From: Andrew Bresticker Subject: Re: [PATCH V3 1/2] crypto: Add Imagination Technologies hw hash accelerator Date: Sun, 8 Mar 2015 22:42:02 -0700 Message-ID: References: <1425610912-9930-1-git-send-email-james.hartley@imgtec.com> <1425610912-9930-2-git-send-email-james.hartley@imgtec.com> Mime-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Cc: linux-crypto@vger.kernel.org To: James Hartley Return-path: Received: from mail-qg0-f47.google.com ([209.85.192.47]:39867 "EHLO mail-qg0-f47.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1750967AbbCIFmD (ORCPT ); Mon, 9 Mar 2015 01:42:03 -0400 Received: by qgdq107 with SMTP id q107so25993378qgd.6 for ; Sun, 08 Mar 2015 22:42:02 -0700 (PDT) In-Reply-To: <1425610912-9930-2-git-send-email-james.hartley@imgtec.com> Sender: linux-crypto-owner@vger.kernel.org List-ID: Hi James, On Thu, Mar 5, 2015 at 7:01 PM, James Hartley wrote: > This adds support for the Imagination Technologies hash > accelerator which provides hardware acceleration for > SHA1 SHA224 SHA256 and MD5 hashes. > > Signed-off-by: James Hartley Some general comments below, I'll leave review of the crypto-specific stuff to Herbert. > diff --git a/drivers/crypto/Makefile b/drivers/crypto/Makefile > index 3924f93..fb84be7 100644 > --- a/drivers/crypto/Makefile > +++ b/drivers/crypto/Makefile > @@ -6,6 +6,7 @@ obj-$(CONFIG_CRYPTO_DEV_CCP) += ccp/ > obj-$(CONFIG_CRYPTO_DEV_FSL_CAAM) += caam/ > obj-$(CONFIG_CRYPTO_DEV_GEODE) += geode-aes.o > obj-$(CONFIG_CRYPTO_DEV_HIFN_795X) += hifn_795x.o > +obj-$(CONFIG_CRYPTO_DEV_IMGTEC_HASH) += img-hash.o > obj-$(CONFIG_CRYPTO_DEV_IXP4XX) += ixp4xx_crypto.o > obj-$(CONFIG_CRYPTO_DEV_MV_CESA) += mv_cesa.o > obj-$(CONFIG_CRYPTO_DEV_MXS_DCP) += mxs-dcp.o > @@ -25,3 +26,4 @@ obj-$(CONFIG_CRYPTO_DEV_TALITOS) += talitos.o > obj-$(CONFIG_CRYPTO_DEV_UX500) += ux500/ > obj-$(CONFIG_CRYPTO_DEV_QAT) += qat/ > obj-$(CONFIG_CRYPTO_DEV_QCE) += qce/ > +obj-$(CONFIG_CRYPTO_DEV_VMX) += vmx/ Unrelated change - perhaps a bad merge conflict resolution? > diff --git a/drivers/crypto/img-hash.c b/drivers/crypto/img-hash.c > new file mode 100644 > index 0000000..94a3a6f > --- /dev/null > +++ b/drivers/crypto/img-hash.c > @@ -0,0 +1,1037 @@ > +/* > + * Copyright (c) 2014 Imagination Technologies > + * Authors: Will Thomas, James Hartley > + * > + * This program is free software; you can redistribute it and/or modify > + * it under the terms of the GNU General Public License version 2 as published > + * by the Free Software Foundation. > + * > + * Interface structure taken from omap-sham driver > + */ > + > +#include > +#include > +#include > +#include > +#include > +#include > +#include > +#include > +#include > + > +#include > +#include > +#include > + > +#define CR_RESET 0 > +#define CR_RESET_SET 1 > +#define CR_RESET_UNSET 0 > + > +#define CR_MESSAGE_LENGTH_H 0x4 > +#define CR_MESSAGE_LENGTH_L 0x8 > + > +#define CR_CONTROL 0xc > +#define CR_CONTROL_BYTE_ORDER_3210 0 > +#define CR_CONTROL_BYTE_ORDER_0123 1 > +#define CR_CONTROL_BYTE_ORDER_2310 2 > +#define CR_CONTROL_BYTE_ORDER_1032 3 > +#define CR_CONTROL_BYTE_ORDER_SHIFT 8 > +#define CR_CONTROL_ALGO_MD5 0 > +#define CR_CONTROL_ALGO_SHA1 1 > +#define CR_CONTROL_ALGO_SHA224 2 > +#define CR_CONTROL_ALGO_SHA256 3 > + > +#define CR_INTSTAT 0x10 > +#define CR_INTENAB 0x14 > +#define CR_INTCLEAR 0x18 > +#define CR_INT_RESULTS_AVAILABLE BIT(0) > +#define CR_INT_NEW_RESULTS_SET BIT(1) > +#define CR_INT_RESULT_READ_ERR BIT(2) > +#define CR_INT_MESSAGE_WRITE_ERROR BIT(3) > +#define CR_INT_STATUS BIT(8) > + > +#define CR_RESULT_QUEUE 0x1c > +#define CR_RSD0 0x40 > +#define CR_CORE_REV 0x50 > +#define CR_CORE_DES1 0x60 > +#define CR_CORE_DES2 0x70 > + > +#define DRIVER_FLAGS_BUSY BIT(0) > +#define DRIVER_FLAGS_FINAL BIT(1) > +#define DRIVER_FLAGS_DMA_ACTIVE BIT(2) > +#define DRIVER_FLAGS_OUTPUT_READY BIT(3) > +#define DRIVER_FLAGS_INIT BIT(4) > +#define DRIVER_FLAGS_CPU BIT(5) > +#define DRIVER_FLAGS_DMA_READY BIT(6) > +#define DRIVER_FLAGS_ERROR