Nim crashes randomly #16613
Comments
ghost
added
the
|
@yglukhov can you:
to help narrow down, several flags can help:
EDIT: it's not impossible this is related to #17527, and that it'd be caused by a GC issue, can you read this comment #17527 (comment) and try with |
@timotheecour were you unable to reproduce the crash? |
that's odd, it should; did you re-build nim with those flags (which is what's needed) or compile your program with those flags? next step would be to minimize to remove any module dependency, since it's a compiler bug |
Does it mean that you tried to reproduce it and failed? |
|
with which does show the stacktrace messages (it crashed after running for 2 minutes)
it means it's hard to investigate without the bug minimized |
|
TLDR: see workaround below, this would be fixed by nim-lang/RFCs#276
if regs[ra].nodeAddr == nil:
stackTrace(c, tos, pc, "D20210729T134630")
let n2 = regs[ra].nodeAddr[]
if n2 == nil:
stackTrace(c, tos, pc, "D20210729T134630.2")
if n == nil:
stackTrace(c, tos, pc, "D20210729T134630.3")
if (nfIsRef notin regs[ra].nodeAddr[].flags and
nfIsRef notin n.flags):
regs[ra].nodeAddr[][] = n[]
import iface, unittest
# import std/vmutils
iface Animal:
proc say(): string
proc test2()
type
Dog = ref object of RootRef
testCalled: bool
proc say(d: Dog): string = "bark"
proc test2(d: Dog) =
d.testCalled = true
proc doSmth(a: Animal): string =
a.test2()
result = a.say()
proc createVoldemortObject(): Animal =
type Hidden = ref object of RootRef
proc test2(h: Hidden) = discard
suite "iface":
test "runtime":
let d = Dog.new()
doAssert doSmth(d) == "bark"
doAssert d.testCalled
block:
# static: vmTrace(true)
when defined case1:
const s = static:
let d = Dog.new()
let res = doSmth(d)
doAssert(d.testCalled)
0
when defined case2:
const s = (proc(): auto =
let d = Dog.new()
let res = doSmth(d)
doAssert(d.testCalled)
0)()=> crashes with -d:case1
# const s = static:
const s = (proc(): auto =
let d = Dog.new()
let res = doSmth(d)
doAssert(d.testCalled)
0)()=> no crash regardless of nim-lang/RFCs#276, this should be further investigated, ideally by finding a reduced example that exhibits the bug all the time instead of just once every 2 minutes |
|
The crash happens, because PNode gets collected as no refs point to it. It is accessed though Line 680 in bbe05c1 |
possibly, but can you show a reduced test case (with no imports) to demonstrate this? Once we have a reduced test case, we can think about ways to fix this, possibly by adding gc'd references to all PNode's for which we've taken the address, storing in some |
I don't have any, and I'm not sure it'll help, judging from the randomness of this issue. However, I've added some code to nim that proves my point and shows 100% reproducibility. |
|
that's great to have a way to reproduce at 100%, I can indeed reproduce via #18628 However, we still need a minimized test case (now that we have a way to reproduce at 100% via that PR), to make it easier to understand the issue and fix it (even if it's indeed due to gc collecting nodeAddr) A potential fix is by maintaining a |
I've been thinking of a similar approach, but not to use a |
|
Ideally though it would be great to get rid of unsafe code ( |
|
#19515 looks a lot like it could fix this issue, doesn't it? |
…im-lang#16780 nim-lang#16613 nim-lang#14553 nim-lang#19909 nim-lang#18641 (nim-lang#19902) [backport] * revert nim-lang#12217 since the root problem seems to have been fixed; fix nim-lang#15974;fix nim-lang#12551; fix nim-lang#19464 * fix nim-lang#16020; fix nim-lang#16780 * fix tests and nim-lang#16613 * fix nim-lang#14553 * fix nim-lang#19909; skip skipRegisterAddr * fix nim-lang#18641
Steps to reproduce:
Wait for 5 minutes
Actual result:
Nim eventually crashes with stack trace:
Nim version: bbe05c1
The text was updated successfully, but these errors were encountered: