]> git.saurik.com Git - apple/xnu.git/blob - osfmk/i386/fpu.c
xnu-3789.70.16.tar.gz
[apple/xnu.git] / osfmk / i386 / fpu.c
1 /*
2 * Copyright (c) 2000-2012 Apple Inc. All rights reserved.
3 *
4 * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
5 *
6 * This file contains Original Code and/or Modifications of Original Code
7 * as defined in and that are subject to the Apple Public Source License
8 * Version 2.0 (the 'License'). You may not use this file except in
9 * compliance with the License. The rights granted to you under the License
10 * may not be used to create, or enable the creation or redistribution of,
11 * unlawful or unlicensed copies of an Apple operating system, or to
12 * circumvent, violate, or enable the circumvention or violation of, any
13 * terms of an Apple operating system software license agreement.
14 *
15 * Please obtain a copy of the License at
16 * http://www.opensource.apple.com/apsl/ and read it before using this file.
17 *
18 * The Original Code and all software distributed under the License are
19 * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
20 * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
21 * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
22 * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
23 * Please see the License for the specific language governing rights and
24 * limitations under the License.
25 *
26 * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
27 */
28 /*
29 * @OSF_COPYRIGHT@
30 */
31 /*
32 * Mach Operating System
33 * Copyright (c) 1992-1990 Carnegie Mellon University
34 * All Rights Reserved.
35 *
36 * Permission to use, copy, modify and distribute this software and its
37 * documentation is hereby granted, provided that both the copyright
38 * notice and this permission notice appear in all copies of the
39 * software, derivative works or modified versions, and any portions
40 * thereof, and that both notices appear in supporting documentation.
41 *
42 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
43 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND FOR
44 * ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
45 *
46 * Carnegie Mellon requests users of this software to return to
47 *
48 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
49 * School of Computer Science
50 * Carnegie Mellon University
51 * Pittsburgh PA 15213-3890
52 *
53 * any improvements or extensions that they make and grant Carnegie Mellon
54 * the rights to redistribute these changes.
55 */
56 /*
57 */
58
59
60 #include <mach/exception_types.h>
61 #include <mach/i386/thread_status.h>
62 #include <mach/i386/fp_reg.h>
63 #include <mach/branch_predicates.h>
64
65 #include <kern/mach_param.h>
66 #include <kern/processor.h>
67 #include <kern/thread.h>
68 #include <kern/zalloc.h>
69 #include <kern/misc_protos.h>
70 #include <kern/spl.h>
71 #include <kern/assert.h>
72
73 #include <libkern/OSAtomic.h>
74
75 #include <architecture/i386/pio.h>
76 #include <i386/cpuid.h>
77 #include <i386/fpu.h>
78 #include <i386/proc_reg.h>
79 #include <i386/misc_protos.h>
80 #include <i386/thread.h>
81 #include <i386/trap.h>
82
83 int fp_kind = FP_NO; /* not inited */
84 zone_t ifps_zone; /* zone for FPU save area */
85
86 #define ALIGNED(addr,size) (((uintptr_t)(addr)&((size)-1))==0)
87
88 /* Forward */
89
90 extern void fpinit(void);
91 extern void fp_save(
92 thread_t thr_act);
93 extern void fp_load(
94 thread_t thr_act);
95
96 static void configure_mxcsr_capability_mask(struct x86_avx_thread_state *fps);
