2 * Copyright (c) 1990 William Jolitz.
3 * Copyright (c) 1991 The Regents of the University of California.
4 * Copyright (c) 2006 The DragonFly Project.
5 * Copyright (c) 2006 Matthew Dillon.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in
16 * the documentation and/or other materials provided with the
18 * 3. Neither the name of The DragonFly Project nor the names of its
19 * contributors may be used to endorse or promote products derived
20 * from this software without specific, prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
23 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
24 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
25 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
26 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
27 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
28 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
29 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
30 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
31 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
32 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 * from: @(#)npx.c 7.2 (Berkeley) 5/12/91
36 * $FreeBSD: src/sys/i386/isa/npx.c,v 1.80.2.3 2001/10/20 19:04:38 tegge Exp $
41 #include <sys/param.h>
42 #include <sys/systm.h>
44 #include <sys/kernel.h>
45 #include <sys/malloc.h>
46 #include <sys/module.h>
47 #include <sys/sysctl.h>
50 #include <sys/signalvar.h>
52 #include <sys/thread2.h>
54 #include <machine/cputypes.h>
55 #include <machine/frame.h>
56 #include <machine/md_var.h>
57 #include <machine/pcb.h>
58 #include <machine/psl.h>
59 #include <machine/specialreg.h>
60 #include <machine/segments.h>
61 #include <machine/globaldata.h>
63 #define fldcw(addr) __asm("fldcw %0" : : "m" (*(addr)))
64 #define fnclex() __asm("fnclex")
65 #define fninit() __asm("fninit")
66 #define fnop() __asm("fnop")
67 #define fnsave(addr) __asm __volatile("fnsave %0" : "=m" (*(addr)))
68 #define fnstcw(addr) __asm __volatile("fnstcw %0" : "=m" (*(addr)))
69 #define fnstsw(addr) __asm __volatile("fnstsw %0" : "=m" (*(addr)))
70 #define frstor(addr) __asm("frstor %0" : : "m" (*(addr)))
71 #define fxrstor(addr) __asm("fxrstor %0" : : "m" (*(addr)))
72 #define fxsave(addr) __asm __volatile("fxsave %0" : "=m" (*(addr)))
73 #ifndef CPU_DISABLE_AVX
75 xrstor(const void *addr, uint64_t mask)
82 __asm __volatile(".byte 0x0f,0xae,0x2f" : : "D" (addr), "a" (low), "d" (high));
85 xsave(void *addr, uint64_t mask)
92 __asm __volatile(".byte 0x0f,0xae,0x27" : : "D" (addr), "a" (low), "d" (high) : "memory");
95 #define start_emulating() __asm("smsw %%ax; orb %0,%%al; lmsw %%ax" \
96 : : "n" (CR0_TS) : "ax")
97 #define stop_emulating() __asm("clts")
99 typedef u_char bool_t;
100 static void fpu_clean_state(void);
101 #define ldmxcsr(csr) __asm __volatile("ldmxcsr %0" : : "m" (csr))
103 static struct krate badfprate = { 1 };
105 static void fpusave (union savefpu *);
106 static void fpurstor (union savefpu *);
108 __read_mostly uint32_t npx_mxcsr_mask = 0xFFBF; /* this is the default */
109 __read_mostly uint64_t npx_xcr0_mask = 0;
112 * Probe the npx_mxcsr_mask as described in the intel document
113 * "Intel processor identification and the CPUID instruction" Section 7
114 * "Denormals are Zero".
115 * Note that for fxsave to work reliably, the os support bit for
116 * FXSAVE/FXRESTORE operations in CR4 has to be set as per
117 * Intel 64 and IA-32 Architectures Developer's Manual: Vol. 1,
120 void npxprobemask(void)
122 /*64-Byte alignment required for xsave*/
123 static union savefpu dummy __aligned(64);
127 load_cr4(rcr4() | CR4_FXSR);
129 npx_mxcsr_mask = ((uint32_t *)&dummy)[7];
135 * Initialize the floating point unit.
139 /*64-Byte alignment required for xsave*/
140 static union savefpu dummy __aligned(64);
141 u_short control = __INITIAL_FPUCW__;
142 u_int mxcsr = __INITIAL_MXCSR__;
145 * fninit has the same h/w bugs as fnsave. Use the detoxified
146 * fnsave to throw away any junk in the fpu. npxsave() initializes
147 * the fpu and sets npxthread = NULL as important side effects.
