arch/mips/kernel/r4k-bugs64.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * Copyright (C) 2003, 2004, 2007  Maciej W. Rozycki
   4  */
   5 #include <linux/context_tracking.h>
   6 #include <linux/init.h>
   7 #include <linux/kernel.h>
   8 #include <linux/ptrace.h>
   9 #include <linux/stddef.h>
  10
  11 #include <asm/bugs.h>
  12 #include <asm/compiler.h>
  13 #include <asm/cpu.h>
  14 #include <asm/fpu.h>
  15 #include <asm/mipsregs.h>
  16 #include <asm/setup.h>
  17
  18 static char bug64hit[] __initdata =
  19         "reliable operation impossible!\n%s";
  20 static char nowar[] __initdata =
  21         "Please report to <linux-mips@linux-mips.org>.";
  22 static char r4kwar[] __initdata =
  23         "Enable CPU_R4000_WORKAROUNDS to rectify.";
  24 static char daddiwar[] __initdata =
  25         "Enable CPU_DADDI_WORKAROUNDS to rectify.";
  26
  27 static __always_inline __init
  28 void align_mod(const int align, const int mod)
  29 {
  30         asm volatile(
  31                 ".set   push\n\t"
  32                 ".set   noreorder\n\t"
  33                 ".balign %0\n\t"
  34                 ".rept  %1\n\t"
  35                 "nop\n\t"
  36                 ".endr\n\t"
  37                 ".set   pop"
  38                 :
  39                 : "n"(align), "n"(mod));
  40 }
  41
  42 static __always_inline __init
  43 void mult_sh_align_mod(long *v1, long *v2, long *w,
  44                        const int align, const int mod)
  45 {
  46         unsigned long flags;
  47         int m1, m2;
  48         long p, s, lv1, lv2, lw;
  49
  50         /*
  51          * We want the multiply and the shift to be isolated from the
  52          * rest of the code to disable gcc optimizations.  Hence the
  53          * asm statements that execute nothing, but make gcc not know
  54          * what the values of m1, m2 and s are and what lv2 and p are
  55          * used for.
  56          */
  57
  58         local_irq_save(flags);
  59         /*
  60          * The following code leads to a wrong result of the first
  61          * dsll32 when executed on R4000 rev. 2.2 or 3.0 (PRId
  62          * 00000422 or 00000430, respectively).
  63          *
  64          * See "MIPS R4000PC/SC Errata, Processor Revision 2.2 and
  65          * 3.0" by MIPS Technologies, Inc., errata #16 and #28 for
  66          * details.  I got no permission to duplicate them here,
  67          * sigh... --macro
  68          */
  69         asm volatile(
  70                 ""
  71                 : "=r" (m1), "=r" (m2), "=r" (s)
  72                 : "0" (5), "1" (8), "2" (5));
  73         align_mod(align, mod);
  74         /*
  75          * The trailing nop is needed to fulfill the two-instruction
  76          * requirement between reading hi/lo and staring a mult/div.
  77          * Leaving it out may cause gas insert a nop itself breaking
  78          * the desired alignment of the next chunk.
  79          */
  80         asm volatile(
  81                 ".set   push\n\t"
  82                 ".set   noat\n\t"
  83                 ".set   noreorder\n\t"
  84                 ".set   nomacro\n\t"
  85                 "mult   %2, %3\n\t"
  86                 "dsll32 %0, %4, %5\n\t"
  87                 "mflo   $0\n\t"
  88                 "dsll32 %1, %4, %5\n\t"
  89                 "nop\n\t"
  90                 ".set   pop"
  91                 : "=&r" (lv1), "=r" (lw)
  92                 : "r" (m1), "r" (m2), "r" (s), "I" (0)
  93                 : "hi", "lo", "$0");
  94         /* We have to use single integers for m1 and m2 and a double
  95          * one for p to be sure the mulsidi3 gcc's RTL multiplication
  96          * instruction has the workaround applied.  Older versions of
  97          * gcc have correct umulsi3 and mulsi3, but other
  98          * multiplication variants lack the workaround.
