb_ocs_hcu_hmac_cra_init(struct crypto_tfm *tfm)
> > +{
> > + struct ocs_hcu_ctx *ctx = crypto_tfm_ctx(tfm);
> > +
> > + __cra_init(tfm, ctx);
> > +
> > + ctx->is_hmac_tfm = true;
> > +
> > + return 0;
> > +}
> > +
> > +/* Function called when 'tfm' is de-initialized. */
> > +static void kmb_ocs_hcu_hmac_cra_exit(struct crypto_tfm *tfm)
> > +{
> > + struct ocs_hcu_ctx *ctx = crypto_tfm_ctx(tfm);
> > +
> > + /* Clear the key. */
> > + memzero_explicit(ctx->key, sizeof(ctx->key));
> > +}
> > +
> > +static struct ahash_alg ocs_hcu_algs[] = {
> > +#ifdef CONFIG_CRYPTO_DEV_KEEMBAY_OCS_HCU_HMAC_SHA224
> > +{
> > + .init = kmb_ocs_hcu_init,
> > + .update = kmb_ocs_hcu_update,
> > + .final = kmb_ocs_hcu_final,
> > + .finup = kmb_ocs_hcu_finup,
> > + .digest = kmb_ocs_hcu_digest,
> > + .export = kmb_ocs_hcu_export,
> > + .import = kmb_ocs_hcu_import,
> > + .halg = {
> > + .digestsize = SHA224_DIGEST_SIZE,
> > + .statesize = sizeof(struct ocs_hcu_rctx),
> > + .base = {
> > + .cra_name = "sha224",
> > + .cra_driver_name = "sha224-keembay-ocs",
> > + .cra_priority = 255,
> > + .cra_flags = CRYPTO_ALG_ASYNC,
> > + .cra_blocksize = SHA224_BLOCK_SIZE,
> > + .cra_ctxsize = sizeof(struct ocs_hcu_ctx),
> > + .cra_alignmask = KMB_OCS_HCU_ALIGN_MASK,
> > + .cra_module = THIS_MODULE,
> > + .cra_init = kmb_ocs_hcu_sha_cra_init,
> > + }
> > + }
> > +},
> > +{
> > + .init = kmb_ocs_hcu_init,
> > + .update = kmb_ocs_hcu_update,
> > + .final = kmb_ocs_hcu_final,
> > + .finup = kmb_ocs_hcu_finup,
> > + .digest = kmb_ocs_hcu_digest,
> > + .export = kmb_ocs_hcu_export,
> > + .import = kmb_ocs_hcu_import,
> > + .setkey = kmb_ocs_hcu_setkey,
> > + .halg = {
> > + .digestsize = SHA224_DIGEST_SIZE,
> > + .statesize = sizeof(struct ocs_hcu_rctx),
> > + .base = {
> > + .cra_name = "hmac(sha224)",
> > + .cra_driver_name = "hmac-sha224-keembay-ocs",
> > + .cra_priority = 255,
> > + .cra_flags = CRYPTO_ALG_ASYNC,
> > + .cra_blocksize = SHA224_BLOCK_SIZE,
> > + .cra_ctxsize = sizeof(struct ocs_hcu_ctx),
> > + .cra_alignmask = KMB_OCS_HCU_ALIGN_MASK,
> > + .cra_module = THIS_MODULE,
> > + .cra_init = kmb_ocs_hcu_hmac_cra_init,
> > + .cra_exit = kmb_ocs_hcu_hmac_cra_exit,
> > + }
> > + }
> > +},
> > +#endif /* CONFIG_CRYPTO_DEV_KEEMBAY_OCS_HCU_HMAC_SHA224 */
> > +{
> > + .init = kmb_ocs_hcu_init,
> > + .update = kmb_ocs_hcu_update,
> > + .final = kmb_ocs_hcu_final,
> > + .finup = kmb_ocs_hcu_finup,
> > + .digest = kmb_ocs_hcu_digest,
> > + .export = kmb_ocs_hcu_export,
> > + .import = kmb_ocs_hcu_import,
> > + .halg = {
> > + .digestsize = SHA256_DIGEST_SIZE,
> > + .statesize = sizeof(struct ocs_hcu_rctx),
> > + .base = {
> > + .cra_name = "sha256",
> > + .cra_driver_name = "sha256-keembay-ocs",
> > + .cra_priority = 255,
> > + .cra_flags = CRYPTO_ALG_ASYNC,
> > + .cra_blocksize = SHA256_BLOCK_SIZE,
> > + .cra_ctxsize = sizeof(struct ocs_hcu_ctx),
> > + .cra_alignmask = KMB_OCS_HCU_ALIGN_MASK,
> > + .cra_module = THIS_MODULE,
> > + .cra_init = kmb_ocs_hcu_sha_cra_init,
> > + }
> > + }
> > +},
> > +{
> > + .init = kmb_ocs_hcu_init,
> > + .update = kmb_ocs_hcu_update,
> > + .final = kmb_ocs_hcu_final,
> > + .finup = kmb_ocs_hcu_finup,
> > + .digest = kmb_ocs_hcu_digest,
> > + .export = kmb_ocs_hcu_export,
> > + .import = kmb_ocs_hcu_import,
> > + .setkey = kmb_ocs_hcu_setkey,
> > + .halg = {
> > + .digestsize = SHA256_DIGEST_SIZE,
> > + .statesize = sizeof(struct ocs_hcu_rctx),
> > + .base = {
> > + .cra_name = "hmac(sha256)",
> > + .cra_driver_name = "hmac-sha256-keembay-ocs",
> > + .cra_priority = 255,
> > + .cra_flags = CRYPTO_ALG_ASYNC,
> > + .cra_blocksize = SHA256_BLOCK_SIZE,
> > + .cra_ctxsize = sizeof(struct ocs_hcu_ctx),
> > + .cra_alignmask = KMB_OCS_HCU_ALIGN_MASK,
> > + .cra_module = THIS_MODULE,
> > + .cra_init = kmb_ocs_hcu_hmac_cra_init,
> > + .cra_exit = kmb_ocs_hcu_hmac_cra_exit,
> > + }
> > + }
> > +},
> > +{
> > + .init = kmb_ocs_hcu_init,
> > + .update = kmb_ocs_hcu_update,
> > + .final = kmb_ocs_hcu_final,
> > + .finup = kmb_ocs_hcu_finup,
> > + .digest = kmb_ocs_hcu_digest,
> > + .export = kmb_ocs_hcu_export,
> > + .import = kmb_ocs_hcu_import,
> > + .halg = {
> > + .digestsize = SM3_DIGEST_SIZE,
> > + .statesize = sizeof(struct ocs_hcu_rctx),
> > + .base = {
> > + .cra_name = "sm3",
> > + .cra_driver_name = "sm3-keembay-ocs",
> > + .cra_priority = 255,
> > + .cra_flags = CRYPTO_ALG_ASYNC,
> > + .cra_blocksize = SM3_BLOCK_SIZE,
> > + .cra_ctxsize = sizeof(struct ocs_hcu_ctx),
> > + .cra_alignmask = KMB_OCS_HCU_ALIGN_MASK,
> > + .cra_module = THIS_MODULE,
> > + .cra_init = kmb_ocs_hcu_sm3_cra_init,
> > + }
> > + }
> > +},
> > +{
> > + .init = kmb_ocs_hcu_init,
> > + .update = kmb_ocs_hcu_update,
> > + .final = kmb_ocs_hcu_final,
> > + .finup = kmb_ocs_hcu_finup,
