drivers/crypto/nx/nx-842-pseries.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 <asm/vio.h>
  25
  26 #include "nx-842.h"
  27 #include "nx_csbcpb.h" /* struct nx_csbcpb */
  28
  29 MODULE_LICENSE("GPL");
  30 MODULE_AUTHOR("Robert Jennings <rcj@linux.vnet.ibm.com>");
  31 MODULE_DESCRIPTION("842 H/W Compression driver for IBM Power processors");
  32 MODULE_ALIAS_CRYPTO("842");
  33 MODULE_ALIAS_CRYPTO("842-nx");
  34
  35 static struct nx842_constraints nx842_pseries_constraints = {
  36         .alignment =    DDE_BUFFER_ALIGN,
  37         .multiple =     DDE_BUFFER_LAST_MULT,
  38         .minimum =      DDE_BUFFER_LAST_MULT,
  39         .maximum =      PAGE_SIZE, /* dynamic, max_sync_size */
  40 };
  41
  42 static int check_constraints(unsigned long buf, unsigned int *len, bool in)
  43 {
  44         if (!IS_ALIGNED(buf, nx842_pseries_constraints.alignment)) {
  45                 pr_debug("%s buffer 0x%lx not aligned to 0x%x\n",
  46                          in ? "input" : "output", buf,
  47                          nx842_pseries_constraints.alignment);
  48                 return -EINVAL;
  49         }
  50         if (*len % nx842_pseries_constraints.multiple) {
  51                 pr_debug("%s buffer len 0x%x not multiple of 0x%x\n",
  52                          in ? "input" : "output", *len,
  53                          nx842_pseries_constraints.multiple);
  54                 if (in)
  55                         return -EINVAL;
  56                 *len = round_down(*len, nx842_pseries_constraints.multiple);
  57         }
  58         if (*len < nx842_pseries_constraints.minimum) {
  59                 pr_debug("%s buffer len 0x%x under minimum 0x%x\n",
  60                          in ? "input" : "output", *len,
  61                          nx842_pseries_constraints.minimum);
  62                 return -EINVAL;
  63         }
  64         if (*len > nx842_pseries_constraints.maximum) {
  65                 pr_debug("%s buffer len 0x%x over maximum 0x%x\n",
  66                          in ? "input" : "output", *len,
  67                          nx842_pseries_constraints.maximum);
  68                 if (in)
  69                         return -EINVAL;
  70                 *len = nx842_pseries_constraints.maximum;
  71         }
  72         return 0;
  73 }
  74
  75 /* I assume we need to align the CSB? */
  76 #define WORKMEM_ALIGN   (256)
  77
  78 struct nx842_workmem {
  79         /* scatterlist */
  80         char slin[4096];
  81         char slout[4096];
  82         /* coprocessor status/parameter block */
  83         struct nx_csbcpb csbcpb;
  84
  85         char padding[WORKMEM_ALIGN];
  86 } __aligned(WORKMEM_ALIGN);
  87
  88 /* Macros for fields within nx_csbcpb */
  89 /* Check the valid bit within the csbcpb valid field */
  90 #define NX842_CSBCBP_VALID_CHK(x) (x & BIT_MASK(7))
  91
  92 /* CE macros operate on the completion_extension field bits in the csbcpb.
  93  * CE0 0=full completion, 1=partial completion
  94  * CE1 0=CE0 indicates completion, 1=termination (output may be modified)
  95  * CE2 0=processed_bytes is source bytes, 1=processed_bytes is target bytes */
  96 #define NX842_CSBCPB_CE0(x)     (x & BIT_MASK(7))
  97 #define NX842_CSBCPB_CE1(x)     (x & BIT_MASK(6))
  98 #define NX842_CSBCPB_CE2(x)     (x & BIT_MASK(5))
  99
 100 /* The NX unit accepts data only on 4K page boundaries */
 101 #define NX842_HW_PAGE_SIZE      (4096)
 102 #define NX842_HW_PAGE_MASK      (~(NX842_HW_PAGE_SIZE-1))
 103
 104 struct ibm_nx842_counters {
 105         atomic64_t comp_complete;
 106         atomic64_t comp_failed;
 107         atomic64_t decomp_complete;
 108         atomic64_t decomp_failed;
 109         atomic64_t swdecomp;
 110         atomic64_t comp_times[32];
 111         atomic64_t decomp_times[32];
 112 };
 113
 114 static struct nx842_devdata {
 115         struct vio_dev *vdev;
 116         struct device *dev;
 117         struct ibm_nx842_counters *counters;
 118         unsigned int max_sg_len;
 119         unsigned int max_sync_size;
 120         unsigned int max_sync_sg;
 121 } __rcu *devdata;
 122 static DEFINE_SPINLOCK(devdata_mutex);
 123
 124 #define NX842_COUNTER_INC(_x) \
 125 static inline void nx842_inc_##_x( \
 126         const struct nx842_devdata *dev) { \
 127         if (dev) \
 128                 atomic64_inc(&dev->counters->_x); \
 129 }
 130 NX842_COUNTER_INC(comp_complete);
 131 NX842_COUNTER_INC(comp_failed);
 132 NX842_COUNTER_INC(decomp_complete);
 133 NX842_COUNTER_INC(decomp_failed);
 134 NX842_COUNTER_INC(swdecomp);
 135
 136 #define NX842_HIST_SLOTS 16
 137
 138 static void ibm_nx842_incr_hist(atomic64_t *times, unsigned int time)
 139 {
 140         int bucket = fls(time);
 141
 142         if (bucket)
 143                 bucket = min((NX842_HIST_SLOTS - 1), bucket - 1);
 144
 145         atomic64_inc(&times[bucket]);
 146 }
 147
 148 /* NX unit operation flags */
 149 #define NX842_OP_COMPRESS       0x0
 150 #define NX842_OP_CRC            0x1
 151 #define NX842_OP_DECOMPRESS     0x2
 152 #define NX842_OP_COMPRESS_CRC   (NX842_OP_COMPRESS | NX842_OP_CRC)
 153 #define NX842_OP_DECOMPRESS_CRC (NX842_OP_DECOMPRESS | NX842_OP_CRC)
