2 * Copyright (c) 2006 The DragonFly Project. All rights reserved.
3 * Copyright (c) 1990 William Jolitz.
4 * Copyright (c) 1991 The Regents of the University of California.
7 * This code is derived from software contributed to The DragonFly Project
8 * by Matthew Dillon <dillon@backplane.com>
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
14 * 1. Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
16 * 2. Redistributions in binary form must reproduce the above copyright
17 * notice, this list of conditions and the following disclaimer in
18 * the documentation and/or other materials provided with the
20 * 3. Neither the name of The DragonFly Project nor the names of its
21 * contributors may be used to endorse or promote products derived
22 * from this software without specific, prior written permission.
24 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
25 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
26 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
27 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
28 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
29 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
30 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
31 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
32 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
33 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
34 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
37 * from: @(#)npx.c 7.2 (Berkeley) 5/12/91
38 * $FreeBSD: src/sys/i386/isa/npx.c,v 1.80.2.3 2001/10/20 19:04:38 tegge Exp $
39 * $DragonFly: src/sys/platform/vkernel/i386/npx.c,v 1.8 2008/01/29 19:54:56 dillon Exp $
42 #include "opt_debug_npx.h"
44 #include <sys/param.h>
45 #include <sys/systm.h>
47 #include <sys/kernel.h>
48 #include <sys/malloc.h>
49 #include <sys/module.h>
50 #include <sys/sysctl.h>
54 #include <sys/syslog.h>
56 #include <sys/signalvar.h>
58 #include <sys/thread2.h>
59 #include <sys/mplock2.h>
62 #include <machine/asmacros.h>
64 #include <machine/cputypes.h>
65 #include <machine/frame.h>
66 #include <machine/md_var.h>
67 #include <machine/pcb.h>
68 #include <machine/psl.h>
70 #include <machine/clock.h>
72 #include <machine/specialreg.h>
73 #include <machine/segments.h>
74 #include <machine/globaldata.h>
76 #define fldcw(addr) __asm("fldcw %0" : : "m" (*(addr)))
77 #define fnclex() __asm("fnclex")
78 #define fninit() __asm("fninit")
79 #define fnop() __asm("fnop")
80 #define fnsave(addr) __asm __volatile("fnsave %0" : "=m" (*(addr)))
81 #define fnstcw(addr) __asm __volatile("fnstcw %0" : "=m" (*(addr)))
82 #define fnstsw(addr) __asm __volatile("fnstsw %0" : "=m" (*(addr)))
83 #define frstor(addr) __asm("frstor %0" : : "m" (*(addr)))
84 #ifndef CPU_DISABLE_SSE
85 #define fxrstor(addr) __asm("fxrstor %0" : : "m" (*(addr)))
86 #define fxsave(addr) __asm __volatile("fxsave %0" : "=m" (*(addr)))
89 #ifndef CPU_DISABLE_SSE
90 #define GET_FPU_EXSW_PTR(td) \
92 &(td)->td_savefpu->sv_xmm.sv_ex_sw : \
93 &(td)->td_savefpu->sv_87.sv_ex_sw)
94 #else /* CPU_DISABLE_SSE */
95 #define GET_FPU_EXSW_PTR(td) \
96 (&(td)->td_savefpu->sv_87.sv_ex_sw)
97 #endif /* CPU_DISABLE_SSE */
99 typedef u_char bool_t
;
100 #ifndef CPU_DISABLE_SSE
101 static void fpu_clean_state(void);
106 static struct krate badfprate
= { 1 };
108 /*static int npx_attach (device_t dev);*/
109 static void fpusave (union savefpu
*);
110 static void fpurstor (union savefpu
*);
112 #if (defined(I586_CPU) || defined(I686_CPU)) && !defined(CPU_DISABLE_SSE)
114 SYSCTL_INT(_kern
, OID_AUTO
, mmxopt
, CTLFLAG_RD
, &mmxopt
, 0,
115 "MMX/XMM optimized bcopy/copyin/copyout support");
118 static int hw_instruction_sse
;
119 SYSCTL_INT(_hw
, OID_AUTO
, instruction_sse
, CTLFLAG_RD
,
120 &hw_instruction_sse
, 0, "SIMD/MMX2 instructions available in CPU");
124 * Attach routine - announce which it is, and wire into system
127 npx_attach(device_t dev
)
129 npxinit(__INITIAL_NPXCW__
);
135 init_fpu(int supports_sse
)
137 cpu_fxsr
= hw_instruction_sse
= supports_sse
;
141 * Initialize the floating point unit.
