diff options
author | Father Chrysostomos <sprout@cpan.org> | 2014-02-14 18:05:47 -0800 |
---|---|---|
committer | Father Chrysostomos <sprout@cpan.org> | 2014-02-15 07:01:39 -0800 |
commit | 795eb8c825e0362c8f90071503f74e21f38873cc (patch) | |
tree | b2aae68dc7809102173f17e5782f4c1da4a629f3 /pp.c | |
parent | 1babc821f551f5bfc02e98a5ed5391b9ebf2fefb (diff) | |
download | perl-795eb8c825e0362c8f90071503f74e21f38873cc.tar.gz |
[perl #121242] Fix crash in gp_free when gv is freed
Commit 4571f4a caused the gp to have a refcount of 1, not 0, in
gp_free when the contents of the glob are freed. This makes
gv_try_downgrade see the gv as a candidate for downgrading in
this example:
sub Fred::AUTOLOAD { $Fred::AUTOLOAD }
undef *{"Fred::AUTOLOAD"};
When the glob is undefined, the sub inside it is freed, and the
gvop ($Fred::AUTOLOAD), when freed, tries to downgrade the glob
(*Fred::AUTOLOAD). Since it is empty, it deletes it completely from
the containing stash, so the GV is freed out from under gp_free, which
is still using it, causing an assertion failure.
We can trigger a similar condition more explicitly:
$ ./miniperl -e 'DESTROY{delete $::{foo}} ${"foo"} = bless []; undef *{"foo"}'
This bug is nothing new. On a non-debugging 5.18.2, I get this:
$ perl5.18.2 -e 'DESTROY{delete $::{foo}} ${"foo"} = bless []; undef *{"foo"}'
Attempt to free unreferenced glob pointers at -e line 1.
Segmentation fault: 11
That crashes in pp_undef after the call to gp_free, becaues pp_undef
continues to manipulate the GV.
The problem occurs not only with pp_undef, but also with other func-
tions calling gp_free:
sv_setsv_flags:
$ ./miniperl -e 'DESTROY{delete $::{foo}} ${"foo"} = bless []; *{"foo"}="bar"'
glob_assign_glob:
$ ./miniperl -e 'DESTROY{delete $::{foo}} ${"foo"} = bless []; *{"foo"}=*bar'
sv_unglob, reached through various paths:
$ ./miniperl -e 'DESTROY{delete $::{foo}} ${"foo"} = do {local *bar}; $${"foo"} = bless []; ${"foo"} = 3'
$ ./miniperl -e 'DESTROY{delete $::{foo}} ${"foo"} = do {local *bar}; $${"foo"} = bless []; utf8::encode(${"foo"})'
$ ./miniperl -e 'DESTROY{delete $::{foo}} ${"foo"} = do {local *bar}; $${"foo"} = bless []; open bar, "t/TEST"; ${"foo"} .= <bar>'
$ ./miniperl -e 'DESTROY{delete $::{foo}} ${"foo"} = do {local *bar}; $${"foo"} = bless []; ${"foo"}++'
$ ./miniperl -e 'DESTROY{delete $::{foo}} ${"foo"} = do {local *bar}; $${"foo"} = bless []; undef ${"foo"}'
$ ./miniperl -e 'DESTROY{delete $::{foo}} ${"foo"} = 3; ${"foo"} =~ s/3/${"foo"} = do {local *bar}; $${"foo"} = bless []; 4/e'
And there are probably more ways to trigger this through sv_unglob.
(I stopped looking when I thought of the fix.)
This patch fixes the problem by protecting the GV using the mortals
stack in functions that call gp_free. I did not change gp_free
itself, since it is an API function that as yet does not touch the
mortals stack, and I am not sure that should change. All of its
callers that this patch touches already do sv_2mortal in some cir-
cumstances.
Diffstat (limited to 'pp.c')
-rw-r--r-- | pp.c | 1 |
1 files changed, 1 insertions, 0 deletions
@@ -1024,6 +1024,7 @@ PP(pp_undef) else stash = NULL; } + SvREFCNT_inc_simple_void_NN(sv_2mortal(sv)); gp_free(MUTABLE_GV(sv)); Newxz(gp, 1, GP); GvGP_set(sv, gp_ref(gp)); |