Merge remote-tracking branch 'moduleh/module.h-split'
[linux-2.6/next.git] / arch / arm / vfp / vfpmodule.c
blob2d17f264847cb02f38177dacad1d968c69cc3cbc
1 /*
2 * linux/arch/arm/vfp/vfpmodule.c
4 * Copyright (C) 2004 ARM Limited.
5 * Written by Deep Blue Solutions Limited.
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.
11 #include <linux/types.h>
12 #include <linux/cpu.h>
13 #include <linux/kernel.h>
14 #include <linux/notifier.h>
15 #include <linux/signal.h>
16 #include <linux/sched.h>
17 #include <linux/smp.h>
18 #include <linux/init.h>
20 #include <asm/cputype.h>
21 #include <asm/thread_notify.h>
22 #include <asm/vfp.h>
24 #include "vfpinstr.h"
25 #include "vfp.h"
28 * Our undef handlers (in entry.S)
30 void vfp_testing_entry(void);
31 void vfp_support_entry(void);
32 void vfp_null_entry(void);
34 void (*vfp_vector)(void) = vfp_null_entry;
37 * Dual-use variable.
38 * Used in startup: set to non-zero if VFP checks fail
39 * After startup, holds VFP architecture
41 unsigned int VFP_arch;
44 * The pointer to the vfpstate structure of the thread which currently
45 * owns the context held in the VFP hardware, or NULL if the hardware
46 * context is invalid.
48 * For UP, this is sufficient to tell which thread owns the VFP context.
49 * However, for SMP, we also need to check the CPU number stored in the
50 * saved state too to catch migrations.
52 union vfp_state *vfp_current_hw_state[NR_CPUS];
55 * Is 'thread's most up to date state stored in this CPUs hardware?
56 * Must be called from non-preemptible context.
58 static bool vfp_state_in_hw(unsigned int cpu, struct thread_info *thread)
60 #ifdef CONFIG_SMP
61 if (thread->vfpstate.hard.cpu != cpu)
62 return false;
63 #endif
64 return vfp_current_hw_state[cpu] == &thread->vfpstate;
68 * Force a reload of the VFP context from the thread structure. We do
69 * this by ensuring that access to the VFP hardware is disabled, and
70 * clear last_VFP_context. Must be called from non-preemptible context.
72 static void vfp_force_reload(unsigned int cpu, struct thread_info *thread)
74 if (vfp_state_in_hw(cpu, thread)) {
75 fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
76 vfp_current_hw_state[cpu] = NULL;
78 #ifdef CONFIG_SMP
79 thread->vfpstate.hard.cpu = NR_CPUS;
80 #endif
84 * Per-thread VFP initialization.
86 static void vfp_thread_flush(struct thread_info *thread)
88 union vfp_state *vfp = &thread->vfpstate;
89 unsigned int cpu;
92 * Disable VFP to ensure we initialize it first. We must ensure
93 * that the modification of vfp_current_hw_state[] and hardware
94 * disable are done for the same CPU and without preemption.
96 * Do this first to ensure that preemption won't overwrite our
97 * state saving should access to the VFP be enabled at this point.
99 cpu = get_cpu();
100 if (vfp_current_hw_state[cpu] == vfp)
101 vfp_current_hw_state[cpu] = NULL;
102 fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
103 put_cpu();
105 memset(vfp, 0, sizeof(union vfp_state));
107 vfp->hard.fpexc = FPEXC_EN;
108 vfp->hard.fpscr = FPSCR_ROUND_NEAREST;
109 #ifdef CONFIG_SMP
110 vfp->hard.cpu = NR_CPUS;
111 #endif
114 static void vfp_thread_exit(struct thread_info *thread)
116 /* release case: Per-thread VFP cleanup. */
117 union vfp_state *vfp = &thread->vfpstate;
118 unsigned int cpu = get_cpu();
120 if (vfp_current_hw_state[cpu] == vfp)
121 vfp_current_hw_state[cpu] = NULL;
122 put_cpu();
125 static void vfp_thread_copy(struct thread_info *thread)
127 struct thread_info *parent = current_thread_info();
129 vfp_sync_hwstate(parent);
130 thread->vfpstate = parent->vfpstate;
131 #ifdef CONFIG_SMP
132 thread->vfpstate.hard.cpu = NR_CPUS;
133 #endif
137 * When this function is called with the following 'cmd's, the following
138 * is true while this function is being run:
139 * THREAD_NOFTIFY_SWTICH:
140 * - the previously running thread will not be scheduled onto another CPU.
141 * - the next thread to be run (v) will not be running on another CPU.
142 * - thread->cpu is the local CPU number
143 * - not preemptible as we're called in the middle of a thread switch
144 * THREAD_NOTIFY_FLUSH:
145 * - the thread (v) will be running on the local CPU, so
146 * v === current_thread_info()
147 * - thread->cpu is the local CPU number at the time it is accessed,
148 * but may change at any time.
149 * - we could be preempted if tree preempt rcu is enabled, so
150 * it is unsafe to use thread->cpu.
151 * THREAD_NOTIFY_EXIT
152 * - the thread (v) will be running on the local CPU, so
153 * v === current_thread_info()
154 * - thread->cpu is the local CPU number at the time it is accessed,
155 * but may change at any time.
156 * - we could be preempted if tree preempt rcu is enabled, so
157 * it is unsafe to use thread->cpu.
159 static int vfp_notifier(struct notifier_block *self, unsigned long cmd, void *v)
161 struct thread_info *thread = v;
162 u32 fpexc;
163 #ifdef CONFIG_SMP
164 unsigned int cpu;
165 #endif
167 switch (cmd) {
168 case THREAD_NOTIFY_SWITCH:
169 fpexc = fmrx(FPEXC);
171 #ifdef CONFIG_SMP
172 cpu = thread->cpu;
175 * On SMP, if VFP is enabled, save the old state in
176 * case the thread migrates to a different CPU. The
177 * restoring is done lazily.
179 if ((fpexc & FPEXC_EN) && vfp_current_hw_state[cpu])
180 vfp_save_state(vfp_current_hw_state[cpu], fpexc);
181 #endif
184 * Always disable VFP so we can lazily save/restore the
185 * old state.
187 fmxr(FPEXC, fpexc & ~FPEXC_EN);
188 break;
190 case THREAD_NOTIFY_FLUSH:
191 vfp_thread_flush(thread);
192 break;
194 case THREAD_NOTIFY_EXIT:
195 vfp_thread_exit(thread);
196 break;
198 case THREAD_NOTIFY_COPY:
199 vfp_thread_copy(thread);
200 break;
203 return NOTIFY_DONE;
206 static struct notifier_block vfp_notifier_block = {
207 .notifier_call = vfp_notifier,
211 * Raise a SIGFPE for the current process.
212 * sicode describes the signal being raised.
214 static void vfp_raise_sigfpe(unsigned int sicode, struct pt_regs *regs)
216 siginfo_t info;
218 memset(&info, 0, sizeof(info));
220 info.si_signo = SIGFPE;
221 info.si_code = sicode;
222 info.si_addr = (void __user *)(instruction_pointer(regs) - 4);
225 * This is the same as NWFPE, because it's not clear what
226 * this is used for
228 current->thread.error_code = 0;
229 current->thread.trap_no = 6;
231 send_sig_info(SIGFPE, &info, current);
234 static void vfp_panic(char *reason, u32 inst)
236 int i;
238 printk(KERN_ERR "VFP: Error: %s\n", reason);
239 printk(KERN_ERR "VFP: EXC 0x%08x SCR 0x%08x INST 0x%08x\n",
240 fmrx(FPEXC), fmrx(FPSCR), inst);
241 for (i = 0; i < 32; i += 2)
242 printk(KERN_ERR "VFP: s%2u: 0x%08x s%2u: 0x%08x\n",
243 i, vfp_get_float(i), i+1, vfp_get_float(i+1));
247 * Process bitmask of exception conditions.