154 fpusave(curthread->td_savefpu);
155 mdcpu->gd_npxthread = NULL;
161 * Free coprocessor (if we have it).
166 if (curthread == mdcpu->gd_npxthread)
167 npxsave(curthread->td_savefpu);
172 * The following mechanism is used to ensure that the FPE_... value
173 * that is passed as a trapcode to the signal handler of the user
174 * process does not have more than one bit set.
176 * Multiple bits may be set if the user process modifies the control
177 * word while a status word bit is already set. While this is a sign
178 * of bad coding, we have no choise than to narrow them down to one
179 * bit, since we must not send a trapcode that is not exactly one of
182 * The mechanism has a static table with 127 entries. Each combination
183 * of the 7 FPU status word exception bits directly translates to a
184 * position in this table, where a single FPE_... value is stored.
185 * This FPE_... value stored there is considered the "most important"
186 * of the exception bits and will be sent as the signal code. The
187 * precedence of the bits is based upon Intel Document "Numerical
188 * Applications", Chapter "Special Computational Situations".
190 * The macro to choose one of these values does these steps: 1) Throw
191 * away status word bits that cannot be masked. 2) Throw away the bits
192 * currently masked in the control word, assuming the user isn't
193 * interested in them anymore. 3) Reinsert status word bit 7 (stack
194 * fault) if it is set, which cannot be masked but must be presered.
195 * 4) Use the remaining bits to point into the trapcode table.
197 * The 6 maskable bits in order of their preference, as stated in the
198 * above referenced Intel manual:
199 * 1 Invalid operation (FP_X_INV)
202 * 1c Operand of unsupported format
204 * 2 QNaN operand (not an exception, irrelavant here)
205 * 3 Any other invalid-operation not mentioned above or zero divide
206 * (FP_X_INV, FP_X_DZ)
207 * 4 Denormal operand (FP_X_DNML)
208 * 5 Numeric over/underflow (FP_X_OFL, FP_X_UFL)
209 * 6 Inexact result (FP_X_IMP)
211 static char fpetable[128] = {
213 FPE_FLTINV, /* 1 - INV */
214 FPE_FLTUND, /* 2 - DNML */
215 FPE_FLTINV, /* 3 - INV | DNML */
216 FPE_FLTDIV, /* 4 - DZ */
217 FPE_FLTINV, /* 5 - INV | DZ */
218 FPE_FLTDIV, /* 6 - DNML | DZ */
219 FPE_FLTINV, /* 7 - INV | DNML | DZ */
220 FPE_FLTOVF, /* 8 - OFL */
221 FPE_FLTINV, /* 9 - INV | OFL */
222 FPE_FLTUND, /* A - DNML | OFL */
223 FPE_FLTINV, /* B - INV | DNML | OFL */
224 FPE_FLTDIV, /* C - DZ | OFL */
225 FPE_FLTINV, /* D - INV | DZ | OFL */
226 FPE_FLTDIV, /* E - DNML | DZ | OFL */
227 FPE_FLTINV, /* F - INV | DNML | DZ | OFL */
228 FPE_FLTUND, /* 10 - UFL */
229 FPE_FLTINV, /* 11 - INV | UFL */
230 FPE_FLTUND, /* 12 - DNML | UFL */
231 FPE_FLTINV, /* 13 - INV | DNML | UFL */
232 FPE_FLTDIV, /* 14 - DZ | UFL */
233 FPE_FLTINV, /* 15 - INV | DZ | UFL */
