Recursive mutexes now go as fast as normal mutexes. The tradeoff is they
are no longer safe to use in signal handlers. However you can still have
signal safe mutexes if you set your mutex to both recursive and pshared.
You can also make functions that use recursive mutexes signal safe using
sigprocmask to ensure recursion doesn't happen due to any signal handler
The impact of this change is that, on Windows, many functions which edit
the file descriptor table rely on recursive mutexes, e.g. open(). If you
develop your app so it uses pread() and pwrite() then your app should go
very fast when performing a heavily multithreaded and contended workload
For example, when scaling to 40+ cores, *NSYNC mutexes can go as much as
1000x faster (in CPU time) than the naive recursive lock implementation.
Now recursive will use *NSYNC under the hood when it's possible to do so
It turns out sched_getcpu() didn't work on many platforms. So the system
call now has tests and is well documented. We now employ new workarounds
on platforms where it isn't supported in our malloc() implementation. It
was previously the case that malloc() was only scalable on Linux/Windows
for x86-64. Now the other platforms are scalable too.
This change solves an issue where many threads attempting to spawn forks
at once would cause fork() performance to degrade with the thread count.
Things got real nasty on NetBSD, which slowed down the whole test fleet,
because there's no vfork() and we're forced to use fork() in our server.
threads count task
1 1062 fork+exit+wait
2 668 fork+exit+wait
4 66 fork+exit+wait
8 19 fork+exit+wait
16 22 fork+exit+wait
32 16 fork+exit+wait
Things are now much less bad on NetBSD, but not great, since it does not
have futexes; we rely on its semaphore file descriptors to do conditions
threads count task
1 1085 fork+exit+wait
2 842 fork+exit+wait
4 532 fork+exit+wait
8 400 fork+exit+wait
16 276 fork+exit+wait
32 66 fork+exit+wait
With OpenBSD which also lacks vfork(), things were just as bad as NetBSD
threads count task
1 584 fork+exit+wait
2 687 fork+exit+wait
4 206 fork+exit+wait
8 24 fork+exit+wait
16 33 fork+exit+wait
32 26 fork+exit+wait
But since OpenBSD has futexes fork() works terrifically thanks to *NSYNC
threads count task
1 525 fork+exit+wait
2 580 fork+exit+wait
4 451 fork+exit+wait
8 479 fork+exit+wait
16 408 fork+exit+wait
32 373 fork+exit+wait
This issue would most likely only manifest itself, when pthread_atfork()
callers manage to slip a spin lock into the outermost position of fork's
list of locks. Since fork() is very slow, a spin lock can be devastating
Needless to say vfork() rules and anyone who says differently is kidding
themselves. Look at what a FreeBSD 14.1 virtual machine with equal specs
can do over the course of three hundred milliseconds.
threads count task
1 2559 vfork+exit+wait
2 5389 vfork+exit+wait
4 34933 vfork+exit+wait
8 43273 vfork+exit+wait
16 49648 vfork+exit+wait
32 40247 vfork+exit+wait
So it's a shame that so few OSes support vfork(). It creates an unsavory
situation, where someone wanting to build a server that spawns processes
would be better served to not use threads and favor a multiprocess model
Cosmopolitan now supports mremap(), which is only supported on Linux and
NetBSD. First, it allows memory mappings to be relocated without copying
them; this can dramatically speed up data structures like std::vector if
the array size grows larger than 256kb. The mremap() system call is also
10x faster than munmap() when shrinking large memory mappings.
There's now two functions, getpagesize() and getgransize() which help to
write portable code that uses mmap(MAP_FIXED). Alternative sysconf() may
be called with our new _SC_GRANSIZE. The madvise() system call now has a
better wrapper with improved documentation.
We now have a C++ red-black tree implementation that implements standard
template library compatible APIs while compiling 10x faster than libcxx.
It's not as beautiful as the red-black tree implementation in Plinko but
this will get the job done and the test proves it upholds all invariants
This change also restores CheckForMemoryLeaks() support and fixes a real
actual bug I discovered with Doug Lea's dlmalloc_inspect_all() function.
It hasn't been helpful enough to be justify the maintenance burden. What
actually does help is mprotect(), kprintf(), --ftrace and --strace which
can always be counted upon to work correctly. We aren't losing much with
this change. Support for ASAN on AARCH64 was never implemented. Applying
ASAN to the core libc runtimes was disabled many months ago. If there is
some way to have an ASAN runtime for user programs that is less invasive
we can potentially consider reintroducing support. But now is premature.
