cosmopolitan/third_party/nsync/README.cosmo
Justine Tunney 2f48a02b44
Make recursive mutexes faster
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
2024-09-10 00:08:59 -07:00

47 lines
1.7 KiB
Text

DESCRIPTION
*NSYNC is a synchronization primitives library.
LICENSE
Apache 2.0
ORIGIN
git@github.com:google/nsync
commit ac5489682760393fe21bd2a8e038b528442412a7
Author: Mike Burrows <m3b@google.com>
Date: Wed Jun 1 16:47:52 2022 -0700
LOCAL CHANGES
- Fix nsync_mu_unlock() on Apple Silicon
- Add clock parameter to many NSYNC wait APIs
- Time APIs were so good that they're now in libc
- Double linked list API was so good that it's now in libc
- Max delay on sleep should be 20ms (not 4ms) on OpenBSD and NetBSD
- Support Apple's ulock futexes which are internal but nicer than GCD
- Ensure resources such as POSIX semaphores are are released on fork.
- Modified *NSYNC to allocate waiter objects on the stack. We need it
because we use *NSYNC mutexes to implement POSIX mutexes, which are
too low-level to safely depend on malloc, or even mmap in our case.
- Rewrote most of the semaphore and futex system call support code so
it works well with Cosmopolitan's fat runtime portability. *NSYNC's
unit test suite passes on all supported platforms. However the BSDs
currently appear to be overutilizing CPU time compared with others.
This appears to be the fault of the OSes rather than *NSYNC / Cosmo
- Support POSIX thread cancellation. APIs that wait on condition vars
are now cancellation points. In PTHREAD_CANCEL_MASKED mode they may
return ECANCELED. In PTHREAD_CANCEL_DEFERRED mode the POSIX threads
library will unwind the stack to re-acquire locks and free waiters.
On the other hand the *NSYNC APIs for mutexes will now safely block
thread cancellation, but you can still use *NSYNC notes to do that.