Skip to contentSkip to conversation

perf record: collect user registers set jointly with dwarf stacks

PATCH Superseded Alexey Budankov 2 participants 7 messages May 29, 2019 – May 30, 2019 perf events

v4 of 5 Superseded Latest is v5 →

lore
Share & tools
lore
#0 lore raw

PATCH v4 1/1 perf record: collect user registers set jointly with dwarf stacks

When dwarf stacks are collected jointly with user specified register set using --user-regs option like below the full register context is still captured on a sample:

  $ perf record -g --call-graph dwarf,1024 --user-regs=IP,SP,BP -- stack_test2.g.O3

188143843893585 0x6b48 [0x4f8]: PERF_RECORD_SAMPLE(IP, 0x4002): 23828/23828: 0x401236 period: 1363819 addr: 0x7ffedbdd51ac ... FP chain: nr:0
... user regs: mask 0xff0fff ABI 64-bit

  .... AX    0x53b
  .... BX    0x7ffedbdd3cc0
  .... CX    0xffffffff
  .... DX    0x33d3a
  .... SI    0x7f09b74c38d0
  .... DI    0x0
  .... BP    0x401260
  .... SP    0x7ffedbdd3cc0
  .... IP    0x401236
  .... FLAGS 0x20a
  .... CS    0x33
  .... SS    0x2b
  .... R8    0x7f09b74c3800
  .... R9    0x7f09b74c2da0
  .... R10   0xfffffffffffff3ce
  .... R11   0x246
  .... R12   0x401070
  .... R13   0x7ffedbdd5db0
  .... R14   0x0
  .... R15   0x0

... ustack: size 1024, offset 0xe0
. data_src: 0x5080021
... thread: stack_test2.g.O:23828
...... dso: /root/abudanko/stacks/stack_test2.g.O3

After applying the change suggested in the patch the sample data contain only user specified register values. IP and SP registers (dwarf_regs) are collected anyways regardless of the --user-regs option value provided from the command line:

  -g call-graph dwarf,K                         full_regs
  -g call-graph dwarf,K --user-regs=user_regs	user_regs + dwarf_regs
  --user-regs=user_regs                         user_regs
  $ perf record -g --call-graph dwarf,1024 --user-regs=BP -- ls
  WARNING: specified --user-regs register set doesn't include registers needed by also specified --call-graph=dwarf, auto adding IP, SP registers.
  arch   COPYING	Documentation  include	Kbuild	 lbuild    MAINTAINERS	modules.builtin		 Module.symvers  perf.data.old	scripts   System.map  virt
  block  CREDITS	drivers        init	Kconfig  lib	   Makefile	modules.builtin.modinfo  net		 README		security  tools       vmlinux
  certs  crypto	fs	       ipc	kernel	 LICENSES  mm		modules.order		 perf.data	 samples	sound	  usr	      vmlinux.o
  [ perf record: Woken up 1 times to write data ]
  [ perf record: Captured and wrote 0.030 MB perf.data (10 samples) ]

188368474305373 0x5e40 [0x470]: PERF_RECORD_SAMPLE(IP, 0x4002): 23839/23839: 0x401236 period: 1260507 addr: 0x7ffd3d85e96c ... FP chain: nr:0
... user regs: mask 0x1c0 ABI 64-bit

  .... BP    0x401260
  .... SP    0x7ffd3d85cc20
  .... IP    0x401236

... ustack: size 1024, offset 0x58
. data_src: 0x5080021

---
Changes in v4:
- added warning message about dwarf registers unconditionally included into the collected registers set

Changes in v3:
- avoid changes in platform specific header files

Changes in v2:
- implemented dwarf register set to avoid corrupted trace when --user-regs option value omits IP,SP

---
 tools/perf/util/evsel.c | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)
tools/perf/util/evsel.c +10 -1
index a6f572a40deb..426dfefeecda 100644
@@ -669,6 +669,9 @@ int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 	return ret;
 }
 
