arch/powerpc/cpu/mpc8xxx/ddr/lc_common_dimm_params.c

   1 /*
   2  * Copyright 2008 Freescale Semiconductor, Inc.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public License
   6  * Version 2 as published by the Free Software Foundation.
   7  */
   8
   9 #include <common.h>
  10 #include <asm/fsl_ddr_sdram.h>
  11
  12 #include "ddr.h"
  13
  14 unsigned int
  15 compute_cas_latency_ddr3(const dimm_params_t *dimm_params,
  16                          common_timing_params_t *outpdimm,
  17                          unsigned int number_of_dimms)
  18 {
  19         unsigned int i;
  20         unsigned int tAAmin_ps = 0;
  21         unsigned int tCKmin_X_ps = 0;
  22         unsigned int common_caslat;
  23         unsigned int caslat_actual;
  24         unsigned int retry = 16;
  25         unsigned int tmp;
  26         const unsigned int mclk_ps = get_memory_clk_period_ps();
  27
  28         /* compute the common CAS latency supported between slots */
  29         tmp = dimm_params[0].caslat_X;
  30         for (i = 1; i < number_of_dimms; i++)
  31                  tmp &= dimm_params[i].caslat_X;
  32         common_caslat = tmp;
  33
  34         /* compute the max tAAmin tCKmin between slots */
  35         for (i = 0; i < number_of_dimms; i++) {
  36                 tAAmin_ps = max(tAAmin_ps, dimm_params[i].tAA_ps);
  37                 tCKmin_X_ps = max(tCKmin_X_ps, dimm_params[i].tCKmin_X_ps);
  38         }
  39         /* validate if the memory clk is in the range of dimms */
  40         if (mclk_ps < tCKmin_X_ps) {
  41                 printf("The DIMM max tCKmin is %d ps,"
  42                         "doesn't support the MCLK cycle %d ps\n",
  43                         tCKmin_X_ps, mclk_ps);
  44                 return 1;
  45         }
  46         /* determine the acutal cas latency */
  47         caslat_actual = (tAAmin_ps + mclk_ps - 1) / mclk_ps;
  48         /* check if the dimms support the CAS latency */
  49         while (!(common_caslat & (1 << caslat_actual)) && retry > 0) {
  50                 caslat_actual++;
  51                 retry--;
  52         }
  53         /* once the caculation of caslat_actual is completed
  54          * we must verify that this CAS latency value does not
  55          * exceed tAAmax, which is 20 ns for all DDR3 speed grades
  56          */
  57         if (caslat_actual * mclk_ps > 20000) {
  58                 printf("The choosen cas latency %d is too large\n",
  59                         caslat_actual);
  60                 return 1;
  61         }
  62         outpdimm->lowest_common_SPD_caslat = caslat_actual;
  63
  64         return 0;
  65 }
  66
  67 /*
  68  * compute_lowest_common_dimm_parameters()
  69  *
  70  * Determine the worst-case DIMM timing parameters from the set of DIMMs
  71  * whose parameters have been computed into the array pointed to
  72  * by dimm_params.
  73  */
  74 unsigned int
  75 compute_lowest_common_dimm_parameters(const dimm_params_t *dimm_params,
  76                                       common_timing_params_t *outpdimm,
  77                                       unsigned int number_of_dimms)
  78 {
  79         unsigned int i, j;
  80
  81         unsigned int tCKmin_X_ps = 0;
  82         unsigned int tCKmax_ps = 0xFFFFFFFF;
  83         unsigned int tCKmax_max_ps = 0;
  84         unsigned int tRCD_ps = 0;
  85         unsigned int tRP_ps = 0;
  86         unsigned int tRAS_ps = 0;
  87         unsigned int tWR_ps = 0;
  88         unsigned int tWTR_ps = 0;
  89         unsigned int tRFC_ps = 0;
  90         unsigned int tRRD_ps = 0;
  91         unsigned int tRC_ps = 0;
  92         unsigned int refresh_rate_ps = 0;
  93         unsigned int tIS_ps = 0;
  94         unsigned int tIH_ps = 0;
  95         unsigned int tDS_ps = 0;
  96         unsigned int tDH_ps = 0;
  97         unsigned int tRTP_ps = 0;
  98         unsigned int tDQSQ_max_ps = 0;
  99         unsigned int tQHS_ps = 0;
 100
 101         unsigned int temp1, temp2, temp3;
 102         unsigned int additive_latency = 0;
 103 #if !defined(CONFIG_FSL_DDR3)
 104         const unsigned int mclk_ps = get_memory_clk_period_ps();
 105         unsigned int lowest_good_caslat;
 106         unsigned int not_ok;
 107
 108         debug("using mclk_ps = %u\n", mclk_ps);
 109 #endif
 110
 111         temp1 = 0;
 112         for (i = 0; i < number_of_dimms; i++) {
 113                 /*
 114                  * If there are no ranks on this DIMM,
 115                  * it probably doesn't exist, so skip it.
