xref: /openbmc/linux/arch/arm/vfp/vfpmodule.c (revision f220d3eb)
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 	siginfo_t info;
220 
221 	clear_siginfo(&info);
222 	info.si_signo = SIGFPE;
223 	info.si_code = sicode;
224 	info.si_addr = (void __user *)(instruction_pointer(regs) - 4);
225 
226 	/*
227 	 * This is the same as NWFPE, because it's not clear what
228 	 * this is used for
229 	 */
230 	current->thread.error_code = 0;
231 	current->thread.trap_no = 6;
232 
233 	send_sig_info(SIGFPE, &info, current);
234 }
235 
236 static void vfp_panic(char *reason, u32 inst)
237 {
238 	int i;
239 
240 	pr_err("VFP: Error: %s\n", reason);
241 	pr_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 		pr_err("VFP: s%2u: 0x%08x s%2u: 0x%08x\n",
245 		       i, vfp_get_float(i), i+1, vfp_get_float(i+1));
246 }
247 
248 /*
249  * Process bitmask of exception conditions.
250  */
251 static void vfp_raise_exceptions(u32 exceptions, u32 inst, u32 fpscr, struct pt_regs *regs)
252 {
253 	int si_code = 0;
254 
255 	pr_debug("VFP: raising exceptions %08x\n", exceptions);
256 
257 	if (exceptions == VFP_EXCEPTION_ERROR) {
258 		vfp_panic("unhandled bounce", inst);
259 		vfp_raise_sigfpe(FPE_FLTINV, regs);
260 		return;
261 	}
262 
263 	/*
264 	 * If any of the status flags are set, update the FPSCR.
265 	 * Comparison instructions always return at least one of
266 	 * these flags set.
267 	 */
268 	if (exceptions & (FPSCR_N|FPSCR_Z|FPSCR_C|FPSCR_V))
269 		fpscr &= ~(FPSCR_N|FPSCR_Z|FPSCR_C|FPSCR_V);
270 
271 	fpscr |= exceptions;
272 
273 	fmxr(FPSCR, fpscr);
274 
275 #define RAISE(stat,en,sig)				\
276 	if (exceptions & stat && fpscr & en)		\
277 		si_code = sig;
278 
279 	/*
280 	 * These are arranged in priority order, least to highest.
281 	 */
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);
287 
288 	if (si_code)
289 		vfp_raise_sigfpe(si_code, regs);
290 }
291 
292 /*
293  * Emulate a VFP instruction.
294  */
295 static u32 vfp_emulate_instruction(u32 inst, u32 fpscr, struct pt_regs *regs)
296 {
297 	u32 exceptions = VFP_EXCEPTION_ERROR;
298 
299 	pr_debug("VFP: emulate: INST=0x%08x SCR=0x%08x\n", inst, fpscr);
300 
301 	if (INST_CPRTDO(inst)) {
302 		if (!INST_CPRT(inst)) {
303 			/*
304 			 * CPDO
305 			 */
306 			if (vfp_single(inst)) {
307 				exceptions = vfp_single_cpdo(inst, fpscr);
308 			} else {
309 				exceptions = vfp_double_cpdo(inst, fpscr);
310 			}
311 		} else {
312 			/*
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.
316 			 */
317 		}
318 	} else {
319 		/*
320 		 * A CPDT instruction can not appear in FPINST2, nor can
321 		 * it cause an exception.  Therefore, we do not have to
322 		 * emulate it.
323 		 */
324 	}
325 	return exceptions & ~VFP_NAN_FLAG;
326 }
327 
328 /*
329  * Package up a bounce condition.
330  */
331 void VFP_bounce(u32 trigger, u32 fpexc, struct pt_regs *regs)
332 {
333 	u32 fpscr, orig_fpscr, fpsid, exceptions;
334 
335 	pr_debug("VFP: bounce: trigger %08x fpexc %08x\n", trigger, fpexc);
336 
337 	/*
338 	 * At this point, FPEXC can have the following configuration:
339 	 *
340 	 *  EX DEX IXE
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
346 	 *
347 	 * Clear various bits and enable access to the VFP so we can
348 	 * handle the bounce.
