1 \input texinfo @c -*- texinfo -*-
3 @setfilename qemu-tech.info
4 @settitle QEMU Internals
11 * QEMU Internals: (qemu-tech). The QEMU Emulator Internals.
18 @center @titlefont{QEMU Internals}
41 * intro_features:: Features
42 * intro_x86_emulation:: x86 and x86-64 emulation
43 * intro_arm_emulation:: ARM emulation
44 * intro_mips_emulation:: MIPS emulation
45 * intro_ppc_emulation:: PowerPC emulation
46 * intro_sparc_emulation:: Sparc32 and Sparc64 emulation
47 * intro_other_emulation:: Other CPU emulation
53 QEMU is a FAST! processor emulator using a portable dynamic
56 QEMU has two operating modes:
61 Full system emulation. In this mode (full platform virtualization),
62 QEMU emulates a full system (usually a PC), including a processor and
63 various peripherals. It can be used to launch several different
64 Operating Systems at once without rebooting the host machine or to
68 User mode emulation. In this mode (application level virtualization),
69 QEMU can launch processes compiled for one CPU on another CPU, however
70 the Operating Systems must match. This can be used for example to ease
71 cross-compilation and cross-debugging.
74 As QEMU requires no host kernel driver to run, it is very safe and
77 QEMU generic features:
81 @item User space only or full system emulation.
83 @item Using dynamic translation to native code for reasonable speed.
86 Working on x86, x86_64 and PowerPC32/64 hosts. Being tested on ARM,
87 HPPA, Sparc32 and Sparc64. Previous versions had some support for
88 Alpha and S390 hosts, but TCG (see below) doesn't support those yet.
90 @item Self-modifying code support.
92 @item Precise exceptions support.
94 @item The virtual CPU is a library (@code{libqemu}) which can be used
95 in other projects (look at @file{qemu/tests/qruncom.c} to have an
96 example of user mode @code{libqemu} usage).
99 Floating point library supporting both full software emulation and
100 native host FPU instructions.
104 QEMU user mode emulation features:
106 @item Generic Linux system call converter, including most ioctls.
108 @item clone() emulation using native CPU clone() to use Linux scheduler for threads.
110 @item Accurate signal handling by remapping host signals to target signals.
113 Linux user emulator (Linux host only) can be used to launch the Wine
114 Windows API emulator (@url{http://www.winehq.org}). A Darwin user
115 emulator (Darwin hosts only) exists and a BSD user emulator for BSD
116 hosts is under development. It would also be possible to develop a
117 similar user emulator for Solaris.
119 QEMU full system emulation features:
122 QEMU uses a full software MMU for maximum portability.
125 QEMU can optionally use an in-kernel accelerator, like kvm. The accelerators
126 execute some of the guest code natively, while
127 continuing to emulate the rest of the machine.
130 Various hardware devices can be emulated and in some cases, host
131 devices (e.g. serial and parallel ports, USB, drives) can be used
132 transparently by the guest Operating System. Host device passthrough
133 can be used for talking to external physical peripherals (e.g. a
134 webcam, modem or tape drive).
137 Symmetric multiprocessing (SMP) even on a host with a single CPU. On a
138 SMP host system, QEMU can use only one CPU fully due to difficulty in
139 implementing atomic memory accesses efficiently.
143 @node intro_x86_emulation
144 @section x86 and x86-64 emulation
146 QEMU x86 target features:
150 @item The virtual x86 CPU supports 16 bit and 32 bit addressing with segmentation.
151 LDT/GDT and IDT are emulated. VM86 mode is also supported to run
152 DOSEMU. There is some support for MMX/3DNow!, SSE, SSE2, SSE3, SSSE3,
153 and SSE4 as well as x86-64 SVM.
155 @item Support of host page sizes bigger than 4KB in user mode emulation.
157 @item QEMU can emulate itself on x86.
159 @item An extensive Linux x86 CPU test program is included @file{tests/test-i386}.
160 It can be used to test other x86 virtual CPUs.
164 Current QEMU limitations:
168 @item Limited x86-64 support.
170 @item IPC syscalls are missing.