97
98 struct x86_avx_thread_state initial_fp_state __attribute((aligned(64)));
99
100
101 /* Global MXCSR capability bitmask */
102 static unsigned int mxcsr_capability_mask;
103
104 #define fninit() \
105 __asm__ volatile("fninit")
106
107 #define fnstcw(control) \
108 __asm__("fnstcw %0" : "=m" (*(unsigned short *)(control)))
109
110 #define fldcw(control) \
111 __asm__ volatile("fldcw %0" : : "m" (*(unsigned short *) &(control)) )
112
113 #define fnclex() \
114 __asm__ volatile("fnclex")
115
116 #define fnsave(state) \
117 __asm__ volatile("fnsave %0" : "=m" (*state))
118
119 #define frstor(state) \
120 __asm__ volatile("frstor %0" : : "m" (state))
121
122 #define fwait() \
123 __asm__("fwait");
124
125 #define fxrstor(addr) __asm__ __volatile__("fxrstor %0" : : "m" (*(addr)))
126 #define fxsave(addr) __asm__ __volatile__("fxsave %0" : "=m" (*(addr)))
127
128 static uint32_t fp_register_state_size = 0;
129 static uint32_t fpu_YMM_present = FALSE;
130 static uint32_t cpuid_reevaluated = 0;
131
132 static void fpu_store_registers(void *, boolean_t);
133 static void fpu_load_registers(void *);
134
135 extern void xsave64o(void);
136 extern void xrstor64o(void);
137
138 #define XMASK ((uint32_t) (XFEM_X87 | XFEM_SSE | XFEM_YMM))
139
140 static inline void xsetbv(uint32_t mask_hi, uint32_t mask_lo) {
141 __asm__ __volatile__("xsetbv" :: "a"(mask_lo), "d"(mask_hi), "c" (XCR0));
142 }
143
144 static inline void xsave(struct x86_fx_thread_state *a) {
145 __asm__ __volatile__("xsave %0" :"=m" (*a) : "a"(XMASK), "d"(0));
146 }
147
148 static inline void xrstor(struct x86_fx_thread_state *a) {
149 __asm__ __volatile__("xrstor %0" :: "m" (*a), "a"(XMASK), "d"(0));
150 }
151
152 #if DEBUG
153 static inline unsigned short
154 fnstsw(void)
155 {
156 unsigned short status;
157 __asm__ volatile("fnstsw %0" : "=ma" (status));
158 return(status);
159 }
160 #endif
161
162 /*
163 * Configure the initial FPU state presented to new threads.
164 * Determine the MXCSR capability mask, which allows us to mask off any
165 * potentially unsafe "reserved" bits before restoring the FPU context.
166 * *Not* per-cpu, assumes symmetry.
167 */
168
169 static void
170 configure_mxcsr_capability_mask(struct x86_avx_thread_state *fps)
171 {
172 /* XSAVE requires a 64 byte aligned store */
173 assert(ALIGNED(fps, 64));
174 /* Clear, to prepare for the diagnostic FXSAVE */
175 bzero(fps, sizeof(*fps));
176
177 fpinit();
178 fpu_store_registers(fps, FALSE);
179
180 mxcsr_capability_mask = fps->fx_MXCSR_MASK;
181
182 /* Set default mask value if necessary */
183 if (mxcsr_capability_mask == 0)
184 mxcsr_capability_mask = 0xffbf;
185
186 /* Clear vector register store */
187 bzero(&fps->fx_XMM_reg[0][0], sizeof(fps->fx_XMM_reg));
188 bzero(&fps->x_YMMH_reg[0][0], sizeof(fps->x_YMMH_reg));
189
190 fps->fp_valid = TRUE;
191 fps->fp_save_layout = fpu_YMM_present ? XSAVE32: FXSAVE32;
192 fpu_load_registers(fps);
193
194 /* Poison values to trap unsafe usage */
195 fps->fp_valid = 0xFFFFFFFF;
196 fps->fp_save_layout = FP_UNUSED;
197
198 /* Re-enable FPU/SSE DNA exceptions */
199 set_ts();
200 }
201
202
203 /*
204 * Look for FPU and initialize it.
205 * Called on each CPU.
206 */
207 void
208 init_fpu(void)
209 {
210 #if DEBUG
211 unsigned short status;
212 unsigned short control;
213 #endif
214 /*
215 * Check for FPU by initializing it,
216 * then trying to read the correct bit patterns from
217 * the control and status registers.
