2019-06-17 16:35:30 +02:00
|
|
|
|
====================
|
|
|
|
|
Translator Internals
|
|
|
|
|
====================
|
|
|
|
|
|
|
|
|
|
QEMU is a dynamic translator. When it first encounters a piece of code,
|
|
|
|
|
it converts it to the host instruction set. Usually dynamic translators
|
|
|
|
|
are very complicated and highly CPU dependent. QEMU uses some tricks
|
|
|
|
|
which make it relatively easily portable and simple while achieving good
|
|
|
|
|
performances.
|
|
|
|
|
|
|
|
|
|
QEMU's dynamic translation backend is called TCG, for "Tiny Code
|
|
|
|
|
Generator". For more information, please take a look at ``tcg/README``.
|
|
|
|
|
|
2021-06-01 14:51:43 +02:00
|
|
|
|
The following sections outline some notable features and implementation
|
|
|
|
|
details of QEMU's dynamic translator.
|
2019-06-17 16:35:30 +02:00
|
|
|
|
|
|
|
|
|
CPU state optimisations
|
|
|
|
|
-----------------------
|
|
|
|
|
|
2021-06-01 14:51:43 +02:00
|
|
|
|
The target CPUs have many internal states which change the way they
|
|
|
|
|
evaluate instructions. In order to achieve a good speed, the
|
2019-06-17 16:35:30 +02:00
|
|
|
|
translation phase considers that some state information of the virtual
|
|
|
|
|
CPU cannot change in it. The state is recorded in the Translation
|
|
|
|
|
Block (TB). If the state changes (e.g. privilege level), a new TB will
|
|
|
|
|
be generated and the previous TB won't be used anymore until the state
|
|
|
|
|
matches the state recorded in the previous TB. The same idea can be applied
|
|
|
|
|
to other aspects of the CPU state. For example, on x86, if the SS,
|
|
|
|
|
DS and ES segments have a zero base, then the translator does not even
|
|
|
|
|
generate an addition for the segment base.
|
|
|
|
|
|
|
|
|
|
Direct block chaining
|
|
|
|
|
---------------------
|
|
|
|
|
|
|
|
|
|
After each translated basic block is executed, QEMU uses the simulated
|
2021-06-01 14:51:43 +02:00
|
|
|
|
Program Counter (PC) and other CPU state information (such as the CS
|
2019-06-17 16:35:30 +02:00
|
|
|
|
segment base value) to find the next basic block.
|
|
|
|
|
|
2021-06-01 14:51:43 +02:00
|
|
|
|
In its simplest, less optimized form, this is done by exiting from the
|
|
|
|
|
current TB, going through the TB epilogue, and then back to the
|
|
|
|
|
main loop. That’s where QEMU looks for the next TB to execute,
|
|
|
|
|
translating it from the guest architecture if it isn’t already available
|
|
|
|
|
in memory. Then QEMU proceeds to execute this next TB, starting at the
|
|
|
|
|
prologue and then moving on to the translated instructions.
|
2019-06-17 16:35:30 +02:00
|
|
|
|
|
2021-06-01 14:51:43 +02:00
|
|
|
|
Exiting from the TB this way will cause the ``cpu_exec_interrupt()``
|
|
|
|
|
callback to be re-evaluated before executing additional instructions.
|
|
|
|
|
It is mandatory to exit this way after any CPU state changes that may
|
|
|
|
|
unmask interrupts.
|
|
|
|
|
|
|
|
|
|
In order to accelerate the cases where the TB for the new
|
|
|
|
|
simulated PC is already available, QEMU has mechanisms that allow
|
|
|
|
|
multiple TBs to be chained directly, without having to go back to the
|
|
|
|
|
main loop as described above. These mechanisms are:
|
|
|
|
|
|
|
|
|
|
``lookup_and_goto_ptr``
|
|
|
|
|
^^^^^^^^^^^^^^^^^^^^^^^
|
|
|
|
|
|
|
|
|
|
Calling ``tcg_gen_lookup_and_goto_ptr()`` will emit a call to
|
|
|
|
|
``helper_lookup_tb_ptr``. This helper will look for an existing TB that
|
|
|
|
|
matches the current CPU state. If the destination TB is available its
|
|
|
|
|
code address is returned, otherwise the address of the JIT epilogue is
|
|
|
|
|
returned. The call to the helper is always followed by the tcg ``goto_ptr``
|
|
|
|
|
opcode, which branches to the returned address. In this way, we either
|
|
|
|
|
branch to the next TB or return to the main loop.
