FreeBSD/Linux Kernel Cross Reference
sys/amd64/amd64/fpu.c
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 * 4. 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 * from: @(#)npx.c 7.2 (Berkeley) 5/12/91
31 */
32
33 #include <sys/cdefs.h>
34 __FBSDID("$FreeBSD$");
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/sysctl.h>
47 #include <machine/bus.h>
48 #include <sys/rman.h>
49 #include <sys/signalvar.h>
50 #include <vm/uma.h>
51
52 #include <machine/cputypes.h>
53 #include <machine/frame.h>
54 #include <machine/intr_machdep.h>
55 #include <machine/md_var.h>
56 #include <machine/pcb.h>
57 #include <machine/psl.h>
58 #include <machine/resource.h>
59 #include <machine/specialreg.h>
60 #include <machine/segments.h>
61 #include <machine/ucontext.h>
62
63 /*
64 * Floating point support.
65 */
66
67 #if defined(__GNUCLIKE_ASM) && !defined(lint)
68
69 #define fldcw(cw) __asm __volatile("fldcw %0" : : "m" (cw))
70 #define fnclex() __asm __volatile("fnclex")
71 #define fninit() __asm __volatile("fninit")
72 #define fnstcw(addr) __asm __volatile("fnstcw %0" : "=m" (*(addr)))
73 #define fnstsw(addr) __asm __volatile("fnstsw %0" : "=am" (*(addr)))
74 #define fxrstor(addr) __asm __volatile("fxrstor %0" : : "m" (*(addr)))
75 #define fxsave(addr) __asm __volatile("fxsave %0" : "=m" (*(addr)))
76 #define ldmxcsr(csr) __asm __volatile("ldmxcsr %0" : : "m" (csr))
77 #define stmxcsr(addr) __asm __volatile("stmxcsr %0" : : "m" (*(addr)))
78
79 static __inline void
80 xrstor(char *addr, uint64_t mask)
81 {
82 uint32_t low, hi;
83
84 low = mask;
85 hi = mask >> 32;
86 __asm __volatile("xrstor %0" : : "m" (*addr), "a" (low), "d" (hi));
87 }
88
89 static __inline void
90 xsave(char *addr, uint64_t mask)
91 {
92 uint32_t low, hi;
93
94 low = mask;
95 hi = mask >> 32;
96 __asm __volatile("xsave %0" : "=m" (*addr) : "a" (low), "d" (hi) :
97 "memory");
98 }
99
100 #else /* !(__GNUCLIKE_ASM && !lint) */
101
102 void fldcw(u_short cw);
103 void fnclex(void);
104 void fninit(void);
105 void fnstcw(caddr_t addr);
106 void fnstsw(caddr_t addr);
107 void fxsave(caddr_t addr);
108 void fxrstor(caddr_t addr);
109 void ldmxcsr(u_int csr);
110 void stmxcsr(u_int *csr);
111 void xrstor(char *addr, uint64_t mask);
112 void xsave(char *addr, uint64_t mask);
113
114 #endif /* __GNUCLIKE_ASM && !lint */
115
116 #define start_emulating() load_cr0(rcr0() | CR0_TS)
117 #define stop_emulating() clts()
118
119 CTASSERT(sizeof(struct savefpu) == 512);
120 CTASSERT(sizeof(struct xstate_hdr) == 64);
121 CTASSERT(sizeof(struct savefpu_ymm) == 832);
122
123 /*
124 * This requirement is to make it easier for asm code to calculate
125 * offset of the fpu save area from the pcb address. FPU save area
126 * must be 64-byte aligned.
127 */
128 CTASSERT(sizeof(struct pcb) % XSAVE_AREA_ALIGN == 0);
129
130 /*
131 * Ensure the copy of XCR0 saved in a core is contained in the padding
132 * area.
133 */
134 CTASSERT(X86_XSTATE_XCR0_OFFSET >= offsetof(struct savefpu, sv_pad) &&
135 X86_XSTATE_XCR0_OFFSET + sizeof(uint64_t) <= sizeof(struct savefpu));
136
137 static void fpu_clean_state(void);
138
139 SYSCTL_INT(_hw, HW_FLOATINGPT, floatingpoint, CTLFLAG_RD,
140 SYSCTL_NULL_INT_PTR, 1, "Floating point instructions executed in hardware");
141
142 int lazy_fpu_switch = 0;
143 SYSCTL_INT(_hw, OID_AUTO, lazy_fpu_switch, CTLFLAG_RWTUN | CTLFLAG_NOFETCH,
144 &lazy_fpu_switch, 0,
145 "Lazily load FPU context after context switch");
146
147 int use_xsave; /* non-static for cpu_switch.S */
148 uint64_t xsave_mask; /* the same */
149 static uma_zone_t fpu_save_area_zone;
150 static struct savefpu *fpu_initialstate;
151
152 struct xsave_area_elm_descr {
153 u_int offset;
154 u_int size;
155 } *xsave_area_desc;
156
157 void
158 fpusave(void *addr)
159 {
160
161 if (use_xsave)
162 xsave((char *)addr, xsave_mask);
163 else
164 fxsave((char *)addr);
165 }
166
167 void
168 fpurestore(void *addr)
169 {
170
171 if (use_xsave)
172 xrstor((char *)addr, xsave_mask);
173 else
174 fxrstor((char *)addr);
175 }
176
177 void
178 fpususpend(void *addr)
179 {
180 u_long cr0;
181
182 cr0 = rcr0();
183 stop_emulating();
184 fpusave(addr);
185 load_cr0(cr0);
186 }
187
188 void
189 fpuresume(void *addr)
190 {
191 u_long cr0;
192
193 cr0 = rcr0();
194 stop_emulating();
195 fninit();
196 if (use_xsave)
197 load_xcr(XCR0, xsave_mask);
198 fpurestore(addr);
199 load_cr0(cr0);
200 }
201
202 /*
203 * Enable XSAVE if supported and allowed by user.
