For example, one kind of program organization that used to be more common in Unix but that remains legal involves keeping signals masked all the time except around certain blocking system calls, e.g. ppoll, that atomically unblock signals and wait. In such a program, a signal can arrive only inside the blocking system call and so the handler can call regular functions without the usual strictures of asynchronous signal safety. (Consider responding to SIGWINCH, which tells you about changing terminal size.)
Synchronous signals are also special in that they're, well, synchronous. That means that in the signal handler you can examine the target memory address or instruction pointer and take action specific to a given spot in your program --- e.g. longjmp to an error handler.
All of this is useful, safe, and legal under POSIX. The main problems with POSIX signals are that, 1) as this thread underscores, most people don't understand them, and 2) signal handlers are process global and hard to share. (Consider if you're running two VMs in a process and each wants to use a SIGSEGV GC safe point trick.)
(userfaultfd, sadly, requires more system calls than a synchronous signal handler to handle anomalous memory access.)
We should be enhancing POSIX signals to make them easier to share, not casting aspersions on them.
It's not. So what?
> The defining characteristics of signals is that, like interrupts and multi-threaded execution
No, they can't. You control when they are masked and unmasked. Certain signals are delivered in response only to certain actions (e.g. floating point errors when enabled, or memory access failures). It's not the chaos you think.
... As long as the called functions are fully async-signal-safe/reentrant. It used to be even more sensitive, in that not all register state was correctly saved/restored on Linux.
(On the fabled plan9, where signals are replaced with arbitrary-text "notes", the issue is even bigger as floating point registers are not saved/restored)
> We should be enhancing POSIX signals to make them easier to share, not casting aspersions on them.
That's fair, but they have already been fixed with signalfd, which more modern processes use to deal with the few POSIX-isms that require it, but many of the original uses have been superseded.
E.g., few modern applications use `timer_create` and SIGALRM, as they can instead use timerfd or poll with strategic timeout (although usually abstracted away by their eventloop). SIGBUS is a POSIX-way to deal with mmap truncation, but sealed memfds can be used to avoid the issue altogether. SIGUSR1/2 can be replaced by any IPC mechanism to give much more flexible controls than a single "reload"/"toggle" signal.
(The POSIX ways can still be useful in certain simpler programs of course.)
That is not accurate. If I have a single-threaded program sitting around a ppoll(2) loop and a signal can arrive only inside my main loop ppoll(), then I know a priori that my signal handler can't be interrupting non-reentrant code and so I can call whatever I want inside it.
> It used to be even more sensitive, in that not all register state was correctly saved/restored on Linux.
Don't confuse architectural flaws with implementation bugs. I'm reminded of an argument on emacs-devel in which someone argued that we couldn't call malloc(3) in a multi-threaded program because one beta version of glibc once had a thread safety bug in the malloc implementation.
> SIGBUS is a POSIX-way to deal with mmap truncation, but sealed memfds can be used to avoid the issue altogether.
Sealing a file descriptor doesn't physically seal a USB key into a USB port. :-) You also get SIGBUS on IO failures on mmaped files, and surprise removal is an IO failure. There's really no alternative to a synchronous signal here --- a regular file being mmap()ed isn't a userfaultfd.
IMHO, I think it's crying shame that Hotspot (last time I checked --- maybe it's fixed?) doesn't watch for SIGBUS on access to MappedByteBuffer and translate surprise USB device file removal into a nice clear VM-level exception.
Even for cases for which userfaultfd can work, a synchronous signal can be more efficient because it involves just one entry into the kernel. (sigreturn is optional.) I'd really hate to give up the conventional signal mechanism entirely, although of course I approve of things like signalfd and userfault that reduce the need for signal handling.
If you have a single-threaded program sitting in ppoll, you cannot ever receive a SIGSEGV during said ppoll unless you passed it bugus fds, timeout or sigmask pointers. A sleeping process cannot segfault.
If you register a SIGSEGV handler, you have zero guarantees that it will only fire at a particular time in your code as it is delivered on any pagefault from any code accidentally generating it. This is why the async-signal-safe rules apply.
If you try handling such fault, what needs to be reentrant is the full call stack that lead to the fault, and every call made from the handler. If, for example, the fault is generated from an event loop handler (in case of a single-threaded example, most things run off event loop handlers), the signal handler must in turn not touch the event loop (no adding/removing/adjusting events, no dispatch) unless the event loop is fully reentrant.
Unless someone sends one with kill(2). Also, I was talking about signals in general, not SIGSEGV in particular. Who uses SIGSEGV as an async work dispatch mechanism?
> unless you passed it bugus fds, timeout or sigmask pointers. A sleeping process cannot segfault.
No, you get EFAULT. System calls don't work that way.
> If you register a SIGSEGV handler, you have zero guarantees that it will only fire at a particular time in your code as it is delivered on any pagefault from any code accidentally generating it
That's why you use sigaction(2) to register signal handlers --- your callback gets both a siginfo_t and a ucontext_t you can use to figure out whether your segfault came from a region of code you know about or some other random thing going wrong in your process. In principle, you can do the non-reentrant thing after having checked that the signal came from the context you expect, and you can do this checking in an async-signal-safe manner.
