]> git.ipfire.org Git - people/ms/u-boot.git/blob - drivers/ddr/altera/sequencer.c
ddr: altera: Clean up scc_mgr_*_delay() args
[people/ms/u-boot.git] / drivers / ddr / altera / sequencer.c
1 /*
2 * Copyright Altera Corporation (C) 2012-2015
3 *
4 * SPDX-License-Identifier: BSD-3-Clause
5 */
6
7 #include <common.h>
8 #include <asm/io.h>
9 #include <asm/arch/sdram.h>
10 #include "sequencer.h"
11 #include "sequencer_auto.h"
12 #include "sequencer_auto_ac_init.h"
13 #include "sequencer_auto_inst_init.h"
14 #include "sequencer_defines.h"
15
16 static struct socfpga_sdr_rw_load_manager *sdr_rw_load_mgr_regs =
17 (struct socfpga_sdr_rw_load_manager *)(SDR_PHYGRP_RWMGRGRP_ADDRESS | 0x800);
18
19 static struct socfpga_sdr_rw_load_jump_manager *sdr_rw_load_jump_mgr_regs =
20 (struct socfpga_sdr_rw_load_jump_manager *)(SDR_PHYGRP_RWMGRGRP_ADDRESS | 0xC00);
21
22 static struct socfpga_sdr_reg_file *sdr_reg_file =
23 (struct socfpga_sdr_reg_file *)SDR_PHYGRP_REGFILEGRP_ADDRESS;
24
25 static struct socfpga_sdr_scc_mgr *sdr_scc_mgr =
26 (struct socfpga_sdr_scc_mgr *)(SDR_PHYGRP_SCCGRP_ADDRESS | 0xe00);
27
28 static struct socfpga_phy_mgr_cmd *phy_mgr_cmd =
29 (struct socfpga_phy_mgr_cmd *)SDR_PHYGRP_PHYMGRGRP_ADDRESS;
30
31 static struct socfpga_phy_mgr_cfg *phy_mgr_cfg =
32 (struct socfpga_phy_mgr_cfg *)(SDR_PHYGRP_PHYMGRGRP_ADDRESS | 0x40);
33
34 static struct socfpga_data_mgr *data_mgr =
35 (struct socfpga_data_mgr *)SDR_PHYGRP_DATAMGRGRP_ADDRESS;
36
37 static struct socfpga_sdr_ctrl *sdr_ctrl =
38 (struct socfpga_sdr_ctrl *)SDR_CTRLGRP_ADDRESS;
39
40 #define DELTA_D 1
41
42 /*
43 * In order to reduce ROM size, most of the selectable calibration steps are
44 * decided at compile time based on the user's calibration mode selection,
45 * as captured by the STATIC_CALIB_STEPS selection below.
46 *
47 * However, to support simulation-time selection of fast simulation mode, where
48 * we skip everything except the bare minimum, we need a few of the steps to
49 * be dynamic. In those cases, we either use the DYNAMIC_CALIB_STEPS for the
50 * check, which is based on the rtl-supplied value, or we dynamically compute
51 * the value to use based on the dynamically-chosen calibration mode
52 */
53
54 #define DLEVEL 0
55 #define STATIC_IN_RTL_SIM 0
56 #define STATIC_SKIP_DELAY_LOOPS 0
57
58 #define STATIC_CALIB_STEPS (STATIC_IN_RTL_SIM | CALIB_SKIP_FULL_TEST | \
59 STATIC_SKIP_DELAY_LOOPS)
60
61 /* calibration steps requested by the rtl */
62 uint16_t dyn_calib_steps;
63
64 /*
65 * To make CALIB_SKIP_DELAY_LOOPS a dynamic conditional option
66 * instead of static, we use boolean logic to select between
67 * non-skip and skip values
68 *
69 * The mask is set to include all bits when not-skipping, but is
70 * zero when skipping
71 */
72
73 uint16_t skip_delay_mask; /* mask off bits when skipping/not-skipping */
74
75 #define SKIP_DELAY_LOOP_VALUE_OR_ZERO(non_skip_value) \
76 ((non_skip_value) & skip_delay_mask)
77
78 struct gbl_type *gbl;
79 struct param_type *param;
80 uint32_t curr_shadow_reg;
81
82 static uint32_t rw_mgr_mem_calibrate_write_test(uint32_t rank_bgn,
83 uint32_t write_group, uint32_t use_dm,
84 uint32_t all_correct, uint32_t *bit_chk, uint32_t all_ranks);
85
86 static void set_failing_group_stage(uint32_t group, uint32_t stage,
87 uint32_t substage)
88 {
89 /*
90 * Only set the global stage if there was not been any other
91 * failing group
92 */
93 if (gbl->error_stage == CAL_STAGE_NIL) {
94 gbl->error_substage = substage;
95 gbl->error_stage = stage;
96 gbl->error_group = group;
97 }
98 }
99
100 static void reg_file_set_group(u16 set_group)
101 {
102 clrsetbits_le32(&sdr_reg_file->cur_stage, 0xffff0000, set_group << 16);
103 }
104
105 static void reg_file_set_stage(u8 set_stage)
106 {
107 clrsetbits_le32(&sdr_reg_file->cur_stage, 0xffff, set_stage & 0xff);
108 }
109
110 static void reg_file_set_sub_stage(u8 set_sub_stage)
111 {
112 set_sub_stage &= 0xff;
113 clrsetbits_le32(&sdr_reg_file->cur_stage, 0xff00, set_sub_stage << 8);
114 }
115
116 static void initialize(void)
117 {
118 debug("%s:%d\n", __func__, __LINE__);
119 /* USER calibration has control over path to memory */
120 /*
121 * In Hard PHY this is a 2-bit control:
122 * 0: AFI Mux Select
123 * 1: DDIO Mux Select
124 */
125 writel(0x3, &phy_mgr_cfg->mux_sel);
126
127 /* USER memory clock is not stable we begin initialization */
128 writel(0, &phy_mgr_cfg->reset_mem_stbl);
129
130 /* USER calibration status all set to zero */
131 writel(0, &phy_mgr_cfg->cal_status);
132
133 writel(0, &phy_mgr_cfg->cal_debug_info);
134
135 if ((dyn_calib_steps & CALIB_SKIP_ALL) != CALIB_SKIP_ALL) {
136 param->read_correct_mask_vg = ((uint32_t)1 <<
137 (RW_MGR_MEM_DQ_PER_READ_DQS /
138 RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS)) - 1;
139 param->write_correct_mask_vg = ((uint32_t)1 <<
140 (RW_MGR_MEM_DQ_PER_READ_DQS /
141 RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS)) - 1;
142 param->read_correct_mask = ((uint32_t)1 <<
143 RW_MGR_MEM_DQ_PER_READ_DQS) - 1;
144 param->write_correct_mask = ((uint32_t)1 <<
145 RW_MGR_MEM_DQ_PER_WRITE_DQS) - 1;
146 param->dm_correct_mask = ((uint32_t)1 <<
147 (RW_MGR_MEM_DATA_WIDTH / RW_MGR_MEM_DATA_MASK_WIDTH))
148 - 1;
149 }
150 }
151
152 static void set_rank_and_odt_mask(uint32_t rank, uint32_t odt_mode)
153 {
154 uint32_t odt_mask_0 = 0;
155 uint32_t odt_mask_1 = 0;
156 uint32_t cs_and_odt_mask;
157
158 if (odt_mode == RW_MGR_ODT_MODE_READ_WRITE) {
159 if (RW_MGR_MEM_NUMBER_OF_RANKS == 1) {
160 /*
161 * 1 Rank
162 * Read: ODT = 0
163 * Write: ODT = 1
164 */
165 odt_mask_0 = 0x0;
166 odt_mask_1 = 0x1;
167 } else if (RW_MGR_MEM_NUMBER_OF_RANKS == 2) {
168 /* 2 Ranks */
169 if (RW_MGR_MEM_NUMBER_OF_CS_PER_DIMM == 1) {
170 /* - Dual-Slot , Single-Rank
171 * (1 chip-select per DIMM)
172 * OR
173 * - RDIMM, 4 total CS (2 CS per DIMM)
174 * means 2 DIMM
175 * Since MEM_NUMBER_OF_RANKS is 2 they are
176 * both single rank
177 * with 2 CS each (special for RDIMM)
178 * Read: Turn on ODT on the opposite rank
179 * Write: Turn on ODT on all ranks
180 */
181 odt_mask_0 = 0x3 & ~(1 << rank);
182 odt_mask_1 = 0x3;
183 } else {
184 /*
185 * USER - Single-Slot , Dual-rank DIMMs
186 * (2 chip-selects per DIMM)
187 * USER Read: Turn on ODT off on all ranks
188 * USER Write: Turn on ODT on active rank
189 */
190 odt_mask_0 = 0x0;
191 odt_mask_1 = 0x3 & (1 << rank);
192 }
193 } else {
194 /* 4 Ranks
195 * Read:
196 * ----------+-----------------------+
197 * | |
198 * | ODT |
199 * Read From +-----------------------+
200 * Rank | 3 | 2 | 1 | 0 |
201 * ----------+-----+-----+-----+-----+
202 * 0 | 0 | 1 | 0 | 0 |
203 * 1 | 1 | 0 | 0 | 0 |
204 * 2 | 0 | 0 | 0 | 1 |
205 * 3 | 0 | 0 | 1 | 0 |
206 * ----------+-----+-----+-----+-----+
207 *
208 * Write:
209 * ----------+-----------------------+
210 * | |
211 * | ODT |
212 * Write To +-----------------------+
213 * Rank | 3 | 2 | 1 | 0 |
214 * ----------+-----+-----+-----+-----+
215 * 0 | 0 | 1 | 0 | 1 |
216 * 1 | 1 | 0 | 1 | 0 |
217 * 2 | 0 | 1 | 0 | 1 |
218 * 3 | 1 | 0 | 1 | 0 |
219 * ----------+-----+-----+-----+-----+
220 */
221 switch (rank) {
222 case 0:
223 odt_mask_0 = 0x4;
224 odt_mask_1 = 0x5;
225 break;
226 case 1:
227 odt_mask_0 = 0x8;
228 odt_mask_1 = 0xA;
229 break;
230 case 2:
231 odt_mask_0 = 0x1;
232 odt_mask_1 = 0x5;
233 break;
234 case 3:
235 odt_mask_0 = 0x2;
236 odt_mask_1 = 0xA;
237 break;
238 }
239 }
240 } else {
241 odt_mask_0 = 0x0;
242 odt_mask_1 = 0x0;
243 }
244
245 cs_and_odt_mask =
246 (0xFF & ~(1 << rank)) |
247 ((0xFF & odt_mask_0) << 8) |
248 ((0xFF & odt_mask_1) << 16);
249 writel(cs_and_odt_mask, SDR_PHYGRP_RWMGRGRP_ADDRESS |
250 RW_MGR_SET_CS_AND_ODT_MASK_OFFSET);
251 }
252
253 /**
254 * scc_mgr_set() - Set SCC Manager register
255 * @off: Base offset in SCC Manager space
256 * @grp: Read/Write group
257 * @val: Value to be set
258 *
259 * This function sets the SCC Manager (Scan Chain Control Manager) register.
260 */
261 static void scc_mgr_set(u32 off, u32 grp, u32 val)
262 {
263 writel(val, SDR_PHYGRP_SCCGRP_ADDRESS | off | (grp << 2));
264 }
265
266 /**
267 * scc_mgr_initialize() - Initialize SCC Manager registers
268 *
269 * Initialize SCC Manager registers.
270 */
271 static void scc_mgr_initialize(void)
272 {
273 /*
274 * Clear register file for HPS. 16 (2^4) is the size of the
275 * full register file in the scc mgr:
276 * RFILE_DEPTH = 1 + log2(MEM_DQ_PER_DQS + 1 + MEM_DM_PER_DQS +
277 * MEM_IF_READ_DQS_WIDTH - 1);
278 */
279 int i;
280
281 for (i = 0; i < 16; i++) {
282 debug_cond(DLEVEL == 1, "%s:%d: Clearing SCC RFILE index %u\n",
283 __func__, __LINE__, i);
284 scc_mgr_set(SCC_MGR_HHP_RFILE_OFFSET, 0, i);
285 }
286 }
287
288 static void scc_mgr_set_dqdqs_output_phase(uint32_t write_group, uint32_t phase)
289 {
290 scc_mgr_set(SCC_MGR_DQDQS_OUT_PHASE_OFFSET, write_group, phase);
291 }
292
293 static void scc_mgr_set_dqs_bus_in_delay(uint32_t read_group, uint32_t delay)
294 {
295 scc_mgr_set(SCC_MGR_DQS_IN_DELAY_OFFSET, read_group, delay);
296 }
297
298 static void scc_mgr_set_dqs_en_phase(uint32_t read_group, uint32_t phase)
299 {
300 scc_mgr_set(SCC_MGR_DQS_EN_PHASE_OFFSET, read_group, phase);
301 }
302
303 static void scc_mgr_set_dqs_en_delay(uint32_t read_group, uint32_t delay)
304 {
305 scc_mgr_set(SCC_MGR_DQS_EN_DELAY_OFFSET, read_group, delay);
306 }
307
308 static void scc_mgr_set_dqs_io_in_delay(uint32_t delay)
309 {
310 scc_mgr_set(SCC_MGR_IO_IN_DELAY_OFFSET, RW_MGR_MEM_DQ_PER_WRITE_DQS,
311 delay);
312 }
313
314 static void scc_mgr_set_dq_in_delay(uint32_t dq_in_group, uint32_t delay)
315 {
316 scc_mgr_set(SCC_MGR_IO_IN_DELAY_OFFSET, dq_in_group, delay);
317 }
318
319 static void scc_mgr_set_dq_out1_delay(uint32_t dq_in_group, uint32_t delay)
320 {
321 scc_mgr_set(SCC_MGR_IO_OUT1_DELAY_OFFSET, dq_in_group, delay);
322 }
323
324 static void scc_mgr_set_dqs_out1_delay(uint32_t delay)
325 {
326 scc_mgr_set(SCC_MGR_IO_OUT1_DELAY_OFFSET, RW_MGR_MEM_DQ_PER_WRITE_DQS,
327 delay);
328 }
329
330 static void scc_mgr_set_dm_out1_delay(uint32_t dm, uint32_t delay)
331 {
332 scc_mgr_set(SCC_MGR_IO_OUT1_DELAY_OFFSET,
333 RW_MGR_MEM_DQ_PER_WRITE_DQS + 1 + dm,
334 delay);
335 }
336
337 /* load up dqs config settings */
338 static void scc_mgr_load_dqs(uint32_t dqs)
339 {
340 writel(dqs, &sdr_scc_mgr->dqs_ena);
341 }
342
343 /* load up dqs io config settings */
344 static void scc_mgr_load_dqs_io(void)
345 {
346 writel(0, &sdr_scc_mgr->dqs_io_ena);
347 }
348
349 /* load up dq config settings */
350 static void scc_mgr_load_dq(uint32_t dq_in_group)
351 {
352 writel(dq_in_group, &sdr_scc_mgr->dq_ena);
353 }
354
355 /* load up dm config settings */
356 static void scc_mgr_load_dm(uint32_t dm)
357 {
358 writel(dm, &sdr_scc_mgr->dm_ena);
359 }
360
361 /**
362 * scc_mgr_set_all_ranks() - Set SCC Manager register for all ranks
363 * @off: Base offset in SCC Manager space
364 * @grp: Read/Write group
365 * @val: Value to be set
366 * @update: If non-zero, trigger SCC Manager update for all ranks
367 *
368 * This function sets the SCC Manager (Scan Chain Control Manager) register
369 * and optionally triggers the SCC update for all ranks.
370 */
371 static void scc_mgr_set_all_ranks(const u32 off, const u32 grp, const u32 val,
372 const int update)
373 {
374 u32 r;
375
376 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
377 r += NUM_RANKS_PER_SHADOW_REG) {
378 scc_mgr_set(off, grp, val);
379
380 if (update || (r == 0)) {
381 writel(grp, &sdr_scc_mgr->dqs_ena);
382 writel(0, &sdr_scc_mgr->update);
383 }
384 }
385 }
386
387 static void scc_mgr_set_dqs_en_phase_all_ranks(u32 read_group, u32 phase)
388 {
389 /*
390 * USER although the h/w doesn't support different phases per
391 * shadow register, for simplicity our scc manager modeling
392 * keeps different phase settings per shadow reg, and it's
393 * important for us to keep them in sync to match h/w.
394 * for efficiency, the scan chain update should occur only
395 * once to sr0.
396 */
397 scc_mgr_set_all_ranks(SCC_MGR_DQS_EN_PHASE_OFFSET,
398 read_group, phase, 0);
399 }
400
401 static void scc_mgr_set_dqdqs_output_phase_all_ranks(uint32_t write_group,
402 uint32_t phase)
403 {
404 /*
405 * USER although the h/w doesn't support different phases per
406 * shadow register, for simplicity our scc manager modeling
407 * keeps different phase settings per shadow reg, and it's
408 * important for us to keep them in sync to match h/w.
409 * for efficiency, the scan chain update should occur only
410 * once to sr0.
411 */
412 scc_mgr_set_all_ranks(SCC_MGR_DQDQS_OUT_PHASE_OFFSET,
413 write_group, phase, 0);
414 }
415
416 static void scc_mgr_set_dqs_en_delay_all_ranks(uint32_t read_group,
417 uint32_t delay)
418 {
419 /*
420 * In shadow register mode, the T11 settings are stored in
421 * registers in the core, which are updated by the DQS_ENA
422 * signals. Not issuing the SCC_MGR_UPD command allows us to
423 * save lots of rank switching overhead, by calling
424 * select_shadow_regs_for_update with update_scan_chains
425 * set to 0.
426 */
427 scc_mgr_set_all_ranks(SCC_MGR_DQS_EN_DELAY_OFFSET,
428 read_group, delay, 1);
429 writel(0, &sdr_scc_mgr->update);
430 }
431
432 /**
433 * scc_mgr_set_oct_out1_delay() - Set OCT output delay
434 * @write_group: Write group
435 * @delay: Delay value
436 *
437 * This function sets the OCT output delay in SCC manager.
438 */
439 static void scc_mgr_set_oct_out1_delay(const u32 write_group, const u32 delay)
440 {
441 const int ratio = RW_MGR_MEM_IF_READ_DQS_WIDTH /
442 RW_MGR_MEM_IF_WRITE_DQS_WIDTH;
443 const int base = write_group * ratio;
444 int i;
445 /*
446 * Load the setting in the SCC manager
447 * Although OCT affects only write data, the OCT delay is controlled
448 * by the DQS logic block which is instantiated once per read group.
449 * For protocols where a write group consists of multiple read groups,
450 * the setting must be set multiple times.
451 */
452 for (i = 0; i < ratio; i++)
453 scc_mgr_set(SCC_MGR_OCT_OUT1_DELAY_OFFSET, base + i, delay);
454 }
455
456 static void scc_mgr_set_hhp_extras(void)
457 {
458 /*
459 * Load the fixed setting in the SCC manager
460 * bits: 0:0 = 1'b1 - dqs bypass
461 * bits: 1:1 = 1'b1 - dq bypass
462 * bits: 4:2 = 3'b001 - rfifo_mode
463 * bits: 6:5 = 2'b01 - rfifo clock_select
464 * bits: 7:7 = 1'b0 - separate gating from ungating setting
465 * bits: 8:8 = 1'b0 - separate OE from Output delay setting
466 */
467 uint32_t value = (0<<8) | (0<<7) | (1<<5) | (1<<2) | (1<<1) | (1<<0);
468 uint32_t addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_HHP_GLOBALS_OFFSET;
469
470 writel(value, addr + SCC_MGR_HHP_EXTRAS_OFFSET);
471 }
472
473 /*
474 * USER Zero all DQS config
475 * TODO: maybe rename to scc_mgr_zero_dqs_config (or something)
476 */
477 static void scc_mgr_zero_all(void)
478 {
479 uint32_t i, r;
480
481 /*
482 * USER Zero all DQS config settings, across all groups and all
483 * shadow registers
484 */
485 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS; r +=
486 NUM_RANKS_PER_SHADOW_REG) {
487 for (i = 0; i < RW_MGR_MEM_IF_READ_DQS_WIDTH; i++) {
488 /*
489 * The phases actually don't exist on a per-rank basis,
490 * but there's no harm updating them several times, so
491 * let's keep the code simple.
