1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * Cryptographic API for the NX-842 hardware compression. 4 * 5 * Copyright (C) IBM Corporation, 2011-2015 6 * 7 * Designer of the Power data compression engine: 8 * Bulent Abali <abali@us.ibm.com> 9 * 10 * Original Authors: Robert Jennings <rcj@linux.vnet.ibm.com> 11 * Seth Jennings <sjenning@linux.vnet.ibm.com> 12 * 13 * Rewrite: Dan Streetman <ddstreet@ieee.org> 14 * 15 * This is an interface to the NX-842 compression hardware in PowerPC 16 * processors. Most of the complexity of this drvier is due to the fact that 17 * the NX-842 compression hardware requires the input and output data buffers 18 * to be specifically aligned, to be a specific multiple in length, and within 19 * specific minimum and maximum lengths. Those restrictions, provided by the 20 * nx-842 driver via nx842_constraints, mean this driver must use bounce 21 * buffers and headers to correct misaligned in or out buffers, and to split 22 * input buffers that are too large. 23 * 24 * This driver will fall back to software decompression if the hardware 25 * decompression fails, so this driver's decompression should never fail as 26 * long as the provided compressed buffer is valid. Any compressed buffer 27 * created by this driver will have a header (except ones where the input 28 * perfectly matches the constraints); so users of this driver cannot simply 29 * pass a compressed buffer created by this driver over to the 842 software 30 * decompression library. Instead, users must use this driver to decompress; 31 * if the hardware fails or is unavailable, the compressed buffer will be 32 * parsed and the header removed, and the raw 842 buffer(s) passed to the 842 33 * software decompression library. 34 * 35 * This does not fall back to software compression, however, since the caller 36 * of this function is specifically requesting hardware compression; if the 37 * hardware compression fails, the caller can fall back to software 38 * compression, and the raw 842 compressed buffer that the software compressor 39 * creates can be passed to this driver for hardware decompression; any 40 * buffer without our specific header magic is assumed to be a raw 842 buffer 41 * and passed directly to the hardware. Note that the software compression 42 * library will produce a compressed buffer that is incompatible with the 43 * hardware decompressor if the original input buffer length is not a multiple 44 * of 8; if such a compressed buffer is passed to this driver for 45 * decompression, the hardware will reject it and this driver will then pass 46 * it over to the software library for decompression. 47 */ 48 49 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 50 51 #include <linux/vmalloc.h> 52 #include <linux/sw842.h> 53 #include <linux/spinlock.h> 54 55 #include "nx-842.h" 56 57 /* The first 5 bits of this magic are 0x1f, which is an invalid 842 5-bit 58 * template (see lib/842/842.h), so this magic number will never appear at 59 * the start of a raw 842 compressed buffer. That is important, as any buffer 60 * passed to us without this magic is assumed to be a raw 842 compressed 61 * buffer, and passed directly to the hardware to decompress. 