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/hardirq.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>
21 #include <linux/uaccess.h>
22 #include <linux/user.h>
23 #include <linux/export.h>
26 #include <asm/cputype.h>
27 #include <asm/system_info.h>
28 #include <asm/thread_notify.h>
35 * Our undef handlers (in entry.S)
37 void vfp_testing_entry(void);
38 void vfp_support_entry(void);
39 void vfp_null_entry(void);
41 void (*vfp_vector)(void) = vfp_null_entry;
45 * Used in startup: set to non-zero if VFP checks fail
46 * After startup, holds VFP architecture
48 unsigned int VFP_arch;
51 * The pointer to the vfpstate structure of the thread which currently
52 * owns the context held in the VFP hardware, or NULL if the hardware
55 * For UP, this is sufficient to tell which thread owns the VFP context.
56 * However, for SMP, we also need to check the CPU number stored in the
57 * saved state too to catch migrations.
59 union vfp_state *vfp_current_hw_state[NR_CPUS];
62 * Is 'thread's most up to date state stored in this CPUs hardware?
63 * Must be called from non-preemptible context.
65 static bool vfp_state_in_hw(unsigned int cpu, struct thread_info *thread)
68 if (thread->vfpstate.hard.cpu != cpu)
71 return vfp_current_hw_state[cpu] == &thread->vfpstate;
75 * Force a reload of the VFP context from the thread structure. We do
76 * this by ensuring that access to the VFP hardware is disabled, and
77 * clear vfp_current_hw_state. Must be called from non-preemptible context.
79 static void vfp_force_reload(unsigned int cpu, struct thread_info *thread)
81 if (vfp_state_in_hw(cpu, thread)) {
82 fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
83 vfp_current_hw_state[cpu] = NULL;
86 thread->vfpstate.hard.cpu = NR_CPUS;
91 * Per-thread VFP initialization.
93 static void vfp_thread_flush(struct thread_info *thread)
95 union vfp_state *vfp = &thread->vfpstate;
99 * Disable VFP to ensure we initialize it first. We must ensure
100 * that the modification of vfp_current_hw_state[] and hardware
101 * disable are done for the same CPU and without preemption.
103 * Do this first to ensure that preemption won't overwrite our
104 * state saving should access to the VFP be enabled at this point.
107 if (vfp_current_hw_state[cpu] == vfp)
108 vfp_current_hw_state[cpu] = NULL;
109 fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
112 memset(vfp, 0, sizeof(union vfp_state));
114 vfp->hard.fpexc = FPEXC_EN;
115 vfp->hard.fpscr = FPSCR_ROUND_NEAREST;
117 vfp->hard.cpu = NR_CPUS;
121 static void vfp_thread_exit(struct thread_info *thread)
123 /* release case: Per-thread VFP cleanup. */
124 union vfp_state *vfp = &thread->vfpstate;
125 unsigned int cpu = get_cpu();
127 if (vfp_current_hw_state[cpu] == vfp)
128 vfp_current_hw_state[cpu] = NULL;
132 static void vfp_thread_copy(struct thread_info *thread)
134 struct thread_info *parent = current_thread_info();
136 vfp_sync_hwstate(parent);
137 thread->vfpstate = parent->vfpstate;
139 thread->vfpstate.hard.cpu = NR_CPUS;
144 * When this function is called with the following 'cmd's, the following
145 * is true while this function is being run:
146 * THREAD_NOFTIFY_SWTICH:
147 * - the previously running thread will not be scheduled onto another CPU.
148 * - the next thread to be run (v) will not be running on another CPU.
149 * - thread->cpu is the local CPU number
150 * - not preemptible as we're called in the middle of a thread switch
151 * THREAD_NOTIFY_FLUSH:
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 * - the thread (v) will be running on the local CPU, so
160 * v === current_thread_info()
161 * - thread->cpu is the local CPU number at the time it is accessed,
162 * but may change at any time.
163 * - we could be preempted if tree preempt rcu is enabled, so
164 * it is unsafe to use thread->cpu.
166 static int vfp_notifier(struct notifier_block *self, unsigned long cmd, void *v)
168 struct thread_info *thread = v;
175 case THREAD_NOTIFY_SWITCH:
182 * On SMP, if VFP is enabled, save the old state in
183 * case the thread migrates to a different CPU. The
184 * restoring is done lazily.