144 npxinit(u_short control
)
146 static union savefpu dummy
__aligned(16);
149 * fninit has the same h/w bugs as fnsave. Use the detoxified
150 * fnsave to throw away any junk in the fpu. npxsave() initializes
151 * the fpu and sets npxthread = NULL as important side effects.
155 /*stop_emulating();*/
157 fpusave(curthread
->td_savefpu
);
158 mdcpu
->gd_npxthread
= NULL
;
159 /*start_emulating();*/
164 * Free coprocessor (if we have it).
169 if (curthread
== mdcpu
->gd_npxthread
)
170 npxsave(curthread
->td_savefpu
);
175 * The following mechanism is used to ensure that the FPE_... value
176 * that is passed as a trapcode to the signal handler of the user
177 * process does not have more than one bit set.
179 * Multiple bits may be set if the user process modifies the control
180 * word while a status word bit is already set. While this is a sign
181 * of bad coding, we have no choise than to narrow them down to one
182 * bit, since we must not send a trapcode that is not exactly one of
185 * The mechanism has a static table with 127 entries. Each combination
186 * of the 7 FPU status word exception bits directly translates to a
187 * position in this table, where a single FPE_... value is stored.
188 * This FPE_... value stored there is considered the "most important"
189 * of the exception bits and will be sent as the signal code. The
190 * precedence of the bits is based upon Intel Document "Numerical
191 * Applications", Chapter "Special Computational Situations".
193 * The macro to choose one of these values does these steps: 1) Throw
194 * away status word bits that cannot be masked. 2) Throw away the bits
195 * currently masked in the control word, assuming the user isn't
196 * interested in them anymore. 3) Reinsert status word bit 7 (stack
197 * fault) if it is set, which cannot be masked but must be presered.
198 * 4) Use the remaining bits to point into the trapcode table.
200 * The 6 maskable bits in order of their preference, as stated in the
201 * above referenced Intel manual:
202 * 1 Invalid operation (FP_X_INV)
205 * 1c Operand of unsupported format
207 * 2 QNaN operand (not an exception, irrelavant here)
208 * 3 Any other invalid-operation not mentioned above or zero divide
209 * (FP_X_INV, FP_X_DZ)
210 * 4 Denormal operand (FP_X_DNML)
211 * 5 Numeric over/underflow (FP_X_OFL, FP_X_UFL)
212 * 6 Inexact result (FP_X_IMP)
214 static char fpetable
[128] = {
216 FPE_FLTINV
, /* 1 - INV */
217 FPE_FLTUND
, /* 2 - DNML */
218 FPE_FLTINV
, /* 3 - INV | DNML */
219 FPE_FLTDIV
, /* 4 - DZ */
220 FPE_FLTINV
, /* 5 - INV | DZ */
221 FPE_FLTDIV
, /* 6 - DNML | DZ */
222 FPE_FLTINV
, /* 7 - INV | DNML | DZ */
223 FPE_FLTOVF
, /* 8 - OFL */
224 FPE_FLTINV
, /* 9 - INV | OFL */
225 FPE_FLTUND
, /* A - DNML | OFL */
226 FPE_FLTINV
, /* B - INV | DNML | OFL */
227 FPE_FLTDIV
