xref: /freebsd/sys/i386/i386/npx.c (revision df615735960370255d3acc4ac2a6f4fd297b7461)
1 /*-
2  * Copyright (c) 1990 William Jolitz.
3  * Copyright (c) 1991 The Regents of the University of California.
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  * 3. Neither the name of the University nor the names of its contributors
15  *    may be used to endorse or promote products derived from this software
16  *    without specific prior written permission.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  */
30 
31 #include <sys/cdefs.h>
32 #include "opt_cpu.h"
33 #include "opt_isa.h"
34 #include "opt_npx.h"
35 
36 #include <sys/param.h>
37 #include <sys/systm.h>
38 #include <sys/bus.h>
39 #include <sys/kernel.h>
40 #include <sys/lock.h>
41 #include <sys/malloc.h>
42 #include <sys/module.h>
43 #include <sys/mutex.h>
44 #include <sys/mutex.h>
45 #include <sys/proc.h>
46 #include <sys/smp.h>
47 #include <sys/sysctl.h>
48 #include <machine/bus.h>
49 #include <sys/rman.h>
50 #ifdef NPX_DEBUG
51 #include <sys/syslog.h>
52 #endif
53 #include <sys/signalvar.h>
54 #include <vm/uma.h>
55 
56 #include <machine/asmacros.h>
57 #include <machine/cputypes.h>
58 #include <machine/frame.h>
59 #include <machine/md_var.h>
60 #include <machine/pcb.h>
61 #include <machine/psl.h>
62 #include <machine/resource.h>
63 #include <machine/specialreg.h>
64 #include <machine/segments.h>
65 #include <machine/ucontext.h>
66 #include <x86/ifunc.h>
67 
68 #include <machine/intr_machdep.h>
69 
70 #ifdef DEV_ISA
71 #include <isa/isavar.h>
72 #endif
73 
74 /*
75  * 387 and 287 Numeric Coprocessor Extension (NPX) Driver.
76  */
77 
78 #define	fldcw(cw)		__asm __volatile("fldcw %0" : : "m" (cw))
79 #define	fnclex()		__asm __volatile("fnclex")
80 #define	fninit()		__asm __volatile("fninit")
81 #define	fnsave(addr)		__asm __volatile("fnsave %0" : "=m" (*(addr)))
82 #define	fnstcw(addr)		__asm __volatile("fnstcw %0" : "=m" (*(addr)))
83 #define	fnstsw(addr)		__asm __volatile("fnstsw %0" : "=am" (*(addr)))
84 #define	fp_divide_by_0()	__asm __volatile( \
85 				    "fldz; fld1; fdiv %st,%st(1); fnop")
86 #define	frstor(addr)		__asm __volatile("frstor %0" : : "m" (*(addr)))
87 #define	fxrstor(addr)		__asm __volatile("fxrstor %0" : : "m" (*(addr)))
88 #define	fxsave(addr)		__asm __volatile("fxsave %0" : "=m" (*(addr)))
89 #define	ldmxcsr(csr)		__asm __volatile("ldmxcsr %0" : : "m" (csr))
90 #define	stmxcsr(addr)		__asm __volatile("stmxcsr %0" : : "m" (*(addr)))
91 
92 static __inline void
xrstor(char * addr,uint64_t mask)93 xrstor(char *addr, uint64_t mask)
94 {
95 	uint32_t low, hi;
96 
97 	low = mask;
98 	hi = mask >> 32;
99 	__asm __volatile("xrstor %0" : : "m" (*addr), "a" (low), "d" (hi));
100 }
101 
102 static __inline void
xsave(char * addr,uint64_t mask)103 xsave(char *addr, uint64_t mask)
104 {
105 	uint32_t low, hi;
106 
107 	low = mask;
108 	hi = mask >> 32;
109 	__asm __volatile("xsave %0" : "=m" (*addr) : "a" (low), "d" (hi) :
110 	    "memory");
111 }
112 
113 static __inline void
xsaveopt(char * addr,uint64_t mask)114 xsaveopt(char *addr, uint64_t mask)
115 {
116 	uint32_t low, hi;
117 
118 	low = mask;
119 	hi = mask >> 32;
120 	__asm __volatile("xsaveopt %0" : "=m" (*addr) : "a" (low), "d" (hi) :
121 	    "memory");
122 }
123 
124 #define GET_FPU_CW(thread) \
125 	(cpu_fxsr ? \
126 		(thread)->td_pcb->pcb_save->sv_xmm.sv_env.en_cw : \
127 		(thread)->td_pcb->pcb_save->sv_87.sv_env.en_cw)
128 #define GET_FPU_SW(thread) \
129 	(cpu_fxsr ? \
130 		(thread)->td_pcb->pcb_save->sv_xmm.sv_env.en_sw : \
131 		(thread)->td_pcb->pcb_save->sv_87.sv_env.en_sw)
132 #define SET_FPU_CW(savefpu, value) do { \
133 	if (cpu_fxsr) \
134 		(savefpu)->sv_xmm.sv_env.en_cw = (value); \
135 	else \
136 		(savefpu)->sv_87.sv_env.en_cw = (value); \
137 } while (0)
138 
139 CTASSERT(sizeof(union savefpu) == 512);
140 CTASSERT(sizeof(struct xstate_hdr) == 64);
141 CTASSERT(sizeof(struct savefpu_ymm) == 832);
142 
143 /*
144  * This requirement is to make it easier for asm code to calculate
145  * offset of the fpu save area from the pcb address. FPU save area
146  * must be 64-byte aligned.
147  */
148 CTASSERT(sizeof(struct pcb) % XSAVE_AREA_ALIGN == 0);
149 
150 /*
151  * Ensure the copy of XCR0 saved in a core is contained in the padding
152  * area.
153  */
154 CTASSERT(X86_XSTATE_XCR0_OFFSET >= offsetof(struct savexmm, sv_pad) &&
155     X86_XSTATE_XCR0_OFFSET + sizeof(uint64_t) <= sizeof(struct savexmm));
156 
157 static	void	fpu_clean_state(void);
158 
159 static	void	fpurstor(union savefpu *);
160 
161 int	hw_float;
162 
163 SYSCTL_INT(_hw, HW_FLOATINGPT, floatingpoint, CTLFLAG_RD,
164     &hw_float, 0, "Floating point instructions executed in hardware");
165 
166 int lazy_fpu_switch = 0;
167 SYSCTL_INT(_hw, OID_AUTO, lazy_fpu_switch, CTLFLAG_RWTUN | CTLFLAG_NOFETCH,
168     &lazy_fpu_switch, 0,
169     "Lazily load FPU context after context switch");
170 
171 u_int cpu_fxsr;		/* SSE enabled */
172 int use_xsave;
173 uint64_t xsave_mask;
174 static	uma_zone_t fpu_save_area_zone;
175 static	union savefpu *npx_initialstate;
176 
177 static struct xsave_area_elm_descr {
178 	u_int	offset;
179 	u_int	size;
180 } *xsave_area_desc;
181 
182 static	volatile u_int		npx_traps_while_probing;
183 
184 alias_for_inthand_t probetrap;
185 __asm("								\n\
186 	.text							\n\
187 	.p2align 2,0x90						\n\
188 	.type	" __XSTRING(CNAME(probetrap)) ",@function	\n\
189 " __XSTRING(CNAME(probetrap)) ":				\n\
190 	ss							\n\
191 	incl	" __XSTRING(CNAME(npx_traps_while_probing)) "	\n\
192 	fnclex							\n\
193 	iret							\n\
194 ");
195 
196 /*
197  * Determine if an FPU is present and how to use it.
198  */
199 static int
npx_probe(void)200 npx_probe(void)
201 {
202 	struct gate_descriptor save_idt_npxtrap;
203 	u_short control, status;
204 
205 	/*
206 	 * Modern CPUs all have an FPU that uses the INT16 interface
207 	 * and provide a simple way to verify that, so handle the
208 	 * common case right away.
209 	 */
210 	if (cpu_feature & CPUID_FPU) {
211 		hw_float = 1;
212 		return (1);
213 	}
214 
215 	save_idt_npxtrap = idt[IDT_MF];
216 	setidt(IDT_MF, probetrap, SDT_SYS386TGT, SEL_KPL,
217 	    GSEL(GCODE_SEL, SEL_KPL));
218 
219 	/*
220 	 * Don't trap while we're probing.
221 	 */
222 	fpu_enable();
223 
224 	/*
225 	 * Finish resetting the coprocessor, if any.  If there is an error
226 	 * pending, then we may get a bogus IRQ13, but npx_intr() will handle
227 	 * it OK.  Bogus halts have never been observed, but we enabled
228 	 * IRQ13 and cleared the BUSY# latch early to handle them anyway.
229 	 */
230 	fninit();
231 
232 	/*
233 	 * Don't use fwait here because it might hang.
234 	 * Don't use fnop here because it usually hangs if there is no FPU.
235 	 */
236 	DELAY(1000);		/* wait for any IRQ13 */
237 #ifdef DIAGNOSTIC
238 	if (npx_traps_while_probing != 0)
239 		printf("fninit caused %u bogus npx trap(s)\n",
240 		       npx_traps_while_probing);
241 #endif
242 	/*
243 	 * Check for a status of mostly zero.
244 	 */
245 	status = 0x5a5a;
246 	fnstsw(&status);
247 	if ((status & 0xb8ff) == 0) {
248 		/*
249 		 * Good, now check for a proper control word.
250 		 */
251 		control = 0x5a5a;
252 		fnstcw(&control);
253 		if ((control & 0x1f3f) == 0x033f) {
254 			/*
255 			 * We have an npx, now divide by 0 to see if exception
256 			 * 16 works.
257 			 */
258 			control &= ~(1 << 2);	/* enable divide by 0 trap */
259 			fldcw(control);
260 			npx_traps_while_probing = 0;
261 			fp_divide_by_0();
262 			if (npx_traps_while_probing != 0) {
263 				/*
264 				 * Good, exception 16 works.
265 				 */
266 				hw_float = 1;
267 				goto cleanup;
268 			}
269 			printf(
270 	"FPU does not use exception 16 for error reporting\n");
271 			goto cleanup;
272 		}
273 	}
274 
275 	/*
276 	 * Probe failed.  Floating point simply won't work.
277 	 * Notify user and disable FPU/MMX/SSE instruction execution.
278 	 */
279 	printf("WARNING: no FPU!\n");
280 	__asm __volatile("smsw %%ax; orb %0,%%al; lmsw %%ax" : :
281 	    "n" (CR0_EM | CR0_MP) : "ax");
282 
283 cleanup:
284 	idt[IDT_MF] = save_idt_npxtrap;
285 	return (hw_float);
286 }
287 
288 static void
fpusave_xsaveopt(union savefpu * addr)289 fpusave_xsaveopt(union savefpu *addr)
290 {
291 
292 	xsaveopt((char *)addr, xsave_mask);
293 }
294 
295 static void
fpusave_xsave(union savefpu * addr)296 fpusave_xsave(union savefpu *addr)
297 {
298 
299 	xsave((char *)addr, xsave_mask);
300 }
301 
302 static void
fpusave_fxsave(union savefpu * addr)303 fpusave_fxsave(union savefpu *addr)
304 {
305 
306 	fxsave((char *)addr);
307 }
308 
309 static void
fpusave_fnsave(union savefpu * addr)310 fpusave_fnsave(union savefpu *addr)
311 {
312 
313 	fnsave((char *)addr);
314 }
315 
316 DEFINE_IFUNC(, void, fpusave, (union savefpu *))
317 {
318 	u_int cp[4];
319 
320 	if (use_xsave) {
321 		cpuid_count(0xd, 0x1, cp);
322 		return ((cp[0] & CPUID_EXTSTATE_XSAVEOPT) != 0 ?
323 		    fpusave_xsaveopt : fpusave_xsave);
324 	}
325 	if (cpu_fxsr)
326 		return (fpusave_fxsave);
327 	return (fpusave_fnsave);
328 }
329 
330 /*
331  * Enable XSAVE if supported and allowed by user.
332  * Calculate the xsave_mask.
333  */
334 static void
npxinit_bsp1(void)335 npxinit_bsp1(void)
336 {
337 	u_int cp[4];
338 	uint64_t xsave_mask_user;
339 
340 	TUNABLE_INT_FETCH("hw.lazy_fpu_switch", &lazy_fpu_switch);
341 	if (!use_xsave)
342 		return;
343 	cpuid_count(0xd, 0x0, cp);
344 	xsave_mask = XFEATURE_ENABLED_X87 | XFEATURE_ENABLED_SSE;
345 	if ((cp[0] & xsave_mask) != xsave_mask)
346 		panic("CPU0 does not support X87 or SSE: %x", cp[0]);
347 	xsave_mask = ((uint64_t)cp[3] << 32) | cp[0];
348 	xsave_mask_user = xsave_mask;
349 	TUNABLE_QUAD_FETCH("hw.xsave_mask", &xsave_mask_user);
350 	xsave_mask_user |= XFEATURE_ENABLED_X87 | XFEATURE_ENABLED_SSE;
351 	xsave_mask &= xsave_mask_user;
352 	if ((xsave_mask & XFEATURE_AVX512) != XFEATURE_AVX512)
353 		xsave_mask &= ~XFEATURE_AVX512;
354 	if ((xsave_mask & XFEATURE_MPX) != XFEATURE_MPX)
355 		xsave_mask &= ~XFEATURE_MPX;
356 }
357 
358 /*
359  * Calculate the fpu save area size.
360  */
361 static void
npxinit_bsp2(void)362 npxinit_bsp2(void)
363 {
364 	u_int cp[4];
365 
366 	if (use_xsave) {
367 		cpuid_count(0xd, 0x0, cp);
368 		cpu_max_ext_state_size = cp[1];
369 
370 		/*
371 		 * Reload the cpu_feature2, since we enabled OSXSAVE.
372 		 */
373 		do_cpuid(1, cp);
374 		cpu_feature2 = cp[2];
375 	} else
376 		cpu_max_ext_state_size = sizeof(union savefpu);
377 }
378 
379 /*
380  * Initialize floating point unit.
381  */
382 void
npxinit(bool bsp)383 npxinit(bool bsp)
384 {
385 	static union savefpu dummy;
386 	register_t saveintr;
387 	u_int mxcsr;
388 	u_short control;
389 
390 	if (bsp) {
391 		if (!npx_probe())
392 			return;
393 		npxinit_bsp1();
394 	}
395 
396 	if (use_xsave) {
397 		load_cr4(rcr4() | CR4_XSAVE);
398 		load_xcr(XCR0, xsave_mask);
399 	}
400 
401 	/*
402 	 * XCR0 shall be set up before CPU can report the save area size.
403 	 */
404 	if (bsp)
405 		npxinit_bsp2();
406 
407 	/*
408 	 * fninit has the same h/w bugs as fnsave.  Use the detoxified
409 	 * fnsave to throw away any junk in the fpu.  fpusave() initializes
410 	 * the fpu.
411 	 *
412 	 * It is too early for critical_enter() to work on AP.
413 	 */
414 	saveintr = intr_disable();
415 	fpu_enable();
416 	if (cpu_fxsr)
417 		fninit();
418 	else
419 		fnsave(&dummy);
420 	control = __INITIAL_NPXCW__;
421 	fldcw(control);
422 	if (cpu_fxsr) {
423 		mxcsr = __INITIAL_MXCSR__;
424 		ldmxcsr(mxcsr);
425 	}
426 	fpu_disable();
427 	intr_restore(saveintr);
428 }
429 
430 /*
431  * On the boot CPU we generate a clean state that is used to
432  * initialize the floating point unit when it is first used by a
433  * process.
434  */
435 static void
npxinitstate(void * arg __unused)436 npxinitstate(void *arg __unused)
437 {
438 	uint64_t *xstate_bv;
439 	register_t saveintr;
440 	int cp[4], i, max_ext_n;
441 
442 	if (!hw_float)
443 		return;
444 
445 	/* Do potentially blocking operations before disabling interrupts. */
446 	fpu_save_area_zone = uma_zcreate("FPU_save_area",
447 	    cpu_max_ext_state_size, NULL, NULL, NULL, NULL,
448 	    XSAVE_AREA_ALIGN - 1, 0);
449 	npx_initialstate = uma_zalloc(fpu_save_area_zone, M_WAITOK | M_ZERO);
450 	if (use_xsave) {
451 		if (xsave_mask >> 32 != 0)
452 			max_ext_n = fls(xsave_mask >> 32) + 32;
453 		else
454 			max_ext_n = fls(xsave_mask);
455 		xsave_area_desc = malloc(max_ext_n * sizeof(struct
456 		    xsave_area_elm_descr), M_DEVBUF, M_WAITOK | M_ZERO);
457 	}
458 
459 	saveintr = intr_disable();
460 	fpu_enable();
461 
462 	if (cpu_fxsr)
463 		fpusave_fxsave(npx_initialstate);
464 	else
465 		fpusave_fnsave(npx_initialstate);
466 	if (cpu_fxsr) {
467 		if (npx_initialstate->sv_xmm.sv_env.en_mxcsr_mask)
468 			cpu_mxcsr_mask =
469 			    npx_initialstate->sv_xmm.sv_env.en_mxcsr_mask;
470 		else
471 			cpu_mxcsr_mask = 0xFFBF;
472 
473 		/*
474 		 * The fninit instruction does not modify XMM
475 		 * registers or x87 registers (MM/ST).  The fpusave
476 		 * call dumped the garbage contained in the registers
477 		 * after reset to the initial state saved.  Clear XMM
478 		 * and x87 registers file image to make the startup
479 		 * program state and signal handler XMM/x87 register
480 		 * content predictable.
