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 void vfp_null_entry(void);
  30
  31 void (*vfp_vector)(void) = vfp_null_entry;
  32 union vfp_state *last_VFP_context[NR_CPUS];
  33
  34 /*
  35  * Dual-use variable.
  36  * Used in startup: set to non-zero if VFP checks fail
  37  * After startup, holds VFP architecture
  38  */
  39 unsigned int VFP_arch;
  40
  41 static int vfp_notifier(struct notifier_block *self, unsigned long cmd, void *v)
  42 {
  43         struct thread_info *thread = v;
  44         union vfp_state *vfp;
  45         __u32 cpu = thread->cpu;
  46
  47         if (likely(cmd == THREAD_NOTIFY_SWITCH)) {
  48                 u32 fpexc = fmrx(FPEXC);
  49
  50 #ifdef CONFIG_SMP
  51                 /*
  52                  * On SMP, if VFP is enabled, save the old state in
  53                  * case the thread migrates to a different CPU. The
  54                  * restoring is done lazily.
  55                  */
  56                 if ((fpexc & FPEXC_EN) && last_VFP_context[cpu]) {
  57                         vfp_save_state(last_VFP_context[cpu], fpexc);
  58                         last_VFP_context[cpu]->hard.cpu = cpu;
  59                 }
  60                 /*
  61                  * Thread migration, just force the reloading of the
  62                  * state on the new CPU in case the VFP registers
  63                  * contain stale data.
  64                  */
  65                 if (thread->vfpstate.hard.cpu != cpu)
  66                         last_VFP_context[cpu] = NULL;
  67 #endif
  68
  69                 /*
  70                  * Always disable VFP so we can lazily save/restore the
  71                  * old state.
  72                  */
  73                 fmxr(FPEXC, fpexc & ~FPEXC_EN);
  74                 return NOTIFY_DONE;
  75         }
  76
  77         vfp = &thread->vfpstate;
  78         if (cmd == THREAD_NOTIFY_FLUSH) {
  79                 /*
  80                  * Per-thread VFP initialisation.
  81                  */
  82                 memset(vfp, 0, sizeof(union vfp_state));
  83
  84                 vfp->hard.fpexc = FPEXC_EN;
  85                 vfp->hard.fpscr = FPSCR_ROUND_NEAREST;
  86
  87                 /*
  88                  * Disable VFP to ensure we initialise it first.
  89                  */
  90                 fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
  91         }
  92
  93         /* flush and release case: Per-thread VFP cleanup. */
  94         if (last_VFP_context[cpu] == vfp)
  95                 last_VFP_context[cpu] = NULL;
  96
  97         return NOTIFY_DONE;
  98 }
  99
 100 static struct notifier_block vfp_notifier_block = {
 101         .notifier_call  = vfp_notifier,
 102 };
 103
 104 /*
 105  * Raise a SIGFPE for the current process.
 106  * sicode describes the signal being raised.
 107  */
 108 void vfp_raise_sigfpe(unsigned int sicode, struct pt_regs *regs)
 109 {
 110         siginfo_t info;
 111
 112         memset(&info, 0, sizeof(info));
 113
 114         info.si_signo = SIGFPE;
 115         info.si_code = sicode;
 116         info.si_addr = (void __user *)(instruction_pointer(regs) - 4);
 117
 118         /*
 119          * This is the same as NWFPE, because it's not clear what
 120          * this is used for
 121          */
 122         current->thread.error_code = 0;
 123         current->thread.trap_no = 6;
 124
 125         send_sig_info(SIGFPE, &info, current);
 126 }
 127
 128 static void vfp_panic(char *reason)
 129 {
 130         int i;
 131
 132         printk(KERN_ERR "VFP: Error: %s\n", reason);
 133         printk(KERN_ERR "VFP: EXC 0x%08x SCR 0x%08x INST 0x%08x\n",
 134                 fmrx(FPEXC), fmrx(FPSCR), fmrx(FPINST));
 135         for (i = 0; i < 32; i += 2)
 136                 printk(KERN_ERR "VFP: s%2u: 0x%08x s%2u: 0x%08x\n",
 137                        i, vfp_get_float(i), i+1, vfp_get_float(i+1));
 138 }
 139
 140 /*
 141  * Process bitmask of exception conditions.