 154 #define NX842_OP_ASYNC          (1<<23)
 155 #define NX842_OP_NOTIFY         (1<<22)
 156 #define NX842_OP_NOTIFY_INT(x)  ((x & 0xff)<<8)
 157
 158 static unsigned long nx842_get_desired_dma(struct vio_dev *viodev)
 159 {
 160         /* No use of DMA mappings within the driver. */
 161         return 0;
 162 }
 163
 164 struct nx842_slentry {
 165         __be64 ptr; /* Real address (use __pa()) */
 166         __be64 len;
 167 };
 168
 169 /* pHyp scatterlist entry */
 170 struct nx842_scatterlist {
 171         int entry_nr; /* number of slentries */
 172         struct nx842_slentry *entries; /* ptr to array of slentries */
 173 };
 174
 175 /* Does not include sizeof(entry_nr) in the size */
 176 static inline unsigned long nx842_get_scatterlist_size(
 177                                 struct nx842_scatterlist *sl)
 178 {
 179         return sl->entry_nr * sizeof(struct nx842_slentry);
 180 }
 181
 182 static int nx842_build_scatterlist(unsigned long buf, int len,
 183                         struct nx842_scatterlist *sl)
 184 {
 185         unsigned long entrylen;
 186         struct nx842_slentry *entry;
 187
 188         sl->entry_nr = 0;
 189
 190         entry = sl->entries;
 191         while (len) {
 192                 entry->ptr = cpu_to_be64(nx842_get_pa((void *)buf));
 193                 entrylen = min_t(int, len,
 194                                  LEN_ON_SIZE(buf, NX842_HW_PAGE_SIZE));
 195                 entry->len = cpu_to_be64(entrylen);
 196
 197                 len -= entrylen;
 198                 buf += entrylen;
 199
 200                 sl->entry_nr++;
 201                 entry++;
 202         }
 203
 204         return 0;
 205 }
 206
 207 static int nx842_validate_result(struct device *dev,
 208         struct cop_status_block *csb)
 209 {
 210         /* The csb must be valid after returning from vio_h_cop_sync */
 211         if (!NX842_CSBCBP_VALID_CHK(csb->valid)) {
 212                 dev_err(dev, "%s: cspcbp not valid upon completion.\n",
 213                                 __func__);
 214                 dev_dbg(dev, "valid:0x%02x cs:0x%02x cc:0x%02x ce:0x%02x\n",
 215                                 csb->valid,
 216                                 csb->crb_seq_number,
 217                                 csb->completion_code,
 218                                 csb->completion_extension);
 219                 dev_dbg(dev, "processed_bytes:%d address:0x%016lx\n",
 220                                 be32_to_cpu(csb->processed_byte_count),
 221                                 (unsigned long)be64_to_cpu(csb->address));
 222                 return -EIO;
 223         }
 224
 225         /* Check return values from the hardware in the CSB */
 226         switch (csb->completion_code) {
 227         case 0: /* Completed without error */
 228                 break;
 229         case 64: /* Compression ok, but output larger than input */
 230                 dev_dbg(dev, "%s: output size larger than input size\n",
 231                                         __func__);
 232                 break;
 233         case 13: /* Output buffer too small */
 234                 dev_dbg(dev, "%s: Out of space in output buffer\n",
 235                                         __func__);
 236                 return -ENOSPC;
 237         case 65: /* Calculated CRC doesn't match the passed value */
 238                 dev_dbg(dev, "%s: CRC mismatch for decompression\n",
 239                                         __func__);
 240                 return -EINVAL;
 241         case 66: /* Input data contains an illegal template field */
 242         case 67: /* Template indicates data past the end of the input stream */
 243                 dev_dbg(dev, "%s: Bad data for decompression (code:%d)\n",
 244                                         __func__, csb->completion_code);
 245                 return -EINVAL;
 246         default:
 247                 dev_dbg(dev, "%s: Unspecified error (code:%d)\n",
 248                                         __func__, csb->completion_code);
 249                 return -EIO;
 250         }
 251
 252         /* Hardware sanity check */
 253         if (!NX842_CSBCPB_CE2(csb->completion_extension)) {
 254                 dev_err(dev, "%s: No error returned by hardware, but "
 255                                 "data returned is unusable, contact support.\n"
 256                                 "(Additional info: csbcbp->processed bytes "
 257                                 "does not specify processed bytes for the "
 258                                 "target buffer.)\n", __func__);
 259                 return -EIO;
 260         }
 261
 262         return 0;
 263 }
 264
 265 /**
 266  * nx842_pseries_compress - Compress data using the 842 algorithm
 267  *
 268  * Compression provide by the NX842 coprocessor on IBM Power systems.
 269  * The input buffer is compressed and the result is stored in the
 270  * provided output buffer.
 271  *
 272  * Upon return from this function @outlen contains the length of the
 273  * compressed data.  If there is an error then @outlen will be 0 and an
 274  * error will be specified by the return code from this function.