249 static void vfp_raise_exceptions(u32 exceptions, u32 inst, u32 fpscr, struct pt_regs *regs)
251 int si_code = 0;
253 pr_debug("VFP: raising exceptions %08x\n", exceptions);
255 if (exceptions == VFP_EXCEPTION_ERROR) {
256 vfp_panic("unhandled bounce", inst);
257 vfp_raise_sigfpe(0, regs);
258 return;
262 * If any of the status flags are set, update the FPSCR.
263 * Comparison instructions always return at least one of
264 * these flags set.
266 if (exceptions & (FPSCR_N|FPSCR_Z|FPSCR_C|FPSCR_V))
267 fpscr &= ~(FPSCR_N|FPSCR_Z|FPSCR_C|FPSCR_V);
269 fpscr |= exceptions;
271 fmxr(FPSCR, fpscr);
273 #define RAISE(stat,en,sig) \
274 if (exceptions & stat && fpscr & en) \
275 si_code = sig;
278 * These are arranged in priority order, least to highest.
280 RAISE(FPSCR_DZC, FPSCR_DZE, FPE_FLTDIV);
281 RAISE(FPSCR_IXC, FPSCR_IXE, FPE_FLTRES);
282 RAISE(FPSCR_UFC, FPSCR_UFE, FPE_FLTUND);
283 RAISE(FPSCR_OFC, FPSCR_OFE, FPE_FLTOVF);
284 RAISE(FPSCR_IOC, FPSCR_IOE, FPE_FLTINV);
286 if (si_code)
287 vfp_raise_sigfpe(si_code, regs);
291 * Emulate a VFP instruction.
293 static u32 vfp_emulate_instruction(u32 inst, u32 fpscr, struct pt_regs *regs)
295 u32 exceptions = VFP_EXCEPTION_ERROR;
297 pr_debug("VFP: emulate: INST=0x%08x SCR=0x%08x\n", inst, fpscr);
299 if (INST_CPRTDO(inst)) {
300 if (!INST_CPRT(inst)) {
302 * CPDO
304 if (vfp_single(inst)) {
305 exceptions = vfp_single_cpdo(inst, fpscr);
306 } else {
307 exceptions = vfp_double_cpdo(inst, fpscr);
309 } else {
311 * A CPRT instruction can not appear in FPINST2, nor
312 * can it cause an exception. Therefore, we do not
313 * have to emulate it.
316 } else {
318 * A CPDT instruction can not appear in FPINST2, nor can
319 * it cause an exception. Therefore, we do not have to
320 * emulate it.
323 return exceptions & ~VFP_NAN_FLAG;
327 * Package up a bounce condition.
329 void VFP_bounce(u32 trigger, u32 fpexc, struct pt_regs *regs)
331 u32 fpscr, orig_fpscr, fpsid, exceptions;
333 pr_debug("VFP: bounce: trigger %08x fpexc %08x\n", trigger, fpexc);
336 * At this point, FPEXC can have the following configuration:
338 * EX DEX IXE
339 * 0 1 x - synchronous exception
340 * 1 x 0 - asynchronous exception
341 * 1 x 1 - sychronous on VFP subarch 1 and asynchronous on later
342 * 0 0 1 - synchronous on VFP9 (non-standard subarch 1
343 * implementation), undefined otherwise
345 * Clear various bits and enable access to the VFP so we can
346 * handle the bounce.
348 fmxr(FPEXC, fpexc & ~(FPEXC_EX|FPEXC_DEX|FPEXC_FP2V|FPEXC_VV|FPEXC_TRAP_MASK));
350 fpsid = fmrx(FPSID);
351 orig_fpscr = fpscr = fmrx(FPSCR);
354 * Check for the special VFP subarch 1 and FPSCR.IXE bit case
356 if ((fpsid & FPSID_ARCH_MASK) == (1 << FPSID_ARCH_BIT)
357 && (fpscr & FPSCR_IXE)) {
359 * Synchronous exception, emulate the trigger instruction
361 goto emulate;
364 if (fpexc & FPEXC_EX) {
365 #ifndef CONFIG_CPU_FEROCEON
367 * Asynchronous exception. The instruction is read from FPINST
368 * and the interrupted instruction has to be restarted.