, /* C - DZ | OFL */
228 FPE_FLTINV
, /* D - INV | DZ | OFL */
229 FPE_FLTDIV
, /* E - DNML | DZ | OFL */
230 FPE_FLTINV
, /* F - INV | DNML | DZ | OFL */
231 FPE_FLTUND
, /* 10 - UFL */
232 FPE_FLTINV
, /* 11 - INV | UFL */
233 FPE_FLTUND
, /* 12 - DNML | UFL */
234 FPE_FLTINV
, /* 13 - INV | DNML | UFL */
235 FPE_FLTDIV
, /* 14 - DZ | UFL */
236 FPE_FLTINV
, /* 15 - INV | DZ | UFL */
237 FPE_FLTDIV
, /* 16 - DNML | DZ | UFL */
238 FPE_FLTINV
, /* 17 - INV | DNML | DZ | UFL */
239 FPE_FLTOVF
, /* 18 - OFL | UFL */
240 FPE_FLTINV
, /* 19 - INV | OFL | UFL */
241 FPE_FLTUND
, /* 1A - DNML | OFL | UFL */
242 FPE_FLTINV
, /* 1B - INV | DNML | OFL | UFL */
243 FPE_FLTDIV
, /* 1C - DZ | OFL | UFL */
244 FPE_FLTINV
, /* 1D - INV | DZ | OFL | UFL */
245 FPE_FLTDIV
, /* 1E - DNML | DZ | OFL | UFL */
246 FPE_FLTINV
, /* 1F - INV | DNML | DZ | OFL | UFL */
247 FPE_FLTRES
, /* 20 - IMP */
248 FPE_FLTINV
, /* 21 - INV | IMP */
249 FPE_FLTUND
, /* 22 - DNML | IMP */
250 FPE_FLTINV
, /* 23 - INV | DNML | IMP */
251 FPE_FLTDIV
, /* 24 - DZ | IMP */
252 FPE_FLTINV
, /* 25 - INV | DZ | IMP */
253 FPE_FLTDIV
, /* 26 - DNML | DZ | IMP */
254 FPE_FLTINV
, /* 27 - INV | DNML | DZ | IMP */
255 FPE_FLTOVF
, /* 28 - OFL | IMP */
256 FPE_FLTINV
, /* 29 - INV | OFL | IMP */
257 FPE_FLTUND
, /* 2A - DNML | OFL | IMP */
258 FPE_FLTINV
, /* 2B - INV | DNML | OFL | IMP */
259 FPE_FLTDIV
, /* 2C - DZ | OFL | IMP */
260 FPE_FLTINV
, /* 2D - INV | DZ | OFL | IMP */
261 FPE_FLTDIV
, /* 2E - DNML | DZ | OFL | IMP */
262 FPE_FLTINV
, /* 2F - INV | DNML | DZ | OFL | IMP */
263 FPE_FLTUND
, /* 30 - UFL | IMP */
264 FPE_FLTINV
, /* 31 - INV | UFL | IMP */
265 FPE_FLTUND
, /* 32 - DNML | UFL | IMP */
266 FPE_FLTINV
, /* 33 - INV | DNML | UFL | IMP */
267 FPE_FLTDIV
, /* 34 - DZ | UFL | IMP */
268 FPE_FLTINV
, /* 35 - INV | DZ | UFL | IMP */
269 FPE_FLTDIV
, /* 36 - DNML | DZ | UFL | IMP */
270 FPE_FLTINV
, /* 37 - INV | DNML | DZ | UFL | IMP */
271 FPE_FLTOVF
, /* 38 - OFL | UFL | IMP */
272 FPE_FLTINV
, /* 39 - INV | OFL | UFL | IMP */
273 FPE_FLTUND
, /* 3A - DNML | OFL | UFL | IMP */
274 FPE_FLTINV
, /* 3B - INV | DNML | OFL | UFL | IMP */
275 FPE_FLTDIV
, /* 3C - DZ | OFL | UFL | IMP */
276 FPE_FLTINV
, /* 3D - INV | DZ | OFL | UFL | IMP */
277 FPE_FLTDIV
, /* 3E - DNML | DZ | OFL | UFL | IMP */
278 FPE_FLTINV
, /* 3F - INV | DNML | DZ | OFL | UFL | IMP */
279 FPE_FLTSUB
, /* 40 - STK */
280 FPE_FLTSUB
, /* 41 - INV | STK */
281 FPE_FLTUND
, /* 42 - DNML | STK */
282 FPE_FLTSUB
, /* 43 - INV | DNML | STK */
283 FPE_FLTDIV
, /* 44 - DZ | STK */
284 FPE_FLTSUB
, /* 45 - INV | DZ | STK */
285 FPE_FLTDIV
, /* 46 - DNML | DZ | STK */
286 FPE_FLTSUB
, /* 47 - INV | DNML | DZ | STK */
287 FPE_FLTOVF
, /* 48 - OFL | STK */
288 FPE_FLTSUB
, /* 49 - INV | OFL | STK */
289 FPE_FLTUND
, /* 4A - DNML | OFL | STK */