62 */ 63 #define NX842_CRYPTO_MAGIC (0xf842) 64 #define NX842_CRYPTO_HEADER_SIZE(g) \ 65 (sizeof(struct nx842_crypto_header) + \ 66 sizeof(struct nx842_crypto_header_group) * (g)) 67 #define NX842_CRYPTO_HEADER_MAX_SIZE \ 68 NX842_CRYPTO_HEADER_SIZE(NX842_CRYPTO_GROUP_MAX) 69 70 /* bounce buffer size */ 71 #define BOUNCE_BUFFER_ORDER (2) 72 #define BOUNCE_BUFFER_SIZE \ 73 ((unsigned int)(PAGE_SIZE << BOUNCE_BUFFER_ORDER)) 74 75 /* try longer on comp because we can fallback to sw decomp if hw is busy */ 76 #define COMP_BUSY_TIMEOUT (250) /* ms */ 77 #define DECOMP_BUSY_TIMEOUT (50) /* ms */ 78 79 struct nx842_crypto_param { 80 u8 *in; 81 unsigned int iremain; 82 u8 *out; 83 unsigned int oremain; 84 unsigned int ototal; 85 }; 86 87 static int update_param(struct nx842_crypto_param *p, 88 unsigned int slen, unsigned int dlen) 89 { 90 if (p->iremain < slen) 91 return -EOVERFLOW; 92 if (p->oremain < dlen) 93 return -ENOSPC; 94 95 p->in += slen; 96 p->iremain -= slen; 97 p->out += dlen; 98 p->oremain -= dlen; 99 p->ototal += dlen; 100 101 return 0; 102 } 103 104 void *nx842_crypto_alloc_ctx(struct nx842_driver *driver) 105 { 106 struct nx842_crypto_ctx *ctx; 107 108 ctx = kzalloc_obj(*ctx); 109 if (!ctx) 110 return ERR_PTR(-ENOMEM); 111 112 spin_lock_init(&ctx->lock); 113 ctx->driver = driver; 114 ctx->wmem = kmalloc(driver->workmem_size, GFP_KERNEL); 115 ctx->sbounce = (u8 *)__get_free_pages(GFP_KERNEL, BOUNCE_BUFFER_ORDER); 116 ctx->dbounce = (u8 *)__get_free_pages(GFP_KERNEL, BOUNCE_BUFFER_ORDER); 117 if (!ctx->wmem || !ctx->sbounce || !ctx->dbounce) { 118 nx842_crypto_free_ctx(ctx); 119 return ERR_PTR(-ENOMEM); 120 } 121 122 return ctx; 123 } 124 EXPORT_SYMBOL_GPL(nx842_crypto_alloc_ctx); 125 126 void nx842_crypto_free_ctx(void *p) 127 { 128 struct nx842_crypto_ctx *ctx = p; 129 130 kfree(ctx->wmem); 131 free_pages((unsigned long)ctx->sbounce, BOUNCE_BUFFER_ORDER); 132 free_pages((unsigned long)ctx->dbounce, BOUNCE_BUFFER_ORDER); 133 kfree(ctx); 134 } 135 EXPORT_SYMBOL_GPL(nx842_crypto_free_ctx); 136 137 static void check_constraints(struct nx842_constraints *c) 138 { 139 /* limit maximum, to always have enough bounce buffer to decompress */ 140 if (c->maximum > BOUNCE_BUFFER_SIZE) 141 c->maximum = BOUNCE_BUFFER_SIZE; 142 } 143 144 static int nx842_crypto_add_header(struct nx842_crypto_header *hdr, u8 *buf) 145 { 146 int s = NX842_CRYPTO_HEADER_SIZE(hdr->groups); 147 148 /* compress should have added space for header */ 149 if (s > be16_to_cpu(hdr->group[0].padding)) { 150 pr_err("Internal error: no space for header\n"); 151 return -EINVAL; 152 } 153 154 memcpy(buf, hdr, s); 155 156 print_hex_dump_debug("header ", DUMP_PREFIX_OFFSET, 16, 1, buf, s, 0); 157 158 return 0; 159 } 160 161 static int compress(struct nx842_crypto_ctx *ctx, 162 struct nx842_crypto_param *p, 163 struct nx842_crypto_header_group *g, 164 struct nx842_constraints *c, 165 u16 *ignore, 166 unsigned int hdrsize) 167 { 168 unsigned int slen = p->iremain, dlen = p->oremain, tmplen; 169 unsigned int adj_slen = slen; 170 u8 *src = p->in, *dst = p->out; 171 int ret, dskip = 0; 172 ktime_t timeout; 173 174 if (p->iremain == 0) 175 return -EOVERFLOW; 176 177 if (p->oremain == 0 || hdrsize + c->minimum > dlen) 178 return -ENOSPC; 179 180 if (slen % c->multiple) 181 adj_slen = round_up(slen, c->multiple); 182 if (slen < c->minimum) 183 adj_slen = c->minimum; 184 if (slen > c->maximum) 185 adj_slen = slen = c->maximum; 186 if (adj_slen > slen || (u64)src % c->alignment) { 187 adj_slen = min(adj_slen, BOUNCE_BUFFER_SIZE); 188 slen = min(slen, BOUNCE_BUFFER_SIZE); 189 if (adj_slen > slen) 190 memset(ctx->sbounce + slen, 0, adj_slen - slen); 191 memcpy(ctx->sbounce, src, slen); 192 src = ctx->sbounce; 193 slen = adj_slen; 194 pr_debug("using comp sbounce buffer, len %x\n", slen); 195 } 196 197 dst += hdrsize; 198 dlen -= hdrsize; 199 200 if ((u64)dst % c->alignment) { 201 dskip = (int)(PTR_ALIGN(dst, c->alignment) - dst); 202 dst += dskip; 203 dlen -= dskip; 204 } 205 if (dlen % c->multiple) 206 dlen = round_down(dlen, c->multiple); 207 if (dlen < c->minimum) { 208 nospc: 209 dst = ctx->dbounce; 210 dlen = min(p->oremain, BOUNCE_BUFFER_SIZE); 211 dlen = round_down(dlen, c->multiple); 212 dskip = 0; 213 pr_debug("using comp dbounce buffer, len %x\n", dlen); 214 } 215 if (dlen > c->maximum) 216 dlen = c->maximum; 217 218 tmplen = dlen; 219 timeout = ktime_add_ms(ktime_get(), COMP_BUSY_TIMEOUT); 220 do { 221 dlen = tmplen; /* reset dlen, if we're retrying */ 222 ret = ctx->driver->compress(src, slen, dst, &dlen, ctx->wmem); 223 /* possibly we should reduce the slen here, instead of 224 * retrying with the dbounce buffer? 225 */ 226 if (ret == -ENOSPC && dst != ctx->dbounce) 227 goto nospc; 228 } while (ret == -EBUSY && ktime_before(ktime_get(), timeout)); 229 if (ret) 230 return ret; 231 232 dskip += hdrsize; 233 234 if (dst == ctx->dbounce) 235 memcpy(p->out + dskip, dst, dlen); 236 237 g->padding = cpu_to_be16(dskip); 238 g->compressed_length = cpu_to_be32(dlen); 239 g->uncompressed_length = cpu_to_be32(slen); 240 241 if (p->iremain < slen) { 242 *ignore = slen - p->iremain; 243 slen = p->iremain; 244 } 245 246 pr_debug("compress slen %x ignore %x dlen %x padding %x\n", 247 slen, *ignore, dlen, dskip); 248 249 return update_param(p, slen, dskip + dlen); 250 } 251 252 int nx842_crypto_compress(struct crypto_scomp *tfm, 253 const u8 *src, unsigned int slen, 254 u8 *dst, unsigned int *dlen, void *pctx) 255 { 256 struct nx842_crypto_ctx *ctx = pctx; 257 struct nx842_crypto_header *hdr = 258 container_of(&ctx->header, 259 struct nx842_crypto_header, hdr); 260 struct nx842_crypto_param p; 261 struct nx842_constraints c = *ctx->driver->constraints; 262 unsigned int groups, hdrsize, h; 263 int ret, n; 264 bool add_header; 265 u16 ignore = 0; 266 267 check_constraints(&c); 268 269 p.in = (u8 *)src; 270 p.iremain = slen; 271 p.out = dst; 272 p.oremain = *dlen; 273 p.ototal = 0; 274 275 *dlen = 0; 276 277 groups = min_t(unsigned int, NX842_CRYPTO_GROUP_MAX, 278 DIV_ROUND_UP(p.iremain, c.maximum)); 279 hdrsize = NX842_CRYPTO_HEADER_SIZE(groups); 280 281 spin_lock_bh(&ctx->lock); 282 283 /* skip adding header if the buffers meet all constraints */ 284 add_header = (p.iremain % c.multiple || 285 p.iremain < c.minimum || 286 p.iremain > c.maximum || 287 (u64)p.in % c.alignment || 288 p.oremain % c.multiple || 289 p.oremain < c.minimum || 290 p.oremain > c.maximum || 291 (u64)p.out % c.alignment); 292 293 hdr->magic = cpu_to_be16(NX842_CRYPTO_MAGIC); 294 hdr->groups = 0; 295 hdr->ignore = 0; 296 297 while (p.iremain > 0) { 298 n = hdr->groups++; 299 ret = -ENOSPC; 300 if (hdr->groups > NX842_CRYPTO_GROUP_MAX) 301 goto unlock; 302 303 /* header goes before first group */ 304 h = !n && add_header ? hdrsize : 0; 305 306 if (ignore) 307 pr_warn("internal error, ignore is set %x\n", ignore); 308 309 ret = compress(ctx, &p, &hdr->group[n], &c, &ignore, h); 310 if (ret) 311 goto unlock; 312 } 313 314 if (!add_header && hdr->groups > 1) { 315 pr_err("Internal error: No header but multiple groups\n"); 316 ret = -EINVAL; 317 goto unlock; 318 } 319 320 /* ignore indicates the input stream needed to be padded */ 321 hdr->ignore = cpu_to_be16(ignore); 322 if (ignore) 323 pr_debug("marked %d bytes as ignore\n", ignore); 324 325 if (add_header) 326 ret = nx842_crypto_add_header(hdr, dst); 327 if (ret) 328 goto unlock; 329 330 *dlen = p.ototal; 331 332 pr_debug("compress total slen %x dlen %x\n", slen, *dlen); 333 334 unlock: 335 spin_unlock_bh(&ctx->lock); 336 return ret; 337 } 338 EXPORT_SYMBOL_GPL(nx842_crypto_compress); 339 340 static int decompress(struct nx842_crypto_ctx *ctx, 341 struct nx842_crypto_param *p, 342 struct nx842_crypto_header_group *g, 343 struct nx842_constraints *c, 344 u16 ignore) 345 { 346 unsigned int slen = be32_to_cpu(g->compressed_length); 347 unsigned int required_len = be32_to_cpu(g->uncompressed_length); 348 unsigned int dlen = p->oremain, tmplen; 349 unsigned int adj_slen = slen; 350 u8 *src = p->in, *dst = p->out; 351 u16 padding = be16_to_cpu(g->padding); 352 int ret, spadding = 0; 353 ktime_t timeout; 354 355 if (!slen || !required_len) 356 return -EINVAL; 357 358 if (p->iremain <= 0 || padding + slen > p->iremain) 359 return -EOVERFLOW; 360 361 if (p->oremain <= 0 || required_len - ignore > p->oremain) 362 return -ENOSPC; 363 364 src += padding; 365 366 if (slen % c->multiple) 367 adj_slen = round_up(slen, c->multiple); 368 if (slen < c->minimum) 369 adj_slen = c->minimum; 370 if (slen > c->maximum) 371 goto usesw; 372 if (slen < adj_slen || (u64)src % c->alignment) { 373 /* we can append padding bytes because the 842 format defines 374 * an "end" template (see lib/842/842_decompress.c) and will 375 * ignore any bytes following it. 376 */ 377 if (slen < adj_slen) 378 memset(ctx->sbounce + slen, 0, adj_slen - slen); 379 memcpy(ctx->sbounce, src, slen); 380 src = ctx->sbounce; 381 spadding = adj_slen - slen; 382 slen = adj_slen; 383 pr_debug("using decomp sbounce buffer, len %x\n", slen); 384 } 385 386 if (dlen % c->multiple) 387 dlen = round_down(dlen, c->multiple); 388 if (dlen < required_len || (u64)dst % c->alignment) { 389 dst = ctx->dbounce; 390 dlen = min(required_len, BOUNCE_BUFFER_SIZE); 391 pr_debug("using decomp dbounce buffer, len %x\n", dlen); 392 } 393 if (dlen < c->minimum) 394 goto usesw; 395 if (dlen > c->maximum) 396 dlen = c->maximum; 397 398 tmplen = dlen; 399 timeout = ktime_add_ms(ktime_get(), DECOMP_BUSY_TIMEOUT); 400 do { 401 dlen = tmplen; /* reset dlen, if we're retrying */ 402 ret = ctx->driver->decompress(src, slen, dst, &dlen, ctx->wmem); 