> If, for example, the fault is generated from an event loop handler (in case of a single-threaded example, most things run off event loop handlers), the signal handler must in turn not touch the event loop (no adding/removing/adjusting events, no dispatch) unless the event loop is fully reentrant.
Of course. That's a matter of program design.
`ppoll(2)` is a C library function you call. On glibc, depending on which implementation you hit on your architecture, the `ppoll` call will segfault before the syscall if you call it with bogus timeout or sigmask pointers, as glibc is dereferencing them.
> That's why you use sigaction(2) to register signal handlers --- your callback gets both a siginfo_t and a ucontext_t you can use to figure out whether your segfault came from a region of code you know about or some other random thing going wrong in your process.
It is true that you could try to figure out whether the faulting address is within a range you thought belonged to a particular thing, but I don't see any reason or benefit. However, this is fragile: Distinguishing between a page fault from a corrupt program state and a page fault in a valid program depends on the program state (specifically, inspecting memory that is likely foo), which in turn makes the output undefined. You won't get false negatives, but will get false positives. You could also try to establish a list of all program counter values (PIE/PIC caveats), and check if the signal originated from near those, but... Ugh.
Now, any other signal I could level with you, but not SIGSEGV. Not until I see some code where it's truly justified: a real, tangible benefit that can only be obtained in this fashion, justifying all the gymnastics and (in my opinion) nastiness of trying to handle a signal that should never be handled in the first place (not even for crash dumps, they're always worse than a proper core).
ppoll is a real system call on all Linux systems and has been for over a decade.
> It is true that you could try to figure out whether the faulting address is within a range you thought belonged to a particular thing,
Which programs do, reliably, all the time.
> You won't get false negatives, but will get false positives.
No, you won't. Once a program wanders into la-la land of memory corruption, all bets are off anyway.
UNIX signals are in no way or form direct translations of hardware CPU traps. The kernel handles hardware traps, which may or may not lead to UNIX signals. Heck, with userfaultfd, a different userspace process could be handling, or injecting, the fault! Not to mention VMs, where the the guest userspace is very far away from any real hardware traps.
There are basically two classes of UNIX signals: Signals that indicate that you might need to take some action (SIGTERM, SIGALRM, SIGUSR1, ...), and signals that indicate that your process did something illegal (SIGILL, SIGSEGV, SIGFP, ...). There is a very, very limited number of cases where handling these errors make sense, and trying to be "clever" to make (faulty and ill-advised) performance optimizations is not one of them.
> This did not improve performance, it was just an unnecessary hack
Well, the Android VM certainly uses a signal mechanism for safepoints and stack overflow checking, for a reason (https://android.googlesource.com/platform/art/+/master/runti...) (both latency and code size), so don't sit there and tell me that the VM running the world's most popular OS is pessimizing itself pointlessly.
> The kernel handles hardware traps, which may or may not lead to UNIX signals.
Not all traps result in signals, and not all signals are traps. Nevertheless, the POSIX signals API is the means through which Unix OSes provide user programs the ability to interact with CPU traps. (Windows does the same thing, morally, with SEH and vectored exception handling --- https://learn.microsoft.com/en-us/windows/win32/debug/vector...). Any decent OS should provide applications with the tools they need to make full use of the underlying hardware. All these anti-signals people are just arguing that programs be bigger and slower, because they can't make full use of all the hardware features of the system, out of, ultimately, an aesthetic objection to signal handling.
Ill-advised as per the authors decision to remove said hack as it brought none of the intended benefits. Or are you suggesting that your opinion is more valuable than the authors whose code we're discussing?
Be careful with fallacies suggesting only one side of an argument is based on opinions. :)
> We have been working on a patch that switches the jit protection mode to EXEC around these potential faulting memory accesses.
Yes, they're changing one aspect of signal handler use to work around this problem. They're not stopping the use of signal handlers in general. Hotspot continues to use signals for efficiency in general. See https://github.com/openjdk/jdk/blob/9059727df135dc90311bd476...
> Be careful with fallacies suggesting only one side of an argument is based on opinions. :)
The wonderful thing about choosing not to care about facts is having whatever opinions you want.
This whole thread is about SIGSEGV, and specifically their SIGSEGV handling. However, catching normal signals is not about efficiency.
Some of their exception handling is still odd: There is no reason for a program that receives SIGILL to ever attempt continuing. But others is fine, like catching SIGFPE to just forward an exception to the calling code.
(Sure, you could construct an argument to say that this is for efficiency if you considered the alternative to be implementing floating point in software so that all exceptions exist in user-space, but hardware floating point is the norm and such alternative would be wholly unreasonable.)
> The wonderful thing about choosing not to care about facts is having whatever opinions you want.
I appreciate the irony of you making such statement, proudly thinking that your opinion equals fact, and therefore any other opinion is not.
This discussion is nothing but subjective opinion vs. subjective opinion. Facts are (hopefully, as I can only speak for myself) inputs to both our opinions, but no opinion about "good" or "bad", "nasty" or not can ever be objective. Objective code quality does not exist.