drivers/crypto/nx/nx-842.c

   1 /*
   2  * Driver for IBM Power 842 compression accelerator
   3  *
   4  * This program is free software; you can redistribute it and/or modify
   5  * it under the terms of the GNU General Public License as published by
   6  * the Free Software Foundation; either version 2 of the License, or
   7  * (at your option) any later version.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write to the Free Software
  16  * Foundation, 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301, USA.
  17  *
  18  * Copyright (C) IBM Corporation, 2012
  19  *
  20  * Authors: Robert Jennings <rcj@linux.vnet.ibm.com>
  21  *          Seth Jennings <sjenning@linux.vnet.ibm.com>
  22  */
  23
  24 #include <linux/kernel.h>
  25 #include <linux/module.h>
  26 #include <linux/nx842.h>
  27 #include <linux/of.h>
  28 #include <linux/slab.h>
  29
  30 #include <asm/page.h>
  31 #include <asm/vio.h>
  32
  33 #include "nx_csbcpb.h" /* struct nx_csbcpb */
  34
  35 #define MODULE_NAME "nx-compress"
  36 MODULE_LICENSE("GPL");
  37 MODULE_AUTHOR("Robert Jennings <rcj@linux.vnet.ibm.com>");
  38 MODULE_DESCRIPTION("842 H/W Compression driver for IBM Power processors");
  39
  40 #define SHIFT_4K 12
  41 #define SHIFT_64K 16
  42 #define SIZE_4K (1UL << SHIFT_4K)
  43 #define SIZE_64K (1UL << SHIFT_64K)
  44
  45 /* IO buffer must be 128 byte aligned */
  46 #define IO_BUFFER_ALIGN 128
  47
  48 struct nx842_header {
  49         int blocks_nr; /* number of compressed blocks */
  50         int offset; /* offset of the first block (from beginning of header) */
  51         int sizes[0]; /* size of compressed blocks */
  52 };
  53
  54 static inline int nx842_header_size(const struct nx842_header *hdr)
  55 {
  56         return sizeof(struct nx842_header) +
  57                         hdr->blocks_nr * sizeof(hdr->sizes[0]);
  58 }
  59
  60 /* Macros for fields within nx_csbcpb */
  61 /* Check the valid bit within the csbcpb valid field */
  62 #define NX842_CSBCBP_VALID_CHK(x) (x & BIT_MASK(7))
  63
  64 /* CE macros operate on the completion_extension field bits in the csbcpb.
  65  * CE0 0=full completion, 1=partial completion
  66  * CE1 0=CE0 indicates completion, 1=termination (output may be modified)
  67  * CE2 0=processed_bytes is source bytes, 1=processed_bytes is target bytes */
  68 #define NX842_CSBCPB_CE0(x)     (x & BIT_MASK(7))
  69 #define NX842_CSBCPB_CE1(x)     (x & BIT_MASK(6))
  70 #define NX842_CSBCPB_CE2(x)     (x & BIT_MASK(5))
  71
  72 /* The NX unit accepts data only on 4K page boundaries */
  73 #define NX842_HW_PAGE_SHIFT     SHIFT_4K
  74 #define NX842_HW_PAGE_SIZE      (ASM_CONST(1) << NX842_HW_PAGE_SHIFT)
  75 #define NX842_HW_PAGE_MASK      (~(NX842_HW_PAGE_SIZE-1))
  76
  77 enum nx842_status {
  78         UNAVAILABLE,
  79         AVAILABLE
  80 };
  81
  82 struct ibm_nx842_counters {
  83         atomic64_t comp_complete;
  84         atomic64_t comp_failed;
  85         atomic64_t decomp_complete;
  86         atomic64_t decomp_failed;
  87         atomic64_t swdecomp;
  88         atomic64_t comp_times[32];
  89         atomic64_t decomp_times[32];
  90 };
  91
  92 static struct nx842_devdata {
  93         struct vio_dev *vdev;
  94         struct device *dev;
  95         struct ibm_nx842_counters *counters;
  96         unsigned int max_sg_len;
  97         unsigned int max_sync_size;
  98         unsigned int max_sync_sg;
  99         enum nx842_status status;
 100 } __rcu *devdata;
 101 static DEFINE_SPINLOCK(devdata_mutex);
 102
 103 #define NX842_COUNTER_INC(_x) \
 104 static inline void nx842_inc_##_x( \
 105         const struct nx842_devdata *dev) { \
 106         if (dev) \
 107                 atomic64_inc(&dev->counters->_x); \
 108 }
 109 NX842_COUNTER_INC(comp_complete);
 110 NX842_COUNTER_INC(comp_failed);
 111 NX842_COUNTER_INC(decomp_complete);
 112 NX842_COUNTER_INC(decomp_failed);
 113 NX842_COUNTER_INC(swdecomp);
 114
 115 #define NX842_HIST_SLOTS 16
 116
 117 static void ibm_nx842_incr_hist(atomic64_t *times, unsigned int time)
 118 {
 119         int bucket = fls(time);
 120
 121         if (bucket)
 122                 bucket = min((NX842_HIST_SLOTS - 1), bucket - 1);
 123
 124         atomic64_inc(&times[bucket]);
 125 }
 126
 127 /* NX unit operation flags */
 128 #define NX842_OP_COMPRESS       0x0
 129 #define NX842_OP_CRC            0x1
 130 #define NX842_OP_DECOMPRESS     0x2
 131 #define NX842_OP_COMPRESS_CRC   (NX842_OP_COMPRESS | NX842_OP_CRC)
 132 #define NX842_OP_DECOMPRESS_CRC (NX842_OP_DECOMPRESS | NX842_OP_CRC)
 133 #define NX842_OP_ASYNC          (1<<23)
 134 #define NX842_OP_NOTIFY         (1<<22)
 135 #define NX842_OP_NOTIFY_INT(x)  ((x & 0xff)<<8)
 136
 137 static unsigned long nx842_get_desired_dma(struct vio_dev *viodev)
 138 {
 139         /* No use of DMA mappings within the driver. */
 140         return 0;
 141 }
 142
 143 struct nx842_slentry {
 144         unsigned long ptr; /* Real address (use __pa()) */
 145         unsigned long len;
 146 };
 147
 148 /* pHyp scatterlist entry */
 149 struct nx842_scatterlist {
 150         int entry_nr; /* number of slentries */
 151         struct nx842_slentry *entries; /* ptr to array of slentries */
 152 };
 153
 154 /* Does not include sizeof(entry_nr) in the size */
 155 static inline unsigned long nx842_get_scatterlist_size(
 156                                 struct nx842_scatterlist *sl)
 157 {
 158         return sl->entry_nr * sizeof(struct nx842_slentry);
 159 }
 160
 161 static int nx842_build_scatterlist(unsigned long buf, int len,
 162                         struct nx842_scatterlist *sl)
 163 {
 164         unsigned long nextpage;
 165         struct nx842_slentry *entry;
 166
 167         sl->entry_nr = 0;
 168
 169         entry = sl->entries;
 170         while (len) {
 171                 entry->ptr = __pa(buf);
 172                 nextpage = ALIGN(buf + 1, NX842_HW_PAGE_SIZE);
 173                 if (nextpage < buf + len) {
 174                         /* we aren't at the end yet */
 175                         if (IS_ALIGNED(buf, NX842_HW_PAGE_SIZE))
 176                                 /* we are in the middle (or beginning) */
 177                                 entry->len = NX842_HW_PAGE_SIZE;
 178                         else
 179                                 /* we are at the beginning */
 180                                 entry->len = nextpage - buf;
 181                 } else {
 182                         /* at the end */
 183                         entry->len = len;
 184                 }
 185
 186                 len -= entry->len;
 187                 buf += entry->len;
 188                 sl->entry_nr++;
 189                 entry++;
 190         }
 191
 192         return 0;
 193 }
 194
 195 /*
 196  * Working memory for software decompression
 197  */
 198 struct sw842_fifo {
 199         union {
 200                 char f8[256][8];
 201                 char f4[512][4];
 202         };
 203         char f2[256][2];
 204         unsigned char f84_full;
 205         unsigned char f2_full;
 206         unsigned char f8_count;
 207         unsigned char f2_count;
 208         unsigned int f4_count;
 209 };
 210
 211 /*
 212  * Working memory for crypto API
 213  */
 214 struct nx842_workmem {
 215         char bounce[PAGE_SIZE]; /* bounce buffer for decompression input */
 216         union {
 217                 /* hardware working memory */
 218                 struct {
 219                         /* scatterlist */
 220                         char slin[SIZE_4K];
 221                         char slout[SIZE_4K];
 222                         /* coprocessor status/parameter block */
 223                         struct nx_csbcpb csbcpb;
 224                 };
 225                 /* software working memory */
 226                 struct sw842_fifo swfifo; /* software decompression fifo */
 227         };
 228 };
 229
 230 int nx842_get_workmem_size(void)
 231 {
 232         return sizeof(struct nx842_workmem) + NX842_HW_PAGE_SIZE;
 233 }
 234 EXPORT_SYMBOL_GPL(nx842_get_workmem_size);
 235
 236 int nx842_get_workmem_size_aligned(void)
 237 {
 238         return sizeof(struct nx842_workmem);
 239 }
 240 EXPORT_SYMBOL_GPL(nx842_get_workmem_size_aligned);
 241
 242 static int nx842_validate_result(struct device *dev,
 243         struct cop_status_block *csb)
 244 {
 245         /* The csb must be valid after returning from vio_h_cop_sync */
 246         if (!NX842_CSBCBP_VALID_CHK(csb->valid)) {
 247                 dev_err(dev, "%s: cspcbp not valid upon completion.\n",
 248                                 __func__);
 249                 dev_dbg(dev, "valid:0x%02x cs:0x%02x cc:0x%02x ce:0x%02x\n",
 250                                 csb->valid,
 251                                 csb->crb_seq_number,
 252                                 csb->completion_code,
 253                                 csb->completion_extension);
 254                 dev_dbg(dev, "processed_bytes:%d address:0x%016lx\n",
 255                                 csb->processed_byte_count,
 256                                 (unsigned long)csb->address);
 257                 return -EIO;
 258         }
 259
 260         /* Check return values from the hardware in the CSB */
 261         switch (csb->completion_code) {
 262         case 0: /* Completed without error */
 263                 break;
 264         case 64: /* Target bytes > Source bytes during compression */
 265         case 13: /* Output buffer too small */
 266                 dev_dbg(dev, "%s: Compression output larger than input\n",
 267                                         __func__);
 268                 return -ENOSPC;
 269         case 66: /* Input data contains an illegal template field */
 270         case 67: /* Template indicates data past the end of the input stream */
 271                 dev_dbg(dev, "%s: Bad data for decompression (code:%d)\n",
 272                                         __func__, csb->completion_code);
 273                 return -EINVAL;
 274         default:
 275                 dev_dbg(dev, "%s: Unspecified error (code:%d)\n",
 276                                         __func__, csb->completion_code);
 277                 return -EIO;
 278         }
 279
 280         /* Hardware sanity check */
 281         if (!NX842_CSBCPB_CE2(csb->completion_extension)) {
 282                 dev_err(dev, "%s: No error returned by hardware, but "
 283                                 "data returned is unusable, contact support.\n"
 284                                 "(Additional info: csbcbp->processed bytes "
 285                                 "does not specify processed bytes for the "
 286                                 "target buffer.)\n", __func__);
 287                 return -EIO;
 288         }
 289
 290         return 0;
 291 }
 292
 293 /**
 294  * nx842_compress - Compress data using the 842 algorithm
 295  *
 296  * Compression provide by the NX842 coprocessor on IBM Power systems.
 297  * The input buffer is compressed and the result is stored in the
 298  * provided output buffer.
 299  *
 300  * Upon return from this function @outlen contains the length of the
 301  * compressed data.  If there is an error then @outlen will be 0 and an
 302  * error will be specified by the return code from this function.
 303  *
 304  * @in: Pointer to input buffer, must be page aligned
 305  * @inlen: Length of input buffer, must be PAGE_SIZE
 306  * @out: Pointer to output buffer
 307  * @outlen: Length of output buffer
 308  * @wrkmem: ptr to buffer for working memory, size determined by
 309  *          nx842_get_workmem_size()
 310  *
 311  * Returns:
 312  *   0          Success, output of length @outlen stored in the buffer at @out
 313  *   -ENOMEM    Unable to allocate internal buffers
 314  *   -ENOSPC    Output buffer is to small
 315  *   -EMSGSIZE  XXX Difficult to describe this limitation
 316  *   -EIO       Internal error
 317  *   -ENODEV    Hardware unavailable
 318  */
 319 int nx842_compress(const unsigned char *in, unsigned int inlen,
 320                        unsigned char *out, unsigned int *outlen, void *wmem)
 321 {
 322         struct nx842_header *hdr;
 323         struct nx842_devdata *local_devdata;
 324         struct device *dev = NULL;
 325         struct nx842_workmem *workmem;
 326         struct nx842_scatterlist slin, slout;
 327         struct nx_csbcpb *csbcpb;
 328         int ret = 0, max_sync_size, i, bytesleft, size, hdrsize;
 329         unsigned long inbuf, outbuf, padding;
 330         struct vio_pfo_op op = {
 331                 .done = NULL,
 332                 .handle = 0,
 333                 .timeout = 0,
 334         };
 335         unsigned long start_time = get_tb();
 336
 337         /*
 338          * Make sure input buffer is 64k page aligned.  This is assumed since
 339          * this driver is designed for page compression only (for now).  This
 340          * is very nice since we can now use direct DDE(s) for the input and
 341          * the alignment is guaranteed.
 342         */
 343         inbuf = (unsigned long)in;
 344         if (!IS_ALIGNED(inbuf, PAGE_SIZE) || inlen != PAGE_SIZE)
 345                 return -EINVAL;
 346
 347         rcu_read_lock();
 348         local_devdata = rcu_dereference(devdata);
 349         if (!local_devdata || !local_devdata->dev) {
 350                 rcu_read_unlock();
 351                 return -ENODEV;
 352         }
 353         max_sync_size = local_devdata->max_sync_size;
 354         dev = local_devdata->dev;
 355
 356         /* Create the header */
 357         hdr = (struct nx842_header *)out;
 358         hdr->blocks_nr = PAGE_SIZE / max_sync_size;
 359         hdrsize = nx842_header_size(hdr);
 360         outbuf = (unsigned long)out + hdrsize;
 361         bytesleft = *outlen - hdrsize;
 362
 363         /* Init scatterlist */
 364         workmem = (struct nx842_workmem *)ALIGN((unsigned long)wmem,
 365                 NX842_HW_PAGE_SIZE);
 366         slin.entries = (struct nx842_slentry *)workmem->slin;
 367         slout.entries = (struct nx842_slentry *)workmem->slout;
 368
 369         /* Init operation */
 370         op.flags = NX842_OP_COMPRESS;
 371         csbcpb = &workmem->csbcpb;
 372         memset(csbcpb, 0, sizeof(*csbcpb));
 373         op.csbcpb = __pa(csbcpb);
 374         op.out = __pa(slout.entries);
 375
 376         for (i = 0; i < hdr->blocks_nr; i++) {
 377                 /*
 378                  * Aligning the output blocks to 128 bytes does waste space,
 379                  * but it prevents the need for bounce buffers and memory
 380                  * copies.  It also simplifies the code a lot.  In the worst
 381                  * case (64k page, 4k max_sync_size), you lose up to
 382                  * (128*16)/64k = ~3% the compression factor. For 64k
 383                  * max_sync_size, the loss would be at most 128/64k = ~0.2%.
 384                  */
 385                 padding = ALIGN(outbuf, IO_BUFFER_ALIGN) - outbuf;
 386                 outbuf += padding;
 387                 bytesleft -= padding;
 388                 if (i == 0)
 389                         /* save offset into first block in header */
 390                         hdr->offset = padding + hdrsize;
 391
 392                 if (bytesleft <= 0) {
 393                         ret = -ENOSPC;
 394                         goto unlock;
 395                 }
 396
 397                 /*
 398                  * NOTE: If the default max_sync_size is changed from 4k
 399                  * to 64k, remove the "likely" case below, since a
 400                  * scatterlist will always be needed.
