1 \input texinfo @c -*- texinfo -*-
3 @setfilename qemu-tech.info
6 @documentencoding UTF-8
8 @settitle QEMU Internals
15 * QEMU Internals: (qemu-tech). The QEMU Emulator Internals.
22 @center @titlefont{QEMU Internals}
44 * intro_features:: Features
45 * intro_x86_emulation:: x86 and x86-64 emulation
46 * intro_arm_emulation:: ARM emulation
47 * intro_mips_emulation:: MIPS emulation
48 * intro_ppc_emulation:: PowerPC emulation
49 * intro_sparc_emulation:: Sparc32 and Sparc64 emulation
50 * intro_xtensa_emulation:: Xtensa emulation
51 * intro_other_emulation:: Other CPU emulation
57 QEMU is a FAST! processor emulator using a portable dynamic
60 QEMU has two operating modes:
65 Full system emulation. In this mode (full platform virtualization),
66 QEMU emulates a full system (usually a PC), including a processor and
67 various peripherals. It can be used to launch several different
68 Operating Systems at once without rebooting the host machine or to
72 User mode emulation. In this mode (application level virtualization),
73 QEMU can launch processes compiled for one CPU on another CPU, however
74 the Operating Systems must match. This can be used for example to ease
75 cross-compilation and cross-debugging.
78 As QEMU requires no host kernel driver to run, it is very safe and
81 QEMU generic features:
85 @item User space only or full system emulation.
87 @item Using dynamic translation to native code for reasonable speed.
90 Working on x86, x86_64 and PowerPC32/64 hosts. Being tested on ARM,
91 S390x, Sparc32 and Sparc64.
93 @item Self-modifying code support.
95 @item Precise exceptions support.
98 Floating point library supporting both full software emulation and
99 native host FPU instructions.
103 QEMU user mode emulation features:
105 @item Generic Linux system call converter, including most ioctls.
107 @item clone() emulation using native CPU clone() to use Linux scheduler for threads.
109 @item Accurate signal handling by remapping host signals to target signals.
112 Linux user emulator (Linux host only) can be used to launch the Wine
113 Windows API emulator (@url{http://www.winehq.org}). A BSD user emulator for BSD
114 hosts is under development. It would also be possible to develop a
115 similar user emulator for Solaris.
117 QEMU full system emulation features:
120 QEMU uses a full software MMU for maximum portability.
123 QEMU can optionally use an in-kernel accelerator, like kvm. The accelerators
124 execute some of the guest code natively, while
125 continuing to emulate the rest of the machine.
128 Various hardware devices can be emulated and in some cases, host
129 devices (e.g. serial and parallel ports, USB, drives) can be used
130 transparently by the guest Operating System. Host device passthrough
131 can be used for talking to external physical peripherals (e.g. a
132 webcam, modem or tape drive).
135 Symmetric multiprocessing (SMP) even on a host with a single CPU. On a
136 SMP host system, QEMU can use only one CPU fully due to difficulty in
137 implementing atomic memory accesses efficiently.
141 @node intro_x86_emulation
142 @section x86 and x86-64 emulation
144 QEMU x86 target features:
148 @item The virtual x86 CPU supports 16 bit and 32 bit addressing with segmentation.
149 LDT/GDT and IDT are emulated. VM86 mode is also supported to run
150 DOSEMU. There is some support for MMX/3DNow!, SSE, SSE2, SSE3, SSSE3,
151 and SSE4 as well as x86-64 SVM.
153 @item Support of host page sizes bigger than 4KB in user mode emulation.
155 @item QEMU can emulate itself on x86.
157 @item An extensive Linux x86 CPU test program is included @file{tests/test-i386}.
158 It can be used to test other x86 virtual CPUs.
162 Current QEMU limitations:
166 @item Limited x86-64 support.
168 @item IPC syscalls are missing.
170 @item The x86 segment limits and access rights are not tested at every
171 memory access (yet). Hopefully, very few OSes seem to rely on that for
176 @node intro_arm_emulation
177 @section ARM emulation
181 @item Full ARM 7 user emulation.
183 @item NWFPE FPU support included in user Linux emulation.
185 @item Can run most ARM Linux binaries.
