bug#17010: Bug#739752: coreutils: ln segfaults when run with --relative and an empty target
On 03/14/2014 03:44 AM, Jim Meyering wrote: On Thu, Mar 13, 2014 at 7:22 PM, Pádraig Brady p...@draigbrady.com wrote: Interesting. So canonicalize_filename_mode() can fail in this case, even with CAN_MISSING. It's unexpected that c_f_m() sets errno=ENOENT when CAN_MISSING is set. I wonder should we change that instead in gnulib? With CAN_MISSING I would expect this function to work on arbitrary strings, including the empty string. What would you have c_f_m(, CAN_MISSING) return? I know of no absolute name corresponding to the dot-relative empty string. Since with CAN_MISSING we should be just degenerating to string processing, I would think it slightly better to return xstrdup(), to avoid special casing that in each caller. I also notice that c_f_m() can return ENOMEM (from areadlink_with_size). It's arguable that we should xalloc_die() within c_f_m() for consistency in that case. However I also notice that c_f_m(, CAN_MISSING) can return ENOENT if the current working dir is unlinked. I'm not sure there is anything else we could do within c_f_m() to handle that. Hence since c_f_m() can validly fail even with CAN_MISSING, I agree your patch is correct. Please push. thanks! Pádraig.
bug#17010: Bug#739752: coreutils: ln segfaults when run with --relative and an empty target
On Fri, Mar 14, 2014 at 4:49 AM, Pádraig Brady p...@draigbrady.com wrote: ... Hence since c_f_m() can validly fail even with CAN_MISSING, I agree your patch is correct. Please push. Done.
bug#17010: Bug#739752: coreutils: ln segfaults when run with --relative and an empty target
On Sat, Feb 22, 2014 at 1:57 AM, Erik Bernstein e...@fscking.org wrote: Package: coreutils Version: 8.21-1 Severity: normal Hi, when ln is run with --relative --symbolic and and empty string as the target, it ungracefully dies with a segmentation fault. The memory violation appears to happen in src/relpath.c:38 when the two input paths are checked for leading slashes: if ((path1[1] == '/') != (path2[1] == '/')) How to reproduce: [1] Open a terminal [2] run: ln -sr '' foobar Result: segmentation fault ln -sr '' foobar Expected result: Some kind of error message ... Thank you for the bug report! That also affected the very latest code in git. Here is a patch: From a6d2db8b6dfe15344aba4aefe9545eb3a4876d45 Mon Sep 17 00:00:00 2001 From: Jim Meyering meyer...@fb.com Date: Thu, 13 Mar 2014 17:05:04 -0700 Subject: [PATCH] ln: with -sr, don't segfault for a TARGET of '' Prior to this change, ln -sr '' F would segfault, attempting to read path2[1] in relpath.c's path_common_prefix function. This problem arises whenever canonicalize_filename_mode returns NULL. * src/ln.c (convert_abs_rel): Call relpath only when both canonicalize_filename_mode calls return non-NULL. * tests/ln/relative.sh: Add a test to trigger this failure. * THANKS.in: List reporter's name/address. * NEWS (Bug fixes): Mention it. Reported by Erik Bernstein in 739...@bugs.debian.org. --- NEWS | 2 ++ THANKS.in| 1 + src/ln.c | 16 ++-- tests/ln/relative.sh | 5 + 4 files changed, 18 insertions(+), 6 deletions(-) diff --git a/NEWS b/NEWS index 62966b2..b3ad65c 100644 --- a/NEWS +++ b/NEWS @@ -25,6 +25,8 @@ GNU coreutils NEWS-*- outline -*- it would display an error, requiring --no-dereference to avoid the issue. [bug introduced in coreutils-5.3.0] + ln -sr '' F no longer segfaults: now it fails with the expected diagnostic + shuf --repeat no longer dumps core if the input is empty. [bug introduced with the --repeat feature in coreutils-8.22] diff --git a/THANKS.in b/THANKS.in index 561d18c..2670265 100644 --- a/THANKS.in +++ b/THANKS.in @@ -193,6 +193,7 @@ Eric G. Miller e...@jps.net Eric Pementepeme...@northpark.edu Eric S. Raymond e...@snark.thyrsus.com Erik Bennettbenn...@cvo.oneworld.com +Erik Bernstein e...@fscking.org Erik Corry e...