1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/arch/arm/vfp/vfpmodule.c
5 * Copyright (C) 2004 ARM Limited.
6 * Written by Deep Blue Solutions Limited.
8 #include <linux/types.h>
10 #include <linux/cpu_pm.h>
11 #include <linux/hardirq.h>
12 #include <linux/kernel.h>
13 #include <linux/notifier.h>
14 #include <linux/signal.h>
15 #include <linux/sched/signal.h>
16 #include <linux/smp.h>
17 #include <linux/init.h>
18 #include <linux/uaccess.h>
19 #include <linux/user.h>
20 #include <linux/export.h>
23 #include <asm/cputype.h>
24 #include <asm/system_info.h>
25 #include <asm/thread_notify.h>
32 * Our undef handlers (in entry.S)
34 asmlinkage
void vfp_testing_entry(void);
35 asmlinkage
void vfp_support_entry(void);
36 asmlinkage
void vfp_null_entry(void);
38 asmlinkage
void (*vfp_vector
)(void) = vfp_null_entry
;
42 * Used in startup: set to non-zero if VFP checks fail
43 * After startup, holds VFP architecture
45 unsigned int VFP_arch
;
48 * The pointer to the vfpstate structure of the thread which currently
49 * owns the context held in the VFP hardware, or NULL if the hardware
52 * For UP, this is sufficient to tell which thread owns the VFP context.
53 * However, for SMP, we also need to check the CPU number stored in the
54 * saved state too to catch migrations.
56 union vfp_state
*vfp_current_hw_state
[NR_CPUS
];
59 * Is 'thread's most up to date state stored in this CPUs hardware?
60 * Must be called from non-preemptible context.
62 static bool vfp_state_in_hw(unsigned int cpu
, struct thread_info
*thread
)
65 if (thread
->vfpstate
.hard
.cpu
!= cpu
)
68 return vfp_current_hw_state
[cpu
] == &thread
->vfpstate
;
72 * Force a reload of the VFP context from the thread structure. We do
73 * this by ensuring that access to the VFP hardware is disabled, and
74 * clear vfp_current_hw_state. Must be called from non-preemptible context.
76 static void vfp_force_reload(unsigned int cpu
, struct thread_info
*thread
)
78 if (vfp_state_in_hw(cpu
, thread
)) {
79 fmxr(FPEXC
, fmrx(FPEXC
) & ~FPEXC_EN
);
80 vfp_current_hw_state
[cpu
] = NULL
;
83 thread
->vfpstate
.hard
.cpu
= NR_CPUS
;
88 * Per-thread VFP initialization.
90 static void vfp_thread_flush(struct thread_info
*thread
)
92 union vfp_state
*vfp
= &thread
->vfpstate
;
96 * Disable VFP to ensure we initialize it first. We must ensure
97 * that the modification of vfp_current_hw_state[] and hardware
98 * disable are done for the same CPU and without preemption.
100 * Do this first to ensure that preemption won't overwrite our
101 * state saving should access to the VFP be enabled at this point.
104 if (vfp_current_hw_state
[cpu
] == vfp
)
105 vfp_current_hw_state
[cpu
] = NULL
;
106 fmxr(FPEXC
, fmrx(FPEXC
) & ~FPEXC_EN
);
109 memset(vfp
, 0, sizeof(union vfp_state
));
111 vfp
->hard
.fpexc
= FPEXC_EN
;
112 vfp
->hard
.fpscr
= FPSCR_ROUND_NEAREST
;
114 vfp
->hard
.cpu
= NR_CPUS
;
118 static void vfp_thread_exit(struct thread_info
*thread
)
120 /* release case: Per-thread VFP cleanup. */
121 union vfp_state
*vfp
= &thread
->vfpstate
;
122 unsigned int cpu
= get_cpu();
124 if (vfp_current_hw_state
[cpu
] == vfp
)
125 vfp_current_hw_state
[cpu
] = NULL
;
129 static void vfp_thread_copy(struct thread_info
*thread
)
131 struct thread_info
*parent
= current_thread_info();
133 vfp_sync_hwstate(parent
);
134 thread
->vfpstate
= parent
->vfpstate
;
136 thread
->vfpstate
.hard
.cpu
= NR_CPUS
;
141 * When this function is called with the following 'cmd's, the following
142 * is true while this function is being run:
143 * THREAD_NOFTIFY_SWTICH:
144 * - the previously running thread will not be scheduled onto another CPU.
145 * - the next thread to be run (v) will not be running on another CPU.