349 	 */
350 	fmxr(FPEXC, fpexc & ~(FPEXC_EX|FPEXC_DEX|FPEXC_FP2V|FPEXC_VV|FPEXC_TRAP_MASK));
351 
352 	fpsid = fmrx(FPSID);
353 	orig_fpscr = fpscr = fmrx(FPSCR);
354 
355 	/*
356 	 * Check for the special VFP subarch 1 and FPSCR.IXE bit case
357 	 */
358 	if ((fpsid & FPSID_ARCH_MASK) == (1 << FPSID_ARCH_BIT)
359 	    && (fpscr & FPSCR_IXE)) {
360 		/*
361 		 * Synchronous exception, emulate the trigger instruction
362 		 */
363 		goto emulate;
364 	}
365 
366 	if (fpexc & FPEXC_EX) {
367 #ifndef CONFIG_CPU_FEROCEON
368 		/*
369 		 * Asynchronous exception. The instruction is read from FPINST
370 		 * and the interrupted instruction has to be restarted.
371 		 */
372 		trigger = fmrx(FPINST);
373 		regs->ARM_pc -= 4;
374 #endif
375 	} else if (!(fpexc & FPEXC_DEX)) {
376 		/*
377 		 * Illegal combination of bits. It can be caused by an
378 		 * unallocated VFP instruction but with FPSCR.IXE set and not
379 		 * on VFP subarch 1.
380 		 */
381 		 vfp_raise_exceptions(VFP_EXCEPTION_ERROR, trigger, fpscr, regs);
382 		goto exit;
383 	}
384 
385 	/*
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.
389 	 */
390 	if (fpexc & (FPEXC_EX | FPEXC_VV)) {
391 		u32 len;
392 
393 		len = fpexc + (1 << FPEXC_LENGTH_BIT);
394 
395 		fpscr &= ~FPSCR_LENGTH_MASK;
396 		fpscr |= (len & FPEXC_LENGTH_MASK) << (FPSCR_LENGTH_BIT - FPEXC_LENGTH_BIT);
397 	}
398 
399 	/*
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.
403 	 */
404 	exceptions = vfp_emulate_instruction(trigger, fpscr, regs);
405 	if (exceptions)
406 		vfp_raise_exceptions(exceptions, trigger, orig_fpscr, regs);
407 
408 	/*
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.
411 	 */
412 	if ((fpexc & (FPEXC_EX | FPEXC_FP2V)) != (FPEXC_EX | FPEXC_FP2V))
413 		goto exit;
414 
415 	/*
416 	 * The barrier() here prevents fpinst2 being read
417 	 * before the condition above.
418 	 */
419 	barrier();
420 	trigger = fmrx(FPINST2);
421 
422  emulate:
423 	exceptions = vfp_emulate_instruction(trigger, orig_fpscr, regs);
424 	if (exceptions)
425 		vfp_raise_exceptions(exceptions, trigger, orig_fpscr, regs);
426  exit:
427 	preempt_enable();
428 }
429 
430 static void vfp_enable(void *unused)
431 {
432 	u32 access;
433 
434 	BUG_ON(preemptible());
435 	access = get_copro_access();
436 
437 	/*
438 	 * Enable full access to VFP (cp10 and cp11)
439 	 */
440 	set_copro_access(access | CPACC_FULL(10) | CPACC_FULL(11));
441 }
442 
443 /* Called by platforms on which we want to disable VFP because it may not be
444  * present on all CPUs within a SMP complex. Needs to be called prior to
445  * vfp_init().
446  */
447 void vfp_disable(void)
448 {
449 	if (VFP_arch) {
450 		pr_debug("%s: should be called prior to vfp_init\n", __func__);
451 		return;
452 	}
453 	VFP_arch = 1;
454 }
455 
456 #ifdef CONFIG_CPU_PM
457 static int vfp_pm_suspend(void)
458 {
459 	struct thread_info *ti = current_thread_info();
460 	u32 fpexc = fmrx(FPEXC);
461 
462 	/* if vfp is on, then save state for resumption */
463 	if (fpexc & FPEXC_EN) {
464 		pr_debug("%s: saving vfp state\n", __func__);
465 		vfp_save_state(&ti->vfpstate, fpexc);
466 
467 		/* disable, just in case */
468 		fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
469 	} else if (vfp_current_hw_state[ti->cpu]) {
470 #ifndef CONFIG_SMP
471 		fmxr(FPEXC, fpexc | FPEXC_EN);
472 		vfp_save_state(vfp_current_hw_state[ti->cpu], fpexc);
473 		fmxr(FPEXC, fpexc);
474 #endif
475 	}
476 
477 	/* clear any information we had about last context state */
478 	vfp_current_hw_state[ti->cpu] = NULL;
479 
480 	return 0;
481 }
482 
483 static void vfp_pm_resume(void)
484 {
485 	/* ensure we have access to the vfp */
486 	vfp_enable(NULL);
487 
488 	/* and disable it to ensure the next usage restores the state */
489 	fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
490 }
491 
492 static int vfp_cpu_pm_notifier(struct notifier_block *self, unsigned long cmd,
493 	void *v)
494 {
495 	switch (cmd) {
496 	case CPU_PM_ENTER:
497 		vfp_pm_suspend();
498 		break;
499 	case CPU_PM_ENTER_FAILED:
500 	case CPU_PM_EXIT:
501 		vfp_pm_resume();
502 		break;
503 	}
504 	return NOTIFY_OK;
505 }
506 
507 static struct notifier_block vfp_cpu_pm_notifier_block = {
508 	.notifier_call = vfp_cpu_pm_notifier,
509 };
510 
511 static void vfp_pm_init(void)
512 {
513 	cpu_pm_register_notifier(&vfp_cpu_pm_notifier_block);
514 }
515 
516 #else
517 static inline void vfp_pm_init(void) { }
518 #endif /* CONFIG_CPU_PM */
519 
520 /*
521  * Ensure that the VFP state stored in 'thread->vfpstate' is up to date
522  * with the hardware state.
