xref: /openbmc/linux/arch/arm/vfp/vfpmodule.c (revision ac8b6f14)
1 /*
2  *  linux/arch/arm/vfp/vfpmodule.c
3  *
4  *  Copyright (C) 2004 ARM Limited.
5  *  Written by Deep Blue Solutions Limited.
6  *
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.
10  */
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>
24 
25 #include <asm/cp15.h>
26 #include <asm/cputype.h>
27 #include <asm/system_info.h>
28 #include <asm/thread_notify.h>
29 #include <asm/vfp.h>
30 
31 #include "vfpinstr.h"
32 #include "vfp.h"
33 
34 /*
35  * Our undef handlers (in entry.S)
36  */
37 asmlinkage void vfp_testing_entry(void);
38 asmlinkage void vfp_support_entry(void);
39 asmlinkage void vfp_null_entry(void);
40 
41 asmlinkage void (*vfp_vector)(void) = vfp_null_entry;
42 
43 /*
44  * Dual-use variable.
45  * Used in startup: set to non-zero if VFP checks fail
46  * After startup, holds VFP architecture
47  */
48 unsigned int VFP_arch;
49 
50 /*
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
53  * context is invalid.
54  *
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.
58  */
59 union vfp_state *vfp_current_hw_state[NR_CPUS];
60 
61 /*
62  * Is 'thread's most up to date state stored in this CPUs hardware?
63  * Must be called from non-preemptible context.
64  */
65 static bool vfp_state_in_hw(unsigned int cpu, struct thread_info *thread)
66 {
67 #ifdef CONFIG_SMP
68 	if (thread->vfpstate.hard.cpu != cpu)
69 		return false;
70 #endif
71 	return vfp_current_hw_state[cpu] == &thread->vfpstate;
72 }
73 
74 /*
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.
78  */
79 static void vfp_force_reload(unsigned int cpu, struct thread_info *thread)
80 {
81 	if (vfp_state_in_hw(cpu, thread)) {
82 		fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
83 		vfp_current_hw_state[cpu] = NULL;
84 	}
85 #ifdef CONFIG_SMP
86 	thread->vfpstate.hard.cpu = NR_CPUS;
87 #endif
88 }
89 
90 /*
91  * Per-thread VFP initialization.
92  */
93 static void vfp_thread_flush(struct thread_info *thread)
94 {
95 	union vfp_state *vfp = &thread->vfpstate;
96 	unsigned int cpu;
97 
98 	/*
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.
102 	 *
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.
105 	 */
106 	cpu = get_cpu();
107 	if (vfp_current_hw_state[cpu] == vfp)
108 		vfp_current_hw_state[cpu] = NULL;
109 	fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
110 	put_cpu();
111 
112 	memset(vfp, 0, sizeof(union vfp_state));
113 
114 	vfp->hard.fpexc = FPEXC_EN;
115 	vfp->hard.fpscr = FPSCR_ROUND_NEAREST;
116 #ifdef CONFIG_SMP
117 	vfp->hard.cpu = NR_CPUS;
118 #endif
119 }
120 
121 static void vfp_thread_exit(struct thread_info *thread)
122 {
123 	/* release case: Per-thread VFP cleanup. */
124 	union vfp_state *vfp = &thread->vfpstate;
125 	unsigned int cpu = get_cpu();
126 
127 	if (vfp_current_hw_state[cpu] == vfp)
128 		vfp_current_hw_state[cpu] = NULL;
129 	put_cpu();
130 }
131 
132 static void vfp_thread_copy(struct thread_info *thread)
133 {
134 	struct thread_info *parent = current_thread_info();
135 
136 	vfp_sync_hwstate(parent);
137 	thread->vfpstate = parent->vfpstate;
138 #ifdef CONFIG_SMP
139 	thread->vfpstate.hard.cpu = NR_CPUS;
140 #endif
141 }
142 
143 /*
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.
158  *  THREAD_NOTIFY_EXIT
159  *   - we could be preempted if tree preempt rcu is enabled, so
160  *	it is unsafe to use thread->cpu.
161  */
162 static int vfp_notifier(struct notifier_block *self, unsigned long cmd, void *v)
163 {
164 	struct thread_info *thread = v;
165 	u32 fpexc;
166 #ifdef CONFIG_SMP
167 	unsigned int cpu;
168 #endif
169 
170 	switch (cmd) {
171 	case THREAD_NOTIFY_SWITCH:
172 		fpexc = fmrx(FPEXC);
173 
174 #ifdef CONFIG_SMP
175 		cpu = thread->cpu;
176 
177 		/*
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.
181 		 */
182 		if ((fpexc & FPEXC_EN) && vfp_current_hw_state[cpu])
183 			vfp_save_state(vfp_current_hw_state[cpu], fpexc);
184 #endif
185 
186 		/*
187 		 * Always disable VFP so we can lazily save/restore the
188 		 * old state.
189 		 */
190 		fmxr(FPEXC, fpexc & ~FPEXC_EN);
191 		break;
192 
193 	case THREAD_NOTIFY_FLUSH:
194 		vfp_thread_flush(thread);
195 		break;
196 
197 	case THREAD_NOTIFY_EXIT:
198 		vfp_thread_exit(thread);
199 		break;
200 
201 	case THREAD_NOTIFY_COPY:
202 		vfp_thread_copy(thread);
203 		break;
204 	}
205 
206 	return NOTIFY_DONE;
207 }
208 
209 static struct notifier_block vfp_notifier_block = {
210 	.notifier_call	= vfp_notifier,
211 };
212 
213 /*
214  * Raise a SIGFPE for the current process.
215  * sicode describes the signal being raised.
216  */
217 static void vfp_raise_sigfpe(unsigned int sicode, struct pt_regs *regs)
218 {
219 	/*
220 	 * This is the same as NWFPE, because it's not clear what
221 	 * this is used for
222 	 */
223 	current->thread.error_code = 0;
224 	current->thread.trap_no = 6;
225 
226 	send_sig_fault(SIGFPE, sicode,
227 		       (void __user *)(instruction_pointer(regs) - 4),
228 		       current);
229 }
230 
231 static void vfp_panic(char *reason, u32 inst)
232 {
233 	int i;
234 
235 	pr_err("VFP: Error: %s\n", reason);
236 	pr_err("VFP: EXC 0x%08x SCR 0x%08x INST 0x%08x\n",
237 		fmrx(FPEXC), fmrx(FPSCR), inst);
238 	for (i = 0; i < 32; i += 2)
239 		pr_err("VFP: s%2u: 0x%08x s%2u: 0x%08x\n",
240 		       i, vfp_get_float(i), i+1, vfp_get_float(i+1));
241 }
242 
243 /*
244  * Process bitmask of exception conditions.
245  */
246 static void vfp_raise_exceptions(u32 exceptions, u32 inst, u32 fpscr, struct pt_regs *regs)
247 {
248 	int si_code = 0;
249 
250 	pr_debug("VFP: raising exceptions %08x\n", exceptions);
251 
252 	if (exceptions == VFP_EXCEPTION_ERROR) {
253 		vfp_panic("unhandled bounce", inst);
254 		vfp_raise_sigfpe(FPE_FLTINV, regs);
255 		return;
256 	}
257 
258 	/*
259 	 * If any of the status flags are set, update the FPSCR.
260 	 * Comparison instructions always return at least one of
261 	 * these flags set.
262 	 */
263 	if (exceptions & (FPSCR_N|FPSCR_Z|FPSCR_C|FPSCR_V))
264 		fpscr &= ~(FPSCR_N|FPSCR_Z|FPSCR_C|FPSCR_V);
265 
266 	fpscr |= exceptions;
267 
268 	fmxr(FPSCR, fpscr);
269 
270 #define RAISE(stat,en,sig)				\
271 	if (exceptions & stat && fpscr & en)		\
272 		si_code = sig;
273 
274 	/*
275 	 * These are arranged in priority order, least to highest.
276 	 */
277 	RAISE(FPSCR_DZC, FPSCR_DZE, FPE_FLTDIV);
278 	RAISE(FPSCR_IXC, FPSCR_IXE, FPE_FLTRES);
279 	RAISE(FPSCR_UFC, FPSCR_UFE, FPE_FLTUND);
280 	RAISE(FPSCR_OFC, FPSCR_OFE, FPE_FLTOVF);
281 	RAISE(FPSCR_IOC, FPSCR_IOE, FPE_FLTINV);
282 
283 	if (si_code)
284 		vfp_raise_sigfpe(si_code, regs);
285 }
286 
287 /*
288  * Emulate a VFP instruction.
