Skip to content

Latest commit

 

History

History
152 lines (116 loc) · 7.06 KB

advanced-gdb-debugging.md

File metadata and controls

152 lines (116 loc) · 7.06 KB
Raw

Embedded Debugging

The autopilots running PX4 support debugging via GDB or LLDB.

Identifying large memory consumers

The command below will list the largest static allocations:

arm-none-eabi-nm --size-sort --print-size --radix=dec build_px4fmu-v2_default/src/firmware/nuttx/firmware_nuttx | grep " [bBdD] "

This NSH command provides the remaining free memory:

free

And the top command shows the stack usage per application:

top

Stack usage is calculated with stack coloring and thus is not the current usage, but the maximum since the start of the task.

Heap allocations

Dynamic heap allocations can be traced on POSIX in SITL with gperftools. Once installed, it can be used with:

  • Run jmavsim: ./Tools/jmavsim_run.sh
  • Then:
cd build_posix_sitl_default/tmp
export HEAPPROFILE=/tmp/heapprofile.hprof
env LD_PRELOAD=/lib64/libtcmalloc.so ../src/firmware/posix/px4 posix-configs/SITL/init/lpe/iris
pprof --pdf ../src/firmware/posix/px4 /tmp/heapprofile.hprof.0001.heap > heap.pdf

It will generate a pdf with a graph of the heap allocations. The numbers in the graph will all be zero, because they are in MB. Just look at the percentages instead. They show the live memory (of the node and the subtree), meaning the memory that was still in use at the end.

If it does not generate heap dumps while running the px4 app you might need to change the settings of the profiler. On some systems it is necessary to set an interval time when to write the dumps:

# Specify interval in seconds
export HEAP_PROFILE_TIME_INTERVAL=10

See the gperftools docs for more information.

Sending MAVLink debug key / value pairs

The code for this tutorial is available here:

All required to set up a debug publication is this code snippet. First add the header file:

#include <uORB/uORB.h>
#include <uORB/topics/debug_key_value.h>

Then advertise the debug value topic (one advertisement for different published names is sufficient). Put this in front of your main loop:

Debugging Hard Faults in NuttX

A hard fault is a state when the operating system detects that it has no valid instructions to execute. This is typically the case when key areas in RAM have been corrupted. A typical scenario is when incorrect memory access smashed the stack and the processor sees that the address in memory is not a valid address for the microprocessors's RAM.