 142  */
 143 static void vfp_raise_exceptions(u32 exceptions, u32 inst, u32 fpscr, struct pt_regs *regs)
 144 {
 145         int si_code = 0;
 146
 147         pr_debug("VFP: raising exceptions %08x\n", exceptions);
 148
 149         if (exceptions == VFP_EXCEPTION_ERROR) {
 150                 vfp_panic("unhandled bounce");
 151                 vfp_raise_sigfpe(0, regs);
 152                 return;
 153         }
 154
 155         /*
 156          * If any of the status flags are set, update the FPSCR.
 157          * Comparison instructions always return at least one of
 158          * these flags set.
 159          */
 160         if (exceptions & (FPSCR_N|FPSCR_Z|FPSCR_C|FPSCR_V))
 161                 fpscr &= ~(FPSCR_N|FPSCR_Z|FPSCR_C|FPSCR_V);
 162
 163         fpscr |= exceptions;
 164
 165         fmxr(FPSCR, fpscr);
 166
 167 #define RAISE(stat,en,sig)                              \
 168         if (exceptions & stat && fpscr & en)            \
 169                 si_code = sig;
 170
 171         /*
 172          * These are arranged in priority order, least to highest.
 173          */
 174         RAISE(FPSCR_DZC, FPSCR_DZE, FPE_FLTDIV);
 175         RAISE(FPSCR_IXC, FPSCR_IXE, FPE_FLTRES);
 176         RAISE(FPSCR_UFC, FPSCR_UFE, FPE_FLTUND);
 177         RAISE(FPSCR_OFC, FPSCR_OFE, FPE_FLTOVF);
 178         RAISE(FPSCR_IOC, FPSCR_IOE, FPE_FLTINV);
 179
 180         if (si_code)
 181                 vfp_raise_sigfpe(si_code, regs);
 182 }
 183
 184 /*
 185  * Emulate a VFP instruction.
 186  */
 187 static u32 vfp_emulate_instruction(u32 inst, u32 fpscr, struct pt_regs *regs)
 188 {
 189         u32 exceptions = VFP_EXCEPTION_ERROR;
 190
 191         pr_debug("VFP: emulate: INST=0x%08x SCR=0x%08x\n", inst, fpscr);
 192
 193         if (INST_CPRTDO(inst)) {
 194                 if (!INST_CPRT(inst)) {
 195                         /*
 196                          * CPDO
 197                          */
 198                         if (vfp_single(inst)) {
 199                                 exceptions = vfp_single_cpdo(inst, fpscr);
 200                         } else {
 201                                 exceptions = vfp_double_cpdo(inst, fpscr);
 202                         }
 203                 } else {
 204                         /*
 205                          * A CPRT instruction can not appear in FPINST2, nor
 206                          * can it cause an exception.  Therefore, we do not
 207                          * have to emulate it.
 208                          */
 209                 }
 210         } else {
 211                 /*
 212                  * A CPDT instruction can not appear in FPINST2, nor can
 213                  * it cause an exception.  Therefore, we do not have to
 214                  * emulate it.
 215                  */
 216         }
 217         return exceptions & ~VFP_NAN_FLAG;
 218 }
 219
 220 /*
 221  * Package up a bounce condition.
 222  */
 223 void VFP9_bounce(u32 trigger, u32 fpexc, struct pt_regs *regs)
 224 {
 225         u32 fpscr, orig_fpscr, exceptions, inst;
 226
 227         pr_debug("VFP: bounce: trigger %08x fpexc %08x\n", trigger, fpexc);
 228
 229         /*
 230          * Enable access to the VFP so we can handle the bounce.
 231          */
 232         fmxr(FPEXC, fpexc & ~(FPEXC_EX|FPEXC_INV|FPEXC_UFC|FPEXC_IOC));
 233
 234         orig_fpscr = fpscr = fmrx(FPSCR);
 235
 236         /*
 237          * If we are running with inexact exceptions enabled, we need to
 238          * emulate the trigger instruction.  Note that as we're emulating
 239          * the trigger instruction, we need to increment PC.
 240          */
 241         if (fpscr & FPSCR_IXE) {
 242                 regs->ARM_pc += 4;
 243                 goto emulate;
 244         }
 245
 246         barrier();
 247
 248         /*
 249          * Modify fpscr to indicate the number of iterations remaining
 250          */
 251         if (fpexc & FPEXC_EX) {
 252                 u32 len;
 253
 254                 len = fpexc + (1 << FPEXC_LENGTH_BIT);
 255
 256                 fpscr &= ~FPSCR_LENGTH_MASK;
 257                 fpscr |= (len & FPEXC_LENGTH_MASK) << (FPSCR_LENGTH_BIT - FPEXC_LENGTH_BIT);
 258         }
 259
 260         /*
 261          * Handle the first FP instruction.  We used to take note of the
 262          * FPEXC bounce reason, but this appears to be unreliable.