146 * - thread->cpu is the local CPU number
147 * - not preemptible as we're called in the middle of a thread switch
148 * THREAD_NOTIFY_FLUSH:
149 * - the thread (v) will be running on the local CPU, so
150 * v === current_thread_info()
151 * - thread->cpu is the local CPU number at the time it is accessed,
152 * but may change at any time.
153 * - we could be preempted if tree preempt rcu is enabled, so
154 * it is unsafe to use thread->cpu.
156 * - we could be preempted if tree preempt rcu is enabled, so
157 * it is unsafe to use thread->cpu.
159 static int vfp_notifier(struct notifier_block
*self
, unsigned long cmd
, void *v
)
161 struct thread_info
*thread
= v
;
168 case THREAD_NOTIFY_SWITCH
:
175 * On SMP, if VFP is enabled, save the old state in
176 * case the thread migrates to a different CPU. The
177 * restoring is done lazily.
179 if ((fpexc
& FPEXC_EN
) && vfp_current_hw_state
[cpu
])
180 vfp_save_state(vfp_current_hw_state
[cpu
], fpexc
);
184 * Always disable VFP so we can lazily save/restore the
187 fmxr(FPEXC
, fpexc
& ~FPEXC_EN
);
190 case THREAD_NOTIFY_FLUSH
:
191 vfp_thread_flush(thread
);
194 case THREAD_NOTIFY_EXIT
:
195 vfp_thread_exit(thread
);
198 case THREAD_NOTIFY_COPY
:
199 vfp_thread_copy(thread
);
206 static struct notifier_block vfp_notifier_block
= {
207 .notifier_call
= vfp_notifier
,
211 * Raise a SIGFPE for the current process.
212 * sicode describes the signal being raised.
214 static void vfp_raise_sigfpe(unsigned int sicode
, struct pt_regs
*regs
)
217 * This is the same as NWFPE, because it's not clear what
220 current
->thread
.error_code
= 0;
221 current
->thread
.trap_no
= 6;
223 send_sig_fault(SIGFPE
, sicode
,
224 (void __user
*)(instruction_pointer(regs
) - 4),
228 static void vfp_panic(char *reason
, u32 inst
)
232 pr_err("VFP: Error: %s\n", reason
);
233 pr_err("VFP: EXC 0x%08x SCR 0x%08x INST 0x%08x\n",
234 fmrx(FPEXC
), fmrx(FPSCR
), inst
);
235 for (i
= 0; i
< 32; i
+= 2)
236 pr_err("VFP: s%2u: 0x%08x s%2u: 0x%08x\n",
237 i
, vfp_get_float(i
), i
+1, vfp_get_float(i
+1));
241 * Process bitmask of exception conditions.
243 static void vfp_raise_exceptions(u32 exceptions
, u32 inst
, u32 fpscr
, struct pt_regs
*regs
)
247 pr_debug("VFP: raising exceptions %08x\n", exceptions
);
249 if (exceptions
== VFP_EXCEPTION_ERROR
) {
250 vfp_panic("unhandled bounce", inst
);
251 vfp_raise_sigfpe(FPE_FLTINV
, regs
);
256 * If any of the status flags are set, update the FPSCR.
257 * Comparison instructions always return at least one of
260 if (exceptions
& (FPSCR_N
|FPSCR_Z
|FPSCR_C
|FPSCR_V
))
261 fpscr
&= ~(FPSCR_N
|FPSCR_Z
|FPSCR_C
|FPSCR_V
);
267 #define RAISE(stat,en,sig) \
268 if (exceptions & stat && fpscr & en) \
272 * These are arranged in priority order, least to highest.
274 RAISE(FPSCR_DZC
, FPSCR_DZE
, FPE_FLTDIV
);
275 RAISE(FPSCR_IXC
, FPSCR_IXE
, FPE_FLTRES
);
276 RAISE(FPSCR_UFC
, FPSCR_UFE
, FPE_FLTUND
);
277 RAISE(FPSCR_OFC
, FPSCR_OFE
, FPE_FLTOVF
);
278 RAISE(FPSCR_IOC
, FPSCR_IOE
, FPE_FLTINV
);
281 vfp_raise_sigfpe(si_code
, regs
);
285 * Emulate a VFP instruction.
287 static u32
vfp_emulate_instruction(u32 inst
, u32 fpscr
, struct pt_regs
*regs
)
289 u32 exceptions
= VFP_EXCEPTION_ERROR
;
291 pr_debug("VFP: emulate: INST=0x%08x SCR=0x%08x\n", inst
, fpscr
);
293 if (INST_CPRTDO(inst
)) {
294 if (!INST_CPRT(inst
)) {
298 if (vfp_single(inst
)) {
299 exceptions
= vfp_single_cpdo(inst
, fpscr
);
301 exceptions
= vfp_double_cpdo(inst
, fpscr
);
305 * A CPRT instruction can not appear in FPINST2, nor
306 * can it cause an exception. Therefore, we do not
307 * have to emulate it.
