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>
25 #include <asm/cputype.h>
26 #include <asm/system_info.h>
27 #include <asm/thread_notify.h>
34 * Our undef handlers (in entry.S)
36 void vfp_testing_entry(void);
37 void vfp_support_entry(void);
38 void vfp_null_entry(void);
40 void (*vfp_vector)(void) = vfp_null_entry;
44 * Used in startup: set to non-zero if VFP checks fail
45 * After startup, holds VFP architecture
47 unsigned int VFP_arch;
50 * The pointer to the vfpstate structure of the thread which currently
51 * owns the context held in the VFP hardware, or NULL if the hardware
54 * For UP, this is sufficient to tell which thread owns the VFP context.
55 * However, for SMP, we also need to check the CPU number stored in the
56 * saved state too to catch migrations.
58 union vfp_state *vfp_current_hw_state[NR_CPUS];
61 * Is 'thread's most up to date state stored in this CPUs hardware?
62 * Must be called from non-preemptible context.
64 static bool vfp_state_in_hw(unsigned int cpu, struct thread_info *thread)
67 if (thread->vfpstate.hard.cpu != cpu)
70 return vfp_current_hw_state[cpu] == &thread->vfpstate;
74 * Force a reload of the VFP context from the thread structure. We do
75 * this by ensuring that access to the VFP hardware is disabled, and
76 * clear vfp_current_hw_state. Must be called from non-preemptible context.
78 static void vfp_force_reload(unsigned int cpu, struct thread_info *thread)
80 if (vfp_state_in_hw(cpu, thread)) {
81 fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
82 vfp_current_hw_state[cpu] = NULL;
85 thread->vfpstate.hard.cpu = NR_CPUS;
90 * Per-thread VFP initialization.
92 static void vfp_thread_flush(struct thread_info *thread)
94 union vfp_state *vfp = &thread->vfpstate;
98 * Disable VFP to ensure we initialize it first. We must ensure
99 * that the modification of vfp_current_hw_state[] and hardware
100 * disable are done for the same CPU and without preemption.
102 * Do this first to ensure that preemption won't overwrite our
103 * state saving should access to the VFP be enabled at this point.
106 if (vfp_current_hw_state[cpu] == vfp)
107 vfp_current_hw_state[cpu] = NULL;
108 fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
111 memset(vfp, 0, sizeof(union vfp_state));
113 vfp->hard.fpexc = FPEXC_EN;
114 vfp->hard.fpscr = FPSCR_ROUND_NEAREST;
116 vfp->hard.cpu = NR_CPUS;
120 static void vfp_thread_exit(struct thread_info *thread)
122 /* release case: Per-thread VFP cleanup. */
123 union vfp_state *vfp = &thread->vfpstate;
124 unsigned int cpu = get_cpu();
126 if (vfp_current_hw_state[cpu] == vfp)
127 vfp_current_hw_state[cpu] = NULL;
131 static void vfp_thread_copy(struct thread_info *thread)
133 struct thread_info *parent = current_thread_info();
135 vfp_sync_hwstate(parent);
136 thread->vfpstate = parent->vfpstate;
138 thread->vfpstate.hard.cpu = NR_CPUS;
143 * When this function is called with the following 'cmd's, the following
144 * is true while this function is being run:
145 * THREAD_NOFTIFY_SWTICH:
146 * - the previously running thread will not be scheduled onto another CPU.
147 * - the next thread to be run (v) will not be running on another CPU.
148 * - thread->cpu is the local CPU number
149 * - not preemptible as we're called in the middle of a thread switch
150 * THREAD_NOTIFY_FLUSH:
151 * - the thread (v) will be running on the local CPU, so
152 * v === current_thread_info()
153 * - thread->cpu is the local CPU number at the time it is accessed,
154 * but may change at any time.
155 * - we could be preempted if tree preempt rcu is enabled, so
156 * it is unsafe to use thread->cpu.
158 * - the thread (v) will be running on the local CPU, so
159 * v === current_thread_info()
160 * - thread->cpu is the local CPU number at the time it is accessed,
161 * but may change at any time.