 116                  */
 117                 if (dimm_params[i].n_ranks == 0) {
 118                         temp1++;
 119                         continue;
 120                 }
 121                 if (dimm_params[i].n_ranks == 4 && i != 0) {
 122                         printf("Found Quad-rank DIMM in wrong bank, ignored."
 123                                 " Software may not run as expected.\n");
 124                         temp1++;
 125                         continue;
 126                 }
 127                 if (dimm_params[i].n_ranks == 4 && \
 128                   CONFIG_CHIP_SELECTS_PER_CTRL/CONFIG_DIMM_SLOTS_PER_CTLR < 4) {
 129                         printf("Found Quad-rank DIMM, not able to support.");
 130                         temp1++;
 131                         continue;
 132                 }
 133
 134                 /*
 135                  * Find minimum tCKmax_ps to find fastest slow speed,
 136                  * i.e., this is the slowest the whole system can go.
 137                  */
 138                 tCKmax_ps = min(tCKmax_ps, dimm_params[i].tCKmax_ps);
 139
 140                 /* Either find maximum value to determine slowest
 141                  * speed, delay, time, period, etc */
 142                 tCKmin_X_ps = max(tCKmin_X_ps, dimm_params[i].tCKmin_X_ps);
 143                 tCKmax_max_ps = max(tCKmax_max_ps, dimm_params[i].tCKmax_ps);
 144                 tRCD_ps = max(tRCD_ps, dimm_params[i].tRCD_ps);
 145                 tRP_ps = max(tRP_ps, dimm_params[i].tRP_ps);
 146                 tRAS_ps = max(tRAS_ps, dimm_params[i].tRAS_ps);
 147                 tWR_ps = max(tWR_ps, dimm_params[i].tWR_ps);
 148                 tWTR_ps = max(tWTR_ps, dimm_params[i].tWTR_ps);
 149                 tRFC_ps = max(tRFC_ps, dimm_params[i].tRFC_ps);
 150                 tRRD_ps = max(tRRD_ps, dimm_params[i].tRRD_ps);
 151                 tRC_ps = max(tRC_ps, dimm_params[i].tRC_ps);
 152                 tIS_ps = max(tIS_ps, dimm_params[i].tIS_ps);
 153                 tIH_ps = max(tIH_ps, dimm_params[i].tIH_ps);
 154                 tDS_ps = max(tDS_ps, dimm_params[i].tDS_ps);
 155                 tDH_ps = max(tDH_ps, dimm_params[i].tDH_ps);
 156                 tRTP_ps = max(tRTP_ps, dimm_params[i].tRTP_ps);
 157                 tQHS_ps = max(tQHS_ps, dimm_params[i].tQHS_ps);
 158                 refresh_rate_ps = max(refresh_rate_ps,
 159                                       dimm_params[i].refresh_rate_ps);
 160
 161                 /*
 162                  * Find maximum tDQSQ_max_ps to find slowest.
 163                  *
 164                  * FIXME: is finding the slowest value the correct
 165                  * strategy for this parameter?