 401                  */
 402                 if (likely(max_sync_size == NX842_HW_PAGE_SIZE)) {
 403                         /* Create direct DDE */
 404                         op.in = __pa(inbuf);
 405                         op.inlen = max_sync_size;
 406
 407                 } else {
 408                         /* Create indirect DDE (scatterlist) */
 409                         nx842_build_scatterlist(inbuf, max_sync_size, &slin);
 410                         op.in = __pa(slin.entries);
 411                         op.inlen = -nx842_get_scatterlist_size(&slin);
 412                 }
 413
 414                 /*
 415                  * If max_sync_size != NX842_HW_PAGE_SIZE, an indirect
 416                  * DDE is required for the outbuf.
 417                  * If max_sync_size == NX842_HW_PAGE_SIZE, outbuf must
 418                  * also be page aligned (1 in 128/4k=32 chance) in order
 419                  * to use a direct DDE.
 420                  * This is unlikely, just use an indirect DDE always.
 421                  */
 422                 nx842_build_scatterlist(outbuf,
 423                         min(bytesleft, max_sync_size), &slout);
 424                 /* op.out set before loop */
 425                 op.outlen = -nx842_get_scatterlist_size(&slout);
 426
 427                 /* Send request to pHyp */
 428                 ret = vio_h_cop_sync(local_devdata->vdev, &op);
 429
 430                 /* Check for pHyp error */
 431                 if (ret) {
 432                         dev_dbg(dev, "%s: vio_h_cop_sync error (ret=%d, hret=%ld)\n",
 433                                 __func__, ret, op.hcall_err);
 434                         ret = -EIO;
 435                         goto unlock;
 436                 }
 437
 438                 /* Check for hardware error */
 439                 ret = nx842_validate_result(dev, &csbcpb->csb);
 440                 if (ret && ret != -ENOSPC)
 441                         goto unlock;
 442
 443                 /* Handle incompressible data */
 444                 if (unlikely(ret == -ENOSPC)) {
 445                         if (bytesleft < max_sync_size) {
 446                                 /*
 447                                  * Not enough space left in the output buffer
 448                                  * to store uncompressed block
 449                                  */
 450                                 goto unlock;
 451                         } else {
 452                                 /* Store incompressible block */
 453                                 memcpy((void *)outbuf, (void *)inbuf,
 454                                         max_sync_size);
 455                                 hdr->sizes[i] = -max_sync_size;
 456                                 outbuf += max_sync_size;
 457                                 bytesleft -= max_sync_size;
 458                                 /* Reset ret, incompressible data handled */
 459                                 ret = 0;
 460                         }
 461                 } else {
 462                         /* Normal case, compression was successful */
 463                         size = csbcpb->csb.processed_byte_count;
 464                         dev_dbg(dev, "%s: processed_bytes=%d\n",
 465                                 __func__, size);
 466                         hdr->sizes[i] = size;
 467                         outbuf += size;
 468                         bytesleft -= size;
 469                 }
 470
 471                 inbuf += max_sync_size;
 472         }
 473
 474         *outlen = (unsigned int)(outbuf - (unsigned long)out);
 475
 476 unlock:
 477         if (ret)
 478                 nx842_inc_comp_failed(local_devdata);
 479         else {
 480                 nx842_inc_comp_complete(local_devdata);
 481                 ibm_nx842_incr_hist(local_devdata->counters->comp_times,
 482                         (get_tb() - start_time) / tb_ticks_per_usec);
 483         }
 484         rcu_read_unlock();
 485         return ret;
 486 }
 487 EXPORT_SYMBOL_GPL(nx842_compress);
 488
 489 static int sw842_decompress(const unsigned char *, int, unsigned char *, int *,
 490                         const void *);
 491
 492 /**
 493  * nx842_decompress - Decompress data using the 842 algorithm
 494  *
 495  * Decompression provide by the NX842 coprocessor on IBM Power systems.
 496  * The input buffer is decompressed and the result is stored in the
 497  * provided output buffer.  The size allocated to the output buffer is
 498  * provided by the caller of this function in @outlen.  Upon return from
 499  * this function @outlen contains the length of the decompressed data.
 500  * If there is an error then @outlen will be 0 and an error will be
 501  * specified by the return code from this function.
 502  *
 503  * @in: Pointer to input buffer, will use bounce buffer if not 128 byte
 504  *      aligned
 505  * @inlen: Length of input buffer
 506  * @out: Pointer to output buffer, must be page aligned
 507  * @outlen: Length of output buffer, must be PAGE_SIZE
 508  * @wrkmem: ptr to buffer for working memory, size determined by
 509  *          nx842_get_workmem_size()
 510  *
 511  * Returns:
 512  *   0          Success, output of length @outlen stored in the buffer at @out
 513  *   -ENODEV    Hardware decompression device is unavailable
 514  *   -ENOMEM    Unable to allocate internal buffers
 515  *   -ENOSPC    Output buffer is to small
 516  *   -EINVAL    Bad input data encountered when attempting decompress
 517  *   -EIO       Internal error
 518  */
 519 int nx842_decompress(const unsigned char *in, unsigned int inlen,
 520                          unsigned char *out, unsigned int *outlen, void *wmem)
 521 {
 522         struct nx842_header *hdr;
 523         struct nx842_devdata *local_devdata;
 524         struct device *dev = NULL;
 525         struct nx842_workmem *workmem;
 526         struct nx842_scatterlist slin, slout;
 527         struct nx_csbcpb *csbcpb;
 528         int ret = 0, i, size, max_sync_size;
 529         unsigned long inbuf, outbuf;
 530         struct vio_pfo_op op = {
 531                 .done = NULL,
 532                 .handle = 0,
 533                 .timeout = 0,
 534         };
 535         unsigned long start_time = get_tb();
 536
 537         /* Ensure page alignment and size */
 538         outbuf = (unsigned long)out;
 539         if (!IS_ALIGNED(outbuf, PAGE_SIZE) || *outlen != PAGE_SIZE)
 540                 return -EINVAL;
 541
 542         rcu_read_lock();
 543         local_devdata = rcu_dereference(devdata);
 544         if (local_devdata)
 545                 dev = local_devdata->dev;
 546
 547         /* Get header */
 548         hdr = (struct nx842_header *)in;
 549
 550         workmem = (struct nx842_workmem *)ALIGN((unsigned long)wmem,
 551                 NX842_HW_PAGE_SIZE);
 552
 553         inbuf = (unsigned long)in + hdr->offset;
 554         if (likely(!IS_ALIGNED(inbuf, IO_BUFFER_ALIGN))) {
 555                 /* Copy block(s) into bounce buffer for alignment */
 556                 memcpy(workmem->bounce, in + hdr->offset, inlen - hdr->offset);
 557                 inbuf = (unsigned long)workmem->bounce;
 558         }
 559
 560         /* Init scatterlist */
 561         slin.entries = (struct nx842_slentry *)workmem->slin;
 562         slout.entries = (struct nx842_slentry *)workmem->slout;
 563
 564         /* Init operation */
 565         op.flags = NX842_OP_DECOMPRESS;
 566         csbcpb = &workmem->csbcpb;
 567         memset(csbcpb, 0, sizeof(*csbcpb));
 568         op.csbcpb = __pa(csbcpb);
 569
 570         /*
 571          * max_sync_size may have changed since compression,
 572          * so we can't read it from the device info. We need
 573          * to derive it from hdr->blocks_nr.
 574          */
 575         max_sync_size = PAGE_SIZE / hdr->blocks_nr;
 576
 577         for (i = 0; i < hdr->blocks_nr; i++) {
 578                 /* Skip padding */
 579                 inbuf = ALIGN(inbuf, IO_BUFFER_ALIGN);
 580
 581                 if (hdr->sizes[i] < 0) {
 582                         /* Negative sizes indicate uncompressed data blocks */
 583                         size = abs(hdr->sizes[i]);
 584                         memcpy((void *)outbuf, (void *)inbuf, size);
 585                         outbuf += size;
 586                         inbuf += size;
 587                         continue;
 588                 }
 589
 590                 if (!dev)
 591                         goto sw;
 592
 593                 /*
 594                  * The better the compression, the more likely the "likely"
 595                  * case becomes.
