Home | History | Annotate | Line # | Download | only in vfp
vfp_init.c revision 1.17
      1  1.17      matt /*      $NetBSD: vfp_init.c,v 1.17 2013/01/28 23:49:13 matt 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   1.1  rearnsha #include <sys/param.h>
     33   1.1  rearnsha #include <sys/types.h>
     34   1.1  rearnsha #include <sys/systm.h>
     35   1.1  rearnsha #include <sys/device.h>
     36   1.1  rearnsha #include <sys/proc.h>
     37   1.4      matt #include <sys/cpu.h>
     38   1.1  rearnsha 
     39   1.5      matt #include <arm/pcb.h>
     40   1.1  rearnsha #include <arm/undefined.h>
     41   1.1  rearnsha #include <arm/vfpreg.h>
     42   1.8      matt #include <arm/mcontext.h>
     43   1.1  rearnsha 
     44  1.12      matt #include <uvm/uvm_extern.h>		/* for pmap.h */
     45  1.12      matt 
     46   1.1  rearnsha /*
     47   1.1  rearnsha  * Use generic co-processor instructions to avoid assembly problems.
     48   1.1  rearnsha  */
     49   1.1  rearnsha 
     50   1.1  rearnsha /* FMRX <X>, fpsid */
     51   1.4      matt static inline uint32_t
     52   1.4      matt read_fpsid(void)
     53   1.4      matt {
     54   1.4      matt 	uint32_t rv;
     55   1.4      matt 	__asm __volatile("mrc p10, 7, %0, c0, c0, 0" : "=r" (rv));
     56   1.4      matt 	return rv;
     57   1.4      matt }
     58   1.4      matt 
     59   1.4      matt /* FMRX <X>, fpexc */
     60   1.4      matt static inline uint32_t
     61   1.4      matt read_fpscr(void)
     62   1.4      matt {
     63   1.4      matt 	uint32_t rv;
     64   1.4      matt 	__asm __volatile("mrc p10, 7, %0, c1, c0, 0" : "=r" (rv));
     65   1.4      matt 	return rv;
     66   1.4      matt }
     67   1.4      matt 
     68   1.1  rearnsha /* FMRX <X>, fpexc */
     69   1.4      matt static inline uint32_t
     70   1.4      matt read_fpexc(void)
     71   1.4      matt {
     72   1.4      matt 	uint32_t rv;
     73   1.4      matt 	__asm __volatile("mrc p10, 7, %0, c8, c0, 0" : "=r" (rv));
     74   1.4      matt 	return rv;
     75   1.4      matt }
     76   1.4      matt 
     77   1.1  rearnsha /* FMRX <X>, fpinst */
     78   1.4      matt static inline uint32_t
     79   1.4      matt read_fpinst(void)
     80   1.4      matt {
     81   1.4      matt 	uint32_t rv;
     82   1.4      matt 	__asm __volatile("mrc p10, 7, %0, c9, c0, 0" : "=r" (rv));
     83   1.4      matt 	return rv;
     84   1.4      matt }
     85   1.4      matt 
     86   1.1  rearnsha /* FMRX <X>, fpinst2 */
     87   1.4      matt static inline uint32_t
     88   1.4      matt read_fpinst2(void)
     89   1.4      matt {
     90   1.4      matt 	uint32_t rv;
     91   1.4      matt 	__asm __volatile("mrc p10, 7, %0, c10, c0, 0" : "=r" (rv));
     92   1.4      matt 	return rv;
     93   1.4      matt }
     94   1.4      matt 
     95   1.1  rearnsha /* FMXR <X>, fpscr */
     96   1.1  rearnsha #define write_fpscr(X)	__asm __volatile("mcr p10, 7, %0, c1, c0, 0" : \
     97   1.1  rearnsha 			    : "r" (X))
     98   1.1  rearnsha /* FMXR <X>, fpexc */
     99   1.1  rearnsha #define write_fpexc(X)	__asm __volatile("mcr p10, 7, %0, c8, c0, 0" : \
    100   1.1  rearnsha 			    : "r" (X))
    101   1.1  rearnsha /* FMXR <X>, fpinst */
    102   1.1  rearnsha #define write_fpinst(X)	__asm __volatile("mcr p10, 7, %0, c9, c0, 0" : \
    103   1.1  rearnsha 			    : "r" (X))
    104   1.1  rearnsha /* FMXR <X>, fpinst2 */
    105   1.1  rearnsha #define write_fpinst2(X) __asm __volatile("mcr p10, 7, %0, c10, c0, 0" : \
    106   1.1  rearnsha 			    : "r" (X))
    107  1.11      matt 
    108  1.11      matt #ifdef FPU_VFP
    109  1.11      matt 
    110   1.1  rearnsha /* FLDMD <X>, {d0-d15} */
    111  1.11      matt static inline void
    112  1.13      matt load_vfpregs_lo(const uint64_t *p)
    113  1.10      matt {
    114  1.10      matt 	/* vldmia rN, {d0-d15} */
    115  1.10      matt 	__asm __volatile("ldc\tp11, c0, [%0], {32}" :: "r" (p) : "memory");
    116  1.10      matt }
    117  1.10      matt 
