The Design and Implementation of the FreeBSD Operating System, Second Edition
Now available: The Design and Implementation of the FreeBSD Operating System (Second Edition)


[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]

FreeBSD/Linux Kernel Cross Reference
sys/amd64/amd64/fpu.c

Version: -  FREEBSD  -  FREEBSD-13-STABLE  -  FREEBSD-13-0  -  FREEBSD-12-STABLE  -  FREEBSD-12-0  -  FREEBSD-11-STABLE  -  FREEBSD-11-0  -  FREEBSD-10-STABLE  -  FREEBSD-10-0  -  FREEBSD-9-STABLE  -  FREEBSD-9-0  -  FREEBSD-8-STABLE  -  FREEBSD-8-0  -  FREEBSD-7-STABLE  -  FREEBSD-7-0  -  FREEBSD-6-STABLE  -  FREEBSD-6-0  -  FREEBSD-5-STABLE  -  FREEBSD-5-0  -  FREEBSD-4-STABLE  -  FREEBSD-3-STABLE  -  FREEBSD22  -  l41  -  OPENBSD  -  linux-2.6  -  MK84  -  PLAN9  -  xnu-8792 
SearchContext: -  none  -  3  -  10 

    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: releng/10.4/sys/amd64/amd64/fpu.c 295148 2016-02-02 14:16:07Z kib $");
   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 use_xsave;                  /* non-static for cpu_switch.S */
  143 uint64_t xsave_mask;            /* the same */
  144 static  uma_zone_t fpu_save_area_zone;
  145 static  struct savefpu *fpu_initialstate;
  146 
  147 struct xsave_area_elm_descr {
  148         u_int   offset;
  149         u_int   size;
  150 } *xsave_area_desc;
  151 
  152 void
  153 fpusave(void *addr)
  154 {
  155 
  156         if (use_xsave)
  157                 xsave((char *)addr, xsave_mask);
  158         else
  159                 fxsave((char *)addr);
  160 }
  161 
  162 void
  163 fpurestore(void *addr)
  164 {
  165 
  166         if (use_xsave)
  167                 xrstor((char *)addr, xsave_mask);
  168         else
  169                 fxrstor((char *)addr);
  170 }
  171 
  172 void
  173 fpususpend(void *addr)
  174 {
  175         u_long cr0;
  176 
  177         cr0 = rcr0();
  178         stop_emulating();
  179         fpusave(addr);
  180         load_cr0(cr0);
  181 }
  182 
  183 void
  184 fpuresume(void *addr)
  185 {
  186         u_long cr0;
  187 
  188         cr0 = rcr0();
  189         stop_emulating();
  190         fninit();
  191         if (use_xsave)
  192                 load_xcr(XCR0, xsave_mask);
  193         fpurestore(addr);
  194         load_cr0(cr0);
  195 }
  196 
  197 /*
  198  * Enable XSAVE if supported and allowed by user.
  199  * Calculate the xsave_mask.
  200  */
  201 static void
  202 fpuinit_bsp1(void)
  203 {
  204         u_int cp[4];
  205         uint64_t xsave_mask_user;
  206 
  207         if ((cpu_feature2 & CPUID2_XSAVE) != 0) {
  208                 use_xsave = 1;
  209                 TUNABLE_INT_FETCH("hw.use_xsave", &use_xsave);
  210         }
  211         if (!use_xsave)
  212                 return;
  213 
  214         cpuid_count(0xd, 0x0, cp);
  215         xsave_mask = XFEATURE_ENABLED_X87 | XFEATURE_ENABLED_SSE;
  216         if ((cp[0] & xsave_mask) != xsave_mask)
  217                 panic("CPU0 does not support X87 or SSE: %x", cp[0]);
  218         xsave_mask = ((uint64_t)cp[3] << 32) | cp[0];
  219         xsave_mask_user = xsave_mask;
  220         TUNABLE_ULONG_FETCH("hw.xsave_mask", &xsave_mask_user);
  221         xsave_mask_user |= XFEATURE_ENABLED_X87 | XFEATURE_ENABLED_SSE;
  222         xsave_mask &= xsave_mask_user;
  223         if ((xsave_mask & XFEATURE_AVX512) != XFEATURE_AVX512)
  224                 xsave_mask &= ~XFEATURE_AVX512;
  225         if ((xsave_mask & XFEATURE_MPX) != XFEATURE_MPX)
  226                 xsave_mask &= ~XFEATURE_MPX;
  227 
  228         cpuid_count(0xd, 0x1, cp);
  229         if ((cp[0] & CPUID_EXTSTATE_XSAVEOPT) != 0) {
  230                 /*
  231                  * Patch the XSAVE instruction in the cpu_switch code
  232                  * to XSAVEOPT.  We assume that XSAVE encoding used
  233                  * REX byte, and set the bit 4 of the r/m byte.
  234                  */
  235                 ctx_switch_xsave[3] |= 0x10;
  236         }
  237 }
  238 
  239 /*
  240  * Calculate the fpu save area size.
  241  */
  242 static void
  243 fpuinit_bsp2(void)
  244 {
  245         u_int cp[4];
  246 
  247         if (use_xsave) {
  248                 cpuid_count(0xd, 0x0, cp);
  249                 cpu_max_ext_state_size = cp[1];
  250 
  251                 /*
  252                  * Reload the cpu_feature2, since we enabled OSXSAVE.
