Re: [Sbcl-bugs] SB-IMPL::RELEASE-FD-STREAM-RESOURCES eats up deadline restarts.

[Lorenz M. <moe...@in...>]

Hi Nikodemus,

I'm definitely not signalling deadlines in a timer :)

In my case, the timeout error occurs when closing a socket. On a
reasonably fast computer and a fast network connection, it almost
never occurs, but on a slow computer or high network load, it happens
way too often. Attached, you find a simple test file that can be
executed using `sbcl --script` which runs a simple server on a socket
and periodically connects to it. I get the following backtrace pretty
reliably on a netbook with an Atom processor:

unhandled SB-SYS:DEADLINE-TIMEOUT in thread #<SB-THREAD:THREAD
                                              "initial thread" RUNNING
                                               {AAA2EB1}>:
  A deadline was reached after 0.002 seconds.

0: (SB-DEBUG::MAP-BACKTRACE #<CLOSURE (LAMBDA #) {B2B798D}> :START 0 :COUNT 128)
1: (SB-DEBUG:BACKTRACE 128 #<SB-SYS:FD-STREAM for "standard error" {AAA3C49}>)
2: (SB-DEBUG::DEBUGGER-DISABLED-HOOK
    #<SB-SYS:DEADLINE-TIMEOUT {B2B4D09}>
    #<unavailable argument>)
3: (SB-DEBUG::RUN-HOOK
    SB-EXT:*INVOKE-DEBUGGER-HOOK*
    #<SB-SYS:DEADLINE-TIMEOUT {B2B4D09}>)
4: (INVOKE-DEBUGGER #<SB-SYS:DEADLINE-TIMEOUT {B2B4D09}>)
5: (ERROR #<SB-SYS:DEADLINE-TIMEOUT {B2B4D09}>)
6: (SB-IMPL::RELEASE-FD-STREAM-RESOURCES
    #<SB-SYS:FD-STREAM for "socket 127.0.0.1:34219, peer: 127.0.0.1:11111"
      {B2B47E1}>)
7: ((SB-PCL::FAST-METHOD SB-GRAY::PCL-CLOSE (SB-KERNEL:ANSI-STREAM))
    #<unused argument>
    #<unused argument>
    #<SB-SYS:FD-STREAM for "socket 127.0.0.1:34219, peer: 127.0.0.1:11111"
      {B2B47E1}>
    :ABORT
    NIL)
8: ((FLET #:CLEANUP-FUN-[PERFORM-REQUEST]4))[:CLEANUP]
9: (PERFORM-REQUEST)
10: (RUN)
11: (SB-INT:SIMPLE-EVAL-IN-LEXENV (RUN) #<NULL-LEXENV>)
12: (SB-EXT:EVAL-TLF (RUN) 8 #<NULL-LEXENV>)
13: ((FLET SB-FASL::EVAL-FORM) (RUN) 8)
14: (SB-INT:LOAD-AS-SOURCE
     #<SB-SYS:FD-STREAM for "file /home/moesenle/tmp/deadlines.lisp" {AAA67C9}>
     :VERBOSE
     NIL
     :PRINT
     NIL
     :CONTEXT
     "loading")
15: ((FLET SB-FASL::LOAD-STREAM)
     #<SB-SYS:FD-STREAM for "file /home/moesenle/tmp/deadlines.lisp" {AAA67C9}>
     NIL)
16: (LOAD
     #<SB-SYS:FD-STREAM for "file /home/moesenle/tmp/deadlines.lisp" {AAA67C9}>
     :VERBOSE
     NIL
     :PRINT
     NIL
     :IF-DOES-NOT-EXIST
     T
     :EXTERNAL-FORMAT
     :DEFAULT)
17: ((FLET SB-IMPL::LOAD-SCRIPT)
     #<SB-SYS:FD-STREAM for "file /home/moesenle/tmp/deadlines.lisp" {AAA67C9}>)
18: ((FLET #:WITHOUT-INTERRUPTS-BODY-[PROCESS-SCRIPT]69))
19: (SB-IMPL::PROCESS-SCRIPT "./deadlines.lisp")
20: (SB-IMPL::TOPLEVEL-INIT)
21: ((FLET #:WITHOUT-INTERRUPTS-BODY-[RESTART-LISP]30))
22: ((LABELS SB-IMPL::RESTART-LISP))


The backtrace is hard to reproduce on a faster computer though. I
don't know enough about how deadlines are implemented internally, I
just know that it happens and I don't see any timers involved :)

Thanks,
Lorenz

On Sun, May 27, 2012 at 5:51 PM, Nikodemus Siivola
<nik...@ra...> wrote:
>>> RELEASE-FD-STREAM-RESOURCES removes the restarts created by
>>> SIGNAL-DEADLINE if a deadline is signaled inside the HANDLER-CASE form
>>> in RELEASE-FD-STREAM-RESOURCES. The reason is that all serious
>>> conditions are caught and re-signaled outside the WITH-INTERRUPTS form
>>> which causes SIGNAL-DEADLINE including its restarts to be unwound.
>>> That means it becomes impossible to defer or cancel the deadline and
>>> DEADLINE-TIMEOUT conditions occur.
>
> Lorenz, having taken a better look, it seems to me that the only way
> to provoke the this is by having something signaling deadlines in a
> timer which fires when WITHOUT-INTERRUPTS is exited.
>
> Am I missing something?
>
> While the forcible unwind is not ideal, and something will be done
> about it, I just wanted to point out that because of exactly this sort
> of reason having timers doing things which may signal deadlines is a
> fundamentally losing proposal, unless those timers handle their own
> deadlines.
>
> Cheers,
>
>  -- Nikodemus
>



-- 
Lorenz Mösenlechner            | moe...@in...
Technische Universität München | Karlstraße 45
80335 München                  | Germany
http://ias.cs.tum.edu/         | Tel: +49 (89) 289-26910