> > + .digest = kmb_ocs_hcu_digest,
> > + .export = kmb_ocs_hcu_export,
> > + .import = kmb_ocs_hcu_import,
> > + .setkey = kmb_ocs_hcu_setkey,
> > + .halg = {
> > + .digestsize = SM3_DIGEST_SIZE,
> > + .statesize = sizeof(struct ocs_hcu_rctx),
> > + .base = {
> > + .cra_name = "hmac(sm3)",
> > + .cra_driver_name = "hmac-sm3-keembay-ocs",
> > + .cra_priority = 255,
> > + .cra_flags = CRYPTO_ALG_ASYNC,
> > + .cra_blocksize = SM3_BLOCK_SIZE,
> > + .cra_ctxsize = sizeof(struct ocs_hcu_ctx),
> > + .cra_alignmask = KMB_OCS_HCU_ALIGN_MASK,
> > + .cra_module = THIS_MODULE,
> > + .cra_init = kmb_ocs_hcu_hmac_sm3_cra_init,
> > + .cra_exit = kmb_ocs_hcu_hmac_cra_exit,
> > + }
> > + }
> > +},
> > +{
> > + .init = kmb_ocs_hcu_init,
> > + .update = kmb_ocs_hcu_update,
> > + .final = kmb_ocs_hcu_final,
> > + .finup = kmb_ocs_hcu_finup,
> > + .digest = kmb_ocs_hcu_digest,
> > + .export = kmb_ocs_hcu_export,
> > + .import = kmb_ocs_hcu_import,
> > + .halg = {
> > + .digestsize = SHA384_DIGEST_SIZE,
> > + .statesize = sizeof(struct ocs_hcu_rctx),
> > + .base = {
> > + .cra_name = "sha384",
> > + .cra_driver_name = "sha384-keembay-ocs",
> > + .cra_priority = 255,
> > + .cra_flags = CRYPTO_ALG_ASYNC,
> > + .cra_blocksize = SHA384_BLOCK_SIZE,
> > + .cra_ctxsize = sizeof(struct ocs_hcu_ctx),
> > + .cra_alignmask = KMB_OCS_HCU_ALIGN_MASK,
> > + .cra_module = THIS_MODULE,
> > + .cra_init = kmb_ocs_hcu_sha_cra_init,
> > + }
> > + }
> > +},
> > +{
> > + .init = kmb_ocs_hcu_init,
> > + .update = kmb_ocs_hcu_update,
> > + .final = kmb_ocs_hcu_final,
> > + .finup = kmb_ocs_hcu_finup,
> > + .digest = kmb_ocs_hcu_digest,
> > + .export = kmb_ocs_hcu_export,
> > + .import = kmb_ocs_hcu_import,
> > + .setkey = kmb_ocs_hcu_setkey,
> > + .halg = {
> > + .digestsize = SHA384_DIGEST_SIZE,
> > + .statesize = sizeof(struct ocs_hcu_rctx),
> > + .base = {
> > + .cra_name = "hmac(sha384)",
> > + .cra_driver_name = "hmac-sha384-keembay-ocs",
> > + .cra_priority = 255,
> > + .cra_flags = CRYPTO_ALG_ASYNC,
> > + .cra_blocksize = SHA384_BLOCK_SIZE,
> > + .cra_ctxsize = sizeof(struct ocs_hcu_ctx),
> > + .cra_alignmask = KMB_OCS_HCU_ALIGN_MASK,
> > + .cra_module = THIS_MODULE,
> > + .cra_init = kmb_ocs_hcu_hmac_cra_init,
> > + .cra_exit = kmb_ocs_hcu_hmac_cra_exit,
> > + }
> > + }
> > +},
> > +{
> > + .init = kmb_ocs_hcu_init,
> > + .update = kmb_ocs_hcu_update,
> > + .final = kmb_ocs_hcu_final,
> > + .finup = kmb_ocs_hcu_finup,
> > + .digest = kmb_ocs_hcu_digest,
> > + .export = kmb_ocs_hcu_export,
> > + .import = kmb_ocs_hcu_import,
> > + .halg = {
> > + .digestsize = SHA512_DIGEST_SIZE,
> > + .statesize = sizeof(struct ocs_hcu_rctx),
> > + .base = {
> > + .cra_name = "sha512",
> > + .cra_driver_name = "sha512-keembay-ocs",
> > + .cra_priority = 255,
> > + .cra_flags = CRYPTO_ALG_ASYNC,
> > + .cra_blocksize = SHA512_BLOCK_SIZE,
> > + .cra_ctxsize = sizeof(struct ocs_hcu_ctx),
> > + .cra_alignmask = KMB_OCS_HCU_ALIGN_MASK,
> > + .cra_module = THIS_MODULE,
> > + .cra_init = kmb_ocs_hcu_sha_cra_init,
> > + }
> > + }
> > +},
> > +{
> > + .init = kmb_ocs_hcu_init,
> > + .update = kmb_ocs_hcu_update,
> > + .final = kmb_ocs_hcu_final,
> > + .finup = kmb_ocs_hcu_finup,
> > + .digest = kmb_ocs_hcu_digest,
> > + .export = kmb_ocs_hcu_export,
> > + .import = kmb_ocs_hcu_import,
> > + .setkey = kmb_ocs_hcu_setkey,
> > + .halg = {
> > + .digestsize = SHA512_DIGEST_SIZE,
> > + .statesize = sizeof(struct ocs_hcu_rctx),
> > + .base = {
> > + .cra_name = "hmac(sha512)",
> > + .cra_driver_name = "hmac-sha512-keembay-ocs",
> > + .cra_priority = 255,
> > + .cra_flags = CRYPTO_ALG_ASYNC,
> > + .cra_blocksize = SHA512_BLOCK_SIZE,
> > + .cra_ctxsize = sizeof(struct ocs_hcu_ctx),
> > + .cra_alignmask = KMB_OCS_HCU_ALIGN_MASK,
> > + .cra_module = THIS_MODULE,
> > + .cra_init = kmb_ocs_hcu_hmac_cra_init,
> > + .cra_exit = kmb_ocs_hcu_hmac_cra_exit,
> > + }
> > + }
> > +}
> > +
> > +};
> > +
> > +static irqreturn_t kmb_ocs_hcu_irq_thread(int irq, void *dev_id)
> > +{
> > + struct ocs_hcu_dev *hcu_dev = (struct ocs_hcu_dev *)dev_id;
> > + struct ocs_hcu_rctx *rctx = ahash_request_ctx(hcu_dev->req);
> > + struct crypto_ahash *tfm = crypto_ahash_reqtfm(hcu_dev->req);
> > + struct ocs_hcu_ctx *ctx = crypto_ahash_ctx(tfm);
> > + int rc = 0;
> > + int i;
> > +
> > + if (hcu_dev->flags & HCU_FLAGS_HCU_ERROR_MASK) {
> > + rc = -EIO;
> > + goto finish;
> > + }
> > +
> > + /*
> > + * If the DO_OPAD flag is set, it means that we are finalizing a SW
> > + * HMAC request and we just computed the result of H(k ^ ipad || m).
> > + *
> > + * We now need to complete the HMAC calculation with the OPAD step,
> > + * that is, we need to compute H(k ^ opad || digest), where digest is
> > + * the digest we just obtained, i.e., H(k ^ ipad || m).
