Re: [PATCH] perf scripts python arm-cs-trace-disasm.py: Skip disasm if address continuity is broken
From: James Clark
Date: Thu Aug 01 2024 - 07:26:40 EST
On 01/08/2024 11:28 am, Al Grant wrote:
-----Original Message-----
From: James Clark <james.clark@xxxxxxxxxx>
Sent: Thursday, August 1, 2024 11:00 AM
To: Ganapatrao Kulkarni <gankulkarni@xxxxxxxxxxxxxxxxxxxxxx>; Mike Leach
<mike.leach@xxxxxxxxxx>
Cc: acme@xxxxxxxxxx; coresight@xxxxxxxxxxxxxxxx; linux-arm-
kernel@xxxxxxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx;
darren@xxxxxxxxxxxxxxxxxxxxxx; scclevenger@xxxxxxxxxxxxxxxxxxxxxx; Leo
Yan <Leo.Yan@xxxxxxx>
Subject: Re: [PATCH] perf scripts python arm-cs-trace-disasm.py: Skip disasm if
address continuity is broken
On 24/07/2024 3:45 pm, James Clark wrote:
On 24/07/2024 7:38 am, Ganapatrao Kulkarni wrote:
On 23-07-2024 09:16 pm, James Clark wrote:
On 23/07/2024 4:26 pm, Ganapatrao Kulkarni wrote:
On 23-07-2024 06:40 pm, James Clark wrote:
On 22/07/2024 11:02 am, Ganapatrao Kulkarni wrote:
Hi James,
On 19-07-2024 08:09 pm, James Clark wrote:
On 19/07/2024 10:26 am, Ganapatrao Kulkarni wrote:
To generate the instruction tracing, script uses 2 contiguous
packets address range. If there a continuity brake due to
discontiguous branch address, it is required to reset the
tracing and start tracing with the new set of contiguous
packets.
Adding change to identify the break and complete the remaining
tracing of current packets and restart tracing from new set of
packets, if continuity is established.
Hi Ganapatrao,
Can you add a before and after example of what's changed to the
commit message? It wasn't immediately obvious to me if this is
adding missing output, or it was correcting the tail end of the
output that was previously wrong.
It is adding tail end of the trace as well avoiding the segfault
of the perf application. With out this change the perf segfaults
with as below log
./perf script
--script=python:./scripts/python/arm-cs-trace-disasm.py -- -d
objdump -k ../../vmlinux -v $* > dump
objdump: error: the stop address should be after the start address
Traceback (most recent call last):
File "./scripts/python/arm-cs-trace-disasm.py", line 271, in
process_event
print_disam(dso_fname, dso_vm_start, start_addr, stop_addr)
File "./scripts/python/arm-cs-trace-disasm.py", line 105, in
print_disam
for line in read_disam(dso_fname, dso_start, start_addr,
stop_addr):
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "./scripts/python/arm-cs-trace-disasm.py", line 99, in
read_disam
disasm_output = check_output(disasm).decode('utf-8').split('\n')
^^^^^^^^^^^^^^^^^^^^
File "/usr/lib64/python3.12/subprocess.py", line 466, in
check_output
return run(*popenargs, stdout=PIPE, timeout=timeout, check=True,
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
^^^^
File "/usr/lib64/python3.12/subprocess.py", line 571, in run
raise CalledProcessError(retcode, process.args,
subprocess.CalledProcessError: Command '['objdump', '-d', '-z',
'--start-address=0xffff80008125b758',
'--stop-address=0xffff80008125a934', '../../vmlinux']' returned
non-zero exit status 1.
Fatal Python error: handler_call_die: problem in Python trace
event handler
Python runtime state: initialized
Current thread 0x0000ffffb05054e0 (most recent call first):
<no Python frame>
Extension modules: perf_trace_context, systemd._journal,
systemd._reader, systemd.id128, report._py3report, _dbus_bindings,
problem._py3abrt (total: 7)
Aborted (core dumped)
Signed-off-by: Ganapatrao Kulkarni
<gankulkarni@xxxxxxxxxxxxxxxxxxxxxx>
---
tools/perf/scripts/python/arm-cs-trace-disasm.py | 10 ++++++++++
1 file changed, 10 insertions(+)
diff --git a/tools/perf/scripts/python/arm-cs-trace-disasm.py
b/tools/perf/scripts/python/arm-cs-trace-disasm.py
index d973c2baed1c..ad10cee2c35e 100755
--- a/tools/perf/scripts/python/arm-cs-trace-disasm.py
+++ b/tools/perf/scripts/python/arm-cs-trace-disasm.py
@@ -198,6 +198,10 @@ def process_event(param_dict):
cpu_data[str(cpu) + 'addr'] = addr
return
+ if (cpu_data.get(str(cpu) + 'ip') == None):
+ cpu_data[str(cpu) + 'ip'] = ip
+
Do you need to write into the global cpu_data here? Doesn't it
get overwritten after you load it back into 'prev_ip'
No, the logic is same as holding the addr of previous packet.
