Replace the USE_DIRECT_JUMP ifdef with a TCG_TARGET_HAS_direct_jump
boolean test. Replace the tb_set_jmp_target1 ifdef with an unconditional
function tb_target_set_jmp_target.
While we're touching all backends, add a parameter for tb->tc_ptr;
we're going to need it shortly for some backends.
Move tb_set_jmp_target and tb_add_jump from exec-all.h to cpu-exec.c.
Backports commit a85833933628384d74ec412024d55cf012640287 from qemu
We use ADRP+ADD to compute the target address for goto_tb. This patch
introduces the NOP instruction which is used to align the above
instruction pair so that we can use one atomic instruction to patch
the destination offsets.
Backports commit b68686bd4bfeb70040b4099df993dfa0b4f37b03 from qemu
translate-all.c will be disabled if tcg is disabled in the build,
so page_size_init() function and related variables will be moved
to exec.c file.
Backports commit a0be0c585f5dcc4d50a37f6a20d3d625c5ef3a2c from qemu
Commit 1f5c00cfdb8114c ("qom/cpu: move tlb_flush to cpu_common_reset")
moved the call to tlb_flush() from the target-specific reset handlers
into the common code qom/cpu.c file, and protected the call with
"#ifdef CONFIG_SOFTMMU" to avoid that it is called for linux-user
only targets. But since qom/cpu.c is common code, CONFIG_SOFTMMU is
*never* defined here, so the tlb_flush() was simply never executed
anymore. Fix it by introducing a wrapper for tlb_flush() in a file
that is re-compiled for each target, i.e. in translate-all.c.
Backports commit 2cd53943115be5118b5b2d4b80ee0a39c94c4f73 from qemu
Some code paths can lead to atomic accesses racing with memset()
on cpu->tb_jmp_cache, which can result in torn reads/writes
and is undefined behaviour in C11.
These torn accesses are unlikely to show up as bugs, but from code
inspection they seem possible. For example, tb_phys_invalidate does:
/* remove the TB from the hash list */
h = tb_jmp_cache_hash_func(tb->pc);
CPU_FOREACH(cpu) {
if (atomic_read(&cpu->tb_jmp_cache[h]) == tb) {
atomic_set(&cpu->tb_jmp_cache[h], NULL);
}
}
Here atomic_set might race with a concurrent memset (such as the
ones scheduled via "unsafe" async work, e.g. tlb_flush_page) and
therefore we might end up with a torn pointer (or who knows what,
because we are under undefined behaviour).
This patch converts parallel accesses to cpu->tb_jmp_cache to use
atomic primitives, thereby bringing these accesses back to defined
behaviour. The price to pay is to potentially execute more instructions
when clearing cpu->tb_jmp_cache, but given how infrequently they happen
and the small size of the cache, the performance impact I have measured
is within noise range when booting debian-arm.
Note that under "safe async" work (e.g. do_tb_flush) we could use memset
because no other vcpus are running. However I'm keeping these accesses
atomic as well to keep things simple and to avoid confusing analysis
tools such as ThreadSanitizer.
Backports commit f3ced3c59287dabc253f83f0c70aa4934470c15e from qemu
Since we're no longer using a direct branch, we have no
limit on the branch distance.
Backports commit acb0b292b6d0f49972dc98f742e79ed53973e438 from qemu
We are partially initializing tb in tb_alloc. Instead, fully
initialize it in tb_gen_code, which is tb_alloc's only caller.
This saves an unnecessary write to tb->cflags.
Backports commit 2b48e10f888059a98043b4816769fa2a326a1d2c from qemu
Allocating an arbitrarily-sized array of tbs results in either
(a) a lot of memory wasted or (b) unnecessary flushes of the code
cache when we run out of TB structs in the array.
An obvious solution would be to just malloc a TB struct when needed,
and keep the TB array as an array of pointers (recall that tb_find_pc()
needs the TB array to run in O(log n)).
