coldboot/coldboot.c

   1 #include <stdint.h>
   2 #include <stdbool.h>
   3
   4 #include "console.h"
   5 #include "microwatt_soc.h"
   6 #include "io.h"
   7
   8 #include <stdlib.h>
   9 #include <stdint.h>
  10 #include <gram.h>
  11
  12 #include "elf64.h"
  13
  14 static inline void mtspr(int sprnum, unsigned long val)
  15 {
  16     __asm__ volatile("mtspr %0,%1" : : "i" (sprnum), "r" (val));
  17 }
  18
  19 static inline uint32_t read32(const void *addr)
  20 {
  21         return *(volatile uint32_t *)addr;
  22 }
  23
  24 static inline void write32(void *addr, uint32_t value)
  25 {
  26         *(volatile uint32_t *)addr = value;
  27 }
  28
  29 struct uart_regs {
  30         uint32_t divisor;
  31         uint32_t rx_data;
  32         uint32_t rx_rdy;
  33         uint32_t rx_err;
  34         uint32_t tx_data;
  35         uint32_t tx_rdy;
  36         uint32_t zero0; // reserved
  37         uint32_t zero1; // reserved
  38         uint32_t ev_status;
  39         uint32_t ev_pending;
  40         uint32_t ev_enable;
  41 };
  42
  43 void uart_writeuint32(uint32_t val) {
  44         const char lut[] = { '0', '1', '2', '3', '4', '5', '6', '7',
  45                          '8', '9', 'A', 'B', 'C', 'D', 'E', 'F' };
  46         uint8_t *val_arr = (uint8_t*)(&val);
  47         size_t i;
  48
  49         for (i = 0; i < 4; i++) {
  50                 putchar(lut[(val_arr[3-i] >> 4) & 0xF]);
  51                 putchar(lut[val_arr[3-i] & 0xF]);
  52         }
  53 }
  54
  55 void memcpy(void *dest, void *src, size_t n) {
  56     int i;
  57     //cast src and dest to char*
  58     char *src_char = (char *)src;
  59     char *dest_char = (char *)dest;
  60     for (i=0; i<n; i++) {
  61 #if 1
  62         if ((i % 4096) == 0) {
  63             puts("memcpy ");
  64             uart_writeuint32(i);
  65             puts("\r\n");
  66         }
  67 #endif
  68         dest_char[i] = src_char[i]; //copy contents byte by byte
  69     }
  70 }
  71
  72 void isr(void) {
  73
  74 }
  75
  76 // WARNING
  77 // KESTREL SPECIFIC
  78 #define TERCEL_SPI_REG_SYS_PHY_CFG1     0x10
  79 #define TERCEL_SPI_REG_SYS_FLASH_CFG5       0x24
  80 #define TERCEL_SPI_PHY_CLOCK_DIVISOR_MASK   0xff
  81 #define TERCEL_SPI_PHY_CLOCK_DIVISOR_SHIFT  0
  82 #define TERCEL_SPI_FLASH_EN_MULTCYC_READ_MASK   0x1
  83 #define TERCEL_SPI_FLASH_EN_MULTCYC_READ_SHIFT  0
  84 static inline uint32_t read_tercel_register(uint8_t reg)
  85 {
  86         return readl((unsigned long)(SPI_FCTRL_BASE+reg));
  87 }
  88
  89 static inline void write_tercel_register(uint8_t reg, uint32_t value)
  90 {
  91         writel(value, (unsigned long)(SPI_FCTRL_BASE+reg));
  92 }
  93
  94 // TODO: need to use this
  95 // https://gitlab.raptorengineering.com/kestrel-collaboration/kestrel-firmware/bare-metal-firmware/-/blob/master/main.c#L2328
  96
  97 /* this is a "level 1" speed-up, which gets an initial improvement of 10-50x
  98  * over the default speed (which is a scant 100 bytes per second).
