-
-
Notifications
You must be signed in to change notification settings - Fork 5.5k
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
signal handling: User-defined interrupt handlers
Interrupt handling is a tricky problem, not just in terms of implementation, but in terms of desired behavior: when an interrupt is received, which code should handle it? Julia's current answer to this is effectively to throw an `InterruptException` to the first task to hit a safepoint. While this seems sensible (the code that's running gets interrupted), it only really works for very basic numerical code. In the case that multiple tasks are running concurrently, or when try-catch handlers are registered, this system breaks down, and results in unpredictable behavior. This unpredictable behavior includes: - Interrupting background/runtime tasks which don't want to be interrupted, as they do little bits of important work (and are critical to library runtime functionality) - Interrupting only one task, when multiple coordinating tasks would want to receive the interrupt to safely terminate a computation - Interrupting only one library's task, when multiple libraries really would want to be notified about the interrupt The above behavior makes it nearly impossible to provide reliable Ctrl-C behavior, and results in very confused users who get stuck hitting Ctrl-C continuously, sometimes getting caught in a hang, sometimes triggering unrelated exception handling code they didn't mean to, sometimes getting a segfault, and very rarely getting the behavior they desire (with unpredictable safety of being able to continue using the active session as intended). This commit provides an alternative behavior for interrupts which is more predictable: user code may now register tasks as "interrupt handlers" (via `Base.register_interrupt_handler`), which will be guaranteed to receive an `InterruptException` whenever the session receives an interrupt signal. Additionally, unlike the previous behavior, no other tasks will receive `InterruptException`s; only explicitly registered handlers may receive them. This behavior allows one or more libraries to register handler tasks which will all be concurrently awoken to handle each interrupt and do whatever is necessary to safely interrupt any running code; the extent to which other tasks are interrupted is arbitrary and library-defined. For example, GPU libraries like AMDGPU.jl can register a handler to safely interrupt GPU kernels running on all GPU queues and do resource cleanup. Concurrently, a complex runtime like the scheduler in Dagger.jl can register a handler to interrupt running tasks on other workers when possible. This commit also adds a more convenient interface for when the REPL is running. When a Ctrl-C is received and the user is not at the REPL prompt, a TerminalMenus-powered prompt will be shown, where the user will have a variety of possible actions, including: - Ignore the interrupt (do nothing) - Activate all module's interrupt handlers - Activate a specific module's interrupt handlers - Disable the interrupt handler (reverting to the old Ctrl-C behavior) - Exit Julia gracefully (with `exit()`) - Exit Julia forcefully (with a `ccall` to `abort`)
- Loading branch information
Showing
14 changed files
with
338 additions
and
8 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Original file line number | Diff line number | Diff line change |
---|---|---|
@@ -0,0 +1,243 @@ | ||
# Internal methods, only to be used to change to a different global interrupt handler | ||
function _register_global_interrupt_handler(handler::Task) | ||
handler_ptr = Base.pointer_from_objref(handler) | ||
slot_ptr = cglobal(:jl_interrupt_handler, Ptr{Cvoid}) | ||
Intrinsics.atomic_pointerset(slot_ptr, handler_ptr, :release) | ||
end | ||
function _unregister_global_interrupt_handler() | ||
slot_ptr = cglobal(:jl_interrupt_handler, Ptr{Cvoid}) | ||
Intrinsics.atomic_pointerset(slot_ptr, C_NULL, :release) | ||
