About using PEBS from the user space

Harald Servat

author

2015-05-03 22:05 UTC #0 lore raw

Dear list,

I'd like to use the perf library to access PEBS so as to collect
referenced memory addresses from the user space. I think I've successfully setup the perf structures (struct perf_event_attr) to configure the performance counters, but I don't see what should I do to access to the captured memory addresses. I've seen that within arch/x86/kernel/cpu/perf_event_intel_ds.c there are the routines alloc_pebs_buffer, alloc_ds_buffer which seems to allocate and setup the necessary buffers using kmalloc_node calls. Question is, how can replicate this from the user space? And how we should connect these buffers to the PEBS infrastructure using perf calls?

Thank you very much in advance.

WARNING / LEGAL TEXT: This message is intended only for the use of the individual or entity to which it is addressed and may contain information which is privileged, confidential, proprietary, or exempt from disclosure under applicable law. If you are not the intended recipient or the person responsible for delivering the message to the intended recipient, you are strictly prohibited from disclosing, distributing, copying, or in any way using this message. If you have received this communication in error, please notify the sender and destroy and delete any copies you may have received.

http://www.bsc.es/disclaimer

Vince Weaver

2015-05-03 22:14 UTC #1 lore raw

Quoted 11 lines · Harald Servat

Dear list,
I'd like to use the perf library to access PEBS so as to collect referenced
memory addresses from the user space. I think I've successfully setup the perf structures (struct perf_event_attr) to configure the performance counters, but I don't see what should I do to access to the captured memory addresses. I've seen that within arch/x86/kernel/cpu/perf_event_intel_ds.c there are the routines alloc_pebs_buffer, alloc_ds_buffer which seems to allocate and setup the necessary buffers using kmalloc_node calls. Question is, how can replicate this from the user space? And how we should connect these buffers to the PEBS infrastructure using perf calls?

You can try looking at the example code in my perf_event_tests code.
https://github.com/deater/perf_event_tests

The stuff you are looking for is probably covered in the
test/record_sample/samples_data_src test/record_sample/sample_weight
and especially the
test/record_sample/sample_regs_intr
tests, although that last one requires a fairly recent kernel to work.

Vince

Harald Servat

author

2015-05-04 08:13 UTC #2 lore raw

Quoted 17 lines · Vince Weaver

Quoted 11 lines · Harald Servat
Dear list,
I'd like to use the perf library to access PEBS so as to collect referenced
memory addresses from the user space. I think I've successfully setup the perf structures (struct perf_event_attr) to configure the performance counters, but I don't see what should I do to access to the captured memory addresses. I've seen that within arch/x86/kernel/cpu/perf_event_intel_ds.c there are the routines alloc_pebs_buffer, alloc_ds_buffer which seems to allocate and setup the necessary buffers using kmalloc_node calls. Question is, how can replicate this from the user space? And how we should connect these buffers to the PEBS infrastructure using perf calls?
You can try looking at the example code in my perf_event_tests code.
https://github.com/deater/perf_event_tests

Quoted 8 lines

The stuff you are looking for is probably covered in the
test/record_sample/samples_data_src test/record_sample/sample_weight
and especially the
test/record_sample/sample_regs_intr
tests, although that last one requires a fairly recent kernel to work.
Vince

Thank you very much for these pointers Vince. I'll take a look at them!

Best regards.

WARNING / LEGAL TEXT: This message is intended only for the use of the individual or entity to which it is addressed and may contain information which is privileged, confidential, proprietary, or exempt from disclosure under applicable law. If you are not the intended recipient or the person responsible for delivering the message to the intended recipient, you are strictly prohibited from disclosing, distributing, copying, or in any way using this message. If you have received this communication in error, please notify the sender and destroy and delete any copies you may have received.

http://www.bsc.es/disclaimer

Harald Servat

author

2015-05-04 10:17 UTC #3 lore raw

In reply to Vince Weaver

Quoted 17 lines · Vince Weaver

Quoted 11 lines · Harald Servat
Dear list,
I'd like to use the perf library to access PEBS so as to collect referenced
memory addresses from the user space. I think I've successfully setup the perf structures (struct perf_event_attr) to configure the performance counters, but I don't see what should I do to access to the captured memory addresses. I've seen that within arch/x86/kernel/cpu/perf_event_intel_ds.c there are the routines alloc_pebs_buffer, alloc_ds_buffer which seems to allocate and setup the necessary buffers using kmalloc_node calls. Question is, how can replicate this from the user space? And how we should connect these buffers to the PEBS infrastructure using perf calls?
You can try looking at the example code in my perf_event_tests code.
https://github.com/deater/perf_event_tests

Quoted 8 lines

The stuff you are looking for is probably covered in the
test/record_sample/samples_data_src test/record_sample/sample_weight
and especially the
test/record_sample/sample_regs_intr
tests, although that last one requires a fairly recent kernel to work.
Vince