 166                  */
 167                 tDQSQ_max_ps = max(tDQSQ_max_ps, dimm_params[i].tDQSQ_max_ps);
 168         }
 169
 170         outpdimm->ndimms_present = number_of_dimms - temp1;
 171
 172         if (temp1 == number_of_dimms) {
 173                 debug("no dimms this memory controller\n");
 174                 return 0;
 175         }
 176
 177         outpdimm->tCKmin_X_ps = tCKmin_X_ps;
 178         outpdimm->tCKmax_ps = tCKmax_ps;
 179         outpdimm->tCKmax_max_ps = tCKmax_max_ps;
 180         outpdimm->tRCD_ps = tRCD_ps;
 181         outpdimm->tRP_ps = tRP_ps;
 182         outpdimm->tRAS_ps = tRAS_ps;
 183         outpdimm->tWR_ps = tWR_ps;
 184         outpdimm->tWTR_ps = tWTR_ps;
 185         outpdimm->tRFC_ps = tRFC_ps;
 186         outpdimm->tRRD_ps = tRRD_ps;
 187         outpdimm->tRC_ps = tRC_ps;
 188         outpdimm->refresh_rate_ps = refresh_rate_ps;
 189         outpdimm->tIS_ps = tIS_ps;
 190         outpdimm->tIH_ps = tIH_ps;
 191         outpdimm->tDS_ps = tDS_ps;
 192         outpdimm->tDH_ps = tDH_ps;
 193         outpdimm->tRTP_ps = tRTP_ps;
 194         outpdimm->tDQSQ_max_ps = tDQSQ_max_ps;
 195         outpdimm->tQHS_ps = tQHS_ps;
 196
 197         /* Determine common burst length for all DIMMs. */
 198         temp1 = 0xff;
 199         for (i = 0; i < number_of_dimms; i++) {
 200                 if (dimm_params[i].n_ranks) {
 201                         temp1 &= dimm_params[i].burst_lengths_bitmask;
 202                 }
 203         }
 204         outpdimm->all_DIMMs_burst_lengths_bitmask = temp1;
 205
 206         /* Determine if all DIMMs registered buffered. */
 207         temp1 = temp2 = 0;
 208         for (i = 0; i < number_of_dimms; i++) {
 209                 if (dimm_params[i].n_ranks) {
 210                         if (dimm_params[i].registered_dimm)
 211                                 temp1 = 1;
 212                         if (!dimm_params[i].registered_dimm)
 213                                 temp2 = 1;
 214                 }
 215         }
 216
 217         outpdimm->all_DIMMs_registered = 0;
 218         if (temp1 && !temp2) {
 219                 outpdimm->all_DIMMs_registered = 1;
 220         }
 221
 222         outpdimm->all_DIMMs_unbuffered = 0;
 223         if (!temp1 && temp2) {
 224                 outpdimm->all_DIMMs_unbuffered = 1;
 225         }
 226
 227         /* CHECKME: */
 228         if (!outpdimm->all_DIMMs_registered
 229             && !outpdimm->all_DIMMs_unbuffered) {
 230                 printf("ERROR:  Mix of registered buffered and unbuffered "
 231                                 "DIMMs detected!\n");
 232         }
 233
 234         temp1 = 0;
 235         if (outpdimm->all_DIMMs_registered)
 236                 for (j = 0; j < 16; j++) {
 237                         outpdimm->rcw[j] = dimm_params[0].rcw[j];
 238                         for (i = 1; i < number_of_dimms; i++)
 239                                 if (dimm_params[i].rcw[j] != dimm_params[0].rcw[j]) {
 240                                         temp3 = 1;
 241                                         break;
 242                                 }
 243                 }
 244
 245         if (temp1 != 0)
 246                 printf("ERROR: Mix different RDIMM detected!\n");
 247
 248 #if defined(CONFIG_FSL_DDR3)
 249         if (compute_cas_latency_ddr3(dimm_params, outpdimm, number_of_dimms))
 250                 return 1;
 251 #else
 252         /*
 253          * Compute a CAS latency suitable for all DIMMs
 254          *
 255          * Strategy for SPD-defined latencies: compute only
 256          * CAS latency defined by all DIMMs.
 257          */
 258
 259         /*
 260          * Step 1: find CAS latency common to all DIMMs using bitwise
 261          * operation.
 262          */
 263         temp1 = 0xFF;
 264         for (i = 0; i < number_of_dimms; i++) {
 265                 if (dimm_params[i].n_ranks) {
 266                         temp2 = 0;
 267                         temp2 |= 1 << dimm_params[i].caslat_X;
 268                         temp2 |= 1 << dimm_params[i].caslat_X_minus_1;
 269                         temp2 |= 1 << dimm_params[i].caslat_X_minus_2;
 270                         /*
 271                          * FIXME: If there was no entry for X-2 (X-1) in
 272                          * the SPD, then caslat_X_minus_2
 273                          * (caslat_X_minus_1) contains either 255 or
 274                          * 0xFFFFFFFF because that's what the glorious
 275                          * __ilog2 function returns for an input of 0.
 276                          * On 32-bit PowerPC, left shift counts with bit
 277                          * 26 set (that the value of 255 or 0xFFFFFFFF
 278                          * will have), cause the destination register to
 279                          * be 0.  That is why this works.