 275  *
 276  * @in: Pointer to input buffer
 277  * @inlen: Length of input buffer
 278  * @out: Pointer to output buffer
 279  * @outlen: Length of output buffer
 280  * @wrkmem: ptr to buffer for working memory, size determined by
 281  *          nx842_pseries_driver.workmem_size
 282  *
 283  * Returns:
 284  *   0          Success, output of length @outlen stored in the buffer at @out
 285  *   -ENOMEM    Unable to allocate internal buffers
 286  *   -ENOSPC    Output buffer is to small
 287  *   -EIO       Internal error
 288  *   -ENODEV    Hardware unavailable
 289  */
 290 static int nx842_pseries_compress(const unsigned char *in, unsigned int inlen,
 291                                   unsigned char *out, unsigned int *outlen,
 292                                   void *wmem)
 293 {
 294         struct nx842_devdata *local_devdata;
 295         struct device *dev = NULL;
 296         struct nx842_workmem *workmem;
 297         struct nx842_scatterlist slin, slout;
 298         struct nx_csbcpb *csbcpb;
 299         int ret = 0;
 300         unsigned long inbuf, outbuf;
 301         struct vio_pfo_op op = {
 302                 .done = NULL,
 303                 .handle = 0,
 304                 .timeout = 0,
 305         };
 306         unsigned long start = get_tb();
 307
 308         inbuf = (unsigned long)in;
 309         if (check_constraints(inbuf, &inlen, true))
 310                 return -EINVAL;
 311
 312         outbuf = (unsigned long)out;
 313         if (check_constraints(outbuf, outlen, false))
 314                 return -EINVAL;
 315
 316         rcu_read_lock();
 317         local_devdata = rcu_dereference(devdata);
 318         if (!local_devdata || !local_devdata->dev) {
 319                 rcu_read_unlock();
 320                 return -ENODEV;
 321         }
 322         dev = local_devdata->dev;
 323
 324         /* Init scatterlist */
 325         workmem = PTR_ALIGN(wmem, WORKMEM_ALIGN);
 326         slin.entries = (struct nx842_slentry *)workmem->slin;
 327         slout.entries = (struct nx842_slentry *)workmem->slout;
 328
 329         /* Init operation */
 330         op.flags = NX842_OP_COMPRESS_CRC;
 331         csbcpb = &workmem->csbcpb;
 332         memset(csbcpb, 0, sizeof(*csbcpb));
 333         op.csbcpb = nx842_get_pa(csbcpb);
 334
 335         if ((inbuf & NX842_HW_PAGE_MASK) ==
 336             ((inbuf + inlen - 1) & NX842_HW_PAGE_MASK)) {
 337                 /* Create direct DDE */
 338                 op.in = nx842_get_pa((void *)inbuf);
 339                 op.inlen = inlen;
 340         } else {
 341                 /* Create indirect DDE (scatterlist) */
 342                 nx842_build_scatterlist(inbuf, inlen, &slin);
 343                 op.in = nx842_get_pa(slin.entries);
 344                 op.inlen = -nx842_get_scatterlist_size(&slin);
 345         }
 346
 347         if ((outbuf & NX842_HW_PAGE_MASK) ==
 348             ((outbuf + *outlen - 1) & NX842_HW_PAGE_MASK)) {
 349                 /* Create direct DDE */
 350                 op.out = nx842_get_pa((void *)outbuf);
 351                 op.outlen = *outlen;
 352         } else {
 353                 /* Create indirect DDE (scatterlist) */
 354                 nx842_build_scatterlist(outbuf, *outlen, &slout);
 355                 op.out = nx842_get_pa(slout.entries);
 356                 op.outlen = -nx842_get_scatterlist_size(&slout);
 357         }
 358
 359         dev_dbg(dev, "%s: op.in %lx op.inlen %ld op.out %lx op.outlen %ld\n",
 360                 __func__, (unsigned long)op.in, (long)op.inlen,
 361                 (unsigned long)op.out, (long)op.outlen);
 362
 363         /* Send request to pHyp */
 364         ret = vio_h_cop_sync(local_devdata->vdev, &op);
 365
 366         /* Check for pHyp error */
 367         if (ret) {
 368                 dev_dbg(dev, "%s: vio_h_cop_sync error (ret=%d, hret=%ld)\n",
 369                         __func__, ret, op.hcall_err);
 370                 ret = -EIO;
 371                 goto unlock;
 372         }
 373
 374         /* Check for hardware error */
 375         ret = nx842_validate_result(dev, &csbcpb->csb);
 376         if (ret)
 377                 goto unlock;
 378
 379         *outlen = be32_to_cpu(csbcpb->csb.processed_byte_count);
 380         dev_dbg(dev, "%s: processed_bytes=%d\n", __func__, *outlen);
 381
 382 unlock:
 383         if (ret)
 384                 nx842_inc_comp_failed(local_devdata);
 385         else {
 386                 nx842_inc_comp_complete(local_devdata);
 387                 ibm_nx842_incr_hist(local_devdata->counters->comp_times,
 388                         (get_tb() - start) / tb_ticks_per_usec);
 389         }
 390         rcu_read_unlock();
 391         return ret;
 392 }
 393
 394 /**
 395  * nx842_pseries_decompress - Decompress data using the 842 algorithm
 396  *
 397  * Decompression provide by the NX842 coprocessor on IBM Power systems.
 398  * The input buffer is decompressed and the result is stored in the
 399  * provided output buffer.  The size allocated to the output buffer is
 400  * provided by the caller of this function in @outlen.  Upon return from
 401  * this function @outlen contains the length of the decompressed data.
 402  * If there is an error then @outlen will be 0 and an error will be
 403  * specified by the return code from this function.
