RIP pthread_cancel
eissing.org
eissing.org
Ironically, our application has also struggled with blocking DNS resolution in the past, so I appreciate the discussion here. In case anyone is interested, here is a quick reference of the different asynchronous DNS resolution methods that you can use in a native code application for some of the most popular platforms:
- Windows/Xbox: GetAddrInfoExW / GetAddrInfoExCancel
- macOS/iOS: CFHostStartInfoResolution / CFHostCancelInfoResolution
- Linux (glibc): getaddrinfo_a / gai_cancel
- Android: android.net.DnsResolver.query (requires the use of JNI)
- PS5: proprietary DNS resolver API - OpenBSD: getaddrinfo_async / asr_abort https://man.openbsd.org/asr_run
- FreeBSD: dnsres_getaddrinfo / found no way to abort https://man.freebsd.org/cgi/man.cgi?query=dnsresOn iOS, its use triggers a local network access popup (it tries to reach your DNS server, which is often on your LAN). If a user denies acess, your app will simply not work.
On Android, it's not compatible with some VPN apps. Those apps are to blame, but your users are going to blame you not them.
So, at my previous company we ended up building libcurl with a threaded DNS resolver on both iOS and Android.
Yes I know in Linux you can set the timeout in a config file.
But really the dns setting should be configurable by the calling code. Some code requires fast lookups and doesn’t mind failing which, while others won’t mind waiting longer. It’s not a one size fits all thing.
The solution is to make it a properly non-blocking api.
As always, on non-Linux Unixen the answer is "screw you!"
There is also POSIX AIO for async I/O on any file descriptor, but at least historically speaking it doesn't work properly on Linux.
Until io_uring the only asynchronous disk IO interface was the io_* syscalls, which were confusingly referred to as Asynchronous IO, though these have nothing to do with POSIX AIO and can only be used bypassing the page cache, and suck for general purpose use.
The first is that getaddrinfo is specified by POSIX, and the POSIX evolve very conservatively and at a glacial pace.
The second reason is that specifying a timeout breaks symmetry with a lot of other functions in Unix/C, both system calls and libc calls. For example, you can't specify a timeout when opening a file, reading from a file, or closing a file, which are all potentially blocking operations. There are ways to do these things in a non-blocking manner with timeouts using aio or io_uring, but those are already relatively complicated APIs for just using simple system calls, and getaddrinfo is much more complicated.
The last reason is that if you use the sockets APIs directly it's not that hard to write a non-blocking DNS resolver (c-ares is one example). The thing is though that if you write your own resolver you have to consider how to do caching, it won't work with NSS on Linux, etc. You can implement these things (systemd-resolved does it, and works with NSS) but they are a lot of work to do properly.
No they're not. Not really, unless you consider disk access and interacting with the page cache/inode cache inside the kernel to be blocking. But if you do that, you should probably also consider scheduling and really any CPU instruction to be blocking. (If the system is too loaded, anything can be slow).
To be fair, network requests can be considered non-blocking in a similar way, but they're depending on other systems that you generally can't control or inspect. In practice you'll see network timeouts. Note that you (at least normally -- there might be tricky exceptions) won't see EINTR from read() to a filesystem file. But you can see EINTR for network sockets. The difference is that, in Unix terminology, disks are not considered "slow devices".
And neither are the tapes. But the pipes, apparently, are.
Well, unfortunately, disk^H^H^H^H large persistent storage I/O is actually slow, or people wouldn't have been writing thread-pools to make it look asynchrnous, or sometimes even process-pools to convert disk I/O to pipe I/O, for the last two decades.
A pipe on the other hand might easily stall for an hour. The kernel generally can't know how long it will have to wait for more data. That's why pipe reads (as well as writes) are interruptible.
The absolute bandwidth number of a harddisk doesn't matter --- in principle you can overload any system such that it fails to schedule and complete all processes in time. Putting aside possible system overload, the "slow device" terminology makes a lot of sense.
I was under the impression that seeking a disk you can generally calculate well with 10ms? Again, it depends on the file system abstractions built on top, and then the cache and the current system load -- how many seeks will be required?
> And neither are the tapes. But the pipes, apparently, are.
The "slow vs fast" language is really unfortunate, I realize it's traditional but it's unnecessarily confusing.
