1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright 2008-2016 Freescale Semiconductor, Inc.
4 * Copyright 2017-2018 NXP Semiconductor
8 #include <fsl_ddr_sdram.h>
10 #include <asm/bitops.h>
14 #if defined(CONFIG_SYS_FSL_DDR3) || defined(CONFIG_SYS_FSL_DDR4)
16 compute_cas_latency(const unsigned int ctrl_num,
17 const dimm_params_t *dimm_params,
18 common_timing_params_t *outpdimm,
19 unsigned int number_of_dimms)
22 unsigned int common_caslat;
23 unsigned int caslat_actual;
24 unsigned int retry = 16;
25 unsigned int tmp = ~0;
26 const unsigned int mclk_ps = get_memory_clk_period_ps(ctrl_num);
27 #ifdef CONFIG_SYS_FSL_DDR3
28 const unsigned int taamax = 20000;
30 const unsigned int taamax = 18000;
33 /* compute the common CAS latency supported between slots */
34 for (i = 0; i < number_of_dimms; i++) {
35 if (dimm_params[i].n_ranks)
36 tmp &= dimm_params[i].caslat_x;
40 /* validate if the memory clk is in the range of dimms */
41 if (mclk_ps < outpdimm->tckmin_x_ps) {
42 printf("DDR clock (MCLK cycle %u ps) is faster than "
43 "the slowest DIMM(s) (tCKmin %u ps) can support.\n",
44 mclk_ps, outpdimm->tckmin_x_ps);
46 #ifdef CONFIG_SYS_FSL_DDR4
47 if (mclk_ps > outpdimm->tckmax_ps) {
48 printf("DDR clock (MCLK cycle %u ps) is slower than DIMM(s) (tCKmax %u ps) can support.\n",
49 mclk_ps, outpdimm->tckmax_ps);
52 /* determine the acutal cas latency */
53 caslat_actual = (outpdimm->taamin_ps + mclk_ps - 1) / mclk_ps;
54 /* check if the dimms support the CAS latency */
55 while (!(common_caslat & (1 << caslat_actual)) && retry > 0) {
59 /* once the caculation of caslat_actual is completed
60 * we must verify that this CAS latency value does not
61 * exceed tAAmax, which is 20 ns for all DDR3 speed grades,
62 * 18ns for all DDR4 speed grades.
64 if (caslat_actual * mclk_ps > taamax) {
65 printf("The chosen cas latency %d is too large\n",
68 outpdimm->lowest_common_spd_caslat = caslat_actual;
69 debug("lowest_common_spd_caslat is 0x%x\n", caslat_actual);
73 #else /* for DDR1 and DDR2 */
75 compute_cas_latency(const unsigned int ctrl_num,
76 const dimm_params_t *dimm_params,
77 common_timing_params_t *outpdimm,
78 unsigned int number_of_dimms)
81 const unsigned int mclk_ps = get_memory_clk_period_ps(ctrl_num);
82 unsigned int lowest_good_caslat;
84 unsigned int temp1, temp2;
86 debug("using mclk_ps = %u\n", mclk_ps);
87 if (mclk_ps > outpdimm->tckmax_ps) {
88 printf("Warning: DDR clock (%u ps) is slower than DIMM(s) (tCKmax %u ps)\n",
89 mclk_ps, outpdimm->tckmax_ps);
93 * Compute a CAS latency suitable for all DIMMs
95 * Strategy for SPD-defined latencies: compute only
96 * CAS latency defined by all DIMMs.
100 * Step 1: find CAS latency common to all DIMMs using bitwise
104 for (i = 0; i < number_of_dimms; i++) {
105 if (dimm_params[i].n_ranks) {
107 temp2 |= 1 << dimm_params[i].caslat_x;
108 temp2 |= 1 << dimm_params[i].caslat_x_minus_1;
109 temp2 |= 1 << dimm_params[i].caslat_x_minus_2;
111 * If there was no entry for X-2 (X-1) in
112 * the SPD, then caslat_x_minus_2
113 * (caslat_x_minus_1) contains either 255 or
114 * 0xFFFFFFFF because that's what the glorious
115 * __ilog2 function returns for an input of 0.
116 * On 32-bit PowerPC, left shift counts with bit
117 * 26 set (that the value of 255 or 0xFFFFFFFF
118 * will have), cause the destination register to
119 * be 0. That is why this works.
