Home | History | Annotate | Line # | Download | only in vfp
vfp_init.c revision 1.58
      1  1.58     skrll /*      $NetBSD: vfp_init.c,v 1.58 2018/08/15 05:52:15 skrll Exp $ */
      2   1.1  rearnsha 
      3   1.1  rearnsha /*
      4   1.1  rearnsha  * Copyright (c) 2008 ARM Ltd
      5   1.1  rearnsha  * All rights reserved.
      6   1.1  rearnsha  *
      7   1.1  rearnsha  * Redistribution and use in source and binary forms, with or without
      8   1.1  rearnsha  * modification, are permitted provided that the following conditions
      9   1.1  rearnsha  * are met:
     10   1.1  rearnsha  * 1. Redistributions of source code must retain the above copyright
     11   1.1  rearnsha  *    notice, this list of conditions and the following disclaimer.
     12   1.1  rearnsha  * 2. Redistributions in binary form must reproduce the above copyright
     13   1.1  rearnsha  *    notice, this list of conditions and the following disclaimer in the
     14   1.1  rearnsha  *    documentation and/or other materials provided with the distribution.
     15   1.1  rearnsha  * 3. The name of the company may not be used to endorse or promote
     16   1.1  rearnsha  *    products derived from this software without specific prior written
     17   1.1  rearnsha  *    permission.
     18   1.1  rearnsha  *
     19   1.1  rearnsha  * THIS SOFTWARE IS PROVIDED BY ARM LTD ``AS IS'' AND ANY EXPRESS OR
     20   1.1  rearnsha  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
     21   1.1  rearnsha  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
     22   1.1  rearnsha  * ARE DISCLAIMED.  IN NO EVENT SHALL ARM LTD BE LIABLE FOR ANY
     23   1.1  rearnsha  * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     24   1.1  rearnsha  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE
     25   1.1  rearnsha  * GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
     26   1.1  rearnsha  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
     27   1.1  rearnsha  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
     28   1.1  rearnsha  * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
     29   1.1  rearnsha  * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
     30   1.1  rearnsha  */
     31   1.1  rearnsha 
     32  1.58     skrll #include <sys/cdefs.h>
     33  1.58     skrll __KERNEL_RCSID(0, "$NetBSD: vfp_init.c,v 1.58 2018/08/15 05:52:15 skrll Exp $");
     34  1.58     skrll 
     35   1.1  rearnsha #include <sys/param.h>
     36   1.1  rearnsha #include <sys/types.h>
     37   1.1  rearnsha #include <sys/systm.h>
     38   1.1  rearnsha #include <sys/device.h>
     39   1.1  rearnsha #include <sys/proc.h>
     40   1.4      matt #include <sys/cpu.h>
     41   1.1  rearnsha 
     42  1.23      matt #include <arm/locore.h>
     43   1.5      matt #include <arm/pcb.h>
     44   1.1  rearnsha #include <arm/undefined.h>
     45   1.1  rearnsha #include <arm/vfpreg.h>
     46   1.8      matt #include <arm/mcontext.h>
     47   1.1  rearnsha 
     48  1.12      matt #include <uvm/uvm_extern.h>		/* for pmap.h */
     49  1.12      matt 
     50  1.11      matt #ifdef FPU_VFP
     51  1.11      matt 
     52  1.29      matt #ifdef CPU_CORTEX
     53  1.56  christos #define SETFPU __asm(".fpu\tvfpv4")
     54  1.29      matt #else
     55  1.56  christos #define SETFPU __asm(".fpu\tvfp")
     56  1.29      matt #endif
     57  1.56  christos SETFPU;
     58  1.29      matt 
     59   1.1  rearnsha /* FLDMD <X>, {d0-d15} */
     60  1.11      matt static inline void
     61  1.13      matt load_vfpregs_lo(const uint64_t *p)
     62  1.10      matt {
     63  1.56  christos 	SETFPU;
     64  1.56  christos 	__asm __volatile("vldmia\t%0, {d0-d15}" :: "r" (p) : "memory");
     65  1.10      matt }
     66  1.10      matt 
     67  1.10      matt /* FSTMD <X>, {d0-d15} */
     68  1.11      matt static inline void
     69  1.10      matt save_vfpregs_lo(uint64_t *p)
     70  1.10      matt {
     71  1.56  christos 	SETFPU;
     72  1.56  christos 	__asm __volatile("vstmia\t%0, {d0-d15}" :: "r" (p) : "memory");
     73  1.10      matt }
     74  1.10      matt 
     75  1.10      matt #ifdef CPU_CORTEX
     76  1.10      matt /* FLDMD <X>, {d16-d31} */
     77  1.11      matt static inline void
