drivers/crypto/atmel-sha.c

   1 /*
   2  * Cryptographic API.
   3  *
   4  * Support for ATMEL SHA1/SHA256 HW acceleration.
   5  *
   6  * Copyright (c) 2012 Eukréa Electromatique - ATMEL
   7  * Author: Nicolas Royer <nicolas@eukrea.com>
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License version 2 as published
  11  * by the Free Software Foundation.
  12  *
  13  * Some ideas are from omap-sham.c drivers.
  14  */
  15
  16
  17 #include <linux/kernel.h>
  18 #include <linux/module.h>
  19 #include <linux/slab.h>
  20 #include <linux/err.h>
  21 #include <linux/clk.h>
  22 #include <linux/io.h>
  23 #include <linux/hw_random.h>
  24 #include <linux/platform_device.h>
  25
  26 #include <linux/device.h>
  27 #include <linux/init.h>
  28 #include <linux/errno.h>
  29 #include <linux/interrupt.h>
  30 #include <linux/irq.h>
  31 #include <linux/scatterlist.h>
  32 #include <linux/dma-mapping.h>
  33 #include <linux/of_device.h>
  34 #include <linux/delay.h>
  35 #include <linux/crypto.h>
  36 #include <linux/cryptohash.h>
  37 #include <crypto/scatterwalk.h>
  38 #include <crypto/algapi.h>
  39 #include <crypto/sha.h>
  40 #include <crypto/hash.h>
  41 #include <crypto/internal/hash.h>
  42 #include <linux/platform_data/crypto-atmel.h>
  43 #include "atmel-sha-regs.h"
  44
  45 /* SHA flags */
  46 #define SHA_FLAGS_BUSY                  BIT(0)
  47 #define SHA_FLAGS_FINAL                 BIT(1)
  48 #define SHA_FLAGS_DMA_ACTIVE    BIT(2)
  49 #define SHA_FLAGS_OUTPUT_READY  BIT(3)
  50 #define SHA_FLAGS_INIT                  BIT(4)
  51 #define SHA_FLAGS_CPU                   BIT(5)
  52 #define SHA_FLAGS_DMA_READY             BIT(6)
  53
  54 #define SHA_FLAGS_FINUP         BIT(16)
  55 #define SHA_FLAGS_SG            BIT(17)
  56 #define SHA_FLAGS_SHA1          BIT(18)
  57 #define SHA_FLAGS_SHA224        BIT(19)
  58 #define SHA_FLAGS_SHA256        BIT(20)
  59 #define SHA_FLAGS_SHA384        BIT(21)
  60 #define SHA_FLAGS_SHA512        BIT(22)
  61 #define SHA_FLAGS_ERROR         BIT(23)
  62 #define SHA_FLAGS_PAD           BIT(24)
  63
  64 #define SHA_OP_UPDATE   1
  65 #define SHA_OP_FINAL    2
  66
  67 #define SHA_BUFFER_LEN          PAGE_SIZE
  68
  69 #define ATMEL_SHA_DMA_THRESHOLD         56
  70
  71 struct atmel_sha_caps {
  72         bool    has_dma;
  73         bool    has_dualbuff;
  74         bool    has_sha224;
  75         bool    has_sha_384_512;
  76 };
  77
  78 struct atmel_sha_dev;
  79
  80 struct atmel_sha_reqctx {
  81         struct atmel_sha_dev    *dd;
  82         unsigned long   flags;
  83         unsigned long   op;
  84
  85         u8      digest[SHA512_DIGEST_SIZE] __aligned(sizeof(u32));
  86         u64     digcnt[2];
  87         size_t  bufcnt;
  88         size_t  buflen;
  89         dma_addr_t      dma_addr;
  90
  91         /* walk state */
  92         struct scatterlist      *sg;
  93         unsigned int    offset; /* offset in current sg */
  94         unsigned int    total;  /* total request */
  95
  96         size_t block_size;
  97
  98         u8      buffer[0] __aligned(sizeof(u32));
  99 };
 100
 101 struct atmel_sha_ctx {
 102         struct atmel_sha_dev    *dd;
 103
 104         unsigned long           flags;
 105
 106         /* fallback stuff */
 107         struct crypto_shash     *fallback;
 108
 109 };
 110
 111 #define ATMEL_SHA_QUEUE_LENGTH  50
 112
 113 struct atmel_sha_dma {
 114         struct dma_chan                 *chan;
 115         struct dma_slave_config dma_conf;
 116 };
 117
 118 struct atmel_sha_dev {
 119         struct list_head        list;
 120         unsigned long           phys_base;
 121         struct device           *dev;
 122         struct clk                      *iclk;
 123         int                                     irq;
 124         void __iomem            *io_base;
 125
 126         spinlock_t              lock;
 127         int                     err;
 128         struct tasklet_struct   done_task;
 129
 130         unsigned long           flags;
 131         struct crypto_queue     queue;
 132         struct ahash_request    *req;
 133
 134         struct atmel_sha_dma    dma_lch_in;
 135
 136         struct atmel_sha_caps   caps;
 137
 138         u32     hw_version;
 139 };
 140
 141 struct atmel_sha_drv {
 142         struct list_head        dev_list;
 143         spinlock_t              lock;
 144 };
 145
 146 static struct atmel_sha_drv atmel_sha = {
 147         .dev_list = LIST_HEAD_INIT(atmel_sha.dev_list),
 148         .lock = __SPIN_LOCK_UNLOCKED(atmel_sha.lock),
 149 };
 150
 151 static inline u32 atmel_sha_read(struct atmel_sha_dev *dd, u32 offset)
 152 {
 153         return readl_relaxed(dd->io_base + offset);
 154 }
 155
 156 static inline void atmel_sha_write(struct atmel_sha_dev *dd,
 157                                         u32 offset, u32 value)
 158 {
 159         writel_relaxed(value, dd->io_base + offset);
 160 }
 161
 162 static size_t atmel_sha_append_sg(struct atmel_sha_reqctx *ctx)
 163 {
 164         size_t count;
 165
 166         while ((ctx->bufcnt < ctx->buflen) && ctx->total) {
 167                 count = min(ctx->sg->length - ctx->offset, ctx->total);
 168                 count = min(count, ctx->buflen - ctx->bufcnt);
 169
 170                 if (count <= 0)
 171                         break;
 172
 173                 scatterwalk_map_and_copy(ctx->buffer + ctx->bufcnt, ctx->sg,
 174                         ctx->offset, count, 0);
 175
 176                 ctx->bufcnt += count;
 177                 ctx->offset += count;
 178                 ctx->total -= count;
 179
 180                 if (ctx->offset == ctx->sg->length) {
 181                         ctx->sg = sg_next(ctx->sg);
 182                         if (ctx->sg)
 183                                 ctx->offset = 0;
 184                         else
 185                                 ctx->total = 0;
 186                 }
 187         }
 188
 189         return 0;
 190 }
 191
 192 /*
 193  * The purpose of this padding is to ensure that the padded message is a
 194  * multiple of 512 bits (SHA1/SHA224/SHA256) or 1024 bits (SHA384/SHA512).
 195  * The bit "1" is appended at the end of the message followed by
 196  * "padlen-1" zero bits. Then a 64 bits block (SHA1/SHA224/SHA256) or
 197  * 128 bits block (SHA384/SHA512) equals to the message length in bits
 198  * is appended.
