Blame SOURCES/gdb-rhbz1842691-corefile-mem-access-14of15.patch

0c1cd1
From FEDORA_PATCHES Mon Sep 17 00:00:00 2001
0c1cd1
From: Keith Seitz <keiths@redhat.com>
0c1cd1
Date: Tue, 28 Jul 2020 09:46:44 -0400
0c1cd1
Subject: gdb-rhbz1842691-corefile-mem-access-14of15.patch
0c1cd1
0c1cd1
;; New core file tests with mappings over existing program memory
0c1cd1
;; Kevin Buettner, RH BZ 1842961
0c1cd1
0c1cd1
   Author: Kevin Buettner <kevinb@redhat.com>
0c1cd1
   Date:   Wed Jun 17 19:25:47 2020 -0700
0c1cd1
0c1cd1
    New core file tests with mappings over existing program memory
0c1cd1
0c1cd1
    This test case was inspired by Pedro's demonstration of a problem
0c1cd1
    with my v2 patches.  It can be found here:
0c1cd1
0c1cd1
        https://sourceware.org/pipermail/gdb-patches/2020-May/168826.html
0c1cd1
0c1cd1
    In a nutshell, my earlier patches could not handle the case in
0c1cd1
    which a read-only mapping created with mmap() was created at
0c1cd1
    an address used by other file-backed read-only memory in use by
0c1cd1
    the process.
0c1cd1
0c1cd1
    This problem has been fixed (for Linux, anyway) by the commit "Use
0c1cd1
    NT_FILE note section for reading core target memory".
0c1cd1
0c1cd1
    When I run this test without any of my recent corefile patches,
0c1cd1
    I see these failures:
0c1cd1
0c1cd1
    FAIL: gdb.base/corefile2.exp: kernel core: print/x mbuf_ro[0]@4
0c1cd1
    FAIL: gdb.base/corefile2.exp: kernel core: print/x mbuf_ro[pagesize-4]@4
0c1cd1
    FAIL: gdb.base/corefile2.exp: kernel core: print/x mbuf_ro[-3]@6
0c1cd1
    FAIL: gdb.base/corefile2.exp: kernel core: print/x mbuf_rw[pagesize-3]@6
0c1cd1
    FAIL: gdb.base/corefile2.exp: kernel core: print/x mbuf_ro[pagesize-3]@6
0c1cd1
    FAIL: gdb.base/corefile2.exp: maint print core-file-backed-mappings
0c1cd1
    FAIL: gdb.base/corefile2.exp: gcore core: print/x mbuf_ro[-3]@6
0c1cd1
0c1cd1
    The ones involving mbuf_ro will almost certainly fail when run on
0c1cd1
    non-Linux systems; I've used setup_xfail on those tests to prevent
0c1cd1
    them from outright FAILing when not run on Linux.  For a time, I
0c1cd1
    had considered skipping these tests altogether when not run on
0c1cd1
    Linux, but I changed my mind due to this failure...
0c1cd1
0c1cd1
    FAIL: gdb.base/corefile2.exp: print/x mbuf_rw[pagesize-3]@6
0c1cd1
0c1cd1
    I think it *should* pass without my recent corefile patches.  The fact
0c1cd1
    that it doesn't is likely due to a bug in GDB.  The following
0c1cd1
    interaction with GDB demonstrates the problem:
0c1cd1
0c1cd1
    (gdb) print/x mbuf_rw[pagesize-3]@6
0c1cd1
    $1 = {0x0, 0x0, 0x0, 0x0, 0x0, 0x0}
0c1cd1
    (gdb) print/x mbuf_rw[pagesize]@3
0c1cd1
    $2 = {0x6b, 0x6b, 0x6b}
0c1cd1
0c1cd1
    The last three values in display of $1 should be the same as those
0c1cd1
    shown by $2.  Like this...
0c1cd1
0c1cd1
    (gdb) print/x mbuf_rw[pagesize-3]@6
0c1cd1
    $1 = {0x0, 0x0, 0x0, 0x6b, 0x6b, 0x6b}
0c1cd1
    (gdb) print/x mbuf_rw[pagesize]@3
0c1cd1
    $2 = {0x6b, 0x6b, 0x6b}
0c1cd1
0c1cd1
    That latter output was obtained with the use of all of my current
0c1cd1
    corefile patches.  I see no failures on Linux when running this test
0c1cd1
    with my current set of corefile patches.  I tested 3 architectures:
0c1cd1
    x86_64, s390x, and aarch64.
