8956 Implement KPTI
Reviewed by: Jerry Jelinek <jerry.jelinek@joyent.com>
Reviewed by: Robert Mustacchi <rm@joyent.com>

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 /*
  22  * Copyright (c) 2004, 2010, Oracle and/or its affiliates. All rights reserved.
  23  * Copyright 2018 Joyent, Inc.
  24  * Copyright (c) 2016 by Delphix. All rights reserved.
  25  */
  26 
  27 #include <sys/asm_linkage.h>
  28 #include <sys/asm_misc.h>
  29 #include <sys/regset.h>
  30 #include <sys/privregs.h>
  31 #include <sys/psw.h>
  32 #include <sys/machbrand.h>
  33 
  34 #if defined(__lint)
  35 
  36 #include <sys/types.h>
  37 #include <sys/thread.h>
  38 #include <sys/systm.h>
  39 
  40 #else   /* __lint */
  41 
  42 #include <sys/segments.h>
  43 #include <sys/pcb.h>
  44 #include <sys/trap.h>
  45 #include <sys/ftrace.h>
  46 #include <sys/traptrace.h>
  47 #include <sys/clock.h>
  48 #include <sys/model.h>
  49 #include <sys/panic.h>
  50 
  51 #if defined(__xpv)
  52 #include <sys/hypervisor.h>
  53 #endif
  54 
  55 #include "assym.h"
  56 
  57 #endif  /* __lint */
  58 
  59 /*
  60  * We implement five flavours of system call entry points
  61  *
  62  * -    syscall/sysretq         (amd64 generic)
  63  * -    syscall/sysretl         (i386 plus SYSC bit)
  64  * -    sysenter/sysexit        (i386 plus SEP bit)
  65  * -    int/iret                (i386 generic)
  66  * -    lcall/iret              (i386 generic)
  67  *
  68  * The current libc included in Solaris uses int/iret as the base unoptimized
  69  * kernel entry method. Older libc implementations and legacy binaries may use
  70  * the lcall call gate, so it must continue to be supported.
  71  *
  72  * System calls that use an lcall call gate are processed in trap() via a
  73  * segment-not-present trap, i.e. lcalls are extremely slow(!).
  74  *
  75  * The basic pattern used in the 32-bit SYSC handler at this point in time is
  76  * to have the bare minimum of assembler, and get to the C handlers as
  77  * quickly as possible.
  78  *
  79  * The 64-bit handler is much closer to the sparcv9 handler; that's
  80  * because of passing arguments in registers.  The 32-bit world still
  81  * passes arguments on the stack -- that makes that handler substantially
  82  * more complex.
  83  *
  84  * The two handlers share a few code fragments which are broken
  85  * out into preprocessor macros below.
  86  *
  87  * XX64 come back and speed all this up later.  The 32-bit stuff looks
  88  * especially easy to speed up the argument copying part ..
  89  *
  90  *
  91  * Notes about segment register usage (c.f. the 32-bit kernel)
  92  *
  93  * In the 32-bit kernel, segment registers are dutifully saved and
  94  * restored on all mode transitions because the kernel uses them directly.
  95  * When the processor is running in 64-bit mode, segment registers are
  96  * largely ignored.
  97  *
  98  * %cs and %ss
  99  *      controlled by the hardware mechanisms that make mode transitions
 100  *
 101  * The remaining segment registers have to either be pointing at a valid
 102  * descriptor i.e. with the 'present' bit set, or they can NULL descriptors
 103  *
 104  * %ds and %es
 105  *      always ignored
 106  *
 107  * %fs and %gs
 108  *      fsbase and gsbase are used to control the place they really point at.
 109  *      The kernel only depends on %gs, and controls its own gsbase via swapgs
 110  *
 111  * Note that loading segment registers is still costly because the GDT
 112  * lookup still happens (this is because the hardware can't know that we're
 113  * not setting up these segment registers for a 32-bit program).  Thus we
 114  * avoid doing this in the syscall path, and defer them to lwp context switch
 115  * handlers, so the register values remain virtualized to the lwp.
 116  */
 117 
 118 #if defined(SYSCALLTRACE)
 119 #define ORL_SYSCALLTRACE(r32)           \
 120         orl     syscalltrace(%rip), r32
 121 #else
 122 #define ORL_SYSCALLTRACE(r32)
 123 #endif
 124 
 125 /*
 126  * In the 32-bit kernel, we do absolutely nothing before getting into the
 127  * brand callback checks.  In 64-bit land, we do swapgs and then come here.
 128  * We assume that the %rsp- and %r15-stashing fields in the CPU structure
 129  * are still unused.
 130  *
 131  * Check if a brand_mach_ops callback is defined for the specified callback_id
 132  * type.  If so invoke it with the kernel's %gs value loaded and the following
 133  * data on the stack:
 134  *
 135  * stack:  --------------------------------------
 136  *      32 | callback pointer                   |
 137  *    | 24 | user (or interrupt) stack pointer  |
 138  *    | 16 | lwp pointer                        |
 139  *    v  8 | userland return address            |
 140  *       0 | callback wrapper return addr       |
 141  *         --------------------------------------
 142  *
 143  * Since we're pushing the userland return address onto the kernel stack
 144  * we need to get that address without accessing the user's stack (since we
 145  * can't trust that data).  There are different ways to get the userland
 146  * return address depending on how the syscall trap was made:
 147  *
 148  * a) For sys_syscall and sys_syscall32 the return address is in %rcx.
 149  * b) For sys_sysenter the return address is in %rdx.
 150  * c) For sys_int80 and sys_syscall_int (int91), upon entry into the macro,
 151  *    the stack pointer points at the state saved when we took the interrupt:
 152  *       ------------------------
 153  *    |  | user's %ss           |
 154  *    |  | user's %esp          |
 155  *    |  | EFLAGS register      |
 156  *    v  | user's %cs           |
 157  *       | user's %eip          |
 158  *       ------------------------
 159  *
 160  * The 2nd parameter to the BRAND_CALLBACK macro is either the
 161  * BRAND_URET_FROM_REG or BRAND_URET_FROM_INTR_STACK macro.  These macros are
 162  * used to generate the proper code to get the userland return address for
 163  * each syscall entry point.
 164  *
 165  * The interface to the brand callbacks on the 64-bit kernel assumes %r15
 166  * is available as a scratch register within the callback.  If the callback
 167  * returns within the kernel then this macro will restore %r15.  If the
 168  * callback is going to return directly to userland then it should restore
 169  * %r15 before returning to userland.
 170  */
 171 #define BRAND_URET_FROM_REG(rip_reg)                                    \
 172         pushq   rip_reg                 /* push the return address      */
 173 
 174 /*
 175  * The interrupt stack pointer we saved on entry to the BRAND_CALLBACK macro
 176  * is currently pointing at the user return address (%eip).