Perhaps a better solution, which is implemented in this patch, is to
allocate TB's right before the translated code they describe. This
results in some memory waste due to padding to have code and TBs in
separate cache lines--for instance, I measured 4.7% of padding in the
used portion of code_gen_buffer when booting aarch64 Linux on a
host with 64-byte cache lines. However, it can allow for optimizations
in some host architectures, since TCG backends could safely assume that
the TB and the corresponding translated code are very close to each
other in memory. See this message by rth for a detailed explanation:
https://lists.gnu.org/archive/html/qemu-devel/2017-03/msg05172.html
Subject: Re: GSoC 2017 Proposal: TCG performance enhancements
Backports commit 6e3b2bfd6af488a896f7936e99ef160f8f37e6f2 from qemu
The translation code uses cpu_ld*_code which can trigger a tlb_fill
which if it fails will erroneously attempts a fault resolution. This
never works during translation as the TB being generated hasn't been
added yet. The target should have checked retaddr before calling
cpu_restore_state but for those that have yet to be fixed we do it
here to avoid a recursive tb_lock() under MTTCG's new locking regime
Backports commit d8b2239bcd8872a5c5f7534d1658fc2365caab2d from qemu
There are a couple of changes that occur at the same time here:
- introduce a single vCPU qemu_tcg_cpu_thread_fn
One of these is spawned per vCPU with its own Thread and Condition
variables. qemu_tcg_rr_cpu_thread_fn is the new name for the old
single threaded function.
- the TLS current_cpu variable is now live for the lifetime of MTTCG
vCPU threads. This is for future work where async jobs need to know
the vCPU context they are operating in.
The user to switch on multi-thread behaviour and spawn a thread
per-vCPU. For a simple test kvm-unit-test like:
./arm/run ./arm/locking-test.flat -smp 4 -accel tcg,thread=multi
Will now use 4 vCPU threads and have an expected FAIL (instead of the
unexpected PASS) as the default mode of the test has no protection when
incrementing a shared variable.
We enable the parallel_cpus flag to ensure we generate correct barrier
and atomic code if supported by the front and backends. This doesn't
automatically enable MTTCG until default_mttcg_enabled() is updated to
check the configuration is supported.
Backports relevant parts of commit 372579427a5040a26dfee78464b50e2bdf27ef26
Enable tcg lock debug asserts in a debug build by default instead of
relying on DEBUG_LOCKING. None of the other DEBUG_* macros have
asserts, so this patch removes DEBUG_LOCKING and enable these asserts
in a debug build.
Backports commit 6ac3d7e845549f08473f020c1c70f14b8911a67e from qemu
gcc 5.3.0 diagnoses
translate-all.c: In function ‘alloc_code_gen_buffer’:
translate-all.c:756:17: error: switch condition has boolean value
switch (buf2 != MAP_FAILED) {
^
Backports commit f68808c7494b38764e1895a9852b994638b86536 from qemu
In the linux-user case all things that involve ''l1_map' and PageDesc
tweaks are protected by the memory lock (mmpa_lock). For SoftMMU mode
we previously relied on single threaded behaviour, with MTTCG we now use
the tb_lock().
As a result we need to do a little re-factoring and push the taking of
this lock up the call tree. This requires a slightly different entry for
the SoftMMU and user-mode cases from tb_invalidate_phys_range.
This also means user-mode breakpoint insertion needs to take two locks
but it hadn't taken any previously so this is an improvement.
Backpoirts commit ba051fb5e56d5ff5e4fa672d37954452e58543b2 from qemu
softmmu requires more functions to be thread-safe, because translation
blocks can be invalidated from e.g. notdirty callbacks. Probably the
same holds for user-mode emulation, it's just that no one has ever
tried to produce a coherent locking there.
This patch will guide the introduction of more tb_lock and tb_unlock
calls for system emulation.
Note that after this patch some (most) of the mentioned functions are
still called outside tb_lock/tb_unlock. The next one will rectify this.