481 		 */
482 		bzero(npx_initialstate->sv_xmm.sv_fp,
483 		    sizeof(npx_initialstate->sv_xmm.sv_fp));
484 		bzero(npx_initialstate->sv_xmm.sv_xmm,
485 		    sizeof(npx_initialstate->sv_xmm.sv_xmm));
486 
487 	} else
488 		bzero(npx_initialstate->sv_87.sv_ac,
489 		    sizeof(npx_initialstate->sv_87.sv_ac));
490 
491 	/*
492 	 * Create a table describing the layout of the CPU Extended
493 	 * Save Area.  See Intel SDM rev. 075 Vol. 1 13.4.1 "Legacy
494 	 * Region of an XSAVE Area" for the source of offsets/sizes.
495 	 * Note that 32bit XSAVE does not use %xmm8-%xmm15, see
496 	 * 10.5.1.2 and 13.5.2 "SSE State".
497 	 */
498 	if (use_xsave) {
499 		xstate_bv = (uint64_t *)((char *)(npx_initialstate + 1) +
500 		    offsetof(struct xstate_hdr, xstate_bv));
501 		*xstate_bv = XFEATURE_ENABLED_X87 | XFEATURE_ENABLED_SSE;
502 
503 		/* x87 state */
504 		xsave_area_desc[0].offset = 0;
505 		xsave_area_desc[0].size = 160;
506 		/* XMM */
507 		xsave_area_desc[1].offset = 160;
508 		xsave_area_desc[1].size = 288 - 160;
509 
510 		for (i = 2; i < max_ext_n; i++) {
511 			cpuid_count(0xd, i, cp);
512 			xsave_area_desc[i].offset = cp[1];
513 			xsave_area_desc[i].size = cp[0];
514 		}
515 	}
516 
517 	fpu_disable();
518 	intr_restore(saveintr);
519 }
520 SYSINIT(npxinitstate, SI_SUB_CPU, SI_ORDER_ANY, npxinitstate, NULL);
521 
522 /*
523  * Free coprocessor (if we have it).
524  */
525 void
npxexit(struct thread * td)526 npxexit(struct thread *td)
527 {
528 
529 	critical_enter();
530 	if (curthread == PCPU_GET(fpcurthread)) {
531 		fpu_enable();
532 		fpusave(curpcb->pcb_save);
533 		fpu_disable();
534 		PCPU_SET(fpcurthread, NULL);
535 	}
536 	critical_exit();
537 #ifdef NPX_DEBUG
538 	if (hw_float) {
539 		u_int	masked_exceptions;
540 
541 		masked_exceptions = GET_FPU_CW(td) & GET_FPU_SW(td) & 0x7f;
542 		/*
543 		 * Log exceptions that would have trapped with the old
544 		 * control word (overflow, divide by 0, and invalid operand).
545 		 */
546 		if (masked_exceptions & 0x0d)
547 			log(LOG_ERR,
548 	"pid %d (%s) exited with masked floating point exceptions 0x%02x\n",
549 			    td->td_proc->p_pid, td->td_proc->p_comm,
550 			    masked_exceptions);
551 	}
552 #endif
553 }
554 
555 int
npxformat(void)556 npxformat(void)
557 {
558 
559 	if (!hw_float)
560 		return (_MC_FPFMT_NODEV);
561 	if (cpu_fxsr)
562 		return (_MC_FPFMT_XMM);
563 	return (_MC_FPFMT_387);
564 }
565 
566 /*
567  * The following mechanism is used to ensure that the FPE_... value
568  * that is passed as a trapcode to the signal handler of the user
569  * process does not have more than one bit set.
570  *
571  * Multiple bits may be set if the user process modifies the control
572  * word while a status word bit is already set.  While this is a sign
573  * of bad coding, we have no choice than to narrow them down to one
574  * bit, since we must not send a trapcode that is not exactly one of
575  * the FPE_ macros.
576  *
577  * The mechanism has a static table with 127 entries.  Each combination
578  * of the 7 FPU status word exception bits directly translates to a
579  * position in this table, where a single FPE_... value is stored.
580  * This FPE_... value stored there is considered the "most important"
581  * of the exception bits and will be sent as the signal code.  The
582  * precedence of the bits is based upon Intel Document "Numerical
583  * Applications", Chapter "Special Computational Situations".
584  *
585  * The macro to choose one of these values does these steps: 1) Throw
586  * away status word bits that cannot be masked.  2) Throw away the bits
587  * currently masked in the control word, assuming the user isn't
588  * interested in them anymore.  3) Reinsert status word bit 7 (stack
589  * fault) if it is set, which cannot be masked but must be presered.
590  * 4) Use the remaining bits to point into the trapcode table.
591  *
592  * The 6 maskable bits in order of their preference, as stated in the
593  * above referenced Intel manual:
594  * 1  Invalid operation (FP_X_INV)
595  * 1a   Stack underflow
596  * 1b   Stack overflow
597  * 1c   Operand of unsupported format
598  * 1d   SNaN operand.
599  * 2  QNaN operand (not an exception, irrelavant here)
600  * 3  Any other invalid-operation not mentioned above or zero divide
601  *      (FP_X_INV, FP_X_DZ)
602  * 4  Denormal operand (FP_X_DNML)
603  * 5  Numeric over/underflow (FP_X_OFL, FP_X_UFL)
604  * 6  Inexact result (FP_X_IMP)
605  */
606 static char fpetable[128] = {
607 	0,
608 	FPE_FLTINV,	/*  1 - INV */
609 	FPE_FLTUND,	/*  2 - DNML */
610 	FPE_FLTINV,	/*  3 - INV | DNML */
611 	FPE_FLTDIV,	/*  4 - DZ */
612 	FPE_FLTINV,	/*  5 - INV | DZ */
613 	FPE_FLTDIV,	/*  6 - DNML | DZ */
614 	FPE_FLTINV,	/*  7 - INV | DNML | DZ */
615 	FPE_FLTOVF,	/*  8 - OFL */
616 	FPE_FLTINV,	/*  9 - INV | OFL */
617 	FPE_FLTUND,	/*  A - DNML | OFL */
618 	FPE_FLTINV,	/*  B - INV | DNML | OFL */
619 	FPE_FLTDIV,	/*  C - DZ | OFL */
620 	FPE_FLTINV,	/*  D - INV | DZ | OFL */
621 	FPE_FLTDIV,	/*  E - DNML | DZ | OFL */
622 	FPE_FLTINV,	/*  F - INV | DNML | DZ | OFL */
623 	FPE_FLTUND,	/* 10 - UFL */
624 	FPE_FLTINV,	/* 11 - INV | UFL */
625 	FPE_FLTUND,	/* 12 - DNML | UFL */
626 	FPE_FLTINV,	/* 13 - INV | DNML | UFL */
627 	FPE_FLTDIV,	/* 14 - DZ | UFL */
628 	FPE_FLTINV,	/* 15 - INV | DZ | UFL */
629 	FPE_FLTDIV,	/* 16 - DNML | DZ | UFL */
630 	FPE_FLTINV,	/* 17 - INV | DNML | DZ | UFL */