 177  */
 178 #define BRAND_URET_FROM_INTR_STACK()                                    \
 179         movq    %gs:CPU_RTMP_RSP, %r15  /* grab the intr. stack pointer */ ;\
 180         pushq   (%r15)                  /* push the return address      */
 181 
 182 #define BRAND_CALLBACK(callback_id, push_userland_ret)                      \
 183         movq    %rsp, %gs:CPU_RTMP_RSP  /* save the stack pointer       */ ;\
 184         movq    %r15, %gs:CPU_RTMP_R15  /* save %r15                    */ ;\
 185         movq    %gs:CPU_THREAD, %r15    /* load the thread pointer      */ ;\
 186         movq    T_STACK(%r15), %rsp     /* switch to the kernel stack   */ ;\
 187         subq    $16, %rsp               /* save space for 2 pointers    */ ;\
 188         pushq   %r14                    /* save %r14                    */ ;\
 189         movq    %gs:CPU_RTMP_RSP, %r14                                     ;\
 190         movq    %r14, 8(%rsp)           /* stash the user stack pointer */ ;\
 191         popq    %r14                    /* restore %r14                 */ ;\
 192         movq    T_LWP(%r15), %r15       /* load the lwp pointer         */ ;\
 193         pushq   %r15                    /* push the lwp pointer         */ ;\
 194         movq    LWP_PROCP(%r15), %r15   /* load the proc pointer        */ ;\
 195         movq    P_BRAND(%r15), %r15     /* load the brand pointer       */ ;\
 196         movq    B_MACHOPS(%r15), %r15   /* load the machops pointer     */ ;\
 197         movq    _CONST(_MUL(callback_id, CPTRSIZE))(%r15), %r15            ;\
 198         cmpq    $0, %r15                                                   ;\
 199         je      1f                                                         ;\
 200         movq    %r15, 16(%rsp)          /* save the callback pointer    */ ;\
 201         push_userland_ret               /* push the return address      */ ;\
 202         call    *24(%rsp)               /* call callback                */ ;\
 203 1:      movq    %gs:CPU_RTMP_R15, %r15  /* restore %r15                 */ ;\
 204         movq    %gs:CPU_RTMP_RSP, %rsp  /* restore the stack pointer    */
 205 
 206 #define MSTATE_TRANSITION(from, to)             \
 207         movl    $from, %edi;                    \
 208         movl    $to, %esi;                      \
 209         call    syscall_mstate
 210 
 211 /*
 212  * Check to see if a simple (direct) return is possible i.e.
 213  *
 214  *      if (t->t_post_sys_ast | syscalltrace |
 215  *          lwp->lwp_pcb.pcb_rupdate == 1)
 216  *              do full version ;
 217  *
 218  * Preconditions:
 219  * -    t is curthread
 220  * Postconditions:
 221  * -    condition code NE is set if post-sys is too complex
 222  * -    rtmp is zeroed if it isn't (we rely on this!)
 223  * -    ltmp is smashed
 224  */
 225 #define CHECK_POSTSYS_NE(t, ltmp, rtmp)                 \
 226         movq    T_LWP(t), ltmp;                         \
 227         movzbl  PCB_RUPDATE(ltmp), rtmp;                \
 228         ORL_SYSCALLTRACE(rtmp);                         \
 229         orl     T_POST_SYS_AST(t), rtmp;                \
 230         cmpl    $0, rtmp
 231 
 232 /*
 233  * Fix up the lwp, thread, and eflags for a successful return
 234  *
 235  * Preconditions:
 236  * -    zwreg contains zero
 237  */
 238 #define SIMPLE_SYSCALL_POSTSYS(t, lwp, zwreg)           \
 239         movb    $LWP_USER, LWP_STATE(lwp);              \
 240         movw    zwreg, T_SYSNUM(t);                     \
 241         andb    $_CONST(0xffff - PS_C), REGOFF_RFL(%rsp)
 242 
 243 /*
 244  * ASSERT(lwptoregs(lwp) == rp);
 245  *
 246  * This may seem obvious, but very odd things happen if this
 247  * assertion is false
 248  *
 249  * Preconditions:
 250  *      (%rsp is ready for normal call sequence)
 251  * Postconditions (if assertion is true):
 252  *      %r11 is smashed
 253  *
 254  * ASSERT(rp->r_cs == descnum)
 255  *
 256  * The code selector is written into the regs structure when the
 257  * lwp stack is created.  We use this ASSERT to validate that
 258  * the regs structure really matches how we came in.
 259  *
 260  * Preconditions:
 261  *      (%rsp is ready for normal call sequence)
 262  * Postconditions (if assertion is true):
 263  *      -none-
 264  *
 265  * ASSERT(lwp->lwp_pcb.pcb_rupdate == 0);
 266  *
 267  * If this is false, it meant that we returned to userland without
 268  * updating the segment registers as we were supposed to.
 269  *
 270  * Note that we must ensure no interrupts or other traps intervene
 271  * between entering privileged mode and performing the assertion,
 272  * otherwise we may perform a context switch on the thread, which
 273  * will end up setting pcb_rupdate to 1 again.
 274  */
 275 #if defined(DEBUG)
 276 
 277 #if !defined(__lint)
 278 
 279 __lwptoregs_msg:
 280         .string "syscall_asm_amd64.s:%d lwptoregs(%p) [%p] != rp [%p]"
 281 
 282 __codesel_msg:
 283         .string "syscall_asm_amd64.s:%d rp->r_cs [%ld] != %ld"
 284 
 285 __no_rupdate_msg:
 286         .string "syscall_asm_amd64.s:%d lwp %p, pcb_rupdate != 0"
 287 
 288 #endif  /* !__lint */
 289 
 290 #define ASSERT_LWPTOREGS(lwp, rp)                       \
 291         movq    LWP_REGS(lwp), %r11;                    \
 292         cmpq    rp, %r11;                               \
 293         je      7f;                                     \
 294         leaq    __lwptoregs_msg(%rip), %rdi;            \
 295         movl    $__LINE__, %esi;                        \
 296         movq    lwp, %rdx;                              \
 297         movq    %r11, %rcx;                             \
 298         movq    rp, %r8;                                \
 299         xorl    %eax, %eax;                             \
 300         call    panic;                                  \
 301 7:
 302 
 303 #define ASSERT_NO_RUPDATE_PENDING(lwp)                  \
 304         testb   $0x1, PCB_RUPDATE(lwp);                 \
 305         je      8f;                                     \
 306         movq    lwp, %rdx;                              \
 307         leaq    __no_rupdate_msg(%rip), %rdi;           \
 308         movl    $__LINE__, %esi;                        \
 309         xorl    %eax, %eax;                             \
 310         call    panic;                                  \
 311 8:
 312 
 313 #else
 314 #define ASSERT_LWPTOREGS(lwp, rp)
 315 #define ASSERT_NO_RUPDATE_PENDING(lwp)
 316 #endif
 317 
 318 /*
 319  * Do the traptrace thing and restore any registers we used
 320  * in situ.  Assumes that %rsp is pointing at the base of
 321  * the struct regs, obviously ..