289  */
290 static u32 vfp_emulate_instruction(u32 inst, u32 fpscr, struct pt_regs *regs)
291 {
292 	u32 exceptions = VFP_EXCEPTION_ERROR;
293 
294 	pr_debug("VFP: emulate: INST=0x%08x SCR=0x%08x\n", inst, fpscr);
295 
296 	if (INST_CPRTDO(inst)) {
297 		if (!INST_CPRT(inst)) {
298 			/*
299 			 * CPDO
300 			 */
301 			if (vfp_single(inst)) {
302 				exceptions = vfp_single_cpdo(inst, fpscr);
303 			} else {
304 				exceptions = vfp_double_cpdo(inst, fpscr);
305 			}
306 		} else {
307 			/*
308 			 * A CPRT instruction can not appear in FPINST2, nor
309 			 * can it cause an exception.  Therefore, we do not
310 			 * have to emulate it.
311 			 */
312 		}
313 	} else {
314 		/*
315 		 * A CPDT instruction can not appear in FPINST2, nor can
316 		 * it cause an exception.  Therefore, we do not have to
317 		 * emulate it.
318 		 */
319 	}
320 	return exceptions & ~VFP_NAN_FLAG;
321 }
322 
323 /*
324  * Package up a bounce condition.
325  */
326 void VFP_bounce(u32 trigger, u32 fpexc, struct pt_regs *regs)
327 {
328 	u32 fpscr, orig_fpscr, fpsid, exceptions;
329 
330 	pr_debug("VFP: bounce: trigger %08x fpexc %08x\n", trigger, fpexc);
331 
332 	/*
333 	 * At this point, FPEXC can have the following configuration:
334 	 *
335 	 *  EX DEX IXE
336 	 *  0   1   x   - synchronous exception
337 	 *  1   x   0   - asynchronous exception
338 	 *  1   x   1   - sychronous on VFP subarch 1 and asynchronous on later
339 	 *  0   0   1   - synchronous on VFP9 (non-standard subarch 1
340 	 *                implementation), undefined otherwise
341 	 *
342 	 * Clear various bits and enable access to the VFP so we can
343 	 * handle the bounce.
344 	 */
345 	fmxr(FPEXC, fpexc & ~(FPEXC_EX|FPEXC_DEX|FPEXC_FP2V|FPEXC_VV|FPEXC_TRAP_MASK));
346 
347 	fpsid = fmrx(FPSID);
348 	orig_fpscr = fpscr = fmrx(FPSCR);
349 
350 	/*
351 	 * Check for the special VFP subarch 1 and FPSCR.IXE bit case
352 	 */
353 	if ((fpsid & FPSID_ARCH_MASK) == (1 << FPSID_ARCH_BIT)
354 	    && (fpscr & FPSCR_IXE)) {
355 		/*
356 		 * Synchronous exception, emulate the trigger instruction
357 		 */
358 		goto emulate;
359 	}
360 
361 	if (fpexc & FPEXC_EX) {
362 #ifndef CONFIG_CPU_FEROCEON
363 		/*
364 		 * Asynchronous exception. The instruction is read from FPINST
365 		 * and the interrupted instruction has to be restarted.
366 		 */
367 		trigger = fmrx(FPINST);
368 		regs->ARM_pc -= 4;
369 #endif
370 	} else if (!(fpexc & FPEXC_DEX)) {
371 		/*
372 		 * Illegal combination of bits. It can be caused by an
373 		 * unallocated VFP instruction but with FPSCR.IXE set and not
374 		 * on VFP subarch 1.
375 		 */
376 		 vfp_raise_exceptions(VFP_EXCEPTION_ERROR, trigger, fpscr, regs);
377 		goto exit;
378 	}
379 
380 	/*
381 	 * Modify fpscr to indicate the number of iterations remaining.
382 	 * If FPEXC.EX is 0, FPEXC.DEX is 1 and the FPEXC.VV bit indicates
383 	 * whether FPEXC.VECITR or FPSCR.LEN is used.