204 * Calculate the xsave_mask.
205 */
206 static void
207 fpuinit_bsp1(void)
208 {
209 u_int cp[4];
210 uint64_t xsave_mask_user;
211
212 TUNABLE_INT_FETCH("hw.lazy_fpu_switch", &lazy_fpu_switch);
213 if ((cpu_feature2 & CPUID2_XSAVE) != 0) {
214 use_xsave = 1;
215 TUNABLE_INT_FETCH("hw.use_xsave", &use_xsave);
216 }
217 if (!use_xsave)
218 return;
219
220 cpuid_count(0xd, 0x0, cp);
221 xsave_mask = XFEATURE_ENABLED_X87 | XFEATURE_ENABLED_SSE;
222 if ((cp[0] & xsave_mask) != xsave_mask)
223 panic("CPU0 does not support X87 or SSE: %x", cp[0]);
224 xsave_mask = ((uint64_t)cp[3] << 32) | cp[0];
225 xsave_mask_user = xsave_mask;
226 TUNABLE_ULONG_FETCH("hw.xsave_mask", &xsave_mask_user);
227 xsave_mask_user |= XFEATURE_ENABLED_X87 | XFEATURE_ENABLED_SSE;
228 xsave_mask &= xsave_mask_user;
229 if ((xsave_mask & XFEATURE_AVX512) != XFEATURE_AVX512)
230 xsave_mask &= ~XFEATURE_AVX512;
231 if ((xsave_mask & XFEATURE_MPX) != XFEATURE_MPX)
232 xsave_mask &= ~XFEATURE_MPX;
233
234 cpuid_count(0xd, 0x1, cp);
235 if ((cp[0] & CPUID_EXTSTATE_XSAVEOPT) != 0) {
236 /*
237 * Patch the XSAVE instruction in the cpu_switch code
238 * to XSAVEOPT. We assume that XSAVE encoding used
239 * REX byte, and set the bit 4 of the r/m byte.
240 */
241 ctx_switch_xsave[3] |= 0x10;
242 }
243 }
244
245 /*
246 * Calculate the fpu save area size.
247 */
248 static void
249 fpuinit_bsp2(void)
250 {
251 u_int cp[4];
252
253 if (use_xsave) {
254 cpuid_count(0xd, 0x0, cp);
255 cpu_max_ext_state_size = cp[1];
256
257 /*
258 * Reload the cpu_feature2, since we enabled OSXSAVE.
259 */
260 do_cpuid(1, cp);
261 cpu_feature2 = cp[2];
262 } else
263 cpu_max_ext_state_size = sizeof(struct savefpu);
264 }
265
266 /*
267 * Initialize the floating point unit.
268 */
269 void
270 fpuinit(void)
271 {
272 register_t saveintr;
273 u_int mxcsr;
274 u_short control;
275
276 if (IS_BSP())
277 fpuinit_bsp1();
278
279 if (use_xsave) {
280 load_cr4(rcr4() | CR4_XSAVE);
281 load_xcr(XCR0, xsave_mask);
282 }
283
284 /*
285 * XCR0 shall be set up before CPU can report the save area size.
286 */
287 if (IS_BSP())
288 fpuinit_bsp2();
289
290 /*
291 * It is too early for critical_enter() to work on AP.
292 */
293 saveintr = intr_disable();
294 stop_emulating();
295 fninit();
296 control = __INITIAL_FPUCW__;
297 fldcw(control);
298 mxcsr = __INITIAL_MXCSR__;
299 ldmxcsr(mxcsr);
300 start_emulating();
301 intr_restore(saveintr);
302 }
303
304 /*
305 * On the boot CPU we generate a clean state that is used to
306 * initialize the floating point unit when it is first used by a
307 * process.
308 */
309 static void
310 fpuinitstate(void *arg __unused)
311 {
312 register_t saveintr;
313 int cp[4], i, max_ext_n;
314
315 fpu_initialstate = malloc(cpu_max_ext_state_size, M_DEVBUF,
316 M_WAITOK | M_ZERO);
317 saveintr = intr_disable();
318 stop_emulating();
319
320 fpusave(fpu_initialstate);
321 if (fpu_initialstate->sv_env.en_mxcsr_mask)
322 cpu_mxcsr_mask = fpu_initialstate->sv_env.en_mxcsr_mask;
323 else
324 cpu_mxcsr_mask = 0xFFBF;
325
326 /*
327 * The fninit instruction does not modify XMM registers or x87
328 * registers (MM/ST). The fpusave call dumped the garbage
329 * contained in the registers after reset to the initial state
330 * saved. Clear XMM and x87 registers file image to make the
331 * startup program state and signal handler XMM/x87 register
332 * content predictable.
333 */
334 bzero(fpu_initialstate->sv_fp, sizeof(fpu_initialstate->sv_fp));
335 bzero(fpu_initialstate->sv_xmm, sizeof(fpu_initialstate->sv_xmm));
336
337 /*
338 * Create a table describing the layout of the CPU Extended
339 * Save Area.
340 */
341 if (use_xsave) {
342 max_ext_n = flsl(xsave_mask);
343 xsave_area_desc = malloc(max_ext_n * sizeof(struct
344 xsave_area_elm_descr), M_DEVBUF, M_WAITOK | M_ZERO);
345 /* x87 state */
346 xsave_area_desc[0].offset = 0;
347 xsave_area_desc[0].size = 160;
348 /* XMM */
349 xsave_area_desc[1].offset = 160;
350 xsave_area_desc[1].size = 288 - 160;
351
352 for (i = 2; i < max_ext_n; i++) {
353 cpuid_count(0xd, i, cp);
354 xsave_area_desc[i].offset = cp[1];
355 xsave_area_desc[i].size = cp[0];
356 }
357 }
358
359 fpu_save_area_zone = uma_zcreate("FPU_save_area",
360 cpu_max_ext_state_size, NULL, NULL, NULL, NULL,
361 XSAVE_AREA_ALIGN - 1, 0);
362
363 start_emulating();
364 intr_restore(saveintr);
365 }
366 /* EFIRT needs this to be initialized before we can enter our EFI environment */
367 SYSINIT(fpuinitstate, SI_SUB_DRIVERS, SI_ORDER_FIRST, fpuinitstate, NULL);
368
369 /*
370 * Free coprocessor (if we have it).