370 trigger = fmrx(FPINST);
371 regs->ARM_pc -= 4;
372 #endif
373 } else if (!(fpexc & FPEXC_DEX)) {
375 * Illegal combination of bits. It can be caused by an
376 * unallocated VFP instruction but with FPSCR.IXE set and not
377 * on VFP subarch 1.
379 vfp_raise_exceptions(VFP_EXCEPTION_ERROR, trigger, fpscr, regs);
380 goto exit;
384 * Modify fpscr to indicate the number of iterations remaining.
385 * If FPEXC.EX is 0, FPEXC.DEX is 1 and the FPEXC.VV bit indicates
386 * whether FPEXC.VECITR or FPSCR.LEN is used.
388 if (fpexc & (FPEXC_EX | FPEXC_VV)) {
389 u32 len;
391 len = fpexc + (1 << FPEXC_LENGTH_BIT);
393 fpscr &= ~FPSCR_LENGTH_MASK;
394 fpscr |= (len & FPEXC_LENGTH_MASK) << (FPSCR_LENGTH_BIT - FPEXC_LENGTH_BIT);
398 * Handle the first FP instruction. We used to take note of the
399 * FPEXC bounce reason, but this appears to be unreliable.
400 * Emulate the bounced instruction instead.
402 exceptions = vfp_emulate_instruction(trigger, fpscr, regs);
403 if (exceptions)
404 vfp_raise_exceptions(exceptions, trigger, orig_fpscr, regs);
407 * If there isn't a second FP instruction, exit now. Note that
408 * the FPEXC.FP2V bit is valid only if FPEXC.EX is 1.
410 if (fpexc ^ (FPEXC_EX | FPEXC_FP2V))
411 goto exit;
414 * The barrier() here prevents fpinst2 being read
415 * before the condition above.
417 barrier();
418 trigger = fmrx(FPINST2);
420 emulate:
421 exceptions = vfp_emulate_instruction(trigger, orig_fpscr, regs);
422 if (exceptions)
423 vfp_raise_exceptions(exceptions, trigger, orig_fpscr, regs);
424 exit:
425 preempt_enable();
428 static void vfp_enable(void *unused)
430 u32 access = get_copro_access();
433 * Enable full access to VFP (cp10 and cp11)
435 set_copro_access(access | CPACC_FULL(10) | CPACC_FULL(11));
438 #ifdef CONFIG_PM
439 #include <linux/syscore_ops.h>
441 static int vfp_pm_suspend(void)
443 struct thread_info *ti = current_thread_info();
444 u32 fpexc = fmrx(FPEXC);
446 /* if vfp is on, then save state for resumption */
447 if (fpexc & FPEXC_EN) {
448 printk(KERN_DEBUG "%s: saving vfp state\n", __func__);
449 vfp_save_state(&ti->vfpstate, fpexc);
451 /* disable, just in case */
452 fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
455 /* clear any information we had about last context state */
456 memset(vfp_current_hw_state, 0, sizeof(vfp_current_hw_state));
458 return 0;
461 static void vfp_pm_resume(void)
463 /* ensure we have access to the vfp */
464 vfp_enable(NULL);
466 /* and disable it to ensure the next usage restores the state */
467 fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
470 static struct syscore_ops vfp_pm_syscore_ops = {
471 .suspend = vfp_pm_suspend,
472 .resume = vfp_pm_resume,
475 static void vfp_pm_init(void)
477 register_syscore_ops(&vfp_pm_syscore_ops);
480 #else
481 static inline void vfp_pm_init(void) { }
482 #endif /* CONFIG_PM */
485 * Ensure that the VFP state stored in 'thread->vfpstate' is up to date
486 * with the hardware state.
488 void vfp_sync_hwstate(struct thread_info *thread)
490 unsigned int cpu = get_cpu();
492 if (vfp_state_in_hw(cpu, thread)) {
493 u32 fpexc = fmrx(FPEXC);
496 * Save the last VFP state on this CPU.
498 fmxr(FPEXC, fpexc | FPEXC_EN);
499 vfp_save_state(&thread->vfpstate, fpexc | FPEXC_EN);
500 fmxr(FPEXC, fpexc);
503 put_cpu();
506 /* Ensure that the thread reloads the hardware VFP state on the next use. */
507 void vfp_flush_hwstate(struct thread_info *thread)
509 unsigned int cpu = get_cpu();
511 vfp_force_reload(cpu, thread);
513 put_cpu();
517 * VFP hardware can lose all context when a CPU goes offline.