    118  1.10      matt /* FSTMD <X>, {d0-d15} */
    119  1.11      matt static inline void
    120  1.10      matt save_vfpregs_lo(uint64_t *p)
    121  1.10      matt {
    122  1.10      matt 	__asm __volatile("stc\tp11, c0, [%0], {32}" :: "r" (p) : "memory");
    123  1.10      matt }
    124  1.10      matt 
    125  1.10      matt #ifdef CPU_CORTEX
    126  1.10      matt /* FLDMD <X>, {d16-d31} */
    127  1.11      matt static inline void
    128  1.13      matt load_vfpregs_hi(const uint64_t *p)
    129  1.10      matt {
    130  1.10      matt 	__asm __volatile("ldcl\tp11, c0, [%0], {32}" :: "r" (&p[16]) : "memory");
    131  1.10      matt }
    132  1.10      matt 
    133  1.10      matt /* FLDMD <X>, {d16-d31} */
    134  1.11      matt static inline void
    135  1.10      matt save_vfpregs_hi(uint64_t *p)
    136  1.10      matt {
    137  1.10      matt 	__asm __volatile("stcl\tp11, c0, [%0], {32}" :: "r" (&p[16]) : "memory");
    138  1.10      matt }
    139  1.10      matt #endif
    140   1.1  rearnsha 
    141  1.13      matt static inline void
    142  1.13      matt load_vfpregs(const struct vfpreg *fregs)
    143  1.13      matt {
    144  1.13      matt 	load_vfpregs_lo(fregs->vfp_regs);
    145  1.13      matt #ifdef CPU_CORTEX
    146  1.13      matt #ifdef CPU_ARM11
    147  1.13      matt 	switch (curcpu()->ci_vfp_id) {
    148  1.13      matt 	case FPU_VFP_CORTEXA5:
    149  1.13      matt 	case FPU_VFP_CORTEXA7:
    150  1.13      matt 	case FPU_VFP_CORTEXA8:
    151  1.13      matt 	case FPU_VFP_CORTEXA9:
    152  1.13      matt #endif
    153  1.13      matt 		load_vfpregs_hi(fregs->vfp_regs);
    154  1.13      matt #ifdef CPU_ARM11
    155  1.13      matt 		break;
    156  1.13      matt 	}
    157  1.13      matt #endif
    158  1.13      matt #endif
    159  1.13      matt }
    160  1.13      matt 
    161  1.13      matt static inline void
    162  1.13      matt save_vfpregs(struct vfpreg *fregs)
    163  1.13      matt {
    164  1.13      matt 	save_vfpregs_lo(fregs->vfp_regs);
    165  1.13      matt #ifdef CPU_CORTEX
    166  1.13      matt #ifdef CPU_ARM11
    167  1.13      matt 	switch (curcpu()->ci_vfp_id) {
    168  1.13      matt 	case FPU_VFP_CORTEXA5:
    169  1.13      matt 	case FPU_VFP_CORTEXA7:
    170  1.13      matt 	case FPU_VFP_CORTEXA8:
    171  1.13      matt 	case FPU_VFP_CORTEXA9:
    172  1.13      matt #endif
    173  1.13      matt 		save_vfpregs_hi(fregs->vfp_regs);
    174  1.13      matt #ifdef CPU_ARM11
    175  1.13      matt 		break;
    176  1.13      matt 	}
    177  1.13      matt #endif
    178  1.13      matt #endif
    179  1.13      matt }
    180  1.13      matt 
    181   1.1  rearnsha /* The real handler for VFP bounces.  */
    182   1.1  rearnsha static int vfp_handler(u_int, u_int, trapframe_t *, int);
    183  1.13      matt #ifdef CPU_CORTEX
    184  1.13      matt static int neon_handler(u_int, u_int, trapframe_t *, int);
    185  1.13      matt #endif
    186   1.1  rearnsha 
    187  1.13      matt static void vfp_state_load(lwp_t *, u_int);
    188  1.13      matt static void vfp_state_save(lwp_t *, u_int);
    189  1.13      matt static void vfp_state_release(lwp_t *, u_int);
    190   1.4      matt 
    191   1.4      matt const pcu_ops_t arm_vfp_ops = {
    192   1.4      matt 	.pcu_id = PCU_FPU,
    193  1.13      matt 	.pcu_state_save = vfp_state_save,
    194   1.4      matt 	.pcu_state_load = vfp_state_load,
    195   1.4      matt 	.pcu_state_release = vfp_state_release,
    196   1.4      matt };
    197   1.1  rearnsha 
    198   1.1  rearnsha struct evcnt vfpevent_use;
    199   1.1  rearnsha struct evcnt vfpevent_reuse;
    200   1.1  rearnsha 
    201   1.1  rearnsha /*
    202   1.1  rearnsha  * Used to test for a VFP. The following function is installed as a coproc10
    203   1.1  rearnsha  * handler on the undefined instruction vector and then we issue a VFP
    204   1.1  rearnsha  * instruction. If undefined_test is non zero then the VFP did not handle
    205   1.1  rearnsha  * the instruction so must be absent, or disabled.