  99          */
 100         asm volatile(
 101                 ""
 102                 : "=r" (m1), "=r" (m2), "=r" (s)
 103                 : "0" (m1), "1" (m2), "2" (s));
 104         align_mod(align, mod);
 105         p = m1 * m2;
 106         lv2 = s << 32;
 107         asm volatile(
 108                 ""
 109                 : "=r" (lv2)
 110                 : "0" (lv2), "r" (p));
 111         local_irq_restore(flags);
 112
 113         *v1 = lv1;
 114         *v2 = lv2;
 115         *w = lw;
 116 }
 117
 118 static __always_inline __init void check_mult_sh(void)
 119 {
 120         long v1[8], v2[8], w[8];
 121         int bug, fix, i;
 122
 123         printk("Checking for the multiply/shift bug... ");
 124
 125         /*
 126          * Testing discovered false negatives for certain code offsets
 127          * into cache lines.  Hence we test all possible offsets for
 128          * the worst assumption of an R4000 I-cache line width of 32
 129          * bytes.
 130          *
 131          * We can't use a loop as alignment directives need to be
 132          * immediates.
 133          */
 134         mult_sh_align_mod(&v1[0], &v2[0], &w[0], 32, 0);
 135         mult_sh_align_mod(&v1[1], &v2[1], &w[1], 32, 1);
 136         mult_sh_align_mod(&v1[2], &v2[2], &w[2], 32, 2);
 137         mult_sh_align_mod(&v1[3], &v2[3], &w[3], 32, 3);
 138         mult_sh_align_mod(&v1[4], &v2[4], &w[4], 32, 4);
 139         mult_sh_align_mod(&v1[5], &v2[5], &w[5], 32, 5);
 140         mult_sh_align_mod(&v1[6], &v2[6], &w[6], 32, 6);
 141         mult_sh_align_mod(&v1[7], &v2[7], &w[7], 32, 7);
 142
 143         bug = 0;
 144         for (i = 0; i < 8; i++)
 145                 if (v1[i] != w[i])
 146                         bug = 1;
 147
 148         if (bug == 0) {
 149                 pr_cont("no.\n");
 150                 return;
 151         }
 152
 153         pr_cont("yes, workaround... ");
 154
 155         fix = 1;
 156         for (i = 0; i < 8; i++)
 157                 if (v2[i] != w[i])
 158                         fix = 0;
 159
 160         if (fix == 1) {
 161                 pr_cont("yes.\n");
 162                 return;
 163         }
 164
 165         pr_cont("no.\n");
 166         panic(bug64hit, !R4000_WAR ? r4kwar : nowar);
 167 }
 168
 169 static volatile int daddi_ov;
 170
 171 asmlinkage void __init do_daddi_ov(struct pt_regs *regs)
 172 {
 173         enum ctx_state prev_state;
 174
 175         prev_state = exception_enter();
 176         daddi_ov = 1;
 177         regs->cp0_epc += 4;
 178         exception_exit(prev_state);
 179 }
 180
 181 static __init void check_daddi(void)
 182 {
 183         extern asmlinkage void handle_daddi_ov(void);
 184         unsigned long flags;
 185         void *handler;
 186         long v, tmp;
 187
 188         printk("Checking for the daddi bug... ");
 189
 190         local_irq_save(flags);
 191         handler = set_except_vector(EXCCODE_OV, handle_daddi_ov);
 192         /*
 193          * The following code fails to trigger an overflow exception
 194          * when executed on R4000 rev. 2.2 or 3.0 (PRId 00000422 or
 195          * 00000430, respectively).
 196          *
 197          * See "MIPS R4000PC/SC Errata, Processor Revision 2.2 and
 198          * 3.0" by MIPS Technologies, Inc., erratum #23 for details.
