xref: /illumos-kvm-cmd/qemu-tech.texi (revision 68396ea9)
1\input texinfo @c -*- texinfo -*-
2@c %**start of header
3@setfilename qemu-tech.info
4
5@documentlanguage en
6@documentencoding UTF-8
7
8@settitle QEMU Internals
9@exampleindent 0
10@paragraphindent 0
11@c %**end of header
12
13@ifinfo
14@direntry
15* QEMU Internals: (qemu-tech).   The QEMU Emulator Internals.
16@end direntry
17@end ifinfo
18
19@iftex
20@titlepage
21@sp 7
22@center @titlefont{QEMU Internals}
23@sp 3
24@end titlepage
25@end iftex
26
27@ifnottex
28@node Top
29@top
30
31@menu
32* Introduction::
33* QEMU Internals::
34* Regression Tests::
35* Index::
36@end menu
37@end ifnottex
38
39@contents
40
41@node Introduction
42@chapter Introduction
43
44@menu
45* intro_features::        Features
46* intro_x86_emulation::   x86 and x86-64 emulation
47* intro_arm_emulation::   ARM emulation
48* intro_mips_emulation::  MIPS emulation
49* intro_ppc_emulation::   PowerPC emulation
50* intro_sparc_emulation:: Sparc32 and Sparc64 emulation
51* intro_other_emulation:: Other CPU emulation
52@end menu
53
54@node intro_features
55@section Features
56
57QEMU is a FAST! processor emulator using a portable dynamic
58translator.
59
60QEMU has two operating modes:
61
62@itemize @minus
63
64@item
65Full system emulation. In this mode (full platform virtualization),
66QEMU emulates a full system (usually a PC), including a processor and
67various peripherals. It can be used to launch several different
68Operating Systems at once without rebooting the host machine or to
69debug system code.
70
71@item
72User mode emulation. In this mode (application level virtualization),
73QEMU can launch processes compiled for one CPU on another CPU, however
74the Operating Systems must match. This can be used for example to ease
75cross-compilation and cross-debugging.
76@end itemize
77
78As QEMU requires no host kernel driver to run, it is very safe and
79easy to use.
80
81QEMU generic features:
82
83@itemize
84
85@item User space only or full system emulation.
86
87@item Using dynamic translation to native code for reasonable speed.
88
89@item
90Working on x86, x86_64 and PowerPC32/64 hosts. Being tested on ARM,
91HPPA, Sparc32 and Sparc64. Previous versions had some support for
92Alpha and S390 hosts, but TCG (see below) doesn't support those yet.
93
94@item Self-modifying code support.
95
96@item Precise exceptions support.
97
98@item The virtual CPU is a library (@code{libqemu}) which can be used
99in other projects (look at @file{qemu/tests/qruncom.c} to have an
100example of user mode @code{libqemu} usage).
101
102@item
103Floating point library supporting both full software emulation and
104native host FPU instructions.
105
106@end itemize
107
108QEMU user mode emulation features:
109@itemize
110@item Generic Linux system call converter, including most ioctls.
111
112@item clone() emulation using native CPU clone() to use Linux scheduler for threads.
113
114@item Accurate signal handling by remapping host signals to target signals.
115@end itemize
116
117Linux user emulator (Linux host only) can be used to launch the Wine
118Windows API emulator (@url{http://www.winehq.org}). A Darwin user
119emulator (Darwin hosts only) exists and a BSD user emulator for BSD
120hosts is under development. It would also be possible to develop a
121similar user emulator for Solaris.
122
123QEMU full system emulation features:
124@itemize
125@item
126QEMU uses a full software MMU for maximum portability.
127
128@item
129QEMU can optionally use an in-kernel accelerator, like kvm. The accelerators
130execute some of the guest code natively, while
131continuing to emulate the rest of the machine.