+#define DWARF_REGS_MASK ((1ULL << PERF_REG_IP) | \
+			 (1ULL << PERF_REG_SP))
+
 static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
 					   struct record_opts *opts,
 					   struct callchain_param *param)
@@ -702,7 +705,13 @@ static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
 		if (!function) {
 			perf_evsel__set_sample_bit(evsel, REGS_USER);
 			perf_evsel__set_sample_bit(evsel, STACK_USER);
-			attr->sample_regs_user |= PERF_REGS_MASK;
+			if (opts->sample_user_regs) {
+				attr->sample_regs_user |= DWARF_REGS_MASK;
+				pr_warning("WARNING: specified --user-regs register set doesn't include registers "
+					   "needed by also specified --call-graph=dwarf, auto adding IP, SP registers.\n");
+			} else {
+				attr->sample_regs_user |= PERF_REGS_MASK;
+			}
 			attr->sample_stack_user = param->dump_size;
 			attr->exclude_callchain_user = 1;
 		} else {
Signature · 1 line
2.20.1
#1 lore raw

Em Wed, May 29, 2019 at 05:30:49PM +0300, Alexey Budankov escreveu:

Quoted 93 lines

When dwarf stacks are collected jointly with user specified register set using --user-regs option like below the full register context is still captured on a sample:

  $ perf record -g --call-graph dwarf,1024 --user-regs=IP,SP,BP -- stack_test2.g.O3

188143843893585 0x6b48 [0x4f8]: PERF_RECORD_SAMPLE(IP, 0x4002): 23828/23828: 0x401236 period: 1363819 addr: 0x7ffedbdd51ac ... FP chain: nr:0
... user regs: mask 0xff0fff ABI 64-bit

  .... AX    0x53b
  .... BX    0x7ffedbdd3cc0
  .... CX    0xffffffff
  .... DX    0x33d3a
  .... SI    0x7f09b74c38d0
  .... DI    0x0
  .... BP    0x401260
  .... SP    0x7ffedbdd3cc0
  .... IP    0x401236
  .... FLAGS 0x20a
  .... CS    0x33
  .... SS    0x2b
  .... R8    0x7f09b74c3800
  .... R9    0x7f09b74c2da0
  .... R10   0xfffffffffffff3ce
  .... R11   0x246
  .... R12   0x401070
  .... R13   0x7ffedbdd5db0
  .... R14   0x0
  .... R15   0x0

... ustack: size 1024, offset 0xe0
. data_src: 0x5080021
... thread: stack_test2.g.O:23828
...... dso: /root/abudanko/stacks/stack_test2.g.O3

After applying the change suggested in the patch the sample data contain only user specified register values. IP and SP registers (dwarf_regs) are collected anyways regardless of the --user-regs option value provided from the command line:

  -g call-graph dwarf,K                         full_regs
  -g call-graph dwarf,K --user-regs=user_regs	user_regs + dwarf_regs
  --user-regs=user_regs                         user_regs
  $ perf record -g --call-graph dwarf,1024 --user-regs=BP -- ls
  WARNING: specified --user-regs register set doesn't include registers needed by also specified --call-graph=dwarf, auto adding IP, SP registers.
  arch   COPYING	Documentation  include	Kbuild	 lbuild    MAINTAINERS	modules.builtin		 Module.symvers  perf.data.old	scripts   System.map  virt
  block  CREDITS	drivers        init	Kconfig  lib	   Makefile	modules.builtin.modinfo  net		 README		security  tools       vmlinux
  certs  crypto	fs	       ipc	kernel	 LICENSES  mm		modules.order		 perf.data	 samples	sound	  usr	      vmlinux.o
  [ perf record: Woken up 1 times to write data ]
  [ perf record: Captured and wrote 0.030 MB perf.data (10 samples) ]

188368474305373 0x5e40 [0x470]: PERF_RECORD_SAMPLE(IP, 0x4002): 23839/23839: 0x401236 period: 1260507 addr: 0x7ffd3d85e96c ... FP chain: nr:0
... user regs: mask 0x1c0 ABI 64-bit