492 */
493 scc_mgr_set_dqs_bus_in_delay(i, IO_DQS_IN_RESERVE);
494 scc_mgr_set_dqs_en_phase(i, 0);
495 scc_mgr_set_dqs_en_delay(i, 0);
496 }
497
498 for (i = 0; i < RW_MGR_MEM_IF_WRITE_DQS_WIDTH; i++) {
499 scc_mgr_set_dqdqs_output_phase(i, 0);
500 /* av/cv don't have out2 */
501 scc_mgr_set_oct_out1_delay(i, IO_DQS_OUT_RESERVE);
502 }
503 }
504
505 /* multicast to all DQS group enables */
506 writel(0xff, &sdr_scc_mgr->dqs_ena);
507 writel(0, &sdr_scc_mgr->update);
508 }
509
510 /**
511 * scc_set_bypass_mode() - Set bypass mode and trigger SCC update
512 * @write_group: Write group
513 *
514 * Set bypass mode and trigger SCC update.
515 */
516 static void scc_set_bypass_mode(const u32 write_group)
517 {
518 /* Only needed once to set all groups, pins, DQ, DQS, DM. */
519 if (write_group == 0) {
520 debug_cond(DLEVEL == 1, "%s:%d Setting HHP Extras\n", __func__,
521 __LINE__);
522 scc_mgr_set_hhp_extras();
523 debug_cond(DLEVEL == 1, "%s:%d Done Setting HHP Extras\n",
524 __func__, __LINE__);
525 }
526
527 /* Multicast to all DQ enables. */
528 writel(0xff, &sdr_scc_mgr->dq_ena);
529 writel(0xff, &sdr_scc_mgr->dm_ena);
530
531 /* Update current DQS IO enable. */
532 writel(0, &sdr_scc_mgr->dqs_io_ena);
533
534 /* Update the DQS logic. */
535 writel(write_group, &sdr_scc_mgr->dqs_ena);
536
537 /* Hit update. */
538 writel(0, &sdr_scc_mgr->update);
539 }
540
541 /**
542 * scc_mgr_load_dqs_for_write_group() - Load DQS settings for Write Group
543 * @write_group: Write group
544 *
545 * Load DQS settings for Write Group, do not trigger SCC update.
546 */
547 static void scc_mgr_load_dqs_for_write_group(const u32 write_group)
548 {
549 const int ratio = RW_MGR_MEM_IF_READ_DQS_WIDTH /
550 RW_MGR_MEM_IF_WRITE_DQS_WIDTH;
551 const int base = write_group * ratio;
552 int i;
553 /*
554 * Load the setting in the SCC manager
555 * Although OCT affects only write data, the OCT delay is controlled
556 * by the DQS logic block which is instantiated once per read group.
557 * For protocols where a write group consists of multiple read groups,
558 * the setting must be set multiple times.
559 */
560 for (i = 0; i < ratio; i++)
561 writel(base + i, &sdr_scc_mgr->dqs_ena);
562 }
563
564 static void scc_mgr_zero_group(uint32_t write_group, uint32_t test_begin,
565 int32_t out_only)
566 {
567 uint32_t i, r;
568
569 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS; r +=
570 NUM_RANKS_PER_SHADOW_REG) {
571 /* Zero all DQ config settings */
572 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
573 scc_mgr_set_dq_out1_delay(i, 0);
574 if (!out_only)
575 scc_mgr_set_dq_in_delay(i, 0);
576 }
577
578 /* multicast to all DQ enables */
579 writel(0xff, &sdr_scc_mgr->dq_ena);
580
581 /* Zero all DM config settings */
582 for (i = 0; i < RW_MGR_NUM_DM_PER_WRITE_GROUP; i++) {
583 scc_mgr_set_dm_out1_delay(i, 0);
584 }
585
586 /* multicast to all DM enables */
587 writel(0xff, &sdr_scc_mgr->dm_ena);
588
589 /* zero all DQS io settings */
590 if (!out_only)
591 scc_mgr_set_dqs_io_in_delay(0);
592 /* av/cv don't have out2 */
593 scc_mgr_set_dqs_out1_delay(IO_DQS_OUT_RESERVE);
594 scc_mgr_set_oct_out1_delay(write_group, IO_DQS_OUT_RESERVE);
595 scc_mgr_load_dqs_for_write_group(write_group);
596
597 /* multicast to all DQS IO enables (only 1) */
598 writel(0, &sdr_scc_mgr->dqs_io_ena);
599
600 /* hit update to zero everything */
601 writel(0, &sdr_scc_mgr->update);
602 }
603 }
604
605 /*
606 * apply and load a particular input delay for the DQ pins in a group
607 * group_bgn is the index of the first dq pin (in the write group)
608 */
609 static void scc_mgr_apply_group_dq_in_delay(uint32_t group_bgn, uint32_t delay)
610 {
611 uint32_t i, p;
612
613 for (i = 0, p = group_bgn; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++, p++) {
614 scc_mgr_set_dq_in_delay(p, delay);
615 scc_mgr_load_dq(p);
616 }
617 }
618
619 /**
620 * scc_mgr_apply_group_dq_out1_delay() - Apply and load an output delay for the DQ pins in a group
621 * @delay: Delay value
622 *
623 * Apply and load a particular output delay for the DQ pins in a group.
624 */
625 static void scc_mgr_apply_group_dq_out1_delay(const u32 delay)
626 {
627 int i;
628
629 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
630 scc_mgr_set_dq_out1_delay(i, delay);
631 scc_mgr_load_dq(i);
632 }
633 }
634
635 /* apply and load a particular output delay for the DM pins in a group */
636 static void scc_mgr_apply_group_dm_out1_delay(uint32_t delay1)
637 {
638 uint32_t i;
639
640 for (i = 0; i < RW_MGR_NUM_DM_PER_WRITE_GROUP; i++) {
641 scc_mgr_set_dm_out1_delay(i, delay1);
642 scc_mgr_load_dm(i);
643 }
644 }
645
646
647 /* apply and load delay on both DQS and OCT out1 */
648 static void scc_mgr_apply_group_dqs_io_and_oct_out1(uint32_t write_group,
649 uint32_t delay)
650 {
651 scc_mgr_set_dqs_out1_delay(delay);
652 scc_mgr_load_dqs_io();
653
654 scc_mgr_set_oct_out1_delay(write_group, delay);
655 scc_mgr_load_dqs_for_write_group(write_group);
656 }
657
658 /* apply a delay to the entire output side: DQ, DM, DQS, OCT */
659 static void scc_mgr_apply_group_all_out_delay_add(uint32_t write_group,
660 uint32_t group_bgn,
661 uint32_t delay)
662 {
663 uint32_t i, p, new_delay;
664
665 /* dq shift */
666 for (i = 0, p = group_bgn; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++, p++) {
667 new_delay = READ_SCC_DQ_OUT2_DELAY;
668 new_delay += delay;
669
670 if (new_delay > IO_IO_OUT2_DELAY_MAX) {
671 debug_cond(DLEVEL == 1, "%s:%d (%u, %u, %u) DQ[%u,%u]:\
672 %u > %lu => %lu", __func__, __LINE__,
673 write_group, group_bgn, delay, i, p, new_delay,
674 (long unsigned int)IO_IO_OUT2_DELAY_MAX,
675 (long unsigned int)IO_IO_OUT2_DELAY_MAX);
676 new_delay = IO_IO_OUT2_DELAY_MAX;
677 }
678
679 scc_mgr_load_dq(i);
680 }
681
682 /* dm shift */
683 for (i = 0; i < RW_MGR_NUM_DM_PER_WRITE_GROUP; i++) {
684 new_delay = READ_SCC_DM_IO_OUT2_DELAY;
685 new_delay += delay;
686
687 if (new_delay > IO_IO_OUT2_DELAY_MAX) {
688 debug_cond(DLEVEL == 1, "%s:%d (%u, %u, %u) DM[%u]:\
689 %u > %lu => %lu\n", __func__, __LINE__,
690 write_group, group_bgn, delay, i, new_delay,
691 (long unsigned int)IO_IO_OUT2_DELAY_MAX,
692 (long unsigned int)IO_IO_OUT2_DELAY_MAX);
693 new_delay = IO_IO_OUT2_DELAY_MAX;
694 }
695
696 scc_mgr_load_dm(i);
697 }
698
699 /* dqs shift */
700 new_delay = READ_SCC_DQS_IO_OUT2_DELAY;
701 new_delay += delay;
702
703 if (new_delay > IO_IO_OUT2_DELAY_MAX) {
704 debug_cond(DLEVEL == 1, "%s:%d (%u, %u, %u) DQS: %u > %d => %d;"
705 " adding %u to OUT1\n", __func__, __LINE__,
706 write_group, group_bgn, delay, new_delay,
707 IO_IO_OUT2_DELAY_MAX, IO_IO_OUT2_DELAY_MAX,
708 new_delay - IO_IO_OUT2_DELAY_MAX);
709 scc_mgr_set_dqs_out1_delay(new_delay -
710 IO_IO_OUT2_DELAY_MAX);
711 new_delay = IO_IO_OUT2_DELAY_MAX;
712 }
713
714 scc_mgr_load_dqs_io();
715
716 /* oct shift */
717 new_delay = READ_SCC_OCT_OUT2_DELAY;
718 new_delay += delay;
719
720 if (new_delay > IO_IO_OUT2_DELAY_MAX) {
721 debug_cond(DLEVEL == 1, "%s:%d (%u, %u, %u) DQS: %u > %d => %d;"
722 " adding %u to OUT1\n", __func__, __LINE__,
723 write_group, group_bgn, delay, new_delay,
724 IO_IO_OUT2_DELAY_MAX, IO_IO_OUT2_DELAY_MAX,
725 new_delay - IO_IO_OUT2_DELAY_MAX);
726 scc_mgr_set_oct_out1_delay(write_group, new_delay -
727 IO_IO_OUT2_DELAY_MAX);
728 new_delay = IO_IO_OUT2_DELAY_MAX;
729 }
730
731 scc_mgr_load_dqs_for_write_group(write_group);
732 }
733
734 /*
735 * USER apply a delay to the entire output side (DQ, DM, DQS, OCT)
736 * and to all ranks
737 */
738 static void scc_mgr_apply_group_all_out_delay_add_all_ranks(
739 uint32_t write_group, uint32_t group_bgn, uint32_t delay)
740 {
741 uint32_t r;
742
743 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
744 r += NUM_RANKS_PER_SHADOW_REG) {
745 scc_mgr_apply_group_all_out_delay_add(write_group,
746 group_bgn, delay);
747 writel(0, &sdr_scc_mgr->update);
748 }
749 }
750
751 /* optimization used to recover some slots in ddr3 inst_rom */
752 /* could be applied to other protocols if we wanted to */
753 static void set_jump_as_return(void)
754 {
755 /*
756 * to save space, we replace return with jump to special shared
757 * RETURN instruction so we set the counter to large value so that
758 * we always jump
759 */
760 writel(0xff, &sdr_rw_load_mgr_regs->load_cntr0);
761 writel(RW_MGR_RETURN, &sdr_rw_load_jump_mgr_regs->load_jump_add0);
762 }
763
764 /*
765 * should always use constants as argument to ensure all computations are
766 * performed at compile time
767 */
768 static void delay_for_n_mem_clocks(const uint32_t clocks)
769 {
770 uint32_t afi_clocks;
771 uint8_t inner = 0;
772 uint8_t outer = 0;
773 uint16_t c_loop = 0;
774
775 debug("%s:%d: clocks=%u ... start\n", __func__, __LINE__, clocks);
776
777
778 afi_clocks = (clocks + AFI_RATE_RATIO-1) / AFI_RATE_RATIO;
779 /* scale (rounding up) to get afi clocks */
780
781 /*
782 * Note, we don't bother accounting for being off a little bit
783 * because of a few extra instructions in outer loops
784 * Note, the loops have a test at the end, and do the test before
785 * the decrement, and so always perform the loop
786 * 1 time more than the counter value
787 */
788 if (afi_clocks == 0) {
789 ;
790 } else if (afi_clocks <= 0x100) {
791 inner = afi_clocks-1;
792 outer = 0;
793 c_loop = 0;
794 } else if (afi_clocks <= 0x10000) {
795 inner = 0xff;
796 outer = (afi_clocks-1) >> 8;
797 c_loop = 0;
798 } else {
799 inner = 0xff;
800 outer = 0xff;
801 c_loop = (afi_clocks-1) >> 16;
802 }
803
804 /*
805 * rom instructions are structured as follows:
806 *
807 * IDLE_LOOP2: jnz cntr0, TARGET_A
808 * IDLE_LOOP1: jnz cntr1, TARGET_B
809 * return
810 *
811 * so, when doing nested loops, TARGET_A is set to IDLE_LOOP2, and
812 * TARGET_B is set to IDLE_LOOP2 as well
813 *
814 * if we have no outer loop, though, then we can use IDLE_LOOP1 only,
815 * and set TARGET_B to IDLE_LOOP1 and we skip IDLE_LOOP2 entirely
816 *
817 * a little confusing, but it helps save precious space in the inst_rom
818 * and sequencer rom and keeps the delays more accurate and reduces
819 * overhead
820 */
821 if (afi_clocks <= 0x100) {
822 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(inner),
823 &sdr_rw_load_mgr_regs->load_cntr1);
824
825 writel(RW_MGR_IDLE_LOOP1,
826 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
827
828 writel(RW_MGR_IDLE_LOOP1, SDR_PHYGRP_RWMGRGRP_ADDRESS |
829 RW_MGR_RUN_SINGLE_GROUP_OFFSET);
830 } else {
831 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(inner),
832 &sdr_rw_load_mgr_regs->load_cntr0);
833
834 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(outer),
835 &sdr_rw_load_mgr_regs->load_cntr1);
836
837 writel(RW_MGR_IDLE_LOOP2,
838 &sdr_rw_load_jump_mgr_regs->load_jump_add0);
839
840 writel(RW_MGR_IDLE_LOOP2,
841 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
842
843 /* hack to get around compiler not being smart enough */
844 if (afi_clocks <= 0x10000) {
845 /* only need to run once */
846 writel(RW_MGR_IDLE_LOOP2, SDR_PHYGRP_RWMGRGRP_ADDRESS |
847 RW_MGR_RUN_SINGLE_GROUP_OFFSET);
848 } else {
849 do {
850 writel(RW_MGR_IDLE_LOOP2,
851 SDR_PHYGRP_RWMGRGRP_ADDRESS |
852 RW_MGR_RUN_SINGLE_GROUP_OFFSET);
853 } while (c_loop-- != 0);
854 }
855 }
856 debug("%s:%d clocks=%u ... end\n", __func__, __LINE__, clocks);
857 }
858
859 static void rw_mgr_mem_initialize(void)
860 {
861 uint32_t r;
862 uint32_t grpaddr = SDR_PHYGRP_RWMGRGRP_ADDRESS |
863 RW_MGR_RUN_SINGLE_GROUP_OFFSET;
864
865 debug("%s:%d\n", __func__, __LINE__);
866
867 /* The reset / cke part of initialization is broadcasted to all ranks */
868 writel(RW_MGR_RANK_ALL, SDR_PHYGRP_RWMGRGRP_ADDRESS |
869 RW_MGR_SET_CS_AND_ODT_MASK_OFFSET);
870
871 /*
872 * Here's how you load register for a loop
873 * Counters are located @ 0x800
874 * Jump address are located @ 0xC00
875 * For both, registers 0 to 3 are selected using bits 3 and 2, like
876 * in 0x800, 0x804, 0x808, 0x80C and 0xC00, 0xC04, 0xC08, 0xC0C
877 * I know this ain't pretty, but Avalon bus throws away the 2 least
878 * significant bits
879 */
880
881 /* start with memory RESET activated */
882
883 /* tINIT = 200us */
884
885 /*
886 * 200us @ 266MHz (3.75 ns) ~ 54000 clock cycles
887 * If a and b are the number of iteration in 2 nested loops
888 * it takes the following number of cycles to complete the operation:
889 * number_of_cycles = ((2 + n) * a + 2) * b
890 * where n is the number of instruction in the inner loop
891 * One possible solution is n = 0 , a = 256 , b = 106 => a = FF,
892 * b = 6A
893 */
894
895 /* Load counters */
896 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(SEQ_TINIT_CNTR0_VAL),
897 &sdr_rw_load_mgr_regs->load_cntr0);
898 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(SEQ_TINIT_CNTR1_VAL),
899 &sdr_rw_load_mgr_regs->load_cntr1);
900 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(SEQ_TINIT_CNTR2_VAL),
901 &sdr_rw_load_mgr_regs->load_cntr2);
902
903 /* Load jump address */
904 writel(RW_MGR_INIT_RESET_0_CKE_0,
905 &sdr_rw_load_jump_mgr_regs->load_jump_add0);
906 writel(RW_MGR_INIT_RESET_0_CKE_0,
907 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
908 writel(RW_MGR_INIT_RESET_0_CKE_0,
909 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
910
911 /* Execute count instruction */
912 writel(RW_MGR_INIT_RESET_0_CKE_0, grpaddr);
913
914 /* indicate that memory is stable */
915 writel(1, &phy_mgr_cfg->reset_mem_stbl);
916
917 /*
918 * transition the RESET to high
919 * Wait for 500us
920 */
921
922 /*
923 * 500us @ 266MHz (3.75 ns) ~ 134000 clock cycles
924 * If a and b are the number of iteration in 2 nested loops
925 * it takes the following number of cycles to complete the operation
926 * number_of_cycles = ((2 + n) * a + 2) * b
927 * where n is the number of instruction in the inner loop
928 * One possible solution is n = 2 , a = 131 , b = 256 => a = 83,
929 * b = FF
930 */
931
932 /* Load counters */
933 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(SEQ_TRESET_CNTR0_VAL),
934 &sdr_rw_load_mgr_regs->load_cntr0);
935 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(SEQ_TRESET_CNTR1_VAL),
936 &sdr_rw_load_mgr_regs->load_cntr1);
937 writel(SKIP_DELAY_LOOP_VALUE_OR_ZERO(SEQ_TRESET_CNTR2_VAL),
938 &sdr_rw_load_mgr_regs->load_cntr2);
939
940 /* Load jump address */
941 writel(RW_MGR_INIT_RESET_1_CKE_0,
942 &sdr_rw_load_jump_mgr_regs->load_jump_add0);
943 writel(RW_MGR_INIT_RESET_1_CKE_0,
944 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
945 writel(RW_MGR_INIT_RESET_1_CKE_0,
946 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
947
948 writel(RW_MGR_INIT_RESET_1_CKE_0, grpaddr);
949
950 /* bring up clock enable */
951
952 /* tXRP < 250 ck cycles */
953 delay_for_n_mem_clocks(250);
954
955 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS; r++) {
956 if (param->skip_ranks[r]) {
957 /* request to skip the rank */
958 continue;
959 }
960
961 /* set rank */
962 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_OFF);
963
964 /*
965 * USER Use Mirror-ed commands for odd ranks if address
966 * mirrorring is on
967 */
968 if ((RW_MGR_MEM_ADDRESS_MIRRORING >> r) & 0x1) {
969 set_jump_as_return();
970 writel(RW_MGR_MRS2_MIRR, grpaddr);
971 delay_for_n_mem_clocks(4);
972 set_jump_as_return();
973 writel(RW_MGR_MRS3_MIRR, grpaddr);
974 delay_for_n_mem_clocks(4);
975 set_jump_as_return();
976 writel(RW_MGR_MRS1_MIRR, grpaddr);
977 delay_for_n_mem_clocks(4);
978 set_jump_as_return();
979 writel(RW_MGR_MRS0_DLL_RESET_MIRR, grpaddr);
980 } else {
981 set_jump_as_return();
982 writel(RW_MGR_MRS2, grpaddr);
983 delay_for_n_mem_clocks(4);
984 set_jump_as_return();
985 writel(RW_MGR_MRS3, grpaddr);
986 delay_for_n_mem_clocks(4);
987 set_jump_as_return();
988 writel(RW_MGR_MRS1, grpaddr);
989 set_jump_as_return();
990 writel(RW_MGR_MRS0_DLL_RESET, grpaddr);
991 }
992 set_jump_as_return();
993 writel(RW_MGR_ZQCL, grpaddr);
994
995 /* tZQinit = tDLLK = 512 ck cycles */
996 delay_for_n_mem_clocks(512);
997 }
998 }
999
1000 /*
1001 * At the end of calibration we have to program the user settings in, and
1002 * USER hand off the memory to the user.