 404  *
 405  * @in: Pointer to input buffer
 406  * @inlen: Length of input buffer
 407  * @out: Pointer to output buffer
 408  * @outlen: Length of output buffer
 409  * @wrkmem: ptr to buffer for working memory, size determined by
 410  *          nx842_pseries_driver.workmem_size
 411  *
 412  * Returns:
 413  *   0          Success, output of length @outlen stored in the buffer at @out
 414  *   -ENODEV    Hardware decompression device is unavailable
 415  *   -ENOMEM    Unable to allocate internal buffers
 416  *   -ENOSPC    Output buffer is to small
 417  *   -EINVAL    Bad input data encountered when attempting decompress
 418  *   -EIO       Internal error
 419  */
 420 static int nx842_pseries_decompress(const unsigned char *in, unsigned int inlen,
 421                                     unsigned char *out, unsigned int *outlen,
 422                                     void *wmem)
 423 {
 424         struct nx842_devdata *local_devdata;
 425         struct device *dev = NULL;
 426         struct nx842_workmem *workmem;
 427         struct nx842_scatterlist slin, slout;
 428         struct nx_csbcpb *csbcpb;
 429         int ret = 0;
 430         unsigned long inbuf, outbuf;
 431         struct vio_pfo_op op = {
 432                 .done = NULL,
 433                 .handle = 0,
 434                 .timeout = 0,
 435         };
 436         unsigned long start = get_tb();
 437
 438         /* Ensure page alignment and size */
 439         inbuf = (unsigned long)in;
 440         if (check_constraints(inbuf, &inlen, true))
 441                 return -EINVAL;
 442
 443         outbuf = (unsigned long)out;
 444         if (check_constraints(outbuf, outlen, false))
 445                 return -EINVAL;
 446
 447         rcu_read_lock();
 448         local_devdata = rcu_dereference(devdata);
 449         if (!local_devdata || !local_devdata->dev) {
 450                 rcu_read_unlock();
 451                 return -ENODEV;
 452         }
 453         dev = local_devdata->dev;
 454
 455         workmem = PTR_ALIGN(wmem, WORKMEM_ALIGN);
 456
 457         /* Init scatterlist */
 458         slin.entries = (struct nx842_slentry *)workmem->slin;
 459         slout.entries = (struct nx842_slentry *)workmem->slout;
 460
 461         /* Init operation */
 462         op.flags = NX842_OP_DECOMPRESS_CRC;
 463         csbcpb = &workmem->csbcpb;
 464         memset(csbcpb, 0, sizeof(*csbcpb));
 465         op.csbcpb = nx842_get_pa(csbcpb);
 466
 467         if ((inbuf & NX842_HW_PAGE_MASK) ==
 468             ((inbuf + inlen - 1) & NX842_HW_PAGE_MASK)) {
 469                 /* Create direct DDE */
 470                 op.in = nx842_get_pa((void *)inbuf);
 471                 op.inlen = inlen;
 472         } else {
 473                 /* Create indirect DDE (scatterlist) */
 474                 nx842_build_scatterlist(inbuf, inlen, &slin);
 475                 op.in = nx842_get_pa(slin.entries);
 476                 op.inlen = -nx842_get_scatterlist_size(&slin);
 477         }
 478
 479         if ((outbuf & NX842_HW_PAGE_MASK) ==
 480             ((outbuf + *outlen - 1) & NX842_HW_PAGE_MASK)) {
 481                 /* Create direct DDE */
 482                 op.out = nx842_get_pa((void *)outbuf);
 483                 op.outlen = *outlen;
 484         } else {
 485                 /* Create indirect DDE (scatterlist) */
 486                 nx842_build_scatterlist(outbuf, *outlen, &slout);
 487                 op.out = nx842_get_pa(slout.entries);
 488                 op.outlen = -nx842_get_scatterlist_size(&slout);
 489         }
 490
 491         dev_dbg(dev, "%s: op.in %lx op.inlen %ld op.out %lx op.outlen %ld\n",
 492                 __func__, (unsigned long)op.in, (long)op.inlen,
 493                 (unsigned long)op.out, (long)op.outlen);
 494
 495         /* Send request to pHyp */
 496         ret = vio_h_cop_sync(local_devdata->vdev, &op);
 497
 498         /* Check for pHyp error */
 499         if (ret) {
 500                 dev_dbg(dev, "%s: vio_h_cop_sync error (ret=%d, hret=%ld)\n",
 501                         __func__, ret, op.hcall_err);
 502                 goto unlock;
 503         }
 504
 505         /* Check for hardware error */
 506         ret = nx842_validate_result(dev, &csbcpb->csb);
 507         if (ret)
 508                 goto unlock;
 509
 510         *outlen = be32_to_cpu(csbcpb->csb.processed_byte_count);
 511
 512 unlock:
 513         if (ret)
 514                 /* decompress fail */
 515                 nx842_inc_decomp_failed(local_devdata);
 516         else {
 517                 nx842_inc_decomp_complete(local_devdata);
 518                 ibm_nx842_incr_hist(local_devdata->counters->decomp_times,
 519                         (get_tb() - start) / tb_ticks_per_usec);
 520         }
 521
 522         rcu_read_unlock();
 523         return ret;
 524 }
 525
 526 /**
 527  * nx842_OF_set_defaults -- Set default (disabled) values for devdata
 528  *
 529  * @devdata - struct nx842_devdata to update
 530  *
 531  * Returns:
 532  *  0 on success
 533  *  -ENOENT if @devdata ptr is NULL
 534  */
 535 static int nx842_OF_set_defaults(struct nx842_devdata *devdata)
 536 {
 537         if (devdata) {
 538                 devdata->max_sync_size = 0;
 539                 devdata->max_sync_sg = 0;
 540                 devdata->max_sg_len = 0;
 541                 return 0;
 542         } else
 543                 return -ENOENT;
 544 }
 545
 546 /**
 547  * nx842_OF_upd_status -- Check the device info from OF status prop
 548  *
 549  * The status property indicates if the accelerator is enabled.  If the
 550  * device is in the OF tree it indicates that the hardware is present.
 551  * The status field indicates if the device is enabled when the status
 552  * is 'okay'.  Otherwise the device driver will be disabled.
 553  *
 554  * @prop - struct property point containing the maxsyncop for the update
 555  *
 556  * Returns:
 557  *  0 - Device is available
 558  *  -ENODEV - Device is not available
 559  */
 560 static int nx842_OF_upd_status(struct property *prop)
 561 {
 562         const char *status = (const char *)prop->value;
 563
 564         if (!strncmp(status, "okay", (size_t)prop->length))
 565                 return 0;
 566         if (!strncmp(status, "disabled", (size_t)prop->length))
 567                 return -ENODEV;
 568         dev_info(devdata->dev, "%s: unknown status '%s'\n", __func__, status);
 569
 570         return -EINVAL;
 571 }
 572
 573 /**
 574  * nx842_OF_upd_maxsglen -- Update the device info from OF maxsglen prop
 575  *
 576  * Definition of the 'ibm,max-sg-len' OF property:
 577  *  This field indicates the maximum byte length of a scatter list
 578  *  for the platform facility. It is a single cell encoded as with encode-int.