 322  */
 323 #ifdef TRAPTRACE
 324 #define SYSCALL_TRAPTRACE(ttype)                                \
 325         TRACE_PTR(%rdi, %rbx, %ebx, %rcx, ttype);               \
 326         TRACE_REGS(%rdi, %rsp, %rbx, %rcx);                     \
 327         TRACE_STAMP(%rdi);      /* rdtsc clobbers %eax, %edx */ \
 328         movq    REGOFF_RAX(%rsp), %rax;                         \
 329         movq    REGOFF_RBX(%rsp), %rbx;                         \
 330         movq    REGOFF_RCX(%rsp), %rcx;                         \
 331         movq    REGOFF_RDX(%rsp), %rdx;                         \
 332         movl    %eax, TTR_SYSNUM(%rdi);                         \
 333         movq    REGOFF_RDI(%rsp), %rdi
 334 
 335 #define SYSCALL_TRAPTRACE32(ttype)                              \
 336         SYSCALL_TRAPTRACE(ttype);                               \
 337         /* paranoia: clean the top 32-bits of the registers */  \
 338         orl     %eax, %eax;                                     \
 339         orl     %ebx, %ebx;                                     \
 340         orl     %ecx, %ecx;                                     \
 341         orl     %edx, %edx;                                     \
 342         orl     %edi, %edi
 343 #else   /* TRAPTRACE */
 344 #define SYSCALL_TRAPTRACE(ttype)
 345 #define SYSCALL_TRAPTRACE32(ttype)
 346 #endif  /* TRAPTRACE */
 347 
 348 /*
 349  * The 64-bit libc syscall wrapper does this:
 350  *
 351  * fn(<args>)
 352  * {
 353  *      movq    %rcx, %r10      -- because syscall smashes %rcx
 354  *      movl    $CODE, %eax
 355  *      syscall
 356  *      <error processing>
 357  * }
 358  *
 359  * Thus when we come into the kernel:
 360  *
 361  *      %rdi, %rsi, %rdx, %r10, %r8, %r9 contain first six args
 362  *      %rax is the syscall number
 363  *      %r12-%r15 contain caller state
 364  *
 365  * The syscall instruction arranges that:
 366  *
 367  *      %rcx contains the return %rip
 368  *      %r11d contains bottom 32-bits of %rflags
 369  *      %rflags is masked (as determined by the SFMASK msr)
 370  *      %cs is set to UCS_SEL (as determined by the STAR msr)
 371  *      %ss is set to UDS_SEL (as determined by the STAR msr)
 372  *      %rip is set to sys_syscall (as determined by the LSTAR msr)
 373  *
 374  * Or in other words, we have no registers available at all.
 375  * Only swapgs can save us!
 376  *
 377  * Under the hypervisor, the swapgs has happened already.  However, the
 378  * state of the world is very different from that we're familiar with.
 379  *
 380  * In particular, we have a stack structure like that for interrupt
 381  * gates, except that the %cs and %ss registers are modified for reasons
 382  * that are not entirely clear.  Critically, the %rcx/%r11 values do
 383  * *not* reflect the usage of those registers under a 'real' syscall[1];
 384  * the stack, therefore, looks like this:
 385  *
 386  *      0x0(rsp)        potentially junk %rcx
 387  *      0x8(rsp)        potentially junk %r11
 388  *      0x10(rsp)       user %rip
 389  *      0x18(rsp)       modified %cs
 390  *      0x20(rsp)       user %rflags
 391  *      0x28(rsp)       user %rsp
 392  *      0x30(rsp)       modified %ss
 393  *
 394  *
 395  * and before continuing on, we must load the %rip into %rcx and the
 396  * %rflags into %r11.
 397  *
 398  * [1] They used to, and we relied on it, but this was broken in 3.1.1.
 399  * Sigh.
 400  */
 401 #if defined(__xpv)
 402 #define XPV_SYSCALL_PROD                                                \
 403         movq    0x10(%rsp), %rcx;                                       \
 404         movq    0x20(%rsp), %r11;                                       \
 405         movq    0x28(%rsp), %rsp
 406 #else
 407 #define XPV_SYSCALL_PROD /* nothing */
 408 #endif
 409 
 410 #if defined(__lint)
 411 
 412 /*ARGSUSED*/
 413 void
 414 sys_syscall()
 415 {}
 416 
 417 void
 418 _allsyscalls()
 419 {}
 420 
 421 size_t _allsyscalls_size;
 422 
 423 #else   /* __lint */
 424 
 425         ENTRY_NP2(brand_sys_syscall,_allsyscalls)
 426         SWAPGS                          /* kernel gsbase */
 427         XPV_SYSCALL_PROD
 428         BRAND_CALLBACK(BRAND_CB_SYSCALL, BRAND_URET_FROM_REG(%rcx))
 429         jmp     noprod_sys_syscall
 430 
 431         ALTENTRY(sys_syscall)
 432         SWAPGS                          /* kernel gsbase */
 433         XPV_SYSCALL_PROD
 434 
 435 noprod_sys_syscall:
 436         movq    %r15, %gs:CPU_RTMP_R15
 437         movq    %rsp, %gs:CPU_RTMP_RSP
 438 
 439         movq    %gs:CPU_THREAD, %r15
 440         movq    T_STACK(%r15), %rsp     /* switch from user to kernel stack */
 441 
 442         ASSERT_UPCALL_MASK_IS_SET
 443 
 444         movl    $UCS_SEL, REGOFF_CS(%rsp)
 445         movq    %rcx, REGOFF_RIP(%rsp)          /* syscall: %rip -> %rcx */
 446         movq    %r11, REGOFF_RFL(%rsp)          /* syscall: %rfl -> %r11d */
 447         movl    $UDS_SEL, REGOFF_SS(%rsp)
 448 
 449         movl    %eax, %eax                      /* wrapper: sysc# -> %eax */
 450         movq    %rdi, REGOFF_RDI(%rsp)
 451         movq    %rsi, REGOFF_RSI(%rsp)
 452         movq    %rdx, REGOFF_RDX(%rsp)
 453         movq    %r10, REGOFF_RCX(%rsp)          /* wrapper: %rcx -> %r10 */
 454         movq    %r10, %rcx                      /* arg[3] for direct calls */
 455 
 456         movq    %r8, REGOFF_R8(%rsp)
 457         movq    %r9, REGOFF_R9(%rsp)
 458         movq    %rax, REGOFF_RAX(%rsp)
 459         movq    %rbx, REGOFF_RBX(%rsp)
 460 
 461         movq    %rbp, REGOFF_RBP(%rsp)
 462         movq    %r10, REGOFF_R10(%rsp)
 463         movq    %gs:CPU_RTMP_RSP, %r11
 464         movq    %r11, REGOFF_RSP(%rsp)
 465         movq    %r12, REGOFF_R12(%rsp)
 466 
 467         movq    %r13, REGOFF_R13(%rsp)
 468         movq    %r14, REGOFF_R14(%rsp)
 469         movq    %gs:CPU_RTMP_R15, %r10
 470         movq    %r10, REGOFF_R15(%rsp)
 471         movq    $0, REGOFF_SAVFP(%rsp)
 472         movq    $0, REGOFF_SAVPC(%rsp)
 473 
 474         /*
 475          * Copy these registers here in case we end up stopped with
 476          * someone (like, say, /proc) messing with our register state.
 477          * We don't -restore- them unless we have to in update_sregs.
 478          *
 479          * Since userland -can't- change fsbase or gsbase directly,
 480          * and capturing them involves two serializing instructions,
 481          * we don't bother to capture them here.
 482          */
 483         xorl    %ebx, %ebx
 484         movw    %ds, %bx
 485         movq    %rbx, REGOFF_DS(%rsp)
 486         movw    %es, %bx
 487         movq    %rbx, REGOFF_ES(%rsp)
 488         movw    %fs, %bx
 489         movq    %rbx, REGOFF_FS(%rsp)
 490         movw    %gs, %bx
 491         movq    %rbx, REGOFF_GS(%rsp)
 492 
 493         /*
 494          * If we're trying to use TRAPTRACE though, I take that back: we're
 495          * probably debugging some problem in the SWAPGS logic and want to know
 496          * what the incoming gsbase was.
