Home | History | Annotate | Line # | Download | only in vfp
vfp_init.c revision 1.64
      1 /*      $NetBSD: vfp_init.c,v 1.64 2019/10/29 16:18:23 joerg Exp $ */
      2 
      3 /*
      4  * Copyright (c) 2008 ARM Ltd
      5  * All rights reserved.
      6  *
      7  * Redistribution and use in source and binary forms, with or without
      8  * modification, are permitted provided that the following conditions
      9  * are met:
     10  * 1. Redistributions of source code must retain the above copyright
     11  *    notice, this list of conditions and the following disclaimer.
     12  * 2. Redistributions in binary form must reproduce the above copyright
     13  *    notice, this list of conditions and the following disclaimer in the
     14  *    documentation and/or other materials provided with the distribution.
     15  * 3. The name of the company may not be used to endorse or promote
     16  *    products derived from this software without specific prior written
     17  *    permission.
     18  *
     19  * THIS SOFTWARE IS PROVIDED BY ARM LTD ``AS IS'' AND ANY EXPRESS OR
     20  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
     21  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
     22  * ARE DISCLAIMED.  IN NO EVENT SHALL ARM LTD BE LIABLE FOR ANY
     23  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE
     25  * GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
     27  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
     28  * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
     29  * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
     30  */
     31 
     32 #include "opt_cputypes.h"
     33 
     34 #include <sys/cdefs.h>
     35 __KERNEL_RCSID(0, "$NetBSD: vfp_init.c,v 1.64 2019/10/29 16:18:23 joerg Exp $");
     36 
     37 #include <sys/param.h>
     38 #include <sys/types.h>
     39 #include <sys/systm.h>
     40 #include <sys/device.h>
     41 #include <sys/proc.h>
     42 #include <sys/cpu.h>
     43 
     44 #include <arm/locore.h>
     45 #include <arm/pcb.h>
     46 #include <arm/undefined.h>
     47 #include <arm/vfpreg.h>
     48 #include <arm/mcontext.h>
     49 
     50 #include <uvm/uvm_extern.h>		/* for pmap.h */
     51 
     52 #ifdef FPU_VFP
     53 
     54 #ifdef CPU_CORTEX
     55 #define SETFPU __asm(".fpu\tvfpv4")
     56 #else
     57 #define SETFPU __asm(".fpu\tvfp")
     58 #endif
     59 SETFPU;
     60 
     61 /* FLDMD <X>, {d0-d15} */
     62 static inline void
     63 load_vfpregs_lo(const uint64_t *p)
     64 {
     65 	SETFPU;
     66 	__asm __volatile(".fpu vfp\n vldmia\t%0, {d0-d15}" :: "r" (p) : "memory");
     67 }
     68 
     69 /* FSTMD <X>, {d0-d15} */
     70 static inline void
     71 save_vfpregs_lo(uint64_t *p)
     72 {
     73 	SETFPU;
     74 	__asm __volatile(".fpu vfp\n vstmia\t%0, {d0-d15}" :: "r" (p) : "memory");
     75 }
     76 
     77 #ifdef CPU_CORTEX
     78 /* FLDMD <X>, {d16-d31} */
     79 static inline void
     80 load_vfpregs_hi(const uint64_t *p)
     81 {
     82 	SETFPU;
     83 	__asm __volatile(".fpu neon-vfpv4\n vldmia\t%0, {d16-d31}" :: "r" (&p[16]) : "memory");
     84 }
     85 
     86 /* FLDMD <X>, {d16-d31} */
     87 static inline void
     88 save_vfpregs_hi(uint64_t *p)
     89 {
     90 	SETFPU;