end | ||
|
||
const INTERRUPT_HANDLERS_LOCK = Threads.ReentrantLock() | ||
const INTERRUPT_HANDLERS = Dict{Module,Vector{Task}}() | ||
const INTERRUPT_HANDLER_RUNNING = Threads.Atomic{Bool}(false) | ||
|
||
""" | ||
register_interrupt_handler(mod::Module, handler::Task) | ||
Registers the task `handler` to handle interrupts (such as from Ctrl-C). | ||
Handlers are expected to sit idly within a `wait()` call or similar. When an | ||
interrupt is received by Ctrl-C or manual SIGINT, one of two actions may | ||
happen: | ||
If the REPL is not running (such as when running `julia myscript.jl`), then all | ||
registered interrupt handlers will be woken with an `InterruptException()`, and | ||
the handler may take whatever actions are necessary to gracefully interrupt any | ||
associated running computations. It is expected that the handler will spawn | ||
tasks to perform the graceful interrupt, so that the handler task may return | ||
quickly to again calling `wait()` to catch future user interrupts. | ||
If the REPL is running, then the user will be presented with a terminal menu | ||
which will allow them to do one of: | ||
- Ignore the interrupt (do nothing) | ||
- Activate all handlers for all modules | ||
- Activate all handlers for a specific module | ||
- Disable this interrupt handler logic (see below for details) | ||
- Exit Julia gracefully (with `exit`) | ||
- Exit Julia forcefully (with a `ccall` to `abort`) | ||
Note that if the interrupt handler logic is disabled by the above menu option, | ||
Julia will fall back to the old Ctrl-C handling behavior, which has the | ||
potential to cause crashes and undefined behavior (but can also interrupt more | ||
kinds of code). If desired, the interrupt handler logic can be re-enabled by | ||
calling `start_repl_interrupt_handler()`, which will disable the old Ctrl-C | ||
handling behavior. | ||
To unregister a previously-registered handler, use | ||
[`unregister_interrupt_handler`](@ref). | ||
!!! warn | ||
Non-yielding tasks may block interrupt handlers from running; this means | ||
that once an interrupt handler is registered, code like `while true end` | ||
may become un-interruptible. | ||
""" | ||
function register_interrupt_handler(mod::Module, handler::Task) | ||
if ccall(:jl_generating_output, Cint, ()) == 1 | ||
throw(ConcurrencyViolationError("Interrupt handlers cannot be registered during precompilation.\nPlease register your handler later (possibly in your module's `__init__`).")) | ||
end | ||
lock(INTERRUPT_HANDLERS_LOCK) do | ||
handlers = get!(Vector{Task}, INTERRUPT_HANDLERS, mod) | ||
push!(handlers, handler) | ||
end | ||
end | ||
|
||
""" | ||
unregister_interrupt_handler(mod::Module, handler::Task) | ||
Unregisters the interrupt handler task `handler`; see | ||
[`register_interrupt_handler`](@ref) for further details. | ||
""" | ||
function unregister_interrupt_handler(mod::Module, handler::Task) | ||
if ccall(:jl_generating_output, Cint, ()) == 1 | ||
throw(ConcurrencyViolationError("Interrupt handlers cannot be unregistered during precompilation.")) | ||
end | ||
lock(INTERRUPT_HANDLERS_LOCK) do | ||
handlers = get!(Vector{Task}, INTERRUPT_HANDLERS, mod) | ||
deleteat!(handlers, findall(==(handler), handlers)) | ||
end | ||
end | ||
|
||
function _throwto_interrupt!(task::Task) | ||
if task.state == :runnable | ||
task._isexception = true | ||
task.result = InterruptException() | ||
try | ||
schedule(task) | ||
catch | ||
end | ||
end | ||
end | ||
|
||
# Simple (no TUI) interrupt handler | ||
|
||
function simple_interrupt_handler() | ||
last_time = 0.0 | ||
while true | ||
try | ||
# Wait to be interrupted | ||
wait() | ||
catch err | ||
if !(err isa InterruptException) | ||
rethrow(err) | ||
end | ||
|
||
# Force-interrupt root task if two interrupts in quick succession (< 1s) | ||
now_time = time() | ||
diff_time = now_time - last_time | ||
last_time = now_time | ||
if diff_time < 1 | ||