631 	FPE_FLTOVF,	/* 18 - OFL | UFL */
632 	FPE_FLTINV,	/* 19 - INV | OFL | UFL */
633 	FPE_FLTUND,	/* 1A - DNML | OFL | UFL */
634 	FPE_FLTINV,	/* 1B - INV | DNML | OFL | UFL */
635 	FPE_FLTDIV,	/* 1C - DZ | OFL | UFL */
636 	FPE_FLTINV,	/* 1D - INV | DZ | OFL | UFL */
637 	FPE_FLTDIV,	/* 1E - DNML | DZ | OFL | UFL */
638 	FPE_FLTINV,	/* 1F - INV | DNML | DZ | OFL | UFL */
639 	FPE_FLTRES,	/* 20 - IMP */
640 	FPE_FLTINV,	/* 21 - INV | IMP */
641 	FPE_FLTUND,	/* 22 - DNML | IMP */
642 	FPE_FLTINV,	/* 23 - INV | DNML | IMP */
643 	FPE_FLTDIV,	/* 24 - DZ | IMP */
644 	FPE_FLTINV,	/* 25 - INV | DZ | IMP */
645 	FPE_FLTDIV,	/* 26 - DNML | DZ | IMP */
646 	FPE_FLTINV,	/* 27 - INV | DNML | DZ | IMP */
647 	FPE_FLTOVF,	/* 28 - OFL | IMP */
648 	FPE_FLTINV,	/* 29 - INV | OFL | IMP */
649 	FPE_FLTUND,	/* 2A - DNML | OFL | IMP */
650 	FPE_FLTINV,	/* 2B - INV | DNML | OFL | IMP */
651 	FPE_FLTDIV,	/* 2C - DZ | OFL | IMP */
652 	FPE_FLTINV,	/* 2D - INV | DZ | OFL | IMP */
653 	FPE_FLTDIV,	/* 2E - DNML | DZ | OFL | IMP */
654 	FPE_FLTINV,	/* 2F - INV | DNML | DZ | OFL | IMP */
655 	FPE_FLTUND,	/* 30 - UFL | IMP */
656 	FPE_FLTINV,	/* 31 - INV | UFL | IMP */
657 	FPE_FLTUND,	/* 32 - DNML | UFL | IMP */
658 	FPE_FLTINV,	/* 33 - INV | DNML | UFL | IMP */
659 	FPE_FLTDIV,	/* 34 - DZ | UFL | IMP */
660 	FPE_FLTINV,	/* 35 - INV | DZ | UFL | IMP */
661 	FPE_FLTDIV,	/* 36 - DNML | DZ | UFL | IMP */
662 	FPE_FLTINV,	/* 37 - INV | DNML | DZ | UFL | IMP */
663 	FPE_FLTOVF,	/* 38 - OFL | UFL | IMP */
664 	FPE_FLTINV,	/* 39 - INV | OFL | UFL | IMP */
665 	FPE_FLTUND,	/* 3A - DNML | OFL | UFL | IMP */
666 	FPE_FLTINV,	/* 3B - INV | DNML | OFL | UFL | IMP */
667 	FPE_FLTDIV,	/* 3C - DZ | OFL | UFL | IMP */
668 	FPE_FLTINV,	/* 3D - INV | DZ | OFL | UFL | IMP */
669 	FPE_FLTDIV,	/* 3E - DNML | DZ | OFL | UFL | IMP */
670 	FPE_FLTINV,	/* 3F - INV | DNML | DZ | OFL | UFL | IMP */
671 	FPE_FLTSUB,	/* 40 - STK */
672 	FPE_FLTSUB,	/* 41 - INV | STK */
673 	FPE_FLTUND,	/* 42 - DNML | STK */
674 	FPE_FLTSUB,	/* 43 - INV | DNML | STK */
675 	FPE_FLTDIV,	/* 44 - DZ | STK */
676 	FPE_FLTSUB,	/* 45 - INV | DZ | STK */
677 	FPE_FLTDIV,	/* 46 - DNML | DZ | STK */
678 	FPE_FLTSUB,	/* 47 - INV | DNML | DZ | STK */
679 	FPE_FLTOVF,	/* 48 - OFL | STK */
680 	FPE_FLTSUB,	/* 49 - INV | OFL | STK */
681 	FPE_FLTUND,	/* 4A - DNML | OFL | STK */
682 	FPE_FLTSUB,	/* 4B - INV | DNML | OFL | STK */
683 	FPE_FLTDIV,	/* 4C - DZ | OFL | STK */
684 	FPE_FLTSUB,	/* 4D - INV | DZ | OFL | STK */
685 	FPE_FLTDIV,	/* 4E - DNML | DZ | OFL | STK */
686 	FPE_FLTSUB,	/* 4F - INV | DNML | DZ | OFL | STK */
687 	FPE_FLTUND,	/* 50 - UFL | STK */
688 	FPE_FLTSUB,	/* 51 - INV | UFL | STK */
689 	FPE_FLTUND,	/* 52 - DNML | UFL | STK */
690 	FPE_FLTSUB,	/* 53 - INV | DNML | UFL | STK */
691 	FPE_FLTDIV,	/* 54 - DZ | UFL | STK */
692 	FPE_FLTSUB,	/* 55 - INV | DZ | UFL | STK */
693 	FPE_FLTDIV,	/* 56 - DNML | DZ | UFL | STK */
694 	FPE_FLTSUB,	/* 57 - INV | DNML | DZ | UFL | STK */
695 	FPE_FLTOVF,	/* 58 - OFL | UFL | STK */
696 	FPE_FLTSUB,	/* 59 - INV | OFL | UFL | STK */
697 	FPE_FLTUND,	/* 5A - DNML | OFL | UFL | STK */
698 	FPE_FLTSUB,	/* 5B - INV | DNML | OFL | UFL | STK */
699 	FPE_FLTDIV,	/* 5C - DZ | OFL | UFL | STK */
700 	FPE_FLTSUB,	/* 5D - INV | DZ | OFL | UFL | STK */
701 	FPE_FLTDIV,	/* 5E - DNML | DZ | OFL | UFL | STK */
702 	FPE_FLTSUB,	/* 5F - INV | DNML | DZ | OFL | UFL | STK */
703 	FPE_FLTRES,	/* 60 - IMP | STK */
704 	FPE_FLTSUB,	/* 61 - INV | IMP | STK */
705 	FPE_FLTUND,	/* 62 - DNML | IMP | STK */
706 	FPE_FLTSUB,	/* 63 - INV | DNML | IMP | STK */
707 	FPE_FLTDIV,	/* 64 - DZ | IMP | STK */
708 	FPE_FLTSUB,	/* 65 - INV | DZ | IMP | STK */
709 	FPE_FLTDIV,	/* 66 - DNML | DZ | IMP | STK */
710 	FPE_FLTSUB,	/* 67 - INV | DNML | DZ | IMP | STK */
711 	FPE_FLTOVF,	/* 68 - OFL | IMP | STK */
712 	FPE_FLTSUB,	/* 69 - INV | OFL | IMP | STK */
713 	FPE_FLTUND,	/* 6A - DNML | OFL | IMP | STK */
714 	FPE_FLTSUB,	/* 6B - INV | DNML | OFL | IMP | STK */
715 	FPE_FLTDIV,	/* 6C - DZ | OFL | IMP | STK */
716 	FPE_FLTSUB,	/* 6D - INV | DZ | OFL | IMP | STK */
717 	FPE_FLTDIV,	/* 6E - DNML | DZ | OFL | IMP | STK */
718 	FPE_FLTSUB,	/* 6F - INV | DNML | DZ | OFL | IMP | STK */
719 	FPE_FLTUND,	/* 70 - UFL | IMP | STK */
720 	FPE_FLTSUB,	/* 71 - INV | UFL | IMP | STK */
721 	FPE_FLTUND,	/* 72 - DNML | UFL | IMP | STK */
722 	FPE_FLTSUB,	/* 73 - INV | DNML | UFL | IMP | STK */
723 	FPE_FLTDIV,	/* 74 - DZ | UFL | IMP | STK */
724 	FPE_FLTSUB,	/* 75 - INV | DZ | UFL | IMP | STK */
725 	FPE_FLTDIV,	/* 76 - DNML | DZ | UFL | IMP | STK */
726 	FPE_FLTSUB,	/* 77 - INV | DNML | DZ | UFL | IMP | STK */
727 	FPE_FLTOVF,	/* 78 - OFL | UFL | IMP | STK */
728 	FPE_FLTSUB,	/* 79 - INV | OFL | UFL | IMP | STK */
729 	FPE_FLTUND,	/* 7A - DNML | OFL | UFL | IMP | STK */
730 	FPE_FLTSUB,	/* 7B - INV | DNML | OFL | UFL | IMP | STK */
731 	FPE_FLTDIV,	/* 7C - DZ | OFL | UFL | IMP | STK */
732 	FPE_FLTSUB,	/* 7D - INV | DZ | OFL | UFL | IMP | STK */
733 	FPE_FLTDIV,	/* 7E - DNML | DZ | OFL | UFL | IMP | STK */
734 	FPE_FLTSUB,	/* 7F - INV | DNML | DZ | OFL | UFL | IMP | STK */
735 };
736 
737 /*
738  * Read the FP status and control words, then generate si_code value
739  * for SIGFPE.  The error code chosen will be one of the
740  * FPE_... macros.  It will be sent as the second argument to old
741  * BSD-style signal handlers and as "siginfo_t->si_code" (second
742  * argument) to SA_SIGINFO signal handlers.