     78  1.13      matt load_vfpregs_hi(const uint64_t *p)
     79  1.10      matt {
     80  1.56  christos 	SETFPU;
     81  1.29      matt 	__asm __volatile("vldmia\t%0, {d16-d31}" :: "r" (&p[16]) : "memory");
     82  1.10      matt }
     83  1.10      matt 
     84  1.10      matt /* FLDMD <X>, {d16-d31} */
     85  1.11      matt static inline void
     86  1.10      matt save_vfpregs_hi(uint64_t *p)
     87  1.10      matt {
     88  1.56  christos 	SETFPU;
     89  1.29      matt 	__asm __volatile("vstmia\t%0, {d16-d31}" :: "r" (&p[16]) : "memory");
     90  1.10      matt }
     91  1.10      matt #endif
     92   1.1  rearnsha 
     93  1.13      matt static inline void
     94  1.13      matt load_vfpregs(const struct vfpreg *fregs)
     95  1.13      matt {
     96  1.13      matt 	load_vfpregs_lo(fregs->vfp_regs);
     97  1.13      matt #ifdef CPU_CORTEX
     98  1.13      matt #ifdef CPU_ARM11
     99  1.13      matt 	switch (curcpu()->ci_vfp_id) {
    100  1.13      matt 	case FPU_VFP_CORTEXA5:
    101  1.13      matt 	case FPU_VFP_CORTEXA7:
    102  1.13      matt 	case FPU_VFP_CORTEXA8:
    103  1.13      matt 	case FPU_VFP_CORTEXA9:
    104  1.20      matt 	case FPU_VFP_CORTEXA15:
    105  1.42       slp 	case FPU_VFP_CORTEXA15_QEMU:
    106  1.50     skrll 	case FPU_VFP_CORTEXA53:
    107  1.53  jmcneill 	case FPU_VFP_CORTEXA57:
    108  1.13      matt #endif
    109  1.13      matt 		load_vfpregs_hi(fregs->vfp_regs);
    110  1.13      matt #ifdef CPU_ARM11
    111  1.13      matt 		break;
    112  1.13      matt 	}
    113  1.13      matt #endif
    114  1.13      matt #endif
    115  1.13      matt }
    116  1.13      matt 
    117  1.13      matt static inline void
    118  1.13      matt save_vfpregs(struct vfpreg *fregs)
    119  1.13      matt {
    120  1.13      matt 	save_vfpregs_lo(fregs->vfp_regs);
    121  1.13      matt #ifdef CPU_CORTEX
    122  1.13      matt #ifdef CPU_ARM11
    123  1.13      matt 	switch (curcpu()->ci_vfp_id) {
    124  1.13      matt 	case FPU_VFP_CORTEXA5:
    125  1.13      matt 	case FPU_VFP_CORTEXA7:
    126  1.13      matt 	case FPU_VFP_CORTEXA8:
    127  1.13      matt 	case FPU_VFP_CORTEXA9:
    128  1.20      matt 	case FPU_VFP_CORTEXA15:
    129  1.42       slp 	case FPU_VFP_CORTEXA15_QEMU:
    130  1.50     skrll 	case FPU_VFP_CORTEXA53:
    131  1.53  jmcneill 	case FPU_VFP_CORTEXA57:
    132  1.13      matt #endif
    133  1.13      matt 		save_vfpregs_hi(fregs->vfp_regs);
    134  1.13      matt #ifdef CPU_ARM11
    135  1.13      matt 		break;
    136  1.13      matt 	}
    137  1.13      matt #endif
    138  1.13      matt #endif
    139  1.13      matt }
    140  1.13      matt 
    141   1.1  rearnsha /* The real handler for VFP bounces.  */
    142   1.1  rearnsha static int vfp_handler(u_int, u_int, trapframe_t *, int);
    143  1.13      matt #ifdef CPU_CORTEX
    144  1.13      matt static int neon_handler(u_int, u_int, trapframe_t *, int);
    145  1.13      matt #endif
    146   1.1  rearnsha 
    147  1.13      matt static void vfp_state_load(lwp_t *, u_int);
    148  1.39     rmind static void vfp_state_save(lwp_t *);
    149  1.39     rmind static void vfp_state_release(lwp_t *);
    150   1.4      matt 
    151   1.4      matt const pcu_ops_t arm_vfp_ops = {
    152   1.4      matt 	.pcu_id = PCU_FPU,
    153  1.13      matt 	.pcu_state_save = vfp_state_save,
    154   1.4      matt 	.pcu_state_load = vfp_state_load,
    155   1.4      matt 	.pcu_state_release = vfp_state_release,
    156   1.4      matt };
    157   1.1  rearnsha 
    158  1.34      matt /* determine what bits can be changed */
    159  1.34      matt uint32_t vfp_fpscr_changable = VFP_FPSCR_CSUM;
    160  1.34      matt /* default to run fast */
    161  1.34      matt uint32_t vfp_fpscr_default = (VFP_FPSCR_DN | VFP_FPSCR_FZ | VFP_FPSCR_RN);
    162  1.34      matt 
    163   1.1  rearnsha /*
    164   1.1  rearnsha  * Used to test for a VFP. The following function is installed as a coproc10
    165   1.1  rearnsha  * handler on the undefined instruction vector and then we issue a VFP
    166   1.1  rearnsha  * instruction. If undefined_test is non zero then the VFP did not handle
    167   1.1  rearnsha  * the instruction so must be absent, or disabled.