172 @item The x86 segment limits and access rights are not tested at every
173 memory access (yet). Hopefully, very few OSes seem to rely on that for
178 @node intro_arm_emulation
179 @section ARM emulation
183 @item Full ARM 7 user emulation.
185 @item NWFPE FPU support included in user Linux emulation.
187 @item Can run most ARM Linux binaries.
191 @node intro_mips_emulation
192 @section MIPS emulation
196 @item The system emulation allows full MIPS32/MIPS64 Release 2 emulation,
197 including privileged instructions, FPU and MMU, in both little and big
200 @item The Linux userland emulation can run many 32 bit MIPS Linux binaries.
204 Current QEMU limitations:
208 @item Self-modifying code is not always handled correctly.
210 @item 64 bit userland emulation is not implemented.
212 @item The system emulation is not complete enough to run real firmware.
214 @item The watchpoint debug facility is not implemented.
218 @node intro_ppc_emulation
219 @section PowerPC emulation
223 @item Full PowerPC 32 bit emulation, including privileged instructions,
226 @item Can run most PowerPC Linux binaries.
230 @node intro_sparc_emulation
231 @section Sparc32 and Sparc64 emulation
235 @item Full SPARC V8 emulation, including privileged
236 instructions, FPU and MMU. SPARC V9 emulation includes most privileged
237 and VIS instructions, FPU and I/D MMU. Alignment is fully enforced.
239 @item Can run most 32-bit SPARC Linux binaries, SPARC32PLUS Linux binaries and
240 some 64-bit SPARC Linux binaries.
244 Current QEMU limitations:
248 @item IPC syscalls are missing.
250 @item Floating point exception support is buggy.
252 @item Atomic instructions are not correctly implemented.
254 @item There are still some problems with Sparc64 emulators.
258 @node intro_other_emulation
259 @section Other CPU emulation
261 In addition to the above, QEMU supports emulation of other CPUs with
262 varying levels of success. These are:
277 @chapter QEMU Internals
280 * QEMU compared to other emulators::
281 * Portable dynamic translation::
282 * Condition code optimisations::
283 * CPU state optimisations::
284 * Translation cache::
285 * Direct block chaining::
286 * Self-modifying code and translated code invalidation::
287 * Exception support::
290 * Hardware interrupts::
291 * User emulation specific details::
295 @node QEMU compared to other emulators
296 @section QEMU compared to other emulators
298 Like bochs [3], QEMU emulates an x86 CPU. But QEMU is much faster than
299 bochs as it uses dynamic compilation. Bochs is closely tied to x86 PC
300 emulation while QEMU can emulate several processors.
302 Like Valgrind [2], QEMU does user space emulation and dynamic
303 translation. Valgrind is mainly a memory debugger while QEMU has no
304 support for it (QEMU could be used to detect out of bound memory
305 accesses as Valgrind, but it has no support to track uninitialised data
306 as Valgrind does). The Valgrind dynamic translator generates better code
307 than QEMU (in particular it does register allocation) but it is closely
308 tied to an x86 host and target and has no support for precise exceptions
309 and system emulation.
311 EM86 [4] is the closest project to user space QEMU (and QEMU still uses
312 some of its code, in particular the ELF file loader). EM86 was limited
313 to an alpha host and used a proprietary and slow interpreter (the
314 interpreter part of the FX!32 Digital Win32 code translator [5]).
316 TWIN [6] is a Windows API emulator like Wine. It is less accurate than
317 Wine but includes a protected mode x86 interpreter to launch x86 Windows
318 executables. Such an approach has greater potential because most of the
319 Windows API is executed natively but it is far more difficult to develop
320 because all the data structures and function parameters exchanged
321 between the API and the x86 code must be converted.
323 User mode Linux [7] was the only solution before QEMU to launch a
324 Linux kernel as a process while not needing any host kernel
325 patches. However, user mode Linux requires heavy kernel patches while
326 QEMU accepts unpatched Linux kernels. The price to pay is that QEMU is
329 The Plex86 [8] PC virtualizer is done in the same spirit as the now
330 obsolete qemu-fast system emulator. It requires a patched Linux kernel
331 to work (you cannot launch the same kernel on your PC), but the
332 patches are really small. As it is a PC virtualizer (no emulation is
333 done except for some privileged instructions), it has the potential of
334 being faster than QEMU. The downside is that a complicated (and
335 potentially unsafe) host kernel patch is needed.