 579  *
 580  * Example:
 581  *  # od -x ibm,max-sg-len
 582  *  0000000 0000 0ff0
 583  *
 584  *  In this example, the maximum byte length of a scatter list is
 585  *  0x0ff0 (4,080).
 586  *
 587  * @devdata - struct nx842_devdata to update
 588  * @prop - struct property point containing the maxsyncop for the update
 589  *
 590  * Returns:
 591  *  0 on success
 592  *  -EINVAL on failure
 593  */
 594 static int nx842_OF_upd_maxsglen(struct nx842_devdata *devdata,
 595                                         struct property *prop) {
 596         int ret = 0;
 597         const unsigned int maxsglen = of_read_number(prop->value, 1);
 598
 599         if (prop->length != sizeof(maxsglen)) {
 600                 dev_err(devdata->dev, "%s: unexpected format for ibm,max-sg-len property\n", __func__);
 601                 dev_dbg(devdata->dev, "%s: ibm,max-sg-len is %d bytes long, expected %lu bytes\n", __func__,
 602                                 prop->length, sizeof(maxsglen));
 603                 ret = -EINVAL;
 604         } else {
 605                 devdata->max_sg_len = min_t(unsigned int,
 606                                             maxsglen, NX842_HW_PAGE_SIZE);
 607         }
 608
 609         return ret;
 610 }
 611
 612 /**
 613  * nx842_OF_upd_maxsyncop -- Update the device info from OF maxsyncop prop
 614  *
 615  * Definition of the 'ibm,max-sync-cop' OF property:
 616  *  Two series of cells.  The first series of cells represents the maximums
 617  *  that can be synchronously compressed. The second series of cells
 618  *  represents the maximums that can be synchronously decompressed.
 619  *  1. The first cell in each series contains the count of the number of
 620  *     data length, scatter list elements pairs that follow – each being
 621  *     of the form
 622  *    a. One cell data byte length
 623  *    b. One cell total number of scatter list elements
 624  *
 625  * Example:
 626  *  # od -x ibm,max-sync-cop
 627  *  0000000 0000 0001 0000 1000 0000 01fe 0000 0001
 628  *  0000020 0000 1000 0000 01fe
 629  *
 630  *  In this example, compression supports 0x1000 (4,096) data byte length
 631  *  and 0x1fe (510) total scatter list elements.  Decompression supports
 632  *  0x1000 (4,096) data byte length and 0x1f3 (510) total scatter list
 633  *  elements.
 634  *
 635  * @devdata - struct nx842_devdata to update
 636  * @prop - struct property point containing the maxsyncop for the update
 637  *
 638  * Returns:
 639  *  0 on success
 640  *  -EINVAL on failure
 641  */
 642 static int nx842_OF_upd_maxsyncop(struct nx842_devdata *devdata,
 643                                         struct property *prop) {
 644         int ret = 0;
 645         unsigned int comp_data_limit, decomp_data_limit;
 646         unsigned int comp_sg_limit, decomp_sg_limit;
 647         const struct maxsynccop_t {
 648                 __be32 comp_elements;
 649                 __be32 comp_data_limit;
 650                 __be32 comp_sg_limit;
 651                 __be32 decomp_elements;
 652                 __be32 decomp_data_limit;
 653                 __be32 decomp_sg_limit;
 654         } *maxsynccop;
 655
 656         if (prop->length != sizeof(*maxsynccop)) {
 657                 dev_err(devdata->dev, "%s: unexpected format for ibm,max-sync-cop property\n", __func__);
 658                 dev_dbg(devdata->dev, "%s: ibm,max-sync-cop is %d bytes long, expected %lu bytes\n", __func__, prop->length,
 659                                 sizeof(*maxsynccop));
 660                 ret = -EINVAL;
 661                 goto out;
 662         }
 663
 664         maxsynccop = (const struct maxsynccop_t *)prop->value;
 665         comp_data_limit = be32_to_cpu(maxsynccop->comp_data_limit);
 666         comp_sg_limit = be32_to_cpu(maxsynccop->comp_sg_limit);
 667         decomp_data_limit = be32_to_cpu(maxsynccop->decomp_data_limit);
 668         decomp_sg_limit = be32_to_cpu(maxsynccop->decomp_sg_limit);
 669
 670         /* Use one limit rather than separate limits for compression and
 671          * decompression. Set a maximum for this so as not to exceed the
 672          * size that the header can support and round the value down to
 673          * the hardware page size (4K) */
 674         devdata->max_sync_size = min(comp_data_limit, decomp_data_limit);
 675
 676         devdata->max_sync_size = min_t(unsigned int, devdata->max_sync_size,
 677                                         65536);
 678
 679         if (devdata->max_sync_size < 4096) {
 680                 dev_err(devdata->dev, "%s: hardware max data size (%u) is "
 681                                 "less than the driver minimum, unable to use "
 682                                 "the hardware device\n",
 683                                 __func__, devdata->max_sync_size);
 684                 ret = -EINVAL;
 685                 goto out;
 686         }
 687
 688         nx842_pseries_constraints.maximum = devdata->max_sync_size;
 689
 690         devdata->max_sync_sg = min(comp_sg_limit, decomp_sg_limit);
 691         if (devdata->max_sync_sg < 1) {
 692                 dev_err(devdata->dev, "%s: hardware max sg size (%u) is "
 693                                 "less than the driver minimum, unable to use "
 694                                 "the hardware device\n",
 695                                 __func__, devdata->max_sync_sg);
 696                 ret = -EINVAL;
 697                 goto out;
 698         }
 699
 700 out:
 701         return ret;
 702 }
 703
 704 /**
 705  *
 706  * nx842_OF_upd -- Handle OF properties updates for the device.
 707  *
 708  * Set all properties from the OF tree.  Optionally, a new property
 709  * can be provided by the @new_prop pointer to overwrite an existing value.