Backports commit 7d7500d99895f888f97397ef32bb536bb0df3b74 from qemu
This adds asserts to check the locking on the various translation
engines structures. There are two sets of structures that are protected
by locks.
The first the l1map and PageDesc structures used to track which
translation blocks are associated with which physical addresses. In
user-mode this is covered by the mmap_lock.
The second case are TB context related structures which are protected by
tb_lock which is also user-mode only.
Currently the asserts do nothing in SoftMMU mode but this will change
for MTTCG.
Backports commit 301e40ed8005306c009978be295ed9a4b725178b from qemu
Make the debug define consistent with the others. The flush operation is
all about invalidating TranslationBlocks on flush events.
Also fix up the commenting on the other DEBUG for the benefit of
checkpatch.
Backports commit 955939a2b51f72bea1c200b559ea39985df5a633 from qemu
When we cannot emulate an atomic operation within a parallel
context, this exception allows us to stop the world and try
again in a serial context.
Backports commit fdbc2b5722f6092e47181a947c90fd4bdcc1c121 from qemu
Also backports parts of commit 02d57ea115b7669f588371c86484a2e8ebc369be
Remove L1 page mapping table properties computing
statically using macros which is dependent on
TARGET_PAGE_BITS. Drop macros V_L1_SIZE, V_L1_SHIFT,
V_L1_BITS macros and replace with variables which are
computed at early stage of VM boot.
Removing dependency can help to make TARGET_PAGE_BITS
dynamic.
Backports commit 66ec9f49399f0a9fa13ee77c472caba0de2773fc from qemu
The return address argument to the softmmu template helpers was
confused. In the legacy case, we wanted to indicate that there
is no return address, and so passed in NULL. However, we then
immediately subtracted GETPC_ADJ from NULL, resulting in a non-zero
value, indicating the presence of an (invalid) return address.
Push the GETPC_ADJ subtraction down to the only point it's required:
immediately before use within cpu_restore_state_from_tb, after all
NULL pointer checks have been completed.
This makes GETPC and GETRA identical. Remove GETRA as the lesser
used macro, replacing all uses with GETPC.
Backports commit 01ecaf438b1eb46abe23392c8ce5b7628b0c8cf5 from qemu
This ensures that if we find the TB on the slow path that tb->page_addr
is correctly set before being tested.
Backports commit 2e1ae44a4f4a6149fbb9dc812243522f07284700 from qemu
When invalidating a translation block, set an invalid flag into the
TranslationBlock structure first. It is also necessary to check whether
the target TB is still valid after acquiring 'tb_lock' but before calling
tb_add_jump() since TB lookup is to be performed out of 'tb_lock' in
future. Note that we don't have to check 'last_tb'; an already invalidated
TB will not be executed anyway and it is thus safe to patch it.
Backports commit 6d21e4208f382dd8ca1f7995a6dd9ea7ca281163 from qemu
Ensure atomicity and ordering of CPU's 'tb_flushed' access for future
translation block lookup out of 'tb_lock'.
This field can only be touched from another thread by tb_flush() in user
mode emulation. So the only access to be sequential atomic is:
* a single write in tb_flush();
* reads/writes out of 'tb_lock'.
In future, before enabling MTTCG in system mode, tb_flush() must be safe
and this field becomes unnecessary.
Backports commit 118b07308a8cedc16ef63d7ab243a95f1701db40 from qemu
Ensure atomicity of CPU's 'tb_jmp_cache' access for future translation
block lookup out of 'tb_lock'.
Note that this patch does *not* make CPU's TLB invalidation safe if it
is done from some other thread while the CPU is in its execution loop.
Backports commit 89a16b1e4294e3664667a151c2f70c84dfac6fd9 from qemu
In user-mode emulation Translation Block can consist of 2 guest pages.
In that case QEMU also mprotects 2 host pages that are dedicated for
guest memory, containing instructions. QEMU detects self-modifying code
with SEGFAULT signal processing.