234 FPE_FLTDIV, /* 16 - DNML | DZ | UFL */
235 FPE_FLTINV, /* 17 - INV | DNML | DZ | UFL */
236 FPE_FLTOVF, /* 18 - OFL | UFL */
237 FPE_FLTINV, /* 19 - INV | OFL | UFL */
238 FPE_FLTUND, /* 1A - DNML | OFL | UFL */
239 FPE_FLTINV, /* 1B - INV | DNML | OFL | UFL */
240 FPE_FLTDIV, /* 1C - DZ | OFL | UFL */
241 FPE_FLTINV, /* 1D - INV | DZ | OFL | UFL */
242 FPE_FLTDIV, /* 1E - DNML | DZ | OFL | UFL */
243 FPE_FLTINV, /* 1F - INV | DNML | DZ | OFL | UFL */
244 FPE_FLTRES, /* 20 - IMP */
245 FPE_FLTINV, /* 21 - INV | IMP */
246 FPE_FLTUND, /* 22 - DNML | IMP */
247 FPE_FLTINV, /* 23 - INV | DNML | IMP */
248 FPE_FLTDIV, /* 24 - DZ | IMP */
249 FPE_FLTINV, /* 25 - INV | DZ | IMP */
250 FPE_FLTDIV, /* 26 - DNML | DZ | IMP */
251 FPE_FLTINV, /* 27 - INV | DNML | DZ | IMP */
252 FPE_FLTOVF, /* 28 - OFL | IMP */
253 FPE_FLTINV, /* 29 - INV | OFL | IMP */
254 FPE_FLTUND, /* 2A - DNML | OFL | IMP */
255 FPE_FLTINV, /* 2B - INV | DNML | OFL | IMP */
256 FPE_FLTDIV, /* 2C - DZ | OFL | IMP */
257 FPE_FLTINV, /* 2D - INV | DZ | OFL | IMP */
258 FPE_FLTDIV, /* 2E - DNML | DZ | OFL | IMP */
259 FPE_FLTINV, /* 2F - INV | DNML | DZ | OFL | IMP */
260 FPE_FLTUND, /* 30 - UFL | IMP */
261 FPE_FLTINV, /* 31 - INV | UFL | IMP */
262 FPE_FLTUND, /* 32 - DNML | UFL | IMP */
263 FPE_FLTINV, /* 33 - INV | DNML | UFL | IMP */
264 FPE_FLTDIV, /* 34 - DZ | UFL | IMP */
265 FPE_FLTINV, /* 35 - INV | DZ | UFL | IMP */
266 FPE_FLTDIV, /* 36 - DNML | DZ | UFL | IMP */
267 FPE_FLTINV, /* 37 - INV | DNML | DZ | UFL | IMP */
268 FPE_FLTOVF, /* 38 - OFL | UFL | IMP */
269 FPE_FLTINV, /* 39 - INV | OFL | UFL | IMP */
270 FPE_FLTUND, /* 3A - DNML | OFL | UFL | IMP */
271 FPE_FLTINV, /* 3B - INV | DNML | OFL | UFL | IMP */
272 FPE_FLTDIV, /* 3C - DZ | OFL | UFL | IMP */
273 FPE_FLTINV, /* 3D - INV | DZ | OFL | UFL | IMP */
274 FPE_FLTDIV, /* 3E - DNML | DZ | OFL | UFL | IMP */
275 FPE_FLTINV, /* 3F - INV | DNML | DZ | OFL | UFL | IMP */
276 FPE_FLTSUB, /* 40 - STK */
277 FPE_FLTSUB, /* 41 - INV | STK */
278 FPE_FLTUND, /* 42 - DNML | STK */
279 FPE_FLTSUB, /* 43 - INV | DNML | STK */
280 FPE_FLTDIV, /* 44 - DZ | STK */
281 FPE_FLTSUB, /* 45 - INV | DZ | STK */
282 FPE_FLTDIV, /* 46 - DNML | DZ | STK */
283 FPE_FLTSUB, /* 47 - INV | DNML | DZ | STK */
284 FPE_FLTOVF, /* 48 - OFL | STK */
285 FPE_FLTSUB, /* 49 - INV | OFL | STK */
286 FPE_FLTUND, /* 4A - DNML | OFL | STK */
287 FPE_FLTSUB, /* 4B - INV | DNML | OFL | STK */
288 FPE_FLTDIV, /* 4C - DZ | OFL | STK */
289 FPE_FLTSUB, /* 4D - INV | DZ | OFL | STK */
290 FPE_FLTDIV, /* 4E - DNML | DZ | OFL | STK */
291 FPE_FLTSUB, /* 4F - INV | DNML | DZ | OFL | STK */
292 FPE_FLTUND, /* 50 - UFL | STK */
293 FPE_FLTSUB, /* 51 - INV | UFL | STK */
294 FPE_FLTUND, /* 52 - DNML | UFL | STK */
295 FPE_FLTSUB, /* 53 - INV | DNML | UFL | STK */
296 FPE_FLTDIV, /* 54 - DZ | UFL | STK */