337 The commercial PC Virtualizers (VMWare [9], VirtualPC [10], TwoOStwo
338 [11]) are faster than QEMU, but they all need specific, proprietary
339 and potentially unsafe host drivers. Moreover, they are unable to
340 provide cycle exact simulation as an emulator can.
342 VirtualBox [12], Xen [13] and KVM [14] are based on QEMU. QEMU-SystemC
343 [15] uses QEMU to simulate a system where some hardware devices are
344 developed in SystemC.
346 @node Portable dynamic translation
347 @section Portable dynamic translation
349 QEMU is a dynamic translator. When it first encounters a piece of code,
350 it converts it to the host instruction set. Usually dynamic translators
351 are very complicated and highly CPU dependent. QEMU uses some tricks
352 which make it relatively easily portable and simple while achieving good
355 After the release of version 0.9.1, QEMU switched to a new method of
356 generating code, Tiny Code Generator or TCG. TCG relaxes the
357 dependency on the exact version of the compiler used. The basic idea
358 is to split every target instruction into a couple of RISC-like TCG
359 ops (see @code{target-i386/translate.c}). Some optimizations can be
360 performed at this stage, including liveness analysis and trivial
361 constant expression evaluation. TCG ops are then implemented in the
362 host CPU back end, also known as TCG target (see
363 @code{tcg/i386/tcg-target.c}). For more information, please take a
364 look at @code{tcg/README}.
366 @node Condition code optimisations
367 @section Condition code optimisations
369 Lazy evaluation of CPU condition codes (@code{EFLAGS} register on x86)
370 is important for CPUs where every instruction sets the condition
371 codes. It tends to be less important on conventional RISC systems
372 where condition codes are only updated when explicitly requested. On
373 Sparc64, costly update of both 32 and 64 bit condition codes can be
374 avoided with lazy evaluation.
376 Instead of computing the condition codes after each x86 instruction,
377 QEMU just stores one operand (called @code{CC_SRC}), the result
378 (called @code{CC_DST}) and the type of operation (called
379 @code{CC_OP}). When the condition codes are needed, the condition
380 codes can be calculated using this information. In addition, an
381 optimized calculation can be performed for some instruction types like
382 conditional branches.
384 @code{CC_OP} is almost never explicitly set in the generated code
385 because it is known at translation time.
387 The lazy condition code evaluation is used on x86, m68k, cris and
388 Sparc. ARM uses a simplified variant for the N and Z flags.
390 @node CPU state optimisations
391 @section CPU state optimisations
393 The target CPUs have many internal states which change the way it
394 evaluates instructions. In order to achieve a good speed, the
395 translation phase considers that some state information of the virtual
396 CPU cannot change in it. The state is recorded in the Translation
397 Block (TB). If the state changes (e.g. privilege level), a new TB will
398 be generated and the previous TB won't be used anymore until the state
399 matches the state recorded in the previous TB. For example, if the SS,
400 DS and ES segments have a zero base, then the translator does not even
401 generate an addition for the segment base.
403 [The FPU stack pointer register is not handled that way yet].
405 @node Translation cache
406 @section Translation cache
408 A 16 MByte cache holds the most recently used translations. For
409 simplicity, it is completely flushed when it is full. A translation unit
410 contains just a single basic block (a block of x86 instructions
411 terminated by a jump or by a virtual CPU state change which the
412 translator cannot deduce statically).
414 @node Direct block chaining
415 @section Direct block chaining
417 After each translated basic block is executed, QEMU uses the simulated
418 Program Counter (PC) and other cpu state informations (such as the CS
419 segment base value) to find the next basic block.
421 In order to accelerate the most common cases where the new simulated PC
422 is known, QEMU can patch a basic block so that it jumps directly to the
425 The most portable code uses an indirect jump. An indirect jump makes
426 it easier to make the jump target modification atomic. On some host
427 architectures (such as x86 or PowerPC), the @code{JUMP} opcode is
428 directly patched so that the block chaining has no overhead.
