Skip to contentSkip to conversation

xfs_scrub: fix strerror_r usage yet again

PATCH In review Darrick J. Wong 5 participants 10 messages Sep 19, 2025 – Oct 3, 2025

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 xfs_scrub: fix strerror_r usage yet again

From: Darrick J. Wong <[email protected]>

In commit 75faf2bc907584, someone tried to fix scrub to use the POSIX version of strerror_r so that the build would work with musl. Unfortunately, neither the author nor myself remembered that GNU libc imposes its own version any time _GNU_SOURCE is defined, which builddefs.in always does. Regrettably, the POSIX and GNU versions have different return types and the GNU version can return any random pointer, so now this code is broken on glibc.

"Fix" this standards body own goal by casting the return value to intptr_t and employing some gross heuristics to guess at the location of the actual error string.

Fixes 75faf2bc907584 ("xfs_scrub: Use POSIX-conformant strerror_r")

---
v2: go the autoconf route

---
 configure.ac          |    1 +
 include/builddefs.in  |    1 +
 m4/package_libcdev.m4 |   46 ++++++++++++++++++++++++++++++++++++++++++++++
 scrub/Makefile        |    4 ++++
 scrub/common.c        |    8 ++++++--
 5 files changed, 58 insertions(+), 2 deletions(-)
configure.ac +1 -0
index d2407cb5de5af2..df19379b02ba55 100644
@@ -183,6 +183,7 @@ AC_CONFIG_CROND_DIR
 AC_CONFIG_UDEV_RULE_DIR
 AC_HAVE_BLKID_TOPO
 AC_HAVE_TRIVIAL_AUTO_VAR_INIT
+AC_STRERROR_R_RETURNS_STRING
 
 if test "$enable_ubsan" = "yes" || test "$enable_ubsan" = "probe"; then
         AC_PACKAGE_CHECK_UBSAN
include/builddefs.in +1 -0
index 93b5c75155c0f4..5aa5742bb31b9e 100644
@@ -241,6 +241,7 @@ CROND_DIR = @crond_dir@
 HAVE_UDEV = @have_udev@
 UDEV_RULE_DIR = @udev_rule_dir@
 HAVE_LIBURCU_ATOMIC64 = @have_liburcu_atomic64@
+STRERROR_R_RETURNS_STRING = @strerror_r_returns_string@
 
 GCCFLAGS = -funsigned-char -fno-strict-aliasing -Wall -Werror -Wextra -Wno-unused-parameter
 #	   -Wbitwise -Wno-transparent-union -Wno-old-initializer -Wno-decl
m4/package_libcdev.m4 +46 -0

49 lines not on this page · raw message

scrub/Makefile +4 -0
index 3636a47942e98e..6375d77a291bcb 100644
@@ -105,6 +105,10 @@ CFILES += unicrash.c
 LCFLAGS += -DHAVE_LIBICU $(LIBICU_CFLAGS)
 endif
 
+ifeq ($(STRERROR_R_RETURNS_STRING),yes)
+LCFLAGS += -DSTRERROR_R_RETURNS_STRING
+endif
+
 # Automatically trigger a media scan once per month
 XFS_SCRUB_ALL_AUTO_MEDIA_SCAN_INTERVAL=1mo
 
scrub/common.c +6 -2
index 9437d0abb8698b..9a33e2a9d54ed4 100644
@@ -126,8 +126,12 @@ __str_out(
 	fprintf(stream, "%s%s: %s: ", stream_start(stream),
 			_(err_levels[level].string), descr);
 	if (error) {
-		strerror_r(error, buf, DESCR_BUFSZ);
-		fprintf(stream, _("%s."), buf);
+#ifdef STRERROR_R_RETURNS_STRING
+		fprintf(stream, _("%s."), strerror_r(error, buf, DESCR_BUFSZ));
+#else
+		if (strerror_r(error, buf, DESCR_BUFSZ) == 0)
+			fprintf(stream, _("%s."), buf);
+#endif
 	} else {
 		va_start(args, format);
 		vfprintf(stream, format, args);
Reviewed-by on v2 1/1
#1 lore raw
Quoted 130 lines · Darrick J. Wong

From: Darrick J. Wong <[email protected]>

In commit 75faf2bc907584, someone tried to fix scrub to use the POSIX version of strerror_r so that the build would work with musl. Unfortunately, neither the author nor myself remembered that GNU libc imposes its own version any time _GNU_SOURCE is defined, which builddefs.in always does. Regrettably, the POSIX and GNU versions have different return types and the GNU version can return any random pointer, so now this code is broken on glibc.