743  *
744  * Some time ago, we cleared the x87 exceptions with FNCLEX there.
745  * Clearing exceptions was necessary mainly to avoid IRQ13 bugs.  The
746  * usermode code which understands the FPU hardware enough to enable
747  * the exceptions, can also handle clearing the exception state in the
748  * handler.  The only consequence of not clearing the exception is the
749  * rethrow of the SIGFPE on return from the signal handler and
750  * reexecution of the corresponding instruction.
751  *
752  * For XMM traps, the exceptions were never cleared.
753  */
754 int
npxtrap_x87(void)755 npxtrap_x87(void)
756 {
757 	u_short control, status;
758 
759 	if (!hw_float) {
760 		printf(
761 	"npxtrap_x87: fpcurthread = %p, curthread = %p, hw_float = %d\n",
762 		       PCPU_GET(fpcurthread), curthread, hw_float);
763 		panic("npxtrap from nowhere");
764 	}
765 	critical_enter();
766 
767 	/*
768 	 * Interrupt handling (for another interrupt) may have pushed the
769 	 * state to memory.  Fetch the relevant parts of the state from
770 	 * wherever they are.
771 	 */
772 	if (PCPU_GET(fpcurthread) != curthread) {
773 		control = GET_FPU_CW(curthread);
774 		status = GET_FPU_SW(curthread);
775 	} else {
776 		fnstcw(&control);
777 		fnstsw(&status);
778 	}
779 	critical_exit();
780 	return (fpetable[status & ((~control & 0x3f) | 0x40)]);
781 }
782 
783 int
npxtrap_sse(void)784 npxtrap_sse(void)
785 {
786 	u_int mxcsr;
787 
788 	if (!hw_float) {
789 		printf(
790 	"npxtrap_sse: fpcurthread = %p, curthread = %p, hw_float = %d\n",
791 		       PCPU_GET(fpcurthread), curthread, hw_float);
792 		panic("npxtrap from nowhere");
793 	}
794 	critical_enter();
795 	if (PCPU_GET(fpcurthread) != curthread)
796 		mxcsr = curthread->td_pcb->pcb_save->sv_xmm.sv_env.en_mxcsr;
797 	else
798 		stmxcsr(&mxcsr);
799 	critical_exit();
800 	return (fpetable[(mxcsr & (~mxcsr >> 7)) & 0x3f]);
801 }
802 
803 static void
restore_npx_curthread(struct thread * td,struct pcb * pcb)804 restore_npx_curthread(struct thread *td, struct pcb *pcb)
805 {
806 
807 	/*
808 	 * Record new context early in case frstor causes a trap.
809 	 */
810 	PCPU_SET(fpcurthread, td);
811 
812 	fpu_enable();
813 	if (cpu_fxsr)
814 		fpu_clean_state();
815 
816 	if ((pcb->pcb_flags & PCB_NPXINITDONE) == 0) {
817 		/*
818 		 * This is the first time this thread has used the FPU or
819 		 * the PCB doesn't contain a clean FPU state.  Explicitly
820 		 * load an initial state.
821 		 *
822 		 * We prefer to restore the state from the actual save
823 		 * area in PCB instead of directly loading from
824 		 * npx_initialstate, to ignite the XSAVEOPT
825 		 * tracking engine.
826 		 */
827 		bcopy(npx_initialstate, pcb->pcb_save, cpu_max_ext_state_size);
828 		fpurstor(pcb->pcb_save);
829 		if (pcb->pcb_initial_npxcw != __INITIAL_NPXCW__)
830 			fldcw(pcb->pcb_initial_npxcw);
831 		pcb->pcb_flags |= PCB_NPXINITDONE;
832 		if (PCB_USER_FPU(pcb))
833 			pcb->pcb_flags |= PCB_NPXUSERINITDONE;
834 	} else {
835 		fpurstor(pcb->pcb_save);
836 	}
837 }
838 
839 /*
840  * Implement device not available (DNA) exception
841  *
842  * It would be better to switch FP context here (if curthread != fpcurthread)
843  * and not necessarily for every context switch, but it is too hard to
844  * access foreign pcb's.
845  */
846 int
npxdna(void)847 npxdna(void)
848 {
849 	struct thread *td;
850 
851 	if (!hw_float)
852 		return (0);
853 	td = curthread;
854 	critical_enter();
855 
856 	KASSERT((curpcb->pcb_flags & PCB_NPXNOSAVE) == 0,
857 	    ("npxdna while in fpu_kern_enter(FPU_KERN_NOCTX)"));
858 	if (__predict_false(PCPU_GET(fpcurthread) == td)) {
859 		/*
860 		 * Some virtual machines seems to set %cr0.TS at
861 		 * arbitrary moments.  Silently clear the TS bit
862 		 * regardless of the eager/lazy FPU context switch
863 		 * mode.
864 		 */
865 		fpu_enable();
866 	} else {
867 		if (__predict_false(PCPU_GET(fpcurthread) != NULL)) {
868 			printf(
869 		    "npxdna: fpcurthread = %p (%d), curthread = %p (%d)\n",
870 			    PCPU_GET(fpcurthread),
871 			    PCPU_GET(fpcurthread)->td_proc->p_pid,
872 			    td, td->td_proc->p_pid);
873 			panic("npxdna");
874 		}
875 		restore_npx_curthread(td, td->td_pcb);
876 	}
877 	critical_exit();
878 	return (1);
879 }
880 
881 /*
882  * Wrapper for fpusave() called from context switch routines.
883  *
884  * npxsave() must be called with interrupts disabled, so that it clears
885  * fpcurthread atomically with saving the state.  We require callers to do the
886  * disabling, since most callers need to disable interrupts anyway to call
887  * npxsave() atomically with checking fpcurthread.
888  */
889 void
npxsave(union savefpu * addr)890 npxsave(union savefpu *addr)
891 {
892 
893 	fpu_enable();
894 	fpusave(addr);
895 }
896 
897 void npxswitch(struct thread *td, struct pcb *pcb);
898 void
npxswitch(struct thread * td,struct pcb * pcb)899 npxswitch(struct thread *td, struct pcb *pcb)
900 {
901 
902 	if (lazy_fpu_switch || (td->td_pflags & TDP_KTHREAD) != 0 ||
903 	    !PCB_USER_FPU(pcb)) {
904 		fpu_disable();
905 		PCPU_SET(fpcurthread, NULL);
906 	} else if (PCPU_GET(fpcurthread) != td) {
907 		restore_npx_curthread(td, pcb);
908 	}
909 }
910 
911 /*
912  * Unconditionally save the current co-processor state across suspend and
913  * resume.
914  */
915 void
npxsuspend(union savefpu * addr)916 npxsuspend(union savefpu *addr)
917 {
918 	register_t cr0;
919 
920 	if (!hw_float)
921 		return;
922 	if (PCPU_GET(fpcurthread) == NULL) {
923 		bcopy(npx_initialstate, addr, cpu_max_ext_state_size);
924 		return;
925 	}
926 	cr0 = rcr0();
927 	fpu_enable();
928 	fpusave(addr);
929 	load_cr0(cr0);
930 }
931 
932 void
npxresume(union savefpu * addr)933 npxresume(union savefpu *addr)
934 {
935 	register_t cr0;
936 
937 	if (!hw_float)
938 		return;
939 
940 	cr0 = rcr0();
941 	npxinit(false);
942 	fpu_enable();
943 	fpurstor(addr);
944 	load_cr0(cr0);
945 }
946 
947 void
npxdrop(void)948 npxdrop(void)
949 {
950 	struct thread *td;
951 
952 	/*
953 	 * Discard pending exceptions in the !cpu_fxsr case so that unmasked
954 	 * ones don't cause a panic on the next frstor.
955 	 */
956 	if (!cpu_fxsr)
957 		fnclex();
958 
959 	td = PCPU_GET(fpcurthread);
960 	KASSERT(td == curthread, ("fpudrop: fpcurthread != curthread"));
961 	CRITICAL_ASSERT(td);
962 	PCPU_SET(fpcurthread, NULL);
963 	td->td_pcb->pcb_flags &= ~PCB_NPXINITDONE;
964 	fpu_disable();
965 }
966 
967 /*
968  * Get the user state of the FPU into pcb->pcb_user_save without
969  * dropping ownership (if possible).  It returns the FPU ownership
970  * status.
