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/module.h>
12 #include <linux/types.h>
13 #include <linux/cpu.h>
14 #include <linux/cpu_pm.h>
15 #include <linux/kernel.h>
16 #include <linux/notifier.h>
17 #include <linux/signal.h>
18 #include <linux/sched.h>
19 #include <linux/smp.h>
20 #include <linux/init.h>
22 #include <asm/cputype.h>
23 #include <asm/thread_notify.h>
30 * Our undef handlers (in entry.S)
32 void vfp_testing_entry(void);
33 void vfp_support_entry(void);
34 void vfp_null_entry(void);
36 void (*vfp_vector)(void) = vfp_null_entry;
40 * Used in startup: set to non-zero if VFP checks fail
41 * After startup, holds VFP architecture
43 unsigned int VFP_arch;
46 * The pointer to the vfpstate structure of the thread which currently
47 * owns the context held in the VFP hardware, or NULL if the hardware
50 * For UP, this is sufficient to tell which thread owns the VFP context.
51 * However, for SMP, we also need to check the CPU number stored in the
52 * saved state too to catch migrations.
54 union vfp_state *vfp_current_hw_state[NR_CPUS];
57 * Is 'thread's most up to date state stored in this CPUs hardware?
58 * Must be called from non-preemptible context.
60 static bool vfp_state_in_hw(unsigned int cpu, struct thread_info *thread)
63 if (thread->vfpstate.hard.cpu != cpu)
66 return vfp_current_hw_state[cpu] == &thread->vfpstate;
70 * Force a reload of the VFP context from the thread structure. We do
71 * this by ensuring that access to the VFP hardware is disabled, and
72 * clear vfp_current_hw_state. Must be called from non-preemptible context.
74 static void vfp_force_reload(unsigned int cpu, struct thread_info *thread)
76 if (vfp_state_in_hw(cpu, thread)) {
77 fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
78 vfp_current_hw_state[cpu] = NULL;
81 thread->vfpstate.hard.cpu = NR_CPUS;
86 * Per-thread VFP initialization.
88 static void vfp_thread_flush(struct thread_info *thread)
90 union vfp_state *vfp = &thread->vfpstate;
94 * Disable VFP to ensure we initialize it first. We must ensure
95 * that the modification of vfp_current_hw_state[] and hardware
96 * disable are done for the same CPU and without preemption.
98 * Do this first to ensure that preemption won't overwrite our
99 * state saving should access to the VFP be enabled at this point.
102 if (vfp_current_hw_state[cpu] == vfp)
103 vfp_current_hw_state[cpu] = NULL;
104 fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
107 memset(vfp, 0, sizeof(union vfp_state));
109 vfp->hard.fpexc = FPEXC_EN;
110 vfp->hard.fpscr = FPSCR_ROUND_NEAREST;
112 vfp->hard.cpu = NR_CPUS;
116 static void vfp_thread_exit(struct thread_info *thread)
118 /* release case: Per-thread VFP cleanup. */
119 union vfp_state *vfp = &thread->vfpstate;
120 unsigned int cpu = get_cpu();
122 if (vfp_current_hw_state[cpu] == vfp)
123 vfp_current_hw_state[cpu] = NULL;
127 static void vfp_thread_copy(struct thread_info *thread)
129 struct thread_info *parent = current_thread_info();
131 vfp_sync_hwstate(parent);
132 thread->vfpstate = parent->vfpstate;
134 thread->vfpstate.hard.cpu = NR_CPUS;
139 * When this function is called with the following 'cmd's, the following
140 * is true while this function is being run:
141 * THREAD_NOFTIFY_SWTICH:
142 * - the previously running thread will not be scheduled onto another CPU.
143 * - the next thread to be run (v) will not be running on another CPU.
144 * - thread->cpu is the local CPU number
145 * - not preemptible as we're called in the middle of a thread switch
146 * THREAD_NOTIFY_FLUSH:
147 * - the thread (v) will be running on the local CPU, so
148 * v === current_thread_info()
149 * - thread->cpu is the local CPU number at the time it is accessed,
150 * but may change at any time.
151 * - we could be preempted if tree preempt rcu is enabled, so
152 * it is unsafe to use thread->cpu.
154 * - the thread (v) will be running on the local CPU, so
155 * v === current_thread_info()
156 * - thread->cpu is the local CPU number at the time it is accessed,
157 * but may change at any time.
158 * - we could be preempted if tree preempt rcu is enabled, so
159 * it is unsafe to use thread->cpu.
161 static int vfp_notifier(struct notifier_block *self, unsigned long cmd, void *v)
163 struct thread_info *thread = v;
170 case THREAD_NOTIFY_SWITCH:
177 * On SMP, if VFP is enabled, save the old state in
178 * case the thread migrates to a different CPU. The
179 * restoring is done lazily.
181 if ((fpexc & FPEXC_EN) && vfp_current_hw_state[cpu])
182 vfp_save_state(vfp_current_hw_state[cpu], fpexc);
186 * Always disable VFP so we can lazily save/restore the
189 fmxr(FPEXC, fpexc & ~FPEXC_EN);
192 case THREAD_NOTIFY_FLUSH:
193 vfp_thread_flush(thread);
196 case THREAD_NOTIFY_EXIT:
197 vfp_thread_exit(thread);
200 case THREAD_NOTIFY_COPY:
201 vfp_thread_copy(thread);
208 static struct notifier_block vfp_notifier_block = {
209 .notifier_call = vfp_notifier,
213 * Raise a SIGFPE for the current process.
214 * sicode describes the signal being raised.
216 static void vfp_raise_sigfpe(unsigned int sicode, struct pt_regs *regs)
220 memset(&info, 0, sizeof(info));
222 info.si_signo = SIGFPE;
223 info.si_code = sicode;
224 info.si_addr = (void __user *)(instruction_pointer(regs) - 4);
227 * This is the same as NWFPE, because it's not clear what
230 current->thread.error_code = 0;
231 current->thread.trap_no = 6;
233 send_sig_info(SIGFPE, &info, current);
236 static void vfp_panic(char *reason, u32 inst)
240 printk(KERN_ERR "VFP: Error: %s\n", reason);
241 printk(KERN_ERR "VFP: EXC 0x%08x SCR 0x%08x INST 0x%08x\n",
242 fmrx(FPEXC), fmrx(FPSCR), inst);
243 for (i = 0; i < 32; i += 2)
244 printk(KERN_ERR "VFP: s%2u: 0x%08x s%2u: 0x%08x\n",
245 i, vfp_get_float(i), i+1, vfp_get_float(i+1));
249 * Process bitmask of exception conditions.
251 static void vfp_raise_exceptions(u32 exceptions, u32 inst, u32 fpscr, struct pt_regs *regs)
255 pr_debug("VFP: raising exceptions %08x\n", exceptions);
257 if (exceptions == VFP_EXCEPTION_ERROR) {
258 vfp_panic("unhandled bounce", inst);
259 vfp_raise_sigfpe(0, regs);
264 * If any of the status flags are set, update the FPSCR.
265 * Comparison instructions always return at least one of
268 if (exceptions & (FPSCR_N|FPSCR_Z|FPSCR_C|FPSCR_V))
269 fpscr &= ~(FPSCR_N|FPSCR_Z|FPSCR_C|FPSCR_V);
275 #define RAISE(stat,en,sig) \
276 if (exceptions & stat && fpscr & en) \
280 * These are arranged in priority order, least to highest.
282 RAISE(FPSCR_DZC, FPSCR_DZE, FPE_FLTDIV);
283 RAISE(FPSCR_IXC, FPSCR_IXE, FPE_FLTRES);
284 RAISE(FPSCR_UFC, FPSCR_UFE, FPE_FLTUND);
285 RAISE(FPSCR_OFC, FPSCR_OFE, FPE_FLTOVF);
286 RAISE(FPSCR_IOC, FPSCR_IOE, FPE_FLTINV);
289 vfp_raise_sigfpe(si_code, regs);
293 * Emulate a VFP instruction.
295 static u32 vfp_emulate_instruction(u32 inst, u32 fpscr, struct pt_regs *regs)
297 u32 exceptions = VFP_EXCEPTION_ERROR;
299 pr_debug("VFP: emulate: INST=0x%08x SCR=0x%08x\n", inst, fpscr);
301 if (INST_CPRTDO(inst)) {
302 if (!INST_CPRT(inst)) {
306 if (vfp_single(inst)) {
307 exceptions = vfp_single_cpdo(inst, fpscr);
309 exceptions = vfp_double_cpdo(inst, fpscr);
313 * A CPRT instruction can not appear in FPINST2, nor
314 * can it cause an exception. Therefore, we do not
315 * have to emulate it.