518 * As we will be running in SMP mode with CPU hotplug, we will save the
519 * hardware state at every thread switch. We clear our held state when
520 * a CPU has been killed, indicating that the VFP hardware doesn't contain
521 * a threads VFP state. When a CPU starts up, we re-enable access to the
522 * VFP hardware.
524 * Both CPU_DYING and CPU_STARTING are called on the CPU which
525 * is being offlined/onlined.
527 static int vfp_hotplug(struct notifier_block *b, unsigned long action,
528 void *hcpu)
530 if (action == CPU_DYING || action == CPU_DYING_FROZEN) {
531 vfp_force_reload((long)hcpu, current_thread_info());
532 } else if (action == CPU_STARTING || action == CPU_STARTING_FROZEN)
533 vfp_enable(NULL);
534 return NOTIFY_OK;
538 * VFP support code initialisation.
540 static int __init vfp_init(void)
542 unsigned int vfpsid;
543 unsigned int cpu_arch = cpu_architecture();
545 if (cpu_arch >= CPU_ARCH_ARMv6)
546 vfp_enable(NULL);
549 * First check that there is a VFP that we can use.
550 * The handler is already setup to just log calls, so
551 * we just need to read the VFPSID register.
553 vfp_vector = vfp_testing_entry;
554 barrier();
555 vfpsid = fmrx(FPSID);
556 barrier();
557 vfp_vector = vfp_null_entry;
559 printk(KERN_INFO "VFP support v0.3: ");
560 if (VFP_arch)
561 printk("not present\n");
562 else if (vfpsid & FPSID_NODOUBLE) {
563 printk("no double precision support\n");
564 } else {
565 hotcpu_notifier(vfp_hotplug, 0);
567 smp_call_function(vfp_enable, NULL, 1);
569 VFP_arch = (vfpsid & FPSID_ARCH_MASK) >> FPSID_ARCH_BIT; /* Extract the architecture version */
570 printk("implementor %02x architecture %d part %02x variant %x rev %x\n",
571 (vfpsid & FPSID_IMPLEMENTER_MASK) >> FPSID_IMPLEMENTER_BIT,
572 (vfpsid & FPSID_ARCH_MASK) >> FPSID_ARCH_BIT,
573 (vfpsid & FPSID_PART_MASK) >> FPSID_PART_BIT,
574 (vfpsid & FPSID_VARIANT_MASK) >> FPSID_VARIANT_BIT,
575 (vfpsid & FPSID_REV_MASK) >> FPSID_REV_BIT);
577 vfp_vector = vfp_support_entry;
579 thread_register_notifier(&vfp_notifier_block);
580 vfp_pm_init();
583 * We detected VFP, and the support code is
584 * in place; report VFP support to userspace.
586 elf_hwcap |= HWCAP_VFP;
587 #ifdef CONFIG_VFPv3
588 if (VFP_arch >= 2) {
589 elf_hwcap |= HWCAP_VFPv3;
592 * Check for VFPv3 D16. CPUs in this configuration
593 * only have 16 x 64bit registers.
595 if (((fmrx(MVFR0) & MVFR0_A_SIMD_MASK)) == 1)
596 elf_hwcap |= HWCAP_VFPv3D16;
598 #endif
600 * Check for the presence of the Advanced SIMD
601 * load/store instructions, integer and single
602 * precision floating point operations. Only check
603 * for NEON if the hardware has the MVFR registers.
605 if ((read_cpuid_id() & 0x000f0000) == 0x000f0000) {
606 #ifdef CONFIG_NEON
607 if ((fmrx(MVFR1) & 0x000fff00) == 0x00011100)
608 elf_hwcap |= HWCAP_NEON;
609 #endif
610 if ((fmrx(MVFR1) & 0xf0000000) == 0x10000000)
611 elf_hwcap |= HWCAP_VFPv4;
614 return 0;
617 late_initcall(vfp_init);