@kroete2.freinet.de Felix Lee f...@teleport.com Felix Rauch Valenti fra...@cse.unsw.edu.au diff --git a/src/ln.c b/src/ln.c index aab9cf2..6726699 100644 --- a/src/ln.c +++ b/src/ln.c @@ -149,13 +149,17 @@ convert_abs_rel (const char *from, const char *target) char *realdest = canonicalize_filename_mode (targetdir, CAN_MISSING); char *realfrom = canonicalize_filename_mode (from, CAN_MISSING); - /* Write to a PATH_MAX buffer. */ - char *relative_from = xmalloc (PATH_MAX); - - if (!relpath (realfrom, realdest, relative_from, PATH_MAX)) + char *relative_from = NULL; + if (realdest realfrom) { - free (relative_from); - relative_from = NULL; + /* Write to a PATH_MAX buffer. */ + relative_from = xmalloc (PATH_MAX); + + if (!relpath (realfrom, realdest, relative_from, PATH_MAX)) +{ + free (relative_from); + relative_from = NULL; +} } free (targetdir); diff --git a/tests/ln/relative.sh b/tests/ln/relative.sh index 7636695..e8c25dc 100755 --- a/tests/ln/relative.sh +++ b/tests/ln/relative.sh @@ -45,4 +45,9 @@ mkdir web ln -sr latest web/latest test $(readlink web/latest) = '../release2' || fail=1 +# Expect this to fail with exit status 1. +# Prior to coreutils-8.23, it would segfault. +ln -sr '' F +test $? = 1 || fail=1 + Exit $fail -- 1.9.0.167.g384364b
bug#17010: Bug#739752: coreutils: ln segfaults when run with --relative and an empty target
On 03/14/2014 01:42 AM, Jim Meyering wrote: From a6d2db8b6dfe15344aba4aefe9545eb3a4876d45 Mon Sep 17 00:00:00 2001 From: Jim Meyering meyer...@fb.com Date: Thu, 13 Mar 2014 17:05:04 -0700 Subject: [PATCH] ln: with -sr, don't segfault for a TARGET of '' Prior to this change, ln -sr '' F would segfault, attempting to read path2[1] in relpath.c's path_common_prefix function. This problem arises whenever canonicalize_filename_mode returns NULL. * src/ln.c (convert_abs_rel): Call relpath only when both canonicalize_filename_mode calls return non-NULL. * tests/ln/relative.sh: Add a test to trigger this failure. * THANKS.in: List reporter's name/address. * NEWS (Bug fixes): Mention it. Reported by Erik Bernstein in 739...@bugs.debian.org. We can amend with the now allocated: Fixes http://bugs.gnu.org/17010 diff --git a/NEWS b/NEWS index 62966b2..b3ad65c 100644 --- a/NEWS +++ b/NEWS @@ -25,6 +25,8 @@ GNU coreutils NEWS-*- outline -*- it would display an error, requiring --no-dereference to avoid the issue. [bug introduced in coreutils-5.3.0] + ln -sr '' F no longer segfaults: now it fails with the expected diagnostic Probably should add: [bug introduced with the --relative feature in coreutils-8.16] diff --git a/src/ln.c b/src/ln.c index aab9cf2..6726699 100644 --- a/src/ln.c +++ b/src/ln.c @@ -149,13 +149,17 @@ convert_abs_rel (const char *from, const char *target) char *realdest = canonicalize_filename_mode (targetdir, CAN_MISSING); char *realfrom = canonicalize_filename_mode (from, CAN_MISSING); Interesting. So canonicalize_filename_mode() can fail in this case, even with CAN_MISSING. It's unexpected that c_f_m() sets errno=ENOENT when CAN_MISSING is set. I wonder should we change that instead in gnulib? With CAN_MISSING I would expect this function to work on arbitrary strings, including the empty string. - /* Write to a PATH_MAX buffer. */ - char *relative_from = xmalloc (PATH_MAX); - - if (!relpath (realfrom, realdest, relative_from, PATH_MAX)) + char *relative_from = NULL; + if (realdest realfrom) { - free (relative_from); - relative_from = NULL; + /* Write to a PATH_MAX buffer. */ + relative_from = xmalloc (PATH_MAX); + + if (!relpath (realfrom, realdest, relative_from, PATH_MAX)) +{ + free (relative_from); + relative_from = NULL; +} } free (targetdir); diff --git a/tests/ln/relative.sh b/tests/ln/relative.sh +# Expect this to fail with exit status 1. +# Prior to coreutils-8.23, it would segfault. +ln -sr '' F +test $? = 1 || fail=1 Won't the ln succeed on FreeBSD as per: http://bugs.gnu.org/13447 thanks, Pádraig.