312 * A CPDT instruction can not appear in FPINST2, nor can
313 * it cause an exception. Therefore, we do not have to
317 return exceptions
& ~VFP_NAN_FLAG
;
321 * Package up a bounce condition.
323 void VFP_bounce(u32 trigger
, u32 fpexc
, struct pt_regs
*regs
)
325 u32 fpscr
, orig_fpscr
, fpsid
, exceptions
;
327 pr_debug("VFP: bounce: trigger %08x fpexc %08x\n", trigger
, fpexc
);
330 * At this point, FPEXC can have the following configuration:
333 * 0 1 x - synchronous exception
334 * 1 x 0 - asynchronous exception
335 * 1 x 1 - sychronous on VFP subarch 1 and asynchronous on later
336 * 0 0 1 - synchronous on VFP9 (non-standard subarch 1
337 * implementation), undefined otherwise
339 * Clear various bits and enable access to the VFP so we can
342 fmxr(FPEXC
, fpexc
& ~(FPEXC_EX
|FPEXC_DEX
|FPEXC_FP2V
|FPEXC_VV
|FPEXC_TRAP_MASK
));
345 orig_fpscr
= fpscr
= fmrx(FPSCR
);
348 * Check for the special VFP subarch 1 and FPSCR.IXE bit case
350 if ((fpsid
& FPSID_ARCH_MASK
) == (1 << FPSID_ARCH_BIT
)
351 && (fpscr
& FPSCR_IXE
)) {
353 * Synchronous exception, emulate the trigger instruction
358 if (fpexc
& FPEXC_EX
) {
359 #ifndef CONFIG_CPU_FEROCEON
361 * Asynchronous exception. The instruction is read from FPINST
362 * and the interrupted instruction has to be restarted.
364 trigger
= fmrx(FPINST
);
367 } else if (!(fpexc
& FPEXC_DEX
)) {
369 * Illegal combination of bits. It can be caused by an
370 * unallocated VFP instruction but with FPSCR.IXE set and not
373 vfp_raise_exceptions(VFP_EXCEPTION_ERROR
, trigger
, fpscr
, regs
);
378 * Modify fpscr to indicate the number of iterations remaining.
379 * If FPEXC.EX is 0, FPEXC.DEX is 1 and the FPEXC.VV bit indicates
380 * whether FPEXC.VECITR or FPSCR.LEN is used.
382 if (fpexc
& (FPEXC_EX
| FPEXC_VV
)) {
385 len
= fpexc
+ (1 << FPEXC_LENGTH_BIT
);
387 fpscr
&= ~FPSCR_LENGTH_MASK
;
388 fpscr
|= (len
& FPEXC_LENGTH_MASK
) << (FPSCR_LENGTH_BIT
- FPEXC_LENGTH_BIT
);
392 * Handle the first FP instruction. We used to take note of the
393 * FPEXC bounce reason, but this appears to be unreliable.
394 * Emulate the bounced instruction instead.
396 exceptions
= vfp_emulate_instruction(trigger
, fpscr
, regs
);
398 vfp_raise_exceptions(exceptions
, trigger
, orig_fpscr
, regs
);
401 * If there isn't a second FP instruction, exit now. Note that
402 * the FPEXC.FP2V bit is valid only if FPEXC.EX is 1.
404 if ((fpexc
& (FPEXC_EX
| FPEXC_FP2V
)) != (FPEXC_EX
| FPEXC_FP2V
))
408 * The barrier() here prevents fpinst2 being read
409 * before the condition above.
412 trigger
= fmrx(FPINST2
);
415 exceptions
= vfp_emulate_instruction(trigger
, orig_fpscr
, regs
);
417 vfp_raise_exceptions(exceptions
, trigger
, orig_fpscr
, regs
);
422 static void vfp_enable(void *unused
)
426 BUG_ON(preemptible());
427 access
= get_copro_access();
430 * Enable full access to VFP (cp10 and cp11)
432 set_copro_access(access
| CPACC_FULL(10) | CPACC_FULL(11));
435 /* Called by platforms on which we want to disable VFP because it may not be
436 * present on all CPUs within a SMP complex. Needs to be called prior to
439 void vfp_disable(void)
442 pr_debug("%s: should be called prior to vfp_init\n", __func__
);
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
*ufp
,
549 struct user_vfp_exc
*ufp_exc
)
551 struct thread_info
*thread
= current_thread_info();
552 struct vfp_hard_struct
*hwstate
= &thread
->vfpstate
.hard
;
554 /* Ensure that the saved hwstate is up-to-date. */
555 vfp_sync_hwstate(thread
);
558 * Copy the floating point registers. There can be unused
559 * registers see asm/hwcap.h for details.
