VirtualBox

source: vbox/trunk/src/VBox/VMM/VMMR0/HMR0A.asm@ 87488

Last change on this file since 87488 was 87487, checked in by vboxsync, 4 years ago

VMM/HMVMX: Moving more stuff to HMR0PERVCPU. bugref:9217

  • Property svn:eol-style set to native
  • Property svn:keywords set to Id Revision
File size: 62.4 KB
Line 
1; $Id: HMR0A.asm 87487 2021-01-29 18:06:39Z vboxsync $
2;; @file
3; HM - Ring-0 VMX, SVM world-switch and helper routines.
4;
5
6;
7; Copyright (C) 2006-2020 Oracle Corporation
8;
9; This file is part of VirtualBox Open Source Edition (OSE), as
10; available from http://www.virtualbox.org. This file is free software;
11; you can redistribute it and/or modify it under the terms of the GNU
12; General Public License (GPL) as published by the Free Software
13; Foundation, in version 2 as it comes in the "COPYING" file of the
14; VirtualBox OSE distribution. VirtualBox OSE is distributed in the
15; hope that it will be useful, but WITHOUT ANY WARRANTY of any kind.
16;
17
18;*********************************************************************************************************************************
19;* Header Files *
20;*********************************************************************************************************************************
21;%define RT_ASM_WITH_SEH64 - trouble with SEH, alignment and (probably) 2nd pass optimizations.
22%define RT_ASM_WITH_SEH64_ALT ; Use asmdefs.mac hackery for manually emitting unwind info.
23%include "VBox/asmdefs.mac"
24%include "VBox/err.mac"
25%include "VBox/vmm/hm_vmx.mac"
26%include "VBox/vmm/cpum.mac"
27%include "VBox/vmm/gvm.mac"
28%include "iprt/x86.mac"
29%include "HMInternal.mac"
30
31%ifndef RT_ARCH_AMD64
32 %error AMD64 only.
33%endif
34
35
36;*********************************************************************************************************************************
37;* Defined Constants And Macros *
38;*********************************************************************************************************************************
39;; The offset of the XMM registers in X86FXSTATE.
40; Use define because I'm too lazy to convert the struct.
41%define XMM_OFF_IN_X86FXSTATE 160
42
43;; Spectre filler for 64-bit mode.
44; Choosen to be an invalid address (also with 5 level paging).
45%define SPECTRE_FILLER 0x02204204207fffff
46
47;;
48; Determine skipping restoring of GDTR, IDTR, TR across VMX non-root operation.
49;
50; @note This is normally done by hmR0VmxExportHostSegmentRegs and VMXRestoreHostState,
51; so much of this is untested code.
52; @{
53%define VMX_SKIP_GDTR
54%define VMX_SKIP_TR
55%define VBOX_SKIP_RESTORE_SEG
56%ifdef RT_OS_DARWIN
57 ; Load the NULL selector into DS, ES, FS and GS on 64-bit darwin so we don't
58 ; risk loading a stale LDT value or something invalid.
59 %define HM_64_BIT_USE_NULL_SEL
60 ; Darwin (Mavericks) uses IDTR limit to store the CPU number so we need to always restore it.
61 ; See @bugref{6875}.
62 %undef VMX_SKIP_IDTR
63%else
64 %define VMX_SKIP_IDTR
65%endif
66;; @}
67
68;; @def CALLEE_PRESERVED_REGISTER_COUNT
69; Number of registers pushed by PUSH_CALLEE_PRESERVED_REGISTERS
70%ifdef ASM_CALL64_GCC
71 %define CALLEE_PRESERVED_REGISTER_COUNT 5
72%else
73 %define CALLEE_PRESERVED_REGISTER_COUNT 7
74%endif
75
76;; @def PUSH_CALLEE_PRESERVED_REGISTERS
77; Macro for pushing all GPRs we must preserve for the caller.
78%macro PUSH_CALLEE_PRESERVED_REGISTERS 0
79 push r15
80 SEH64_PUSH_GREG r15
81 %assign cbFrame cbFrame + 8
82 %assign frm_saved_r15 -cbFrame
83
84 push r14
85 SEH64_PUSH_GREG r14
86 %assign cbFrame cbFrame + 8
87 %assign frm_saved_r14 -cbFrame
88
89 push r13
90 SEH64_PUSH_GREG r13
91 %assign cbFrame cbFrame + 8
92 %assign frm_saved_r13 -cbFrame
93
94 push r12
95 SEH64_PUSH_GREG r12
96 %assign cbFrame cbFrame + 8
97 %assign frm_saved_r12 -cbFrame
98
99 push rbx
100 SEH64_PUSH_GREG rbx
101 %assign cbFrame cbFrame + 8
102 %assign frm_saved_rbx -cbFrame
103
104 %ifdef ASM_CALL64_MSC
105 push rsi
106 SEH64_PUSH_GREG rsi
107 %assign cbFrame cbFrame + 8
108 %assign frm_saved_rsi -cbFrame
109
110 push rdi
111 SEH64_PUSH_GREG rdi
112 %assign cbFrame cbFrame + 8
113 %assign frm_saved_rdi -cbFrame
114 %endif
115%endmacro
116
117;; @def POP_CALLEE_PRESERVED_REGISTERS
118; Counterpart to PUSH_CALLEE_PRESERVED_REGISTERS for use in the epilogue.
119%macro POP_CALLEE_PRESERVED_REGISTERS 0
120 %ifdef ASM_CALL64_MSC
121 pop rdi
122 %assign cbFrame cbFrame - 8
123 %undef frm_saved_rdi
124
125 pop rsi
126 %assign cbFrame cbFrame - 8
127 %undef frm_saved_rsi
128 %endif
129 pop rbx
130 %assign cbFrame cbFrame - 8
131 %undef frm_saved_rbx
132
133 pop r12
134 %assign cbFrame cbFrame - 8
135 %undef frm_saved_r12
136
137 pop r13
138 %assign cbFrame cbFrame - 8
139 %undef frm_saved_r13
140
141 pop r14
142 %assign cbFrame cbFrame - 8
143 %undef frm_saved_r14
144
145 pop r15
146 %assign cbFrame cbFrame - 8
147 %undef frm_saved_r15
148%endmacro
149
150
151;; @def PUSH_RELEVANT_SEGMENT_REGISTERS
152; Macro saving all segment registers on the stack.
153; @param 1 Full width register name.
154; @param 2 16-bit register name for \a 1.
155; @cobbers rax, rdx, rcx
156%macro PUSH_RELEVANT_SEGMENT_REGISTERS 2
157 %ifndef VBOX_SKIP_RESTORE_SEG
158 %error untested code. probably does not work any more!
159 %ifndef HM_64_BIT_USE_NULL_SEL
160 mov %2, es
161 push %1
162 mov %2, ds
163 push %1
164 %endif
165
166 ; Special case for FS; Windows and Linux either don't use it or restore it when leaving kernel mode,
167 ; Solaris OTOH doesn't and we must save it.
168 mov ecx, MSR_K8_FS_BASE
169 rdmsr
170 push rdx
171 push rax
172 %ifndef HM_64_BIT_USE_NULL_SEL
173 push fs
174 %endif
175
176 ; Special case for GS; OSes typically use swapgs to reset the hidden base register for GS on entry into the kernel.
177 ; The same happens on exit.
178 mov ecx, MSR_K8_GS_BASE
179 rdmsr
180 push rdx
181 push rax
182 %ifndef HM_64_BIT_USE_NULL_SEL
183 push gs
184 %endif
185 %endif ; !VBOX_SKIP_RESTORE_SEG
186%endmacro ; PUSH_RELEVANT_SEGMENT_REGISTERS
187
188;; @def POP_RELEVANT_SEGMENT_REGISTERS
189; Macro restoring all segment registers on the stack.
190; @param 1 Full width register name.
191; @param 2 16-bit register name for \a 1.
192; @cobbers rax, rdx, rcx
193%macro POP_RELEVANT_SEGMENT_REGISTERS 2
194 %ifndef VBOX_SKIP_RESTORE_SEG
195 %error untested code. probably does not work any more!
196 ; Note: do not step through this code with a debugger!
197 %ifndef HM_64_BIT_USE_NULL_SEL
198 xor eax, eax
199 mov ds, ax
200 mov es, ax
201 mov fs, ax
202 mov gs, ax
203 %endif
204
205 %ifndef HM_64_BIT_USE_NULL_SEL
206 pop gs
207 %endif
208 pop rax
209 pop rdx
210 mov ecx, MSR_K8_GS_BASE
211 wrmsr
212
213 %ifndef HM_64_BIT_USE_NULL_SEL
214 pop fs
215 %endif
216 pop rax
217 pop rdx
218 mov ecx, MSR_K8_FS_BASE
219 wrmsr
220 ; Now it's safe to step again
221
222 %ifndef HM_64_BIT_USE_NULL_SEL
223 pop %1
224 mov ds, %2
225 pop %1
226 mov es, %2
227 %endif
228 %endif ; !VBOX_SKIP_RESTORE_SEG
229%endmacro ; POP_RELEVANT_SEGMENT_REGISTERS
230
231
232;;
233; Creates an indirect branch prediction barrier on CPUs that need and supports that.
234; @clobbers eax, edx, ecx
235; @param 1 How to address CPUMCTX.
236; @param 2 Which flag to test for (CPUMCTX_WSF_IBPB_ENTRY or CPUMCTX_WSF_IBPB_EXIT)
237%macro INDIRECT_BRANCH_PREDICTION_BARRIER_CTX 2
238 test byte [%1 + CPUMCTX.fWorldSwitcher], %2
239 jz %%no_indirect_branch_barrier
240 mov ecx, MSR_IA32_PRED_CMD
241 mov eax, MSR_IA32_PRED_CMD_F_IBPB
242 xor edx, edx
243 wrmsr
244%%no_indirect_branch_barrier:
245%endmacro
246
247;;
248; Creates an indirect branch prediction barrier on CPUs that need and supports that.
249; @clobbers eax, edx, ecx
250; @param 1 How to address VMCPU.
251; @param 2 Which flag to test for (CPUMCTX_WSF_IBPB_ENTRY or CPUMCTX_WSF_IBPB_EXIT)
252%macro INDIRECT_BRANCH_PREDICTION_BARRIER 2
253 test byte [%1 + VMCPU.cpum.GstCtx + CPUMCTX.fWorldSwitcher], %2
254 jz %%no_indirect_branch_barrier
255 mov ecx, MSR_IA32_PRED_CMD
256 mov eax, MSR_IA32_PRED_CMD_F_IBPB
257 xor edx, edx
258 wrmsr
259%%no_indirect_branch_barrier:
260%endmacro
261
262;;
263; Creates an indirect branch prediction and L1D barrier on CPUs that need and supports that.
264; @clobbers eax, edx, ecx
265; @param 1 How to address CPUMCTX.