Saving the previous packet saved ip in to prev_ip before
overwriting with the current packet.
It's not exactly the same logic as holding the addr of the previous
sample. For addr, we return on the first None, with your change we
now "pretend" that the second one is also the previous one:
if (cpu_data.get(str(cpu) + 'addr') == None):
cpu_data[str(cpu) + 'addr'] = addr
return <----------------------------sample 0 return
if (cpu_data.get(str(cpu) + 'ip') == None):
cpu_data[str(cpu) + 'ip'] = ip <---- sample 1 save but no
return
Then for sample 1 'prev_ip' is actually now the 'current' IP:
Yes, it is dummy for first packet. Added anticipating that we wont
hit the discontinuity for the first packet itself.
Can this be changed to more intuitive like below?
diff --git a/tools/perf/scripts/python/arm-cs-trace-disasm.py
b/tools/perf/scripts/python/arm-cs-trace-disasm.py
index d973c2baed1c..d49f5090059f 100755
--- a/tools/perf/scripts/python/arm-cs-trace-disasm.py
+++ b/tools/perf/scripts/python/arm-cs-trace-disasm.py
@@ -198,6 +198,8 @@ def process_event(param_dict):
cpu_data[str(cpu) + 'addr'] = addr
return
+ if (cpu_data.get(str(cpu) + 'ip') != None):
+ prev_ip = cpu_data[str(cpu) + 'ip']
if (options.verbose == True):
print("Event type: %s" % name)
@@ -243,12 +245,18 @@ def process_event(param_dict):
# Record for previous sample packet
cpu_data[str(cpu) + 'addr'] = addr
+ cpu_data[str(cpu) + 'ip'] = stop_addr
# Handle CS_ETM_TRACE_ON packet if start_addr=0 and
stop_addr=4
if (start_addr == 0 and stop_addr == 4):
print("CPU%d: CS_ETM_TRACE_ON packet is inserted" %
cpu)
return
+ if (stop_addr < start_addr and prev_ip != 0):
+ # Continuity of the Packets broken, set start_addr
to previous
+ # packet ip to complete the remaining tracing of the
address range.
+ start_addr = prev_ip
+
if (start_addr < int(dso_start) or start_addr > int(dso_end)):
print("Start address 0x%x is out of range [ 0x%x ..
0x%x ] for dso %s" % (start_addr, int(dso_start), int(dso_end), dso))
return
Without this patch below is the failure log(with segfault) for
reference.
[root@sut01sys-r214 perf]# timeout 4s ./perf record -e cs_etm// -C 1
dd if=/dev/zero of=/dev/null
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 1.087 MB perf.data ]
[root@sut01sys-r214 perf]# ./perf script
--script=python:./scripts/python/arm-cs-trace-disasm.py -- -d
objdump -k ../../vmlinux -v $* > dump
objdump: error: the stop address should be after the start address
Traceback (most recent call last):
File "./scripts/python/arm-cs-trace-disasm.py", line 271, in
process_event
print_disam(dso_fname, dso_vm_start, start_addr, stop_addr)
File "./scripts/python/arm-cs-trace-disasm.py", line 105, in
print_disam
for line in read_disam(dso_fname, dso_start, start_addr,
stop_addr):
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "./scripts/python/arm-cs-trace-disasm.py", line 99, in
read_disam
disasm_output = check_output(disasm).decode('utf-8').split('\n')
^^^^^^^^^^^^^^^^^^^^
File "/usr/lib64/python3.12/subprocess.py", line 466, in
check_output
return run(*popenargs, stdout=PIPE, timeout=timeout, check=True,
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
^^
File "/usr/lib64/python3.12/subprocess.py", line 571, in run
raise CalledProcessError(retcode, process.args,
subprocess.CalledProcessError: Command '['objdump', '-d', '-z',
'--start-address=0xffff80008125b758',
'--stop-address=0xffff80008125a934', '../../vmlinux']' returned
non-zero exit status 1.