320 * A CPDT instruction can not appear in FPINST2, nor can
321 * it cause an exception. Therefore, we do not have to
325 return exceptions & ~VFP_NAN_FLAG;
329 * Package up a bounce condition.
331 void VFP_bounce(u32 trigger, u32 fpexc, struct pt_regs *regs)
333 u32 fpscr, orig_fpscr, fpsid, exceptions;
335 pr_debug("VFP: bounce: trigger %08x fpexc %08x\n", trigger, fpexc);
338 * At this point, FPEXC can have the following configuration:
341 * 0 1 x - synchronous exception
342 * 1 x 0 - asynchronous exception
343 * 1 x 1 - sychronous on VFP subarch 1 and asynchronous on later
344 * 0 0 1 - synchronous on VFP9 (non-standard subarch 1
345 * implementation), undefined otherwise
347 * Clear various bits and enable access to the VFP so we can
350 fmxr(FPEXC, fpexc & ~(FPEXC_EX|FPEXC_DEX|FPEXC_FP2V|FPEXC_VV|FPEXC_TRAP_MASK));
353 orig_fpscr = fpscr = fmrx(FPSCR);
356 * Check for the special VFP subarch 1 and FPSCR.IXE bit case
358 if ((fpsid & FPSID_ARCH_MASK) == (1 << FPSID_ARCH_BIT)
359 && (fpscr & FPSCR_IXE)) {
361 * Synchronous exception, emulate the trigger instruction
366 if (fpexc & FPEXC_EX) {
367 #ifndef CONFIG_CPU_FEROCEON
369 * Asynchronous exception. The instruction is read from FPINST
370 * and the interrupted instruction has to be restarted.
372 trigger = fmrx(FPINST);
375 } else if (!(fpexc & FPEXC_DEX)) {
377 * Illegal combination of bits. It can be caused by an
378 * unallocated VFP instruction but with FPSCR.IXE set and not
381 vfp_raise_exceptions(VFP_EXCEPTION_ERROR, trigger, fpscr, regs);
386 * Modify fpscr to indicate the number of iterations remaining.
387 * If FPEXC.EX is 0, FPEXC.DEX is 1 and the FPEXC.VV bit indicates
388 * whether FPEXC.VECITR or FPSCR.LEN is used.
390 if (fpexc & (FPEXC_EX | FPEXC_VV)) {
393 len = fpexc + (1 << FPEXC_LENGTH_BIT);
395 fpscr &= ~FPSCR_LENGTH_MASK;
396 fpscr |= (len & FPEXC_LENGTH_MASK) << (FPSCR_LENGTH_BIT - FPEXC_LENGTH_BIT);
400 * Handle the first FP instruction. We used to take note of the
401 * FPEXC bounce reason, but this appears to be unreliable.
402 * Emulate the bounced instruction instead.
404 exceptions = vfp_emulate_instruction(trigger, fpscr, regs);
406 vfp_raise_exceptions(exceptions, trigger, orig_fpscr, regs);
409 * If there isn't a second FP instruction, exit now. Note that
410 * the FPEXC.FP2V bit is valid only if FPEXC.EX is 1.
412 if (fpexc ^ (FPEXC_EX | FPEXC_FP2V))
416 * The barrier() here prevents fpinst2 being read
417 * before the condition above.
420 trigger = fmrx(FPINST2);
423 exceptions = vfp_emulate_instruction(trigger, orig_fpscr, regs);
425 vfp_raise_exceptions(exceptions, trigger, orig_fpscr, regs);
430 static void vfp_enable(void *unused)
432 u32 access = get_copro_access();
435 * Enable full access to VFP (cp10 and cp11)
437 set_copro_access(access | CPACC_FULL(10) | CPACC_FULL(11));
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));
461 static void vfp_pm_resume(void)
463 /* ensure we have access to the vfp */
466 /* and disable it to ensure the next usage restores the state */
467 fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
470 static int vfp_cpu_pm_notifier(struct notifier_block *self, unsigned long cmd,
477 case CPU_PM_ENTER_FAILED:
485 static struct notifier_block vfp_cpu_pm_notifier_block = {
486 .notifier_call = vfp_cpu_pm_notifier,
489 static void vfp_pm_init(void)
491 cpu_pm_register_notifier(&vfp_cpu_pm_notifier_block);
495 static inline void vfp_pm_init(void) { }
496 #endif /* CONFIG_CPU_PM */
499 * Ensure that the VFP state stored in 'thread->vfpstate' is up to date
500 * with the hardware state.