189 @node intro_mips_emulation
190 @section MIPS emulation
194 @item The system emulation allows full MIPS32/MIPS64 Release 2 emulation,
195 including privileged instructions, FPU and MMU, in both little and big
198 @item The Linux userland emulation can run many 32 bit MIPS Linux binaries.
202 Current QEMU limitations:
206 @item Self-modifying code is not always handled correctly.
208 @item 64 bit userland emulation is not implemented.
210 @item The system emulation is not complete enough to run real firmware.
212 @item The watchpoint debug facility is not implemented.
216 @node intro_ppc_emulation
217 @section PowerPC emulation
221 @item Full PowerPC 32 bit emulation, including privileged instructions,
224 @item Can run most PowerPC Linux binaries.
228 @node intro_sparc_emulation
229 @section Sparc32 and Sparc64 emulation
233 @item Full SPARC V8 emulation, including privileged
234 instructions, FPU and MMU. SPARC V9 emulation includes most privileged
235 and VIS instructions, FPU and I/D MMU. Alignment is fully enforced.
237 @item Can run most 32-bit SPARC Linux binaries, SPARC32PLUS Linux binaries and
238 some 64-bit SPARC Linux binaries.
242 Current QEMU limitations:
246 @item IPC syscalls are missing.
248 @item Floating point exception support is buggy.
250 @item Atomic instructions are not correctly implemented.
252 @item There are still some problems with Sparc64 emulators.
256 @node intro_xtensa_emulation
257 @section Xtensa emulation
261 @item Core Xtensa ISA emulation, including most options: code density,
262 loop, extended L32R, 16- and 32-bit multiplication, 32-bit division,
263 MAC16, miscellaneous operations, boolean, FP coprocessor, coprocessor
264 context, debug, multiprocessor synchronization,
265 conditional store, exceptions, relocatable vectors, unaligned exception,
266 interrupts (including high priority and timer), hardware alignment,
267 region protection, region translation, MMU, windowed registers, thread
268 pointer, processor ID.
270 @item Not implemented options: data/instruction cache (including cache
271 prefetch and locking), XLMI, processor interface. Also options not
272 covered by the core ISA (e.g. FLIX, wide branches) are not implemented.
274 @item Can run most Xtensa Linux binaries.
276 @item New core configuration that requires no additional instructions
277 may be created from overlay with minimal amount of hand-written code.
281 @node intro_other_emulation
282 @section Other CPU emulation
284 In addition to the above, QEMU supports emulation of other CPUs with
285 varying levels of success. These are:
300 @chapter QEMU Internals
303 * QEMU compared to other emulators::
304 * Portable dynamic translation::
305 * Condition code optimisations::
306 * CPU state optimisations::
307 * Translation cache::
308 * Direct block chaining::
309 * Self-modifying code and translated code invalidation::
310 * Exception support::
313 * Hardware interrupts::
314 * User emulation specific details::
318 @node QEMU compared to other emulators
319 @section QEMU compared to other emulators
321 Like bochs [1], QEMU emulates an x86 CPU. But QEMU is much faster than
322 bochs as it uses dynamic compilation. Bochs is closely tied to x86 PC
323 emulation while QEMU can emulate several processors.
325 Like Valgrind [2], QEMU does user space emulation and dynamic
326 translation. Valgrind is mainly a memory debugger while QEMU has no
327 support for it (QEMU could be used to detect out of bound memory
328 accesses as Valgrind, but it has no support to track uninitialised data
329 as Valgrind does). The Valgrind dynamic translator generates better code
330 than QEMU (in particular it does register allocation) but it is closely
331 tied to an x86 host and target and has no support for precise exceptions
332 and system emulation.
334 EM86 [3] is the closest project to user space QEMU (and QEMU still uses
335 some of its code, in particular the ELF file loader). EM86 was limited
336 to an alpha host and used a proprietary and slow interpreter (the
337 interpreter part of the FX!32 Digital Win32 code translator [4]).
339 TWIN from Willows Software was a Windows API emulator like Wine. It is less
340 accurate than Wine but includes a protected mode x86 interpreter to launch
341 x86 Windows executables. Such an approach has greater potential because most
342 of the Windows API is executed natively but it is far more difficult to
343 develop because all the data structures and function parameters exchanged
344 between the API and the x86 code must be converted.