126 * Step 2: check each common CAS latency against tCK of each
129 lowest_good_caslat = 0;
133 temp2 = __ilog2(temp1);
134 debug("checking common caslat = %u\n", temp2);
136 /* Check if this CAS latency will work on all DIMMs at tCK. */
137 for (i = 0; i < number_of_dimms; i++) {
138 if (!dimm_params[i].n_ranks)
141 if (dimm_params[i].caslat_x == temp2) {
142 if (mclk_ps >= dimm_params[i].tckmin_x_ps) {
143 debug("CL = %u ok on DIMM %u at tCK=%u ps with tCKmin_X_ps of %u\n",
145 dimm_params[i].tckmin_x_ps);
152 if (dimm_params[i].caslat_x_minus_1 == temp2) {
153 unsigned int tckmin_x_minus_1_ps
154 = dimm_params[i].tckmin_x_minus_1_ps;
155 if (mclk_ps >= tckmin_x_minus_1_ps) {
156 debug("CL = %u ok on DIMM %u at tCK=%u ps with tckmin_x_minus_1_ps of %u\n",
158 tckmin_x_minus_1_ps);
165 if (dimm_params[i].caslat_x_minus_2 == temp2) {
166 unsigned int tckmin_x_minus_2_ps
167 = dimm_params[i].tckmin_x_minus_2_ps;
168 if (mclk_ps >= tckmin_x_minus_2_ps) {
169 debug("CL = %u ok on DIMM %u at tCK=%u ps with tckmin_x_minus_2_ps of %u\n",
171 tckmin_x_minus_2_ps);
180 lowest_good_caslat = temp2;
182 temp1 &= ~(1 << temp2);
185 debug("lowest common SPD-defined CAS latency = %u\n",
187 outpdimm->lowest_common_spd_caslat = lowest_good_caslat;
191 * Compute a common 'de-rated' CAS latency.
193 * The strategy here is to find the *highest* dereated cas latency
194 * with the assumption that all of the DIMMs will support a dereated
195 * CAS latency higher than or equal to their lowest dereated value.
198 for (i = 0; i < number_of_dimms; i++)
199 temp1 = max(temp1, dimm_params[i].caslat_lowest_derated);
201 outpdimm->highest_common_derated_caslat = temp1;
202 debug("highest common dereated CAS latency = %u\n", temp1);
209 * compute_lowest_common_dimm_parameters()
211 * Determine the worst-case DIMM timing parameters from the set of DIMMs
212 * whose parameters have been computed into the array pointed to
216 compute_lowest_common_dimm_parameters(const unsigned int ctrl_num,
217 const dimm_params_t *dimm_params,
218 common_timing_params_t *outpdimm,
219 const unsigned int number_of_dimms)
223 unsigned int tckmin_x_ps = 0;
224 unsigned int tckmax_ps = 0xFFFFFFFF;
225 unsigned int trcd_ps = 0;
226 unsigned int trp_ps = 0;
227 unsigned int tras_ps = 0;
228 #if defined(CONFIG_SYS_FSL_DDR3) || defined(CONFIG_SYS_FSL_DDR4)
229 unsigned int taamin_ps = 0;
231 #ifdef CONFIG_SYS_FSL_DDR4
232 unsigned int twr_ps = 15000;
233 unsigned int trfc1_ps = 0;
234 unsigned int trfc2_ps = 0;
235 unsigned int trfc4_ps = 0;
236 unsigned int trrds_ps = 0;
237 unsigned int trrdl_ps = 0;
238 unsigned int tccdl_ps = 0;
239 unsigned int trfc_slr_ps = 0;
241 unsigned int twr_ps = 0;
242 unsigned int twtr_ps = 0;
243 unsigned int trfc_ps = 0;
244 unsigned int trrd_ps = 0;
245 unsigned int trtp_ps = 0;
247 unsigned int trc_ps = 0;
248 unsigned int refresh_rate_ps = 0;
249 unsigned int extended_op_srt = 1;
250 #if defined(CONFIG_SYS_FSL_DDR1) || defined(CONFIG_SYS_FSL_DDR2)
251 unsigned int tis_ps = 0;
252 unsigned int tih_ps = 0;
253 unsigned int tds_ps = 0;
254 unsigned int tdh_ps = 0;
255 unsigned int tdqsq_max_ps = 0;
256 unsigned int tqhs_ps = 0;
258 unsigned int temp1, temp2;
259 unsigned int additive_latency = 0;
262 for (i = 0; i < number_of_dimms; i++) {
264 * If there are no ranks on this DIMM,
265 * it probably doesn't exist, so skip it.
267 if (dimm_params[i].n_ranks == 0) {
271 if (dimm_params[i].n_ranks == 4 && i != 0) {
272 printf("Found Quad-rank DIMM in wrong bank, ignored."
273 " Software may not run as expected.\n");
279 * check if quad-rank DIMM is plugged if
280 * CONFIG_CHIP_SELECT_QUAD_CAPABLE is not defined
281 * Only the board with proper design is capable
283 #ifndef CONFIG_FSL_DDR_FIRST_SLOT_QUAD_CAPABLE
284 if (dimm_params[i].n_ranks == 4 && \
285 CONFIG_CHIP_SELECTS_PER_CTRL/CONFIG_DIMM_SLOTS_PER_CTLR < 4) {
286 printf("Found Quad-rank DIMM, not able to support.");
292 * Find minimum tckmax_ps to find fastest slow speed,
293 * i.e., this is the slowest the whole system can go.
295 tckmax_ps = min(tckmax_ps,
296 (unsigned int)dimm_params[i].tckmax_ps);
297 #if defined(CONFIG_SYS_FSL_DDR3) || defined(CONFIG_SYS_FSL_DDR4)
298 taamin_ps = max(taamin_ps,
299 (unsigned int)dimm_params[i].taa_ps);
301 tckmin_x_ps = max(tckmin_x_ps,
302 (unsigned int)dimm_params[i].tckmin_x_ps);
303 trcd_ps = max(trcd_ps, (unsigned int)dimm_params[i].trcd_ps);
304 trp_ps = max(trp_ps, (unsigned int)dimm_params[i].trp_ps);
305 tras_ps = max(tras_ps, (unsigned int)dimm_params[i].tras_ps);
306 #ifdef CONFIG_SYS_FSL_DDR4
307 trfc1_ps = max(trfc1_ps,
308 (unsigned int)dimm_params[i].trfc1_ps);
309 trfc2_ps = max(trfc2_ps,
310 (unsigned int)dimm_params[i].trfc2_ps);
311 trfc4_ps = max(trfc4_ps,
312 (unsigned int)dimm_params[i].trfc4_ps);
313 trrds_ps = max(trrds_ps,
314 (unsigned int)dimm_params[i].trrds_ps);
315 trrdl_ps = max(trrdl_ps,
316 (unsigned int)dimm_params[i].trrdl_ps);
317 tccdl_ps = max(tccdl_ps,
318 (unsigned int)dimm_params[i].tccdl_ps);
319 trfc_slr_ps = max(trfc_slr_ps,
320 (unsigned int)dimm_params[i].trfc_slr_ps);
322 twr_ps = max(twr_ps, (unsigned int)dimm_params[i].twr_ps);
323 twtr_ps = max(twtr_ps, (unsigned int)dimm_params[i].twtr_ps);
324 trfc_ps = max(trfc_ps, (unsigned int)dimm_params[i].trfc_ps);
325 trrd_ps = max(trrd_ps, (unsigned int)dimm_params[i].trrd_ps);
326 trtp_ps = max(trtp_ps, (unsigned int)dimm_params[i].trtp_ps);
328 trc_ps = max(trc_ps, (unsigned int)dimm_params[i].trc_ps);
329 #if defined(CONFIG_SYS_FSL_DDR1) || defined(CONFIG_SYS_FSL_DDR2)
330 tis_ps = max(tis_ps, (unsigned int)dimm_params[i].tis_ps);
331 tih_ps = max(tih_ps, (unsigned int)dimm_params[i].tih_ps);
332 tds_ps = max(tds_ps, (unsigned int)dimm_params[i].tds_ps);
333 tdh_ps = max(tdh_ps, (unsigned int)dimm_params[i].tdh_ps);
334 tqhs_ps = max(tqhs_ps, (unsigned int)dimm_params[i].tqhs_ps);
336 * Find maximum tdqsq_max_ps to find slowest.
338 * FIXME: is finding the slowest value the correct
339 * strategy for this parameter?