    168   1.1  rearnsha  */
    169   1.1  rearnsha 
    170   1.1  rearnsha static int undefined_test;
    171   1.1  rearnsha 
    172   1.1  rearnsha static int
    173   1.4      matt vfp_test(u_int address, u_int insn, trapframe_t *frame, int fault_code)
    174   1.1  rearnsha {
    175   1.1  rearnsha 
    176   1.1  rearnsha 	frame->tf_pc += INSN_SIZE;
    177   1.1  rearnsha 	++undefined_test;
    178   1.4      matt 	return 0;
    179   1.4      matt }
    180   1.4      matt 
    181  1.35      matt #else
    182  1.35      matt /* determine what bits can be changed */
    183  1.35      matt uint32_t vfp_fpscr_changable = VFP_FPSCR_CSUM|VFP_FPSCR_ESUM|VFP_FPSCR_RMODE;
    184   1.4      matt #endif /* FPU_VFP */
    185   1.4      matt 
    186   1.4      matt static int
    187   1.4      matt vfp_fpscr_handler(u_int address, u_int insn, trapframe_t *frame, int fault_code)
    188   1.4      matt {
    189   1.4      matt 	struct lwp * const l = curlwp;
    190   1.4      matt 	const u_int regno = (insn >> 12) & 0xf;
    191   1.4      matt 	/*
    192   1.4      matt 	 * Only match move to/from the FPSCR register and we
    193   1.4      matt 	 * can't be using the SP,LR,PC as a source.
    194   1.4      matt 	 */
    195   1.4      matt 	if ((insn & 0xffef0fff) != 0xeee10a10 || regno > 12)
    196   1.4      matt 		return 1;
    197   1.4      matt 
    198   1.4      matt 	struct pcb * const pcb = lwp_getpcb(l);
    199   1.4      matt 
    200   1.4      matt #ifdef FPU_VFP
    201   1.4      matt 	/*
    202   1.4      matt 	 * If FPU is valid somewhere, let's just reenable VFP and
    203   1.4      matt 	 * retry the instruction (only safe thing to do since the
    204   1.4      matt 	 * pcb has a stale copy).
    205   1.4      matt 	 */
    206   1.4      matt 	if (pcb->pcb_vfp.vfp_fpexc & VFP_FPEXC_EN)
    207   1.4      matt 		return 1;
    208   1.4      matt 
    209  1.51       chs 	if (__predict_false(!vfp_used_p(l))) {
    210  1.35      matt 		pcb->pcb_vfp.vfp_fpscr = vfp_fpscr_default;
    211   1.4      matt 	}
    212  1.26      matt #endif
    213   1.4      matt 
    214   1.4      matt 	/*
    215  1.30     skrll 	 * We now know the pcb has the saved copy.
    216   1.4      matt 	 */
    217   1.4      matt 	register_t * const regp = &frame->tf_r0 + regno;
    218   1.4      matt 	if (insn & 0x00100000) {
    219   1.4      matt 		*regp = pcb->pcb_vfp.vfp_fpscr;
    220   1.4      matt 	} else {
    221  1.34      matt 		pcb->pcb_vfp.vfp_fpscr &= ~vfp_fpscr_changable;
    222  1.34      matt 		pcb->pcb_vfp.vfp_fpscr |= *regp & vfp_fpscr_changable;
    223   1.4      matt 	}
    224   1.4      matt 
    225  1.37      matt 	curcpu()->ci_vfp_evs[0].ev_count++;
    226   1.4      matt 
    227   1.4      matt 	frame->tf_pc += INSN_SIZE;
    228   1.4      matt 	return 0;
    229   1.1  rearnsha }
    230   1.1  rearnsha 
    231   1.4      matt #ifndef FPU_VFP
    232   1.4      matt /*
    233   1.4      matt  * If we don't want VFP support, we still need to handle emulating VFP FPSCR
    234   1.4      matt  * instructions.
    235   1.4      matt  */
    236   1.4      matt void
    237  1.37      matt vfp_attach(struct cpu_info *ci)
    238   1.4      matt {
    239  1.37      matt 	if (CPU_IS_PRIMARY(ci)) {
    240  1.37      matt 		install_coproc_handler(VFP_COPROC, vfp_fpscr_handler);
    241  1.37      matt 	}
    242  1.37      matt 	evcnt_attach_dynamic(&ci->ci_vfp_evs[0], EVCNT_TYPE_TRAP, NULL,
    243  1.37      matt 	    ci->ci_cpuname, "vfp fpscr traps");
    244   1.4      matt }
    245   1.4      matt 
    246   1.4      matt #else
    247   1.1  rearnsha void
    248  1.37      matt vfp_attach(struct cpu_info *ci)
    249   1.1  rearnsha {
    250   1.4      matt 	const char *model = NULL;
    251   1.1  rearnsha 
    252  1.37      matt 	if (CPU_ID_ARM11_P(ci->ci_arm_cpuid)
    253  1.37      matt 	    || CPU_ID_MV88SV58XX_P(ci->ci_arm_cpuid)
    254  1.37      matt 	    || CPU_ID_CORTEX_P(ci->ci_arm_cpuid)) {
    255  1.37      matt #if 0
    256  1.37      matt 		const uint32_t nsacr = armreg_nsacr_read();
    257  1.37      matt 		const uint32_t nsacr_vfp = __BITS(VFP_COPROC,VFP_COPROC2);
    258  1.37      matt 		if ((nsacr & nsacr_vfp) != nsacr_vfp) {
    259  1.40      matt 			aprint_normal_dev(ci->ci_dev,
    260  1.40      matt 			    "VFP access denied (NSACR=%#x)\n", nsacr);
    261  1.37      matt 			install_coproc_handler(VFP_COPROC, vfp_fpscr_handler);
    262  1.37      matt 			ci->ci_vfp_id = 0;
    263  1.37      matt 			evcnt_attach_dynamic(&ci->ci_vfp_evs[0],
    264  1.37      matt 			    EVCNT_TYPE_TRAP, NULL, ci->ci_cpuname,
    265  1.37      matt 			    "vfp fpscr traps");
    266  1.37      matt 			return;
    267  1.37      matt 		}
    268  1.37      matt #endif
    269   1.7      matt 		const uint32_t cpacr_vfp = CPACR_CPn(VFP_COPROC);
    270   1.7      matt 		const uint32_t cpacr_vfp2 = CPACR_CPn(VFP_COPROC2);
    271   1.1  rearnsha 
    272   1.7      matt 		/*
    273   1.7      matt 		 * We first need to enable access to the coprocessors.