  99  */
 100 static void crank_up_qspi_level1(void)
 101 {
 102     // WARNING: KESTREL SPECIFIC
 103     // Set SPI clock cycle divider to 1
 104     uint32_t dword;
 105     dword = read_tercel_register(TERCEL_SPI_REG_SYS_PHY_CFG1);
 106     dword &= ~(TERCEL_SPI_PHY_CLOCK_DIVISOR_MASK <<
 107                TERCEL_SPI_PHY_CLOCK_DIVISOR_SHIFT);
 108     dword |= ((1 & TERCEL_SPI_PHY_CLOCK_DIVISOR_MASK) <<
 109                    TERCEL_SPI_PHY_CLOCK_DIVISOR_SHIFT);
 110     write_tercel_register(TERCEL_SPI_REG_SYS_PHY_CFG1, dword);
 111     // Enable read merging
 112     dword = read_tercel_register(TERCEL_SPI_REG_SYS_FLASH_CFG5);
 113     dword |= (TERCEL_SPI_FLASH_EN_MULTCYC_READ_MASK <<
 114               TERCEL_SPI_FLASH_EN_MULTCYC_READ_SHIFT);
 115     write_tercel_register(TERCEL_SPI_REG_SYS_FLASH_CFG5, dword);
 116 }
 117
 118 static bool fl_read(void *dst, uint32_t offset, uint32_t size)
 119 {
 120     uint8_t *d = dst;
 121     memcpy(d, (void *)(unsigned long)(SPI_FLASH_BASE + offset), size);
 122     return true;
 123 }
 124
 125 static unsigned long copy_flash(unsigned int offset, unsigned int dst_offs)
 126 {
 127     Elf64_Ehdr ehdr;
 128     Elf64_Phdr ph;
 129     unsigned int i, poff, size, off;
 130     void *addr;
 131
 132     puts("Trying flash...\r\n");
 133     if (!fl_read(&ehdr, offset, sizeof(ehdr)))
 134         return -1ul;
 135     if (!IS_ELF(ehdr) || ehdr.e_ident[EI_CLASS] != ELFCLASS64) {
 136         puts("Doesn't look like an elf64\r\n");
 137         goto dump;
 138     }
 139     if (ehdr.e_ident[EI_DATA] != ELFDATA2LSB ||
 140         ehdr.e_machine != EM_PPC64) {
 141         puts("Not a ppc64le binary\r\n");
 142         goto dump;
 143     }
 144
 145     poff = offset + ehdr.e_phoff;
 146     for (i = 0; i < ehdr.e_phnum; i++) {
 147         if (!fl_read(&ph, poff, sizeof(ph)))
 148             goto dump;
 149         if (ph.p_type != PT_LOAD)
 150             continue;
 151
 152         /* XXX Add bound checking ! */
 153         size = ph.p_filesz;
 154         addr = (void *)ph.p_vaddr;
 155         off  = offset + ph.p_offset;
 156         //printf("Copy segment %d (0x%x bytes) to %p\n", i, size, addr);
 157         puts("Copy segment ");
 158         uart_writeuint32(i);
 159         puts(" size ");
 160         uart_writeuint32(size);
 161         puts(" addr ");
 162         uart_writeuint32((uint32_t)(unsigned long)addr);
 163         puts("\r\n");
 164         fl_read(addr+dst_offs, off, size);
 165         poff += ehdr.e_phentsize;
 166     }
 167
 168     puts("Booting from DRAM at");
 169     uart_writeuint32((unsigned int)(dst_offs+ehdr.e_entry));
 170     puts("\r\n");
 171
 172     puts("Dump DRAM\r\n");
 173     for (i = 0; i < 64; i++) {
 174         uart_writeuint32(readl(dst_offs+ehdr.e_entry+(i*4)));
 175         puts(" ");
 176         if ((i & 7) == 7) puts("\r\n");
 177     }
 178     puts("\r\n");
 179
 180     //flush_cpu_icache();
 181     return dst_offs+ehdr.e_entry;
 182 dump:
 183     puts("HDR: \r\n");
 184     for (i = 0; i < 8; i++) {
 185         uart_writeuint32(ehdr.e_ident[i]);
 186         puts("\r\n");
 187     }
 188
 189     return -1ul;
 190 }
 191
 192
 193 // XXX
 194 // Defining gram_[read|write] allows a trace of all register
 195 // accesses to be dumped to console for debugging purposes.