BIT(7) > +#define DRIVER_FLAGS_SG BIT(8) > +#define DRIVER_FLAGS_SHA1 BIT(18) > +#define DRIVER_FLAGS_SHA224 BIT(19) > +#define DRIVER_FLAGS_SHA256 BIT(20) > +#define DRIVER_FLAGS_MD5 BIT(21) > + > +#define IMG_HASH_QUEUE_LENGTH 20 > +#define IMG_HASH_DMA_THRESHOLD 64 > + > +#ifdef __LITTLE_ENDIAN > +#define IMG_HASH_BYTE_ORDER (CR_CONTROL_BYTE_ORDER_3210) > +#else > +#define IMG_HASH_BYTE_ORDER (CR_CONTROL_BYTE_ORDER_0123) > +#endif Unnecessary parenthesis. > +struct img_hash_dev; > + > +struct img_hash_request_ctx { > + struct img_hash_dev *hdev; > + u8 digest[SHA256_DIGEST_SIZE] __aligned(sizeof(u32)); > + unsigned long flags; > + size_t digsize; > + > + dma_addr_t dma_addr; > + size_t dma_ct; > + > + /* sg root */ > + struct scatterlist *sgfirst; > + /* walk state */ > + struct scatterlist *sg; > + size_t nents; > + size_t offset; > + unsigned int total; > + size_t sent; > + > + unsigned long op; > + > + size_t bufcnt; > + u8 buffer[0] __aligned(sizeof(u32)); > + struct ahash_request fallback_req; > +}; > + > +struct img_hash_ctx { > + struct img_hash_dev *hdev; > + unsigned long flags; > + struct crypto_ahash *fallback; > +}; > + > +struct img_hash_dev { > + struct list_head list; > + struct device *dev; > + struct clk *iclk; > + struct clk *fclk; Maybe make these names a little more obvious so it's clear which clock is which. > + int irq; This isn't used outside of probe(), so it need not be carried around in this struct. > + void __iomem *io_base; > + > + phys_addr_t bus_addr; > + void __iomem *cpu_addr; > + > + spinlock_t lock; > + int err; > + struct tasklet_struct done_task; > + struct tasklet_struct dma_task; > + > + unsigned long flags; > + struct crypto_queue queue; > + struct ahash_request *req; > + > + struct dma_chan *dma_lch; > +}; > + > +struct img_hash_drv { > + struct list_head dev_list; > + spinlock_t lock; > +}; > + > +static struct img_hash_drv img_hash = { > + .dev_list = LIST_HEAD_INIT(img_hash.dev_list), > + .lock = __SPIN_LOCK_UNLOCKED(img_hash.lock), > +}; > + > +static inline u32 img_hash_read(struct img_hash_dev *hdev, u32 offset) > +{ > + return readl_relaxed(hdev->io_base + offset); > +} > + > +static inline void img_hash_write(struct img_hash_dev *hdev, > + u32 offset, u32 value) > +{ > + writel_relaxed(value, hdev->io_base + offset); > +} > + > +static inline u32 img_hash_read_result_queue(struct img_hash_dev *hdev) > +{ > + return be32_to_cpu(img_hash_read(hdev, CR_RESULT_QUEUE)); > +} > + > +static void img_hash_start(struct img_hash_dev *hdev, bool dma) > +{ > + struct img_hash_request_ctx *ctx = ahash_request_ctx(hdev->req); > + u32 cr = IMG_HASH_BYTE_ORDER << CR_CONTROL_BYTE_ORDER_SHIFT; > + > + if (ctx->flags & DRIVER_FLAGS_MD5) > + cr |= CR_CONTROL_ALGO_MD5; > + else if (ctx->flags & DRIVER_FLAGS_SHA1) > + cr |= CR_CONTROL_ALGO_SHA1; > + else if (ctx->flags & DRIVER_FLAGS_SHA224) > + cr |= CR_CONTROL_ALGO_SHA224; > + else if (ctx->flags & DRIVER_FLAGS_SHA256) > + cr |= CR_CONTROL_ALGO_SHA256; > + dev_dbg(hdev->dev, "Starting hash process\n"); > + img_hash_write(hdev, CR_CONTROL, cr); > + > + /* > + * The hardware block requires two cycles between writing the control > + * register and writing the first word of data in non DMA mode, to > + * ensure the first data write is not grouped in burst with the control > + * register write a read is issued to 'flush' the bus. > + */ > + if (!dma) > + img_hash_read(hdev, CR_CONTROL); > +} > + > +static int img_hash_xmit_cpu(struct img_hash_dev *hdev, const u8 *buf, > + size_t length, int final) > +{ > + u32 count, len32; > + const u32 *buffer = (const u32 *)buf; > + > + dev_dbg(hdev->dev, "xmit_cpu: length: %u bytes\n", length); > + > + if (final) > + hdev->flags |= DRIVER_FLAGS_FINAL; > + > + len32 = DIV_ROUND_UP(length, sizeof(u32)); > + > + for (count = 0; count < len32; count++) > + writel_relaxed(buffer[count], hdev->cpu_addr); > + > + return -EINPROGRESS; > +} > + > +static void img_hash_dma_callback(void *data) > +{ > + struct