162 * - we could be preempted if tree preempt rcu is enabled, so
163 * it is unsafe to use thread->cpu.
165 static int vfp_notifier(struct notifier_block *self, unsigned long cmd, void *v)
167 struct thread_info *thread = v;
174 case THREAD_NOTIFY_SWITCH:
181 * On SMP, if VFP is enabled, save the old state in
182 * case the thread migrates to a different CPU. The
183 * restoring is done lazily.
185 if ((fpexc & FPEXC_EN) && vfp_current_hw_state[cpu])
186 vfp_save_state(vfp_current_hw_state[cpu], fpexc);
190 * Always disable VFP so we can lazily save/restore the
193 fmxr(FPEXC, fpexc & ~FPEXC_EN);
196 case THREAD_NOTIFY_FLUSH:
197 vfp_thread_flush(thread);
200 case THREAD_NOTIFY_EXIT:
201 vfp_thread_exit(thread);
204 case THREAD_NOTIFY_COPY:
205 vfp_thread_copy(thread);
212 static struct notifier_block vfp_notifier_block = {
213 .notifier_call = vfp_notifier,
217 * Raise a SIGFPE for the current process.
218 * sicode describes the signal being raised.
220 static void vfp_raise_sigfpe(unsigned int sicode, struct pt_regs *regs)
224 memset(&info, 0, sizeof(info));
226 info.si_signo = SIGFPE;
227 info.si_code = sicode;
228 info.si_addr = (void __user *)(instruction_pointer(regs) - 4);
231 * This is the same as NWFPE, because it's not clear what
234 current->thread.error_code = 0;
235 current->thread.trap_no = 6;
237 send_sig_info(SIGFPE, &info, current);
240 static void vfp_panic(char *reason, u32 inst)
244 printk(KERN_ERR "VFP: Error: %s\n", reason);
245 printk(KERN_ERR "VFP: EXC 0x%08x SCR 0x%08x INST 0x%08x\n",
246 fmrx(FPEXC), fmrx(FPSCR), inst);
247 for (i = 0; i < 32; i += 2)
248 printk(KERN_ERR "VFP: s%2u: 0x%08x s%2u: 0x%08x\n",
249 i, vfp_get_float(i), i+1, vfp_get_float(i+1));
253 * Process bitmask of exception conditions.
255 static void vfp_raise_exceptions(u32 exceptions, u32 inst, u32 fpscr, struct pt_regs *regs)
259 pr_debug("VFP: raising exceptions %08x\n", exceptions);
261 if (exceptions == VFP_EXCEPTION_ERROR) {
262 vfp_panic("unhandled bounce", inst);
263 vfp_raise_sigfpe(0, regs);
268 * If any of the status flags are set, update the FPSCR.
269 * Comparison instructions always return at least one of
272 if (exceptions & (FPSCR_N|FPSCR_Z|FPSCR_C|FPSCR_V))
273 fpscr &= ~(FPSCR_N|FPSCR_Z|FPSCR_C|FPSCR_V);
279 #define RAISE(stat,en,sig) \
280 if (exceptions & stat && fpscr & en) \
284 * These are arranged in priority order, least to highest.
286 RAISE(FPSCR_DZC, FPSCR_DZE, FPE_FLTDIV);
287 RAISE(FPSCR_IXC, FPSCR_IXE, FPE_FLTRES);
288 RAISE(FPSCR_UFC, FPSCR_UFE, FPE_FLTUND);
289 RAISE(FPSCR_OFC, FPSCR_OFE, FPE_FLTOVF);
290 RAISE(FPSCR_IOC, FPSCR_IOE, FPE_FLTINV);
293 vfp_raise_sigfpe(si_code, regs);
297 * Emulate a VFP instruction.