Fatal Python error: handler_call_die: problem in Python trace event
handler
Python runtime state: initialized
Current thread 0x0000ffffb90d54e0 (most recent call first):
<no Python frame>
Extension modules: perf_trace_context, systemd._journal,
systemd._reader, systemd.id128, report._py3report, _dbus_bindings,
problem._py3abrt (total: 7)
Aborted (core dumped)
dump snippet:
============
Event type: branches
Sample = { cpu: 0001 addr: 0xffff80008030cb00 phys_addr:
0x0000000000000000 ip: 0xffff800080313f0c pid: 12720 tid: 12720
period: 1 time: 5986372298040 }
ffff800080313f04 <__perf_event_header__init_id+0x4c>:
ffff800080313f04: 36100094 tbz w20, #2,
ffff800080313f14 <__perf_event_header__init_id+0x5c>
ffff800080313f08: f941e6a0 ldr x0, [x21,
#968]
ffff800080313f0c: d63f0000 blr x0
perf 12720/12720 [0001] 5986.372298040
__perf_event_header__init_id+0x54
.../coresight/linux/kernel/events/core.c 586 return
event->clock();
Event type: branches
Sample = { cpu: 0001 addr: 0xffff8000801bb4a8 phys_addr:
0x0000000000000000 ip: 0xffff80008030cb0c pid: 12720 tid: 12720
period: 1 time: 5986372298040 }
ffff80008030cb00 <local_clock>:
ffff80008030cb00: d503233f paciasp
ffff80008030cb04: a9bf7bfd stp x29, x30,
[sp, #-16]!
ffff80008030cb08: 910003fd mov x29, sp
ffff80008030cb0c: 97faba67 bl ffff8000801bb4a8
<sched_clock>
perf 12720/12720 [0001] 5986.372298040
local_clock+0xc ...t/linux/./include/linux/sched/clock.h 64 return
sched_clock();
Event type: branches
Sample = { cpu: 0001 addr: 0xffff80008125a8a8 phys_addr:
0x0000000000000000 ip: 0xffff8000801bb4c8 pid: 12720 tid: 12720
period: 1 time: 5986372298040 }
ffff8000801bb4a8 <sched_clock>:
ffff8000801bb4a8: d503233f paciasp
ffff8000801bb4ac: a9be7bfd stp x29, x30,
[sp, #-32]!
ffff8000801bb4b0: 910003fd mov x29, sp
ffff8000801bb4b4: a90153f3 stp x19, x20,
[sp, #16]
ffff8000801bb4b8: d5384113 mrs x19, sp_el0
ffff8000801bb4bc: b9401260 ldr w0, [x19, #16]
ffff8000801bb4c0: 11000400 add w0, w0, #0x1
ffff8000801bb4c4: b9001260 str w0, [x19, #16]
ffff8000801bb4c8: 94427cf8 bl ffff80008125a8a8
<sched_clock_noinstr>
perf 12720/12720 [0001] 5986.372298040
sched_clock+0x20 ...sight/linux/kernel/time/sched_clock.c 105 ns =
sched_clock_noinstr();
Event type: branches
Sample = { cpu: 0001 addr: 0xffff80008125b758 phys_addr:
0x0000000000000000 ip: 0xffff80008125a8e4 pid: 12720 tid: 12720
period: 1 time: 5986372298040 }
ffff80008125a8a8 <sched_clock_noinstr>:
ffff80008125a8a8: d503233f paciasp
ffff80008125a8ac: a9bc7bfd stp x29, x30,
[sp, #-64]!