  253                  */
  254                 do_cpuid(1, cp);
  255                 cpu_feature2 = cp[2];
  256         } else
  257                 cpu_max_ext_state_size = sizeof(struct savefpu);
  258 }
  259 
  260 /*
  261  * Initialize the floating point unit.
  262  */
  263 void
  264 fpuinit(void)
  265 {
  266         register_t saveintr;
  267         u_int mxcsr;
  268         u_short control;
  269 
  270         if (IS_BSP())
  271                 fpuinit_bsp1();
  272 
  273         if (use_xsave) {
  274                 load_cr4(rcr4() | CR4_XSAVE);
  275                 load_xcr(XCR0, xsave_mask);
  276         }
  277 
  278         /*
  279          * XCR0 shall be set up before CPU can report the save area size.
  280          */
  281         if (IS_BSP())
  282                 fpuinit_bsp2();
  283 
  284         /*
  285          * It is too early for critical_enter() to work on AP.
  286          */
  287         saveintr = intr_disable();
  288         stop_emulating();
  289         fninit();
  290         control = __INITIAL_FPUCW__;
  291         fldcw(control);
  292         mxcsr = __INITIAL_MXCSR__;
  293         ldmxcsr(mxcsr);
  294         start_emulating();
  295         intr_restore(saveintr);
  296 }
  297 
  298 /*
  299  * On the boot CPU we generate a clean state that is used to
  300  * initialize the floating point unit when it is first used by a
  301  * process.
  302  */
  303 static void
  304 fpuinitstate(void *arg __unused)
  305 {
  306         register_t saveintr;
  307         int cp[4], i, max_ext_n;
  308 
  309         fpu_initialstate = malloc(cpu_max_ext_state_size, M_DEVBUF,
  310             M_WAITOK | M_ZERO);
  311         saveintr = intr_disable();
  312         stop_emulating();
  313 
  314         fpusave(fpu_initialstate);
  315         if (fpu_initialstate->sv_env.en_mxcsr_mask)
  316                 cpu_mxcsr_mask = fpu_initialstate->sv_env.en_mxcsr_mask;
  317         else
  318                 cpu_mxcsr_mask = 0xFFBF;
  319 
  320         /*
  321          * The fninit instruction does not modify XMM registers or x87
  322          * registers (MM/ST).  The fpusave call dumped the garbage
  323          * contained in the registers after reset to the initial state
  324          * saved.  Clear XMM and x87 registers file image to make the
  325          * startup program state and signal handler XMM/x87 register
  326          * content predictable.
  327          */
  328         bzero(fpu_initialstate->sv_fp, sizeof(fpu_initialstate->sv_fp));
  329         bzero(fpu_initialstate->sv_xmm, sizeof(fpu_initialstate->sv_xmm));
  330 
  331         /*
  332          * Create a table describing the layout of the CPU Extended
  333          * Save Area.
  334          */
  335         if (use_xsave) {
  336                 max_ext_n = flsl(xsave_mask);
  337                 xsave_area_desc = malloc(max_ext_n * sizeof(struct
  338                     xsave_area_elm_descr), M_DEVBUF, M_WAITOK | M_ZERO);
  339                 /* x87 state */
  340                 xsave_area_desc[0].offset = 0;
  341                 xsave_area_desc[0].size = 160;
  342                 /* XMM */
  343                 xsave_area_desc[1].offset = 160;
  344                 xsave_area_desc[1].size = 288 - 160;
  345 
  346                 for (i = 2; i < max_ext_n; i++) {
  347                         cpuid_count(0xd, i, cp);
  348                         xsave_area_desc[i].offset = cp[1];
  349                         xsave_area_desc[i].size = cp[0];
  350                 }
  351         }
  352 
  353         fpu_save_area_zone = uma_zcreate("FPU_save_area",
  354             cpu_max_ext_state_size, NULL, NULL, NULL, NULL,
  355             XSAVE_AREA_ALIGN - 1, 0);
  356 
  357         start_emulating();
  358         intr_restore(saveintr);
  359 }
  360 SYSINIT(fpuinitstate, SI_SUB_DRIVERS, SI_ORDER_ANY, fpuinitstate, NULL);
  361 
  362 /*
  363  * Free coprocessor (if we have it).
  364  */
  365 void
  366 fpuexit(struct thread *td)
  367 {
  368 
  369         critical_enter();
  370         if (curthread == PCPU_GET(fpcurthread)) {
  371                 stop_emulating();
  372                 fpusave(curpcb->pcb_save);
  373                 start_emulating();
  374                 PCPU_SET(fpcurthread, NULL);
  375         }
  376         critical_exit();
  377 }
  378 
  379 int
  380 fpuformat(void)
  381 {
  382 
  383         return (_MC_FPFMT_XMM);
  384 }
  385 
  386 /* 
  387  * The following mechanism is used to ensure that the FPE_... value
  388  * that is passed as a trapcode to the signal handler of the user
  389  * process does not have more than one bit set.
  390  * 
  391  * Multiple bits may be set if the user process modifies the control
  392  * word while a status word bit is already set.  While this is a sign
  393  * of bad coding, we have no choise than to narrow them down to one
  394  * bit, since we must not send a trapcode that is not exactly one of
  395  * the FPE_ macros.
  396  *
  397  * The mechanism has a static table with 127 entries.  Each combination
  398  * of the 7 FPU status word exception bits directly translates to a
  399  * position in this table, where a single FPE_... value is stored.
  400  * This FPE_... value stored there is considered the "most important"
  401  * of the exception bits and will be sent as the signal code.  The
  402  * precedence of the bits is based upon Intel Document "Numerical
  403  * Applications", Chapter "Special Computational Situations".