 497          *
 498          * Since we already did SWAPGS, record the KGSBASE.
 499          */
 500 #if defined(DEBUG) && defined(TRAPTRACE) && !defined(__xpv)
 501         movl    $MSR_AMD_KGSBASE, %ecx
 502         rdmsr
 503         movl    %eax, REGOFF_GSBASE(%rsp)
 504         movl    %edx, REGOFF_GSBASE+4(%rsp)
 505 #endif
 506 
 507         /*
 508          * Machine state saved in the regs structure on the stack
 509          * First six args in %rdi, %rsi, %rdx, %rcx, %r8, %r9
 510          * %eax is the syscall number
 511          * %rsp is the thread's stack, %r15 is curthread
 512          * REG_RSP(%rsp) is the user's stack
 513          */
 514 
 515         SYSCALL_TRAPTRACE($TT_SYSC64)
 516 
 517         movq    %rsp, %rbp
 518 
 519         movq    T_LWP(%r15), %r14
 520         ASSERT_NO_RUPDATE_PENDING(%r14)
 521         ENABLE_INTR_FLAGS
 522 
 523         MSTATE_TRANSITION(LMS_USER, LMS_SYSTEM)
 524         movl    REGOFF_RAX(%rsp), %eax  /* (%rax damaged by mstate call) */
 525 
 526         ASSERT_LWPTOREGS(%r14, %rsp)
 527 
 528         movb    $LWP_SYS, LWP_STATE(%r14)
 529         incq    LWP_RU_SYSC(%r14)
 530         movb    $NORMALRETURN, LWP_EOSYS(%r14)
 531 
 532         incq    %gs:CPU_STATS_SYS_SYSCALL
 533 
 534         movw    %ax, T_SYSNUM(%r15)
 535         movzbl  T_PRE_SYS(%r15), %ebx
 536         ORL_SYSCALLTRACE(%ebx)
 537         testl   %ebx, %ebx
 538         jne     _syscall_pre
 539 
 540 _syscall_invoke:
 541         movq    REGOFF_RDI(%rbp), %rdi
 542         movq    REGOFF_RSI(%rbp), %rsi
 543         movq    REGOFF_RDX(%rbp), %rdx
 544         movq    REGOFF_RCX(%rbp), %rcx
 545         movq    REGOFF_R8(%rbp), %r8
 546         movq    REGOFF_R9(%rbp), %r9
 547 
 548         cmpl    $NSYSCALL, %eax
 549         jae     _syscall_ill
 550         shll    $SYSENT_SIZE_SHIFT, %eax
 551         leaq    sysent(%rax), %rbx
 552 
 553         call    *SY_CALLC(%rbx)
 554 
 555         movq    %rax, %r12
 556         movq    %rdx, %r13
 557 
 558         /*
 559          * If the handler returns two ints, then we need to split the
 560          * 64-bit return value into two 32-bit values.
 561          */
 562         testw   $SE_32RVAL2, SY_FLAGS(%rbx)
 563         je      5f
 564         movq    %r12, %r13
 565         shrq    $32, %r13       /* upper 32-bits into %edx */
 566         movl    %r12d, %r12d    /* lower 32-bits into %eax */
 567 5:
 568         /*
 569          * Optimistically assume that there's no post-syscall
 570          * work to do.  (This is to avoid having to call syscall_mstate()
 571          * with interrupts disabled)
 572          */
 573         MSTATE_TRANSITION(LMS_SYSTEM, LMS_USER)
 574 
 575         /*
 576          * We must protect ourselves from being descheduled here;
 577          * If we were, and we ended up on another cpu, or another
 578          * lwp got in ahead of us, it could change the segment
 579          * registers without us noticing before we return to userland.
 580          */
 581         CLI(%r14)
 582         CHECK_POSTSYS_NE(%r15, %r14, %ebx)
 583         jne     _syscall_post
 584 
 585         /*
 586          * We need to protect ourselves against non-canonical return values
 587          * because Intel doesn't check for them on sysret (AMD does).  Canonical
 588          * addresses on current amd64 processors only use 48-bits for VAs; an
 589          * address is canonical if all upper bits (47-63) are identical. If we
 590          * find a non-canonical %rip, we opt to go through the full
 591          * _syscall_post path which takes us into an iretq which is not
 592          * susceptible to the same problems sysret is.
 593          *
 594          * We're checking for a canonical address by first doing an arithmetic
 595          * shift. This will fill in the remaining bits with the value of bit 63.
 596          * If the address were canonical, the register would now have either all
 597          * zeroes or all ones in it. Therefore we add one (inducing overflow)
 598          * and compare against 1. A canonical address will either be zero or one
 599          * at this point, hence the use of ja.
 600          *
 601          * At this point, r12 and r13 have the return value so we can't use
 602          * those registers.
 603          */
 604         movq    REGOFF_RIP(%rsp), %rcx
 605         sarq    $47, %rcx
 606         incq    %rcx
 607         cmpq    $1, %rcx
 608         ja      _syscall_post
 609 
 610 
 611         SIMPLE_SYSCALL_POSTSYS(%r15, %r14, %bx)
 612 
 613         movq    %r12, REGOFF_RAX(%rsp)
 614         movq    %r13, REGOFF_RDX(%rsp)
 615 
 616         /*
 617          * To get back to userland, we need the return %rip in %rcx and
 618          * the return %rfl in %r11d.  The sysretq instruction also arranges
 619          * to fix up %cs and %ss; everything else is our responsibility.
 620          */
 621         movq    REGOFF_RDI(%rsp), %rdi
 622         movq    REGOFF_RSI(%rsp), %rsi
 623         movq    REGOFF_RDX(%rsp), %rdx
 624         /* %rcx used to restore %rip value */
 625 
 626         movq    REGOFF_R8(%rsp), %r8
 627         movq    REGOFF_R9(%rsp), %r9
 628         movq    REGOFF_RAX(%rsp), %rax
 629         movq    REGOFF_RBX(%rsp), %rbx
 630 
 631         movq    REGOFF_RBP(%rsp), %rbp
 632         movq    REGOFF_R10(%rsp), %r10
 633         /* %r11 used to restore %rfl value */
 634         movq    REGOFF_R12(%rsp), %r12
 635 
 636         movq    REGOFF_R13(%rsp), %r13
 637         movq    REGOFF_R14(%rsp), %r14
 638         movq    REGOFF_R15(%rsp), %r15
 639 
 640         movq    REGOFF_RIP(%rsp), %rcx
 641         movl    REGOFF_RFL(%rsp), %r11d
 642 
 643 #if defined(__xpv)
 644         addq    $REGOFF_RIP, %rsp
 645 #else
 646         movq    REGOFF_RSP(%rsp), %rsp
 647 #endif
 648 
 649         /*
 650          * There can be no instructions between the ALTENTRY below and
 651          * SYSRET or we could end up breaking brand support. See label usage
 652          * in sn1_brand_syscall_callback for an example.
 653          */
 654         ASSERT_UPCALL_MASK_IS_SET
 655 #if defined(__xpv)
 656         SYSRETQ
 657         ALTENTRY(nopop_sys_syscall_swapgs_sysretq)
 658 
 659         /*
 660          * We can only get here after executing a brand syscall
 661          * interposition callback handler and simply need to
 662          * "sysretq" back to userland. On the hypervisor this
 663          * involves the iret hypercall which requires us to construct
 664          * just enough of the stack needed for the hypercall.