132
133@item
134Various hardware devices can be emulated and in some cases, host
135devices (e.g. serial and parallel ports, USB, drives) can be used
136transparently by the guest Operating System. Host device passthrough
137can be used for talking to external physical peripherals (e.g. a
138webcam, modem or tape drive).
139
140@item
141Symmetric multiprocessing (SMP) even on a host with a single CPU. On a
142SMP host system, QEMU can use only one CPU fully due to difficulty in
143implementing atomic memory accesses efficiently.
144
145@end itemize
146
147@node intro_x86_emulation
148@section x86 and x86-64 emulation
149
150QEMU x86 target features:
151
152@itemize
153
154@item The virtual x86 CPU supports 16 bit and 32 bit addressing with segmentation.
155LDT/GDT and IDT are emulated. VM86 mode is also supported to run
156DOSEMU. There is some support for MMX/3DNow!, SSE, SSE2, SSE3, SSSE3,
157and SSE4 as well as x86-64 SVM.
158
159@item Support of host page sizes bigger than 4KB in user mode emulation.
160
161@item QEMU can emulate itself on x86.
162
163@item An extensive Linux x86 CPU test program is included @file{tests/test-i386}.
164It can be used to test other x86 virtual CPUs.
165
166@end itemize
167
168Current QEMU limitations:
169
170@itemize
171
172@item Limited x86-64 support.
173
174@item IPC syscalls are missing.
175
176@item The x86 segment limits and access rights are not tested at every
177memory access (yet). Hopefully, very few OSes seem to rely on that for
178normal use.
179
180@end itemize
181
182@node intro_arm_emulation
183@section ARM emulation
184
185@itemize
186
187@item Full ARM 7 user emulation.
188
189@item NWFPE FPU support included in user Linux emulation.
190
191@item Can run most ARM Linux binaries.
192
193@end itemize
194
195@node intro_mips_emulation
196@section MIPS emulation
197
198@itemize
199
200@item The system emulation allows full MIPS32/MIPS64 Release 2 emulation,
201including privileged instructions, FPU and MMU, in both little and big
202endian modes.
203
204@item The Linux userland emulation can run many 32 bit MIPS Linux binaries.
205
206@end itemize
207
208Current QEMU limitations:
209
210@itemize
211
212@item Self-modifying code is not always handled correctly.
213
214@item 64 bit userland emulation is not implemented.
215
216@item The system emulation is not complete enough to run real firmware.
217
218@item The watchpoint debug facility is not implemented.
219
220@end itemize
221
222@node intro_ppc_emulation
223@section PowerPC emulation
224
225@itemize
226
227@item Full PowerPC 32 bit emulation, including privileged instructions,
228FPU and MMU.
229
230@item Can run most PowerPC Linux binaries.
231
232@end itemize
233
234@node intro_sparc_emulation
235@section Sparc32 and Sparc64 emulation
236
237@itemize
238
239@item Full SPARC V8 emulation, including privileged
240instructions, FPU and MMU. SPARC V9 emulation includes most privileged
241and VIS instructions, FPU and I/D MMU. Alignment is fully enforced.
242
243@item Can run most 32-bit SPARC Linux binaries, SPARC32PLUS Linux binaries and
244some 64-bit SPARC Linux binaries.
245
246@end itemize
247
248Current QEMU limitations:
249
250@itemize
251
252@item IPC syscalls are missing.
253
254@item Floating point exception support is buggy.
255
256@item Atomic instructions are not correctly implemented.
257
258@item There are still some problems with Sparc64 emulators.