     91 	__asm __volatile(".fpu neon-vfpv4\nvstmia\t%0, {d16-d31}" :: "r" (&p[16]) : "memory");
     92 }
     93 #endif
     94 
     95 static inline void
     96 load_vfpregs(const struct vfpreg *fregs)
     97 {
     98 	load_vfpregs_lo(fregs->vfp_regs);
     99 #ifdef CPU_CORTEX
    100 #ifdef CPU_ARM11
    101 	switch (curcpu()->ci_vfp_id) {
    102 	case FPU_VFP_CORTEXA5:
    103 	case FPU_VFP_CORTEXA7:
    104 	case FPU_VFP_CORTEXA8:
    105 	case FPU_VFP_CORTEXA9:
    106 	case FPU_VFP_CORTEXA15:
    107 	case FPU_VFP_CORTEXA15_QEMU:
    108 	case FPU_VFP_CORTEXA53:
    109 	case FPU_VFP_CORTEXA57:
    110 #endif
    111 		load_vfpregs_hi(fregs->vfp_regs);
    112 #ifdef CPU_ARM11
    113 		break;
    114 	}
    115 #endif
    116 #endif
    117 }
    118 
    119 static inline void
    120 save_vfpregs(struct vfpreg *fregs)
    121 {
    122 	save_vfpregs_lo(fregs->vfp_regs);
    123 #ifdef CPU_CORTEX
    124 #ifdef CPU_ARM11
    125 	switch (curcpu()->ci_vfp_id) {
    126 	case FPU_VFP_CORTEXA5:
    127 	case FPU_VFP_CORTEXA7:
    128 	case FPU_VFP_CORTEXA8:
    129 	case FPU_VFP_CORTEXA9:
    130 	case FPU_VFP_CORTEXA15:
    131 	case FPU_VFP_CORTEXA15_QEMU:
    132 	case FPU_VFP_CORTEXA53:
    133 	case FPU_VFP_CORTEXA57:
    134 #endif
    135 		save_vfpregs_hi(fregs->vfp_regs);
    136 #ifdef CPU_ARM11
    137 		break;
    138 	}
    139 #endif
    140 #endif
    141 }
    142 
    143 /* The real handler for VFP bounces.  */
    144 static int vfp_handler(u_int, u_int, trapframe_t *, int);
    145 #ifdef CPU_CORTEX
    146 static int neon_handler(u_int, u_int, trapframe_t *, int);
    147 #endif
    148 
    149 static void vfp_state_load(lwp_t *, u_int);
    150 static void vfp_state_save(lwp_t *);
    151 static void vfp_state_release(lwp_t *);
    152 
    153 const pcu_ops_t arm_vfp_ops = {
    154 	.pcu_id = PCU_FPU,
    155 	.pcu_state_save = vfp_state_save,
    156 	.pcu_state_load = vfp_state_load,
    157 	.pcu_state_release = vfp_state_release,
    158 };
    159 
    160 /* determine what bits can be changed */
    161 uint32_t vfp_fpscr_changable = VFP_FPSCR_CSUM;
    162 /* default to run fast */
    163 uint32_t vfp_fpscr_default = (VFP_FPSCR_DN | VFP_FPSCR_FZ | VFP_FPSCR_RN);
    164 
    165 /*
    166  * Used to test for a VFP. The following function is installed as a coproc10
    167  * handler on the undefined instruction vector and then we issue a VFP
    168  * instruction. If undefined_test is non zero then the VFP did not handle
    169  * the instruction so must be absent, or disabled.
    170  */
    171 
    172 static int undefined_test;
    173 
    174 static int
    175 vfp_test(u_int address, u_int insn, trapframe_t *frame, int fault_code)
    176 {
    177 
    178 	frame->tf_pc += INSN_SIZE;
    179 	++undefined_test;
    180 	return 0;
    181 }
    182 
    183 #else
    184 /* determine what bits can be changed */
    185 uint32_t vfp_fpscr_changable = VFP_FPSCR_CSUM|VFP_FPSCR_ESUM|VFP_FPSCR_RMODE;
    186 #endif /* FPU_VFP */
    187 
    188 static int
    189 vfp_fpscr_handler(u_int address, u_int insn, trapframe_t *frame, int fault_code)
    190 {
    191 	struct lwp * const l = curlwp;
    192 	const u_int regno = (insn >> 12) & 0xf;
    193 	/*
    194 	 * Only match move to/from the FPSCR register and we
    195 	 * can't be using the SP,LR,PC as a source.