0c1cd1
0c1cd1
    I also tested on FreeBSD 12.1-RELEASE.  I see the following results
0c1cd1
    both with and without the current set of core file patches:
0c1cd1
0c1cd1
        # of expected passes		26
0c1cd1
        # of expected failures		8
0c1cd1
0c1cd1
    Of particular interest is that I did *not* see the problematic mbuf_rw
0c1cd1
    failure noted earlier (both with and without the core file patches).
0c1cd1
    I still don't have an explanation for why this failure occurred on
0c1cd1
    Linux.  Prior to running the tests, I had hypothesized that I'd see
0c1cd1
    this failure on FreeBSD too, but testing shows that this is not the
0c1cd1
    case.
0c1cd1
0c1cd1
    Also of importance is that we see no FAILs with this test on FreeBSD
0c1cd1
    which indicates that I XFAILed the correct tests.
0c1cd1
0c1cd1
    This version runs the interesting tests twice, once with a kernel
0c1cd1
    created core file and another time with a gcore created core file.
0c1cd1
0c1cd1
    It also does a very minimal test of the new command "maint print
0c1cd1
    core-file-backed-mappings".
0c1cd1
0c1cd1
    gdb/testsuite/ChangeLog:
0c1cd1
0c1cd1
    	* gdb.base/corefile2.exp: New file.
0c1cd1
    	* gdb.base/coremaker2.exp: New file.
0c1cd1
0c1cd1
diff --git a/gdb/testsuite/gdb.base/corefile2.exp b/gdb/testsuite/gdb.base/corefile2.exp
0c1cd1
new file mode 100644
0c1cd1
--- /dev/null
0c1cd1
+++ b/gdb/testsuite/gdb.base/corefile2.exp
0c1cd1
@@ -0,0 +1,185 @@
0c1cd1
+# Copyright 2020 Free Software Foundation, Inc.
0c1cd1
+
0c1cd1
+# This program is free software; you can redistribute it and/or modify
0c1cd1
+# it under the terms of the GNU General Public License as published by
0c1cd1
+# the Free Software Foundation; either version 3 of the License, or
0c1cd1
+# (at your option) any later version.
0c1cd1
+#
0c1cd1
+# This program is distributed in the hope that it will be useful,
0c1cd1
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
0c1cd1
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
0c1cd1
+# GNU General Public License for more details.
0c1cd1
+#
0c1cd1
+# You should have received a copy of the GNU General Public License
0c1cd1
+# along with this program.  If not, see <http://www.gnu.org/licenses/>.
0c1cd1
+
0c1cd1
+# Tests of core file memory accesses when mmap() has been used to
0c1cd1
+# create a "hole" of zeroes over pre-existing memory regions.  See
0c1cd1
+# coremaker2.c for details.
0c1cd1
+
0c1cd1
+# are we on a target board
0c1cd1
+if ![isnative] then {
0c1cd1
+    return
0c1cd1
+}
0c1cd1
+
0c1cd1
+# Some of these tests will only work on GNU/Linux due to the
0c1cd1
+# fact that Linux core files includes a section describing
0c1cd1
+# memory address to file mappings.  We'll use set_up_xfail for the
0c1cd1
+# affected tests.  As other targets become supported, the condition
0c1cd1
+# can be changed accordingly.
0c1cd1
+
0c1cd1
+set xfail 0
0c1cd1
+if { ![istarget *-linux*] } {
0c1cd1
+    set xfail 1
0c1cd1
+}
0c1cd1
+
0c1cd1
+standard_testfile coremaker2.c
0c1cd1
+
0c1cd1
+if {[build_executable $testfile.exp $testfile $srcfile debug] == -1} {
0c1cd1
+    untested "failed to compile"
0c1cd1
+    return -1
0c1cd1
+}
0c1cd1
+
0c1cd1
+set corefile [core_find $binfile {}]
0c1cd1
+if {$corefile == ""} {
0c1cd1
+    return 0
0c1cd1
+}
0c1cd1
+
0c1cd1
+gdb_start
0c1cd1
+gdb_reinitialize_dir $srcdir/$subdir
0c1cd1
+gdb_load ${binfile}
0c1cd1
+
0c1cd1
+# Attempt to load the core file.