371 */
372 void
373 fpuexit(struct thread *td)
374 {
375
376 critical_enter();
377 if (curthread == PCPU_GET(fpcurthread)) {
378 stop_emulating();
379 fpusave(curpcb->pcb_save);
380 start_emulating();
381 PCPU_SET(fpcurthread, NULL);
382 }
383 critical_exit();
384 }
385
386 int
387 fpuformat(void)
388 {
389
390 return (_MC_FPFMT_XMM);
391 }
392
393 /*
394 * The following mechanism is used to ensure that the FPE_... value
395 * that is passed as a trapcode to the signal handler of the user
396 * process does not have more than one bit set.
397 *
398 * Multiple bits may be set if the user process modifies the control
399 * word while a status word bit is already set. While this is a sign
400 * of bad coding, we have no choise than to narrow them down to one
401 * bit, since we must not send a trapcode that is not exactly one of
402 * the FPE_ macros.
403 *
404 * The mechanism has a static table with 127 entries. Each combination
405 * of the 7 FPU status word exception bits directly translates to a
406 * position in this table, where a single FPE_... value is stored.
407 * This FPE_... value stored there is considered the "most important"
408 * of the exception bits and will be sent as the signal code. The
409 * precedence of the bits is based upon Intel Document "Numerical
410 * Applications", Chapter "Special Computational Situations".
411 *
412 * The macro to choose one of these values does these steps: 1) Throw
413 * away status word bits that cannot be masked. 2) Throw away the bits
414 * currently masked in the control word, assuming the user isn't
415 * interested in them anymore. 3) Reinsert status word bit 7 (stack
416 * fault) if it is set, which cannot be masked but must be presered.
417 * 4) Use the remaining bits to point into the trapcode table.
418 *
419 * The 6 maskable bits in order of their preference, as stated in the
420 * above referenced Intel manual:
421 * 1 Invalid operation (FP_X_INV)
422 * 1a Stack underflow
423 * 1b Stack overflow
424 * 1c Operand of unsupported format
425 * 1d SNaN operand.
426 * 2 QNaN operand (not an exception, irrelavant here)
427 * 3 Any other invalid-operation not mentioned above or zero divide
428 * (FP_X_INV, FP_X_DZ)
429 * 4 Denormal operand (FP_X_DNML)
430 * 5 Numeric over/underflow (FP_X_OFL, FP_X_UFL)
431 * 6 Inexact result (FP_X_IMP)
432 */
433 static char fpetable[128] = {
434 0,
435 FPE_FLTINV, /* 1 - INV */
436 FPE_FLTUND, /* 2 - DNML */
437 FPE_FLTINV, /* 3 - INV | DNML */
438 FPE_FLTDIV, /* 4 - DZ */
439 FPE_FLTINV, /* 5 - INV | DZ */
440 FPE_FLTDIV, /* 6 - DNML | DZ */
441 FPE_FLTINV, /* 7 - INV | DNML | DZ */
442 FPE_FLTOVF, /* 8 - OFL */
443 FPE_FLTINV, /* 9 - INV | OFL */
444 FPE_FLTUND, /* A - DNML | OFL */
445 FPE_FLTINV, /* B - INV | DNML | OFL */
446 FPE_FLTDIV, /* C - DZ | OFL */
447 FPE_FLTINV, /* D - INV | DZ | OFL */
448 FPE_FLTDIV, /* E - DNML | DZ | OFL */
449 FPE_FLTINV, /* F - INV | DNML | DZ | OFL */
450 FPE_FLTUND, /* 10 - UFL */
451 FPE_FLTINV, /* 11 - INV | UFL */
452 FPE_FLTUND, /* 12 - DNML | UFL */
453 FPE_FLTINV, /* 13 - INV | DNML | UFL */
454 FPE_FLTDIV, /* 14 - DZ | UFL */
455 FPE_FLTINV, /* 15 - INV | DZ | UFL */
456 FPE_FLTDIV, /* 16 - DNML | DZ | UFL */
457 FPE_FLTINV, /* 17 - INV | DNML | DZ | UFL */
458 FPE_FLTOVF, /* 18 - OFL | UFL */
459 FPE_FLTINV, /* 19 - INV | OFL | UFL */
460 FPE_FLTUND, /* 1A - DNML | OFL | UFL */
461 FPE_FLTINV, /* 1B - INV | DNML | OFL | UFL */
462 FPE_FLTDIV, /* 1C - DZ | OFL | UFL */
463 FPE_FLTINV, /* 1D - INV | DZ | OFL | UFL */
464 FPE_FLTDIV, /* 1E - DNML | DZ | OFL | UFL */
465 FPE_FLTINV, /* 1F - INV | DNML | DZ | OFL | UFL */
466 FPE_FLTRES, /* 20 - IMP */
467 FPE_FLTINV, /* 21 - INV | IMP */
468 FPE_FLTUND, /* 22 - DNML | IMP */
469 FPE_FLTINV, /* 23 - INV | DNML | IMP */
470 FPE_FLTDIV, /* 24 - DZ | IMP */
471 FPE_FLTINV, /* 25 - INV | DZ | IMP */
472 FPE_FLTDIV, /* 26 - DNML | DZ | IMP */
473 FPE_FLTINV, /* 27 - INV | DNML | DZ | IMP */
474 FPE_FLTOVF, /* 28 - OFL | IMP */
475 FPE_FLTINV, /* 29 - INV | OFL | IMP */
476 FPE_FLTUND, /* 2A - DNML | OFL | IMP */
477 FPE_FLTINV, /* 2B - INV | DNML | OFL | IMP */
478 FPE_FLTDIV, /* 2C - DZ | OFL | IMP */
479 FPE_FLTINV, /* 2D - INV | DZ | OFL | IMP */
480 FPE_FLTDIV, /* 2E - DNML | DZ | OFL | IMP */
481 FPE_FLTINV, /* 2F - INV | DNML | DZ | OFL | IMP */