186 if ((fpexc & FPEXC_EN) && vfp_current_hw_state[cpu])
187 vfp_save_state(vfp_current_hw_state[cpu], fpexc);
191 * Always disable VFP so we can lazily save/restore the
194 fmxr(FPEXC, fpexc & ~FPEXC_EN);
197 case THREAD_NOTIFY_FLUSH:
198 vfp_thread_flush(thread);
201 case THREAD_NOTIFY_EXIT:
202 vfp_thread_exit(thread);
205 case THREAD_NOTIFY_COPY:
206 vfp_thread_copy(thread);
213 static struct notifier_block vfp_notifier_block = {
214 .notifier_call = vfp_notifier,
218 * Raise a SIGFPE for the current process.
219 * sicode describes the signal being raised.
221 static void vfp_raise_sigfpe(unsigned int sicode, struct pt_regs *regs)
225 memset(&info, 0, sizeof(info));
227 info.si_signo = SIGFPE;
228 info.si_code = sicode;
229 info.si_addr = (void __user *)(instruction_pointer(regs) - 4);
232 * This is the same as NWFPE, because it's not clear what
235 current->thread.error_code = 0;
236 current->thread.trap_no = 6;
238 send_sig_info(SIGFPE, &info, current);
241 static void vfp_panic(char *reason, u32 inst)
245 pr_err("VFP: Error: %s\n", reason);
246 pr_err("VFP: EXC 0x%08x SCR 0x%08x INST 0x%08x\n",
247 fmrx(FPEXC), fmrx(FPSCR), inst);
248 for (i = 0; i < 32; i += 2)
249 pr_err("VFP: s%2u: 0x%08x s%2u: 0x%08x\n",
250 i, vfp_get_float(i), i+1, vfp_get_float(i+1));
254 * Process bitmask of exception conditions.
256 static void vfp_raise_exceptions(u32 exceptions, u32 inst, u32 fpscr, struct pt_regs *regs)
260 pr_debug("VFP: raising exceptions %08x\n", exceptions);
262 if (exceptions == VFP_EXCEPTION_ERROR) {
263 vfp_panic("unhandled bounce", inst);
264 vfp_raise_sigfpe(0, regs);
269 * If any of the status flags are set, update the FPSCR.
270 * Comparison instructions always return at least one of
273 if (exceptions & (FPSCR_N|FPSCR_Z|FPSCR_C|FPSCR_V))
274 fpscr &= ~(FPSCR_N|FPSCR_Z|FPSCR_C|FPSCR_V);
280 #define RAISE(stat,en,sig) \
281 if (exceptions & stat && fpscr & en) \
285 * These are arranged in priority order, least to highest.
287 RAISE(FPSCR_DZC, FPSCR_DZE, FPE_FLTDIV);
288 RAISE(FPSCR_IXC, FPSCR_IXE, FPE_FLTRES);
289 RAISE(FPSCR_UFC, FPSCR_UFE, FPE_FLTUND);
290 RAISE(FPSCR_OFC, FPSCR_OFE, FPE_FLTOVF);
291 RAISE(FPSCR_IOC, FPSCR_IOE, FPE_FLTINV);
294 vfp_raise_sigfpe(si_code, regs);
298 * Emulate a VFP instruction.
300 static u32 vfp_emulate_instruction(u32 inst, u32 fpscr, struct pt_regs *regs)
302 u32 exceptions = VFP_EXCEPTION_ERROR;
304 pr_debug("VFP: emulate: INST=0x%08x SCR=0x%08x\n", inst, fpscr);
306 if (INST_CPRTDO(inst)) {
307 if (!INST_CPRT(inst)) {
311 if (vfp_single(inst)) {
312 exceptions = vfp_single_cpdo(inst, fpscr);
314 exceptions = vfp_double_cpdo(inst, fpscr);
318 * A CPRT instruction can not appear in FPINST2, nor
319 * can it cause an exception. Therefore, we do not
320 * have to emulate it.
325 * A CPDT instruction can not appear in FPINST2, nor can
326 * it cause an exception. Therefore, we do not have to
330 return exceptions & ~VFP_NAN_FLAG;
334 * Package up a bounce condition.