    196 	 */
    197 	if ((insn & 0xffef0fff) != 0xeee10a10 || regno > 12)
    198 		return 1;
    199 
    200 	struct pcb * const pcb = lwp_getpcb(l);
    201 
    202 #ifdef FPU_VFP
    203 	/*
    204 	 * If FPU is valid somewhere, let's just reenable VFP and
    205 	 * retry the instruction (only safe thing to do since the
    206 	 * pcb has a stale copy).
    207 	 */
    208 	if (pcb->pcb_vfp.vfp_fpexc & VFP_FPEXC_EN)
    209 		return 1;
    210 
    211 	if (__predict_false(!vfp_used_p(l))) {
    212 		pcb->pcb_vfp.vfp_fpscr = vfp_fpscr_default;
    213 	}
    214 #endif
    215 
    216 	/*
    217 	 * We now know the pcb has the saved copy.
    218 	 */
    219 	register_t * const regp = &frame->tf_r0 + regno;
    220 	if (insn & 0x00100000) {
    221 		*regp = pcb->pcb_vfp.vfp_fpscr;
    222 	} else {
    223 		pcb->pcb_vfp.vfp_fpscr &= ~vfp_fpscr_changable;
    224 		pcb->pcb_vfp.vfp_fpscr |= *regp & vfp_fpscr_changable;
    225 	}
    226 
    227 	curcpu()->ci_vfp_evs[0].ev_count++;
    228 
    229 	frame->tf_pc += INSN_SIZE;
    230 	return 0;
    231 }
    232 
    233 #ifndef FPU_VFP
    234 /*
    235  * If we don't want VFP support, we still need to handle emulating VFP FPSCR
    236  * instructions.
    237  */
    238 void
    239 vfp_attach(struct cpu_info *ci)
    240 {
    241 	if (CPU_IS_PRIMARY(ci)) {
    242 		install_coproc_handler(VFP_COPROC, vfp_fpscr_handler);
    243 	}
    244 	evcnt_attach_dynamic(&ci->ci_vfp_evs[0], EVCNT_TYPE_TRAP, NULL,
    245 	    ci->ci_cpuname, "vfp fpscr traps");
    246 }
    247 
    248 #else
    249 void
    250 vfp_attach(struct cpu_info *ci)
    251 {
    252 	const char *model = NULL;
    253 
    254 	if (CPU_ID_ARM11_P(ci->ci_arm_cpuid)
    255 	    || CPU_ID_MV88SV58XX_P(ci->ci_arm_cpuid)
    256 	    || CPU_ID_CORTEX_P(ci->ci_arm_cpuid)) {
    257 #if 0
    258 		const uint32_t nsacr = armreg_nsacr_read();
    259 		const uint32_t nsacr_vfp = __BITS(VFP_COPROC,VFP_COPROC2);
    260 		if ((nsacr & nsacr_vfp) != nsacr_vfp) {
    261 			aprint_normal_dev(ci->ci_dev,
    262 			    "VFP access denied (NSACR=%#x)\n", nsacr);
    263 			if (CPU_IS_PRIMARY(ci))
    264 				install_coproc_handler(VFP_COPROC, vfp_fpscr_handler);
    265 			ci->ci_vfp_id = 0;
    266 			evcnt_attach_dynamic(&ci->ci_vfp_evs[0],
    267 			    EVCNT_TYPE_TRAP, NULL, ci->ci_cpuname,
    268 			    "vfp fpscr traps");
    269 			return;
    270 		}
    271 #endif
    272 		const uint32_t cpacr_vfp = CPACR_CPn(VFP_COPROC);
    273 		const uint32_t cpacr_vfp2 = CPACR_CPn(VFP_COPROC2);
    274 
    275 		/*
    276 		 * We first need to enable access to the coprocessors.