1003 */
1004 static void rw_mgr_mem_handoff(void)
1005 {
1006 uint32_t r;
1007 uint32_t grpaddr = SDR_PHYGRP_RWMGRGRP_ADDRESS |
1008 RW_MGR_RUN_SINGLE_GROUP_OFFSET;
1009
1010 debug("%s:%d\n", __func__, __LINE__);
1011 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS; r++) {
1012 if (param->skip_ranks[r])
1013 /* request to skip the rank */
1014 continue;
1015 /* set rank */
1016 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_OFF);
1017
1018 /* precharge all banks ... */
1019 writel(RW_MGR_PRECHARGE_ALL, grpaddr);
1020
1021 /* load up MR settings specified by user */
1022
1023 /*
1024 * Use Mirror-ed commands for odd ranks if address
1025 * mirrorring is on
1026 */
1027 if ((RW_MGR_MEM_ADDRESS_MIRRORING >> r) & 0x1) {
1028 set_jump_as_return();
1029 writel(RW_MGR_MRS2_MIRR, grpaddr);
1030 delay_for_n_mem_clocks(4);
1031 set_jump_as_return();
1032 writel(RW_MGR_MRS3_MIRR, grpaddr);
1033 delay_for_n_mem_clocks(4);
1034 set_jump_as_return();
1035 writel(RW_MGR_MRS1_MIRR, grpaddr);
1036 delay_for_n_mem_clocks(4);
1037 set_jump_as_return();
1038 writel(RW_MGR_MRS0_USER_MIRR, grpaddr);
1039 } else {
1040 set_jump_as_return();
1041 writel(RW_MGR_MRS2, grpaddr);
1042 delay_for_n_mem_clocks(4);
1043 set_jump_as_return();
1044 writel(RW_MGR_MRS3, grpaddr);
1045 delay_for_n_mem_clocks(4);
1046 set_jump_as_return();
1047 writel(RW_MGR_MRS1, grpaddr);
1048 delay_for_n_mem_clocks(4);
1049 set_jump_as_return();
1050 writel(RW_MGR_MRS0_USER, grpaddr);
1051 }
1052 /*
1053 * USER need to wait tMOD (12CK or 15ns) time before issuing
1054 * other commands, but we will have plenty of NIOS cycles before
1055 * actual handoff so its okay.
1056 */
1057 }
1058 }
1059
1060 /*
1061 * performs a guaranteed read on the patterns we are going to use during a
1062 * read test to ensure memory works
1063 */
1064 static uint32_t rw_mgr_mem_calibrate_read_test_patterns(uint32_t rank_bgn,
1065 uint32_t group, uint32_t num_tries, uint32_t *bit_chk,
1066 uint32_t all_ranks)
1067 {
1068 uint32_t r, vg;
1069 uint32_t correct_mask_vg;
1070 uint32_t tmp_bit_chk;
1071 uint32_t rank_end = all_ranks ? RW_MGR_MEM_NUMBER_OF_RANKS :
1072 (rank_bgn + NUM_RANKS_PER_SHADOW_REG);
1073 uint32_t addr;
1074 uint32_t base_rw_mgr;
1075
1076 *bit_chk = param->read_correct_mask;
1077 correct_mask_vg = param->read_correct_mask_vg;
1078
1079 for (r = rank_bgn; r < rank_end; r++) {
1080 if (param->skip_ranks[r])
1081 /* request to skip the rank */
1082 continue;
1083
1084 /* set rank */
1085 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_READ_WRITE);
1086
1087 /* Load up a constant bursts of read commands */
1088 writel(0x20, &sdr_rw_load_mgr_regs->load_cntr0);
1089 writel(RW_MGR_GUARANTEED_READ,
1090 &sdr_rw_load_jump_mgr_regs->load_jump_add0);
1091
1092 writel(0x20, &sdr_rw_load_mgr_regs->load_cntr1);
1093 writel(RW_MGR_GUARANTEED_READ_CONT,
1094 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
1095
1096 tmp_bit_chk = 0;
1097 for (vg = RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS-1; ; vg--) {
1098 /* reset the fifos to get pointers to known state */
1099
1100 writel(0, &phy_mgr_cmd->fifo_reset);
1101 writel(0, SDR_PHYGRP_RWMGRGRP_ADDRESS |
1102 RW_MGR_RESET_READ_DATAPATH_OFFSET);
1103
1104 tmp_bit_chk = tmp_bit_chk << (RW_MGR_MEM_DQ_PER_READ_DQS
1105 / RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS);
1106
1107 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_SINGLE_GROUP_OFFSET;
1108 writel(RW_MGR_GUARANTEED_READ, addr +
1109 ((group * RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS +
1110 vg) << 2));
1111
1112 base_rw_mgr = readl(SDR_PHYGRP_RWMGRGRP_ADDRESS);
1113 tmp_bit_chk = tmp_bit_chk | (correct_mask_vg & (~base_rw_mgr));
1114
1115 if (vg == 0)
1116 break;
1117 }
1118 *bit_chk &= tmp_bit_chk;
1119 }
1120
1121 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_SINGLE_GROUP_OFFSET;
1122 writel(RW_MGR_CLEAR_DQS_ENABLE, addr + (group << 2));
1123
1124 set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
1125 debug_cond(DLEVEL == 1, "%s:%d test_load_patterns(%u,ALL) => (%u == %u) =>\
1126 %lu\n", __func__, __LINE__, group, *bit_chk, param->read_correct_mask,
1127 (long unsigned int)(*bit_chk == param->read_correct_mask));
1128 return *bit_chk == param->read_correct_mask;
1129 }
1130
1131 static uint32_t rw_mgr_mem_calibrate_read_test_patterns_all_ranks
1132 (uint32_t group, uint32_t num_tries, uint32_t *bit_chk)
1133 {
1134 return rw_mgr_mem_calibrate_read_test_patterns(0, group,
1135 num_tries, bit_chk, 1);
1136 }
1137
1138 /* load up the patterns we are going to use during a read test */
1139 static void rw_mgr_mem_calibrate_read_load_patterns(uint32_t rank_bgn,
1140 uint32_t all_ranks)
1141 {
1142 uint32_t r;
1143 uint32_t rank_end = all_ranks ? RW_MGR_MEM_NUMBER_OF_RANKS :
1144 (rank_bgn + NUM_RANKS_PER_SHADOW_REG);
1145
1146 debug("%s:%d\n", __func__, __LINE__);
1147 for (r = rank_bgn; r < rank_end; r++) {
1148 if (param->skip_ranks[r])
1149 /* request to skip the rank */
1150 continue;
1151
1152 /* set rank */
1153 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_READ_WRITE);
1154
1155 /* Load up a constant bursts */
1156 writel(0x20, &sdr_rw_load_mgr_regs->load_cntr0);
1157
1158 writel(RW_MGR_GUARANTEED_WRITE_WAIT0,
1159 &sdr_rw_load_jump_mgr_regs->load_jump_add0);
1160
1161 writel(0x20, &sdr_rw_load_mgr_regs->load_cntr1);
1162
1163 writel(RW_MGR_GUARANTEED_WRITE_WAIT1,
1164 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
1165
1166 writel(0x04, &sdr_rw_load_mgr_regs->load_cntr2);
1167
1168 writel(RW_MGR_GUARANTEED_WRITE_WAIT2,
1169 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
1170
1171 writel(0x04, &sdr_rw_load_mgr_regs->load_cntr3);
1172
1173 writel(RW_MGR_GUARANTEED_WRITE_WAIT3,
1174 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
1175
1176 writel(RW_MGR_GUARANTEED_WRITE, SDR_PHYGRP_RWMGRGRP_ADDRESS |
1177 RW_MGR_RUN_SINGLE_GROUP_OFFSET);
1178 }
1179
1180 set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
1181 }
1182
1183 /*
1184 * try a read and see if it returns correct data back. has dummy reads
1185 * inserted into the mix used to align dqs enable. has more thorough checks
1186 * than the regular read test.
1187 */
1188 static uint32_t rw_mgr_mem_calibrate_read_test(uint32_t rank_bgn, uint32_t group,
1189 uint32_t num_tries, uint32_t all_correct, uint32_t *bit_chk,
1190 uint32_t all_groups, uint32_t all_ranks)
1191 {
1192 uint32_t r, vg;
1193 uint32_t correct_mask_vg;
1194 uint32_t tmp_bit_chk;
1195 uint32_t rank_end = all_ranks ? RW_MGR_MEM_NUMBER_OF_RANKS :
1196 (rank_bgn + NUM_RANKS_PER_SHADOW_REG);
1197 uint32_t addr;
1198 uint32_t base_rw_mgr;
1199
1200 *bit_chk = param->read_correct_mask;
1201 correct_mask_vg = param->read_correct_mask_vg;
1202
1203 uint32_t quick_read_mode = (((STATIC_CALIB_STEPS) &
1204 CALIB_SKIP_DELAY_SWEEPS) && ENABLE_SUPER_QUICK_CALIBRATION);
1205
1206 for (r = rank_bgn; r < rank_end; r++) {
1207 if (param->skip_ranks[r])
1208 /* request to skip the rank */
1209 continue;
1210
1211 /* set rank */
1212 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_READ_WRITE);
1213
1214 writel(0x10, &sdr_rw_load_mgr_regs->load_cntr1);
1215
1216 writel(RW_MGR_READ_B2B_WAIT1,
1217 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
1218
1219 writel(0x10, &sdr_rw_load_mgr_regs->load_cntr2);
1220 writel(RW_MGR_READ_B2B_WAIT2,
1221 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
1222
1223 if (quick_read_mode)
1224 writel(0x1, &sdr_rw_load_mgr_regs->load_cntr0);
1225 /* need at least two (1+1) reads to capture failures */
1226 else if (all_groups)
1227 writel(0x06, &sdr_rw_load_mgr_regs->load_cntr0);
1228 else
1229 writel(0x32, &sdr_rw_load_mgr_regs->load_cntr0);
1230
1231 writel(RW_MGR_READ_B2B,
1232 &sdr_rw_load_jump_mgr_regs->load_jump_add0);
1233 if (all_groups)
1234 writel(RW_MGR_MEM_IF_READ_DQS_WIDTH *
1235 RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS - 1,
1236 &sdr_rw_load_mgr_regs->load_cntr3);
1237 else
1238 writel(0x0, &sdr_rw_load_mgr_regs->load_cntr3);
1239
1240 writel(RW_MGR_READ_B2B,
1241 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
1242
1243 tmp_bit_chk = 0;
1244 for (vg = RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS-1; ; vg--) {
1245 /* reset the fifos to get pointers to known state */
1246 writel(0, &phy_mgr_cmd->fifo_reset);
1247 writel(0, SDR_PHYGRP_RWMGRGRP_ADDRESS |
1248 RW_MGR_RESET_READ_DATAPATH_OFFSET);
1249
1250 tmp_bit_chk = tmp_bit_chk << (RW_MGR_MEM_DQ_PER_READ_DQS
1251 / RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS);
1252
1253 if (all_groups)
1254 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_ALL_GROUPS_OFFSET;
1255 else
1256 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_SINGLE_GROUP_OFFSET;
1257
1258 writel(RW_MGR_READ_B2B, addr +
1259 ((group * RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS +
1260 vg) << 2));
1261
1262 base_rw_mgr = readl(SDR_PHYGRP_RWMGRGRP_ADDRESS);
1263 tmp_bit_chk = tmp_bit_chk | (correct_mask_vg & ~(base_rw_mgr));
1264
1265 if (vg == 0)
1266 break;
1267 }
1268 *bit_chk &= tmp_bit_chk;
1269 }
1270
1271 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_SINGLE_GROUP_OFFSET;
1272 writel(RW_MGR_CLEAR_DQS_ENABLE, addr + (group << 2));
1273
1274 if (all_correct) {
1275 set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
1276 debug_cond(DLEVEL == 2, "%s:%d read_test(%u,ALL,%u) =>\
1277 (%u == %u) => %lu", __func__, __LINE__, group,
1278 all_groups, *bit_chk, param->read_correct_mask,
1279 (long unsigned int)(*bit_chk ==
1280 param->read_correct_mask));
1281 return *bit_chk == param->read_correct_mask;
1282 } else {
1283 set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
1284 debug_cond(DLEVEL == 2, "%s:%d read_test(%u,ONE,%u) =>\
1285 (%u != %lu) => %lu\n", __func__, __LINE__,
1286 group, all_groups, *bit_chk, (long unsigned int)0,
1287 (long unsigned int)(*bit_chk != 0x00));
1288 return *bit_chk != 0x00;
1289 }
1290 }
1291
1292 static uint32_t rw_mgr_mem_calibrate_read_test_all_ranks(uint32_t group,
1293 uint32_t num_tries, uint32_t all_correct, uint32_t *bit_chk,
1294 uint32_t all_groups)
1295 {
1296 return rw_mgr_mem_calibrate_read_test(0, group, num_tries, all_correct,
1297 bit_chk, all_groups, 1);
1298 }
1299
1300 static void rw_mgr_incr_vfifo(uint32_t grp, uint32_t *v)
1301 {
1302 writel(grp, &phy_mgr_cmd->inc_vfifo_hard_phy);
1303 (*v)++;
1304 }
1305
1306 static void rw_mgr_decr_vfifo(uint32_t grp, uint32_t *v)
1307 {
1308 uint32_t i;
1309
1310 for (i = 0; i < VFIFO_SIZE-1; i++)
1311 rw_mgr_incr_vfifo(grp, v);
1312 }
1313
1314 static int find_vfifo_read(uint32_t grp, uint32_t *bit_chk)
1315 {
1316 uint32_t v;
1317 uint32_t fail_cnt = 0;
1318 uint32_t test_status;
1319
1320 for (v = 0; v < VFIFO_SIZE; ) {
1321 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: vfifo %u\n",
1322 __func__, __LINE__, v);
1323 test_status = rw_mgr_mem_calibrate_read_test_all_ranks
1324 (grp, 1, PASS_ONE_BIT, bit_chk, 0);
1325 if (!test_status) {
1326 fail_cnt++;
1327
1328 if (fail_cnt == 2)
1329 break;
1330 }
1331
1332 /* fiddle with FIFO */
1333 rw_mgr_incr_vfifo(grp, &v);
1334 }
1335
1336 if (v >= VFIFO_SIZE) {
1337 /* no failing read found!! Something must have gone wrong */
1338 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: vfifo failed\n",
1339 __func__, __LINE__);
1340 return 0;
1341 } else {
1342 return v;
1343 }
1344 }
1345
1346 static int find_working_phase(uint32_t *grp, uint32_t *bit_chk,
1347 uint32_t dtaps_per_ptap, uint32_t *work_bgn,
1348 uint32_t *v, uint32_t *d, uint32_t *p,
1349 uint32_t *i, uint32_t *max_working_cnt)
1350 {
1351 uint32_t found_begin = 0;
1352 uint32_t tmp_delay = 0;
1353 uint32_t test_status;
1354
1355 for (*d = 0; *d <= dtaps_per_ptap; (*d)++, tmp_delay +=
1356 IO_DELAY_PER_DQS_EN_DCHAIN_TAP) {
1357 *work_bgn = tmp_delay;
1358 scc_mgr_set_dqs_en_delay_all_ranks(*grp, *d);
1359
1360 for (*i = 0; *i < VFIFO_SIZE; (*i)++) {
1361 for (*p = 0; *p <= IO_DQS_EN_PHASE_MAX; (*p)++, *work_bgn +=
1362 IO_DELAY_PER_OPA_TAP) {
1363 scc_mgr_set_dqs_en_phase_all_ranks(*grp, *p);
1364
1365 test_status =
1366 rw_mgr_mem_calibrate_read_test_all_ranks
1367 (*grp, 1, PASS_ONE_BIT, bit_chk, 0);
1368
1369 if (test_status) {
1370 *max_working_cnt = 1;
1371 found_begin = 1;
1372 break;
1373 }
1374 }
1375
1376 if (found_begin)
1377 break;
1378
1379 if (*p > IO_DQS_EN_PHASE_MAX)
1380 /* fiddle with FIFO */
1381 rw_mgr_incr_vfifo(*grp, v);
1382 }
1383
1384 if (found_begin)
1385 break;
1386 }
1387
1388 if (*i >= VFIFO_SIZE) {
1389 /* cannot find working solution */
1390 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: no vfifo/\
1391 ptap/dtap\n", __func__, __LINE__);
1392 return 0;
1393 } else {
1394 return 1;
1395 }
1396 }
1397
1398 static void sdr_backup_phase(uint32_t *grp, uint32_t *bit_chk,
1399 uint32_t *work_bgn, uint32_t *v, uint32_t *d,
1400 uint32_t *p, uint32_t *max_working_cnt)
1401 {
1402 uint32_t found_begin = 0;
1403 uint32_t tmp_delay;
1404
1405 /* Special case code for backing up a phase */
1406 if (*p == 0) {
1407 *p = IO_DQS_EN_PHASE_MAX;
1408 rw_mgr_decr_vfifo(*grp, v);
1409 } else {
1410 (*p)--;
1411 }
1412 tmp_delay = *work_bgn - IO_DELAY_PER_OPA_TAP;
1413 scc_mgr_set_dqs_en_phase_all_ranks(*grp, *p);
1414
1415 for (*d = 0; *d <= IO_DQS_EN_DELAY_MAX && tmp_delay < *work_bgn;
1416 (*d)++, tmp_delay += IO_DELAY_PER_DQS_EN_DCHAIN_TAP) {
1417 scc_mgr_set_dqs_en_delay_all_ranks(*grp, *d);
1418
1419 if (rw_mgr_mem_calibrate_read_test_all_ranks(*grp, 1,
1420 PASS_ONE_BIT,
1421 bit_chk, 0)) {
1422 found_begin = 1;
1423 *work_bgn = tmp_delay;
1424 break;
1425 }
1426 }
1427
1428 /* We have found a working dtap before the ptap found above */
1429 if (found_begin == 1)
1430 (*max_working_cnt)++;
1431
1432 /*
1433 * Restore VFIFO to old state before we decremented it
1434 * (if needed).
