Currently, we check the values on the machine stack & register state to
see if they're actually a pointer to an ASAN fake stack, and mark the
values on the fake stack too if required. However, we are only doing
that for the _current_ thread (the one actually running the GC), not for
any other thread in the program.
Make rb_gc_mark_machine_context (which is called for marking non-current
threads) perform the same ASAN fake stack handling that
mark_current_machine_context performs.
[Bug #20310]
This `st_table` is used to both mark and pin classes
defined from the C API. But `vm->mark_object_ary` already
does both much more efficiently.
Currently a Ruby process starts with 252 rooted classes,
which uses `7224B` in an `st_table` or `2016B` in an `RArray`.
So a baseline of 5kB saved, but since `mark_object_ary` is
preallocated with `1024` slots but only use `405` of them,
it's a net `7kB` save.
`vm->mark_object_ary` is also being refactored.
Prior to this changes, `mark_object_ary` was a regular `RArray`, but
since this allows for references to be moved, it was marked a second
time from `rb_vm_mark()` to pin these objects.
This has the detrimental effect of marking these references on every
minors even though it's a mostly append only list.
But using a custom TypedData we can save from having to mark
all the references on minor GC runs.
Addtionally, immediate values are now ignored and not appended
to `vm->mark_object_ary` as it's just wasted space.
Rather than exposing that an imemo has a flag and four fields, this
changes the implementation to only expose one field (the klass) and
fills the rest with 0. The type will have to fill in the values themselves.
Previously every call to vm_ci_new (when the CI was not packable) would
result in a different callinfo being returned this meant that every
kwarg callsite had its own CI.
When calling, different CIs result in different CCs. These CIs and CCs
both end up persisted on the T_CLASS inside cc_tbl. So in an eval loop
this resulted in a memory leak of both types of object. This also likely
resulted in extra memory used, and extra time searching, in non-eval
cases.
For simplicity in this commit I always allocate a CI object inside
rb_vm_ci_lookup, but ideally we would lazily allocate it only when
needed. I hope to do that as a follow up in the future.
Ruby makes it easy to delegate all arguments from one method to another:
```ruby
def f(*args, **kw)
g(*args, **kw)
end
```
Unfortunately, this indirection decreases performance. One reason it
decreases performance is that this allocates an array and a hash per
call to `f`, even if `args` and `kw` are not modified.
Due to Ruby's ability to modify almost anything at runtime, it's
difficult to avoid the array allocation in the general case. For
example, it's not safe to avoid the allocation in a case like this:
```ruby
def f(*args, **kw)
foo(bar)
g(*args, **kw)
end
```
Because `foo` may be `eval` and `bar` may be a string referencing `args`
or `kw`.
To fix this correctly, you need to perform something similar to escape
analysis on the variables. However, there is a case where you can
avoid the allocation without doing escape analysis, and that is when
the splat variables are anonymous:
```ruby
def f(*, **)
g(*, **)
end
```
When splat variables are anonymous, it is not possible to reference
them directly, it is only possible to use them as splats to other
methods. Since that is the case, if `f` is called with a regular
splat and a keyword splat, it can pass the arguments directly to
`g` without copying them, avoiding allocation. For example:
```ruby
def g(a, b:)
a + b
end
def f(*, **)
g(*, **)
end
a = [1]
kw = {b: 2}
f(*a, **kw)
```
I call this technique: Allocationless Anonymous Splat Forwarding.
This is implemented using a couple additional iseq param flags,
anon_rest and anon_kwrest. If anon_rest is set, and an array splat
is passed when calling the method when the array splat can be used
without modification, `setup_parameters_complex` does not duplicate
it. Similarly, if anon_kwest is set, and a keyword splat is passed
when calling the method, `setup_parameters_complex` does not
duplicate it.
The order of iseq may differ from the order of tokens, typically
`while`/`until` conditions are put after the body.
These orders can match by using line numbers as builtin-indexes, but
at the same time, it introduces the restriction that multiple `cexpr!`
and `cstmt!` cannot appear in the same line.