> > + */
> > + if (rctx->flags & REQ_FLAGS_HMAC_SW_DO_OPAD) {
> > + /* Gets the intermediate data (which includes the digest). */
> > + rc = ocs_hcu_get_intermediate_data(hcu_dev, &rctx->idata,
> > + rctx->algo);
> > + if (rc)
> > + goto finish;
> > + /*
> > + * Compute k ^ opad and store it in the request buffer (which
> > + * is not used anymore at this point).
> > + * Note: key has been padded / hashed already (so keylen ==
> > + * blksz) .
> > + */
> > + WARN_ON(ctx->key_len != rctx->blk_sz);
> is the output of the actionable by an admin?
An admin should just report the bug to the driver maintainers.
Anyway, v4 tries to improve this message.
>
>
> > + for (i = 0; i < rctx->blk_sz; i++)
> > + rctx->buffer[i] = ctx->key[i] ^ HMAC_OPAD_VALUE;
> > + /* Now append the digest to the rest of the buffer. */
> > + for (i = 0; (i < rctx->dig_sz); i++)
> > + rctx->buffer[rctx->blk_sz + i] = rctx->idata.digest[i];
> > + /*
> > + * Remove the DO_OPAD flag (so we won't end up here again after
> > + * final hash is computed).
> > + */
> > + rctx->flags &= ~(REQ_FLAGS_HMAC_SW_DO_OPAD);
> > + /* Now hash the buffer to obtain the final HMAC. */
> > + rc = kmb_ocs_hcu_hash_pio(hcu_dev, rctx->buffer,
> > + rctx->blk_sz + rctx->dig_sz,
> > + rctx->algo, true);
> > + if (rc)
> > + goto finish;
> > +
> > + return IRQ_HANDLED;
> > + }
> > +
> > +finish:
> > + kmb_ocs_hcu_finish_request(hcu_dev, rc);
> > +
> > + return IRQ_HANDLED;
> > +}
> > +
> > +/* Device tree driver match. */
> > +static const struct of_device_id kmb_ocs_hcu_of_match[] = {
> > + {
> > + .compatible = "intel,keembay-ocs-hcu",
> > + },
> > + {}
> > +};
> > +
> > +static int kmb_ocs_hcu_remove(struct platform_device *pdev)
> > +{
> > + struct ocs_hcu_dev *hcu_dev;
> > + int rc;
> > +
> > + hcu_dev = platform_get_drvdata(pdev);
> > + if (!hcu_dev)
> > + return -ENODEV;
> > +
> > + crypto_unregister_ahashes(ocs_hcu_algs, ARRAY_SIZE(ocs_hcu_algs));
> > +
> > + spin_lock(&ocs_hcu.lock);
> my not _bh this time?
> would a mutext work just as well?
> (note: I'm just adking. I'm not sure.)
Other crypto device drivers are using the simple spin_lock() /
spin_unlock() here. Not sure why, I'll change it to spin_lock_bh()
because I think it's more correct.
>
> > + list_del(&hcu_dev->list);
> > + spin_unlock(&ocs_hcu.lock);
> > +
> > + rc = crypto_engine_exit(hcu_dev->engine);
> > +
> > + ocs_hcu_hw_disable(hcu_dev);
> > +
> > + return rc;
> > +}
> > +
> > +static int kmb_ocs_hcu_probe(struct platform_device *pdev)
> > +{
> > + struct device *dev = &pdev->dev;
> > + struct ocs_hcu_dev *hcu_dev;
> > + struct resource *hcu_mem;
> > + int rc;
> > +
> > + hcu_dev = devm_kzalloc(dev, sizeof(*hcu_dev), GFP_KERNEL);
> > + if (!hcu_dev)
> > + return -ENOMEM;
> > +
> > + hcu_dev->dev = dev;
> > +
> > + platform_set_drvdata(pdev, hcu_dev);
> > + rc = dma_set_mask_and_coherent(&pdev->dev, OCS_HCU_DMA_BIT_MASK);
> > + if (rc)
> > + return rc;
> > +
> > + INIT_LIST_HEAD(&hcu_dev->list);
> > +
> > + /* Get the memory address and remap. */
> > + hcu_mem = platform_get_resource(pdev, IORESOURCE_MEM, 0);
> > + if (!hcu_mem) {
> > + dev_err(dev, "Could not retrieve io mem resource.\n");
> > + rc = -ENODEV;
> > + goto list_del;
> > + }
> > +
> > + hcu_dev->io_base = devm_ioremap_resource(dev, hcu_mem);
> > + if (IS_ERR(hcu_dev->io_base)) {
> > + dev_err(dev, "Could not io-remap mem resource.\n");
> > + rc = PTR_ERR(hcu_dev->io_base);
> > + goto list_del;
> > + }
> > +
> > + /* Get and request IRQ. */
> > + hcu_dev->irq = platform_get_irq(pdev, 0);
> > + if (hcu_dev->irq < 0) {
> > + dev_err(dev, "Could not retrieve IRQ.\n");
> > + rc = hcu_dev->irq;
> > + goto list_del;
> > + }
> > +
> > + rc = devm_request_threaded_irq(&pdev->dev, hcu_dev->irq,
> > + ocs_hcu_irq_handler,
> > + kmb_ocs_hcu_irq_thread,
> > + 0, "keembay-ocs-hcu", hcu_dev);
> > + if (rc < 0) {
> > + dev_err(dev, "Could not request IRQ.\n");
> > + goto list_del;
> > + }
> > +
> > + spin_lock(&ocs_hcu.lock);
> Its good practice to hold the lock with addint to the list but, is it
> possible to have a concurancy issue in a probe function?
Same as above. I think we can have a concurrency issue if another
device is already probed and running (this is only theoretical in Keem
Bay, since there is only one OCS HCU device).
>
> > + list_add_tail(&hcu_dev->list, &ocs_hcu.dev_list);
> > + spin_unlock(&ocs_hcu.lock);
> > +
> > + /* Initialize crypto engine */
> > + hcu_dev->engine = crypto_engine_alloc_init(dev, 1);
> > + if (!hcu_dev->engine)
> > + goto list_del;
> > +
> > + rc = crypto_engine_start(hcu_dev->engine);
> > + if (rc) {
> > + dev_err(dev, "Could not start engine.\n");
> > + goto cleanup;
> > + }
> > +
> > + /* Security infrastructure guarantees OCS clock is enabled. */
> > +
> > + rc = crypto_register_ahashes(ocs_hcu_algs, ARRAY_SIZE(ocs_hcu_algs));
> > + if (rc) {
> > + dev_err(dev, "Could not register algorithms.\n");
> > + goto cleanup;
> > + }
> > +
> > + return 0;
> > +cleanup:
> > + crypto_engine_exit(hcu_dev->engine);
> > +list_del:
> > + spin_lock(&ocs_hcu.lock);
> > + list_del(&hcu_dev->list);
> > + spin_unlock(&ocs_hcu.lock);
> > +
> > + return rc;
> > +}
> > +
> > +/* The OCS driver is a platform device. */
> > +static struct platform_driver kmb_ocs_hcu_driver = {
> > + .probe = kmb_ocs_hcu_probe,
> > + .remove = kmb_ocs_hcu_remove,
> > + .driver = {
> > + .name = DRV_NAME,
> > + .of_match_table = kmb_ocs_hcu_of_match,
> > + },
> > +};
> > +
> > +module_platform_driver(kmb_ocs_hcu_driver);
> > +
> > +MODULE_LICENSE("GPL");
> > diff --git a/drivers/crypto/keembay/ocs-hcu.c b/drivers/crypto/keembay/ocs-hcu.c
> > new file mode 100644
> > index 000000000000..5523a03975a7
> > --- /dev/null
> > +++ b/drivers/crypto/keembay/ocs-hcu.c
> > @@ -0,0 +1,593 @@
> > +// SPDX-License-Identifier: GPL-2.0-only
> > +/*
> > + * Intel Keem Bay OCS HCU Crypto Driver.