0c1cd1
+
0c1cd1
+gdb_test_multiple "core-file $corefile" "core-file command" {
0c1cd1
+    -re ".* program is being debugged already.*y or n. $" {
0c1cd1
+	# gdb_load may connect us to a gdbserver.
0c1cd1
+	send_gdb "y\n"
0c1cd1
+	exp_continue
0c1cd1
+    }
0c1cd1
+    -re "Core was generated by .*corefile.*\r\n\#0  .*\(\).*\r\n$gdb_prompt $" {
0c1cd1
+	pass "core-file command"
0c1cd1
+    }
0c1cd1
+    -re "Core was generated by .*\r\n\#0  .*\(\).*\r\n$gdb_prompt $" {
0c1cd1
+	pass "core-file command (with bad program name)"
0c1cd1
+    }
0c1cd1
+    -re ".*registers from core file: File in wrong format.* $" {
0c1cd1
+	fail "core-file command (could not read registers from core file)"
0c1cd1
+    }
0c1cd1
+}
0c1cd1
+
0c1cd1
+# Perform the "interesting" tests which check the contents of certain
0c1cd1
+# memory regions.
0c1cd1
+
0c1cd1
+proc do_tests { } {
0c1cd1
+    global xfail
0c1cd1
+
0c1cd1
+    # Check contents of beginning of buf_rw and buf_ro.
0c1cd1
+
0c1cd1
+    gdb_test {print/x buf_rw[0]@4} {\{0x6b, 0x6b, 0x6b, 0x6b\}}
0c1cd1
+    gdb_test {print/x buf_ro[0]@4} {\{0xc5, 0xc5, 0xc5, 0xc5\}}
0c1cd1
+
0c1cd1
+    # Check for correct contents at beginning of mbuf_rw and mbuf_ro.
0c1cd1
+
0c1cd1
+    gdb_test {print/x mbuf_rw[0]@4} {\{0x0, 0x0, 0x0, 0x0\}}
0c1cd1
+
0c1cd1
+    if { $xfail } { setup_xfail "*-*-*" }
0c1cd1
+    gdb_test {print/x mbuf_ro[0]@4} {\{0x0, 0x0, 0x0, 0x0\}}
0c1cd1
+
0c1cd1
+    # Check contents of mbuf_rw and mbuf_ro at the end of these regions.
0c1cd1
+
0c1cd1
+    gdb_test {print/x mbuf_rw[pagesize-4]@4} {\{0x0, 0x0, 0x0, 0x0\}}
0c1cd1
+
0c1cd1
+    if { $xfail } { setup_xfail "*-*-*" }
0c1cd1
+    gdb_test {print/x mbuf_ro[pagesize-4]@4} {\{0x0, 0x0, 0x0, 0x0\}}
0c1cd1
+
0c1cd1
+    # Check contents of mbuf_rw and mbuf_ro, right before the hole,
0c1cd1
+    # overlapping into the beginning of these mmap'd regions.
0c1cd1
+
0c1cd1
+    gdb_test {print/x mbuf_rw[-3]@6} {\{0x6b, 0x6b, 0x6b, 0x0, 0x0, 0x0\}}
0c1cd1
+
0c1cd1
+    if { $xfail } { setup_xfail "*-*-*" }
0c1cd1
+    gdb_test {print/x mbuf_ro[-3]@6} {\{0xc5, 0xc5, 0xc5, 0x0, 0x0, 0x0\}}
0c1cd1
+
0c1cd1
+    # Likewise, at the end of the mbuf_rw and mbuf_ro, with overlap.
0c1cd1
+
0c1cd1
+    # If this test FAILs, it's probably a genuine bug unrelated to whether
0c1cd1
+    # the core file includes a section describing memory address to file
0c1cd1
+    # mappings or not.  (So don't xfail it!)
