[gdb/tdep] Handle mxcsr kernel bug on Intel Skylake CPUs

On my openSUSE Leap 15.1 x86_64 Skylake system with the default (4.12) kernel,
I run into:
...
FAIL: gdb.base/gcore.exp: corefile restored all registers
...

The problem is that there's a difference in the mxcsr register value before
and after the gcore command:
...
- mxcsr          0x0                 [ ]
+ mxcsr          0x400440            [ DAZ OM ]
...

This can be traced back to amd64_linux_nat_target::fetch_registers, where
xstateregs is partially initialized by the ptrace call:
...
          char xstateregs[X86_XSTATE_MAX_SIZE];
          struct iovec iov;

          amd64_collect_xsave (regcache, -1, xstateregs, 0);
          iov.iov_base = xstateregs;
          iov.iov_len = sizeof (xstateregs);
          if (ptrace (PTRACE_GETREGSET, tid,
                      (unsigned int) NT_X86_XSTATE, (long) &iov) < 0)
            perror_with_name (_("Couldn't get extended state status"));

          amd64_supply_xsave (regcache, -1, xstateregs);
...
after which amd64_supply_xsave is called.

The amd64_supply_xsave call is supposed to only use initialized parts of
xstateregs, but due to a kernel bug on intel skylake (fixed from 4.14 onwards
by commit 0852b374173b "x86/fpu: Add FPU state copying quirk to handle XRSTOR
failure on Intel Skylake CPUs") it can happen that the mxcsr part of
xstateregs is not initialized, while amd64_supply_xsave expects it to be
initialized, which explains the FAIL mentioned above.

Fix the undetermined behaviour by initializing xstateregs before calling
ptrace, which makes sure we get a 0x0 for mxcsr when this kernel bug occurs,
and which also happens to fix the FAIL.

Furthermore, add an xfail for this FAIL which triggers the same kernel bug:
...
FAIL: gdb.arch/amd64-init-x87-values.exp: check_setting_mxcsr_before_enable: \
  check new value of MXCSR is still in place
...

Both FAILs pass when using a 5.3 kernel instead on the system mentioned above.

Tested on x86_64-linux.

gdb/ChangeLog:

2019-09-24  Tom de Vries  <tdevries@suse.de>

	PR gdb/23815
	* amd64-linux-nat.c (amd64_linux_nat_target::fetch_registers):
	Initialize xstateregs before ptrace PTRACE_GETREGSET call.

gdb/testsuite/ChangeLog:

2019-09-24  Tom de Vries  <tdevries@suse.de>

	PR gdb/24598
	* gdb.arch/amd64-init-x87-values.exp: Add xfail.

diff --git a/gdb/ChangeLog b/gdb/ChangeLog
index 77aab7649247b646e8a714b10e096e99865776be..ee53e9c00a56ca30ad539cf518661760bf26d746 100644 (file)
--- a/gdb/ChangeLog
+++ b/gdb/ChangeLog
@@ -1,3 +1,9 @@
+2019-09-24  Tom de Vries  <tdevries@suse.de>
+
+       PR gdb/23815
+       * amd64-linux-nat.c (amd64_linux_nat_target::fetch_registers):
+       Initialize xstateregs before ptrace PTRACE_GETREGSET call.
+
 2019-09-23  Dimitar Dimitrov  <dimitar@dinux.eu>
 
        * NEWS: Mention new simulator port for PRU.

diff --git a/gdb/amd64-linux-nat.c b/gdb/amd64-linux-nat.c
index 4f1c98a0d1e17125b02d64453230eb6e06989aa5..d0328b677d598456909c817c6e9969dc9961645f 100644 (file)
--- a/gdb/amd64-linux-nat.c
+++ b/gdb/amd64-linux-nat.c
@@ -238,6 +238,12 @@ amd64_linux_nat_target::fetch_registers (struct regcache *regcache, int regnum)
          char xstateregs[X86_XSTATE_MAX_SIZE];
          struct iovec iov;
 
+         /* Pre-4.14 kernels have a bug (fixed by commit 0852b374173b
+            "x86/fpu: Add FPU state copying quirk to handle XRSTOR failure on
+            Intel Skylake CPUs") that sometimes causes the mxcsr location in
+            xstateregs not to be copied by PTRACE_GETREGSET.  Make sure that
+            the location is at least initialized with a defined value.  */
+         memset (xstateregs, 0, sizeof (xstateregs));
          iov.iov_base = xstateregs;
          iov.iov_len = sizeof (xstateregs);
          if (ptrace (PTRACE_GETREGSET, tid,

diff --git a/gdb/testsuite/ChangeLog b/gdb/testsuite/ChangeLog
index 37e323f747c69804939755d5c171283847953afc..706c5da420afa8ac5acb94d262f299798431334e 100644 (file)
--- a/gdb/testsuite/ChangeLog
+++ b/gdb/testsuite/ChangeLog
@@ -1,3 +1,8 @@
+2019-09-24  Tom de Vries  <tdevries@suse.de>
+
+       PR gdb/24598
+       * gdb.arch/amd64-init-x87-values.exp: Add xfail.
+
 2019-09-22  Tom de Vries  <tdevries@suse.de>
 
        * gdb.base/restore.exp: Allow register variables to be optimized out at

diff --git a/gdb/testsuite/gdb.arch/amd64-init-x87-values.exp b/gdb/testsuite/gdb.arch/amd64-init-x87-values.exp
index cdf92dcd37761c7d6f5ee7306d8419f8893f5105..5fd18dbb7942e8fb6d936294be7b1cd9e355f8f8 100644 (file)
--- a/gdb/testsuite/gdb.arch/amd64-init-x87-values.exp
+++ b/gdb/testsuite/gdb.arch/amd64-init-x87-values.exp
@@ -116,7 +116,7 @@ proc_with_prefix check_x87_regs_around_init {} {
 # nop that does not enable any FP features).  Finally check that the
 # mxcsr register still has the value we set.
 proc_with_prefix check_setting_mxcsr_before_enable {} {
-    global binfile
+    global binfile gdb_prompt
 
     clean_restart ${binfile}
 
@@ -127,7 +127,22 @@ proc_with_prefix check_setting_mxcsr_before_enable {} {
 
     gdb_test_no_output "set \$mxcsr=0x9f80" "set a new value for MXCSR"
     gdb_test "stepi" "fwait" "step forward one instruction for mxcsr test"
-    gdb_test "p/x \$mxcsr" " = 0x9f80" "check new value of MXCSR is still in place"
+
+    set test "check new value of MXCSR is still in place"
+    set pass_pattern " = 0x9f80"
+    # Pre-4.14 kernels have a bug (fixed by commit 0852b374173b "x86/fpu:
+    # Add FPU state copying quirk to handle XRSTOR failure on Intel Skylake
+    # CPUs") that causes mxcsr not to be copied, in which case we get 0 instead of
+    # the just saved value.
+    set xfail_pattern " = 0x0"
+    gdb_test_multiple "p/x \$mxcsr" $test {
+       -re "\[\r\n\]*(?:$pass_pattern)\[\r\n\]+$gdb_prompt $" {
+           pass $test
+       }
+       -re "\[\r\n\]*(?:$xfail_pattern)\[\r\n\]+$gdb_prompt $" {
+           xfail $test
+       }
+    }
 }
 
 # Start the test file, all FP features will be disabled.  Set new