290 FPE_FLTSUB
, /* 4B - INV | DNML | OFL | STK */
291 FPE_FLTDIV
, /* 4C - DZ | OFL | STK */
292 FPE_FLTSUB
, /* 4D - INV | DZ | OFL | STK */
293 FPE_FLTDIV
, /* 4E - DNML | DZ | OFL | STK */
294 FPE_FLTSUB
, /* 4F - INV | DNML | DZ | OFL | STK */
295 FPE_FLTUND
, /* 50 - UFL | STK */
296 FPE_FLTSUB
, /* 51 - INV | UFL | STK */
297 FPE_FLTUND
, /* 52 - DNML | UFL | STK */
298 FPE_FLTSUB
, /* 53 - INV | DNML | UFL | STK */
299 FPE_FLTDIV
, /* 54 - DZ | UFL | STK */
300 FPE_FLTSUB
, /* 55 - INV | DZ | UFL | STK */
301 FPE_FLTDIV
, /* 56 - DNML | DZ | UFL | STK */
302 FPE_FLTSUB
, /* 57 - INV | DNML | DZ | UFL | STK */
303 FPE_FLTOVF
, /* 58 - OFL | UFL | STK */
304 FPE_FLTSUB
, /* 59 - INV | OFL | UFL | STK */
305 FPE_FLTUND
, /* 5A - DNML | OFL | UFL | STK */
306 FPE_FLTSUB
, /* 5B - INV | DNML | OFL | UFL | STK */
307 FPE_FLTDIV
, /* 5C - DZ | OFL | UFL | STK */
308 FPE_FLTSUB
, /* 5D - INV | DZ | OFL | UFL | STK */
309 FPE_FLTDIV
, /* 5E - DNML | DZ | OFL | UFL | STK */
310 FPE_FLTSUB
, /* 5F - INV | DNML | DZ | OFL | UFL | STK */
311 FPE_FLTRES
, /* 60 - IMP | STK */
312 FPE_FLTSUB
, /* 61 - INV | IMP | STK */
313 FPE_FLTUND
, /* 62 - DNML | IMP | STK */
314 FPE_FLTSUB
, /* 63 - INV | DNML | IMP | STK */
315 FPE_FLTDIV
, /* 64 - DZ | IMP | STK */
316 FPE_FLTSUB
, /* 65 - INV | DZ | IMP | STK */
317 FPE_FLTDIV
, /* 66 - DNML | DZ | IMP | STK */
318 FPE_FLTSUB
, /* 67 - INV | DNML | DZ | IMP | STK */
319 FPE_FLTOVF
, /* 68 - OFL | IMP | STK */
320 FPE_FLTSUB
, /* 69 - INV | OFL | IMP | STK */
321 FPE_FLTUND
, /* 6A - DNML | OFL | IMP | STK */
322 FPE_FLTSUB
, /* 6B - INV | DNML | OFL | IMP | STK */
323 FPE_FLTDIV
, /* 6C - DZ | OFL | IMP | STK */
324 FPE_FLTSUB
, /* 6D - INV | DZ | OFL | IMP | STK */
325 FPE_FLTDIV
, /* 6E - DNML | DZ | OFL | IMP | STK */
326 FPE_FLTSUB
, /* 6F - INV | DNML | DZ | OFL | IMP | STK */
327 FPE_FLTUND
, /* 70 - UFL | IMP | STK */
328 FPE_FLTSUB
, /* 71 - INV | UFL | IMP | STK */
329 FPE_FLTUND
, /* 72 - DNML | UFL | IMP | STK */
330 FPE_FLTSUB
, /* 73 - INV | DNML | UFL | IMP | STK */
331 FPE_FLTDIV
, /* 74 - DZ | UFL | IMP | STK */
332 FPE_FLTSUB
, /* 75 - INV | DZ | UFL | IMP | STK */
333 FPE_FLTDIV
, /* 76 - DNML | DZ | UFL | IMP | STK */
334 FPE_FLTSUB
, /* 77 - INV | DNML | DZ | UFL | IMP | STK */
335 FPE_FLTOVF
, /* 78 - OFL | UFL | IMP | STK */
336 FPE_FLTSUB
, /* 79 - INV | OFL | UFL | IMP | STK */
337 FPE_FLTUND
, /* 7A - DNML | OFL | UFL | IMP | STK */
338 FPE_FLTSUB
, /* 7B - INV | DNML | OFL | UFL | IMP | STK */
339 FPE_FLTDIV
, /* 7C - DZ | OFL | UFL | IMP | STK */
340 FPE_FLTSUB
, /* 7D - INV | DZ | OFL | UFL | IMP | STK */
341 FPE_FLTDIV
, /* 7E - DNML | DZ | OFL | UFL | IMP | STK */
342 FPE_FLTSUB
, /* 7F - INV | DNML | DZ | OFL | UFL | IMP | STK */
349 * Preserve the FP status word, clear FP exceptions, then generate a SIGFPE.