img_hash_dev *hdev = (struct img_hash_dev *)data; > + struct img_hash_request_ctx *ctx = ahash_request_ctx(hdev->req); > + > + if (ctx->bufcnt) { > + img_hash_xmit_cpu(hdev, ctx->buffer, ctx->bufcnt, 0); > + ctx->bufcnt = 0; > + } > + if (ctx->sg) > + tasklet_schedule(&hdev->dma_task); > +} > + > +static int img_hash_xmit_dma(struct img_hash_dev *hdev, struct scatterlist *sg) > +{ > + struct dma_async_tx_descriptor *desc; > + struct img_hash_request_ctx *ctx = ahash_request_ctx(hdev->req); > + > + ctx->dma_ct = dma_map_sg(hdev->dev, sg, 1, DMA_MEM_TO_DEV); > + if (ctx->dma_ct == 0) { > + dev_err(hdev->dev, "Invalid DMA sg\n"); > + hdev->err = -EINVAL; > + return -EINVAL; > + } > + > + desc = dmaengine_prep_slave_sg(hdev->dma_lch, > + sg, Line this up (and the rest of this statement) with the open paren of dmaengine_prep_slave_sg(). > + ctx->dma_ct, > + DMA_MEM_TO_DEV, > + DMA_PREP_INTERRUPT | DMA_CTRL_ACK); > + if (!desc) { > + dev_err(hdev->dev, "Null DMA descriptor\n"); > + hdev->err = -EINVAL; > + dma_unmap_sg(hdev->dev, sg, 1, DMA_MEM_TO_DEV); > + return -EINVAL; > + } > + desc->callback = img_hash_dma_callback; > + desc->callback_param = hdev; > + dmaengine_submit(desc); > + dma_async_issue_pending(hdev->dma_lch); > + > + return 0; > +} > + > +static int img_hash_write_via_cpu(struct img_hash_dev *hdev) > +{ > + struct img_hash_request_ctx *ctx = ahash_request_ctx(hdev->req); > + int ret = 0; > + > + ctx->bufcnt = sg_copy_to_buffer(hdev->req->src, sg_nents(ctx->sg), > + ctx->buffer, hdev->req->nbytes); > + > + ctx->total = hdev->req->nbytes; > + ctx->bufcnt = 0; > + > + hdev->flags |= (DRIVER_FLAGS_CPU | DRIVER_FLAGS_FINAL); > + > + img_hash_start(hdev, false); > + > + if (ctx->total) > + ret = img_hash_xmit_cpu(hdev, ctx->buffer, ctx->total, 1); > + else > + ret = 0; The else block here (and possibly the entire check) seems unnecessary. img_hash_xmit_cpu() won't do anything if the length is 0. > + > + return ret; > +} > + > +static int img_hash_finish(struct ahash_request *req) > +{ > + struct img_hash_request_ctx *ctx = ahash_request_ctx(req); > + > + if (!req->result) > + return -EINVAL; > + > + memcpy(req->result, ctx->digest, ctx->digsize); > + > + return 0; > +} > + > +static void img_hash_copy_hash(struct ahash_request *req) > +{ > + struct img_hash_request_ctx *ctx = ahash_request_ctx(req); > + u32 *hash = (u32 *)ctx->digest; > + int i; > + > + for (i = (ctx->digsize / sizeof(u32)) - 1; i >= 0; i--) > + hash[i] = img_hash_read_result_queue(ctx->hdev); > +} > + > +static void img_hash_finish_req(struct ahash_request *req, int err) > +{ > + struct img_hash_request_ctx *ctx = ahash_request_ctx(req); > + struct img_hash_dev *hdev = ctx->hdev; > + > + if (!err) { > + img_hash_copy_hash(req); > + if (DRIVER_FLAGS_FINAL & hdev->flags) > + err = img_hash_finish(req); > + } else { > + dev_warn(hdev->dev, "Hash failed with error %d\n", err); > + ctx->flags |= DRIVER_FLAGS_ERROR; > + } > + > + hdev->flags &= ~(DRIVER_FLAGS_DMA_READY | DRIVER_FLAGS_OUTPUT_READY | > + DRIVER_FLAGS_CPU | DRIVER_FLAGS_BUSY | DRIVER_FLAGS_FINAL); > + > + if (req->base.complete) > + req->base.complete(&req->base, err); > +} > + > +static int img_hash_write_via_dma(struct img_hash_dev *hdev) > +{ > + struct img_hash_request_ctx *ctx = ahash_request_ctx(hdev->req); > + > + img_hash_start(hdev, true); > + > + dev_dbg(hdev->dev, "xmit dma size: %d\n", ctx->total); > + > + if (!ctx->total) > + hdev->flags |= DRIVER_FLAGS_FINAL; > + > + hdev->flags |= DRIVER_FLAGS_DMA_ACTIVE | DRIVER_FLAGS_FINAL; > + > + tasklet_schedule(&hdev->dma_task); > + > + return -EINPROGRESS; > +} > + > +static int img_hash_dma_init(struct img_hash_dev *hdev) > +{ > + struct dma_slave_config dma_conf; > + int err = -EINVAL; > + > + hdev->dma_lch = dma_request_slave_channel(hdev->dev, "tx"); > + if (!hdev->dma_lch) { > + dev_err(hdev->dev, "Couldn't aquire a slave DMA channel.