Based on kernel version 6.13
. Page generated on 2025-01-21 08:21 EST
.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 | .. SPDX-License-Identifier: GPL-2.0 ========================== Userspace debugging advice ========================== This document provides a brief overview of common tools to debug the Linux Kernel from userspace. For debugging advice aimed at driver developers go :doc:`here </process/debugging/driver_development_debugging_guide>`. For general debugging advice, see :doc:`general advice document </process/debugging/index>`. .. contents:: :depth: 3 The following sections show you the available tools. Dynamic debug ------------- Mechanism to filter what ends up in the kernel log by dis-/en-abling log messages. Prerequisite: ``CONFIG_DYNAMIC_DEBUG`` Dynamic debug is only able to target: - pr_debug() - dev_dbg() - print_hex_dump_debug() - print_hex_dump_bytes() Therefore the usability of this tool is, as of now, quite limited as there is no uniform rule for adding debug prints to the codebase, resulting in a variety of ways these prints are implemented. Also, note that most debug statements are implemented as a variation of dprintk(), which have to be activated via a parameter in respective module, dynamic debug is unable to do that step for you. Here is one example, that enables all available pr_debug()'s within the file:: $ alias ddcmd='echo $* > /proc/dynamic_debug/control' $ ddcmd '-p; file v4l2-h264.c +p' $ grep =p /proc/dynamic_debug/control drivers/media/v4l2-core/v4l2-h264.c:372 [v4l2_h264]print_ref_list_b =p "ref_pic_list_b%u (cur_poc %u%c) %s" drivers/media/v4l2-core/v4l2-h264.c:333 [v4l2_h264]print_ref_list_p =p "ref_pic_list_p (cur_poc %u%c) %s\n" **When should you use this over Ftrace ?** - When the code contains one of the valid print statements (see above) or when you have added multiple pr_debug() statements during development - When timing is not an issue, meaning if multiple pr_debug() statements in the code won't cause delays - When you care more about receiving specific log messages than tracing the pattern of how a function is called For the full documentation see :doc:`/admin-guide/dynamic-debug-howto` Ftrace ------ Prerequisite: ``CONFIG_DYNAMIC_FTRACE`` This tool uses the tracefs file system for the control files and output files. That file system will be mounted as a ``tracing`` directory, which can be found in either ``/sys/kernel/`` or ``/sys/debug/kernel/``. Some of the most important operations for debugging are: - You can perform a function trace by adding a function name to the ``set_ftrace_filter`` file (which accepts any function name found within the ``available_filter_functions`` file) or you can specifically disable certain functions by adding their names to the ``set_ftrace_notrace`` file (more info at: :ref:`trace/ftrace:dynamic ftrace`). - In order to find out where calls originate from you can activate the ``func_stack_trace`` option under ``options/func_stack_trace``. - Tracing the children of a function call and showing the return values are possible by adding the desired function in the ``set_graph_function`` file (requires config ``FUNCTION_GRAPH_RETVAL``); more info at :ref:`trace/ftrace:dynamic ftrace with the function graph tracer`. For the full Ftrace documentation see :doc:`/trace/ftrace` Or you could also trace for specific events by :ref:`using event tracing <trace/events:2. using event tracing>`, which can be defined as described here: :ref:`Creating a custom Ftrace tracepoint <process/debugging/driver_development_debugging_guide:ftrace>`. For the full Ftrace event tracing documentation see :doc:`/trace/events` .. _read_ftrace_log: Reading the ftrace log ~~~~~~~~~~~~~~~~~~~~~~ The ``trace`` file can be read just like any other file (``cat``, ``tail``, ``head``, ``vim``, etc.), the size of the file is limited by the ``buffer_size_kb`` (``echo 1000 > buffer_size_kb``). The :ref:`trace/ftrace:trace_pipe` will behave similarly to the ``trace`` file, but whenever you read from the file the content is consumed. Kernelshark ~~~~~~~~~~~ A GUI interface to visualize the traces as a graph and list view from the output of the `trace-cmd <https://git.kernel.org/pub/scm/utils/trace-cmd/trace-cmd.git/>`__ application. For the full documentation see `<https://kernelshark.org/Documentation.html>`__ Perf & alternatives ------------------- The tools mentioned above provide ways to inspect kernel code, results, variable values, etc. Sometimes you have to find out first where to look and for those cases, a box of performance tracking tools can help you to frame the issue. Why should you do a performance analysis? ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ A performance analysis is a good first step when among other reasons: - you cannot define the issue - you do not know where it occurs - the running system should not be interrupted or it is a remote system, where you cannot install a new module/kernel How to do a simple analysis with linux tools? ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ For the start of a performance analysis, you can start with the usual tools like: - ``top`` / ``htop`` / ``atop`` (*get an overview of the system load, see spikes on specific processes*) - ``mpstat -P ALL`` (*look at the load distribution among CPUs*) - ``iostat -x`` (*observe input and output devices utilization and performance*) - ``vmstat`` (*overview of memory usage on the system*) - ``pidstat`` (*similar to* ``vmstat`` *but per process, to dial it down to the target*) - ``strace -tp $PID`` (*once you know the process, you can figure out how it communicates with the Kernel*) These should help to narrow down the areas to look at sufficiently. Diving deeper with perf ~~~~~~~~~~~~~~~~~~~~~~~ The **perf** tool provides a series of metrics and events to further dial down on issues. Prerequisite: build or install perf on your system Gather statistics data for finding all files starting with ``gcc`` in ``/usr``:: # perf stat -d find /usr -name 'gcc*' | wc -l Performance counter stats for 'find /usr -name gcc*': 1277.81 msec task-clock # 0.997 CPUs utilized 9 context-switches # 7.043 /sec 1 cpu-migrations # 0.783 /sec 704 page-faults # 550.943 /sec 766548897 cycles # 0.600 GHz (97.15%) 798285467 instructions # 1.04 insn per cycle (97.15%) 57582731 branches # 45.064 M/sec (2.85%) 3842573 branch-misses # 6.67% of all branches (97.15%) 281616097 L1-dcache-loads # 220.390 M/sec (97.15%) 4220975 L1-dcache-load-misses # 1.50% of all L1-dcache accesses (97.15%) <not supported> LLC-loads <not supported> LLC-load-misses 1.281746009 seconds time elapsed 0.508796000 seconds user 0.773209000 seconds sys 52 The availability of events and metrics depends on the system you are running. For the full documentation see `<https://perf.wiki.kernel.org/index.php/Main_Page>`__ Perfetto ~~~~~~~~ A set of tools to measure and analyze how well applications and systems perform. You can use it to: * identify bottlenecks * optimize code * make software run faster and more efficiently. **What is the difference between perfetto and perf?** * perf is tool as part of and specialized for the Linux Kernel and has CLI user interface. * perfetto cross-platform performance analysis stack, has extended functionality into userspace and provides a WEB user interface. For the full documentation see `<https://perfetto.dev/docs/>`__ Kernel panic analysis tools --------------------------- To capture the crash dump please use ``Kdump`` & ``Kexec``. Below you can find some advice for analysing the data. For the full documentation see the :doc:`/admin-guide/kdump/kdump` In order to find the corresponding line in the code you can use `faddr2line <https://elixir.bootlin.com/linux/v6.11.6/source/scripts/faddr2line>`__; note that you need to enable ``CONFIG_DEBUG_INFO`` for that to work. An alternative to using ``faddr2line`` is the use of ``objdump`` (and its derivatives for the different platforms like ``aarch64-linux-gnu-objdump``). Take this line as an example: ``[ +0.000240] rkvdec_device_run+0x50/0x138 [rockchip_vdec]``. We can find the corresponding line of code by executing:: aarch64-linux-gnu-objdump -dS drivers/staging/media/rkvdec/rockchip-vdec.ko | grep rkvdec_device_run\>: -A 40 0000000000000ac8 <rkvdec_device_run>: ac8: d503201f nop acc: d503201f nop { ad0: d503233f paciasp ad4: a9bd7bfd stp x29, x30, [sp, #-48]! ad8: 910003fd mov x29, sp adc: a90153f3 stp x19, x20, [sp, #16] ae0: a9025bf5 stp x21, x22, [sp, #32] const struct rkvdec_coded_fmt_desc *desc = ctx->coded_fmt_desc; ae4: f9411814 ldr x20, [x0, #560] struct rkvdec_dev *rkvdec = ctx->dev; ae8: f9418015 ldr x21, [x0, #768] if (WARN_ON(!desc)) aec: b4000654 cbz x20, bb4 <rkvdec_device_run+0xec> ret = pm_runtime_resume_and_get(rkvdec->dev); af0: f943d2b6 ldr x22, [x21, #1952] ret = __pm_runtime_resume(dev, RPM_GET_PUT); af4: aa0003f3 mov x19, x0 af8: 52800081 mov w1, #0x4 // #4 afc: aa1603e0 mov x0, x22 b00: 94000000 bl 0 <__pm_runtime_resume> if (ret < 0) { b04: 37f80340 tbnz w0, #31, b6c <rkvdec_device_run+0xa4> dev_warn(rkvdec->dev, "Not good\n"); b08: f943d2a0 ldr x0, [x21, #1952] b0c: 90000001 adrp x1, 0 <rkvdec_try_ctrl-0x8> b10: 91000021 add x1, x1, #0x0 b14: 94000000 bl 0 <_dev_warn> *bad = 1; b18: d2800001 mov x1, #0x0 // #0 ... Meaning, in this line from the crash dump:: [ +0.000240] rkvdec_device_run+0x50/0x138 [rockchip_vdec] I can take the ``0x50`` as offset, which I have to add to the base address of the corresponding function, which I find in this line:: 0000000000000ac8 <rkvdec_device_run>: The result of ``0xac8 + 0x50 = 0xb18`` And when I search for that address within the function I get the following line:: *bad = 1; b18: d2800001 mov x1, #0x0 **Copyright** ©2024 : Collabora |