971  */
972 int
npxgetregs(struct thread * td)973 npxgetregs(struct thread *td)
974 {
975 	struct pcb *pcb;
976 	uint64_t *xstate_bv, bit;
977 	char *sa;
978 	union savefpu *s;
979 	uint32_t mxcsr, mxcsr_mask;
980 	int max_ext_n, i;
981 	int owned;
982 	bool do_mxcsr;
983 
984 	if (!hw_float)
985 		return (_MC_FPOWNED_NONE);
986 
987 	pcb = td->td_pcb;
988 	critical_enter();
989 	if ((pcb->pcb_flags & PCB_NPXINITDONE) == 0) {
990 		bcopy(npx_initialstate, get_pcb_user_save_pcb(pcb),
991 		    cpu_max_ext_state_size);
992 		SET_FPU_CW(get_pcb_user_save_pcb(pcb), pcb->pcb_initial_npxcw);
993 		npxuserinited(td);
994 		critical_exit();
995 		return (_MC_FPOWNED_PCB);
996 	}
997 	if (td == PCPU_GET(fpcurthread)) {
998 		fpusave(get_pcb_user_save_pcb(pcb));
999 		if (!cpu_fxsr)
1000 			/*
1001 			 * fnsave initializes the FPU and destroys whatever
1002 			 * context it contains.  Make sure the FPU owner
1003 			 * starts with a clean state next time.
1004 			 */
1005 			npxdrop();
1006 		owned = _MC_FPOWNED_FPU;
1007 	} else {
1008 		owned = _MC_FPOWNED_PCB;
1009 	}
1010 	if (use_xsave) {
1011 		/*
1012 		 * Handle partially saved state.
1013 		 */
1014 		sa = (char *)get_pcb_user_save_pcb(pcb);
1015 		xstate_bv = (uint64_t *)(sa + sizeof(union savefpu) +
1016 		    offsetof(struct xstate_hdr, xstate_bv));
1017 		if (xsave_mask >> 32 != 0)
1018 			max_ext_n = fls(xsave_mask >> 32) + 32;
1019 		else
1020 			max_ext_n = fls(xsave_mask);
1021 		for (i = 0; i < max_ext_n; i++) {
1022 			bit = 1ULL << i;
1023 			if ((xsave_mask & bit) == 0 || (*xstate_bv & bit) != 0)
1024 				continue;
1025 			do_mxcsr = false;
1026 			if (i == 0 && (*xstate_bv & (XFEATURE_ENABLED_SSE |
1027 			    XFEATURE_ENABLED_AVX)) != 0) {
1028 				/*
1029 				 * x87 area was not saved by XSAVEOPT,
1030 				 * but one of XMM or AVX was.  Then we need
1031 				 * to preserve MXCSR from being overwritten
1032 				 * with the default value.
1033 				 */
1034 				s = (union savefpu *)sa;
1035 				mxcsr = s->sv_xmm.sv_env.en_mxcsr;
1036 				mxcsr_mask = s->sv_xmm.sv_env.en_mxcsr_mask;
1037 				do_mxcsr = true;
1038 			}
1039 			bcopy((char *)npx_initialstate +
1040 			    xsave_area_desc[i].offset,
1041 			    sa + xsave_area_desc[i].offset,
1042 			    xsave_area_desc[i].size);
1043 			if (do_mxcsr) {
1044 				s->sv_xmm.sv_env.en_mxcsr = mxcsr;
1045 				s->sv_xmm.sv_env.en_mxcsr_mask = mxcsr_mask;
1046 			}
1047 			*xstate_bv |= bit;
1048 		}
1049 	}
1050 	critical_exit();
1051 	return (owned);
1052 }
1053 
1054 void
npxuserinited(struct thread * td)1055 npxuserinited(struct thread *td)
1056 {
1057 	struct pcb *pcb;
1058 
1059 	CRITICAL_ASSERT(td);
1060 	pcb = td->td_pcb;
1061 	if (PCB_USER_FPU(pcb))
1062 		pcb->pcb_flags |= PCB_NPXINITDONE;
1063 	pcb->pcb_flags |= PCB_NPXUSERINITDONE;
1064 }
1065 
1066 int
npxsetxstate(struct thread * td,char * xfpustate,size_t xfpustate_size)1067 npxsetxstate(struct thread *td, char *xfpustate, size_t xfpustate_size)
1068 {
1069 	struct xstate_hdr *hdr, *ehdr;
1070 	size_t len, max_len;
1071 	uint64_t bv;
1072 
1073 	/* XXXKIB should we clear all extended state in xstate_bv instead ? */
1074 	if (xfpustate == NULL)
1075 		return (0);
1076 	if (!use_xsave)
1077 		return (EOPNOTSUPP);
1078 
1079 	len = xfpustate_size;
1080 	if (len < sizeof(struct xstate_hdr))
1081 		return (EINVAL);
1082 	max_len = cpu_max_ext_state_size - sizeof(union savefpu);
1083 	if (len > max_len)
1084 		return (EINVAL);
1085 
1086 	ehdr = (struct xstate_hdr *)xfpustate;
1087 	bv = ehdr->xstate_bv;
1088 
1089 	/*
1090 	 * Avoid #gp.
1091 	 */
1092 	if (bv & ~xsave_mask)
1093 		return (EINVAL);
1094 
1095 	hdr = (struct xstate_hdr *)(get_pcb_user_save_td(td) + 1);
1096 
1097 	hdr->xstate_bv = bv;
1098 	bcopy(xfpustate + sizeof(struct xstate_hdr),
1099 	    (char *)(hdr + 1), len - sizeof(struct xstate_hdr));
1100 
1101 	return (0);
1102 }
1103 
1104 int
npxsetregs(struct thread * td,union savefpu * addr,char * xfpustate,size_t xfpustate_size)1105 npxsetregs(struct thread *td, union savefpu *addr, char *xfpustate,
1106 	size_t xfpustate_size)
1107 {
1108 	struct pcb *pcb;
1109 	int error;
1110 
1111 	if (!hw_float)
1112 		return (ENXIO);
1113 
1114 	if (cpu_fxsr)
1115 		addr->sv_xmm.sv_env.en_mxcsr &= cpu_mxcsr_mask;
1116 	pcb = td->td_pcb;
1117 	error = 0;
1118 	critical_enter();
1119 	if (td == PCPU_GET(fpcurthread) && PCB_USER_FPU(pcb)) {
1120 		error = npxsetxstate(td, xfpustate, xfpustate_size);
1121 		if (error == 0) {
1122 			if (!cpu_fxsr)
1123 				fnclex();	/* As in npxdrop(). */
1124 			bcopy(addr, get_pcb_user_save_td(td), sizeof(*addr));
1125 			fpurstor(get_pcb_user_save_td(td));
1126 			pcb->pcb_flags |= PCB_NPXUSERINITDONE | PCB_NPXINITDONE;
1127 		}
1128 	} else {
1129 		error = npxsetxstate(td, xfpustate, xfpustate_size);
1130 		if (error == 0) {
1131 			bcopy(addr, get_pcb_user_save_td(td), sizeof(*addr));
1132 			npxuserinited(td);
1133 		}
1134 	}
1135 	critical_exit();
1136 	return (error);
1137 }
1138 
1139 static void
npx_fill_fpregs_xmm1(struct savexmm * sv_xmm,struct save87 * sv_87)1140 npx_fill_fpregs_xmm1(struct savexmm *sv_xmm, struct save87 *sv_87)
1141 {
1142 	struct env87 *penv_87;
1143 	struct envxmm *penv_xmm;
1144 	struct fpacc87 *fx_reg;
1145 	int i, st;
1146 	uint64_t mantissa;
1147 	uint16_t tw, exp;
1148 	uint8_t ab_tw;
1149 
1150 	penv_87 = &sv_87->sv_env;
1151 	penv_xmm = &sv_xmm->sv_env;
1152 
1153 	/* FPU control/status */
1154 	penv_87->en_cw = penv_xmm->en_cw;
1155 	penv_87->en_sw = penv_xmm->en_sw;
1156 	penv_87->en_fip = penv_xmm->en_fip;
1157 	penv_87->en_fcs = penv_xmm->en_fcs;
1158 	penv_87->en_opcode = penv_xmm->en_opcode;
1159 	penv_87->en_foo = penv_xmm->en_foo;
1160 	penv_87->en_fos = penv_xmm->en_fos;
1161 
1162 	/*
1163 	 * FPU registers and tags.
1164 	 * For ST(i), i = fpu_reg - top; we start with fpu_reg=7.