> > + *
> > + * Copyright (C) 2018-2020 Intel Corporation
> > + */
> > +
> > +#include <linux/delay.h>
> > +#include <linux/device.h>
> > +#include <linux/iopoll.h>
> > +#include <linux/irq.h>
> > +#include <linux/module.h>
> > +
> > +#include <crypto/sha.h>
> > +
> > +#include "ocs-hcu.h"
> > +
> > +#define OCS_HCU_MODE 0x00
> > +#define OCS_HCU_CHAIN 0x04
> > +#define OCS_HCU_OPERATION 0x08
> > +#define OCS_HCU_KEY_0 0x0C
> > +#define OCS_HCU_KEY_1 0x10
> > +#define OCS_HCU_KEY_2 0x14
> > +#define OCS_HCU_KEY_3 0x18
> > +#define OCS_HCU_KEY_4 0x1C
> > +#define OCS_HCU_KEY_5 0x20
> > +#define OCS_HCU_KEY_6 0x24
> > +#define OCS_HCU_KEY_7 0x28
> > +#define OCS_HCU_KEY_8 0x2C
> > +#define OCS_HCU_KEY_9 0x30
> > +#define OCS_HCU_KEY_10 0x34
> > +#define OCS_HCU_KEY_11 0x38
> > +#define OCS_HCU_KEY_12 0x3C
> > +#define OCS_HCU_KEY_13 0x40
> > +#define OCS_HCU_KEY_14 0x44
> > +#define OCS_HCU_KEY_15 0x48
> > +#define OCS_HCU_ISR 0x50
> > +#define OCS_HCU_IER 0x54
> > +#define OCS_HCU_STATUS 0x58
> > +#define OCS_HCU_MSG_LEN_LO 0x60
> > +#define OCS_HCU_MSG_LEN_HI 0x64
> > +#define OCS_HCU_KEY_BYTE_ORDER_CFG 0x80
> > +#define OCS_HCU_DMA_SRC_ADDR 0x400
> > +#define OCS_HCU_DMA_DST_ADDR 0x404
> > +#define OCS_HCU_DMA_SRC_SIZE 0x408
> > +#define OCS_HCU_DMA_DST_SIZE 0x40C
> > +#define OCS_HCU_DMA_DMA_MODE 0x410
> > +#define OCS_HCU_DMA_OTHER_MODE 0x414
> > +#define OCS_HCU_DMA_NEXT_SRC_DESCR 0x418
> > +#define OCS_HCU_DMA_NEXT_DST_DESCR 0x41C
> > +#define OCS_HCU_DMA_WHILE_ACTIVE_MODE 0x420
> > +#define OCS_HCU_DMA_LOG 0x424
> > +#define OCS_HCU_DMA_STATUS 0x428
> > +#define OCS_HCU_DMA_PERF_CNTR 0x42C
> > +#define OCS_HCU_DMA_VALID_SAI_31_0 0x440
> > +#define OCS_HCU_DMA_VALID_SAI_63_32 0x444
> > +#define OCS_HCU_DMA_VALID_SAI_95_64 0x448
> > +#define OCS_HCU_DMA_VALID_SAI_127_96 0x44C
> > +#define OCS_HCU_DMA_VALID_SAI_159_128 0x450
> > +#define OCS_HCU_DMA_VALID_SAI_191_160 0x454
> > +#define OCS_HCU_DMA_VALID_SAI_223_192 0x458
> > +#define OCS_HCU_DMA_VALID_SAI_255_224 0x45C
> > +#define OCS_HCU_DMA_MSI_ISR 0x480
> > +#define OCS_HCU_DMA_MSI_IER 0x484
> > +#define OCS_HCU_DMA_MSI_MASK 0x488
> > +#define OCS_HCU_DMA_MSI_MA 0x800
> > +#define OCS_HCU_DMA_MSI_DA 0x804
> > +#define OCS_HCU_DMA_MSI_EN 0x808
> > +#define OCS_HCU_DMA_INBUFFER_WRITE_FIFO 0x600
> > +#define OCS_HCU_DMA_OUTBUFFER_READ_FIFO 0x700
> > +
> > +/* Register bit definitions. */
> > +#define HCU_MODE_ALGO_SHIFT 16
> > +#define HCU_MODE_ALGO_MASK (OCS_HCU_ALGO_MASK \
> > + << HCU_MODE_ALGO_SHIFT)
> > +
> > +#define HCU_STATUS_BUSY_MASK BIT(0)
> > +
> > +#define HCU_BYTE_ORDER_SWAP BIT(0)
> > +
> > +#define HCU_IRQ_HASH_DONE BIT(2)
> > +#define HCU_IRQ_HASH_ERR (BIT(3) | BIT(1) | BIT(0))
> > +
> > +#define HCU_DMA_IRQ_SRC_DONE BIT(0)
> > +#define HCU_DMA_IRQ_DST_DONE BIT(1)
> > +#define HCU_DMA_IRQ_SAI_ERR BIT(2)
> > +#define HCU_DMA_IRQ_BAD_COMP_ERR BIT(3)
> > +#define HCU_DMA_IRQ_INBUF_RD_ERR BIT(4)
> > +#define HCU_DMA_IRQ_INBUF_WD_ERR BIT(5)
> > +#define HCU_DMA_IRQ_OUTBUF_WR_ERR BIT(6)
> > +#define HCU_DMA_IRQ_OUTBUF_RD_ERR BIT(7)
> > +#define HCU_DMA_IRQ_CRD_ERR BIT(8)
> > +#define HCU_DMA_IRQ_ERR_MASK (HCU_DMA_IRQ_SAI_ERR | \
> > + HCU_DMA_IRQ_BAD_COMP_ERR | \
> > + HCU_DMA_IRQ_INBUF_RD_ERR | \
> > + HCU_DMA_IRQ_INBUF_WD_ERR | \
> > + HCU_DMA_IRQ_OUTBUF_WR_ERR | \
> > + HCU_DMA_IRQ_OUTBUF_RD_ERR | \
> > + HCU_DMA_IRQ_CRD_ERR)
> > +
> > +#define HCU_DMA_SNOOP_MASK (0x7 << 28)
> > +#define HCU_DMA_SRC_LL_EN BIT(25)
> > +#define HCU_DMA_EN BIT(31)
> > +#define HCU_DMA_STAT_SRC_DONE BIT(15)
> > +
> > +#define HCU_MODE_HMAC_SHIFT 22
> > +#define HCU_MODE_HMAC_MASK BIT(HCU_MODE_HMAC_SHIFT)
> > +
> > +#define KMB_HCU_ENDIANNESS_VALUE (0x2A)
> > +
> > +#define HCU_DMA_MSI_UNMASK BIT(0)
> > +#define HCU_DMA_MSI_DISABLE 0
> > +#define HCU_IRQ_DISABLE 0
> > +
> > +#define OCS_HCU_START BIT(0)
> > +#define OCS_HCU_TERMINATE BIT(1)
> > +
> > +#define HCU_CHAIN_WRITE_ENDIANNESS_OFFSET 30
> > +#define HCU_CHAIN_READ_ENDIANNESS_OFFSET 28
> > +#define HCU_DATA_WRITE_ENDIANNESS_OFFSET 26
> > +
> > +#define OCS_HCU_NUM_CHAINS_SHA256_224_SM3 8
> > +#define OCS_HCU_NUM_CHAINS_SHA384_512 16
> > +
> > +#define OCS_HCU_HASH_FINAL_CPU_RETRIES 1000
> > +/*
> > + * While polling on a busy HCU, wait maximum 200us between one check and the
> > + * other.
