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/signal.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 asmlinkage void vfp_testing_entry(void);
38 asmlinkage void vfp_support_entry(void);
39 asmlinkage void vfp_null_entry(void);
41 asmlinkage 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 * - 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 pr_err("VFP: Error: %s\n", reason);
242 pr_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 pr_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(FPE_FIXME, 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)) != (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)
435 BUG_ON(preemptible());
436 access = get_copro_access();
439 * Enable full access to VFP (cp10 and cp11)
441 set_copro_access(access | CPACC_FULL(10) | CPACC_FULL(11));
444 /* Called by platforms on which we want to disable VFP because it may not be
445 * present on all CPUs within a SMP complex. Needs to be called prior to
448 void vfp_disable(void)
451 pr_debug("%s: should be called prior to vfp_init\n", __func__);
458 static int vfp_pm_suspend(void)
460 struct thread_info *ti = current_thread_info();
461 u32 fpexc = fmrx(FPEXC);
463 /* if vfp is on, then save state for resumption */
464 if (fpexc & FPEXC_EN) {
465 pr_debug("%s: saving vfp state\n", __func__);
466 vfp_save_state(&ti->vfpstate, fpexc);
468 /* disable, just in case */
469 fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
470 } else if (vfp_current_hw_state[ti->cpu]) {
472 fmxr(FPEXC, fpexc | FPEXC_EN);
473 vfp_save_state(vfp_current_hw_state[ti->cpu], fpexc);
478 /* clear any information we had about last context state */
479 vfp_current_hw_state[ti->cpu] = NULL;
484 static void vfp_pm_resume(void)
486 /* ensure we have access to the vfp */
489 /* and disable it to ensure the next usage restores the state */
490 fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
493 static int vfp_cpu_pm_notifier(struct notifier_block *self, unsigned long cmd,
500 case CPU_PM_ENTER_FAILED:
508 static struct notifier_block vfp_cpu_pm_notifier_block = {
509 .notifier_call = vfp_cpu_pm_notifier,
512 static void vfp_pm_init(void)
514 cpu_pm_register_notifier(&vfp_cpu_pm_notifier_block);
518 static inline void vfp_pm_init(void) { }
519 #endif /* CONFIG_CPU_PM */
522 * Ensure that the VFP state stored in 'thread->vfpstate' is up to date
523 * with the hardware state.
525 void vfp_sync_hwstate(struct thread_info *thread)
527 unsigned int cpu = get_cpu();
529 if (vfp_state_in_hw(cpu, thread)) {
530 u32 fpexc = fmrx(FPEXC);
533 * Save the last VFP state on this CPU.
535 fmxr(FPEXC, fpexc | FPEXC_EN);
536 vfp_save_state(&thread->vfpstate, fpexc | FPEXC_EN);
543 /* Ensure that the thread reloads the hardware VFP state on the next use. */
544 void vfp_flush_hwstate(struct thread_info *thread)
546 unsigned int cpu = get_cpu();
548 vfp_force_reload(cpu, thread);
554 * Save the current VFP state into the provided structures and prepare
555 * for entry into a new function (signal handler).
557 int vfp_preserve_user_clear_hwstate(struct user_vfp *ufp,
558 struct user_vfp_exc *ufp_exc)
560 struct thread_info *thread = current_thread_info();
561 struct vfp_hard_struct *hwstate = &thread->vfpstate.hard;
563 /* Ensure that the saved hwstate is up-to-date. */
564 vfp_sync_hwstate(thread);
567 * Copy the floating point registers. There can be unused
568 * registers see asm/hwcap.h for details.
570 memcpy(&ufp->fpregs, &hwstate->fpregs, sizeof(hwstate->fpregs));
573 * Copy the status and control register.
575 ufp->fpscr = hwstate->fpscr;
578 * Copy the exception registers.
