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/cpu_pm.h>
14 #include <linux/kernel.h>
15 #include <linux/notifier.h>
16 #include <linux/signal.h>
17 #include <linux/sched.h>
18 #include <linux/smp.h>
19 #include <linux/init.h>
22 #include <asm/cputype.h>
23 #include <asm/system_info.h>
24 #include <asm/thread_notify.h>
31 * Our undef handlers (in entry.S)
33 void vfp_testing_entry(void);
34 void vfp_support_entry(void);
35 void vfp_null_entry(void);
37 void (*vfp_vector)(void) = vfp_null_entry;
41 * Used in startup: set to non-zero if VFP checks fail
42 * After startup, holds VFP architecture
44 unsigned int VFP_arch;
47 * The pointer to the vfpstate structure of the thread which currently
48 * owns the context held in the VFP hardware, or NULL if the hardware
51 * For UP, this is sufficient to tell which thread owns the VFP context.
52 * However, for SMP, we also need to check the CPU number stored in the
53 * saved state too to catch migrations.
55 union vfp_state *vfp_current_hw_state[NR_CPUS];
58 * Is 'thread's most up to date state stored in this CPUs hardware?
59 * Must be called from non-preemptible context.
61 static bool vfp_state_in_hw(unsigned int cpu, struct thread_info *thread)
64 if (thread->vfpstate.hard.cpu != cpu)
67 return vfp_current_hw_state[cpu] == &thread->vfpstate;
71 * Force a reload of the VFP context from the thread structure. We do
72 * this by ensuring that access to the VFP hardware is disabled, and
73 * clear vfp_current_hw_state. Must be called from non-preemptible context.
75 static void vfp_force_reload(unsigned int cpu, struct thread_info *thread)
77 if (vfp_state_in_hw(cpu, thread)) {
78 fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
79 vfp_current_hw_state[cpu] = NULL;
82 thread->vfpstate.hard.cpu = NR_CPUS;
87 * Per-thread VFP initialization.
89 static void vfp_thread_flush(struct thread_info *thread)
91 union vfp_state *vfp = &thread->vfpstate;
95 * Disable VFP to ensure we initialize it first. We must ensure
96 * that the modification of vfp_current_hw_state[] and hardware
97 * disable are done for the same CPU and without preemption.
99 * Do this first to ensure that preemption won't overwrite our
100 * state saving should access to the VFP be enabled at this point.
103 if (vfp_current_hw_state[cpu] == vfp)
104 vfp_current_hw_state[cpu] = NULL;
105 fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
108 memset(vfp, 0, sizeof(union vfp_state));
110 vfp->hard.fpexc = FPEXC_EN;
111 vfp->hard.fpscr = FPSCR_ROUND_NEAREST;
113 vfp->hard.cpu = NR_CPUS;
117 static void vfp_thread_exit(struct thread_info *thread)
119 /* release case: Per-thread VFP cleanup. */
120 union vfp_state *vfp = &thread->vfpstate;
121 unsigned int cpu = get_cpu();
123 if (vfp_current_hw_state[cpu] == vfp)
124 vfp_current_hw_state[cpu] = NULL;
128 static void vfp_thread_copy(struct thread_info *thread)
130 struct thread_info *parent = current_thread_info();
132 vfp_sync_hwstate(parent);
133 thread->vfpstate = parent->vfpstate;
135 thread->vfpstate.hard.cpu = NR_CPUS;
140 * When this function is called with the following 'cmd's, the following
141 * is true while this function is being run:
142 * THREAD_NOFTIFY_SWTICH:
143 * - the previously running thread will not be scheduled onto another CPU.
144 * - the next thread to be run (v) will not be running on another CPU.
145 * - thread->cpu is the local CPU number
146 * - not preemptible as we're called in the middle of a thread switch
147 * THREAD_NOTIFY_FLUSH:
148 * - the thread (v) will be running on the local CPU, so
149 * v === current_thread_info()
150 * - thread->cpu is the local CPU number at the time it is accessed,
151 * but may change at any time.
152 * - we could be preempted if tree preempt rcu is enabled, so
153 * it is unsafe to use thread->cpu.
155 * - the thread (v) will be running on the local CPU, so
156 * v === current_thread_info()
157 * - thread->cpu is the local CPU number at the time it is accessed,
158 * but may change at any time.
159 * - we could be preempted if tree preempt rcu is enabled, so
160 * it is unsafe to use thread->cpu.