 710  * The device will remain disabled until all values are valid, this function
 711  * will return an error for updates unless all values are valid.
 712  *
 713  * @new_prop: If not NULL, this property is being updated.  If NULL, update
 714  *  all properties from the current values in the OF tree.
 715  *
 716  * Returns:
 717  *  0 - Success
 718  *  -ENOMEM - Could not allocate memory for new devdata structure
 719  *  -EINVAL - property value not found, new_prop is not a recognized
 720  *      property for the device or property value is not valid.
 721  *  -ENODEV - Device is not available
 722  */
 723 static int nx842_OF_upd(struct property *new_prop)
 724 {
 725         struct nx842_devdata *old_devdata = NULL;
 726         struct nx842_devdata *new_devdata = NULL;
 727         struct device_node *of_node = NULL;
 728         struct property *status = NULL;
 729         struct property *maxsglen = NULL;
 730         struct property *maxsyncop = NULL;
 731         int ret = 0;
 732         unsigned long flags;
 733
 734         new_devdata = kzalloc(sizeof(*new_devdata), GFP_NOFS);
 735         if (!new_devdata)
 736                 return -ENOMEM;
 737
 738         spin_lock_irqsave(&devdata_mutex, flags);
 739         old_devdata = rcu_dereference_check(devdata,
 740                         lockdep_is_held(&devdata_mutex));
 741         if (old_devdata)
 742                 of_node = old_devdata->dev->of_node;
 743
 744         if (!old_devdata || !of_node) {
 745                 pr_err("%s: device is not available\n", __func__);
 746                 spin_unlock_irqrestore(&devdata_mutex, flags);
 747                 kfree(new_devdata);
 748                 return -ENODEV;
 749         }
 750
 751         memcpy(new_devdata, old_devdata, sizeof(*old_devdata));
 752         new_devdata->counters = old_devdata->counters;
 753
 754         /* Set ptrs for existing properties */
 755         status = of_find_property(of_node, "status", NULL);
 756         maxsglen = of_find_property(of_node, "ibm,max-sg-len", NULL);
 757         maxsyncop = of_find_property(of_node, "ibm,max-sync-cop", NULL);
 758         if (!status || !maxsglen || !maxsyncop) {
 759                 dev_err(old_devdata->dev, "%s: Could not locate device properties\n", __func__);
 760                 ret = -EINVAL;
 761                 goto error_out;
 762         }
 763
 764         /*
 765          * If this is a property update, there are only certain properties that
 766          * we care about. Bail if it isn't in the below list
 767          */
 768         if (new_prop && (strncmp(new_prop->name, "status", new_prop->length) ||
 769                          strncmp(new_prop->name, "ibm,max-sg-len", new_prop->length) ||
 770                          strncmp(new_prop->name, "ibm,max-sync-cop", new_prop->length)))
 771                 goto out;
 772
 773         /* Perform property updates */
 774         ret = nx842_OF_upd_status(status);
 775         if (ret)
 776                 goto error_out;
 777
 778         ret = nx842_OF_upd_maxsglen(new_devdata, maxsglen);
 779         if (ret)
 780                 goto error_out;
 781
 782         ret = nx842_OF_upd_maxsyncop(new_devdata, maxsyncop);
 783         if (ret)
 784                 goto error_out;
 785
 786 out:
 787         dev_info(old_devdata->dev, "%s: max_sync_size new:%u old:%u\n",
 788                         __func__, new_devdata->max_sync_size,
 789                         old_devdata->max_sync_size);
 790         dev_info(old_devdata->dev, "%s: max_sync_sg new:%u old:%u\n",
 791                         __func__, new_devdata->max_sync_sg,
 792                         old_devdata->max_sync_sg);
 793         dev_info(old_devdata->dev, "%s: max_sg_len new:%u old:%u\n",
 794                         __func__, new_devdata->max_sg_len,
 795                         old_devdata->max_sg_len);
 796
 797         rcu_assign_pointer(devdata, new_devdata);
 798         spin_unlock_irqrestore(&devdata_mutex, flags);
 799         synchronize_rcu();
 800         dev_set_drvdata(new_devdata->dev, new_devdata);
 801         kfree(old_devdata);
 802         return 0;
 803
 804 error_out:
 805         if (new_devdata) {
 806                 dev_info(old_devdata->dev, "%s: device disabled\n", __func__);
 807                 nx842_OF_set_defaults(new_devdata);
 808                 rcu_assign_pointer(devdata, new_devdata);
 809                 spin_unlock_irqrestore(&devdata_mutex, flags);
 810                 synchronize_rcu();
 811                 dev_set_drvdata(new_devdata->dev, new_devdata);
 812                 kfree(old_devdata);
 813         } else {
 814                 dev_err(old_devdata->dev, "%s: could not update driver from hardware\n", __func__);
 815                 spin_unlock_irqrestore(&devdata_mutex, flags);
 816         }
 817
 818         if (!ret)
 819                 ret = -EINVAL;
 820         return ret;
 821 }
 822
 823 /**
 824  * nx842_OF_notifier - Process updates to OF properties for the device
 825  *
 826  * @np: notifier block
 827  * @action: notifier action
 828  * @update: struct pSeries_reconfig_prop_update pointer if action is
 829  *      PSERIES_UPDATE_PROPERTY
 830  *
 831  * Returns:
 832  *      NOTIFY_OK on success
 833  *      NOTIFY_BAD encoded with error number on failure, use
 834  *              notifier_to_errno() to decode this value
 835  */
 836 static int nx842_OF_notifier(struct notifier_block *np, unsigned long action,
 837                              void *data)
 838 {
 839         struct of_reconfig_data *upd = data;
 840         struct nx842_devdata *local_devdata;
 841         struct device_node *node = NULL;
 842
 843         rcu_read_lock();
 844         local_devdata = rcu_dereference(devdata);