299 static u32 vfp_emulate_instruction(u32 inst, u32 fpscr, struct pt_regs *regs)
301 u32 exceptions = VFP_EXCEPTION_ERROR;
303 pr_debug("VFP: emulate: INST=0x%08x SCR=0x%08x\n", inst, fpscr);
305 if (INST_CPRTDO(inst)) {
306 if (!INST_CPRT(inst)) {
310 if (vfp_single(inst)) {
311 exceptions = vfp_single_cpdo(inst, fpscr);
313 exceptions = vfp_double_cpdo(inst, fpscr);
317 * A CPRT instruction can not appear in FPINST2, nor
318 * can it cause an exception. Therefore, we do not
319 * have to emulate it.
324 * A CPDT instruction can not appear in FPINST2, nor can
325 * it cause an exception. Therefore, we do not have to
329 return exceptions & ~VFP_NAN_FLAG;
333 * Package up a bounce condition.
335 void VFP_bounce(u32 trigger, u32 fpexc, struct pt_regs *regs)
337 u32 fpscr, orig_fpscr, fpsid, exceptions;
339 pr_debug("VFP: bounce: trigger %08x fpexc %08x\n", trigger, fpexc);
342 * At this point, FPEXC can have the following configuration:
345 * 0 1 x - synchronous exception
346 * 1 x 0 - asynchronous exception
347 * 1 x 1 - sychronous on VFP subarch 1 and asynchronous on later
348 * 0 0 1 - synchronous on VFP9 (non-standard subarch 1
349 * implementation), undefined otherwise
351 * Clear various bits and enable access to the VFP so we can
354 fmxr(FPEXC, fpexc & ~(FPEXC_EX|FPEXC_DEX|FPEXC_FP2V|FPEXC_VV|FPEXC_TRAP_MASK));
357 orig_fpscr = fpscr = fmrx(FPSCR);
360 * Check for the special VFP subarch 1 and FPSCR.IXE bit case
362 if ((fpsid & FPSID_ARCH_MASK) == (1 << FPSID_ARCH_BIT)
363 && (fpscr & FPSCR_IXE)) {
365 * Synchronous exception, emulate the trigger instruction
370 if (fpexc & FPEXC_EX) {
371 #ifndef CONFIG_CPU_FEROCEON
373 * Asynchronous exception. The instruction is read from FPINST
374 * and the interrupted instruction has to be restarted.
376 trigger = fmrx(FPINST);
379 } else if (!(fpexc & FPEXC_DEX)) {
381 * Illegal combination of bits. It can be caused by an
382 * unallocated VFP instruction but with FPSCR.IXE set and not
385 vfp_raise_exceptions(VFP_EXCEPTION_ERROR, trigger, fpscr, regs);
390 * Modify fpscr to indicate the number of iterations remaining.
391 * If FPEXC.EX is 0, FPEXC.DEX is 1 and the FPEXC.VV bit indicates
392 * whether FPEXC.VECITR or FPSCR.LEN is used.
394 if (fpexc & (FPEXC_EX | FPEXC_VV)) {
397 len = fpexc + (1 << FPEXC_LENGTH_BIT);
399 fpscr &= ~FPSCR_LENGTH_MASK;
400 fpscr |= (len & FPEXC_LENGTH_MASK) << (FPSCR_LENGTH_BIT - FPEXC_LENGTH_BIT);
404 * Handle the first FP instruction. We used to take note of the
405 * FPEXC bounce reason, but this appears to be unreliable.
406 * Emulate the bounced instruction instead.
408 exceptions = vfp_emulate_instruction(trigger, fpscr, regs);
410 vfp_raise_exceptions(exceptions, trigger, orig_fpscr, regs);
413 * If there isn't a second FP instruction, exit now. Note that
414 * the FPEXC.FP2V bit is valid only if FPEXC.EX is 1.
416 if (fpexc ^ (FPEXC_EX | FPEXC_FP2V))
420 * The barrier() here prevents fpinst2 being read
421 * before the condition above.