    206   1.1  rearnsha  */
    207   1.1  rearnsha 
    208   1.1  rearnsha static int undefined_test;
    209   1.1  rearnsha 
    210   1.1  rearnsha static int
    211   1.4      matt vfp_test(u_int address, u_int insn, trapframe_t *frame, int fault_code)
    212   1.1  rearnsha {
    213   1.1  rearnsha 
    214   1.1  rearnsha 	frame->tf_pc += INSN_SIZE;
    215   1.1  rearnsha 	++undefined_test;
    216   1.4      matt 	return 0;
    217   1.4      matt }
    218   1.4      matt 
    219   1.4      matt #endif /* FPU_VFP */
    220   1.4      matt 
    221   1.4      matt struct evcnt vfp_fpscr_ev =
    222   1.4      matt     EVCNT_INITIALIZER(EVCNT_TYPE_TRAP, NULL, "VFP", "FPSCR traps");
    223   1.4      matt EVCNT_ATTACH_STATIC(vfp_fpscr_ev);
    224   1.4      matt 
    225   1.4      matt static int
    226   1.4      matt vfp_fpscr_handler(u_int address, u_int insn, trapframe_t *frame, int fault_code)
    227   1.4      matt {
    228   1.4      matt 	struct lwp * const l = curlwp;
    229   1.4      matt 	const u_int regno = (insn >> 12) & 0xf;
    230   1.4      matt 	/*
    231   1.4      matt 	 * Only match move to/from the FPSCR register and we
    232   1.4      matt 	 * can't be using the SP,LR,PC as a source.
    233   1.4      matt 	 */
    234   1.4      matt 	if ((insn & 0xffef0fff) != 0xeee10a10 || regno > 12)
    235   1.4      matt 		return 1;
    236   1.4      matt 
    237   1.4      matt 	struct pcb * const pcb = lwp_getpcb(l);
    238   1.4      matt 
    239   1.4      matt #ifdef FPU_VFP
    240   1.4      matt 	/*
    241   1.4      matt 	 * If FPU is valid somewhere, let's just reenable VFP and
    242   1.4      matt 	 * retry the instruction (only safe thing to do since the
    243   1.4      matt 	 * pcb has a stale copy).
    244   1.4      matt 	 */
    245   1.4      matt 	if (pcb->pcb_vfp.vfp_fpexc & VFP_FPEXC_EN)
    246   1.4      matt 		return 1;
    247   1.4      matt #endif
    248   1.4      matt 
    249   1.4      matt 	if (__predict_false((l->l_md.md_flags & MDLWP_VFPUSED) == 0)) {
    250   1.4      matt 		l->l_md.md_flags |= MDLWP_VFPUSED;
    251   1.4      matt 		pcb->pcb_vfp.vfp_fpscr =
    252   1.4      matt 		    (VFP_FPSCR_DN | VFP_FPSCR_FZ);	/* Runfast */
    253   1.4      matt 	}
    254   1.4      matt 
    255   1.4      matt 	/*
    256   1.4      matt 	 * We know know the pcb has the saved copy.
    257   1.4      matt 	 */
    258   1.4      matt 	register_t * const regp = &frame->tf_r0 + regno;
    259   1.4      matt 	if (insn & 0x00100000) {
    260   1.4      matt 		*regp = pcb->pcb_vfp.vfp_fpscr;
    261   1.4      matt 	} else {
    262   1.4      matt 		pcb->pcb_vfp.vfp_fpscr = *regp;
    263   1.4      matt 	}
    264   1.4      matt 
    265   1.4      matt 	vfp_fpscr_ev.ev_count++;
    266   1.4      matt 
    267   1.4      matt 	frame->tf_pc += INSN_SIZE;
    268   1.4      matt 	return 0;
    269   1.1  rearnsha }
    270   1.1  rearnsha 
    271   1.4      matt #ifndef FPU_VFP
    272   1.4      matt /*
    273   1.4      matt  * If we don't want VFP support, we still need to handle emulating VFP FPSCR
    274   1.4      matt  * instructions.