  404  *
  405  * The macro to choose one of these values does these steps: 1) Throw
  406  * away status word bits that cannot be masked.  2) Throw away the bits
  407  * currently masked in the control word, assuming the user isn't
  408  * interested in them anymore.  3) Reinsert status word bit 7 (stack
  409  * fault) if it is set, which cannot be masked but must be presered.
  410  * 4) Use the remaining bits to point into the trapcode table.
  411  *
  412  * The 6 maskable bits in order of their preference, as stated in the
  413  * above referenced Intel manual:
  414  * 1  Invalid operation (FP_X_INV)
  415  * 1a   Stack underflow
  416  * 1b   Stack overflow
  417  * 1c   Operand of unsupported format
  418  * 1d   SNaN operand.
  419  * 2  QNaN operand (not an exception, irrelavant here)
  420  * 3  Any other invalid-operation not mentioned above or zero divide
  421  *      (FP_X_INV, FP_X_DZ)
  422  * 4  Denormal operand (FP_X_DNML)
  423  * 5  Numeric over/underflow (FP_X_OFL, FP_X_UFL)
  424  * 6  Inexact result (FP_X_IMP) 
  425  */
  426 static char fpetable[128] = {
  427         0,
  428         FPE_FLTINV,     /*  1 - INV */
  429         FPE_FLTUND,     /*  2 - DNML */
  430         FPE_FLTINV,     /*  3 - INV | DNML */
  431         FPE_FLTDIV,     /*  4 - DZ */
  432         FPE_FLTINV,     /*  5 - INV | DZ */
  433         FPE_FLTDIV,     /*  6 - DNML | DZ */
  434         FPE_FLTINV,     /*  7 - INV | DNML | DZ */
  435         FPE_FLTOVF,     /*  8 - OFL */
  436         FPE_FLTINV,     /*  9 - INV | OFL */
  437         FPE_FLTUND,     /*  A - DNML | OFL */
  438         FPE_FLTINV,     /*  B - INV | DNML | OFL */
  439         FPE_FLTDIV,     /*  C - DZ | OFL */
  440         FPE_FLTINV,     /*  D - INV | DZ | OFL */
  441         FPE_FLTDIV,     /*  E - DNML | DZ | OFL */
  442         FPE_FLTINV,     /*  F - INV | DNML | DZ | OFL */
  443         FPE_FLTUND,     /* 10 - UFL */
  444         FPE_FLTINV,     /* 11 - INV | UFL */
  445         FPE_FLTUND,     /* 12 - DNML | UFL */
  446         FPE_FLTINV,     /* 13 - INV | DNML | UFL */
  447         FPE_FLTDIV,     /* 14 - DZ | UFL */
  448         FPE_FLTINV,     /* 15 - INV | DZ | UFL */
  449         FPE_FLTDIV,     /* 16 - DNML | DZ | UFL */
  450         FPE_FLTINV,     /* 17 - INV | DNML | DZ | UFL */
  451         FPE_FLTOVF,     /* 18 - OFL | UFL */
  452         FPE_FLTINV,     /* 19 - INV | OFL | UFL */
  453         FPE_FLTUND,     /* 1A - DNML | OFL | UFL */
  454         FPE_FLTINV,     /* 1B - INV | DNML | OFL | UFL */
  455         FPE_FLTDIV,     /* 1C - DZ | OFL | UFL */
  456         FPE_FLTINV,     /* 1D - INV | DZ | OFL | UFL */
  457         FPE_FLTDIV,     /* 1E - DNML | DZ | OFL | UFL */
  458         FPE_FLTINV,     /* 1F - INV | DNML | DZ | OFL | UFL */
  459         FPE_FLTRES,     /* 20 - IMP */
  460         FPE_FLTINV,     /* 21 - INV | IMP */
  461         FPE_FLTUND,     /* 22 - DNML | IMP */
  462         FPE_FLTINV,     /* 23 - INV | DNML | IMP */
  463         FPE_FLTDIV,     /* 24 - DZ | IMP */
  464         FPE_FLTINV,     /* 25 - INV | DZ | IMP */
  465         FPE_FLTDIV,     /* 26 - DNML | DZ | IMP */
  466         FPE_FLTINV,     /* 27 - INV | DNML | DZ | IMP */
  467         FPE_FLTOVF,     /* 28 - OFL | IMP */
  468         FPE_FLTINV,     /* 29 - INV | OFL | IMP */
  469         FPE_FLTUND,     /* 2A - DNML | OFL | IMP */
  470         FPE_FLTINV,     /* 2B - INV | DNML | OFL | IMP */
  471         FPE_FLTDIV,     /* 2C - DZ | OFL | IMP */
  472         FPE_FLTINV,     /* 2D - INV | DZ | OFL | IMP */
  473         FPE_FLTDIV,     /* 2E - DNML | DZ | OFL | IMP */
  474         FPE_FLTINV,     /* 2F - INV | DNML | DZ | OFL | IMP */
  475         FPE_FLTUND,     /* 30 - UFL | IMP */
  476         FPE_FLTINV,     /* 31 - INV | UFL | IMP */
  477         FPE_FLTUND,     /* 32 - DNML | UFL | IMP */
  478         FPE_FLTINV,     /* 33 - INV | DNML | UFL | IMP */
  479         FPE_FLTDIV,     /* 34 - DZ | UFL | IMP */
  480         FPE_FLTINV,     /* 35 - INV | DZ | UFL | IMP */
  481         FPE_FLTDIV,     /* 36 - DNML | DZ | UFL | IMP */
  482         FPE_FLTINV,     /* 37 - INV | DNML | DZ | UFL | IMP */
  483         FPE_FLTOVF,     /* 38 - OFL | UFL | IMP */
  484         FPE_FLTINV,     /* 39 - INV | OFL | UFL | IMP */