 845         if (local_devdata)
 846                 node = local_devdata->dev->of_node;
 847
 848         if (local_devdata &&
 849                         action == OF_RECONFIG_UPDATE_PROPERTY &&
 850                         !strcmp(upd->dn->name, node->name)) {
 851                 rcu_read_unlock();
 852                 nx842_OF_upd(upd->prop);
 853         } else
 854                 rcu_read_unlock();
 855
 856         return NOTIFY_OK;
 857 }
 858
 859 static struct notifier_block nx842_of_nb = {
 860         .notifier_call = nx842_OF_notifier,
 861 };
 862
 863 #define nx842_counter_read(_name)                                       \
 864 static ssize_t nx842_##_name##_show(struct device *dev,         \
 865                 struct device_attribute *attr,                          \
 866                 char *buf) {                                            \
 867         struct nx842_devdata *local_devdata;                    \
 868         int p = 0;                                                      \
 869         rcu_read_lock();                                                \
 870         local_devdata = rcu_dereference(devdata);                       \
 871         if (local_devdata)                                              \
 872                 p = snprintf(buf, PAGE_SIZE, "%ld\n",                   \
 873                        atomic64_read(&local_devdata->counters->_name)); \
 874         rcu_read_unlock();                                              \
 875         return p;                                                       \
 876 }
 877
 878 #define NX842DEV_COUNTER_ATTR_RO(_name)                                 \
 879         nx842_counter_read(_name);                                      \
 880         static struct device_attribute dev_attr_##_name = __ATTR(_name, \
 881                                                 0444,                   \
 882                                                 nx842_##_name##_show,\
 883                                                 NULL);
 884
 885 NX842DEV_COUNTER_ATTR_RO(comp_complete);
 886 NX842DEV_COUNTER_ATTR_RO(comp_failed);
 887 NX842DEV_COUNTER_ATTR_RO(decomp_complete);
 888 NX842DEV_COUNTER_ATTR_RO(decomp_failed);
 889 NX842DEV_COUNTER_ATTR_RO(swdecomp);
 890
 891 static ssize_t nx842_timehist_show(struct device *,
 892                 struct device_attribute *, char *);
 893
 894 static struct device_attribute dev_attr_comp_times = __ATTR(comp_times, 0444,
 895                 nx842_timehist_show, NULL);
 896 static struct device_attribute dev_attr_decomp_times = __ATTR(decomp_times,
 897                 0444, nx842_timehist_show, NULL);
 898
 899 static ssize_t nx842_timehist_show(struct device *dev,
 900                 struct device_attribute *attr, char *buf) {
 901         char *p = buf;
 902         struct nx842_devdata *local_devdata;
 903         atomic64_t *times;
 904         int bytes_remain = PAGE_SIZE;
 905         int bytes;
 906         int i;
 907
 908         rcu_read_lock();
 909         local_devdata = rcu_dereference(devdata);
 910         if (!local_devdata) {
 911                 rcu_read_unlock();
 912                 return 0;
 913         }
 914
 915         if (attr == &dev_attr_comp_times)
 916                 times = local_devdata->counters->comp_times;
 917         else if (attr == &dev_attr_decomp_times)
 918                 times = local_devdata->counters->decomp_times;
 919         else {
 920                 rcu_read_unlock();
 921                 return 0;
 922         }
 923
 924         for (i = 0; i < (NX842_HIST_SLOTS - 2); i++) {
 925                 bytes = snprintf(p, bytes_remain, "%u-%uus:\t%ld\n",
 926                                i ? (2<<(i-1)) : 0, (2<<i)-1,
 927                                atomic64_read(&times[i]));
 928                 bytes_remain -= bytes;
 929                 p += bytes;
 930         }
 931         /* The last bucket holds everything over
 932          * 2<<(NX842_HIST_SLOTS - 2) us */
 933         bytes = snprintf(p, bytes_remain, "%uus - :\t%ld\n",
 934                         2<<(NX842_HIST_SLOTS - 2),
 935                         atomic64_read(&times[(NX842_HIST_SLOTS - 1)]));
 936         p += bytes;
 937
 938         rcu_read_unlock();
 939         return p - buf;
 940 }
 941
 942 static struct attribute *nx842_sysfs_entries[] = {
 943         &dev_attr_comp_complete.attr,
 944         &dev_attr_comp_failed.attr,
 945         &dev_attr_decomp_complete.attr,
 946         &dev_attr_decomp_failed.attr,
 947         &dev_attr_swdecomp.attr,
 948         &dev_attr_comp_times.attr,
 949         &dev_attr_decomp_times.attr,
 950         NULL,
 951 };
 952
 953 static struct attribute_group nx842_attribute_group = {
 954         .name = NULL,           /* put in device directory */
 955         .attrs = nx842_sysfs_entries,
 956 };
 957
 958 static struct nx842_driver nx842_pseries_driver = {
 959         .name =         KBUILD_MODNAME,
 960         .owner =        THIS_MODULE,
 961         .workmem_size = sizeof(struct nx842_workmem),
 962         .constraints =  &nx842_pseries_constraints,
 963         .compress =     nx842_pseries_compress,
 964         .decompress =   nx842_pseries_decompress,
 965 };
 966
 967 static int nx842_pseries_crypto_init(struct crypto_tfm *tfm)
 968 {
 969         return nx842_crypto_init(tfm, &nx842_pseries_driver);
 970 }
 971
 972 static struct crypto_alg nx842_pseries_alg = {
 973         .cra_name               = "842",
 974         .cra_driver_name        = "842-nx",
 975         .cra_priority           = 300,
 976         .cra_flags              = CRYPTO_ALG_TYPE_COMPRESS,
 977         .cra_ctxsize            = sizeof(struct nx842_crypto_ctx),
 978         .cra_module             = THIS_MODULE,