_throwto_interrupt!(Base.roottask) | ||
end | ||
|
||
# Interrupt all handlers | ||
lock(INTERRUPT_HANDLERS_LOCK) do | ||
for mod in keys(INTERRUPT_HANDLERS) | ||
for handler in INTERRUPT_HANDLERS[mod] | ||
if handler.state == :runnable | ||
_throwto_interrupt!(handler) | ||
end | ||
end | ||
end | ||
end | ||
end | ||
end | ||
end | ||
function simple_interrupt_handler_checked() | ||
try | ||
simple_interrupt_handler() | ||
catch err | ||
# Some internal error, make sure we start a new handler | ||
Threads.atomic_xchg!(INTERRUPT_HANDLER_RUNNING, false) | ||
_unregister_global_interrupt_handler() | ||
start_simple_interrupt_handler() | ||
rethrow() | ||
end | ||
# Clean exit | ||
Threads.atomic_xchg!(INTERRUPT_HANDLER_RUNNING, false) | ||
_unregister_global_interrupt_handler() | ||
end | ||
function start_simple_interrupt_handler(; force::Bool=false) | ||
if (Threads.atomic_cas!(INTERRUPT_HANDLER_RUNNING, false, true) == false) || force | ||
simple_interrupt_handler_task = errormonitor(Threads.@spawn simple_interrupt_handler_checked()) | ||
_register_global_interrupt_handler(simple_interrupt_handler_task) | ||
end | ||
end | ||
|
||
# REPL (TUI) interrupt handler | ||
|
||
function repl_interrupt_handler() | ||
invokelatest(REPL_MODULE_REF[]) do REPL | ||
TerminalMenus = REPL.TerminalMenus | ||
|
||
root_menu = TerminalMenus.RadioMenu( | ||
[ | ||
"Interrupt all", | ||
"Interrupt only...", | ||
"Interrupt root task (REPL/script)", | ||
"Ignore it", | ||
"Stop handling interrupts", | ||
"Exit Julia", | ||
"Force-exit Julia", | ||
] | ||
) | ||
|
||
while true | ||
try | ||
# Wait to be interrupted | ||
wait() | ||
catch err | ||
if !(err isa InterruptException) | ||
rethrow(err) | ||
end | ||
|
||
# Display root menu | ||
@label display_root | ||
choice = TerminalMenus.request("Interrupt received, select an action:", root_menu) | ||
if choice == 1 | ||
lock(INTERRUPT_HANDLERS_LOCK) do | ||
for mod in keys(INTERRUPT_HANDLERS) | ||
for handler in INTERRUPT_HANDLERS[mod] | ||
if handler.state == :runnable | ||
_throwto_interrupt!(handler) | ||
end | ||
end | ||
end | ||
end | ||
elseif choice == 2 | ||
# Display modules menu | ||
mods = lock(INTERRUPT_HANDLERS_LOCK) do | ||
collect(keys(INTERRUPT_HANDLERS)) | ||
end | ||
mod_menu = TerminalMenus.RadioMenu(vcat(map(string, mods), "Go Back")) | ||
@label display_mods | ||
choice = TerminalMenus.request("Select a library to interrupt:", mod_menu) | ||
if choice > length(mods) || choice == -1 | ||
@goto display_root | ||
else | ||
lock(INTERRUPT_HANDLERS_LOCK) do | ||
for handler in INTERRUPT_HANDLERS[mods[choice]] | ||
_throwto_interrupt!(handler) | ||
end | ||
end | ||
@goto display_mods | ||
end | ||
elseif choice == 3 | ||
# Force-interrupt root task | ||
_throwto_interrupt!(Base.roottask) | ||
elseif choice == 4 || choice == -1 | ||
# Do nothing | ||
elseif choice == 5 | ||
# Exit handler (caller will unregister us) | ||
return | ||
elseif choice == 6 | ||
# Exit Julia cleanly | ||
exit() | ||
elseif choice == 7 | ||
# Force an exit | ||
ccall(:abort, Cvoid, ()) | ||
end | ||
end | ||
end | ||
end | ||
end | ||
function repl_interrupt_handler_checked() | ||
try | ||
repl_interrupt_handler() | ||
catch err | ||
# Some internal error, make sure we start a new handler | ||
Threads.atomic_xchg!(INTERRUPT_HANDLER_RUNNING, false) | ||
_unregister_global_interrupt_handler() | ||
start_repl_interrupt_handler() | ||
rethrow() | ||
end | ||
# Clean exit | ||
Threads.atomic_xchg!(INTERRUPT_HANDLER_RUNNING, false) | ||
_unregister_global_interrupt_handler() | ||
end | ||
function start_repl_interrupt_handler(; force::Bool=false) | ||
if (Threads.atomic_cas!(INTERRUPT_HANDLER_RUNNING, false, true) == false) || force | ||
repl_interrupt_handler_task = errormonitor(Threads.@spawn repl_interrupt_handler_checked()) | ||
_register_global_interrupt_handler(repl_interrupt_handler_task) | ||
end | ||
end |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Oops, something went wrong.