482 FPE_FLTUND, /* 30 - UFL | IMP */
483 FPE_FLTINV, /* 31 - INV | UFL | IMP */
484 FPE_FLTUND, /* 32 - DNML | UFL | IMP */
485 FPE_FLTINV, /* 33 - INV | DNML | UFL | IMP */
486 FPE_FLTDIV, /* 34 - DZ | UFL | IMP */
487 FPE_FLTINV, /* 35 - INV | DZ | UFL | IMP */
488 FPE_FLTDIV, /* 36 - DNML | DZ | UFL | IMP */
489 FPE_FLTINV, /* 37 - INV | DNML | DZ | UFL | IMP */
490 FPE_FLTOVF, /* 38 - OFL | UFL | IMP */
491 FPE_FLTINV, /* 39 - INV | OFL | UFL | IMP */
492 FPE_FLTUND, /* 3A - DNML | OFL | UFL | IMP */
493 FPE_FLTINV, /* 3B - INV | DNML | OFL | UFL | IMP */
494 FPE_FLTDIV, /* 3C - DZ | OFL | UFL | IMP */
495 FPE_FLTINV, /* 3D - INV | DZ | OFL | UFL | IMP */
496 FPE_FLTDIV, /* 3E - DNML | DZ | OFL | UFL | IMP */
497 FPE_FLTINV, /* 3F - INV | DNML | DZ | OFL | UFL | IMP */
498 FPE_FLTSUB, /* 40 - STK */
499 FPE_FLTSUB, /* 41 - INV | STK */
500 FPE_FLTUND, /* 42 - DNML | STK */
501 FPE_FLTSUB, /* 43 - INV | DNML | STK */
502 FPE_FLTDIV, /* 44 - DZ | STK */
503 FPE_FLTSUB, /* 45 - INV | DZ | STK */
504 FPE_FLTDIV, /* 46 - DNML | DZ | STK */
505 FPE_FLTSUB, /* 47 - INV | DNML | DZ | STK */
506 FPE_FLTOVF, /* 48 - OFL | STK */
507 FPE_FLTSUB, /* 49 - INV | OFL | STK */
508 FPE_FLTUND, /* 4A - DNML | OFL | STK */
509 FPE_FLTSUB, /* 4B - INV | DNML | OFL | STK */
510 FPE_FLTDIV, /* 4C - DZ | OFL | STK */
511 FPE_FLTSUB, /* 4D - INV | DZ | OFL | STK */
512 FPE_FLTDIV, /* 4E - DNML | DZ | OFL | STK */
513 FPE_FLTSUB, /* 4F - INV | DNML | DZ | OFL | STK */
514 FPE_FLTUND, /* 50 - UFL | STK */
515 FPE_FLTSUB, /* 51 - INV | UFL | STK */
516 FPE_FLTUND, /* 52 - DNML | UFL | STK */
517 FPE_FLTSUB, /* 53 - INV | DNML | UFL | STK */
518 FPE_FLTDIV, /* 54 - DZ | UFL | STK */
519 FPE_FLTSUB, /* 55 - INV | DZ | UFL | STK */
520 FPE_FLTDIV, /* 56 - DNML | DZ | UFL | STK */
521 FPE_FLTSUB, /* 57 - INV | DNML | DZ | UFL | STK */
522 FPE_FLTOVF, /* 58 - OFL | UFL | STK */
523 FPE_FLTSUB, /* 59 - INV | OFL | UFL | STK */
524 FPE_FLTUND, /* 5A - DNML | OFL | UFL | STK */
525 FPE_FLTSUB, /* 5B - INV | DNML | OFL | UFL | STK */
526 FPE_FLTDIV, /* 5C - DZ | OFL | UFL | STK */
527 FPE_FLTSUB, /* 5D - INV | DZ | OFL | UFL | STK */
528 FPE_FLTDIV, /* 5E - DNML | DZ | OFL | UFL | STK */
529 FPE_FLTSUB, /* 5F - INV | DNML | DZ | OFL | UFL | STK */
530 FPE_FLTRES, /* 60 - IMP | STK */
531 FPE_FLTSUB, /* 61 - INV | IMP | STK */
532 FPE_FLTUND, /* 62 - DNML | IMP | STK */
533 FPE_FLTSUB, /* 63 - INV | DNML | IMP | STK */
534 FPE_FLTDIV, /* 64 - DZ | IMP | STK */
535 FPE_FLTSUB, /* 65 - INV | DZ | IMP | STK */
536 FPE_FLTDIV, /* 66 - DNML | DZ | IMP | STK */
537 FPE_FLTSUB, /* 67 - INV | DNML | DZ | IMP | STK */
538 FPE_FLTOVF, /* 68 - OFL | IMP | STK */
539 FPE_FLTSUB, /* 69 - INV | OFL | IMP | STK */
540 FPE_FLTUND, /* 6A - DNML | OFL | IMP | STK */
541 FPE_FLTSUB, /* 6B - INV | DNML | OFL | IMP | STK */
542 FPE_FLTDIV, /* 6C - DZ | OFL | IMP | STK */
543 FPE_FLTSUB, /* 6D - INV | DZ | OFL | IMP | STK */
544 FPE_FLTDIV, /* 6E - DNML | DZ | OFL | IMP | STK */
545 FPE_FLTSUB, /* 6F - INV | DNML | DZ | OFL | IMP | STK */
546 FPE_FLTUND, /* 70 - UFL | IMP | STK */
547 FPE_FLTSUB, /* 71 - INV | UFL | IMP | STK */
548 FPE_FLTUND, /* 72 - DNML | UFL | IMP | STK */
549 FPE_FLTSUB, /* 73 - INV | DNML | UFL | IMP | STK */
550 FPE_FLTDIV, /* 74 - DZ | UFL | IMP | STK */
551 FPE_FLTSUB, /* 75 - INV | DZ | UFL | IMP | STK */
552 FPE_FLTDIV, /* 76 - DNML | DZ | UFL | IMP | STK */
553 FPE_FLTSUB, /* 77 - INV | DNML | DZ | UFL | IMP | STK */
554 FPE_FLTOVF, /* 78 - OFL | UFL | IMP | STK */
555 FPE_FLTSUB, /* 79 - INV | OFL | UFL | IMP | STK */
556 FPE_FLTUND, /* 7A - DNML | OFL | UFL | IMP | STK */
557 FPE_FLTSUB, /* 7B - INV | DNML | OFL | UFL | IMP | STK */
558 FPE_FLTDIV, /* 7C - DZ | OFL | UFL | IMP | STK */
559 FPE_FLTSUB, /* 7D - INV | DZ | OFL | UFL | IMP | STK */
560 FPE_FLTDIV, /* 7E - DNML | DZ | OFL | UFL | IMP | STK */
561 FPE_FLTSUB, /* 7F - INV | DNML | DZ | OFL | UFL | IMP | STK */
562 };
563
564 /*
565 * Read the FP status and control words, then generate si_code value
566 * for SIGFPE. The error code chosen will be one of the
567 * FPE_... macros. It will be sent as the second argument to old
568 * BSD-style signal handlers and as "siginfo_t->si_code" (second
569 * argument) to SA_SIGINFO signal handlers.