346 User mode Linux [5] was the only solution before QEMU to launch a
347 Linux kernel as a process while not needing any host kernel
348 patches. However, user mode Linux requires heavy kernel patches while
349 QEMU accepts unpatched Linux kernels. The price to pay is that QEMU is
352 The Plex86 [6] PC virtualizer is done in the same spirit as the now
353 obsolete qemu-fast system emulator. It requires a patched Linux kernel
354 to work (you cannot launch the same kernel on your PC), but the
355 patches are really small. As it is a PC virtualizer (no emulation is
356 done except for some privileged instructions), it has the potential of
357 being faster than QEMU. The downside is that a complicated (and
358 potentially unsafe) host kernel patch is needed.
360 The commercial PC Virtualizers (VMWare [7], VirtualPC [8]) are faster
361 than QEMU (without virtualization), but they all need specific, proprietary
362 and potentially unsafe host drivers. Moreover, they are unable to
363 provide cycle exact simulation as an emulator can.
365 VirtualBox [9], Xen [10] and KVM [11] are based on QEMU. QEMU-SystemC
366 [12] uses QEMU to simulate a system where some hardware devices are
367 developed in SystemC.
369 @node Portable dynamic translation
370 @section Portable dynamic translation
372 QEMU is a dynamic translator. When it first encounters a piece of code,
373 it converts it to the host instruction set. Usually dynamic translators
374 are very complicated and highly CPU dependent. QEMU uses some tricks
375 which make it relatively easily portable and simple while achieving good
378 After the release of version 0.9.1, QEMU switched to a new method of
379 generating code, Tiny Code Generator or TCG. TCG relaxes the
380 dependency on the exact version of the compiler used. The basic idea
381 is to split every target instruction into a couple of RISC-like TCG
382 ops (see @code{target-i386/translate.c}). Some optimizations can be
383 performed at this stage, including liveness analysis and trivial
384 constant expression evaluation. TCG ops are then implemented in the
385 host CPU back end, also known as TCG target (see
386 @code{tcg/i386/tcg-target.inc.c}). For more information, please take a
387 look at @code{tcg/README}.
389 @node Condition code optimisations
390 @section Condition code optimisations
392 Lazy evaluation of CPU condition codes (@code{EFLAGS} register on x86)
393 is important for CPUs where every instruction sets the condition
394 codes. It tends to be less important on conventional RISC systems
395 where condition codes are only updated when explicitly requested. On
396 Sparc64, costly update of both 32 and 64 bit condition codes can be
397 avoided with lazy evaluation.
399 Instead of computing the condition codes after each x86 instruction,
400 QEMU just stores one operand (called @code{CC_SRC}), the result
401 (called @code{CC_DST}) and the type of operation (called
402 @code{CC_OP}). When the condition codes are needed, the condition
403 codes can be calculated using this information. In addition, an
404 optimized calculation can be performed for some instruction types like
405 conditional branches.
407 @code{CC_OP} is almost never explicitly set in the generated code
408 because it is known at translation time.
410 The lazy condition code evaluation is used on x86, m68k, cris and
411 Sparc. ARM uses a simplified variant for the N and Z flags.
413 @node CPU state optimisations
414 @section CPU state optimisations
416 The target CPUs have many internal states which change the way it
417 evaluates instructions. In order to achieve a good speed, the
418 translation phase considers that some state information of the virtual
419 CPU cannot change in it. The state is recorded in the Translation
420 Block (TB). If the state changes (e.g. privilege level), a new TB will
421 be generated and the previous TB won't be used anymore until the state
422 matches the state recorded in the previous TB. For example, if the SS,
423 DS and ES segments have a zero base, then the translator does not even
424 generate an addition for the segment base.
426 [The FPU stack pointer register is not handled that way yet].
428 @node Translation cache
429 @section Translation cache
431 A 32 MByte cache holds the most recently used translations. For
432 simplicity, it is completely flushed when it is full. A translation unit
433 contains just a single basic block (a block of x86 instructions
434 terminated by a jump or by a virtual CPU state change which the
435 translator cannot deduce statically).
437 @node Direct block chaining
438 @section Direct block chaining
440 After each translated basic block is executed, QEMU uses the simulated
441 Program Counter (PC) and other cpu state information (such as the CS
442 segment base value) to find the next basic block.
