i386-linux-user wine exception regression tests fail

Bug #1915682 reported by Dirk A Niggemann
6
This bug affects 1 person
Affects Status Importance Assigned to Milestone
QEMU
Expired
Undecided
Unassigned

Bug Description

When trying to run wine (latest devel from git) regression tests for ntdll in a statically linked qemu-i386 (commit 392b9a74b9b621c52d05e37bc6f41f1bbab5c6f8) on arm32 (raspberry pi 4) in a debian buster chroot, the exception tests fail at the first test with an infinite exception loop.

WINEDEBUG=+seh wine wine/dlls/ntdll/tests/ntdll_test.exe exception

Working x86_64 system running 32-bit code

0024:warn:seh:dispatch_exception EXCEPTION_ACCESS_VIOLATION exception (code=c0000005) raised
0024:trace:seh:dispatch_exception eax=00000000 ebx=7ffc2000 ecx=004e0ef4 edx=003c0004 esi=003c0000 edi=00000000
0024:trace:seh:dispatch_exception ebp=0085fa08 esp=0085f9ac cs=0023 ds=002b es=002b fs=0063 gs=006b flags=00010246
0024:trace:seh:call_vectored_handlers calling handler at 7B00B460 code=c0000005 flags=0
0024:trace:seh:call_vectored_handlers handler at 7B00B460 returned 0
0024:trace:seh:call_stack_handlers calling handler at 004178B0 code=c0000005 flags=0
0024:trace:seh:call_stack_handlers handler at 004178B0 returned 0
0024:trace:seh:dispatch_exception call_stack_handlers continuing
0024:trace:seh:NtGetContextThread 0xfffffffe: dr0=42424240 dr1=00000000 dr2=126bb070 dr3=0badbad0 dr6=00000000 dr7=ffff0115

Non-working qemu

0024:warn:seh:dispatch_exception EXCEPTION_ACCESS_VIOLATION exception (code=c0000005) raised
0024:trace:seh:dispatch_exception eax=00000000 ebx=3ffe2000 ecx=004e0ef4 edx=003c0004 esi=003c0000 edi=00000000
0024:trace:seh:dispatch_exception ebp=0085fa08 esp=0085f9ac cs=0023 ds=002b es=002b fs=003b gs=0033 flags=00000246
0024:trace:seh:call_vectored_handlers calling handler at 7B00B460 code=c0000005 flags=0
0024:trace:seh:call_vectored_handlers handler at 7B00B460 returned 0
0024:trace:seh:call_stack_handlers calling handler at 004178B0 code=c0000005 flags=0
0024:trace:seh:call_stack_handlers handler at 004178B0 returned 0
0024:trace:seh:dispatch_exception call_stack_handlers continuing
0024:trace:seh:dispatch_exception call_stack_handlers ret status = 0
0024:trace:seh:dispatch_exception code=0 flags=1 addr=7BC2389C ip=7bc2389c tid=0024

The non-working verion is never managing to set the CPU context using NtContinue/SetContextThread back to the correct running thread stack and IP. It executes as if the context restore just returns to the function that called NtContinue() (dispatch_exception(), not the function that raised the exception or one of its parent exception handlers).

It looks like NtSetContextThread(), specifically the asm function set_full_cpu_context() is being handled incorrectly.

wine code below. note interesting use of iret with no previous interrupt call. The exception handler is called with a jmp.