218 */
219 set_cr0((get_cr0() & ~(CR0_EM|CR0_TS)) | CR0_NE); /* allow use of FPU */
220 fninit();
221 #if DEBUG
222 status = fnstsw();
223 fnstcw(&control);
224
225 assert(((status & 0xff) == 0) && ((control & 0x103f) == 0x3f));
226 #endif
227 /* Advertise SSE support */
228 if (cpuid_features() & CPUID_FEATURE_FXSR) {
229 fp_kind = FP_FXSR;
230 set_cr4(get_cr4() | CR4_OSFXS);
231 /* And allow SIMD exceptions if present */
232 if (cpuid_features() & CPUID_FEATURE_SSE) {
233 set_cr4(get_cr4() | CR4_OSXMM);
234 }
235 fp_register_state_size = sizeof(struct x86_fx_thread_state);
236
237 } else
238 panic("fpu is not FP_FXSR");
239
240 /* Configure the XSAVE context mechanism if the processor supports
241 * AVX/YMM registers
242 */
243 if (cpuid_features() & CPUID_FEATURE_XSAVE) {
244 cpuid_xsave_leaf_t *xsp = &cpuid_info()->cpuid_xsave_leaf[0];
245 if (xsp->extended_state[0] & (uint32_t)XFEM_YMM) {
246 assert(xsp->extended_state[0] & (uint32_t) XFEM_SSE);
247 /* XSAVE container size for all features */
248 fp_register_state_size = sizeof(struct x86_avx_thread_state);
249 fpu_YMM_present = TRUE;
250 set_cr4(get_cr4() | CR4_OSXSAVE);
251 xsetbv(0, XMASK);
252 /* Re-evaluate CPUID, once, to reflect OSXSAVE */
253 if (OSCompareAndSwap(0, 1, &cpuid_reevaluated))
254 cpuid_set_info();
255 /* Verify that now selected state can be accommodated */
256 assert(xsp->extended_state[1] == fp_register_state_size);
257 }
258 }
259 else
260 fpu_YMM_present = FALSE;
261
262 fpinit();
263
264 /*
265 * Trap wait instructions. Turn off FPU for now.
266 */
267 set_cr0(get_cr0() | CR0_TS | CR0_MP);
268 }
269
270 /*
271 * Allocate and initialize FP state for current thread.
272 * Don't load state.
273 */
274 static void *
275 fp_state_alloc(void)
276 {
277 struct x86_fx_thread_state *ifps = zalloc(ifps_zone);
278
279 #if DEBUG
280 if (!(ALIGNED(ifps,64))) {
281 panic("fp_state_alloc: %p, %u, %p, %u", ifps, (unsigned) ifps_zone->elem_size, (void *) ifps_zone->free_elements, (unsigned) ifps_zone->alloc_size);
282 }
283 #endif
284 bzero(ifps, sizeof(*ifps));
285 return ifps;
286 }
287
288 static inline void
289 fp_state_free(void *ifps)
290 {
291 zfree(ifps_zone, ifps);
292 }
293
294 void clear_fpu(void)
295 {
296 set_ts();
297 }
298
299
300 static void fpu_load_registers(void *fstate) {
301 struct x86_fx_thread_state *ifps = fstate;
302 fp_save_layout_t layout = ifps->fp_save_layout;
303
304 assert(layout == FXSAVE32 || layout == FXSAVE64 || layout == XSAVE32 || layout == XSAVE64);
305 assert(ALIGNED(ifps, 64));
306 assert(ml_get_interrupts_enabled() == FALSE);
307
308 #if DEBUG
309 if (layout == XSAVE32 || layout == XSAVE64) {
310 struct x86_avx_thread_state *iavx = fstate;
311 unsigned i;
312 /* Verify reserved bits in the XSAVE header*/
313 if (iavx->_xh.xsbv & ~7)
314 panic("iavx->_xh.xsbv: 0x%llx", iavx->_xh.xsbv);
315 for (i = 0; i < sizeof(iavx->_xh.xhrsvd); i++)
316 if (iavx->_xh.xhrsvd[i])
317 panic("Reserved bit set");
318 }
319 if (fpu_YMM_present) {
320 if (layout != XSAVE32 && layout != XSAVE64)
321 panic("Inappropriate layout: %u\n", layout);
322 }
323 #endif /* DEBUG */
324
325 if ((layout == XSAVE64) || (layout == XSAVE32))
326 xrstor(ifps);
327 else
328 fxrstor(ifps);
329 }
330
331 static void fpu_store_registers(void *fstate, boolean_t is64) {
332 struct x86_fx_thread_state *ifps = fstate;
333 assert(ALIGNED(ifps, 64));
334 if (fpu_YMM_present) {
335 xsave(ifps);
336 ifps->fp_save_layout = is64 ? XSAVE64 : XSAVE32;
337 }
338 else {
339 fxsave(ifps);
340 ifps->fp_save_layout = is64 ? FXSAVE64 : FXSAVE32;
341 }
342 }
343
344 /*
345 * Initialize FP handling.
