Bug 1347993

+++ This bug was initially created as a clone of Bug #1261564 +++

Description of problem:
ISTM that sometimes hardware watchpoints are not set correctly on armv8.


#1 - kernel Bug 1342821: aarch64: ptrace: BAS: Support any contiguous range
GDB can return false positive hardware watchpoints (rwatch+awatch)

TCMS 118567: /tools/gdb/Sanity/basic-smoke-test
  https://tcms.engineering.redhat.com/case/118567/

--- Additional comment from Martin Cermak on 2016-05-16 16:56:22 CEST ---

So I've been testing with gdb-7.6.1-91.el7.aarch64.

However, there is yet another QA testcase linked to this bug, which, I believe, originally spotted this bug: http://pkgs.devel.redhat.com/cgit/tests/gdb/tree/Sanity/basic-smoke-test.  This testcase still fails.  The scenario using the aforementioned files is as follows:

=======
gcc -o testcase testcase.c -g
export ARCH=${ARCH:-"`uname -n`"}
./use-case.exp
=======

--- Additional comment from Jan Kratochvil on 2016-05-25 00:11:21 CEST ---

FSF GDB trunk also does FAIL: GNU gdb (GDB) 7.11.50.20160524-git

The testcase
  http://pkgs.devel.redhat.com/cgit/tests/gdb/plain/Sanity/basic-smoke-test/use-case.exp
  http://pkgs.devel.redhat.com/cgit/tests/gdb/plain/Sanity/basic-smoke-test/testcase.c
has never been tested/ported to aarch64.  It needs:
  perl -pi-orig -e 's/&& \$arch!="ppc64le"/$& && \$arch!="aarch64"/' use-case.exp 

Although it correctly found out there is a hardware watchpoint bug on aarch64:
  Aarch64: watchpoints set on non-8-byte-aligned addresses are always missed
  https://sourceware.org/bugzilla/show_bug.cgi?id=19806

Without the 'perl' substitution above:
-=x86_64
+=aarch64

One can see it on:
 Hardware read watchpoint 6: r^M
 ^M
 Value = 8^M
-0x00000000004006bf in main () at testcase.c:60^M
-60       d=r;^M
+main () at testcase.c:59^M
+59       d=p;^M

x86_64 is right, aarch64 is wrong, there is no 'r' variable read to stop at line 59 (on line 58 variable 'r' is written to, not read from).

This can be reproduced in a simplified test reusing 'testcase.c':

cat >cmd2 <<EOH
set verbose
set pagination off
b 56
run
p &p
p &r
watch p
rwatch r
continue
#  delete 2
continue
EOH
gcc -o testcase testcase.c -g -Wall
gdb -x cmd2 ./testcase

$1 = (int *) 0x42002c <p>
$2 = (int *) 0x420028 <r>
Hardware watchpoint 2: p
Hardware read watchpoint 3: r
Hardware watchpoint 2: p
Old value = 1
New value = 2
main () at testcase.c:58
58        r=d;
Hardware read watchpoint 3: r
Value = 8
main () at testcase.c:59
59        d=p;
(gdb)

testcase.c:
      int p,r,d,e,l;
->
      long p,r,d,e,l;

$1 = (long *) 0x420040 <p>
$2 = (long *) 0x420038 <r>
Hardware watchpoint 2: p
Hardware read watchpoint 3: r
Hardware watchpoint 2: p
Old value = 1
New value = 2
main () at testcase.c:58
58        r=d;
Hardware read watchpoint 3: r
Value = 8
0x00000000004007cc in main () at testcase.c:60
60        d=r;
(gdb)

By changing int->long we workaround the GDB bug of a shared 8-byte aligned memory location by two variables 'p' and 'r' and so the false stopping at line 59 no longer occurs.

One can also workaround the bug affecting the 'r' watchpoint by deleting the 'p' watchpoint at that cmd2 line:
  #  delete 2


#2 - Upstream GDB Bug: 

GDB can miss a watchpoint (watch):
https://sourceware.org/ml/gdb-patches/2016-06/msg00315.html


Version-Release number of selected component (if applicable):
gdb-7.6.1-78.el7.aarch64
gdb-7.6.1-91.el7.aarch64
kernel-4.2.0-0.rc6.19.el7.aarch64

How reproducible:
always