0c1cd1
+    gdb_test {print/x mbuf_rw[pagesize-3]@6} {\{0x0, 0x0, 0x0, 0x6b, 0x6b, 0x6b\}}
0c1cd1
+
0c1cd1
+    if { $xfail } { setup_xfail "*-*-*" }
0c1cd1
+    gdb_test {print/x mbuf_ro[pagesize-3]@6} {\{0x0, 0x0, 0x0, 0xc5, 0xc5, 0xc5\}}
0c1cd1
+
0c1cd1
+    # Check contents of (what should be) buf_rw and buf_ro immediately after
0c1cd1
+    # mbuf_rw and mbuf_ro holes.
0c1cd1
+
0c1cd1
+    gdb_test {print/x mbuf_rw[pagesize]@4} {\{0x6b, 0x6b, 0x6b, 0x6b\}}
0c1cd1
+    gdb_test {print/x mbuf_ro[pagesize]@4} {\{0xc5, 0xc5, 0xc5, 0xc5\}}
0c1cd1
+
0c1cd1
+    # Check contents at ends of buf_rw and buf_rw.
0c1cd1
+
0c1cd1
+    gdb_test {print/x buf_rw[sizeof(buf_rw)-4]@4} {\{0x6b, 0x6b, 0x6b, 0x6b\}}
0c1cd1
+    gdb_test {print/x buf_ro[sizeof(buf_ro)-4]@4} {\{0xc5, 0xc5, 0xc5, 0xc5\}}
0c1cd1
+}
0c1cd1
+
0c1cd1
+# Run tests with kernel-produced core file.
0c1cd1
+
0c1cd1
+with_test_prefix "kernel core" {
0c1cd1
+    do_tests
0c1cd1
+}
0c1cd1
+
0c1cd1
+# Verify that "maint print core-file-backed-mappings" exists and does
0c1cd1
+# not crash GDB.  If it produces any output at all, make sure that
0c1cd1
+# that output at least mentions binfile.
0c1cd1
+
0c1cd1
+set test "maint print core-file-backed-mappings"
0c1cd1
+gdb_test_multiple $test "" {
0c1cd1
+    -re ".*$binfile.*$gdb_prompt $" {
0c1cd1
+	pass $test
0c1cd1
+    }
0c1cd1
+    -re "^$test\[\r\n\]*$gdb_prompt $" {
0c1cd1
+	pass "$test (no output)"
0c1cd1
+    }
0c1cd1
+}
0c1cd1
+
0c1cd1
+# Restart and run to the abort call.
0c1cd1
+
0c1cd1
+clean_restart $binfile
0c1cd1
+
0c1cd1
+if ![runto_main] then {
0c1cd1
+    fail "can't run to main"
0c1cd1
+    return
0c1cd1
+}
0c1cd1
+
0c1cd1
+gdb_breakpoint [gdb_get_line_number "abort"]
0c1cd1
+gdb_continue_to_breakpoint "at abort"
0c1cd1
+
0c1cd1
+# Do not execute abort call; instead, invoke gcore command to make a
0c1cd1
+# gdb-produced core file.
0c1cd1
+
0c1cd1
+set corefile [standard_output_file gcore.test]
0c1cd1
+set core_supported [gdb_gcore_cmd "$corefile" "save a corefile"]
0c1cd1
+if {!$core_supported} {
0c1cd1
+  return
0c1cd1
+}
0c1cd1
+
0c1cd1
+# maint print-core-file-backed-mappings shouldn't produce any output
0c1cd1
+# when not debugging a core file.
0c1cd1
+
0c1cd1
+gdb_test_no_output "maint print core-file-backed-mappings" \
0c1cd1
+    "maint print core-file-backed-mapping with no core file"
0c1cd1
+
0c1cd1
+clean_restart $binfile
0c1cd1
+
0c1cd1
+set core_loaded [gdb_core_cmd "$corefile" "re-load generated corefile"]
0c1cd1
+if { $core_loaded == -1 } {
0c1cd1
+    # No use proceeding from here.
0c1cd1
+    return
0c1cd1
+}
0c1cd1
+
0c1cd1
+# Run tests using gcore-produced core file.