 596                  */
 597                 if (likely((inbuf & NX842_HW_PAGE_MASK) ==
 598                         ((inbuf + hdr->sizes[i] - 1) & NX842_HW_PAGE_MASK))) {
 599                         /* Create direct DDE */
 600                         op.in = __pa(inbuf);
 601                         op.inlen = hdr->sizes[i];
 602                 } else {
 603                         /* Create indirect DDE (scatterlist) */
 604                         nx842_build_scatterlist(inbuf, hdr->sizes[i] , &slin);
 605                         op.in = __pa(slin.entries);
 606                         op.inlen = -nx842_get_scatterlist_size(&slin);
 607                 }
 608
 609                 /*
 610                  * NOTE: If the default max_sync_size is changed from 4k
 611                  * to 64k, remove the "likely" case below, since a
 612                  * scatterlist will always be needed.
 613                  */
 614                 if (likely(max_sync_size == NX842_HW_PAGE_SIZE)) {
 615                         /* Create direct DDE */
 616                         op.out = __pa(outbuf);
 617                         op.outlen = max_sync_size;
 618                 } else {
 619                         /* Create indirect DDE (scatterlist) */
 620                         nx842_build_scatterlist(outbuf, max_sync_size, &slout);
 621                         op.out = __pa(slout.entries);
 622                         op.outlen = -nx842_get_scatterlist_size(&slout);
 623                 }
 624
 625                 /* Send request to pHyp */
 626                 ret = vio_h_cop_sync(local_devdata->vdev, &op);
 627
 628                 /* Check for pHyp error */
 629                 if (ret) {
 630                         dev_dbg(dev, "%s: vio_h_cop_sync error (ret=%d, hret=%ld)\n",
 631                                 __func__, ret, op.hcall_err);
 632                         dev = NULL;
 633                         goto sw;
 634                 }
 635
 636                 /* Check for hardware error */
 637                 ret = nx842_validate_result(dev, &csbcpb->csb);
 638                 if (ret) {
 639                         dev = NULL;
 640                         goto sw;
 641                 }
 642
 643                 /* HW decompression success */
 644                 inbuf += hdr->sizes[i];
 645                 outbuf += csbcpb->csb.processed_byte_count;
 646                 continue;
 647
 648 sw:
 649                 /* software decompression */
 650                 size = max_sync_size;
 651                 ret = sw842_decompress(
 652                         (unsigned char *)inbuf, hdr->sizes[i],
 653                         (unsigned char *)outbuf, &size, wmem);
 654                 if (ret)
 655                         pr_debug("%s: sw842_decompress failed with %d\n",
 656                                 __func__, ret);
 657
 658                 if (ret) {
 659                         if (ret != -ENOSPC && ret != -EINVAL &&
 660                                         ret != -EMSGSIZE)
 661                                 ret = -EIO;
 662                         goto unlock;
 663                 }
 664
 665                 /* SW decompression success */
 666                 inbuf += hdr->sizes[i];
 667                 outbuf += size;
 668         }
 669
 670         *outlen = (unsigned int)(outbuf - (unsigned long)out);
 671
 672 unlock:
 673         if (ret)
 674                 /* decompress fail */
 675                 nx842_inc_decomp_failed(local_devdata);
 676         else {
 677                 if (!dev)
 678                         /* software decompress */
 679                         nx842_inc_swdecomp(local_devdata);
 680                 nx842_inc_decomp_complete(local_devdata);
 681                 ibm_nx842_incr_hist(local_devdata->counters->decomp_times,
 682                         (get_tb() - start_time) / tb_ticks_per_usec);
 683         }
 684
 685         rcu_read_unlock();
 686         return ret;
 687 }
 688 EXPORT_SYMBOL_GPL(nx842_decompress);
 689
 690 /**
 691  * nx842_OF_set_defaults -- Set default (disabled) values for devdata
 692  *
 693  * @devdata - struct nx842_devdata to update
 694  *
 695  * Returns:
 696  *  0 on success
 697  *  -ENOENT if @devdata ptr is NULL
 698  */
 699 static int nx842_OF_set_defaults(struct nx842_devdata *devdata)
 700 {
 701         if (devdata) {
 702                 devdata->max_sync_size = 0;
 703                 devdata->max_sync_sg = 0;
 704                 devdata->max_sg_len = 0;
 705                 devdata->status = UNAVAILABLE;
 706                 return 0;
 707         } else
 708                 return -ENOENT;
 709 }
 710
 711 /**
 712  * nx842_OF_upd_status -- Update the device info from OF status prop
 713  *
 714  * The status property indicates if the accelerator is enabled.  If the
 715  * device is in the OF tree it indicates that the hardware is present.
 716  * The status field indicates if the device is enabled when the status
 717  * is 'okay'.  Otherwise the device driver will be disabled.
 718  *
 719  * @devdata - struct nx842_devdata to update
 720  * @prop - struct property point containing the maxsyncop for the update
 721  *
 722  * Returns:
 723  *  0 - Device is available
 724  *  -EINVAL - Device is not available
 725  */
 726 static int nx842_OF_upd_status(struct nx842_devdata *devdata,
 727                                         struct property *prop) {
 728         int ret = 0;
 729         const char *status = (const char *)prop->value;
 730
 731         if (!strncmp(status, "okay", (size_t)prop->length)) {
 732                 devdata->status = AVAILABLE;
 733         } else {
 734                 dev_info(devdata->dev, "%s: status '%s' is not 'okay'\n",
 735                                 __func__, status);
 736                 devdata->status = UNAVAILABLE;
 737         }
 738
 739         return ret;
 740 }
 741
 742 /**
 743  * nx842_OF_upd_maxsglen -- Update the device info from OF maxsglen prop
 744  *
 745  * Definition of the 'ibm,max-sg-len' OF property:
 746  *  This field indicates the maximum byte length of a scatter list
 747  *  for the platform facility. It is a single cell encoded as with encode-int.
 748  *
 749  * Example:
 750  *  # od -x ibm,max-sg-len
 751  *  0000000 0000 0ff0
 752  *
 753  *  In this example, the maximum byte length of a scatter list is
 754  *  0x0ff0 (4,080).
 755  *
 756  * @devdata - struct nx842_devdata to update
 757  * @prop - struct property point containing the maxsyncop for the update
 758  *
 759  * Returns:
 760  *  0 on success
 761  *  -EINVAL on failure
 762  */
 763 static int nx842_OF_upd_maxsglen(struct nx842_devdata *devdata,
 764                                         struct property *prop) {
 765         int ret = 0;
 766         const int *maxsglen = prop->value;
 767
 768         if (prop->length != sizeof(*maxsglen)) {
 769                 dev_err(devdata->dev, "%s: unexpected format for ibm,max-sg-len property\n", __func__);
 770                 dev_dbg(devdata->dev, "%s: ibm,max-sg-len is %d bytes long, expected %lu bytes\n", __func__,
 771                                 prop->length, sizeof(*maxsglen));
 772                 ret = -EINVAL;
 773         } else {
 774                 devdata->max_sg_len = (unsigned int)min(*maxsglen,
 775                                 (int)NX842_HW_PAGE_SIZE);
 776         }
 777
 778         return ret;
 779 }
 780
 781 /**
 782  * nx842_OF_upd_maxsyncop -- Update the device info from OF maxsyncop prop
 783  *
 784  * Definition of the 'ibm,max-sync-cop' OF property:
 785  *  Two series of cells.  The first series of cells represents the maximums
 786  *  that can be synchronously compressed. The second series of cells
 787  *  represents the maximums that can be synchronously decompressed.
 788  *  1. The first cell in each series contains the count of the number of
 789  *     data length, scatter list elements pairs that follow – each being
 790  *     of the form
 791  *    a. One cell data byte length
 792  *    b. One cell total number of scatter list elements
 793  *
 794  * Example:
 795  *  # od -x ibm,max-sync-cop
 796  *  0000000 0000 0001 0000 1000 0000 01fe 0000 0001
 797  *  0000020 0000 1000 0000 01fe
 798  *
 799  *  In this example, compression supports 0x1000 (4,096) data byte length
 800  *  and 0x1fe (510) total scatter list elements.  Decompression supports
 801  *  0x1000 (4,096) data byte length and 0x1f3 (510) total scatter list
 802  *  elements.