\n"); > + return -EBUSY; > + } > + dma_conf.direction = DMA_MEM_TO_DEV; > + dma_conf.dst_addr = hdev->bus_addr; > + dma_conf.dst_addr_width = DMA_SLAVE_BUSWIDTH_4_BYTES; > + dma_conf.dst_maxburst = 16; > + dma_conf.device_fc = false; > + > + err = dmaengine_slave_config(hdev->dma_lch, &dma_conf); > + > + if (err) { > + dev_err(hdev->dev, "Couldn't configure DMA slave.\n"); > + dma_release_channel(hdev->dma_lch); > + return err; > + } > + return 0; nit: I would expect there to be a newline before the return rather than before the if(). > +} > + > +static void img_hash_dma_task(unsigned long d) > +{ > + struct img_hash_dev *hdev = (struct img_hash_dev *)d; > + struct img_hash_request_ctx *ctx = ahash_request_ctx(hdev->req); > + u8 *addr; > + size_t nbytes, bleft, wsend, len, tbc; > + struct scatterlist tsg; > + > + if (!ctx->sg) > + return; > + > + addr = sg_virt(ctx->sg); > + nbytes = ctx->sg->length - ctx->offset; > + > + /* The hash accelerator does not support a data valid mask. This means > + * that if each dma (i.e. per page) is not a multiple of 4 bytes, the > + * padding bytes in the last word written by that dma would erroneously > + * be included in the hash. To avoid this we round down the transfer, > + * and add the excess to the start of the next dma. It does not matter > + * that the final dma may not be a multiple of 4 bytes as the hashing > + * block is programmed to accept the correct number of bytes. > + */ nit: Multi-line comment style is: /* * blah blah blah */ > + > + bleft = nbytes % 4; > + wsend = (nbytes / 4); > + > + if (wsend) { > + sg_init_one(&tsg, addr + ctx->offset, wsend * 4); > + if (img_hash_xmit_dma(hdev, &tsg)) { > + dev_err(hdev->dev, "DMA failed, falling back to CPU"); > + ctx->flags |= DRIVER_FLAGS_CPU; > + hdev->err = 0; > + img_hash_xmit_cpu(hdev, addr + ctx->offset, > + wsend * 4, 0); > + ctx->sent += wsend * 4; > + wsend = 0; > + } else { > + ctx->sent += wsend * 4; > + } > + } > + > + if (bleft) { > + ctx->bufcnt = sg_pcopy_to_buffer(ctx->sgfirst, ctx->nents, > + ctx->buffer, bleft, ctx->sent); > + tbc = 0; > + ctx->sg = sg_next(ctx->sg); > + while (ctx->sg && (ctx->bufcnt < 4)) { > + len = ctx->sg->length; > + if (likely(len > (4 - ctx->bufcnt))) > + len = 4 - ctx->bufcnt; > + tbc = sg_pcopy_to_buffer(ctx->sgfirst, ctx->nents, > + ctx->buffer + ctx->bufcnt, len, > + ctx->sent + ctx->bufcnt); > + ctx->bufcnt += tbc; > + if (tbc >= ctx->sg->length) { > + ctx->sg = sg_next(ctx->sg); > + tbc = 0; > + } > + } > + > + ctx->sent += ctx->bufcnt; > + ctx->offset = tbc; > + > + if (!wsend) > + img_hash_dma_callback(hdev); > + } else { > + ctx->offset = 0; > + ctx->sg = sg_next(ctx->sg); > + } > +} > + > +static int img_hash_write_via_dma_stop(struct img_hash_dev *hdev) > +{ > + struct img_hash_request_ctx *ctx = ahash_request_ctx(hdev->req); > + > + if (ctx->flags & DRIVER_FLAGS_SG) > + dma_unmap_sg(hdev->dev, ctx->sg, ctx->dma_ct, DMA_TO_DEVICE); > + > + return 0; > +} > + > +static int img_hash_process_data(struct img_hash_dev *hdev) > +{ > + struct ahash_request *req = hdev->req; > + struct img_hash_request_ctx *ctx = ahash_request_ctx(req); > + int err = 0; > + > + ctx->bufcnt = 0; > + > + if (req->nbytes >= IMG_HASH_DMA_THRESHOLD) { > + dev_dbg(hdev->dev, "process data request(%d bytes) using DMA\n", > + req->nbytes); > + err = img_hash_write_via_dma(hdev); > + } else { > + dev_dbg(hdev->dev, "process data request(%d bytes) using CPU\n", > + req->nbytes); > + err = img_hash_write_via_cpu(hdev); > + } > + return err; > +} > + > +static int img_hash_hw_init(struct img_hash_dev *hdev) > +{ > + unsigned long long nbits; > + u32 u, l; > + int ret; > + > + ret = clk_prepare_enable(hdev->iclk); It looks like there's no corresponding clk_disable_unprepare() to this, so the prepare/enable counts will just keep increasing. Perhaps it would be better to manage enabling/disabling the clocks with runtime PM? > + if (ret) > + return ret; > + > + img_hash_write(hdev, CR_RESET, CR_RESET_SET); > + img_hash_write(hdev, CR_RESET, CR_RESET_UNSET); > + img_hash_write(hdev, CR_INTENAB, CR_INT_NEW_RESULTS_SET); > + > + nbits = (hdev->req->nbytes << 3); > + u = nbits >> 32; > + l = nbits; > + img_hash_write(hdev, CR_MESSAGE_LENGTH_H, u); > + img_hash_write(hdev, CR_MESSAGE_LENGTH_L, l); > + > + if (!