    274   1.7      matt 		 */
    275   1.7      matt 		uint32_t cpacr = armreg_cpacr_read();
    276   1.7      matt 		cpacr |= __SHIFTIN(CPACR_ALL, cpacr_vfp);
    277   1.7      matt 		cpacr |= __SHIFTIN(CPACR_ALL, cpacr_vfp2);
    278   1.7      matt 		armreg_cpacr_write(cpacr);
    279   1.1  rearnsha 
    280  1.48  jmcneill 		arm_isb();
    281  1.48  jmcneill 
    282   1.7      matt 		/*
    283   1.7      matt 		 * If we could enable them, then they exist.
    284   1.7      matt 		 */
    285   1.7      matt 		cpacr = armreg_cpacr_read();
    286  1.40      matt 		bool vfp_p = __SHIFTOUT(cpacr, cpacr_vfp2) == CPACR_ALL
    287  1.40      matt 		    && __SHIFTOUT(cpacr, cpacr_vfp) == CPACR_ALL;
    288  1.28      matt 		if (!vfp_p) {
    289  1.40      matt 			aprint_normal_dev(ci->ci_dev,
    290  1.40      matt 			    "VFP access denied (CPACR=%#x)\n", cpacr);
    291  1.28      matt 			install_coproc_handler(VFP_COPROC, vfp_fpscr_handler);
    292  1.28      matt 			ci->ci_vfp_id = 0;
    293  1.37      matt 			evcnt_attach_dynamic(&ci->ci_vfp_evs[0],
    294  1.37      matt 			    EVCNT_TYPE_TRAP, NULL, ci->ci_cpuname,
    295  1.37      matt 			    "vfp fpscr traps");
    296  1.28      matt 			return;
    297  1.28      matt 		}
    298   1.6      matt 	}
    299   1.6      matt 
    300   1.7      matt 	void *uh = install_coproc_handler(VFP_COPROC, vfp_test);
    301   1.7      matt 
    302   1.7      matt 	undefined_test = 0;
    303   1.7      matt 
    304  1.21      matt 	const uint32_t fpsid = armreg_fpsid_read();
    305   1.1  rearnsha 
    306   1.1  rearnsha 	remove_coproc_handler(uh);
    307   1.1  rearnsha 
    308   1.1  rearnsha 	if (undefined_test != 0) {
    309   1.4      matt 		aprint_normal_dev(ci->ci_dev, "No VFP detected\n");
    310   1.4      matt 		install_coproc_handler(VFP_COPROC, vfp_fpscr_handler);
    311   1.4      matt 		ci->ci_vfp_id = 0;
    312   1.1  rearnsha 		return;
    313   1.1  rearnsha 	}
    314   1.1  rearnsha 
    315   1.4      matt 	ci->ci_vfp_id = fpsid;
    316   1.4      matt 	switch (fpsid & ~ VFP_FPSID_REV_MSK) {
    317   1.4      matt 	case FPU_VFP10_ARM10E:
    318   1.4      matt 		model = "VFP10 R1";
    319   1.4      matt 		break;
    320   1.4      matt 	case FPU_VFP11_ARM11:
    321   1.4      matt 		model = "VFP11";
    322   1.4      matt 		break;
    323  1.36      matt 	case FPU_VFP_MV88SV58XX:
    324  1.36      matt 		model = "VFP3";
    325  1.36      matt 		break;
    326   1.7      matt 	case FPU_VFP_CORTEXA5:
    327   1.7      matt 	case FPU_VFP_CORTEXA7:
    328   1.7      matt 	case FPU_VFP_CORTEXA8:
    329   1.7      matt 	case FPU_VFP_CORTEXA9:
    330  1.20      matt 	case FPU_VFP_CORTEXA15:
    331  1.42       slp 	case FPU_VFP_CORTEXA15_QEMU:
    332  1.50     skrll 	case FPU_VFP_CORTEXA53:
    333  1.53  jmcneill 	case FPU_VFP_CORTEXA57:
    334  1.37      matt 		if (armreg_cpacr_read() & CPACR_V7_ASEDIS) {
    335  1.37      matt 			model = "VFP 4.0+";
    336  1.37      matt 		} else {
    337  1.37      matt 			model = "NEON MPE (VFP 3.0+)";
    338  1.37      matt 			cpu_neon_present = 1;
    339  1.37      matt 		}
    340   1.6      matt 		break;
    341   1.4      matt 	default:
    342  1.36      matt 		aprint_normal_dev(ci->ci_dev, "unrecognized VFP version %#x\n",
    343   1.4      matt 		    fpsid);
    344   1.4      matt 		install_coproc_handler(VFP_COPROC, vfp_fpscr_handler);
    345  1.35      matt 		vfp_fpscr_changable = VFP_FPSCR_CSUM|VFP_FPSCR_ESUM
    346  1.35      matt 		    |VFP_FPSCR_RMODE;
    347  1.35      matt 		vfp_fpscr_default = 0;
    348   1.4      matt 		return;
    349   1.4      matt 	}
    350   1.1  rearnsha 
    351  1.17      matt 	cpu_fpu_present = 1;
    352  1.21      matt 	cpu_media_and_vfp_features[0] = armreg_mvfr0_read();