  485         FPE_FLTUND,     /* 3A - DNML | OFL | UFL | IMP */
  486         FPE_FLTINV,     /* 3B - INV | DNML | OFL | UFL | IMP */
  487         FPE_FLTDIV,     /* 3C - DZ | OFL | UFL | IMP */
  488         FPE_FLTINV,     /* 3D - INV | DZ | OFL | UFL | IMP */
  489         FPE_FLTDIV,     /* 3E - DNML | DZ | OFL | UFL | IMP */
  490         FPE_FLTINV,     /* 3F - INV | DNML | DZ | OFL | UFL | IMP */
  491         FPE_FLTSUB,     /* 40 - STK */
  492         FPE_FLTSUB,     /* 41 - INV | STK */
  493         FPE_FLTUND,     /* 42 - DNML | STK */
  494         FPE_FLTSUB,     /* 43 - INV | DNML | STK */
  495         FPE_FLTDIV,     /* 44 - DZ | STK */
  496         FPE_FLTSUB,     /* 45 - INV | DZ | STK */
  497         FPE_FLTDIV,     /* 46 - DNML | DZ | STK */
  498         FPE_FLTSUB,     /* 47 - INV | DNML | DZ | STK */
  499         FPE_FLTOVF,     /* 48 - OFL | STK */
  500         FPE_FLTSUB,     /* 49 - INV | OFL | STK */
  501         FPE_FLTUND,     /* 4A - DNML | OFL | STK */
  502         FPE_FLTSUB,     /* 4B - INV | DNML | OFL | STK */
  503         FPE_FLTDIV,     /* 4C - DZ | OFL | STK */
  504         FPE_FLTSUB,     /* 4D - INV | DZ | OFL | STK */
  505         FPE_FLTDIV,     /* 4E - DNML | DZ | OFL | STK */
  506         FPE_FLTSUB,     /* 4F - INV | DNML | DZ | OFL | STK */
  507         FPE_FLTUND,     /* 50 - UFL | STK */
  508         FPE_FLTSUB,     /* 51 - INV | UFL | STK */
  509         FPE_FLTUND,     /* 52 - DNML | UFL | STK */
  510         FPE_FLTSUB,     /* 53 - INV | DNML | UFL | STK */
  511         FPE_FLTDIV,     /* 54 - DZ | UFL | STK */
  512         FPE_FLTSUB,     /* 55 - INV | DZ | UFL | STK */
  513         FPE_FLTDIV,     /* 56 - DNML | DZ | UFL | STK */
  514         FPE_FLTSUB,     /* 57 - INV | DNML | DZ | UFL | STK */
  515         FPE_FLTOVF,     /* 58 - OFL | UFL | STK */
  516         FPE_FLTSUB,     /* 59 - INV | OFL | UFL | STK */
  517         FPE_FLTUND,     /* 5A - DNML | OFL | UFL | STK */
  518         FPE_FLTSUB,     /* 5B - INV | DNML | OFL | UFL | STK */
  519         FPE_FLTDIV,     /* 5C - DZ | OFL | UFL | STK */
  520         FPE_FLTSUB,     /* 5D - INV | DZ | OFL | UFL | STK */
  521         FPE_FLTDIV,     /* 5E - DNML | DZ | OFL | UFL | STK */
  522         FPE_FLTSUB,     /* 5F - INV | DNML | DZ | OFL | UFL | STK */
  523         FPE_FLTRES,     /* 60 - IMP | STK */
  524         FPE_FLTSUB,     /* 61 - INV | IMP | STK */
  525         FPE_FLTUND,     /* 62 - DNML | IMP | STK */
  526         FPE_FLTSUB,     /* 63 - INV | DNML | IMP | STK */
  527         FPE_FLTDIV,     /* 64 - DZ | IMP | STK */
  528         FPE_FLTSUB,     /* 65 - INV | DZ | IMP | STK */
  529         FPE_FLTDIV,     /* 66 - DNML | DZ | IMP | STK */
  530         FPE_FLTSUB,     /* 67 - INV | DNML | DZ | IMP | STK */
  531         FPE_FLTOVF,     /* 68 - OFL | IMP | STK */
  532         FPE_FLTSUB,     /* 69 - INV | OFL | IMP | STK */
  533         FPE_FLTUND,     /* 6A - DNML | OFL | IMP | STK */
  534         FPE_FLTSUB,     /* 6B - INV | DNML | OFL | IMP | STK */
  535         FPE_FLTDIV,     /* 6C - DZ | OFL | IMP | STK */
  536         FPE_FLTSUB,     /* 6D - INV | DZ | OFL | IMP | STK */
  537         FPE_FLTDIV,     /* 6E - DNML | DZ | OFL | IMP | STK */
  538         FPE_FLTSUB,     /* 6F - INV | DNML | DZ | OFL | IMP | STK */
  539         FPE_FLTUND,     /* 70 - UFL | IMP | STK */
  540         FPE_FLTSUB,     /* 71 - INV | UFL | IMP | STK */
  541         FPE_FLTUND,     /* 72 - DNML | UFL | IMP | STK */
  542         FPE_FLTSUB,     /* 73 - INV | DNML | UFL | IMP | STK */
  543         FPE_FLTDIV,     /* 74 - DZ | UFL | IMP | STK */
  544         FPE_FLTSUB,     /* 75 - INV | DZ | UFL | IMP | STK */
  545         FPE_FLTDIV,     /* 76 - DNML | DZ | UFL | IMP | STK */
  546         FPE_FLTSUB,     /* 77 - INV | DNML | DZ | UFL | IMP | STK */
  547         FPE_FLTOVF,     /* 78 - OFL | UFL | IMP | STK */
  548         FPE_FLTSUB,     /* 79 - INV | OFL | UFL | IMP | STK */
  549         FPE_FLTUND,     /* 7A - DNML | OFL | UFL | IMP | STK */
  550         FPE_FLTSUB,     /* 7B - INV | DNML | OFL | UFL | IMP | STK */
  551         FPE_FLTDIV,     /* 7C - DZ | OFL | UFL | IMP | STK */
  552         FPE_FLTSUB,     /* 7D - INV | DZ | OFL | UFL | IMP | STK */