384 	 */
385 	if (fpexc & (FPEXC_EX | FPEXC_VV)) {
386 		u32 len;
387 
388 		len = fpexc + (1 << FPEXC_LENGTH_BIT);
389 
390 		fpscr &= ~FPSCR_LENGTH_MASK;
391 		fpscr |= (len & FPEXC_LENGTH_MASK) << (FPSCR_LENGTH_BIT - FPEXC_LENGTH_BIT);
392 	}
393 
394 	/*
395 	 * Handle the first FP instruction.  We used to take note of the
396 	 * FPEXC bounce reason, but this appears to be unreliable.
397 	 * Emulate the bounced instruction instead.
398 	 */
399 	exceptions = vfp_emulate_instruction(trigger, fpscr, regs);
400 	if (exceptions)
401 		vfp_raise_exceptions(exceptions, trigger, orig_fpscr, regs);
402 
403 	/*
404 	 * If there isn't a second FP instruction, exit now. Note that
405 	 * the FPEXC.FP2V bit is valid only if FPEXC.EX is 1.
406 	 */
407 	if ((fpexc & (FPEXC_EX | FPEXC_FP2V)) != (FPEXC_EX | FPEXC_FP2V))
408 		goto exit;
409 
410 	/*
411 	 * The barrier() here prevents fpinst2 being read
412 	 * before the condition above.
413 	 */
414 	barrier();
415 	trigger = fmrx(FPINST2);
416 
417  emulate:
418 	exceptions = vfp_emulate_instruction(trigger, orig_fpscr, regs);
419 	if (exceptions)
420 		vfp_raise_exceptions(exceptions, trigger, orig_fpscr, regs);
421  exit:
422 	preempt_enable();
423 }
424 
425 static void vfp_enable(void *unused)
426 {
427 	u32 access;
428 
429 	BUG_ON(preemptible());
430 	access = get_copro_access();
431 
432 	/*
433 	 * Enable full access to VFP (cp10 and cp11)
434 	 */
435 	set_copro_access(access | CPACC_FULL(10) | CPACC_FULL(11));
436 }
437 
438 /* Called by platforms on which we want to disable VFP because it may not be
439  * present on all CPUs within a SMP complex. Needs to be called prior to
440  * vfp_init().
441  */
442 void vfp_disable(void)
443 {
444 	if (VFP_arch) {
445 		pr_debug("%s: should be called prior to vfp_init\n", __func__);
446 		return;
447 	}
448 	VFP_arch = 1;
449 }
450 
451 #ifdef CONFIG_CPU_PM
452 static int vfp_pm_suspend(void)
453 {
454 	struct thread_info *ti = current_thread_info();
455 	u32 fpexc = fmrx(FPEXC);
456 
457 	/* if vfp is on, then save state for resumption */
458 	if (fpexc & FPEXC_EN) {
459 		pr_debug("%s: saving vfp state\n", __func__);
460 		vfp_save_state(&ti->vfpstate, fpexc);
461 
462 		/* disable, just in case */
463 		fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
464 	} else if (vfp_current_hw_state[ti->cpu]) {
465 #ifndef CONFIG_SMP
466 		fmxr(FPEXC, fpexc | FPEXC_EN);
467 		vfp_save_state(vfp_current_hw_state[ti->cpu], fpexc);
468 		fmxr(FPEXC, fpexc);
469 #endif
470 	}
471 
472 	/* clear any information we had about last context state */
473 	vfp_current_hw_state[ti->cpu] = NULL;
474 
475 	return 0;
476 }
477 
478 static void vfp_pm_resume(void)
479 {
480 	/* ensure we have access to the vfp */
481 	vfp_enable(NULL);
482 
483 	/* and disable it to ensure the next usage restores the state */
484 	fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
485 }
486 
487 static int vfp_cpu_pm_notifier(struct notifier_block *self, unsigned long cmd,
488 	void *v)
489 {
490 	switch (cmd) {
491 	case CPU_PM_ENTER:
492 		vfp_pm_suspend();
493 		break;
494 	case CPU_PM_ENTER_FAILED:
495 	case CPU_PM_EXIT:
496 		vfp_pm_resume();
497 		break;
498 	}
499 	return NOTIFY_OK;
500 }
501 
502 static struct notifier_block vfp_cpu_pm_notifier_block = {
503 	.notifier_call = vfp_cpu_pm_notifier,
504 };
505 
506 static void vfp_pm_init(void)
507 {
508 	cpu_pm_register_notifier(&vfp_cpu_pm_notifier_block);
509 }
510 
511 #else
512 static inline void vfp_pm_init(void) { }
513 #endif /* CONFIG_CPU_PM */
514 
515 /*
516  * Ensure that the VFP state stored in 'thread->vfpstate' is up to date
517  * with the hardware state.
