[mono] Mono coredumps when built with `_FORTIFY_SOURCE=2` forced.

[ayakael]

Description

Building runtime on current edge of Alpine Linux using mono generates a broken runtime.

Reproduction Steps

On Alpine Linux Edge, build a mono-flavored runtime, and try to execute ./dotnet build. Core dump will occur.

Expected behavior

Good execution

Actual behavior

Coredump

Regression?

Bug occurs on dotnet6 as well.

Known Workarounds

Undefining _FORTIFY_SOURCE in src/mono/mono/utils/mono-thread-coop.c

Configuration

7.0.103 and 6.0.114
Alpine Linux Edge
s390x and ppc64le, presumably any runtime built with mono-flavor

Other information

Running dotnet build through gdb yields the following:

0x000003fffd70fc8e in ?? () at /home/build/aports/community/dotnet7-build/src/dotnet-v7.0.103/src/runtime/artifacts/source-build/self/src/src/mono/mono/utils/mono-threads-coop.c:123 from /home/build/aports/community/dotnet7-build/src/dotnet-bunny-12/release/shared/Microsoft.NETCore.App/7.0.3/libcoreclr.so
123             g_assert (mono_threads_is_blocking_transition_enabled ());

Runtimes built previously still work, so something occurs during build for runtime to fail with src/mono/mono/utils/mono-threads-coop.c on line 123.

Tagging subscribers to this area: @dotnet/area-system-io
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Description

Building runtime on current edge of Alpine Linux using mono generates a broken runtime.

Reproduction Steps

On Alpine Linux Edge, build a mono-flavored runtime, and try to execute ./dotnet build. Core dump will occur.

Expected behavior

Good execution

Actual behavior

Coredump

Regression?

Bug occurs on dotnet6 as well.

Known Workarounds

Building on Alpine Linux v3.17

Configuration

7.0.103 and 6.0.114
Alpine Linux Edge
s390x and ppc64le, presumably any runtime built with mono-flavor

Other information

Running dotnet build through gdb yields the following:

0x000003fffd70fc8e in ?? () at /home/build/aports/community/dotnet7-build/src/dotnet-v7.0.103/src/runtime/artifacts/source-build/self/src/src/mono/mono/utils/mono-threads-coop.c:123 from /home/build/aports/community/dotnet7-build/src/dotnet-bunny-12/release/shared/Microsoft.NETCore.App/7.0.3/libcoreclr.so
123     /home/build/aports/community/dotnet7-build/src/dotnet-v7.0.103/src/runtime/artifacts/source-build/self/src/src/mono/mono/utils/mono-threads-coop.c: No such file or directory.

This thus points to a library issue. The following diff compares the library versions of last successful build and this one:

