Hi Zeid,

On Sat, 31 Oct 2020 at 23:11, Zied Guermazi zied.guermazi@trande.de wrote:

hi,

while testing the implementation in gdb of branch tracing on arm processors using etm, I faced the the situation where a breakpoint was set, was hit and then the execution of the program was continued. While decoding generated traces, I got the address of the breakpoint (0x400552) executed twice, and then the following address (0x400554) also executed twice. the instruction at (0x400554) is a BL ( a function call) and the second execution corrupts the function history.

here is a dump of generated trace elements

---

trace_chan_id: 18 isa: CS_ETM_ISA_T32 start addr = 0x400552 end addr = 0x400554 instructions count = 1 last_i_type: OCSD_INSTR_OTHER last_i_subtype: OCSD_S_INSTR_NONE last instruction was executed last instruction size: 2

---

trace_chan_id: 18 isa: CS_ETM_ISA_T32 start addr = 0x400552 end addr = 0x400554 instructions count = 1 last_i_type: OCSD_INSTR_OTHER last_i_subtype: OCSD_S_INSTR_NONE last instruction was executed last instruction size: 2

---

trace_chan_id: 18 isa: CS_ETM_ISA_T32 start addr = 0x400554 end addr = 0x400558 instructions count = 1 last_i_type: OCSD_INSTR_BR last_i_subtype: OCSD_S_INSTR_BR_LINK last instruction was executed last instruction size: 4

---

trace_chan_id: 18 isa: CS_ETM_ISA_T32 start addr = 0x400554 end addr = 0x400558 instructions count = 1 last_i_type: OCSD_INSTR_BR last_i_subtype: OCSD_S_INSTR_BR_LINK last instruction was executed last instruction size: 4

the explanation I have for this behavior is that :

-when setting the software breakpoint, the memory content of the instruction (at 0x400552) was altered to the instruction BKPT,

-when the breakpoint was hit, the original opcode was set at (0x400552) and a BKPT was set to the next instruction address (0x400554), then the execution was continued

-when the second breakpoint (0x400554) was hit, the a BKPT opcode was set at (0x400552) and the original opcode was set at (0x400554) then the execution was continued

I am using the function "int target_read_code (CORE_ADDR memaddr, gdb_byte *myaddr, ssize_t len)" to give program memory content to the decoder. so the collected etm traces are correct, but, as memory was altered in between, the decoder is "cheated".

I need to identify the re-execution of code due to breakpoint handling, and roll back its impact on etm decoding.

is there a mean to get the actual content of program memory including patched addresses?

is there a means of getting the history of patched addresses during the debugging of a program?

what is the type and subtype of a BKPT instruction in a decoded trace elements?

I can only really comment on this question. The type / subtype information in the output from the decoder is generated from the decoder walking the memory image of the executed trace - not from the trace packets themselves. The decoder classifies instructions according to how they will affect trace flow with the "other" category being set for the majority of instructions. The categories are: other, branch, indirect branch, ISB / DSB / DMB / WFI / WFE. These are important in program flow trace (PTM 1.x, ETM 4.x) as these determine which instruction we attach the E/N atoms to. BKPT will be classified as "other", if it is seen, as it has no effect on normal program flow. It will cause an exception which has a specific trace packet format.

Regards

Mike

do you have any other idea for handling this situation?

I am attaching the source code of the program as well as the disassembled binary. the code was compiled as an application running on linux on an ARMv7 A (STM32MP157 SoC). the breakpoint was set at line 43 in the source code (line 238 in the disassembled code)

Kind Regards

Zied Guermazi

---

CoreSight mailing list CoreSight@lists.linaro.org https://lists.linaro.org/mailman/listinfo/coresight

Hi Zied,

On Mon, 2 Nov 2020 at 15:52, Zied Guermazi zied.guermazi@trande.de wrote:

hi

Thanks Mike for your support, it was very helpful.

to put everything together, on arm, gdb inserts a sw breakpoint by patching the code with an undefined instruction ( see comments in arm-tdep.c line7687) when a breakpoint is hit, an exception number 9 "Undefined Instruction exception" is raised and a branch packet with this info is generated in etm traces, the trap is get handled by the kernel and it sends the appropriate signal to gdb process.

Looks like that code was designed for very early architectures. I wonder if it should not use the architected BKPT instruction when available (arch v5T onwards I think).

when the user continues the execution, gdb patches back the code and executes the instruction. this leads to the instruction traced twice with an exception in between, the same happens for next executed instruction

here is the log of decoded packets

[btrace] [ftrace] update insn: fun = main, file = ./function_call_history.c, level = 0, insn = [1; 2) cs_etm_decoder_trace_element_callback: elem->elem_type OCSD_GEN_TRC_ELEM_INSTR_RANGE <= first execution attempt that raises an undefined instruction exception trace_chan_id: 18 isa: CS_ETM_ISA_T32 start addr = 0x400534 end addr = 0x400536 instructions count = 1 last_i_type: OCSD_INSTR_OTHER last_i_subtype: OCSD_S_INSTR_NONE last instruction was executed last instruction size: 2 [btrace] [ftrace] update insn: fun = main, file = ./function_call_history.c, level = 0, insn = [1; 3) cs_etm_decoder_trace_element_callback: elem->elem_type OCSD_GEN_TRC_ELEM_EXCEPTION <= the exception is traced trace_chan_id: 18 exception number: 9 <= undefined instruction exception cs_etm_decoder_trace_element_callback: elem->elem_type OCSD_GEN_TRC_ELEM_TRACE_ON cs_etm_decoder_trace_element_callback: elem->elem_type OCSD_GEN_TRC_ELEM_PE_CONTEXT cs_etm_decoder_trace_element_callback: elem->elem_type OCSD_GEN_TRC_ELEM_INSTR_RANGE <= execution of the original instruction trace_chan_id: 18 isa: CS_ETM_ISA_T32 start addr = 0x400534 end addr = 0x400536 instructions count = 1 last_i_type: OCSD_INSTR_OTHER last_i_subtype: OCSD_S_INSTR_NONE last instruction was executed last instruction size: 2

as the code was changed during execution, it can not be reconstructed during traces decoding.