1435 */
1436 (*p)++;
1437 if (*p > IO_DQS_EN_PHASE_MAX) {
1438 *p = 0;
1439 rw_mgr_incr_vfifo(*grp, v);
1440 }
1441
1442 scc_mgr_set_dqs_en_delay_all_ranks(*grp, 0);
1443 }
1444
1445 static int sdr_nonworking_phase(uint32_t *grp, uint32_t *bit_chk,
1446 uint32_t *work_bgn, uint32_t *v, uint32_t *d,
1447 uint32_t *p, uint32_t *i, uint32_t *max_working_cnt,
1448 uint32_t *work_end)
1449 {
1450 uint32_t found_end = 0;
1451
1452 (*p)++;
1453 *work_end += IO_DELAY_PER_OPA_TAP;
1454 if (*p > IO_DQS_EN_PHASE_MAX) {
1455 /* fiddle with FIFO */
1456 *p = 0;
1457 rw_mgr_incr_vfifo(*grp, v);
1458 }
1459
1460 for (; *i < VFIFO_SIZE + 1; (*i)++) {
1461 for (; *p <= IO_DQS_EN_PHASE_MAX; (*p)++, *work_end
1462 += IO_DELAY_PER_OPA_TAP) {
1463 scc_mgr_set_dqs_en_phase_all_ranks(*grp, *p);
1464
1465 if (!rw_mgr_mem_calibrate_read_test_all_ranks
1466 (*grp, 1, PASS_ONE_BIT, bit_chk, 0)) {
1467 found_end = 1;
1468 break;
1469 } else {
1470 (*max_working_cnt)++;
1471 }
1472 }
1473
1474 if (found_end)
1475 break;
1476
1477 if (*p > IO_DQS_EN_PHASE_MAX) {
1478 /* fiddle with FIFO */
1479 rw_mgr_incr_vfifo(*grp, v);
1480 *p = 0;
1481 }
1482 }
1483
1484 if (*i >= VFIFO_SIZE + 1) {
1485 /* cannot see edge of failing read */
1486 debug_cond(DLEVEL == 2, "%s:%d sdr_nonworking_phase: end:\
1487 failed\n", __func__, __LINE__);
1488 return 0;
1489 } else {
1490 return 1;
1491 }
1492 }
1493
1494 static int sdr_find_window_centre(uint32_t *grp, uint32_t *bit_chk,
1495 uint32_t *work_bgn, uint32_t *v, uint32_t *d,
1496 uint32_t *p, uint32_t *work_mid,
1497 uint32_t *work_end)
1498 {
1499 int i;
1500 int tmp_delay = 0;
1501
1502 *work_mid = (*work_bgn + *work_end) / 2;
1503
1504 debug_cond(DLEVEL == 2, "work_bgn=%d work_end=%d work_mid=%d\n",
1505 *work_bgn, *work_end, *work_mid);
1506 /* Get the middle delay to be less than a VFIFO delay */
1507 for (*p = 0; *p <= IO_DQS_EN_PHASE_MAX;
1508 (*p)++, tmp_delay += IO_DELAY_PER_OPA_TAP)
1509 ;
1510 debug_cond(DLEVEL == 2, "vfifo ptap delay %d\n", tmp_delay);
1511 while (*work_mid > tmp_delay)
1512 *work_mid -= tmp_delay;
1513 debug_cond(DLEVEL == 2, "new work_mid %d\n", *work_mid);
1514
1515 tmp_delay = 0;
1516 for (*p = 0; *p <= IO_DQS_EN_PHASE_MAX && tmp_delay < *work_mid;
1517 (*p)++, tmp_delay += IO_DELAY_PER_OPA_TAP)
1518 ;
1519 tmp_delay -= IO_DELAY_PER_OPA_TAP;
1520 debug_cond(DLEVEL == 2, "new p %d, tmp_delay=%d\n", (*p) - 1, tmp_delay);
1521 for (*d = 0; *d <= IO_DQS_EN_DELAY_MAX && tmp_delay < *work_mid; (*d)++,
1522 tmp_delay += IO_DELAY_PER_DQS_EN_DCHAIN_TAP)
1523 ;
1524 debug_cond(DLEVEL == 2, "new d %d, tmp_delay=%d\n", *d, tmp_delay);
1525
1526 scc_mgr_set_dqs_en_phase_all_ranks(*grp, (*p) - 1);
1527 scc_mgr_set_dqs_en_delay_all_ranks(*grp, *d);
1528
1529 /*
1530 * push vfifo until we can successfully calibrate. We can do this
1531 * because the largest possible margin in 1 VFIFO cycle.
1532 */
1533 for (i = 0; i < VFIFO_SIZE; i++) {
1534 debug_cond(DLEVEL == 2, "find_dqs_en_phase: center: vfifo=%u\n",
1535 *v);
1536 if (rw_mgr_mem_calibrate_read_test_all_ranks(*grp, 1,
1537 PASS_ONE_BIT,
1538 bit_chk, 0)) {
1539 break;
1540 }
1541
1542 /* fiddle with FIFO */
1543 rw_mgr_incr_vfifo(*grp, v);
1544 }
1545
1546 if (i >= VFIFO_SIZE) {
1547 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: center: \
1548 failed\n", __func__, __LINE__);
1549 return 0;
1550 } else {
1551 return 1;
1552 }
1553 }
1554
1555 /* find a good dqs enable to use */
1556 static uint32_t rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase(uint32_t grp)
1557 {
1558 uint32_t v, d, p, i;
1559 uint32_t max_working_cnt;
1560 uint32_t bit_chk;
1561 uint32_t dtaps_per_ptap;
1562 uint32_t work_bgn, work_mid, work_end;
1563 uint32_t found_passing_read, found_failing_read, initial_failing_dtap;
1564
1565 debug("%s:%d %u\n", __func__, __LINE__, grp);
1566
1567 reg_file_set_sub_stage(CAL_SUBSTAGE_VFIFO_CENTER);
1568
1569 scc_mgr_set_dqs_en_delay_all_ranks(grp, 0);
1570 scc_mgr_set_dqs_en_phase_all_ranks(grp, 0);
1571
1572 /* ************************************************************** */
1573 /* * Step 0 : Determine number of delay taps for each phase tap * */
1574 dtaps_per_ptap = IO_DELAY_PER_OPA_TAP/IO_DELAY_PER_DQS_EN_DCHAIN_TAP;
1575
1576 /* ********************************************************* */
1577 /* * Step 1 : First push vfifo until we get a failing read * */
1578 v = find_vfifo_read(grp, &bit_chk);
1579
1580 max_working_cnt = 0;
1581
1582 /* ******************************************************** */
1583 /* * step 2: find first working phase, increment in ptaps * */
1584 work_bgn = 0;
1585 if (find_working_phase(&grp, &bit_chk, dtaps_per_ptap, &work_bgn, &v, &d,
1586 &p, &i, &max_working_cnt) == 0)
1587 return 0;
1588
1589 work_end = work_bgn;
1590
1591 /*
1592 * If d is 0 then the working window covers a phase tap and
1593 * we can follow the old procedure otherwise, we've found the beginning,
1594 * and we need to increment the dtaps until we find the end.
1595 */
1596 if (d == 0) {
1597 /* ********************************************************* */
1598 /* * step 3a: if we have room, back off by one and
1599 increment in dtaps * */
1600
1601 sdr_backup_phase(&grp, &bit_chk, &work_bgn, &v, &d, &p,
1602 &max_working_cnt);
1603
1604 /* ********************************************************* */
1605 /* * step 4a: go forward from working phase to non working
1606 phase, increment in ptaps * */
1607 if (sdr_nonworking_phase(&grp, &bit_chk, &work_bgn, &v, &d, &p,
1608 &i, &max_working_cnt, &work_end) == 0)
1609 return 0;
1610
1611 /* ********************************************************* */
1612 /* * step 5a: back off one from last, increment in dtaps * */
1613
1614 /* Special case code for backing up a phase */
1615 if (p == 0) {
1616 p = IO_DQS_EN_PHASE_MAX;
1617 rw_mgr_decr_vfifo(grp, &v);
1618 } else {
1619 p = p - 1;
1620 }
1621
1622 work_end -= IO_DELAY_PER_OPA_TAP;
1623 scc_mgr_set_dqs_en_phase_all_ranks(grp, p);
1624
1625 /* * The actual increment of dtaps is done outside of
1626 the if/else loop to share code */
1627 d = 0;
1628
1629 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: v/p: \
1630 vfifo=%u ptap=%u\n", __func__, __LINE__,
1631 v, p);
1632 } else {
1633 /* ******************************************************* */
1634 /* * step 3-5b: Find the right edge of the window using
1635 delay taps * */
1636 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase:vfifo=%u \
1637 ptap=%u dtap=%u bgn=%u\n", __func__, __LINE__,
1638 v, p, d, work_bgn);
1639
1640 work_end = work_bgn;
1641
1642 /* * The actual increment of dtaps is done outside of the
1643 if/else loop to share code */
1644
1645 /* Only here to counterbalance a subtract later on which is
1646 not needed if this branch of the algorithm is taken */
1647 max_working_cnt++;
1648 }
1649
1650 /* The dtap increment to find the failing edge is done here */
1651 for (; d <= IO_DQS_EN_DELAY_MAX; d++, work_end +=
1652 IO_DELAY_PER_DQS_EN_DCHAIN_TAP) {
1653 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: \
1654 end-2: dtap=%u\n", __func__, __LINE__, d);
1655 scc_mgr_set_dqs_en_delay_all_ranks(grp, d);
1656
1657 if (!rw_mgr_mem_calibrate_read_test_all_ranks(grp, 1,
1658 PASS_ONE_BIT,
1659 &bit_chk, 0)) {
1660 break;
1661 }
1662 }
1663
1664 /* Go back to working dtap */
1665 if (d != 0)
1666 work_end -= IO_DELAY_PER_DQS_EN_DCHAIN_TAP;
1667
1668 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: v/p/d: vfifo=%u \
1669 ptap=%u dtap=%u end=%u\n", __func__, __LINE__,
1670 v, p, d-1, work_end);
1671
1672 if (work_end < work_bgn) {
1673 /* nil range */
1674 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: end-2: \
1675 failed\n", __func__, __LINE__);
1676 return 0;
1677 }
1678
1679 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: found range [%u,%u]\n",
1680 __func__, __LINE__, work_bgn, work_end);
1681
1682 /* *************************************************************** */
1683 /*
1684 * * We need to calculate the number of dtaps that equal a ptap
1685 * * To do that we'll back up a ptap and re-find the edge of the
1686 * * window using dtaps
1687 */
1688
1689 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: calculate dtaps_per_ptap \
1690 for tracking\n", __func__, __LINE__);
1691
1692 /* Special case code for backing up a phase */
1693 if (p == 0) {
1694 p = IO_DQS_EN_PHASE_MAX;
1695 rw_mgr_decr_vfifo(grp, &v);
1696 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: backedup \
1697 cycle/phase: v=%u p=%u\n", __func__, __LINE__,
1698 v, p);
1699 } else {
1700 p = p - 1;
1701 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: backedup \
1702 phase only: v=%u p=%u", __func__, __LINE__,
1703 v, p);
1704 }
1705
1706 scc_mgr_set_dqs_en_phase_all_ranks(grp, p);
1707
1708 /*
1709 * Increase dtap until we first see a passing read (in case the
1710 * window is smaller than a ptap),
1711 * and then a failing read to mark the edge of the window again
1712 */
1713
1714 /* Find a passing read */
1715 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: find passing read\n",
1716 __func__, __LINE__);
1717 found_passing_read = 0;
1718 found_failing_read = 0;
1719 initial_failing_dtap = d;
1720 for (; d <= IO_DQS_EN_DELAY_MAX; d++) {
1721 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: testing \
1722 read d=%u\n", __func__, __LINE__, d);
1723 scc_mgr_set_dqs_en_delay_all_ranks(grp, d);
1724
1725 if (rw_mgr_mem_calibrate_read_test_all_ranks(grp, 1,
1726 PASS_ONE_BIT,
1727 &bit_chk, 0)) {
1728 found_passing_read = 1;
1729 break;
1730 }
1731 }
1732
1733 if (found_passing_read) {
1734 /* Find a failing read */
1735 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: find failing \
1736 read\n", __func__, __LINE__);
1737 for (d = d + 1; d <= IO_DQS_EN_DELAY_MAX; d++) {
1738 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: \
1739 testing read d=%u\n", __func__, __LINE__, d);
1740 scc_mgr_set_dqs_en_delay_all_ranks(grp, d);
1741
1742 if (!rw_mgr_mem_calibrate_read_test_all_ranks
1743 (grp, 1, PASS_ONE_BIT, &bit_chk, 0)) {
1744 found_failing_read = 1;
1745 break;
1746 }
1747 }
1748 } else {
1749 debug_cond(DLEVEL == 1, "%s:%d find_dqs_en_phase: failed to \
1750 calculate dtaps", __func__, __LINE__);
1751 debug_cond(DLEVEL == 1, "per ptap. Fall back on static value\n");
1752 }
1753
1754 /*
1755 * The dynamically calculated dtaps_per_ptap is only valid if we
1756 * found a passing/failing read. If we didn't, it means d hit the max
1757 * (IO_DQS_EN_DELAY_MAX). Otherwise, dtaps_per_ptap retains its
1758 * statically calculated value.
1759 */
1760 if (found_passing_read && found_failing_read)
1761 dtaps_per_ptap = d - initial_failing_dtap;
1762
1763 writel(dtaps_per_ptap, &sdr_reg_file->dtaps_per_ptap);
1764 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: dtaps_per_ptap=%u \
1765 - %u = %u", __func__, __LINE__, d,
1766 initial_failing_dtap, dtaps_per_ptap);
1767
1768 /* ******************************************** */
1769 /* * step 6: Find the centre of the window * */
1770 if (sdr_find_window_centre(&grp, &bit_chk, &work_bgn, &v, &d, &p,
1771 &work_mid, &work_end) == 0)
1772 return 0;
1773
1774 debug_cond(DLEVEL == 2, "%s:%d find_dqs_en_phase: center found: \
1775 vfifo=%u ptap=%u dtap=%u\n", __func__, __LINE__,
1776 v, p-1, d);
1777 return 1;
1778 }
1779
1780 /*
1781 * Try rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase across different
1782 * dq_in_delay values
1783 */
1784 static uint32_t
1785 rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase_sweep_dq_in_delay
1786 (uint32_t write_group, uint32_t read_group, uint32_t test_bgn)
1787 {
1788 uint32_t found;
1789 uint32_t i;
1790 uint32_t p;
1791 uint32_t d;
1792 uint32_t r;
1793
1794 const uint32_t delay_step = IO_IO_IN_DELAY_MAX /
1795 (RW_MGR_MEM_DQ_PER_READ_DQS-1);
1796 /* we start at zero, so have one less dq to devide among */
1797
1798 debug("%s:%d (%u,%u,%u)", __func__, __LINE__, write_group, read_group,
1799 test_bgn);
1800
1801 /* try different dq_in_delays since the dq path is shorter than dqs */
1802
1803 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
1804 r += NUM_RANKS_PER_SHADOW_REG) {
1805 for (i = 0, p = test_bgn, d = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++, p++, d += delay_step) {
1806 debug_cond(DLEVEL == 1, "%s:%d rw_mgr_mem_calibrate_\
1807 vfifo_find_dqs_", __func__, __LINE__);
1808 debug_cond(DLEVEL == 1, "en_phase_sweep_dq_in_delay: g=%u/%u ",
1809 write_group, read_group);
1810 debug_cond(DLEVEL == 1, "r=%u, i=%u p=%u d=%u\n", r, i , p, d);
1811 scc_mgr_set_dq_in_delay(p, d);
1812 scc_mgr_load_dq(p);
1813 }
1814 writel(0, &sdr_scc_mgr->update);
1815 }
1816
1817 found = rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase(read_group);
1818
1819 debug_cond(DLEVEL == 1, "%s:%d rw_mgr_mem_calibrate_vfifo_find_dqs_\
1820 en_phase_sweep_dq", __func__, __LINE__);
1821 debug_cond(DLEVEL == 1, "_in_delay: g=%u/%u found=%u; Reseting delay \
1822 chain to zero\n", write_group, read_group, found);
1823
1824 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
1825 r += NUM_RANKS_PER_SHADOW_REG) {
1826 for (i = 0, p = test_bgn; i < RW_MGR_MEM_DQ_PER_READ_DQS;
1827 i++, p++) {
1828 scc_mgr_set_dq_in_delay(p, 0);
1829 scc_mgr_load_dq(p);
1830 }
1831 writel(0, &sdr_scc_mgr->update);
1832 }
1833
1834 return found;
1835 }
1836
1837 /* per-bit deskew DQ and center */
1838 static uint32_t rw_mgr_mem_calibrate_vfifo_center(uint32_t rank_bgn,
1839 uint32_t write_group, uint32_t read_group, uint32_t test_bgn,
1840 uint32_t use_read_test, uint32_t update_fom)
1841 {
1842 uint32_t i, p, d, min_index;
1843 /*
1844 * Store these as signed since there are comparisons with
1845 * signed numbers.
1846 */
1847 uint32_t bit_chk;
1848 uint32_t sticky_bit_chk;
1849 int32_t left_edge[RW_MGR_MEM_DQ_PER_READ_DQS];
1850 int32_t right_edge[RW_MGR_MEM_DQ_PER_READ_DQS];
1851 int32_t final_dq[RW_MGR_MEM_DQ_PER_READ_DQS];
1852 int32_t mid;
1853 int32_t orig_mid_min, mid_min;
1854 int32_t new_dqs, start_dqs, start_dqs_en, shift_dq, final_dqs,
1855 final_dqs_en;
1856 int32_t dq_margin, dqs_margin;
1857 uint32_t stop;
1858 uint32_t temp_dq_in_delay1, temp_dq_in_delay2;
1859 uint32_t addr;
1860
1861 debug("%s:%d: %u %u", __func__, __LINE__, read_group, test_bgn);
1862
1863 addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_DQS_IN_DELAY_OFFSET;
1864 start_dqs = readl(addr + (read_group << 2));
1865 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS)
1866 start_dqs_en = readl(addr + ((read_group << 2)
1867 - IO_DQS_EN_DELAY_OFFSET));
1868
1869 /* set the left and right edge of each bit to an illegal value */
1870 /* use (IO_IO_IN_DELAY_MAX + 1) as an illegal value */
1871 sticky_bit_chk = 0;
1872 for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
1873 left_edge[i] = IO_IO_IN_DELAY_MAX + 1;
1874 right_edge[i] = IO_IO_IN_DELAY_MAX + 1;
1875 }
1876
1877 /* Search for the left edge of the window for each bit */
1878 for (d = 0; d <= IO_IO_IN_DELAY_MAX; d++) {
1879 scc_mgr_apply_group_dq_in_delay(write_group, test_bgn, d);
1880
1881 writel(0, &sdr_scc_mgr->update);
1882
1883 /*
1884 * Stop searching when the read test doesn't pass AND when
1885 * we've seen a passing read on every bit.
1886 */
1887 if (use_read_test) {
1888 stop = !rw_mgr_mem_calibrate_read_test(rank_bgn,
1889 read_group, NUM_READ_PB_TESTS, PASS_ONE_BIT,
1890 &bit_chk, 0, 0);
1891 } else {
1892 rw_mgr_mem_calibrate_write_test(rank_bgn, write_group,
1893 0, PASS_ONE_BIT,
1894 &bit_chk, 0);
1895 bit_chk = bit_chk >> (RW_MGR_MEM_DQ_PER_READ_DQS *
1896 (read_group - (write_group *
1897 RW_MGR_MEM_IF_READ_DQS_WIDTH /
1898 RW_MGR_MEM_IF_WRITE_DQS_WIDTH)));
1899 stop = (bit_chk == 0);
1900 }
1901 sticky_bit_chk = sticky_bit_chk | bit_chk;
1902 stop = stop && (sticky_bit_chk == param->read_correct_mask);
1903 debug_cond(DLEVEL == 2, "%s:%d vfifo_center(left): dtap=%u => %u == %u \
1904 && %u", __func__, __LINE__, d,
1905 sticky_bit_chk,
1906 param->read_correct_mask, stop);
1907
1908 if (stop == 1) {
1909 break;
1910 } else {
1911 for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
1912 if (bit_chk & 1) {
1913 /* Remember a passing test as the
1914 left_edge */
1915 left_edge[i] = d;
1916 } else {
1917 /* If a left edge has not been seen yet,
1918 then a future passing test will mark
1919 this edge as the right edge */
1920 if (left_edge[i] ==
1921 IO_IO_IN_DELAY_MAX + 1) {
1922 right_edge[i] = -(d + 1);
1923 }
1924 }
1925 bit_chk = bit_chk >> 1;
1926 }
1927 }
1928 }
1929
1930 /* Reset DQ delay chains to 0 */
1931 scc_mgr_apply_group_dq_in_delay(test_bgn, 0);
1932 sticky_bit_chk = 0;
1933 for (i = RW_MGR_MEM_DQ_PER_READ_DQS - 1;; i--) {
1934 debug_cond(DLEVEL == 2, "%s:%d vfifo_center: left_edge[%u]: \
1935 %d right_edge[%u]: %d\n", __func__, __LINE__,
1936 i, left_edge[i], i, right_edge[i]);
1937
1938 /*
1939 * Check for cases where we haven't found the left edge,
1940 * which makes our assignment of the the right edge invalid.
1941 * Reset it to the illegal value.