 199  *
 200  * For SHA1/SHA224/SHA256, padlen is calculated as followed:
 201  *  - if message length < 56 bytes then padlen = 56 - message length
 202  *  - else padlen = 64 + 56 - message length
 203  *
 204  * For SHA384/SHA512, padlen is calculated as followed:
 205  *  - if message length < 112 bytes then padlen = 112 - message length
 206  *  - else padlen = 128 + 112 - message length
 207  */
 208 static void atmel_sha_fill_padding(struct atmel_sha_reqctx *ctx, int length)
 209 {
 210         unsigned int index, padlen;
 211         u64 bits[2];
 212         u64 size[2];
 213
 214         size[0] = ctx->digcnt[0];
 215         size[1] = ctx->digcnt[1];
 216
 217         size[0] += ctx->bufcnt;
 218         if (size[0] < ctx->bufcnt)
 219                 size[1]++;
 220
 221         size[0] += length;
 222         if (size[0]  < length)
 223                 size[1]++;
 224
 225         bits[1] = cpu_to_be64(size[0] << 3);
 226         bits[0] = cpu_to_be64(size[1] << 3 | size[0] >> 61);
 227
 228         if (ctx->flags & (SHA_FLAGS_SHA384 | SHA_FLAGS_SHA512)) {
 229                 index = ctx->bufcnt & 0x7f;
 230                 padlen = (index < 112) ? (112 - index) : ((128+112) - index);
 231                 *(ctx->buffer + ctx->bufcnt) = 0x80;
 232                 memset(ctx->buffer + ctx->bufcnt + 1, 0, padlen-1);
 233                 memcpy(ctx->buffer + ctx->bufcnt + padlen, bits, 16);
 234                 ctx->bufcnt += padlen + 16;
 235                 ctx->flags |= SHA_FLAGS_PAD;
 236         } else {
 237                 index = ctx->bufcnt & 0x3f;
 238                 padlen = (index < 56) ? (56 - index) : ((64+56) - index);
 239                 *(ctx->buffer + ctx->bufcnt) = 0x80;
 240                 memset(ctx->buffer + ctx->bufcnt + 1, 0, padlen-1);
 241                 memcpy(ctx->buffer + ctx->bufcnt + padlen, &bits[1], 8);
 242                 ctx->bufcnt += padlen + 8;
 243                 ctx->flags |= SHA_FLAGS_PAD;
 244         }
 245 }
 246
 247 static int atmel_sha_init(struct ahash_request *req)
 248 {
 249         struct crypto_ahash *tfm = crypto_ahash_reqtfm(req);
 250         struct atmel_sha_ctx *tctx = crypto_ahash_ctx(tfm);
 251         struct atmel_sha_reqctx *ctx = ahash_request_ctx(req);
 252         struct atmel_sha_dev *dd = NULL;
 253         struct atmel_sha_dev *tmp;
 254
 255         spin_lock_bh(&atmel_sha.lock);
 256         if (!tctx->dd) {
 257                 list_for_each_entry(tmp, &atmel_sha.dev_list, list) {
 258                         dd = tmp;
 259                         break;
 260                 }
 261                 tctx->dd = dd;
 262         } else {
 263                 dd = tctx->dd;
 264         }
 265
 266         spin_unlock_bh(&atmel_sha.lock);
 267
 268         ctx->dd = dd;
 269
 270         ctx->flags = 0;
 271
 272         dev_dbg(dd->dev, "init: digest size: %d\n",
 273                 crypto_ahash_digestsize(tfm));
 274
 275         switch (crypto_ahash_digestsize(tfm)) {
 276         case SHA1_DIGEST_SIZE:
 277                 ctx->flags |= SHA_FLAGS_SHA1;
 278                 ctx->block_size = SHA1_BLOCK_SIZE;
 279                 break;
 280         case SHA224_DIGEST_SIZE:
 281                 ctx->flags |= SHA_FLAGS_SHA224;
 282                 ctx->block_size = SHA224_BLOCK_SIZE;
 283                 break;
 284         case SHA256_DIGEST_SIZE:
 285                 ctx->flags |= SHA_FLAGS_SHA256;
 286                 ctx->block_size = SHA256_BLOCK_SIZE;
 287                 break;
 288         case SHA384_DIGEST_SIZE:
 289                 ctx->flags |= SHA_FLAGS_SHA384;
 290                 ctx->block_size = SHA384_BLOCK_SIZE;
 291                 break;
 292         case SHA512_DIGEST_SIZE:
 293                 ctx->flags |= SHA_FLAGS_SHA512;
 294                 ctx->block_size = SHA512_BLOCK_SIZE;
 295                 break;
 296         default:
 297                 return -EINVAL;
 298                 break;
 299         }
 300
 301         ctx->bufcnt = 0;
 302         ctx->digcnt[0] = 0;
 303         ctx->digcnt[1] = 0;
 304         ctx->buflen = SHA_BUFFER_LEN;
 305
 306         return 0;
 307 }
 308
 309 static void atmel_sha_write_ctrl(struct atmel_sha_dev *dd, int dma)
 310 {
 311         struct atmel_sha_reqctx *ctx = ahash_request_ctx(dd->req);
 312         u32 valcr = 0, valmr = SHA_MR_MODE_AUTO;
 313
 314         if (likely(dma)) {
 315                 if (!dd->caps.has_dma)
 316                         atmel_sha_write(dd, SHA_IER, SHA_INT_TXBUFE);
 317                 valmr = SHA_MR_MODE_PDC;
 318                 if (dd->caps.has_dualbuff)
 319                         valmr |= SHA_MR_DUALBUFF;
 320         } else {
 321                 atmel_sha_write(dd, SHA_IER, SHA_INT_DATARDY);
 322         }
 323
 324         if (ctx->flags & SHA_FLAGS_SHA1)
 325                 valmr |= SHA_MR_ALGO_SHA1;
 326         else if (ctx->flags & SHA_FLAGS_SHA224)
 327                 valmr |= SHA_MR_ALGO_SHA224;
 328         else if (ctx->flags & SHA_FLAGS_SHA256)
 329                 valmr |= SHA_MR_ALGO_SHA256;
 330         else if (ctx->flags & SHA_FLAGS_SHA384)
 331                 valmr |= SHA_MR_ALGO_SHA384;
 332         else if (ctx->flags & SHA_FLAGS_SHA512)
 333                 valmr |= SHA_MR_ALGO_SHA512;
 334
 335         /* Setting CR_FIRST only for the first iteration */
 336         if (!(ctx->digcnt[0] || ctx->digcnt[1]))
 337                 valcr = SHA_CR_FIRST;
 338
 339         atmel_sha_write(dd, SHA_CR, valcr);
 340         atmel_sha_write(dd, SHA_MR, valmr);
 341 }
 342
 343 static int atmel_sha_xmit_cpu(struct atmel_sha_dev *dd, const u8 *buf,
 344                               size_t length, int final)
 345 {
 346         struct atmel_sha_reqctx *ctx = ahash_request_ctx(dd->req);
 347         int count, len32;
 348         const u32 *buffer = (const u32 *)buf;
 349
 350         dev_dbg(dd->dev, "xmit_cpu: digcnt: 0x%llx 0x%llx, length: %d, final: %d\n",
 351                 ctx->digcnt[1], ctx->digcnt[0], length, final);
 352
 353         atmel_sha_write_ctrl(dd, 0);
 354
 355         /* should be non-zero before next lines to disable clocks later */
 356         ctx->digcnt[0] += length;
 357         if (ctx->digcnt[0] < length)
 358                 ctx->digcnt[1]++;
 359
 360         if (final)
 361                 dd->flags |= SHA_FLAGS_FINAL; /* catch last interrupt */
 362
 363         len32 = DIV_ROUND_UP(length, sizeof(u32));
 364
 365         dd->flags |= SHA_FLAGS_CPU;
 366
 367         for (count = 0; count < len32; count++)
 368                 atmel_sha_write(dd, SHA_REG_DIN(count), buffer[count]);
 369
 370         return -EINPROGRESS;
 371 }
 372
 373 static int atmel_sha_xmit_pdc(struct atmel_sha_dev *dd, dma_addr_t dma_addr1,
 374                 size_t length1, dma_addr_t dma_addr2, size_t length2, int final)
 375 {
 376         struct atmel_sha_reqctx *ctx = ahash_request_ctx(dd->req);
 377         int len32;
 378
 379         dev_dbg(dd->dev, "xmit_pdc: digcnt: 0x%llx 0x%llx, length: %d, final: %d\n",
 380                 ctx->digcnt[1], ctx->digcnt[0], length1, final);
 381
 382         len32 = DIV_ROUND_UP(length1, sizeof(u32));
 383         atmel_sha_write(dd, SHA_PTCR, SHA_PTCR_TXTDIS);
 384         atmel_sha_write(dd, SHA_TPR, dma_addr1);
 385         atmel_sha_write(dd, SHA_TCR, len32);
 386
 387         len32 = DIV_ROUND_UP(length2, sizeof(u32));
 388         atmel_sha_write(dd, SHA_TNPR, dma_addr2);