580 ufp_exc->fpexc = hwstate->fpexc;
581 ufp_exc->fpinst = hwstate->fpinst;
582 ufp_exc->fpinst2 = ufp_exc->fpinst2;
584 /* Ensure that VFP is disabled. */
585 vfp_flush_hwstate(thread);
588 * As per the PCS, clear the length and stride bits for function
591 hwstate->fpscr &= ~(FPSCR_LENGTH_MASK | FPSCR_STRIDE_MASK);
595 /* Sanitise and restore the current VFP state from the provided structures. */
596 int vfp_restore_user_hwstate(struct user_vfp *ufp, struct user_vfp_exc *ufp_exc)
598 struct thread_info *thread = current_thread_info();
599 struct vfp_hard_struct *hwstate = &thread->vfpstate.hard;
602 /* Disable VFP to avoid corrupting the new thread state. */
603 vfp_flush_hwstate(thread);
606 * Copy the floating point registers. There can be unused
607 * registers see asm/hwcap.h for details.
609 memcpy(&hwstate->fpregs, &ufp->fpregs, sizeof(hwstate->fpregs));
611 * Copy the status and control register.
613 hwstate->fpscr = ufp->fpscr;
616 * Sanitise and restore the exception registers.
618 fpexc = ufp_exc->fpexc;
620 /* Ensure the VFP is enabled. */
623 /* Ensure FPINST2 is invalid and the exception flag is cleared. */
624 fpexc &= ~(FPEXC_EX | FPEXC_FP2V);
625 hwstate->fpexc = fpexc;
627 hwstate->fpinst = ufp_exc->fpinst;
628 hwstate->fpinst2 = ufp_exc->fpinst2;
634 * VFP hardware can lose all context when a CPU goes offline.
635 * As we will be running in SMP mode with CPU hotplug, we will save the
636 * hardware state at every thread switch. We clear our held state when
637 * a CPU has been killed, indicating that the VFP hardware doesn't contain
638 * a threads VFP state. When a CPU starts up, we re-enable access to the
639 * VFP hardware. The callbacks below are called on the CPU which
640 * is being offlined/onlined.
642 static int vfp_dying_cpu(unsigned int cpu)
644 vfp_current_hw_state[cpu] = NULL;
648 static int vfp_starting_cpu(unsigned int unused)
654 void vfp_kmode_exception(void)
657 * If we reach this point, a floating point exception has been raised
658 * while running in kernel mode. If the NEON/VFP unit was enabled at the
659 * time, it means a VFP instruction has been issued that requires
660 * software assistance to complete, something which is not currently
661 * supported in kernel mode.
662 * If the NEON/VFP unit was disabled, and the location pointed to below
663 * is properly preceded by a call to kernel_neon_begin(), something has
664 * caused the task to be scheduled out and back in again. In this case,
665 * rebuilding and running with CONFIG_DEBUG_ATOMIC_SLEEP enabled should
666 * be helpful in localizing the problem.
668 if (fmrx(FPEXC) & FPEXC_EN)
669 pr_crit("BUG: unsupported FP instruction in kernel mode\n");
671 pr_crit("BUG: FP instruction issued in kernel mode with FP unit disabled\n");
674 #ifdef CONFIG_KERNEL_MODE_NEON
677 * Kernel-side NEON support functions
679 void kernel_neon_begin(void)
681 struct thread_info *thread = current_thread_info();
686 * Kernel mode NEON is only allowed outside of interrupt context
687 * with preemption disabled. This will make sure that the kernel
688 * mode NEON register contents never need to be preserved.
690 BUG_ON(in_interrupt());
693 fpexc = fmrx(FPEXC) | FPEXC_EN;
697 * Save the userland NEON/VFP state. Under UP,
698 * the owner could be a task other than 'current'
700 if (vfp_state_in_hw(cpu, thread))
701 vfp_save_state(&thread->vfpstate, fpexc);
703 else if (vfp_current_hw_state[cpu] != NULL)
704 vfp_save_state(vfp_current_hw_state[cpu], fpexc);
706 vfp_current_hw_state[cpu] = NULL;
708 EXPORT_SYMBOL(kernel_neon_begin);
710 void kernel_neon_end(void)
712 /* Disable the NEON/VFP unit. */
713 fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
716 EXPORT_SYMBOL(kernel_neon_end);
718 #endif /* CONFIG_KERNEL_MODE_NEON */
721 * VFP support code initialisation.