Actually Portable Executable now supports Android. Cosmo's old mmap code
required a 47 bit address space. The new implementation is very agnostic
and supports both smaller address spaces (e.g. embedded) and even modern
56-bit PML5T paging for x86 which finally came true on Zen4 Threadripper
Cosmopolitan no longer requires UNIX systems to observe the Windows 64kb
granularity; i.e. sysconf(_SC_PAGE_SIZE) will now report the host native
page size. This fixes a longstanding POSIX conformance issue, concerning
file mappings that overlap the end of file. Other aspects of conformance
have been improved too, such as the subtleties of address assignment and
and the various subtleties surrounding MAP_FIXED and MAP_FIXED_NOREPLACE
On Windows, mappings larger than 100 megabytes won't be broken down into
thousands of independent 64kb mappings. Support for MAP_STACK is removed
by this change; please use NewCosmoStack() instead.
Stack overflow avoidance is now being implemented using the POSIX thread
APIs. Please use GetStackBottom() and GetStackAddr(), instead of the old
error-prone GetStackAddr() and HaveStackMemory() APIs which are removed.
If pthread_create() is linked into the binary, then the cosmo runtime
will create an independent dlmalloc arena for each core. Whenever the
malloc() function is used it will index `g_heaps[sched_getcpu() / 2]`
to find the arena with the greatest hyperthread / numa locality. This
may be configured via an environment variable. For example if you say
`export COSMOPOLITAN_HEAP_COUNT=1` then you can restore the old ways.
Your process may be configured to have anywhere between 1 - 128 heaps
We need this revision because it makes multithreaded C++ applications
faster. For example, an HTTP server I'm working on that makes extreme
use of the STL went from 16k to 2000k requests per second, after this
change was made. To understand why, try out the malloc_test benchmark
which calls malloc() + realloc() in a loop across many threads, which
sees a a 250x improvement in process clock time and 200x on wall time
The tradeoff is this adds ~25ns of latency to individual malloc calls
compared to MODE=tiny, once the cosmo runtime has transitioned into a
fully multi-threaded state. If you don't need malloc() to be scalable
then cosmo provides many options for you. For starters the heap count
variable above can be set to put the process back in single heap mode
plus you can go even faster still, if you include tinymalloc.inc like
many of the programs in tool/build/.. are already doing since that'll
shave tens of kb off your binary footprint too. Theres also MODE=tiny
which is configured to use just 1 plain old dlmalloc arena by default
Another tradeoff is we need more memory now (except in MODE=tiny), to
track the provenance of memory allocation. This is so allocations can
be freely shared across threads, and because OSes can reschedule code
to different CPUs at any time.
This change adds a TLS freelist for small dynamic memory allocations.
Cosmopolitan's TIB is now 512 bytes in size. Single-threaded malloc()
performance isn't impacted by this, until pthread_create() is called.
Single-threaded programs may also want to consider using:
#include "libc/mem/tinymalloc.inc"
Which will shave 30k off the executable size and sometimes go faster.
Now that these functions are behind _COSMO_SOURCE there's no reason for
having the ugly underscore anymore. To use these functions, you need to
pass -mcosmo to cosmocc.
This change upgrades to GCC 12.3 and GNU binutils 2.42. The GNU linker
appears to have changed things so that only a single de-duplicated str
table is present in the binary, and it gets placed wherever the linker
wants, regardless of what the linker script says. To cope with that we
need to stop using .ident to embed licenses. As such, this change does
significant work to revamp how third party licenses are defined in the
codebase, using `.section .notice,"aR",@progbits`.
This new GCC 12.3 toolchain has support for GNU indirect functions. It
lets us support __target_clones__ for the first time. This is used for
optimizing the performance of libc string functions such as strlen and
friends so far on x86, by ensuring AVX systems favor a second codepath
that uses VEX encoding. It shaves some latency off certain operations.
It's a useful feature to have for scientific computing for the reasons
explained by the test/libcxx/openmp_test.cc example which compiles for
fifteen different microarchitectures. Thanks to the upgrades, it's now
also possible to use newer instruction sets, such as AVX512FP16, VNNI.
Cosmo now uses the %gs register on x86 by default for TLS. Doing it is
helpful for any program that links `cosmo_dlopen()`. Such programs had
to recompile their binaries at startup to change the TLS instructions.
That's not great, since it means every page in the executable needs to
be faulted. The work of rewriting TLS-related x86 opcodes, is moved to
fixupobj.com instead. This is great news for MacOS x86 users, since we
previously needed to morph the binary every time for that platform but
now that's no longer necessary. The only platforms where we need fixup
of TLS x86 opcodes at runtime are now Windows, OpenBSD, and NetBSD. On
Windows we morph TLS to point deeper into the TIB, based on a TlsAlloc
assignment, and on OpenBSD/NetBSD we morph %gs back into %fs since the
kernels do not allow us to specify a value for the %gs register.