424 trigger = fmrx(FPINST2);
427 exceptions = vfp_emulate_instruction(trigger, orig_fpscr, regs);
429 vfp_raise_exceptions(exceptions, trigger, orig_fpscr, regs);
434 static void vfp_enable(void *unused)
438 BUG_ON(preemptible());
439 access = get_copro_access();
442 * Enable full access to VFP (cp10 and cp11)
444 set_copro_access(access | CPACC_FULL(10) | CPACC_FULL(11));
448 static int vfp_pm_suspend(void)
450 struct thread_info *ti = current_thread_info();
451 u32 fpexc = fmrx(FPEXC);
453 /* if vfp is on, then save state for resumption */
454 if (fpexc & FPEXC_EN) {
455 printk(KERN_DEBUG "%s: saving vfp state\n", __func__);
456 vfp_save_state(&ti->vfpstate, fpexc);
458 /* disable, just in case */
459 fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
462 /* clear any information we had about last context state */
463 memset(vfp_current_hw_state, 0, sizeof(vfp_current_hw_state));
468 static void vfp_pm_resume(void)
470 /* ensure we have access to the vfp */
473 /* and disable it to ensure the next usage restores the state */
474 fmxr(FPEXC, fmrx(FPEXC) & ~FPEXC_EN);
477 static int vfp_cpu_pm_notifier(struct notifier_block *self, unsigned long cmd,
484 case CPU_PM_ENTER_FAILED:
492 static struct notifier_block vfp_cpu_pm_notifier_block = {
493 .notifier_call = vfp_cpu_pm_notifier,
496 static void vfp_pm_init(void)
498 cpu_pm_register_notifier(&vfp_cpu_pm_notifier_block);
502 static inline void vfp_pm_init(void) { }
503 #endif /* CONFIG_CPU_PM */
506 * Ensure that the VFP state stored in 'thread->vfpstate' is up to date
507 * with the hardware state.
509 void vfp_sync_hwstate(struct thread_info *thread)
511 unsigned int cpu = get_cpu();
513 if (vfp_state_in_hw(cpu, thread)) {
514 u32 fpexc = fmrx(FPEXC);
517 * Save the last VFP state on this CPU.
519 fmxr(FPEXC, fpexc | FPEXC_EN);
520 vfp_save_state(&thread->vfpstate, fpexc | FPEXC_EN);
527 /* Ensure that the thread reloads the hardware VFP state on the next use. */
528 void vfp_flush_hwstate(struct thread_info *thread)
530 unsigned int cpu = get_cpu();
532 vfp_force_reload(cpu, thread);
538 * Save the current VFP state into the provided structures and prepare
539 * for entry into a new function (signal handler).
541 int vfp_preserve_user_clear_hwstate(struct user_vfp __user *ufp,
542 struct user_vfp_exc __user *ufp_exc)
544 struct thread_info *thread = current_thread_info();
545 struct vfp_hard_struct *hwstate = &thread->vfpstate.hard;
548 /* Ensure that the saved hwstate is up-to-date. */
549 vfp_sync_hwstate(thread);
552 * Copy the floating point registers. There can be unused
553 * registers see asm/hwcap.h for details.
555 err |= __copy_to_user(&ufp->fpregs, &hwstate->fpregs,
556 sizeof(hwstate->fpregs));
558 * Copy the status and control register.
560 __put_user_error(hwstate->fpscr, &ufp->fpscr, err);
563 * Copy the exception registers.
565 __put_user_error(hwstate->fpexc, &ufp_exc->fpexc, err);
566 __put_user_error(hwstate->fpinst, &ufp_exc->fpinst, err);
567 __put_user_error(hwstate->fpinst2, &ufp_exc->fpinst2, err);
572 /* Ensure that VFP is disabled. */
573 vfp_flush_hwstate(thread);
576 * As per the PCS, clear the length and stride bits for function
579 hwstate->fpscr &= ~(FPSCR_LENGTH_MASK | FPSCR_STRIDE_MASK);
583 /* Sanitise and restore the current VFP state from the provided structures. */
584 int vfp_restore_user_hwstate(struct user_vfp __user *ufp,
585 struct user_vfp_exc __user *ufp_exc)
587 struct thread_info *thread = current_thread_info();
588 struct vfp_hard_struct *hwstate = &thread->vfpstate.hard;
592 /* Disable VFP to avoid corrupting the new thread state. */
593 vfp_flush_hwstate(thread);
596 * Copy the floating point registers. There can be unused
597 * registers see asm/hwcap.h for details.