 196 // To use, define GRAM_RW_FUNC in gram.h
 197 uint32_t gram_read(const struct gramCtx *ctx, void *addr) {
 198         uint32_t dword;
 199
 200         puts("gram_read: ");
 201         uart_writeuint32((unsigned long)addr);
 202         dword = readl((unsigned long)addr);
 203         puts(": ");
 204         uart_writeuint32((unsigned long)dword);
 205         puts("\n");
 206
 207         return dword;
 208 }
 209
 210 int gram_write(const struct gramCtx *ctx, void *addr, uint32_t value) {
 211         puts("gram_write: ");
 212         uart_writeuint32((unsigned long)addr);
 213         puts(": ");
 214         uart_writeuint32((unsigned long)value);
 215         writel(value, (unsigned long)addr);
 216         puts("\n");
 217
 218         return 0;
 219 }
 220
 221 int main(void) {
 222         const int kNumIterations = 14;
 223         int res, failcnt = 0;
 224         uint32_t tmp;
 225     unsigned long ftr, spi_offs=0x0;
 226         volatile uint32_t *ram = (uint32_t*)MEMORY_BASE;
 227
 228         console_init();
 229         //puts("Firmware launched...\n");
 230
 231 #if 1
 232     puts(" Soc signature: ");
 233     tmp = readl(SYSCON_BASE + SYS_REG_SIGNATURE);
 234     uart_writeuint32(tmp);
 235     puts("  Soc features: ");
 236     ftr = readl(SYSCON_BASE + SYS_REG_INFO);
 237     if (ftr & SYS_REG_INFO_HAS_UART)
 238         puts("UART ");
 239     if (ftr & SYS_REG_INFO_HAS_DRAM)
 240         puts("DRAM ");
 241     if (ftr & SYS_REG_INFO_HAS_BRAM)
 242         puts("BRAM ");
 243     if (ftr & SYS_REG_INFO_HAS_SPI_FLASH)
 244         puts("SPIFLASH ");
 245     if (ftr & SYS_REG_INFO_HAS_LITEETH)
 246         puts("ETHERNET ");
 247     puts("\r\n");
 248
 249     if (ftr & SYS_REG_INFO_HAS_SPI_FLASH) {
 250         // speed up the QSPI to at least a sane level
 251         crank_up_qspi_level1();
 252
 253         puts("SPI Offset: ");
 254         spi_offs = readl(SYSCON_BASE + SYS_REG_SPI_INFO);
 255         uart_writeuint32(spi_offs);
 256         puts("\r\n");
 257     }
 258
 259 #endif
 260
 261 #if 1
 262 #if 1
 263     if (ftr & SYS_REG_INFO_HAS_SPI_FLASH) {
 264         // print out configuration parameters for QSPI
 265         volatile uint32_t *qspi_cfg = (uint32_t*)SPI_FCTRL_BASE;
 266         for (int k=0; k < 2; k++) {
 267             tmp = readl((unsigned long)&(qspi_cfg[k]));
 268             puts("cfg");
 269             uart_writeuint32(k);
 270             puts(" ");
 271             uart_writeuint32(tmp);
 272             puts("\n");
 273         }
 274     }
 275 #endif
 276     if (ftr & SYS_REG_INFO_HAS_SPI_FLASH) {
 277         volatile uint32_t *qspi = (uint32_t*)SPI_FLASH_BASE+0x600000;
 278         //volatile uint8_t *qspi_bytes = (uint8_t*)spi_offs;
 279          // let's not, eh? writel(0xDEAF0123, (unsigned long)&(qspi[0]));
 280          // tmp = readl((unsigned long)&(qspi[0]));
 281         for (int i=0;i<256;i++) {
 282           tmp = readl((unsigned long)&(qspi[i]));
 283           uart_writeuint32(tmp);
 284           puts(" ");
 285           if ((i & 0x7) == 0x7) puts("\r\n");
 286         }
 287         puts("\r\n");
 288         /*
 289         for (i=0;i<256;i++) {
 290           tmp = readb((unsigned long)&(qspi_bytes[i]));
 291           uart_writeuint32(tmp);
 292           puts(" ");
 293         }
 294         */
 295 #if 0
 296         while (1) {
 297             // quick read
 298             tmp = readl((unsigned long)&(qspi[0x1000/4]));
 299             puts("read 0x1000");
 300             uart_writeuint32(tmp);
 301             putchar(10);
 302         }
 303         while (1) {
 304             unsigned char c = getchar();
 305             putchar(c);
 306             if (c == 13) { // if CR send LF
 307
 308                 // quick read
 309                 tmp = readl((unsigned long)&(qspi[1<<i]));
 310                 puts("read ");
 311                 uart_writeuint32(1<<i);
 312                 puts(" ");
 313                 uart_writeuint32(tmp);
 314                 putchar(10);
 315                 i++;