518  */
519 void vfp_sync_hwstate(struct thread_info *thread)
520 {
521 	unsigned int cpu = get_cpu();
522 
523 	if (vfp_state_in_hw(cpu, thread)) {
524 		u32 fpexc = fmrx(FPEXC);
525 
526 		/*
527 		 * Save the last VFP state on this CPU.
528 		 */
529 		fmxr(FPEXC, fpexc | FPEXC_EN);
530 		vfp_save_state(&thread->vfpstate, fpexc | FPEXC_EN);
531 		fmxr(FPEXC, fpexc);
532 	}
533 
534 	put_cpu();
535 }
536 
537 /* Ensure that the thread reloads the hardware VFP state on the next use. */
538 void vfp_flush_hwstate(struct thread_info *thread)
539 {
540 	unsigned int cpu = get_cpu();
541 
542 	vfp_force_reload(cpu, thread);
543 
544 	put_cpu();
545 }
546 
547 /*
548  * Save the current VFP state into the provided structures and prepare
549  * for entry into a new function (signal handler).
550  */
551 int vfp_preserve_user_clear_hwstate(struct user_vfp *ufp,
552 				    struct user_vfp_exc *ufp_exc)
553 {
554 	struct thread_info *thread = current_thread_info();
555 	struct vfp_hard_struct *hwstate = &thread->vfpstate.hard;
556 
557 	/* Ensure that the saved hwstate is up-to-date. */
558 	vfp_sync_hwstate(thread);
559 
560 	/*
561 	 * Copy the floating point registers. There can be unused
562 	 * registers see asm/hwcap.h for details.
563 	 */
564 	memcpy(&ufp->fpregs, &hwstate->fpregs, sizeof(hwstate->fpregs));
565 
566 	/*
567 	 * Copy the status and control register.
568 	 */
569 	ufp->fpscr = hwstate->fpscr;
570 
571 	/*
572 	 * Copy the exception registers.
573 	 */
574 	ufp_exc->fpexc = hwstate->fpexc;
575 	ufp_exc->fpinst = hwstate->fpinst;
576 	ufp_exc->fpinst2 = hwstate->fpinst2;
577 
578 	/* Ensure that VFP is disabled. */
579 	vfp_flush_hwstate(thread);
580 
581 	/*
582 	 * As per the PCS, clear the length and stride bits for function
583 	 * entry.
584 	 */
585 	hwstate->fpscr &= ~(FPSCR_LENGTH_MASK | FPSCR_STRIDE_MASK);
586 	return 0;
587 }
588 
589 /* Sanitise and restore the current VFP state from the provided structures. */
590 int vfp_restore_user_hwstate(struct user_vfp *ufp, struct user_vfp_exc *ufp_exc)
591 {
592 	struct thread_info *thread = current_thread_info();
593 	struct vfp_hard_struct *hwstate = &thread->vfpstate.hard;
594 	unsigned long fpexc;
595 
596 	/* Disable VFP to avoid corrupting the new thread state. */
597 	vfp_flush_hwstate(thread);
598 
599 	/*
600 	 * Copy the floating point registers. There can be unused
601 	 * registers see asm/hwcap.h for details.
602 	 */
603 	memcpy(&hwstate->fpregs, &ufp->fpregs, sizeof(hwstate->fpregs));
604 	/*
605 	 * Copy the status and control register.
606 	 */
607 	hwstate->fpscr = ufp->fpscr;
608 
609 	/*
610 	 * Sanitise and restore the exception registers.
