Re: [qemu-s390x] [Qemu-devel] linux-user: s390x issue on Fedora 30 (dyna

Hi, David.

I can't repro the problem either, but I do have a patch authored by an engineer that left the company a while ago that seems to be at least related to your scenario. May I ask you to test it (just apply it to QEMU ToT and rerun the scenario to see if it changes the outcome) (feel free to experiment and modify the changes)? I never managed to understand the purpose of that patch (I lost the contact with the former engineer, and the patch is without any comment), so I never dared to send it to the list, but I hope your scenario may actually explain the purpose and the origin of the patch.

I am attaching the patch both as a file and inline.

Yours,

Aleksandar

From 377f99e807f4aa42ece9f0cd437f50af11611b4c Mon Sep 17 00:00:00 2001
From: Aleksandar Markovic <address@hidden>
Date: Mon, 19 Aug 2019 15:20:29 +0200
Subject: [PATCH] linux-user: Special case /etc/ld.cache.so and pretend it does
not exist

Prevent target executables from opening host ld.so.cache.
---
linux-user/syscall.c | 28 ++++++++++++++++++++++++++++
1 file changed, 28 insertions(+)

diff --git a/linux-user/syscall.c b/linux-user/syscall.c
index 8367cb1..f5bae6e 100644
--- a/linux-user/syscall.c
+++ b/linux-user/syscall.c
@@ -8820,6 +8820,10 @@ static abi_long do_syscall1(void *cpu_env, int num, abi_long arg1,
if (!(p = lock_user_string(arg1))) {
return -TARGET_EFAULT;
}
+ if (strcmp("/etc/ld.so.cache", path(p)) == 0) {
+ unlock_user(p, arg1, 0);
+ return -TARGET_ENONET;
+ }
ret = get_errno(statfs(path(p), &stfs));
unlock_user(p, arg1, 0);
convert_statfs:
@@ -8859,6 +8863,10 @@ static abi_long do_syscall1(void *cpu_env, int num, abi_long arg1,
if (!(p = lock_user_string(arg1))) {
return -TARGET_EFAULT;
}
+ if (strcmp("/etc/ld.so.cache", path(p)) == 0) {
+ unlock_user(p, arg1, 0);
+ return -TARGET_ENONET;
+ }
ret = get_errno(statfs(path(p), &stfs));
unlock_user(p, arg1, 0);
convert_statfs64:
@@ -9059,6 +9067,10 @@ static abi_long do_syscall1(void *cpu_env, int num, abi_long arg1,
if (!(p = lock_user_string(arg1))) {
return -TARGET_EFAULT;
}
+ if (strcmp("/etc/ld.so.cache", path(p)) == 0) {
+ unlock_user(p, arg1, 0);
+ return -TARGET_ENONET;
+ }
ret = get_errno(stat(path(p), &st));
unlock_user(p, arg1, 0);
goto do_stat;
@@ -9068,6 +9080,10 @@ static abi_long do_syscall1(void *cpu_env, int num, abi_long arg1,
if (!(p = lock_user_string(arg1))) {
return -TARGET_EFAULT;
}
+ if (strcmp("/etc/ld.so.cache", path(p)) == 0) {
+ unlock_user(p, arg1, 0);
+ return -TARGET_ENONET;
+ }
ret = get_errno(lstat(path(p), &st));
unlock_user(p, arg1, 0);
goto do_stat;
@@ -10268,6 +10284,10 @@ static abi_long do_syscall1(void *cpu_env, int num, abi_long arg1,
if (!(p = lock_user_string(arg1))) {
return -TARGET_EFAULT;
}
+ if (strcmp("/etc/ld.so.cache", path(p)) == 0) {
+ unlock_user(p, arg1, 0);
+ return -TARGET_ENONET;
+ }
ret = get_errno(stat(path(p), &st));
unlock_user(p, arg1, 0);
if (!is_error(ret))
@@ -10279,6 +10299,10 @@ static abi_long do_syscall1(void *cpu_env, int num, abi_long arg1,
if (!(p = lock_user_string(arg1))) {
return -TARGET_EFAULT;
}
+ if (strcmp("/etc/ld.so.cache", path(p)) == 0) {
+ unlock_user(p, arg1, 0);
+ return -TARGET_ENONET;
+ }
ret = get_errno(lstat(path(p), &st));
unlock_user(p, arg1, 0);
if (!is_error(ret))
@@ -10319,6 +10343,10 @@ static abi_long do_syscall1(void *cpu_env, int num, abi_long arg1,
if (p == NULL) {
return -TARGET_EFAULT;
}
+ if (strcmp("/etc/ld.so.cache", path(p)) == 0) {
+ unlock_user(p, arg1, 0);
+ return -TARGET_ENONET;
+ }
#if defined(__NR_statx)
{
/*
--
2.7.4

On Mon, Aug 19, 2019 at 3:08 PM David Hildenbrand <address@hidden> wrote:

On 19.08.19 14:11, Peter Maydell wrote:
> On Sat, 17 Aug 2019 at 17:14, David Hildenbrand <address@hidden> wrote:
>>
>> On 17.08.19 17:59, David Hildenbrand wrote:
>>> Hi everybody,
>>>
>>> I was just trying to run qemu-s390x (linux-user) with a very simple
>>> binary (gzip + lib/ld64.so.1, compiled under Fedora 27). This used to
>>> work just fine a while ago (especially when I was working on vector
>>> instructions using QEMU v3.1). However, now I can't get past a SEGFAULT
>>> in the dynamic library loader (I assume it is trying to locate glibc). I
>>> tried a couple of other binaries that definitely used to work (from
>>> Fedora 30).
>>>
>>> I checked QEMU v4.1, v4.0 and v3.1. All are broken for me. Which is
>>> weird - because it used to work :/
>>>
>>> I remember that I was running Fedora 29 the last time I had it running,
>>> so my gut feeling is that this is related to some other system library
>>> (but which?). I am running on an up-to-date Fedora 30 x86-64 now.
>>>
>>> Any ideas? Has this been reported already? (not sure if this is a Fedora
>>> 30 issue)
>
> I'm pretty sure the problem you've run into is a long standing
> bug in the glibc dynamic loader. It cannot cope with the ld.so.cache
> being for the wrong endianness. (Correct endianness but incorrect
> architecture it correctly detects and ignores). The result is that
> running a linux-user QEMU dynamic binary for big-endian on little-endian
> like this will crash in the dynamic loader unless you arrange that it can't
> find the host's ld.so.cache somehow, eg:
> (a) run inside a chroot
> (b) create an empty /etc/ld.so.cache file inside the -L directory
>
> The ideal fix would be if somebody cared enough to track down
> and fix the ld.so bug.
>
> Compare:
> https://bugs.launchpad.net/qemu/+bug/1701798
> https://bugs.launchpad.net/qemu/+bug/1835693
>
> thanks
> -- PMM
>

Thanks, running

"ldconfig -c etc/ld.so.cache -r ."

Seems to fix the issue for me. So you are sure the bug resides in glic
and not in the qemu-user pieces of the library loader?

--

Thanks,

David / dhildenb

From:	Aleksandar Markovic
Subject:	Re: [qemu-s390x] [Qemu-devel] linux-user: s390x issue on Fedora 30 (dynamic library loader?)
Date:	Mon, 19 Aug 2019 15:34:27 +0200