 389         atmel_sha_write(dd, SHA_TNCR, len32);
 390
 391         atmel_sha_write_ctrl(dd, 1);
 392
 393         /* should be non-zero before next lines to disable clocks later */
 394         ctx->digcnt[0] += length1;
 395         if (ctx->digcnt[0] < length1)
 396                 ctx->digcnt[1]++;
 397
 398         if (final)
 399                 dd->flags |= SHA_FLAGS_FINAL; /* catch last interrupt */
 400
 401         dd->flags |=  SHA_FLAGS_DMA_ACTIVE;
 402
 403         /* Start DMA transfer */
 404         atmel_sha_write(dd, SHA_PTCR, SHA_PTCR_TXTEN);
 405
 406         return -EINPROGRESS;
 407 }
 408
 409 static void atmel_sha_dma_callback(void *data)
 410 {
 411         struct atmel_sha_dev *dd = data;
 412
 413         /* dma_lch_in - completed - wait DATRDY */
 414         atmel_sha_write(dd, SHA_IER, SHA_INT_DATARDY);
 415 }
 416
 417 static int atmel_sha_xmit_dma(struct atmel_sha_dev *dd, dma_addr_t dma_addr1,
 418                 size_t length1, dma_addr_t dma_addr2, size_t length2, int final)
 419 {
 420         struct atmel_sha_reqctx *ctx = ahash_request_ctx(dd->req);
 421         struct dma_async_tx_descriptor  *in_desc;
 422         struct scatterlist sg[2];
 423
 424         dev_dbg(dd->dev, "xmit_dma: digcnt: 0x%llx 0x%llx, length: %d, final: %d\n",
 425                 ctx->digcnt[1], ctx->digcnt[0], length1, final);
 426
 427         if (ctx->flags & (SHA_FLAGS_SHA1 | SHA_FLAGS_SHA224 |
 428                         SHA_FLAGS_SHA256)) {
 429                 dd->dma_lch_in.dma_conf.src_maxburst = 16;
 430                 dd->dma_lch_in.dma_conf.dst_maxburst = 16;
 431         } else {
 432                 dd->dma_lch_in.dma_conf.src_maxburst = 32;
 433                 dd->dma_lch_in.dma_conf.dst_maxburst = 32;
 434         }
 435
 436         dmaengine_slave_config(dd->dma_lch_in.chan, &dd->dma_lch_in.dma_conf);
 437
 438         if (length2) {
 439                 sg_init_table(sg, 2);
 440                 sg_dma_address(&sg[0]) = dma_addr1;
 441                 sg_dma_len(&sg[0]) = length1;
 442                 sg_dma_address(&sg[1]) = dma_addr2;
 443                 sg_dma_len(&sg[1]) = length2;
 444                 in_desc = dmaengine_prep_slave_sg(dd->dma_lch_in.chan, sg, 2,
 445                         DMA_MEM_TO_DEV, DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
 446         } else {
 447                 sg_init_table(sg, 1);
 448                 sg_dma_address(&sg[0]) = dma_addr1;
 449                 sg_dma_len(&sg[0]) = length1;
 450                 in_desc = dmaengine_prep_slave_sg(dd->dma_lch_in.chan, sg, 1,
 451                         DMA_MEM_TO_DEV, DMA_PREP_INTERRUPT | DMA_CTRL_ACK);
 452         }
 453         if (!in_desc)
 454                 return -EINVAL;
 455
 456         in_desc->callback = atmel_sha_dma_callback;
 457         in_desc->callback_param = dd;
 458
 459         atmel_sha_write_ctrl(dd, 1);
 460
 461         /* should be non-zero before next lines to disable clocks later */
 462         ctx->digcnt[0] += length1;
 463         if (ctx->digcnt[0] < length1)
 464                 ctx->digcnt[1]++;
 465
 466         if (final)
 467                 dd->flags |= SHA_FLAGS_FINAL; /* catch last interrupt */
 468
 469         dd->flags |=  SHA_FLAGS_DMA_ACTIVE;
 470
 471         /* Start DMA transfer */
 472         dmaengine_submit(in_desc);
 473         dma_async_issue_pending(dd->dma_lch_in.chan);
 474
 475         return -EINPROGRESS;
 476 }
 477
 478 static int atmel_sha_xmit_start(struct atmel_sha_dev *dd, dma_addr_t dma_addr1,
 479                 size_t length1, dma_addr_t dma_addr2, size_t length2, int final)
 480 {
 481         if (dd->caps.has_dma)
 482                 return atmel_sha_xmit_dma(dd, dma_addr1, length1,
 483                                 dma_addr2, length2, final);
 484         else
 485                 return atmel_sha_xmit_pdc(dd, dma_addr1, length1,
 486                                 dma_addr2, length2, final);
 487 }
 488
 489 static int atmel_sha_update_cpu(struct atmel_sha_dev *dd)
 490 {
 491         struct atmel_sha_reqctx *ctx = ahash_request_ctx(dd->req);
 492         int bufcnt;
 493
 494         atmel_sha_append_sg(ctx);
 495         atmel_sha_fill_padding(ctx, 0);
 496         bufcnt = ctx->bufcnt;
 497         ctx->bufcnt = 0;
 498
 499         return atmel_sha_xmit_cpu(dd, ctx->buffer, bufcnt, 1);
 500 }
 501
 502 static int atmel_sha_xmit_dma_map(struct atmel_sha_dev *dd,
 503                                         struct atmel_sha_reqctx *ctx,
 504                                         size_t length, int final)
 505 {
 506         ctx->dma_addr = dma_map_single(dd->dev, ctx->buffer,
 507                                 ctx->buflen + ctx->block_size, DMA_TO_DEVICE);
 508         if (dma_mapping_error(dd->dev, ctx->dma_addr)) {
 509                 dev_err(dd->dev, "dma %u bytes error\n", ctx->buflen +
 510                                 ctx->block_size);
 511                 return -EINVAL;
 512         }
 513
 514         ctx->flags &= ~SHA_FLAGS_SG;
 515
 516         /* next call does not fail... so no unmap in the case of error */
 517         return atmel_sha_xmit_start(dd, ctx->dma_addr, length, 0, 0, final);
 518 }
 519
 520 static int atmel_sha_update_dma_slow(struct atmel_sha_dev *dd)
 521 {
 522         struct atmel_sha_reqctx *ctx = ahash_request_ctx(dd->req);
 523         unsigned int final;
 524         size_t count;
 525
 526         atmel_sha_append_sg(ctx);
 527
 528         final = (ctx->flags & SHA_FLAGS_FINUP) && !ctx->total;
 529
 530         dev_dbg(dd->dev, "slow: bufcnt: %u, digcnt: 0x%llx 0x%llx, final: %d\n",
 531                  ctx->bufcnt, ctx->digcnt[1], ctx->digcnt[0], final);
 532
 533         if (final)
 534                 atmel_sha_fill_padding(ctx, 0);
 535
 536         if (final || (ctx->bufcnt == ctx->buflen && ctx->total)) {
 537                 count = ctx->bufcnt;
 538                 ctx->bufcnt = 0;
 539                 return atmel_sha_xmit_dma_map(dd, ctx, count, final);
 540         }
 541
 542         return 0;
 543 }
 544
 545 static int atmel_sha_update_dma_start(struct atmel_sha_dev *dd)
 546 {
 547         struct atmel_sha_reqctx *ctx = ahash_request_ctx(dd->req);
 548         unsigned int length, final, tail;
 549         struct scatterlist *sg;
 550         unsigned int count;
 551
 552         if (!ctx->total)
 553                 return 0;
 554
 555         if (ctx->bufcnt || ctx->offset)
 556                 return atmel_sha_update_dma_slow(dd);
 557
 558         dev_dbg(dd->dev, "fast: digcnt: 0x%llx 0x%llx, bufcnt: %u, total: %u\n",
 559                 ctx->digcnt[1], ctx->digcnt[0], ctx->bufcnt, ctx->total);
 560
 561         sg = ctx->sg;
 562
 563         if (!IS_ALIGNED(sg->offset, sizeof(u32)))
 564                 return atmel_sha_update_dma_slow(dd);
 565
 566         if (!sg_is_last(sg) && !IS_ALIGNED(sg->length, ctx->block_size))
 567                 /* size is not ctx->block_size aligned */
 568                 return atmel_sha_update_dma_slow(dd);
 569
 570         length = min(ctx->total, sg->length);
 571
 572         if (sg_is_last(sg)) {
 573                 if (!(ctx->flags & SHA_FLAGS_FINUP)) {
 574                         /* not last sg must be ctx->block_size aligned */
 575                         tail = length & (ctx->block_size - 1);
 576                         length -= tail;
 577                 }
 578         }
 579
 580         ctx->total -= length;