162 static int vfp_notifier(struct notifier_block *self, unsigned long cmd, void *v)
164 struct thread_info *thread = v;
171 case THREAD_NOTIFY_SWITCH:
178 * On SMP, if VFP is enabled, save the old state in
179 * case the thread migrates to a different CPU. The
180 * restoring is done lazily.
182 if ((fpexc & FPEXC_EN) && vfp_current_hw_state[cpu])
183 vfp_save_state(vfp_current_hw_state[cpu], fpexc);
187 * Always disable VFP so we can lazily save/restore the
190 fmxr(FPEXC, fpexc & ~FPEXC_EN);
193 case THREAD_NOTIFY_FLUSH:
194 vfp_thread_flush(thread);
197 case THREAD_NOTIFY_EXIT:
198 vfp_thread_exit(thread);
201 case THREAD_NOTIFY_COPY:
202 vfp_thread_copy(thread);
209 static struct notifier_block vfp_notifier_block = {
210 .notifier_call = vfp_notifier,
214 * Raise a SIGFPE for the current process.
215 * sicode describes the signal being raised.
217 static void vfp_raise_sigfpe(unsigned int sicode, struct pt_regs *regs)
221 memset(&info, 0, sizeof(info));
223 info.si_signo = SIGFPE;
224 info.si_code = sicode;
225 info.si_addr = (void __user *)(instruction_pointer(regs) - 4);
228 * This is the same as NWFPE, because it's not clear what
231 current->thread.error_code = 0;
232 current->thread.trap_no = 6;
234 send_sig_info(SIGFPE, &info, current);
237 static void vfp_panic(char *reason, u32 inst)
241 printk(KERN_ERR "VFP: Error: %s\n", reason);
242 printk(KERN_ERR "VFP: EXC 0x%08x SCR 0x%08x INST 0x%08x\n",
243 fmrx(FPEXC), fmrx(FPSCR), inst);
244 for (i = 0; i < 32; i += 2)
245 printk(KERN_ERR "VFP: s%2u: 0x%08x s%2u: 0x%08x\n",
246 i, vfp_get_float(i), i+1, vfp_get_float(i+1));
250 * Process bitmask of exception conditions.
252 static void vfp_raise_exceptions(u32 exceptions, u32 inst, u32 fpscr, struct pt_regs *regs)
256 pr_debug("VFP: raising exceptions %08x\n", exceptions);
258 if (exceptions == VFP_EXCEPTION_ERROR) {
259 vfp_panic("unhandled bounce", inst);
260 vfp_raise_sigfpe(0, regs);
265 * If any of the status flags are set, update the FPSCR.
266 * Comparison instructions always return at least one of
269 if (exceptions & (FPSCR_N|FPSCR_Z|FPSCR_C|FPSCR_V))
270 fpscr &= ~(FPSCR_N|FPSCR_Z|FPSCR_C|FPSCR_V);
276 #define RAISE(stat,en,sig) \
277 if (exceptions & stat && fpscr & en) \
281 * These are arranged in priority order, least to highest.
283 RAISE(FPSCR_DZC, FPSCR_DZE, FPE_FLTDIV);
284 RAISE(FPSCR_IXC, FPSCR_IXE, FPE_FLTRES);
285 RAISE(FPSCR_UFC, FPSCR_UFE, FPE_FLTUND);
286 RAISE(FPSCR_OFC, FPSCR_OFE, FPE_FLTOVF);
287 RAISE(FPSCR_IOC, FPSCR_IOE, FPE_FLTINV);
290 vfp_raise_sigfpe(si_code, regs);
294 * Emulate a VFP instruction.
296 static u32 vfp_emulate_instruction(u32 inst, u32 fpscr, struct pt_regs *regs)
298 u32 exceptions = VFP_EXCEPTION_ERROR;
300 pr_debug("VFP: emulate: INST=0x%08x SCR=0x%08x\n", inst, fpscr);
302 if (INST_CPRTDO(inst)) {
303 if (!INST_CPRT(inst)) {
307 if (vfp_single(inst)) {
308 exceptions = vfp_single_cpdo(inst, fpscr);
310 exceptions = vfp_double_cpdo(inst, fpscr);
314 * A CPRT instruction can not appear in FPINST2, nor
315 * can it cause an exception. Therefore, we do not
316 * have to emulate it.
321 * A CPDT instruction can not appear in FPINST2, nor can
322 * it cause an exception. Therefore, we do not have to
326 return exceptions & ~VFP_NAN_FLAG;
330 * Package up a bounce condition.