 803  *
 804  * @devdata - struct nx842_devdata to update
 805  * @prop - struct property point containing the maxsyncop for the update
 806  *
 807  * Returns:
 808  *  0 on success
 809  *  -EINVAL on failure
 810  */
 811 static int nx842_OF_upd_maxsyncop(struct nx842_devdata *devdata,
 812                                         struct property *prop) {
 813         int ret = 0;
 814         const struct maxsynccop_t {
 815                 int comp_elements;
 816                 int comp_data_limit;
 817                 int comp_sg_limit;
 818                 int decomp_elements;
 819                 int decomp_data_limit;
 820                 int decomp_sg_limit;
 821         } *maxsynccop;
 822
 823         if (prop->length != sizeof(*maxsynccop)) {
 824                 dev_err(devdata->dev, "%s: unexpected format for ibm,max-sync-cop property\n", __func__);
 825                 dev_dbg(devdata->dev, "%s: ibm,max-sync-cop is %d bytes long, expected %lu bytes\n", __func__, prop->length,
 826                                 sizeof(*maxsynccop));
 827                 ret = -EINVAL;
 828                 goto out;
 829         }
 830
 831         maxsynccop = (const struct maxsynccop_t *)prop->value;
 832
 833         /* Use one limit rather than separate limits for compression and
 834          * decompression. Set a maximum for this so as not to exceed the
 835          * size that the header can support and round the value down to
 836          * the hardware page size (4K) */
 837         devdata->max_sync_size =
 838                         (unsigned int)min(maxsynccop->comp_data_limit,
 839                                         maxsynccop->decomp_data_limit);
 840
 841         devdata->max_sync_size = min_t(unsigned int, devdata->max_sync_size,
 842                                         SIZE_64K);
 843
 844         if (devdata->max_sync_size < SIZE_4K) {
 845                 dev_err(devdata->dev, "%s: hardware max data size (%u) is "
 846                                 "less than the driver minimum, unable to use "
 847                                 "the hardware device\n",
 848                                 __func__, devdata->max_sync_size);
 849                 ret = -EINVAL;
 850                 goto out;
 851         }
 852
 853         devdata->max_sync_sg = (unsigned int)min(maxsynccop->comp_sg_limit,
 854                                                 maxsynccop->decomp_sg_limit);
 855         if (devdata->max_sync_sg < 1) {
 856                 dev_err(devdata->dev, "%s: hardware max sg size (%u) is "
 857                                 "less than the driver minimum, unable to use "
 858                                 "the hardware device\n",
 859                                 __func__, devdata->max_sync_sg);
 860                 ret = -EINVAL;
 861                 goto out;
 862         }
 863
 864 out:
 865         return ret;
 866 }
 867
 868 /**
 869  *
 870  * nx842_OF_upd -- Handle OF properties updates for the device.
 871  *
 872  * Set all properties from the OF tree.  Optionally, a new property
 873  * can be provided by the @new_prop pointer to overwrite an existing value.
 874  * The device will remain disabled until all values are valid, this function
 875  * will return an error for updates unless all values are valid.
 876  *
 877  * @new_prop: If not NULL, this property is being updated.  If NULL, update
 878  *  all properties from the current values in the OF tree.
 879  *
 880  * Returns:
 881  *  0 - Success
 882  *  -ENOMEM - Could not allocate memory for new devdata structure
 883  *  -EINVAL - property value not found, new_prop is not a recognized
 884  *      property for the device or property value is not valid.
 885  *  -ENODEV - Device is not available
 886  */
 887 static int nx842_OF_upd(struct property *new_prop)
 888 {
 889         struct nx842_devdata *old_devdata = NULL;
 890         struct nx842_devdata *new_devdata = NULL;
 891         struct device_node *of_node = NULL;
 892         struct property *status = NULL;
 893         struct property *maxsglen = NULL;
 894         struct property *maxsyncop = NULL;
 895         int ret = 0;
 896         unsigned long flags;
 897
 898         spin_lock_irqsave(&devdata_mutex, flags);
 899         old_devdata = rcu_dereference_check(devdata,
 900                         lockdep_is_held(&devdata_mutex));
 901         if (old_devdata)
 902                 of_node = old_devdata->dev->of_node;
 903
 904         if (!old_devdata || !of_node) {
 905                 pr_err("%s: device is not available\n", __func__);
 906                 spin_unlock_irqrestore(&devdata_mutex, flags);
 907                 return -ENODEV;
 908         }
 909
 910         new_devdata = kzalloc(sizeof(*new_devdata), GFP_NOFS);
 911         if (!new_devdata) {
 912                 dev_err(old_devdata->dev, "%s: Could not allocate memory for device data\n", __func__);
 913                 ret = -ENOMEM;
 914                 goto error_out;
 915         }
 916
 917         memcpy(new_devdata, old_devdata, sizeof(*old_devdata));
 918         new_devdata->counters = old_devdata->counters;
 919
 920         /* Set ptrs for existing properties */
 921         status = of_find_property(of_node, "status", NULL);
 922         maxsglen = of_find_property(of_node, "ibm,max-sg-len", NULL);
 923         maxsyncop = of_find_property(of_node, "ibm,max-sync-cop", NULL);
 924         if (!status || !maxsglen || !maxsyncop) {
 925                 dev_err(old_devdata->dev, "%s: Could not locate device properties\n", __func__);
 926                 ret = -EINVAL;
 927                 goto error_out;
 928         }
 929
 930         /*
 931          * If this is a property update, there are only certain properties that
 932          * we care about. Bail if it isn't in the below list
 933          */
 934         if (new_prop && (strncmp(new_prop->name, "status", new_prop->length) ||
 935                          strncmp(new_prop->name, "ibm,max-sg-len", new_prop->length) ||
 936                          strncmp(new_prop->name, "ibm,max-sync-cop", new_prop->length)))
 937                 goto out;
 938
 939         /* Perform property updates */
 940         ret = nx842_OF_upd_status(new_devdata, status);
 941         if (ret)
 942                 goto error_out;
 943
 944         ret = nx842_OF_upd_maxsglen(new_devdata, maxsglen);
 945         if (ret)
 946                 goto error_out;
 947
 948         ret = nx842_OF_upd_maxsyncop(new_devdata, maxsyncop);
 949         if (ret)
 950                 goto error_out;
 951
 952 out:
 953         dev_info(old_devdata->dev, "%s: max_sync_size new:%u old:%u\n",
 954                         __func__, new_devdata->max_sync_size,
 955                         old_devdata->max_sync_size);
 956         dev_info(old_devdata->dev, "%s: max_sync_sg new:%u old:%u\n",
 957                         __func__, new_devdata->max_sync_sg,
 958                         old_devdata->max_sync_sg);
 959         dev_info(old_devdata->dev, "%s: max_sg_len new:%u old:%u\n",
 960                         __func__, new_devdata->max_sg_len,
 961                         old_devdata->max_sg_len);
 962
 963         rcu_assign_pointer(devdata, new_devdata);
 964         spin_unlock_irqrestore(&devdata_mutex, flags);
 965         synchronize_rcu();
 966         dev_set_drvdata(new_devdata->dev, new_devdata);
 967         kfree(old_devdata);
 968         return 0;
 969
 970 error_out:
 971         if (new_devdata) {
 972                 dev_info(old_devdata->dev, "%s: device disabled\n", __func__);
 973                 nx842_OF_set_defaults(new_devdata);
 974                 rcu_assign_pointer(devdata, new_devdata);
 975                 spin_unlock_irqrestore(&devdata_mutex, flags);
 976                 synchronize_rcu();
 977                 dev_set_drvdata(new_devdata->dev, new_devdata);
 978                 kfree(old_devdata);
 979         } else {
 980                 dev_err(old_devdata->dev, "%s: could not update driver from hardware\n", __func__);
