TheRock/debug-tools/rocgdb/source/gdb/testsuite/gdb.rocm/snapshot-objfile-on-load.exp

# Copyright (C) 2022-2024 Free Software Foundation, Inc.
# Copyright (C) 2022-2024 Advanced Micro Devices, Inc. All rights reserved.

# This file is part of GDB.

# This program is free software; you can redistribute it and/or modify
# it under the terms of the GNU General Public License as published by
# the Free Software Foundation; either version 3 of the License, or
# (at your option) any later version.

# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
# GNU General Public License for more details.

# You should have received a copy of the GNU General Public License
# along with this program.  If not, see <http://www.gnu.org/licenses/>.

# Code running on a GPU can be loaded directly from a memory buffer.  In such
# situation no actual file is ever used.  To support this, GDB accesses the
# objfile directly from the buffer in the memory of the debuggee.  However,
# some programs might be tempted to dispose of (or reuse) the memory buffer
# containing the objfile's image as soon as they transferred the machine code
# to the device.  If they do that, they might destroy debug information GDB
# requires, without GDB being aware of it.  This can lead to crashes if GDB
# tries to decode random data.
#
# To prevent that, GDB makes a copy of the entire objfile when it first
# encounters it.
#
# This test checks that GDB can handle a program which disposes of its objfile
# buffer while the associated code is still running on the device.
#
# In addition, it forces GDB to reload the objfile after it has been
# corrupted, to make sure GDB won't crash if it attaches to a running
# program and encounters already-corrupted images.
#
# Test different variants of corrupted images:
#
#  - fully zeroed-out buffer
#  - corrupted section offset in ELF header

load_lib rocm.exp

require allow_hip_tests

standard_testfile .cpp

set hipmodule_path [standard_output_file ${testfile}.hipfb]

if { [gdb_compile  $srcdir/$subdir/$srcfile \
	$hipmodule_path object \
	{ debug hip additional_flags=--cuda-device-only } ] != "" } {
    return -1
}

if { [build_executable "failed to prepare"  ${testfile} ${srcfile} \
	{debug hip}] } {
    return -1
}

with_rocm_gpu_lock {
    foreach_with_prefix mode {zero shoff} {
	clean_restart $::binfile
	gdb_test_no_output "set args $mode $hipmodule_path" "set args"

	runto "printId" allow-pending
	# Frame #1 normally looks like "#1  0x... in kernel () at ...",
	# but under -O1/-O2/-O3 the return address can fall on the
	# entry instruction of kernel(), in which case GDB omits the
	# leading "0x... in " prefix and prints just "#1  kernel () at
	# ...".  Accept either form.
	gdb_test "bt" \
	    [multi_line \
		 "#0\[ \t\]*printId \\(\\) at \[^\r\n\]+" \
		 "#1\[ \t\]*(${::hex} in )?kernel \\(\\) at \[^\r\n\]+"]

	# Now force GDB to reload the shared library, to make sure it
	# doesn't crash when it finds the corrupted code object image.
	#
	# Depending on the runtime version, GDB might be given the address
	# of either 1) the original code object given to the runtime, or 2) a
	# copy of the original code done by the runtime during load.
	#
	# In case #1, the client has scrambled the memory after the code object
	# has been loaded, making it an invalid ELF.  We need to make sure that
	# reloading it does not cause GDB to crash.  In case #2, we just
	# reload the same binary image, which will work fine.
	set any "\[^\r\n\]"
	gdb_test_no_output "nosharedlibrary"
	gdb_test "info shared" "memory://.*"
    }
}