Another possible idea is to use `RubyVM::AbstractSyntaxTree` and
`node_id` instead of ripper, with making BASERUBY 3.1 or later.
ASAN leaves a pointer to the fake frame on the stack; we can use the
__asan_addr_is_in_fake_stack API to work out the extent of the fake
stack and thus mark any VALUEs contained therein.
[Bug #20001]
This commit changes how stack extents are calculated for both the main
thread and other threads. Ruby uses the address of a local variable as
part of the calculation for machine stack extents:
* pthreads uses it as a lower-bound on the start of the stack, because
glibc (and maybe other libcs) can store its own data on the stack
before calling into user code on thread creation.
* win32 uses it as an argument to VirtualQuery, which gets the extent of
the memory mapping which contains the variable
However, the local being used for this is actually too low (too close to
the leaf function call) in both the main thread case and the new thread
case.
In the main thread case, we have the `INIT_STACK` macro, which is used
for pthreads to set the `native_main_thread->stack_start` value. This
value is correctly captured at the very top level of the program (in
main.c). However, this is _not_ what's used to set the execution context
machine stack (`th->ec->machine_stack.stack_start`); that gets set as
part of a call to `ruby_thread_init_stack` in `Init_BareVM`, using the
address of a local variable allocated _inside_ `Init_BareVM`. This is
too low; we need to use a local allocated closer to the top of the
program.
In the new thread case, the lolcal is allocated inside
`native_thread_init_stack`, which is, again, too low.
In both cases, this means that we might have VALUEs lying outside the
bounds of `th->ec->machine.stack_{start,end}`, which won't be marked
correctly by the GC machinery.
To fix this,
* In the main thread case: We already have `INIT_STACK` at the right
level, so just pass that local var to `ruby_thread_init_stack`.
* In the new thread case: Allocate the local one level above the call to
`native_thread_init_stack` in `call_thread_start_func2`.
[Bug #20001]
fix
to BUILTIN_ATTR_SINGLE_NOARG_LEAF
The attribute was created when the other attribute was called BUILTIN_ATTR_INLINE.
Now that the original attribute is renamed to BUILTIN_ATTR_LEAF, it's
only confusing that we call it "_INLINE".
ASAN leaves a pointer to the fake frame on the stack; we can use the
__asan_addr_is_in_fake_stack API to work out the extent of the fake
stack and thus mark any VALUEs contained therein.
[Bug #20001]
The implementation of `native_thread_init_stack` for the various
threading models can use the address of a local variable as part of the
calculation of the machine stack extents:
* pthreads uses it as a lower-bound on the start of the stack, because
glibc (and maybe other libcs) can store its own data on the stack
before calling into user code on thread creation.
* win32 uses it as an argument to VirtualQuery, which gets the extent of
the memory mapping which contains the variable
However, the local being used for this is actually allocated _inside_
the `native_thread_init_stack` frame; that means the caller might
allocate a VALUE on the stack that actually lies outside the bounds
stored in machine.stack_{start,end}.
A local variable from one level above the topmost frame that stores
VALUEs on the stack must be drilled down into the call to
`native_thread_init_stack` to be used in the calculation. This probably
doesn't _really_ matter for the win32 case (they'll be in the same
memory mapping so VirtualQuery should return the same thing), but
definitely could matter for the pthreads case.
[Bug #20001]
Before this patch, the MN scheduler waits for the IO with the
following steps:
1. `poll(fd, timeout=0)` to check fd is ready or not.
2. if fd is not ready, waits with MN thread scheduler
3. call `func` to issue the blocking I/O call
The advantage of advanced `poll()` is we can wait for the
IO ready for any fds. However `poll()` becomes overhead
for already ready fds.
This patch changes the steps like:
1. call `func` to issue the blocking I/O call
2. if the `func` returns `EWOULDBLOCK` the fd is `O_NONBLOCK`
and we need to wait for fd is ready so that waits with MN
thread scheduler.
In this case, we can wait only for `O_NONBLOCK` fds. Otherwise
it waits with blocking operations such as `read()` system call.