> > + */
> > +#define OCS_HCU_WAIT_BUSY_RETRY_DELAY_US 200
> > +/* Wait on a busy HCU for maximum 1 second. */
> > +#define OCS_HCU_WAIT_BUSY_TIMEOUT_US 1000000
> > +
> > +static inline u32 ocs_hcu_num_chains(u32 algo)
> > +{
> > + switch (algo & OCS_HCU_ALGO_MASK) {
> > + case OCS_HCU_ALGO_SHA224:
> why not gard with and #ifdef config blabla sha224 here?
Good question. I've tried to separate the code handling the HW from the
one implementing the crypto API. The ocs-hcu.c driver only contains the
fist kind of code, so I didn't put any define related to the second.
I'd prefer to keep it this way if possible.
>
> > + case OCS_HCU_ALGO_SHA256:
> > + case OCS_HCU_ALGO_SM3:
> > + return OCS_HCU_NUM_CHAINS_SHA256_224_SM3;
> > + case OCS_HCU_ALGO_SHA384:
> > + case OCS_HCU_ALGO_SHA512:
> > + return OCS_HCU_NUM_CHAINS_SHA384_512;
> > + default:
> > + return 0;
> > + };
> > +}
> > +
> > +static inline u32 ocs_hcu_block_size(u32 algo)
> > +{
> > + switch (algo & OCS_HCU_ALGO_MASK) {
> > + case OCS_HCU_ALGO_SHA224:
> > + return SHA224_BLOCK_SIZE;
> why not gard with and #ifdef config blabla sha224 here?
Same as above.
>
> > + case OCS_HCU_ALGO_SHA256:
> > + case OCS_HCU_ALGO_SM3:
> > + /* SM3 shares the same block size. */
> > + return SHA256_BLOCK_SIZE;
> > + case OCS_HCU_ALGO_SHA384:
> > + return SHA384_BLOCK_SIZE;
> > + case OCS_HCU_ALGO_SHA512:
> > + return SHA512_BLOCK_SIZE;
> > + default:
> > + return 0;
> > + }
> > +}
> > +
> > +/**
> > + * ocs_hcu_wait_busy() - Wait for HCU OCS hardware to became usable.
> > + * @hcu_dev: OCS HCU device to wait for.
> > + *
> > + * Return: 0 if device free, -ETIMEOUT if device busy and internal timeout has
> > + * expired.
> > + */
> > +int ocs_hcu_wait_busy(struct ocs_hcu_dev *hcu_dev)
> > +{
> > + long val;
> > +
> > + return readl_poll_timeout(hcu_dev->io_base + OCS_HCU_STATUS, val,
> > + !(val & HCU_STATUS_BUSY_MASK),
> > + OCS_HCU_WAIT_BUSY_RETRY_DELAY_US,
> > + OCS_HCU_WAIT_BUSY_TIMEOUT_US);
> > +}
> > +
> > +static void ocs_hcu_done_irq_en(struct ocs_hcu_dev *hcu_dev)
> > +{
> > + /* Clear any pending interrupts. */
> > + writel(0xFFFFFFFF, hcu_dev->io_base + OCS_HCU_ISR);
> > + /* Enable error and HCU done interrupts. */
> > + writel(HCU_IRQ_HASH_DONE | HCU_IRQ_HASH_ERR,
> > + hcu_dev->io_base + OCS_HCU_IER);
> > +}
> > +
> > +static void ocs_hcu_dma_irq_en(struct ocs_hcu_dev *hcu_dev)
> > +{
> > + /* Clear any pending interrupts. */
> > + writel(0xFFFFFFFF, hcu_dev->io_base + OCS_HCU_DMA_MSI_ISR);
> > + /* Only operating on DMA source completion and error interrupts. */
> > + writel(HCU_DMA_IRQ_ERR_MASK | HCU_DMA_IRQ_SRC_DONE,
> > + hcu_dev->io_base + OCS_HCU_DMA_MSI_IER);
> > + /* Unmask */
> > + writel(HCU_DMA_MSI_UNMASK, hcu_dev->io_base + OCS_HCU_DMA_MSI_MASK);
> > +}
> > +
> > +static void ocs_hcu_irq_dis(struct ocs_hcu_dev *hcu_dev)
> > +{
> > + writel(HCU_IRQ_DISABLE, hcu_dev->io_base + OCS_HCU_IER);
> > + writel(HCU_DMA_MSI_DISABLE, hcu_dev->io_base + OCS_HCU_DMA_MSI_IER);
> > +}
> > +
> > +/**
> > + * ocs_hcu_get_intermediate_data() - Get intermediate data.
> > + * @hcu_dev: The target HCU device.
> > + * @data: Where to store the intermediate.
> > + * @algo: The algorithm being used.
> > + *
> > + * This function is used to save the current hashing process state in order to
> > + * continue it in the future.
> > + *
> > + * Note: once all data has been processed, the intermediate data actually
> > + * contains the hashing result. So this function is also used to retrieve the
> > + * final result of a hashing process.
> > + *
> > + * Return: 0 on success, negative error code otherwise.
> > + */
> > +int ocs_hcu_get_intermediate_data(struct ocs_hcu_dev *hcu_dev,
> > + struct ocs_hcu_idata *data, u32 algo)
> > +{
> > + const int n = ocs_hcu_num_chains(algo);
> > + u32 *chain;
> > + int busy;
> > + int i;
> > +
> > + /* Data not requested. */
> > + if (!data)
> > + return -EINVAL;
> > +
> > + chain = (u32 *)data->digest;
> > +
> > + /* Ensure that the OCS is no longer busy before reading the chains. */
> > + busy = ocs_hcu_wait_busy(hcu_dev);
> > +
> > + if (busy)
> > + return -EBUSY;
> > +
> > + for (i = 0; i < n; i++)
> > + chain[i] = readl(hcu_dev->io_base + OCS_HCU_CHAIN);
>
> How do we know the size of the chain array (data->direst) is >=
> ocs_hcu_num_chains? looks like a bug waiting to happen.
> is it worth a check? maybe add a comment to explain why the loop is safe WRT
> the number of chains the data-> digest passed in.
idata->digest is an array of SHA512_DIGEST_SIZE bytes and the maximum
value othat ocs_hcu_num_chains() can return
is OCS_HCU_NUM_CHAINS_SHA384_512 which is SHA512_DIGEST_SIZE / 4.
But you are right, I need a comment to explain this.
>
> > +
> > + data->msg_len_lo = readl(hcu_dev->io_base + OCS_HCU_MSG_LEN_LO);
> > + data->msg_len_hi = readl(hcu_dev->io_base + OCS_HCU_MSG_LEN_HI);
> > +
> > + return 0;
> > +}
> > +
> > +/**
> > + * ocs_hcu_set_intermediate_data() - Set intermediate data.
> > + * @hcu_dev: The target HCU device.
> > + * @data: The intermediate data to be set.