346 */
347
348 void
349 fpu_module_init(void)
350 {
351 if ((fp_register_state_size != sizeof(struct x86_fx_thread_state)) &&
352 (fp_register_state_size != sizeof(struct x86_avx_thread_state)))
353 panic("fpu_module_init: incorrect savearea size %u\n", fp_register_state_size);
354
355 assert(fpu_YMM_present != 0xFFFFFFFF);
356
357 /* We explicitly choose an allocation size of 64
358 * to eliminate waste for the 832 byte sized
359 * AVX XSAVE register save area.
360 */
361 ifps_zone = zinit(fp_register_state_size,
362 thread_max * fp_register_state_size,
363 64 * fp_register_state_size,
364 "x86 fpsave state");
365
366 /* To maintain the required alignment, disable
367 * zone debugging for this zone as that appends
368 * 16 bytes to each element.
369 */
370 zone_change(ifps_zone, Z_ALIGNMENT_REQUIRED, TRUE);
371 /* Determine MXCSR reserved bits and configure initial FPU state*/
372 configure_mxcsr_capability_mask(&initial_fp_state);
373 }
374
375 /*
376 * Save thread`s FPU context.
377 */
378 void
379 fpu_save_context(thread_t thread)
380 {
381 struct x86_fx_thread_state *ifps;
382
383 assert(ml_get_interrupts_enabled() == FALSE);
384 ifps = (thread)->machine.ifps;
385 #if DEBUG
386 if (ifps && ((ifps->fp_valid != FALSE) && (ifps->fp_valid != TRUE))) {
387 panic("ifps->fp_valid: %u\n", ifps->fp_valid);
388 }
389 #endif
390 if (ifps != 0 && (ifps->fp_valid == FALSE)) {
391 /* Clear CR0.TS in preparation for the FP context save. In
392 * theory, this shouldn't be necessary since a live FPU should
393 * indicate that TS is clear. However, various routines
394 * (such as sendsig & sigreturn) manipulate TS directly.
395 */
396 clear_ts();
397 /* registers are in FPU - save to memory */
398 fpu_store_registers(ifps, (thread_is_64bit(thread) && is_saved_state64(thread->machine.iss)));
399 ifps->fp_valid = TRUE;
400 }
401 set_ts();
402 }
403
404
405 /*
406 * Free a FPU save area.
407 * Called only when thread terminating - no locking necessary.
408 */
409 void
410 fpu_free(void *fps)
411 {
412 fp_state_free(fps);
413 }
414
415 /*
416 * Set the floating-point state for a thread based
417 * on the FXSave formatted data. This is basically
418 * the same as fpu_set_state except it uses the
419 * expanded data structure.
420 * If the thread is not the current thread, it is
421 * not running (held). Locking needed against
422 * concurrent fpu_set_state or fpu_get_state.
423 */
424 kern_return_t
425 fpu_set_fxstate(
426 thread_t thr_act,
427 thread_state_t tstate,
428 thread_flavor_t f)
429 {
430 struct x86_fx_thread_state *ifps;
431 struct x86_fx_thread_state *new_ifps;
432 x86_float_state64_t *state;
433 pcb_t pcb;
434 size_t state_size = sizeof(struct x86_fx_thread_state);
435 boolean_t old_valid, fresh_state = FALSE;
436
437 if (fp_kind == FP_NO)
438 return KERN_FAILURE;
439
440 if ((f == x86_AVX_STATE32 || f == x86_AVX_STATE64) &&
441 !ml_fpu_avx_enabled())
442 return KERN_FAILURE;
443
444 state = (x86_float_state64_t *)tstate;
445
446 assert(thr_act != THREAD_NULL);
447 pcb = THREAD_TO_PCB(thr_act);
448
449 if (state == NULL) {
450 /*
451 * new FPU state is 'invalid'.
452 * Deallocate the fp state if it exists.
453 */
454 simple_lock(&pcb->lock);
455
456 ifps = pcb->ifps;
457 pcb->ifps = 0;
458
459 simple_unlock(&pcb->lock);
460
461 if (ifps != 0) {
462 fp_state_free(ifps);
463 }
464 } else {
465 /*
466 * Valid incoming state. Allocate the fp state if there is none.
467 */
468 new_ifps = 0;
469 Retry:
470 simple_lock(&pcb->lock);
471
472 ifps = pcb->ifps;
473 if (ifps == 0) {
474 if (new_ifps == 0) {
475 simple_unlock(&pcb->lock);
476 new_ifps = fp_state_alloc();
477 goto Retry;
478 }
479 ifps = new_ifps;
480 new_ifps = 0;
481 pcb->ifps = ifps;
482 fresh_state = TRUE;
483 }
484
485 /*
486 * now copy over the new data.
487 */
488
489 old_valid = ifps->fp_valid;
490
491 #if DEBUG || DEVELOPMENT
492 if ((fresh_state == FALSE) && (old_valid == FALSE) && (thr_act != current_thread())) {
493 panic("fpu_set_fxstate inconsistency, thread: %p not stopped", thr_act);
494 }
495 #endif
496 /*
497 * Clear any reserved bits in the MXCSR to prevent a GPF
498 * when issuing an FXRSTOR.
499 */
500
501 state->fpu_mxcsr &= mxcsr_capability_mask;
502
503 bcopy((char *)&state->fpu_fcw, (char *)ifps, state_size);
504
505 if (fpu_YMM_present) {
506 struct x86_avx_thread_state *iavx = (void *) ifps;
507 uint32_t fpu_nyreg = 0;
508
509 if (f == x86_AVX_STATE32)
510 fpu_nyreg = 8;
511 else if (f == x86_AVX_STATE64)
512 fpu_nyreg = 16;
513
514 if (fpu_nyreg) {
515 x86_avx_state64_t *ystate = (x86_avx_state64_t *) state;
516 bcopy(&ystate->__fpu_ymmh0, &iavx->x_YMMH_reg[0][0], fpu_nyreg * sizeof(_STRUCT_XMM_REG));
517 }
518
519 iavx->fp_save_layout = thread_is_64bit(thr_act) ? XSAVE64 : XSAVE32;
520 /* Sanitize XSAVE header */
521 bzero(&iavx->_xh.xhrsvd[0], sizeof(iavx->_xh.xhrsvd));
522 if (fpu_nyreg)
523 iavx->_xh.xsbv = (XFEM_YMM | XFEM_SSE | XFEM_X87);
524 else
525 iavx->_xh.xsbv = (XFEM_SSE | XFEM_X87);
526 } else {
527 ifps->fp_save_layout = thread_is_64bit(thr_act) ? FXSAVE64 : FXSAVE32;
528 }
529 ifps->fp_valid = old_valid;
530
531 if (old_valid == FALSE) {
532 boolean_t istate = ml_set_interrupts_enabled(FALSE);
533 ifps->fp_valid = TRUE;
534 /* If altering the current thread's state, disable FPU */
535 if (thr_act == current_thread())
536 set_ts();
537
538 ml_set_interrupts_enabled(istate);
539 }
540
541 simple_unlock(&pcb->lock);
542
543 if (new_ifps != 0)
544 fp_state_free(new_ifps);
545 }
546 return KERN_SUCCESS;
547 }
548
549 /*
550 * Get the floating-point state for a thread.
551 * If the thread is not the current thread, it is
552 * not running (held). Locking needed against
553 * concurrent fpu_set_state or fpu_get_state.
554 */
555 kern_return_t
556 fpu_get_fxstate(
557 thread_t thr_act,
558 thread_state_t tstate,
559 thread_flavor_t f)
560 {
561 struct x86_fx_thread_state *ifps;
562 x86_float_state64_t *state;
563 kern_return_t ret = KERN_FAILURE;
564 pcb_t pcb;
565 size_t state_size = sizeof(struct x86_fx_thread_state);
566
567 if (fp_kind == FP_NO)
568 return KERN_FAILURE;
569
570 if ((f == x86_AVX_STATE32 || f == x86_AVX_STATE64) &&
571 !ml_fpu_avx_enabled())
572 return KERN_FAILURE;
573
574 state = (x86_float_state64_t *)tstate;
575
576 assert(thr_act != THREAD_NULL);
577 pcb = THREAD_TO_PCB(thr_act);
578
579 simple_lock(&pcb->lock);
580
581 ifps = pcb->ifps;
582 if (ifps == 0) {
583 /*
584 * No valid floating-point state.