266; @param 2 Which IBPB flag to test for (CPUMCTX_WSF_IBPB_ENTRY or CPUMCTX_WSF_IBPB_EXIT)
267; @param 3 Which FLUSH flag to test for (CPUMCTX_WSF_L1D_ENTRY)
268; @param 4 Which MDS flag to test for (CPUMCTX_WSF_MDS_ENTRY)
269%macro INDIRECT_BRANCH_PREDICTION_AND_L1_CACHE_BARRIER 4
270 ; Only one test+jmp when disabled CPUs.
271 test byte [%1 + CPUMCTX.fWorldSwitcher], (%2 | %3 | %4)
272 jz %%no_barrier_needed
273
274 ; The eax:edx value is the same for both.
275 AssertCompile(MSR_IA32_PRED_CMD_F_IBPB == MSR_IA32_FLUSH_CMD_F_L1D)
276 mov eax, MSR_IA32_PRED_CMD_F_IBPB
277 xor edx, edx
278
279 ; Indirect branch barrier.
280 test byte [%1 + CPUMCTX.fWorldSwitcher], %2
281 jz %%no_indirect_branch_barrier
282 mov ecx, MSR_IA32_PRED_CMD
283 wrmsr
284%%no_indirect_branch_barrier:
285
286 ; Level 1 data cache flush.
287 test byte [%1 + CPUMCTX.fWorldSwitcher], %3
288 jz %%no_cache_flush_barrier
289 mov ecx, MSR_IA32_FLUSH_CMD
290 wrmsr
291 jmp %%no_mds_buffer_flushing ; MDS flushing is included in L1D_FLUSH
292%%no_cache_flush_barrier:
293
294 ; MDS buffer flushing.
295 test byte [%1 + CPUMCTX.fWorldSwitcher], %4
296 jz %%no_mds_buffer_flushing
297 sub xSP, xSP
298 mov [xSP], ds
299 verw [xSP]
300 add xSP, xSP
301%%no_mds_buffer_flushing:
302
303%%no_barrier_needed:
304%endmacro
305
306
307;*********************************************************************************************************************************
308;* External Symbols *
309;*********************************************************************************************************************************
310%ifdef VBOX_WITH_KERNEL_USING_XMM
311extern NAME(CPUMIsGuestFPUStateActive)
312%endif
313
314
315BEGINCODE
316
317
318;;
319; Used on platforms with poor inline assembly support to retrieve all the
320; info from the CPU and put it in the @a pRestoreHost structure.
321;
322; @returns VBox status code
323; @param pRestoreHost msc: rcx gcc: rdi Pointer to the RestoreHost struct.
324; @param fHaveFsGsBase msc: dl gcc: sil Whether we can use rdfsbase or not.
325;
326ALIGNCODE(64)
327BEGINPROC hmR0VmxExportHostSegmentRegsAsmHlp
328%ifdef ASM_CALL64_MSC
329 %define pRestoreHost rcx
330%elifdef ASM_CALL64_GCC
331 %define pRestoreHost rdi
332%else
333 %error Unknown calling convension.
334%endif
335 SEH64_END_PROLOGUE
336
337 ; Start with the FS and GS base so we can trash DL/SIL.
338%ifdef ASM_CALL64_MSC
339 or dl, dl
340%else
341 or sil, sil
342%endif
343 jz .use_rdmsr_for_fs_and_gs_base
344 rdfsbase rax
345 mov [pRestoreHost + VMXRESTOREHOST.uHostFSBase], rax
346 rdgsbase rax
347 mov [pRestoreHost + VMXRESTOREHOST.uHostGSBase], rax
348.done_fs_and_gs_base:
349
350 ; TR, GDTR and IDTR
351 str [pRestoreHost + VMXRESTOREHOST.uHostSelTR]
352 sgdt [pRestoreHost + VMXRESTOREHOST.HostGdtr]
353 sidt [pRestoreHost + VMXRESTOREHOST.HostIdtr]
354
355 ; Segment registers.
356 xor eax, eax
357 mov eax, cs
358 mov [pRestoreHost + VMXRESTOREHOST.uHostSelCS], ax
359
360 mov eax, ss
361 mov [pRestoreHost + VMXRESTOREHOST.uHostSelSS], ax
362
363 mov eax, gs
364 mov [pRestoreHost + VMXRESTOREHOST.uHostSelGS], ax
365
366 mov eax, fs
367 mov [pRestoreHost + VMXRESTOREHOST.uHostSelFS], ax
368
369 mov eax, es
370 mov [pRestoreHost + VMXRESTOREHOST.uHostSelES], ax
371
372 mov eax, ds
373 mov [pRestoreHost + VMXRESTOREHOST.uHostSelDS], ax
374
375 ret
376
377ALIGNCODE(16)
378.use_rdmsr_for_fs_and_gs_base:
379%ifdef ASM_CALL64_MSC
380 mov r8, pRestoreHost
381%endif
382
383 mov ecx, MSR_K8_FS_BASE
384 rdmsr
385 shl rdx, 32
386 or rdx, rax
387 mov [r8 + VMXRESTOREHOST.uHostFSBase], rax
388
389 mov ecx, MSR_K8_GS_BASE
390 rdmsr
391 shl rdx, 32
392 or rdx, rax
393 mov [r8 + VMXRESTOREHOST.uHostGSBase], rax
394
395%ifdef ASM_CALL64_MSC
396 mov pRestoreHost, r8
397%endif
398 jmp .done_fs_and_gs_base
399%undef pRestoreHost
400ENDPROC hmR0VmxExportHostSegmentRegsAsmHlp
401
402
403;;
404; Restores host-state fields.
405;
406; @returns VBox status code
407; @param f32RestoreHost msc: ecx gcc: edi RestoreHost flags.
408; @param pRestoreHost msc: rdx gcc: rsi Pointer to the RestoreHost struct.
409;
410ALIGNCODE(64)
411BEGINPROC VMXRestoreHostState
412%ifndef ASM_CALL64_GCC
413 ; Use GCC's input registers since we'll be needing both rcx and rdx further
414 ; down with the wrmsr instruction. Use the R10 and R11 register for saving
415 ; RDI and RSI since MSC preserve the two latter registers.
416 mov r10, rdi
417 mov r11, rsi
418 mov rdi, rcx
419 mov rsi, rdx
420%endif
421 SEH64_END_PROLOGUE
422
423.restore_gdtr:
424 test edi, VMX_RESTORE_HOST_GDTR
425 jz .restore_idtr
426 lgdt [rsi + VMXRESTOREHOST.HostGdtr]
427
428.restore_idtr:
429 test edi, VMX_RESTORE_HOST_IDTR
430 jz .restore_ds
431 lidt [rsi + VMXRESTOREHOST.HostIdtr]
432
433.restore_ds:
434 test edi, VMX_RESTORE_HOST_SEL_DS
435 jz .restore_es
436 mov ax, [rsi + VMXRESTOREHOST.uHostSelDS]
437 mov ds, eax
438
439.restore_es:
440 test edi, VMX_RESTORE_HOST_SEL_ES
441 jz .restore_tr
442 mov ax, [rsi + VMXRESTOREHOST.uHostSelES]
443 mov es, eax
444
445.restore_tr:
446 test edi, VMX_RESTORE_HOST_SEL_TR
447 jz .restore_fs
448 ; When restoring the TR, we must first clear the busy flag or we'll end up faulting.
449 mov dx, [rsi + VMXRESTOREHOST.uHostSelTR]
450 mov ax, dx
451 and eax, X86_SEL_MASK_OFF_RPL ; mask away TI and RPL bits leaving only the descriptor offset
452 test edi, VMX_RESTORE_HOST_GDT_READ_ONLY | VMX_RESTORE_HOST_GDT_NEED_WRITABLE
453 jnz .gdt_readonly_or_need_writable
454 add rax, qword [rsi + VMXRESTOREHOST.HostGdtr + 2] ; xAX <- descriptor offset + GDTR.pGdt.
455 and dword [rax + 4], ~RT_BIT(9) ; clear the busy flag in TSS desc (bits 0-7=base, bit 9=busy bit)
456 ltr dx
457
458.restore_fs:
459 ;
460 ; When restoring the selector values for FS and GS, we'll temporarily trash
461 ; the base address (at least the high 32-bit bits, but quite possibly the
462 ; whole base address), the wrmsr will restore it correctly. (VT-x actually
463 ; restores the base correctly when leaving guest mode, but not the selector
464 ; value, so there is little problem with interrupts being enabled prior to
465 ; this restore job.)
466 ; We'll disable ints once for both FS and GS as that's probably faster.
467 ;
468 test edi, VMX_RESTORE_HOST_SEL_FS | VMX_RESTORE_HOST_SEL_GS
469 jz .restore_success
470 pushfq
471 cli ; (see above)
472
473 test edi, VMX_RESTORE_HOST_CAN_USE_WRFSBASE_AND_WRGSBASE
474 jz .restore_fs_using_wrmsr
475
476.restore_fs_using_wrfsbase:
477 test edi, VMX_RESTORE_HOST_SEL_FS
478 jz .restore_gs_using_wrgsbase
479 mov rax, qword [rsi + VMXRESTOREHOST.uHostFSBase]
480 mov cx, word [rsi + VMXRESTOREHOST.uHostSelFS]
481 mov fs, ecx
482 wrfsbase rax
483
484.restore_gs_using_wrgsbase:
485 test edi, VMX_RESTORE_HOST_SEL_GS
486 jz .restore_flags
487 mov rax, qword [rsi + VMXRESTOREHOST.uHostGSBase]
488 mov cx, word [rsi + VMXRESTOREHOST.uHostSelGS]
489 mov gs, ecx
490 wrgsbase rax
491
492.restore_flags:
493 popfq
494
495.restore_success:
496 mov eax, VINF_SUCCESS
497%ifndef ASM_CALL64_GCC
498 ; Restore RDI and RSI on MSC.
499 mov rdi, r10
500 mov rsi, r11
501%endif
502 ret
503
504ALIGNCODE(8)
505.gdt_readonly_or_need_writable:
506 test edi, VMX_RESTORE_HOST_GDT_NEED_WRITABLE
507 jnz .gdt_readonly_need_writable
508.gdt_readonly:
509 mov rcx, cr0
510 mov r9, rcx
511 add rax, qword [rsi + VMXRESTOREHOST.HostGdtr + 2] ; xAX <- descriptor offset + GDTR.pGdt.
512 and rcx, ~X86_CR0_WP
513 mov cr0, rcx
514 and dword [rax + 4], ~RT_BIT(9) ; clear the busy flag in TSS desc (bits 0-7=base, bit 9=busy bit)
515 ltr dx
516 mov cr0, r9
517 jmp .restore_fs
518
519ALIGNCODE(8)
520.gdt_readonly_need_writable:
521 add rax, qword [rsi + VMXRESTOREHOST.HostGdtrRw + 2] ; xAX <- descriptor offset + GDTR.pGdtRw
522 and dword [rax + 4], ~RT_BIT(9) ; clear the busy flag in TSS desc (bits 0-7=base, bit 9=busy bit)
523 lgdt [rsi + VMXRESTOREHOST.HostGdtrRw]
524 ltr dx
525 lgdt [rsi + VMXRESTOREHOST.HostGdtr] ; load the original GDT
526 jmp .restore_fs
527
528ALIGNCODE(8)
529.restore_fs_using_wrmsr:
530 test edi, VMX_RESTORE_HOST_SEL_FS
531 jz .restore_gs_using_wrmsr
532 mov eax, dword [rsi + VMXRESTOREHOST.uHostFSBase] ; uHostFSBase - Lo
533 mov edx, dword [rsi + VMXRESTOREHOST.uHostFSBase + 4h] ; uHostFSBase - Hi
534 mov cx, word [rsi + VMXRESTOREHOST.uHostSelFS]
535 mov fs, ecx
536 mov ecx, MSR_K8_FS_BASE
537 wrmsr
538
539.restore_gs_using_wrmsr:
540 test edi, VMX_RESTORE_HOST_SEL_GS
541 jz .restore_flags
542 mov eax, dword [rsi + VMXRESTOREHOST.uHostGSBase] ; uHostGSBase - Lo
543 mov edx, dword [rsi + VMXRESTOREHOST.uHostGSBase + 4h] ; uHostGSBase - Hi
544 mov cx, word [rsi + VMXRESTOREHOST.uHostSelGS]
545 mov gs, ecx
546 mov ecx, MSR_K8_GS_BASE
547 wrmsr
548 jmp .restore_flags
549ENDPROC VMXRestoreHostState
550
551
552;;
553; Dispatches an NMI to the host.
554;
555ALIGNCODE(16)
556BEGINPROC VMXDispatchHostNmi
557 ; NMI is always vector 2. The IDT[2] IRQ handler cannot be anything else. See Intel spec. 6.3.1 "External Interrupts".
558 SEH64_END_PROLOGUE
559 int 2
560 ret
561ENDPROC VMXDispatchHostNmi
562
563
564;;
565; Common restore logic for success and error paths. We duplicate this because we
566; don't want to waste writing the VINF_SUCCESS return value to the stack in the
567; regular code path.
568;
569; @param 1 Zero if regular return, non-zero if error return. Controls label emission.
570; @param 2 fLoadSaveGuestXcr0 value
571; @param 3 The (CPUMCTX_WSF_IBPB_ENTRY | CPUMCTX_WSF_L1D_ENTRY | CPUMCTX_WSF_MDS_ENTRY) + CPUMCTX_WSF_IBPB_EXIT value.
572; The entry values are either all set or not at all, as we're too lazy to flesh out all the variants.
573; @param 4 The SSE saving/restoring: 0 to do nothing, 1 to do it manually, 2 to use xsave/xrstor.
574;
575; @note Important that this does not modify cbFrame or rsp.
576%macro RESTORE_STATE_VMX 4
577 ; Restore base and limit of the IDTR & GDTR.
578 %ifndef VMX_SKIP_IDTR
579 lidt [rsp + cbFrame + frm_saved_idtr]
580 %endif
581 %ifndef VMX_SKIP_GDTR
582 lgdt [rsp + cbFrame + frm_saved_gdtr]
583 %endif
584
585 ; Save the guest state and restore the non-volatile registers. We use rax=pGstCtx here.
586 mov [rsp + cbFrame + frm_guest_rax], rax
587 mov rax, [rsp + cbFrame + frm_pGstCtx]
588
589 mov qword [rax + CPUMCTX.ebp], rbp
590 lea rbp, [rsp + cbFrame] ; re-establish the frame pointer as early as possible.
591 mov qword [rax + CPUMCTX.ecx], rcx
592 mov rcx, SPECTRE_FILLER
593 mov qword [rax + CPUMCTX.edx], rdx
594 mov rdx, [rbp + frm_guest_rax]
595 mov qword [rax + CPUMCTX.eax], rdx
596 mov rdx, rcx
597 mov qword [rax + CPUMCTX.r8], r8
598 mov r8, rcx
599 mov qword [rax + CPUMCTX.r9], r9
600 mov r9, rcx
601 mov qword [rax + CPUMCTX.r10], r10
602 mov r10, rcx
603 mov qword [rax + CPUMCTX.r11], r11
604 mov r11, rcx
605 mov qword [rax + CPUMCTX.esi], rsi
606 %ifdef ASM_CALL64_MSC
607 mov rsi, [rbp + frm_saved_rsi]
608 %else
609 mov rsi, rcx
610 %endif
611 mov qword [rax + CPUMCTX.edi], rdi
612 %ifdef ASM_CALL64_MSC
613 mov rdi, [rbp + frm_saved_rdi]
614 %else
615 mov rdi, rcx
616 %endif
617 mov qword [rax + CPUMCTX.ebx], rbx
618 mov rbx, [rbp + frm_saved_rbx]
619 mov qword [rax + CPUMCTX.r12], r12
620 mov r12, [rbp + frm_saved_r12]
621 mov qword [rax + CPUMCTX.r13], r13
622 mov r13, [rbp + frm_saved_r13]
623 mov qword [rax + CPUMCTX.r14], r14
624 mov r14, [rbp + frm_saved_r14]
625 mov qword [rax + CPUMCTX.r15], r15
626 mov r15, [rbp + frm_saved_r15]
627
628 mov rdx, cr2
629 mov qword [rax + CPUMCTX.cr2], rdx
630 mov rdx, rcx
631
632 %if %4 != 0
633 ; Save the context pointer in r8 for the SSE save/restore.
634 mov r8, rax
635 %endif
636
637 %if %3 & CPUMCTX_WSF_IBPB_EXIT
638 ; Fight spectre (trashes rax, rdx and rcx).
639 %if %1 = 0 ; Skip this in failure branch (=> guru)
640 mov ecx, MSR_IA32_PRED_CMD
641 mov eax, MSR_IA32_PRED_CMD_F_IBPB
642 xor edx, edx
643 wrmsr
644 %endif
645 %endif
646
647 %ifndef VMX_SKIP_TR
648 ; Restore TSS selector; must mark it as not busy before using ltr!
649 ; ASSUME that this is supposed to be 'BUSY' (saves 20-30 ticks on the T42p).
650 %ifndef VMX_SKIP_GDTR
651 lgdt [rbp + frm_saved_gdtr]
652 %endif
653 movzx eax, word [rbp + frm_saved_tr]
654 mov ecx, eax
655 and eax, X86_SEL_MASK_OFF_RPL ; mask away TI and RPL bits leaving only the descriptor offset
656 add rax, [rbp + frm_saved_gdtr + 2] ; eax <- GDTR.address + descriptor offset
657 and dword [rax + 4], ~RT_BIT(9) ; clear the busy flag in TSS desc (bits 0-7=base, bit 9=busy bit)
658 ltr cx
659 %endif
660 movzx edx, word [rbp + frm_saved_ldtr]
661 test edx, edx
662 jz %%skip_ldt_write
663 lldt dx
664%%skip_ldt_write:
665
666 %if %1 != 0
667.return_after_vmwrite_error:
668 %endif
669 ; Restore segment registers.
670 ;POP_RELEVANT_SEGMENT_REGISTERS rax, ax - currently broken.
671
672 %if %2 != 0
673 ; Restore the host XCR0.
674 xor ecx, ecx
675 mov eax, [rbp + frm_uHostXcr0]
676 mov edx, [rbp + frm_uHostXcr0 + 4]
677 xsetbv
678 %endif
679%endmacro ; RESTORE_STATE_VMX
680
681
682;;
683; hmR0VmxStartVm template
684;
685; @param 1 The suffix of the variation.
686; @param 2 fLoadSaveGuestXcr0 value
687; @param 3 The CPUMCTX_WSF_IBPB_ENTRY + CPUMCTX_WSF_IBPB_EXIT value.
688; @param 4 The SSE saving/restoring: 0 to do nothing, 1 to do it manually, 2 to use xsave/xrstor.
689; Drivers shouldn't use AVX registers without saving+loading:
690; https://msdn.microsoft.com/en-us/library/windows/hardware/ff545910%28v=vs.85%29.aspx?f=255&MSPPError=-2147217396
691; However the compiler docs have different idea:
692; https://msdn.microsoft.com/en-us/library/9z1stfyw.aspx
693; We'll go with the former for now.
694;
695%macro hmR0VmxStartVmTemplate 4
696
697;;
698; Prepares for and executes VMLAUNCH/VMRESUME (64 bits guest mode)
699;
700; @returns VBox status code
701; @param NULL msc:rcx, gcc:rdi The cross context VM structure. (unused, passing NULL now)
702; @param pVCpu msc:rdx, gcc:rsi The cross context virtual CPU structure of the calling EMT.
703; @param fResume msc:r8l, gcc:dl Whether to use vmlauch/vmresume.
704;
705ALIGNCODE(64)
706BEGINPROC RT_CONCAT(hmR0VmxStartVm,%1)
707 %ifdef VBOX_WITH_KERNEL_USING_XMM
708 %if %4 = 0
709 ;
710 ; The non-saving variant will currently check the two SSE preconditions and pick
711 ; the right variant to continue with. Later we can see if we can't manage to
712 ; move these decisions into hmR0VmxUpdateStartVmFunction().
713 ;
714 %ifdef ASM_CALL64_MSC
715 test byte [rdx + VMCPU.cpum.GstCtx + CPUMCTX.fUsedFpuGuest], 1
716 %else
717 test byte [rsi + VMCPU.cpum.GstCtx + CPUMCTX.fUsedFpuGuest], 1
718 %endif
719 jz .save_xmm_no_need
720 %ifdef ASM_CALL64_MSC
721 cmp dword [rdx + VMCPU.cpum.GstCtx + CPUMCTX.fXStateMask], 0
722 %else
723 cmp dword [rsi + VMCPU.cpum.GstCtx + CPUMCTX.fXStateMask], 0
724 %endif
725 je RT_CONCAT3(hmR0VmxStartVm,%1,_SseManual)
726 jmp RT_CONCAT3(hmR0VmxStartVm,%1,_SseXSave)
727.save_xmm_no_need:
728 %endif
729 %endif
730 push xBP
731 SEH64_PUSH_xBP
732 mov xBP, xSP
733 SEH64_SET_FRAME_xBP 0
734 pushf
735 cli
736
737 %define frm_fRFlags -008h
738 %define frm_pGstCtx -010h ; Where we stash guest CPU context for use after the vmrun.
739 %define frm_uHostXcr0 -020h ; 128-bit
740 %define frm_saved_gdtr -02ah ; 16+64: Only used when VMX_SKIP_GDTR isn't defined
741 %define frm_saved_tr -02ch ; 16-bit: Only used when VMX_SKIP_TR isn't defined
742 %define frm_MDS_seg -030h ; 16-bit: Temporary storage for the MDS flushing.