259
260@end itemize
261
262@node intro_other_emulation
263@section Other CPU emulation
264
265In addition to the above, QEMU supports emulation of other CPUs with
266varying levels of success. These are:
267
268@itemize
269
270@item
271Alpha
272@item
273CRIS
274@item
275M68k
276@item
277SH4
278@end itemize
279
280@node QEMU Internals
281@chapter QEMU Internals
282
283@menu
284* QEMU compared to other emulators::
285* Portable dynamic translation::
286* Condition code optimisations::
287* CPU state optimisations::
288* Translation cache::
289* Direct block chaining::
290* Self-modifying code and translated code invalidation::
291* Exception support::
292* MMU emulation::
293* Device emulation::
294* Hardware interrupts::
295* User emulation specific details::
296* Bibliography::
297@end menu
298
299@node QEMU compared to other emulators
300@section QEMU compared to other emulators
301
302Like bochs [3], QEMU emulates an x86 CPU. But QEMU is much faster than
303bochs as it uses dynamic compilation. Bochs is closely tied to x86 PC
304emulation while QEMU can emulate several processors.
305
306Like Valgrind [2], QEMU does user space emulation and dynamic
307translation. Valgrind is mainly a memory debugger while QEMU has no
308support for it (QEMU could be used to detect out of bound memory
309accesses as Valgrind, but it has no support to track uninitialised data
310as Valgrind does). The Valgrind dynamic translator generates better code
311than QEMU (in particular it does register allocation) but it is closely
312tied to an x86 host and target and has no support for precise exceptions
313and system emulation.
314
315EM86 [4] is the closest project to user space QEMU (and QEMU still uses
316some of its code, in particular the ELF file loader). EM86 was limited
317to an alpha host and used a proprietary and slow interpreter (the
318interpreter part of the FX!32 Digital Win32 code translator [5]).
319
320TWIN [6] is a Windows API emulator like Wine. It is less accurate than
321Wine but includes a protected mode x86 interpreter to launch x86 Windows
322executables. Such an approach has greater potential because most of the
323Windows API is executed natively but it is far more difficult to develop
324because all the data structures and function parameters exchanged
325between the API and the x86 code must be converted.
326
327User mode Linux [7] was the only solution before QEMU to launch a
328Linux kernel as a process while not needing any host kernel
329patches. However, user mode Linux requires heavy kernel patches while
330QEMU accepts unpatched Linux kernels. The price to pay is that QEMU is
331slower.
332
333The Plex86 [8] PC virtualizer is done in the same spirit as the now
334obsolete qemu-fast system emulator. It requires a patched Linux kernel
335to work (you cannot launch the same kernel on your PC), but the
336patches are really small. As it is a PC virtualizer (no emulation is
337done except for some privileged instructions), it has the potential of
338being faster than QEMU. The downside is that a complicated (and
339potentially unsafe) host kernel patch is needed.
340
341The commercial PC Virtualizers (VMWare [9], VirtualPC [10], TwoOStwo
342[11]) are faster than QEMU, but they all need specific, proprietary
343and potentially unsafe host drivers. Moreover, they are unable to
344provide cycle exact simulation as an emulator can.
345
346VirtualBox [12], Xen [13] and KVM [14] are based on QEMU. QEMU-SystemC
347[15] uses QEMU to simulate a system where some hardware devices are
348developed in SystemC.
349
350@node Portable dynamic translation
351@section Portable dynamic translation
352
353QEMU is a dynamic translator. When it first encounters a piece of code,
354it converts it to the host instruction set. Usually dynamic translators
355are very complicated and highly CPU dependent. QEMU uses some tricks
356which make it relatively easily portable and simple while achieving good
357performances.
358
359After the release of version 0.9.1, QEMU switched to a new method of
360generating code, Tiny Code Generator or TCG. TCG relaxes the
361dependency on the exact version of the compiler used. The basic idea
362is to split every target instruction into a couple of RISC-like TCG
363ops (see @code{target-i386/translate.c}). Some optimizations can be
364performed at this stage, including liveness analysis and trivial
365constant expression evaluation. TCG ops are then implemented in the
366host CPU back end, also known as TCG target (see
367@code{tcg/i386/tcg-target.c}). For more information, please take a
368look at @code{tcg/README}.