    275   1.4      matt  */
    276   1.4      matt void
    277   1.4      matt vfp_attach(void)
    278   1.4      matt {
    279   1.4      matt 	install_coproc_handler(VFP_COPROC, vfp_fpscr_handler);
    280   1.4      matt }
    281   1.4      matt 
    282   1.4      matt #else
    283  1.16      matt #if 0
    284  1.12      matt static bool
    285  1.12      matt vfp_patch_branch(uintptr_t code, uintptr_t func, uintptr_t newfunc)
    286  1.12      matt {
    287  1.12      matt 	for (;; code += sizeof(uint32_t)) {
    288  1.12      matt 		uint32_t insn = *(uint32_t *)code;
    289  1.12      matt 		if ((insn & 0xffd08000) == 0xe8908000)	/* ldm ... { pc } */
    290  1.12      matt 			return false;
    291  1.12      matt 		if ((insn & 0xfffffff0) == 0xe12fff10)	/* bx rN */
    292  1.12      matt 			return false;
    293  1.12      matt 		if ((insn & 0xf1a0f000) == 0xe1a0f000)	/* mov pc, ... */
    294  1.12      matt 			return false;
    295  1.12      matt 		if ((insn >> 25) != 0x75)		/* not b/bl insn */
    296  1.12      matt 			continue;
    297  1.12      matt 		intptr_t imm26 = ((int32_t)insn << 8) >> 6;
    298  1.12      matt 		if (code + imm26 + 8 == func) {
    299  1.12      matt 			int32_t imm24 = (newfunc - (code + 8)) >> 2;
    300  1.12      matt 			uint32_t new_insn = (insn & 0xff000000)
    301  1.12      matt 			   | (imm24 & 0xffffff);
    302  1.12      matt 			KASSERTMSG((uint32_t)((imm24 >> 24) + 1) <= 1, "%x",
    303  1.12      matt 			    ((imm24 >> 24) + 1));
    304  1.12      matt 			*(uint32_t *)code = new_insn;
    305  1.12      matt 			cpu_idcache_wbinv_range(code, sizeof(uint32_t));
    306  1.12      matt 			return true;
    307  1.12      matt 		}
    308  1.12      matt 	}
    309  1.12      matt }
    310  1.16      matt #endif
    311  1.12      matt 
    312   1.1  rearnsha void
    313   1.2    cegger vfp_attach(void)
    314   1.1  rearnsha {
    315   1.4      matt 	struct cpu_info * const ci = curcpu();
    316   1.4      matt 	const char *model = NULL;
    317   1.7      matt 	bool vfp_p = false;
    318   1.1  rearnsha 
    319   1.7      matt 	if (CPU_ID_ARM11_P(curcpu()->ci_arm_cpuid)
    320   1.7      matt 	    || CPU_ID_CORTEX_P(curcpu()->ci_arm_cpuid)) {
    321   1.7      matt 		const uint32_t cpacr_vfp = CPACR_CPn(VFP_COPROC);
    322   1.7      matt 		const uint32_t cpacr_vfp2 = CPACR_CPn(VFP_COPROC2);
    323   1.1  rearnsha 
    324   1.7      matt 		/*
    325   1.7      matt 		 * We first need to enable access to the coprocessors.
    326   1.7      matt 		 */
    327   1.7      matt 		uint32_t cpacr = armreg_cpacr_read();
    328   1.7      matt 		cpacr |= __SHIFTIN(CPACR_ALL, cpacr_vfp);
    329   1.7      matt 		cpacr |= __SHIFTIN(CPACR_ALL, cpacr_vfp2);
    330  1.10      matt #if 0
    331   1.9      matt 		if (CPU_ID_CORTEX_P(curcpu()->ci_arm_cpuid)) {
    332   1.9      matt 			/*
    333  1.10      matt 			 * Disable access to the upper 16 FP registers and NEON.
    334   1.9      matt 			 */
    335   1.9      matt 			cpacr |= CPACR_V7_D32DIS;
    336  1.10      matt 			cpacr |= CPACR_V7_ASEDIS;
    337   1.9      matt 		}
    338  1.10      matt #endif
    339   1.7      matt 		armreg_cpacr_write(cpacr);
    340   1.1  rearnsha 
    341   1.7      matt 		/*
    342   1.7      matt 		 * If we could enable them, then they exist.
    343   1.7      matt 		 */
    344   1.7      matt 		cpacr = armreg_cpacr_read();
    345   1.7      matt 		vfp_p = __SHIFTOUT(cpacr, cpacr_vfp2) != CPACR_NOACCESS
    346   1.7      matt 		    || __SHIFTOUT(cpacr, cpacr_vfp) != CPACR_NOACCESS;
    347   1.6      matt 	}
    348   1.6      matt 
    349   1.7      matt 	void *uh = install_coproc_handler(VFP_COPROC, vfp_test);
    350   1.7      matt 
    351   1.7      matt 	undefined_test = 0;
    352   1.7      matt 
    353   1.4      matt 	const uint32_t fpsid = read_fpsid();
    354   1.1  rearnsha 
    355   1.1  rearnsha 	remove_coproc_handler(uh);
    356   1.1  rearnsha 
    357   1.1  rearnsha 	if (undefined_test != 0) {
    358   1.4      matt 		aprint_normal_dev(ci->ci_dev, "No VFP detected\n");
    359   1.4      matt 		install_coproc_handler(VFP_COPROC, vfp_fpscr_handler);
    360   1.4      matt 		ci->ci_vfp_id = 0;
    361   1.1  rearnsha 		return;
    362   1.1  rearnsha 	}
    363   1.1  rearnsha 
    364   1.4      matt 	ci->ci_vfp_id = fpsid;
    365   1.4      matt 	switch (fpsid & ~ VFP_FPSID_REV_MSK) {
    366   1.4      matt 	case FPU_VFP10_ARM10E:
    367   1.4      matt 		model = "VFP10 R1";
    368   1.4      matt 		break;
    369   1.4      matt 	case FPU_VFP11_ARM11:
    370   1.4      matt 		model = "VFP11";
    371   1.4      matt 		break;
    372   1.7      matt 	case FPU_VFP_CORTEXA5:
    373   1.7      matt 	case FPU_VFP_CORTEXA7:
    374   1.7      matt 	case FPU_VFP_CORTEXA8:
    375   1.7      matt 	case FPU_VFP_CORTEXA9:
    376   1.7      matt 		model = "NEON MPE (VFP 3.0+)";
    377   1.6      matt 		break;
    378   1.4      matt 	default:
    379   1.4      matt 		aprint_normal_dev(ci->ci_dev, "unrecognized VFP version %x\n",
    380   1.4      matt 		    fpsid);
    381   1.4      matt 		install_coproc_handler(VFP_COPROC, vfp_fpscr_handler);
    382   1.4      matt 		return;
    383   1.4      matt 	}
    384   1.1  rearnsha 
    385  1.17      matt 	cpu_fpu_present = 1;
    386   1.1  rearnsha 	if (fpsid != 0) {
    387   1.1  rearnsha 		aprint_normal("vfp%d at %s: %s\n",
    388   1.6      matt 		    device_unit(curcpu()->ci_dev), device_xname(curcpu()->ci_dev),
    389   1.1  rearnsha 		    model);
    390   1.1  rearnsha 	}
    391   1.1  rearnsha 	evcnt_attach_dynamic(&vfpevent_use, EVCNT_TYPE_MISC, NULL,
    392  1.12      matt 	    "VFP", "coproc use");
    393   1.1  rearnsha 	evcnt_attach_dynamic(&vfpevent_reuse, EVCNT_TYPE_MISC, NULL,
    394  1.12      matt 	    "VFP", "coproc re-use");
    395   1.1  rearnsha 	install_coproc_handler(VFP_COPROC, vfp_handler);
    396   1.1  rearnsha 	install_coproc_handler(VFP_COPROC2, vfp_handler);
    397  1.13      matt #ifdef CPU_CORTEX
    398  1.13      matt 	install_coproc_handler(CORE_UNKNOWN_HANDLER, neon_handler);
    399  1.13      matt #endif
    400  1.12      matt 
    401  1.16      matt #if 0
    402  1.12      matt 	vfp_patch_branch((uintptr_t)pmap_copy_page_generic,
    403  1.12      matt 	   (uintptr_t)bcopy_page, (uintptr_t)bcopy_page_vfp);
    404  1.12      matt 	vfp_patch_branch((uintptr_t)pmap_zero_page_generic,
    405  1.12      matt 	   (uintptr_t)bzero_page, (uintptr_t)bzero_page_vfp);
    406  1.16      matt #endif
    407   1.1  rearnsha }
    408   1.1  rearnsha 
    409   1.1  rearnsha /* The real handler for VFP bounces.  */
    410   1.4      matt static int
    411   1.4      matt vfp_handler(u_int address, u_int insn, trapframe_t *frame,
    412   1.1  rearnsha     int fault_code)
    413   1.1  rearnsha {
    414   1.4      matt 	struct cpu_info * const ci = curcpu();
    415   1.1  rearnsha 
    416   1.1  rearnsha 	/* This shouldn't ever happen.  */
    417   1.1  rearnsha 	if (fault_code != FAULT_USER)
    418  1.14      matt 		panic("VFP fault at %#x in non-user mode", frame->tf_pc);
    419   1.1  rearnsha 
    420   1.4      matt 	if (ci->ci_vfp_id == 0)
    421   1.1  rearnsha 		/* No VFP detected, just fault.  */
    422   1.1  rearnsha 		return 1;
    423   1.1  rearnsha 
    424   1.4      matt 	/*
    425   1.4      matt 	 * If we are just changing/fetching FPSCR, don't bother loading it.
    426   1.4      matt 	 */
    427   1.4      matt 	if (!vfp_fpscr_handler(address, insn, frame, fault_code))
    428   1.4      matt 		return 0;
    429   1.1  rearnsha 
    430   1.4      matt 	pcu_load(&arm_vfp_ops);
    431   1.3     rmind 
    432   1.4      matt 	/* Need to restart the faulted instruction.  */
    433   1.4      matt //	frame->tf_pc -= INSN_SIZE;
    434   1.4      matt 	return 0;
    435   1.4      matt }
    436   1.1  rearnsha 
    437  1.13      matt #ifdef CPU_CORTEX
    438  1.13      matt /* The real handler for NEON bounces.  */
    439  1.13      matt static int
    440  1.13      matt neon_handler(u_int address, u_int insn, trapframe_t *frame,
    441  1.13      matt     int fault_code)
    442  1.13      matt {
    443  1.13      matt 	struct cpu_info * const ci = curcpu();
    444  1.13      matt 
    445  1.13      matt 	if (ci->ci_vfp_id == 0)
    446  1.13      matt 		/* No VFP detected, just fault.  */
    447  1.13      matt 		return 1;
    448  1.13      matt 
    449  1.13      matt 	if ((insn & 0xfe000000) != 0xf2000000
    450  1.13      matt 	    && (insn & 0xfe000000) != 0xf4000000)
    451  1.13      matt 		/* Not NEON instruction, just fault.  */
    452  1.13      matt 		return 1;
    453  1.13      matt 
    454  1.13      matt 	/* This shouldn't ever happen.  */
    455  1.13      matt 	if (fault_code != FAULT_USER)
    456  1.13      matt 		panic("NEON fault in non-user mode");
    457  1.13      matt 
    458  1.13      matt 	pcu_load(&arm_vfp_ops);
    459  1.13      matt 
    460  1.13      matt 	/* Need to restart the faulted instruction.  */
    461  1.13      matt //	frame->tf_pc -= INSN_SIZE;
    462  1.13      matt 	return 0;
    463  1.13      matt }
    464  1.13      matt #endif
    465  1.13      matt 
    466   1.4      matt static void
    467  1.13      matt vfp_state_load(lwp_t *l, u_int flags)
    468   1.4      matt {
    469   1.4      matt 	struct pcb * const pcb = lwp_getpcb(l);
    470  1.13      matt 
    471  1.13      matt 	KASSERT(flags & PCU_ENABLE);
    472  1.13      matt 
    473  1.13      matt 	if (flags & PCU_KERNEL) {
    474  1.13      matt 		if ((flags & PCU_LOADED) == 0) {
    475  1.13      matt 			pcb->pcb_kernel_vfp.vfp_fpexc = pcb->pcb_vfp.vfp_fpexc;
    476  1.13      matt 		}
    477  1.15      matt 		pcb->pcb_vfp.vfp_fpexc = VFP_FPEXC_EN;
    478  1.13      matt 		write_fpexc(pcb->pcb_vfp.vfp_fpexc);
    479  1.13      matt 		/*
    480  1.13      matt 		 * Load the kernel registers (just the first 16) if they've
    481  1.13      matt 		 * been used..