< readline(8.2.0-r0)
> ncurses-terminfo-base(6.4_p20230211-r0)
> ncurses-libs(6.4_p20230211-r0)
> readline(8.2.001-r0)
< xz-libs(5.4.0-r1)
> xz-libs(5.4.1-r0)
< llvm15-libs(15.0.6-r1)
< clang15-libs(15.0.6-r2)
< clang15(15.0.6-r2)
> llvm15-libs(15.0.7-r0)
> clang15-libs(15.0.7-r2)
> clang15(15.0.7-r2)
> lz4-libs(1.9.4-r1)
< cmake(3.25.1-r0)
> cmake(3.25.2-r0)
< dotnet7-build(7.0.101-r0)
< dotnet7-artifacts(7.0.101-r0)
> dotnet7-build(7.0.102-r0)
> dotnet7-artifacts(7.0.102-r0)
< findutils(4.9.0-r2)
> findutils(4.9.0-r3)
< icu-data-full(72.1-r1)
< icu-libs(72.1-r1)
< icu(72.1-r1)
< icu-dev(72.1-r1)
> icu-data-full(72.1-r2)
> icu-libs(72.1-r2)
> icu(72.1-r2)
> icu-dev(72.1-r2)
< libblkid(2.38.1-r2)
< libuuid(2.38.1-r2)
< libfdisk(2.38.1-r2)
< libmount(2.38.1-r2)
< libsmartcols(2.38.1-r2)
< util-linux-dev(2.38.1-r2)
> libblkid(2.38.1-r4)
> libuuid(2.38.1-r4)
> libfdisk(2.38.1-r4)
> libmount(2.38.1-r4)
> libsmartcols(2.38.1-r4)
> util-linux-dev(2.38.1-r4)
< libcom_err(1.46.5-r5)
< e2fsprogs-libs(1.46.5-r5)
< e2fsprogs-dev(1.46.5-r5)
> libcom_err(1.47.0-r0)
> e2fsprogs-libs(1.47.0-r0)
> e2fsprogs-dev(1.47.0-r0)
< glib(2.74.4-r0)
> glib(2.74.5-r0)
< libldap(2.6.3-r6)
> libldap(2.6.4-r0)
< openssl-dev(3.0.7-r2)
< nghttp2-dev(1.51.0-r0)
> openssl-dev(3.0.8-r0)
> nghttp2-dev(1.52.0-r0)
< curl-dev(7.87.0-r2)
> curl-dev(7.88.0-r1)
< libgit2(1.5.0-r2)
< bsd-compat-headers(0.7.2-r3)
< ncurses-dev(6.4_p20230107-r0)
> libgit2(1.5.1-r0)
> bsd-compat-headers(0.7.2-r4)
> ncurses-dev(6.4_p20230211-r0)
> libedit(20221030.3.1-r0)
< libgit2-dev(1.5.0-r2)
> libgit2-dev(1.5.1-r0)
< xz(5.4.0-r1)
< xz-dev(5.4.0-r1)
> xz(5.4.1-r0)
> xz-dev(5.4.1-r0)
< linux-headers(6.1.0-r0)
> linux-headers(6.1.11-r0)
< python3(3.11.1-r2)
< lldb(15.0.6-r0)
< lldb-dev(15.0.6-r0)
< llvm15(15.0.6-r1)
> python3(3.11.2-r0)
> lldb(15.0.7-r0)
> lldb-dev(15.0.7-r0)
> llvm15(15.0.7-r0)
< zstd(1.5.2-r10)
< zstd-dev(1.5.2-r10)
> zstd(1.5.4-r0)
> zstd-dev(1.5.4-r0)
< py3-setuptools(65.6.3-r0)
< llvm15-test-utils(15.0.6-r1)
< llvm15-dev(15.0.6-r1)
< userspace-rcu(0.13.2-r0)
< userspace-rcu-dev(0.13.2-r0)
> py3-setuptools(67.3.2-r0)
> llvm15-test-utils(15.0.7-r0)
> llvm15-dev(15.0.7-r0)
> userspace-rcu(0.14.0-r0)
> userspace-rcu-dev(0.14.0-r0)
< nodejs(18.12.1-r0)
< numactl(2.0.16-r0)
< numactl-dev(2.0.16-r0)
> c-ares(1.19.0-r0)
> nodejs(18.14.0-r0)
> numactl(2.0.16-r2)
> numactl-dev(2.0.16-r2)
> popt(1.19-r1)
> rsync(3.2.7-r0)
< skalibs(2.12.0.1-r0)
< utmps-libs(0.1.2.0-r1)
> skalibs(2.13.0.0-r0)
> utmps-libs(0.1.2.1-r0)
< npm(9.2.0-r0)
> npm(9.4.2-r0)
< libelf(0.188-r0)
> libelf(0.188-r1)< readline(8.2.0-r0)
> ncurses-terminfo-base(6.4_p20230211-r0)
> ncurses-libs(6.4_p20230211-r0)
> readline(8.2.001-r0)