    277 		 */
    278 		uint32_t cpacr = armreg_cpacr_read();
    279 		cpacr |= __SHIFTIN(CPACR_ALL, cpacr_vfp);
    280 		cpacr |= __SHIFTIN(CPACR_ALL, cpacr_vfp2);
    281 		armreg_cpacr_write(cpacr);
    282 
    283 		arm_isb();
    284 
    285 		/*
    286 		 * If we could enable them, then they exist.
    287 		 */
    288 		cpacr = armreg_cpacr_read();
    289 		bool vfp_p = __SHIFTOUT(cpacr, cpacr_vfp2) == CPACR_ALL
    290 		    && __SHIFTOUT(cpacr, cpacr_vfp) == CPACR_ALL;
    291 		if (!vfp_p) {
    292 			aprint_normal_dev(ci->ci_dev,
    293 			    "VFP access denied (CPACR=%#x)\n", cpacr);
    294 			if (CPU_IS_PRIMARY(ci))
    295 				install_coproc_handler(VFP_COPROC, vfp_fpscr_handler);
    296 			ci->ci_vfp_id = 0;
    297 			evcnt_attach_dynamic(&ci->ci_vfp_evs[0],
    298 			    EVCNT_TYPE_TRAP, NULL, ci->ci_cpuname,
    299 			    "vfp fpscr traps");
    300 			return;
    301 		}
    302 	}
    303 
    304 	void *uh = install_coproc_handler(VFP_COPROC, vfp_test);
    305 
    306 	undefined_test = 0;
    307 
    308 	const uint32_t fpsid = armreg_fpsid_read();
    309 
    310 	remove_coproc_handler(uh);
    311 
    312 	if (undefined_test != 0) {
    313 		aprint_normal_dev(ci->ci_dev, "No VFP detected\n");
    314 		if (CPU_IS_PRIMARY(ci))
    315 			install_coproc_handler(VFP_COPROC, vfp_fpscr_handler);
    316 		ci->ci_vfp_id = 0;
    317 		return;
    318 	}
    319 
    320 	ci->ci_vfp_id = fpsid;
    321 	switch (fpsid & ~ VFP_FPSID_REV_MSK) {
    322 	case FPU_VFP10_ARM10E:
    323 		model = "VFP10 R1";
    324 		break;
    325 	case FPU_VFP11_ARM11:
    326 		model = "VFP11";
    327 		break;
    328 	case FPU_VFP_MV88SV58XX:
    329 		model = "VFP3";
    330 		break;
    331 	case FPU_VFP_CORTEXA5:
    332 	case FPU_VFP_CORTEXA7:
    333 	case FPU_VFP_CORTEXA8:
    334 	case FPU_VFP_CORTEXA9:
    335 	case FPU_VFP_CORTEXA12:
    336 	case FPU_VFP_CORTEXA15:
    337 	case FPU_VFP_CORTEXA15_QEMU:
    338 	case FPU_VFP_CORTEXA17:
    339 	case FPU_VFP_CORTEXA53:
    340 	case FPU_VFP_CORTEXA57:
    341 		if (armreg_cpacr_read() & CPACR_V7_ASEDIS) {
    342 			model = "VFP 4.0+";
    343 		} else {
    344 			model = "NEON MPE (VFP 3.0+)";
    345 			cpu_neon_present = 1;
    346 		}
    347 		break;
    348 	default:
    349 		aprint_normal_dev(ci->ci_dev, "unrecognized VFP version %#x\n",
    350 		    fpsid);
    351 		if (CPU_IS_PRIMARY(ci))
    352 			install_coproc_handler(VFP_COPROC, vfp_fpscr_handler);
    353 		vfp_fpscr_changable = VFP_FPSCR_CSUM|VFP_FPSCR_ESUM
    354 		    |VFP_FPSCR_RMODE;
    355 		vfp_fpscr_default = 0;
    356 		return;
    357 	}
    358 
    359 	cpu_fpu_present = 1;
    360 	cpu_media_and_vfp_features[0] = armreg_mvfr0_read();
    361 	cpu_media_and_vfp_features[1] = armreg_mvfr1_read();