561 memcpy(&ufp
->fpregs
, &hwstate
->fpregs
, sizeof(hwstate
->fpregs
));
564 * Copy the status and control register.
566 ufp
->fpscr
= hwstate
->fpscr
;
569 * Copy the exception registers.
571 ufp_exc
->fpexc
= hwstate
->fpexc
;
572 ufp_exc
->fpinst
= hwstate
->fpinst
;
573 ufp_exc
->fpinst2
= hwstate
->fpinst2
;
575 /* Ensure that VFP is disabled. */
576 vfp_flush_hwstate(thread
);
579 * As per the PCS, clear the length and stride bits for function
582 hwstate
->fpscr
&= ~(FPSCR_LENGTH_MASK
| FPSCR_STRIDE_MASK
);
586 /* Sanitise and restore the current VFP state from the provided structures. */
587 int vfp_restore_user_hwstate(struct user_vfp
*ufp
, struct user_vfp_exc
*ufp_exc
)
589 struct thread_info
*thread
= current_thread_info();
590 struct vfp_hard_struct
*hwstate
= &thread
->vfpstate
.hard
;
593 /* Disable VFP to avoid corrupting the new thread state. */
594 vfp_flush_hwstate(thread
);
597 * Copy the floating point registers. There can be unused
598 * registers see asm/hwcap.h for details.
600 memcpy(&hwstate
->fpregs
, &ufp
->fpregs
, sizeof(hwstate
->fpregs
));
602 * Copy the status and control register.
604 hwstate
->fpscr
= ufp
->fpscr
;
607 * Sanitise and restore the exception registers.
609 fpexc
= ufp_exc
->fpexc
;
611 /* Ensure the VFP is enabled. */
614 /* Ensure FPINST2 is invalid and the exception flag is cleared. */
615 fpexc
&= ~(FPEXC_EX
| FPEXC_FP2V
);
616 hwstate
->fpexc
= fpexc
;
618 hwstate
->fpinst
= ufp_exc
->fpinst
;
619 hwstate
->fpinst2
= ufp_exc
->fpinst2
;
625 * VFP hardware can lose all context when a CPU goes offline.
626 * As we will be running in SMP mode with CPU hotplug, we will save the
627 * hardware state at every thread switch. We clear our held state when
628 * a CPU has been killed, indicating that the VFP hardware doesn't contain
629 * a threads VFP state. When a CPU starts up, we re-enable access to the
630 * VFP hardware. The callbacks below are called on the CPU which
631 * is being offlined/onlined.
633 static int vfp_dying_cpu(unsigned int cpu
)
635 vfp_current_hw_state
[cpu
] = NULL
;
639 static int vfp_starting_cpu(unsigned int unused
)
645 void vfp_kmode_exception(void)
648 * If we reach this point, a floating point exception has been raised
649 * while running in kernel mode. If the NEON/VFP unit was enabled at the
650 * time, it means a VFP instruction has been issued that requires
651 * software assistance to complete, something which is not currently
652 * supported in kernel mode.
653 * If the NEON/VFP unit was disabled, and the location pointed to below
654 * is properly preceded by a call to kernel_neon_begin(), something has
655 * caused the task to be scheduled out and back in again. In this case,
656 * rebuilding and running with CONFIG_DEBUG_ATOMIC_SLEEP enabled should
657 * be helpful in localizing the problem.
659 if (fmrx(FPEXC
) & FPEXC_EN
)
660 pr_crit("BUG: unsupported FP instruction in kernel mode\n");
662 pr_crit("BUG: FP instruction issued in kernel mode with FP unit disabled\n");
665 #ifdef CONFIG_KERNEL_MODE_NEON
668 * Kernel-side NEON support functions
670 void kernel_neon_begin(void)
672 struct thread_info
*thread
= current_thread_info();
677 * Kernel mode NEON is only allowed outside of interrupt context
678 * with preemption disabled. This will make sure that the kernel
679 * mode NEON register contents never need to be preserved.