 581         ctx->offset = length; /* offset where to start slow */
 582
 583         final = (ctx->flags & SHA_FLAGS_FINUP) && !ctx->total;
 584
 585         /* Add padding */
 586         if (final) {
 587                 tail = length & (ctx->block_size - 1);
 588                 length -= tail;
 589                 ctx->total += tail;
 590                 ctx->offset = length; /* offset where to start slow */
 591
 592                 sg = ctx->sg;
 593                 atmel_sha_append_sg(ctx);
 594
 595                 atmel_sha_fill_padding(ctx, length);
 596
 597                 ctx->dma_addr = dma_map_single(dd->dev, ctx->buffer,
 598                         ctx->buflen + ctx->block_size, DMA_TO_DEVICE);
 599                 if (dma_mapping_error(dd->dev, ctx->dma_addr)) {
 600                         dev_err(dd->dev, "dma %u bytes error\n",
 601                                 ctx->buflen + ctx->block_size);
 602                         return -EINVAL;
 603                 }
 604
 605                 if (length == 0) {
 606                         ctx->flags &= ~SHA_FLAGS_SG;
 607                         count = ctx->bufcnt;
 608                         ctx->bufcnt = 0;
 609                         return atmel_sha_xmit_start(dd, ctx->dma_addr, count, 0,
 610                                         0, final);
 611                 } else {
 612                         ctx->sg = sg;
 613                         if (!dma_map_sg(dd->dev, ctx->sg, 1,
 614                                 DMA_TO_DEVICE)) {
 615                                         dev_err(dd->dev, "dma_map_sg  error\n");
 616                                         return -EINVAL;
 617                         }
 618
 619                         ctx->flags |= SHA_FLAGS_SG;
 620
 621                         count = ctx->bufcnt;
 622                         ctx->bufcnt = 0;
 623                         return atmel_sha_xmit_start(dd, sg_dma_address(ctx->sg),
 624                                         length, ctx->dma_addr, count, final);
 625                 }
 626         }
 627
 628         if (!dma_map_sg(dd->dev, ctx->sg, 1, DMA_TO_DEVICE)) {
 629                 dev_err(dd->dev, "dma_map_sg  error\n");
 630                 return -EINVAL;
 631         }
 632
 633         ctx->flags |= SHA_FLAGS_SG;
 634
 635         /* next call does not fail... so no unmap in the case of error */
 636         return atmel_sha_xmit_start(dd, sg_dma_address(ctx->sg), length, 0,
 637                                                                 0, final);
 638 }
 639
 640 static int atmel_sha_update_dma_stop(struct atmel_sha_dev *dd)
 641 {
 642         struct atmel_sha_reqctx *ctx = ahash_request_ctx(dd->req);
 643
 644         if (ctx->flags & SHA_FLAGS_SG) {
 645                 dma_unmap_sg(dd->dev, ctx->sg, 1, DMA_TO_DEVICE);
 646                 if (ctx->sg->length == ctx->offset) {
 647                         ctx->sg = sg_next(ctx->sg);
 648                         if (ctx->sg)
 649                                 ctx->offset = 0;
 650                 }
 651                 if (ctx->flags & SHA_FLAGS_PAD) {
 652                         dma_unmap_single(dd->dev, ctx->dma_addr,
 653                                 ctx->buflen + ctx->block_size, DMA_TO_DEVICE);
 654                 }
 655         } else {
 656                 dma_unmap_single(dd->dev, ctx->dma_addr, ctx->buflen +
 657                                                 ctx->block_size, DMA_TO_DEVICE);
 658         }
 659
 660         return 0;
 661 }
 662
 663 static int atmel_sha_update_req(struct atmel_sha_dev *dd)
 664 {
 665         struct ahash_request *req = dd->req;
 666         struct atmel_sha_reqctx *ctx = ahash_request_ctx(req);
 667         int err;
 668
 669         dev_dbg(dd->dev, "update_req: total: %u, digcnt: 0x%llx 0x%llx\n",
 670                 ctx->total, ctx->digcnt[1], ctx->digcnt[0]);
 671
 672         if (ctx->flags & SHA_FLAGS_CPU)
 673                 err = atmel_sha_update_cpu(dd);
 674         else
 675                 err = atmel_sha_update_dma_start(dd);
 676
 677         /* wait for dma completion before can take more data */
 678         dev_dbg(dd->dev, "update: err: %d, digcnt: 0x%llx 0%llx\n",
 679                         err, ctx->digcnt[1], ctx->digcnt[0]);
 680
 681         return err;
 682 }
 683
 684 static int atmel_sha_final_req(struct atmel_sha_dev *dd)
 685 {
 686         struct ahash_request *req = dd->req;
 687         struct atmel_sha_reqctx *ctx = ahash_request_ctx(req);
 688         int err = 0;
 689         int count;
 690
 691         if (ctx->bufcnt >= ATMEL_SHA_DMA_THRESHOLD) {
 692                 atmel_sha_fill_padding(ctx, 0);
 693                 count = ctx->bufcnt;
 694                 ctx->bufcnt = 0;
 695                 err = atmel_sha_xmit_dma_map(dd, ctx, count, 1);
 696         }
 697         /* faster to handle last block with cpu */
 698         else {
 699                 atmel_sha_fill_padding(ctx, 0);
 700                 count = ctx->bufcnt;
 701                 ctx->bufcnt = 0;
 702                 err = atmel_sha_xmit_cpu(dd, ctx->buffer, count, 1);
 703         }
 704
 705         dev_dbg(dd->dev, "final_req: err: %d\n", err);
 706
 707         return err;
 708 }
 709
 710 static void atmel_sha_copy_hash(struct ahash_request *req)
 711 {
 712         struct atmel_sha_reqctx *ctx = ahash_request_ctx(req);
 713         u32 *hash = (u32 *)ctx->digest;
 714         int i;
 715
 716         if (ctx->flags & SHA_FLAGS_SHA1)
 717                 for (i = 0; i < SHA1_DIGEST_SIZE / sizeof(u32); i++)
 718                         hash[i] = atmel_sha_read(ctx->dd, SHA_REG_DIGEST(i));
 719         else if (ctx->flags & SHA_FLAGS_SHA224)
 720                 for (i = 0; i < SHA224_DIGEST_SIZE / sizeof(u32); i++)
 721                         hash[i] = atmel_sha_read(ctx->dd, SHA_REG_DIGEST(i));
 722         else if (ctx->flags & SHA_FLAGS_SHA256)
 723                 for (i = 0; i < SHA256_DIGEST_SIZE / sizeof(u32); i++)
 724                         hash[i] = atmel_sha_read(ctx->dd, SHA_REG_DIGEST(i));
 725         else if (ctx->flags & SHA_FLAGS_SHA384)
 726                 for (i = 0; i < SHA384_DIGEST_SIZE / sizeof(u32); i++)
 727                         hash[i] = atmel_sha_read(ctx->dd, SHA_REG_DIGEST(i));
 728         else
 729                 for (i = 0; i < SHA512_DIGEST_SIZE / sizeof(u32); i++)
 730                         hash[i] = atmel_sha_read(ctx->dd, SHA_REG_DIGEST(i));
 731 }
 732
 733 static void atmel_sha_copy_ready_hash(struct ahash_request *req)
 734 {
 735         struct atmel_sha_reqctx *ctx = ahash_request_ctx(req);
 736
 737         if (!req->result)
 738                 return;
 739
 740         if (ctx->flags & SHA_FLAGS_SHA1)
 741                 memcpy(req->result, ctx->digest, SHA1_DIGEST_SIZE);
 742         else if (ctx->flags & SHA_FLAGS_SHA224)
 743                 memcpy(req->result, ctx->digest, SHA224_DIGEST_SIZE);
 744         else if (ctx->flags & SHA_FLAGS_SHA256)
 745                 memcpy(req->result, ctx->digest, SHA256_DIGEST_SIZE);
 746         else if (ctx->flags & SHA_FLAGS_SHA384)
 747                 memcpy(req->result, ctx->digest, SHA384_DIGEST_SIZE);
 748         else
 749                 memcpy(req->result, ctx->digest, SHA512_DIGEST_SIZE);
 750 }
 751
 752 static int atmel_sha_finish(struct ahash_request *req)
 753 {
 754         struct atmel_sha_reqctx *ctx = ahash_request_ctx(req);