332 void VFP_bounce(u32 trigger, u32 fpexc, struct pt_regs *regs)
334 u32 fpscr, orig_fpscr, fpsid, exceptions;
336 pr_debug("VFP: bounce: trigger %08x fpexc %08x\n", trigger, fpexc);
339 * At this point, FPEXC can have the following configuration:
342 * 0 1 x - synchronous exception
343 * 1 x 0 - asynchronous exception
344 * 1 x 1 - sychronous on VFP subarch 1 and asynchronous on later
345 * 0 0 1 - synchronous on VFP9 (non-standard subarch 1
346 * implementation), undefined otherwise
348 * Clear various bits and enable access to the VFP so we can
351 fmxr(FPEXC, fpexc & ~(FPEXC_EX|FPEXC_DEX|FPEXC_FP2V|FPEXC_VV|FPEXC_TRAP_MASK));
354 orig_fpscr = fpscr = fmrx(FPSCR);
357 * Check for the special VFP subarch 1 and FPSCR.IXE bit case
359 if ((fpsid & FPSID_ARCH_MASK) == (1 << FPSID_ARCH_BIT)
360 && (fpscr & FPSCR_IXE)) {
362 * Synchronous exception, emulate the trigger instruction
367 if (fpexc & FPEXC_EX) {
368 #ifndef CONFIG_CPU_FEROCEON
370 * Asynchronous exception. The instruction is read from FPINST
371 * and the interrupted instruction has to be restarted.
373 trigger = fmrx(FPINST);
376 } else if (!(fpexc & FPEXC_DEX)) {
378 * Illegal combination of bits. It can be caused by an
379 * unallocated VFP instruction but with FPSCR.IXE set and not
382 vfp_raise_exceptions(VFP_EXCEPTION_ERROR, trigger, fpscr, regs);
387 * Modify fpscr to indicate the number of iterations remaining.
388 * If FPEXC.EX is 0, FPEXC.DEX is 1 and the FPEXC.VV bit indicates
389 * whether FPEXC.VECITR or FPSCR.LEN is used.
391 if (fpexc & (FPEXC_EX | FPEXC_VV)) {
394 len = fpexc + (1 << FPEXC_LENGTH_BIT);
396 fpscr &= ~FPSCR_LENGTH_MASK;
397 fpscr |= (len & FPEXC_LENGTH_MASK) << (FPSCR_LENGTH_BIT - FPEXC_LENGTH_BIT);
401 * Handle the first FP instruction. We used to take note of the
402 * FPEXC bounce reason, but this appears to be unreliable.
403 * Emulate the bounced instruction instead.
405 exceptions = vfp_emulate_instruction(trigger, fpscr, regs);
407 vfp_raise_exceptions(exceptions, trigger, orig_fpscr, regs);
410 * If there isn't a second FP instruction, exit now. Note that
411 * the FPEXC.FP2V bit is valid only if FPEXC.EX is 1.
413 if (fpexc ^ (FPEXC_EX | FPEXC_FP2V))
417 * The barrier() here prevents fpinst2 being read
418 * before the condition above.
421 trigger = fmrx(FPINST2);
424 exceptions = vfp_emulate_instruction(trigger, orig_fpscr, regs);
426 vfp_raise_exceptions(exceptions, trigger, orig_fpscr, regs);
431 static void vfp_enable(void *unused)
433 u32 access = get_copro_access();
436 * Enable full access to VFP (cp10 and cp11)
438 set_copro_access(access | CPACC_FULL(10) | CPACC_FULL(11));
442 static int vfp_pm_suspend(void)
444 struct thread_info *ti = current_thread_info();
445 u32 fpexc = fmrx(FPEXC);
447 /* if vfp is on, then save state for resumption */
448 if (fpexc & FPEXC_EN) {
449 printk(KERN_DEBUG "%s: saving vfp state\n", __func__);
450 vfp_save_state(&ti->vfpstate, fpexc);
452 /* disable, just in case */
453 fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
456 /* clear any information we had about last context state */
457 memset(vfp_current_hw_state, 0, sizeof(vfp_current_hw_state));
462 static void vfp_pm_resume(void)
464 /* ensure we have access to the vfp */
467 /* and disable it to ensure the next usage restores the state */
468 fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
471 static int vfp_cpu_pm_notifier(struct notifier_block *self, unsigned long cmd,
478 case CPU_PM_ENTER_FAILED:
486 static struct notifier_block vfp_cpu_pm_notifier_block = {
487 .notifier_call = vfp_cpu_pm_notifier,
490 static void vfp_pm_init(void)
492 cpu_pm_register_notifier(&vfp_cpu_pm_notifier_block);
496 static inline void vfp_pm_init(void) { }
497 #endif /* CONFIG_CPU_PM */
500 * Ensure that the VFP state stored in 'thread->vfpstate' is up to date
501 * with the hardware state.