1942 */
1943 if ((left_edge[i] == IO_IO_IN_DELAY_MAX + 1) && (
1944 right_edge[i] != IO_IO_IN_DELAY_MAX + 1)) {
1945 right_edge[i] = IO_IO_IN_DELAY_MAX + 1;
1946 debug_cond(DLEVEL == 2, "%s:%d vfifo_center: reset \
1947 right_edge[%u]: %d\n", __func__, __LINE__,
1948 i, right_edge[i]);
1949 }
1950
1951 /*
1952 * Reset sticky bit (except for bits where we have seen
1953 * both the left and right edge).
1954 */
1955 sticky_bit_chk = sticky_bit_chk << 1;
1956 if ((left_edge[i] != IO_IO_IN_DELAY_MAX + 1) &&
1957 (right_edge[i] != IO_IO_IN_DELAY_MAX + 1)) {
1958 sticky_bit_chk = sticky_bit_chk | 1;
1959 }
1960
1961 if (i == 0)
1962 break;
1963 }
1964
1965 /* Search for the right edge of the window for each bit */
1966 for (d = 0; d <= IO_DQS_IN_DELAY_MAX - start_dqs; d++) {
1967 scc_mgr_set_dqs_bus_in_delay(read_group, d + start_dqs);
1968 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
1969 uint32_t delay = d + start_dqs_en;
1970 if (delay > IO_DQS_EN_DELAY_MAX)
1971 delay = IO_DQS_EN_DELAY_MAX;
1972 scc_mgr_set_dqs_en_delay(read_group, delay);
1973 }
1974 scc_mgr_load_dqs(read_group);
1975
1976 writel(0, &sdr_scc_mgr->update);
1977
1978 /*
1979 * Stop searching when the read test doesn't pass AND when
1980 * we've seen a passing read on every bit.
1981 */
1982 if (use_read_test) {
1983 stop = !rw_mgr_mem_calibrate_read_test(rank_bgn,
1984 read_group, NUM_READ_PB_TESTS, PASS_ONE_BIT,
1985 &bit_chk, 0, 0);
1986 } else {
1987 rw_mgr_mem_calibrate_write_test(rank_bgn, write_group,
1988 0, PASS_ONE_BIT,
1989 &bit_chk, 0);
1990 bit_chk = bit_chk >> (RW_MGR_MEM_DQ_PER_READ_DQS *
1991 (read_group - (write_group *
1992 RW_MGR_MEM_IF_READ_DQS_WIDTH /
1993 RW_MGR_MEM_IF_WRITE_DQS_WIDTH)));
1994 stop = (bit_chk == 0);
1995 }
1996 sticky_bit_chk = sticky_bit_chk | bit_chk;
1997 stop = stop && (sticky_bit_chk == param->read_correct_mask);
1998
1999 debug_cond(DLEVEL == 2, "%s:%d vfifo_center(right): dtap=%u => %u == \
2000 %u && %u", __func__, __LINE__, d,
2001 sticky_bit_chk, param->read_correct_mask, stop);
2002
2003 if (stop == 1) {
2004 break;
2005 } else {
2006 for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
2007 if (bit_chk & 1) {
2008 /* Remember a passing test as
2009 the right_edge */
2010 right_edge[i] = d;
2011 } else {
2012 if (d != 0) {
2013 /* If a right edge has not been
2014 seen yet, then a future passing
2015 test will mark this edge as the
2016 left edge */
2017 if (right_edge[i] ==
2018 IO_IO_IN_DELAY_MAX + 1) {
2019 left_edge[i] = -(d + 1);
2020 }
2021 } else {
2022 /* d = 0 failed, but it passed
2023 when testing the left edge,
2024 so it must be marginal,
2025 set it to -1 */
2026 if (right_edge[i] ==
2027 IO_IO_IN_DELAY_MAX + 1 &&
2028 left_edge[i] !=
2029 IO_IO_IN_DELAY_MAX
2030 + 1) {
2031 right_edge[i] = -1;
2032 }
2033 /* If a right edge has not been
2034 seen yet, then a future passing
2035 test will mark this edge as the
2036 left edge */
2037 else if (right_edge[i] ==
2038 IO_IO_IN_DELAY_MAX +
2039 1) {
2040 left_edge[i] = -(d + 1);
2041 }
2042 }
2043 }
2044
2045 debug_cond(DLEVEL == 2, "%s:%d vfifo_center[r,\
2046 d=%u]: ", __func__, __LINE__, d);
2047 debug_cond(DLEVEL == 2, "bit_chk_test=%d left_edge[%u]: %d ",
2048 (int)(bit_chk & 1), i, left_edge[i]);
2049 debug_cond(DLEVEL == 2, "right_edge[%u]: %d\n", i,
2050 right_edge[i]);
2051 bit_chk = bit_chk >> 1;
2052 }
2053 }
2054 }
2055
2056 /* Check that all bits have a window */
2057 for (i = 0; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
2058 debug_cond(DLEVEL == 2, "%s:%d vfifo_center: left_edge[%u]: \
2059 %d right_edge[%u]: %d", __func__, __LINE__,
2060 i, left_edge[i], i, right_edge[i]);
2061 if ((left_edge[i] == IO_IO_IN_DELAY_MAX + 1) || (right_edge[i]
2062 == IO_IO_IN_DELAY_MAX + 1)) {
2063 /*
2064 * Restore delay chain settings before letting the loop
2065 * in rw_mgr_mem_calibrate_vfifo to retry different
2066 * dqs/ck relationships.
2067 */
2068 scc_mgr_set_dqs_bus_in_delay(read_group, start_dqs);
2069 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
2070 scc_mgr_set_dqs_en_delay(read_group,
2071 start_dqs_en);
2072 }
2073 scc_mgr_load_dqs(read_group);
2074 writel(0, &sdr_scc_mgr->update);
2075
2076 debug_cond(DLEVEL == 1, "%s:%d vfifo_center: failed to \
2077 find edge [%u]: %d %d", __func__, __LINE__,
2078 i, left_edge[i], right_edge[i]);
2079 if (use_read_test) {
2080 set_failing_group_stage(read_group *
2081 RW_MGR_MEM_DQ_PER_READ_DQS + i,
2082 CAL_STAGE_VFIFO,
2083 CAL_SUBSTAGE_VFIFO_CENTER);
2084 } else {
2085 set_failing_group_stage(read_group *
2086 RW_MGR_MEM_DQ_PER_READ_DQS + i,
2087 CAL_STAGE_VFIFO_AFTER_WRITES,
2088 CAL_SUBSTAGE_VFIFO_CENTER);
2089 }
2090 return 0;
2091 }
2092 }
2093
2094 /* Find middle of window for each DQ bit */
2095 mid_min = left_edge[0] - right_edge[0];
2096 min_index = 0;
2097 for (i = 1; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++) {
2098 mid = left_edge[i] - right_edge[i];
2099 if (mid < mid_min) {
2100 mid_min = mid;
2101 min_index = i;
2102 }
2103 }
2104
2105 /*
2106 * -mid_min/2 represents the amount that we need to move DQS.
2107 * If mid_min is odd and positive we'll need to add one to
2108 * make sure the rounding in further calculations is correct
2109 * (always bias to the right), so just add 1 for all positive values.
2110 */
2111 if (mid_min > 0)
2112 mid_min++;
2113
2114 mid_min = mid_min / 2;
2115
2116 debug_cond(DLEVEL == 1, "%s:%d vfifo_center: mid_min=%d (index=%u)\n",
2117 __func__, __LINE__, mid_min, min_index);
2118
2119 /* Determine the amount we can change DQS (which is -mid_min) */
2120 orig_mid_min = mid_min;
2121 new_dqs = start_dqs - mid_min;
2122 if (new_dqs > IO_DQS_IN_DELAY_MAX)
2123 new_dqs = IO_DQS_IN_DELAY_MAX;
2124 else if (new_dqs < 0)
2125 new_dqs = 0;
2126
2127 mid_min = start_dqs - new_dqs;
2128 debug_cond(DLEVEL == 1, "vfifo_center: new mid_min=%d new_dqs=%d\n",
2129 mid_min, new_dqs);
2130
2131 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
2132 if (start_dqs_en - mid_min > IO_DQS_EN_DELAY_MAX)
2133 mid_min += start_dqs_en - mid_min - IO_DQS_EN_DELAY_MAX;
2134 else if (start_dqs_en - mid_min < 0)
2135 mid_min += start_dqs_en - mid_min;
2136 }
2137 new_dqs = start_dqs - mid_min;
2138
2139 debug_cond(DLEVEL == 1, "vfifo_center: start_dqs=%d start_dqs_en=%d \
2140 new_dqs=%d mid_min=%d\n", start_dqs,
2141 IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS ? start_dqs_en : -1,
2142 new_dqs, mid_min);
2143
2144 /* Initialize data for export structures */
2145 dqs_margin = IO_IO_IN_DELAY_MAX + 1;
2146 dq_margin = IO_IO_IN_DELAY_MAX + 1;
2147
2148 /* add delay to bring centre of all DQ windows to the same "level" */
2149 for (i = 0, p = test_bgn; i < RW_MGR_MEM_DQ_PER_READ_DQS; i++, p++) {
2150 /* Use values before divide by 2 to reduce round off error */
2151 shift_dq = (left_edge[i] - right_edge[i] -
2152 (left_edge[min_index] - right_edge[min_index]))/2 +
2153 (orig_mid_min - mid_min);
2154
2155 debug_cond(DLEVEL == 2, "vfifo_center: before: \
2156 shift_dq[%u]=%d\n", i, shift_dq);
2157
2158 addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_IO_IN_DELAY_OFFSET;
2159 temp_dq_in_delay1 = readl(addr + (p << 2));
2160 temp_dq_in_delay2 = readl(addr + (i << 2));
2161
2162 if (shift_dq + (int32_t)temp_dq_in_delay1 >
2163 (int32_t)IO_IO_IN_DELAY_MAX) {
2164 shift_dq = (int32_t)IO_IO_IN_DELAY_MAX - temp_dq_in_delay2;
2165 } else if (shift_dq + (int32_t)temp_dq_in_delay1 < 0) {
2166 shift_dq = -(int32_t)temp_dq_in_delay1;
2167 }
2168 debug_cond(DLEVEL == 2, "vfifo_center: after: \
2169 shift_dq[%u]=%d\n", i, shift_dq);
2170 final_dq[i] = temp_dq_in_delay1 + shift_dq;
2171 scc_mgr_set_dq_in_delay(p, final_dq[i]);
2172 scc_mgr_load_dq(p);
2173
2174 debug_cond(DLEVEL == 2, "vfifo_center: margin[%u]=[%d,%d]\n", i,
2175 left_edge[i] - shift_dq + (-mid_min),
2176 right_edge[i] + shift_dq - (-mid_min));
2177 /* To determine values for export structures */
2178 if (left_edge[i] - shift_dq + (-mid_min) < dq_margin)
2179 dq_margin = left_edge[i] - shift_dq + (-mid_min);
2180
2181 if (right_edge[i] + shift_dq - (-mid_min) < dqs_margin)
2182 dqs_margin = right_edge[i] + shift_dq - (-mid_min);
2183 }
2184
2185 final_dqs = new_dqs;
2186 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS)
2187 final_dqs_en = start_dqs_en - mid_min;
2188
2189 /* Move DQS-en */
2190 if (IO_SHIFT_DQS_EN_WHEN_SHIFT_DQS) {
2191 scc_mgr_set_dqs_en_delay(read_group, final_dqs_en);
2192 scc_mgr_load_dqs(read_group);
2193 }
2194
2195 /* Move DQS */
2196 scc_mgr_set_dqs_bus_in_delay(read_group, final_dqs);
2197 scc_mgr_load_dqs(read_group);
2198 debug_cond(DLEVEL == 2, "%s:%d vfifo_center: dq_margin=%d \
2199 dqs_margin=%d", __func__, __LINE__,
2200 dq_margin, dqs_margin);
2201
2202 /*
2203 * Do not remove this line as it makes sure all of our decisions
2204 * have been applied. Apply the update bit.
2205 */
2206 writel(0, &sdr_scc_mgr->update);
2207
2208 return (dq_margin >= 0) && (dqs_margin >= 0);
2209 }
2210
2211 /*
2212 * calibrate the read valid prediction FIFO.
2213 *
2214 * - read valid prediction will consist of finding a good DQS enable phase,
2215 * DQS enable delay, DQS input phase, and DQS input delay.
2216 * - we also do a per-bit deskew on the DQ lines.
2217 */
2218 static uint32_t rw_mgr_mem_calibrate_vfifo(uint32_t read_group,
2219 uint32_t test_bgn)
2220 {
2221 uint32_t p, d, rank_bgn, sr;
2222 uint32_t dtaps_per_ptap;
2223 uint32_t tmp_delay;
2224 uint32_t bit_chk;
2225 uint32_t grp_calibrated;
2226 uint32_t write_group, write_test_bgn;
2227 uint32_t failed_substage;
2228
2229 debug("%s:%d: %u %u\n", __func__, __LINE__, read_group, test_bgn);
2230
2231 /* update info for sims */
2232 reg_file_set_stage(CAL_STAGE_VFIFO);
2233
2234 write_group = read_group;
2235 write_test_bgn = test_bgn;
2236
2237 /* USER Determine number of delay taps for each phase tap */
2238 dtaps_per_ptap = 0;
2239 tmp_delay = 0;
2240 while (tmp_delay < IO_DELAY_PER_OPA_TAP) {
2241 dtaps_per_ptap++;
2242 tmp_delay += IO_DELAY_PER_DQS_EN_DCHAIN_TAP;
2243 }
2244 dtaps_per_ptap--;
2245 tmp_delay = 0;
2246
2247 /* update info for sims */
2248 reg_file_set_group(read_group);
2249
2250 grp_calibrated = 0;
2251
2252 reg_file_set_sub_stage(CAL_SUBSTAGE_GUARANTEED_READ);
2253 failed_substage = CAL_SUBSTAGE_GUARANTEED_READ;
2254
2255 for (d = 0; d <= dtaps_per_ptap && grp_calibrated == 0; d += 2) {
2256 /*
2257 * In RLDRAMX we may be messing the delay of pins in
2258 * the same write group but outside of the current read
2259 * the group, but that's ok because we haven't
2260 * calibrated output side yet.
2261 */
2262 if (d > 0) {
2263 scc_mgr_apply_group_all_out_delay_add_all_ranks
2264 (write_group, write_test_bgn, d);
2265 }
2266
2267 for (p = 0; p <= IO_DQDQS_OUT_PHASE_MAX && grp_calibrated == 0;
2268 p++) {
2269 /* set a particular dqdqs phase */
2270 scc_mgr_set_dqdqs_output_phase_all_ranks(read_group, p);
2271
2272 debug_cond(DLEVEL == 1, "%s:%d calibrate_vfifo: g=%u \
2273 p=%u d=%u\n", __func__, __LINE__,
2274 read_group, p, d);
2275
2276 /*
2277 * Load up the patterns used by read calibration
2278 * using current DQDQS phase.
2279 */
2280 rw_mgr_mem_calibrate_read_load_patterns(0, 1);
2281 if (!(gbl->phy_debug_mode_flags &
2282 PHY_DEBUG_DISABLE_GUARANTEED_READ)) {
2283 if (!rw_mgr_mem_calibrate_read_test_patterns_all_ranks
2284 (read_group, 1, &bit_chk)) {
2285 debug_cond(DLEVEL == 1, "%s:%d Guaranteed read test failed:",
2286 __func__, __LINE__);
2287 debug_cond(DLEVEL == 1, " g=%u p=%u d=%u\n",
2288 read_group, p, d);
2289 break;
2290 }
2291 }
2292
2293 /* case:56390 */
2294 grp_calibrated = 1;
2295 if (rw_mgr_mem_calibrate_vfifo_find_dqs_en_phase_sweep_dq_in_delay
2296 (write_group, read_group, test_bgn)) {
2297 /*
2298 * USER Read per-bit deskew can be done on a
2299 * per shadow register basis.
2300 */
2301 for (rank_bgn = 0, sr = 0;
2302 rank_bgn < RW_MGR_MEM_NUMBER_OF_RANKS;
2303 rank_bgn += NUM_RANKS_PER_SHADOW_REG,
2304 ++sr) {
2305 /*
2306 * Determine if this set of ranks
2307 * should be skipped entirely.
2308 */
2309 if (!param->skip_shadow_regs[sr]) {
2310 /*
2311 * If doing read after write
2312 * calibration, do not update
2313 * FOM, now - do it then.
2314 */
2315 if (!rw_mgr_mem_calibrate_vfifo_center
2316 (rank_bgn, write_group,
2317 read_group, test_bgn, 1, 0)) {
2318 grp_calibrated = 0;
2319 failed_substage =
2320 CAL_SUBSTAGE_VFIFO_CENTER;
2321 }
2322 }
2323 }
2324 } else {
2325 grp_calibrated = 0;
2326 failed_substage = CAL_SUBSTAGE_DQS_EN_PHASE;
2327 }
2328 }
2329 }
2330
2331 if (grp_calibrated == 0) {
2332 set_failing_group_stage(write_group, CAL_STAGE_VFIFO,
2333 failed_substage);
2334 return 0;
2335 }
2336
2337 /*
2338 * Reset the delay chains back to zero if they have moved > 1
2339 * (check for > 1 because loop will increase d even when pass in
2340 * first case).
2341 */
2342 if (d > 2)
2343 scc_mgr_zero_group(write_group, write_test_bgn, 1);
2344
2345 return 1;
2346 }
2347
2348 /* VFIFO Calibration -- Read Deskew Calibration after write deskew */
2349 static uint32_t rw_mgr_mem_calibrate_vfifo_end(uint32_t read_group,
2350 uint32_t test_bgn)
2351 {
2352 uint32_t rank_bgn, sr;
2353 uint32_t grp_calibrated;
2354 uint32_t write_group;
2355
2356 debug("%s:%d %u %u", __func__, __LINE__, read_group, test_bgn);
2357
2358 /* update info for sims */
2359
2360 reg_file_set_stage(CAL_STAGE_VFIFO_AFTER_WRITES);
2361 reg_file_set_sub_stage(CAL_SUBSTAGE_VFIFO_CENTER);
2362
2363 write_group = read_group;
2364
2365 /* update info for sims */
2366 reg_file_set_group(read_group);
2367
2368 grp_calibrated = 1;
2369 /* Read per-bit deskew can be done on a per shadow register basis */
2370 for (rank_bgn = 0, sr = 0; rank_bgn < RW_MGR_MEM_NUMBER_OF_RANKS;
2371 rank_bgn += NUM_RANKS_PER_SHADOW_REG, ++sr) {
2372 /* Determine if this set of ranks should be skipped entirely */
2373 if (!param->skip_shadow_regs[sr]) {
2374 /* This is the last calibration round, update FOM here */
2375 if (!rw_mgr_mem_calibrate_vfifo_center(rank_bgn,
2376 write_group,
2377 read_group,
2378 test_bgn, 0,
2379 1)) {
2380 grp_calibrated = 0;
2381 }
2382 }
2383 }
2384
2385
2386 if (grp_calibrated == 0) {
2387 set_failing_group_stage(write_group,
2388 CAL_STAGE_VFIFO_AFTER_WRITES,
2389 CAL_SUBSTAGE_VFIFO_CENTER);
2390 return 0;
2391 }
2392
2393 return 1;
2394 }
2395
2396 /* Calibrate LFIFO to find smallest read latency */
2397 static uint32_t rw_mgr_mem_calibrate_lfifo(void)
2398 {
2399 uint32_t found_one;
2400 uint32_t bit_chk;
2401
2402 debug("%s:%d\n", __func__, __LINE__);
2403
2404 /* update info for sims */
2405 reg_file_set_stage(CAL_STAGE_LFIFO);
2406 reg_file_set_sub_stage(CAL_SUBSTAGE_READ_LATENCY);
2407
2408 /* Load up the patterns used by read calibration for all ranks */
2409 rw_mgr_mem_calibrate_read_load_patterns(0, 1);
2410 found_one = 0;
2411
2412 do {
2413 writel(gbl->curr_read_lat, &phy_mgr_cfg->phy_rlat);
2414 debug_cond(DLEVEL == 2, "%s:%d lfifo: read_lat=%u",
2415 __func__, __LINE__, gbl->curr_read_lat);
2416
2417 if (!rw_mgr_mem_calibrate_read_test_all_ranks(0,
2418 NUM_READ_TESTS,
2419 PASS_ALL_BITS,
2420 &bit_chk, 1)) {
2421 break;
2422 }
2423
2424 found_one = 1;
2425 /* reduce read latency and see if things are working */
2426 /* correctly */
2427 gbl->curr_read_lat--;
2428 } while (gbl->curr_read_lat > 0);
2429
2430 /* reset the fifos to get pointers to known state */
2431
2432 writel(0, &phy_mgr_cmd->fifo_reset);
2433
2434 if (found_one) {
2435 /* add a fudge factor to the read latency that was determined */
2436 gbl->curr_read_lat += 2;
2437 writel(gbl->curr_read_lat, &phy_mgr_cfg->phy_rlat);
2438 debug_cond(DLEVEL == 2, "%s:%d lfifo: success: using \
2439 read_lat=%u\n", __func__, __LINE__,
2440 gbl->curr_read_lat);
2441 return 1;
2442 } else {
2443 set_failing_group_stage(0xff, CAL_STAGE_LFIFO,
2444 CAL_SUBSTAGE_READ_LATENCY);
2445
2446 debug_cond(DLEVEL == 2, "%s:%d lfifo: failed at initial \
2447 read_lat=%u\n", __func__, __LINE__,
2448 gbl->curr_read_lat);
2449 return 0;
2450 }
2451 }
2452
2453 /*
2454 * issue write test command.