 755         struct atmel_sha_dev *dd = ctx->dd;
 756         int err = 0;
 757
 758         if (ctx->digcnt[0] || ctx->digcnt[1])
 759                 atmel_sha_copy_ready_hash(req);
 760
 761         dev_dbg(dd->dev, "digcnt: 0x%llx 0x%llx, bufcnt: %d\n", ctx->digcnt[1],
 762                 ctx->digcnt[0], ctx->bufcnt);
 763
 764         return err;
 765 }
 766
 767 static void atmel_sha_finish_req(struct ahash_request *req, int err)
 768 {
 769         struct atmel_sha_reqctx *ctx = ahash_request_ctx(req);
 770         struct atmel_sha_dev *dd = ctx->dd;
 771
 772         if (!err) {
 773                 atmel_sha_copy_hash(req);
 774                 if (SHA_FLAGS_FINAL & dd->flags)
 775                         err = atmel_sha_finish(req);
 776         } else {
 777                 ctx->flags |= SHA_FLAGS_ERROR;
 778         }
 779
 780         /* atomic operation is not needed here */
 781         dd->flags &= ~(SHA_FLAGS_BUSY | SHA_FLAGS_FINAL | SHA_FLAGS_CPU |
 782                         SHA_FLAGS_DMA_READY | SHA_FLAGS_OUTPUT_READY);
 783
 784         clk_disable_unprepare(dd->iclk);
 785
 786         if (req->base.complete)
 787                 req->base.complete(&req->base, err);
 788
 789         /* handle new request */
 790         tasklet_schedule(&dd->done_task);
 791 }
 792
 793 static int atmel_sha_hw_init(struct atmel_sha_dev *dd)
 794 {
 795         clk_prepare_enable(dd->iclk);
 796
 797         if (!(SHA_FLAGS_INIT & dd->flags)) {
 798                 atmel_sha_write(dd, SHA_CR, SHA_CR_SWRST);
 799                 dd->flags |= SHA_FLAGS_INIT;
 800                 dd->err = 0;
 801         }
 802
 803         return 0;
 804 }
 805
 806 static inline unsigned int atmel_sha_get_version(struct atmel_sha_dev *dd)
 807 {
 808         return atmel_sha_read(dd, SHA_HW_VERSION) & 0x00000fff;
 809 }
 810
 811 static void atmel_sha_hw_version_init(struct atmel_sha_dev *dd)
 812 {
 813         atmel_sha_hw_init(dd);
 814
 815         dd->hw_version = atmel_sha_get_version(dd);
 816
 817         dev_info(dd->dev,
 818                         "version: 0x%x\n", dd->hw_version);
 819
 820         clk_disable_unprepare(dd->iclk);
 821 }
 822
 823 static int atmel_sha_handle_queue(struct atmel_sha_dev *dd,
 824                                   struct ahash_request *req)
 825 {
 826         struct crypto_async_request *async_req, *backlog;
 827         struct atmel_sha_reqctx *ctx;
 828         unsigned long flags;
 829         int err = 0, ret = 0;
 830
 831         spin_lock_irqsave(&dd->lock, flags);
 832         if (req)
 833                 ret = ahash_enqueue_request(&dd->queue, req);
 834
 835         if (SHA_FLAGS_BUSY & dd->flags) {
 836                 spin_unlock_irqrestore(&dd->lock, flags);
 837                 return ret;
 838         }
 839
 840         backlog = crypto_get_backlog(&dd->queue);
 841         async_req = crypto_dequeue_request(&dd->queue);
 842         if (async_req)
 843                 dd->flags |= SHA_FLAGS_BUSY;
 844
 845         spin_unlock_irqrestore(&dd->lock, flags);
 846
 847         if (!async_req)
 848                 return ret;
 849
 850         if (backlog)
 851                 backlog->complete(backlog, -EINPROGRESS);
 852
 853         req = ahash_request_cast(async_req);
 854         dd->req = req;
 855         ctx = ahash_request_ctx(req);
 856
 857         dev_dbg(dd->dev, "handling new req, op: %lu, nbytes: %d\n",
 858                                                 ctx->op, req->nbytes);
 859
 860         err = atmel_sha_hw_init(dd);
 861
 862         if (err)
 863                 goto err1;
 864
 865         if (ctx->op == SHA_OP_UPDATE) {
 866                 err = atmel_sha_update_req(dd);
 867                 if (err != -EINPROGRESS && (ctx->flags & SHA_FLAGS_FINUP))
 868                         /* no final() after finup() */
 869                         err = atmel_sha_final_req(dd);
 870         } else if (ctx->op == SHA_OP_FINAL) {
 871                 err = atmel_sha_final_req(dd);
 872         }
 873
 874 err1:
 875         if (err != -EINPROGRESS)
 876                 /* done_task will not finish it, so do it here */
 877                 atmel_sha_finish_req(req, err);
 878
 879         dev_dbg(dd->dev, "exit, err: %d\n", err);
 880
 881         return ret;
 882 }
 883
 884 static int atmel_sha_enqueue(struct ahash_request *req, unsigned int op)
 885 {
 886         struct atmel_sha_reqctx *ctx = ahash_request_ctx(req);
 887         struct atmel_sha_ctx *tctx = crypto_tfm_ctx(req->base.tfm);
 888         struct atmel_sha_dev *dd = tctx->dd;
 889
 890         ctx->op = op;
 891
 892         return atmel_sha_handle_queue(dd, req);
 893 }
 894
 895 static int atmel_sha_update(struct ahash_request *req)
 896 {
 897         struct atmel_sha_reqctx *ctx = ahash_request_ctx(req);
 898
 899         if (!req->nbytes)
 900                 return 0;
 901
 902         ctx->total = req->nbytes;
 903         ctx->sg = req->src;
 904         ctx->offset = 0;
 905
 906         if (ctx->flags & SHA_FLAGS_FINUP) {
 907                 if (ctx->bufcnt + ctx->total < ATMEL_SHA_DMA_THRESHOLD)
 908                         /* faster to use CPU for short transfers */
 909                         ctx->flags |= SHA_FLAGS_CPU;
 910         } else if (ctx->bufcnt + ctx->total < ctx->buflen) {
 911                 atmel_sha_append_sg(ctx);
 912                 return 0;
 913         }
 914         return atmel_sha_enqueue(req, SHA_OP_UPDATE);
 915 }
 916
 917 static int atmel_sha_final(struct ahash_request *req)
 918 {
 919         struct atmel_sha_reqctx *ctx = ahash_request_ctx(req);
 920         struct atmel_sha_ctx *tctx = crypto_tfm_ctx(req->base.tfm);
 921         struct atmel_sha_dev *dd = tctx->dd;
 922
 923         int err = 0;
 924
 925         ctx->flags |= SHA_FLAGS_FINUP;
 926
 927         if (ctx->flags & SHA_FLAGS_ERROR)
 928                 return 0; /* uncompleted hash is not needed */
 929
 930         if (ctx->bufcnt) {
 931                 return atmel_sha_enqueue(req, SHA_OP_FINAL);
 932         } else if (!(ctx->flags & SHA_FLAGS_PAD)) { /* add padding */
 933                 err = atmel_sha_hw_init(dd);
 934                 if (err)
 935                         goto err1;
 936
 937                 dd->flags |= SHA_FLAGS_BUSY;
 938                 err = atmel_sha_final_req(dd);
 939         } else {
 940                 /* copy ready hash (+ finalize hmac) */
 941                 return atmel_sha_finish(req);
 942         }
 943
 944 err1:
 945         if (err != -EINPROGRESS)
 946                 /* done_task will not finish it, so do it here */
 947                 atmel_sha_finish_req(req, err);
 948
 949         return err;
 950 }
 951
 952 static int atmel_sha_finup(struct ahash_request *req)
 953 {
 954         struct atmel_sha_reqctx *ctx = ahash_request_ctx(req);
 955         int err1, err2;
 956
 957         ctx->flags |= SHA_FLAGS_FINUP;
 958
 959         err1 = atmel_sha_update(req);
 960         if (err1 == -EINPROGRESS ||
 961             (err1 == -EBUSY && (ahash_request_flags(req) &
 962                                 CRYPTO_TFM_REQ_MAY_BACKLOG)))
 963                 return err1;
 964
 965         /*
 966          * final() has to be always called to cleanup resources
 967          * even if udpate() failed, except EINPROGRESS
 968          */
 969         err2 = atmel_sha_final(req);
 970
 971         return err1 ?: err2;
 972 }
 973