In case if instruction in 1st page is modifying memory of 2nd
page (or vice versa) QEMU will mark 2nd page with PAGE_WRITE,
invalidate TB, generate new TB contatining 1 guest instruction and
exit to CPU loop. QEMU won't call mprotect, and new TB will cause
same SEGFAULT. Page will have both PAGE_WRITE_ORG and PAGE_WRITE
flags, so QEMU will handle the signal as guest binary problem,
and exit with guest SEGFAULT.
Solution is to do following: In case if current TB was invalidated
continue to invalidate TBs from remaining guest pages and mark pages
as PAGE_WRITE. After that disable host page protection with mprotect.
If current tb was invalidated longjmp to main loop. That is more
efficient, since we won't get SEGFAULT when executing new TB.
Backports commit 7399a337e4126f7c8c8af3336726f001378c4798 from qemu
Information is tracked inside the TCGContext structure, and later used
by tracing events with the 'tcg' and 'vcpu' properties.
The 'cpu' field is used to check tracing of translation-time
events ("*_trans"). The 'tcg_env' field is used to pass it to
execution-time events ("*_exec").
Backports commit 7c2550432abe62f53e6df878ceba6ceaf71f0e7e from qemu
For some workloads such as arm bootup, tb_phys_hash is performance-critical.
The is due to the high frequency of accesses to the hash table, originated
by (frequent) TLB flushes that wipe out the cpu-private tb_jmp_cache's.
More info:
https://lists.nongnu.org/archive/html/qemu-devel/2016-03/msg05098.html
To dig further into this I modified an arm image booting debian jessie to
immediately shut down after boot. Analysis revealed that quite a bit of time
is unnecessarily spent in tb_phys_hash: the cause is poor hashing that
results in very uneven loading of chains in the hash table's buckets;
the longest observed chain had ~550 elements.
The appended addresses this with two changes:
1) Use xxhash as the hash table's hash function. xxhash is a fast,
high-quality hashing function.
2) Feed the hashing function with not just tb_phys, but also pc and flags.
This improves performance over using just tb_phys for hashing, since that
resulted in some hash buckets having many TB's, while others getting very few;
with these changes, the longest observed chain on a single hash bucket is
brought down from ~550 to ~40.
Tests show that the other element checked for in tb_find_physical,
cs_base, is always a match when tb_phys+pc+flags are a match,
so hashing cs_base is wasteful. It could be that this is an ARM-only
thing, though. UPDATE:
On Tue, Apr 05, 2016 at 08:41:43 -0700, Richard Henderson wrote:
> The cs_base field is only used by i386 (in 16-bit modes), and sparc (for a TB
> consisting of only a delay slot).
> It may well still turn out to be reasonable to ignore cs_base for hashing.
BTW, after this change the hash table should not be called "tb_hash_phys"
anymore; this is addressed later in this series.
This change gives consistent bootup time improvements. I tested two
host machines:
- Intel Xeon E5-2690: 11.6% less time
- Intel i7-4790K: 19.2% less time
Increasing the number of hash buckets yields further improvements. However,
using a larger, fixed number of buckets can degrade performance for other
workloads that do not translate as many blocks (600K+ for debian-jessie arm
bootup). This is dealt with later in this series.
Backports commit 42bd32287f3a18d823f2258b813824a39ed7c6d9 from qemu
The function cpu_resume_from_signal() is now always called with a
NULL puc argument, and is rather misnamed since it is never called
from a signal handler. It is essentially forcing an exit to the
top level cpu loop but without raising any exception, so rename
it to cpu_loop_exit_noexc() and drop the useless unused argument.
Backports commit 6886b98036a8f8f5bce8b10756ce080084cef11b from qemu
Since the only caller of page_unprotect() which might cause it to
need to call cpu_resume_from_signal() is handle_cpu_signal() in
the user-mode code, push the longjump handling out to that function.