351 * Clearing exceptions is necessary mainly to avoid IRQ13 bugs. We now
352 * depend on longjmp() restoring a usable state. Restoring the state
353 * or examining it might fail if we didn't clear exceptions.
355 * The error code chosen will be one of the FPE_... macros. It will be
356 * sent as the second argument to old BSD-style signal handlers and as
357 * "siginfo_t->si_code" (second argument) to SA_SIGINFO signal handlers.
359 * XXX the FP state is not preserved across signal handlers. So signal
360 * handlers cannot afford to do FP unless they preserve the state or
361 * longjmp() out. Both preserving the state and longjmp()ing may be
362 * destroyed by IRQ13 bugs. Clearing FP exceptions is not an acceptable
363 * solution for signals other than SIGFPE.
365 * The MP lock is not held on entry (see i386/i386/exception.s) and
366 * should not be held on exit. Interrupts are enabled. We must enter
367 * a critical section to stabilize the FP system and prevent an interrupt
368 * or preemption from changing the FP state out from under us.
371 npx_intr(void *dummy
)
375 struct intrframe
*frame
;
381 * This exception can only occur with CR0_TS clear, otherwise we
382 * would get a DNA exception. However, since interrupts were
383 * enabled a preemption could have sneaked in and used the FP system
384 * before we entered our critical section. If that occured, the
385 * TS bit will be set and npxthread will be NULL.
387 panic("npx_intr: not coded");
388 /* XXX FP STATE FLAG MUST BE PART OF CONTEXT SUPPLIED BY REAL KERNEL */
390 if (rcr0() & CR0_TS
) {
391 KASSERT(mdcpu
->gd_npxthread
== NULL
, ("gd_npxthread was %p with TS set!", mdcpu
->gd_npxthread
));
397 if (mdcpu
->gd_npxthread
== NULL
) {
399 kprintf("npxintr: npxthread = %p, curthread = %p\n",
400 mdcpu
->gd_npxthread
, curthread
);
401 panic("npxintr from nowhere");
403 if (mdcpu
->gd_npxthread
!= curthread
) {
405 kprintf("npxintr: npxthread = %p, curthread = %p\n",
406 mdcpu
->gd_npxthread
, curthread
);
407 panic("npxintr from non-current process");
410 exstat
= GET_FPU_EXSW_PTR(curthread
);
419 * Pass exception to process.
421 frame
= (struct intrframe
*)&dummy
; /* XXX */
422 if ((ISPL(frame
->if_cs
) == SEL_UPL
) /*||(frame->if_eflags&PSL_VM)*/) {
424 * Interrupt is essentially a trap, so we can afford to call
425 * the SIGFPE handler (if any) as soon as the interrupt
428 * XXX little or nothing is gained from this, and plenty is
429 * lost - the interrupt frame has to contain the trap frame
430 * (this is otherwise only necessary for the rescheduling trap
431 * in doreti, and the frame for that could easily be set up
432 * just before it is used).
434 curthread
->td_lwp
->lwp_md
.md_regs
= INTR_TO_TRAPFRAME(frame
);
436 * Encode the appropriate code for detailed information on
440 fpetable
[(*exstat
& ~control
& 0x3f) | (*exstat
& 0x40)];
441 trapsignal(curthread
->td_lwp
, SIGFPE
, code
);
444 * Nested interrupt. These losers occur when:
445 * o an IRQ13 is bogusly generated at a bogus time, e.g.:
446 * o immediately after an fnsave or frstor of an
448 * o a couple of 386 instructions after
449 * "fstpl _memvar" causes a stack overflow.