OpenBSD users are now required to use APE Loader to run Cosmo binaries
and assimilation is no longer possible. OpenBSD kernel needs to change
to allow programs to specify a value for the %gs register, or it needs
to stop marking executable pages loaded by the kernel as mimmutable().
This release fixes __constructor__, .ctor, .init_array, and lastly the
.preinit_array so they behave the exact same way as glibc.
We no longer use hex constants to define math.h symbols like M_PI.
Somehow or another, I previously had missed `BUILD.mk` files.
In the process I found a few straggler cases where the modeline was
different from the file, including one very involved manual fix where a
file had been treated like it was ts=2 and ts=8 on separate occasions.
The commit history in the PR shows the gory details; the BUILD.mk was
automated, everything else was mostly manual.
At least in neovim, `│vi:` is not recognized as a modeline because it
has no preceding whitespace. After fixing this, opening a file yields
an error because `net` is not an option. (`noet`, however, is.)
We torture test dlmalloc() in test/libc/stdio/memory_test.c. That test
was crashing on occasion on Apple M1 microprocessors when dlmalloc was
using *NSYNC locks. It was relatively easy to spot the cause, which is
this one particular compare and swap operation, which needed to change
to use sequentially-consistent ordering rather than an acquire barrier
Our makefile generator now accepts badly formatted include lines. It's
now more hermetic with better error checking in the cosmo repo, and it
can be configured to not be hermetic at all.
- On Windows connect() can now be interrupted by a signal; connect() w/
O_NONBLOCK will now raise EINPROGRESS; and connect() with SO_SNDTIMEO
will raise ETIMEDOUT after the interval has elapsed.
- We now get the AcceptEx(), ConnectEx(), and TransmitFile() functions
from the WIN32 API the officially blessed way, using WSAIoctl().
- Do nothing on Windows when fsync() is called on a directory handle.
This was raising EACCES earlier becaues GENERIC_WRITE is required on
the handle. It's possible to FlushFileBuffers() a directory handle if
it's opened with write access but MSDN doesn't document what it does.
If you have any idea, please let us know!
- Prefer manual reset event objects for read() and write() on Windows.
- Do some code cleanup on our dlmalloc customizations.
- Fix errno type error in Windows blocking routines.
- Make the futex polyfill simpler and faster.
- Every unit test now passes on Apple Silicon. The final piece of this
puzzle was porting our POSIX threads cancelation support, since that
works differently on ARM64 XNU vs. AMD64. Our semaphore support on
Apple Silicon is also superior now compared to AMD64, thanks to the
grand central dispatch library which lets *NSYNC locks go faster.
- The Cosmopolitan runtime is now more stable, particularly on Windows.
To do this, thread local storage is mandatory at all runtime levels,
and the innermost packages of the C library is no longer being built
using ASAN. TLS is being bootstrapped with a 128-byte TIB during the
process startup phase, and then later on the runtime re-allocates it
either statically or dynamically to support code using _Thread_local.
fork() and execve() now do a better job cooperating with threads. We
can now check how much stack memory is left in the process or thread
when functions like kprintf() / execve() etc. call alloca(), so that
ENOMEM can be raised, reduce a buffer size, or just print a warning.
- POSIX signal emulation is now implemented the same way kernels do it
with pthread_kill() and raise(). Any thread can interrupt any other
thread, regardless of what it's doing. If it's blocked on read/write
then the killer thread will cancel its i/o operation so that EINTR can
be returned in the mark thread immediately. If it's doing a tight CPU
bound operation, then that's also interrupted by the signal delivery.
Signal delivery works now by suspending a thread and pushing context
data structures onto its stack, and redirecting its execution to a
trampoline function, which calls SetThreadContext(GetCurrentThread())
when it's done.
- We're now doing a better job managing locks and handles. On NetBSD we
now close semaphore file descriptors in forked children. Semaphores on
Windows can now be canceled immediately, which means mutexes/condition
variables will now go faster. Apple Silicon semaphores can be canceled
too. We're now using Apple's pthread_yield() funciton. Apple _nocancel
syscalls are now used on XNU when appropriate to ensure pthread_cancel
requests aren't lost. The MbedTLS library has been updated to support
POSIX thread cancelations. See tool/build/runitd.c for an example of
how it can be used for production multi-threaded tls servers. Handles
on Windows now leak less often across processes. All i/o operations on
Windows are now overlapped, which means file pointers can no longer be
inherited across dup() and fork() for the time being.