611 	 */
612 	fpexc = ufp_exc->fpexc;
613 
614 	/* Ensure the VFP is enabled. */
615 	fpexc |= FPEXC_EN;
616 
617 	/* Ensure FPINST2 is invalid and the exception flag is cleared. */
618 	fpexc &= ~(FPEXC_EX | FPEXC_FP2V);
619 	hwstate->fpexc = fpexc;
620 
621 	hwstate->fpinst = ufp_exc->fpinst;
622 	hwstate->fpinst2 = ufp_exc->fpinst2;
623 
624 	return 0;
625 }
626 
627 /*
628  * VFP hardware can lose all context when a CPU goes offline.
629  * As we will be running in SMP mode with CPU hotplug, we will save the
630  * hardware state at every thread switch.  We clear our held state when
631  * a CPU has been killed, indicating that the VFP hardware doesn't contain
632  * a threads VFP state.  When a CPU starts up, we re-enable access to the
633  * VFP hardware. The callbacks below are called on the CPU which
634  * is being offlined/onlined.
635  */
636 static int vfp_dying_cpu(unsigned int cpu)
637 {
638 	vfp_current_hw_state[cpu] = NULL;
639 	return 0;
640 }
641 
642 static int vfp_starting_cpu(unsigned int unused)
643 {
644 	vfp_enable(NULL);
645 	return 0;
646 }
647 
648 void vfp_kmode_exception(void)
649 {
650 	/*
651 	 * If we reach this point, a floating point exception has been raised
652 	 * while running in kernel mode. If the NEON/VFP unit was enabled at the
653 	 * time, it means a VFP instruction has been issued that requires
654 	 * software assistance to complete, something which is not currently
655 	 * supported in kernel mode.
656 	 * If the NEON/VFP unit was disabled, and the location pointed to below
657 	 * is properly preceded by a call to kernel_neon_begin(), something has
658 	 * caused the task to be scheduled out and back in again. In this case,
659 	 * rebuilding and running with CONFIG_DEBUG_ATOMIC_SLEEP enabled should
660 	 * be helpful in localizing the problem.
661 	 */
662 	if (fmrx(FPEXC) & FPEXC_EN)
663 		pr_crit("BUG: unsupported FP instruction in kernel mode\n");
664 	else
665 		pr_crit("BUG: FP instruction issued in kernel mode with FP unit disabled\n");
666 }
667 
668 #ifdef CONFIG_KERNEL_MODE_NEON
669 
670 /*
671  * Kernel-side NEON support functions
672  */
673 void kernel_neon_begin(void)
674 {
675 	struct thread_info *thread = current_thread_info();
676 	unsigned int cpu;
677 	u32 fpexc;
678 
679 	/*
680 	 * Kernel mode NEON is only allowed outside of interrupt context
681 	 * with preemption disabled. This will make sure that the kernel
682 	 * mode NEON register contents never need to be preserved.
683 	 */
684 	BUG_ON(in_interrupt());
685 	cpu = get_cpu();
686 
687 	fpexc = fmrx(FPEXC) | FPEXC_EN;
688 	fmxr(FPEXC, fpexc);
689 
690 	/*
691 	 * Save the userland NEON/VFP state. Under UP,
692 	 * the owner could be a task other than 'current'
693 	 */
694 	if (vfp_state_in_hw(cpu, thread))
695 		vfp_save_state(&thread->vfpstate, fpexc);
696 #ifndef CONFIG_SMP
697 	else if (vfp_current_hw_state[cpu] != NULL)
698 		vfp_save_state(vfp_current_hw_state[cpu], fpexc);
699 #endif
700 	vfp_current_hw_state[cpu] = NULL;
701 }
702 EXPORT_SYMBOL(kernel_neon_begin);
703 
704 void kernel_neon_end(void)
705 {
706 	/* Disable the NEON/VFP unit. */
707 	fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
708 	put_cpu();
709 }
710 EXPORT_SYMBOL(kernel_neon_end);
711 
712 #endif /* CONFIG_KERNEL_MODE_NEON */
713 
714 /*
715  * VFP support code initialisation.
716  */
717 static int __init vfp_init(void)
718 {
719 	unsigned int vfpsid;
720 	unsigned int cpu_arch = cpu_architecture();
721 
722 	/*
723 	 * Enable the access to the VFP on all online CPUs so the
724 	 * following test on FPSID will succeed.