> > + * @algo: The algorithm being used.
> > + *
> > + * This function is used to continue a previous hashing process.
> > + */
> > +void ocs_hcu_set_intermediate_data(struct ocs_hcu_dev *hcu_dev,
> > + struct ocs_hcu_idata *data, u32 algo)
> > +{
> > + const int n = ocs_hcu_num_chains(algo);
> > + u32 *chain = (u32 *)data->digest;
> > + int i;
> > +
> > + for (i = 0; i < n; i++)
> > + writel(chain[i], hcu_dev->io_base + OCS_HCU_CHAIN);
> same issue here how can we be confident the loop will not go past the end of
> the chain[].
I'll add a comment here too.
>
> > +
> > + writel(data->msg_len_lo, hcu_dev->io_base + OCS_HCU_MSG_LEN_LO);
> > + writel(data->msg_len_hi, hcu_dev->io_base + OCS_HCU_MSG_LEN_HI);
> > +}
> > +
> > +/**
> > + * ocs_hcu_hw_init() - Initialize the HCU device.
> > + * @hcu_dev: The HCU device to initialize.
> > + */
> > +void ocs_hcu_hw_init(struct ocs_hcu_dev *hcu_dev)
> > +{
> > + u32 cfg = 0;
> looks like this initialzation is a waist of cycles. remove it.
I reworked the code and removed this function all together in v4.
>
> > +
> > + /* Return if HW is already initialized. */
> > + if (hcu_dev->flags & HCU_FLAGS_HCU_INIT)
> > + return;
> > + /* Initialize hardware. */
> > + cfg = KMB_HCU_ENDIANNESS_VALUE << HCU_DATA_WRITE_ENDIANNESS_OFFSET;
> > + writel(cfg, hcu_dev->io_base + OCS_HCU_MODE);
> > + hcu_dev->flags |= HCU_FLAGS_HCU_INIT;
> > +}
> > +
> > +/**
> > + * ocs_hcu_hw_cfg() - Configure the HCU hardware.
> > + * @hcu_dev: The HCU device to configure.
> > + * @algo: The algorithm to be used by the HCU device.
> > + *
> > + * NOTE: This function must be called after ocs_hcu_hw_init().
> > + *
> > + * Return: 0 on success, negative error code otherwise.
> > + */
> > +int ocs_hcu_hw_cfg(struct ocs_hcu_dev *hcu_dev, u32 algo)
> > +{
> > + u32 cfg = readl(hcu_dev->io_base + OCS_HCU_MODE);
> > + u32 ocs_algo = algo & OCS_HCU_ALGO_MASK;
> > +
> > + if (ocs_algo != OCS_HCU_ALGO_SHA256 &&
> > + ocs_algo != OCS_HCU_ALGO_SHA224 &&
> > + ocs_algo != OCS_HCU_ALGO_SHA384 &&
> > + ocs_algo != OCS_HCU_ALGO_SHA512 &&
> > + ocs_algo != OCS_HCU_ALGO_SM3)
> > + return -EINVAL;
> > +
> > + if ((hcu_dev->flags & HCU_FLAGS_HCU_INIT) == 0)
> > + return -EPERM;
> > +
> > + cfg &= ~HCU_MODE_ALGO_MASK;
> > + cfg |= ocs_algo << HCU_MODE_ALGO_SHIFT;
> > + cfg &= ~HCU_MODE_HMAC_MASK;
> > + if (algo & OCS_HCU_ALGO_HMAC_MASK)
> > + cfg |= BIT(HCU_MODE_HMAC_SHIFT);
> > +
> > + writel(cfg, hcu_dev->io_base + OCS_HCU_MODE);
> > +
> > + return 0;
> > +}
> > +
> > +/**
> > + * ocs_hcu_hw_disable() - Disable the HCU hardware.
> > + * @hcu_dev: The HCU device to disable.
> > + */
> > +void ocs_hcu_hw_disable(struct ocs_hcu_dev *hcu_dev)
> > +{
> > + if ((hcu_dev->flags & HCU_FLAGS_HCU_INIT) == HCU_FLAGS_HCU_INIT) {
> > + /* Clear hardware. */
> > + writel(0, hcu_dev->io_base + OCS_HCU_MODE);
> > + ocs_hcu_irq_dis(hcu_dev);
> > + hcu_dev->flags &= ~HCU_FLAGS_HCU_INIT;
> > + }
> > +}
> > +
> > +/**
> > + * ocs_hcu_tx_data_done() - Request the hardware to compute the final hash.
> > + * @hcu_dev: The HCU device to use.
> > + */
> > +void ocs_hcu_tx_data_done(struct ocs_hcu_dev *hcu_dev)
> > +{
> > + /*
> > + * Enable HCU interrupts, so that HCU_DONE will be triggered once the
> > + * final hash is computed.
> > + */
> > + ocs_hcu_done_irq_en(hcu_dev);
> > + writel(OCS_HCU_TERMINATE, hcu_dev->io_base + OCS_HCU_OPERATION);
> > +}
> > +
> > +static unsigned int ocs_hcu_hash_final_cpu(struct ocs_hcu_dev *hcu_dev,
> what's a final_cpu?
Function removed in v4.
>
> > + u8 *buf, u32 sz)
> > +{
> > + int retries = OCS_HCU_HASH_FINAL_CPU_RETRIES;
> > + u32 sz_32 = sz / sizeof(u32);
> > + u32 *buf_32 = (u32 *)buf;
> > + int i;
> > +
> > + /* Write in using full register size. */
> > + for (i = 0; i < sz_32; i++)
> > + writel(buf_32[i], hcu_dev->io_base +
> > + OCS_HCU_DMA_INBUFFER_WRITE_FIFO);
> > +
> > + /* Write final bytes into buffer. */
> > + for (i = sz_32 * sizeof(u32); i < sz; i++)
> > + writeb(buf[i], hcu_dev->io_base +
> > + OCS_HCU_DMA_INBUFFER_WRITE_FIFO);
> > +
> > + /* Wait until the writes are complete. */
> maybe you mean poll for completion? I mean the wrights are done already.
Fixed in v4.
> > + do {
> > + if ((readl(hcu_dev->io_base + OCS_HCU_DMA_STATUS) &
> > + HCU_DMA_STAT_SRC_DONE))
> > + break;
> > +
> > + if ((readl(hcu_dev->io_base + OCS_HCU_DMA_MSI_ISR) &
> > + HCU_DMA_IRQ_ERR_MASK))
> > + return 0;
> > +
> > + usleep_range(100, 200);
> > + } while (retries--);
> > +
> > + return sz;
> > +}
> > +
> > +static unsigned int ocs_hcu_hash_block_aligned_cpu(struct ocs_hcu_dev *hcu_dev,
> > + u8 *buf, u32 sz, u32 algo)
> > +{
> > + u32 blk_sz = ocs_hcu_block_size(algo);
> > + u32 *buf_32 = (u32 *)buf;
> > + u32 blk_sz_32;
> > + u32 num_blks;
> > + int i;
> > +
> > + if (blk_sz == 0)
> > + return 0;
> > +
> > + blk_sz_32 = blk_sz / sizeof(u32);
> > + num_blks = sz / blk_sz;
> > +
> > + for (i = 0; i < (blk_sz_32 * num_blks); i++)
> > + writel(buf_32[i], hcu_dev->io_base +
> > + OCS_HCU_DMA_INBUFFER_WRITE_FIFO);
> > +
> > + if (readl(hcu_dev->io_base + OCS_HCU_DMA_MSI_ISR) &
> > + HCU_DMA_IRQ_ERR_MASK)
> > + return 0;
> is this worth an error text?