    482  1.13      matt 		 */
    483  1.13      matt 		if (flags & PCU_LOADED) {
    484  1.13      matt 			load_vfpregs_lo(pcb->pcb_kernel_vfp.vfp_regs);
    485  1.13      matt 		}
    486  1.13      matt 		return;
    487  1.13      matt 	}
    488   1.4      matt 	struct vfpreg * const fregs = &pcb->pcb_vfp;
    489   1.1  rearnsha 
    490   1.1  rearnsha 	/*
    491   1.1  rearnsha 	 * Instrument VFP usage -- if a process has not previously
    492   1.1  rearnsha 	 * used the VFP, mark it as having used VFP for the first time,
    493   1.1  rearnsha 	 * and count this event.
    494   1.1  rearnsha 	 *
    495   1.1  rearnsha 	 * If a process has used the VFP, count a "used VFP, and took
    496   1.1  rearnsha 	 * a trap to use it again" event.
    497   1.1  rearnsha 	 */
    498   1.4      matt 	if (__predict_false((l->l_md.md_flags & MDLWP_VFPUSED) == 0)) {
    499   1.1  rearnsha 		vfpevent_use.ev_count++;
    500   1.4      matt 		l->l_md.md_flags |= MDLWP_VFPUSED;
    501   1.3     rmind 		pcb->pcb_vfp.vfp_fpscr =
    502   1.1  rearnsha 		    (VFP_FPSCR_DN | VFP_FPSCR_FZ);	/* Runfast */
    503   1.4      matt 	} else {
    504   1.1  rearnsha 		vfpevent_reuse.ev_count++;
    505   1.4      matt 	}
    506   1.1  rearnsha 
    507   1.4      matt 	if (fregs->vfp_fpexc & VFP_FPEXC_EN) {
    508   1.4      matt 		/*
    509   1.4      matt 		 * If we think the VFP is enabled, it must have be disabled by
    510   1.4      matt 		 * vfp_state_release for another LWP so we can just restore
    511   1.4      matt 		 * FPEXC and return since our VFP state is still loaded.
    512   1.4      matt 		 */
    513   1.4      matt 		write_fpexc(fregs->vfp_fpexc);
    514   1.4      matt 		return;
    515   1.4      matt 	}
    516   1.1  rearnsha 
    517  1.13      matt 	/* Load and Enable the VFP (so that we can write the registers).  */
    518  1.13      matt 	if (flags & PCU_RELOAD) {
    519  1.13      matt 		uint32_t fpexc = read_fpexc();
    520  1.13      matt 		KDASSERT((fpexc & VFP_FPEXC_EX) == 0);
    521  1.13      matt 		write_fpexc(fpexc | VFP_FPEXC_EN);
    522  1.13      matt 
    523  1.13      matt 		load_vfpregs(fregs);
    524  1.13      matt 		write_fpscr(fregs->vfp_fpscr);
    525  1.13      matt 
    526  1.13      matt 		if (fregs->vfp_fpexc & VFP_FPEXC_EX) {
    527  1.13      matt 			struct cpu_info * const ci = curcpu();
    528  1.13      matt 			/* Need to restore the exception handling state.  */
    529  1.13      matt 			switch (ci->ci_vfp_id) {
    530  1.13      matt 			case FPU_VFP10_ARM10E:
    531  1.13      matt 			case FPU_VFP11_ARM11:
    532  1.13      matt 			case FPU_VFP_CORTEXA5:
    533  1.13      matt 			case FPU_VFP_CORTEXA7:
    534  1.13      matt 			case FPU_VFP_CORTEXA8:
    535  1.13      matt 			case FPU_VFP_CORTEXA9:
    536  1.13      matt 				write_fpinst2(fregs->vfp_fpinst2);
    537  1.13      matt 				write_fpinst(fregs->vfp_fpinst);
    538  1.13      matt 				break;
    539  1.13      matt 			default:
    540  1.13      matt 				panic("%s: Unsupported VFP %#x",
    541  1.13      matt 				    __func__, ci->ci_vfp_id);
    542  1.13      matt 			}
    543   1.1  rearnsha 		}
    544   1.1  rearnsha 	}
    545   1.4      matt 
    546   1.4      matt 	/* Finally, restore the FPEXC but don't enable the VFP. */
    547   1.4      matt 	fregs->vfp_fpexc |= VFP_FPEXC_EN;
    548   1.4      matt 	write_fpexc(fregs->vfp_fpexc);
    549   1.1  rearnsha }
    550   1.1  rearnsha 
    551   1.1  rearnsha void
    552  1.13      matt vfp_state_save(lwp_t *l, u_int flags)
    553   1.1  rearnsha {
    554   1.4      matt 	struct pcb * const pcb = lwp_getpcb(l);
    555  1.13      matt 	uint32_t fpexc = read_fpexc();
    556  1.13      matt 	write_fpexc((fpexc | VFP_FPEXC_EN) & ~VFP_FPEXC_EX);
    557   1.1  rearnsha 
    558  1.13      matt 	if (flags & PCU_KERNEL) {
    559  1.13      matt 		/*
    560  1.13      matt 		 * Save the kernel set of VFP registers.