341 tdqsq_max_ps = max(tdqsq_max_ps,
342 (unsigned int)dimm_params[i].tdqsq_max_ps);
344 refresh_rate_ps = max(refresh_rate_ps,
345 (unsigned int)dimm_params[i].refresh_rate_ps);
346 /* extended_op_srt is either 0 or 1, 0 having priority */
347 extended_op_srt = min(extended_op_srt,
348 (unsigned int)dimm_params[i].extended_op_srt);
351 outpdimm->ndimms_present = number_of_dimms - temp1;
353 if (temp1 == number_of_dimms) {
354 debug("no dimms this memory controller\n");
358 outpdimm->tckmin_x_ps = tckmin_x_ps;
359 outpdimm->tckmax_ps = tckmax_ps;
360 #if defined(CONFIG_SYS_FSL_DDR3) || defined(CONFIG_SYS_FSL_DDR4)
361 outpdimm->taamin_ps = taamin_ps;
363 outpdimm->trcd_ps = trcd_ps;
364 outpdimm->trp_ps = trp_ps;
365 outpdimm->tras_ps = tras_ps;
366 #ifdef CONFIG_SYS_FSL_DDR4
367 outpdimm->trfc1_ps = trfc1_ps;
368 outpdimm->trfc2_ps = trfc2_ps;
369 outpdimm->trfc4_ps = trfc4_ps;
370 outpdimm->trrds_ps = trrds_ps;
371 outpdimm->trrdl_ps = trrdl_ps;
372 outpdimm->tccdl_ps = tccdl_ps;
373 outpdimm->trfc_slr_ps = trfc_slr_ps;
375 outpdimm->twtr_ps = twtr_ps;
376 outpdimm->trfc_ps = trfc_ps;
377 outpdimm->trrd_ps = trrd_ps;
378 outpdimm->trtp_ps = trtp_ps;
380 outpdimm->twr_ps = twr_ps;
381 outpdimm->trc_ps = trc_ps;
382 outpdimm->refresh_rate_ps = refresh_rate_ps;
383 outpdimm->extended_op_srt = extended_op_srt;
384 #if defined(CONFIG_SYS_FSL_DDR1) || defined(CONFIG_SYS_FSL_DDR2)
385 outpdimm->tis_ps = tis_ps;
386 outpdimm->tih_ps = tih_ps;
387 outpdimm->tds_ps = tds_ps;
388 outpdimm->tdh_ps = tdh_ps;
389 outpdimm->tdqsq_max_ps = tdqsq_max_ps;
390 outpdimm->tqhs_ps = tqhs_ps;
393 /* Determine common burst length for all DIMMs. */
395 for (i = 0; i < number_of_dimms; i++) {
396 if (dimm_params[i].n_ranks) {
397 temp1 &= dimm_params[i].burst_lengths_bitmask;
400 outpdimm->all_dimms_burst_lengths_bitmask = temp1;
402 /* Determine if all DIMMs registered buffered. */
404 for (i = 0; i < number_of_dimms; i++) {
405 if (dimm_params[i].n_ranks) {
406 if (dimm_params[i].registered_dimm) {
408 #ifndef CONFIG_SPL_BUILD
409 printf("Detected RDIMM %s\n",
410 dimm_params[i].mpart);
414 #ifndef CONFIG_SPL_BUILD
415 printf("Detected UDIMM %s\n",
416 dimm_params[i].mpart);
422 outpdimm->all_dimms_registered = 0;
423 outpdimm->all_dimms_unbuffered = 0;
424 if (temp1 && !temp2) {
425 outpdimm->all_dimms_registered = 1;
426 } else if (!temp1 && temp2) {
427 outpdimm->all_dimms_unbuffered = 1;
429 printf("ERROR: Mix of registered buffered and unbuffered "
430 "DIMMs detected!\n");
434 if (outpdimm->all_dimms_registered)
435 for (j = 0; j < 16; j++) {
436 outpdimm->rcw[j] = dimm_params[0].rcw[j];
437 for (i = 1; i < number_of_dimms; i++) {
438 if (!dimm_params[i].n_ranks)
440 if (dimm_params[i].rcw[j] != dimm_params[0].rcw[j]) {
448 printf("ERROR: Mix different RDIMM detected!\n");
450 /* calculate cas latency for all DDR types */
451 if (compute_cas_latency(ctrl_num, dimm_params,
452 outpdimm, number_of_dimms))
455 /* Determine if all DIMMs ECC capable. */
457 for (i = 0; i < number_of_dimms; i++) {
458 if (dimm_params[i].n_ranks &&
459 !(dimm_params[i].edc_config & EDC_ECC)) {
465 debug("all DIMMs ECC capable\n");
467 debug("Warning: not all DIMMs ECC capable, cant enable ECC\n");
469 outpdimm->all_dimms_ecc_capable = temp1;
472 * Compute additive latency.