 981                 spin_unlock_irqrestore(&devdata_mutex, flags);
 982         }
 983
 984         if (!ret)
 985                 ret = -EINVAL;
 986         return ret;
 987 }
 988
 989 /**
 990  * nx842_OF_notifier - Process updates to OF properties for the device
 991  *
 992  * @np: notifier block
 993  * @action: notifier action
 994  * @update: struct pSeries_reconfig_prop_update pointer if action is
 995  *      PSERIES_UPDATE_PROPERTY
 996  *
 997  * Returns:
 998  *      NOTIFY_OK on success
 999  *      NOTIFY_BAD encoded with error number on failure, use
1000  *              notifier_to_errno() to decode this value
1001  */
1002 static int nx842_OF_notifier(struct notifier_block *np, unsigned long action,
1003                              void *update)
1004 {
1005         struct of_prop_reconfig *upd = update;
1006         struct nx842_devdata *local_devdata;
1007         struct device_node *node = NULL;
1008
1009         rcu_read_lock();
1010         local_devdata = rcu_dereference(devdata);
1011         if (local_devdata)
1012                 node = local_devdata->dev->of_node;
1013
1014         if (local_devdata &&
1015                         action == OF_RECONFIG_UPDATE_PROPERTY &&
1016                         !strcmp(upd->dn->name, node->name)) {
1017                 rcu_read_unlock();
1018                 nx842_OF_upd(upd->prop);
1019         } else
1020                 rcu_read_unlock();
1021
1022         return NOTIFY_OK;
1023 }
1024
1025 static struct notifier_block nx842_of_nb = {
1026         .notifier_call = nx842_OF_notifier,
1027 };
1028
1029 #define nx842_counter_read(_name)                                       \
1030 static ssize_t nx842_##_name##_show(struct device *dev,         \
1031                 struct device_attribute *attr,                          \
1032                 char *buf) {                                            \
1033         struct nx842_devdata *local_devdata;                    \
1034         int p = 0;                                                      \
1035         rcu_read_lock();                                                \
1036         local_devdata = rcu_dereference(devdata);                       \
1037         if (local_devdata)                                              \
1038                 p = snprintf(buf, PAGE_SIZE, "%ld\n",                   \
1039                        atomic64_read(&local_devdata->counters->_name)); \
1040         rcu_read_unlock();                                              \
1041         return p;                                                       \
1042 }
1043
1044 #define NX842DEV_COUNTER_ATTR_RO(_name)                                 \
1045         nx842_counter_read(_name);                                      \
1046         static struct device_attribute dev_attr_##_name = __ATTR(_name, \
1047                                                 0444,                   \
1048                                                 nx842_##_name##_show,\
1049                                                 NULL);
1050
1051 NX842DEV_COUNTER_ATTR_RO(comp_complete);
1052 NX842DEV_COUNTER_ATTR_RO(comp_failed);
1053 NX842DEV_COUNTER_ATTR_RO(decomp_complete);
1054 NX842DEV_COUNTER_ATTR_RO(decomp_failed);
1055 NX842DEV_COUNTER_ATTR_RO(swdecomp);
1056
1057 static ssize_t nx842_timehist_show(struct device *,
1058                 struct device_attribute *, char *);
1059
1060 static struct device_attribute dev_attr_comp_times = __ATTR(comp_times, 0444,
1061                 nx842_timehist_show, NULL);
1062 static struct device_attribute dev_attr_decomp_times = __ATTR(decomp_times,
1063                 0444, nx842_timehist_show, NULL);
1064
1065 static ssize_t nx842_timehist_show(struct device *dev,
1066                 struct device_attribute *attr, char *buf) {
1067         char *p = buf;
1068         struct nx842_devdata *local_devdata;
1069         atomic64_t *times;
1070         int bytes_remain = PAGE_SIZE;
1071         int bytes;
1072         int i;
1073
1074         rcu_read_lock();
1075         local_devdata = rcu_dereference(devdata);
1076         if (!local_devdata) {
1077                 rcu_read_unlock();
1078                 return 0;
1079         }
1080
1081         if (attr == &dev_attr_comp_times)
1082                 times = local_devdata->counters->comp_times;
1083         else if (attr == &dev_attr_decomp_times)
1084                 times = local_devdata->counters->decomp_times;
1085         else {
1086                 rcu_read_unlock();
1087                 return 0;
1088         }
1089
1090         for (i = 0; i < (NX842_HIST_SLOTS - 2); i++) {
1091                 bytes = snprintf(p, bytes_remain, "%u-%uus:\t%ld\n",
1092                                i ? (2<<(i-1)) : 0, (2<<i)-1,
1093                                atomic64_read(&times[i]));
1094                 bytes_remain -= bytes;
1095                 p += bytes;
1096         }
1097         /* The last bucket holds everything over
1098          * 2<<(NX842_HIST_SLOTS - 2) us */
1099         bytes = snprintf(p, bytes_remain, "%uus - :\t%ld\n",
1100                         2<<(NX842_HIST_SLOTS - 2),
1101                         atomic64_read(&times[(NX842_HIST_SLOTS - 1)]));
1102         p += bytes;
1103
1104         rcu_read_unlock();
1105         return p - buf;
1106 }
1107
1108 static struct attribute *nx842_sysfs_entries[] = {
1109         &dev_attr_comp_complete.attr,
1110         &dev_attr_comp_failed.attr,
1111         &dev_attr_decomp_complete.attr,
1112         &dev_attr_decomp_failed.attr,
1113         &dev_attr_swdecomp.attr,
1114         &dev_attr_comp_times.attr,
1115         &dev_attr_decomp_times.attr,
1116         NULL,
1117 };
1118
1119 static struct attribute_group nx842_attribute_group = {
1120         .name = NULL,           /* put in device directory */
1121         .attrs = nx842_sysfs_entries,
1122 };
1123
1124 static int __init nx842_probe(struct vio_dev *viodev,
1125                                   const struct vio_device_id *id)
1126 {
1127         struct nx842_devdata *old_devdata, *new_devdata = NULL;
1128         unsigned long flags;
1129         int ret = 0;
1130
1131         spin_lock_irqsave(&devdata_mutex, flags);
1132         old_devdata = rcu_dereference_check(devdata,
1133                         lockdep_is_held(&devdata_mutex));
1134
1135         if (old_devdata && old_devdata->vdev != NULL) {
1136                 dev_err(&viodev->dev, "%s: Attempt to register more than one instance of the hardware\n", __func__);
1137                 ret = -1;
1138                 goto error_unlock;
1139         }
1140
1141         dev_set_drvdata(&viodev->dev, NULL);
1142
1143         new_devdata = kzalloc(sizeof(*new_devdata), GFP_NOFS);
1144         if (!new_devdata) {
1145                 dev_err(&viodev->dev, "%s: Could not allocate memory for device data\n", __func__);
1146                 ret = -ENOMEM;
1147                 goto error_unlock;
1148         }
1149
1150         new_devdata->counters = kzalloc(sizeof(*new_devdata->counters),
1151                         GFP_NOFS);
1152         if (!new_devdata->counters) {
1153                 dev_err(&viodev->dev, "%s: Could not allocate memory for performance counters\n", __func__);
1154                 ret = -ENOMEM;
1155                 goto error_unlock;
1156         }
1157
1158         new_devdata->vdev = viodev;
1159         new_devdata->dev = &viodev->dev;
1160         nx842_OF_set_defaults(new_devdata);
1161
1162         rcu_assign_pointer(devdata, new_devdata);
1163         spin_unlock_irqrestore(&devdata_mutex, flags);
1164         synchronize_rcu();
1165         kfree(old_devdata);
1166
1167         of_reconfig_notifier_register(&nx842_of_nb);
1168
1169         ret = nx842_OF_upd(NULL);
1170         if (ret && ret != -ENODEV) {
1171                 dev_err(&viodev->dev, "could not parse device tree. %d\n", ret);
1172                 ret = -1;
1173                 goto error;
1174         }
1175
1176         rcu_read_lock();