    561  1.13      matt 		 * (just the first 16).
    562  1.13      matt 		 */
    563  1.13      matt 		save_vfpregs_lo(pcb->pcb_kernel_vfp.vfp_regs);
    564   1.1  rearnsha 		return;
    565  1.13      matt 	}
    566  1.13      matt 
    567  1.13      matt 	struct vfpreg * const fregs = &pcb->pcb_vfp;
    568   1.1  rearnsha 
    569   1.4      matt 	/*
    570   1.4      matt 	 * Enable the VFP (so we can read the registers).
    571   1.4      matt 	 * Make sure the exception bit is cleared so that we can
    572   1.4      matt 	 * safely dump the registers.
    573   1.4      matt 	 */
    574   1.4      matt 	fregs->vfp_fpexc = fpexc;
    575   1.4      matt 	if (fpexc & VFP_FPEXC_EX) {
    576   1.4      matt 		struct cpu_info * const ci = curcpu();
    577   1.4      matt 		/* Need to save the exception handling state */
    578   1.4      matt 		switch (ci->ci_vfp_id) {
    579   1.4      matt 		case FPU_VFP10_ARM10E:
    580   1.4      matt 		case FPU_VFP11_ARM11:
    581   1.8      matt 		case FPU_VFP_CORTEXA5:
    582   1.8      matt 		case FPU_VFP_CORTEXA7:
    583   1.8      matt 		case FPU_VFP_CORTEXA8:
    584   1.8      matt 		case FPU_VFP_CORTEXA9:
    585   1.4      matt 			fregs->vfp_fpinst = read_fpinst();
    586   1.4      matt 			fregs->vfp_fpinst2 = read_fpinst2();
    587   1.4      matt 			break;
    588   1.4      matt 		default:
    589   1.4      matt 			panic("%s: Unsupported VFP %#x",
    590   1.4      matt 			    __func__, ci->ci_vfp_id);
    591   1.1  rearnsha 		}
    592   1.1  rearnsha 	}
    593   1.4      matt 	fregs->vfp_fpscr = read_fpscr();
    594  1.13      matt 	save_vfpregs(fregs);
    595   1.4      matt 
    596   1.1  rearnsha 	/* Disable the VFP.  */
    597   1.4      matt 	write_fpexc(fpexc);
    598   1.1  rearnsha }
    599   1.1  rearnsha 
    600   1.1  rearnsha void
    601  1.13      matt vfp_state_release(lwp_t *l, u_int flags)
    602   1.1  rearnsha {
    603   1.4      matt 	struct pcb * const pcb = lwp_getpcb(l);
    604   1.1  rearnsha 
    605  1.13      matt 	if (flags & PCU_KERNEL) {
    606  1.13      matt 		/*
    607  1.13      matt 		 * Restore the FPEXC since we borrowed that field.
    608  1.13      matt 		 */
    609  1.13      matt 		pcb->pcb_vfp.vfp_fpexc = pcb->pcb_kernel_vfp.vfp_fpexc;
    610  1.13      matt 	} else {
    611  1.13      matt 		/*
    612  1.13      matt 		 * Now mark the VFP as disabled (and our state
    613  1.13      matt 		 * has been already saved or is being discarded).
    614  1.13      matt 		 */
    615  1.13      matt 		pcb->pcb_vfp.vfp_fpexc &= ~VFP_FPEXC_EN;
    616  1.13      matt 	}
    617   1.1  rearnsha 
    618   1.1  rearnsha 	/*
    619   1.4      matt 	 * Turn off the FPU so the next time a VFP instruction is issued
    620   1.4      matt 	 * an exception happens.  We don't know if this LWP's state was
    621   1.4      matt 	 * loaded but if we turned off the FPU for some other LWP, when
    622   1.4      matt 	 * pcu_load invokes vfp_state_load it will see that VFP_FPEXC_EN
    623  1.13      matt 	 * is still set so it just restore fpexc and return since its
    624   1.4      matt 	 * contents are still sitting in the VFP.