 665          * (rip, cs, rflags, rsp, ss).
 666          */
 667         movq    %rsp, %gs:CPU_RTMP_RSP          /* save user's rsp */
 668         movq    %gs:CPU_THREAD, %r11
 669         movq    T_STACK(%r11), %rsp
 670 
 671         movq    %rcx, REGOFF_RIP(%rsp)
 672         movl    $UCS_SEL, REGOFF_CS(%rsp)
 673         movq    %gs:CPU_RTMP_RSP, %r11
 674         movq    %r11, REGOFF_RSP(%rsp)
 675         pushfq
 676         popq    %r11                            /* hypercall enables ints */
 677         movq    %r11, REGOFF_RFL(%rsp)
 678         movl    $UDS_SEL, REGOFF_SS(%rsp)
 679         addq    $REGOFF_RIP, %rsp
 680         /*
 681          * XXPV: see comment in SYSRETQ definition for future optimization
 682          *       we could take.
 683          */
 684         ASSERT_UPCALL_MASK_IS_SET
 685         SYSRETQ
 686 #else
 687         ALTENTRY(nopop_sys_syscall_swapgs_sysretq)
 688         jmp     tr_sysretq

 689 #endif
 690         /*NOTREACHED*/
 691         SET_SIZE(nopop_sys_syscall_swapgs_sysretq)
 692 
 693 _syscall_pre:
 694         call    pre_syscall
 695         movl    %eax, %r12d
 696         testl   %eax, %eax
 697         jne     _syscall_post_call
 698         /*
 699          * Didn't abort, so reload the syscall args and invoke the handler.
 700          */
 701         movzwl  T_SYSNUM(%r15), %eax
 702         jmp     _syscall_invoke
 703 
 704 _syscall_ill:
 705         call    nosys
 706         movq    %rax, %r12
 707         movq    %rdx, %r13
 708         jmp     _syscall_post_call
 709 
 710 _syscall_post:
 711         STI
 712         /*
 713          * Sigh, our optimism wasn't justified, put it back to LMS_SYSTEM
 714          * so that we can account for the extra work it takes us to finish.
 715          */
 716         MSTATE_TRANSITION(LMS_USER, LMS_SYSTEM)
 717 _syscall_post_call:
 718         movq    %r12, %rdi
 719         movq    %r13, %rsi
 720         call    post_syscall
 721         MSTATE_TRANSITION(LMS_SYSTEM, LMS_USER)
 722         jmp     _sys_rtt
 723         SET_SIZE(sys_syscall)
 724         SET_SIZE(brand_sys_syscall)
 725 
 726 #endif  /* __lint */
 727 
 728 #if defined(__lint)
 729 
 730 /*ARGSUSED*/
 731 void
 732 sys_syscall32()
 733 {}
 734 
 735 #else   /* __lint */
 736 
 737         ENTRY_NP(brand_sys_syscall32)
 738         SWAPGS                          /* kernel gsbase */
 739         XPV_TRAP_POP
 740         BRAND_CALLBACK(BRAND_CB_SYSCALL32, BRAND_URET_FROM_REG(%rcx))
 741         jmp     nopop_sys_syscall32
 742 
 743         ALTENTRY(sys_syscall32)
 744         SWAPGS                          /* kernel gsbase */
 745         XPV_TRAP_POP
 746 
 747 nopop_sys_syscall32:
 748         movl    %esp, %r10d
 749         movq    %gs:CPU_THREAD, %r15
 750         movq    T_STACK(%r15), %rsp
 751         movl    %eax, %eax
 752 
 753         movl    $U32CS_SEL, REGOFF_CS(%rsp)
 754         movl    %ecx, REGOFF_RIP(%rsp)          /* syscall: %rip -> %rcx */
 755         movq    %r11, REGOFF_RFL(%rsp)          /* syscall: %rfl -> %r11d */
 756         movq    %r10, REGOFF_RSP(%rsp)
 757         movl    $UDS_SEL, REGOFF_SS(%rsp)
 758 
 759 _syscall32_save:
 760         movl    %edi, REGOFF_RDI(%rsp)
 761         movl    %esi, REGOFF_RSI(%rsp)
 762         movl    %ebp, REGOFF_RBP(%rsp)
 763         movl    %ebx, REGOFF_RBX(%rsp)
 764         movl    %edx, REGOFF_RDX(%rsp)
 765         movl    %ecx, REGOFF_RCX(%rsp)
 766         movl    %eax, REGOFF_RAX(%rsp)          /* wrapper: sysc# -> %eax */
 767         movq    $0, REGOFF_SAVFP(%rsp)
 768         movq    $0, REGOFF_SAVPC(%rsp)
 769 
 770         /*
 771          * Copy these registers here in case we end up stopped with
 772          * someone (like, say, /proc) messing with our register state.
 773          * We don't -restore- them unless we have to in update_sregs.
 774          *
 775          * Since userland -can't- change fsbase or gsbase directly,
 776          * we don't bother to capture them here.
 777          */
 778         xorl    %ebx, %ebx
 779         movw    %ds, %bx
 780         movq    %rbx, REGOFF_DS(%rsp)
 781         movw    %es, %bx
 782         movq    %rbx, REGOFF_ES(%rsp)
 783         movw    %fs, %bx
 784         movq    %rbx, REGOFF_FS(%rsp)
 785         movw    %gs, %bx
 786         movq    %rbx, REGOFF_GS(%rsp)
 787 
 788         /*
 789          * If we're trying to use TRAPTRACE though, I take that back: we're
 790          * probably debugging some problem in the SWAPGS logic and want to know
 791          * what the incoming gsbase was.
 792          *
 793          * Since we already did SWAPGS, record the KGSBASE.
 794          */
 795 #if defined(DEBUG) && defined(TRAPTRACE) && !defined(__xpv)
 796         movl    $MSR_AMD_KGSBASE, %ecx
 797         rdmsr
 798         movl    %eax, REGOFF_GSBASE(%rsp)
 799         movl    %edx, REGOFF_GSBASE+4(%rsp)
 800 #endif
 801 
 802         /*
 803          * Application state saved in the regs structure on the stack
 804          * %eax is the syscall number
 805          * %rsp is the thread's stack, %r15 is curthread
 806          * REG_RSP(%rsp) is the user's stack
 807          */
 808 
 809         SYSCALL_TRAPTRACE32($TT_SYSC)
 810 
 811         movq    %rsp, %rbp
 812 
 813         movq    T_LWP(%r15), %r14
 814         ASSERT_NO_RUPDATE_PENDING(%r14)
 815 
 816         ENABLE_INTR_FLAGS
 817 
 818         MSTATE_TRANSITION(LMS_USER, LMS_SYSTEM)
 819         movl    REGOFF_RAX(%rsp), %eax  /* (%rax damaged by mstate call) */
 820 
 821         ASSERT_LWPTOREGS(%r14, %rsp)
 822 
 823         incq     %gs:CPU_STATS_SYS_SYSCALL
 824 
 825         /*
 826          * Make some space for MAXSYSARGS (currently 8) 32-bit args placed
 827          * into 64-bit (long) arg slots, maintaining 16 byte alignment.  Or
 828          * more succinctly:
 829          *
 830          *      SA(MAXSYSARGS * sizeof (long)) == 64
 831          */
 832 #define SYS_DROP        64                      /* drop for args */
 833         subq    $SYS_DROP, %rsp
 834         movb    $LWP_SYS, LWP_STATE(%r14)
 835         movq    %r15, %rdi
 836         movq    %rsp, %rsi
 837         call    syscall_entry
 838 
 839         /*
 840          * Fetch the arguments copied onto the kernel stack and put
 841          * them in the right registers to invoke a C-style syscall handler.