    362 	if (fpsid != 0) {
    363 		uint32_t f0 = armreg_mvfr0_read();
    364 		uint32_t f1 = armreg_mvfr1_read();
    365 		aprint_normal("vfp%d at %s: %s%s%s%s%s\n",
    366 		    device_unit(ci->ci_dev),
    367 		    device_xname(ci->ci_dev),
    368 		    model,
    369 		    ((f0 & ARM_MVFR0_ROUNDING_MASK) ? ", rounding" : ""),
    370 		    ((f0 & ARM_MVFR0_EXCEPT_MASK) ? ", exceptions" : ""),
    371 		    ((f1 & ARM_MVFR1_D_NAN_MASK) ? ", NaN propagation" : ""),
    372 		    ((f1 & ARM_MVFR1_FTZ_MASK) ? ", denormals" : ""));
    373 		aprint_debug("vfp%d: mvfr: [0]=%#x [1]=%#x\n",
    374 		    device_unit(ci->ci_dev), f0, f1);
    375 		if (CPU_IS_PRIMARY(ci)) {
    376 			if (f0 & ARM_MVFR0_ROUNDING_MASK) {
    377 				vfp_fpscr_changable |= VFP_FPSCR_RMODE;
    378 			}
    379 			if (f1 & ARM_MVFR0_EXCEPT_MASK) {
    380 				vfp_fpscr_changable |= VFP_FPSCR_ESUM;
    381 			}
    382 			// If hardware supports propagation of NaNs, select it.
    383 			if (f1 & ARM_MVFR1_D_NAN_MASK) {
    384 				vfp_fpscr_default &= ~VFP_FPSCR_DN;
    385 				vfp_fpscr_changable |= VFP_FPSCR_DN;
    386 			}
    387 			// If hardware supports denormalized numbers, use it.
    388 			if (cpu_media_and_vfp_features[1] & ARM_MVFR1_FTZ_MASK) {
    389 				vfp_fpscr_default &= ~VFP_FPSCR_FZ;
    390 				vfp_fpscr_changable |= VFP_FPSCR_FZ;
    391 			}
    392 		}
    393 	}
    394 	evcnt_attach_dynamic(&ci->ci_vfp_evs[0], EVCNT_TYPE_MISC, NULL,
    395 	    ci->ci_cpuname, "vfp coproc use");
    396 	evcnt_attach_dynamic(&ci->ci_vfp_evs[1], EVCNT_TYPE_MISC, NULL,
    397 	    ci->ci_cpuname, "vfp coproc re-use");
    398 	evcnt_attach_dynamic(&ci->ci_vfp_evs[2], EVCNT_TYPE_TRAP, NULL,
    399 	    ci->ci_cpuname, "vfp coproc fault");
    400 	if (CPU_IS_PRIMARY(ci)) {
    401 		install_coproc_handler(VFP_COPROC, vfp_handler);
    402 		install_coproc_handler(VFP_COPROC2, vfp_handler);
    403 #ifdef CPU_CORTEX
    404 		if (cpu_neon_present)
    405 			install_coproc_handler(CORE_UNKNOWN_HANDLER, neon_handler);
    406 #endif
    407 	}
    408 }
    409 
    410 /* The real handler for VFP bounces.  */
    411 static int
    412 vfp_handler(u_int address, u_int insn, trapframe_t *frame, int fault_code)
    413 {
    414 	struct cpu_info * const ci = curcpu();
    415 
    416 	/* This shouldn't ever happen.  */
    417 	if (fault_code != FAULT_USER)
    418 		panic("VFP fault at %#x in non-user mode", frame->tf_pc);
    419 
    420 	if (ci->ci_vfp_id == 0) {
    421 		/* No VFP detected, just fault.  */
    422 		return 1;
    423 	}
    424 
    425 	/*
    426 	 * If we already own the FPU and it's enabled (and no exception), raise
    427 	 * SIGILL.  If there is an exception, drop through to raise a SIGFPE.