|
|
|
|
|
|
|
|
|
|
``goto_tb + exit_tb``
|
|
|
|
|
^^^^^^^^^^^^^^^^^^^^^
|
|
|
|
|
|
|
|
|
|
The translation code usually implements branching by performing the
|
|
|
|
|
following steps:
|
|
|
|
|
|
|
|
|
|
1. Call ``tcg_gen_goto_tb()`` passing a jump slot index (either 0 or 1)
|
|
|
|
|
as a parameter.
|
|
|
|
|
|
|
|
|
|
2. Emit TCG instructions to update the CPU state with any information
|
|
|
|
|
that has been assumed constant and is required by the main loop to
|
|
|
|
|
correctly locate and execute the next TB. For most guests, this is
|
|
|
|
|
just the PC of the branch destination, but others may store additional
|
|
|
|
|
data. The information updated in this step must be inferable from both
|
|
|
|
|
``cpu_get_tb_cpu_state()`` and ``cpu_restore_state()``.
|
|
|
|
|
|
|
|
|
|
3. Call ``tcg_gen_exit_tb()`` passing the address of the current TB and
|
|
|
|
|
the jump slot index again.
|
|
|
|
|
|
|
|
|
|
Step 1, ``tcg_gen_goto_tb()``, will emit a ``goto_tb`` TCG
|
|
|
|
|
instruction that later on gets translated to a jump to an address
|
|
|
|
|
associated with the specified jump slot. Initially, this is the address
|
|
|
|
|
of step 2's instructions, which update the CPU state information. Step 3,
|
|
|
|
|
``tcg_gen_exit_tb()``, exits from the current TB returning a tagged
|
|
|
|
|
pointer composed of the last executed TB’s address and the jump slot
|
|
|
|
|
index.
|
|
|
|
|
|
|
|
|
|
The first time this whole sequence is executed, step 1 simply jumps
|
|
|
|
|
to step 2. Then the CPU state information gets updated and we exit from
|
|
|
|
|
the current TB. As a result, the behavior is very similar to the less
|
|
|
|
|
optimized form described earlier in this section.
|
|
|
|
|
|
|
|
|
|
Next, the main loop looks for the next TB to execute using the
|
|
|
|
|
current CPU state information (creating the TB if it wasn’t already
|
|
|
|
|
available) and, before starting to execute the new TB’s instructions,
|
|
|
|
|
patches the previously executed TB by associating one of its jump
|
|
|
|
|
slots (the one specified in the call to ``tcg_gen_exit_tb()``) with the
|
|
|
|
|
address of the new TB.
|
|
|
|
|
|
|
|
|
|
The next time this previous TB is executed and we get to that same
|
|
|
|
|
``goto_tb`` step, it will already be patched (assuming the destination TB
|
|
|
|
|
is still in memory) and will jump directly to the first instruction of
|
|
|
|
|
the destination TB, without going back to the main loop.
|
|
|
|
|
|
|
|
|
|
For the ``goto_tb + exit_tb`` mechanism to be used, the following
|
|
|
|
|
conditions need to be satisfied:
|
|
|
|
|
|
|
|
|
|
* The change in CPU state must be constant, e.g., a direct branch and
|
|
|
|
|
not an indirect branch.
|
|
|
|
|
|
|
|
|
|
* The direct branch cannot cross a page boundary. Memory mappings
|
|
|
|
|
may change, causing the code at the destination address to change.
|
|
|
|
|
|
|
|
|
|
Note that, on step 3 (``tcg_gen_exit_tb()``), in addition to the
|
|
|
|
|
jump slot index, the address of the TB just executed is also returned.
|
|
|
|
|
This address corresponds to the TB that will be patched; it may be
|
|
|
|
|
different than the one that was directly executed from the main loop
|
|
|
|
|
if the latter had already been chained to other TBs.