(DRIVER_FLAGS_INIT & hdev->flags)) { > + hdev->flags |= DRIVER_FLAGS_INIT; > + hdev->err = 0; > + } > + dev_dbg(hdev->dev, "hw initialized, nbits: %llx\n", nbits); > + return 0; > +} > + > +static int img_hash_init(struct ahash_request *req) > +{ > + struct crypto_ahash *tfm = crypto_ahash_reqtfm(req); > + struct img_hash_request_ctx *rctx = ahash_request_ctx(req); > + struct img_hash_ctx *ctx = crypto_ahash_ctx(tfm); > + > + ahash_request_set_tfm(&rctx->fallback_req, ctx->fallback); > + rctx->fallback_req.base.flags = req->base.flags > + & CRYPTO_TFM_REQ_MAY_SLEEP; > + > + return crypto_ahash_init(&rctx->fallback_req); > +} > + > +static int img_hash_handle_queue(struct img_hash_dev *hdev, > + struct ahash_request *req) > +{ > + struct crypto_async_request *async_req, *backlog; > + struct img_hash_request_ctx *ctx; > + unsigned long flags; > + int err = 0, res = 0; > + > + spin_lock_irqsave(&hdev->lock, flags); > + > + if (req) > + res = ahash_enqueue_request(&hdev->queue, req); > + > + if (DRIVER_FLAGS_BUSY & hdev->flags) { > + spin_unlock_irqrestore(&hdev->lock, flags); > + return res; > + } > + > + backlog = crypto_get_backlog(&hdev->queue); > + async_req = crypto_dequeue_request(&hdev->queue); > + if (async_req) > + hdev->flags |= DRIVER_FLAGS_BUSY; > + > + spin_unlock_irqrestore(&hdev->lock, flags); > + > + if (!async_req) > + return res; > + > + if (backlog) > + backlog->complete(backlog, -EINPROGRESS); > + > + req = ahash_request_cast(async_req); > + hdev->req = req; > + > + ctx = ahash_request_ctx(req); > + > + dev_info(hdev->dev, "processing req, op: %lu, bytes: %d\n", > + ctx->op, req->nbytes); > + > + err = img_hash_hw_init(hdev); > + > + if (!err) > + err = img_hash_process_data(hdev); > + > + if (err != -EINPROGRESS) { > + /* done_task will not finish so do it here */ > + img_hash_finish_req(req, err); > + } > + return res; > +} > + > +static int img_hash_update(struct ahash_request *req) > +{ > + struct img_hash_request_ctx *rctx = ahash_request_ctx(req); > + struct crypto_ahash *tfm = crypto_ahash_reqtfm(req); > + struct img_hash_ctx *ctx = crypto_ahash_ctx(tfm); > + > + ahash_request_set_tfm(&rctx->fallback_req, ctx->fallback); > + rctx->fallback_req.base.flags = req->base.flags > + & CRYPTO_TFM_REQ_MAY_SLEEP; > + rctx->fallback_req.nbytes = req->nbytes; > + rctx->fallback_req.src = req->src; > + > + return crypto_ahash_update(&rctx->fallback_req); > +} > + > +static int img_hash_final(struct ahash_request *req) > +{ > + struct img_hash_request_ctx *rctx = ahash_request_ctx(req); > + struct crypto_ahash *tfm = crypto_ahash_reqtfm(req); > + struct img_hash_ctx *ctx = crypto_ahash_ctx(tfm); > + > + ahash_request_set_tfm(&rctx->fallback_req, ctx->fallback); > + rctx->fallback_req.base.flags = req->base.flags > + & CRYPTO_TFM_REQ_MAY_SLEEP; > + rctx->fallback_req.result = req->result; > + > + return crypto_ahash_final(&rctx->fallback_req); > +} > + > +static int img_hash_finup(struct ahash_request *req) > +{ > + struct img_hash_request_ctx *rctx = ahash_request_ctx(req); > + struct crypto_ahash *tfm = crypto_ahash_reqtfm(req); > + struct img_hash_ctx *ctx = crypto_ahash_ctx(tfm); > + > + ahash_request_set_tfm(&rctx->fallback_req, ctx->fallback); > + rctx->fallback_req.base.flags = req->base.flags > + & CRYPTO_TFM_REQ_MAY_SLEEP; > + rctx->fallback_req.nbytes = req->nbytes; > + rctx->fallback_req.src = req->src; > + rctx->fallback_req.result = req->result; > + > + return crypto_ahash_finup(&rctx->fallback_req); > +} > + > +static int img_hash_digest(struct ahash_request *req) > +{ > + struct crypto_ahash *tfm = crypto_ahash_reqtfm(req); > + struct img_hash_ctx *tctx = crypto_ahash_ctx(tfm); > + struct img_hash_request_ctx *ctx = ahash_request_ctx(req); > + struct img_hash_dev *hdev = NULL; > + struct img_hash_dev *tmp; > + int err; > + > + spin_lock_bh(&img_hash.lock); Why spin_{lock,unlock}_bh() here? > + if (!tctx->hdev) { > + list_for_each_entry(tmp, &img_hash.dev_list, list) { > + hdev = tmp; > + break; > + } > + tctx->hdev = hdev; > + > + } else { > + hdev = tctx->hdev; > + } > + > + spin_unlock_bh(&img_hash.lock); > + ctx->hdev = hdev; > + ctx->flags = 0; > + ctx->digsize = crypto_ahash_digestsize(tfm); > + > + switch (ctx->digsize) { > + case SHA1_DIGEST_SIZE: > + ctx->flags |= DRIVER_FLAGS_SHA1; > + break; > + case SHA256_DIGEST_SIZE: > + ctx->flags |= DRIVER_FLAGS_SHA256; > + break; > + case SHA224_DIGEST_SIZE: > + ctx->flags |= DRIVER_FLAGS_SHA224; > + break; > + case MD5_DIGEST_SIZE: > + ctx->flags |= DRIVER_FLAGS_MD5; > + break; > + default: > + return -EINVAL; > + } > + > + ctx->bufcnt = 0; > + ctx->offset = 0; > + ctx->sent = 0; > + ctx->total = req->nbytes; > + ctx->sg = req->src; > + ctx->sgfirst = req->src; > + ctx->nents = sg_nents(ctx->sg); > + > + err = img_hash_handle_queue(tctx->hdev, req); > + > + return err; > +} > + > +static int img_hash_cra_init(struct crypto_tfm *tfm) > +{ > + struct img_hash_ctx *ctx = crypto_tfm_ctx(tfm); > + const char *alg_name = crypto_tfm_alg_name(tfm); > + int err = -ENOMEM; > + > + ctx->fallback = crypto_alloc_ahash(alg_name, 0, > + CRYPTO_ALG_NEED_FALLBACK); > + if (IS_ERR(ctx->fallback)) { > + pr_err("img_hash: Could not load fallback driver.\n"); > + err = PTR_ERR(ctx->fallback); > + goto err; > + } > + crypto_ahash_set_reqsize(__crypto_ahash_cast(tfm), > + sizeof(struct img_hash_request_ctx) + > + IMG_HASH_DMA_THRESHOLD); > + > + return 0; > + > +err: > + return err; > +} > + > +static void img_hash_cra_exit(struct crypto_tfm *tfm) > +{ > + struct img_hash_ctx *tctx = crypto_tfm_ctx(tfm); > + > + crypto_free_ahash(tctx->fallback); > +} > + > +static irqreturn_t img_irq_handler(int irq, void *dev_id) > +{ > + struct img_hash_dev *hdev = dev_id; > + u32 reg; > + > + reg = img_hash_read(hdev, CR_INTSTAT); > + img_hash_write(hdev, CR_INTCLEAR, reg); > + > + if (reg & CR_INT_NEW_RESULTS_SET) { > + dev_dbg(hdev->dev, "IRQ CR_INT_NEW_RESULTS_SET\n"); > + if (DRIVER_FLAGS_BUSY & hdev->flags) { > + hdev->flags |= DRIVER_FLAGS_OUTPUT_READY; > + if (!(DRIVER_FLAGS_CPU & hdev->flags)) > + hdev->flags |= DRIVER_FLAGS_DMA_READY; > + tasklet_schedule(&hdev->done_task); Since this done_task only gets scheduled from here, why not make this a threaded IRQ handler and just do the work here instead of separating it between a hard IRQ handler and a tasklet? > + } else { > + dev_warn(hdev->dev, > + "HASH interrupt when no active requests.\n"); > + } > + } else if (reg & CR_INT_RESULTS_AVAILABLE) { > + dev_warn(hdev->dev, > + "IRQ triggered before the hash had completed\n"); > + } else if (reg & CR_INT_RESULT_READ_ERR) { > + dev_warn(hdev->dev, > + "Attempt to read from an empty result queue\n"); > + } else if (reg & CR_INT_MESSAGE_WRITE_ERROR) { > + dev_warn(hdev->dev, > + "Data written before the hardware was configured\n"); > + } > + return IRQ_HANDLED; > +} > + > +static struct ahash_alg img_algs[] = { > + { > + .init = img_hash_init, > + .update = img_hash_update, > + .final = img_hash_final, > + .finup = img_hash_finup, > + .digest = img_hash_digest, > + .halg = { > + .digestsize = MD5_DIGEST_SIZE, > + .base = { > + .cra_name = "md5", > + .cra_driver_name = "img-md5", > + .cra_priority = 301, > + .cra_flags = > + CRYPTO_ALG_ASYNC | > + CRYPTO_ALG_NEED_FALLBACK, > + .cra_blocksize = MD5_HMAC_BLOCK_SIZE, > + .cra_ctxsize = sizeof(struct img_hash_ctx), > + .cra_init = img_hash_cra_init, > + .cra_exit = img_hash_cra_exit, > + .cra_module = THIS_MODULE, > + } > + } > + }, > + { > + .init = img_hash_init, > + .update = img_hash_update, > + .final = img_hash_final, > + .finup = img_hash_finup, > + .digest = img_hash_digest, > + .halg = { > + .digestsize = SHA1_DIGEST_SIZE, > + .base = { > + .cra_name = "sha1", > + .cra_driver_name = "img-sha1", > + .cra_priority = 3000, > + .cra_flags = > + CRYPTO_ALG_ASYNC | > + CRYPTO_ALG_NEED_FALLBACK, > + .cra_blocksize = SHA1_BLOCK_SIZE, > + .cra_ctxsize = sizeof(struct img_hash_ctx), > + .cra_init = img_hash_cra_init, > + .cra_exit = img_hash_cra_exit, > + .cra_module = THIS_MODULE, > + } > + } > + }, > + { > + .init = img_hash_init, > + .update = img_hash_update, > + .final = img_hash_final, > + .finup = img_hash_finup, > + .digest = img_hash_digest, > + .halg = { > + .digestsize = SHA224_DIGEST_SIZE, > + .base = { > + .cra_name = "sha224", > + .cra_driver_name = "img-sha224", > + .cra_priority = 3000, > + .cra_flags = > + CRYPTO_ALG_ASYNC | > + CRYPTO_ALG_NEED_FALLBACK, > + .cra_blocksize = SHA224_BLOCK_SIZE, > + .cra_ctxsize = sizeof(struct img_hash_ctx), > + .cra_init = img_hash_cra_init, > + .cra_exit = img_hash_cra_exit, > + .cra_module = THIS_MODULE, > + } > + } > + }, > + { > + .init = img_hash_init, > + .update = img_hash_update, > + .final = img_hash_final, > + .finup = img_hash_finup, > + .digest = img_hash_digest, > + .halg = { > + .digestsize = SHA256_DIGEST_SIZE, > + .base = { > + .cra_name = "sha256", > + .cra_driver_name = "img-sha256", > + .cra_priority = 301, > + .cra_flags = > + CRYPTO_ALG_ASYNC | > + CRYPTO_ALG_NEED_FALLBACK, > + .cra_blocksize = SHA256_BLOCK_SIZE, > + .cra_ctxsize = sizeof(struct img_hash_ctx), > + .cra_init = img_hash_cra_init, > + .cra_exit = img_hash_cra_exit, > + .cra_module = THIS_MODULE, > + } > + } > + } > +}; > + > +static int img_register_algs(struct img_hash_dev *hdev) > +{ > + int i, err; > + > + for (i = 0; i < ARRAY_SIZE(img_algs); i++) { > + err = crypto_register_ahash(&img_algs[i]); > + if (err) > + goto err_reg; > + } > + return 0; > + > +err_reg: > + for (; i--; ) > + crypto_unregister_ahash(&img_algs[i]); > + > + return err; > +} > + > +static int img_unregister_algs(struct img_hash_dev *hdev) > +{ > + int i; > + > + for (i = 0; i < ARRAY_SIZE(img_algs); i++) > + crypto_unregister_ahash(&img_algs[i]); > + return 0; > +} > + > +static void img_hash_done_task(unsigned long data) > +{ > + struct img_hash_dev *hdev = (struct img_hash_dev *)data; > + int err = 0; > + > + if (hdev->err == -EINVAL) { > + err = hdev->err; > + goto finish; > + } > + > + if (!(DRIVER_FLAGS_BUSY & hdev->flags)) { > + img_hash_handle_queue(hdev, NULL); > + return; > + } > + > + if (DRIVER_FLAGS_CPU & hdev->flags) { > + if (DRIVER_FLAGS_OUTPUT_READY & hdev->flags) { > + hdev->flags &= ~DRIVER_FLAGS_OUTPUT_READY; > + goto finish; > + } > + } else if (DRIVER_FLAGS_DMA_READY & hdev->flags) { > + if (DRIVER_FLAGS_DMA_ACTIVE & hdev->flags) { > + hdev->flags &= ~DRIVER_FLAGS_DMA_ACTIVE; > + img_hash_write_via_dma_stop(hdev); > + if (hdev->err) { > + err = hdev->err; > + goto finish; > + } > + } > + if (DRIVER_FLAGS_OUTPUT_READY & hdev->flags) { > + hdev->flags &= ~(DRIVER_FLAGS_DMA_READY | > + DRIVER_FLAGS_OUTPUT_READY); > + goto finish; > + } > + } > + return; > + > +finish: > + img_hash_finish_req(hdev->req, err); > +} > + > +static const struct of_device_id img_hash_match[] = { > + { .compatible = "img,hash-accelerator" }, > + {} > +}; > +MODULE_DEVICE_TABLE(of, img_hash_match) > + > +static int img_hash_probe(struct platform_device *pdev) > +{ > + struct img_hash_dev *hdev; > + struct device *dev = &pdev->dev; > + struct resource *hash_res; > + int err; > + > + hdev = devm_kzalloc(dev, sizeof(*hdev), GFP_KERNEL); > + if (hdev == NULL) { > + err = -ENOMEM; > + goto sha_dev_err; Why not just "return -ENOMEM" here? There should be no cleanup necessary at this point? > + } > + spin_lock_init(&hdev->lock); > + > + hdev->dev = dev; > + > + platform_set_drvdata(pdev, hdev); > + > + INIT_LIST_HEAD(&hdev->list); > + > + tasklet_init(&hdev->done_task, img_hash_done_task, (unsigned long)hdev); > + tasklet_init(&hdev->dma_task, img_hash_dma_task, (unsigned long)hdev); > + > + crypto_init_queue(&hdev->queue, IMG_HASH_QUEUE_LENGTH); > + > + /* Register bank */ > + hash_res = platform_get_resource(pdev, IORESOURCE_MEM, 