ffff80008125a8b0: 910003fd mov x29, sp
ffff80008125a8b4: a90153f3 stp x19, x20,
[sp, #16]
ffff80008125a8b8: b000e354 adrp x20,
ffff800082ec3000 <tick_bc_dev+0x140>
ffff80008125a8bc: 910d0294 add x20, x20,
#0x340
ffff80008125a8c0: a90363f7 stp x23, x24,
[sp, #48]
ffff80008125a8c4: 91002297 add x23, x20, #0x8
ffff80008125a8c8: 52800518 mov w24, #0x28
// #40
ffff80008125a8cc: a9025bf5 stp x21, x22,
[sp, #32]
ffff80008125a8d0: b9400296 ldr w22, [x20]
ffff80008125a8d4: 120002d5 and w21, w22, #0x1
ffff80008125a8d8: 9bb87eb5 umull x21, w21, w24
ffff80008125a8dc: 8b1502f3 add x19, x23, x21
ffff80008125a8e0: f9400e60 ldr x0, [x19, #24]
ffff80008125a8e4: d63f0000 blr x0
perf 12720/12720 [0001] 5986.372298040
sched_clock_noinstr+0x3c ...sight/linux/kernel/time/sched_clock.c
93 cyc = (rd->read_sched_clock() - rd->epoch_cyc) &
Event type: branches
Sample = { cpu: 0001 addr: 0xffff8000801bb4cc phys_addr:
0x0000000000000000 ip: 0xffff80008125a930 pid: 12720 tid: 12720
period: 1 time: 5986372298040 }
With fix:
=========
Event type: branches
Sample = { cpu: 0001 addr: 0xffff80008030cb00 phys_addr:
0x0000000000000000 ip: 0xffff800080313f0c pid: 12720 tid: 12720
period: 1 time: 5986372298040 }
ffff800080313f04 <__perf_event_header__init_id+0x4c>:
ffff800080313f04: 36100094 tbz w20, #2,
ffff800080313f14 <__perf_event_header__init_id+0x5c>
ffff800080313f08: f941e6a0 ldr x0, [x21,
#968]
ffff800080313f0c: d63f0000 blr x0
perf 12720/12720 [0001] 5986.372298040
__perf_event_header__init_id+0x54
.../coresight/linux/kernel/events/core.c 586 return
event->clock();
Event type: branches
Sample = { cpu: 0001 addr: 0xffff8000801bb4a8 phys_addr:
0x0000000000000000 ip: 0xffff80008030cb0c pid: 12720 tid: 12720
period: 1 time: 5986372298040 }
ffff80008030cb00 <local_clock>:
ffff80008030cb00: d503233f paciasp
ffff80008030cb04: a9bf7bfd stp x29, x30,
[sp, #-16]!
ffff80008030cb08: 910003fd mov x29, sp
ffff80008030cb0c: 97faba67 bl ffff8000801bb4a8
<sched_clock>
perf 12720/12720 [0001] 5986.372298040
local_clock+0xc ...t/linux/./include/linux/sched/clock.h 64 return
sched_clock();
Event type: branches
Sample = { cpu: 0001 addr: 0xffff80008125a8a8 phys_addr:
0x0000000000000000 ip: 0xffff8000801bb4c8 pid: 12720 tid: 12720
period: 1 time: 5986372298040 }
ffff8000801bb4a8 <sched_clock>:
ffff8000801bb4a8: d503233f paciasp
ffff8000801bb4ac: a9be7bfd stp x29, x30,
[sp, #-32]!
ffff8000801bb4b0: 910003fd mov x29, sp
ffff8000801bb4b4: a90153f3 stp x19, x20,
[sp, #16]
ffff8000801bb4b8: d5384113 mrs x19, sp_el0
ffff8000801bb4bc: b9401260 ldr w0, [x19, #16]
ffff8000801bb4c0: 11000400 add w0, w0, #0x1
ffff8000801bb4c4: b9001260 str w0, [x19, #16]
ffff8000801bb4c8: 94427cf8 bl ffff80008125a8a8
<sched_clock_noinstr>
perf 12720/12720 [0001] 5986.372298040
sched_clock+0x20 ...sight/linux/kernel/time/sched_clock.c 105 ns =
sched_clock_noinstr();
Event type: branches
Sample = { cpu: 0001 addr: 0xffff80008125b758 phys_addr:
0x0000000000000000 ip: 0xffff80008125a8e4 pid: 12720 tid: 12720
period: 1 time: 5986372298040 }
ffff80008125a8a8 <sched_clock_noinstr>:
ffff80008125a8a8: d503233f paciasp
ffff80008125a8ac: a9bc7bfd stp x29, x30,
[sp, #-64]!
ffff80008125a8b0: 910003fd mov x29, sp
ffff80008125a8b4: a90153f3 stp x19, x20,
[sp, #16]
ffff80008125a8b8: b000e354 adrp x20,
ffff800082ec3000 <tick_bc_dev+0x140>
ffff80008125a8bc: 910d0294 add x20, x20,
#0x340
ffff80008125a8c0: a90363f7 stp x23, x24,
[sp, #48]
ffff80008125a8c4: 91002297 add x23, x20, #0x8
ffff80008125a8c8: 52800518 mov w24, #0x28
// #40
ffff80008125a8cc: a9025bf5 stp x21, x22,
[sp, #32]
ffff80008125a8d0: b9400296 ldr w22, [x20]
ffff80008125a8d4: 120002d5 and w21, w22, #0x1
ffff80008125a8d8: 9bb87eb5 umull x21, w21, w24
ffff80008125a8dc: 8b1502f3 add x19, x23, x21
ffff80008125a8e0: f9400e60 ldr x0, [x19, #24]
ffff80008125a8e4: d63f0000 blr x0
It looks like the disassembly now assumes this BLR wasn't taken. We
go from ffff80008125a8e4 straight through to ...