297 FPE_FLTSUB, /* 55 - INV | DZ | UFL | STK */
298 FPE_FLTDIV, /* 56 - DNML | DZ | UFL | STK */
299 FPE_FLTSUB, /* 57 - INV | DNML | DZ | UFL | STK */
300 FPE_FLTOVF, /* 58 - OFL | UFL | STK */
301 FPE_FLTSUB, /* 59 - INV | OFL | UFL | STK */
302 FPE_FLTUND, /* 5A - DNML | OFL | UFL | STK */
303 FPE_FLTSUB, /* 5B - INV | DNML | OFL | UFL | STK */
304 FPE_FLTDIV, /* 5C - DZ | OFL | UFL | STK */
305 FPE_FLTSUB, /* 5D - INV | DZ | OFL | UFL | STK */
306 FPE_FLTDIV, /* 5E - DNML | DZ | OFL | UFL | STK */
307 FPE_FLTSUB, /* 5F - INV | DNML | DZ | OFL | UFL | STK */
308 FPE_FLTRES, /* 60 - IMP | STK */
309 FPE_FLTSUB, /* 61 - INV | IMP | STK */
310 FPE_FLTUND, /* 62 - DNML | IMP | STK */
311 FPE_FLTSUB, /* 63 - INV | DNML | IMP | STK */
312 FPE_FLTDIV, /* 64 - DZ | IMP | STK */
313 FPE_FLTSUB, /* 65 - INV | DZ | IMP | STK */
314 FPE_FLTDIV, /* 66 - DNML | DZ | IMP | STK */
315 FPE_FLTSUB, /* 67 - INV | DNML | DZ | IMP | STK */
316 FPE_FLTOVF, /* 68 - OFL | IMP | STK */
317 FPE_FLTSUB, /* 69 - INV | OFL | IMP | STK */
318 FPE_FLTUND, /* 6A - DNML | OFL | IMP | STK */
319 FPE_FLTSUB, /* 6B - INV | DNML | OFL | IMP | STK */
320 FPE_FLTDIV, /* 6C - DZ | OFL | IMP | STK */
321 FPE_FLTSUB, /* 6D - INV | DZ | OFL | IMP | STK */
322 FPE_FLTDIV, /* 6E - DNML | DZ | OFL | IMP | STK */
323 FPE_FLTSUB, /* 6F - INV | DNML | DZ | OFL | IMP | STK */
324 FPE_FLTUND, /* 70 - UFL | IMP | STK */
325 FPE_FLTSUB, /* 71 - INV | UFL | IMP | STK */
326 FPE_FLTUND, /* 72 - DNML | UFL | IMP | STK */
327 FPE_FLTSUB, /* 73 - INV | DNML | UFL | IMP | STK */
328 FPE_FLTDIV, /* 74 - DZ | UFL | IMP | STK */
329 FPE_FLTSUB, /* 75 - INV | DZ | UFL | IMP | STK */
330 FPE_FLTDIV, /* 76 - DNML | DZ | UFL | IMP | STK */
331 FPE_FLTSUB, /* 77 - INV | DNML | DZ | UFL | IMP | STK */
332 FPE_FLTOVF, /* 78 - OFL | UFL | IMP | STK */
333 FPE_FLTSUB, /* 79 - INV | OFL | UFL | IMP | STK */
334 FPE_FLTUND, /* 7A - DNML | OFL | UFL | IMP | STK */
335 FPE_FLTSUB, /* 7B - INV | DNML | OFL | UFL | IMP | STK */
336 FPE_FLTDIV, /* 7C - DZ | OFL | UFL | IMP | STK */
337 FPE_FLTSUB, /* 7D - INV | DZ | OFL | UFL | IMP | STK */
338 FPE_FLTDIV, /* 7E - DNML | DZ | OFL | UFL | IMP | STK */
339 FPE_FLTSUB, /* 7F - INV | DNML | DZ | OFL | UFL | IMP | STK */
345 * Implement the device not available (DNA) exception. gd_npxthread had
346 * better be NULL. Restore the current thread's FP state and set gd_npxthread
349 * Interrupts are enabled and preemption can occur. Enter a critical
350 * section to stabilize the FP state.
355 struct mdglobaldata *md = mdcpu;
359 td = md->mi.gd_curthread;
362 * npxthread is almost always NULL. When it isn't NULL it can
363 * only be exactly equal to 'td'. This case occurs when the switch
364 * code pro-actively restores the FPU state due to the trap() code
365 * being interruptable (e.g. such as by an interrupt thread).