 974 static int atmel_sha_digest(struct ahash_request *req)
 975 {
 976         return atmel_sha_init(req) ?: atmel_sha_finup(req);
 977 }
 978
 979 static int atmel_sha_cra_init_alg(struct crypto_tfm *tfm, const char *alg_base)
 980 {
 981         struct atmel_sha_ctx *tctx = crypto_tfm_ctx(tfm);
 982         const char *alg_name = crypto_tfm_alg_name(tfm);
 983
 984         /* Allocate a fallback and abort if it failed. */
 985         tctx->fallback = crypto_alloc_shash(alg_name, 0,
 986                                             CRYPTO_ALG_NEED_FALLBACK);
 987         if (IS_ERR(tctx->fallback)) {
 988                 pr_err("atmel-sha: fallback driver '%s' could not be loaded.\n",
 989                                 alg_name);
 990                 return PTR_ERR(tctx->fallback);
 991         }
 992         crypto_ahash_set_reqsize(__crypto_ahash_cast(tfm),
 993                                  sizeof(struct atmel_sha_reqctx) +
 994                                  SHA_BUFFER_LEN + SHA512_BLOCK_SIZE);
 995
 996         return 0;
 997 }
 998
 999 static int atmel_sha_cra_init(struct crypto_tfm *tfm)
1000 {
1001         return atmel_sha_cra_init_alg(tfm, NULL);
1002 }
1003
1004 static void atmel_sha_cra_exit(struct crypto_tfm *tfm)
1005 {
1006         struct atmel_sha_ctx *tctx = crypto_tfm_ctx(tfm);
1007
1008         crypto_free_shash(tctx->fallback);
1009         tctx->fallback = NULL;
1010 }
1011
1012 static struct ahash_alg sha_1_256_algs[] = {
1013 {
1014         .init           = atmel_sha_init,
1015         .update         = atmel_sha_update,
1016         .final          = atmel_sha_final,
1017         .finup          = atmel_sha_finup,
1018         .digest         = atmel_sha_digest,
1019         .halg = {
1020                 .digestsize     = SHA1_DIGEST_SIZE,
1021                 .base   = {
1022                         .cra_name               = "sha1",
1023                         .cra_driver_name        = "atmel-sha1",
1024                         .cra_priority           = 100,
1025                         .cra_flags              = CRYPTO_ALG_ASYNC |
1026                                                 CRYPTO_ALG_NEED_FALLBACK,
1027                         .cra_blocksize          = SHA1_BLOCK_SIZE,
1028                         .cra_ctxsize            = sizeof(struct atmel_sha_ctx),
1029                         .cra_alignmask          = 0,
1030                         .cra_module             = THIS_MODULE,
1031                         .cra_init               = atmel_sha_cra_init,
1032                         .cra_exit               = atmel_sha_cra_exit,
1033                 }
1034         }
1035 },
1036 {
1037         .init           = atmel_sha_init,
1038         .update         = atmel_sha_update,
1039         .final          = atmel_sha_final,
1040         .finup          = atmel_sha_finup,
1041         .digest         = atmel_sha_digest,
1042         .halg = {
1043                 .digestsize     = SHA256_DIGEST_SIZE,
1044                 .base   = {
1045                         .cra_name               = "sha256",
1046                         .cra_driver_name        = "atmel-sha256",
1047                         .cra_priority           = 100,
1048                         .cra_flags              = CRYPTO_ALG_ASYNC |
1049                                                 CRYPTO_ALG_NEED_FALLBACK,
1050                         .cra_blocksize          = SHA256_BLOCK_SIZE,
1051                         .cra_ctxsize            = sizeof(struct atmel_sha_ctx),
1052                         .cra_alignmask          = 0,
1053                         .cra_module             = THIS_MODULE,
1054                         .cra_init               = atmel_sha_cra_init,
1055                         .cra_exit               = atmel_sha_cra_exit,
1056                 }
1057         }
1058 },
1059 };
1060
1061 static struct ahash_alg sha_224_alg = {
1062         .init           = atmel_sha_init,
1063         .update         = atmel_sha_update,
1064         .final          = atmel_sha_final,
1065         .finup          = atmel_sha_finup,
1066         .digest         = atmel_sha_digest,
1067         .halg = {
1068                 .digestsize     = SHA224_DIGEST_SIZE,
1069                 .base   = {
1070                         .cra_name               = "sha224",
1071                         .cra_driver_name        = "atmel-sha224",
1072                         .cra_priority           = 100,
1073                         .cra_flags              = CRYPTO_ALG_ASYNC |
1074                                                 CRYPTO_ALG_NEED_FALLBACK,
1075                         .cra_blocksize          = SHA224_BLOCK_SIZE,
1076                         .cra_ctxsize            = sizeof(struct atmel_sha_ctx),
1077                         .cra_alignmask          = 0,
1078                         .cra_module             = THIS_MODULE,
1079                         .cra_init               = atmel_sha_cra_init,
1080                         .cra_exit               = atmel_sha_cra_exit,
1081                 }
1082         }
1083 };
1084
1085 static struct ahash_alg sha_384_512_algs[] = {
1086 {
1087         .init           = atmel_sha_init,
1088         .update         = atmel_sha_update,
1089         .final          = atmel_sha_final,
1090         .finup          = atmel_sha_finup,
1091         .digest         = atmel_sha_digest,
1092         .halg = {
1093                 .digestsize     = SHA384_DIGEST_SIZE,
1094                 .base   = {
1095                         .cra_name               = "sha384",
1096                         .cra_driver_name        = "atmel-sha384",
1097                         .cra_priority           = 100,
1098                         .cra_flags              = CRYPTO_ALG_ASYNC |
1099                                                 CRYPTO_ALG_NEED_FALLBACK,
1100                         .cra_blocksize          = SHA384_BLOCK_SIZE,
1101                         .cra_ctxsize            = sizeof(struct atmel_sha_ctx),
1102                         .cra_alignmask          = 0x3,
1103                         .cra_module             = THIS_MODULE,
1104                         .cra_init               = atmel_sha_cra_init,
1105                         .cra_exit               = atmel_sha_cra_exit,
1106                 }
1107         }
1108 },
1109 {
1110         .init           = atmel_sha_init,
1111         .update         = atmel_sha_update,
1112         .final          = atmel_sha_final,
1113         .finup          = atmel_sha_finup,
1114         .digest         = atmel_sha_digest,
1115         .halg = {
1116                 .digestsize     = SHA512_DIGEST_SIZE,
1117                 .base   = {
1118                         .cra_name               = "sha512",
1119                         .cra_driver_name        = "atmel-sha512",
1120                         .cra_priority           = 100,
1121                         .cra_flags              = CRYPTO_ALG_ASYNC |
1122                                                 CRYPTO_ALG_NEED_FALLBACK,
1123                         .cra_blocksize          = SHA512_BLOCK_SIZE,
1124                         .cra_ctxsize            = sizeof(struct atmel_sha_ctx),
1125                         .cra_alignmask          = 0x3,
1126                         .cra_module             = THIS_MODULE,
1127                         .cra_init               = atmel_sha_cra_init,
1128                         .cra_exit               = atmel_sha_cra_exit,