523  */
524 void vfp_sync_hwstate(struct thread_info *thread)
525 {
526 	unsigned int cpu = get_cpu();
527 
528 	if (vfp_state_in_hw(cpu, thread)) {
529 		u32 fpexc = fmrx(FPEXC);
530 
531 		/*
532 		 * Save the last VFP state on this CPU.
533 		 */
534 		fmxr(FPEXC, fpexc | FPEXC_EN);
535 		vfp_save_state(&thread->vfpstate, fpexc | FPEXC_EN);
536 		fmxr(FPEXC, fpexc);
537 	}
538 
539 	put_cpu();
540 }
541 
542 /* Ensure that the thread reloads the hardware VFP state on the next use. */
543 void vfp_flush_hwstate(struct thread_info *thread)
544 {
545 	unsigned int cpu = get_cpu();
546 
547 	vfp_force_reload(cpu, thread);
548 
549 	put_cpu();
550 }
551 
552 /*
553  * Save the current VFP state into the provided structures and prepare
554  * for entry into a new function (signal handler).
555  */
556 int vfp_preserve_user_clear_hwstate(struct user_vfp __user *ufp,
557 				    struct user_vfp_exc __user *ufp_exc)
558 {
559 	struct thread_info *thread = current_thread_info();
560 	struct vfp_hard_struct *hwstate = &thread->vfpstate.hard;
561 	int err = 0;
562 
563 	/* Ensure that the saved hwstate is up-to-date. */
564 	vfp_sync_hwstate(thread);
565 
566 	/*
567 	 * Copy the floating point registers. There can be unused
568 	 * registers see asm/hwcap.h for details.
569 	 */
570 	err |= __copy_to_user(&ufp->fpregs, &hwstate->fpregs,
571 			      sizeof(hwstate->fpregs));
572 	/*
573 	 * Copy the status and control register.
574 	 */
575 	__put_user_error(hwstate->fpscr, &ufp->fpscr, err);
576 
577 	/*
578 	 * Copy the exception registers.
579 	 */
580 	__put_user_error(hwstate->fpexc, &ufp_exc->fpexc, err);
581 	__put_user_error(hwstate->fpinst, &ufp_exc->fpinst, err);
582 	__put_user_error(hwstate->fpinst2, &ufp_exc->fpinst2, err);
583 
584 	if (err)
585 		return -EFAULT;
586 
587 	/* Ensure that VFP is disabled. */
588 	vfp_flush_hwstate(thread);
589 
590 	/*
591 	 * As per the PCS, clear the length and stride bits for function
592 	 * entry.
593 	 */
594 	hwstate->fpscr &= ~(FPSCR_LENGTH_MASK | FPSCR_STRIDE_MASK);
595 	return 0;
596 }
597 
598 /* Sanitise and restore the current VFP state from the provided structures. */
599 int vfp_restore_user_hwstate(struct user_vfp *ufp, struct user_vfp_exc *ufp_exc)
600 {
601 	struct thread_info *thread = current_thread_info();
602 	struct vfp_hard_struct *hwstate = &thread->vfpstate.hard;
603 	unsigned long fpexc;
604 
605 	/* Disable VFP to avoid corrupting the new thread state. */
606 	vfp_flush_hwstate(thread);
607 
608 	/*
609 	 * Copy the floating point registers. There can be unused
610 	 * registers see asm/hwcap.h for details.