Vince,

that is really interesting work, thank you! However, for some reason
the PEBS collected address is never shown (and maybe not captured) according to the following partial output

PERF_RECORD_SAMPLE [c001], MISC=16386 (PERF_RECORD_MISC_USER,PERF_RECORD_MISC_EXACT_IP ), Size=32

	PERF_SAMPLE_IP, IP: 404510
	PERF_SAMPLE_WEIGHT, Weight: 30
	PERF_SAMPLE_DATA_SRC, Raw: 68100242
		Load Hit Line fill buffer No snoop Hit Level 1 TLB Level 2 TLB
PERF_RECORD_SAMPLE [c001], MISC=16386 (PERF_RECORD_MISC_USER,PERF_RECORD_MISC_EXACT_IP ), Size=32
	PERF_SAMPLE_IP, IP: 404524
	PERF_SAMPLE_WEIGHT, Weight: 76
	PERF_SAMPLE_DATA_SRC, Raw: 68100442
		Load Hit L2 cache No snoop Hit Level 1 TLB Level 2 TLB
PERF_RECORD_SAMPLE [c001], MISC=16386 (PERF_RECORD_MISC_USER,PERF_RECORD_MISC_EXACT_IP ), Size=32
	PERF_SAMPLE_IP, IP: 404524
	PERF_SAMPLE_WEIGHT, Weight: 77
	PERF_SAMPLE_DATA_SRC, Raw: 68100842
		Load Hit L3 cache No snoop Hit Level 1 TLB Level 2 TLB

I've looked into your code in order to compare it with the perf tool
and it looks like you support it in the code in the following lines of lib/parse_record.c, so it looks that sample_type is not tagged with PERF_SAMPLE_ADDR bit.

511 if (sample_type & PERF_SAMPLE_ADDR) {

512                 long long addr;
513                 memcpy(&addr,&data[offset],sizeof(long long));
514                 if (!quiet) printf("\tPERF_SAMPLE_ADDR, addr: %llx\n",addr);
515                 offset+=8;

516 }

I tweaked a bit the perf dump application so that it prints the
sample_type and it always shows c10f whereas your binary prints c001. If I simply go to one of the tests (say sample_data_src.c) and force pe.sample = 0xc10f then it works (printing additional data). So, would the following change make sense into your code to emit the sampled address?

Best regards.

WARNING / LEGAL TEXT: This message is intended only for the use of the individual or entity to which it is addressed and may contain information which is privileged, confidential, proprietary, or exempt from disclosure under applicable law. If you are not the intended recipient or the person responsible for delivering the message to the intended recipient, you are strictly prohibited from disclosing, distributing, copying, or in any way using this message. If you have received this communication in error, please notify the sender and destroy and delete any copies you may have received.

http://www.bsc.es/disclaimer

tests/record_sample/sample_data_src.c +1 -1

index 1c3d260..35c7e81 100644
@@ -113,7 +113,7 @@ int main(int argc, char **argv) {
         pe.size=sizeof(struct perf_event_attr);
         pe.sample_period=SAMPLE_FREQUENCY;
         pe.sample_type=PERF_SAMPLE_IP | PERF_SAMPLE_WEIGHT |
-			PERF_SAMPLE_DATA_SRC;
+			PERF_SAMPLE_DATA_SRC | PERF_SAMPLE_ADDR ;
 
 	global_sample_type=pe.sample_type;

Harald Servat

author

2015-06-25 16:06 UTC #4 lore raw

In reply to Vince Weaver

Quoted 17 lines · Vince Weaver

Quoted 11 lines · Harald Servat
Dear list,
I'd like to use the perf library to access PEBS so as to collect referenced
memory addresses from the user space. I think I've successfully setup the perf structures (struct perf_event_attr) to configure the performance counters, but I don't see what should I do to access to the captured memory addresses. I've seen that within arch/x86/kernel/cpu/perf_event_intel_ds.c there are the routines alloc_pebs_buffer, alloc_ds_buffer which seems to allocate and setup the necessary buffers using kmalloc_node calls. Question is, how can replicate this from the user space? And how we should connect these buffers to the PEBS infrastructure using perf calls?
You can try looking at the example code in my perf_event_tests code.
https://github.com/deater/perf_event_tests

Quoted 8 lines

The stuff you are looking for is probably covered in the
test/record_sample/samples_data_src test/record_sample/sample_weight
and especially the
test/record_sample/sample_regs_intr
tests, although that last one requires a fairly recent kernel to work.
Vince

Hello,

is there a way to setup PEBS to collect information from loads that
take at least X cycles to complete using MSR_PEBS_LD_LAT_THRESHOLD through this approach?

Thank you!