430 @node Self-modifying code and translated code invalidation
431 @section Self-modifying code and translated code invalidation
433 Self-modifying code is a special challenge in x86 emulation because no
434 instruction cache invalidation is signaled by the application when code
437 When translated code is generated for a basic block, the corresponding
438 host page is write protected if it is not already read-only. Then, if
439 a write access is done to the page, Linux raises a SEGV signal. QEMU
440 then invalidates all the translated code in the page and enables write
441 accesses to the page.
443 Correct translated code invalidation is done efficiently by maintaining
444 a linked list of every translated block contained in a given page. Other
445 linked lists are also maintained to undo direct block chaining.
447 On RISC targets, correctly written software uses memory barriers and
448 cache flushes, so some of the protection above would not be
449 necessary. However, QEMU still requires that the generated code always
450 matches the target instructions in memory in order to handle
451 exceptions correctly.
453 @node Exception support
454 @section Exception support
456 longjmp() is used when an exception such as division by zero is
459 The host SIGSEGV and SIGBUS signal handlers are used to get invalid
460 memory accesses. The simulated program counter is found by
461 retranslating the corresponding basic block and by looking where the
462 host program counter was at the exception point.
464 The virtual CPU cannot retrieve the exact @code{EFLAGS} register because
465 in some cases it is not computed because of condition code
466 optimisations. It is not a big concern because the emulated code can
467 still be restarted in any cases.
470 @section MMU emulation
472 For system emulation QEMU supports a soft MMU. In that mode, the MMU
473 virtual to physical address translation is done at every memory
474 access. QEMU uses an address translation cache to speed up the
477 In order to avoid flushing the translated code each time the MMU
478 mappings change, QEMU uses a physically indexed translation cache. It
479 means that each basic block is indexed with its physical address.
481 When MMU mappings change, only the chaining of the basic blocks is
482 reset (i.e. a basic block can no longer jump directly to another one).
484 @node Device emulation
485 @section Device emulation
487 Systems emulated by QEMU are organized by boards. At initialization
488 phase, each board instantiates a number of CPUs, devices, RAM and
489 ROM. Each device in turn can assign I/O ports or memory areas (for
490 MMIO) to its handlers. When the emulation starts, an access to the
491 ports or MMIO memory areas assigned to the device causes the
492 corresponding handler to be called.
494 RAM and ROM are handled more optimally, only the offset to the host
495 memory needs to be added to the guest address.
497 The video RAM of VGA and other display cards is special: it can be
498 read or written directly like RAM, but write accesses cause the memory
499 to be marked with VGA_DIRTY flag as well.
501 QEMU supports some device classes like serial and parallel ports, USB,
502 drives and network devices, by providing APIs for easier connection to
503 the generic, higher level implementations. The API hides the
504 implementation details from the devices, like native device use or
505 advanced block device formats like QCOW.
507 Usually the devices implement a reset method and register support for
508 saving and loading of the device state. The devices can also use
509 timers, especially together with the use of bottom halves (BHs).
511 @node Hardware interrupts
512 @section Hardware interrupts
514 In order to be faster, QEMU does not check at every basic block if an
515 hardware interrupt is pending. Instead, the user must asynchrously
516 call a specific function to tell that an interrupt is pending. This
517 function resets the chaining of the currently executing basic
518 block. It ensures that the execution will return soon in the main loop
519 of the CPU emulator. Then the main loop can test if the interrupt is
520 pending and handle it.
522 @node User emulation specific details
523 @section User emulation specific details
525 @subsection Linux system call translation
527 QEMU includes a generic system call translator for Linux. It means that
528 the parameters of the system calls can be converted to fix the
529 endianness and 32/64 bit issues. The IOCTLs are converted with a generic
530 type description system (see @file{ioctls.h} and @file{thunk.c}).
532 QEMU supports host CPUs which have pages bigger than 4KB. It records all
533 the mappings the process does and try to emulated the @code{mmap()}
534 system calls in cases where the host @code{mmap()} call would fail
535 because of bad page alignment.
537 @subsection Linux signals
539 Normal and real-time signals are queued along with their information
540 (@code{siginfo_t}) as it is done in the Linux kernel. Then an interrupt
541 request is done to the virtual CPU. When it is interrupted, one queued
542 signal is handled by generating a stack frame in the virtual CPU as the
543 Linux kernel does. The @code{sigreturn()} system call is emulated to return
544 from the virtual signal handler.