    353  1.21      matt 	cpu_media_and_vfp_features[1] = armreg_mvfr1_read();
    354   1.1  rearnsha 	if (fpsid != 0) {
    355  1.34      matt 		uint32_t f0 = armreg_mvfr0_read();
    356  1.41      matt 		uint32_t f1 = armreg_mvfr1_read();
    357  1.34      matt 		aprint_normal("vfp%d at %s: %s%s%s%s%s\n",
    358  1.37      matt 		    device_unit(ci->ci_dev),
    359  1.37      matt 		    device_xname(ci->ci_dev),
    360  1.34      matt 		    model,
    361  1.34      matt 		    ((f0 & ARM_MVFR0_ROUNDING_MASK) ? ", rounding" : ""),
    362  1.34      matt 		    ((f0 & ARM_MVFR0_EXCEPT_MASK) ? ", exceptions" : ""),
    363  1.38      matt 		    ((f1 & ARM_MVFR1_D_NAN_MASK) ? ", NaN propagation" : ""),
    364  1.34      matt 		    ((f1 & ARM_MVFR1_FTZ_MASK) ? ", denormals" : ""));
    365  1.49  jmcneill 		aprint_debug("vfp%d: mvfr: [0]=%#x [1]=%#x\n",
    366  1.37      matt 		    device_unit(ci->ci_dev), f0, f1);
    367  1.37      matt 		if (CPU_IS_PRIMARY(ci)) {
    368  1.37      matt 			if (f0 & ARM_MVFR0_ROUNDING_MASK) {
    369  1.37      matt 				vfp_fpscr_changable |= VFP_FPSCR_RMODE;
    370  1.37      matt 			}
    371  1.37      matt 			if (f1 & ARM_MVFR0_EXCEPT_MASK) {
    372  1.37      matt 				vfp_fpscr_changable |= VFP_FPSCR_ESUM;
    373  1.37      matt 			}
    374  1.38      matt 			// If hardware supports propagation of NaNs, select it.
    375  1.37      matt 			if (f1 & ARM_MVFR1_D_NAN_MASK) {
    376  1.37      matt 				vfp_fpscr_default &= ~VFP_FPSCR_DN;
    377  1.37      matt 				vfp_fpscr_changable |= VFP_FPSCR_DN;
    378  1.37      matt 			}
    379  1.37      matt 			// If hardware supports denormalized numbers, use it.
    380  1.37      matt 			if (cpu_media_and_vfp_features[1] & ARM_MVFR1_FTZ_MASK) {
    381  1.37      matt 				vfp_fpscr_default &= ~VFP_FPSCR_FZ;
    382  1.37      matt 				vfp_fpscr_changable |= VFP_FPSCR_FZ;
    383  1.37      matt 			}
    384  1.37      matt 		}
    385  1.37      matt 	}
    386  1.37      matt 	evcnt_attach_dynamic(&ci->ci_vfp_evs[0], EVCNT_TYPE_MISC, NULL,
    387  1.37      matt 	    ci->ci_cpuname, "vfp coproc use");
    388  1.37      matt 	evcnt_attach_dynamic(&ci->ci_vfp_evs[1], EVCNT_TYPE_MISC, NULL,
    389  1.37      matt 	    ci->ci_cpuname, "vfp coproc re-use");
    390  1.37      matt 	evcnt_attach_dynamic(&ci->ci_vfp_evs[2], EVCNT_TYPE_TRAP, NULL,
    391  1.37      matt 	    ci->ci_cpuname, "vfp coproc fault");
    392   1.1  rearnsha 	install_coproc_handler(VFP_COPROC, vfp_handler);
    393   1.1  rearnsha 	install_coproc_handler(VFP_COPROC2, vfp_handler);
    394  1.13      matt #ifdef CPU_CORTEX
    395  1.43      matt 	if (cpu_neon_present)
    396  1.43      matt 		install_coproc_handler(CORE_UNKNOWN_HANDLER, neon_handler);
    397  1.13      matt #endif
    398   1.1  rearnsha }
    399   1.1  rearnsha 
    400   1.1  rearnsha /* The real handler for VFP bounces.  */
    401   1.4      matt static int
    402  1.21      matt vfp_handler(u_int address, u_int insn, trapframe_t *frame, int fault_code)
    403   1.1  rearnsha {
    404   1.4      matt 	struct cpu_info * const ci = curcpu();
    405   1.1  rearnsha 
    406   1.1  rearnsha 	/* This shouldn't ever happen.  */
    407   1.1  rearnsha 	if (fault_code != FAULT_USER)
    408  1.14      matt 		panic("VFP fault at %#x in non-user mode", frame->tf_pc);
    409   1.1  rearnsha 
    410  1.27      matt 	if (ci->ci_vfp_id == 0) {
    411   1.1  rearnsha 		/* No VFP detected, just fault.  */
    412   1.1  rearnsha 		return 1;
    413  1.27      matt 	}
    414  1.27      matt 
    415  1.47      matt 	/*
    416  1.47      matt 	 * If we already own the FPU and it's enabled (and no exception), raise
    417  1.47      matt 	 * SIGILL.  If there is an exception, drop through to raise a SIGFPE.