1177         if (dev_set_drvdata(&viodev->dev, rcu_dereference(devdata))) {
1178                 rcu_read_unlock();
1179                 dev_err(&viodev->dev, "failed to set driver data for device\n");
1180                 ret = -1;
1181                 goto error;
1182         }
1183         rcu_read_unlock();
1184
1185         if (sysfs_create_group(&viodev->dev.kobj, &nx842_attribute_group)) {
1186                 dev_err(&viodev->dev, "could not create sysfs device attributes\n");
1187                 ret = -1;
1188                 goto error;
1189         }
1190
1191         return 0;
1192
1193 error_unlock:
1194         spin_unlock_irqrestore(&devdata_mutex, flags);
1195         if (new_devdata)
1196                 kfree(new_devdata->counters);
1197         kfree(new_devdata);
1198 error:
1199         return ret;
1200 }
1201
1202 static int __exit nx842_remove(struct vio_dev *viodev)
1203 {
1204         struct nx842_devdata *old_devdata;
1205         unsigned long flags;
1206
1207         pr_info("Removing IBM Power 842 compression device\n");
1208         sysfs_remove_group(&viodev->dev.kobj, &nx842_attribute_group);
1209
1210         spin_lock_irqsave(&devdata_mutex, flags);
1211         old_devdata = rcu_dereference_check(devdata,
1212                         lockdep_is_held(&devdata_mutex));
1213         of_reconfig_notifier_unregister(&nx842_of_nb);
1214         rcu_assign_pointer(devdata, NULL);
1215         spin_unlock_irqrestore(&devdata_mutex, flags);
1216         synchronize_rcu();
1217         dev_set_drvdata(&viodev->dev, NULL);
1218         if (old_devdata)
1219                 kfree(old_devdata->counters);
1220         kfree(old_devdata);
1221         return 0;
1222 }
1223
1224 static struct vio_device_id nx842_driver_ids[] = {
1225         {"ibm,compression-v1", "ibm,compression"},
1226         {"", ""},
1227 };
1228
1229 static struct vio_driver nx842_driver = {
1230         .name = MODULE_NAME,
1231         .probe = nx842_probe,
1232         .remove = nx842_remove,
1233         .get_desired_dma = nx842_get_desired_dma,
1234         .id_table = nx842_driver_ids,
1235 };
1236
1237 static int __init nx842_init(void)
1238 {
1239         struct nx842_devdata *new_devdata;
1240         pr_info("Registering IBM Power 842 compression driver\n");
1241
1242         RCU_INIT_POINTER(devdata, NULL);
1243         new_devdata = kzalloc(sizeof(*new_devdata), GFP_KERNEL);
1244         if (!new_devdata) {
1245                 pr_err("Could not allocate memory for device data\n");
1246                 return -ENOMEM;
1247         }
1248         new_devdata->status = UNAVAILABLE;
1249         RCU_INIT_POINTER(devdata, new_devdata);
1250
1251         return vio_register_driver(&nx842_driver);
1252 }
1253
1254 module_init(nx842_init);
1255
1256 static void __exit nx842_exit(void)
1257 {
1258         struct nx842_devdata *old_devdata;
1259         unsigned long flags;
1260
1261         pr_info("Exiting IBM Power 842 compression driver\n");
1262         spin_lock_irqsave(&devdata_mutex, flags);
1263         old_devdata = rcu_dereference_check(devdata,
1264                         lockdep_is_held(&devdata_mutex));
1265         rcu_assign_pointer(devdata, NULL);
1266         spin_unlock_irqrestore(&devdata_mutex, flags);
1267         synchronize_rcu();
1268         if (old_devdata)
1269                 dev_set_drvdata(old_devdata->dev, NULL);
1270         kfree(old_devdata);
1271         vio_unregister_driver(&nx842_driver);
1272 }
1273
1274 module_exit(nx842_exit);
1275
1276 /*********************************
1277  * 842 software decompressor
1278 *********************************/
1279 typedef int (*sw842_template_op)(const char **, int *, unsigned char **,
1280                                                 struct sw842_fifo *);
1281
1282 static int sw842_data8(const char **, int *, unsigned char **,
1283                                                 struct sw842_fifo *);
1284 static int sw842_data4(const char **, int *, unsigned char **,
1285                                                 struct sw842_fifo *);
1286 static int sw842_data2(const char **, int *, unsigned char **,
1287                                                 struct sw842_fifo *);
1288 static int sw842_ptr8(const char **, int *, unsigned char **,
1289                                                 struct sw842_fifo *);
1290 static int sw842_ptr4(const char **, int *, unsigned char **,
1291                                                 struct sw842_fifo *);
1292 static int sw842_ptr2(const char **, int *, unsigned char **,
1293                                                 struct sw842_fifo *);
1294
1295 /* special templates */
1296 #define SW842_TMPL_REPEAT 0x1B
1297 #define SW842_TMPL_ZEROS 0x1C
1298 #define SW842_TMPL_EOF 0x1E
1299
1300 static sw842_template_op sw842_tmpl_ops[26][4] = {
1301         { sw842_data8, NULL}, /* 0 (00000) */
1302         { sw842_data4, sw842_data2, sw842_ptr2,  NULL},
1303         { sw842_data4, sw842_ptr2,  sw842_data2, NULL},
1304         { sw842_data4, sw842_ptr2,  sw842_ptr2,  NULL},
1305         { sw842_data4, sw842_ptr4,  NULL},
1306         { sw842_data2, sw842_ptr2,  sw842_data4, NULL},
1307         { sw842_data2, sw842_ptr2,  sw842_data2, sw842_ptr2},
1308         { sw842_data2, sw842_ptr2,  sw842_ptr2,  sw842_data2},
1309         { sw842_data2, sw842_ptr2,  sw842_ptr2,  sw842_ptr2,},
1310         { sw842_data2, sw842_ptr2,  sw842_ptr4,  NULL},
1311         { sw842_ptr2,  sw842_data2, sw842_data4, NULL}, /* 10 (01010) */
1312         { sw842_ptr2,  sw842_data4, sw842_ptr2,  NULL},
1313         { sw842_ptr2,  sw842_data2, sw842_ptr2,  sw842_data2},
1314         { sw842_ptr2,  sw842_data2, sw842_ptr2,  sw842_ptr2},
1315         { sw842_ptr2,  sw842_data2, sw842_ptr4,  NULL},
1316         { sw842_ptr2,  sw842_ptr2,  sw842_data4, NULL},
1317         { sw842_ptr2,  sw842_ptr2,  sw842_data2, sw842_ptr2},
1318         { sw842_ptr2,  sw842_ptr2,  sw842_ptr2,  sw842_data2},
1319         { sw842_ptr2,  sw842_ptr2,  sw842_ptr2,  sw842_ptr2},
1320         { sw842_ptr2,  sw842_ptr2,  sw842_ptr4,  NULL},
1321         { sw842_ptr4,  sw842_data4, NULL}, /* 20 (10100) */
1322         { sw842_ptr4,  sw842_data2, sw842_ptr2,  NULL},
1323         { sw842_ptr4,  sw842_ptr2,  sw842_data2, NULL},
1324         { sw842_ptr4,  sw842_ptr2,  sw842_ptr2,  NULL},
1325         { sw842_ptr4,  sw842_ptr4,  NULL},
1326         { sw842_ptr8,  NULL}
1327 };
1328
1329 /* Software decompress helpers */
1330
1331 static uint8_t sw842_get_byte(const char *buf, int bit)
1332 {
1333         uint8_t tmpl;
1334         uint16_t tmp;
1335         tmp = htons(*(uint16_t *)(buf));
1336         tmp = (uint16_t)(tmp << bit);
1337         tmp = ntohs(tmp);
1338         memcpy(&tmpl, &tmp, 1);
1339         return tmpl;
1340 }
1341
1342 static uint8_t sw842_get_template(const char **buf, int *bit)
1343 {
1344         uint8_t byte;
1345         byte = sw842_get_byte(*buf, *bit);
1346         byte = byte >> 3;
1347         byte &= 0x1F;
1348         *buf += (*bit + 5) / 8;
1349         *bit = (*bit + 5) % 8;
1350         return byte;
1351 }
1352
1353 /* repeat_count happens to be 5-bit too (like the template) */
1354 static uint8_t sw842_get_repeat_count(const char **buf, int *bit)
1355 {
1356         uint8_t byte;
1357         byte = sw842_get_byte(*buf, *bit);
1358         byte = byte >> 2;
1359         byte &= 0x3F;
1360         *buf += (*bit + 6) / 8;
1361         *bit = (*bit + 6) % 8;
1362         return byte;
1363 }
1364
1365 static uint8_t sw842_get_ptr2(const char **buf, int *bit)
1366 {
1367         uint8_t ptr;
1368         ptr = sw842_get_byte(*buf, *bit);
1369         (*buf)++;
1370         return ptr;
1371 }
1372
1373 static uint16_t sw842_get_ptr4(const char **buf, int *bit,
1374                 struct sw842_fifo *fifo)
1375 {
1376         uint16_t ptr;
1377         ptr = htons(*(uint16_t *)(*buf));
1378         ptr = (uint16_t)(ptr << *bit);
1379         ptr = ptr >> 7;
1380         ptr &= 0x01FF;
1381         *buf += (*bit + 9) / 8;