1129                 }
1130         }
1131 },
1132 };
1133
1134 static void atmel_sha_done_task(unsigned long data)
1135 {
1136         struct atmel_sha_dev *dd = (struct atmel_sha_dev *)data;
1137         int err = 0;
1138
1139         if (!(SHA_FLAGS_BUSY & dd->flags)) {
1140                 atmel_sha_handle_queue(dd, NULL);
1141                 return;
1142         }
1143
1144         if (SHA_FLAGS_CPU & dd->flags) {
1145                 if (SHA_FLAGS_OUTPUT_READY & dd->flags) {
1146                         dd->flags &= ~SHA_FLAGS_OUTPUT_READY;
1147                         goto finish;
1148                 }
1149         } else if (SHA_FLAGS_DMA_READY & dd->flags) {
1150                 if (SHA_FLAGS_DMA_ACTIVE & dd->flags) {
1151                         dd->flags &= ~SHA_FLAGS_DMA_ACTIVE;
1152                         atmel_sha_update_dma_stop(dd);
1153                         if (dd->err) {
1154                                 err = dd->err;
1155                                 goto finish;
1156                         }
1157                 }
1158                 if (SHA_FLAGS_OUTPUT_READY & dd->flags) {
1159                         /* hash or semi-hash ready */
1160                         dd->flags &= ~(SHA_FLAGS_DMA_READY |
1161                                                 SHA_FLAGS_OUTPUT_READY);
1162                         err = atmel_sha_update_dma_start(dd);
1163                         if (err != -EINPROGRESS)
1164                                 goto finish;
1165                 }
1166         }
1167         return;
1168
1169 finish:
1170         /* finish curent request */
1171         atmel_sha_finish_req(dd->req, err);
1172 }
1173
1174 static irqreturn_t atmel_sha_irq(int irq, void *dev_id)
1175 {
1176         struct atmel_sha_dev *sha_dd = dev_id;
1177         u32 reg;
1178
1179         reg = atmel_sha_read(sha_dd, SHA_ISR);
1180         if (reg & atmel_sha_read(sha_dd, SHA_IMR)) {
1181                 atmel_sha_write(sha_dd, SHA_IDR, reg);
1182                 if (SHA_FLAGS_BUSY & sha_dd->flags) {
1183                         sha_dd->flags |= SHA_FLAGS_OUTPUT_READY;
1184                         if (!(SHA_FLAGS_CPU & sha_dd->flags))
1185                                 sha_dd->flags |= SHA_FLAGS_DMA_READY;
1186                         tasklet_schedule(&sha_dd->done_task);
1187                 } else {
1188                         dev_warn(sha_dd->dev, "SHA interrupt when no active requests.\n");
1189                 }
1190                 return IRQ_HANDLED;
1191         }
1192
1193         return IRQ_NONE;
1194 }
1195
1196 static void atmel_sha_unregister_algs(struct atmel_sha_dev *dd)
1197 {
1198         int i;
1199
1200         for (i = 0; i < ARRAY_SIZE(sha_1_256_algs); i++)
1201                 crypto_unregister_ahash(&sha_1_256_algs[i]);
1202
1203         if (dd->caps.has_sha224)
1204                 crypto_unregister_ahash(&sha_224_alg);
1205
1206         if (dd->caps.has_sha_384_512) {
1207                 for (i = 0; i < ARRAY_SIZE(sha_384_512_algs); i++)
1208                         crypto_unregister_ahash(&sha_384_512_algs[i]);
1209         }
1210 }
1211
1212 static int atmel_sha_register_algs(struct atmel_sha_dev *dd)
1213 {
1214         int err, i, j;
1215
1216         for (i = 0; i < ARRAY_SIZE(sha_1_256_algs); i++) {
1217                 err = crypto_register_ahash(&sha_1_256_algs[i]);
1218                 if (err)
1219                         goto err_sha_1_256_algs;
1220         }
1221
1222         if (dd->caps.has_sha224) {
1223                 err = crypto_register_ahash(&sha_224_alg);
1224                 if (err)
1225                         goto err_sha_224_algs;
1226         }
1227
1228         if (dd->caps.has_sha_384_512) {
1229                 for (i = 0; i < ARRAY_SIZE(sha_384_512_algs); i++) {
1230                         err = crypto_register_ahash(&sha_384_512_algs[i]);
1231                         if (err)
1232                                 goto err_sha_384_512_algs;
1233                 }
1234         }
1235
1236         return 0;
1237
1238 err_sha_384_512_algs:
1239         for (j = 0; j < i; j++)
1240                 crypto_unregister_ahash(&sha_384_512_algs[j]);
1241         crypto_unregister_ahash(&sha_224_alg);
1242 err_sha_224_algs:
1243         i = ARRAY_SIZE(sha_1_256_algs);
1244 err_sha_1_256_algs:
1245         for (j = 0; j < i; j++)
1246                 crypto_unregister_ahash(&sha_1_256_algs[j]);
1247
1248         return err;
1249 }
1250
1251 static bool atmel_sha_filter(struct dma_chan *chan, void *slave)
1252 {
1253         struct at_dma_slave     *sl = slave;
1254
1255         if (sl && sl->dma_dev == chan->device->dev) {
1256                 chan->private = sl;
1257                 return true;
1258         } else {
1259                 return false;
1260         }
1261 }
1262
1263 static int atmel_sha_dma_init(struct atmel_sha_dev *dd,
1264                                 struct crypto_platform_data *pdata)
1265 {
1266         int err = -ENOMEM;
1267         dma_cap_mask_t mask_in;
1268
1269         /* Try to grab DMA channel */
1270         dma_cap_zero(mask_in);
1271         dma_cap_set(DMA_SLAVE, mask_in);
1272
1273         dd->dma_lch_in.chan = dma_request_slave_channel_compat(mask_in,
1274                         atmel_sha_filter, &pdata->dma_slave->rxdata, dd->dev, "tx");
1275         if (!dd->dma_lch_in.chan) {
1276                 dev_warn(dd->dev, "no DMA channel available\n");
1277                 return err;
1278         }
1279
1280         dd->dma_lch_in.dma_conf.direction = DMA_MEM_TO_DEV;
1281         dd->dma_lch_in.dma_conf.dst_addr = dd->phys_base +
1282                 SHA_REG_DIN(0);
1283         dd->dma_lch_in.dma_conf.src_maxburst = 1;
1284         dd->dma_lch_in.dma_conf.src_addr_width =
1285                 DMA_SLAVE_BUSWIDTH_4_BYTES;
1286         dd->dma_lch_in.dma_conf.dst_maxburst = 1;
1287         dd->dma_lch_in.dma_conf.dst_addr_width =
1288                 DMA_SLAVE_BUSWIDTH_4_BYTES;
1289         dd->dma_lch_in.dma_conf.device_fc = false;
1290
1291         return 0;
1292 }
1293
1294 static void atmel_sha_dma_cleanup(struct atmel_sha_dev *dd)
1295 {
1296         dma_release_channel(dd->dma_lch_in.chan);
1297 }
1298
1299 static void atmel_sha_get_cap(struct atmel_sha_dev *dd)
1300 {
1301
1302         dd->caps.has_dma = 0;
1303         dd->caps.has_dualbuff = 0;
1304         dd->caps.has_sha224 = 0;
1305         dd->caps.has_sha_384_512 = 0;
1306
1307         /* keep only major version number */
1308         switch (dd->hw_version & 0xff0) {
1309         case 0x410:
1310                 dd->caps.has_dma = 1;
1311                 dd->caps.has_dualbuff = 1;
1312                 dd->caps.has_sha224 = 1;
1313                 dd->caps.has_sha_384_512 = 1;
1314                 break;
1315         case 0x400:
1316                 dd->caps.has_dma = 1;
1317                 dd->caps.has_dualbuff = 1;
1318                 dd->caps.has_sha224 = 1;
1319                 break;
1320         case 0x320:
1321                 break;
1322         default:
1323                 dev_warn(dd->dev,
1324                                 "Unmanaged sha version, set minimum capabilities\n");
1325                 break;
1326         }
1327 }
1328
1329 #if defined(CONFIG_OF)
1330 static const struct of_device_id atmel_sha_dt_ids[] = {
1331         { .compatible = "atmel,at91sam9g46-sha" },
1332         { /* sentinel */ }
1333 };
1334
1335 MODULE_DEVICE_TABLE(of, atmel_sha_dt_ids);