perf 12720/12720 [0001] 5986.372298040
sched_clock_noinstr+0x3c ...sight/linux/kernel/time/sched_clock.c
93 cyc = (rd->read_sched_clock() - rd->epoch_cyc) &
Event type: branches
Sample = { cpu: 0001 addr: 0xffff8000801bb4cc phys_addr:
0x0000000000000000 ip: 0xffff80008125a930 pid: 12720 tid: 12720
period: 1 time: 5986372298040 }
ffff80008125a8e8 <sched_clock_noinstr+0x40>:
ffff80008125a8e8: f8756ae3 ldr x3, [x23, x21]
ffff80008125a8e4 which is just the previous one +4. Isn't your issue
actually a decode issue in Perf itself? Why is there a discontinuity
without branch samples being generated where either the source or
destination address is 0?
What are your record options to create this issue? As I mentioned in
the previous reply I haven't been able to reproduce it.
I am using below perf record command.
timeout 4s ./perf record -e cs_etm// -C 1 dd if=/dev/zero of=/dev/null
Thanks I managed to reproduce it. I'll take a look to see if I think the
issue is somewhere else.
At least for the failures I encountered, the issue is due to the
alternatives runtime instruction patching mechanism. vmlinux ends up
being the wrong image to decode with because a load of branches are
actually turned into nops.
Can you confirm if you use --kcore instead of vmlinux that you still get
failures:
sudo perf record -e cs_etm// -C 1 --kcore -o <output-folder.data> -- \
dd if=/dev/zero of=/dev/null
perf script -i <output-folder.data> \
tools/perf/scripts/python/arm-cs-trace-disasm.py -d llvm-objdump \
-k <output-folder.data>/kcore_dir/kcore
But I still think bad decode detection should be moved as much as
possible into OpenCSD and Perf rather than this script. Otherwise every
tool will have to re-implement it, and OpenCSD has a lot more info to
make decisions with.
One change we can make is to desynchronize when an N atom is an
unconditional branch:
There's a CPU hardware erratum affecting multiple CPU types and
generations (including Neoverse N1 and V1), where a branch to the
next instruction will be traced as an N atom regardless of whether it's
unconditional, taken conditional, indirect etc. This was detected by
a similar check in one of our other ETM decoders and we root-caused
it to incorrect ETM implementation.
The safe check for current silicon is that it's an unconditional branch
that is direct and whose target is not the next instruction.
You can't infer that an N atom on an unconditional indirect branch is
a synchronization error, since it may have actually branched to the
next instruction, e.g. in a switch-like construction.
Maybe OpenCSD could make the stricter check (as written below)
configurable so you could enable it if you knew for sure that the trace
wasn't affected by this erratum, but that's not a safe default.
Al
That's good to know. In that case it would be better to exclude branches
to the next instruction from the check then rather than make it
configurable. At least for direct branches, that way it "just works".
Indirect looks a bit more complicated because you have to wait for the
address, but I'm sure it can be done one way or another.
diff --git a/decoder/source/etmv4/trc_pkt_decode_etmv4i.cpp
b/decoder/source/etmv4/trc_pkt_decode_etmv4i.cpp
index c557998..3eefd5d 100644
--- a/decoder/source/etmv4/trc_pkt_decode_etmv4i.cpp
+++ b/decoder/source/etmv4/trc_pkt_decode_etmv4i.cpp
@@ -1341,6 +1341,14 @@ ocsd_err_t TrcPktDecodeEtmV4I::processAtom(const
ocsd_atm_val atom)
// save recorded next instuction address
ocsd_vaddr_t nextAddr = m_instr_info.instr_addr;
+ // must have lost sync if an unconditional branch wasn't taken
+ if (atom == ATOM_N && !m_instr_info.is_conditional) {
+ m_need_addr = true;
+ m_out_elem.addElemType(m_index_curr_pkt,
OCSD_GEN_TRC_ELEM_NO_SYNC);
+ // wait for next address
+ return OCSD_OK;
+ }
+
Another one we can spot is when a new address comes that is before the
current decode address (basically the backwards check that you added).
There are probably others that can be spotted like an address appearing
after a direct branch that doesn't match the branch target.
I think at that point, desynchronising should cause the disassembly
script to throw away the last bit, rather than force it to be printed as
in this patch. As I mentioned above in the thread, it leads to printing
disassembly that's implausible and misleading (where an unconditional
branch wasn't taken).
_______________________________________________
CoreSight mailing list -- coresight@xxxxxxxxxxxxxxxx
To unsubscribe send an email to coresight-leave@xxxxxxxxxxxxxxxx