 199          * I got no permission to duplicate it here, sigh... --macro
 200          */
 201         asm volatile(
 202                 ".set   push\n\t"
 203                 ".set   noat\n\t"
 204                 ".set   noreorder\n\t"
 205                 ".set   nomacro\n\t"
 206                 "addiu  %1, $0, %2\n\t"
 207                 "dsrl   %1, %1, 1\n\t"
 208 #ifdef HAVE_AS_SET_DADDI
 209                 ".set   daddi\n\t"
 210 #endif
 211                 "daddi  %0, %1, %3\n\t"
 212                 ".set   pop"
 213                 : "=r" (v), "=&r" (tmp)
 214                 : "I" (0xffffffffffffdb9aUL), "I" (0x1234));
 215         set_except_vector(EXCCODE_OV, handler);
 216         local_irq_restore(flags);
 217
 218         if (daddi_ov) {
 219                 pr_cont("no.\n");
 220                 return;
 221         }
 222
 223         pr_cont("yes, workaround... ");
 224
 225         local_irq_save(flags);
 226         handler = set_except_vector(EXCCODE_OV, handle_daddi_ov);
 227         asm volatile(
 228                 "addiu  %1, $0, %2\n\t"
 229                 "dsrl   %1, %1, 1\n\t"
 230                 "daddi  %0, %1, %3"
 231                 : "=r" (v), "=&r" (tmp)
 232                 : "I" (0xffffffffffffdb9aUL), "I" (0x1234));
 233         set_except_vector(EXCCODE_OV, handler);
 234         local_irq_restore(flags);
 235
 236         if (daddi_ov) {
 237                 pr_cont("yes.\n");
 238                 return;
 239         }
 240
 241         pr_cont("no.\n");
 242         panic(bug64hit, !DADDI_WAR ? daddiwar : nowar);
 243 }
 244
 245 int daddiu_bug  = -1;
 246
 247 static __init void check_daddiu(void)
 248 {
 249         long v, w, tmp;
 250
 251         printk("Checking for the daddiu bug... ");
 252
 253         /*
 254          * The following code leads to a wrong result of daddiu when
 255          * executed on R4400 rev. 1.0 (PRId 00000440).
 256          *
 257          * See "MIPS R4400PC/SC Errata, Processor Revision 1.0" by
 258          * MIPS Technologies, Inc., erratum #7 for details.
 259          *
 260          * According to "MIPS R4000PC/SC Errata, Processor Revision
 261          * 2.2 and 3.0" by MIPS Technologies, Inc., erratum #41 this
 262          * problem affects R4000 rev. 2.2 and 3.0 (PRId 00000422 and
 263          * 00000430, respectively), too.  Testing failed to trigger it
 264          * so far.
 265          *
 266          * I got no permission to duplicate the errata here, sigh...
 267          * --macro
 268          */
 269         asm volatile(
 270                 ".set   push\n\t"
 271                 ".set   noat\n\t"
 272                 ".set   noreorder\n\t"
 273                 ".set   nomacro\n\t"
 274                 "addiu  %2, $0, %3\n\t"
 275                 "dsrl   %2, %2, 1\n\t"
 276 #ifdef HAVE_AS_SET_DADDI
 277                 ".set   daddi\n\t"
 278 #endif
 279                 "daddiu %0, %2, %4\n\t"
 280                 "addiu  %1, $0, %4\n\t"
 281                 "daddu  %1, %2\n\t"
 282                 ".set   pop"
 283                 : "=&r" (v), "=&r" (w), "=&r" (tmp)
 284                 : "I" (0xffffffffffffdb9aUL), "I" (0x1234));
 285
 286         daddiu_bug = v != w;
 287
 288         if (!daddiu_bug) {
 289                 pr_cont("no.\n");
 290                 return;
 291         }
 292
 293         pr_cont("yes, workaround... ");
 294
 295         asm volatile(
 296                 "addiu  %2, $0, %3\n\t"
 297                 "dsrl   %2, %2, 1\n\t"
 298                 "daddiu %0, %2, %4\n\t"
 299                 "addiu  %1, $0, %4\n\t"
 300                 "daddu  %1, %2"
 301                 : "=&r" (v), "=&r" (w), "=&r" (tmp)
 302                 : "I" (0xffffffffffffdb9aUL), "I" (0x1234));
 303
 304         if (v == w) {
 305                 pr_cont("yes.\n");
 306                 return;
 307         }
 308
 309         pr_cont("no.\n");
 310         panic(bug64hit, !DADDI_WAR ? daddiwar : nowar);
 311 }
 312
 313 void __init check_bugs64_early(void)
 314 {
 315         check_mult_sh();
 316         check_daddiu();
 317 }
 318
 319 void __init check_bugs64(void)
 320 {
 321         check_daddi();
 322 }