"Fix" this standards body own goal by casting the return value to intptr_t and employing some gross heuristics to guess at the location of the actual error string.

Fixes 75faf2bc907584 ("xfs_scrub: Use POSIX-conformant strerror_r")

---
v2: go the autoconf route
---

configure.ac          |    1 +
include/builddefs.in  |    1 +
m4/package_libcdev.m4 |   46 ++++++++++++++++++++++++++++++++++++++++++++++
scrub/Makefile        |    4 ++++
scrub/common.c        |    8 ++++++--
5 files changed, 58 insertions(+), 2 deletions(-)
configure.ac +0 -0
index d2407cb5de5af2..df19379b02ba55 100644
@@ -183,6 +183,7 @@ AC_CONFIG_CROND_DIR

AC_CONFIG_UDEV_RULE_DIR
AC_HAVE_BLKID_TOPO
AC_HAVE_TRIVIAL_AUTO_VAR_INIT
+AC_STRERROR_R_RETURNS_STRING

if test "$enable_ubsan" = "yes" || test "$enable_ubsan" = "probe"; then
AC_PACKAGE_CHECK_UBSAN

include/builddefs.in +0 -0
index 93b5c75155c0f4..5aa5742bb31b9e 100644
@@ -241,6 +241,7 @@ CROND_DIR = @crond_dir@

HAVE_UDEV = @have_udev@
UDEV_RULE_DIR = @udev_rule_dir@ HAVE_LIBURCU_ATOMIC64 = @have_liburcu_atomic64@ +STRERROR_R_RETURNS_STRING = @strerror_r_returns_string@

GCCFLAGS = -funsigned-char -fno-strict-aliasing -Wall -Werror -Wextra -Wno-unused-parameter
# -Wbitwise -Wno-transparent-union -Wno-old-initializer -Wno-decl

m4/package_libcdev.m4 +46 -0
index ce1ba47264659c..c5538c30d2518a 100644
@@ -301,3 +301,49 @@ syscall(__NR_file_getattr, 0, 0, 0, 0, 0);
       AC_MSG_RESULT(no))
    AC_SUBST(have_file_getattr)
  ])
+
+#
+# Check if strerror_r returns an int, as opposed to a char *, because there are
+# two versions of this function, with differences that are hard to detect.
+#
+# GNU strerror_r returns a pointer to a string on success, but the returned
+# pointer might point to a static buffer and not buf, so you have to use the
+# return value.  The declaration has the __warn_unused_result__ attribute to
+# enforce this.
+#
+# XSI strerror_r always writes to buf and returns 0 on success, -1 on error.
+#
+# How do you select a particular version?  By defining macros, of course!
+# _GNU_SOURCE always gets you the GNU version, and _POSIX_C_SOURCE >= 200112L
+# gets you the XSI version but only if _GNU_SOURCE isn't defined.
+#
+# The build system #defines _GNU_SOURCE unconditionally, so when compiling
+# against glibc we get the GNU version.  However, when compiling against musl,
+# the _GNU_SOURCE definition does nothing and we get the XSI version anyway.
+# Not definining _GNU_SOURCE breaks the build in many areas, so we'll create
+# yet another #define for just this weird quirk so that we can patch around it
+# in the one place we need it.
+#
+# Note that we have to force erroring out on the int conversion warnings
+# because C doesn't consider it a hard error to cast a char pointer to an int
+# even when CFLAGS contains -std=gnu11.
+AC_DEFUN([AC_STRERROR_R_RETURNS_STRING],
+  [AC_MSG_CHECKING([if strerror_r returns char *])
+    OLD_CFLAGS="$CFLAGS"
+    CFLAGS="$CFLAGS -Wall -Werror"
+    AC_LINK_IFELSE(
+    [AC_LANG_PROGRAM([[
+#define _GNU_SOURCE
+#include <stdio.h>
+#include <string.h>
+  ]], [[
+char buf[1024];
+puts(strerror_r(0, buf, sizeof(buf)));
+  ]])
+    ],
+       strerror_r_returns_string=yes
+       AC_MSG_RESULT(yes),
+       AC_MSG_RESULT(no))
+    CFLAGS="$OLD_CFLAGS"
+    AC_SUBST(strerror_r_returns_string)
+  ])
scrub/Makefile +0 -0
index 3636a47942e98e..6375d77a291bcb 100644
@@ -105,6 +105,10 @@ CFILES += unicrash.c
 LCFLAGS += -DHAVE_LIBICU $(LIBICU_CFLAGS)

