arch/arm/vfp/vfpmodule.c

   1 /*
   2  *  linux/arch/arm/vfp/vfpmodule.c
   3  *
   4  *  Copyright (C) 2004 ARM Limited.
   5  *  Written by Deep Blue Solutions Limited.
   6  *
   7  * This program is free software; you can redistribute it and/or modify
   8  * it under the terms of the GNU General Public License version 2 as
   9  * published by the Free Software Foundation.
  10  */
  11 #include <linux/module.h>
  12 #include <linux/types.h>
  13 #include <linux/kernel.h>
  14 #include <linux/signal.h>
  15 #include <linux/sched.h>
  16 #include <linux/init.h>
  17
  18 #include <asm/thread_notify.h>
  19 #include <asm/vfp.h>
  20
  21 #include "vfpinstr.h"
  22 #include "vfp.h"
  23
  24 /*
  25  * Our undef handlers (in entry.S)
  26  */
  27 void vfp_testing_entry(void);
  28 void vfp_support_entry(void);
  29
  30 void (*vfp_vector)(void) = vfp_testing_entry;
  31 union vfp_state *last_VFP_context[NR_CPUS];
  32
  33 /*
  34  * Dual-use variable.
  35  * Used in startup: set to non-zero if VFP checks fail
  36  * After startup, holds VFP architecture
  37  */
  38 unsigned int VFP_arch;
  39
  40 static int vfp_notifier(struct notifier_block *self, unsigned long cmd, void *v)
  41 {
  42         struct thread_info *thread = v;
  43         union vfp_state *vfp;
  44         __u32 cpu = thread->cpu;
  45
  46         if (likely(cmd == THREAD_NOTIFY_SWITCH)) {
  47                 u32 fpexc = fmrx(FPEXC);
  48
  49 #ifdef CONFIG_SMP
  50                 /*
  51                  * On SMP, if VFP is enabled, save the old state in
  52                  * case the thread migrates to a different CPU. The
  53                  * restoring is done lazily.
  54                  */
  55                 if ((fpexc & FPEXC_ENABLE) && last_VFP_context[cpu]) {
  56                         vfp_save_state(last_VFP_context[cpu], fpexc);
  57                         last_VFP_context[cpu]->hard.cpu = cpu;
  58                 }
  59                 /*
  60                  * Thread migration, just force the reloading of the
  61                  * state on the new CPU in case the VFP registers
  62                  * contain stale data.
  63                  */
  64                 if (thread->vfpstate.hard.cpu != cpu)
  65                         last_VFP_context[cpu] = NULL;
  66 #endif
  67
  68                 /*
  69                  * Always disable VFP so we can lazily save/restore the
  70                  * old state.
  71                  */
  72                 fmxr(FPEXC, fpexc & ~FPEXC_ENABLE);
  73                 return NOTIFY_DONE;
  74         }
  75
  76         vfp = &thread->vfpstate;
  77         if (cmd == THREAD_NOTIFY_FLUSH) {
  78                 /*
  79                  * Per-thread VFP initialisation.
  80                  */
  81                 memset(vfp, 0, sizeof(union vfp_state));
  82
  83                 vfp->hard.fpexc = FPEXC_ENABLE;
  84                 vfp->hard.fpscr = FPSCR_ROUND_NEAREST;
  85
  86                 /*
  87                  * Disable VFP to ensure we initialise it first.
  88                  */
  89                 fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_ENABLE);
  90         }
  91
  92         /* flush and release case: Per-thread VFP cleanup. */
  93         if (last_VFP_context[cpu] == vfp)
  94                 last_VFP_context[cpu] = NULL;
  95
  96         return NOTIFY_DONE;
  97 }
  98
  99 static struct notifier_block vfp_notifier_block = {
 100         .notifier_call  = vfp_notifier,
 101 };
 102
 103 /*
 104  * Raise a SIGFPE for the current process.
 105  * sicode describes the signal being raised.
 106  */
 107 void vfp_raise_sigfpe(unsigned int sicode, struct pt_regs *regs)
 108 {
 109         siginfo_t info;
 110
 111         memset(&info, 0, sizeof(info));
 112
 113         info.si_signo = SIGFPE;
 114         info.si_code = sicode;
 115         info.si_addr = (void __user *)(instruction_pointer(regs) - 4);
 116
 117         /*
 118          * This is the same as NWFPE, because it's not clear what
 119          * this is used for
 120          */
 121         current->thread.error_code = 0;
 122         current->thread.trap_no = 6;
 123
 124         send_sig_info(SIGFPE, &info, current);
 125 }
 126
 127 static void vfp_panic(char *reason)
 128 {
 129         int i;
 130
 131         printk(KERN_ERR "VFP: Error: %s\n", reason);
 132         printk(KERN_ERR "VFP: EXC 0x%08x SCR 0x%08x INST 0x%08x\n",
 133                 fmrx(FPEXC), fmrx(FPSCR), fmrx(FPINST));
 134         for (i = 0; i < 32; i += 2)
 135                 printk(KERN_ERR "VFP: s%2u: 0x%08x s%2u: 0x%08x\n",
 136                        i, vfp_get_float(i), i+1, vfp_get_float(i+1));
 137 }
 138
 139 /*
 140  * Process bitmask of exception conditions.
 141  */
 142 static void vfp_raise_exceptions(u32 exceptions, u32 inst, u32 fpscr, struct pt_regs *regs)
 143 {
 144         int si_code = 0;
 145
 146         pr_debug("VFP: raising exceptions %08x\n", exceptions);
 147
 148         if (exceptions == VFP_EXCEPTION_ERROR) {
 149                 vfp_panic("unhandled bounce");
 150                 vfp_raise_sigfpe(0, regs);
 151                 return;
 152         }
 153
 154         /*
 155          * If any of the status flags are set, update the FPSCR.
 156          * Comparison instructions always return at least one of
 157          * these flags set.
 158          */
 159         if (exceptions & (FPSCR_N|FPSCR_Z|FPSCR_C|FPSCR_V))
 160                 fpscr &= ~(FPSCR_N|FPSCR_Z|FPSCR_C|FPSCR_V);
 161
 162         fpscr |= exceptions;
 163
 164         fmxr(FPSCR, fpscr);
 165
 166 #define RAISE(stat,en,sig)                              \
 167         if (exceptions & stat && fpscr & en)            \
 168                 si_code = sig;
 169
 170         /*
 171          * These are arranged in priority order, least to highest.
 172          */
 173         RAISE(FPSCR_DZC, FPSCR_DZE, FPE_FLTDIV);
 174         RAISE(FPSCR_IXC, FPSCR_IXE, FPE_FLTRES);
 175         RAISE(FPSCR_UFC, FPSCR_UFE, FPE_FLTUND);
 176         RAISE(FPSCR_OFC, FPSCR_OFE, FPE_FLTOVF);
 177         RAISE(FPSCR_IOC, FPSCR_IOE, FPE_FLTINV);
 178
 179         if (si_code)
 180                 vfp_raise_sigfpe(si_code, regs);
 181 }
 182
 183 /*
 184  * Emulate a VFP instruction.
 185  */
 186 static u32 vfp_emulate_instruction(u32 inst, u32 fpscr, struct pt_regs *regs)
 187 {
 188         u32 exceptions = VFP_EXCEPTION_ERROR;
 189
 190         pr_debug("VFP: emulate: INST=0x%08x SCR=0x%08x\n", inst, fpscr);
 191
 192         if (INST_CPRTDO(inst)) {
 193                 if (!INST_CPRT(inst)) {
 194                         /*
 195                          * CPDO
 196                          */
 197                         if (vfp_single(inst)) {
 198                                 exceptions = vfp_single_cpdo(inst, fpscr);
 199                         } else {
 200                                 exceptions = vfp_double_cpdo(inst, fpscr);
 201                         }
 202                 } else {
 203                         /*
 204                          * A CPRT instruction can not appear in FPINST2, nor
 205                          * can it cause an exception.  Therefore, we do not
 206                          * have to emulate it.
 207                          */
 208                 }
 209         } else {
 210                 /*
 211                  * A CPDT instruction can not appear in FPINST2, nor can
 212                  * it cause an exception.  Therefore, we do not have to
 213                  * emulate it.
 214                  */
 215         }
 216         return exceptions & ~VFP_NAN_FLAG;
 217 }
 218
 219 /*
 220  * Package up a bounce condition.
 221  */
 222 void VFP9_bounce(u32 trigger, u32 fpexc, struct pt_regs *regs)
 223 {
 224         u32 fpscr, orig_fpscr, exceptions, inst;
 225
 226         pr_debug("VFP: bounce: trigger %08x fpexc %08x\n", trigger, fpexc);
 227
 228         /*
 229          * Enable access to the VFP so we can handle the bounce.
 230          */
 231         fmxr(FPEXC, fpexc & ~(FPEXC_EXCEPTION|FPEXC_INV|FPEXC_UFC|FPEXC_IOC));
 232
 233         orig_fpscr = fpscr = fmrx(FPSCR);
 234
 235         /*
 236          * If we are running with inexact exceptions enabled, we need to
 237          * emulate the trigger instruction.  Note that as we're emulating
 238          * the trigger instruction, we need to increment PC.