585 */
586
587 bcopy((char *)&initial_fp_state, (char *)&state->fpu_fcw,
588 state_size);
589
590 simple_unlock(&pcb->lock);
591
592 return KERN_SUCCESS;
593 }
594 /*
595 * Make sure we`ve got the latest fp state info
596 * If the live fpu state belongs to our target
597 */
598 if (thr_act == current_thread()) {
599 boolean_t intr;
600
601 intr = ml_set_interrupts_enabled(FALSE);
602
603 clear_ts();
604 fp_save(thr_act);
605 clear_fpu();
606
607 (void)ml_set_interrupts_enabled(intr);
608 }
609 if (ifps->fp_valid) {
610 bcopy((char *)ifps, (char *)&state->fpu_fcw, state_size);
611 if (fpu_YMM_present) {
612 struct x86_avx_thread_state *iavx = (void *) ifps;
613 uint32_t fpu_nyreg = 0;
614
615 if (f == x86_AVX_STATE32)
616 fpu_nyreg = 8;
617 else if (f == x86_AVX_STATE64)
618 fpu_nyreg = 16;
619
620 if (fpu_nyreg) {
621 x86_avx_state64_t *ystate = (x86_avx_state64_t *) state;
622 bcopy(&iavx->x_YMMH_reg[0][0], &ystate->__fpu_ymmh0, fpu_nyreg * sizeof(_STRUCT_XMM_REG));
623 }
624 }
625
626 ret = KERN_SUCCESS;
627 }
628 simple_unlock(&pcb->lock);
629
630 return ret;
631 }
632
633
634
635 /*
636 * the child thread is 'stopped' with the thread
637 * mutex held and is currently not known by anyone
638 * so no way for fpu state to get manipulated by an
639 * outside agency -> no need for pcb lock
640 */
641
642 void
643 fpu_dup_fxstate(
644 thread_t parent,
645 thread_t child)
646 {
647 struct x86_fx_thread_state *new_ifps = NULL;
648 boolean_t intr;
649 pcb_t ppcb;
650
651 ppcb = THREAD_TO_PCB(parent);
652
653 if (ppcb->ifps == NULL)
654 return;
655
656 if (child->machine.ifps)
657 panic("fpu_dup_fxstate: child's ifps non-null");
658
659 new_ifps = fp_state_alloc();
660
661 simple_lock(&ppcb->lock);
662
663 if (ppcb->ifps != NULL) {
664 struct x86_fx_thread_state *ifps = ppcb->ifps;
665 /*
666 * Make sure we`ve got the latest fp state info
667 */
668 if (current_thread() == parent) {
669 intr = ml_set_interrupts_enabled(FALSE);
670 assert(current_thread() == parent);
671 clear_ts();
672 fp_save(parent);
673 clear_fpu();
674
675 (void)ml_set_interrupts_enabled(intr);
676 }
677
678 if (ifps->fp_valid) {
679 child->machine.ifps = new_ifps;
680 assert((fp_register_state_size == sizeof(struct x86_fx_thread_state)) ||
681 (fp_register_state_size == sizeof(struct x86_avx_thread_state)));
682 bcopy((char *)(ppcb->ifps),
683 (char *)(child->machine.ifps), fp_register_state_size);
684
685 /* Mark the new fp saved state as non-live. */
686 /* Temporarily disabled: radar 4647827
687 * new_ifps->fp_valid = TRUE;
688 */
689
690 /*
691 * Clear any reserved bits in the MXCSR to prevent a GPF
692 * when issuing an FXRSTOR.
693 */
694 new_ifps->fx_MXCSR &= mxcsr_capability_mask;
695 new_ifps = NULL;
696 }
697 }
698 simple_unlock(&ppcb->lock);
699
700 if (new_ifps != NULL)
701 fp_state_free(new_ifps);
702 }
703
704
705 /*
706 * Initialize FPU.