- We now spawn a thread on Windows to deliver SIGCHLD and wakeup wait4()
which means, for example, that posix_spawn() now goes 3x faster. POSIX
spawn is also now more correct. Like Musl, it's now able to report the
failure code of execve() via a pipe although our approach favors using
shared memory to do that on systems that have a true vfork() function.
- We now spawn a thread to deliver SIGALRM to threads when setitimer()
is used. This enables the most precise wakeups the OS makes possible.
- The Cosmopolitan runtime now uses less memory. On NetBSD for example,
it turned out the kernel would actually commit the PT_GNU_STACK size
which caused RSS to be 6mb for every process. Now it's down to ~4kb.
On Apple Silicon, we reduce the mandatory upstream thread size to the
smallest possible size to reduce the memory overhead of Cosmo threads.
The examples directory has a program called greenbean which can spawn
a web server on Linux with 10,000 worker threads and have the memory
usage of the process be ~77mb. The 1024 byte overhead of POSIX-style
thread-local storage is now optional; it won't be allocated until the
pthread_setspecific/getspecific functions are called. On Windows, the
threads that get spawned which are internal to the libc implementation
use reserve rather than commit memory, which shaves a few hundred kb.
- sigaltstack() is now supported on Windows, however it's currently not
able to be used to handle stack overflows, since crash signals are
still generated by WIN32. However the crash handler will still switch
to the alt stack, which is helpful in environments with tiny threads.
- Test binaries are now smaller. Many of the mandatory dependencies of
the test runner have been removed. This ensures many programs can do a
better job only linking the the thing they're testing. This caused the
test binaries for LIBC_FMT for example, to decrease from 200kb to 50kb
- long double is no longer used in the implementation details of libc,
except in the APIs that define it. The old code that used long double
for time (instead of struct timespec) has now been thoroughly removed.
- ShowCrashReports() is now much tinier in MODE=tiny. Instead of doing
backtraces itself, it'll just print a command you can run on the shell
using our new `cosmoaddr2line` program to view the backtrace.
- Crash report signal handling now works in a much better way. Instead
of terminating the process, it now relies on SA_RESETHAND so that the
default SIG_IGN behavior can terminate the process if necessary.
- Our pledge() functionality has now been fully ported to AARCH64 Linux.
This change fixes Cosmopolitan so it has fewer opinions about compiler
warnings. The whole repository had to be cleaned up to be buildable in
-Werror -Wall mode. This lets us benefit from things like strict const
checking. Some actual bugs might have been caught too.
- Remove PAGESIZE constant
- Fix realloc() documentation
- Fix ttyname_r() error reporting
- Make forking more reliable on Windows
- Make execvp() a few microseconds faster
- Make system() a few microseconds faster
- Tighten up the socket-related magic numbers
- Loosen restrictions on mmap() offset alignment
- Improve GetProgramExecutableName() with getenv("_")
- Use mkstemp() as basis for mktemp(), tmpfile(), tmpfd()
- Fix flakes in pthread_cancel_test, unix_test, fork_test
- Fix recently introduced futex stack overflow regression
- Let sockets be passed as stdio to subprocesses on Windows
- Improve security of bind() on Windows w/ SO_EXCLUSIVEADDRUSE
This change ports APE Loader to Linux AARCH64, so that Raspberry Pi
users can run programs like redbean, without the executable needing
to modify itself. Progress has also slipped into this change on the
issue of making progress better conforming to user expectations and
industry standards regarding which symbols we're allowed to declare
- Found some bugs in LLVM compiler-rt library
- The useless LIBC_STUBS package is now deleted
- Improve the overflow checking story even further
- Get chibicc tests working in MODE=dbg mode again
- The libc/isystem/ headers now have correctly named guards
This change takes an entirely new approach to the incremental linking of
pkzip executables. The assets created by zipobj.com are now treated like
debug data. After a .com.dbg is compiled, fixupobj.com should be run, so
it can apply fixups to the offsets and move the zip directory to the end
of the file. Since debug data doesn't get objcopy'd, a new tool has been
introduced called zipcopy.com which should be run after objcopy whenever
a .com file is created. This is all automated by the `cosmocc` toolchain
which is rapidly becoming the new recommended approach.
This change also introduces the new C23 checked arithmetic macros.
This change implements a new approach to function call logging, that's
based on the GCC flag: -fpatchable-function-entry. Read the commentary
in build/config.mk to learn how it works.
- ASAN memory morgue is now lockless
- Make C11 atomics header more portable
- Rewrote pthread keys support to be lockless
- Simplify Python's unicode table unpacking code
- Make crash report write(2) closer to being atomic
- Make it possible to strace/ftrace a single thread
- ASAN now checks nul-terminated strings fast and properly
- Windows fork() now restores TLS memory of calling thread