    625   1.1  rearnsha 	 */
    626   1.4      matt 	write_fpexc(read_fpexc() & ~VFP_FPEXC_EN);
    627   1.1  rearnsha }
    628   1.1  rearnsha 
    629   1.1  rearnsha void
    630   1.2    cegger vfp_savecontext(void)
    631   1.1  rearnsha {
    632   1.4      matt 	pcu_save(&arm_vfp_ops);
    633   1.1  rearnsha }
    634   1.1  rearnsha 
    635   1.1  rearnsha void
    636   1.4      matt vfp_discardcontext(void)
    637   1.1  rearnsha {
    638   1.4      matt 	pcu_discard(&arm_vfp_ops);
    639   1.4      matt }
    640   1.1  rearnsha 
    641   1.8      matt void
    642  1.13      matt vfp_kernel_acquire(void)
    643  1.13      matt {
    644  1.13      matt 	if (__predict_false(cpu_intr_p())) {
    645  1.13      matt 		write_fpexc(VFP_FPEXC_EN);
    646  1.13      matt 		if (curcpu()->ci_data.cpu_pcu_curlwp[PCU_FPU] != NULL) {
    647  1.13      matt 			lwp_t * const l = curlwp;
    648  1.13      matt 			struct pcb * const pcb = lwp_getpcb(l);
    649  1.13      matt 			KASSERT((l->l_md.md_flags & MDLWP_VFPINTR) == 0);
    650  1.13      matt 			l->l_md.md_flags |= MDLWP_VFPINTR;
    651  1.13      matt 			save_vfpregs_lo(&pcb->pcb_kernel_vfp.vfp_regs[16]);
    652  1.13      matt 		}
    653  1.13      matt 	} else {
    654  1.13      matt 		pcu_kernel_acquire(&arm_vfp_ops);
    655  1.13      matt 	}
    656  1.13      matt }
    657  1.13      matt 
    658  1.13      matt void
    659  1.13      matt vfp_kernel_release(void)
    660  1.13      matt {
    661  1.13      matt 	if (__predict_false(cpu_intr_p())) {
    662  1.13      matt 		uint32_t fpexc = 0;
    663  1.13      matt 		if (curcpu()->ci_data.cpu_pcu_curlwp[PCU_FPU] != NULL) {
    664  1.13      matt 			lwp_t * const l = curlwp;
    665  1.13      matt 			struct pcb * const pcb = lwp_getpcb(l);
    666  1.13      matt 			KASSERT(l->l_md.md_flags & MDLWP_VFPINTR);
    667  1.13      matt 			load_vfpregs_lo(&pcb->pcb_kernel_vfp.vfp_regs[16]);
    668  1.13      matt 			l->l_md.md_flags &= ~MDLWP_VFPINTR;
    669  1.13      matt 			fpexc = pcb->pcb_vfp.vfp_fpexc;
    670  1.13      matt 		}
    671  1.13      matt 		write_fpexc(fpexc);
    672  1.13      matt 	} else {
    673  1.13      matt 		pcu_kernel_release(&arm_vfp_ops);
    674  1.13      matt 	}
    675  1.13      matt }
    676  1.13      matt 
    677  1.13      matt void
    678   1.8      matt vfp_getcontext(struct lwp *l, mcontext_t *mcp, int *flagsp)
    679   1.8      matt {
    680   1.8      matt 	if (l->l_md.md_flags & MDLWP_VFPUSED) {
    681   1.8      matt 		const struct pcb * const pcb = lwp_getpcb(l);
    682   1.8      matt 		pcu_save(&arm_vfp_ops);
    683   1.8      matt 		mcp->__fpu.__vfpregs.__vfp_fpscr = pcb->pcb_vfp.vfp_fpscr;
    684   1.8      matt 		memcpy(mcp->__fpu.__vfpregs.__vfp_fstmx, pcb->pcb_vfp.vfp_regs,
    685   1.8      matt 		    sizeof(mcp->__fpu.__vfpregs.__vfp_fstmx));
    686  1.10      matt 		*flagsp |= _UC_FPU|_UC_ARM_VFP;
    687   1.8      matt 	}
    688   1.8      matt }
    689   1.8      matt 
    690   1.8      matt void
    691   1.8      matt vfp_setcontext(struct lwp *l, const mcontext_t *mcp)
    692   1.8      matt {
    693   1.8      matt 	pcu_discard(&arm_vfp_ops);
    694   1.8      matt 	struct pcb * const pcb = lwp_getpcb(l);
    695   1.8      matt 	l->l_md.md_flags |= MDLWP_VFPUSED;
    696   1.8      matt 	pcb->pcb_vfp.vfp_fpscr = mcp->__fpu.__vfpregs.__vfp_fpscr;
    697   1.8      matt 	memcpy(pcb->pcb_vfp.vfp_regs, mcp->__fpu.__vfpregs.__vfp_fstmx,
    698   1.8      matt 	    sizeof(mcp->__fpu.__vfpregs.__vfp_fstmx));
    699   1.8      matt }
    700   1.8      matt 
    701   1.4      matt #endif /* FPU_VFP */
    702