    428 	 */
    429 	if (curcpu()->ci_pcu_curlwp[PCU_FPU] == curlwp
    430 	    && (armreg_fpexc_read() & (VFP_FPEXC_EX|VFP_FPEXC_EN)) == VFP_FPEXC_EN)
    431 		return 1;
    432 
    433 	/*
    434 	 * Make sure we own the FP.
    435 	 */
    436 	pcu_load(&arm_vfp_ops);
    437 
    438 	uint32_t fpexc = armreg_fpexc_read();
    439 	if (fpexc & VFP_FPEXC_EX) {
    440 		ksiginfo_t ksi;
    441 		KASSERT(fpexc & VFP_FPEXC_EN);
    442 
    443 		curcpu()->ci_vfp_evs[2].ev_count++;
    444 
    445 		/*
    446 		 * Need the clear the exception condition so any signal
    447 		 * and future use can proceed.
    448 		 */
    449 		armreg_fpexc_write(fpexc & ~(VFP_FPEXC_EX|VFP_FPEXC_FSUM));
    450 
    451 		pcu_save(&arm_vfp_ops, curlwp);
    452 
    453 		/*
    454 		 * XXX Need to emulate bounce instructions here to get correct
    455 		 * XXX exception codes, etc.
    456 		 */
    457 		KSI_INIT_TRAP(&ksi);
    458 		ksi.ksi_signo = SIGFPE;
    459 		if (fpexc & VFP_FPEXC_IXF)
    460 			ksi.ksi_code = FPE_FLTRES;
    461 		else if (fpexc & VFP_FPEXC_UFF)
    462 			ksi.ksi_code = FPE_FLTUND;
    463 		else if (fpexc & VFP_FPEXC_OFF)
    464 			ksi.ksi_code = FPE_FLTOVF;
    465 		else if (fpexc & VFP_FPEXC_DZF)
    466 			ksi.ksi_code = FPE_FLTDIV;
    467 		else if (fpexc & VFP_FPEXC_IOF)
    468 			ksi.ksi_code = FPE_FLTINV;
    469 		ksi.ksi_addr = (uint32_t *)address;
    470 		ksi.ksi_trap = 0;
    471 		trapsignal(curlwp, &ksi);
    472 		return 0;
    473 	}
    474 
    475 	/* Need to restart the faulted instruction.  */
    476 //	frame->tf_pc -= INSN_SIZE;
    477 	return 0;
    478 }
    479 
    480 #ifdef CPU_CORTEX
    481 /* The real handler for NEON bounces.  */
    482 static int
    483 neon_handler(u_int address, u_int insn, trapframe_t *frame, int fault_code)
    484 {
    485 	struct cpu_info * const ci = curcpu();
    486 
    487 	if (ci->ci_vfp_id == 0)
    488 		/* No VFP detected, just fault.  */
    489 		return 1;
    490 
    491 	if ((insn & 0xfe000000) != 0xf2000000
    492 	    && (insn & 0xfe000000) != 0xf4000000)
    493 		/* Not NEON instruction, just fault.  */
    494 		return 1;
    495 
    496 	/* This shouldn't ever happen.  */
    497 	if (fault_code != FAULT_USER)
    498 		panic("NEON fault in non-user mode");
    499 
    500 	/* if we already own the FPU and it's enabled, raise SIGILL */
    501 	if (curcpu()->ci_pcu_curlwp[PCU_FPU] == curlwp
    502 	    && (armreg_fpexc_read() & VFP_FPEXC_EN) != 0)
    503 		return 1;
    504 
    505 	pcu_load(&arm_vfp_ops);
    506 
    507 	/* Need to restart the faulted instruction.  */
    508 //	frame->tf_pc -= INSN_SIZE;
    509 	return 0;
    510 }
    511 #endif
    512 
    513 static void
    514 vfp_state_load(lwp_t *l, u_int flags)
    515 {
    516 	struct pcb * const pcb = lwp_getpcb(l);
    517 	struct vfpreg * const fregs = &pcb->pcb_vfp;
    518 
    519 	/*
    520 	 * Instrument VFP usage -- if a process has not previously
    521 	 * used the VFP, mark it as having used VFP for the first time,
    522 	 * and count this event.