723 static int __init vfp_init(void)
726 unsigned int cpu_arch = cpu_architecture();
729 * Enable the access to the VFP on all online CPUs so the
730 * following test on FPSID will succeed.
732 if (cpu_arch >= CPU_ARCH_ARMv6)
733 on_each_cpu(vfp_enable, NULL, 1);
736 * First check that there is a VFP that we can use.
737 * The handler is already setup to just log calls, so
738 * we just need to read the VFPSID register.
740 vfp_vector = vfp_testing_entry;
742 vfpsid = fmrx(FPSID);
744 vfp_vector = vfp_null_entry;
746 pr_info("VFP support v0.3: ");
748 pr_cont("not present\n");
750 /* Extract the architecture on CPUID scheme */
751 } else if ((read_cpuid_id() & 0x000f0000) == 0x000f0000) {
752 VFP_arch = vfpsid & FPSID_CPUID_ARCH_MASK;
753 VFP_arch >>= FPSID_ARCH_BIT;
755 * Check for the presence of the Advanced SIMD
756 * load/store instructions, integer and single
757 * precision floating point operations. Only check
758 * for NEON if the hardware has the MVFR registers.
760 if (IS_ENABLED(CONFIG_NEON) &&
761 (fmrx(MVFR1) & 0x000fff00) == 0x00011100)
762 elf_hwcap |= HWCAP_NEON;
764 if (IS_ENABLED(CONFIG_VFPv3)) {
765 u32 mvfr0 = fmrx(MVFR0);
766 if (((mvfr0 & MVFR0_DP_MASK) >> MVFR0_DP_BIT) == 0x2 ||
767 ((mvfr0 & MVFR0_SP_MASK) >> MVFR0_SP_BIT) == 0x2) {
768 elf_hwcap |= HWCAP_VFPv3;
770 * Check for VFPv3 D16 and VFPv4 D16. CPUs in
771 * this configuration only have 16 x 64bit
774 if ((mvfr0 & MVFR0_A_SIMD_MASK) == 1)
776 elf_hwcap |= HWCAP_VFPv3D16;
778 elf_hwcap |= HWCAP_VFPD32;
781 if ((fmrx(MVFR1) & 0xf0000000) == 0x10000000)
782 elf_hwcap |= HWCAP_VFPv4;
784 /* Extract the architecture version on pre-cpuid scheme */
786 if (vfpsid & FPSID_NODOUBLE) {
787 pr_cont("no double precision support\n");
791 VFP_arch = (vfpsid & FPSID_ARCH_MASK) >> FPSID_ARCH_BIT;
794 cpuhp_setup_state_nocalls(CPUHP_AP_ARM_VFP_STARTING,
795 "arm/vfp:starting", vfp_starting_cpu,
798 vfp_vector = vfp_support_entry;
800 thread_register_notifier(&vfp_notifier_block);
804 * We detected VFP, and the support code is
805 * in place; report VFP support to userspace.
807 elf_hwcap |= HWCAP_VFP;
809 pr_cont("implementor %02x architecture %d part %02x variant %x rev %x\n",
810 (vfpsid & FPSID_IMPLEMENTER_MASK) >> FPSID_IMPLEMENTER_BIT,
812 (vfpsid & FPSID_PART_MASK) >> FPSID_PART_BIT,
813 (vfpsid & FPSID_VARIANT_MASK) >> FPSID_VARIANT_BIT,
814 (vfpsid & FPSID_REV_MASK) >> FPSID_REV_BIT);
819 core_initcall(vfp_init);