 280                          */
 281                         temp1 &= temp2;
 282                 }
 283         }
 284
 285         /*
 286          * Step 2: check each common CAS latency against tCK of each
 287          * DIMM's SPD.
 288          */
 289         lowest_good_caslat = 0;
 290         temp2 = 0;
 291         while (temp1) {
 292                 not_ok = 0;
 293                 temp2 =  __ilog2(temp1);
 294                 debug("checking common caslat = %u\n", temp2);
 295
 296                 /* Check if this CAS latency will work on all DIMMs at tCK. */
 297                 for (i = 0; i < number_of_dimms; i++) {
 298                         if (!dimm_params[i].n_ranks) {
 299                                 continue;
 300                         }
 301                         if (dimm_params[i].caslat_X == temp2) {
 302                                 if (mclk_ps >= dimm_params[i].tCKmin_X_ps) {
 303                                         debug("CL = %u ok on DIMM %u at tCK=%u"
 304                                             " ps with its tCKmin_X_ps of %u\n",
 305                                                temp2, i, mclk_ps,
 306                                                dimm_params[i].tCKmin_X_ps);
 307                                         continue;
 308                                 } else {
 309                                         not_ok++;
 310                                 }
 311                         }
 312
 313                         if (dimm_params[i].caslat_X_minus_1 == temp2) {
 314                                 unsigned int tCKmin_X_minus_1_ps
 315                                         = dimm_params[i].tCKmin_X_minus_1_ps;
 316                                 if (mclk_ps >= tCKmin_X_minus_1_ps) {
 317                                         debug("CL = %u ok on DIMM %u at "
 318                                                 "tCK=%u ps with its "
 319                                                 "tCKmin_X_minus_1_ps of %u\n",
 320                                                temp2, i, mclk_ps,
 321                                                tCKmin_X_minus_1_ps);
 322                                         continue;
 323                                 } else {
 324                                         not_ok++;
 325                                 }
 326                         }
 327
 328                         if (dimm_params[i].caslat_X_minus_2 == temp2) {
 329                                 unsigned int tCKmin_X_minus_2_ps
 330                                         = dimm_params[i].tCKmin_X_minus_2_ps;
 331                                 if (mclk_ps >= tCKmin_X_minus_2_ps) {
 332                                         debug("CL = %u ok on DIMM %u at "
 333                                                 "tCK=%u ps with its "
 334                                                 "tCKmin_X_minus_2_ps of %u\n",
 335                                                temp2, i, mclk_ps,
 336                                                tCKmin_X_minus_2_ps);
 337                                         continue;
 338                                 } else {
 339                                         not_ok++;
 340                                 }
 341                         }
 342                 }
 343
 344                 if (!not_ok) {
 345                         lowest_good_caslat = temp2;
 346                 }
 347
 348                 temp1 &= ~(1 << temp2);
 349         }
 350
 351         debug("lowest common SPD-defined CAS latency = %u\n",
 352                lowest_good_caslat);
 353         outpdimm->lowest_common_SPD_caslat = lowest_good_caslat;
 354
 355
 356         /*
 357          * Compute a common 'de-rated' CAS latency.
 358          *
 359          * The strategy here is to find the *highest* dereated cas latency
 360          * with the assumption that all of the DIMMs will support a dereated
 361          * CAS latency higher than or equal to their lowest dereated value.
 362          */
 363         temp1 = 0;
 364         for (i = 0; i < number_of_dimms; i++) {
 365                 temp1 = max(temp1, dimm_params[i].caslat_lowest_derated);
 366         }
 367         outpdimm->highest_common_derated_caslat = temp1;
 368         debug("highest common dereated CAS latency = %u\n", temp1);
 369 #endif /* #if defined(CONFIG_FSL_DDR3) */
 370
 371         /* Determine if all DIMMs ECC capable. */
 372         temp1 = 1;
 373         for (i = 0; i < number_of_dimms; i++) {
 374                 if (dimm_params[i].n_ranks && dimm_params[i].edc_config != 2) {
 375                         temp1 = 0;
 376                         break;
 377                 }
 378         }
 379         if (temp1) {
 380                 debug("all DIMMs ECC capable\n");
 381         } else {
 382                 debug("Warning: not all DIMMs ECC capable, cant enable ECC\n");
 383         }
 384         outpdimm->all_DIMMs_ECC_capable = temp1;
 385
 386 #ifndef CONFIG_FSL_DDR3
 387         /* FIXME: move to somewhere else to validate. */
 388         if (mclk_ps > tCKmax_max_ps) {
 389                 printf("Warning: some of the installed DIMMs "
 390                                 "can not operate this slowly.\n");
 391                 return 1;
 392         }
 393 #endif
 394         /*
 395          * Compute additive latency.