707 *
708 */
709
710 void
711 fpinit(void)
712 {
713 unsigned short control;
714
715 clear_ts();
716 fninit();
717 fnstcw(&control);
718 control &= ~(FPC_PC|FPC_RC); /* Clear precision & rounding control */
719 control |= (FPC_PC_64 | /* Set precision */
720 FPC_RC_RN | /* round-to-nearest */
721 FPC_ZE | /* Suppress zero-divide */
722 FPC_OE | /* and overflow */
723 FPC_UE | /* underflow */
724 FPC_IE | /* Allow NaNQs and +-INF */
725 FPC_DE | /* Allow denorms as operands */
726 FPC_PE); /* No trap for precision loss */
727 fldcw(control);
728
729 /* Initialize SSE/SSE2 */
730 __builtin_ia32_ldmxcsr(0x1f80);
731 }
732
733 /*
734 * Coprocessor not present.
735 */
736
737 uint64_t x86_isr_fp_simd_use;
738
739 void
740 fpnoextflt(void)
741 {
742 boolean_t intr;
743 thread_t thr_act;
744 pcb_t pcb;
745 struct x86_fx_thread_state *ifps = 0;
746
747 thr_act = current_thread();
748 pcb = THREAD_TO_PCB(thr_act);
749
750 assert(fp_register_state_size != 0);
751
752 if (pcb->ifps == 0 && !get_interrupt_level()) {
753 ifps = fp_state_alloc();
754 bcopy((char *)&initial_fp_state, (char *)ifps,
755 fp_register_state_size);
756 if (!thread_is_64bit(thr_act)) {
757 ifps->fp_save_layout = fpu_YMM_present ? XSAVE32 : FXSAVE32;
758 }
759 else
760 ifps->fp_save_layout = fpu_YMM_present ? XSAVE64 : FXSAVE64;
761 ifps->fp_valid = TRUE;
762 }
763 intr = ml_set_interrupts_enabled(FALSE);
764
765 clear_ts(); /* Enable FPU use */
766
767 if (__improbable(get_interrupt_level())) {
768 /* Track number of #DNA traps at interrupt context,
769 * which is likely suboptimal. Racy, but good enough.
770 */
771 x86_isr_fp_simd_use++;
772 /*
773 * Save current FP/SIMD context if valid
774 * Initialize live FP/SIMD registers
775 */
776 if (pcb->ifps) {
777 fp_save(thr_act);
778 }
779 fpinit();
780 } else {
781 if (pcb->ifps == 0) {
782 pcb->ifps = ifps;
783 ifps = 0;
784 }
785 /*
786 * Load this thread`s state into coprocessor live context.
787 */
788 fp_load(thr_act);
789 }
790 (void)ml_set_interrupts_enabled(intr);
791
792 if (ifps)
793 fp_state_free(ifps);
794 }
795
796 /*
797 * FPU overran end of segment.
798 * Re-initialize FPU. Floating point state is not valid.
799 */
800
801 void
802 fpextovrflt(void)
803 {
804 thread_t thr_act = current_thread();
805 pcb_t pcb;
806 struct x86_fx_thread_state *ifps;
807 boolean_t intr;
808
809 intr = ml_set_interrupts_enabled(FALSE);
810
811 if (get_interrupt_level())
812 panic("FPU segment overrun exception at interrupt context\n");
813 if (current_task() == kernel_task)
814 panic("FPU segment overrun exception in kernel thread context\n");
815
816 /*
817 * This is a non-recoverable error.
818 * Invalidate the thread`s FPU state.
819 */
820 pcb = THREAD_TO_PCB(thr_act);
821 simple_lock(&pcb->lock);
822 ifps = pcb->ifps;
823 pcb->ifps = 0;
824 simple_unlock(&pcb->lock);
825
826 /*
827 * Re-initialize the FPU.
828 */
829 clear_ts();
830 fninit();
831
832 /*
833 * And disable access.
834 */
835 clear_fpu();
836
837 (void)ml_set_interrupts_enabled(intr);
838
839 if (ifps)
840 zfree(ifps_zone, ifps);
841
842 /*
843 * Raise exception.
844 */
845 i386_exception(EXC_BAD_ACCESS, VM_PROT_READ|VM_PROT_EXECUTE, 0);
846 /*NOTREACHED*/
847 }
848
849 /*
850 * FPU error. Called by AST.