 239          */
 240         if (fpscr & FPSCR_IXE) {
 241                 regs->ARM_pc += 4;
 242                 goto emulate;
 243         }
 244
 245         barrier();
 246
 247         /*
 248          * Modify fpscr to indicate the number of iterations remaining
 249          */
 250         if (fpexc & FPEXC_EXCEPTION) {
 251                 u32 len;
 252
 253                 len = fpexc + (1 << FPEXC_LENGTH_BIT);
 254
 255                 fpscr &= ~FPSCR_LENGTH_MASK;
 256                 fpscr |= (len & FPEXC_LENGTH_MASK) << (FPSCR_LENGTH_BIT - FPEXC_LENGTH_BIT);
 257         }
 258
 259         /*
 260          * Handle the first FP instruction.  We used to take note of the
 261          * FPEXC bounce reason, but this appears to be unreliable.
 262          * Emulate the bounced instruction instead.
 263          */
 264         inst = fmrx(FPINST);
 265         exceptions = vfp_emulate_instruction(inst, fpscr, regs);
 266         if (exceptions)
 267                 vfp_raise_exceptions(exceptions, inst, orig_fpscr, regs);
 268
 269         /*
 270          * If there isn't a second FP instruction, exit now.
 271          */
 272         if (!(fpexc & FPEXC_FPV2))
 273                 return;
 274
 275         /*
 276          * The barrier() here prevents fpinst2 being read
 277          * before the condition above.
 278          */
 279         barrier();
 280         trigger = fmrx(FPINST2);
 281         orig_fpscr = fpscr = fmrx(FPSCR);
 282
 283  emulate:
 284         exceptions = vfp_emulate_instruction(trigger, fpscr, regs);
 285         if (exceptions)
 286                 vfp_raise_exceptions(exceptions, trigger, orig_fpscr, regs);
 287 }
 288
 289 static void vfp_enable(void *unused)
 290 {
 291         u32 access = get_copro_access();
 292
 293         /*
 294          * Enable full access to VFP (cp10 and cp11)
 295          */
 296         set_copro_access(access | CPACC_FULL(10) | CPACC_FULL(11));
 297 }
 298
 299 #include <linux/smp.h>
 300
 301 /*
 302  * VFP support code initialisation.
 303  */
 304 static int __init vfp_init(void)
 305 {
 306         unsigned int vfpsid;
 307         unsigned int cpu_arch = cpu_architecture();
 308         u32 access = 0;
 309
 310         if (cpu_arch >= CPU_ARCH_ARMv6) {
 311                 access = get_copro_access();
 312
 313                 /*
 314                  * Enable full access to VFP (cp10 and cp11)
 315                  */
 316                 set_copro_access(access | CPACC_FULL(10) | CPACC_FULL(11));
 317         }
 318
 319         /*
 320          * First check that there is a VFP that we can use.
 321          * The handler is already setup to just log calls, so
 322          * we just need to read the VFPSID register.
 323          */
 324         vfpsid = fmrx(FPSID);
 325         barrier();
 326
 327         printk(KERN_INFO "VFP support v0.3: ");
 328         if (VFP_arch) {
 329                 printk("not present\n");
 330
 331                 /*
 332                  * Restore the copro access register.
 333                  */
 334                 if (cpu_arch >= CPU_ARCH_ARMv6)
 335                         set_copro_access(access);
 336         } else if (vfpsid & FPSID_NODOUBLE) {
 337                 printk("no double precision support\n");
 338         } else {
 339                 smp_call_function(vfp_enable, NULL, 1, 1);
 340
 341                 VFP_arch = (vfpsid & FPSID_ARCH_MASK) >> FPSID_ARCH_BIT;  /* Extract the architecture version */
 342                 printk("implementor %02x architecture %d part %02x variant %x rev %x\n",
 343                         (vfpsid & FPSID_IMPLEMENTER_MASK) >> FPSID_IMPLEMENTER_BIT,
 344                         (vfpsid & FPSID_ARCH_MASK) >> FPSID_ARCH_BIT,
 345                         (vfpsid & FPSID_PART_MASK) >> FPSID_PART_BIT,
 346                         (vfpsid & FPSID_VARIANT_MASK) >> FPSID_VARIANT_BIT,
 347                         (vfpsid & FPSID_REV_MASK) >> FPSID_REV_BIT);
 348
 349                 vfp_vector = vfp_support_entry;
 350
 351                 thread_register_notifier(&vfp_notifier_block);
 352
 353                 /*
 354                  * We detected VFP, and the support code is
 355                  * in place; report VFP support to userspace.
 356                  */
 357                 elf_hwcap |= HWCAP_VFP;
 358         }
 359         return 0;
 360 }
 361
 362 late_initcall(vfp_init);