endif

scrub/Makefile +4 -0
+ifeq ($(STRERROR_R_RETURNS_STRING),yes)
+LCFLAGS += -DSTRERROR_R_RETURNS_STRING
+endif
+

# Automatically trigger a media scan once per month XFS_SCRUB_ALL_AUTO_MEDIA_SCAN_INTERVAL=1mo

scrub/common.c +0 -1
index 9437d0abb8698b..9a33e2a9d54ed4 100644
@@ -126,8 +126,12 @@ __str_out(
 fprintf(stream, "%s%s: %s: ", stream_start(stream),
 _(err_levels[level].string), descr);
 if (error) {
- strerror_r(error, buf, DESCR_BUFSZ);

scrub/common.c __str_out( · line 137

Quoted 10 lines
scrub/common.c +6 -1
- fprintf(stream, _("%s."), buf);
+#ifdef STRERROR_R_RETURNS_STRING
+ fprintf(stream, _("%s."), strerror_r(error, buf, DESCR_BUFSZ));
+#else
+ if (strerror_r(error, buf, DESCR_BUFSZ) == 0)
+ fprintf(stream, _("%s."), buf);
+#endif
 } else {
 va_start(args, format);
 vfprintf(stream, format, args);

I did check *build* on glibc, but I don’t think mine had warn_unused_result yet. (It’s an older version.)

Thanks for fixing this, looks good to me.

The commit message isn’t accurate any more though.

#2 lore raw
In reply to Darrick J. Wong

The autoconf maigc looks good (as good as autoconf can look anyway), but why is this code using strerror_r to start with? AFAIK on Linux strerror is using thread local storage since the damn of time, so just doing this as:

	fprintf(stream, _("%s."), strerror(error));

should be perfectly fine, while also simpler and more portable.

#3 lore raw
Quoted 8 lines · Christoph Hellwig

The autoconf maigc looks good (as good as autoconf can look anyway), but why is this code using strerror_r to start with? AFAIK on Linux strerror is using thread local storage since the damn of time, so just doing this as:

	fprintf(stream, _("%s."), strerror(error));

should be perfectly fine, while also simpler and more portable.

But there's no guarantee that the implementation does this, is there? The manpage doesn't say anything like that.

--D

#4 lore raw
Quoted 6 lines · Darrick J. Wong
Quoted 5 lines · Christoph Hellwig

The autoconf maigc looks good (as good as autoconf can look anyway), but why is this code using strerror_r to start with? AFAIK on Linux strerror is using thread local storage since the damn of time, so just doing this as:

Quoted 6 lines
Quoted 3 lines
	fprintf(stream, _("%s."), strerror(error));

should be perfectly fine, while also simpler and more portable.

But there's no guarantee that the implementation does this, is there? The manpage doesn't say anything like that.

To me this pretty clear reads that the return string is stable until the next call to strerror as long as you only use it in the calling thread:

"The strerror() function returns a pointer to a string that describes the error code passed in the argument errnum, ...
This string must not be modified by the application, and the returned pointer will be invalidated on a subsequent call to strerror() or strerror_l(), or if the thread that obtained the string exits. No other library function, including perror(3), will modify this string."

#5 lore raw
Quoted 18 lines · Christoph Hellwig
Quoted 6 lines · Darrick J. Wong
Quoted 5 lines · Christoph Hellwig

The autoconf maigc looks good (as good as autoconf can look anyway), but why is this code using strerror_r to start with? AFAIK on Linux strerror is using thread local storage since the damn of time, so just doing this as:

Quoted 6 lines
Quoted 3 lines
	fprintf(stream, _("%s."), strerror(error));

should be perfectly fine, while also simpler and more portable.

But there's no guarantee that the implementation does this, is there? The manpage doesn't say anything like that.

To me this pretty clear reads that the return string is stable until the next call to strerror as long as you only use it in the calling thread:

Quoted 6 lines

"The strerror() function returns a pointer to a string that describes the error code passed in the argument errnum, ...
This string must not be modified by the application, and the returned pointer will be invalidated on a subsequent call to strerror() or strerror_l(), or if the thread that obtained the string exits. No other library function, including perror(3), will modify this string."

Huh, I don't see "...or if the thread that obtained the string exits" in my version of the strerror(3) manpage on Debian 12. For that matter, it still lists strerror() as MT-Unsafe.

Looking through the manpages git repo, I see that strerror actually does get changed to MT-Safe in commit 01576f703f820e54602:

strerror.3: Change strerror() reference from MT-Unsafe to MT-Safe

The information in this patch was obtained from a glibc upstream patch, commit 28aff047818eb1726394296d27b9c7885340bead ("string: Implement strerror in terms of strerror_l").

According to the patch above, for glibc versions >=2.32, strerror() is MT-Safe.

Then looking through the glibc repository, that 28aff047818e commit comes from:

https://sourceware.org/git/?p=glibc.git;a=commitdiff;h=28aff047818eb1726394296d27b9c7885340bead;hp=9deec7c8bab24659e78172dd850f4ca37c57940c

which turns strerror into strerror_l. This commit only appears in glibc 2.32, which appeared in August 2020. That version first appears in Debian 12 and RHEL 9, which are still fairly new. musl apparently has had strerror call strerror_l since at least 2011.

Has strerror() been designated as thread-safe at a POSIX level, or is this just an implementation quirk of these two C libraries? strerror definitely wasn't thread-safe on glibc when I wrote this program.

--D

#7 lore raw
Quoted 9 lines · Holger Hoffstätte
Quoted 3 lines · Darrick J. Wong

Has strerror() been designated as thread-safe at a POSIX level, or is this just an implementation quirk of these two C libraries? strerror definitely wasn't thread-safe on glibc when I wrote this program.

It still is not:
https://pubs.opengroup.org/onlinepubs/9799919799/functions/strerror.html

Pretty safe to say that this particular train has sailed.

Sailed off the end of the pier, yeah. ;)

Andrey: could you pick this one up, please?

--D

Quoted 3 lines

cheers
Holger

#8 lore raw
Quoted 6 lines · Darrick J. Wong
Quoted 5 lines · Holger Hoffstätte
Quoted 3 lines · Darrick J. Wong

Has strerror() been designated as thread-safe at a POSIX level, or is this just an implementation quirk of these two C libraries? strerror definitely wasn't thread-safe on glibc when I wrote this program.

Quoted 8 lines
Quoted 4 lines

It still is not:
https://pubs.opengroup.org/onlinepubs/9799919799/functions/strerror.html

Pretty safe to say that this particular train has sailed.

Sailed off the end of the pier, yeah. ;)

Andrey: could you pick this one up, please?

I will include this one in next for-next

Signature · 1 line
- Andrey
#9 lore raw
In reply to Darrick J. Wong
Quoted 5 lines · Darrick J. Wong

Then looking through the glibc repository, that 28aff047818e commit comes from:

https://sourceware.org/git/?p=glibc.git;a=commitdiff;h=28aff047818eb1726394296d27b9c7885340bead;hp=9deec7c8bab24659e78172dd850f4ca37c57940c

Quoted 8 lines

which turns strerror into strerror_l. This commit only appears in glibc 2.32, which appeared in August 2020. That version first appears in Debian 12 and RHEL 9, which are still fairly new. musl apparently has had strerror call strerror_l since at least 2011.

Has strerror() been designated as thread-safe at a POSIX level, or is this just an implementation quirk of these two C libraries? strerror definitely wasn't thread-safe on glibc when I wrote this program.

No POSIX gurantee (and I don't think that matters for xfs_scrub), but glibc before wasn't exactly thread unsafe before. It just simply leaked the error buffer when it had to print "unknown erorr.." for errors now known to it.