743 %define frm_saved_idtr -03ah ; 16+64: Only used when VMX_SKIP_IDTR isn't defined
744 %define frm_saved_ldtr -03ch ; 16-bit: always saved.
745 %define frm_rcError -040h ; 32-bit: Error status code (not used in the success path)
746 %define frm_guest_rax -048h ; Temporary storage slot for guest RAX.
747 %if %4 = 0
748 %assign cbFrame 048h
749 %else
750 %define frm_saved_xmm6 -050h
751 %define frm_saved_xmm7 -060h
752 %define frm_saved_xmm8 -070h
753 %define frm_saved_xmm9 -080h
754 %define frm_saved_xmm10 -090h
755 %define frm_saved_xmm11 -0a0h
756 %define frm_saved_xmm12 -0b0h
757 %define frm_saved_xmm13 -0c0h
758 %define frm_saved_xmm14 -0d0h
759 %define frm_saved_xmm15 -0e0h
760 %define frm_saved_mxcsr -0f0h
761 %assign cbFrame 0f0h
762 %endif
763 %assign cbBaseFrame cbFrame
764 sub rsp, cbFrame - 8h
765 SEH64_ALLOCATE_STACK cbFrame
766
767 ; Save all general purpose host registers.
768 PUSH_CALLEE_PRESERVED_REGISTERS
769 ;PUSH_RELEVANT_SEGMENT_REGISTERS xAX, ax - currently broken
770 SEH64_END_PROLOGUE
771
772 ;
773 ; Unify the input parameter registers: rsi=pVCpu, bl=fResume, rdi=&pVCpu->cpum.GstCtx;
774 ;
775 %ifdef ASM_CALL64_GCC
776 mov ebx, edx ; fResume
777 %else
778 mov rsi, rdx ; pVCpu
779 mov ebx, r8d ; fResume
780 %endif
781 lea rdi, [rsi + VMCPU.cpum.GstCtx]
782 mov [rbp + frm_pGstCtx], rdi
783
784 %ifdef VBOX_STRICT
785 ;
786 ; Verify template preconditions / parameters to ensure HMSVM.cpp didn't miss some state change.
787 ;
788 cmp byte [rsi + GVMCPU.hmr0 + HMR0PERVCPU.fLoadSaveGuestXcr0], %2
789 mov eax, VERR_VMX_STARTVM_PRECOND_0
790 jne NAME(RT_CONCAT(hmR0VmxStartVmHostRIP,%1).precond_failure_return)
791
792 mov eax, [rsi + VMCPU.cpum.GstCtx + CPUMCTX.fWorldSwitcher]
793 and eax, CPUMCTX_WSF_IBPB_ENTRY | CPUMCTX_WSF_L1D_ENTRY | CPUMCTX_WSF_MDS_ENTRY | CPUMCTX_WSF_IBPB_EXIT
794 cmp eax, %3
795 mov eax, VERR_VMX_STARTVM_PRECOND_1
796 jne NAME(RT_CONCAT(hmR0VmxStartVmHostRIP,%1).precond_failure_return)
797
798 %ifdef VBOX_WITH_KERNEL_USING_XMM
799 mov eax, VERR_VMX_STARTVM_PRECOND_2
800 test byte [rsi + VMCPU.cpum.GstCtx + CPUMCTX.fUsedFpuGuest], 1
801 %if %4 = 0
802 jnz NAME(RT_CONCAT(hmR0VmxStartVmHostRIP,%1).precond_failure_return)
803 %else
804 jz NAME(RT_CONCAT(hmR0VmxStartVmHostRIP,%1).precond_failure_return)
805
806 mov eax, VERR_VMX_STARTVM_PRECOND_3
807 cmp dword [rsi + VMCPU.cpum.GstCtx + CPUMCTX.fXStateMask], 0
808 %if %4 = 1
809 jne NAME(RT_CONCAT(hmR0VmxStartVmHostRIP,%1).precond_failure_return)
810 %elif %4 = 2
811 je NAME(RT_CONCAT(hmR0VmxStartVmHostRIP,%1).precond_failure_return)
812 %else
813 %error Invalid template parameter 4.
814 %endif
815 %endif
816 %endif
817 %endif ; VBOX_STRICT
818
819 %if %4 != 0
820 ; Save the non-volatile SSE host register state.
821 movdqa [rbp + frm_saved_xmm6 ], xmm6
822 movdqa [rbp + frm_saved_xmm7 ], xmm7
823 movdqa [rbp + frm_saved_xmm8 ], xmm8
824 movdqa [rbp + frm_saved_xmm9 ], xmm9
825 movdqa [rbp + frm_saved_xmm10], xmm10
826 movdqa [rbp + frm_saved_xmm11], xmm11
827 movdqa [rbp + frm_saved_xmm12], xmm12
828 movdqa [rbp + frm_saved_xmm13], xmm13
829 movdqa [rbp + frm_saved_xmm14], xmm14
830 movdqa [rbp + frm_saved_xmm15], xmm15
831 stmxcsr [rbp + frm_saved_mxcsr]
832
833 ; Load the guest state related to the above non-volatile and volatile SSE registers. Trashes rcx, eax and edx.
834 mov rcx, [rdi + CPUMCTX.pXStateR0]
835 %if %4 = 1 ; manual
836 movdqa xmm0, [rcx + XMM_OFF_IN_X86FXSTATE + 000h]
837 movdqa xmm1, [rcx + XMM_OFF_IN_X86FXSTATE + 010h]
838 movdqa xmm2, [rcx + XMM_OFF_IN_X86FXSTATE + 020h]
839 movdqa xmm3, [rcx + XMM_OFF_IN_X86FXSTATE + 030h]
840 movdqa xmm4, [rcx + XMM_OFF_IN_X86FXSTATE + 040h]
841 movdqa xmm5, [rcx + XMM_OFF_IN_X86FXSTATE + 050h]
842 movdqa xmm6, [rcx + XMM_OFF_IN_X86FXSTATE + 060h]
843 movdqa xmm7, [rcx + XMM_OFF_IN_X86FXSTATE + 070h]
844 movdqa xmm8, [rcx + XMM_OFF_IN_X86FXSTATE + 080h]
845 movdqa xmm9, [rcx + XMM_OFF_IN_X86FXSTATE + 090h]
846 movdqa xmm10, [rcx + XMM_OFF_IN_X86FXSTATE + 0a0h]
847 movdqa xmm11, [rcx + XMM_OFF_IN_X86FXSTATE + 0b0h]
848 movdqa xmm12, [rcx + XMM_OFF_IN_X86FXSTATE + 0c0h]
849 movdqa xmm13, [rcx + XMM_OFF_IN_X86FXSTATE + 0d0h]
850 movdqa xmm14, [rcx + XMM_OFF_IN_X86FXSTATE + 0e0h]
851 movdqa xmm15, [rcx + XMM_OFF_IN_X86FXSTATE + 0f0h]
852 ldmxcsr [rcx + X86FXSTATE.MXCSR]
853 %elif %4 = 2 ; use xrstor/xsave
854 mov eax, [rsi + VMCPU.cpum.GstCtx + CPUMCTX.fXStateMask]
855 and eax, CPUM_VOLATILE_XSAVE_GUEST_COMPONENTS
856 xor edx, edx
857 xrstor [rcx]
858 %else
859 %error invalid template parameter 4
860 %endif
861 %endif
862
863 %if %2 != 0
864 ; Save the host XCR0 and load the guest one if necessary.
865 ; Note! Trashes rax, rdx and rcx.
866 xor ecx, ecx
867 xgetbv ; save the host one on the stack
868 mov [rbp + frm_uHostXcr0], eax
869 mov [rbp + frm_uHostXcr0 + 4], edx
870
871 mov eax, [rdi + CPUMCTX.aXcr] ; load the guest one
872 mov edx, [rdi + CPUMCTX.aXcr + 4]
873 xor ecx, ecx ; paranoia; indicate that we must restore XCR0 (popped into ecx, thus 0)
874 xsetbv
875 %endif
876
877 ; Save host LDTR.
878 sldt word [rbp + frm_saved_ldtr]
879
880 %ifndef VMX_SKIP_TR
881 ; The host TR limit is reset to 0x67; save & restore it manually.
882 str word [rbp + frm_saved_tr]
883 %endif
884
885 %ifndef VMX_SKIP_GDTR
886 ; VT-x only saves the base of the GDTR & IDTR and resets the limit to 0xffff; we must restore the limit correctly!
887 sgdt [rbp + frm_saved_gdtr]
888 %endif
889 %ifndef VMX_SKIP_IDTR
890 sidt [rbp + frm_saved_idtr]
891 %endif
892
893 ; Load CR2 if necessary (expensive as writing CR2 is a synchronizing instruction - (bird: still expensive on 10980xe)).
894 mov rcx, qword [rdi + CPUMCTX.cr2]
895 mov rdx, cr2
896 cmp rcx, rdx
897 je .skip_cr2_write
898 mov cr2, rcx
899.skip_cr2_write:
900
901 ; Set the vmlaunch/vmresume "return" host RIP and RSP values if they've changed (unlikly).
902 ; The vmwrite isn't quite for free (on an 10980xe at least), thus we check if anything changed
903 ; before writing here.
904 lea rcx, [NAME(RT_CONCAT(hmR0VmxStartVmHostRIP,%1)) wrt rip]
905 cmp rcx, [rsi + GVMCPU.hmr0 + HMR0PERVCPU.u + HMR0CPUVMX.uHostRIP]
906 jne .write_host_rip
907.wrote_host_rip:
908 cmp rsp, [rsi + GVMCPU.hmr0 + HMR0PERVCPU.u + HMR0CPUVMX.uHostRSP]
909 jne .write_host_rsp
910.wrote_host_rsp:
911
912 ;
913 ; Fight spectre and similar. Trashes rax, rcx, and rdx.
914 ;
915 %if %3 & (CPUMCTX_WSF_IBPB_ENTRY | CPUMCTX_WSF_L1D_ENTRY) ; The eax:edx value is the same for the first two.
916 AssertCompile(MSR_IA32_PRED_CMD_F_IBPB == MSR_IA32_FLUSH_CMD_F_L1D)
917 mov eax, MSR_IA32_PRED_CMD_F_IBPB
918 xor edx, edx
919 %endif
920 %if %3 & CPUMCTX_WSF_IBPB_ENTRY ; Indirect branch barrier.
921 mov ecx, MSR_IA32_PRED_CMD
922 wrmsr
923 %endif
924 %if %3 & CPUMCTX_WSF_L1D_ENTRY ; Level 1 data cache flush.
925 mov ecx, MSR_IA32_FLUSH_CMD
926 wrmsr
927 %elif %3 & CPUMCTX_WSF_MDS_ENTRY ; MDS flushing is included in L1D_FLUSH
928 mov word [rbp + frm_MDS_seg], ds
929 verw word [rbp + frm_MDS_seg]
930 %endif
931
932 ; Resume or start VM?
933 cmp bl, 0 ; fResume
934
935 ; Load guest general purpose registers.