 396          *
 397          * For DDR1, additive latency should be 0.
 398          *
 399          * For DDR2, with ODT enabled, use "a value" less than ACTTORW,
 400          *      which comes from Trcd, and also note that:
 401          *          add_lat + caslat must be >= 4
 402          *
 403          * For DDR3, we use the AL=0
 404          *
 405          * When to use additive latency for DDR2:
 406          *
 407          * I. Because you are using CL=3 and need to do ODT on writes and
 408          *    want functionality.
 409          *    1. Are you going to use ODT? (Does your board not have
 410          *      additional termination circuitry for DQ, DQS, DQS_,
 411          *      DM, RDQS, RDQS_ for x4/x8 configs?)
 412          *    2. If so, is your lowest supported CL going to be 3?
 413          *    3. If so, then you must set AL=1 because
 414          *
 415          *       WL >= 3 for ODT on writes
 416          *       RL = AL + CL
 417          *       WL = RL - 1
 418          *       ->
 419          *       WL = AL + CL - 1
 420          *       AL + CL - 1 >= 3
 421          *       AL + CL >= 4
 422          *  QED
 423          *
 424          *  RL >= 3 for ODT on reads
 425          *  RL = AL + CL
 426          *
 427          *  Since CL aren't usually less than 2, AL=0 is a minimum,
 428          *  so the WL-derived AL should be the  -- FIXME?
 429          *
 430          * II. Because you are using auto-precharge globally and want to
 431          *     use additive latency (posted CAS) to get more bandwidth.
 432          *     1. Are you going to use auto-precharge mode globally?
 433          *
 434          *        Use addtivie latency and compute AL to be 1 cycle less than
 435          *        tRCD, i.e. the READ or WRITE command is in the cycle
 436          *        immediately following the ACTIVATE command..
 437          *
 438          * III. Because you feel like it or want to do some sort of
 439          *      degraded-performance experiment.
 440          *     1.  Do you just want to use additive latency because you feel
 441          *         like it?
 442          *
 443          * Validation:  AL is less than tRCD, and within the other
 444          * read-to-precharge constraints.
 445          */
 446
 447         additive_latency = 0;
 448
 449 #if defined(CONFIG_FSL_DDR2)
 450         if (lowest_good_caslat < 4) {
 451                 additive_latency = picos_to_mclk(tRCD_ps) - lowest_good_caslat;
 452                 if (mclk_to_picos(additive_latency) > tRCD_ps) {
 453                         additive_latency = picos_to_mclk(tRCD_ps);
 454                         debug("setting additive_latency to %u because it was "
 455                                 " greater than tRCD_ps\n", additive_latency);
 456                 }
 457         }
 458
 459 #elif defined(CONFIG_FSL_DDR3)
 460         /*
 461          * The system will not use the global auto-precharge mode.
 462          * However, it uses the page mode, so we set AL=0
 463          */
 464         additive_latency = 0;
 465 #endif
 466
 467         /*
 468          * Validate additive latency
 469          * FIXME: move to somewhere else to validate
 470          *
 471          * AL <= tRCD(min)
 472          */
 473         if (mclk_to_picos(additive_latency) > tRCD_ps) {
 474                 printf("Error: invalid additive latency exceeds tRCD(min).\n");
 475                 return 1;
 476         }
 477
 478         /*
 479          * RL = CL + AL;  RL >= 3 for ODT_RD_CFG to be enabled
 480          * WL = RL - 1;  WL >= 3 for ODT_WL_CFG to be enabled
 481          * ADD_LAT (the register) must be set to a value less
 482          * than ACTTORW if WL = 1, then AL must be set to 1
 483          * RD_TO_PRE (the register) must be set to a minimum
 484          * tRTP + AL if AL is nonzero
 485          */
 486
 487         /*
 488          * Additive latency will be applied only if the memctl option to
 489          * use it.
 490          */
 491         outpdimm->additive_latency = additive_latency;
 492
 493         return 0;
 494 }