Since this is the only caller of cpu_resume_from_signal() which
passes a non-NULL puc argument, split the non-NULL handling into
a new cpu_exit_tb_from_sighandler() function. This allows us
to merge the softmmu and usermode implementations of the
cpu_resume_from_signal() function, which are now identical.
Backports commit f213e72f2356b77768b9cb73814a3b26ad5a0099 from qemu
The user-mode-only function tb_invalidate_phys_page() is only
called from two places:
* page_unprotect(), which passes in a non-zero pc, a puc pointer
and the value 'true' for the locked argument
* page_set_flags(), which passes in a zero pc, a NULL puc pointer
and a 'false' locked argument
If the pc is non-zero then we may call cpu_resume_from_signal(),
which does a longjmp out of the calling code (and out of the
signal handler); this is to cover the case of a target CPU with
"precise self-modifying code" (currently only x86) executing
a store instruction which modifies code in the same TB as the
store itself. Rather than doing the longjump directly here,
return a flag to the caller which indicates whether the current
TB was modified, and move the longjump to page_unprotect.
Backports commit 75809229bbf28b371afce14921ff5be98ddc5faa from qemu
mr->ram_block->offset is already aligned to both host and target size
(see qemu_ram_alloc_internal). Remove further masking as it is
unnecessary.
Backports commit e4e697940dff612b789b0858270c20a8b680f78d from qemu
exec-all.h contains TCG-specific definitions. It is not needed outside
TCG-specific files such as translate.c, exec.c or *helper.c.
One generic function had snuck into include/exec/exec-all.h; move it to
include/qom/cpu.h.
Backports commit 63c915526d6a54a95919ebece83fa9ca631b2508 from qemu
This field was used for telling cpu_interrupt() to unlink a chain of TBs
being executed when it worked that way. Now, cpu_interrupt() don't do
this anymore. So we don't need this field anymore.
Backports commit 3213525f8ab48742db09dab18cb9ae6f36a6c921 from qemu
'tb_invalidated_flag' was meant to catch two events:
* some TB has been invalidated by tb_phys_invalidate();
* the whole translation buffer has been flushed by tb_flush().
Then it was checked:
* in cpu_exec() to ensure that the last executed TB can be safely
linked to directly call the next one;
* in cpu_exec_nocache() to decide if the original TB should be provided
for further possible invalidation along with the temporarily
generated TB.
It is always safe to patch an invalidated TB since it is not going to be
used anyway. It is also safe to call tb_phys_invalidate() for an already
invalidated TB. Thus, setting this flag in tb_phys_invalidate() is
simply unnecessary. Moreover, it can prevent from pretty proper linking
of TBs, if any arbitrary TB has been invalidated. So just don't touch it
in tb_phys_invalidate().
If this flag is only used to catch whether tb_flush() has been called
then rename it to 'tb_flushed'. Declare it as 'bool' and stick to using
only 'true' and 'false' to set its value. Also, instead of setting it in
tb_gen_code(), just after tb_flush() has been called, do it right inside
of tb_flush().
In cpu_exec(), this flag is used to track if tb_flush() has been called
and have made 'next_tb' (a reference to the last executed TB) invalid
for linking it to directly call the next TB. tb_flush() can be called
during the CPU execution loop from tb_gen_code(), during TB execution or
by another thread while 'tb_lock' is released. Catch for translation
buffer flush reliably by resetting this flag once before first TB lookup
and each time we find it set before trying to add a direct jump. Don't
touch in in tb_find_physical().
Each vCPU has its own execution loop in multithreaded mode and thus
should have its own copy of the flag to be able to reset it with its own
'next_tb' and don't affect any other vCPU execution thread. So make this
flag per-vCPU and move it to CPUState.
In cpu_exec_nocache(), we only need to check if tb_flush() has been
called from tb_gen_code() called by cpu_exec_nocache() itself. To do
this reliably, preserve the old value of the flag, reset it before
calling tb_gen_code(), check afterwards, and combine the saved value
back to the flag.