570 *
571 * Some time ago, we cleared the x87 exceptions with FNCLEX there.
572 * Clearing exceptions was necessary mainly to avoid IRQ13 bugs. The
573 * usermode code which understands the FPU hardware enough to enable
574 * the exceptions, can also handle clearing the exception state in the
575 * handler. The only consequence of not clearing the exception is the
576 * rethrow of the SIGFPE on return from the signal handler and
577 * reexecution of the corresponding instruction.
578 *
579 * For XMM traps, the exceptions were never cleared.
580 */
581 int
582 fputrap_x87(void)
583 {
584 struct savefpu *pcb_save;
585 u_short control, status;
586
587 critical_enter();
588
589 /*
590 * Interrupt handling (for another interrupt) may have pushed the
591 * state to memory. Fetch the relevant parts of the state from
592 * wherever they are.
593 */
594 if (PCPU_GET(fpcurthread) != curthread) {
595 pcb_save = curpcb->pcb_save;
596 control = pcb_save->sv_env.en_cw;
597 status = pcb_save->sv_env.en_sw;
598 } else {
599 fnstcw(&control);
600 fnstsw(&status);
601 }
602
603 critical_exit();
604 return (fpetable[status & ((~control & 0x3f) | 0x40)]);
605 }
606
607 int
608 fputrap_sse(void)
609 {
610 u_int mxcsr;
611
612 critical_enter();
613 if (PCPU_GET(fpcurthread) != curthread)
614 mxcsr = curpcb->pcb_save->sv_env.en_mxcsr;
615 else
616 stmxcsr(&mxcsr);
617 critical_exit();
618 return (fpetable[(mxcsr & (~mxcsr >> 7)) & 0x3f]);
619 }
620
621 static void
622 restore_fpu_curthread(struct thread *td)
623 {
624 struct pcb *pcb;
625
626 /*
627 * Record new context early in case frstor causes a trap.
628 */
629 PCPU_SET(fpcurthread, td);
630
631 stop_emulating();
632 fpu_clean_state();
633 pcb = td->td_pcb;
634
635 if ((pcb->pcb_flags & PCB_FPUINITDONE) == 0) {
636 /*
637 * This is the first time this thread has used the FPU or
638 * the PCB doesn't contain a clean FPU state. Explicitly
639 * load an initial state.
640 *
641 * We prefer to restore the state from the actual save
642 * area in PCB instead of directly loading from
643 * fpu_initialstate, to ignite the XSAVEOPT
644 * tracking engine.
645 */
646 bcopy(fpu_initialstate, pcb->pcb_save,
647 cpu_max_ext_state_size);
648 fpurestore(pcb->pcb_save);
649 if (pcb->pcb_initial_fpucw != __INITIAL_FPUCW__)
650 fldcw(pcb->pcb_initial_fpucw);
651 if (PCB_USER_FPU(pcb))
652 set_pcb_flags(pcb, PCB_FPUINITDONE |
653 PCB_USERFPUINITDONE);
654 else
655 set_pcb_flags(pcb, PCB_FPUINITDONE);
656 } else
657 fpurestore(pcb->pcb_save);
658 }
659
660 /*
661 * Device Not Available (DNA, #NM) exception handler.
662 *
663 * It would be better to switch FP context here (if curthread !=
664 * fpcurthread) and not necessarily for every context switch, but it
665 * is too hard to access foreign pcb's.
666 */
667 void
668 fpudna(void)
669 {
670 struct thread *td;
671
672 td = curthread;
673 /*
674 * This handler is entered with interrupts enabled, so context
675 * switches may occur before critical_enter() is executed. If
676 * a context switch occurs, then when we regain control, our
677 * state will have been completely restored. The CPU may
678 * change underneath us, but the only part of our context that
679 * lives in the CPU is CR0.TS and that will be "restored" by
680 * setting it on the new CPU.
681 */
682 critical_enter();
683
684 KASSERT((curpcb->pcb_flags & PCB_FPUNOSAVE) == 0,
685 ("fpudna while in fpu_kern_enter(FPU_KERN_NOCTX)"));
686 if (__predict_false(PCPU_GET(fpcurthread) == td)) {
687 /*
688 * Some virtual machines seems to set %cr0.TS at
689 * arbitrary moments. Silently clear the TS bit
690 * regardless of the eager/lazy FPU context switch
691 * mode.