444 In order to accelerate the most common cases where the new simulated PC
445 is known, QEMU can patch a basic block so that it jumps directly to the
448 The most portable code uses an indirect jump. An indirect jump makes
449 it easier to make the jump target modification atomic. On some host
450 architectures (such as x86 or PowerPC), the @code{JUMP} opcode is
451 directly patched so that the block chaining has no overhead.
453 @node Self-modifying code and translated code invalidation
454 @section Self-modifying code and translated code invalidation
456 Self-modifying code is a special challenge in x86 emulation because no
457 instruction cache invalidation is signaled by the application when code
460 When translated code is generated for a basic block, the corresponding
461 host page is write protected if it is not already read-only. Then, if
462 a write access is done to the page, Linux raises a SEGV signal. QEMU
463 then invalidates all the translated code in the page and enables write
464 accesses to the page.
466 Correct translated code invalidation is done efficiently by maintaining
467 a linked list of every translated block contained in a given page. Other
468 linked lists are also maintained to undo direct block chaining.
470 On RISC targets, correctly written software uses memory barriers and
471 cache flushes, so some of the protection above would not be
472 necessary. However, QEMU still requires that the generated code always
473 matches the target instructions in memory in order to handle
474 exceptions correctly.
476 @node Exception support
477 @section Exception support
479 longjmp() is used when an exception such as division by zero is
482 The host SIGSEGV and SIGBUS signal handlers are used to get invalid
483 memory accesses. The simulated program counter is found by
484 retranslating the corresponding basic block and by looking where the
485 host program counter was at the exception point.
487 The virtual CPU cannot retrieve the exact @code{EFLAGS} register because
488 in some cases it is not computed because of condition code
489 optimisations. It is not a big concern because the emulated code can
490 still be restarted in any cases.
493 @section MMU emulation
495 For system emulation QEMU supports a soft MMU. In that mode, the MMU
496 virtual to physical address translation is done at every memory
497 access. QEMU uses an address translation cache to speed up the
500 In order to avoid flushing the translated code each time the MMU
501 mappings change, QEMU uses a physically indexed translation cache. It
502 means that each basic block is indexed with its physical address.
504 When MMU mappings change, only the chaining of the basic blocks is
505 reset (i.e. a basic block can no longer jump directly to another one).
507 @node Device emulation
508 @section Device emulation
510 Systems emulated by QEMU are organized by boards. At initialization
511 phase, each board instantiates a number of CPUs, devices, RAM and
512 ROM. Each device in turn can assign I/O ports or memory areas (for
513 MMIO) to its handlers. When the emulation starts, an access to the
514 ports or MMIO memory areas assigned to the device causes the
515 corresponding handler to be called.
517 RAM and ROM are handled more optimally, only the offset to the host
518 memory needs to be added to the guest address.
520 The video RAM of VGA and other display cards is special: it can be
521 read or written directly like RAM, but write accesses cause the memory
522 to be marked with VGA_DIRTY flag as well.
524 QEMU supports some device classes like serial and parallel ports, USB,
525 drives and network devices, by providing APIs for easier connection to
526 the generic, higher level implementations. The API hides the
527 implementation details from the devices, like native device use or
528 advanced block device formats like QCOW.
530 Usually the devices implement a reset method and register support for
531 saving and loading of the device state. The devices can also use
532 timers, especially together with the use of bottom halves (BHs).
534 @node Hardware interrupts
535 @section Hardware interrupts
537 In order to be faster, QEMU does not check at every basic block if a
538 hardware interrupt is pending. Instead, the user must asynchronously
539 call a specific function to tell that an interrupt is pending. This
540 function resets the chaining of the currently executing basic
541 block. It ensures that the execution will return soon in the main loop
542 of the CPU emulator. Then the main loop can test if the interrupt is
543 pending and handle it.
545 @node User emulation specific details
546 @section User emulation specific details
548 @subsection Linux system call translation
550 QEMU includes a generic system call translator for Linux. It means that
551 the parameters of the system calls can be converted to fix the
552 endianness and 32/64 bit issues. The IOCTLs are converted with a generic
553 type description system (see @file{ioctls.h} and @file{thunk.c}).