336 void VFP_bounce(u32 trigger, u32 fpexc, struct pt_regs *regs)
338 u32 fpscr, orig_fpscr, fpsid, exceptions;
340 pr_debug("VFP: bounce: trigger %08x fpexc %08x\n", trigger, fpexc);
343 * At this point, FPEXC can have the following configuration:
346 * 0 1 x - synchronous exception
347 * 1 x 0 - asynchronous exception
348 * 1 x 1 - sychronous on VFP subarch 1 and asynchronous on later
349 * 0 0 1 - synchronous on VFP9 (non-standard subarch 1
350 * implementation), undefined otherwise
352 * Clear various bits and enable access to the VFP so we can
355 fmxr(FPEXC, fpexc & ~(FPEXC_EX|FPEXC_DEX|FPEXC_FP2V|FPEXC_VV|FPEXC_TRAP_MASK));
358 orig_fpscr = fpscr = fmrx(FPSCR);
361 * Check for the special VFP subarch 1 and FPSCR.IXE bit case
363 if ((fpsid & FPSID_ARCH_MASK) == (1 << FPSID_ARCH_BIT)
364 && (fpscr & FPSCR_IXE)) {
366 * Synchronous exception, emulate the trigger instruction
371 if (fpexc & FPEXC_EX) {
372 #ifndef CONFIG_CPU_FEROCEON
374 * Asynchronous exception. The instruction is read from FPINST
375 * and the interrupted instruction has to be restarted.
377 trigger = fmrx(FPINST);
380 } else if (!(fpexc & FPEXC_DEX)) {
382 * Illegal combination of bits. It can be caused by an
383 * unallocated VFP instruction but with FPSCR.IXE set and not
386 vfp_raise_exceptions(VFP_EXCEPTION_ERROR, trigger, fpscr, regs);
391 * Modify fpscr to indicate the number of iterations remaining.
392 * If FPEXC.EX is 0, FPEXC.DEX is 1 and the FPEXC.VV bit indicates
393 * whether FPEXC.VECITR or FPSCR.LEN is used.
395 if (fpexc & (FPEXC_EX | FPEXC_VV)) {
398 len = fpexc + (1 << FPEXC_LENGTH_BIT);
400 fpscr &= ~FPSCR_LENGTH_MASK;
401 fpscr |= (len & FPEXC_LENGTH_MASK) << (FPSCR_LENGTH_BIT - FPEXC_LENGTH_BIT);
405 * Handle the first FP instruction. We used to take note of the
406 * FPEXC bounce reason, but this appears to be unreliable.
407 * Emulate the bounced instruction instead.
409 exceptions = vfp_emulate_instruction(trigger, fpscr, regs);
411 vfp_raise_exceptions(exceptions, trigger, orig_fpscr, regs);
414 * If there isn't a second FP instruction, exit now. Note that
415 * the FPEXC.FP2V bit is valid only if FPEXC.EX is 1.
417 if ((fpexc & (FPEXC_EX | FPEXC_FP2V)) != (FPEXC_EX | FPEXC_FP2V))
421 * The barrier() here prevents fpinst2 being read
422 * before the condition above.
425 trigger = fmrx(FPINST2);
428 exceptions = vfp_emulate_instruction(trigger, orig_fpscr, regs);
430 vfp_raise_exceptions(exceptions, trigger, orig_fpscr, regs);
435 static void vfp_enable(void *unused)
439 BUG_ON(preemptible());
440 access = get_copro_access();
443 * Enable full access to VFP (cp10 and cp11)
445 set_copro_access(access | CPACC_FULL(10) | CPACC_FULL(11));
449 static int vfp_pm_suspend(void)
451 struct thread_info *ti = current_thread_info();
452 u32 fpexc = fmrx(FPEXC);
454 /* if vfp is on, then save state for resumption */
455 if (fpexc & FPEXC_EN) {
456 pr_debug("%s: saving vfp state\n", __func__);
457 vfp_save_state(&ti->vfpstate, fpexc);
459 /* disable, just in case */
460 fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
461 } else if (vfp_current_hw_state[ti->cpu]) {
463 fmxr(FPEXC, fpexc | FPEXC_EN);
464 vfp_save_state(vfp_current_hw_state[ti->cpu], fpexc);
469 /* clear any information we had about last context state */
470 vfp_current_hw_state[ti->cpu] = NULL;
475 static void vfp_pm_resume(void)
477 /* ensure we have access to the vfp */
480 /* and disable it to ensure the next usage restores the state */
481 fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
484 static int vfp_cpu_pm_notifier(struct notifier_block *self, unsigned long cmd,
491 case CPU_PM_ENTER_FAILED:
499 static struct notifier_block vfp_cpu_pm_notifier_block = {
500 .notifier_call = vfp_cpu_pm_notifier,
503 static void vfp_pm_init(void)
505 cpu_pm_register_notifier(&vfp_cpu_pm_notifier_block);
509 static inline void vfp_pm_init(void) { }
510 #endif /* CONFIG_CPU_PM */
513 * Ensure that the VFP state stored in 'thread->vfpstate' is up to date
514 * with the hardware state.