936 mov rax, qword [rdi + CPUMCTX.eax]
937 mov rbx, qword [rdi + CPUMCTX.ebx]
938 mov rcx, qword [rdi + CPUMCTX.ecx]
939 mov rdx, qword [rdi + CPUMCTX.edx]
940 mov rbp, qword [rdi + CPUMCTX.ebp]
941 mov rsi, qword [rdi + CPUMCTX.esi]
942 mov r8, qword [rdi + CPUMCTX.r8]
943 mov r9, qword [rdi + CPUMCTX.r9]
944 mov r10, qword [rdi + CPUMCTX.r10]
945 mov r11, qword [rdi + CPUMCTX.r11]
946 mov r12, qword [rdi + CPUMCTX.r12]
947 mov r13, qword [rdi + CPUMCTX.r13]
948 mov r14, qword [rdi + CPUMCTX.r14]
949 mov r15, qword [rdi + CPUMCTX.r15]
950 mov rdi, qword [rdi + CPUMCTX.edi]
951
952 je .vmlaunch64_launch
953
954 vmresume
955 jc NAME(RT_CONCAT(hmR0VmxStartVmHostRIP,%1).vmxstart64_invalid_vmcs_ptr)
956 jz NAME(RT_CONCAT(hmR0VmxStartVmHostRIP,%1).vmxstart64_start_failed)
957 jmp NAME(RT_CONCAT(hmR0VmxStartVmHostRIP,%1)) ; here if vmresume detected a failure
958
959.vmlaunch64_launch:
960 vmlaunch
961 jc NAME(RT_CONCAT(hmR0VmxStartVmHostRIP,%1).vmxstart64_invalid_vmcs_ptr)
962 jz NAME(RT_CONCAT(hmR0VmxStartVmHostRIP,%1).vmxstart64_start_failed)
963 jmp NAME(RT_CONCAT(hmR0VmxStartVmHostRIP,%1)) ; here if vmlaunch detected a failure
964
965
966; Put these two outside the normal code path as they should rarely change.
967ALIGNCODE(8)
968.write_host_rip:
969 mov [rsi + GVMCPU.hmr0 + HMR0PERVCPU.u + HMR0CPUVMX.uHostRIP], rcx
970 mov eax, VMX_VMCS_HOST_RIP ;; @todo It is only strictly necessary to write VMX_VMCS_HOST_RIP when
971 vmwrite rax, rcx ;; the VMXVMCSINFO::pfnStartVM function changes (eventually
972 %ifdef VBOX_STRICT ;; take the Windows/SSE stuff into account then)...
973 jna NAME(RT_CONCAT(hmR0VmxStartVmHostRIP,%1).vmwrite_failed)
974 %endif
975 jmp .wrote_host_rip
976
977ALIGNCODE(8)
978.write_host_rsp:
979 mov [rsi + GVMCPU.hmr0 + HMR0PERVCPU.u + HMR0CPUVMX.uHostRSP], rsp
980 mov eax, VMX_VMCS_HOST_RSP
981 vmwrite rax, rsp
982 %ifdef VBOX_STRICT
983 jna NAME(RT_CONCAT(hmR0VmxStartVmHostRIP,%1).vmwrite_failed)
984 %endif
985 jmp .wrote_host_rsp
986
987ALIGNCODE(64)
988GLOBALNAME RT_CONCAT(hmR0VmxStartVmHostRIP,%1)
989 RESTORE_STATE_VMX 0, %2, %3, %4
990 mov eax, VINF_SUCCESS
991
992.vmstart64_end:
993 %if %4 != 0
994 mov r11d, eax ; save the return code.
995
996 ; Save the guest SSE state related to non-volatile and volatile SSE registers.
997 mov rcx, [r8 + CPUMCTX.pXStateR0]
998 %if %4 = 1 ; manual
999 stmxcsr [rcx + X86FXSTATE.MXCSR]
1000 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 000h], xmm0
1001 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 010h], xmm1
1002 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 020h], xmm2
1003 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 030h], xmm3
1004 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 040h], xmm4
1005 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 050h], xmm5
1006 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 060h], xmm6
1007 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 070h], xmm7
1008 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 080h], xmm8
1009 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 090h], xmm9
1010 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 0a0h], xmm10
1011 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 0b0h], xmm11
1012 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 0c0h], xmm12
1013 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 0d0h], xmm13
1014 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 0e0h], xmm14
1015 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 0f0h], xmm15
1016 %elif %4 = 2 ; use xrstor/xsave
1017 mov eax, [r8 + CPUMCTX.fXStateMask]
1018 and eax, CPUM_VOLATILE_XSAVE_GUEST_COMPONENTS
1019 xor edx, edx
1020 xsave [rcx]
1021 %else
1022 %error invalid template parameter 4
1023 %endif
1024
1025 ; Restore the host non-volatile SSE register state.
1026 ldmxcsr [rbp + frm_saved_mxcsr]
1027 movdqa xmm6, [rbp + frm_saved_xmm6 ]
1028 movdqa xmm7, [rbp + frm_saved_xmm7 ]
1029 movdqa xmm8, [rbp + frm_saved_xmm8 ]
1030 movdqa xmm9, [rbp + frm_saved_xmm9 ]
1031 movdqa xmm10, [rbp + frm_saved_xmm10]
1032 movdqa xmm11, [rbp + frm_saved_xmm11]
1033 movdqa xmm12, [rbp + frm_saved_xmm12]
1034 movdqa xmm13, [rbp + frm_saved_xmm13]
1035 movdqa xmm14, [rbp + frm_saved_xmm14]
1036 movdqa xmm15, [rbp + frm_saved_xmm15]
1037
1038 mov eax, r11d
1039 %endif ; %4 != 0
1040
1041 lea rsp, [rbp + frm_fRFlags]
1042 popf
1043 leave
1044 ret
1045
1046 ;
1047 ; Error returns.
1048 ;
1049 %ifdef VBOX_STRICT
1050.vmwrite_failed:
1051 mov dword [rsp + cbFrame + frm_rcError], VERR_VMX_INVALID_VMCS_FIELD
1052 jz .return_after_vmwrite_error
1053 mov dword [rsp + cbFrame + frm_rcError], VERR_VMX_INVALID_VMCS_PTR
1054 jmp .return_after_vmwrite_error
1055 %endif
1056.vmxstart64_invalid_vmcs_ptr:
1057 mov dword [rsp + cbFrame + frm_rcError], VERR_VMX_INVALID_VMCS_PTR_TO_START_VM
1058 jmp .vmstart64_error_return
1059.vmxstart64_start_failed:
1060 mov dword [rsp + cbFrame + frm_rcError], VERR_VMX_UNABLE_TO_START_VM
1061.vmstart64_error_return:
1062 RESTORE_STATE_VMX 1, %2, %3, %4
1063 mov eax, [rbp + frm_rcError]
1064 jmp .vmstart64_end
1065
1066 %ifdef VBOX_STRICT
1067 ; Precondition checks failed.
1068.precond_failure_return:
1069 POP_CALLEE_PRESERVED_REGISTERS
1070 %if cbFrame != cbBaseFrame
1071 %error Bad frame size value: cbFrame, expected cbBaseFrame
1072 %endif
1073 jmp .vmstart64_end
1074 %endif
1075
1076 %undef frm_fRFlags
1077 %undef frm_pGstCtx
1078 %undef frm_uHostXcr0
1079 %undef frm_saved_gdtr
1080 %undef frm_saved_tr
1081 %undef frm_fNoRestoreXcr0
1082 %undef frm_saved_idtr
1083 %undef frm_saved_ldtr
1084 %undef frm_rcError
1085 %undef frm_guest_rax
1086 %undef cbFrame
1087ENDPROC RT_CONCAT(hmR0VmxStartVm,%1)
1088
1089%endmacro ; hmR0VmxStartVmTemplate
1090
1091%macro hmR0VmxStartVmSseTemplate 3
1092hmR0VmxStartVmTemplate _SansXcr0_SansIbpbEntry_SansL1dEntry_SansMdsEntry_SansIbpbExit %+ %2, 0, 0 | 0 | 0 | 0 , %1
1093hmR0VmxStartVmTemplate _WithXcr0_SansIbpbEntry_SansL1dEntry_SansMdsEntry_SansIbpbExit %+ %2, 1, 0 | 0 | 0 | 0 , %1
1094hmR0VmxStartVmTemplate _SansXcr0_WithIbpbEntry_SansL1dEntry_SansMdsEntry_SansIbpbExit %+ %2, 0, CPUMCTX_WSF_IBPB_ENTRY | 0 | 0 | 0 , %1
1095hmR0VmxStartVmTemplate _WithXcr0_WithIbpbEntry_SansL1dEntry_SansMdsEntry_SansIbpbExit %+ %2, 1, CPUMCTX_WSF_IBPB_ENTRY | 0 | 0 | 0 , %1
1096hmR0VmxStartVmTemplate _SansXcr0_SansIbpbEntry_WithL1dEntry_SansMdsEntry_SansIbpbExit %+ %2, 0, 0 | CPUMCTX_WSF_L1D_ENTRY | 0 | 0 , %1
1097hmR0VmxStartVmTemplate _WithXcr0_SansIbpbEntry_WithL1dEntry_SansMdsEntry_SansIbpbExit %+ %2, 1, 0 | CPUMCTX_WSF_L1D_ENTRY | 0 | 0 , %1
1098hmR0VmxStartVmTemplate _SansXcr0_WithIbpbEntry_WithL1dEntry_SansMdsEntry_SansIbpbExit %+ %2, 0, CPUMCTX_WSF_IBPB_ENTRY | CPUMCTX_WSF_L1D_ENTRY | 0 | 0 , %1
1099hmR0VmxStartVmTemplate _WithXcr0_WithIbpbEntry_WithL1dEntry_SansMdsEntry_SansIbpbExit %+ %2, 1, CPUMCTX_WSF_IBPB_ENTRY | CPUMCTX_WSF_L1D_ENTRY | 0 | 0 , %1
1100hmR0VmxStartVmTemplate _SansXcr0_SansIbpbEntry_SansL1dEntry_WithMdsEntry_SansIbpbExit %+ %2, 0, 0 | 0 | CPUMCTX_WSF_MDS_ENTRY | 0 , %1