    418  1.47      matt 	 */
    419  1.46      matt 	if (curcpu()->ci_pcu_curlwp[PCU_FPU] == curlwp
    420  1.47      matt 	    && (armreg_fpexc_read() & (VFP_FPEXC_EX|VFP_FPEXC_EN)) == VFP_FPEXC_EN)
    421  1.47      matt 		return 1;
    422  1.44      matt 
    423  1.27      matt 	/*
    424  1.27      matt 	 * Make sure we own the FP.
    425  1.27      matt 	 */
    426  1.27      matt 	pcu_load(&arm_vfp_ops);
    427   1.1  rearnsha 
    428  1.21      matt 	uint32_t fpexc = armreg_fpexc_read();
    429  1.21      matt 	if (fpexc & VFP_FPEXC_EX) {
    430  1.21      matt 		ksiginfo_t ksi;
    431  1.21      matt 		KASSERT(fpexc & VFP_FPEXC_EN);
    432  1.21      matt 
    433  1.37      matt 		curcpu()->ci_vfp_evs[2].ev_count++;
    434  1.21      matt 
    435  1.21      matt 		/*
    436  1.21      matt 		 * Need the clear the exception condition so any signal
    437  1.33     skrll 		 * and future use can proceed.
    438  1.21      matt 		 */
    439  1.31     skrll 		armreg_fpexc_write(fpexc & ~(VFP_FPEXC_EX|VFP_FPEXC_FSUM));
    440  1.21      matt 
    441  1.51       chs 		pcu_save(&arm_vfp_ops, curlwp);
    442  1.33     skrll 
    443  1.33     skrll 		/*
    444  1.33     skrll 		 * XXX Need to emulate bounce instructions here to get correct
    445  1.33     skrll 		 * XXX exception codes, etc.
    446  1.33     skrll 		 */
    447  1.21      matt 		KSI_INIT_TRAP(&ksi);
    448  1.21      matt 		ksi.ksi_signo = SIGFPE;
    449  1.21      matt 		if (fpexc & VFP_FPEXC_IXF)
    450  1.21      matt 			ksi.ksi_code = FPE_FLTRES;
    451  1.21      matt 		else if (fpexc & VFP_FPEXC_UFF)
    452  1.21      matt 			ksi.ksi_code = FPE_FLTUND;
    453  1.21      matt 		else if (fpexc & VFP_FPEXC_OFF)
    454  1.21      matt 			ksi.ksi_code = FPE_FLTOVF;
    455  1.21      matt 		else if (fpexc & VFP_FPEXC_DZF)
    456  1.21      matt 			ksi.ksi_code = FPE_FLTDIV;
    457  1.21      matt 		else if (fpexc & VFP_FPEXC_IOF)
    458  1.21      matt 			ksi.ksi_code = FPE_FLTINV;
    459  1.21      matt 		ksi.ksi_addr = (uint32_t *)address;
    460  1.21      matt 		ksi.ksi_trap = 0;
    461  1.21      matt 		trapsignal(curlwp, &ksi);
    462  1.21      matt 		return 0;
    463  1.21      matt 	}
    464  1.21      matt 
    465   1.4      matt 	/* Need to restart the faulted instruction.  */
    466   1.4      matt //	frame->tf_pc -= INSN_SIZE;
    467   1.4      matt 	return 0;
    468   1.4      matt }
    469   1.1  rearnsha 
    470  1.13      matt #ifdef CPU_CORTEX
    471  1.13      matt /* The real handler for NEON bounces.  */
    472  1.13      matt static int
    473  1.21      matt neon_handler(u_int address, u_int insn, trapframe_t *frame, int fault_code)
    474  1.13      matt {
    475  1.13      matt 	struct cpu_info * const ci = curcpu();
    476  1.13      matt 
    477  1.13      matt 	if (ci->ci_vfp_id == 0)
    478  1.13      matt 		/* No VFP detected, just fault.  */
    479  1.13      matt 		return 1;
    480  1.13      matt 
    481  1.13      matt 	if ((insn & 0xfe000000) != 0xf2000000
    482  1.13      matt 	    && (insn & 0xfe000000) != 0xf4000000)
    483  1.13      matt 		/* Not NEON instruction, just fault.  */
    484  1.13      matt 		return 1;
    485  1.13      matt 
    486  1.13      matt 	/* This shouldn't ever happen.  */
    487  1.13      matt 	if (fault_code != FAULT_USER)
    488  1.13      matt 		panic("NEON fault in non-user mode");
    489  1.13      matt 
    490  1.45      matt 	/* if we already own the FPU and it's enabled, raise SIGILL */
    491  1.45      matt 	if (curcpu()->ci_pcu_curlwp[PCU_FPU] == curlwp
    492  1.45      matt 	    && (armreg_fpexc_read() & VFP_FPEXC_EN) != 0)
    493  1.47      matt 		return 1;
    494  1.43      matt 
    495  1.13      matt 	pcu_load(&arm_vfp_ops);
    496  1.13      matt 
    497  1.13      matt 	/* Need to restart the faulted instruction.  */
    498  1.13      matt //	frame->tf_pc -= INSN_SIZE;
    499  1.13      matt 	return 0;
    500  1.13      matt }
    501  1.13      matt #endif
    502  1.13      matt 
    503   1.4      matt static void
    504  1.13      matt vfp_state_load(lwp_t *l, u_int flags)
    505   1.4      matt {
    506   1.4      matt 	struct pcb * const pcb = lwp_getpcb(l);
    507   1.4      matt 	struct vfpreg * const fregs = &pcb->pcb_vfp;
    508   1.1  rearnsha 
    509   1.1  rearnsha 	/*
    510   1.1  rearnsha 	 * Instrument VFP usage -- if a process has not previously
    511   1.1  rearnsha 	 * used the VFP, mark it as having used VFP for the first time,
    512   1.1  rearnsha 	 * and count this event.
    513   1.1  rearnsha 	 *
    514   1.1  rearnsha 	 * If a process has used the VFP, count a "used VFP, and took
    515   1.1  rearnsha 	 * a trap to use it again" event.
    516   1.1  rearnsha 	 */
    517  1.39     rmind 	if (__predict_false((flags & PCU_VALID) == 0)) {
    518  1.37      matt 		curcpu()->ci_vfp_evs[0].ev_count++;
    519  1.34      matt 		pcb->pcb_vfp.vfp_fpscr = vfp_fpscr_default;
    520   1.4      matt 	} else {
    521  1.37      matt 		curcpu()->ci_vfp_evs[1].ev_count++;
    522   1.4      matt 	}
    523   1.1  rearnsha 
    524  1.54    bouyer 	KASSERT((armreg_fpexc_read() & VFP_FPEXC_EN) == 0);
    525  1.39     rmind 	/*
    526  1.39     rmind 	 * If the VFP is already enabled we must be bouncing an instruction.
    527  1.39     rmind 	 */
    528  1.39     rmind 	if (flags & PCU_REENABLE) {
    529  1.39     rmind 		uint32_t fpexc = armreg_fpexc_read();
    530  1.39     rmind 		armreg_fpexc_write(fpexc | VFP_FPEXC_EN);
    531  1.54    bouyer 		fregs->vfp_fpexc |= VFP_FPEXC_EN;
    532  1.39     rmind 		return;
    533  1.39     rmind 	}
    534  1.54    bouyer 	KASSERT((fregs->vfp_fpexc & VFP_FPEXC_EN) == 0);
    535  1.33     skrll 
    536  1.39     rmind 	/*
    537  1.39     rmind 	 * Load and Enable the VFP (so that we can write the registers).
    538  1.39     rmind 	 */
    539  1.39     rmind 	fregs->vfp_fpexc |= VFP_FPEXC_EN;
    540  1.39     rmind 	armreg_fpexc_write(fregs->vfp_fpexc);
    541  1.54    bouyer 	KASSERT(curcpu()->ci_pcu_curlwp[PCU_FPU] == NULL);
    542  1.54    bouyer 	KASSERT(l->l_pcu_cpu[PCU_FPU] == NULL);
    543  1.13      matt 
    544  1.39     rmind 	load_vfpregs(fregs);
    545  1.39     rmind 	armreg_fpscr_write(fregs->vfp_fpscr);
    546  1.13      matt 
    547  1.39     rmind 	if (fregs->vfp_fpexc & VFP_FPEXC_EX) {
    548  1.39     rmind 		/* Need to restore the exception handling state.  */
    549  1.52       chs 		armreg_fpinst_write(fregs->vfp_fpinst);
    550  1.39     rmind 		if (fregs->vfp_fpexc & VFP_FPEXC_FP2V)
    551  1.52       chs 			armreg_fpinst2_write(fregs->vfp_fpinst2);
    552   1.1  rearnsha 	}
    553   1.1  rearnsha }
    554   1.1  rearnsha 
    555   1.1  rearnsha void
    556  1.39     rmind vfp_state_save(lwp_t *l)
    557   1.1  rearnsha {
    558   1.4      matt 	struct pcb * const pcb = lwp_getpcb(l);
    559  1.39     rmind 	struct vfpreg * const fregs = &pcb->pcb_vfp;
    560  1.21      matt 	uint32_t fpexc = armreg_fpexc_read();
    561  1.33     skrll 
    562  1.54    bouyer 	KASSERT(curcpu()->ci_pcu_curlwp[PCU_FPU] == l);
    563  1.54    bouyer 	KASSERT(curcpu() == l->l_pcu_cpu[PCU_FPU]);
    564  1.54    bouyer 	KASSERT(curlwp == l || curlwp->l_pcu_cpu[PCU_FPU] != curcpu());
    565  1.33     skrll 	/*
    566  1.33     skrll 	 * Enable the VFP (so we can read the registers).
    567  1.33     skrll 	 * Make sure the exception bit is cleared so that we can
    568  1.33     skrll 	 * safely dump the registers.