692 */
693 stop_emulating();
694 } else {
695 if (__predict_false(PCPU_GET(fpcurthread) != NULL)) {
696 panic(
697 "fpudna: fpcurthread = %p (%d), curthread = %p (%d)\n",
698 PCPU_GET(fpcurthread),
699 PCPU_GET(fpcurthread)->td_tid, td, td->td_tid);
700 }
701 restore_fpu_curthread(td);
702 }
703 critical_exit();
704 }
705
706 void fpu_activate_sw(struct thread *td); /* Called from the context switch */
707 void
708 fpu_activate_sw(struct thread *td)
709 {
710
711 if (lazy_fpu_switch || (td->td_pflags & TDP_KTHREAD) != 0 ||
712 !PCB_USER_FPU(td->td_pcb)) {
713 PCPU_SET(fpcurthread, NULL);
714 start_emulating();
715 } else if (PCPU_GET(fpcurthread) != td) {
716 restore_fpu_curthread(td);
717 }
718 }
719
720 void
721 fpudrop(void)
722 {
723 struct thread *td;
724
725 td = PCPU_GET(fpcurthread);
726 KASSERT(td == curthread, ("fpudrop: fpcurthread != curthread"));
727 CRITICAL_ASSERT(td);
728 PCPU_SET(fpcurthread, NULL);
729 clear_pcb_flags(td->td_pcb, PCB_FPUINITDONE);
730 start_emulating();
731 }
732
733 /*
734 * Get the user state of the FPU into pcb->pcb_user_save without
735 * dropping ownership (if possible). It returns the FPU ownership
736 * status.
737 */
738 int
739 fpugetregs(struct thread *td)
740 {
741 struct pcb *pcb;
742 uint64_t *xstate_bv, bit;
743 char *sa;
744 int max_ext_n, i, owned;
745
746 pcb = td->td_pcb;
747 critical_enter();
748 if ((pcb->pcb_flags & PCB_USERFPUINITDONE) == 0) {
749 bcopy(fpu_initialstate, get_pcb_user_save_pcb(pcb),
750 cpu_max_ext_state_size);
751 get_pcb_user_save_pcb(pcb)->sv_env.en_cw =
752 pcb->pcb_initial_fpucw;
753 fpuuserinited(td);
754 critical_exit();
755 return (_MC_FPOWNED_PCB);
756 }
757 if (td == PCPU_GET(fpcurthread) && PCB_USER_FPU(pcb)) {
758 fpusave(get_pcb_user_save_pcb(pcb));
759 owned = _MC_FPOWNED_FPU;
760 } else {
761 owned = _MC_FPOWNED_PCB;
762 }
763 if (use_xsave) {
764 /*
765 * Handle partially saved state.
766 */
767 sa = (char *)get_pcb_user_save_pcb(pcb);
768 xstate_bv = (uint64_t *)(sa + sizeof(struct savefpu) +
769 offsetof(struct xstate_hdr, xstate_bv));
770 max_ext_n = flsl(xsave_mask);
771 for (i = 0; i < max_ext_n; i++) {
772 bit = 1ULL << i;
773 if ((xsave_mask & bit) == 0 || (*xstate_bv & bit) != 0)
774 continue;
775 bcopy((char *)fpu_initialstate +
776 xsave_area_desc[i].offset,
777 sa + xsave_area_desc[i].offset,
778 xsave_area_desc[i].size);
779 *xstate_bv |= bit;
780 }
781 }
782 critical_exit();
783 return (owned);
784 }
785
786 void
787 fpuuserinited(struct thread *td)
788 {
789 struct pcb *pcb;
790
791 CRITICAL_ASSERT(td);
792 pcb = td->td_pcb;
793 if (PCB_USER_FPU(pcb))
794 set_pcb_flags(pcb,
795 PCB_FPUINITDONE | PCB_USERFPUINITDONE);
796 else
797 set_pcb_flags(pcb, PCB_FPUINITDONE);
798 }
799
800 int
801 fpusetxstate(struct thread *td, char *xfpustate, size_t xfpustate_size)
802 {
803 struct xstate_hdr *hdr, *ehdr;
804 size_t len, max_len;
805 uint64_t bv;
806
807 /* XXXKIB should we clear all extended state in xstate_bv instead ? */
808 if (xfpustate == NULL)
809 return (0);
810 if (!use_xsave)
811 return (EOPNOTSUPP);
812
813 len = xfpustate_size;
814 if (len < sizeof(struct xstate_hdr))
815 return (EINVAL);
816 max_len = cpu_max_ext_state_size - sizeof(struct savefpu);
817 if (len > max_len)
818 return (EINVAL);
819
820 ehdr = (struct xstate_hdr *)xfpustate;
821 bv = ehdr->xstate_bv;
822
823 /*
824 * Avoid #gp.
825 */
826 if (bv & ~xsave_mask)
827 return (EINVAL);
828
829 hdr = (struct xstate_hdr *)(get_pcb_user_save_td(td) + 1);
830
831 hdr->xstate_bv = bv;
832 bcopy(xfpustate + sizeof(struct xstate_hdr),
833 (char *)(hdr + 1), len - sizeof(struct xstate_hdr));
834
835 return (0);
836 }
837
838 /*
839 * Set the state of the FPU.