  553         FPE_FLTDIV,     /* 7E - DNML | DZ | OFL | UFL | IMP | STK */
  554         FPE_FLTSUB,     /* 7F - INV | DNML | DZ | OFL | UFL | IMP | STK */
  555 };
  556 
  557 /*
  558  * Read the FP status and control words, then generate si_code value
  559  * for SIGFPE.  The error code chosen will be one of the
  560  * FPE_... macros.  It will be sent as the second argument to old
  561  * BSD-style signal handlers and as "siginfo_t->si_code" (second
  562  * argument) to SA_SIGINFO signal handlers.
  563  *
  564  * Some time ago, we cleared the x87 exceptions with FNCLEX there.
  565  * Clearing exceptions was necessary mainly to avoid IRQ13 bugs.  The
  566  * usermode code which understands the FPU hardware enough to enable
  567  * the exceptions, can also handle clearing the exception state in the
  568  * handler.  The only consequence of not clearing the exception is the
  569  * rethrow of the SIGFPE on return from the signal handler and
  570  * reexecution of the corresponding instruction.
  571  *
  572  * For XMM traps, the exceptions were never cleared.
  573  */
  574 int
  575 fputrap_x87(void)
  576 {
  577         struct savefpu *pcb_save;
  578         u_short control, status;
  579 
  580         critical_enter();
  581 
  582         /*
  583          * Interrupt handling (for another interrupt) may have pushed the
  584          * state to memory.  Fetch the relevant parts of the state from
  585          * wherever they are.
  586          */
  587         if (PCPU_GET(fpcurthread) != curthread) {
  588                 pcb_save = curpcb->pcb_save;
  589                 control = pcb_save->sv_env.en_cw;
  590                 status = pcb_save->sv_env.en_sw;
  591         } else {
  592                 fnstcw(&control);
  593                 fnstsw(&status);
  594         }
  595 
  596         critical_exit();
  597         return (fpetable[status & ((~control & 0x3f) | 0x40)]);
  598 }
  599 
  600 int
  601 fputrap_sse(void)
  602 {
  603         u_int mxcsr;
  604 
  605         critical_enter();
  606         if (PCPU_GET(fpcurthread) != curthread)
  607                 mxcsr = curpcb->pcb_save->sv_env.en_mxcsr;
  608         else
  609                 stmxcsr(&mxcsr);
  610         critical_exit();
  611         return (fpetable[(mxcsr & (~mxcsr >> 7)) & 0x3f]);
  612 }
  613 
  614 /*
  615  * Device Not Available (DNA, #NM) exception handler.
  616  *
  617  * It would be better to switch FP context here (if curthread !=
  618  * fpcurthread) and not necessarily for every context switch, but it
  619  * is too hard to access foreign pcb's.
  620  */
  621 void
  622 fpudna(void)
  623 {
  624 
  625         /*
  626          * This handler is entered with interrupts enabled, so context
  627          * switches may occur before critical_enter() is executed.  If
  628          * a context switch occurs, then when we regain control, our
  629          * state will have been completely restored.  The CPU may
  630          * change underneath us, but the only part of our context that
  631          * lives in the CPU is CR0.TS and that will be "restored" by
  632          * setting it on the new CPU.
  633          */
  634         critical_enter();
  635 
  636         if (PCPU_GET(fpcurthread) == curthread) {
  637                 printf("fpudna: fpcurthread == curthread\n");
  638                 stop_emulating();
  639                 critical_exit();
  640                 return;
  641         }
  642         if (PCPU_GET(fpcurthread) != NULL) {
  643                 panic("fpudna: fpcurthread = %p (%d), curthread = %p (%d)\n",
  644                     PCPU_GET(fpcurthread), PCPU_GET(fpcurthread)->td_tid,
  645                     curthread, curthread->td_tid);
  646         }
  647         stop_emulating();
  648         /*
  649          * Record new context early in case frstor causes a trap.
  650          */
  651         PCPU_SET(fpcurthread, curthread);
  652 
  653         fpu_clean_state();
  654 
  655         if ((curpcb->pcb_flags & PCB_FPUINITDONE) == 0) {
  656                 /*
  657                  * This is the first time this thread has used the FPU or
  658                  * the PCB doesn't contain a clean FPU state.  Explicitly
  659                  * load an initial state.