 263          * Emulate the bounced instruction instead.
 264          */
 265         inst = fmrx(FPINST);
 266         exceptions = vfp_emulate_instruction(inst, fpscr, regs);
 267         if (exceptions)
 268                 vfp_raise_exceptions(exceptions, inst, orig_fpscr, regs);
 269
 270         /*
 271          * If there isn't a second FP instruction, exit now.
 272          */
 273         if (!(fpexc & FPEXC_FPV2))
 274                 return;
 275
 276         /*
 277          * The barrier() here prevents fpinst2 being read
 278          * before the condition above.
 279          */
 280         barrier();
 281         trigger = fmrx(FPINST2);
 282         orig_fpscr = fpscr = fmrx(FPSCR);
 283
 284  emulate:
 285         exceptions = vfp_emulate_instruction(trigger, fpscr, regs);
 286         if (exceptions)
 287                 vfp_raise_exceptions(exceptions, trigger, orig_fpscr, regs);
 288 }
 289
 290 static void vfp_enable(void *unused)
 291 {
 292         u32 access = get_copro_access();
 293
 294         /*
 295          * Enable full access to VFP (cp10 and cp11)
 296          */
 297         set_copro_access(access | CPACC_FULL(10) | CPACC_FULL(11));
 298 }
 299
 300 #include <linux/smp.h>
 301
 302 /*
 303  * VFP support code initialisation.
 304  */
 305 static int __init vfp_init(void)
 306 {
 307         unsigned int vfpsid;
 308         unsigned int cpu_arch = cpu_architecture();
 309         u32 access = 0;
 310
 311         if (cpu_arch >= CPU_ARCH_ARMv6) {
 312                 access = get_copro_access();
 313
 314                 /*
 315                  * Enable full access to VFP (cp10 and cp11)
 316                  */
 317                 set_copro_access(access | CPACC_FULL(10) | CPACC_FULL(11));
 318         }
 319
 320         /*
 321          * First check that there is a VFP that we can use.
 322          * The handler is already setup to just log calls, so
 323          * we just need to read the VFPSID register.
 324          */
 325         vfp_vector = vfp_testing_entry;
 326         vfpsid = fmrx(FPSID);
 327         barrier();
 328         vfp_vector = vfp_null_entry;
 329
 330         printk(KERN_INFO "VFP support v0.3: ");
 331         if (VFP_arch) {
 332                 printk("not present\n");
 333
 334                 /*
 335                  * Restore the copro access register.
 336                  */
 337                 if (cpu_arch >= CPU_ARCH_ARMv6)
 338                         set_copro_access(access);
 339         } else if (vfpsid & FPSID_NODOUBLE) {
 340                 printk("no double precision support\n");
 341         } else {
 342                 smp_call_function(vfp_enable, NULL, 1, 1);
 343
 344                 VFP_arch = (vfpsid & FPSID_ARCH_MASK) >> FPSID_ARCH_BIT;  /* Extract the architecture version */
 345                 printk("implementor %02x architecture %d part %02x variant %x rev %x\n",
 346                         (vfpsid & FPSID_IMPLEMENTER_MASK) >> FPSID_IMPLEMENTER_BIT,
 347                         (vfpsid & FPSID_ARCH_MASK) >> FPSID_ARCH_BIT,
 348                         (vfpsid & FPSID_PART_MASK) >> FPSID_PART_BIT,
 349                         (vfpsid & FPSID_VARIANT_MASK) >> FPSID_VARIANT_BIT,
 350                         (vfpsid & FPSID_REV_MASK) >> FPSID_REV_BIT);
 351
 352                 vfp_vector = vfp_support_entry;
 353
 354                 thread_register_notifier(&vfp_notifier_block);
 355
 356                 /*
 357                  * We detected VFP, and the support code is
 358                  * in place; report VFP support to userspace.
 359                  */
 360                 elf_hwcap |= HWCAP_VFP;
 361         }
 362         return 0;
 363 }
 364
 365 late_initcall(vfp_init);