However we don't need to call `poll()` to check fd is ready
in advance.
With this patch we can observe performance improvement
on microbenchmark which repeats blocking I/O (not
`O_NONBLOCK` fd) with and without MN thread scheduler.
```ruby
require 'benchmark'
f = open('/dev/null', 'w')
f.sync = true
TN = 1
N = 1_000_000 / TN
Benchmark.bm{|x|
x.report{
TN.times.map{
Thread.new{
N.times{f.print '.'}
}
}.each(&:join)
}
}
__END__
TN = 1
user system total real
ruby32 0.393966 0.101122 0.495088 ( 0.495235)
ruby33 0.493963 0.089521 0.583484 ( 0.584091)
ruby33+MN 0.639333 0.200843 0.840176 ( 0.840291) <- Slow
this+MN 0.512231 0.099091 0.611322 ( 0.611074) <- Good
```
Our current implementation of rb_postponed_job_register suffers from
some safety issues that can lead to interpreter crashes (see bug #1991).
Essentially, the issue is that jobs can be called with the wrong
arguments.
We made two attempts to fix this whilst keeping the promised semantics,
but:
* The first one involved masking/unmasking when flushing jobs, which
was believed to be too expensive
* The second one involved a lock-free, multi-producer, single-consumer
ringbuffer, which was too complex
The critical insight behind this third solution is that essentially the
only user of these APIs are a) internal, or b) profiling gems.
For a), none of the usages actually require variable data; they will
work just fine with the preregistration interface.
For b), generally profiling gems only call a single callback with a
single piece of data (which is actually usually just zero) for the life
of the program. The ringbuffer is complex because it needs to support
multi-word inserts of job & data (which can't be atomic); but nobody
actually even needs that functionality, really.
So, this comit:
* Introduces a pre-registration API for jobs, with a GVL-requiring
rb_postponed_job_prereigster, which returns a handle which can be
used with an async-signal-safe rb_postponed_job_trigger.
* Deprecates rb_postponed_job_register (and re-implements it on top of
the preregister function for compatability)
* Moves all the internal usages of postponed job register
pre-registration
This patch introduces thread specific storage APIs
for tools which use `rb_internal_thread_event_hook` APIs.
* `rb_internal_thread_specific_key_create()` to create a tool specific
thread local storage key and allocate the storage if not available.
* `rb_internal_thread_specific_set()` sets a data to thread and tool
specific storage.
* `rb_internal_thread_specific_get()` gets a data in thread and tool
specific storage.
Note that `rb_internal_thread_specific_get|set(thread_val, key)`
can be called without GVL and safe for async signal and safe for
multi-threading (native threads). So you can call it in any internal
thread event hooks. Further more you can call it from other native
threads. Of course `thread_val` should be living while accessing the
data from this function.
Note that you should not forget to clean up the set data.
`rb_vm_tag_jmpbuf_{init,deinit}` are safe to raise exception since the
given tag is not yet pushed to `ec->tag` or already popped from it at
the time, so `ec->tag` is always valid and it's safe to raise exception
when xmalloc fails.
`rb_jmpbuf_t` type is considerably large due to inline-allocated
Asyncify buffer, and it leads to stack overflow even with small number
of C-method call frames. This commit allocates the Asyncify buffer used
by `rb_wasm_setjmp` in heap to mitigate the issue.
This patch introduces a new type `rb_vm_tag_jmpbuf_t` to abstract the
representation of a jump buffer, and init/deinit hook points to manage
lifetime of the buffer. These changes are effectively NFC for non-wasm
platforms.
* Port call threshold logic from Rust to C for performance
* Prefix global/field names with yjit_
* Fix linker error
* Fix preprocessor condition for rb_yjit_threshold_hit
* Fix third linker issue
* Exclude yjit_calls_at_interv from RJIT bindgen
---------
Co-authored-by: Takashi Kokubun <takashikkbn@gmail.com>
This patch introduce M:N thread scheduler for Ractor system.
In general, M:N thread scheduler employs N native threads (OS threads)
to manage M user-level threads (Ruby threads in this case).