 316             }
 317         }
 318
 319         return 0;
 320 #endif
 321     }
 322 #endif
 323 #if 0
 324         volatile uint32_t *hyperram = (uint32_t*)0xa0000000;
 325     writel(0xDEAF0123, (unsigned long)&(hyperram[0]));
 326     tmp = readl((unsigned long)&(hyperram[0]));
 327     while (1) {
 328         unsigned char c = getchar();
 329         putchar(c);
 330         if (c == 13) { // if CR send LF
 331
 332             // quick write/read
 333             writel(0xDEAF0123+i, (unsigned long)&(hyperram[1<<i]));
 334             tmp = readl((unsigned long)&(hyperram[1<<i]));
 335             puts("read ");
 336             uart_writeuint32(1<<i);
 337             puts(" ");
 338             uart_writeuint32(tmp);
 339             putchar(10);
 340             i++;
 341         }
 342     }
 343
 344     return 0;
 345 #endif
 346
 347     // init DRAM only if SYSCON says it exists (duh)
 348     if (ftr & SYS_REG_INFO_HAS_DRAM)
 349     {
 350         puts("DRAM init... ");
 351
 352         struct gramCtx ctx;
 353 #if 1
 354         struct gramProfile profile = {
 355             .mode_registers = {
 356                 0xb20, 0x806, 0x200, 0x0
 357             },
 358             .rdly_p0 = 2,
 359             .rdly_p1 = 2,
 360         };
 361 #endif
 362 #if 0
 363         struct gramProfile profile = {
 364             .mode_registers = {
 365                 0x0320, 0x0006, 0x0200, 0x0000
 366             },
 367             .rdly_p0 = 1,
 368             .rdly_p1 = 1,
 369         };
 370 #endif
 371         struct gramProfile profile2;
 372         gram_init(&ctx, &profile, (void*)MEMORY_BASE,
 373                                   (void*)DRAM_CTRL_BASE,
 374                                   (void*)DRAM_INIT_BASE);
 375         puts("done\n");
 376
 377         puts("MR profile: ");
 378         uart_writeuint32(profile.mode_registers[0]);
 379         puts(" ");
 380         uart_writeuint32(profile.mode_registers[1]);
 381         puts(" ");
 382         uart_writeuint32(profile.mode_registers[2]);
 383         puts(" ");
 384         uart_writeuint32(profile.mode_registers[3]);
 385         puts("\n");
 386
 387         // FIXME
 388         // Early read test for WB access sim
 389         //uart_writeuint32(*ram);
 390
 391 #if 1
 392         puts("Rdly\np0: ");
 393         for (size_t i = 0; i < 8; i++) {
 394             profile2.rdly_p0 = i;
 395             gram_load_calibration(&ctx, &profile2);
 396             gram_reset_burstdet(&ctx);
 397
 398             for (size_t j = 0; j < 128; j++) {
 399                 tmp = readl((unsigned long)&(ram[i]));
 400             }
 401             if (gram_read_burstdet(&ctx, 0)) {
 402                 puts("1");
 403             } else {
 404                 puts("0");
 405             }
 406         }
 407         puts("\n");
 408
 409         puts("Rdly\np1: ");
 410         for (size_t i = 0; i < 8; i++) {
 411             profile2.rdly_p1 = i;
 412             gram_load_calibration(&ctx, &profile2);
 413             gram_reset_burstdet(&ctx);
 414             for (size_t j = 0; j < 128; j++) {
 415                 tmp = readl((unsigned long)&(ram[i]));
 416             }
 417             if (gram_read_burstdet(&ctx, 1)) {
 418                 puts("1");
 419             } else {
 420                 puts("0");
 421             }
 422         }
 423         puts("\n");
 424
 425         puts("Auto calibrating... ");
 426         res = gram_generate_calibration(&ctx, &profile2);
 427         if (res != GRAM_ERR_NONE) {
 428             puts("failed\n");
 429             gram_load_calibration(&ctx, &profile);
 430         } else {
 431             gram_load_calibration(&ctx, &profile2);
 432         }
 433         puts("done\n");
 434
 435         puts("Auto calibration profile:");
 436         puts("p0 rdly:");
 437         uart_writeuint32(profile2.rdly_p0);
 438         puts(" p1 rdly:");
 439         uart_writeuint32(profile2.rdly_p1);
 440         puts("\n");
 441 #endif
 442