0); > + > + hdev->io_base = devm_ioremap_resource(dev, hash_res); > + if (IS_ERR(hdev->io_base)) { > + dev_err(dev, "can't ioremap\n"); nit: When printing error messages, it's helpful to print the error code as well. > + err = PTR_ERR(hdev->io_base); > + goto hash_io_err; > + } > + > + /* Write port (DMA or CPU) */ > + hash_res = platform_get_resource(pdev, IORESOURCE_MEM, 1); > + if (!hash_res) { > + dev_err(dev, "no write port resource info\n"); > + err = -ENODEV; > + goto res_err; > + } > + hdev->bus_addr = hash_res->start; Maybe set this after devm_ioremap_resource() succeeds and avoid the extra error check? > + hdev->cpu_addr = devm_ioremap_resource(dev, hash_res); > + if (IS_ERR(hdev->cpu_addr)) { > + dev_err(dev, "can't ioremap write port\n"); > + err = PTR_ERR(hdev->cpu_addr); > + goto res_err; > + } > + > + hdev->irq = platform_get_irq(pdev, 0); > + if (hdev->irq < 0) { > + dev_err(dev, "no IRQ resource info\n"); > + err = hdev->irq; > + goto res_err; > + } > + > + err = devm_request_irq(dev, hdev->irq, img_irq_handler, 0, > + dev_name(dev), hdev); > + if (err) { > + dev_err(dev, "unable to request %s irq\n", dev_name(dev)); dev_* already prints dev_name(). > + goto res_err; > + } > + dev_dbg(dev, "using IRQ channel %d\n", hdev->irq); > + > + hdev->iclk = devm_clk_get(&pdev->dev, "hash_clk"); > + if (IS_ERR(hdev->iclk)) { > + dev_err(dev, "clock initialization failed.\n"); > + err = PTR_ERR(hdev->iclk); > + goto res_err; > + } > + > + hdev->fclk = devm_clk_get(&pdev->dev, "hash_reg_clk"); I don't see this clock getting enabled anywhere. > + if (IS_ERR(hdev->fclk)) { > + dev_err(dev, "clock initialization failed.\n"); > + err = PTR_ERR(hdev->fclk); > + goto res_err; > + } > + > + err = img_hash_dma_init(hdev); > + if (err) > + goto err_dma; > + > + dev_dbg(dev, "using %s for DMA transfers\n", > + dma_chan_name(hdev->dma_lch)); > + > + spin_lock(&img_hash.lock); > + list_add_tail(&hdev->list, &img_hash.dev_list); > + spin_unlock(&img_hash.lock); > + > + err = img_register_algs(hdev); > + if (err) > + goto err_algs; > + dev_info(dev, "data bus: 0x%x;\tsize:0x%x\n", hdev->bus_addr, 0x4); dev_dbg()? (or maybe not at all) > + dev_info(dev, "Img MD5/SHA1/SHA224/SHA256 Hardware accelerator initialized\n"); > + > + return 0; > + > +err_algs: > + spin_lock(&img_hash.lock); > + list_del(&hdev->list); > + spin_unlock(&img_hash.lock); > + dma_release_channel(hdev->dma_lch); > +err_dma: > +hash_io_err: > + devm_clk_put(dev, hdev->iclk); > + devm_clk_put(dev, hdev->fclk); Since the clocks were acquired via the devm_* API, there's no need to explicitly put() them. > +res_err: > + tasklet_kill(&hdev->done_task); > + tasklet_kill(&hdev->dma_task); > +sha_dev_err: > + dev_err(dev, "initialization failed.\n"); The driver core will print an error message if a device probe fails (for reasons other than -ENODEV, at least), so there's no need to print here. > + return err; > +} > + > +static int img_hash_remove(struct platform_device *pdev) > +{ > + static struct img_hash_dev *hdev; > + > + hdev = platform_get_drvdata(pdev); > + spin_lock(&img_hash.lock); > + list_del(&hdev->list); > + spin_unlock(&img_hash.lock); > + > + img_unregister_algs(hdev); > + > + tasklet_kill(&hdev->done_task); > + tasklet_kill(&hdev->dma_task); > + > + dma_release_channel(hdev->dma_lch); > + > + clk_disable_unprepare(hdev->iclk); > + clk_disable_unprepare(hdev->fclk); Unbalanced prepare_enable()/disable_unprepare(). > + return 0; > +} > + > +static struct platform_driver img_hash_driver = { > + .probe = img_hash_probe, > + .remove = img_hash_remove, > + .driver = { > + .name = "img-hash-accelerator", > + .of_match_table = of_match_ptr(img_hash_match), > + } > +}; > +module_platform_driver(img_hash_driver); > + > +MODULE_LICENSE("GPL v2"); > +MODULE_DESCRIPTION("Imgtec SHA1/224/256 & MD5 hw accelerator driver"); > +MODULE_AUTHOR("Will Thomas."); > +MODULE_AUTHOR("James Hartley."); It's helpful to include email addresses in the MODULE_AUTHOR tags.