 842          * %rax contains the handler address.
 843          *
 844          * Ideas for making all this go faster of course include simply
 845          * forcibly fetching 6 arguments from the user stack under lofault
 846          * protection, reverting to copyin_args only when watchpoints
 847          * are in effect.
 848          *
 849          * (If we do this, make sure that exec and libthread leave
 850          * enough space at the top of the stack to ensure that we'll
 851          * never do a fetch from an invalid page.)
 852          *
 853          * Lots of ideas here, but they won't really help with bringup B-)
 854          * Correctness can't wait, performance can wait a little longer ..
 855          */
 856 
 857         movq    %rax, %rbx
 858         movl    0(%rsp), %edi
 859         movl    8(%rsp), %esi
 860         movl    0x10(%rsp), %edx
 861         movl    0x18(%rsp), %ecx
 862         movl    0x20(%rsp), %r8d
 863         movl    0x28(%rsp), %r9d
 864 
 865         call    *SY_CALLC(%rbx)
 866 
 867         movq    %rbp, %rsp      /* pop the args */
 868 
 869         /*
 870          * amd64 syscall handlers -always- return a 64-bit value in %rax.
 871          * On the 32-bit kernel, they always return that value in %eax:%edx
 872          * as required by the 32-bit ABI.
 873          *
 874          * Simulate the same behaviour by unconditionally splitting the
 875          * return value in the same way.
 876          */
 877         movq    %rax, %r13
 878         shrq    $32, %r13       /* upper 32-bits into %edx */
 879         movl    %eax, %r12d     /* lower 32-bits into %eax */
 880 
 881         /*
 882          * Optimistically assume that there's no post-syscall
 883          * work to do.  (This is to avoid having to call syscall_mstate()
 884          * with interrupts disabled)
 885          */
 886         MSTATE_TRANSITION(LMS_SYSTEM, LMS_USER)
 887 
 888         /*
 889          * We must protect ourselves from being descheduled here;
 890          * If we were, and we ended up on another cpu, or another
 891          * lwp got in ahead of us, it could change the segment
 892          * registers without us noticing before we return to userland.
 893          */
 894         CLI(%r14)
 895         CHECK_POSTSYS_NE(%r15, %r14, %ebx)
 896         jne     _full_syscall_postsys32
 897         SIMPLE_SYSCALL_POSTSYS(%r15, %r14, %bx)
 898 
 899         /*
 900          * To get back to userland, we need to put the return %rip in %rcx and
 901          * the return %rfl in %r11d.  The sysret instruction also arranges
 902          * to fix up %cs and %ss; everything else is our responsibility.
 903          */
 904 
 905         movl    %r12d, %eax                     /* %eax: rval1 */
 906         movl    REGOFF_RBX(%rsp), %ebx
 907         /* %ecx used for return pointer */
 908         movl    %r13d, %edx                     /* %edx: rval2 */
 909         movl    REGOFF_RBP(%rsp), %ebp
 910         movl    REGOFF_RSI(%rsp), %esi
 911         movl    REGOFF_RDI(%rsp), %edi
 912 
 913         movl    REGOFF_RFL(%rsp), %r11d         /* %r11 -> eflags */
 914         movl    REGOFF_RIP(%rsp), %ecx          /* %ecx -> %eip */
 915         movl    REGOFF_RSP(%rsp), %esp
 916 
 917         ASSERT_UPCALL_MASK_IS_SET
 918         ALTENTRY(nopop_sys_syscall32_swapgs_sysretl)
 919         jmp     tr_sysretl

 920         SET_SIZE(nopop_sys_syscall32_swapgs_sysretl)
 921         /*NOTREACHED*/
 922 
 923 _full_syscall_postsys32:
 924         STI
 925         /*
 926          * Sigh, our optimism wasn't justified, put it back to LMS_SYSTEM
 927          * so that we can account for the extra work it takes us to finish.
 928          */
 929         MSTATE_TRANSITION(LMS_USER, LMS_SYSTEM)
 930         movq    %r15, %rdi
 931         movq    %r12, %rsi                      /* rval1 - %eax */
 932         movq    %r13, %rdx                      /* rval2 - %edx */
 933         call    syscall_exit
 934         MSTATE_TRANSITION(LMS_SYSTEM, LMS_USER)
 935         jmp     _sys_rtt
 936         SET_SIZE(sys_syscall32)
 937         SET_SIZE(brand_sys_syscall32)
 938 
 939 #endif  /* __lint */
 940 
 941 /*
 942  * System call handler via the sysenter instruction
 943  * Used only for 32-bit system calls on the 64-bit kernel.
 944  *
 945  * The caller in userland has arranged that:
 946  *
 947  * -    %eax contains the syscall number
 948  * -    %ecx contains the user %esp
 949  * -    %edx contains the return %eip
 950  * -    the user stack contains the args to the syscall
 951  *
 952  * Hardware and (privileged) initialization code have arranged that by
 953  * the time the sysenter instructions completes:
 954  *
 955  * - %rip is pointing to sys_sysenter (below).
 956  * - %cs and %ss are set to kernel text and stack (data) selectors.
 957  * - %rsp is pointing at the lwp's stack
 958  * - interrupts have been disabled.
 959  *
 960  * Note that we are unable to return both "rvals" to userland with
 961  * this call, as %edx is used by the sysexit instruction.
 962  *
 963  * One final complication in this routine is its interaction with
 964  * single-stepping in a debugger.  For most of the system call mechanisms, the
 965  * CPU automatically clears the single-step flag before we enter the kernel.
 966  * The sysenter mechanism does not clear the flag, so a user single-stepping
 967  * through a libc routine may suddenly find themself single-stepping through the
 968  * kernel.  To detect this, kmdb and trap() both compare the trap %pc to the
 969  * [brand_]sys_enter addresses on each single-step trap.  If it finds that we
 970  * have single-stepped to a sysenter entry point, it explicitly clears the flag
 971  * and executes the sys_sysenter routine.
 972  *
 973  * One final complication in this final complication is the fact that we have
 974  * two different entry points for sysenter: brand_sys_sysenter and sys_sysenter.
 975  * If we enter at brand_sys_sysenter and start single-stepping through the
 976  * kernel with kmdb, we will eventually hit the instruction at sys_sysenter.
 977  * kmdb cannot distinguish between that valid single-step and the undesirable
 978  * one mentioned above.  To avoid this situation, we simply add a jump over the
 979  * instruction at sys_sysenter to make it impossible to single-step to it.

 980  */
 981 #if defined(__lint)
 982 
 983 void
 984 sys_sysenter()
 985 {}
 986 
 987 #else   /* __lint */
 988 
 989         ENTRY_NP(brand_sys_sysenter)
 990         SWAPGS                          /* kernel gsbase */
 991         ALTENTRY(_brand_sys_sysenter_post_swapgs)
 992 
 993         BRAND_CALLBACK(BRAND_CB_SYSENTER, BRAND_URET_FROM_REG(%rdx))
 994         /*
 995          * Jump over sys_sysenter to allow single-stepping as described
 996          * above.