503 void vfp_sync_hwstate(struct thread_info *thread)
505 unsigned int cpu = get_cpu();
507 if (vfp_state_in_hw(cpu, thread)) {
508 u32 fpexc = fmrx(FPEXC);
511 * Save the last VFP state on this CPU.
513 fmxr(FPEXC, fpexc | FPEXC_EN);
514 vfp_save_state(&thread->vfpstate, fpexc | FPEXC_EN);
521 /* Ensure that the thread reloads the hardware VFP state on the next use. */
522 void vfp_flush_hwstate(struct thread_info *thread)
524 unsigned int cpu = get_cpu();
526 vfp_force_reload(cpu, thread);
532 * VFP hardware can lose all context when a CPU goes offline.
533 * As we will be running in SMP mode with CPU hotplug, we will save the
534 * hardware state at every thread switch. We clear our held state when
535 * a CPU has been killed, indicating that the VFP hardware doesn't contain
536 * a threads VFP state. When a CPU starts up, we re-enable access to the
539 * Both CPU_DYING and CPU_STARTING are called on the CPU which
540 * is being offlined/onlined.
542 static int vfp_hotplug(struct notifier_block *b, unsigned long action,
545 if (action == CPU_DYING || action == CPU_DYING_FROZEN) {
546 vfp_force_reload((long)hcpu, current_thread_info());
547 } else if (action == CPU_STARTING || action == CPU_STARTING_FROZEN)
553 * VFP support code initialisation.
555 static int __init vfp_init(void)
558 unsigned int cpu_arch = cpu_architecture();
560 if (cpu_arch >= CPU_ARCH_ARMv6)
564 * First check that there is a VFP that we can use.
565 * The handler is already setup to just log calls, so
566 * we just need to read the VFPSID register.
568 vfp_vector = vfp_testing_entry;
570 vfpsid = fmrx(FPSID);
572 vfp_vector = vfp_null_entry;
574 printk(KERN_INFO "VFP support v0.3: ");
576 printk("not present\n");
577 else if (vfpsid & FPSID_NODOUBLE) {
578 printk("no double precision support\n");
580 hotcpu_notifier(vfp_hotplug, 0);
582 smp_call_function(vfp_enable, NULL, 1);
584 VFP_arch = (vfpsid & FPSID_ARCH_MASK) >> FPSID_ARCH_BIT; /* Extract the architecture version */
585 printk("implementor %02x architecture %d part %02x variant %x rev %x\n",
586 (vfpsid & FPSID_IMPLEMENTER_MASK) >> FPSID_IMPLEMENTER_BIT,
587 (vfpsid & FPSID_ARCH_MASK) >> FPSID_ARCH_BIT,
588 (vfpsid & FPSID_PART_MASK) >> FPSID_PART_BIT,
589 (vfpsid & FPSID_VARIANT_MASK) >> FPSID_VARIANT_BIT,
590 (vfpsid & FPSID_REV_MASK) >> FPSID_REV_BIT);
592 vfp_vector = vfp_support_entry;
594 thread_register_notifier(&vfp_notifier_block);
598 * We detected VFP, and the support code is
599 * in place; report VFP support to userspace.
601 elf_hwcap |= HWCAP_VFP;
604 elf_hwcap |= HWCAP_VFPv3;
607 * Check for VFPv3 D16. CPUs in this configuration
608 * only have 16 x 64bit registers.
610 if (((fmrx(MVFR0) & MVFR0_A_SIMD_MASK)) == 1)
611 elf_hwcap |= HWCAP_VFPv3D16;
615 * Check for the presence of the Advanced SIMD
616 * load/store instructions, integer and single
617 * precision floating point operations. Only check
618 * for NEON if the hardware has the MVFR registers.
620 if ((read_cpuid_id() & 0x000f0000) == 0x000f0000) {
622 if ((fmrx(MVFR1) & 0x000fff00) == 0x00011100)
623 elf_hwcap |= HWCAP_NEON;
625 if ((fmrx(MVFR1) & 0xf0000000) == 0x10000000)
626 elf_hwcap |= HWCAP_VFPv4;
632 late_initcall(vfp_init);