This patch is based on the patch "tcg: move tb_invalidated_flag to
CPUState" from Paolo Bonzini <pbonzini@redhat.com>.
Backports commit 6f789be56d3f38e9214dafcfab3bf9be7191f370 from qemu
Unify the code of this function with tb_jmp_remove_from_list(). Making
these functions similar improves their readability. Also this could be a
step towards making this function thread-safe.
Backports commit f9c5b66f487a04d3747dc6997b1503f9258df945 from qemu
Move the code for removing jumps to a TB out of tb_phys_invalidate() to
a separate static inline function tb_jmp_unlink(). This simplifies
tb_phys_invalidate() and improves code structure.
Backports commit 89bba496322d4cf996d42cdd4bb0912231656c3d from qemu
tb_jmp_remove() was only used to remove the TB from a list of all TBs
jumping to the same TB which is n-th jump destination of the given TB.
Put a comment briefly describing the function behavior and rename it to
better reflect its purpose.
Backports commit 133626783aa5a1bf86332fa3e6f7b8efe005f924 from qemu
Initialize TB's direct jump list data fields and reset the jumps before
tb_link_page() puts it into the physical hash table and the physical
page list. So TB is completely initialized before it becomes visible.
This is pure rearrangement of code to a more suitable place, though it
could be a preparation for relaxing the locking scheme in future.
Backports commit 901bc3deb43bf37c85e43955905d003be7ae5fa5 from qemu
These fields do not contain pure pointers to a TranslationBlock
structure. So uintptr_t is the most appropriate type for them.
Also put some asserts to assure that the two least significant bits of
the pointer are always zero before assigning it to jmp_list_first.
Backports commit c37e6d7e3589ecb96914faa21025ad7ba6654aea from qemu
Briefly describe in a comment how direct block chaining is done. It
should help in understanding of the following data fields.
Rename some fields in TranslationBlock and TCGContext structures to
better reflect their purpose (dropping excessive 'tb_' prefix in
TranslationBlock but keeping it in TCGContext):
tb_next_offset => jmp_reset_offset
tb_jmp_offset => jmp_insn_offset
tb_next => jmp_target_addr
jmp_next => jmp_list_next
jmp_first => jmp_list_first
Avoid using a magic constant as an invalid offset which is used to
indicate that there's no n-th jump generated.
Backports commit f309101c26b59641fc1aa8fb2a98a5441cdaea03 from qemu
The setting of tcg_ctx.code_gen_buffer_size is done by the only caller of
size_code_gen_buffer(), which is code_gen_alloc():
$ git grep size_code_gen_buffer
translate-all.c:static inline size_t size_code_gen_buffer(size_t tb_size)
translate-all.c: tcg_ctx.code_gen_buffer_size = size_code_gen_buffer(tb_size);
Backports commit 835154b6e2200460f04719d0028716a37c178368 from qemu
We are inconsistent with the type of tb->flags: usage varies loosely
between int and uint64_t. Settle to uint32_t everywhere, which is
superior to both: at least one target (aarch64) uses the most significant
bit in the u32, and uint64_t is wasteful.
Compile-tested for all targets.
Backports commit 89fee74a0f066dfd73830a7b5fa137e87888c870 from qemu
Since 5e5f07e08 "TCG: Move translation block variables
to new context inside tcg_ctx: tb_ctx" on Feb 1 2013, compilation
of usermode + TB_DEBUG_CHECK has been broken. Fix it.
Backports commit 7e6bd36d61129feb7f667cb09ffec1b7b54b971c from qemu
qemu-log: dfilter-ise exec, out_asm, op and opt_op
This ensures the code generation debug code will honour -dfilter if set.
For the "exec" tracing I've added a new inline macro for efficiency's
sake.
Backports commit d977e1c2dbc9e63454b2000f91954d02543bf43b from qemu