516 void vfp_sync_hwstate(struct thread_info *thread)
518 unsigned int cpu = get_cpu();
520 if (vfp_state_in_hw(cpu, thread)) {
521 u32 fpexc = fmrx(FPEXC);
524 * Save the last VFP state on this CPU.
526 fmxr(FPEXC, fpexc | FPEXC_EN);
527 vfp_save_state(&thread->vfpstate, fpexc | FPEXC_EN);
534 /* Ensure that the thread reloads the hardware VFP state on the next use. */
535 void vfp_flush_hwstate(struct thread_info *thread)
537 unsigned int cpu = get_cpu();
539 vfp_force_reload(cpu, thread);
545 * Save the current VFP state into the provided structures and prepare
546 * for entry into a new function (signal handler).
548 int vfp_preserve_user_clear_hwstate(struct user_vfp __user *ufp,
549 struct user_vfp_exc __user *ufp_exc)
551 struct thread_info *thread = current_thread_info();
552 struct vfp_hard_struct *hwstate = &thread->vfpstate.hard;
555 /* Ensure that the saved hwstate is up-to-date. */
556 vfp_sync_hwstate(thread);
559 * Copy the floating point registers. There can be unused
560 * registers see asm/hwcap.h for details.
562 err |= __copy_to_user(&ufp->fpregs, &hwstate->fpregs,
563 sizeof(hwstate->fpregs));
565 * Copy the status and control register.
567 __put_user_error(hwstate->fpscr, &ufp->fpscr, err);
570 * Copy the exception registers.
572 __put_user_error(hwstate->fpexc, &ufp_exc->fpexc, err);
573 __put_user_error(hwstate->fpinst, &ufp_exc->fpinst, err);
574 __put_user_error(hwstate->fpinst2, &ufp_exc->fpinst2, err);
579 /* Ensure that VFP is disabled. */
580 vfp_flush_hwstate(thread);
583 * As per the PCS, clear the length and stride bits for function
586 hwstate->fpscr &= ~(FPSCR_LENGTH_MASK | FPSCR_STRIDE_MASK);
590 /* Sanitise and restore the current VFP state from the provided structures. */
591 int vfp_restore_user_hwstate(struct user_vfp __user *ufp,
592 struct user_vfp_exc __user *ufp_exc)
594 struct thread_info *thread = current_thread_info();
595 struct vfp_hard_struct *hwstate = &thread->vfpstate.hard;
599 /* Disable VFP to avoid corrupting the new thread state. */
600 vfp_flush_hwstate(thread);
603 * Copy the floating point registers. There can be unused
604 * registers see asm/hwcap.h for details.
606 err |= __copy_from_user(&hwstate->fpregs, &ufp->fpregs,
607 sizeof(hwstate->fpregs));
609 * Copy the status and control register.
611 __get_user_error(hwstate->fpscr, &ufp->fpscr, err);
614 * Sanitise and restore the exception registers.
616 __get_user_error(fpexc, &ufp_exc->fpexc, err);
618 /* Ensure the VFP is enabled. */
621 /* Ensure FPINST2 is invalid and the exception flag is cleared. */
622 fpexc &= ~(FPEXC_EX | FPEXC_FP2V);
623 hwstate->fpexc = fpexc;
625 __get_user_error(hwstate->fpinst, &ufp_exc->fpinst, err);
626 __get_user_error(hwstate->fpinst2, &ufp_exc->fpinst2, err);
628 return err ? -EFAULT : 0;
632 * VFP hardware can lose all context when a CPU goes offline.
633 * As we will be running in SMP mode with CPU hotplug, we will save the
634 * hardware state at every thread switch. We clear our held state when
635 * a CPU has been killed, indicating that the VFP hardware doesn't contain
636 * a threads VFP state. When a CPU starts up, we re-enable access to the
639 * Both CPU_DYING and CPU_STARTING are called on the CPU which
640 * is being offlined/onlined.