  660                  *
  661                  * We prefer to restore the state from the actual save
  662                  * area in PCB instead of directly loading from
  663                  * fpu_initialstate, to ignite the XSAVEOPT
  664                  * tracking engine.
  665                  */
  666                 bcopy(fpu_initialstate, curpcb->pcb_save,
  667                     cpu_max_ext_state_size);
  668                 fpurestore(curpcb->pcb_save);
  669                 if (curpcb->pcb_initial_fpucw != __INITIAL_FPUCW__)
  670                         fldcw(curpcb->pcb_initial_fpucw);
  671                 if (PCB_USER_FPU(curpcb))
  672                         set_pcb_flags(curpcb,
  673                             PCB_FPUINITDONE | PCB_USERFPUINITDONE);
  674                 else
  675                         set_pcb_flags(curpcb, PCB_FPUINITDONE);
  676         } else
  677                 fpurestore(curpcb->pcb_save);
  678         critical_exit();
  679 }
  680 
  681 void
  682 fpudrop(void)
  683 {
  684         struct thread *td;
  685 
  686         td = PCPU_GET(fpcurthread);
  687         KASSERT(td == curthread, ("fpudrop: fpcurthread != curthread"));
  688         CRITICAL_ASSERT(td);
  689         PCPU_SET(fpcurthread, NULL);
  690         clear_pcb_flags(td->td_pcb, PCB_FPUINITDONE);
  691         start_emulating();
  692 }
  693 
  694 /*
  695  * Get the user state of the FPU into pcb->pcb_user_save without
  696  * dropping ownership (if possible).  It returns the FPU ownership
  697  * status.
  698  */
  699 int
  700 fpugetregs(struct thread *td)
  701 {
  702         struct pcb *pcb;
  703         uint64_t *xstate_bv, bit;
  704         char *sa;
  705         int max_ext_n, i, owned;
  706 
  707         pcb = td->td_pcb;
  708         if ((pcb->pcb_flags & PCB_USERFPUINITDONE) == 0) {
  709                 bcopy(fpu_initialstate, get_pcb_user_save_pcb(pcb),
  710                     cpu_max_ext_state_size);
  711                 get_pcb_user_save_pcb(pcb)->sv_env.en_cw =
  712                     pcb->pcb_initial_fpucw;
  713                 fpuuserinited(td);
  714                 return (_MC_FPOWNED_PCB);
  715         }
  716         critical_enter();
  717         if (td == PCPU_GET(fpcurthread) && PCB_USER_FPU(pcb)) {
  718                 fpusave(get_pcb_user_save_pcb(pcb));
  719                 owned = _MC_FPOWNED_FPU;
  720         } else {
  721                 owned = _MC_FPOWNED_PCB;
  722         }
  723         critical_exit();
  724         if (use_xsave) {
  725                 /*
  726                  * Handle partially saved state.
  727                  */
  728                 sa = (char *)get_pcb_user_save_pcb(pcb);
  729                 xstate_bv = (uint64_t *)(sa + sizeof(struct savefpu) +
  730                     offsetof(struct xstate_hdr, xstate_bv));
  731                 max_ext_n = flsl(xsave_mask);
  732                 for (i = 0; i < max_ext_n; i++) {
  733                         bit = 1ULL << i;
  734                         if ((xsave_mask & bit) == 0 || (*xstate_bv & bit) != 0)
  735                                 continue;
  736                         bcopy((char *)fpu_initialstate +
  737                             xsave_area_desc[i].offset,
  738                             sa + xsave_area_desc[i].offset,
  739                             xsave_area_desc[i].size);
  740                         *xstate_bv |= bit;
  741                 }
  742         }
  743         return (owned);
  744 }
  745 
  746 void
  747 fpuuserinited(struct thread *td)
  748 {
  749         struct pcb *pcb;
  750 
  751         pcb = td->td_pcb;
  752         if (PCB_USER_FPU(pcb))
  753                 set_pcb_flags(pcb,
  754                     PCB_FPUINITDONE | PCB_USERFPUINITDONE);
  755         else
  756                 set_pcb_flags(pcb, PCB_FPUINITDONE);
  757 }
  758 
  759 int
  760 fpusetxstate(struct thread *td, char *xfpustate, size_t xfpustate_size)
  761 {
  762         struct xstate_hdr *hdr, *ehdr;
  763         size_t len, max_len;
  764         uint64_t bv;
  765 
  766         /* XXXKIB should we clear all extended state in xstate_bv instead ? */
  767         if (xfpustate == NULL)
  768                 return (0);
  769         if (!use_xsave)
  770                 return (EOPNOTSUPP);
  771 
  772         len = xfpustate_size;
  773         if (len < sizeof(struct xstate_hdr))
  774                 return (EINVAL);
  775         max_len = cpu_max_ext_state_size - sizeof(struct savefpu);
  776         if (len > max_len)
  777                 return (EINVAL);
  778 
  779         ehdr = (struct xstate_hdr *)xfpustate;
  780         bv = ehdr->xstate_bv;
  781 
  782         /*
  783          * Avoid #gp.
  784          */
  785         if (bv & ~xsave_mask)
  786                 return (EINVAL);
  787 
  788         hdr = (struct xstate_hdr *)(get_pcb_user_save_td(td) + 1);
  789 
  790         hdr->xstate_bv = bv;
  791         bcopy(xfpustate + sizeof(struct xstate_hdr),
  792             (char *)(hdr + 1), len - sizeof(struct xstate_hdr));
  793 
  794         return (0);
  795 }
  796 
  797 /*
  798  * Set the state of the FPU.