    569  1.33     skrll 	 */
    570  1.21      matt 	armreg_fpexc_write((fpexc | VFP_FPEXC_EN) & ~VFP_FPEXC_EX);
    571   1.1  rearnsha 
    572   1.4      matt 	fregs->vfp_fpexc = fpexc;
    573   1.4      matt 	if (fpexc & VFP_FPEXC_EX) {
    574   1.4      matt 		/* Need to save the exception handling state */
    575  1.21      matt 		fregs->vfp_fpinst = armreg_fpinst_read();
    576  1.21      matt 		if (fpexc & VFP_FPEXC_FP2V)
    577  1.21      matt 			fregs->vfp_fpinst2 = armreg_fpinst2_read();
    578   1.1  rearnsha 	}
    579  1.21      matt 	fregs->vfp_fpscr = armreg_fpscr_read();
    580  1.13      matt 	save_vfpregs(fregs);
    581   1.4      matt 
    582   1.1  rearnsha 	/* Disable the VFP.  */
    583  1.33     skrll 	armreg_fpexc_write(fpexc & ~VFP_FPEXC_EN);
    584   1.1  rearnsha }
    585   1.1  rearnsha 
    586   1.1  rearnsha void
    587  1.39     rmind vfp_state_release(lwp_t *l)
    588   1.1  rearnsha {
    589   1.4      matt 	struct pcb * const pcb = lwp_getpcb(l);
    590   1.1  rearnsha 
    591  1.39     rmind 	/*
    592  1.39     rmind 	 * Now mark the VFP as disabled (and our state
    593  1.39     rmind 	 * has been already saved or is being discarded).
    594  1.39     rmind 	 */
    595  1.39     rmind 	pcb->pcb_vfp.vfp_fpexc &= ~VFP_FPEXC_EN;
    596   1.1  rearnsha 
    597   1.1  rearnsha 	/*
    598   1.4      matt 	 * Turn off the FPU so the next time a VFP instruction is issued
    599   1.4      matt 	 * an exception happens.  We don't know if this LWP's state was
    600   1.4      matt 	 * loaded but if we turned off the FPU for some other LWP, when
    601   1.4      matt 	 * pcu_load invokes vfp_state_load it will see that VFP_FPEXC_EN
    602  1.13      matt 	 * is still set so it just restore fpexc and return since its
    603   1.4      matt 	 * contents are still sitting in the VFP.
    604   1.1  rearnsha 	 */
    605  1.21      matt 	armreg_fpexc_write(armreg_fpexc_read() & ~VFP_FPEXC_EN);
    606   1.1  rearnsha }
    607   1.1  rearnsha 
    608   1.1  rearnsha void
    609  1.51       chs vfp_savecontext(lwp_t *l)
    610   1.1  rearnsha {
    611  1.51       chs 	pcu_save(&arm_vfp_ops, l);
    612   1.1  rearnsha }
    613   1.1  rearnsha 
    614   1.1  rearnsha void
    615  1.51       chs vfp_discardcontext(lwp_t *l, bool used_p)
    616   1.1  rearnsha {
    617  1.51       chs 	pcu_discard(&arm_vfp_ops, l, used_p);
    618  1.25      matt }
    619  1.25      matt 
    620  1.25      matt bool
    621  1.51       chs vfp_used_p(const lwp_t *l)
    622  1.25      matt {
    623  1.51       chs 	return pcu_valid_p(&arm_vfp_ops, l);
    624  1.13      matt }
    625  1.13      matt 
    626  1.13      matt void
    627   1.8      matt vfp_getcontext(struct lwp *l, mcontext_t *mcp, int *flagsp)
    628   1.8      matt {
    629  1.51       chs 	if (vfp_used_p(l)) {
    630   1.8      matt 		const struct pcb * const pcb = lwp_getpcb(l);
    631  1.51       chs 
    632  1.51       chs 		pcu_save(&arm_vfp_ops, l);
    633   1.8      matt 		mcp->__fpu.__vfpregs.__vfp_fpscr = pcb->pcb_vfp.vfp_fpscr;
    634   1.8      matt 		memcpy(mcp->__fpu.__vfpregs.__vfp_fstmx, pcb->pcb_vfp.vfp_regs,
    635   1.8      matt 		    sizeof(mcp->__fpu.__vfpregs.__vfp_fstmx));
    636  1.10      matt 		*flagsp |= _UC_FPU|_UC_ARM_VFP;
    637   1.8      matt 	}
    638   1.8      matt }
    639   1.8      matt 
    640   1.8      matt void
    641   1.8      matt vfp_setcontext(struct lwp *l, const mcontext_t *mcp)
    642   1.8      matt {
    643   1.8      matt 	struct pcb * const pcb = lwp_getpcb(l);
    644  1.51       chs 
    645  1.51       chs 	pcu_discard(&arm_vfp_ops, l, true);
    646   1.8      matt 	pcb->pcb_vfp.vfp_fpscr = mcp->__fpu.__vfpregs.__vfp_fpscr;
    647   1.8      matt 	memcpy(pcb->pcb_vfp.vfp_regs, mcp->__fpu.__vfpregs.__vfp_fstmx,
    648   1.8      matt 	    sizeof(mcp->__fpu.__vfpregs.__vfp_fstmx));
    649   1.8      matt }
    650   1.8      matt 
    651   1.4      matt #endif /* FPU_VFP */
    652