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