369
370@node Condition code optimisations
371@section Condition code optimisations
372
373Lazy evaluation of CPU condition codes (@code{EFLAGS} register on x86)
374is important for CPUs where every instruction sets the condition
375codes. It tends to be less important on conventional RISC systems
376where condition codes are only updated when explicitly requested. On
377Sparc64, costly update of both 32 and 64 bit condition codes can be
378avoided with lazy evaluation.
379
380Instead of computing the condition codes after each x86 instruction,
381QEMU just stores one operand (called @code{CC_SRC}), the result
382(called @code{CC_DST}) and the type of operation (called
383@code{CC_OP}). When the condition codes are needed, the condition
384codes can be calculated using this information. In addition, an
385optimized calculation can be performed for some instruction types like
386conditional branches.
387
388@code{CC_OP} is almost never explicitly set in the generated code
389because it is known at translation time.
390
391The lazy condition code evaluation is used on x86, m68k, cris and
392Sparc. ARM uses a simplified variant for the N and Z flags.
393
394@node CPU state optimisations
395@section CPU state optimisations
396
397The target CPUs have many internal states which change the way it
398evaluates instructions. In order to achieve a good speed, the
399translation phase considers that some state information of the virtual
400CPU cannot change in it. The state is recorded in the Translation
401Block (TB). If the state changes (e.g. privilege level), a new TB will
402be generated and the previous TB won't be used anymore until the state
403matches the state recorded in the previous TB. For example, if the SS,
404DS and ES segments have a zero base, then the translator does not even
405generate an addition for the segment base.
406
407[The FPU stack pointer register is not handled that way yet].
408
409@node Translation cache
410@section Translation cache
411
412A 16 MByte cache holds the most recently used translations. For
413simplicity, it is completely flushed when it is full. A translation unit
414contains just a single basic block (a block of x86 instructions
415terminated by a jump or by a virtual CPU state change which the
416translator cannot deduce statically).
417
418@node Direct block chaining
419@section Direct block chaining
420
421After each translated basic block is executed, QEMU uses the simulated
422Program Counter (PC) and other cpu state informations (such as the CS
423segment base value) to find the next basic block.
424
425In order to accelerate the most common cases where the new simulated PC
426is known, QEMU can patch a basic block so that it jumps directly to the
427next one.
428
429The most portable code uses an indirect jump. An indirect jump makes
430it easier to make the jump target modification atomic. On some host
431architectures (such as x86 or PowerPC), the @code{JUMP} opcode is
432directly patched so that the block chaining has no overhead.
433
434@node Self-modifying code and translated code invalidation
435@section Self-modifying code and translated code invalidation
436
437Self-modifying code is a special challenge in x86 emulation because no
438instruction cache invalidation is signaled by the application when code
439is modified.
440
441When translated code is generated for a basic block, the corresponding
442host page is write protected if it is not already read-only. Then, if
443a write access is done to the page, Linux raises a SEGV signal. QEMU
444then invalidates all the translated code in the page and enables write
445accesses to the page.
446
447Correct translated code invalidation is done efficiently by maintaining
448a linked list of every translated block contained in a given page. Other
449linked lists are also maintained to undo direct block chaining.
450
451On RISC targets, correctly written software uses memory barriers and
452cache flushes, so some of the protection above would not be
453necessary. However, QEMU still requires that the generated code always
454matches the target instructions in memory in order to handle
455exceptions correctly.
456
457@node Exception support
458@section Exception support
459
460longjmp() is used when an exception such as division by zero is
461encountered.
462
463The host SIGSEGV and SIGBUS signal handlers are used to get invalid
464memory accesses. The simulated program counter is found by
465retranslating the corresponding basic block and by looking where the
466host program counter was at the exception point.
467
468The virtual CPU cannot retrieve the exact @code{EFLAGS} register because
469in some cases it is not computed because of condition code
470optimisations. It is not a big concern because the emulated code can
471still be restarted in any cases.