367 if (__predict_false(md->gd_npxthread != NULL)) {
368 if (md->gd_npxthread == td) {
371 kprintf("npxdna: npxthread = %p, curthread = %p\n",
372 md->gd_npxthread, td);
377 * Setup the initial saved state if the thread has never before
378 * used the FP unit. This also occurs when a thread pushes a
379 * signal handler and uses FP in the handler.
382 if ((td->td_flags & (TDF_USINGFP | TDF_KERNELFP)) == 0) {
383 td->td_flags |= TDF_USINGFP;
389 * The setting of gd_npxthread and the call to fpurstor() must not
390 * be preempted by an interrupt thread or we will take an npxdna
391 * trap and potentially save our current fpstate (which is garbage)
392 * and then restore the garbage rather then the originally saved
398 * Record new context early in case frstor causes an IRQ13.
400 md->gd_npxthread = td;
403 * The following frstor may cause an IRQ13 when the state being
404 * restored has a pending error. The error will appear to have been
405 * triggered by the current (npx) user instruction even when that
406 * instruction is a no-wait instruction that should not trigger an
407 * error (e.g., fnclex). On at least one 486 system all of the
408 * no-wait instructions are broken the same as frstor, so our
409 * treatment does not amplify the breakage. On at least one
410 * 386/Cyrix 387 system, fnclex works correctly while frstor and
411 * fnsave are broken, so our treatment breaks fnclex if it is the
412 * first FPU instruction after a context switch.
414 if ((td->td_savefpu->sv_xmm.sv_env.en_mxcsr & ~npx_mxcsr_mask) &&
416 krateprintf(&badfprate,
417 "%s: FXRSTR: illegal FP MXCSR %08x didinit = %d\n",
418 td->td_comm, td->td_savefpu->sv_xmm.sv_env.en_mxcsr,
420 td->td_savefpu->sv_xmm.sv_env.en_mxcsr &= npx_mxcsr_mask;
421 lwpsignal(td->td_proc, td->td_lwp, SIGFPE);
423 fpurstor(td->td_savefpu);
430 * From cpu heavy restore (already in critical section, gd_npxthread is NULL),
431 * and TDF_USINGFP is already set. Actively restore the FPU state to avoid
432 * excessive npxdna traps.
435 npxdna_quick(thread_t newtd)
438 mdcpu->gd_npxthread = newtd;
439 if ((newtd->td_savefpu->sv_xmm.sv_env.en_mxcsr & ~npx_mxcsr_mask) &&
441 krateprintf(&badfprate,
442 "%s: FXRSTR: illegal FP MXCSR %08x\n",
444 newtd->td_savefpu->sv_xmm.sv_env.en_mxcsr);
445 newtd->td_savefpu->sv_xmm.sv_env.en_mxcsr &= npx_mxcsr_mask;
446 lwpsignal(newtd->td_proc, newtd->td_lwp, SIGFPE);
448 fpurstor(newtd->td_savefpu);
452 * Wrapper for the fnsave instruction to handle h/w bugs. If there is an error
453 * pending, then fnsave generates a bogus IRQ13 on some systems. Force
454 * any IRQ13 to be handled immediately, and then ignore it. This routine is
455 * often called at splhigh so it must not use many system services. In
456 * particular, it's much easier to install a special handler than to
457 * guarantee that it's safe to use npxintr() and its supporting code.
459 * WARNING! This call is made during a switch and the MP lock will be
460 * setup for the new target thread rather then the current thread, so we
461 * cannot do anything here that depends on the *_mplock() functions as
462 * we may trip over their assertions.
464 * WARNING! When using fxsave we MUST fninit after saving the FP state. The
465 * kernel will always assume that the FP state is 'safe' (will not cause
466 * exceptions) for mmx/xmm use if npxthread is NULL. The kernel must still
467 * setup a custom save area before actually using the FP unit, but it will
468 * not bother calling fninit. This greatly improves kernel performance when
469 * it wishes to use the FP unit.
472 npxsave(union savefpu *addr)
474 struct mdglobaldata *md;
480 md->gd_npxthread = NULL;
482 fpurstor(&md->gd_zerofpu); /* security wipe */
488 fpusave(union savefpu *addr)
490 #ifndef CPU_DISABLE_AVX
492 xsave(addr, npx_xcr0_mask);
502 * Save the FP state to the mcontext structure.