642 static int vfp_hotplug(struct notifier_block *b, unsigned long action,
645 if (action == CPU_DYING || action == CPU_DYING_FROZEN) {
646 vfp_force_reload((long)hcpu, current_thread_info());
647 } else if (action == CPU_STARTING || action == CPU_STARTING_FROZEN)
652 #ifdef CONFIG_KERNEL_MODE_NEON
655 * Kernel-side NEON support functions
657 void kernel_neon_begin(void)
659 struct thread_info *thread = current_thread_info();
664 * Kernel mode NEON is only allowed outside of interrupt context
665 * with preemption disabled. This will make sure that the kernel
666 * mode NEON register contents never need to be preserved.
668 BUG_ON(in_interrupt());
671 fpexc = fmrx(FPEXC) | FPEXC_EN;
675 * Save the userland NEON/VFP state. Under UP,
676 * the owner could be a task other than 'current'
678 if (vfp_state_in_hw(cpu, thread))
679 vfp_save_state(&thread->vfpstate, fpexc);
681 else if (vfp_current_hw_state[cpu] != NULL)
682 vfp_save_state(vfp_current_hw_state[cpu], fpexc);
684 vfp_current_hw_state[cpu] = NULL;
686 EXPORT_SYMBOL(kernel_neon_begin);
688 void kernel_neon_end(void)
690 /* Disable the NEON/VFP unit. */
691 fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
694 EXPORT_SYMBOL(kernel_neon_end);
696 #endif /* CONFIG_KERNEL_MODE_NEON */
699 * VFP support code initialisation.
701 static int __init vfp_init(void)
704 unsigned int cpu_arch = cpu_architecture();
706 if (cpu_arch >= CPU_ARCH_ARMv6)
707 on_each_cpu(vfp_enable, NULL, 1);
710 * First check that there is a VFP that we can use.
711 * The handler is already setup to just log calls, so
712 * we just need to read the VFPSID register.
714 vfp_vector = vfp_testing_entry;
716 vfpsid = fmrx(FPSID);
718 vfp_vector = vfp_null_entry;
720 pr_info("VFP support v0.3: ");
722 pr_cont("not present\n");
723 else if (vfpsid & FPSID_NODOUBLE) {
724 pr_cont("no double precision support\n");
726 hotcpu_notifier(vfp_hotplug, 0);
728 VFP_arch = (vfpsid & FPSID_ARCH_MASK) >> FPSID_ARCH_BIT; /* Extract the architecture version */
729 pr_cont("implementor %02x architecture %d part %02x variant %x rev %x\n",
730 (vfpsid & FPSID_IMPLEMENTER_MASK) >> FPSID_IMPLEMENTER_BIT,
731 (vfpsid & FPSID_ARCH_MASK) >> FPSID_ARCH_BIT,
732 (vfpsid & FPSID_PART_MASK) >> FPSID_PART_BIT,
733 (vfpsid & FPSID_VARIANT_MASK) >> FPSID_VARIANT_BIT,
734 (vfpsid & FPSID_REV_MASK) >> FPSID_REV_BIT);
736 vfp_vector = vfp_support_entry;
738 thread_register_notifier(&vfp_notifier_block);
742 * We detected VFP, and the support code is
743 * in place; report VFP support to userspace.
745 elf_hwcap |= HWCAP_VFP;
748 elf_hwcap |= HWCAP_VFPv3;
751 * Check for VFPv3 D16 and VFPv4 D16. CPUs in
752 * this configuration only have 16 x 64bit
755 if (((fmrx(MVFR0) & MVFR0_A_SIMD_MASK)) == 1)
756 elf_hwcap |= HWCAP_VFPv3D16; /* also v4-D16 */
758 elf_hwcap |= HWCAP_VFPD32;
762 * Check for the presence of the Advanced SIMD
763 * load/store instructions, integer and single
764 * precision floating point operations. Only check
765 * for NEON if the hardware has the MVFR registers.
767 if ((read_cpuid_id() & 0x000f0000) == 0x000f0000) {
769 if ((fmrx(MVFR1) & 0x000fff00) == 0x00011100)
770 elf_hwcap |= HWCAP_NEON;
773 if ((fmrx(MVFR1) & 0xf0000000) == 0x10000000)
774 elf_hwcap |= HWCAP_VFPv4;
781 late_initcall(vfp_init);