    523 	 *
    524 	 * If a process has used the VFP, count a "used VFP, and took
    525 	 * a trap to use it again" event.
    526 	 */
    527 	if (__predict_false((flags & PCU_VALID) == 0)) {
    528 		curcpu()->ci_vfp_evs[0].ev_count++;
    529 		pcb->pcb_vfp.vfp_fpscr = vfp_fpscr_default;
    530 	} else {
    531 		curcpu()->ci_vfp_evs[1].ev_count++;
    532 	}
    533 
    534 	KASSERT((armreg_fpexc_read() & VFP_FPEXC_EN) == 0);
    535 	/*
    536 	 * If the VFP is already enabled we must be bouncing an instruction.
    537 	 */
    538 	if (flags & PCU_REENABLE) {
    539 		uint32_t fpexc = armreg_fpexc_read();
    540 		armreg_fpexc_write(fpexc | VFP_FPEXC_EN);
    541 		fregs->vfp_fpexc |= VFP_FPEXC_EN;
    542 		return;
    543 	}
    544 	KASSERT((fregs->vfp_fpexc & VFP_FPEXC_EN) == 0);
    545 
    546 	/*
    547 	 * Load and Enable the VFP (so that we can write the registers).
    548 	 */
    549 	fregs->vfp_fpexc |= VFP_FPEXC_EN;
    550 	armreg_fpexc_write(fregs->vfp_fpexc);
    551 	KASSERT(curcpu()->ci_pcu_curlwp[PCU_FPU] == NULL);
    552 	KASSERT(l->l_pcu_cpu[PCU_FPU] == NULL);
    553 
    554 	load_vfpregs(fregs);
    555 	armreg_fpscr_write(fregs->vfp_fpscr);
    556 
    557 	if (fregs->vfp_fpexc & VFP_FPEXC_EX) {
    558 		/* Need to restore the exception handling state.  */
    559 		armreg_fpinst_write(fregs->vfp_fpinst);
    560 		if (fregs->vfp_fpexc & VFP_FPEXC_FP2V)
    561 			armreg_fpinst2_write(fregs->vfp_fpinst2);
    562 	}
    563 }
    564 
    565 void
    566 vfp_state_save(lwp_t *l)
    567 {
    568 	struct pcb * const pcb = lwp_getpcb(l);
    569 	struct vfpreg * const fregs = &pcb->pcb_vfp;
    570 	uint32_t fpexc = armreg_fpexc_read();
    571 
    572 	KASSERT(curcpu()->ci_pcu_curlwp[PCU_FPU] == l);
    573 	KASSERT(curcpu() == l->l_pcu_cpu[PCU_FPU]);
    574 	KASSERT(curlwp == l || curlwp->l_pcu_cpu[PCU_FPU] != curcpu());
    575 	/*
    576 	 * Enable the VFP (so we can read the registers).
    577 	 * Make sure the exception bit is cleared so that we can
    578 	 * safely dump the registers.