555 QEMU supports host CPUs which have pages bigger than 4KB. It records all
556 the mappings the process does and try to emulated the @code{mmap()}
557 system calls in cases where the host @code{mmap()} call would fail
558 because of bad page alignment.
560 @subsection Linux signals
562 Normal and real-time signals are queued along with their information
563 (@code{siginfo_t}) as it is done in the Linux kernel. Then an interrupt
564 request is done to the virtual CPU. When it is interrupted, one queued
565 signal is handled by generating a stack frame in the virtual CPU as the
566 Linux kernel does. The @code{sigreturn()} system call is emulated to return
567 from the virtual signal handler.
569 Some signals (such as SIGALRM) directly come from the host. Other
570 signals are synthesized from the virtual CPU exceptions such as SIGFPE
571 when a division by zero is done (see @code{main.c:cpu_loop()}).
573 The blocked signal mask is still handled by the host Linux kernel so
574 that most signal system calls can be redirected directly to the host
575 Linux kernel. Only the @code{sigaction()} and @code{sigreturn()} system
576 calls need to be fully emulated (see @file{signal.c}).
578 @subsection clone() system call and threads
580 The Linux clone() system call is usually used to create a thread. QEMU
581 uses the host clone() system call so that real host threads are created
582 for each emulated thread. One virtual CPU instance is created for each
585 The virtual x86 CPU atomic operations are emulated with a global lock so
586 that their semantic is preserved.
588 Note that currently there are still some locking issues in QEMU. In
589 particular, the translated cache flush is not protected yet against
592 @subsection Self-virtualization
594 QEMU was conceived so that ultimately it can emulate itself. Although
595 it is not very useful, it is an important test to show the power of the
598 Achieving self-virtualization is not easy because there may be address
599 space conflicts. QEMU user emulators solve this problem by being an
600 executable ELF shared object as the ld-linux.so ELF interpreter. That
601 way, it can be relocated at load time.
604 @section Bibliography
609 @url{http://bochs.sourceforge.net/}, the Bochs IA-32 Emulator Project,
610 by Kevin Lawton et al.
613 @url{http://www.valgrind.org/}, Valgrind, an open-source memory debugger
617 @url{http://ftp.dreamtime.org/pub/linux/Linux-Alpha/em86/v0.2/docs/em86.html},
618 the EM86 x86 emulator on Alpha-Linux.
621 @url{http://www.usenix.org/publications/library/proceedings/usenix-nt97/@/full_papers/chernoff/chernoff.pdf},
622 DIGITAL FX!32: Running 32-Bit x86 Applications on Alpha NT, by Anton
623 Chernoff and Ray Hookway.
626 @url{http://user-mode-linux.sourceforge.net/},
627 The User-mode Linux Kernel.
630 @url{http://www.plex86.org/},
631 The new Plex86 project.
634 @url{http://www.vmware.com/},
635 The VMWare PC virtualizer.
638 @url{https://www.microsoft.com/download/details.aspx?id=3702},
639 The VirtualPC PC virtualizer.
642 @url{http://virtualbox.org/},
643 The VirtualBox PC virtualizer.
646 @url{http://www.xen.org/},
650 @url{http://www.linux-kvm.org/},
651 Kernel Based Virtual Machine (KVM).
654 @url{http://www.greensocs.com/projects/QEMUSystemC},
655 QEMU-SystemC, a hardware co-simulator.
659 @node Regression Tests
660 @chapter Regression Tests
662 In the directory @file{tests/}, various interesting testing programs
663 are available. They are used for regression testing.
671 @section @file{test-i386}
673 This program executes most of the 16 bit and 32 bit x86 instructions and
674 generates a text output. It can be compared with the output obtained with
675 a real CPU or another emulator. The target @code{make test} runs this
676 program and a @code{diff} on the generated output.
678 The Linux system call @code{modify_ldt()} is used to create x86 selectors
679 to test some 16 bit addressing and 32 bit with segmentation cases.
681 The Linux system call @code{vm86()} is used to test vm86 emulation.
683 Various exceptions are raised to test most of the x86 user space
687 @section @file{linux-test}
689 This program tests various Linux system calls. It is used to verify
690 that the system call parameters are correctly converted between target