546 Some signals (such as SIGALRM) directly come from the host. Other
547 signals are synthetized from the virtual CPU exceptions such as SIGFPE
548 when a division by zero is done (see @code{main.c:cpu_loop()}).
550 The blocked signal mask is still handled by the host Linux kernel so
551 that most signal system calls can be redirected directly to the host
552 Linux kernel. Only the @code{sigaction()} and @code{sigreturn()} system
553 calls need to be fully emulated (see @file{signal.c}).
555 @subsection clone() system call and threads
557 The Linux clone() system call is usually used to create a thread. QEMU
558 uses the host clone() system call so that real host threads are created
559 for each emulated thread. One virtual CPU instance is created for each
562 The virtual x86 CPU atomic operations are emulated with a global lock so
563 that their semantic is preserved.
565 Note that currently there are still some locking issues in QEMU. In
566 particular, the translated cache flush is not protected yet against
569 @subsection Self-virtualization
571 QEMU was conceived so that ultimately it can emulate itself. Although
572 it is not very useful, it is an important test to show the power of the
575 Achieving self-virtualization is not easy because there may be address
576 space conflicts. QEMU user emulators solve this problem by being an
577 executable ELF shared object as the ld-linux.so ELF interpreter. That
578 way, it can be relocated at load time.
581 @section Bibliography
586 @url{http://citeseer.nj.nec.com/piumarta98optimizing.html}, Optimizing
587 direct threaded code by selective inlining (1998) by Ian Piumarta, Fabio
591 @url{http://developer.kde.org/~sewardj/}, Valgrind, an open-source
592 memory debugger for x86-GNU/Linux, by Julian Seward.
595 @url{http://bochs.sourceforge.net/}, the Bochs IA-32 Emulator Project,
596 by Kevin Lawton et al.
599 @url{http://www.cs.rose-hulman.edu/~donaldlf/em86/index.html}, the EM86
600 x86 emulator on Alpha-Linux.
603 @url{http://www.usenix.org/publications/library/proceedings/usenix-nt97/@/full_papers/chernoff/chernoff.pdf},
604 DIGITAL FX!32: Running 32-Bit x86 Applications on Alpha NT, by Anton
605 Chernoff and Ray Hookway.
608 @url{http://www.willows.com/}, Windows API library emulation from
612 @url{http://user-mode-linux.sourceforge.net/},
613 The User-mode Linux Kernel.
616 @url{http://www.plex86.org/},
617 The new Plex86 project.
620 @url{http://www.vmware.com/},
621 The VMWare PC virtualizer.
624 @url{http://www.microsoft.com/windowsxp/virtualpc/},
625 The VirtualPC PC virtualizer.
628 @url{http://www.twoostwo.org/},
629 The TwoOStwo PC virtualizer.
632 @url{http://virtualbox.org/},
633 The VirtualBox PC virtualizer.
636 @url{http://www.xen.org/},
640 @url{http://kvm.qumranet.com/kvmwiki/Front_Page},
641 Kernel Based Virtual Machine (KVM).
644 @url{http://www.greensocs.com/projects/QEMUSystemC},
645 QEMU-SystemC, a hardware co-simulator.
649 @node Regression Tests
650 @chapter Regression Tests
652 In the directory @file{tests/}, various interesting testing programs
653 are available. They are used for regression testing.
662 @section @file{test-i386}
664 This program executes most of the 16 bit and 32 bit x86 instructions and
665 generates a text output. It can be compared with the output obtained with
666 a real CPU or another emulator. The target @code{make test} runs this
667 program and a @code{diff} on the generated output.
669 The Linux system call @code{modify_ldt()} is used to create x86 selectors
670 to test some 16 bit addressing and 32 bit with segmentation cases.
672 The Linux system call @code{vm86()} is used to test vm86 emulation.
674 Various exceptions are raised to test most of the x86 user space
678 @section @file{linux-test}
680 This program tests various Linux system calls. It is used to verify
681 that the system call parameters are correctly converted between target
685 @section @file{qruncom.c}
687 Example of usage of @code{libqemu} to emulate a user mode i386 CPU.