1165 	 */
1166 	st = 7 - ((penv_xmm->en_sw >> 11) & 7);
1167 	ab_tw = penv_xmm->en_tw;
1168 	tw = 0;
1169 	for (i = 0x80; i != 0; i >>= 1) {
1170 		sv_87->sv_ac[st] = sv_xmm->sv_fp[st].fp_acc;
1171 		tw <<= 2;
1172 		if (ab_tw & i) {
1173 			/* Non-empty - we need to check ST(i) */
1174 			fx_reg = &sv_xmm->sv_fp[st].fp_acc;
1175 			/* The first 64 bits contain the mantissa. */
1176 			mantissa = *((uint64_t *)fx_reg->fp_bytes);
1177 			/*
1178 			 * The final 16 bits contain the sign bit and the exponent.
1179 			 * Mask the sign bit since it is of no consequence to these
1180 			 * tests.
1181 			 */
1182 			exp = *((uint16_t *)&fx_reg->fp_bytes[8]) & 0x7fff;
1183 			if (exp == 0) {
1184 				if (mantissa == 0)
1185 					tw |= 1; /* Zero */
1186 				else
1187 					tw |= 2; /* Denormal */
1188 			} else if (exp == 0x7fff)
1189 				tw |= 2; /* Infinity or NaN */
1190 		} else
1191 			tw |= 3; /* Empty */
1192 		st = (st - 1) & 7;
1193 	}
1194 	penv_87->en_tw = tw;
1195 }
1196 
1197 void
npx_fill_fpregs_xmm(struct savexmm * sv_xmm,struct save87 * sv_87)1198 npx_fill_fpregs_xmm(struct savexmm *sv_xmm, struct save87 *sv_87)
1199 {
1200 
1201 	bzero(sv_87, sizeof(*sv_87));
1202 	npx_fill_fpregs_xmm1(sv_xmm, sv_87);
1203 }
1204 
1205 void
npx_set_fpregs_xmm(struct save87 * sv_87,struct savexmm * sv_xmm)1206 npx_set_fpregs_xmm(struct save87 *sv_87, struct savexmm *sv_xmm)
1207 {
1208 	struct env87 *penv_87;
1209 	struct envxmm *penv_xmm;
1210 	int i;
1211 
1212 	penv_87 = &sv_87->sv_env;
1213 	penv_xmm = &sv_xmm->sv_env;
1214 
1215 	/* FPU control/status */
1216 	penv_xmm->en_cw = penv_87->en_cw;
1217 	penv_xmm->en_sw = penv_87->en_sw;
1218 	penv_xmm->en_fip = penv_87->en_fip;
1219 	penv_xmm->en_fcs = penv_87->en_fcs;
1220 	penv_xmm->en_opcode = penv_87->en_opcode;
1221 	penv_xmm->en_foo = penv_87->en_foo;
1222 	penv_xmm->en_fos = penv_87->en_fos;
1223 
1224 	/*
1225 	 * FPU registers and tags.
1226 	 * Abridged  /  Full translation (values in binary), see FXSAVE spec.
1227 	 * 0		11
1228 	 * 1		00, 01, 10
1229 	 */
1230 	penv_xmm->en_tw = 0;
1231 	for (i = 0; i < 8; ++i) {
1232 		sv_xmm->sv_fp[i].fp_acc = sv_87->sv_ac[i];
1233 		if ((penv_87->en_tw & (3 << i * 2)) != (3 << i * 2))
1234 			penv_xmm->en_tw |= 1 << i;
1235 	}
1236 }
1237 
1238 void
npx_get_fsave(void * addr)1239 npx_get_fsave(void *addr)
1240 {
1241 	struct thread *td;
1242 	union savefpu *sv;
1243 
1244 	td = curthread;
1245 	npxgetregs(td);
1246 	sv = get_pcb_user_save_td(td);
1247 	if (cpu_fxsr)
1248 		npx_fill_fpregs_xmm1(&sv->sv_xmm, addr);
1249 	else
1250 		bcopy(sv, addr, sizeof(struct env87) +
1251 		    sizeof(struct fpacc87[8]));
1252 }
1253 
1254 int
npx_set_fsave(void * addr)1255 npx_set_fsave(void *addr)
1256 {
1257 	union savefpu sv;
1258 	int error;
1259 
1260 	bzero(&sv, sizeof(sv));
1261 	if (cpu_fxsr)
1262 		npx_set_fpregs_xmm(addr, &sv.sv_xmm);
1263 	else
1264 		bcopy(addr, &sv, sizeof(struct env87) +
1265 		    sizeof(struct fpacc87[8]));
1266 	error = npxsetregs(curthread, &sv, NULL, 0);
1267 	return (error);
1268 }
1269 
1270 /*
1271  * On AuthenticAMD processors, the fxrstor instruction does not restore
1272  * the x87's stored last instruction pointer, last data pointer, and last
1273  * opcode values, except in the rare case in which the exception summary
1274  * (ES) bit in the x87 status word is set to 1.
1275  *
1276  * In order to avoid leaking this information across processes, we clean
1277  * these values by performing a dummy load before executing fxrstor().
1278  */
1279 static void
fpu_clean_state(void)1280 fpu_clean_state(void)
1281 {
1282 	static float dummy_variable = 0.0;
1283 	u_short status;
1284 
1285 	/*
1286 	 * Clear the ES bit in the x87 status word if it is currently
1287 	 * set, in order to avoid causing a fault in the upcoming load.
1288 	 */
1289 	fnstsw(&status);
1290 	if (status & 0x80)
1291 		fnclex();
1292 
1293 	/*
1294 	 * Load the dummy variable into the x87 stack.  This mangles
1295 	 * the x87 stack, but we don't care since we're about to call
1296 	 * fxrstor() anyway.
1297 	 */
1298 	__asm __volatile("ffree %%st(7); flds %0" : : "m" (dummy_variable));
1299 }
1300 
1301 static void
fpurstor(union savefpu * addr)1302 fpurstor(union savefpu *addr)
1303 {
1304 
1305 	if (use_xsave)
1306 		xrstor((char *)addr, xsave_mask);
1307 	else if (cpu_fxsr)
1308 		fxrstor(addr);
1309 	else
1310 		frstor(addr);
1311 }
1312 
1313 #ifdef DEV_ISA
1314 /*
1315  * This sucks up the legacy ISA support assignments from PNPBIOS/ACPI.
1316  */
1317 static struct isa_pnp_id npxisa_ids[] = {
1318 	{ 0x040cd041, "Legacy ISA coprocessor support" }, /* PNP0C04 */
1319 	{ 0 }
1320 };
1321 
1322 static int
npxisa_probe(device_t dev)1323 npxisa_probe(device_t dev)
1324 {
1325 	int result;
1326 	if ((result = ISA_PNP_PROBE(device_get_parent(dev), dev, npxisa_ids)) <= 0) {
1327 		device_quiet(dev);
1328 	}
1329 	return(result);
1330 }
1331 
1332 static int
npxisa_attach(device_t dev)1333 npxisa_attach(device_t dev)
1334 {
1335 	return (0);
1336 }
1337 
1338 static device_method_t npxisa_methods[] = {
1339 	/* Device interface */
1340 	DEVMETHOD(device_probe,		npxisa_probe),
1341 	DEVMETHOD(device_attach,	npxisa_attach),
1342 	{ 0, 0 }
1343 };
1344 
1345 static driver_t npxisa_driver = {
1346 	"npxisa",
1347 	npxisa_methods,
1348 	1,			/* no softc */
1349 };
1350 
1351 DRIVER_MODULE(npxisa, isa, npxisa_driver, 0, 0);
1352 DRIVER_MODULE(npxisa, acpi, npxisa_driver, 0, 0);
1353 ISA_PNP_INFO(npxisa_ids);
1354 #endif /* DEV_ISA */
1355 
1356 static MALLOC_DEFINE(M_FPUKERN_CTX, "fpukern_ctx",
1357     "Kernel contexts for FPU state");
1358 
1359 #define	FPU_KERN_CTX_NPXINITDONE 0x01
1360 #define	FPU_KERN_CTX_DUMMY	 0x02
1361 #define	FPU_KERN_CTX_INUSE	 0x04
1362 
1363 struct fpu_kern_ctx {
1364 	union savefpu *prev;
1365 	uint32_t flags;
1366 	char hwstate1[];
1367 };
1368 
1369 struct fpu_kern_ctx *
fpu_kern_alloc_ctx(u_int flags)1370 fpu_kern_alloc_ctx(u_int flags)
1371 {
1372 	struct fpu_kern_ctx *res;
1373 	size_t sz;
1374 
1375 	sz = sizeof(struct fpu_kern_ctx) + XSAVE_AREA_ALIGN +
1376 	    cpu_max_ext_state_size;
1377 	res = malloc(sz, M_FPUKERN_CTX, ((flags & FPU_KERN_NOWAIT) ?