> also, isn't returning zero implies success? the error mask bit is true. Is
> that not an error?
Function removed in v4 (all PIO ashing code has been removed in favor
of DMA operations).
>
> > +
> > + return i * sizeof(u32);
> > +}
> > +
> > +static void ocs_hcu_start_hash(struct ocs_hcu_dev *hcu_dev)
> > +{
> > + writel(OCS_HCU_START, hcu_dev->io_base + OCS_HCU_OPERATION);
> > +}
> > +
> > +/**
> > + * ocs_hcu_hash_cpu() - Perform OCS HCU hashing without using DMA.
> > + * @hcu_dev: The OCS HCU device to use.
> > + * @buf: The data to hash.
> > + * @sz: The size of the data to hash.
> > + * @algo: The hashing algorithm to use.
> > + * @finalize: Whether or not this is the last hashing operation and therefore
> > + * the final hash should be compute even if data is not
> > + * block-aligned.
> > + *
> > + * Return: the number of bytes hashed.
> > + */
> > +unsigned int ocs_hcu_hash_cpu(struct ocs_hcu_dev *hcu_dev,
> > + u8 *buf, u32 sz, u32 algo, bool finalize)
> > +{
> > + unsigned int written;
> > +
> > + if (!buf)
> > + return 0;
> > +
> > + written = ocs_hcu_hash_block_aligned_cpu(hcu_dev, buf, sz, algo);
> > +
> > + if (finalize) {
> > + sz -= written;
> > + if (sz) {
> > + buf += written;
> > + written += ocs_hcu_hash_final_cpu(hcu_dev, buf, sz);
> > + }
> > + ocs_hcu_tx_data_done(hcu_dev);
> > + }
> > +
> > + return written;
> > +}
> > +
> > +/**
> > + * ocs_hcu_ll_dma_start() - Start OCS HCU hashing via DMA
> > + * @hcu_dev: The OCS HCU device to use.
> > + * @head: The head of the OCS DMA list to hash.
> > + * @finalize: Whether or not this is the last hashing operation and therefore
> > + * the final hash should be compute even if data is not
> > + * block-aligned.
> > + *
> > + * Return: 0 on success, negative error code otherwise.
> > + */
> > +int ocs_hcu_ll_dma_start(struct ocs_hcu_dev *hcu_dev, dma_addr_t head,
> > + bool finalize)
> > +{
> > + u32 cfg = HCU_DMA_SNOOP_MASK | HCU_DMA_SRC_LL_EN | HCU_DMA_EN;
> > +
> > + if (head == DMA_MAPPING_ERROR || head & ~OCS_HCU_DMA_BIT_MASK)
> > + return -EINVAL;
> > +
> > + /*
> > + * For final requests we use HCU_DONE IRQ to be notified when all input
> > + * data has been processed by the HCU; however, we cannot do so for
> > + * non-final requests, because we don't get a HCU_DONE IRQ when we
> > + * don't terminate the operation.
> > + *
> > + * Therefore, for non-final requests, we use the DMA IRQ, which
> > + * triggers when DMA has finishing feeding all the input data to the
> > + * HCU, but the HCU may still be processing it. This is fine, since we
> > + * will wait for the HCU processing to be completed when we try to read
> > + * intermediate results, in ocs_hcu_get_intermediate_data().
> > + */
> > + if (finalize)
> > + ocs_hcu_done_irq_en(hcu_dev);
> > + else
> > + ocs_hcu_dma_irq_en(hcu_dev);
> > +
> > + writel(head, hcu_dev->io_base + OCS_HCU_DMA_NEXT_SRC_DESCR);
> > + writel(0, hcu_dev->io_base + OCS_HCU_DMA_SRC_SIZE);
> > + writel(0, hcu_dev->io_base + OCS_HCU_DMA_DST_SIZE);
> > + ocs_hcu_start_hash(hcu_dev);
> > + writel(cfg, hcu_dev->io_base + OCS_HCU_DMA_DMA_MODE);
> > +
> > + if (finalize)
> > + writel(OCS_HCU_TERMINATE, hcu_dev->io_base + OCS_HCU_OPERATION);
> > +
> > + return 0;
> > +}
> > +
> > +/**
> > + * ocs_hcu_write_key() - Write key to OCS HMAC KEY registers.
> > + * @hcu_dev: The OCS HCU device the key should be written to.
> > + * @key: The key to be written.
> > + * @len: The size of the key to write. It must be HCU_HW_KEY_LEN.
> > + *
> > + * Return: 0 on success, negative error code otherwise.
> > + */
> > +int ocs_hcu_write_key(struct ocs_hcu_dev *hcu_dev, u8 *key, size_t len)
> > +{
> > + u32 *key_32 = (u32 *)key;
> > + size_t len32;
> > + int i = 0;
> > +
> > + if (len != HCU_HW_KEY_LEN)
> > + return -EINVAL;
> > +
> > + /*
> > + * OCS hardware expects the MSB of the key to be written at the highest
> > + * address of the HCU Key vector; in other word, the key must be
> > + * written in reverse order.
> ick.
>
> > + *
> > + * Therefore, we first enable byte swapping for the HCU key vector;
> > + * so that bytes of 32-bit word written to OCS_HCU_KEY_[0..15] will be
> > + * swapped:
> > + * 3 <---> 0, 2 <---> 1.
> > + */
> > + writel(HCU_BYTE_ORDER_SWAP,
> > + hcu_dev->io_base + OCS_HCU_KEY_BYTE_ORDER_CFG);
> > + /*
> > + * And then we write the 32-bit words composing the key starting from
> > + * the end of the key.
> > + */
> > + len32 = HCU_HW_KEY_LEN / sizeof(u32);
> > + for (i = 0; i < len32; i++)
> > + writel(key_32[len32 - 1 - i],
> > + hcu_dev->io_base + OCS_HCU_KEY_0 + (sizeof(u32) * i));
> > +
> > + return 0;
> > +}
> > +
> > +/**
> > + * ocs_hcu_write_key() - Clear key stored in OCS HMAC KEY registers.
> > + * @hcu_dev: The OCS HCU device whose key registers should be cleared.
> > + */
> > +void ocs_hcu_clear_key(struct ocs_hcu_dev *hcu_dev)
> > +{
> > + int reg_off;
> > +
> > + /* Clear OCS_HCU_KEY_[0..15] */
> > + for (reg_off = 0; reg_off < HCU_HW_KEY_LEN; reg_off += sizeof(u32))
> > + writel(0, hcu_dev->io_base + OCS_HCU_KEY_0 + reg_off);
> > +}
> > +
> > +irqreturn_t ocs_hcu_irq_handler(int irq, void *dev_id)
> > +{
> > + struct ocs_hcu_dev *hcu_dev = dev_id;
> > + u32 hcu_irq = readl(hcu_dev->io_base + OCS_HCU_ISR);
> > + u32 dma_irq = readl(hcu_dev->io_base + OCS_HCU_DMA_MSI_ISR);
> > + irqreturn_t rc = IRQ_NONE;
> > +
> > + /* Check the HCU status. */
> > + if (hcu_irq & HCU_IRQ_HASH_ERR) {
> > + hcu_dev->flags |= HCU_FLAGS_HCU_ERR;
> > + rc = IRQ_WAKE_THREAD;
> > + } else if (hcu_irq & HCU_IRQ_HASH_DONE) {
> > + rc = IRQ_WAKE_THREAD;
> > + }
> > + /* Clear the HCU interrupt. */
> > + writel(hcu_irq, hcu_dev->io_base + OCS_HCU_ISR);
> > +
> > + /* Check the DMA status. */
> > + if (dma_irq & HCU_DMA_IRQ_ERR_MASK) {
> > + hcu_dev->flags |= HCU_FLAGS_HCU_DMA_ERR;
> > + rc = IRQ_WAKE_THREAD;
> > + goto exit;
> > + }
> > + if (dma_irq & HCU_DMA_IRQ_SRC_DONE) {
> > + /* DMA is complete, indicate that the HCU is done this
> > + * transaction.