472
473@node MMU emulation
474@section MMU emulation
475
476For system emulation QEMU supports a soft MMU. In that mode, the MMU
477virtual to physical address translation is done at every memory
478access. QEMU uses an address translation cache to speed up the
479translation.
480
481In order to avoid flushing the translated code each time the MMU
482mappings change, QEMU uses a physically indexed translation cache. It
483means that each basic block is indexed with its physical address.
484
485When MMU mappings change, only the chaining of the basic blocks is
486reset (i.e. a basic block can no longer jump directly to another one).
487
488@node Device emulation
489@section Device emulation
490
491Systems emulated by QEMU are organized by boards. At initialization
492phase, each board instantiates a number of CPUs, devices, RAM and
493ROM. Each device in turn can assign I/O ports or memory areas (for
494MMIO) to its handlers. When the emulation starts, an access to the
495ports or MMIO memory areas assigned to the device causes the
496corresponding handler to be called.
497
498RAM and ROM are handled more optimally, only the offset to the host
499memory needs to be added to the guest address.
500
501The video RAM of VGA and other display cards is special: it can be
502read or written directly like RAM, but write accesses cause the memory
503to be marked with VGA_DIRTY flag as well.
504
505QEMU supports some device classes like serial and parallel ports, USB,
506drives and network devices, by providing APIs for easier connection to
507the generic, higher level implementations. The API hides the
508implementation details from the devices, like native device use or
509advanced block device formats like QCOW.
510
511Usually the devices implement a reset method and register support for
512saving and loading of the device state. The devices can also use
513timers, especially together with the use of bottom halves (BHs).
514
515@node Hardware interrupts
516@section Hardware interrupts
517
518In order to be faster, QEMU does not check at every basic block if an
519hardware interrupt is pending. Instead, the user must asynchronously
520call a specific function to tell that an interrupt is pending. This
521function resets the chaining of the currently executing basic
522block. It ensures that the execution will return soon in the main loop
523of the CPU emulator. Then the main loop can test if the interrupt is
524pending and handle it.
525
526@node User emulation specific details
527@section User emulation specific details
528
529@subsection Linux system call translation
530
531QEMU includes a generic system call translator for Linux. It means that
532the parameters of the system calls can be converted to fix the
533endianness and 32/64 bit issues. The IOCTLs are converted with a generic
534type description system (see @file{ioctls.h} and @file{thunk.c}).
535
536QEMU supports host CPUs which have pages bigger than 4KB. It records all
537the mappings the process does and try to emulated the @code{mmap()}
538system calls in cases where the host @code{mmap()} call would fail
539because of bad page alignment.
540
541@subsection Linux signals
542
543Normal and real-time signals are queued along with their information
544(@code{siginfo_t}) as it is done in the Linux kernel. Then an interrupt
545request is done to the virtual CPU. When it is interrupted, one queued
546signal is handled by generating a stack frame in the virtual CPU as the
547Linux kernel does. The @code{sigreturn()} system call is emulated to return
548from the virtual signal handler.
549
550Some signals (such as SIGALRM) directly come from the host. Other
551signals are synthesized from the virtual CPU exceptions such as SIGFPE
552when a division by zero is done (see @code{main.c:cpu_loop()}).
553
554The blocked signal mask is still handled by the host Linux kernel so
555that most signal system calls can be redirected directly to the host
556Linux kernel. Only the @code{sigaction()} and @code{sigreturn()} system
557calls need to be fully emulated (see @file{signal.c}).
558
559@subsection clone() system call and threads
560
561The Linux clone() system call is usually used to create a thread. QEMU
562uses the host clone() system call so that real host threads are created
563for each emulated thread. One virtual CPU instance is created for each
564thread.
565
566The virtual x86 CPU atomic operations are emulated with a global lock so
567that their semantic is preserved.
568
569Note that currently there are still some locking issues in QEMU. In
570particular, the translated cache flush is not protected yet against
571reentrancy.