2455 * two variants are provided. one that just tests a write pattern and
2456 * another that tests datamask functionality.
2457 */
2458 static void rw_mgr_mem_calibrate_write_test_issue(uint32_t group,
2459 uint32_t test_dm)
2460 {
2461 uint32_t mcc_instruction;
2462 uint32_t quick_write_mode = (((STATIC_CALIB_STEPS) & CALIB_SKIP_WRITES) &&
2463 ENABLE_SUPER_QUICK_CALIBRATION);
2464 uint32_t rw_wl_nop_cycles;
2465 uint32_t addr;
2466
2467 /*
2468 * Set counter and jump addresses for the right
2469 * number of NOP cycles.
2470 * The number of supported NOP cycles can range from -1 to infinity
2471 * Three different cases are handled:
2472 *
2473 * 1. For a number of NOP cycles greater than 0, the RW Mgr looping
2474 * mechanism will be used to insert the right number of NOPs
2475 *
2476 * 2. For a number of NOP cycles equals to 0, the micro-instruction
2477 * issuing the write command will jump straight to the
2478 * micro-instruction that turns on DQS (for DDRx), or outputs write
2479 * data (for RLD), skipping
2480 * the NOP micro-instruction all together
2481 *
2482 * 3. A number of NOP cycles equal to -1 indicates that DQS must be
2483 * turned on in the same micro-instruction that issues the write
2484 * command. Then we need
2485 * to directly jump to the micro-instruction that sends out the data
2486 *
2487 * NOTE: Implementing this mechanism uses 2 RW Mgr jump-counters
2488 * (2 and 3). One jump-counter (0) is used to perform multiple
2489 * write-read operations.
2490 * one counter left to issue this command in "multiple-group" mode
2491 */
2492
2493 rw_wl_nop_cycles = gbl->rw_wl_nop_cycles;
2494
2495 if (rw_wl_nop_cycles == -1) {
2496 /*
2497 * CNTR 2 - We want to execute the special write operation that
2498 * turns on DQS right away and then skip directly to the
2499 * instruction that sends out the data. We set the counter to a
2500 * large number so that the jump is always taken.
2501 */
2502 writel(0xFF, &sdr_rw_load_mgr_regs->load_cntr2);
2503
2504 /* CNTR 3 - Not used */
2505 if (test_dm) {
2506 mcc_instruction = RW_MGR_LFSR_WR_RD_DM_BANK_0_WL_1;
2507 writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_DATA,
2508 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2509 writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_NOP,
2510 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
2511 } else {
2512 mcc_instruction = RW_MGR_LFSR_WR_RD_BANK_0_WL_1;
2513 writel(RW_MGR_LFSR_WR_RD_BANK_0_DATA,
2514 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2515 writel(RW_MGR_LFSR_WR_RD_BANK_0_NOP,
2516 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
2517 }
2518 } else if (rw_wl_nop_cycles == 0) {
2519 /*
2520 * CNTR 2 - We want to skip the NOP operation and go straight
2521 * to the DQS enable instruction. We set the counter to a large
2522 * number so that the jump is always taken.
2523 */
2524 writel(0xFF, &sdr_rw_load_mgr_regs->load_cntr2);
2525
2526 /* CNTR 3 - Not used */
2527 if (test_dm) {
2528 mcc_instruction = RW_MGR_LFSR_WR_RD_DM_BANK_0;
2529 writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_DQS,
2530 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2531 } else {
2532 mcc_instruction = RW_MGR_LFSR_WR_RD_BANK_0;
2533 writel(RW_MGR_LFSR_WR_RD_BANK_0_DQS,
2534 &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2535 }
2536 } else {
2537 /*
2538 * CNTR 2 - In this case we want to execute the next instruction
2539 * and NOT take the jump. So we set the counter to 0. The jump
2540 * address doesn't count.
2541 */
2542 writel(0x0, &sdr_rw_load_mgr_regs->load_cntr2);
2543 writel(0x0, &sdr_rw_load_jump_mgr_regs->load_jump_add2);
2544
2545 /*
2546 * CNTR 3 - Set the nop counter to the number of cycles we
2547 * need to loop for, minus 1.
2548 */
2549 writel(rw_wl_nop_cycles - 1, &sdr_rw_load_mgr_regs->load_cntr3);
2550 if (test_dm) {
2551 mcc_instruction = RW_MGR_LFSR_WR_RD_DM_BANK_0;
2552 writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_NOP,
2553 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
2554 } else {
2555 mcc_instruction = RW_MGR_LFSR_WR_RD_BANK_0;
2556 writel(RW_MGR_LFSR_WR_RD_BANK_0_NOP,
2557 &sdr_rw_load_jump_mgr_regs->load_jump_add3);
2558 }
2559 }
2560
2561 writel(0, SDR_PHYGRP_RWMGRGRP_ADDRESS |
2562 RW_MGR_RESET_READ_DATAPATH_OFFSET);
2563
2564 if (quick_write_mode)
2565 writel(0x08, &sdr_rw_load_mgr_regs->load_cntr0);
2566 else
2567 writel(0x40, &sdr_rw_load_mgr_regs->load_cntr0);
2568
2569 writel(mcc_instruction, &sdr_rw_load_jump_mgr_regs->load_jump_add0);
2570
2571 /*
2572 * CNTR 1 - This is used to ensure enough time elapses
2573 * for read data to come back.
2574 */
2575 writel(0x30, &sdr_rw_load_mgr_regs->load_cntr1);
2576
2577 if (test_dm) {
2578 writel(RW_MGR_LFSR_WR_RD_DM_BANK_0_WAIT,
2579 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
2580 } else {
2581 writel(RW_MGR_LFSR_WR_RD_BANK_0_WAIT,
2582 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
2583 }
2584
2585 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_RUN_SINGLE_GROUP_OFFSET;
2586 writel(mcc_instruction, addr + (group << 2));
2587 }
2588
2589 /* Test writes, can check for a single bit pass or multiple bit pass */
2590 static uint32_t rw_mgr_mem_calibrate_write_test(uint32_t rank_bgn,
2591 uint32_t write_group, uint32_t use_dm, uint32_t all_correct,
2592 uint32_t *bit_chk, uint32_t all_ranks)
2593 {
2594 uint32_t r;
2595 uint32_t correct_mask_vg;
2596 uint32_t tmp_bit_chk;
2597 uint32_t vg;
2598 uint32_t rank_end = all_ranks ? RW_MGR_MEM_NUMBER_OF_RANKS :
2599 (rank_bgn + NUM_RANKS_PER_SHADOW_REG);
2600 uint32_t addr_rw_mgr;
2601 uint32_t base_rw_mgr;
2602
2603 *bit_chk = param->write_correct_mask;
2604 correct_mask_vg = param->write_correct_mask_vg;
2605
2606 for (r = rank_bgn; r < rank_end; r++) {
2607 if (param->skip_ranks[r]) {
2608 /* request to skip the rank */
2609 continue;
2610 }
2611
2612 /* set rank */
2613 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_READ_WRITE);
2614
2615 tmp_bit_chk = 0;
2616 addr_rw_mgr = SDR_PHYGRP_RWMGRGRP_ADDRESS;
2617 for (vg = RW_MGR_MEM_VIRTUAL_GROUPS_PER_WRITE_DQS-1; ; vg--) {
2618 /* reset the fifos to get pointers to known state */
2619 writel(0, &phy_mgr_cmd->fifo_reset);
2620
2621 tmp_bit_chk = tmp_bit_chk <<
2622 (RW_MGR_MEM_DQ_PER_WRITE_DQS /
2623 RW_MGR_MEM_VIRTUAL_GROUPS_PER_WRITE_DQS);
2624 rw_mgr_mem_calibrate_write_test_issue(write_group *
2625 RW_MGR_MEM_VIRTUAL_GROUPS_PER_WRITE_DQS+vg,
2626 use_dm);
2627
2628 base_rw_mgr = readl(addr_rw_mgr);
2629 tmp_bit_chk = tmp_bit_chk | (correct_mask_vg & ~(base_rw_mgr));
2630 if (vg == 0)
2631 break;
2632 }
2633 *bit_chk &= tmp_bit_chk;
2634 }
2635
2636 if (all_correct) {
2637 set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
2638 debug_cond(DLEVEL == 2, "write_test(%u,%u,ALL) : %u == \
2639 %u => %lu", write_group, use_dm,
2640 *bit_chk, param->write_correct_mask,
2641 (long unsigned int)(*bit_chk ==
2642 param->write_correct_mask));
2643 return *bit_chk == param->write_correct_mask;
2644 } else {
2645 set_rank_and_odt_mask(0, RW_MGR_ODT_MODE_OFF);
2646 debug_cond(DLEVEL == 2, "write_test(%u,%u,ONE) : %u != ",
2647 write_group, use_dm, *bit_chk);
2648 debug_cond(DLEVEL == 2, "%lu" " => %lu", (long unsigned int)0,
2649 (long unsigned int)(*bit_chk != 0));
2650 return *bit_chk != 0x00;
2651 }
2652 }
2653
2654 /*
2655 * center all windows. do per-bit-deskew to possibly increase size of
2656 * certain windows.
2657 */
2658 static uint32_t rw_mgr_mem_calibrate_writes_center(uint32_t rank_bgn,
2659 uint32_t write_group, uint32_t test_bgn)
2660 {
2661 uint32_t i, p, min_index;
2662 int32_t d;
2663 /*
2664 * Store these as signed since there are comparisons with
2665 * signed numbers.
2666 */
2667 uint32_t bit_chk;
2668 uint32_t sticky_bit_chk;
2669 int32_t left_edge[RW_MGR_MEM_DQ_PER_WRITE_DQS];
2670 int32_t right_edge[RW_MGR_MEM_DQ_PER_WRITE_DQS];
2671 int32_t mid;
2672 int32_t mid_min, orig_mid_min;
2673 int32_t new_dqs, start_dqs, shift_dq;
2674 int32_t dq_margin, dqs_margin, dm_margin;
2675 uint32_t stop;
2676 uint32_t temp_dq_out1_delay;
2677 uint32_t addr;
2678
2679 debug("%s:%d %u %u", __func__, __LINE__, write_group, test_bgn);
2680
2681 dm_margin = 0;
2682
2683 addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_IO_OUT1_DELAY_OFFSET;
2684 start_dqs = readl(addr +
2685 (RW_MGR_MEM_DQ_PER_WRITE_DQS << 2));
2686
2687 /* per-bit deskew */
2688
2689 /*
2690 * set the left and right edge of each bit to an illegal value
2691 * use (IO_IO_OUT1_DELAY_MAX + 1) as an illegal value.
2692 */
2693 sticky_bit_chk = 0;
2694 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2695 left_edge[i] = IO_IO_OUT1_DELAY_MAX + 1;
2696 right_edge[i] = IO_IO_OUT1_DELAY_MAX + 1;
2697 }
2698
2699 /* Search for the left edge of the window for each bit */
2700 for (d = 0; d <= IO_IO_OUT1_DELAY_MAX; d++) {
2701 scc_mgr_apply_group_dq_out1_delay(write_group, d);
2702
2703 writel(0, &sdr_scc_mgr->update);
2704
2705 /*
2706 * Stop searching when the read test doesn't pass AND when
2707 * we've seen a passing read on every bit.
2708 */
2709 stop = !rw_mgr_mem_calibrate_write_test(rank_bgn, write_group,
2710 0, PASS_ONE_BIT, &bit_chk, 0);
2711 sticky_bit_chk = sticky_bit_chk | bit_chk;
2712 stop = stop && (sticky_bit_chk == param->write_correct_mask);
2713 debug_cond(DLEVEL == 2, "write_center(left): dtap=%d => %u \
2714 == %u && %u [bit_chk= %u ]\n",
2715 d, sticky_bit_chk, param->write_correct_mask,
2716 stop, bit_chk);
2717
2718 if (stop == 1) {
2719 break;
2720 } else {
2721 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2722 if (bit_chk & 1) {
2723 /*
2724 * Remember a passing test as the
2725 * left_edge.
2726 */
2727 left_edge[i] = d;
2728 } else {
2729 /*
2730 * If a left edge has not been seen
2731 * yet, then a future passing test will
2732 * mark this edge as the right edge.
2733 */
2734 if (left_edge[i] ==
2735 IO_IO_OUT1_DELAY_MAX + 1) {
2736 right_edge[i] = -(d + 1);
2737 }
2738 }
2739 debug_cond(DLEVEL == 2, "write_center[l,d=%d):", d);
2740 debug_cond(DLEVEL == 2, "bit_chk_test=%d left_edge[%u]: %d",
2741 (int)(bit_chk & 1), i, left_edge[i]);
2742 debug_cond(DLEVEL == 2, "right_edge[%u]: %d\n", i,
2743 right_edge[i]);
2744 bit_chk = bit_chk >> 1;
2745 }
2746 }
2747 }
2748
2749 /* Reset DQ delay chains to 0 */
2750 scc_mgr_apply_group_dq_out1_delay(0);
2751 sticky_bit_chk = 0;
2752 for (i = RW_MGR_MEM_DQ_PER_WRITE_DQS - 1;; i--) {
2753 debug_cond(DLEVEL == 2, "%s:%d write_center: left_edge[%u]: \
2754 %d right_edge[%u]: %d\n", __func__, __LINE__,
2755 i, left_edge[i], i, right_edge[i]);
2756
2757 /*
2758 * Check for cases where we haven't found the left edge,
2759 * which makes our assignment of the the right edge invalid.
2760 * Reset it to the illegal value.
2761 */
2762 if ((left_edge[i] == IO_IO_OUT1_DELAY_MAX + 1) &&
2763 (right_edge[i] != IO_IO_OUT1_DELAY_MAX + 1)) {
2764 right_edge[i] = IO_IO_OUT1_DELAY_MAX + 1;
2765 debug_cond(DLEVEL == 2, "%s:%d write_center: reset \
2766 right_edge[%u]: %d\n", __func__, __LINE__,
2767 i, right_edge[i]);
2768 }
2769
2770 /*
2771 * Reset sticky bit (except for bits where we have
2772 * seen the left edge).
2773 */
2774 sticky_bit_chk = sticky_bit_chk << 1;
2775 if ((left_edge[i] != IO_IO_OUT1_DELAY_MAX + 1))
2776 sticky_bit_chk = sticky_bit_chk | 1;
2777
2778 if (i == 0)
2779 break;
2780 }
2781
2782 /* Search for the right edge of the window for each bit */
2783 for (d = 0; d <= IO_IO_OUT1_DELAY_MAX - start_dqs; d++) {
2784 scc_mgr_apply_group_dqs_io_and_oct_out1(write_group,
2785 d + start_dqs);
2786
2787 writel(0, &sdr_scc_mgr->update);
2788
2789 /*
2790 * Stop searching when the read test doesn't pass AND when
2791 * we've seen a passing read on every bit.
2792 */
2793 stop = !rw_mgr_mem_calibrate_write_test(rank_bgn, write_group,
2794 0, PASS_ONE_BIT, &bit_chk, 0);
2795
2796 sticky_bit_chk = sticky_bit_chk | bit_chk;
2797 stop = stop && (sticky_bit_chk == param->write_correct_mask);
2798
2799 debug_cond(DLEVEL == 2, "write_center (right): dtap=%u => %u == \
2800 %u && %u\n", d, sticky_bit_chk,
2801 param->write_correct_mask, stop);
2802
2803 if (stop == 1) {
2804 if (d == 0) {
2805 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS;
2806 i++) {
2807 /* d = 0 failed, but it passed when
2808 testing the left edge, so it must be
2809 marginal, set it to -1 */
2810 if (right_edge[i] ==
2811 IO_IO_OUT1_DELAY_MAX + 1 &&
2812 left_edge[i] !=
2813 IO_IO_OUT1_DELAY_MAX + 1) {
2814 right_edge[i] = -1;
2815 }
2816 }
2817 }
2818 break;
2819 } else {
2820 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2821 if (bit_chk & 1) {
2822 /*
2823 * Remember a passing test as
2824 * the right_edge.
2825 */
2826 right_edge[i] = d;
2827 } else {
2828 if (d != 0) {
2829 /*
2830 * If a right edge has not
2831 * been seen yet, then a future
2832 * passing test will mark this
2833 * edge as the left edge.
2834 */
2835 if (right_edge[i] ==
2836 IO_IO_OUT1_DELAY_MAX + 1)
2837 left_edge[i] = -(d + 1);
2838 } else {
2839 /*
2840 * d = 0 failed, but it passed
2841 * when testing the left edge,
2842 * so it must be marginal, set
2843 * it to -1.
2844 */
2845 if (right_edge[i] ==
2846 IO_IO_OUT1_DELAY_MAX + 1 &&
2847 left_edge[i] !=
2848 IO_IO_OUT1_DELAY_MAX + 1)
2849 right_edge[i] = -1;
2850 /*
2851 * If a right edge has not been
2852 * seen yet, then a future
2853 * passing test will mark this
2854 * edge as the left edge.
2855 */
2856 else if (right_edge[i] ==
2857 IO_IO_OUT1_DELAY_MAX +
2858 1)
2859 left_edge[i] = -(d + 1);
2860 }
2861 }
2862 debug_cond(DLEVEL == 2, "write_center[r,d=%d):", d);
2863 debug_cond(DLEVEL == 2, "bit_chk_test=%d left_edge[%u]: %d",
2864 (int)(bit_chk & 1), i, left_edge[i]);
2865 debug_cond(DLEVEL == 2, "right_edge[%u]: %d\n", i,
2866 right_edge[i]);
2867 bit_chk = bit_chk >> 1;
2868 }
2869 }
2870 }
2871
2872 /* Check that all bits have a window */
2873 for (i = 0; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2874 debug_cond(DLEVEL == 2, "%s:%d write_center: left_edge[%u]: \
2875 %d right_edge[%u]: %d", __func__, __LINE__,
2876 i, left_edge[i], i, right_edge[i]);
2877 if ((left_edge[i] == IO_IO_OUT1_DELAY_MAX + 1) ||
2878 (right_edge[i] == IO_IO_OUT1_DELAY_MAX + 1)) {
2879 set_failing_group_stage(test_bgn + i,
2880 CAL_STAGE_WRITES,
2881 CAL_SUBSTAGE_WRITES_CENTER);
2882 return 0;
2883 }
2884 }
2885
2886 /* Find middle of window for each DQ bit */
2887 mid_min = left_edge[0] - right_edge[0];
2888 min_index = 0;
2889 for (i = 1; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++) {
2890 mid = left_edge[i] - right_edge[i];
2891 if (mid < mid_min) {
2892 mid_min = mid;
2893 min_index = i;
2894 }
2895 }
2896
2897 /*
2898 * -mid_min/2 represents the amount that we need to move DQS.
2899 * If mid_min is odd and positive we'll need to add one to
2900 * make sure the rounding in further calculations is correct
2901 * (always bias to the right), so just add 1 for all positive values.
2902 */
2903 if (mid_min > 0)
2904 mid_min++;
2905 mid_min = mid_min / 2;
2906 debug_cond(DLEVEL == 1, "%s:%d write_center: mid_min=%d\n", __func__,
2907 __LINE__, mid_min);
2908
2909 /* Determine the amount we can change DQS (which is -mid_min) */
2910 orig_mid_min = mid_min;
2911 new_dqs = start_dqs;
2912 mid_min = 0;
2913 debug_cond(DLEVEL == 1, "%s:%d write_center: start_dqs=%d new_dqs=%d \
2914 mid_min=%d\n", __func__, __LINE__, start_dqs, new_dqs, mid_min);
2915 /* Initialize data for export structures */
2916 dqs_margin = IO_IO_OUT1_DELAY_MAX + 1;
2917 dq_margin = IO_IO_OUT1_DELAY_MAX + 1;
2918
2919 /* add delay to bring centre of all DQ windows to the same "level" */
2920 for (i = 0, p = test_bgn; i < RW_MGR_MEM_DQ_PER_WRITE_DQS; i++, p++) {
2921 /* Use values before divide by 2 to reduce round off error */
2922 shift_dq = (left_edge[i] - right_edge[i] -
2923 (left_edge[min_index] - right_edge[min_index]))/2 +
2924 (orig_mid_min - mid_min);
2925
2926 debug_cond(DLEVEL == 2, "%s:%d write_center: before: shift_dq \
2927 [%u]=%d\n", __func__, __LINE__, i, shift_dq);
2928
2929 addr = SDR_PHYGRP_SCCGRP_ADDRESS | SCC_MGR_IO_OUT1_DELAY_OFFSET;
2930 temp_dq_out1_delay = readl(addr + (i << 2));
2931 if (shift_dq + (int32_t)temp_dq_out1_delay >
2932 (int32_t)IO_IO_OUT1_DELAY_MAX) {
2933 shift_dq = (int32_t)IO_IO_OUT1_DELAY_MAX - temp_dq_out1_delay;
2934 } else if (shift_dq + (int32_t)temp_dq_out1_delay < 0) {
2935 shift_dq = -(int32_t)temp_dq_out1_delay;
2936 }
2937 debug_cond(DLEVEL == 2, "write_center: after: shift_dq[%u]=%d\n",
2938 i, shift_dq);
2939 scc_mgr_set_dq_out1_delay(i, temp_dq_out1_delay + shift_dq);
2940 scc_mgr_load_dq(i);
2941
2942 debug_cond(DLEVEL == 2, "write_center: margin[%u]=[%d,%d]\n", i,
2943 left_edge[i] - shift_dq + (-mid_min),
2944 right_edge[i] + shift_dq - (-mid_min));
2945 /* To determine values for export structures */
2946 if (left_edge[i] - shift_dq + (-mid_min) < dq_margin)
2947 dq_margin = left_edge[i] - shift_dq + (-mid_min);
2948
2949 if (right_edge[i] + shift_dq - (-mid_min) < dqs_margin)
2950 dqs_margin = right_edge[i] + shift_dq - (-mid_min);
2951 }
2952
2953 /* Move DQS */
2954 scc_mgr_apply_group_dqs_io_and_oct_out1(write_group, new_dqs);
2955 writel(0, &sdr_scc_mgr->update);
2956
2957 /* Centre DM */
2958 debug_cond(DLEVEL == 2, "%s:%d write_center: DM\n", __func__, __LINE__);
2959
2960 /*
2961 * set the left and right edge of each bit to an illegal value,
2962 * use (IO_IO_OUT1_DELAY_MAX + 1) as an illegal value,
2963 */
2964 left_edge[0] = IO_IO_OUT1_DELAY_MAX + 1;
2965 right_edge[0] = IO_IO_OUT1_DELAY_MAX + 1;
2966 int32_t bgn_curr = IO_IO_OUT1_DELAY_MAX + 1;
2967 int32_t end_curr = IO_IO_OUT1_DELAY_MAX + 1;
2968 int32_t bgn_best = IO_IO_OUT1_DELAY_MAX + 1;
2969 int32_t end_best = IO_IO_OUT1_DELAY_MAX + 1;
2970 int32_t win_best = 0;
2971
2972 /* Search for the/part of the window with DM shift */
2973 for (d = IO_IO_OUT1_DELAY_MAX; d >= 0; d -= DELTA_D) {
2974 scc_mgr_apply_group_dm_out1_delay(d);
2975 writel(0, &sdr_scc_mgr->update);
2976
2977 if (rw_mgr_mem_calibrate_write_test(rank_bgn, write_group, 1,
2978 PASS_ALL_BITS, &bit_chk,
2979 0)) {
2980 /* USE Set current end of the window */
2981 end_curr = -d;
2982 /*
2983 * If a starting edge of our window has not been seen
2984 * this is our current start of the DM window.
2985 */
2986 if (bgn_curr == IO_IO_OUT1_DELAY_MAX + 1)
2987 bgn_curr = -d;
2988
2989 /*
2990 * If current window is bigger than best seen.
2991 * Set best seen to be current window.
2992 */
2993 if ((end_curr-bgn_curr+1) > win_best) {
2994 win_best = end_curr-bgn_curr+1;
2995 bgn_best = bgn_curr;
2996 end_best = end_curr;
2997 }
2998 } else {
2999 /* We just saw a failing test. Reset temp edge */
3000 bgn_curr = IO_IO_OUT1_DELAY_MAX + 1;
3001 end_curr = IO_IO_OUT1_DELAY_MAX + 1;
3002 }
3003 }
3004
3005
3006 /* Reset DM delay chains to 0 */
3007 scc_mgr_apply_group_dm_out1_delay(0);
3008
3009 /*
3010 * Check to see if the current window nudges up aganist 0 delay.
3011 * If so we need to continue the search by shifting DQS otherwise DQS
3012 * search begins as a new search. */
3013 if (end_curr != 0) {
3014 bgn_curr = IO_IO_OUT1_DELAY_MAX + 1;
3015 end_curr = IO_IO_OUT1_DELAY_MAX + 1;
3016 }
3017
3018 /* Search for the/part of the window with DQS shifts */
3019 for (d = 0; d <= IO_IO_OUT1_DELAY_MAX - new_dqs; d += DELTA_D) {
3020 /*
3021 * Note: This only shifts DQS, so are we limiting ourselve to
3022 * width of DQ unnecessarily.
3023 */
3024 scc_mgr_apply_group_dqs_io_and_oct_out1(write_group,
3025 d + new_dqs);
3026
3027 writel(0, &sdr_scc_mgr->update);
3028 if (rw_mgr_mem_calibrate_write_test(rank_bgn, write_group, 1,
3029 PASS_ALL_BITS, &bit_chk,
3030 0)) {
3031 /* USE Set current end of the window */
3032 end_curr = d;
3033 /*
3034 * If a beginning edge of our window has not been seen
3035 * this is our current begin of the DM window.
3036 */
3037 if (bgn_curr == IO_IO_OUT1_DELAY_MAX + 1)
3038 bgn_curr = d;
3039
3040 /*
3041 * If current window is bigger than best seen. Set best
3042 * seen to be current window.
3043 */
3044 if ((end_curr-bgn_curr+1) > win_best) {
3045 win_best = end_curr-bgn_curr+1;
3046 bgn_best = bgn_curr;
3047 end_best = end_curr;
3048 }
3049 } else {
3050 /* We just saw a failing test. Reset temp edge */
3051 bgn_curr = IO_IO_OUT1_DELAY_MAX + 1;
3052 end_curr = IO_IO_OUT1_DELAY_MAX + 1;
3053
3054 /* Early exit optimization: if ther remaining delay
3055 chain space is less than already seen largest window
3056 we can exit */
3057 if ((win_best-1) >
3058 (IO_IO_OUT1_DELAY_MAX - new_dqs - d)) {
3059 break;
3060 }
3061 }
3062 }
3063
3064 /* assign left and right edge for cal and reporting; */
3065 left_edge[0] = -1*bgn_best;
3066 right_edge[0] = end_best;
3067
3068 debug_cond(DLEVEL == 2, "%s:%d dm_calib: left=%d right=%d\n", __func__,
3069 __LINE__, left_edge[0], right_edge[0]);
3070
3071 /* Move DQS (back to orig) */
3072 scc_mgr_apply_group_dqs_io_and_oct_out1(write_group, new_dqs);
3073
3074 /* Move DM */
3075
3076 /* Find middle of window for the DM bit */
3077 mid = (left_edge[0] - right_edge[0]) / 2;
3078
3079 /* only move right, since we are not moving DQS/DQ */
3080 if (mid < 0)
3081 mid = 0;
3082
3083 /* dm_marign should fail if we never find a window */
3084 if (win_best == 0)
3085 dm_margin = -1;
3086 else
3087 dm_margin = left_edge[0] - mid;
3088
3089 scc_mgr_apply_group_dm_out1_delay(mid);
3090 writel(0, &sdr_scc_mgr->update);
3091
3092 debug_cond(DLEVEL == 2, "%s:%d dm_calib: left=%d right=%d mid=%d \
3093 dm_margin=%d\n", __func__, __LINE__, left_edge[0],
3094 right_edge[0], mid, dm_margin);
3095 /* Export values */
3096 gbl->fom_out += dq_margin + dqs_margin;
3097
3098 debug_cond(DLEVEL == 2, "%s:%d write_center: dq_margin=%d \
3099 dqs_margin=%d dm_margin=%d\n", __func__, __LINE__,
3100 dq_margin, dqs_margin, dm_margin);
3101
3102 /*
3103 * Do not remove this line as it makes sure all of our
3104 * decisions have been applied.
3105 */
3106 writel(0, &sdr_scc_mgr->update);
3107 return (dq_margin >= 0) && (dqs_margin >= 0) && (dm_margin >= 0);
3108 }
3109
3110 /* calibrate the write operations */
3111 static uint32_t rw_mgr_mem_calibrate_writes(uint32_t rank_bgn, uint32_t g,
3112 uint32_t test_bgn)
3113 {
3114 /* update info for sims */
3115 debug("%s:%d %u %u\n", __func__, __LINE__, g, test_bgn);
3116
3117 reg_file_set_stage(CAL_STAGE_WRITES);
3118 reg_file_set_sub_stage(CAL_SUBSTAGE_WRITES_CENTER);
3119
3120 reg_file_set_group(g);
3121
3122 if (!rw_mgr_mem_calibrate_writes_center(rank_bgn, g, test_bgn)) {
3123 set_failing_group_stage(g, CAL_STAGE_WRITES,
3124 CAL_SUBSTAGE_WRITES_CENTER);
3125 return 0;
3126 }
3127
3128 return 1;
3129 }
3130
3131 /* precharge all banks and activate row 0 in bank "000..." and bank "111..." */
3132 static void mem_precharge_and_activate(void)
3133 {
3134 uint32_t r;
3135
3136 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS; r++) {
3137 if (param->skip_ranks[r]) {
3138 /* request to skip the rank */
3139 continue;
3140 }
3141
3142 /* set rank */
3143 set_rank_and_odt_mask(r, RW_MGR_ODT_MODE_OFF);
3144
3145 /* precharge all banks ... */
3146 writel(RW_MGR_PRECHARGE_ALL, SDR_PHYGRP_RWMGRGRP_ADDRESS |
3147 RW_MGR_RUN_SINGLE_GROUP_OFFSET);
3148
3149 writel(0x0F, &sdr_rw_load_mgr_regs->load_cntr0);
3150 writel(RW_MGR_ACTIVATE_0_AND_1_WAIT1,
3151 &sdr_rw_load_jump_mgr_regs->load_jump_add0);
3152
3153 writel(0x0F, &sdr_rw_load_mgr_regs->load_cntr1);
3154 writel(RW_MGR_ACTIVATE_0_AND_1_WAIT2,
3155 &sdr_rw_load_jump_mgr_regs->load_jump_add1);
3156
3157 /* activate rows */
3158 writel(RW_MGR_ACTIVATE_0_AND_1, SDR_PHYGRP_RWMGRGRP_ADDRESS |
3159 RW_MGR_RUN_SINGLE_GROUP_OFFSET);
3160 }
3161 }
3162
3163 /* Configure various memory related parameters. */
3164 static void mem_config(void)
3165 {
3166 uint32_t rlat, wlat;
3167 uint32_t rw_wl_nop_cycles;
3168 uint32_t max_latency;
3169
3170 debug("%s:%d\n", __func__, __LINE__);
3171 /* read in write and read latency */
3172 wlat = readl(&data_mgr->t_wl_add);
3173 wlat += readl(&data_mgr->mem_t_add);
3174
3175 /* WL for hard phy does not include additive latency */
3176
3177 /*
3178 * add addtional write latency to offset the address/command extra
3179 * clock cycle. We change the AC mux setting causing AC to be delayed
3180 * by one mem clock cycle. Only do this for DDR3
3181 */
3182 wlat = wlat + 1;
3183
3184 rlat = readl(&data_mgr->t_rl_add);
3185
3186 rw_wl_nop_cycles = wlat - 2;
3187 gbl->rw_wl_nop_cycles = rw_wl_nop_cycles;
3188
3189 /*
3190 * For AV/CV, lfifo is hardened and always runs at full rate so
3191 * max latency in AFI clocks, used here, is correspondingly smaller.
3192 */
3193 max_latency = (1<<MAX_LATENCY_COUNT_WIDTH)/1 - 1;
3194 /* configure for a burst length of 8 */
3195
3196 /* write latency */
3197 /* Adjust Write Latency for Hard PHY */
3198 wlat = wlat + 1;
3199
3200 /* set a pretty high read latency initially */
3201 gbl->curr_read_lat = rlat + 16;
3202
3203 if (gbl->curr_read_lat > max_latency)
3204 gbl->curr_read_lat = max_latency;
3205
3206 writel(gbl->curr_read_lat, &phy_mgr_cfg->phy_rlat);
3207
3208 /* advertise write latency */
3209 gbl->curr_write_lat = wlat;
3210 writel(wlat - 2, &phy_mgr_cfg->afi_wlat);
3211
3212 /* initialize bit slips */
3213 mem_precharge_and_activate();
3214 }
3215
3216 /* Set VFIFO and LFIFO to instant-on settings in skip calibration mode */
3217 static void mem_skip_calibrate(void)
3218 {
3219 uint32_t vfifo_offset;
3220 uint32_t i, j, r;
3221
3222 debug("%s:%d\n", __func__, __LINE__);
3223 /* Need to update every shadow register set used by the interface */
3224 for (r = 0; r < RW_MGR_MEM_NUMBER_OF_RANKS;
3225 r += NUM_RANKS_PER_SHADOW_REG) {
3226 /*
3227 * Set output phase alignment settings appropriate for
3228 * skip calibration.
3229 */
3230 for (i = 0; i < RW_MGR_MEM_IF_READ_DQS_WIDTH; i++) {
3231 scc_mgr_set_dqs_en_phase(i, 0);
3232 #if IO_DLL_CHAIN_LENGTH == 6
3233 scc_mgr_set_dqdqs_output_phase(i, 6);
3234 #else
3235 scc_mgr_set_dqdqs_output_phase(i, 7);
3236 #endif
3237 /*
3238 * Case:33398
3239 *
3240 * Write data arrives to the I/O two cycles before write
3241 * latency is reached (720 deg).
3242 * -> due to bit-slip in a/c bus
3243 * -> to allow board skew where dqs is longer than ck
3244 * -> how often can this happen!?
3245 * -> can claim back some ptaps for high freq
3246 * support if we can relax this, but i digress...
3247 *
3248 * The write_clk leads mem_ck by 90 deg
3249 * The minimum ptap of the OPA is 180 deg
3250 * Each ptap has (360 / IO_DLL_CHAIN_LENGH) deg of delay
3251 * The write_clk is always delayed by 2 ptaps
3252 *
3253 * Hence, to make DQS aligned to CK, we need to delay
3254 * DQS by:
3255 * (720 - 90 - 180 - 2 * (360 / IO_DLL_CHAIN_LENGTH))
3256 *
3257 * Dividing the above by (360 / IO_DLL_CHAIN_LENGTH)
3258 * gives us the number of ptaps, which simplies to:
3259 *
3260 * (1.25 * IO_DLL_CHAIN_LENGTH - 2)
3261 */
3262 scc_mgr_set_dqdqs_output_phase(i, (1.25 *
3263 IO_DLL_CHAIN_LENGTH - 2));
3264 }
3265 writel(0xff, &sdr_scc_mgr->dqs_ena);
3266 writel(0xff, &sdr_scc_mgr->dqs_io_ena);
3267
3268 for (i = 0; i < RW_MGR_MEM_IF_WRITE_DQS_WIDTH; i++) {
3269 writel(i, SDR_PHYGRP_SCCGRP_ADDRESS |
3270 SCC_MGR_GROUP_COUNTER_OFFSET);
3271 }
3272 writel(0xff, &sdr_scc_mgr->dq_ena);
3273 writel(0xff, &sdr_scc_mgr->dm_ena);
3274 writel(0, &sdr_scc_mgr->update);
3275 }
3276
3277 /* Compensate for simulation model behaviour */
3278 for (i = 0; i < RW_MGR_MEM_IF_READ_DQS_WIDTH; i++) {
3279 scc_mgr_set_dqs_bus_in_delay(i, 10);
3280 scc_mgr_load_dqs(i);
3281 }
3282 writel(0, &sdr_scc_mgr->update);
3283
3284 /*
3285 * ArriaV has hard FIFOs that can only be initialized by incrementing
3286 * in sequencer.
3287 */
3288 vfifo_offset = CALIB_VFIFO_OFFSET;
3289 for (j = 0; j < vfifo_offset; j++) {
3290 writel(0xff, &phy_mgr_cmd->inc_vfifo_hard_phy);
3291 }
3292 writel(0, &phy_mgr_cmd->fifo_reset);
3293
3294 /*
3295 * For ACV with hard lfifo, we get the skip-cal setting from
3296 * generation-time constant.
3297 */
3298 gbl->curr_read_lat = CALIB_LFIFO_OFFSET;
3299 writel(gbl->curr_read_lat, &phy_mgr_cfg->phy_rlat);
3300 }
3301
3302 /* Memory calibration entry point */
3303 static uint32_t mem_calibrate(void)
3304 {
3305 uint32_t i;
3306 uint32_t rank_bgn, sr;
3307 uint32_t write_group, write_test_bgn;
3308 uint32_t read_group, read_test_bgn;
3309 uint32_t run_groups, current_run;
3310 uint32_t failing_groups = 0;
3311 uint32_t group_failed = 0;
3312 uint32_t sr_failed = 0;
3313
3314 debug("%s:%d\n", __func__, __LINE__);
3315 /* Initialize the data settings */
3316
3317 gbl->error_substage = CAL_SUBSTAGE_NIL;
3318 gbl->error_stage = CAL_STAGE_NIL;
3319 gbl->error_group = 0xff;
3320 gbl->fom_in = 0;
3321 gbl->fom_out = 0;
3322
3323 mem_config();
3324
3325 for (i = 0; i < RW_MGR_MEM_IF_READ_DQS_WIDTH; i++) {
3326 writel(i, SDR_PHYGRP_SCCGRP_ADDRESS |
3327 SCC_MGR_GROUP_COUNTER_OFFSET);
3328 scc_set_bypass_mode(i);
3329 }
3330
3331 if ((dyn_calib_steps & CALIB_SKIP_ALL) == CALIB_SKIP_ALL) {
3332 /*
3333 * Set VFIFO and LFIFO to instant-on settings in skip
3334 * calibration mode.
3335 */
3336 mem_skip_calibrate();
3337 } else {
3338 for (i = 0; i < NUM_CALIB_REPEAT; i++) {
3339 /*
3340 * Zero all delay chain/phase settings for all
3341 * groups and all shadow register sets.