A better way to conceptualize it IMHO is bounded vs unbounded: a file or a tape contains a fixed amount of data known a priori, a socket or a pipe does not.
The important point is that the kernel takes locks during all those operations, and will wait an unbounded amount of time if those locks are contended.
So really and truly, yes, any synchronous syscall can schedule out for an arbitrary amount of time, no matter what you do.
It's sort of semantic. The word "block" isn't a synonym for "sleep", it has a specific meaning in POSIX. In that meaning, you're correct, they never "block". But in the generic way most people use the term "block", they absolutely do.
Calling a separate (non-cancellable) thread to perform the lookup sounds a like viable solution...
Windows 8 and above also have their own asynchronous DNS API on NON-POSIX land.
[0] https://learn.microsoft.com/en-us/windows/win32/api/ws2tcpip...
To be precise, even on Linux getaddrinfo_a is not guaranteed to be present. It's a glibc extension. musl doesn't have it.
I've always thought APIs like `pthread_cancel` are too nasty to use. Glad to see well documented evidence of my crank opinion
Imagine cpu-bound worker threads that do nothing but consume work via condition variables and spend long periods of time in hot compute-only loops working on said work... Instead of adding a conditional in the compute you're probably not interested in slowing down at all, you turn on async cancellation and pthread_cancel() the workers when you need to interrupt what's going on.
But it's worth noting pthread_cancel() is also rarely supported anywhere outside first-class pthreads-capable systems like modern linux. So if you have any intention of running elsewhere, forget about it. Thread cancellation support in general is actually somewhat rare IME.
Having written some of the implementation for a non x86 commercial Unix well over 30 years ago now (yeah, I know), pthread_cancel is not that rare. A carve out like “modern linux” is io_uring or even inotify and epoll. AIX and HP-UX, fuck even OSF/1 had pthread_cancel.
Windows has TerminateThread. Most RTOS have some kind of thread level task killing interface.
While they have different semantics than pthread_cancel, that doesn’t really affect the example you’re giving - they can all be used for the “cpu-bound worker”
`pthread_cancel` is different
> Originally, there was no TerminateThread function. The original designers felt strongly that no such function should exist because there was no safe way to terminate a thread, and there’s no point having a function that cannot be called safely. But people screamed that they needed the TerminateThread function, even though it wasn’t safe, so the operating system designers caved and added the function because people demanded it. Of course, those people who insisted that they needed TerminateThread now regret having been given it.
If your code is running in a hot loop, you would have to insert that branch into the hot loop (even well-predicted branches add a few cycles, and can do things like break up decode groups) or have the hot loop bail out every once in a while to go execute the branch and code, which would mean tiling your interior hot loop and thus adding probably significant overhead that way.
Also, you say "cached memory address" but I can almost guarantee that unless you're doing that load a lot more frequently than once every 10 milliseconds the inner loop is going to knock that address out of l1 and probably l2 by the time you get back around to it.
Also, don’t you have to hit a pthread cancellation point for pthread_cancel to take effect?
Those are way more expensive than a branch, but if you want the exact behavior, you could do “if (done) { break; } else { pthread_??? }”
No, the whole point here is async cancellation - you don't test for it and you don't enter a cancellation point.
Excerpt from pthread_setcancelstate(3):
> Asynchronous cancelability
> Setting the cancelability type to PTHREAD_CANCEL_ASYNCHRONOUS is rarely useful.
> Since the thread could be canceled at any time, it cannot safely reserve resources
> (e.g., allocating memory with malloc(3)), acquire mutexes, semaphores, or locks, and
> so on. Reserving resources is unsafe because the application has no way of knowing
> what the state of these resources is when the thread is canceled; that is, did cance-
> lation occur before the resources were reserved, while they were reserved, or after
> they were released? Furthermore, some internal data structures (e.g., the linked
> list of free blocks managed by the malloc(3) family of functions) may be left in an
> inconsistent state if cancelation occurs in the middle of the function call. Conse-
> quently, clean-up handlers cease to be useful.
>
> Functions that can be safely asynchronously canceled are called async-cancel-safe
> functions. POSIX.1-2001 and POSIX.1-2008 require only that pthread_cancel(3),
> pthread_setcancelstate(), and pthread_setcanceltype() be async-cancel-safe. In gen-
> eral, other library functions can't be safely called from an asynchronously cance-
> lable thread.