611 	 */
612 	memcpy(&hwstate->fpregs, &ufp->fpregs, sizeof(hwstate->fpregs));
613 	/*
614 	 * Copy the status and control register.
615 	 */
616 	hwstate->fpscr = ufp->fpscr;
617 
618 	/*
619 	 * Sanitise and restore the exception registers.
620 	 */
621 	fpexc = ufp_exc->fpexc;
622 
623 	/* Ensure the VFP is enabled. */
624 	fpexc |= FPEXC_EN;
625 
626 	/* Ensure FPINST2 is invalid and the exception flag is cleared. */
627 	fpexc &= ~(FPEXC_EX | FPEXC_FP2V);
628 	hwstate->fpexc = fpexc;
629 
630 	hwstate->fpinst = ufp_exc->fpinst;
631 	hwstate->fpinst2 = ufp_exc->fpinst2;
632 
633 	return 0;
634 }
635 
636 /*
637  * VFP hardware can lose all context when a CPU goes offline.
638  * As we will be running in SMP mode with CPU hotplug, we will save the
639  * hardware state at every thread switch.  We clear our held state when
640  * a CPU has been killed, indicating that the VFP hardware doesn't contain
641  * a threads VFP state.  When a CPU starts up, we re-enable access to the
642  * VFP hardware. The callbacks below are called on the CPU which
643  * is being offlined/onlined.
644  */
645 static int vfp_dying_cpu(unsigned int cpu)
646 {
647 	vfp_current_hw_state[cpu] = NULL;
648 	return 0;
649 }
650 
651 static int vfp_starting_cpu(unsigned int unused)
652 {
653 	vfp_enable(NULL);
654 	return 0;
655 }
656 
657 void vfp_kmode_exception(void)
658 {
659 	/*
660 	 * If we reach this point, a floating point exception has been raised
661 	 * while running in kernel mode. If the NEON/VFP unit was enabled at the
662 	 * time, it means a VFP instruction has been issued that requires
663 	 * software assistance to complete, something which is not currently
664 	 * supported in kernel mode.
665 	 * If the NEON/VFP unit was disabled, and the location pointed to below
666 	 * is properly preceded by a call to kernel_neon_begin(), something has
667 	 * caused the task to be scheduled out and back in again. In this case,
668 	 * rebuilding and running with CONFIG_DEBUG_ATOMIC_SLEEP enabled should
669 	 * be helpful in localizing the problem.
670 	 */
671 	if (fmrx(FPEXC) & FPEXC_EN)
672 		pr_crit("BUG: unsupported FP instruction in kernel mode\n");
673 	else
674 		pr_crit("BUG: FP instruction issued in kernel mode with FP unit disabled\n");
675 }
676 
677 #ifdef CONFIG_KERNEL_MODE_NEON
678 
679 /*
680  * Kernel-side NEON support functions
681  */
682 void kernel_neon_begin(void)
683 {
684 	struct thread_info *thread = current_thread_info();
685 	unsigned int cpu;
686 	u32 fpexc;
687 
688 	/*
689 	 * Kernel mode NEON is only allowed outside of interrupt context
690 	 * with preemption disabled. This will make sure that the kernel
691 	 * mode NEON register contents never need to be preserved.
692 	 */
693 	BUG_ON(in_interrupt());
694 	cpu = get_cpu();
695 
696 	fpexc = fmrx(FPEXC) | FPEXC_EN;
697 	fmxr(FPEXC, fpexc);
698 
699 	/*
700 	 * Save the userland NEON/VFP state. Under UP,
701 	 * the owner could be a task other than 'current'
702 	 */
703 	if (vfp_state_in_hw(cpu, thread))
704 		vfp_save_state(&thread->vfpstate, fpexc);
705 #ifndef CONFIG_SMP
706 	else if (vfp_current_hw_state[cpu] != NULL)
707 		vfp_save_state(vfp_current_hw_state[cpu], fpexc);
708 #endif
709 	vfp_current_hw_state[cpu] = NULL;
710 }
711 EXPORT_SYMBOL(kernel_neon_begin);
712 
713 void kernel_neon_end(void)
714 {
715 	/* Disable the NEON/VFP unit. */
716 	fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
717 	put_cpu();
718 }
719 EXPORT_SYMBOL(kernel_neon_end);
720 
721 #endif /* CONFIG_KERNEL_MODE_NEON */
722 
723 /*
724  * VFP support code initialisation.