bug#17010: Bug#739752: coreutils: ln segfaults when run with --relative and an empty target
On Thu, Mar 13, 2014 at 7:22 PM, Pádraig Brady p...@draigbrady.com wrote: On 03/14/2014 01:42 AM, Jim Meyering wrote: From a6d2db8b6dfe15344aba4aefe9545eb3a4876d45 Mon Sep 17 00:00:00 2001 From: Jim Meyering meyer...@fb.com Date: Thu, 13 Mar 2014 17:05:04 -0700 Subject: [PATCH] ln: with -sr, don't segfault for a TARGET of '' Prior to this change, ln -sr '' F would segfault, attempting to read path2[1] in relpath.c's path_common_prefix function. This problem arises whenever canonicalize_filename_mode returns NULL. * src/ln.c (convert_abs_rel): Call relpath only when both canonicalize_filename_mode calls return non-NULL. * tests/ln/relative.sh: Add a test to trigger this failure. * THANKS.in: List reporter's name/address. * NEWS (Bug fixes): Mention it. Reported by Erik Bernstein in 739...@bugs.debian.org. We can amend with the now allocated: Fixes http://bugs.gnu.org/17010 Done. diff --git a/NEWS b/NEWS ... + ln -sr '' F no longer segfaults: now it fails with the expected diagnostic Probably should add: [bug introduced with the --relative feature in coreutils-8.16] Definitely. Thanks. diff --git a/src/ln.c b/src/ln.c index aab9cf2..6726699 100644 --- a/src/ln.c +++ b/src/ln.c @@ -149,13 +149,17 @@ convert_abs_rel (const char *from, const char *target) char *realdest = canonicalize_filename_mode (targetdir, CAN_MISSING); char *realfrom = canonicalize_filename_mode (from, CAN_MISSING); Interesting. So canonicalize_filename_mode() can fail in this case, even with CAN_MISSING. It's unexpected that c_f_m() sets errno=ENOENT when CAN_MISSING is set. I wonder should we change that instead in gnulib? With CAN_MISSING I would expect this function to work on arbitrary strings, including the empty string. What would you have c_f_m(, CAN_MISSING) return? I know of no absolute name corresponding to the dot-relative empty string. diff --git a/tests/ln/relative.sh b/tests/ln/relative.sh +# Expect this to fail with exit status 1. +# Prior to coreutils-8.23, it would segfault. +ln -sr '' F +test $? = 1 || fail=1 Won't the ln succeed on FreeBSD as per: http://bugs.gnu.org/13447 You're right. Good catch. Adjusted as well. Thanks for the speedy and thorough review. Here's a revised patch: From 0093ac8d57a0f1a16fd09d98f6a524dddb6053e7 Mon Sep 17 00:00:00 2001 From: Jim Meyering meyer...@fb.com Date: Thu, 13 Mar 2014 17:05:04 -0700 Subject: [PATCH] ln: with -sr, don't segfault for a TARGET of '' Prior to this change, ln -sr '' F would segfault, attempting to read path2[1] in relpath.c's path_common_prefix function. This problem arises whenever canonicalize_filename_mode returns NULL. * src/ln.c (convert_abs_rel): Call relpath only when both canonicalize_filename_mode calls return non-NULL. * tests/ln/relative.sh: Add a test to trigger this failure. * THANKS.in: List reporter's name/address. * NEWS (Bug fixes): Mention it. Reported by Erik Bernstein in 739...@bugs.debian.org. Fixes http://bugs.gnu.org/17010. --- NEWS | 3 +++ THANKS.in| 1 + src/ln.c | 16 ++-- tests/ln/relative.sh | 5 + 4 files changed, 19 insertions(+), 6 deletions(-) diff --git a/NEWS b/NEWS index 62966b2..35d48e5 100644 --- a/NEWS +++ b/NEWS @@ -25,6 +25,9 @@ GNU coreutils NEWS-*- outline -*- it would display an error, requiring --no-dereference to avoid the issue. [bug introduced in coreutils-5.3.0] + ln -sr '' F no longer segfaults. Now works as expected. + [bug introduced with the --relative feature in coreutils-8.16] + shuf --repeat no longer dumps core if the input is empty. [bug introduced with the --repeat feature in coreutils-8.22] diff --git a/THANKS.in b/THANKS.in index 561d18c..2670265 100644 --- a/THANKS.in +++ b/THANKS.in @@ -193,6 +193,7 @@ Eric G. Miller e...@jps.net Eric Pementepeme...@northpark.edu Eric S. Raymond e...@snark.thyrsus.com Erik Bennettbenn...@cvo.oneworld.com +Erik Bernstein e...@fscking.org Erik Corry e...@kroete2.freinet.de Felix Lee f...@teleport.com Felix Rauch Valenti fra...@cse.unsw.edu.au diff --git a/src/ln.c b/src/ln.c index aab9cf2..6726699 100644 --- a/src/ln.c +++ b/src/ln.c @@ -149,13 +149,17 @@ convert_abs_rel (const char *from, const char *target) char *realdest = canonicalize_filename_mode (targetdir, CAN_MISSING); char *realfrom = canonicalize_filename_mode (from, CAN_MISSING); - /* Write to a PATH_MAX buffer. */ - char *relative_from = xmalloc (PATH_MAX); - - if (!relpath (realfrom, realdest, relative_from, PATH_MAX)) + char *relative_from = NULL; + if (realdest realfrom) { - free (relative_from); - relative_from = NULL; + /* Write to a PATH_MAX buffer. */ + relative_from = xmalloc (PATH_MAX); + + if