1101hmR0VmxStartVmTemplate _WithXcr0_SansIbpbEntry_SansL1dEntry_WithMdsEntry_SansIbpbExit %+ %2, 1, 0 | 0 | CPUMCTX_WSF_MDS_ENTRY | 0 , %1
1102hmR0VmxStartVmTemplate _SansXcr0_WithIbpbEntry_SansL1dEntry_WithMdsEntry_SansIbpbExit %+ %2, 0, CPUMCTX_WSF_IBPB_ENTRY | 0 | CPUMCTX_WSF_MDS_ENTRY | 0 , %1
1103hmR0VmxStartVmTemplate _WithXcr0_WithIbpbEntry_SansL1dEntry_WithMdsEntry_SansIbpbExit %+ %2, 1, CPUMCTX_WSF_IBPB_ENTRY | 0 | CPUMCTX_WSF_MDS_ENTRY | 0 , %1
1104hmR0VmxStartVmTemplate _SansXcr0_SansIbpbEntry_WithL1dEntry_WithMdsEntry_SansIbpbExit %+ %2, 0, 0 | CPUMCTX_WSF_L1D_ENTRY | CPUMCTX_WSF_MDS_ENTRY | 0 , %1
1105hmR0VmxStartVmTemplate _WithXcr0_SansIbpbEntry_WithL1dEntry_WithMdsEntry_SansIbpbExit %+ %2, 1, 0 | CPUMCTX_WSF_L1D_ENTRY | CPUMCTX_WSF_MDS_ENTRY | 0 , %1
1106hmR0VmxStartVmTemplate _SansXcr0_WithIbpbEntry_WithL1dEntry_WithMdsEntry_SansIbpbExit %+ %2, 0, CPUMCTX_WSF_IBPB_ENTRY | CPUMCTX_WSF_L1D_ENTRY | CPUMCTX_WSF_MDS_ENTRY | 0 , %1
1107hmR0VmxStartVmTemplate _WithXcr0_WithIbpbEntry_WithL1dEntry_WithMdsEntry_SansIbpbExit %+ %2, 1, CPUMCTX_WSF_IBPB_ENTRY | CPUMCTX_WSF_L1D_ENTRY | CPUMCTX_WSF_MDS_ENTRY | 0 , %1
1108hmR0VmxStartVmTemplate _SansXcr0_SansIbpbEntry_SansL1dEntry_SansMdsEntry_WithIbpbExit %+ %2, 0, 0 | 0 | 0 | CPUMCTX_WSF_IBPB_EXIT, %1
1109hmR0VmxStartVmTemplate _WithXcr0_SansIbpbEntry_SansL1dEntry_SansMdsEntry_WithIbpbExit %+ %2, 1, 0 | 0 | 0 | CPUMCTX_WSF_IBPB_EXIT, %1
1110hmR0VmxStartVmTemplate _SansXcr0_WithIbpbEntry_SansL1dEntry_SansMdsEntry_WithIbpbExit %+ %2, 0, CPUMCTX_WSF_IBPB_ENTRY | 0 | 0 | CPUMCTX_WSF_IBPB_EXIT, %1
1111hmR0VmxStartVmTemplate _WithXcr0_WithIbpbEntry_SansL1dEntry_SansMdsEntry_WithIbpbExit %+ %2, 1, CPUMCTX_WSF_IBPB_ENTRY | 0 | 0 | CPUMCTX_WSF_IBPB_EXIT, %1
1112hmR0VmxStartVmTemplate _SansXcr0_SansIbpbEntry_WithL1dEntry_SansMdsEntry_WithIbpbExit %+ %2, 0, 0 | CPUMCTX_WSF_L1D_ENTRY | 0 | CPUMCTX_WSF_IBPB_EXIT, %1
1113hmR0VmxStartVmTemplate _WithXcr0_SansIbpbEntry_WithL1dEntry_SansMdsEntry_WithIbpbExit %+ %2, 1, 0 | CPUMCTX_WSF_L1D_ENTRY | 0 | CPUMCTX_WSF_IBPB_EXIT, %1
1114hmR0VmxStartVmTemplate _SansXcr0_WithIbpbEntry_WithL1dEntry_SansMdsEntry_WithIbpbExit %+ %2, 0, CPUMCTX_WSF_IBPB_ENTRY | CPUMCTX_WSF_L1D_ENTRY | 0 | CPUMCTX_WSF_IBPB_EXIT, %1
1115hmR0VmxStartVmTemplate _WithXcr0_WithIbpbEntry_WithL1dEntry_SansMdsEntry_WithIbpbExit %+ %2, 1, CPUMCTX_WSF_IBPB_ENTRY | CPUMCTX_WSF_L1D_ENTRY | 0 | CPUMCTX_WSF_IBPB_EXIT, %1
1116hmR0VmxStartVmTemplate _SansXcr0_SansIbpbEntry_SansL1dEntry_WithMdsEntry_WithIbpbExit %+ %2, 0, 0 | 0 | CPUMCTX_WSF_MDS_ENTRY | CPUMCTX_WSF_IBPB_EXIT, %1
1117hmR0VmxStartVmTemplate _WithXcr0_SansIbpbEntry_SansL1dEntry_WithMdsEntry_WithIbpbExit %+ %2, 1, 0 | 0 | CPUMCTX_WSF_MDS_ENTRY | CPUMCTX_WSF_IBPB_EXIT, %1
1118hmR0VmxStartVmTemplate _SansXcr0_WithIbpbEntry_SansL1dEntry_WithMdsEntry_WithIbpbExit %+ %2, 0, CPUMCTX_WSF_IBPB_ENTRY | 0 | CPUMCTX_WSF_MDS_ENTRY | CPUMCTX_WSF_IBPB_EXIT, %1
1119hmR0VmxStartVmTemplate _WithXcr0_WithIbpbEntry_SansL1dEntry_WithMdsEntry_WithIbpbExit %+ %2, 1, CPUMCTX_WSF_IBPB_ENTRY | 0 | CPUMCTX_WSF_MDS_ENTRY | CPUMCTX_WSF_IBPB_EXIT, %1
1120hmR0VmxStartVmTemplate _SansXcr0_SansIbpbEntry_WithL1dEntry_WithMdsEntry_WithIbpbExit %+ %2, 0, 0 | CPUMCTX_WSF_L1D_ENTRY | CPUMCTX_WSF_MDS_ENTRY | CPUMCTX_WSF_IBPB_EXIT, %1
1121hmR0VmxStartVmTemplate _WithXcr0_SansIbpbEntry_WithL1dEntry_WithMdsEntry_WithIbpbExit %+ %2, 1, 0 | CPUMCTX_WSF_L1D_ENTRY | CPUMCTX_WSF_MDS_ENTRY | CPUMCTX_WSF_IBPB_EXIT, %1
1122hmR0VmxStartVmTemplate _SansXcr0_WithIbpbEntry_WithL1dEntry_WithMdsEntry_WithIbpbExit %+ %2, 0, CPUMCTX_WSF_IBPB_ENTRY | CPUMCTX_WSF_L1D_ENTRY | CPUMCTX_WSF_MDS_ENTRY | CPUMCTX_WSF_IBPB_EXIT, %1
1123hmR0VmxStartVmTemplate _WithXcr0_WithIbpbEntry_WithL1dEntry_WithMdsEntry_WithIbpbExit %+ %2, 1, CPUMCTX_WSF_IBPB_ENTRY | CPUMCTX_WSF_L1D_ENTRY | CPUMCTX_WSF_MDS_ENTRY | CPUMCTX_WSF_IBPB_EXIT, %1
1124%endmacro
1125
1126hmR0VmxStartVmSseTemplate 0,,RT_NOTHING
1127%ifdef VBOX_WITH_KERNEL_USING_XMM
1128hmR0VmxStartVmSseTemplate 1,_SseManual,RT_NOTHING
1129hmR0VmxStartVmSseTemplate 2,_SseXSave,RT_NOTHING
1130%endif
1131
1132
1133;;
1134; Clears the MDS buffers using VERW.
1135ALIGNCODE(16)
1136BEGINPROC hmR0MdsClear
1137 SEH64_END_PROLOGUE
1138 sub xSP, xCB
1139 mov [xSP], ds
1140 verw [xSP]
1141 add xSP, xCB
1142 ret
1143ENDPROC hmR0MdsClear
1144
1145
1146;;
1147; hmR0SvmVmRun template
1148;
1149; @param 1 The suffix of the variation.
1150; @param 2 fLoadSaveGuestXcr0 value
1151; @param 3 The CPUMCTX_WSF_IBPB_ENTRY + CPUMCTX_WSF_IBPB_EXIT value.
1152; @param 4 The SSE saving/restoring: 0 to do nothing, 1 to do it manually, 2 to use xsave/xrstor.
1153; Drivers shouldn't use AVX registers without saving+loading:
1154; https://msdn.microsoft.com/en-us/library/windows/hardware/ff545910%28v=vs.85%29.aspx?f=255&MSPPError=-2147217396
1155; However the compiler docs have different idea:
1156; https://msdn.microsoft.com/en-us/library/9z1stfyw.aspx
1157; We'll go with the former for now.
1158;
1159%macro hmR0SvmVmRunTemplate 4
1160
1161;;
1162; Prepares for and executes VMRUN (32-bit and 64-bit guests).
1163;
1164; @returns VBox status code
1165; @param pVM msc:rcx,gcc:rdi The cross context VM structure (unused).
1166; @param pVCpu msc:rdx,gcc:rsi The cross context virtual CPU structure of the calling EMT.
1167; @param HCPhysVmcb msc:r8, gcc:rdx Physical address of guest VMCB.
1168;
1169ALIGNCODE(64) ; This + immediate optimizations causes serious trouble for yasm and the SEH frames: prologue -28 bytes, must be <256
1170 ; So the SEH64_XXX stuff is currently not operational.
1171BEGINPROC RT_CONCAT(hmR0SvmVmRun,%1)
1172 %ifdef VBOX_WITH_KERNEL_USING_XMM
1173 %if %4 = 0
1174 ;
1175 ; The non-saving variant will currently check the two SSE preconditions and pick
1176 ; the right variant to continue with. Later we can see if we can't manage to
1177 ; move these decisions into hmR0SvmUpdateVmRunFunction().
1178 ;
1179 %ifdef ASM_CALL64_MSC
1180 test byte [rdx + VMCPU.cpum.GstCtx + CPUMCTX.fUsedFpuGuest], 1
1181 %else
1182 test byte [rsi + VMCPU.cpum.GstCtx + CPUMCTX.fUsedFpuGuest], 1
1183 %endif
1184 jz .save_xmm_no_need
1185 %ifdef ASM_CALL64_MSC
1186 cmp dword [rdx + VMCPU.cpum.GstCtx + CPUMCTX.fXStateMask], 0
1187 %else
1188 cmp dword [rsi + VMCPU.cpum.GstCtx + CPUMCTX.fXStateMask], 0
1189 %endif
1190 je RT_CONCAT3(hmR0SvmVmRun,%1,_SseManual)
1191 jmp RT_CONCAT3(hmR0SvmVmRun,%1,_SseXSave)
1192.save_xmm_no_need:
1193 %endif
1194 %endif
1195 push rbp
1196 SEH64_PUSH_xBP
1197 mov rbp, rsp
1198 SEH64_SET_FRAME_xBP 0
1199 pushf
1200 %assign cbFrame 30h
1201 %if %4 != 0
1202 %assign cbFrame cbFrame + 16 * 11 ; Reserve space for 10x 128-bit XMM registers and MXCSR (32-bit)
1203 %endif
1204 %assign cbBaseFrame cbFrame
1205 sub rsp, cbFrame - 8h ; We subtract 8 bytes for the above pushf
1206 SEH64_ALLOCATE_STACK cbFrame ; And we have CALLEE_PRESERVED_REGISTER_COUNT following it.
1207
1208 %define frm_fRFlags -008h
1209 %define frm_uHostXcr0 -018h ; 128-bit
1210 ;%define frm_fNoRestoreXcr0 -020h ; Non-zero if we should skip XCR0 restoring.