/***********************************************************************
 * set_full_cpu_context
 *
 * Set the new CPU context.
 */
extern void set_full_cpu_context( const CONTEXT *context );
__ASM_GLOBAL_FUNC( set_full_cpu_context,
                   "movl $0,%fs:0x1f8\n\t" /* x86_thread_data()->syscall_frame = NULL */
                   "movl 4(%esp),%ecx\n\t"
                   "movw 0x8c(%ecx),%gs\n\t" /* SegGs */
                   "movw 0x90(%ecx),%fs\n\t" /* SegFs */
                   "movw 0x94(%ecx),%es\n\t" /* SegEs */
                   "movl 0x9c(%ecx),%edi\n\t" /* Edi */
                   "movl 0xa0(%ecx),%esi\n\t" /* Esi */
                   "movl 0xa4(%ecx),%ebx\n\t" /* Ebx */
                   "movl 0xb4(%ecx),%ebp\n\t" /* Ebp */
                   "movw %ss,%ax\n\t"
                   "cmpw 0xc8(%ecx),%ax\n\t" /* SegSs */
                   "jne 1f\n\t"
                   /* As soon as we have switched stacks the context structure could
                    * be invalid (when signal handlers are executed for example). Copy
                    * values on the target stack before changing ESP. */
                   "movl 0xc4(%ecx),%eax\n\t" /* Esp */
                   "leal -4*4(%eax),%eax\n\t"
                   "movl 0xc0(%ecx),%edx\n\t" /* EFlags */
                   ".byte 0x36\n\t"
                   "movl %edx,3*4(%eax)\n\t"
                   "movl 0xbc(%ecx),%edx\n\t" /* SegCs */
                   ".byte 0x36\n\t"
                   "movl %edx,2*4(%eax)\n\t"
                   "movl 0xb8(%ecx),%edx\n\t" /* Eip */
                   ".byte 0x36\n\t"
                   "movl %edx,1*4(%eax)\n\t"
                   "movl 0xb0(%ecx),%edx\n\t" /* Eax */
                   ".byte 0x36\n\t"
                   "movl %edx,0*4(%eax)\n\t"
                   "pushl 0x98(%ecx)\n\t" /* SegDs */
                   "movl 0xa8(%ecx),%edx\n\t" /* Edx */
                   "movl 0xac(%ecx),%ecx\n\t" /* Ecx */
                   "popl %ds\n\t"
                   "movl %eax,%esp\n\t"
                   "popl %eax\n\t"
                   "iret\n"
                   /* Restore the context when the stack segment changes. We can't use
                    * the same code as above because we do not know if the stack segment
                    * is 16 or 32 bit, and 'movl' will throw an exception when we try to
                    * access memory above the limit. */
                   "1:\n\t"
                   "movl 0xa8(%ecx),%edx\n\t" /* Edx */
                   "movl 0xb0(%ecx),%eax\n\t" /* Eax */
                   "movw 0xc8(%ecx),%ss\n\t" /* SegSs */
                   "movl 0xc4(%ecx),%esp\n\t" /* Esp */
                   "pushl 0xc0(%ecx)\n\t" /* EFlags */
                   "pushl 0xbc(%ecx)\n\t" /* SegCs */
                   "pushl 0xb8(%ecx)\n\t" /* Eip */
                   "pushl 0x98(%ecx)\n\t" /* SegDs */
                   "movl 0xac(%ecx),%ecx\n\t" /* Ecx */
                   "popl %ds\n\t"
                   "iret" )

Tags: i386
Revision history for this message
Dirk A Niggemann (dniggema) wrote :

Be aware that most of the regression test failures are caused by lack of ptrace() support.

The wine traces above show one of these cases. I will provide traces of an actual relevant failure.

However the failure to restart the correct thread is in some way related to the use of iret in set_full_cpu_context().

Revision history for this message
Dirk A Niggemann (dniggema) wrote :

That was a complete misdiaagnsis. The IRET works fine in user space, at the lowest privlege level.

The wrong thread is not restarted. The correct thread continues execution, at the correct address, but with the wrong thread-local data.

Thread-local data is accessed via the segment loaded in the fs register in wine. The segment selector is set the same in each thread, but each thread should have a unique GDT entry pointing to its thread-local data.

The issue is that clone() copies the pointer reference to the process's GDT without creating a new GDT area. All threads wind up sharing the same GDT. As the selector is the same, all threads share the same threa-local data. Oddly enough, this has fewer ill effects than excpected. Unless SEH (Structured Exception Handling) is invoked, at which point threads start unwinding the stack frames of other threads...

Copying the GDT during clone() results in a workig system with correct exception handling.

Revision history for this message
Thomas Huth (th-huth) wrote :

The QEMU project is currently moving its bug tracking to another system.
For this we need to know which bugs are still valid and which could be
closed already. Thus we are setting the bug state to "Incomplete" now.

If the bug has already been fixed in the latest upstream version of QEMU,
then please close this ticket as "Fix released".

If it is not fixed yet and you think that this bug report here is still
valid, then you have two options:

1) If you already have an account on gitlab.com, please open a new ticket
for this problem in our new tracker here:

    https://gitlab.com/qemu-project/qemu/-/issues

and then close this ticket here on Launchpad (or let it expire auto-
matically after 60 days). Please mention the URL of this bug ticket on
Launchpad in the new ticket on GitLab.

2) If you don't have an account on gitlab.com and don't intend to get
one, but still would like to keep this ticket opened, then please switch
the state back to "New" or "Confirmed" within the next 60 days (other-
wise it will get closed as "Expired"). We will then eventually migrate
the ticket automatically to the new system (but you won't be the reporter
of the bug in the new system and thus you won't get notified on changes
anymore).

Thank you and sorry for the inconvenience.

Changed in qemu:
status: New → Incomplete
tags: added: i386
Revision history for this message
Launchpad Janitor (janitor) wrote :

[Expired for QEMU because there has been no activity for 60 days.]

Changed in qemu:
status: Incomplete → Expired
To post a comment you must log in.
This report contains Public information  
Everyone can see this information.

Other bug subscribers