504 * WARNING: If you want to try to npxsave() directly to mctx->mc_fpregs,
505 * then it MUST be 16-byte aligned. Currently this is not guarenteed.
508 npxpush(mcontext_t *mctx)
510 thread_t td = curthread;
512 KKASSERT((td->td_flags & TDF_KERNELFP) == 0);
514 if (td->td_flags & TDF_USINGFP) {
515 if (mdcpu->gd_npxthread == td) {
517 * XXX Note: This is a bit inefficient if the signal
518 * handler uses floating point, extra faults will
521 mctx->mc_ownedfp = _MC_FPOWNED_FPU;
522 npxsave(td->td_savefpu);
524 mctx->mc_ownedfp = _MC_FPOWNED_PCB;
526 KKASSERT(sizeof(*td->td_savefpu) <= sizeof(mctx->mc_fpregs));
527 bcopy(td->td_savefpu, mctx->mc_fpregs, sizeof(*td->td_savefpu));
528 td->td_flags &= ~TDF_USINGFP;
529 #ifndef CPU_DISABLE_AVX
530 if (npx_xcr0_mask & CPU_XFEATURE_YMM)
531 mctx->mc_fpformat = _MC_FPFMT_YMM;
536 mctx->mc_fpformat = _MC_FPFMT_XMM;
538 mctx->mc_fpformat = _MC_FPFMT_387;
541 mctx->mc_ownedfp = _MC_FPOWNED_NONE;
542 mctx->mc_fpformat = _MC_FPFMT_NODEV;
547 * Restore the FP state from the mcontext structure.
550 npxpop(mcontext_t *mctx)
552 thread_t td = curthread;
554 switch(mctx->mc_ownedfp) {
555 case _MC_FPOWNED_NONE:
557 * If the signal handler used the FP unit but the interrupted
558 * code did not, release the FP unit. Clear TDF_USINGFP will
559 * force the FP unit to reinit so the interrupted code sees
562 if (td->td_flags & TDF_USINGFP) {
563 if (td == mdcpu->gd_npxthread)
564 npxsave(td->td_savefpu);
565 td->td_flags &= ~TDF_USINGFP;
568 case _MC_FPOWNED_FPU:
569 case _MC_FPOWNED_PCB:
571 * Clear ownership of the FP unit and restore our saved state.
573 * NOTE: The signal handler may have set-up some FP state and
574 * enabled the FP unit, so we have to restore no matter what.
576 * XXX: This is bit inefficient, if the code being returned
577 * to is actively using the FP this results in multiple
580 * WARNING: The saved state was exposed to userland and may
581 * have to be sanitized to avoid a GP fault in the kernel.
583 if (td == mdcpu->gd_npxthread)
584 npxsave(td->td_savefpu);
585 KKASSERT(sizeof(*td->td_savefpu) <= sizeof(mctx->mc_fpregs));
586 bcopy(mctx->mc_fpregs, td->td_savefpu, sizeof(*td->td_savefpu));
587 if ((td->td_savefpu->sv_xmm.sv_env.en_mxcsr & ~npx_mxcsr_mask) &&
589 krateprintf(&badfprate,
590 "pid %d (%s) signal return from user: "
591 "illegal FP MXCSR %08x\n",
594 td->td_savefpu->sv_xmm.sv_env.en_mxcsr);
596 td->td_flags |= TDF_USINGFP;
603 * On AuthenticAMD processors, the fxrstor instruction does not restore
604 * the x87's stored last instruction pointer, last data pointer, and last
605 * opcode values, except in the rare case in which the exception summary
606 * (ES) bit in the x87 status word is set to 1.
608 * In order to avoid leaking this information across processes, we clean
609 * these values by performing a dummy load before executing fxrstor().
611 static double dummy_variable = 0.0;
613 fpu_clean_state(void)
618 * Clear the ES bit in the x87 status word if it is currently
619 * set, in order to avoid causing a fault in the upcoming load.
626 * Load the dummy variable into the x87 stack. This mangles
627 * the x87 stack, but we don't care since we're about to call
630 __asm __volatile("ffree %%st(7); flds %0" : : "m" (dummy_variable));
634 fpurstor(union savefpu *addr)
636 #ifndef CPU_DISABLE_AVX
638 xrstor(addr, npx_xcr0_mask);