< xz-libs(5.4.0-r1)
> xz-libs(5.4.1-r0)
< llvm15-libs(15.0.6-r1)
< clang15-libs(15.0.6-r2)
< clang15(15.0.6-r2)
> llvm15-libs(15.0.7-r0)
> clang15-libs(15.0.7-r2)
> clang15(15.0.7-r2)
> lz4-libs(1.9.4-r1)
< cmake(3.25.1-r0)
> cmake(3.25.2-r0)
< dotnet7-build(7.0.101-r0)
< dotnet7-artifacts(7.0.101-r0)
> dotnet7-build(7.0.102-r0)
> dotnet7-artifacts(7.0.102-r0)
< findutils(4.9.0-r2)
> findutils(4.9.0-r3)
< icu-data-full(72.1-r1)
< icu-libs(72.1-r1)
< icu(72.1-r1)
< icu-dev(72.1-r1)
> icu-data-full(72.1-r2)
> icu-libs(72.1-r2)
> icu(72.1-r2)
> icu-dev(72.1-r2)
< libblkid(2.38.1-r2)
< libuuid(2.38.1-r2)
< libfdisk(2.38.1-r2)
< libmount(2.38.1-r2)
< libsmartcols(2.38.1-r2)
< util-linux-dev(2.38.1-r2)
> libblkid(2.38.1-r4)
> libuuid(2.38.1-r4)
> libfdisk(2.38.1-r4)
> libmount(2.38.1-r4)
> libsmartcols(2.38.1-r4)
> util-linux-dev(2.38.1-r4)
< libcom_err(1.46.5-r5)
< e2fsprogs-libs(1.46.5-r5)
< e2fsprogs-dev(1.46.5-r5)
> libcom_err(1.47.0-r0)
> e2fsprogs-libs(1.47.0-r0)
> e2fsprogs-dev(1.47.0-r0)
< glib(2.74.4-r0)
> glib(2.74.5-r0)
< libldap(2.6.3-r6)
> libldap(2.6.4-r0)
< openssl-dev(3.0.7-r2)
< nghttp2-dev(1.51.0-r0)
> openssl-dev(3.0.8-r0)
> nghttp2-dev(1.52.0-r0)
< curl-dev(7.87.0-r2)
> curl-dev(7.88.0-r1)
< libgit2(1.5.0-r2)
< bsd-compat-headers(0.7.2-r3)
< ncurses-dev(6.4_p20230107-r0)
> libgit2(1.5.1-r0)
> bsd-compat-headers(0.7.2-r4)
> ncurses-dev(6.4_p20230211-r0)
> libedit(20221030.3.1-r0)
< libgit2-dev(1.5.0-r2)
> libgit2-dev(1.5.1-r0)
< xz(5.4.0-r1)
< xz-dev(5.4.0-r1)
> xz(5.4.1-r0)
> xz-dev(5.4.1-r0)
< linux-headers(6.1.0-r0)
> linux-headers(6.1.11-r0)
< python3(3.11.1-r2)
< lldb(15.0.6-r0)
< lldb-dev(15.0.6-r0)
< llvm15(15.0.6-r1)
> python3(3.11.2-r0)
> lldb(15.0.7-r0)
> lldb-dev(15.0.7-r0)
> llvm15(15.0.7-r0)
< zstd(1.5.2-r10)
< zstd-dev(1.5.2-r10)
> zstd(1.5.4-r0)
> zstd-dev(1.5.4-r0)
< py3-setuptools(65.6.3-r0)
< llvm15-test-utils(15.0.6-r1)
< llvm15-dev(15.0.6-r1)
< userspace-rcu(0.13.2-r0)
< userspace-rcu-dev(0.13.2-r0)
> py3-setuptools(67.3.2-r0)
> llvm15-test-utils(15.0.7-r0)
> llvm15-dev(15.0.7-r0)
> userspace-rcu(0.14.0-r0)
> userspace-rcu-dev(0.14.0-r0)
< nodejs(18.12.1-r0)
< numactl(2.0.16-r0)
< numactl-dev(2.0.16-r0)
> c-ares(1.19.0-r0)
> nodejs(18.14.0-r0)
> numactl(2.0.16-r2)
> numactl-dev(2.0.16-r2)
> popt(1.19-r1)
> rsync(3.2.7-r0)
< skalibs(2.12.0.1-r0)
< utmps-libs(0.1.2.0-r1)
> skalibs(2.13.0.0-r0)
> utmps-libs(0.1.2.1-r0)
< npm(9.2.0-r0)
> npm(9.4.2-r0)
< libelf(0.188-r0)
> libelf(0.188-r1)
< util-linux-misc(2.38.1-r2)
> setarch(2.38.1-r4)
> util-linux-misc(2.38.1-r4)
< inetutils-syslogd-openrc(2.4-r1)
< util-linux-misc(2.38.1-r2)
> setarch(2.38.1-r4)
> util-linux-misc(2.38.1-r4)
< inetutils-syslogd-openrc(2.4-r1)

Runtimes built previously still work, so something occurs during build for linking to fail with src/mono/mono/utils/mono-threads-coop.c on line 123. Any library updates that look suspicious?

Author:	ayakael
Assignees:	-
Labels:	area-System.IO
Milestone:	-

[ayakael]

Of note, both builds use the backported version of #76500. Indeed, without this, mono flavored runtime doesn't even build on musl.

[ayakael]

Further investigation reveals a more telling error:

Reading symbols from ./dotnet...
(gdb) run
Starting program: /var/build/dotnet7/community/dotnet7-stage0/src/test/dotnet build

Program received signal SIGILL, Illegal instruction.
copy_stack_data_internal (stackdata_begin=0x7fffffffdb98, info=<optimized out>, wrapper_data1=<optimized out>, wrapper_data2=<optimized out>) at /var/build/dotnet7/community/dotnet7-stage0/src/dotnet-v7.0.101-source-build/src/runtime/src/mono/mono/utils/mono-threads-coop.c:193
193		memcpy (state->gc_stackdata, stackdata_end, stackdata_size);

A similar bug occured in another package I maintain, and was brough to the surface by https://gitlab.alpinelinux.org/alpine/aports/-/merge_requests/43463, which makes clang15 enable fortify-source by default. The actual bug is a target mismatch bug introduced in gcc12. It is tracked here: https://gitlab.alpinelinux.org/alpine/aports/-/issues/14105

Applying the following patch does the trick as a workaround:

From 98054ea87ce70247bb09ceafd2ad1a0b36d2fef4 Mon Sep 17 00:00:00 2001
Patch-Source: https://github.com/dotnet/runtime/issues/82269
From: Antoine Martin <dev@ayakael.net>
Date: Sat, 1 Oct 2022 09:21:58 -0400
Subject: [PATCH] Undefine fortify-source on mono-thread-coop

When _FORTIFY_SOURCE=2, there is a bug relating to memcpy that expresses itself.
See: https://gitlab.alpinelinux.org/alpine/aports/-/issues/14105. Alpine Linux
now sets this by default since https://gitlab.alpinelinux.org/alpine/aports/-/merge_requests/43463,
which makes mono-flavored runtime dump its core. This patch offers a workaround
by undefining _FORTIFY_SOURCE in the problematic file.

---
diff --git a/src/runtime/src/mono/mono/utils/mono-threads-coop.c b/src/runtime/src/mono/mono/utils/mono-threads-coop.c
index 4ed659d6605..34bb5785fba 100644
--- a/src/runtime/src/mono/mono/utils/mono-threads-coop.c
+++ b/src/runtime/src/mono/mono/utils/mono-threads-coop.c
@@ -15,6 +15,7 @@
 #ifdef TARGET_MACH
 #define _DARWIN_C_SOURCE
 #endif
+#undef _FORTIFY_SOURCE
 
 #include <mono/utils/mono-compiler.h>
 #include <mono/utils/mono-threads.h>

[nekopsykose]

copying what i wrote downstream for reference:

it's not; this is an actual fortify violation. clang is also not affected by that inlining issue, which manifests as a compilation failure, not a runtime sigill.
that issue is about gcc failing to compile code at all in certain situations with fortify. this one is that the code is wrong, and crashes on the fortify assertion. breaking fortify hits e.g. a ud2 instruction (SIGILL), because the code did something incorrect.
so, strictly speaking, in this case it is (probably) not a compiler bug and undefining fortify is not correct. the code itself does something wrong in copy_stack_data_internal with the memcpy. the most common cause is that it fails to uphold the requirement for memcpy: the start and end don't overlap.

from man 3 memcpy:

 DESCRIPTION
        The memcpy() function copies n bytes from memory area src to memory
        area dest.  The memory areas must not overlap.  Use memmove(3) if the
        memory areas do overlap.

if you change that memcpy to memmove on line 193, it will probably pass. and if you write some sample code where the memory overlaps, you will reproduce the same crash (with either gcc or clang utilising the fortify headers). they're working as intended :)

[ayakael]

Changing to memcpy to memmove still yields an error:

Reading symbols from ./dotnet...
(gdb) run
Starting program: /var/build/dotnet7/community/dotnet7-stage0/src/test/dotnet build

Program received signal SIGILL, Illegal instruction.
copy_stack_data_internal (stackdata_begin=0x7fffffffd848, info=<optimized out>, wrapper_data1=<optimized out>, wrapper_data2=<optimized out>) at /var/build/dotnet7/community/dotnet7-stage0/src/dotnet-v7.0.101-source-build/src/runtime/src/mono/mono/utils/mono-threads-coop.c:193
193             memmove (state->gc_stackdata, stackdata_end, stackdata_size);

[nekopsykose]

the memmove

_FORTIFY_FN(memmove) void *memmove(void * _FORTIFY_POS0 __d,
                                   const void * _FORTIFY_POS0 __s, size_t __n)
{
	size_t __bd = __builtin_object_size(__d, 0);
	size_t __bs = __builtin_object_size(__s, 0);

	if (__n > __bd || __n > __bs)
		__builtin_trap();
	return __orig_memmove(__d, __s, __n);
}

__builtin_object_size would return (size_t)-1 if it fails, so then the check would pass. meaning, the assertion is real, and the __n is probably wrong

the memcpy fortify is that, with an extra overlap check.