|
2019-06-17 16:35:30 +02:00
|
|
|
|
|
|
|
|
|
Self-modifying code and translated code invalidation
|
|
|
|
|
----------------------------------------------------
|
|
|
|
|
|
|
|
|
|
Self-modifying code is a special challenge in x86 emulation because no
|
|
|
|
|
instruction cache invalidation is signaled by the application when code
|
|
|
|
|
is modified.
|
|
|
|
|
|
|
|
|
|
User-mode emulation marks a host page as write-protected (if it is
|
|
|
|
|
not already read-only) every time translated code is generated for a
|
|
|
|
|
basic block. Then, if a write access is done to the page, Linux raises
|
|
|
|
|
a SEGV signal. QEMU then invalidates all the translated code in the page
|
|
|
|
|
and enables write accesses to the page. For system emulation, write
|
|
|
|
|
protection is achieved through the software MMU.
|
|
|
|
|
|
|
|
|
|
Correct translated code invalidation is done efficiently by maintaining
|
|
|
|
|
a linked list of every translated block contained in a given page. Other
|
|
|
|
|
linked lists are also maintained to undo direct block chaining.
|
|
|
|
|
|
|
|
|
|
On RISC targets, correctly written software uses memory barriers and
|
|
|
|
|
cache flushes, so some of the protection above would not be
|
|
|
|
|
necessary. However, QEMU still requires that the generated code always
|
|
|
|
|
matches the target instructions in memory in order to handle
|
|
|
|
|
exceptions correctly.
|
|
|
|
|
|
|
|
|
|
Exception support
|
|
|
|
|
-----------------
|
|
|
|
|
|
|
|
|
|
longjmp() is used when an exception such as division by zero is
|
|
|
|
|
encountered.
|
|
|
|
|
|
|
|
|
|
The host SIGSEGV and SIGBUS signal handlers are used to get invalid
|
|
|
|
|
memory accesses. QEMU keeps a map from host program counter to
|
|
|
|
|
target program counter, and looks up where the exception happened
|
|
|
|
|
based on the host program counter at the exception point.
|
|
|
|
|
|
|
|
|
|
On some targets, some bits of the virtual CPU's state are not flushed to the
|
|
|
|
|
memory until the end of the translation block. This is done for internal
|
|
|
|
|
emulation state that is rarely accessed directly by the program and/or changes
|
|
|
|
|
very often throughout the execution of a translation block---this includes
|
|
|
|
|
condition codes on x86, delay slots on SPARC, conditional execution on
|
2020-03-09 22:58:18 +01:00
|
|
|
|
Arm, and so on. This state is stored for each target instruction, and
|
2019-06-17 16:35:30 +02:00
|
|
|
|
looked up on exceptions.
|
|
|
|
|
|
|
|
|
|
MMU emulation
|
|
|
|
|
-------------
|
|
|
|
|
|
|
|
|
|
For system emulation QEMU uses a software MMU. In that mode, the MMU
|
|
|
|
|
virtual to physical address translation is done at every memory
|
|
|
|
|
access.
|
|
|
|
|
|
|
|
|
|
QEMU uses an address translation cache (TLB) to speed up the translation.
|
|
|
|
|
In order to avoid flushing the translated code each time the MMU
|
|
|
|
|
mappings change, all caches in QEMU are physically indexed. This
|
|
|
|
|
means that each basic block is indexed with its physical address.
|
|
|
|
|
|
|
|
|
|
In order to avoid invalidating the basic block chain when MMU mappings
|
|
|
|
|
change, chaining is only performed when the destination of the jump
|
|
|
|
|
shares a page with the basic block that is performing the jump.
|
|
|
|
|
|
|
|
|
|
The MMU can also distinguish RAM and ROM memory areas from MMIO memory
|
|
|
|
|
areas. Access is faster for RAM and ROM because the translation cache also
|
|
|
|
|
hosts the offset between guest address and host memory. Accessing MMIO
|
|
|
|
|
memory areas instead calls out to C code for device emulation.
|
|
|
|
|
Finally, the MMU helps tracking dirty pages and pages pointed to by
|
|
|
|
|
translation blocks.
|
|
|
|
|
|