 443         puts("Reloading built-in calibration profile...");
 444         gram_load_calibration(&ctx, &profile);
 445
 446         puts("DRAM test... \n");
 447         for (size_t i = 0; i < kNumIterations; i++) {
 448             writel(0xDEAF0000 | i*4, (unsigned long)&(ram[i]));
 449         }
 450
 451 #if 0
 452         for (int dly = 0; dly < 8; dly++) {
 453             failcnt = 0;
 454             profile2.rdly_p0 = dly;
 455             profile2.rdly_p1 = dly;
 456             puts("p0 rdly:");
 457             uart_writeuint32(profile2.rdly_p0);
 458             puts(" p1 rdly:");
 459             uart_writeuint32(profile2.rdly_p1);
 460             gram_load_calibration(&ctx, &profile2);
 461             for (size_t i = 0; i < kNumIterations; i++) {
 462                 if (readl((unsigned long)&(ram[i])) != (0xDEAF0000 | i*4)) {
 463                     puts("fail : *(0x");
 464                     uart_writeuint32((unsigned long)(&ram[i]));
 465                     puts(") = ");
 466                     uart_writeuint32(readl((unsigned long)&(ram[i])));
 467                     puts("\n");
 468                     failcnt++;
 469
 470                     if (failcnt > 10) {
 471                         puts("Test canceled (more than 10 errors)\n");
 472                         break;
 473                     }
 474                 }
 475             }
 476         }
 477 #else
 478         failcnt = 0;
 479         for (size_t i = 0; i < kNumIterations; i++) {
 480             if (readl((unsigned long)&(ram[i])) != (0xDEAF0000 | i*4)) {
 481                 puts("fail : *(0x");
 482                 uart_writeuint32((unsigned long)(&ram[i]));
 483                 puts(") = ");
 484                 uart_writeuint32(readl((unsigned long)&(ram[i])));
 485                 puts("\n");
 486                 failcnt++;
 487
 488                 if (failcnt > 10) {
 489                     puts("Test canceled (more than 10 errors)\n");
 490                     break;
 491                 }
 492             }
 493         }
 494     }
 495 #endif
 496         puts("done\n");
 497
 498 #if 0 // ooo, annoying: won't work. no idea why
 499     // temporary hard-hack: boot directly from QSPI. really
 500     // should do something like detect at least... something
 501     if ((ftr & SYS_REG_INFO_HAS_SPI_FLASH))
 502     {
 503         // jump to absolute address
 504         mtspr(8, SPI_FLASH_BASE); // move address to LR
 505         __asm__ volatile("blr");
 506         return 0;
 507     }
 508 #endif
 509
 510     // memcpy from SPI Flash then boot
 511     if ((ftr & SYS_REG_INFO_HAS_SPI_FLASH) &&
 512         (failcnt == 0))
 513     {
 514 /*
 515         puts("ELF @ QSPI\n");
 516         // identify ELF, copy if present, and get the start address
 517         unsigned long faddr = copy_flash(spi_offs,
 518                                          0x600000); // hack!
 519         if (faddr != -1ul) {
 520             // jump to absolute address
 521             mtspr(8, faddr); // move address to LR
 522             __asm__ volatile("blr");
 523
 524             // works with head.S which copies r3 into ctr then does bctr
 525             return faddr;
 526         }
 527         puts("copy QSPI\n");
 528 */
 529         // another terrible hack: copy from flash at offset 0x600000
 530         // a block of size 0x600000 into mem address 0x600000, then
 531         // jump to it.  this allows a dtb image to be executed
 532         puts("copy QSPI\n");
 533         volatile uint32_t *mem = (uint32_t*)0x600000;
 534         fl_read(mem, 0x600000, 0x600000); // shorter (testing) 0x8000);
 535         puts("dump mem\n");
 536         for (int i=0;i<256;i++) {
 537           tmp = readl((unsigned long)&(mem[i]));
 538           uart_writeuint32(tmp);
 539           puts(" ");
 540           if ((i & 0x7) == 0x7) puts("\r\n");
 541         }
 542         puts("\r\n");
 543         mtspr(8, 0x600000); // move address to LR
 544         __asm__ volatile("blr");
 545     }
 546
 547         return 0;
 548 }
 549