502 void vfp_sync_hwstate(struct thread_info *thread)
504 unsigned int cpu = get_cpu();
506 if (vfp_state_in_hw(cpu, thread)) {
507 u32 fpexc = fmrx(FPEXC);
510 * Save the last VFP state on this CPU.
512 fmxr(FPEXC, fpexc | FPEXC_EN);
513 vfp_save_state(&thread->vfpstate, fpexc | FPEXC_EN);
520 /* Ensure that the thread reloads the hardware VFP state on the next use. */
521 void vfp_flush_hwstate(struct thread_info *thread)
523 unsigned int cpu = get_cpu();
525 vfp_force_reload(cpu, thread);
531 * VFP hardware can lose all context when a CPU goes offline.
532 * As we will be running in SMP mode with CPU hotplug, we will save the
533 * hardware state at every thread switch. We clear our held state when
534 * a CPU has been killed, indicating that the VFP hardware doesn't contain
535 * a threads VFP state. When a CPU starts up, we re-enable access to the
538 * Both CPU_DYING and CPU_STARTING are called on the CPU which
539 * is being offlined/onlined.
541 static int vfp_hotplug(struct notifier_block *b, unsigned long action,
544 if (action == CPU_DYING || action == CPU_DYING_FROZEN) {
545 vfp_force_reload((long)hcpu, current_thread_info());
546 } else if (action == CPU_STARTING || action == CPU_STARTING_FROZEN)
552 * VFP support code initialisation.
554 static int __init vfp_init(void)
557 unsigned int cpu_arch = cpu_architecture();
559 if (cpu_arch >= CPU_ARCH_ARMv6)
563 * First check that there is a VFP that we can use.
564 * The handler is already setup to just log calls, so
565 * we just need to read the VFPSID register.
567 vfp_vector = vfp_testing_entry;
569 vfpsid = fmrx(FPSID);
571 vfp_vector = vfp_null_entry;
573 printk(KERN_INFO "VFP support v0.3: ");
575 printk("not present\n");
576 else if (vfpsid & FPSID_NODOUBLE) {
577 printk("no double precision support\n");
579 hotcpu_notifier(vfp_hotplug, 0);
581 smp_call_function(vfp_enable, NULL, 1);
583 VFP_arch = (vfpsid & FPSID_ARCH_MASK) >> FPSID_ARCH_BIT; /* Extract the architecture version */
584 printk("implementor %02x architecture %d part %02x variant %x rev %x\n",
585 (vfpsid & FPSID_IMPLEMENTER_MASK) >> FPSID_IMPLEMENTER_BIT,
586 (vfpsid & FPSID_ARCH_MASK) >> FPSID_ARCH_BIT,
587 (vfpsid & FPSID_PART_MASK) >> FPSID_PART_BIT,
588 (vfpsid & FPSID_VARIANT_MASK) >> FPSID_VARIANT_BIT,
589 (vfpsid & FPSID_REV_MASK) >> FPSID_REV_BIT);
591 vfp_vector = vfp_support_entry;
593 thread_register_notifier(&vfp_notifier_block);
597 * We detected VFP, and the support code is
598 * in place; report VFP support to userspace.
600 elf_hwcap |= HWCAP_VFP;
603 elf_hwcap |= HWCAP_VFPv3;
606 * Check for VFPv3 D16. CPUs in this configuration
607 * only have 16 x 64bit registers.
609 if (((fmrx(MVFR0) & MVFR0_A_SIMD_MASK)) == 1)
610 elf_hwcap |= HWCAP_VFPv3D16;
614 * Check for the presence of the Advanced SIMD
615 * load/store instructions, integer and single
616 * precision floating point operations. Only check
617 * for NEON if the hardware has the MVFR registers.
619 if ((read_cpuid_id() & 0x000f0000) == 0x000f0000) {
621 if ((fmrx(MVFR1) & 0x000fff00) == 0x00011100)
622 elf_hwcap |= HWCAP_NEON;
624 if ((fmrx(MVFR1) & 0xf0000000) == 0x10000000)
625 elf_hwcap |= HWCAP_VFPv4;
631 late_initcall(vfp_init);