403 } while (ret == -EBUSY && ktime_before(ktime_get(), timeout)); 404 if (ret) { 405 usesw: 406 /* reset everything, sw doesn't have constraints */ 407 src = p->in + padding; 408 slen = be32_to_cpu(g->compressed_length); 409 spadding = 0; 410 dst = p->out; 411 dlen = p->oremain; 412 if (dlen < required_len) { /* have ignore bytes */ 413 dst = ctx->dbounce; 414 dlen = BOUNCE_BUFFER_SIZE; 415 } 416 pr_info_ratelimited("using software 842 decompression\n"); 417 ret = sw842_decompress(src, slen, dst, &dlen); 418 } 419 if (ret) 420 return ret; 421 422 slen -= spadding; 423 424 dlen -= ignore; 425 if (ignore) 426 pr_debug("ignoring last %x bytes\n", ignore); 427 428 if (dst == ctx->dbounce) 429 memcpy(p->out, dst, dlen); 430 431 pr_debug("decompress slen %x padding %x dlen %x ignore %x\n", 432 slen, padding, dlen, ignore); 433 434 return update_param(p, slen + padding, dlen); 435 } 436 437 int nx842_crypto_decompress(struct crypto_scomp *tfm, 438 const u8 *src, unsigned int slen, 439 u8 *dst, unsigned int *dlen, void *pctx) 440 { 441 struct nx842_crypto_ctx *ctx = pctx; 442 struct nx842_crypto_header *hdr; 443 struct nx842_crypto_param p; 444 struct nx842_constraints c = *ctx->driver->constraints; 445 int n, ret, hdr_len; 446 u16 ignore = 0; 447 448 check_constraints(&c); 449 450 p.in = (u8 *)src; 451 p.iremain = slen; 452 p.out = dst; 453 p.oremain = *dlen; 454 p.ototal = 0; 455 456 *dlen = 0; 457 458 hdr = (struct nx842_crypto_header *)src; 459 460 spin_lock_bh(&ctx->lock); 461 462 /* If it doesn't start with our header magic number, assume it's a raw 463 * 842 compressed buffer and pass it directly to the hardware driver 464 */ 465 if (be16_to_cpu(hdr->magic) != NX842_CRYPTO_MAGIC) { 466 struct nx842_crypto_header_group g = { 467 .padding = 0, 468 .compressed_length = cpu_to_be32(p.iremain), 469 .uncompressed_length = cpu_to_be32(p.oremain), 470 }; 471 472 ret = decompress(ctx, &p, &g, &c, 0); 473 if (ret) 474 goto unlock; 475 476 goto success; 477 } 478 479 if (!hdr->groups) { 480 pr_err("header has no groups\n"); 481 ret = -EINVAL; 482 goto unlock; 483 } 484 if (hdr->groups > NX842_CRYPTO_GROUP_MAX) { 485 pr_err("header has too many groups %x, max %x\n", 486 hdr->groups, NX842_CRYPTO_GROUP_MAX); 487 ret = -EINVAL; 488 goto unlock; 489 } 490 491 hdr_len = NX842_CRYPTO_HEADER_SIZE(hdr->groups); 492 if (hdr_len > slen) { 493 ret = -EOVERFLOW; 494 goto unlock; 495 } 496 497 memcpy(&ctx->header, src, hdr_len); 498 hdr = container_of(&ctx->header, struct nx842_crypto_header, hdr); 499 500 for (n = 0; n < hdr->groups; n++) { 501 /* ignore applies to last group */ 502 if (n + 1 == hdr->groups) 503 ignore = be16_to_cpu(hdr->ignore); 504 505 ret = decompress(ctx, &p, &hdr->group[n], &c, ignore); 506 if (ret) 507 goto unlock; 508 } 509 510 success: 511 *dlen = p.ototal; 512 513 pr_debug("decompress total slen %x dlen %x\n", slen, *dlen); 514 515 ret = 0; 516 517 unlock: 518 spin_unlock_bh(&ctx->lock); 519 520 return ret; 521 } 522 EXPORT_SYMBOL_GPL(nx842_crypto_decompress); 523 524 MODULE_LICENSE("GPL"); 525 MODULE_DESCRIPTION("IBM PowerPC Nest (NX) 842 Hardware Compression Driver"); 526 MODULE_AUTHOR("Dan Streetman <ddstreet@ieee.org>"); 527