>
> One of the few circumstances in which asynchronous cancelability is useful is for
> cancelation of a thread that is in a pure compute-bound loop.A better approach would have been to mimic how kernels internally handle signals received during syscalls. Receiving a signal is supposed to cancel the syscall. But from the kernel’s perspective, a syscall implementation is just some code. It can call other functions, acquire locks, wait for conditions, and do anything else you would expect code to do. All of that needs to be cleanly cancelled and unwound to avoid breaking the rest of the system.
So it works like this: when a signal is sent to a thread, a persistent “interrupted” flag is set for that thread. Like with pthread_cancel, this doesn’t immediately interrupt the thread, but only has an effect once the thread calls one of a specific set of functions. For pthread_cancel, that set consists of a bunch of syscalls and other “cancellation points”. For kernel-internal code, it consists of most functions that wait for a condition. The difference is in what happens afterwards. In pthread_cancel’s case, the thread is immediately aborted with only designated cleanups running. In the kernel, the condition-waiting function simply returns an error code. The caller is expected to handle this like any other error code, i.e. by performing any necessary cleanup and then returning the same error code itself. This continues until the entire chain of calls has been unwound. Classic C manual error handling. It’s nothing special, but because interruption works the same way as regular error handling, it‘s more likely to “just work”. Once everything is unwound, the “interrupted” flag is cleared and the original signal can be handled.
(The error code for interruption is usually EINTR, but don’t confuse this with EINTR handling in userspace, which is a mess. The difference is because userspace generally doesn’t want to abort operations upon receiving EINTR, and because from userspace’s perspective there’s no persistent flag.)
pthread_cancel could have been designed the same way: cancellation points return an error code rather than forcibly unwinding. Admittedly, this system might not work quite as well in userspace as it does in kernels. Kernel code already needs to be scrupulous about proper error handling, whereas userspace code often just aborts if a syscall fails. Still, the system would work fine for well-written userspace code, which is more than can be said for pthread_cancel.
Or just use some standalone DNS resolve code or library (which basically replicates getaddrinfo but supports this in an async way)?
See also here the discussion: https://github.com/crystal-lang/crystal/issues/13619
On modern Linux systems, it is: systemd-resolved (https://www.freedesktop.org/software/systemd/man/latest/syst...) is a system service which can be queried through RPC (using dbus or varlink), through the traditional glibc APIs (using a NSS plugin), or even by being queried on the loopback interface as if it were a normal recursive DNS server (for compatibility with software which bypasses glibc and does direct DNS queries).
Unless of course you're making use of shared memory between processes, then you could still leak even when the entire process is shut down.
Thread.destroy() Thread.stop() Thread.suspend()
so much potential for corrupted state.
In general, what you really want is for the API call to be nonblocking so you’re not forced to burn a thread.
I don’t see why glibc would have to do that inside a call to getaddrinfo. can’t it do that once at library initialization? If it has to react to changes to that file while a process is running, couldn’t it have a separate thread for polling that file for changes, or use inotify for a separate thread to be called when it changes? Swapping in the new config atomically might be problematic, but I would think that is solvable.
Even ignoring the issue mentioned it seems wasteful to open, parse, and close that file repeatedly.
“The GNU C Library is designed to be a backwards compatible, portable, and high performance ISO C library. It aims to follow all relevant standards including ISO C11, POSIX.1-2008, and IEEE 754-2008.”
⇒ I don’t think they make that argument.
It could read and parse the file the first time a thread gets created, too.
A cheaper alternative is to check the modification time of the config file, and only reparse it in pthread_cancel when that changes. That doesn’t 10% fix the problem in theory, but would do it in practice.
If it were a library dedicated to DNS, sure, but glibc is used by nearly every process in the system, including many which will never call getaddrinfo.
C libraries are not black magic. Nor are they holy code. We needn't fear them. It's the simplest part of the software stack.
When someone kills off the curl context surely you simply set a suicide flag on the thread and wake it up so it can be joined.
it makes cleanup logic convoluted when you can’t easily pthread_cleanup_push, forcing you to either block on the join or signal cancellation and detach