599 err |= __copy_from_user(&hwstate->fpregs, &ufp->fpregs,
600 sizeof(hwstate->fpregs));
602 * Copy the status and control register.
604 __get_user_error(hwstate->fpscr, &ufp->fpscr, err);
607 * Sanitise and restore the exception registers.
609 __get_user_error(fpexc, &ufp_exc->fpexc, err);
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 __get_user_error(hwstate->fpinst, &ufp_exc->fpinst, err);
619 __get_user_error(hwstate->fpinst2, &ufp_exc->fpinst2, err);
621 return err ? -EFAULT : 0;
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
632 * Both CPU_DYING and CPU_STARTING are called on the CPU which
633 * is being offlined/onlined.
635 static int vfp_hotplug(struct notifier_block *b, unsigned long action,
638 if (action == CPU_DYING || action == CPU_DYING_FROZEN) {
639 vfp_force_reload((long)hcpu, current_thread_info());
640 } else if (action == CPU_STARTING || action == CPU_STARTING_FROZEN)
646 * VFP support code initialisation.
648 static int __init vfp_init(void)
651 unsigned int cpu_arch = cpu_architecture();
653 if (cpu_arch >= CPU_ARCH_ARMv6)
654 on_each_cpu(vfp_enable, NULL, 1);
657 * First check that there is a VFP that we can use.
658 * The handler is already setup to just log calls, so
659 * we just need to read the VFPSID register.
661 vfp_vector = vfp_testing_entry;
663 vfpsid = fmrx(FPSID);
665 vfp_vector = vfp_null_entry;
667 printk(KERN_INFO "VFP support v0.3: ");
669 printk("not present\n");
670 else if (vfpsid & FPSID_NODOUBLE) {
671 printk("no double precision support\n");
673 hotcpu_notifier(vfp_hotplug, 0);
675 VFP_arch = (vfpsid & FPSID_ARCH_MASK) >> FPSID_ARCH_BIT; /* Extract the architecture version */
676 printk("implementor %02x architecture %d part %02x variant %x rev %x\n",
677 (vfpsid & FPSID_IMPLEMENTER_MASK) >> FPSID_IMPLEMENTER_BIT,
678 (vfpsid & FPSID_ARCH_MASK) >> FPSID_ARCH_BIT,
679 (vfpsid & FPSID_PART_MASK) >> FPSID_PART_BIT,
680 (vfpsid & FPSID_VARIANT_MASK) >> FPSID_VARIANT_BIT,
681 (vfpsid & FPSID_REV_MASK) >> FPSID_REV_BIT);
683 vfp_vector = vfp_support_entry;
685 thread_register_notifier(&vfp_notifier_block);
689 * We detected VFP, and the support code is
690 * in place; report VFP support to userspace.
692 elf_hwcap |= HWCAP_VFP;
695 elf_hwcap |= HWCAP_VFPv3;
698 * Check for VFPv3 D16. CPUs in this configuration
699 * only have 16 x 64bit registers.
701 if (((fmrx(MVFR0) & MVFR0_A_SIMD_MASK)) == 1)
702 elf_hwcap |= HWCAP_VFPv3D16;
706 * Check for the presence of the Advanced SIMD
707 * load/store instructions, integer and single
708 * precision floating point operations. Only check
709 * for NEON if the hardware has the MVFR registers.
711 if ((read_cpuid_id() & 0x000f0000) == 0x000f0000) {
713 if ((fmrx(MVFR1) & 0x000fff00) == 0x00011100)
714 elf_hwcap |= HWCAP_NEON;
716 if ((fmrx(MVFR1) & 0xf0000000) == 0x10000000)
717 elf_hwcap |= HWCAP_VFPv4;
723 late_initcall(vfp_init);