in addition, and for tracing applications running on Linux, we are not interested in capturing raised exceptions, we can consider rolling back last instruction in ftraces. As this is not obvious, we can consider ignoring the repeated instruction as a workaround.

for tracing bare metal software, we need to keep tracing exception, so we can have a flag for ignoring exceptions, and activate or dis-activate it according to the context.

what do you think about it, shall I go for implementing it as described above?

I assume that in this scenario, trace collection is ongoing over the BKPT hit / restart sequence and is decoded at some point later. Otherwise spotting the breakpoint would be easy.

I cannot think of many circumstances where an instruction would be executed - or appear to be executed in the trace twice in succession* - other than being restarted after an exception. This debug case is one of those occasions - I would check that there are not other exceptions that might mimic this.

Other than that it would appear that the execute / exception / execute again pattern can be used to spot a break and the 1st execute could be dropped since this was the breakpoint. If it was set on a conditional then you are interested in the actual trace result which could be either executed or not.

Regards

Mike

* branch to self might appear like this, as would setting the trace address filters to a include just single instruction - but there would be no intervening exceptions in these cases.

Kind Regards

Zied Guermazi

On 02.11.20 12:59, Mike Leach wrote:

Hi Zeid,

On Sat, 31 Oct 2020 at 23:11, Zied Guermazi zied.guermazi@trande.de wrote:

hi,

while testing the implementation in gdb of branch tracing on arm processors using etm, I faced the the situation where a breakpoint was set, was hit and then the execution of the program was continued. While decoding generated traces, I got the address of the breakpoint (0x400552) executed twice, and then the following address (0x400554) also executed twice. the instruction at (0x400554) is a BL ( a function call) and the second execution corrupts the function history.

here is a dump of generated trace elements

---

trace_chan_id: 18 isa: CS_ETM_ISA_T32 start addr = 0x400552 end addr = 0x400554 instructions count = 1 last_i_type: OCSD_INSTR_OTHER last_i_subtype: OCSD_S_INSTR_NONE last instruction was executed last instruction size: 2

---

trace_chan_id: 18 isa: CS_ETM_ISA_T32 start addr = 0x400552 end addr = 0x400554 instructions count = 1 last_i_type: OCSD_INSTR_OTHER last_i_subtype: OCSD_S_INSTR_NONE last instruction was executed last instruction size: 2

---

trace_chan_id: 18 isa: CS_ETM_ISA_T32 start addr = 0x400554 end addr = 0x400558 instructions count = 1 last_i_type: OCSD_INSTR_BR last_i_subtype: OCSD_S_INSTR_BR_LINK last instruction was executed last instruction size: 4

---

trace_chan_id: 18 isa: CS_ETM_ISA_T32 start addr = 0x400554 end addr = 0x400558 instructions count = 1 last_i_type: OCSD_INSTR_BR last_i_subtype: OCSD_S_INSTR_BR_LINK last instruction was executed last instruction size: 4

the explanation I have for this behavior is that :

-when setting the software breakpoint, the memory content of the instruction (at 0x400552) was altered to the instruction BKPT,

-when the breakpoint was hit, the original opcode was set at (0x400552) and a BKPT was set to the next instruction address (0x400554), then the execution was continued

-when the second breakpoint (0x400554) was hit, the a BKPT opcode was set at (0x400552) and the original opcode was set at (0x400554) then the execution was continued

I am using the function "int target_read_code (CORE_ADDR memaddr, gdb_byte *myaddr, ssize_t len)" to give program memory content to the decoder. so the collected etm traces are correct, but, as memory was altered in between, the decoder is "cheated".

I need to identify the re-execution of code due to breakpoint handling, and roll back its impact on etm decoding.

is there a mean to get the actual content of program memory including patched addresses?

is there a means of getting the history of patched addresses during the debugging of a program?

what is the type and subtype of a BKPT instruction in a decoded trace elements?

I can only really comment on this question. The type / subtype information in the output from the decoder is generated from the decoder walking the memory image of the executed trace - not from the trace packets themselves. The decoder classifies instructions according to how they will affect trace flow with the "other" category being set for the majority of instructions. The categories are: other, branch, indirect branch, ISB / DSB / DMB / WFI / WFE. These are important in program flow trace (PTM 1.x, ETM 4.x) as these determine which instruction we attach the E/N atoms to. BKPT will be classified as "other", if it is seen, as it has no effect on normal program flow. It will cause an exception which has a specific trace packet format.

Regards

Mike

do you have any other idea for handling this situation?

I am attaching the source code of the program as well as the disassembled binary. the code was compiled as an application running on linux on an ARMv7 A (STM32MP157 SoC). the breakpoint was set at line 43 in the source code (line 238 in the disassembled code)

Kind Regards

Zied Guermazi

---

CoreSight mailing list CoreSight@lists.linaro.org https://lists.linaro.org/mailman/listinfo/coresight