450 * These are especially nasty when combined with a
452 * o an IRQ13 occurs at the same time as another higher-
453 * priority interrupt.
455 * Treat them like a true async interrupt.
457 lwpsignal(curproc
, curthread
->td_lwp
, SIGFPE
);
466 * Implement the device not available (DNA) exception. gd_npxthread had
467 * better be NULL. Restore the current thread's FP state and set gd_npxthread
470 * Interrupts are enabled and preemption can occur. Enter a critical
471 * section to stabilize the FP state.
474 npxdna(struct trapframe
*frame
)
476 thread_t td
= curthread
;
480 if (mdcpu
->gd_npxthread
!= NULL
) {
481 kprintf("npxdna: npxthread = %p, curthread = %p\n",
482 mdcpu
->gd_npxthread
, td
);
487 * Setup the initial saved state if the thread has never before
488 * used the FP unit. This also occurs when a thread pushes a
489 * signal handler and uses FP in the handler.
491 if ((curthread
->td_flags
& TDF_USINGFP
) == 0) {
492 curthread
->td_flags
|= TDF_USINGFP
;
493 npxinit(__INITIAL_NPXCW__
);
498 * The setting of gd_npxthread and the call to fpurstor() must not
499 * be preempted by an interrupt thread or we will take an npxdna
500 * trap and potentially save our current fpstate (which is garbage)
501 * and then restore the garbage rather then the originally saved
505 /*stop_emulating();*/
507 * Record new context early in case frstor causes an IRQ13.
509 mdcpu
->gd_npxthread
= td
;
510 exstat
= GET_FPU_EXSW_PTR(td
);
513 * The following frstor may cause an IRQ13 when the state being
514 * restored has a pending error. The error will appear to have been
515 * triggered by the current (npx) user instruction even when that
516 * instruction is a no-wait instruction that should not trigger an
517 * error (e.g., fnclex). On at least one 486 system all of the
518 * no-wait instructions are broken the same as frstor, so our
519 * treatment does not amplify the breakage. On at least one
520 * 386/Cyrix 387 system, fnclex works correctly while frstor and
521 * fnsave are broken, so our treatment breaks fnclex if it is the
522 * first FPU instruction after a context switch.
524 if ((td
->td_savefpu
->sv_xmm
.sv_env
.en_mxcsr
& ~0xFFBF) && cpu_fxsr
) {
525 krateprintf(&badfprate
,
526 "FXRSTR: illegal FP MXCSR %08x didinit = %d\n",
527 td
->td_savefpu
->sv_xmm
.sv_env
.en_mxcsr
, didinit
);
528 td
->td_savefpu
->sv_xmm
.sv_env
.en_mxcsr
&= 0xFFBF;
529 lwpsignal(curproc
, curthread
->td_lwp
, SIGFPE
);
531 fpurstor(curthread
->td_savefpu
);
538 * Wrapper for the fnsave instruction to handle h/w bugs. If there is an error
539 * pending, then fnsave generates a bogus IRQ13 on some systems. Force
540 * any IRQ13 to be handled immediately, and then ignore it. This routine is
541 * often called at splhigh so it must not use many system services. In
542 * particular, it's much easier to install a special handler than to
543 * guarantee that it's safe to use npxintr() and its supporting code.
545 * WARNING! This call is made during a switch and the MP lock will be
546 * setup for the new target thread rather then the current thread, so we
547 * cannot do anything here that depends on the *_mplock() functions as
548 * we may trip over their assertions.
550 * WARNING! When using fxsave we MUST fninit after saving the FP state. The
551 * kernel will always assume that the FP state is 'safe' (will not cause
552 * exceptions) for mmx/xmm use if npxthread is NULL. The kernel must still
553 * setup a custom save area before actually using the FP unit, but it will
554 * not bother calling fninit. This greatly improves kernel performance when
555 * it wishes to use the FP unit.
558 npxsave(union savefpu
*addr
)
561 /*stop_emulating();*/
563 mdcpu
->gd_npxthread
= NULL
;
565 /*start_emulating();*/
570 fpusave(union savefpu
*addr
)
579 * Save the FP state to the mcontext structure.