On the Ruby interpreter, 1 native thread is provided for 1 Ractor
and all Ruby threads are managed by the native thread.
From Ruby 1.9, the interpreter uses 1:1 thread scheduler which means
1 Ruby thread has 1 native thread. M:N scheduler change this strategy.
Because of compatibility issue (and stableness issue of the implementation)
main Ractor doesn't use M:N scheduler on default. On the other words,
threads on the main Ractor will be managed with 1:1 thread scheduler.
There are additional settings by environment variables:
`RUBY_MN_THREADS=1` enables M:N thread scheduler on the main ractor.
Note that non-main ractors use the M:N scheduler without this
configuration. With this configuration, single ractor applications
run threads on M:1 thread scheduler (green threads, user-level threads).
`RUBY_MAX_CPU=n` specifies maximum number of native threads for
M:N scheduler (default: 8).
This patch will be reverted soon if non-easy issues are found.
[Bug #19842]
Previously, Kernel#lambda returned a non-lambda proc when given a
non-literal block and issued a warning under the `:deprecated` category.
With this change, Kernel#lambda will always return a lambda proc, if it
returns without raising.
Due to interactions with block passing optimizations, we previously had
two separate code paths for detecting whether Kernel#lambda got a
literal block. This change allows us to remove one path, the hack done
with rb_control_frame_t::block_code introduced in 85a337f for supporting
situations where Kernel#lambda returned a non-lambda proc.
[Feature #19777]
Co-authored-by: Takashi Kokubun <takashikkbn@gmail.com>
`struct rb_calling_info::cd` is introduced and `rb_calling_info::ci`
is replaced with it to manipulate the inline cache of iseq while
method invocation process. So that `ci` can be acessed with
`calling->cd->ci`. It adds one indirection but it can be justified
by the following points:
1) `vm_search_method_fastpath()` doesn't need `ci` and also
`vm_call_iseq_setup_normal()` doesn't need `ci`. It means
reducing `cd->ci` access in `vm_sendish()` can make it faster.
2) most of method types need to access `ci` once in theory
so that 1 additional indirection doesn't matter.
Remove rb_control_frame_t::__bp__ and optimize bmethod calls
This commit removes the __bp__ field from rb_control_frame_t. It was
introduced to help MJIT, but since MJIT was replaced by RJIT, we can use
vm_base_ptr() to compute it from the SP of the previous control frame
instead. Removing the field avoids needing to set it up when pushing new
frames.
Simply removing __bp__ would cause crashes since RJIT and YJIT used a
slightly different stack layout for bmethod calls than the interpreter.
At the moment of the call, the two layouts looked as follows:
┌────────────┐ ┌────────────┐
│ frame_base │ │ frame_base │
├────────────┤ ├────────────┤
│ ... │ │ ... │
├────────────┤ ├────────────┤
│ args │ │ args │
├────────────┤ └────────────┘<─prev_frame_sp
│ receiver │
prev_frame_sp─>└────────────┘
RJIT & YJIT interpreter
Essentially, vm_base_ptr() needs to compute the address to frame_base
given prev_frame_sp in the diagrams. The presence of the receiver
created an off-by-one situation.
Make the interpreter use the layout the JITs use for iseq-to-iseq
bmethod calls. Doing so removes unnecessary argument shifting and
vm_exec_core() re-entry from the interpreter, yielding a speed
improvement visible through `benchmark/vm_defined_method.yml`:
patched: 7578743.1 i/s
master: 4796596.3 i/s - 1.58x slower
C-to-iseq bmethod calls now store one more VALUE than before, but that
should have negligible impact on overall performance.
Note that re-entering vm_exec_core() used to be necessary for firing
TracePoint events, but that's no longer the case since
9121e57a5f.
Closesruby/ruby#6428
After [1], using ext/Setup to link some, but not all extensions failed
during linking. I did not know about this option, and had assumed that
only `--with-static-linked-ext` builds can include statically linked
extensions.
Include the support code for statically linked extensions in all
configurations like before [1]. Initialize the table lazily to minimize
footprint on builds that have no statically linked extensions.