1336
1337 static struct crypto_platform_data *atmel_sha_of_init(struct platform_device *pdev)
1338 {
1339         struct device_node *np = pdev->dev.of_node;
1340         struct crypto_platform_data *pdata;
1341
1342         if (!np) {
1343                 dev_err(&pdev->dev, "device node not found\n");
1344                 return ERR_PTR(-EINVAL);
1345         }
1346
1347         pdata = devm_kzalloc(&pdev->dev, sizeof(*pdata), GFP_KERNEL);
1348         if (!pdata) {
1349                 dev_err(&pdev->dev, "could not allocate memory for pdata\n");
1350                 return ERR_PTR(-ENOMEM);
1351         }
1352
1353         pdata->dma_slave = devm_kzalloc(&pdev->dev,
1354                                         sizeof(*(pdata->dma_slave)),
1355                                         GFP_KERNEL);
1356         if (!pdata->dma_slave) {
1357                 dev_err(&pdev->dev, "could not allocate memory for dma_slave\n");
1358                 devm_kfree(&pdev->dev, pdata);
1359                 return ERR_PTR(-ENOMEM);
1360         }
1361
1362         return pdata;
1363 }
1364 #else /* CONFIG_OF */
1365 static inline struct crypto_platform_data *atmel_sha_of_init(struct platform_device *dev)
1366 {
1367         return ERR_PTR(-EINVAL);
1368 }
1369 #endif
1370
1371 static int atmel_sha_probe(struct platform_device *pdev)
1372 {
1373         struct atmel_sha_dev *sha_dd;
1374         struct crypto_platform_data     *pdata;
1375         struct device *dev = &pdev->dev;
1376         struct resource *sha_res;
1377         unsigned long sha_phys_size;
1378         int err;
1379
1380         sha_dd = kzalloc(sizeof(struct atmel_sha_dev), GFP_KERNEL);
1381         if (sha_dd == NULL) {
1382                 dev_err(dev, "unable to alloc data struct.\n");
1383                 err = -ENOMEM;
1384                 goto sha_dd_err;
1385         }
1386
1387         sha_dd->dev = dev;
1388
1389         platform_set_drvdata(pdev, sha_dd);
1390
1391         INIT_LIST_HEAD(&sha_dd->list);
1392
1393         tasklet_init(&sha_dd->done_task, atmel_sha_done_task,
1394                                         (unsigned long)sha_dd);
1395
1396         crypto_init_queue(&sha_dd->queue, ATMEL_SHA_QUEUE_LENGTH);
1397
1398         sha_dd->irq = -1;
1399
1400         /* Get the base address */
1401         sha_res = platform_get_resource(pdev, IORESOURCE_MEM, 0);
1402         if (!sha_res) {
1403                 dev_err(dev, "no MEM resource info\n");
1404                 err = -ENODEV;
1405                 goto res_err;
1406         }
1407         sha_dd->phys_base = sha_res->start;
1408         sha_phys_size = resource_size(sha_res);
1409
1410         /* Get the IRQ */
1411         sha_dd->irq = platform_get_irq(pdev,  0);
1412         if (sha_dd->irq < 0) {
1413                 dev_err(dev, "no IRQ resource info\n");
1414                 err = sha_dd->irq;
1415                 goto res_err;
1416         }
1417
1418         err = request_irq(sha_dd->irq, atmel_sha_irq, IRQF_SHARED, "atmel-sha",
1419                                                 sha_dd);
1420         if (err) {
1421                 dev_err(dev, "unable to request sha irq.\n");
1422                 goto res_err;
1423         }
1424
1425         /* Initializing the clock */
1426         sha_dd->iclk = clk_get(&pdev->dev, "sha_clk");
1427         if (IS_ERR(sha_dd->iclk)) {
1428                 dev_err(dev, "clock intialization failed.\n");
1429                 err = PTR_ERR(sha_dd->iclk);
1430                 goto clk_err;
1431         }
1432
1433         sha_dd->io_base = ioremap(sha_dd->phys_base, sha_phys_size);
1434         if (!sha_dd->io_base) {
1435                 dev_err(dev, "can't ioremap\n");
1436                 err = -ENOMEM;
1437                 goto sha_io_err;
1438         }
1439
1440         atmel_sha_hw_version_init(sha_dd);
1441
1442         atmel_sha_get_cap(sha_dd);
1443
1444         if (sha_dd->caps.has_dma) {
1445                 pdata = pdev->dev.platform_data;
1446                 if (!pdata) {
1447                         pdata = atmel_sha_of_init(pdev);
1448                         if (IS_ERR(pdata)) {
1449                                 dev_err(&pdev->dev, "platform data not available\n");
1450                                 err = PTR_ERR(pdata);
1451                                 goto err_pdata;
1452                         }
1453                 }
1454                 if (!pdata->dma_slave) {
1455                         err = -ENXIO;
1456                         goto err_pdata;
1457                 }
1458                 err = atmel_sha_dma_init(sha_dd, pdata);
1459                 if (err)
1460                         goto err_sha_dma;
1461
1462                 dev_info(dev, "using %s for DMA transfers\n",
1463                                 dma_chan_name(sha_dd->dma_lch_in.chan));
1464         }
1465
1466         spin_lock(&atmel_sha.lock);
1467         list_add_tail(&sha_dd->list, &atmel_sha.dev_list);
1468         spin_unlock(&atmel_sha.lock);
1469
1470         err = atmel_sha_register_algs(sha_dd);
1471         if (err)
1472                 goto err_algs;
1473
1474         dev_info(dev, "Atmel SHA1/SHA256%s%s\n",
1475                         sha_dd->caps.has_sha224 ? "/SHA224" : "",
1476                         sha_dd->caps.has_sha_384_512 ? "/SHA384/SHA512" : "");
1477
1478         return 0;
1479
1480 err_algs:
1481         spin_lock(&atmel_sha.lock);
1482         list_del(&sha_dd->list);
1483         spin_unlock(&atmel_sha.lock);
1484         if (sha_dd->caps.has_dma)
1485                 atmel_sha_dma_cleanup(sha_dd);
1486 err_sha_dma:
1487 err_pdata:
1488         iounmap(sha_dd->io_base);
1489 sha_io_err:
1490         clk_put(sha_dd->iclk);
1491 clk_err:
1492         free_irq(sha_dd->irq, sha_dd);
1493 res_err:
1494         tasklet_kill(&sha_dd->done_task);
1495         kfree(sha_dd);
1496         sha_dd = NULL;
1497 sha_dd_err:
1498         dev_err(dev, "initialization failed.\n");
1499
1500         return err;
1501 }
1502
1503 static int atmel_sha_remove(struct platform_device *pdev)
1504 {
1505         static struct atmel_sha_dev *sha_dd;
1506
1507         sha_dd = platform_get_drvdata(pdev);
1508         if (!sha_dd)
1509                 return -ENODEV;
1510         spin_lock(&atmel_sha.lock);
1511         list_del(&sha_dd->list);
1512         spin_unlock(&atmel_sha.lock);
1513
1514         atmel_sha_unregister_algs(sha_dd);
1515
1516         tasklet_kill(&sha_dd->done_task);
1517
1518         if (sha_dd->caps.has_dma)
1519                 atmel_sha_dma_cleanup(sha_dd);
1520
1521         iounmap(sha_dd->io_base);
1522
1523         clk_put(sha_dd->iclk);
1524
1525         if (sha_dd->irq >= 0)
1526                 free_irq(sha_dd->irq, sha_dd);
1527
1528         kfree(sha_dd);
1529         sha_dd = NULL;
1530
1531         return 0;
1532 }
1533
1534 static struct platform_driver atmel_sha_driver = {
1535         .probe          = atmel_sha_probe,
1536         .remove         = atmel_sha_remove,
1537         .driver         = {
1538                 .name   = "atmel_sha",
1539                 .owner  = THIS_MODULE,
1540                 .of_match_table = of_match_ptr(atmel_sha_dt_ids),
1541         },
1542 };
1543
1544 module_platform_driver(atmel_sha_driver);
1545
1546 MODULE_DESCRIPTION("Atmel SHA (1/256/224/384/512) hw acceleration support.");
1547 MODULE_LICENSE("GPL v2");
1548 MODULE_AUTHOR("Nicolas Royer - Eukréa Electromatique");