0c1cd1
+
0c1cd1
+with_test_prefix "gcore core" {
0c1cd1
+    do_tests
0c1cd1
+}
0c1cd1
diff --git a/gdb/testsuite/gdb.base/coremaker2.c b/gdb/testsuite/gdb.base/coremaker2.c
0c1cd1
new file mode 100644
0c1cd1
--- /dev/null
0c1cd1
+++ b/gdb/testsuite/gdb.base/coremaker2.c
0c1cd1
@@ -0,0 +1,150 @@
0c1cd1
+/* Copyright 1992-2020 Free Software Foundation, Inc.
0c1cd1
+
0c1cd1
+   This file is part of GDB.
0c1cd1
+
0c1cd1
+   This program is free software; you can redistribute it and/or modify
0c1cd1
+   it under the terms of the GNU General Public License as published by
0c1cd1
+   the Free Software Foundation; either version 3 of the License, or
0c1cd1
+   (at your option) any later version.
0c1cd1
+
0c1cd1
+   This program is distributed in the hope that it will be useful,
0c1cd1
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
0c1cd1
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
0c1cd1
+   GNU General Public License for more details.
0c1cd1
+
0c1cd1
+   You should have received a copy of the GNU General Public License
0c1cd1
+   along with this program.  If not, see <http://www.gnu.org/licenses/>.  */
0c1cd1
+
0c1cd1
+/*  This test has two large memory areas buf_rw and buf_ro. 
0c1cd1
+
0c1cd1
+    buf_rw is written to by the program while buf_ro is initialized at
0c1cd1
+    compile / load time.  Thus, when a core file is created, buf_rw's
0c1cd1
+    memory should reside in the core file, but buf_ro probably won't be.
0c1cd1
+    Instead, the contents of buf_ro are available from the executable.
0c1cd1
+
0c1cd1
+    Now, for the wrinkle:  We create a one page read-only mapping over
0c1cd1
+    both of these areas.  This will create a one page "hole" of all
0c1cd1
+    zeros in each area.
0c1cd1
+
0c1cd1
+    Will GDB be able to correctly read memory from each of the four
0c1cd1
+    (or six, if you count the regions on the other side of each hole)
0c1cd1
+    memory regions?  */
0c1cd1
+
0c1cd1
+#include <stdio.h>
0c1cd1
+#include <sys/types.h>
0c1cd1
+#include <fcntl.h>
0c1cd1
+#include <sys/mman.h>
0c1cd1
+#include <signal.h>
0c1cd1
+#include <stdlib.h>
0c1cd1
+#include <unistd.h>
0c1cd1
+#include <string.h>
0c1cd1
+#include <errno.h>
0c1cd1
+
0c1cd1
+/* These are globals so that we can find them easily when debugging
0c1cd1
+   the core file.  */
0c1cd1
+long pagesize;
0c1cd1
+unsigned long long addr;
0c1cd1
+char *mbuf_ro;
0c1cd1
+char *mbuf_rw;
0c1cd1
+
0c1cd1
+/* 24 KiB buffer.  */
0c1cd1
+char buf_rw[24 * 1024];
0c1cd1
+
0c1cd1
+/* 24 KiB worth of data.  For this test case, we can't allocate a
0c1cd1
+   buffer and then fill it; we want GDB to have to read this data
0c1cd1
+   from the executable; it should NOT find it in the core file.  */
0c1cd1
+
0c1cd1
+#define C5_16 \
0c1cd1
+  0xc5, 0xc5, 0xc5, 0xc5, \
0c1cd1
+  0xc5, 0xc5, 0xc5, 0xc5, \
0c1cd1
+  0xc5, 0xc5, 0xc5, 0xc5, \
0c1cd1
+  0xc5, 0xc5, 0xc5, 0xc5
0c1cd1
+
0c1cd1
+#define C5_256 \
0c1cd1