> > + */
> > + rc = IRQ_WAKE_THREAD;
> > + }
> > +
> > +exit:
> > + /* Clear the HCU DMA interrupt. */
> > + writel(dma_irq, hcu_dev->io_base + OCS_HCU_DMA_MSI_ISR);
> > +
> > + return rc;
> > +}
> > +
> > +MODULE_LICENSE("GPL v2");
> > diff --git a/drivers/crypto/keembay/ocs-hcu.h b/drivers/crypto/keembay/ocs-hcu.h
> > new file mode 100644
> > index 000000000000..3c7e3881791c
> > --- /dev/null
> > +++ b/drivers/crypto/keembay/ocs-hcu.h
> > @@ -0,0 +1,115 @@
> > +/* SPDX-License-Identifier: GPL-2.0-only */
> > +/*
> > + * Intel Keem Bay OCS HCU Crypto Driver.
> > + *
> > + * Copyright (C) 2018-2020 Intel Corporation
> > + */
> > +
> > +#include <linux/dma-mapping.h>
> > +
> > +#ifndef _CRYPTO_OCS_HCU_H
> > +#define _CRYPTO_OCS_HCU_H
> > +
> > +#define OCS_LL_DMA_FLAG_TERMINATE BIT(31)
> > +#define OCS_LL_DMA_FLAG_FREEZE BIT(30)
> > +#define OCS_LL_DMA_FLAG_RESERVED (BIT(30) - 1)
> > +
> > +#define OCS_HCU_ALGO_SHA256 2
> > +#define OCS_HCU_ALGO_SHA224 3
> > +#define OCS_HCU_ALGO_SHA384 4
> > +#define OCS_HCU_ALGO_SHA512 5
> > +#define OCS_HCU_ALGO_SM3 6
> > +
> > +#define OCS_HCU_ALGO_MASK (BIT(3) - 1)
> > +
> > +#define OCS_HCU_ALGO_HMAC_SHIFT 4
> > +#define OCS_HCU_ALGO_HMAC_MASK BIT(OCS_HCU_ALGO_HMAC_SHIFT)
> > +
> > +#define OCS_HCU_DMA_NO_SNOOP 1
> > +#define OCS_HCU_DMA_SNOOP 0
> > +#define OCS_HCU_DMA_BTF_SWAP 1
> > +#define OCS_HCU_DMA_BTF_NO_SWAP 0
> > +#define OCS_HCU_DMA_ADDR_MODE_FIXED 1
> > +#define OCS_HCU_DMA_ADDR_MODE_LINEAR 0
> > +#define OCS_HCU_DMA_BIT_MASK DMA_BIT_MASK(32)
> > +
> > +#define OCS_HCU_MAX_CHAIN_NUM 16
> > +
> > +/* Device flags */
> > +#define HCU_FLAGS_HCU_INIT BIT(0)
> > +#define HCU_FLAGS_HCU_ERR BIT(17)
> > +#define HCU_FLAGS_HCU_DMA_ERR BIT(18)
> > +#define HCU_FLAGS_HCU_ERROR_MASK (HCU_FLAGS_HCU_DMA_ERR | \
> > + HCU_FLAGS_HCU_ERR)
> > +
> > +#define HCU_HW_KEY_LEN 64
> > +
> > +/**
> > + * HCU device context.
> > + * @list: List of device contexts.
> > + * @dev: OCS HCU device.
> > + * @irq: IRQ number.
> > + * @io_base: IO Base of HCU.
> > + * @flags: HW flags indicating state.
> > + * @req: Request being operated on.
> > + * @engine: Crypto engine for the device.
> > + */
> > +struct ocs_hcu_dev {
> > + struct list_head list;
> > + struct device *dev;
> > + int irq;
> > + /* Base address of OCS HCU */
> > + void __iomem *io_base;
> > + /* Status of the OCS HCU device */
> > + u32 flags;
> > + /* Active request. */
> > + struct ahash_request *req;
> > + struct crypto_engine *engine;
> > +};
> > +
> > +/**
> > + * struct ocs_dma_desc - OCS DMA linked list descriptor.
> > + * @src_addr: Source address of the data.
> > + * @src_len: Length of data to be fetched.
> > + * @nxt_desc: Next descriptor to fetch.
> > + * @ll_flags: Flags (Freeze @ terminate) for the DMA engine.
> > + */
> > +struct ocs_dma_desc {
> > + u32 src_adr;
> > + u32 src_len;
> > + u32 nxt_desc;
> > + u32 ll_flags;
> > +};
> > +
> > +/**
> > + * Structure to contain the intermediate data generated by the HCU.
> > + * @msg_len_lo: Length of data the HCU has operated on in bits, low 32b.
> > + * @msg_len_hi: Length of data the HCU has operated on in bits, high 32b.
> > + * @digest: The digest read from the HCU. If the HCU is terminated, it will
> > + * contain the actual hash digest. Otherwise it is the intermediate
> > + * state.
> > + */
> > +struct ocs_hcu_idata {
> > + u32 msg_len_lo;
> > + u32 msg_len_hi;
> > + u8 digest[SHA512_DIGEST_SIZE] __aligned(sizeof(u32));
> > +};
> > +
> > +irqreturn_t ocs_hcu_irq_handler(int irq, void *dev_id);
> > +int ocs_hcu_get_intermediate_data(struct ocs_hcu_dev *hcu_dev,
> > + struct ocs_hcu_idata *data, u32 algo);
> > +void ocs_hcu_set_intermediate_data(struct ocs_hcu_dev *hcu_dev,
> > + struct ocs_hcu_idata *data, u32 algo);
> > +void ocs_hcu_hw_init(struct ocs_hcu_dev *hcu_dev);
> > +void ocs_hcu_hw_disable(struct ocs_hcu_dev *hcu_dev);
> > +int ocs_hcu_hw_cfg(struct ocs_hcu_dev *hcu_dev, u32 algo);
> > +unsigned int ocs_hcu_hash_cpu(struct ocs_hcu_dev *hcu_dev,
> > + u8 *buf, u32 sz, u32 algo, bool finalize);
> > +int ocs_hcu_ll_dma_start(struct ocs_hcu_dev *hcu_dev, dma_addr_t head,
> > + bool finalize);
> > +void ocs_hcu_tx_data_done(struct ocs_hcu_dev *hcu_dev);
> > +int ocs_hcu_wait_busy(struct ocs_hcu_dev *hcu_dev);
> > +int ocs_hcu_write_key(struct ocs_hcu_dev *hcu_dev, u8 *key, size_t len);
> > +void ocs_hcu_clear_key(struct ocs_hcu_dev *hcu_dev);
> > +
> > +#endif /* _CRYPTO_OCS_HCU_H */
> > --
> > 2.26.2
> >