725 	 */
726 	if (cpu_arch >= CPU_ARCH_ARMv6)
727 		on_each_cpu(vfp_enable, NULL, 1);
728 
729 	/*
730 	 * First check that there is a VFP that we can use.
731 	 * The handler is already setup to just log calls, so
732 	 * we just need to read the VFPSID register.
733 	 */
734 	vfp_vector = vfp_testing_entry;
735 	barrier();
736 	vfpsid = fmrx(FPSID);
737 	barrier();
738 	vfp_vector = vfp_null_entry;
739 
740 	pr_info("VFP support v0.3: ");
741 	if (VFP_arch) {
742 		pr_cont("not present\n");
743 		return 0;
744 	/* Extract the architecture on CPUID scheme */
745 	} else if ((read_cpuid_id() & 0x000f0000) == 0x000f0000) {
746 		VFP_arch = vfpsid & FPSID_CPUID_ARCH_MASK;
747 		VFP_arch >>= FPSID_ARCH_BIT;
748 		/*
749 		 * Check for the presence of the Advanced SIMD
750 		 * load/store instructions, integer and single
751 		 * precision floating point operations. Only check
752 		 * for NEON if the hardware has the MVFR registers.
753 		 */
754 		if (IS_ENABLED(CONFIG_NEON) &&
755 		   (fmrx(MVFR1) & 0x000fff00) == 0x00011100)
756 			elf_hwcap |= HWCAP_NEON;
757 
758 		if (IS_ENABLED(CONFIG_VFPv3)) {
759 			u32 mvfr0 = fmrx(MVFR0);
760 			if (((mvfr0 & MVFR0_DP_MASK) >> MVFR0_DP_BIT) == 0x2 ||
761 			    ((mvfr0 & MVFR0_SP_MASK) >> MVFR0_SP_BIT) == 0x2) {
762 				elf_hwcap |= HWCAP_VFPv3;
763 				/*
764 				 * Check for VFPv3 D16 and VFPv4 D16.  CPUs in
765 				 * this configuration only have 16 x 64bit
766 				 * registers.
767 				 */
768 				if ((mvfr0 & MVFR0_A_SIMD_MASK) == 1)
769 					/* also v4-D16 */
770 					elf_hwcap |= HWCAP_VFPv3D16;
771 				else
772 					elf_hwcap |= HWCAP_VFPD32;
773 			}
774 
775 			if ((fmrx(MVFR1) & 0xf0000000) == 0x10000000)
776 				elf_hwcap |= HWCAP_VFPv4;
777 		}
778 	/* Extract the architecture version on pre-cpuid scheme */
779 	} else {
780 		if (vfpsid & FPSID_NODOUBLE) {
781 			pr_cont("no double precision support\n");
782 			return 0;
783 		}
784 
785 		VFP_arch = (vfpsid & FPSID_ARCH_MASK) >> FPSID_ARCH_BIT;
786 	}
787 
788 	cpuhp_setup_state_nocalls(CPUHP_AP_ARM_VFP_STARTING,
789 				  "arm/vfp:starting", vfp_starting_cpu,
790 				  vfp_dying_cpu);
791 
792 	vfp_vector = vfp_support_entry;
793 
794 	thread_register_notifier(&vfp_notifier_block);
795 	vfp_pm_init();
796 
797 	/*
798 	 * We detected VFP, and the support code is
799 	 * in place; report VFP support to userspace.
800 	 */
801 	elf_hwcap |= HWCAP_VFP;
802 
803 	pr_cont("implementor %02x architecture %d part %02x variant %x rev %x\n",
804 		(vfpsid & FPSID_IMPLEMENTER_MASK) >> FPSID_IMPLEMENTER_BIT,
805 		VFP_arch,
806 		(vfpsid & FPSID_PART_MASK) >> FPSID_PART_BIT,
807 		(vfpsid & FPSID_VARIANT_MASK) >> FPSID_VARIANT_BIT,
808 		(vfpsid & FPSID_REV_MASK) >> FPSID_REV_BIT);
809 
810 	return 0;
811 }
812 
813 core_initcall(vfp_init);
814