474 * For DDR1, additive latency should be 0.
476 * For DDR2, with ODT enabled, use "a value" less than ACTTORW,
477 * which comes from Trcd, and also note that:
478 * add_lat + caslat must be >= 4
480 * For DDR3, we use the AL=0
482 * When to use additive latency for DDR2:
484 * I. Because you are using CL=3 and need to do ODT on writes and
485 * want functionality.
486 * 1. Are you going to use ODT? (Does your board not have
487 * additional termination circuitry for DQ, DQS, DQS_,
488 * DM, RDQS, RDQS_ for x4/x8 configs?)
489 * 2. If so, is your lowest supported CL going to be 3?
490 * 3. If so, then you must set AL=1 because
492 * WL >= 3 for ODT on writes
501 * RL >= 3 for ODT on reads
504 * Since CL aren't usually less than 2, AL=0 is a minimum,
505 * so the WL-derived AL should be the -- FIXME?
507 * II. Because you are using auto-precharge globally and want to
508 * use additive latency (posted CAS) to get more bandwidth.
509 * 1. Are you going to use auto-precharge mode globally?
511 * Use addtivie latency and compute AL to be 1 cycle less than
512 * tRCD, i.e. the READ or WRITE command is in the cycle
513 * immediately following the ACTIVATE command..
515 * III. Because you feel like it or want to do some sort of
516 * degraded-performance experiment.
517 * 1. Do you just want to use additive latency because you feel
520 * Validation: AL is less than tRCD, and within the other
521 * read-to-precharge constraints.
524 additive_latency = 0;
526 #if defined(CONFIG_SYS_FSL_DDR2)
527 if ((outpdimm->lowest_common_spd_caslat < 4) &&
528 (picos_to_mclk(ctrl_num, trcd_ps) >
529 outpdimm->lowest_common_spd_caslat)) {
530 additive_latency = picos_to_mclk(ctrl_num, trcd_ps) -
531 outpdimm->lowest_common_spd_caslat;
532 if (mclk_to_picos(ctrl_num, additive_latency) > trcd_ps) {
533 additive_latency = picos_to_mclk(ctrl_num, trcd_ps);
534 debug("setting additive_latency to %u because it was "
535 " greater than tRCD_ps\n", additive_latency);
541 * Validate additive latency
545 if (mclk_to_picos(ctrl_num, additive_latency) > trcd_ps) {
546 printf("Error: invalid additive latency exceeds tRCD(min).\n");
551 * RL = CL + AL; RL >= 3 for ODT_RD_CFG to be enabled
552 * WL = RL - 1; WL >= 3 for ODT_WL_CFG to be enabled
553 * ADD_LAT (the register) must be set to a value less
554 * than ACTTORW if WL = 1, then AL must be set to 1
555 * RD_TO_PRE (the register) must be set to a minimum
556 * tRTP + AL if AL is nonzero
560 * Additive latency will be applied only if the memctl option to
563 outpdimm->additive_latency = additive_latency;
565 debug("tCKmin_ps = %u\n", outpdimm->tckmin_x_ps);
566 debug("trcd_ps = %u\n", outpdimm->trcd_ps);
567 debug("trp_ps = %u\n", outpdimm->trp_ps);
568 debug("tras_ps = %u\n", outpdimm->tras_ps);
569 #ifdef CONFIG_SYS_FSL_DDR4
570 debug("trfc1_ps = %u\n", trfc1_ps);
571 debug("trfc2_ps = %u\n", trfc2_ps);
572 debug("trfc4_ps = %u\n", trfc4_ps);
573 debug("trrds_ps = %u\n", trrds_ps);
574 debug("trrdl_ps = %u\n", trrdl_ps);
575 debug("tccdl_ps = %u\n", tccdl_ps);
576 debug("trfc_slr_ps = %u\n", trfc_slr_ps);
578 debug("twtr_ps = %u\n", outpdimm->twtr_ps);
579 debug("trfc_ps = %u\n", outpdimm->trfc_ps);
580 debug("trrd_ps = %u\n", outpdimm->trrd_ps);
582 debug("twr_ps = %u\n", outpdimm->twr_ps);
583 debug("trc_ps = %u\n", outpdimm->trc_ps);