  799  */
  800 int
  801 fpusetregs(struct thread *td, struct savefpu *addr, char *xfpustate,
  802     size_t xfpustate_size)
  803 {
  804         struct pcb *pcb;
  805         int error;
  806 
  807         pcb = td->td_pcb;
  808         critical_enter();
  809         if (td == PCPU_GET(fpcurthread) && PCB_USER_FPU(pcb)) {
  810                 error = fpusetxstate(td, xfpustate, xfpustate_size);
  811                 if (error != 0) {
  812                         critical_exit();
  813                         return (error);
  814                 }
  815                 bcopy(addr, get_pcb_user_save_td(td), sizeof(*addr));
  816                 fpurestore(get_pcb_user_save_td(td));
  817                 critical_exit();
  818                 set_pcb_flags(pcb, PCB_FPUINITDONE | PCB_USERFPUINITDONE);
  819         } else {
  820                 critical_exit();
  821                 error = fpusetxstate(td, xfpustate, xfpustate_size);
  822                 if (error != 0)
  823                         return (error);
  824                 bcopy(addr, get_pcb_user_save_td(td), sizeof(*addr));
  825                 fpuuserinited(td);
  826         }
  827         return (0);
  828 }
  829 
  830 /*
  831  * On AuthenticAMD processors, the fxrstor instruction does not restore
  832  * the x87's stored last instruction pointer, last data pointer, and last
  833  * opcode values, except in the rare case in which the exception summary
  834  * (ES) bit in the x87 status word is set to 1.
  835  *
  836  * In order to avoid leaking this information across processes, we clean
  837  * these values by performing a dummy load before executing fxrstor().
  838  */
  839 static void
  840 fpu_clean_state(void)
  841 {
  842         static float dummy_variable = 0.0;
  843         u_short status;
  844 
  845         /*
  846          * Clear the ES bit in the x87 status word if it is currently
  847          * set, in order to avoid causing a fault in the upcoming load.
  848          */
  849         fnstsw(&status);
  850         if (status & 0x80)
  851                 fnclex();
  852 
  853         /*
  854          * Load the dummy variable into the x87 stack.  This mangles
  855          * the x87 stack, but we don't care since we're about to call
  856          * fxrstor() anyway.
  857          */
  858         __asm __volatile("ffree %%st(7); flds %0" : : "m" (dummy_variable));
  859 }
  860 
  861 /*
  862  * This really sucks.  We want the acpi version only, but it requires
  863  * the isa_if.h file in order to get the definitions.
  864  */
  865 #include "opt_isa.h"
  866 #ifdef DEV_ISA
  867 #include <isa/isavar.h>
  868 /*
  869  * This sucks up the legacy ISA support assignments from PNPBIOS/ACPI.
  870  */
  871 static struct isa_pnp_id fpupnp_ids[] = {
  872         { 0x040cd041, "Legacy ISA coprocessor support" }, /* PNP0C04 */
  873         { 0 }
  874 };
  875 
  876 static int
  877 fpupnp_probe(device_t dev)
  878 {
  879         int result;
  880 
  881         result = ISA_PNP_PROBE(device_get_parent(dev), dev, fpupnp_ids);
  882         if (result <= 0)
  883                 device_quiet(dev);
  884         return (result);
  885 }
  886 
  887 static int
  888 fpupnp_attach(device_t dev)
  889 {
  890 
  891         return (0);
  892 }
  893 
  894 static device_method_t fpupnp_methods[] = {
  895         /* Device interface */
  896         DEVMETHOD(device_probe,         fpupnp_probe),
  897         DEVMETHOD(device_attach,        fpupnp_attach),
  898         DEVMETHOD(device_detach,        bus_generic_detach),
  899         DEVMETHOD(device_shutdown,      bus_generic_shutdown),
  900         DEVMETHOD(device_suspend,       bus_generic_suspend),
  901         DEVMETHOD(device_resume,        bus_generic_resume),
  902         
  903         { 0, 0 }
  904 };
  905 
  906 static driver_t fpupnp_driver = {
  907         "fpupnp",
  908         fpupnp_methods,
  909         1,                      /* no softc */
  910 };
  911 
  912 static devclass_t fpupnp_devclass;
  913 
  914 DRIVER_MODULE(fpupnp, acpi, fpupnp_driver, fpupnp_devclass, 0, 0);
  915 #endif  /* DEV_ISA */
  916 
  917 static MALLOC_DEFINE(M_FPUKERN_CTX, "fpukern_ctx",
  918     "Kernel contexts for FPU state");
  919 
  920 #define FPU_KERN_CTX_FPUINITDONE 0x01
  921 #define FPU_KERN_CTX_DUMMY       0x02   /* avoided save for the kern thread */
  922 
  923 struct fpu_kern_ctx {
  924         struct savefpu *prev;
  925         uint32_t flags;
  926         char hwstate1[];
  927 };
  928 
  929 struct fpu_kern_ctx *
  930 fpu_kern_alloc_ctx(u_int flags)
  931 {
  932         struct fpu_kern_ctx *res;
  933         size_t sz;
  934 
  935         sz = sizeof(struct fpu_kern_ctx) + XSAVE_AREA_ALIGN +
  936             cpu_max_ext_state_size;
  937         res = malloc(sz, M_FPUKERN_CTX, ((flags & FPU_KERN_NOWAIT) ?
