Blob Blame History Raw
http://sourceware.org/ml/gdb-patches/2009-11/msg00592.html
Subject: [patch] Fix syscall restarts for amd64->i386 biarch

Hi,

tested only on recent Linux kernels, it should apply also on vanilla ones.
There were various changes of the kernels behavior in the past.

FSF GDB HEAD state:
kernel debugger inferior state
x86_64 x86_64   x86_64   PASS
x86_64 x86_64   i386     FAIL without this patch, PASS with this patch
x86_64 i386     i386     PASS on recent kernels
                         (FAIL: kernel-2.6.31.5-127.fc12.x86_64 - Fedora 12)
                         (PASS: kernel-2.6.32-0.55.rc8.git1.fc13.x86_64)
i386   i386     i386     PASS


Currently gdb.base/interrupt.exp fails on amd64 host running under
--target_board unix/-m32 with:
	continue
	Continuing.
	Unknown error 512

<linux/errno.h>:
/*
 * These should never be seen by user programs.  To return one of ERESTART*
 * codes, signal_pending() MUST be set.  Note that ptrace can observe these
 * at syscall exit tracing, but they will never be left for the debugged user
 * process to see.
 */
#define ERESTARTSYS     512

"Unknown error 512" printed above is printed by the inferior itself, not by GDB.

It is because GDB reads it as 0xfffffffffffffe00 but writes it back as
0xfffffe00.
+      /* Sign-extend %eax as during return from a syscall it is being checked
+	 for -ERESTART* values -512 being above 0xfffffffffffffe00; tested by
+	 interrupt.exp.  */


Quote of Roland McGrath from IRC:

roland: in the user_regset model, there are 64-bit user_regset flavors and
32-bit user_regset flavors, so at the kabi level the (kernel) caller can say
what it means: calls on the 32-bit user_regset flavor will behave as if on
a 32-bit kernel/userland.  in ptrace, there is no way for x86_64 ptrace calls
to say "i think of the inferior as being 32 bits, so act accordingly" (tho ppc
and/or sparc have ptr
roland: ace requests that do that iirc)
roland: ergo 64-bit ptrace callers must either save/restore full 64-bits so
the kernel's sign-extension choices are preserved, or else grok magic ways to
expand stored 32-bit register contents to 64-bit values to stuff via 64-bit
ptrace
[...]
roland: there is a "32-bit-flavored task", but it's not really true that it
has 32-bit registers.  there is no 32-bit-only userland condition.  any task
can always ljmp to the 64-bit code segment and run 64-bit insns including
a 64-bit syscall
roland: so a 64-bit debugger should see and be able to fiddle the full
registers.  it can even change cs via ptrace to force the inferior into
running 32 or 64 bit code.


Saving whole 64bits for i386 targets on x86_64 hosts does not much match the
GDB architecture as `struct type' for these registers still should be 32bit
etc.   Therefore provided just this exception.

The problem is reproducible only if one does an inferior call during the
interruption to do full inferior save/restore from GDB regcache.

Regression tested on {x86_64,x86_64-m32,i686}-fedora12-linux-gnu.


Thanks,
Jan


gdb/
2009-11-29  Jan Kratochvil  <jan.kratochvil@redhat.com>

	* amd64-nat.c (amd64_collect_native_gregset): Do not pre-clear %eax.
	Sign extend it afterwards.

--- a/gdb/amd64-nat.c
+++ b/gdb/amd64-nat.c
@@ -131,9 +131,9 @@ amd64_collect_native_gregset (const struct regcache *regcache,
     {
       num_regs = amd64_native_gregset32_num_regs;
 
-      /* Make sure %eax, %ebx, %ecx, %edx, %esi, %edi, %ebp, %esp and
+      /* Make sure %ebx, %ecx, %edx, %esi, %edi, %ebp, %esp and
          %eip get zero-extended to 64 bits.  */
-      for (i = 0; i <= I386_EIP_REGNUM; i++)
+      for (i = I386_ECX_REGNUM; i <= I386_EIP_REGNUM; i++)
 	{
 	  if (regnum == -1 || regnum == i)
 	    memset (regs + amd64_native_gregset_reg_offset (gdbarch, i), 0, 8);
@@ -159,4 +159,20 @@ amd64_collect_native_gregset (const struct regcache *regcache,
 	    regcache_raw_collect (regcache, i, regs + offset);
 	}
     }
+
+  if (gdbarch_bfd_arch_info (gdbarch)->bits_per_word == 32)
+    {
+      /* Sign-extend %eax as during return from a syscall it is being checked
+	 for -ERESTART* values -512 being above 0xfffffffffffffe00; tested by
+	 interrupt.exp.  */
+
+      int i = I386_EAX_REGNUM;
+
+      if (regnum == -1 || regnum == i)
+	{
+	  void *ptr = regs + amd64_native_gregset_reg_offset (gdbarch, i);
+
+	  *(int64_t *) ptr = *(int32_t *) ptr;
+	}
+    }
 }