WARNING / LEGAL TEXT: This message is intended only for the use of the individual or entity to which it is addressed and may contain information which is privileged, confidential, proprietary, or exempt from disclosure under applicable law. If you are not the intended recipient or the person responsible for delivering the message to the intended recipient, you are strictly prohibited from disclosing, distributing, copying, or in any way using this message. If you have received this communication in error, please notify the sender and destroy and delete any copies you may have received.

http://www.bsc.es/disclaimer

Manuel Selva

2015-06-25 19:06 UTC #5 lore raw

Hello Harald,

Setting the config1 field of the perf_event_attr field (as done by the perf mem record tool) to X will instruct the PMU to only sample loads with latency higher than X. Is it what you are looking for ?

----
Manu

Quoted 60 lines · Harald Servat

Quoted 24 lines · Vince Weaver
Quoted 18 lines · Harald Servat
Dear list,
I'd like to use the perf library to access PEBS so as to collect
referenced
memory addresses from the user space. I think I've successfully setup the perf
structures (struct perf_event_attr) to configure the performance counters, but
I don't see what should I do to access to the captured memory addresses. I've
seen that within arch/x86/kernel/cpu/perf_event_intel_ds.c there are the routines alloc_pebs_buffer, alloc_ds_buffer which seems to allocate and setup
the necessary buffers using kmalloc_node calls. Question is, how can replicate
this from the user space? And how we should connect these buffers to the PEBS
infrastructure using perf calls?
You can try looking at the example code in my perf_event_tests code.
https://github.com/deater/perf_event_tests
Quoted 8 lines
The stuff you are looking for is probably covered in the
test/record_sample/samples_data_src test/record_sample/sample_weight
and especially the
test/record_sample/sample_regs_intr
tests, although that last one requires a fairly recent kernel to work.
Vince
Hello,
is there a way to setup PEBS to collect information from loads that
take at least X cycles to complete using MSR_PEBS_LD_LAT_THRESHOLD through this approach?
Thank you!
WARNING / LEGAL TEXT: This message is intended only for the use of the individual or entity to which it is addressed and may contain information which is privileged, confidential, proprietary, or exempt from disclosure under applicable law. If you are not the intended recipient or the person responsible for delivering the message to the intended recipient, you are strictly prohibited from disclosing, distributing, copying, or in any way using this message. If you have received this communication in error, please notify the sender and destroy and delete any copies you may have received.
http://www.bsc.es/disclaimer
--
To unsubscribe from this list: send the line "unsubscribe linux-perf-users" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

Harald Servat

author

2015-06-26 08:13 UTC #6 lore raw

Ah! Great, thank you Manuel!!

Quoted 71 lines · Manuel Selva

Hello Harald,
Setting the config1 field of the perf_event_attr field (as done by the perf mem record tool) to X will instruct the PMU to only sample loads with latency higher than X. Is it what you are looking for ?
----
Manu
Quoted 61 lines · Harald Servat
Quoted 25 lines · Vince Weaver
Quoted 11 lines · Harald Servat
Dear list,
I'd like to use the perf library to access PEBS so as to collect
referenced
memory addresses from the user space. I think I've successfully setup the perf
structures (struct perf_event_attr) to configure the performance counters, but
I don't see what should I do to access to the captured memory addresses. I've
seen that within arch/x86/kernel/cpu/perf_event_intel_ds.c there are
Quoted 8 lines
the
routines alloc_pebs_buffer, alloc_ds_buffer which seems to allocate and setup
the necessary buffers using kmalloc_node calls. Question is, how can replicate
this from the user space? And how we should connect these buffers to the PEBS
infrastructure using perf calls?
You can try looking at the example code in my perf_event_tests code.
https://github.com/deater/perf_event_tests
Quoted 8 lines
The stuff you are looking for is probably covered in the
test/record_sample/samples_data_src test/record_sample/sample_weight
and especially the
test/record_sample/sample_regs_intr
tests, although that last one requires a fairly recent kernel to work.
Vince
Hello,
is there a way to setup PEBS to collect information from loads that
take at least X cycles to complete using MSR_PEBS_LD_LAT_THRESHOLD through this approach?
Thank you!
WARNING / LEGAL TEXT: This message is intended only for the use of the individual or entity to which it is addressed and may contain information which is privileged, confidential, proprietary, or exempt from disclosure under applicable law. If you are not the intended recipient or the person responsible for delivering the message to the intended recipient, you are strictly prohibited from disclosing, distributing, copying, or in any way using this message. If you have received this communication in error, please notify the sender and destroy and delete any copies you may have received.
http://www.bsc.es/disclaimer
--
To unsubscribe from this list: send the line "unsubscribe linux-perf-users" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

WARNING / LEGAL TEXT: This message is intended only for the use of the individual or entity to which it is addressed and may contain information which is privileged, confidential, proprietary, or exempt from disclosure under applicable law. If you are not the intended recipient or the person responsible for delivering the message to the intended recipient, you are strictly prohibited from disclosing, distributing, copying, or in any way using this message. If you have received this communication in error, please notify the sender and destroy and delete any copies you may have received.

http://www.bsc.es/disclaimer