1378 	    M_NOWAIT : M_WAITOK) | M_ZERO);
1379 	return (res);
1380 }
1381 
1382 void
fpu_kern_free_ctx(struct fpu_kern_ctx * ctx)1383 fpu_kern_free_ctx(struct fpu_kern_ctx *ctx)
1384 {
1385 
1386 	KASSERT((ctx->flags & FPU_KERN_CTX_INUSE) == 0, ("free'ing inuse ctx"));
1387 	/* XXXKIB clear the memory ? */
1388 	free(ctx, M_FPUKERN_CTX);
1389 }
1390 
1391 static union savefpu *
fpu_kern_ctx_savefpu(struct fpu_kern_ctx * ctx)1392 fpu_kern_ctx_savefpu(struct fpu_kern_ctx *ctx)
1393 {
1394 	vm_offset_t p;
1395 
1396 	p = (vm_offset_t)&ctx->hwstate1;
1397 	p = roundup2(p, XSAVE_AREA_ALIGN);
1398 	return ((union savefpu *)p);
1399 }
1400 
1401 void
fpu_kern_enter(struct thread * td,struct fpu_kern_ctx * ctx,u_int flags)1402 fpu_kern_enter(struct thread *td, struct fpu_kern_ctx *ctx, u_int flags)
1403 {
1404 	struct pcb *pcb;
1405 
1406 	pcb = td->td_pcb;
1407 	KASSERT((flags & FPU_KERN_NOCTX) != 0 || ctx != NULL,
1408 	    ("ctx is required when !FPU_KERN_NOCTX"));
1409 	KASSERT(ctx == NULL || (ctx->flags & FPU_KERN_CTX_INUSE) == 0,
1410 	    ("using inuse ctx"));
1411 	KASSERT((pcb->pcb_flags & PCB_NPXNOSAVE) == 0,
1412 	    ("recursive fpu_kern_enter while in PCB_NPXNOSAVE state"));
1413 
1414 	if ((flags & FPU_KERN_NOCTX) != 0) {
1415 		critical_enter();
1416 		fpu_enable();
1417 		if (curthread == PCPU_GET(fpcurthread)) {
1418 			fpusave(curpcb->pcb_save);
1419 			PCPU_SET(fpcurthread, NULL);
1420 		} else {
1421 			KASSERT(PCPU_GET(fpcurthread) == NULL,
1422 			    ("invalid fpcurthread"));
1423 		}
1424 
1425 		/*
1426 		 * This breaks XSAVEOPT tracker, but
1427 		 * PCB_NPXNOSAVE state is supposed to never need to
1428 		 * save FPU context at all.
1429 		 */
1430 		fpurstor(npx_initialstate);
1431 		pcb->pcb_flags |= PCB_KERNNPX | PCB_NPXNOSAVE | PCB_NPXINITDONE;
1432 		return;
1433 	}
1434 	if ((flags & FPU_KERN_KTHR) != 0 && is_fpu_kern_thread(0)) {
1435 		ctx->flags = FPU_KERN_CTX_DUMMY | FPU_KERN_CTX_INUSE;
1436 		return;
1437 	}
1438 	pcb = td->td_pcb;
1439 	critical_enter();
1440 	KASSERT(!PCB_USER_FPU(pcb) || pcb->pcb_save ==
1441 	    get_pcb_user_save_pcb(pcb), ("mangled pcb_save"));
1442 	ctx->flags = FPU_KERN_CTX_INUSE;
1443 	if ((pcb->pcb_flags & PCB_NPXINITDONE) != 0)
1444 		ctx->flags |= FPU_KERN_CTX_NPXINITDONE;
1445 	npxexit(td);
1446 	ctx->prev = pcb->pcb_save;
1447 	pcb->pcb_save = fpu_kern_ctx_savefpu(ctx);
1448 	pcb->pcb_flags |= PCB_KERNNPX;
1449 	pcb->pcb_flags &= ~PCB_NPXINITDONE;
1450 	critical_exit();
1451 }
1452 
1453 int
fpu_kern_leave(struct thread * td,struct fpu_kern_ctx * ctx)1454 fpu_kern_leave(struct thread *td, struct fpu_kern_ctx *ctx)
1455 {
1456 	struct pcb *pcb;
1457 
1458 	pcb = td->td_pcb;
1459 
1460 	if ((pcb->pcb_flags & PCB_NPXNOSAVE) != 0) {
1461 		KASSERT(ctx == NULL, ("non-null ctx after FPU_KERN_NOCTX"));
1462 		KASSERT(PCPU_GET(fpcurthread) == NULL,
1463 		    ("non-NULL fpcurthread for PCB_NPXNOSAVE"));
1464 		CRITICAL_ASSERT(td);
1465 
1466 		pcb->pcb_flags &= ~(PCB_NPXNOSAVE | PCB_NPXINITDONE);
1467 		fpu_disable();
1468 	} else {
1469 		KASSERT((ctx->flags & FPU_KERN_CTX_INUSE) != 0,
1470 		    ("leaving not inuse ctx"));
1471 		ctx->flags &= ~FPU_KERN_CTX_INUSE;
1472 
1473 		if (is_fpu_kern_thread(0) &&
1474 		    (ctx->flags & FPU_KERN_CTX_DUMMY) != 0)
1475 			return (0);
1476 		KASSERT((ctx->flags & FPU_KERN_CTX_DUMMY) == 0,
1477 		    ("dummy ctx"));
1478 		critical_enter();
1479 		if (curthread == PCPU_GET(fpcurthread))
1480 			npxdrop();
1481 		pcb->pcb_save = ctx->prev;
1482 	}
1483 
1484 	if (pcb->pcb_save == get_pcb_user_save_pcb(pcb)) {
1485 		if ((pcb->pcb_flags & PCB_NPXUSERINITDONE) != 0) {
1486 			pcb->pcb_flags |= PCB_NPXINITDONE;
1487 			if ((pcb->pcb_flags & PCB_KERNNPX_THR) == 0)
1488 				pcb->pcb_flags &= ~PCB_KERNNPX;
1489 		} else if ((pcb->pcb_flags & PCB_KERNNPX_THR) == 0)
1490 			pcb->pcb_flags &= ~(PCB_NPXINITDONE | PCB_KERNNPX);
1491 	} else {
1492 		if ((ctx->flags & FPU_KERN_CTX_NPXINITDONE) != 0)
1493 			pcb->pcb_flags |= PCB_NPXINITDONE;
1494 		else
1495 			pcb->pcb_flags &= ~PCB_NPXINITDONE;
1496 		KASSERT(!PCB_USER_FPU(pcb), ("unpaired fpu_kern_leave"));
1497 	}
1498 	critical_exit();
1499 	return (0);
1500 }
1501 
1502 int
fpu_kern_thread(u_int flags)1503 fpu_kern_thread(u_int flags)
1504 {
1505 
1506 	KASSERT((curthread->td_pflags & TDP_KTHREAD) != 0,
1507 	    ("Only kthread may use fpu_kern_thread"));
1508 	KASSERT(curpcb->pcb_save == get_pcb_user_save_pcb(curpcb),
1509 	    ("mangled pcb_save"));
1510 	KASSERT(PCB_USER_FPU(curpcb), ("recursive call"));
1511 
1512 	curpcb->pcb_flags |= PCB_KERNNPX | PCB_KERNNPX_THR;
1513 	return (0);
1514 }
1515 
1516 int
is_fpu_kern_thread(u_int flags)1517 is_fpu_kern_thread(u_int flags)
1518 {
1519 
1520 	if ((curthread->td_pflags & TDP_KTHREAD) == 0)
1521 		return (0);
1522 	return ((curpcb->pcb_flags & PCB_KERNNPX_THR) != 0);
1523 }
1524 
1525 /*
1526  * FPU save area alloc/free/init utility routines
1527  */
1528 union savefpu *
fpu_save_area_alloc(void)1529 fpu_save_area_alloc(void)
1530 {
1531 
1532 	return (uma_zalloc(fpu_save_area_zone, M_WAITOK));
1533 }
1534 
1535 void
fpu_save_area_free(union savefpu * fsa)1536 fpu_save_area_free(union savefpu *fsa)
1537 {
1538 
1539 	uma_zfree(fpu_save_area_zone, fsa);
1540 }
1541 
1542 void
fpu_save_area_reset(union savefpu * fsa)1543 fpu_save_area_reset(union savefpu *fsa)
1544 {
1545 
1546 	bcopy(npx_initialstate, fsa, cpu_max_ext_state_size);
1547 }
1548