840 */
841 int
842 fpusetregs(struct thread *td, struct savefpu *addr, char *xfpustate,
843 size_t xfpustate_size)
844 {
845 struct pcb *pcb;
846 int error;
847
848 addr->sv_env.en_mxcsr &= cpu_mxcsr_mask;
849 pcb = td->td_pcb;
850 error = 0;
851 critical_enter();
852 if (td == PCPU_GET(fpcurthread) && PCB_USER_FPU(pcb)) {
853 error = fpusetxstate(td, xfpustate, xfpustate_size);
854 if (error == 0) {
855 bcopy(addr, get_pcb_user_save_td(td), sizeof(*addr));
856 fpurestore(get_pcb_user_save_td(td));
857 set_pcb_flags(pcb, PCB_FPUINITDONE |
858 PCB_USERFPUINITDONE);
859 }
860 } else {
861 error = fpusetxstate(td, xfpustate, xfpustate_size);
862 if (error == 0) {
863 bcopy(addr, get_pcb_user_save_td(td), sizeof(*addr));
864 fpuuserinited(td);
865 }
866 }
867 critical_exit();
868 return (error);
869 }
870
871 /*
872 * On AuthenticAMD processors, the fxrstor instruction does not restore
873 * the x87's stored last instruction pointer, last data pointer, and last
874 * opcode values, except in the rare case in which the exception summary
875 * (ES) bit in the x87 status word is set to 1.
876 *
877 * In order to avoid leaking this information across processes, we clean
878 * these values by performing a dummy load before executing fxrstor().
879 */
880 static void
881 fpu_clean_state(void)
882 {
883 static float dummy_variable = 0.0;
884 u_short status;
885
886 /*
887 * Clear the ES bit in the x87 status word if it is currently
888 * set, in order to avoid causing a fault in the upcoming load.
889 */
890 fnstsw(&status);
891 if (status & 0x80)
892 fnclex();
893
894 /*
895 * Load the dummy variable into the x87 stack. This mangles
896 * the x87 stack, but we don't care since we're about to call
897 * fxrstor() anyway.
898 */
899 __asm __volatile("ffree %%st(7); flds %0" : : "m" (dummy_variable));
900 }
901
902 /*
903 * This really sucks. We want the acpi version only, but it requires
904 * the isa_if.h file in order to get the definitions.
905 */
906 #include "opt_isa.h"
907 #ifdef DEV_ISA
908 #include <isa/isavar.h>
909 /*
910 * This sucks up the legacy ISA support assignments from PNPBIOS/ACPI.
911 */
912 static struct isa_pnp_id fpupnp_ids[] = {
913 { 0x040cd041, "Legacy ISA coprocessor support" }, /* PNP0C04 */
914 { 0 }
915 };
916
917 static int
918 fpupnp_probe(device_t dev)
919 {
920 int result;
921
922 result = ISA_PNP_PROBE(device_get_parent(dev), dev, fpupnp_ids);
923 if (result <= 0)
924 device_quiet(dev);
925 return (result);
926 }
927
928 static int
929 fpupnp_attach(device_t dev)
930 {
931
932 return (0);
933 }
934
935 static device_method_t fpupnp_methods[] = {
936 /* Device interface */
937 DEVMETHOD(device_probe, fpupnp_probe),
938 DEVMETHOD(device_attach, fpupnp_attach),
939 DEVMETHOD(device_detach, bus_generic_detach),
940 DEVMETHOD(device_shutdown, bus_generic_shutdown),
941 DEVMETHOD(device_suspend, bus_generic_suspend),
942 DEVMETHOD(device_resume, bus_generic_resume),
943
944 { 0, 0 }
945 };
946
947 static driver_t fpupnp_driver = {
948 "fpupnp",
949 fpupnp_methods,
950 1, /* no softc */
951 };
952
953 static devclass_t fpupnp_devclass;
954
955 DRIVER_MODULE(fpupnp, acpi, fpupnp_driver, fpupnp_devclass, 0, 0);
956 #endif /* DEV_ISA */
957
958 static MALLOC_DEFINE(M_FPUKERN_CTX, "fpukern_ctx",
959 "Kernel contexts for FPU state");
960
961 #define FPU_KERN_CTX_FPUINITDONE 0x01
962 #define FPU_KERN_CTX_DUMMY 0x02 /* avoided save for the kern thread */
963 #define FPU_KERN_CTX_INUSE 0x04
964
965 struct fpu_kern_ctx {
966 struct savefpu *prev;
967 uint32_t flags;
968 char hwstate1[];
969 };
970
971 struct fpu_kern_ctx *
972 fpu_kern_alloc_ctx(u_int flags)
973 {
974 struct fpu_kern_ctx *res;
975 size_t sz;
976
977 sz = sizeof(struct fpu_kern_ctx) + XSAVE_AREA_ALIGN +
978 cpu_max_ext_state_size;
979 res = malloc(sz, M_FPUKERN_CTX, ((flags & FPU_KERN_NOWAIT) ?
980 M_NOWAIT : M_WAITOK) | M_ZERO);
981 return (res);
982 }
983
984 void
985 fpu_kern_free_ctx(struct fpu_kern_ctx *ctx)
986 {
987
988 KASSERT((ctx->flags & FPU_KERN_CTX_INUSE) == 0, ("free'ing inuse ctx"));
989 /* XXXKIB clear the memory ? */
990 free(ctx, M_FPUKERN_CTX);
991 }
992
993 static struct savefpu *
994 fpu_kern_ctx_savefpu(struct fpu_kern_ctx *ctx)
995 {
996 vm_offset_t p;
997
998 p = (vm_offset_t)&ctx->hwstate1;
999 p = roundup2(p, XSAVE_AREA_ALIGN);
1000 return ((struct savefpu *)p);
1001 }
1002
1003 int
1004 fpu_kern_enter(struct thread *td, struct fpu_kern_ctx *ctx, u_int flags)
1005 {
1006 struct pcb *pcb;
1007
1008 pcb = td->td_pcb;
1009 KASSERT((flags & FPU_KERN_NOCTX) != 0 || ctx != NULL,
1010 ("ctx is required when !FPU_KERN_NOCTX"));
1011 KASSERT(ctx == NULL || (ctx->flags & FPU_KERN_CTX_INUSE) == 0,
1012 ("using inuse ctx"));
1013 KASSERT((pcb->pcb_flags & PCB_FPUNOSAVE) == 0,
1014 ("recursive fpu_kern_enter while in PCB_FPUNOSAVE state"));
1015
1016 if ((flags & FPU_KERN_NOCTX) != 0) {
1017 critical_enter();
1018 stop_emulating();
1019 if (curthread == PCPU_GET(fpcurthread)) {
1020 fpusave(curpcb->pcb_save);
1021 PCPU_SET(fpcurthread, NULL);
1022 } else {
1023 KASSERT(PCPU_GET(fpcurthread) == NULL,
1024 ("invalid fpcurthread"));
1025 }
1026
1027 /*
1028 * This breaks XSAVEOPT tracker, but
1029 * PCB_FPUNOSAVE state is supposed to never need to
1030 * save FPU context at all.