572
573@subsection Self-virtualization
574
575QEMU was conceived so that ultimately it can emulate itself. Although
576it is not very useful, it is an important test to show the power of the
577emulator.
578
579Achieving self-virtualization is not easy because there may be address
580space conflicts. QEMU user emulators solve this problem by being an
581executable ELF shared object as the ld-linux.so ELF interpreter. That
582way, it can be relocated at load time.
583
584@node Bibliography
585@section Bibliography
586
587@table @asis
588
589@item [1]
590@url{http://citeseer.nj.nec.com/piumarta98optimizing.html}, Optimizing
591direct threaded code by selective inlining (1998) by Ian Piumarta, Fabio
592Riccardi.
593
594@item [2]
595@url{http://developer.kde.org/~sewardj/}, Valgrind, an open-source
596memory debugger for x86-GNU/Linux, by Julian Seward.
597
598@item [3]
599@url{http://bochs.sourceforge.net/}, the Bochs IA-32 Emulator Project,
600by Kevin Lawton et al.
601
602@item [4]
603@url{http://www.cs.rose-hulman.edu/~donaldlf/em86/index.html}, the EM86
604x86 emulator on Alpha-Linux.
605
606@item [5]
607@url{http://www.usenix.org/publications/library/proceedings/usenix-nt97/@/full_papers/chernoff/chernoff.pdf},
608DIGITAL FX!32: Running 32-Bit x86 Applications on Alpha NT, by Anton
609Chernoff and Ray Hookway.
610
611@item [6]
612@url{http://www.willows.com/}, Windows API library emulation from
613Willows Software.
614
615@item [7]
616@url{http://user-mode-linux.sourceforge.net/},
617The User-mode Linux Kernel.
618
619@item [8]
620@url{http://www.plex86.org/},
621The new Plex86 project.
622
623@item [9]
624@url{http://www.vmware.com/},
625The VMWare PC virtualizer.
626
627@item [10]
628@url{http://www.microsoft.com/windowsxp/virtualpc/},
629The VirtualPC PC virtualizer.
630
631@item [11]
632@url{http://www.twoostwo.org/},
633The TwoOStwo PC virtualizer.
634
635@item [12]
636@url{http://virtualbox.org/},
637The VirtualBox PC virtualizer.
638
639@item [13]
640@url{http://www.xen.org/},
641The Xen hypervisor.
642
643@item [14]
644@url{http://kvm.qumranet.com/kvmwiki/Front_Page},
645Kernel Based Virtual Machine (KVM).
646
647@item [15]
648@url{http://www.greensocs.com/projects/QEMUSystemC},
649QEMU-SystemC, a hardware co-simulator.
650
651@end table
652
653@node Regression Tests
654@chapter Regression Tests
655
656In the directory @file{tests/}, various interesting testing programs
657are available. They are used for regression testing.
658
659@menu
660* test-i386::
661* linux-test::
662* qruncom.c::
663@end menu
664
665@node test-i386
666@section @file{test-i386}
667
668This program executes most of the 16 bit and 32 bit x86 instructions and
669generates a text output. It can be compared with the output obtained with
670a real CPU or another emulator. The target @code{make test} runs this
671program and a @code{diff} on the generated output.
672
673The Linux system call @code{modify_ldt()} is used to create x86 selectors
674to test some 16 bit addressing and 32 bit with segmentation cases.
675
676The Linux system call @code{vm86()} is used to test vm86 emulation.
677
678Various exceptions are raised to test most of the x86 user space
679exception reporting.
680
681@node linux-test
682@section @file{linux-test}
683
684This program tests various Linux system calls. It is used to verify
685that the system call parameters are correctly converted between target
686and host CPUs.
687
688@node qruncom.c
689@section @file{qruncom.c}
690
691Example of usage of @code{libqemu} to emulate a user mode i386 CPU.
692
693@node Index
694@chapter Index
695@printindex cp
696
697@bye
698