725  */
726 static int __init vfp_init(void)
727 {
728 	unsigned int vfpsid;
729 	unsigned int cpu_arch = cpu_architecture();
730 
731 	/*
732 	 * Enable the access to the VFP on all online CPUs so the
733 	 * following test on FPSID will succeed.
734 	 */
735 	if (cpu_arch >= CPU_ARCH_ARMv6)
736 		on_each_cpu(vfp_enable, NULL, 1);
737 
738 	/*
739 	 * First check that there is a VFP that we can use.
740 	 * The handler is already setup to just log calls, so
741 	 * we just need to read the VFPSID register.
742 	 */
743 	vfp_vector = vfp_testing_entry;
744 	barrier();
745 	vfpsid = fmrx(FPSID);
746 	barrier();
747 	vfp_vector = vfp_null_entry;
748 
749 	pr_info("VFP support v0.3: ");
750 	if (VFP_arch) {
751 		pr_cont("not present\n");
752 		return 0;
753 	/* Extract the architecture on CPUID scheme */
754 	} else if ((read_cpuid_id() & 0x000f0000) == 0x000f0000) {
755 		VFP_arch = vfpsid & FPSID_CPUID_ARCH_MASK;
756 		VFP_arch >>= FPSID_ARCH_BIT;
757 		/*
758 		 * Check for the presence of the Advanced SIMD
759 		 * load/store instructions, integer and single
760 		 * precision floating point operations. Only check
761 		 * for NEON if the hardware has the MVFR registers.
762 		 */
763 		if (IS_ENABLED(CONFIG_NEON) &&
764 		   (fmrx(MVFR1) & 0x000fff00) == 0x00011100)
765 			elf_hwcap |= HWCAP_NEON;
766 
767 		if (IS_ENABLED(CONFIG_VFPv3)) {
768 			u32 mvfr0 = fmrx(MVFR0);
769 			if (((mvfr0 & MVFR0_DP_MASK) >> MVFR0_DP_BIT) == 0x2 ||
770 			    ((mvfr0 & MVFR0_SP_MASK) >> MVFR0_SP_BIT) == 0x2) {
771 				elf_hwcap |= HWCAP_VFPv3;
772 				/*
773 				 * Check for VFPv3 D16 and VFPv4 D16.  CPUs in
774 				 * this configuration only have 16 x 64bit
775 				 * registers.
776 				 */
777 				if ((mvfr0 & MVFR0_A_SIMD_MASK) == 1)
778 					/* also v4-D16 */
779 					elf_hwcap |= HWCAP_VFPv3D16;
780 				else
781 					elf_hwcap |= HWCAP_VFPD32;
782 			}
783 
784 			if ((fmrx(MVFR1) & 0xf0000000) == 0x10000000)
785 				elf_hwcap |= HWCAP_VFPv4;
786 		}
787 	/* Extract the architecture version on pre-cpuid scheme */
788 	} else {
789 		if (vfpsid & FPSID_NODOUBLE) {
790 			pr_cont("no double precision support\n");
791 			return 0;
792 		}
793 
794 		VFP_arch = (vfpsid & FPSID_ARCH_MASK) >> FPSID_ARCH_BIT;
795 	}
796 
797 	cpuhp_setup_state_nocalls(CPUHP_AP_ARM_VFP_STARTING,
798 				  "arm/vfp:starting", vfp_starting_cpu,
799 				  vfp_dying_cpu);
800 
801 	vfp_vector = vfp_support_entry;
802 
803 	thread_register_notifier(&vfp_notifier_block);
804 	vfp_pm_init();
805 
806 	/*
807 	 * We detected VFP, and the support code is
808 	 * in place; report VFP support to userspace.
809 	 */
810 	elf_hwcap |= HWCAP_VFP;
811 
812 	pr_cont("implementor %02x architecture %d part %02x variant %x rev %x\n",
813 		(vfpsid & FPSID_IMPLEMENTER_MASK) >> FPSID_IMPLEMENTER_BIT,
814 		VFP_arch,
815 		(vfpsid & FPSID_PART_MASK) >> FPSID_PART_BIT,
816 		(vfpsid & FPSID_VARIANT_MASK) >> FPSID_VARIANT_BIT,
817 		(vfpsid & FPSID_REV_MASK) >> FPSID_REV_BIT);
818 
819 	return 0;
820 }
821 
822 core_initcall(vfp_init);
823