  938             M_NOWAIT : M_WAITOK) | M_ZERO);
  939         return (res);
  940 }
  941 
  942 void
  943 fpu_kern_free_ctx(struct fpu_kern_ctx *ctx)
  944 {
  945 
  946         /* XXXKIB clear the memory ? */
  947         free(ctx, M_FPUKERN_CTX);
  948 }
  949 
  950 static struct savefpu *
  951 fpu_kern_ctx_savefpu(struct fpu_kern_ctx *ctx)
  952 {
  953         vm_offset_t p;
  954 
  955         p = (vm_offset_t)&ctx->hwstate1;
  956         p = roundup2(p, XSAVE_AREA_ALIGN);
  957         return ((struct savefpu *)p);
  958 }
  959 
  960 int
  961 fpu_kern_enter(struct thread *td, struct fpu_kern_ctx *ctx, u_int flags)
  962 {
  963         struct pcb *pcb;
  964 
  965         if ((flags & FPU_KERN_KTHR) != 0 && is_fpu_kern_thread(0)) {
  966                 ctx->flags = FPU_KERN_CTX_DUMMY;
  967                 return (0);
  968         }
  969         pcb = td->td_pcb;
  970         KASSERT(!PCB_USER_FPU(pcb) || pcb->pcb_save ==
  971             get_pcb_user_save_pcb(pcb), ("mangled pcb_save"));
  972         ctx->flags = 0;
  973         if ((pcb->pcb_flags & PCB_FPUINITDONE) != 0)
  974                 ctx->flags |= FPU_KERN_CTX_FPUINITDONE;
  975         fpuexit(td);
  976         ctx->prev = pcb->pcb_save;
  977         pcb->pcb_save = fpu_kern_ctx_savefpu(ctx);
  978         set_pcb_flags(pcb, PCB_KERNFPU);
  979         clear_pcb_flags(pcb, PCB_FPUINITDONE);
  980         return (0);
  981 }
  982 
  983 int
  984 fpu_kern_leave(struct thread *td, struct fpu_kern_ctx *ctx)
  985 {
  986         struct pcb *pcb;
  987 
  988         if (is_fpu_kern_thread(0) && (ctx->flags & FPU_KERN_CTX_DUMMY) != 0)
  989                 return (0);
  990         KASSERT((ctx->flags & FPU_KERN_CTX_DUMMY) == 0, ("dummy ctx"));
  991         pcb = td->td_pcb;
  992         critical_enter();
  993         if (curthread == PCPU_GET(fpcurthread))
  994                 fpudrop();
  995         critical_exit();
  996         pcb->pcb_save = ctx->prev;
  997         if (pcb->pcb_save == get_pcb_user_save_pcb(pcb)) {
  998                 if ((pcb->pcb_flags & PCB_USERFPUINITDONE) != 0) {
  999                         set_pcb_flags(pcb, PCB_FPUINITDONE);
 1000                         clear_pcb_flags(pcb, PCB_KERNFPU);
 1001                 } else
 1002                         clear_pcb_flags(pcb, PCB_FPUINITDONE | PCB_KERNFPU);
 1003         } else {
 1004                 if ((ctx->flags & FPU_KERN_CTX_FPUINITDONE) != 0)
 1005                         set_pcb_flags(pcb, PCB_FPUINITDONE);
 1006                 else
 1007                         clear_pcb_flags(pcb, PCB_FPUINITDONE);
 1008                 KASSERT(!PCB_USER_FPU(pcb), ("unpaired fpu_kern_leave"));
 1009         }
 1010         return (0);
 1011 }
 1012 
 1013 int
 1014 fpu_kern_thread(u_int flags)
 1015 {
 1016 
 1017         KASSERT((curthread->td_pflags & TDP_KTHREAD) != 0,
 1018             ("Only kthread may use fpu_kern_thread"));
 1019         KASSERT(curpcb->pcb_save == get_pcb_user_save_pcb(curpcb),
 1020             ("mangled pcb_save"));
 1021         KASSERT(PCB_USER_FPU(curpcb), ("recursive call"));
 1022 
 1023         set_pcb_flags(curpcb, PCB_KERNFPU);
 1024         return (0);
 1025 }
 1026 
 1027 int
 1028 is_fpu_kern_thread(u_int flags)
 1029 {
 1030 
 1031         if ((curthread->td_pflags & TDP_KTHREAD) == 0)
 1032                 return (0);
 1033         return ((curpcb->pcb_flags & PCB_KERNFPU) != 0);
 1034 }
 1035 
 1036 /*
 1037  * FPU save area alloc/free/init utility routines
 1038  */
 1039 struct savefpu *
 1040 fpu_save_area_alloc(void)
 1041 {
 1042 
 1043         return (uma_zalloc(fpu_save_area_zone, 0));
 1044 }
 1045 
 1046 void
 1047 fpu_save_area_free(struct savefpu *fsa)
 1048 {
 1049 
 1050         uma_zfree(fpu_save_area_zone, fsa);
 1051 }
 1052 
 1053 void
 1054 fpu_save_area_reset(struct savefpu *fsa)
 1055 {
 1056 
 1057         bcopy(fpu_initialstate, fsa, cpu_max_ext_state_size);
 1058 }

Cache object: 9612084209b6e385b941a81ea14437c2


[ source navigation ] [ diff markup ] [ identifier search ] [ freetext search ] [ file search ] [ list types ] [ track identifier ]


This page is part of the FreeBSD/Linux Linux Kernel Cross-Reference, and was automatically generated using a modified version of the LXR engine.