  • NuttX maintains two stacks: The IRQ stack for interrupt processing and the user stack
  • The stack grows downward. So the highest address in the example below is 0x20021060, the size is 0x11f4 (4596 bytes) and consequently the lowest address is 0x2001fe6c.
Assertion failed at file:armv7-m/up_hardfault.c line: 184 task: ekf_att_pos_estimator
sp:     20003f90
IRQ stack:
  base: 20003fdc
  size: 000002e8
20003f80: 080d27c6 20003f90 20021060 0809b8d5 080d288c 000000b8 08097155 00000010
20003fa0: 20003ce0 00000003 00000000 0809bb61 0809bb4d 080a6857 e000ed24 080a3879
20003fc0: 00000000 2001f578 080ca038 000182b8 20017cc0 0809bad1 20020c14 00000000
sp:     20020ce8
User stack:
  base: 20021060
  size: 000011f4
20020ce0: 60000010 2001f578 2001f578 080ca038 000182b8 0808439f 2001fb88 20020d4c
20020d00: 20020d44 080a1073 666b655b 65686320 205d6b63 6f6c6576 79746963 76696420
20020d20: 65747265 63202c64 6b636568 63636120 63206c65 69666e6f 08020067 0805c4eb
20020d40: 080ca9d4 0805c21b 080ca1cc 080ca9d4 385833fb 38217db9 00000000 080ca964
20020d60: 080ca980 080ca9a0 080ca9bc 080ca9d4 080ca9fc 080caa14 20022824 00000002
20020d80: 2002218c 0806a30f 08069ab2 81000000 3f7fffec 00000000 3b4ae00c 3b12eaa6
20020da0: 00000000 00000000 080ca010 4281fb70 20020f78 20017cc0 20020f98 20017cdc
20020dc0: 2001ee0c 0808d7ff 080ca010 00000000 3f800000 00000000 080ca020 3aa35c4e
20020de0: 3834d331 00000000 01010101 00000000 01010001 000d4f89 000d4f89 000f9fda
20020e00: 3f7d8df4 3bac67ea 3ca594e6 be0b9299 40b643aa 41ebe4ed bcc04e1b 43e89c96
20020e20: 448f3bc9 c3c50317 b4c8d827 362d3366 b49d74cf ba966159 00000000 00000000
20020e40: 3eb4da7b 3b96b9b7 3eead66a 00000000 00000000 00000000 00000000 00000000
20020e60: 00000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000
20020e80: 00000016 00000000 00000000 00010000 00000000 3c23d70a 00000000 00000000
20020ea0: 00000000 20020f78 00000000 2001ed20 20020fa4 2001f498 2001f1a8 2001f500
20020ec0: 2001f520 00000003 2001f170 ffffffe9 3b831ad2 3c23d70a 00000000 00000000
20020ee0: 00000000 00000000 00000000 00000000 00000000 00000000 00000000 00000000
20020f00: 00000000 00000000 00000000 00000000 2001f4f0 2001f4a0 3d093964 00000001
20020f20: 00000000 0808ae91 20012d10 2001da40 0000260b 2001f577 2001da40 0000260b
20020f40: 2001f1a8 08087fd7 08087f9d 080cf448 0000260b 080afab1 080afa9d 00000003
20020f60: 2001f577 0809c577 2001ed20 2001f4d8 2001f498 0805e077 2001f568 20024540
20020f80: 00000000 00000000 00000000 0000260b 3d093a57 00000000 2001f540 2001f4f0
20020fa0: 0000260b 3ea5b000 3ddbf5fa 00000000 3c23d70a 00000000 00000000 000f423f
20020fc0: 00000000 000182b8 20017cc0 2001ed20 2001f4e8 00000000 2001f120 0805ea0d
20020fe0: 2001f090 2001f120 2001eda8 ffffffff 000182b8 00000000 00000000 00000000
20021000: 00000000 00000000 00000009 00000000 08090001 2001f93c 0000000c 00000000
20021020: 00000101 2001f96c 00000000 00000000 00000000 00000000 00000000 00000000
20021040: 00000000 00000000 00000000 00000000 00000000 0809866d 00000000 00000000
R0: 20000f48 0a91ae0c 20020d00 20020d00 2001f578 080ca038 000182b8 20017cc0
R8: 2001ed20 2001f4e8 2001ed20 00000005 20020d20 20020ce8 0808439f 08087c4e
xPSR: 61000000 BASEPRI: 00000000 CONTROL: 00000000
EXC_RETURN: ffffffe9

To decode the hardfault, load the exact binary into the debugger:

arm-none-eabi-gdb build_px4fmu-v2_default/src/firmware/nuttx/firmware_nuttx

Then in the GDB prompt, start with the last instructions in R8, with the first address in flash (recognizable because it starts with 0x080, the first is 0x0808439f). The execution is left to right. So one of the last steps before the hard fault was when mavlink_log.c tried to publish something,

(gdb) info line *0x0808439f
Line 77 of "../src/modules/systemlib/mavlink_log.c" starts at address 0x8084398 <mavlink_vasprintf+36>
   and ends at 0x80843a0 <mavlink_vasprintf+44>.
(gdb) info line *0x08087c4e
Line 311 of "../src/modules/uORB/uORBDevices_nuttx.cpp"
   starts at address 0x8087c4e <uORB::DeviceNode::publish(orb_metadata const*, void*, void const*)+2>
   and ends at 0x8087c52 <uORB::DeviceNode::publish(orb_metadata const*, void*, void const*)+6>.