681 BUG_ON(in_interrupt());
684 fpexc
= fmrx(FPEXC
) | FPEXC_EN
;
688 * Save the userland NEON/VFP state. Under UP,
689 * the owner could be a task other than 'current'
691 if (vfp_state_in_hw(cpu
, thread
))
692 vfp_save_state(&thread
->vfpstate
, fpexc
);
694 else if (vfp_current_hw_state
[cpu
] != NULL
)
695 vfp_save_state(vfp_current_hw_state
[cpu
], fpexc
);
697 vfp_current_hw_state
[cpu
] = NULL
;
699 EXPORT_SYMBOL(kernel_neon_begin
);
701 void kernel_neon_end(void)
703 /* Disable the NEON/VFP unit. */
704 fmxr(FPEXC
, fmrx(FPEXC
) & ~FPEXC_EN
);
707 EXPORT_SYMBOL(kernel_neon_end
);
709 #endif /* CONFIG_KERNEL_MODE_NEON */
712 * VFP support code initialisation.
714 static int __init
vfp_init(void)
717 unsigned int cpu_arch
= cpu_architecture();
720 * Enable the access to the VFP on all online CPUs so the
721 * following test on FPSID will succeed.
723 if (cpu_arch
>= CPU_ARCH_ARMv6
)
724 on_each_cpu(vfp_enable
, NULL
, 1);
727 * First check that there is a VFP that we can use.
728 * The handler is already setup to just log calls, so
729 * we just need to read the VFPSID register.
731 vfp_vector
= vfp_testing_entry
;
733 vfpsid
= fmrx(FPSID
);
735 vfp_vector
= vfp_null_entry
;
737 pr_info("VFP support v0.3: ");
739 pr_cont("not present\n");
741 /* Extract the architecture on CPUID scheme */
742 } else if ((read_cpuid_id() & 0x000f0000) == 0x000f0000) {
743 VFP_arch
= vfpsid
& FPSID_CPUID_ARCH_MASK
;
744 VFP_arch
>>= FPSID_ARCH_BIT
;
746 * Check for the presence of the Advanced SIMD
747 * load/store instructions, integer and single
748 * precision floating point operations. Only check
749 * for NEON if the hardware has the MVFR registers.
751 if (IS_ENABLED(CONFIG_NEON
) &&
752 (fmrx(MVFR1
) & 0x000fff00) == 0x00011100)
753 elf_hwcap
|= HWCAP_NEON
;
755 if (IS_ENABLED(CONFIG_VFPv3
)) {
756 u32 mvfr0
= fmrx(MVFR0
);
757 if (((mvfr0
& MVFR0_DP_MASK
) >> MVFR0_DP_BIT
) == 0x2 ||
758 ((mvfr0
& MVFR0_SP_MASK
) >> MVFR0_SP_BIT
) == 0x2) {
759 elf_hwcap
|= HWCAP_VFPv3
;
761 * Check for VFPv3 D16 and VFPv4 D16. CPUs in
762 * this configuration only have 16 x 64bit
765 if ((mvfr0
& MVFR0_A_SIMD_MASK
) == 1)
767 elf_hwcap
|= HWCAP_VFPv3D16
;
769 elf_hwcap
|= HWCAP_VFPD32
;
772 if ((fmrx(MVFR1
) & 0xf0000000) == 0x10000000)
773 elf_hwcap
|= HWCAP_VFPv4
;
775 /* Extract the architecture version on pre-cpuid scheme */
777 if (vfpsid
& FPSID_NODOUBLE
) {
778 pr_cont("no double precision support\n");
782 VFP_arch
= (vfpsid
& FPSID_ARCH_MASK
) >> FPSID_ARCH_BIT
;
785 cpuhp_setup_state_nocalls(CPUHP_AP_ARM_VFP_STARTING
,
786 "arm/vfp:starting", vfp_starting_cpu
,
789 vfp_vector
= vfp_support_entry
;
791 thread_register_notifier(&vfp_notifier_block
);
795 * We detected VFP, and the support code is
796 * in place; report VFP support to userspace.
798 elf_hwcap
|= HWCAP_VFP
;
800 pr_cont("implementor %02x architecture %d part %02x variant %x rev %x\n",
801 (vfpsid
& FPSID_IMPLEMENTER_MASK
) >> FPSID_IMPLEMENTER_BIT
,
803 (vfpsid
& FPSID_PART_MASK
) >> FPSID_PART_BIT
,
804 (vfpsid
& FPSID_VARIANT_MASK
) >> FPSID_VARIANT_BIT
,
805 (vfpsid
& FPSID_REV_MASK
) >> FPSID_REV_BIT
);
810 core_initcall(vfp_init
);