851 */
852
853 void
854 fpexterrflt(void)
855 {
856 thread_t thr_act = current_thread();
857 struct x86_fx_thread_state *ifps = thr_act->machine.ifps;
858 boolean_t intr;
859
860 intr = ml_set_interrupts_enabled(FALSE);
861
862 if (get_interrupt_level())
863 panic("FPU error exception at interrupt context\n");
864 if (current_task() == kernel_task)
865 panic("FPU error exception in kernel thread context\n");
866
867 /*
868 * Save the FPU state and turn off the FPU.
869 */
870 fp_save(thr_act);
871
872 (void)ml_set_interrupts_enabled(intr);
873
874 /*
875 * Raise FPU exception.
876 * Locking not needed on pcb->ifps,
877 * since thread is running.
878 */
879 i386_exception(EXC_ARITHMETIC,
880 EXC_I386_EXTERR,
881 ifps->fx_status);
882
883 /*NOTREACHED*/
884 }
885
886 /*
887 * Save FPU state.
888 *
889 * Locking not needed:
890 * . if called from fpu_get_state, pcb already locked.
891 * . if called from fpnoextflt or fp_intr, we are single-cpu
892 * . otherwise, thread is running.
893 * N.B.: Must be called with interrupts disabled
894 */
895
896 void
897 fp_save(
898 thread_t thr_act)
899 {
900 pcb_t pcb = THREAD_TO_PCB(thr_act);
901 struct x86_fx_thread_state *ifps = pcb->ifps;
902
903 assert(ifps != 0);
904 if (ifps != 0 && !ifps->fp_valid) {
905 assert((get_cr0() & CR0_TS) == 0);
906 /* registers are in FPU */
907 ifps->fp_valid = TRUE;
908 fpu_store_registers(ifps, thread_is_64bit(thr_act));
909 }
910 }
911
912 /*
913 * Restore FPU state from PCB.
914 *
915 * Locking not needed; always called on the current thread.
916 */
917
918 void
919 fp_load(
920 thread_t thr_act)
921 {
922 pcb_t pcb = THREAD_TO_PCB(thr_act);
923 struct x86_fx_thread_state *ifps = pcb->ifps;
924
925 assert(ifps);
926 #if DEBUG
927 if (ifps->fp_valid != FALSE && ifps->fp_valid != TRUE) {
928 panic("fp_load() invalid fp_valid: %u, fp_save_layout: %u\n",
929 ifps->fp_valid, ifps->fp_save_layout);
930 }
931 #endif
932
933 if (ifps->fp_valid == FALSE) {
934 fpinit();
935 } else {
936 fpu_load_registers(ifps);
937 }
938 ifps->fp_valid = FALSE; /* in FPU */
939 }
940
941 /*
942 * SSE arithmetic exception handling code.
943 * Basically the same as the x87 exception handler with a different subtype
944 */
945
946 void
947 fpSSEexterrflt(void)
948 {
949 thread_t thr_act = current_thread();
950 struct x86_fx_thread_state *ifps = thr_act->machine.ifps;
951 boolean_t intr;
952
953 intr = ml_set_interrupts_enabled(FALSE);
954
955 if (get_interrupt_level())
956 panic("SSE exception at interrupt context\n");
957 if (current_task() == kernel_task)
958 panic("SSE exception in kernel thread context\n");
959
960 /*
961 * Save the FPU state and turn off the FPU.
962 */
963 fp_save(thr_act);
964
965 (void)ml_set_interrupts_enabled(intr);
966 /*
967 * Raise FPU exception.
968 * Locking not needed on pcb->ifps,
969 * since thread is running.
970 */
971
972 i386_exception(EXC_ARITHMETIC,
973 EXC_I386_SSEEXTERR,
974 ifps->fx_MXCSR);
975 /*NOTREACHED*/
976 }
977
978 void
979 fp_setvalid(boolean_t value) {
980 thread_t thr_act = current_thread();
981 struct x86_fx_thread_state *ifps = thr_act->machine.ifps;
982
983 if (ifps) {
984 ifps->fp_valid = value;
985
986 if (value == TRUE) {
987 boolean_t istate = ml_set_interrupts_enabled(FALSE);
988 clear_fpu();
989 ml_set_interrupts_enabled(istate);
990 }
991 }
992 }
993
994 boolean_t
995 ml_fpu_avx_enabled(void) {
996 return (fpu_YMM_present == TRUE);
997 }