1031 */
1032 fpurestore(fpu_initialstate);
1033 set_pcb_flags(pcb, PCB_KERNFPU | PCB_FPUNOSAVE |
1034 PCB_FPUINITDONE);
1035 return (0);
1036 }
1037 if ((flags & FPU_KERN_KTHR) != 0 && is_fpu_kern_thread(0)) {
1038 ctx->flags = FPU_KERN_CTX_DUMMY | FPU_KERN_CTX_INUSE;
1039 return (0);
1040 }
1041 critical_enter();
1042 KASSERT(!PCB_USER_FPU(pcb) || pcb->pcb_save ==
1043 get_pcb_user_save_pcb(pcb), ("mangled pcb_save"));
1044 ctx->flags = FPU_KERN_CTX_INUSE;
1045 if ((pcb->pcb_flags & PCB_FPUINITDONE) != 0)
1046 ctx->flags |= FPU_KERN_CTX_FPUINITDONE;
1047 fpuexit(td);
1048 ctx->prev = pcb->pcb_save;
1049 pcb->pcb_save = fpu_kern_ctx_savefpu(ctx);
1050 set_pcb_flags(pcb, PCB_KERNFPU);
1051 clear_pcb_flags(pcb, PCB_FPUINITDONE);
1052 critical_exit();
1053 return (0);
1054 }
1055
1056 int
1057 fpu_kern_leave(struct thread *td, struct fpu_kern_ctx *ctx)
1058 {
1059 struct pcb *pcb;
1060
1061 pcb = td->td_pcb;
1062
1063 if ((pcb->pcb_flags & PCB_FPUNOSAVE) != 0) {
1064 KASSERT(ctx == NULL, ("non-null ctx after FPU_KERN_NOCTX"));
1065 KASSERT(PCPU_GET(fpcurthread) == NULL,
1066 ("non-NULL fpcurthread for PCB_FPUNOSAVE"));
1067 CRITICAL_ASSERT(td);
1068
1069 clear_pcb_flags(pcb, PCB_FPUNOSAVE | PCB_FPUINITDONE);
1070 start_emulating();
1071 } else {
1072 KASSERT((ctx->flags & FPU_KERN_CTX_INUSE) != 0,
1073 ("leaving not inuse ctx"));
1074 ctx->flags &= ~FPU_KERN_CTX_INUSE;
1075
1076 if (is_fpu_kern_thread(0) &&
1077 (ctx->flags & FPU_KERN_CTX_DUMMY) != 0)
1078 return (0);
1079 KASSERT((ctx->flags & FPU_KERN_CTX_DUMMY) == 0,
1080 ("dummy ctx"));
1081 critical_enter();
1082 if (curthread == PCPU_GET(fpcurthread))
1083 fpudrop();
1084 pcb->pcb_save = ctx->prev;
1085 }
1086
1087 if (pcb->pcb_save == get_pcb_user_save_pcb(pcb)) {
1088 if ((pcb->pcb_flags & PCB_USERFPUINITDONE) != 0) {
1089 set_pcb_flags(pcb, PCB_FPUINITDONE);
1090 clear_pcb_flags(pcb, PCB_KERNFPU);
1091 } else
1092 clear_pcb_flags(pcb, PCB_FPUINITDONE | PCB_KERNFPU);
1093 } else {
1094 if ((ctx->flags & FPU_KERN_CTX_FPUINITDONE) != 0)
1095 set_pcb_flags(pcb, PCB_FPUINITDONE);
1096 else
1097 clear_pcb_flags(pcb, PCB_FPUINITDONE);
1098 KASSERT(!PCB_USER_FPU(pcb), ("unpaired fpu_kern_leave"));
1099 }
1100 critical_exit();
1101 return (0);
1102 }
1103
1104 int
1105 fpu_kern_thread(u_int flags)
1106 {
1107
1108 KASSERT((curthread->td_pflags & TDP_KTHREAD) != 0,
1109 ("Only kthread may use fpu_kern_thread"));
1110 KASSERT(curpcb->pcb_save == get_pcb_user_save_pcb(curpcb),
1111 ("mangled pcb_save"));
1112 KASSERT(PCB_USER_FPU(curpcb), ("recursive call"));
1113
1114 set_pcb_flags(curpcb, PCB_KERNFPU);
1115 return (0);
1116 }
1117
1118 int
1119 is_fpu_kern_thread(u_int flags)
1120 {
1121
1122 if ((curthread->td_pflags & TDP_KTHREAD) == 0)
1123 return (0);
1124 return ((curpcb->pcb_flags & PCB_KERNFPU) != 0);
1125 }
1126
1127 /*
1128 * FPU save area alloc/free/init utility routines
1129 */
1130 struct savefpu *
1131 fpu_save_area_alloc(void)
1132 {
1133
1134 return (uma_zalloc(fpu_save_area_zone, 0));
1135 }
1136
1137 void
1138 fpu_save_area_free(struct savefpu *fsa)
1139 {
1140
1141 uma_zfree(fpu_save_area_zone, fsa);
1142 }
1143
1144 void
1145 fpu_save_area_reset(struct savefpu *fsa)
1146 {
1147
1148 bcopy(fpu_initialstate, fsa, cpu_max_ext_state_size);
1149 }
Cache object: 81d38efbc5937f8fdc2117e425e166d2
|