Segmentation fault with a program compiled by OCaml 3.11.1, 3.11.2, 3.12.0beta1 (bytecode, native, Linux AMD64 and MacOS X)

[vicuna]

Original bug ID: 5109
Reporter: monate
Assigned to: @xavierleroy
Status: closed (set by @xavierleroy on 2012-03-24T14:01:41Z)
Resolution: fixed
Priority: normal
Severity: crash
Version: 3.12.0+beta1 or 3.12.0+rc1
Fixed in version: 3.12.1+dev
Category: ~DO NOT USE (was: OCaml general)
Monitored by: phillyg mehdi "Pascal Cuoq" @glondu "Julien Signoles" prevosto

Bug description

Dear OCaml developers,

The code contained in the tar file available at
http://frama-c.com/framac_stripped_bug.tar.bz2
produces a deterministic segmentation fault.
It can be reproduced using:
./configure
make
bin/toplevel.byte bug.c

I tried to recompile OCaml 3.11.1 with #define DEBUG added in byterun/gc_ctrl.c but the heap consistency checks are ok.
Using ocamldebug a backtrace is available and seems to indicate that the call to Cil_const.CurrentLoc fails.
Strangely if one removes the line 63 of cil/src/cil_const.ml the bug disappears.

Do not hesitate to ask for more advanced tests from me.
Cheers,
Benjamin

File attachments

• restore_lval.diff

[vicuna]

Comment author: @ygrek

404

[vicuna]

Comment author: monate

Sorry: the correct link is

http://frama-c.com/frama-c_stripped.tar.bz2

[vicuna]

Comment author: Pascal Cuoq

Note: all references to Obj functions have been erased from this "stripped" reproduction case. There are references to Obj.t, but the expressions that take these types are all "raise exn;"

[vicuna]

Comment author: @ygrek

cil/src/fronc/cparser.mly
insert in line 1372 (before doFunctionDef)
Printf.printf "parser %d %d %d\n%!" (Obj.magic (fst3 $2) : int) (Obj.magic (snd3 $2) : int) (Obj.magic (trd3 $2) : int);
$ bin/toplevel.byte bug.c | grep parser
parser 70006161577636 70006161585852 0
Adding
Format.printf "%a@." d_cabsloc (trd3 $2);
will crash right there in the parser

[vicuna]

Comment author: monate

I'm impressed! How did you track this behavior?
If I understand correctly this looks like a bug of ocamlyacc: the representation of the value returned by Parsing.peek_val does not match the type it gets.

[vicuna]

Comment author: @ygrek

The key is to note that v passed to CurrentLoc.set is bad (reason of a crash). Than trace it back to the point of origin with grep+sed+printf (or better camlp4 :)
(trd3 $2) corresponds to RBRACE, but ocamllex emits correct value, so indeed it seems like the bug in ocamlyacc, hard to believe though..

[vicuna]

Comment author: Julien Signoles

The above link does not work anymore :-(.

See http://bts.frama-c.com/view.php?id=626 for an easy way to reproduce something which looks like the very same bug (but with an unstripped Frama-C version which contains references to Obj, sorry OCaml Folks).

[vicuna]

Comment author: phillyg

Benjamin, I'm wondering whether you can still reproduce the bug on your linux machine. if so, please email me at philip@pgbovine.net ... i might be able to help debug.

[vicuna]

Comment author: prevosto

A more or less minimal code (with only the two parsers and lexers for C and ACSL) that reproduces the issue is at

http://dl.dropbox.com/u/6149559/bug_5109_minimal.tar.gz

just type make to build it and run the test case

It seems that there is somehow an interference between Cparser and Logic_parser, which causes in the end Cparser to get a null pointer instead of the real CODE_ANNOT.
Note that there is no segfault anymore, as the program does not do anything with the value, still, we shouldn't get as output

In parser, annot is 0,

[vicuna]

Comment author: @ygrek

There are two parsers used recursively (i.e. cparser calls clexer which calls logic_parser). IIUC this is not supported by ocamlyacc - see let env = in stdlib/parsing.ml - so either coalesce the parsers or use menhir..

[vicuna]

Comment author: @xavierleroy

The repro case works, thanks. In response to ygrek: It is true that the Parsing module has some global state, but it tries to support nested parsers nonetheless (see function Parsing.yyparse). This said, the bug might very well be in this support for reentrancy.

[vicuna]

Comment author: @ygrek

Looks like restoring env.lval solves this issue

[vicuna]

Comment author: @xavierleroy

Thanks for the "restore_lval" patch, which I applied in the 3.12 bugfix branch. It will go in release 3.12.1.

For the record: the original code in Parsing envisioned the case where a parser is called from a semantic action of another parser, but not the case where a parser is called from the lexer of another parser, as in the case in the Frama-C example. In the former case, env.lval is irrelevant, but in the latter case, it is necessary to preserve its value.