1211 %define frm_pGstCtx -028h ; Where we stash guest CPU context for use after the vmrun.
1212 %define frm_HCPhysVmcbHost -030h ; Where we stash HCPhysVmcbHost for the vmload after vmrun.
1213 %if %4 != 0
1214 %define frm_saved_xmm6 -040h
1215 %define frm_saved_xmm7 -050h
1216 %define frm_saved_xmm8 -060h
1217 %define frm_saved_xmm9 -070h
1218 %define frm_saved_xmm10 -080h
1219 %define frm_saved_xmm11 -090h
1220 %define frm_saved_xmm12 -0a0h
1221 %define frm_saved_xmm13 -0b0h
1222 %define frm_saved_xmm14 -0c0h
1223 %define frm_saved_xmm15 -0d0h
1224 %define frm_saved_mxcsr -0e0h
1225 %endif
1226
1227 ; Manual save and restore:
1228 ; - General purpose registers except RIP, RSP, RAX
1229 ;
1230 ; Trashed:
1231 ; - CR2 (we don't care)
1232 ; - LDTR (reset to 0)
1233 ; - DRx (presumably not changed at all)
1234 ; - DR7 (reset to 0x400)
1235
1236 ; Save all general purpose host registers.
1237 PUSH_CALLEE_PRESERVED_REGISTERS
1238 SEH64_END_PROLOGUE
1239 %if cbFrame != (cbBaseFrame + 8 * CALLEE_PRESERVED_REGISTER_COUNT)
1240 %error Bad cbFrame value
1241 %endif
1242
1243 ; Shuffle parameter registers so that r8=HCPhysVmcb and rsi=pVCpu. (rdx & rcx will soon be trashed.)
1244 %ifdef ASM_CALL64_GCC
1245 mov r8, rdx ; Put HCPhysVmcb in r8 like on MSC as rdx is trashed below.
1246 %else
1247 mov rsi, rdx ; Put pVCpu in rsi like on GCC as rdx is trashed below.
1248 ;mov rdi, rcx ; Put pVM in rdi like on GCC as rcx is trashed below.
1249 %endif
1250
1251 %ifdef VBOX_STRICT
1252 ;
1253 ; Verify template preconditions / parameters to ensure HMSVM.cpp didn't miss some state change.
1254 ;
1255 cmp byte [rsi + GVMCPU.hmr0 + HMR0PERVCPU.fLoadSaveGuestXcr0], %2
1256 mov eax, VERR_SVM_VMRUN_PRECOND_0
1257 jne .failure_return
1258
1259 mov eax, [rsi + VMCPU.cpum.GstCtx + CPUMCTX.fWorldSwitcher]
1260 and eax, CPUMCTX_WSF_IBPB_ENTRY | CPUMCTX_WSF_IBPB_EXIT
1261 cmp eax, %3
1262 mov eax, VERR_SVM_VMRUN_PRECOND_1
1263 jne .failure_return
1264
1265 %ifdef VBOX_WITH_KERNEL_USING_XMM
1266 mov eax, VERR_SVM_VMRUN_PRECOND_2
1267 test byte [rsi + VMCPU.cpum.GstCtx + CPUMCTX.fUsedFpuGuest], 1
1268 %if %4 = 0
1269 jnz .failure_return
1270 %else
1271 jz .failure_return
1272
1273 mov eax, VERR_SVM_VMRUN_PRECOND_3
1274 cmp dword [rsi + VMCPU.cpum.GstCtx + CPUMCTX.fXStateMask], 0
1275 %if %4 = 1
1276 jne .failure_return
1277 %elif %4 = 2
1278 je .failure_return
1279 %else
1280 %error Invalid template parameter 4.
1281 %endif
1282 %endif
1283 %endif
1284 %endif ; VBOX_STRICT
1285
1286 %if %4 != 0
1287 ; Save the non-volatile SSE host register state.
1288 movdqa [rbp + frm_saved_xmm6 ], xmm6
1289 movdqa [rbp + frm_saved_xmm7 ], xmm7
1290 movdqa [rbp + frm_saved_xmm8 ], xmm8
1291 movdqa [rbp + frm_saved_xmm9 ], xmm9
1292 movdqa [rbp + frm_saved_xmm10], xmm10
1293 movdqa [rbp + frm_saved_xmm11], xmm11
1294 movdqa [rbp + frm_saved_xmm12], xmm12
1295 movdqa [rbp + frm_saved_xmm13], xmm13
1296 movdqa [rbp + frm_saved_xmm14], xmm14
1297 movdqa [rbp + frm_saved_xmm15], xmm15
1298 stmxcsr [rbp + frm_saved_mxcsr]
1299
1300 ; Load the guest state related to the above non-volatile and volatile SSE registers. Trashes rcx, eax and edx.
1301 mov rcx, [rsi + VMCPU.cpum.GstCtx + CPUMCTX.pXStateR0]
1302 %if %4 = 1 ; manual
1303 movdqa xmm0, [rcx + XMM_OFF_IN_X86FXSTATE + 000h]
1304 movdqa xmm1, [rcx + XMM_OFF_IN_X86FXSTATE + 010h]
1305 movdqa xmm2, [rcx + XMM_OFF_IN_X86FXSTATE + 020h]
1306 movdqa xmm3, [rcx + XMM_OFF_IN_X86FXSTATE + 030h]
1307 movdqa xmm4, [rcx + XMM_OFF_IN_X86FXSTATE + 040h]
1308 movdqa xmm5, [rcx + XMM_OFF_IN_X86FXSTATE + 050h]
1309 movdqa xmm6, [rcx + XMM_OFF_IN_X86FXSTATE + 060h]
1310 movdqa xmm7, [rcx + XMM_OFF_IN_X86FXSTATE + 070h]
1311 movdqa xmm8, [rcx + XMM_OFF_IN_X86FXSTATE + 080h]
1312 movdqa xmm9, [rcx + XMM_OFF_IN_X86FXSTATE + 090h]
1313 movdqa xmm10, [rcx + XMM_OFF_IN_X86FXSTATE + 0a0h]
1314 movdqa xmm11, [rcx + XMM_OFF_IN_X86FXSTATE + 0b0h]
1315 movdqa xmm12, [rcx + XMM_OFF_IN_X86FXSTATE + 0c0h]
1316 movdqa xmm13, [rcx + XMM_OFF_IN_X86FXSTATE + 0d0h]
1317 movdqa xmm14, [rcx + XMM_OFF_IN_X86FXSTATE + 0e0h]
1318 movdqa xmm15, [rcx + XMM_OFF_IN_X86FXSTATE + 0f0h]
1319 ldmxcsr [rcx + X86FXSTATE.MXCSR]
1320 %elif %4 = 2 ; use xrstor/xsave
1321 mov eax, [rsi + VMCPU.cpum.GstCtx + CPUMCTX.fXStateMask]
1322 and eax, CPUM_VOLATILE_XSAVE_GUEST_COMPONENTS
1323 xor edx, edx
1324 xrstor [rcx]
1325 %else
1326 %error invalid template parameter 4
1327 %endif
1328 %endif
1329
1330 %if %2 != 0
1331 ; Save the host XCR0 and load the guest one if necessary.
1332 xor ecx, ecx
1333 xgetbv ; save the host XCR0 on the stack
1334 mov [rbp + frm_uHostXcr0 + 8], rdx
1335 mov [rbp + frm_uHostXcr0 ], rax
1336
1337 mov eax, [rsi + VMCPU.cpum.GstCtx + CPUMCTX.aXcr] ; load the guest XCR0
1338 mov edx, [rsi + VMCPU.cpum.GstCtx + CPUMCTX.aXcr + 4]
1339 xor ecx, ecx ; paranoia
1340 xsetbv
1341 %endif
1342
1343 ; Save host fs, gs, sysenter msr etc.
1344 mov rax, [rsi + VMCPU.hm + HMCPU.u + HMCPUSVM.HCPhysVmcbHost]
1345 mov qword [rbp + frm_HCPhysVmcbHost], rax ; save for the vmload after vmrun
1346 lea rsi, [rsi + VMCPU.cpum.GstCtx]
1347 mov qword [rbp + frm_pGstCtx], rsi
1348 vmsave
1349
1350 %if %3 & CPUMCTX_WSF_IBPB_ENTRY
1351 ; Fight spectre (trashes rax, rdx and rcx).
1352 mov ecx, MSR_IA32_PRED_CMD
1353 mov eax, MSR_IA32_PRED_CMD_F_IBPB
1354 xor edx, edx
1355 wrmsr
1356 %endif
1357
1358 ; Setup rax for VMLOAD.
1359 mov rax, r8 ; HCPhysVmcb (64 bits physical address; take low dword only)
1360
1361 ; Load guest general purpose registers (rax is loaded from the VMCB by VMRUN).
1362 mov rbx, qword [rsi + CPUMCTX.ebx]
1363 mov rcx, qword [rsi + CPUMCTX.ecx]
1364 mov rdx, qword [rsi + CPUMCTX.edx]
1365 mov rdi, qword [rsi + CPUMCTX.edi]
1366 mov rbp, qword [rsi + CPUMCTX.ebp]
1367 mov r8, qword [rsi + CPUMCTX.r8]
1368 mov r9, qword [rsi + CPUMCTX.r9]
1369 mov r10, qword [rsi + CPUMCTX.r10]
1370 mov r11, qword [rsi + CPUMCTX.r11]
1371 mov r12, qword [rsi + CPUMCTX.r12]
1372 mov r13, qword [rsi + CPUMCTX.r13]
1373 mov r14, qword [rsi + CPUMCTX.r14]
1374 mov r15, qword [rsi + CPUMCTX.r15]
1375 mov rsi, qword [rsi + CPUMCTX.esi]
1376
1377 ; Clear the global interrupt flag & execute sti to make sure external interrupts cause a world switch.
1378 clgi
1379 sti
1380
1381 ; Load guest FS, GS, Sysenter MSRs etc.
1382 vmload
1383
1384 ; Run the VM.
1385 vmrun
1386
1387 ; Save guest fs, gs, sysenter msr etc.
1388 vmsave
1389
1390 ; Load host fs, gs, sysenter msr etc.
1391 mov rax, [rsp + cbFrame + frm_HCPhysVmcbHost] ; load HCPhysVmcbHost (rbp is not operational yet, thus rsp)
1392 vmload
1393
1394 ; Set the global interrupt flag again, but execute cli to make sure IF=0.