    579 	 */
    580 	armreg_fpexc_write((fpexc | VFP_FPEXC_EN) & ~VFP_FPEXC_EX);
    581 
    582 	fregs->vfp_fpexc = fpexc;
    583 	if (fpexc & VFP_FPEXC_EX) {
    584 		/* Need to save the exception handling state */
    585 		fregs->vfp_fpinst = armreg_fpinst_read();
    586 		if (fpexc & VFP_FPEXC_FP2V)
    587 			fregs->vfp_fpinst2 = armreg_fpinst2_read();
    588 	}
    589 	fregs->vfp_fpscr = armreg_fpscr_read();
    590 	save_vfpregs(fregs);
    591 
    592 	/* Disable the VFP.  */
    593 	armreg_fpexc_write(fpexc & ~VFP_FPEXC_EN);
    594 }
    595 
    596 void
    597 vfp_state_release(lwp_t *l)
    598 {
    599 	struct pcb * const pcb = lwp_getpcb(l);
    600 
    601 	/*
    602 	 * Now mark the VFP as disabled (and our state
    603 	 * has been already saved or is being discarded).
    604 	 */
    605 	pcb->pcb_vfp.vfp_fpexc &= ~VFP_FPEXC_EN;
    606 
    607 	/*
    608 	 * Turn off the FPU so the next time a VFP instruction is issued
    609 	 * an exception happens.  We don't know if this LWP's state was
    610 	 * loaded but if we turned off the FPU for some other LWP, when
    611 	 * pcu_load invokes vfp_state_load it will see that VFP_FPEXC_EN
    612 	 * is still set so it just restore fpexc and return since its
    613 	 * contents are still sitting in the VFP.
    614 	 */
    615 	armreg_fpexc_write(armreg_fpexc_read() & ~VFP_FPEXC_EN);
    616 }
    617 
    618 void
    619 vfp_savecontext(lwp_t *l)
    620 {
    621 	pcu_save(&arm_vfp_ops, l);
    622 }
    623 
    624 void
    625 vfp_discardcontext(lwp_t *l, bool used_p)
    626 {
    627 	pcu_discard(&arm_vfp_ops, l, used_p);
    628 }
    629 
    630 bool
    631 vfp_used_p(const lwp_t *l)
    632 {
    633 	return pcu_valid_p(&arm_vfp_ops, l);
    634 }
    635 
    636 void
    637 vfp_getcontext(struct lwp *l, mcontext_t *mcp, int *flagsp)
    638 {
    639 	if (vfp_used_p(l)) {
    640 		const struct pcb * const pcb = lwp_getpcb(l);
    641 
    642 		pcu_save(&arm_vfp_ops, l);
    643 		mcp->__fpu.__vfpregs.__vfp_fpscr = pcb->pcb_vfp.vfp_fpscr;
    644 		memcpy(mcp->__fpu.__vfpregs.__vfp_fstmx, pcb->pcb_vfp.vfp_regs,
    645 		    sizeof(mcp->__fpu.__vfpregs.__vfp_fstmx));
    646 		*flagsp |= _UC_FPU|_UC_ARM_VFP;
    647 	}
    648 }
    649 
    650 void
    651 vfp_setcontext(struct lwp *l, const mcontext_t *mcp)
    652 {
    653 	struct pcb * const pcb = lwp_getpcb(l);
    654 
    655 	pcu_discard(&arm_vfp_ops, l, true);
    656 	pcb->pcb_vfp.vfp_fpscr = mcp->__fpu.__vfpregs.__vfp_fpscr;
    657 	memcpy(pcb->pcb_vfp.vfp_regs, mcp->__fpu.__vfpregs.__vfp_fstmx,
    658 	    sizeof(mcp->__fpu.__vfpregs.__vfp_fstmx));
    659 }
    660 
    661 #endif /* FPU_VFP */
    662