581 * WARNING: If you want to try to npxsave() directly to mctx->mc_fpregs,
582 * then it MUST be 16-byte aligned. Currently this is not guarenteed.
585 npxpush(mcontext_t
*mctx
)
587 thread_t td
= curthread
;
589 if (td
->td_flags
& TDF_USINGFP
) {
590 if (mdcpu
->gd_npxthread
== td
) {
592 * XXX Note: This is a bit inefficient if the signal
593 * handler uses floating point, extra faults will
596 mctx
->mc_ownedfp
= _MC_FPOWNED_FPU
;
597 npxsave(td
->td_savefpu
);
599 mctx
->mc_ownedfp
= _MC_FPOWNED_PCB
;
601 bcopy(td
->td_savefpu
, mctx
->mc_fpregs
, sizeof(mctx
->mc_fpregs
));
602 td
->td_flags
&= ~TDF_USINGFP
;
604 #ifndef CPU_DISABLE_SSE
605 (cpu_fxsr
) ? _MC_FPFMT_XMM
:
609 mctx
->mc_ownedfp
= _MC_FPOWNED_NONE
;
610 mctx
->mc_fpformat
= _MC_FPFMT_NODEV
;
615 * Restore the FP state from the mcontext structure.
618 npxpop(mcontext_t
*mctx
)
620 thread_t td
= curthread
;
622 switch(mctx
->mc_ownedfp
) {
623 case _MC_FPOWNED_NONE
:
625 * If the signal handler used the FP unit but the interrupted
626 * code did not, release the FP unit. Clear TDF_USINGFP will
627 * force the FP unit to reinit so the interrupted code sees
630 if (td
->td_flags
& TDF_USINGFP
) {
631 if (td
== mdcpu
->gd_npxthread
)
632 npxsave(td
->td_savefpu
);
633 td
->td_flags
&= ~TDF_USINGFP
;
636 case _MC_FPOWNED_FPU
:
637 case _MC_FPOWNED_PCB
:
639 * Clear ownership of the FP unit and restore our saved state.
641 * NOTE: The signal handler may have set-up some FP state and
642 * enabled the FP unit, so we have to restore no matter what.
644 * XXX: This is bit inefficient, if the code being returned
645 * to is actively using the FP this results in multiple
648 * WARNING: The saved state was exposed to userland and may
649 * have to be sanitized to avoid a GP fault in the kernel.
651 if (td
== mdcpu
->gd_npxthread
)
652 npxsave(td
->td_savefpu
);
653 bcopy(mctx
->mc_fpregs
, td
->td_savefpu
, sizeof(*td
->td_savefpu
));
654 if ((td
->td_savefpu
->sv_xmm
.sv_env
.en_mxcsr
& ~0xFFBF) &&
656 krateprintf(&badfprate
,
657 "pid %d (%s) signal return from user: "
658 "illegal FP MXCSR %08x\n",
661 td
->td_savefpu
->sv_xmm
.sv_env
.en_mxcsr
);
662 td
->td_savefpu
->sv_xmm
.sv_env
.en_mxcsr
&= 0xFFBF;
664 td
->td_flags
|= TDF_USINGFP
;
670 #ifndef CPU_DISABLE_SSE
672 * On AuthenticAMD processors, the fxrstor instruction does not restore
673 * the x87's stored last instruction pointer, last data pointer, and last
674 * opcode values, except in the rare case in which the exception summary
675 * (ES) bit in the x87 status word is set to 1.
677 * In order to avoid leaking this information across processes, we clean
678 * these values by performing a dummy load before executing fxrstor().
680 static double dummy_variable
= 0.0;
682 fpu_clean_state(void)
687 * Clear the ES bit in the x87 status word if it is currently
688 * set, in order to avoid causing a fault in the upcoming load.
695 * Load the dummy variable into the x87 stack. This mangles
696 * the x87 stack, but we don't care since we're about to call
699 __asm
__volatile("ffree %%st(7); fld %0" : : "m" (dummy_variable
));
701 #endif /* CPU_DISABLE_SSE */
704 fpurstor(union savefpu
*addr
)
706 #ifndef CPU_DISABLE_SSE