1395 cli
1396 stgi
1397
1398 ; Pop pVCpu (saved above) and save the guest GPRs (sans RSP and RAX).
1399 mov rax, [rsp + cbFrame + frm_pGstCtx] ; (rbp still not operational)
1400
1401 mov qword [rax + CPUMCTX.ebp], rbp
1402 lea rbp, [rsp + cbFrame]
1403 mov qword [rax + CPUMCTX.ecx], rcx
1404 mov rcx, SPECTRE_FILLER
1405 mov qword [rax + CPUMCTX.edx], rdx
1406 mov rdx, rcx
1407 mov qword [rax + CPUMCTX.r8], r8
1408 mov r8, rcx
1409 mov qword [rax + CPUMCTX.r9], r9
1410 mov r9, rcx
1411 mov qword [rax + CPUMCTX.r10], r10
1412 mov r10, rcx
1413 mov qword [rax + CPUMCTX.r11], r11
1414 mov r11, rcx
1415 mov qword [rax + CPUMCTX.edi], rdi
1416 %ifdef ASM_CALL64_MSC
1417 mov rdi, [rbp + frm_saved_rdi]
1418 %else
1419 mov rdi, rcx
1420 %endif
1421 mov qword [rax + CPUMCTX.esi], rsi
1422 %ifdef ASM_CALL64_MSC
1423 mov rsi, [rbp + frm_saved_rsi]
1424 %else
1425 mov rsi, rcx
1426 %endif
1427 mov qword [rax + CPUMCTX.ebx], rbx
1428 mov rbx, [rbp + frm_saved_rbx]
1429 mov qword [rax + CPUMCTX.r12], r12
1430 mov r12, [rbp + frm_saved_r12]
1431 mov qword [rax + CPUMCTX.r13], r13
1432 mov r13, [rbp + frm_saved_r13]
1433 mov qword [rax + CPUMCTX.r14], r14
1434 mov r14, [rbp + frm_saved_r14]
1435 mov qword [rax + CPUMCTX.r15], r15
1436 mov r15, [rbp + frm_saved_r15]
1437
1438 %if %4 != 0
1439 ; Set r8 = &pVCpu->cpum.GstCtx; for use below when saving and restoring SSE state.
1440 mov r8, rax
1441 %endif
1442
1443 %if %3 & CPUMCTX_WSF_IBPB_EXIT
1444 ; Fight spectre (trashes rax, rdx and rcx).
1445 mov ecx, MSR_IA32_PRED_CMD
1446 mov eax, MSR_IA32_PRED_CMD_F_IBPB
1447 xor edx, edx
1448 wrmsr
1449 %endif
1450
1451 %if %2 != 0
1452 ; Restore the host xcr0.
1453 xor ecx, ecx
1454 mov rdx, [rbp + frm_uHostXcr0 + 8]
1455 mov rax, [rbp + frm_uHostXcr0]
1456 xsetbv
1457 %endif
1458
1459 %if %4 != 0
1460 ; Save the guest SSE state related to non-volatile and volatile SSE registers.
1461 mov rcx, [r8 + CPUMCTX.pXStateR0]
1462 %if %4 = 1 ; manual
1463 stmxcsr [rcx + X86FXSTATE.MXCSR]
1464 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 000h], xmm0
1465 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 010h], xmm1
1466 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 020h], xmm2
1467 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 030h], xmm3
1468 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 040h], xmm4
1469 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 050h], xmm5
1470 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 060h], xmm6
1471 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 070h], xmm7
1472 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 080h], xmm8
1473 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 090h], xmm9
1474 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 0a0h], xmm10
1475 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 0b0h], xmm11
1476 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 0c0h], xmm12
1477 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 0d0h], xmm13
1478 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 0e0h], xmm14
1479 movdqa [rcx + XMM_OFF_IN_X86FXSTATE + 0f0h], xmm15
1480 %elif %4 = 2 ; use xrstor/xsave
1481 mov eax, [r8 + CPUMCTX.fXStateMask]
1482 and eax, CPUM_VOLATILE_XSAVE_GUEST_COMPONENTS
1483 xor edx, edx
1484 xsave [rcx]
1485 %else
1486 %error invalid template parameter 4
1487 %endif
1488
1489 ; Restore the host non-volatile SSE register state.
1490 ldmxcsr [rbp + frm_saved_mxcsr]
1491 movdqa xmm6, [rbp + frm_saved_xmm6 ]
1492 movdqa xmm7, [rbp + frm_saved_xmm7 ]
1493 movdqa xmm8, [rbp + frm_saved_xmm8 ]
1494 movdqa xmm9, [rbp + frm_saved_xmm9 ]
1495 movdqa xmm10, [rbp + frm_saved_xmm10]
1496 movdqa xmm11, [rbp + frm_saved_xmm11]
1497 movdqa xmm12, [rbp + frm_saved_xmm12]
1498 movdqa xmm13, [rbp + frm_saved_xmm13]
1499 movdqa xmm14, [rbp + frm_saved_xmm14]
1500 movdqa xmm15, [rbp + frm_saved_xmm15]
1501 %endif ; %4 != 0
1502
1503 ; Epilogue (assumes we restored volatile registers above when saving the guest GPRs).
1504 mov eax, VINF_SUCCESS
1505 add rsp, cbFrame - 8h
1506 popf
1507 leave
1508 ret
1509
1510 %ifdef VBOX_STRICT
1511 ; Precondition checks failed.
1512.failure_return:
1513 POP_CALLEE_PRESERVED_REGISTERS
1514 %if cbFrame != cbBaseFrame
1515 %error Bad frame size value: cbFrame
1516 %endif
1517 add rsp, cbFrame - 8h
1518 popf
1519 leave
1520 ret
1521 %endif
1522
1523%undef frm_uHostXcr0
1524%undef frm_fNoRestoreXcr0
1525%undef frm_pVCpu
1526%undef frm_HCPhysVmcbHost
1527%undef cbFrame
1528ENDPROC RT_CONCAT(hmR0SvmVmRun,%1)
1529
1530%endmacro ; hmR0SvmVmRunTemplate
1531
1532;
1533; Instantiate the hmR0SvmVmRun various variations.
1534;
1535hmR0SvmVmRunTemplate _SansXcr0_SansIbpbEntry_SansIbpbExit, 0, 0, 0
1536hmR0SvmVmRunTemplate _WithXcr0_SansIbpbEntry_SansIbpbExit, 1, 0, 0
1537hmR0SvmVmRunTemplate _SansXcr0_WithIbpbEntry_SansIbpbExit, 0, CPUMCTX_WSF_IBPB_ENTRY, 0
1538hmR0SvmVmRunTemplate _WithXcr0_WithIbpbEntry_SansIbpbExit, 1, CPUMCTX_WSF_IBPB_ENTRY, 0
1539hmR0SvmVmRunTemplate _SansXcr0_SansIbpbEntry_WithIbpbExit, 0, CPUMCTX_WSF_IBPB_EXIT, 0
1540hmR0SvmVmRunTemplate _WithXcr0_SansIbpbEntry_WithIbpbExit, 1, CPUMCTX_WSF_IBPB_EXIT, 0
1541hmR0SvmVmRunTemplate _SansXcr0_WithIbpbEntry_WithIbpbExit, 0, CPUMCTX_WSF_IBPB_ENTRY | CPUMCTX_WSF_IBPB_EXIT, 0
1542hmR0SvmVmRunTemplate _WithXcr0_WithIbpbEntry_WithIbpbExit, 1, CPUMCTX_WSF_IBPB_ENTRY | CPUMCTX_WSF_IBPB_EXIT, 0
1543%ifdef VBOX_WITH_KERNEL_USING_XMM
1544hmR0SvmVmRunTemplate _SansXcr0_SansIbpbEntry_SansIbpbExit_SseManual, 0, 0, 1
1545hmR0SvmVmRunTemplate _WithXcr0_SansIbpbEntry_SansIbpbExit_SseManual, 1, 0, 1
1546hmR0SvmVmRunTemplate _SansXcr0_WithIbpbEntry_SansIbpbExit_SseManual, 0, CPUMCTX_WSF_IBPB_ENTRY, 1
1547hmR0SvmVmRunTemplate _WithXcr0_WithIbpbEntry_SansIbpbExit_SseManual, 1, CPUMCTX_WSF_IBPB_ENTRY, 1
1548hmR0SvmVmRunTemplate _SansXcr0_SansIbpbEntry_WithIbpbExit_SseManual, 0, CPUMCTX_WSF_IBPB_EXIT, 1
1549hmR0SvmVmRunTemplate _WithXcr0_SansIbpbEntry_WithIbpbExit_SseManual, 1, CPUMCTX_WSF_IBPB_EXIT, 1
1550hmR0SvmVmRunTemplate _SansXcr0_WithIbpbEntry_WithIbpbExit_SseManual, 0, CPUMCTX_WSF_IBPB_ENTRY | CPUMCTX_WSF_IBPB_EXIT, 1
1551hmR0SvmVmRunTemplate _WithXcr0_WithIbpbEntry_WithIbpbExit_SseManual, 1, CPUMCTX_WSF_IBPB_ENTRY | CPUMCTX_WSF_IBPB_EXIT, 1
1552
1553hmR0SvmVmRunTemplate _SansXcr0_SansIbpbEntry_SansIbpbExit_SseXSave, 0, 0, 2
1554hmR0SvmVmRunTemplate _WithXcr0_SansIbpbEntry_SansIbpbExit_SseXSave, 1, 0, 2
1555hmR0SvmVmRunTemplate _SansXcr0_WithIbpbEntry_SansIbpbExit_SseXSave, 0, CPUMCTX_WSF_IBPB_ENTRY, 2
1556hmR0SvmVmRunTemplate _WithXcr0_WithIbpbEntry_SansIbpbExit_SseXSave, 1, CPUMCTX_WSF_IBPB_ENTRY, 2
1557hmR0SvmVmRunTemplate _SansXcr0_SansIbpbEntry_WithIbpbExit_SseXSave, 0, CPUMCTX_WSF_IBPB_EXIT, 2
1558hmR0SvmVmRunTemplate _WithXcr0_SansIbpbEntry_WithIbpbExit_SseXSave, 1, CPUMCTX_WSF_IBPB_EXIT, 2
1559hmR0SvmVmRunTemplate _SansXcr0_WithIbpbEntry_WithIbpbExit_SseXSave, 0, CPUMCTX_WSF_IBPB_ENTRY | CPUMCTX_WSF_IBPB_EXIT, 2
1560hmR0SvmVmRunTemplate _WithXcr0_WithIbpbEntry_WithIbpbExit_SseXSave, 1, CPUMCTX_WSF_IBPB_ENTRY | CPUMCTX_WSF_IBPB_EXIT, 2
1561%endif
1562
Note: See TracBrowser for help on using the repository browser.

© 2025 Oracle Support Privacy / Do Not Sell My Info Terms of Use Trademark Policy Automated Access Etiquette