3342 */
3343 scc_mgr_zero_all();
3344
3345 run_groups = ~param->skip_groups;
3346
3347 for (write_group = 0, write_test_bgn = 0; write_group
3348 < RW_MGR_MEM_IF_WRITE_DQS_WIDTH; write_group++,
3349 write_test_bgn += RW_MGR_MEM_DQ_PER_WRITE_DQS) {
3350 /* Initialized the group failure */
3351 group_failed = 0;
3352
3353 current_run = run_groups & ((1 <<
3354 RW_MGR_NUM_DQS_PER_WRITE_GROUP) - 1);
3355 run_groups = run_groups >>
3356 RW_MGR_NUM_DQS_PER_WRITE_GROUP;
3357
3358 if (current_run == 0)
3359 continue;
3360
3361 writel(write_group, SDR_PHYGRP_SCCGRP_ADDRESS |
3362 SCC_MGR_GROUP_COUNTER_OFFSET);
3363 scc_mgr_zero_group(write_group, write_test_bgn,
3364 0);
3365
3366 for (read_group = write_group *
3367 RW_MGR_MEM_IF_READ_DQS_WIDTH /
3368 RW_MGR_MEM_IF_WRITE_DQS_WIDTH,
3369 read_test_bgn = 0;
3370 read_group < (write_group + 1) *
3371 RW_MGR_MEM_IF_READ_DQS_WIDTH /
3372 RW_MGR_MEM_IF_WRITE_DQS_WIDTH &&
3373 group_failed == 0;
3374 read_group++, read_test_bgn +=
3375 RW_MGR_MEM_DQ_PER_READ_DQS) {
3376 /* Calibrate the VFIFO */
3377 if (!((STATIC_CALIB_STEPS) &
3378 CALIB_SKIP_VFIFO)) {
3379 if (!rw_mgr_mem_calibrate_vfifo
3380 (read_group,
3381 read_test_bgn)) {
3382 group_failed = 1;
3383
3384 if (!(gbl->
3385 phy_debug_mode_flags &
3386 PHY_DEBUG_SWEEP_ALL_GROUPS)) {
3387 return 0;
3388 }
3389 }
3390 }
3391 }
3392
3393 /* Calibrate the output side */
3394 if (group_failed == 0) {
3395 for (rank_bgn = 0, sr = 0; rank_bgn
3396 < RW_MGR_MEM_NUMBER_OF_RANKS;
3397 rank_bgn +=
3398 NUM_RANKS_PER_SHADOW_REG,
3399 ++sr) {
3400 sr_failed = 0;
3401 if (!((STATIC_CALIB_STEPS) &
3402 CALIB_SKIP_WRITES)) {
3403 if ((STATIC_CALIB_STEPS)
3404 & CALIB_SKIP_DELAY_SWEEPS) {
3405 /* not needed in quick mode! */
3406 } else {
3407 /*
3408 * Determine if this set of
3409 * ranks should be skipped
3410 * entirely.
3411 */
3412 if (!param->skip_shadow_regs[sr]) {
3413 if (!rw_mgr_mem_calibrate_writes
3414 (rank_bgn, write_group,
3415 write_test_bgn)) {
3416 sr_failed = 1;
3417 if (!(gbl->
3418 phy_debug_mode_flags &
3419 PHY_DEBUG_SWEEP_ALL_GROUPS)) {
3420 return 0;
3421 }
3422 }
3423 }
3424 }
3425 }
3426 if (sr_failed != 0)
3427 group_failed = 1;
3428 }
3429 }
3430
3431 if (group_failed == 0) {
3432 for (read_group = write_group *
3433 RW_MGR_MEM_IF_READ_DQS_WIDTH /
3434 RW_MGR_MEM_IF_WRITE_DQS_WIDTH,
3435 read_test_bgn = 0;
3436 read_group < (write_group + 1)
3437 * RW_MGR_MEM_IF_READ_DQS_WIDTH
3438 / RW_MGR_MEM_IF_WRITE_DQS_WIDTH &&
3439 group_failed == 0;
3440 read_group++, read_test_bgn +=
3441 RW_MGR_MEM_DQ_PER_READ_DQS) {
3442 if (!((STATIC_CALIB_STEPS) &
3443 CALIB_SKIP_WRITES)) {
3444 if (!rw_mgr_mem_calibrate_vfifo_end
3445 (read_group, read_test_bgn)) {
3446 group_failed = 1;
3447
3448 if (!(gbl->phy_debug_mode_flags
3449 & PHY_DEBUG_SWEEP_ALL_GROUPS)) {
3450 return 0;
3451 }
3452 }
3453 }
3454 }
3455 }
3456
3457 if (group_failed != 0)
3458 failing_groups++;
3459 }
3460
3461 /*
3462 * USER If there are any failing groups then report
3463 * the failure.
3464 */
3465 if (failing_groups != 0)
3466 return 0;
3467
3468 /* Calibrate the LFIFO */
3469 if (!((STATIC_CALIB_STEPS) & CALIB_SKIP_LFIFO)) {
3470 /*
3471 * If we're skipping groups as part of debug,
3472 * don't calibrate LFIFO.
3473 */
3474 if (param->skip_groups == 0) {
3475 if (!rw_mgr_mem_calibrate_lfifo())
3476 return 0;
3477 }
3478 }
3479 }
3480 }
3481
3482 /*
3483 * Do not remove this line as it makes sure all of our decisions
3484 * have been applied.
3485 */
3486 writel(0, &sdr_scc_mgr->update);
3487 return 1;
3488 }
3489
3490 static uint32_t run_mem_calibrate(void)
3491 {
3492 uint32_t pass;
3493 uint32_t debug_info;
3494
3495 debug("%s:%d\n", __func__, __LINE__);
3496
3497 /* Reset pass/fail status shown on afi_cal_success/fail */
3498 writel(PHY_MGR_CAL_RESET, &phy_mgr_cfg->cal_status);
3499
3500 /* stop tracking manger */
3501 uint32_t ctrlcfg = readl(&sdr_ctrl->ctrl_cfg);
3502
3503 writel(ctrlcfg & 0xFFBFFFFF, &sdr_ctrl->ctrl_cfg);
3504
3505 initialize();
3506 rw_mgr_mem_initialize();
3507
3508 pass = mem_calibrate();
3509
3510 mem_precharge_and_activate();
3511 writel(0, &phy_mgr_cmd->fifo_reset);
3512
3513 /*
3514 * Handoff:
3515 * Don't return control of the PHY back to AFI when in debug mode.
3516 */
3517 if ((gbl->phy_debug_mode_flags & PHY_DEBUG_IN_DEBUG_MODE) == 0) {
3518 rw_mgr_mem_handoff();
3519 /*
3520 * In Hard PHY this is a 2-bit control:
3521 * 0: AFI Mux Select
3522 * 1: DDIO Mux Select
3523 */
3524 writel(0x2, &phy_mgr_cfg->mux_sel);
3525 }
3526
3527 writel(ctrlcfg, &sdr_ctrl->ctrl_cfg);
3528
3529 if (pass) {
3530 printf("%s: CALIBRATION PASSED\n", __FILE__);
3531
3532 gbl->fom_in /= 2;
3533 gbl->fom_out /= 2;
3534
3535 if (gbl->fom_in > 0xff)
3536 gbl->fom_in = 0xff;
3537
3538 if (gbl->fom_out > 0xff)
3539 gbl->fom_out = 0xff;
3540
3541 /* Update the FOM in the register file */
3542 debug_info = gbl->fom_in;
3543 debug_info |= gbl->fom_out << 8;
3544 writel(debug_info, &sdr_reg_file->fom);
3545
3546 writel(debug_info, &phy_mgr_cfg->cal_debug_info);
3547 writel(PHY_MGR_CAL_SUCCESS, &phy_mgr_cfg->cal_status);
3548 } else {
3549 printf("%s: CALIBRATION FAILED\n", __FILE__);
3550
3551 debug_info = gbl->error_stage;
3552 debug_info |= gbl->error_substage << 8;
3553 debug_info |= gbl->error_group << 16;
3554
3555 writel(debug_info, &sdr_reg_file->failing_stage);
3556 writel(debug_info, &phy_mgr_cfg->cal_debug_info);
3557 writel(PHY_MGR_CAL_FAIL, &phy_mgr_cfg->cal_status);
3558
3559 /* Update the failing group/stage in the register file */
3560 debug_info = gbl->error_stage;
3561 debug_info |= gbl->error_substage << 8;
3562 debug_info |= gbl->error_group << 16;
3563 writel(debug_info, &sdr_reg_file->failing_stage);
3564 }
3565
3566 return pass;
3567 }
3568
3569 /**
3570 * hc_initialize_rom_data() - Initialize ROM data
3571 *
3572 * Initialize ROM data.
3573 */
3574 static void hc_initialize_rom_data(void)
3575 {
3576 u32 i, addr;
3577
3578 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_INST_ROM_WRITE_OFFSET;
3579 for (i = 0; i < ARRAY_SIZE(inst_rom_init); i++)
3580 writel(inst_rom_init[i], addr + (i << 2));
3581
3582 addr = SDR_PHYGRP_RWMGRGRP_ADDRESS | RW_MGR_AC_ROM_WRITE_OFFSET;
3583 for (i = 0; i < ARRAY_SIZE(ac_rom_init); i++)
3584 writel(ac_rom_init[i], addr + (i << 2));
3585 }
3586
3587 /**
3588 * initialize_reg_file() - Initialize SDR register file
3589 *
3590 * Initialize SDR register file.
3591 */
3592 static void initialize_reg_file(void)
3593 {
3594 /* Initialize the register file with the correct data */
3595 writel(REG_FILE_INIT_SEQ_SIGNATURE, &sdr_reg_file->signature);
3596 writel(0, &sdr_reg_file->debug_data_addr);
3597 writel(0, &sdr_reg_file->cur_stage);
3598 writel(0, &sdr_reg_file->fom);
3599 writel(0, &sdr_reg_file->failing_stage);
3600 writel(0, &sdr_reg_file->debug1);
3601 writel(0, &sdr_reg_file->debug2);
3602 }
3603
3604 /**
3605 * initialize_hps_phy() - Initialize HPS PHY
3606 *
3607 * Initialize HPS PHY.
3608 */
3609 static void initialize_hps_phy(void)
3610 {
3611 uint32_t reg;
3612 /*
3613 * Tracking also gets configured here because it's in the
3614 * same register.
3615 */
3616 uint32_t trk_sample_count = 7500;
3617 uint32_t trk_long_idle_sample_count = (10 << 16) | 100;
3618 /*
3619 * Format is number of outer loops in the 16 MSB, sample
3620 * count in 16 LSB.
3621 */
3622
3623 reg = 0;
3624 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_ACDELAYEN_SET(2);
3625 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_DQDELAYEN_SET(1);
3626 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_DQSDELAYEN_SET(1);
3627 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_DQSLOGICDELAYEN_SET(1);
3628 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_RESETDELAYEN_SET(0);
3629 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_LPDDRDIS_SET(1);
3630 /*
3631 * This field selects the intrinsic latency to RDATA_EN/FULL path.
3632 * 00-bypass, 01- add 5 cycles, 10- add 10 cycles, 11- add 15 cycles.
3633 */
3634 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_ADDLATSEL_SET(0);
3635 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_SAMPLECOUNT_19_0_SET(
3636 trk_sample_count);
3637 writel(reg, &sdr_ctrl->phy_ctrl0);
3638
3639 reg = 0;
3640 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_1_SAMPLECOUNT_31_20_SET(
3641 trk_sample_count >>
3642 SDR_CTRLGRP_PHYCTRL_PHYCTRL_0_SAMPLECOUNT_19_0_WIDTH);
3643 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_1_LONGIDLESAMPLECOUNT_19_0_SET(
3644 trk_long_idle_sample_count);
3645 writel(reg, &sdr_ctrl->phy_ctrl1);
3646
3647 reg = 0;
3648 reg |= SDR_CTRLGRP_PHYCTRL_PHYCTRL_2_LONGIDLESAMPLECOUNT_31_20_SET(
3649 trk_long_idle_sample_count >>
3650 SDR_CTRLGRP_PHYCTRL_PHYCTRL_1_LONGIDLESAMPLECOUNT_19_0_WIDTH);
3651 writel(reg, &sdr_ctrl->phy_ctrl2);
3652 }
3653
3654 static void initialize_tracking(void)
3655 {
3656 uint32_t concatenated_longidle = 0x0;
3657 uint32_t concatenated_delays = 0x0;
3658 uint32_t concatenated_rw_addr = 0x0;
3659 uint32_t concatenated_refresh = 0x0;
3660 uint32_t trk_sample_count = 7500;
3661 uint32_t dtaps_per_ptap;
3662 uint32_t tmp_delay;
3663
3664 /*
3665 * compute usable version of value in case we skip full
3666 * computation later
3667 */
3668 dtaps_per_ptap = 0;
3669 tmp_delay = 0;
3670 while (tmp_delay < IO_DELAY_PER_OPA_TAP) {
3671 dtaps_per_ptap++;
3672 tmp_delay += IO_DELAY_PER_DCHAIN_TAP;
3673 }
3674 dtaps_per_ptap--;
3675
3676 concatenated_longidle = concatenated_longidle ^ 10;
3677 /*longidle outer loop */
3678 concatenated_longidle = concatenated_longidle << 16;
3679 concatenated_longidle = concatenated_longidle ^ 100;
3680 /*longidle sample count */
3681 concatenated_delays = concatenated_delays ^ 243;
3682 /* trfc, worst case of 933Mhz 4Gb */
3683 concatenated_delays = concatenated_delays << 8;
3684 concatenated_delays = concatenated_delays ^ 14;
3685 /* trcd, worst case */
3686 concatenated_delays = concatenated_delays << 8;
3687 concatenated_delays = concatenated_delays ^ 10;
3688 /* vfifo wait */
3689 concatenated_delays = concatenated_delays << 8;
3690 concatenated_delays = concatenated_delays ^ 4;
3691 /* mux delay */
3692
3693 concatenated_rw_addr = concatenated_rw_addr ^ RW_MGR_IDLE;
3694 concatenated_rw_addr = concatenated_rw_addr << 8;
3695 concatenated_rw_addr = concatenated_rw_addr ^ RW_MGR_ACTIVATE_1;
3696 concatenated_rw_addr = concatenated_rw_addr << 8;
3697 concatenated_rw_addr = concatenated_rw_addr ^ RW_MGR_SGLE_READ;
3698 concatenated_rw_addr = concatenated_rw_addr << 8;
3699 concatenated_rw_addr = concatenated_rw_addr ^ RW_MGR_PRECHARGE_ALL;
3700
3701 concatenated_refresh = concatenated_refresh ^ RW_MGR_REFRESH_ALL;
3702 concatenated_refresh = concatenated_refresh << 24;
3703 concatenated_refresh = concatenated_refresh ^ 1000; /* trefi */
3704
3705 /* Initialize the register file with the correct data */
3706 writel(dtaps_per_ptap, &sdr_reg_file->dtaps_per_ptap);
3707 writel(trk_sample_count, &sdr_reg_file->trk_sample_count);
3708 writel(concatenated_longidle, &sdr_reg_file->trk_longidle);
3709 writel(concatenated_delays, &sdr_reg_file->delays);
3710 writel(concatenated_rw_addr, &sdr_reg_file->trk_rw_mgr_addr);
3711 writel(RW_MGR_MEM_IF_READ_DQS_WIDTH, &sdr_reg_file->trk_read_dqs_width);
3712 writel(concatenated_refresh, &sdr_reg_file->trk_rfsh);
3713 }
3714
3715 int sdram_calibration_full(void)
3716 {
3717 struct param_type my_param;
3718 struct gbl_type my_gbl;
3719 uint32_t pass;
3720 uint32_t i;
3721
3722 param = &my_param;
3723 gbl = &my_gbl;
3724
3725 /* Initialize the debug mode flags */
3726 gbl->phy_debug_mode_flags = 0;
3727 /* Set the calibration enabled by default */
3728 gbl->phy_debug_mode_flags |= PHY_DEBUG_ENABLE_CAL_RPT;
3729 /*
3730 * Only sweep all groups (regardless of fail state) by default
3731 * Set enabled read test by default.
3732 */
3733 #if DISABLE_GUARANTEED_READ
3734 gbl->phy_debug_mode_flags |= PHY_DEBUG_DISABLE_GUARANTEED_READ;
3735 #endif
3736 /* Initialize the register file */
3737 initialize_reg_file();
3738
3739 /* Initialize any PHY CSR */
3740 initialize_hps_phy();
3741
3742 scc_mgr_initialize();
3743
3744 initialize_tracking();
3745
3746 /* USER Enable all ranks, groups */
3747 for (i = 0; i < RW_MGR_MEM_NUMBER_OF_RANKS; i++)
3748 param->skip_ranks[i] = 0;
3749 for (i = 0; i < NUM_SHADOW_REGS; ++i)
3750 param->skip_shadow_regs[i] = 0;
3751 param->skip_groups = 0;
3752
3753 printf("%s: Preparing to start memory calibration\n", __FILE__);
3754
3755 debug("%s:%d\n", __func__, __LINE__);
3756 debug_cond(DLEVEL == 1,
3757 "DDR3 FULL_RATE ranks=%u cs/dimm=%u dq/dqs=%u,%u vg/dqs=%u,%u ",
3758 RW_MGR_MEM_NUMBER_OF_RANKS, RW_MGR_MEM_NUMBER_OF_CS_PER_DIMM,
3759 RW_MGR_MEM_DQ_PER_READ_DQS, RW_MGR_MEM_DQ_PER_WRITE_DQS,
3760 RW_MGR_MEM_VIRTUAL_GROUPS_PER_READ_DQS,
3761 RW_MGR_MEM_VIRTUAL_GROUPS_PER_WRITE_DQS);
3762 debug_cond(DLEVEL == 1,
3763 "dqs=%u,%u dq=%u dm=%u ptap_delay=%u dtap_delay=%u ",
3764 RW_MGR_MEM_IF_READ_DQS_WIDTH, RW_MGR_MEM_IF_WRITE_DQS_WIDTH,
3765 RW_MGR_MEM_DATA_WIDTH, RW_MGR_MEM_DATA_MASK_WIDTH,
3766 IO_DELAY_PER_OPA_TAP, IO_DELAY_PER_DCHAIN_TAP);
3767 debug_cond(DLEVEL == 1, "dtap_dqsen_delay=%u, dll=%u",
3768 IO_DELAY_PER_DQS_EN_DCHAIN_TAP, IO_DLL_CHAIN_LENGTH);
3769 debug_cond(DLEVEL == 1, "max values: en_p=%u dqdqs_p=%u en_d=%u dqs_in_d=%u ",
3770 IO_DQS_EN_PHASE_MAX, IO_DQDQS_OUT_PHASE_MAX,
3771 IO_DQS_EN_DELAY_MAX, IO_DQS_IN_DELAY_MAX);
3772 debug_cond(DLEVEL == 1, "io_in_d=%u io_out1_d=%u io_out2_d=%u ",
3773 IO_IO_IN_DELAY_MAX, IO_IO_OUT1_DELAY_MAX,
3774 IO_IO_OUT2_DELAY_MAX);
3775 debug_cond(DLEVEL == 1, "dqs_in_reserve=%u dqs_out_reserve=%u\n",
3776 IO_DQS_IN_RESERVE, IO_DQS_OUT_RESERVE);
3777
3778 hc_initialize_rom_data();
3779
3780 /* update info for sims */
3781 reg_file_set_stage(CAL_STAGE_NIL);
3782 reg_file_set_group(0);
3783
3784 /*
3785 * Load global needed for those actions that require
3786 * some dynamic calibration support.
3787 */
3788 dyn_calib_steps = STATIC_CALIB_STEPS;
3789 /*
3790 * Load global to allow dynamic selection of delay loop settings
3791 * based on calibration mode.
3792 */
3793 if (!(dyn_calib_steps & CALIB_SKIP_DELAY_LOOPS))
3794 skip_delay_mask = 0xff;
3795 else
3796 skip_delay_mask = 0x0;
3797
3798 pass = run_mem_calibrate();
3799
3800 printf("%s: Calibration complete\n", __FILE__);
3801 return pass;
3802 }