1382         *bit = (*bit + 9) % 8;
1383         return ptr;
1384 }
1385
1386 static uint8_t sw842_get_ptr8(const char **buf, int *bit,
1387                 struct sw842_fifo *fifo)
1388 {
1389         return sw842_get_ptr2(buf, bit);
1390 }
1391
1392 /* Software decompress template ops */
1393
1394 static int sw842_data8(const char **inbuf, int *inbit,
1395                 unsigned char **outbuf, struct sw842_fifo *fifo)
1396 {
1397         int ret;
1398
1399         ret = sw842_data4(inbuf, inbit, outbuf, fifo);
1400         if (ret)
1401                 return ret;
1402         ret = sw842_data4(inbuf, inbit, outbuf, fifo);
1403         return ret;
1404 }
1405
1406 static int sw842_data4(const char **inbuf, int *inbit,
1407                 unsigned char **outbuf, struct sw842_fifo *fifo)
1408 {
1409         int ret;
1410
1411         ret = sw842_data2(inbuf, inbit, outbuf, fifo);
1412         if (ret)
1413                 return ret;
1414         ret = sw842_data2(inbuf, inbit, outbuf, fifo);
1415         return ret;
1416 }
1417
1418 static int sw842_data2(const char **inbuf, int *inbit,
1419                 unsigned char **outbuf, struct sw842_fifo *fifo)
1420 {
1421         **outbuf = sw842_get_byte(*inbuf, *inbit);
1422         (*inbuf)++;
1423         (*outbuf)++;
1424         **outbuf = sw842_get_byte(*inbuf, *inbit);
1425         (*inbuf)++;
1426         (*outbuf)++;
1427         return 0;
1428 }
1429
1430 static int sw842_ptr8(const char **inbuf, int *inbit,
1431                 unsigned char **outbuf, struct sw842_fifo *fifo)
1432 {
1433         uint8_t ptr;
1434         ptr = sw842_get_ptr8(inbuf, inbit, fifo);
1435         if (!fifo->f84_full && (ptr >= fifo->f8_count))
1436                 return 1;
1437         memcpy(*outbuf, fifo->f8[ptr], 8);
1438         *outbuf += 8;
1439         return 0;
1440 }
1441
1442 static int sw842_ptr4(const char **inbuf, int *inbit,
1443                 unsigned char **outbuf, struct sw842_fifo *fifo)
1444 {
1445         uint16_t ptr;
1446         ptr = sw842_get_ptr4(inbuf, inbit, fifo);
1447         if (!fifo->f84_full && (ptr >= fifo->f4_count))
1448                 return 1;
1449         memcpy(*outbuf, fifo->f4[ptr], 4);
1450         *outbuf += 4;
1451         return 0;
1452 }
1453
1454 static int sw842_ptr2(const char **inbuf, int *inbit,
1455                 unsigned char **outbuf, struct sw842_fifo *fifo)
1456 {
1457         uint8_t ptr;
1458         ptr = sw842_get_ptr2(inbuf, inbit);
1459         if (!fifo->f2_full && (ptr >= fifo->f2_count))
1460                 return 1;
1461         memcpy(*outbuf, fifo->f2[ptr], 2);
1462         *outbuf += 2;
1463         return 0;
1464 }
1465
1466 static void sw842_copy_to_fifo(const char *buf, struct sw842_fifo *fifo)
1467 {
1468         unsigned char initial_f2count = fifo->f2_count;
1469
1470         memcpy(fifo->f8[fifo->f8_count], buf, 8);
1471         fifo->f4_count += 2;
1472         fifo->f8_count += 1;
1473
1474         if (!fifo->f84_full && fifo->f4_count >= 512) {
1475                 fifo->f84_full = 1;
1476                 fifo->f4_count /= 512;
1477         }
1478
1479         memcpy(fifo->f2[fifo->f2_count++], buf, 2);
1480         memcpy(fifo->f2[fifo->f2_count++], buf + 2, 2);
1481         memcpy(fifo->f2[fifo->f2_count++], buf + 4, 2);
1482         memcpy(fifo->f2[fifo->f2_count++], buf + 6, 2);
1483         if (fifo->f2_count < initial_f2count)
1484                 fifo->f2_full = 1;
1485 }
1486
1487 static int sw842_decompress(const unsigned char *src, int srclen,
1488                         unsigned char *dst, int *destlen,
1489                         const void *wrkmem)
1490 {
1491         uint8_t tmpl;
1492         const char *inbuf;
1493         int inbit = 0;
1494         unsigned char *outbuf, *outbuf_end, *origbuf, *prevbuf;
1495         const char *inbuf_end;
1496         sw842_template_op op;
1497         int opindex;
1498         int i, repeat_count;
1499         struct sw842_fifo *fifo;
1500         int ret = 0;
1501
1502         fifo = &((struct nx842_workmem *)(wrkmem))->swfifo;
1503         memset(fifo, 0, sizeof(*fifo));
1504
1505         origbuf = NULL;
1506         inbuf = src;
1507         inbuf_end = src + srclen;
1508         outbuf = dst;
1509         outbuf_end = dst + *destlen;
1510
1511         while ((tmpl = sw842_get_template(&inbuf, &inbit)) != SW842_TMPL_EOF) {
1512                 if (inbuf >= inbuf_end) {
1513                         ret = -EINVAL;
1514                         goto out;
1515                 }
1516
1517                 opindex = 0;
1518                 prevbuf = origbuf;
1519                 origbuf = outbuf;
1520                 switch (tmpl) {
1521                 case SW842_TMPL_REPEAT:
1522                         if (prevbuf == NULL) {
1523                                 ret = -EINVAL;
1524                                 goto out;
1525                         }
1526
1527                         repeat_count = sw842_get_repeat_count(&inbuf,
1528                                                                 &inbit) + 1;
1529
1530                         /* Did the repeat count advance past the end of input */
1531                         if (inbuf > inbuf_end) {
1532                                 ret = -EINVAL;
1533                                 goto out;
1534                         }
1535
1536                         for (i = 0; i < repeat_count; i++) {
1537                                 /* Would this overflow the output buffer */
1538                                 if ((outbuf + 8) > outbuf_end) {
1539                                         ret = -ENOSPC;
1540                                         goto out;
1541                                 }
1542
1543                                 memcpy(outbuf, prevbuf, 8);
1544                                 sw842_copy_to_fifo(outbuf, fifo);
1545                                 outbuf += 8;
1546                         }
1547                         break;
1548
1549                 case SW842_TMPL_ZEROS:
1550                         /* Would this overflow the output buffer */
1551                         if ((outbuf + 8) > outbuf_end) {
1552                                 ret = -ENOSPC;
1553                                 goto out;
1554                         }
1555
1556                         memset(outbuf, 0, 8);
1557                         sw842_copy_to_fifo(outbuf, fifo);
1558                         outbuf += 8;
1559                         break;
1560
1561                 default:
1562                         if (tmpl > 25) {
1563                                 ret = -EINVAL;
1564                                 goto out;
1565                         }
1566
1567                         /* Does this go past the end of the input buffer */
1568                         if ((inbuf + 2) > inbuf_end) {
1569                                 ret = -EINVAL;
1570                                 goto out;
1571                         }
1572
1573                         /* Would this overflow the output buffer */
1574                         if ((outbuf + 8) > outbuf_end) {
1575                                 ret = -ENOSPC;
1576                                 goto out;
1577                         }
1578
1579                         while (opindex < 4 &&
1580                                 (op = sw842_tmpl_ops[tmpl][opindex++])
1581                                         != NULL) {
1582                                 ret = (*op)(&inbuf, &inbit, &outbuf, fifo);
1583                                 if (ret) {
1584                                         ret = -EINVAL;
1585                                         goto out;
1586                                 }
1587                                 sw842_copy_to_fifo(origbuf, fifo);
1588                         }
1589                 }
1590         }
1591
1592 out:
1593         if (!ret)
1594                 *destlen = (unsigned int)(outbuf - dst);
1595         else
1596                 *destlen = 0;
1597
1598         return ret;
1599 }