+  C5_16, C5_16, C5_16, C5_16, \
0c1cd1
+  C5_16, C5_16, C5_16, C5_16, \
0c1cd1
+  C5_16, C5_16, C5_16, C5_16, \
0c1cd1
+  C5_16, C5_16, C5_16, C5_16
0c1cd1
+
0c1cd1
+#define C5_1k \
0c1cd1
+  C5_256, C5_256, C5_256, C5_256
0c1cd1
+
0c1cd1
+#define C5_24k \
0c1cd1
+  C5_1k, C5_1k, C5_1k, C5_1k, \
0c1cd1
+  C5_1k, C5_1k, C5_1k, C5_1k, \
0c1cd1
+  C5_1k, C5_1k, C5_1k, C5_1k, \
0c1cd1
+  C5_1k, C5_1k, C5_1k, C5_1k, \
0c1cd1
+  C5_1k, C5_1k, C5_1k, C5_1k, \
0c1cd1
+  C5_1k, C5_1k, C5_1k, C5_1k
0c1cd1
+
0c1cd1
+const char buf_ro[] = { C5_24k };
0c1cd1
+
0c1cd1
+int
0c1cd1
+main (int argc, char **argv)
0c1cd1
+{
0c1cd1
+  int i, bitcount;
0c1cd1
+
0c1cd1
+#ifdef _SC_PAGESIZE
0c1cd1
+  pagesize = sysconf (_SC_PAGESIZE);
0c1cd1
+#else
0c1cd1
+  pagesize = 8192;
0c1cd1
+#endif
0c1cd1
+
0c1cd1
+  /* Verify that pagesize is a power of 2.  */
0c1cd1
+  bitcount = 0;
0c1cd1
+  for (i = 0; i < 4 * sizeof (pagesize); i++)
0c1cd1
+    if (pagesize & (1 << i))
0c1cd1
+      bitcount++;
0c1cd1
+
0c1cd1
+  if (bitcount != 1)
0c1cd1
+    {
0c1cd1
+      fprintf (stderr, "pagesize is not a power of 2.\n");
0c1cd1
+      exit (1);
0c1cd1
+    }
0c1cd1
+
0c1cd1
+  /* Compute an address that should be within buf_ro.  Complain if not.  */
0c1cd1
+  addr = ((unsigned long long) buf_ro + pagesize) & ~(pagesize - 1);
0c1cd1
+
0c1cd1
+  if (addr <= (unsigned long long) buf_ro
0c1cd1
+      || addr >= (unsigned long long) buf_ro + sizeof (buf_ro))
0c1cd1
+    {
0c1cd1
+      fprintf (stderr, "Unable to compute a suitable address within buf_ro.\n");
0c1cd1
+      exit (1);
0c1cd1
+    }
0c1cd1
+
0c1cd1
+  mbuf_ro = mmap ((void *) addr, pagesize, PROT_READ,
0c1cd1
+               MAP_ANONYMOUS | MAP_PRIVATE | MAP_FIXED, -1, 0);
0c1cd1
+
0c1cd1
+  if (mbuf_ro == MAP_FAILED)
0c1cd1
+    {
0c1cd1
+      fprintf (stderr, "mmap #1 failed: %s.\n", strerror (errno));
0c1cd1
+      exit (1);
0c1cd1
+    }
0c1cd1
+
0c1cd1
+  /* Write (and fill) the R/W region.  */
0c1cd1
+  for (i = 0; i < sizeof (buf_rw); i++)
0c1cd1
+    buf_rw[i] = 0x6b;
0c1cd1
+
0c1cd1
+  /* Compute an mmap address within buf_rw.  Complain if it's somewhere
0c1cd1
+     else.  */
0c1cd1
+  addr = ((unsigned long long) buf_rw + pagesize) & ~(pagesize - 1);
0c1cd1
+
0c1cd1
+  if (addr <= (unsigned long long) buf_rw
0c1cd1
+      || addr >= (unsigned long long) buf_rw + sizeof (buf_rw))
0c1cd1
+    {
0c1cd1
+      fprintf (stderr, "Unable to compute a suitable address within buf_rw.\n");
0c1cd1
+      exit (1);
0c1cd1
+    }
0c1cd1
+
0c1cd1
+  mbuf_rw = mmap ((void *) addr, pagesize, PROT_READ,
0c1cd1
+               MAP_ANONYMOUS | MAP_PRIVATE | MAP_FIXED, -1, 0);
0c1cd1
+
0c1cd1
+  if (mbuf_rw == MAP_FAILED)
0c1cd1
+    {
0c1cd1
+      fprintf (stderr, "mmap #2 failed: %s.\n", strerror (errno));
0c1cd1
+      exit (1);
0c1cd1
+    }
0c1cd1
+
0c1cd1
+  /* With correct ulimit, etc. this should cause a core dump.  */
0c1cd1
+  abort ();
0c1cd1
+}