[1]: 790cf4b6d0 "Fix autoload status of
statically linked extensions"
Originally, when 2e7bceb34e fixed cfuncs to no
longer use the VM stack for large array splats, it was thought to have fully
fixed Bug #4040, since the issue was fixed for methods defined in Ruby (iseqs)
back in Ruby 2.2.
After additional research, I determined that same issue affects almost all
types of method calls, not just iseq and cfunc calls. There were two main
types of remaining issues, important cases (where large array splat should
work) and pedantic cases (where large array splat raised SystemStackError
instead of ArgumentError).
Important cases:
```ruby
define_method(:a){|*a|}
a(*1380888.times)
def b(*a); end
send(:b, *1380888.times)
:b.to_proc.call(self, *1380888.times)
def d; yield(*1380888.times) end
d(&method(:b))
def self.method_missing(*a); end
not_a_method(*1380888.times)
```
Pedantic cases:
```ruby
def a; end
a(*1380888.times)
def b(_); end
b(*1380888.times)
def c(_=nil); end
c(*1380888.times)
c = Class.new do
attr_accessor :a
alias b a=
end.new
c.a(*1380888.times)
c.b(*1380888.times)
c = Struct.new(:a) do
alias b a=
end.new
c.a(*1380888.times)
c.b(*1380888.times)
```
This patch fixes all usage of CALLER_SETUP_ARG with splatting a large
number of arguments, and required similar fixes to use a temporary
hidden array in three other cases where the VM would use the VM stack
for handling a large number of arguments. However, it is possible
there may be additional cases where splatting a large number
of arguments still causes a SystemStackError.
This has a measurable performance impact, as it requires additional
checks for a large number of arguments in many additional cases.
This change is fairly invasive, as there were many different VM
functions that needed to be modified to support this. To avoid
too much API change, I modified struct rb_calling_info to add a
heap_argv member for storing the array, so I would not have to
thread it through many functions. This struct is always stack
allocated, which helps ensure sure GC doesn't collect it early.
Because of how invasive the changes are, and how rarely large
arrays are actually splatted in Ruby code, the existing test/spec
suites are not great at testing for correct behavior. To try to
find and fix all issues, I tested this in CI with
VM_ARGC_STACK_MAX to -1, ensuring that a temporary array is used
for all array splat method calls. This was very helpful in
finding breaking cases, especially ones involving flagged keyword
hashes.
Fixes [Bug #4040]
Co-authored-by: Jimmy Miller <jimmy.miller@shopify.com>
Rebuilding the loaded feature index slowed down with the bug fix
for #17885 in 79a4484a07. The
slowdown was extreme if realpath emulation was used, but even when
not emulated, it could be about 10x slower.
This adds loaded_features_realpath_map to rb_vm_struct. This is a
hidden hash mapping loaded feature paths to realpaths. When
rebuilding the loaded feature index, look at this hash to get
cached realpath values, and skip calling rb_check_realpath if a
cached value is found.
Fixes [Bug #19246]
The `catch_except_p` flag is used for communicating between parent and
child iseq's that a throw instruction was emitted. So for example if a
child iseq has a throw in it and the parent wants to catch the throw, we
use this flag to communicate to the parent iseq that a throw instruction
was emitted.
This flag is only useful at compile time, it only impacts the
compilation process so it seems to be fine to move it from the iseq body
to the compile_data struct.
Co-authored-by: Aaron Patterson <tenderlove@ruby-lang.org>
`rb_current_ractor()` expects it has valid `ec` and `r`.
`rb_current_ractor_raw()` with a parameter `false` allows to return
NULL if `ec` is not available.
KJ Tsanaktsidis
Peter Zhu
Jean Boussier
Kevin Newton
John Hawthorn
Nobuyoshi Nakada
Jeremy Evans
Takashi Kokubun
KJ Tsanaktsidis
Koichi Sasada
Yuta Saito
Maxime Chevalier-Boisvert
Alan Wu
Samuel Williams
eileencodes
Matt Valentine-House