 979         .cra_init               = nx842_pseries_crypto_init,
 980         .cra_exit               = nx842_crypto_exit,
 981         .cra_u                  = { .compress = {
 982         .coa_compress           = nx842_crypto_compress,
 983         .coa_decompress         = nx842_crypto_decompress } }
 984 };
 985
 986 static int nx842_probe(struct vio_dev *viodev,
 987                        const struct vio_device_id *id)
 988 {
 989         struct nx842_devdata *old_devdata, *new_devdata = NULL;
 990         unsigned long flags;
 991         int ret = 0;
 992
 993         new_devdata = kzalloc(sizeof(*new_devdata), GFP_NOFS);
 994         if (!new_devdata)
 995                 return -ENOMEM;
 996
 997         new_devdata->counters = kzalloc(sizeof(*new_devdata->counters),
 998                         GFP_NOFS);
 999         if (!new_devdata->counters) {
1000                 kfree(new_devdata);
1001                 return -ENOMEM;
1002         }
1003
1004         spin_lock_irqsave(&devdata_mutex, flags);
1005         old_devdata = rcu_dereference_check(devdata,
1006                         lockdep_is_held(&devdata_mutex));
1007
1008         if (old_devdata && old_devdata->vdev != NULL) {
1009                 dev_err(&viodev->dev, "%s: Attempt to register more than one instance of the hardware\n", __func__);
1010                 ret = -1;
1011                 goto error_unlock;
1012         }
1013
1014         dev_set_drvdata(&viodev->dev, NULL);
1015
1016         new_devdata->vdev = viodev;
1017         new_devdata->dev = &viodev->dev;
1018         nx842_OF_set_defaults(new_devdata);
1019
1020         rcu_assign_pointer(devdata, new_devdata);
1021         spin_unlock_irqrestore(&devdata_mutex, flags);
1022         synchronize_rcu();
1023         kfree(old_devdata);
1024
1025         of_reconfig_notifier_register(&nx842_of_nb);
1026
1027         ret = nx842_OF_upd(NULL);
1028         if (ret)
1029                 goto error;
1030
1031         ret = crypto_register_alg(&nx842_pseries_alg);
1032         if (ret) {
1033                 dev_err(&viodev->dev, "could not register comp alg: %d\n", ret);
1034                 goto error;
1035         }
1036
1037         rcu_read_lock();
1038         dev_set_drvdata(&viodev->dev, rcu_dereference(devdata));
1039         rcu_read_unlock();
1040
1041         if (sysfs_create_group(&viodev->dev.kobj, &nx842_attribute_group)) {
1042                 dev_err(&viodev->dev, "could not create sysfs device attributes\n");
1043                 ret = -1;
1044                 goto error;
1045         }
1046
1047         return 0;
1048
1049 error_unlock:
1050         spin_unlock_irqrestore(&devdata_mutex, flags);
1051         if (new_devdata)
1052                 kfree(new_devdata->counters);
1053         kfree(new_devdata);
1054 error:
1055         return ret;
1056 }
1057
1058 static int nx842_remove(struct vio_dev *viodev)
1059 {
1060         struct nx842_devdata *old_devdata;
1061         unsigned long flags;
1062
1063         pr_info("Removing IBM Power 842 compression device\n");
1064         sysfs_remove_group(&viodev->dev.kobj, &nx842_attribute_group);
1065
1066         crypto_unregister_alg(&nx842_pseries_alg);
1067
1068         spin_lock_irqsave(&devdata_mutex, flags);
1069         old_devdata = rcu_dereference_check(devdata,
1070                         lockdep_is_held(&devdata_mutex));
1071         of_reconfig_notifier_unregister(&nx842_of_nb);
1072         RCU_INIT_POINTER(devdata, NULL);
1073         spin_unlock_irqrestore(&devdata_mutex, flags);
1074         synchronize_rcu();
1075         dev_set_drvdata(&viodev->dev, NULL);
1076         if (old_devdata)
1077                 kfree(old_devdata->counters);
1078         kfree(old_devdata);
1079
1080         return 0;
1081 }
1082
1083 static const struct vio_device_id nx842_vio_driver_ids[] = {
1084         {"ibm,compression-v1", "ibm,compression"},
1085         {"", ""},
1086 };
1087
1088 static struct vio_driver nx842_vio_driver = {
1089         .name = KBUILD_MODNAME,
1090         .probe = nx842_probe,
1091         .remove = nx842_remove,
1092         .get_desired_dma = nx842_get_desired_dma,
1093         .id_table = nx842_vio_driver_ids,
1094 };
1095
1096 static int __init nx842_pseries_init(void)
1097 {
1098         struct nx842_devdata *new_devdata;
1099         int ret;
1100
1101         if (!of_find_compatible_node(NULL, NULL, "ibm,compression"))
1102                 return -ENODEV;
1103
1104         RCU_INIT_POINTER(devdata, NULL);
1105         new_devdata = kzalloc(sizeof(*new_devdata), GFP_KERNEL);
1106         if (!new_devdata)
1107                 return -ENOMEM;
1108
1109         RCU_INIT_POINTER(devdata, new_devdata);
1110
1111         ret = vio_register_driver(&nx842_vio_driver);
1112         if (ret) {
1113                 pr_err("Could not register VIO driver %d\n", ret);
1114
1115                 kfree(new_devdata);
1116                 return ret;
1117         }
1118
1119         return 0;
1120 }
1121
1122 module_init(nx842_pseries_init);
1123
1124 static void __exit nx842_pseries_exit(void)
1125 {
1126         struct nx842_devdata *old_devdata;
1127         unsigned long flags;
1128
1129         crypto_unregister_alg(&nx842_pseries_alg);
1130
1131         spin_lock_irqsave(&devdata_mutex, flags);
1132         old_devdata = rcu_dereference_check(devdata,
1133                         lockdep_is_held(&devdata_mutex));
1134         RCU_INIT_POINTER(devdata, NULL);
1135         spin_unlock_irqrestore(&devdata_mutex, flags);
1136         synchronize_rcu();
1137         if (old_devdata && old_devdata->dev)
1138                 dev_set_drvdata(old_devdata->dev, NULL);
1139         kfree(old_devdata);
1140         vio_unregister_driver(&nx842_vio_driver);
1141 }
1142
1143 module_exit(nx842_pseries_exit);
1144