 997          */
 998         jmp     _sys_sysenter_post_swapgs
 999 
1000         ALTENTRY(sys_sysenter)
1001         SWAPGS                          /* kernel gsbase */

1002         ALTENTRY(_sys_sysenter_post_swapgs)
1003 
1004         movq    %gs:CPU_THREAD, %r15
1005 
1006         movl    $U32CS_SEL, REGOFF_CS(%rsp)
1007         movl    %ecx, REGOFF_RSP(%rsp)          /* wrapper: %esp -> %ecx */
1008         movl    %edx, REGOFF_RIP(%rsp)          /* wrapper: %eip -> %edx */
1009         /*
1010          * NOTE: none of the instructions that run before we get here should
1011          * clobber bits in (R)FLAGS! This includes the kpti trampoline.
1012          */
1013         pushfq
1014         popq    %r10
1015         movl    $UDS_SEL, REGOFF_SS(%rsp)
1016 
1017         /*
1018          * Set the interrupt flag before storing the flags to the
1019          * flags image on the stack so we can return to user with
1020          * interrupts enabled if we return via sys_rtt_syscall32
1021          */
1022         orq     $PS_IE, %r10
1023         movq    %r10, REGOFF_RFL(%rsp)
1024 
1025         movl    %edi, REGOFF_RDI(%rsp)
1026         movl    %esi, REGOFF_RSI(%rsp)
1027         movl    %ebp, REGOFF_RBP(%rsp)
1028         movl    %ebx, REGOFF_RBX(%rsp)
1029         movl    %edx, REGOFF_RDX(%rsp)
1030         movl    %ecx, REGOFF_RCX(%rsp)
1031         movl    %eax, REGOFF_RAX(%rsp)          /* wrapper: sysc# -> %eax */
1032         movq    $0, REGOFF_SAVFP(%rsp)
1033         movq    $0, REGOFF_SAVPC(%rsp)
1034 
1035         /*
1036          * Copy these registers here in case we end up stopped with
1037          * someone (like, say, /proc) messing with our register state.
1038          * We don't -restore- them unless we have to in update_sregs.
1039          *
1040          * Since userland -can't- change fsbase or gsbase directly,
1041          * we don't bother to capture them here.
1042          */
1043         xorl    %ebx, %ebx
1044         movw    %ds, %bx
1045         movq    %rbx, REGOFF_DS(%rsp)
1046         movw    %es, %bx
1047         movq    %rbx, REGOFF_ES(%rsp)
1048         movw    %fs, %bx
1049         movq    %rbx, REGOFF_FS(%rsp)
1050         movw    %gs, %bx
1051         movq    %rbx, REGOFF_GS(%rsp)
1052 
1053         /*
1054          * If we're trying to use TRAPTRACE though, I take that back: we're
1055          * probably debugging some problem in the SWAPGS logic and want to know
1056          * what the incoming gsbase was.
1057          *
1058          * Since we already did SWAPGS, record the KGSBASE.
1059          */
1060 #if defined(DEBUG) && defined(TRAPTRACE) && !defined(__xpv)
1061         movl    $MSR_AMD_KGSBASE, %ecx
1062         rdmsr
1063         movl    %eax, REGOFF_GSBASE(%rsp)
1064         movl    %edx, REGOFF_GSBASE+4(%rsp)
1065 #endif
1066 
1067         /*
1068          * Application state saved in the regs structure on the stack
1069          * %eax is the syscall number
1070          * %rsp is the thread's stack, %r15 is curthread
1071          * REG_RSP(%rsp) is the user's stack
1072          */
1073 
1074         SYSCALL_TRAPTRACE($TT_SYSENTER)
1075 
1076         movq    %rsp, %rbp
1077 
1078         movq    T_LWP(%r15), %r14
1079         ASSERT_NO_RUPDATE_PENDING(%r14)
1080 
1081         ENABLE_INTR_FLAGS
1082 
1083         /*
1084          * Catch 64-bit process trying to issue sysenter instruction
1085          * on Nocona based systems.
1086          */
1087         movq    LWP_PROCP(%r14), %rax
1088         cmpq    $DATAMODEL_ILP32, P_MODEL(%rax)
1089         je      7f
1090 
1091         /*
1092          * For a non-32-bit process, simulate a #ud, since that's what
1093          * native hardware does.  The traptrace entry (above) will
1094          * let you know what really happened.
1095          */
1096         movq    $T_ILLINST, REGOFF_TRAPNO(%rsp)
1097         movq    REGOFF_CS(%rsp), %rdi
1098         movq    %rdi, REGOFF_ERR(%rsp)
1099         movq    %rsp, %rdi
1100         movq    REGOFF_RIP(%rsp), %rsi
1101         movl    %gs:CPU_ID, %edx
1102         call    trap
1103         jmp     _sys_rtt
1104 7:
1105 
1106         MSTATE_TRANSITION(LMS_USER, LMS_SYSTEM)
1107         movl    REGOFF_RAX(%rsp), %eax  /* (%rax damaged by mstate calls) */
1108 
1109         ASSERT_LWPTOREGS(%r14, %rsp)
1110 
1111         incq    %gs:CPU_STATS_SYS_SYSCALL
1112 
1113         /*
1114          * Make some space for MAXSYSARGS (currently 8) 32-bit args
1115          * placed into 64-bit (long) arg slots, plus one 64-bit
1116          * (long) arg count, maintaining 16 byte alignment.
1117          */
1118         subq    $SYS_DROP, %rsp
1119         movb    $LWP_SYS, LWP_STATE(%r14)
1120         movq    %r15, %rdi
1121         movq    %rsp, %rsi
1122         call    syscall_entry
1123 
1124         /*
1125          * Fetch the arguments copied onto the kernel stack and put
1126          * them in the right registers to invoke a C-style syscall handler.
1127          * %rax contains the handler address.
1128          */
1129         movq    %rax, %rbx
1130         movl    0(%rsp), %edi
1131         movl    8(%rsp), %esi
1132         movl    0x10(%rsp), %edx
1133         movl    0x18(%rsp), %ecx
1134         movl    0x20(%rsp), %r8d
1135         movl    0x28(%rsp), %r9d
1136 
1137         call    *SY_CALLC(%rbx)
1138 
1139         movq    %rbp, %rsp      /* pop the args */
1140 
1141         /*
1142          * amd64 syscall handlers -always- return a 64-bit value in %rax.
1143          * On the 32-bit kernel, the always return that value in %eax:%edx
1144          * as required by the 32-bit ABI.
1145          *
1146          * Simulate the same behaviour by unconditionally splitting the
1147          * return value in the same way.
1148          */
1149         movq    %rax, %r13
1150         shrq    $32, %r13       /* upper 32-bits into %edx */
1151         movl    %eax, %r12d     /* lower 32-bits into %eax */
1152 
1153         /*
1154          * Optimistically assume that there's no post-syscall
1155          * work to do.  (This is to avoid having to call syscall_mstate()
1156          * with interrupts disabled)
1157          */
1158         MSTATE_TRANSITION(LMS_SYSTEM, LMS_USER)
1159 
1160         /*
1161          * We must protect ourselves from being descheduled here;
1162          * If we were, and we ended up on another cpu, or another
1163          * lwp got int ahead of us, it could change the segment
1164          * registers without us noticing before we return to userland.
1165          *
1166          * This cli is undone in the tr_sysexit trampoline code.
1167          */
1168         cli
1169         CHECK_POSTSYS_NE(%r15, %r14, %ebx)
1170         jne     _full_syscall_postsys32
1171         SIMPLE_SYSCALL_POSTSYS(%r15, %r14, %bx)
1172 
1173         /*
1174          * To get back to userland, load up the 32-bit registers and
1175          * sysexit back where we came from.
1176          */
1177 
1178         /*
1179          * Interrupts will be turned on by the 'sti' executed just before
1180          * sysexit.  The following ensures that restoring the user's rflags
1181          * doesn't enable interrupts too soon.
1182          */
1183         andq    $_BITNOT(PS_IE), REGOFF_RFL(%rsp)
1184 
1185         /*
1186          * (There's no point in loading up %edx because the sysexit
1187          * mechanism smashes it.)
1188          */
1189         movl    %r12d, %eax
1190         movl    REGOFF_RBX(%rsp), %ebx
1191         movl    REGOFF_RBP(%rsp), %ebp
1192         movl    REGOFF_RSI(%rsp), %esi
1193         movl    REGOFF_RDI(%rsp), %edi
1194 
1195         movl    REGOFF_RIP(%rsp), %edx  /* sysexit: %edx -> %eip */
1196         pushq   REGOFF_RFL(%rsp)
1197         popfq
1198         movl    REGOFF_RSP(%rsp), %ecx  /* sysexit: %ecx -> %esp */
1199         ALTENTRY(sys_sysenter_swapgs_sysexit)
1200         jmp     tr_sysexit


1201         SET_SIZE(sys_sysenter_swapgs_sysexit)
1202         SET_SIZE(sys_sysenter)
1203         SET_SIZE(_sys_sysenter_post_swapgs)
1204         SET_SIZE(brand_sys_sysenter)
1205 
1206 #endif  /* __lint */
1207 
1208 /*
1209  * This is the destination of the "int $T_SYSCALLINT" interrupt gate, used by
1210  * the generic i386 libc to do system calls. We do a small amount of setup
1211  * before jumping into the existing sys_syscall32 path.
1212  */
1213 #if defined(__lint)
1214 
1215 /*ARGSUSED*/
1216 void
1217 sys_syscall_int()
1218 {}
1219 
1220 #else   /* __lint */
1221 
1222         ENTRY_NP(brand_sys_syscall_int)
1223         SWAPGS                          /* kernel gsbase */
1224         XPV_TRAP_POP
1225         call    smap_enable
1226         BRAND_CALLBACK(BRAND_CB_INT91, BRAND_URET_FROM_INTR_STACK())
1227         jmp     nopop_syscall_int
1228 
1229         ALTENTRY(sys_syscall_int)
1230         SWAPGS                          /* kernel gsbase */
1231         XPV_TRAP_POP
1232         call    smap_enable
1233 
1234 nopop_syscall_int:
1235         movq    %gs:CPU_THREAD, %r15
1236         movq    T_STACK(%r15), %rsp
1237         movl    %eax, %eax
1238         /*
1239          * Set t_post_sys on this thread to force ourselves out via the slow
1240          * path. It might be possible at some later date to optimize this out
1241          * and use a faster return mechanism.
1242          */
1243         movb    $1, T_POST_SYS(%r15)
1244         CLEAN_CS
1245         jmp     _syscall32_save
1246         /*
1247          * There should be no instructions between this label and SWAPGS/IRET
1248          * or we could end up breaking branded zone support. See the usage of
1249          * this label in lx_brand_int80_callback and sn1_brand_int91_callback
1250          * for examples.
1251          *
1252          * We want to swapgs to maintain the invariant that all entries into
1253          * tr_iret_user are done on the user gsbase.
1254          */
1255         ALTENTRY(sys_sysint_swapgs_iret)
1256         SWAPGS
1257         jmp     tr_iret_user
1258         /*NOTREACHED*/
1259         SET_SIZE(sys_sysint_swapgs_iret)
1260         SET_SIZE(sys_syscall_int)
1261         SET_SIZE(brand_sys_syscall_int)
1262 
1263 #endif  /* __lint */
1264 
1265 /*
1266  * Legacy 32-bit applications and old libc implementations do lcalls;
1267  * we should never get here because the LDT entry containing the syscall
1268  * segment descriptor has the "segment present" bit cleared, which means
1269  * we end up processing those system calls in trap() via a not-present trap.
1270  *
1271  * We do it this way because a call gate unhelpfully does -nothing- to the
1272  * interrupt flag bit, so an interrupt can run us just after the lcall
1273  * completes, but just before the swapgs takes effect.   Thus the INTR_PUSH and
1274  * INTR_POP paths would have to be slightly more complex to dance around
1275  * this problem, and end up depending explicitly on the first
1276  * instruction of this handler being either swapgs or cli.
1277  */
1278 
1279 #if defined(__lint)
1280 
1281 /*ARGSUSED*/
1282 void
1283 sys_lcall32()
1284 {}
1285 
1286 #else   /* __lint */
1287 
1288         ENTRY_NP(sys_lcall32)
1289         SWAPGS                          /* kernel gsbase */
1290         pushq   $0
1291         pushq   %rbp
1292         movq    %rsp, %rbp
1293         leaq    __lcall_panic_str(%rip), %rdi
1294         xorl    %eax, %eax
1295         call    panic
1296         SET_SIZE(sys_lcall32)
1297 
1298 __lcall_panic_str:
1299         .string "sys_lcall32: shouldn't be here!"
1300 
1301 /*
1302  * Declare a uintptr_t which covers the entire pc range of syscall
1303  * handlers for the stack walkers that need this.
1304  */
1305         .align  CPTRSIZE
1306         .globl  _allsyscalls_size
1307         .type   _allsyscalls_size, @object
1308 _allsyscalls_size:
1309         .NWORD  . - _allsyscalls
1310         SET_SIZE(_allsyscalls_size)
1311 
1312 #endif  /* __lint */
1313 
1314 /*
1315  * These are the thread context handlers for lwps using sysenter/sysexit.
1316  */
1317 
1318 #if defined(__lint)
1319 
1320 /*ARGSUSED*/
1321 void
1322 sep_save(void *ksp)
1323 {}
1324 
1325 /*ARGSUSED*/
1326 void
1327 sep_restore(void *ksp)
1328 {}
1329 
1330 #else   /* __lint */
1331 
1332         /*
1333          * setting this value to zero as we switch away causes the
1334          * stack-pointer-on-sysenter to be NULL, ensuring that we
1335          * don't silently corrupt another (preempted) thread stack
1336          * when running an lwp that (somehow) didn't get sep_restore'd
1337          */
1338         ENTRY_NP(sep_save)
1339         xorl    %edx, %edx
1340         xorl    %eax, %eax
1341         movl    $MSR_INTC_SEP_ESP, %ecx
1342         wrmsr
1343         ret
1344         SET_SIZE(sep_save)
1345 
1346         /*
1347          * Update the kernel stack pointer as we resume onto this cpu.
1348          */
1349         ENTRY_NP(sep_restore)
1350         movq    %rdi, %rdx
1351         shrq    $32, %rdx
1352         movl    %edi, %eax
1353         movl    $MSR_INTC_SEP_ESP, %ecx
1354         wrmsr
1355         ret
1356         SET_SIZE(sep_restore)
1357 
1358 #endif  /* __lint */
--- EOF ---