  .... BP    0x401260
  .... SP    0x7ffd3d85cc20
  .... IP    0x401236

... ustack: size 1024, offset 0x58
. data_src: 0x5080021

---
Changes in v4:
- added warning message about dwarf registers unconditionally included into the collected registers set

Changes in v3:
- avoid changes in platform specific header files

Changes in v2:
- implemented dwarf register set to avoid corrupted trace when --user-regs option value omits IP,SP

---
 tools/perf/util/evsel.c | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)
tools/perf/util/evsel.c +3 -0
index a6f572a40deb..426dfefeecda 100644
@@ -669,6 +669,9 @@ int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 	return ret;
 }
 
+#define DWARF_REGS_MASK ((1ULL << PERF_REG_IP) | \
+			 (1ULL << PERF_REG_SP))
+
 static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
 					   struct record_opts *opts,
 					   struct callchain_param *param)

tools/perf/util/evsel.c static void __perf_evsel__config_callchain(struct perf_evsel *evsel, · line 708

Quoted 5 lines
tools/perf/util/evsel.c +1 -1
@@ -702,7 +705,13 @@ static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
 		if (!function) {
 			perf_evsel__set_sample_bit(evsel, REGS_USER);
 			perf_evsel__set_sample_bit(evsel, STACK_USER);
-			attr->sample_regs_user |= PERF_REGS_MASK;
+			if (opts->sample_user_regs) {

Where are you checking that opts->sample_user_regs doesn't have either IP or SP?

So, __perf_evsel__config_callchain its the routine that sets up the attr->sample_regs_user when callchains are asked for, and what was it doing? Asking for _all_ user regs, right?

I.e. what you're saying is that when --callgraph-dwarf is asked for, then only IP and BP are needed, and we should stop doing that, so that would be a first patch, if that is the case. I.e. a patch that doesn't even mention opts->sample_user_regs.

Then, a second patch would fix the opt->sample_user_regs request clash with --callgraph dwarf, i.e. it would do something like:

	      if ((opts->sample_regs_user & DWARF_REGS_MASK) != DWARF_REGS_MASK) {
	      		char * ip = (opts->sample_regs_user & (1ULL << PERF_REG_IP)) ? NULL : "IP",
	      		     * sp = (opts->sample_regs_user & (1ULL << PERF_REG_SP)) ? NULL : "SP",
			     * all = (!ip && !sp) ?  "s" : "";

			pr_warning("WARNING: specified --user-regs register set doesn't include register%s "
				   "needed by also specified --call-graph=dwarf, auto adding %s%s%s register%s.\n",
				   all, ip, all : ", " : "", sp, all);
		}

This if and only if all the registers that are needed to do DWARF unwinding are just IP and BP, which doesn't look like its true, since when no --user_regs is set (i.e. opts->user_regs is not set) then we continue asking for PERF_REGS_MASK...

Can you check where I'm missing something?

Jiri DWARF unwind uses just IP and SP? Looking at tools/perf/util/unwind-libunwind-local.c's access_reg() I don't think so, right?

- Arnaldo

Quoted 11 lines
tools/perf/util/evsel.c +6 -0
+				attr->sample_regs_user |= DWARF_REGS_MASK;
+				pr_warning("WARNING: specified --user-regs register set doesn't include registers "
+					   "needed by also specified --call-graph=dwarf, auto adding IP, SP registers.\n");
+			} else {
+				attr->sample_regs_user |= PERF_REGS_MASK;
+			}
 			attr->sample_stack_user = param->dump_size;
 			attr->exclude_callchain_user = 1;
 		} else {
Signature · 1 line
2.20.1
Signature · 1 line
- Arnaldo
#2 lore raw
Quoted 2 lines · Arnaldo Carvalho de Melo

Em Wed, May 29, 2019 at 05:30:49PM +0300, Alexey Budankov escreveu:

Quoted 1 line

<SNIP>

Quoted 12 lines
Quoted 12 lines
---
 tools/perf/util/evsel.c | 11 ++++++++++-
 1 file changed, 10 insertions(+), 1 deletion(-)
tools/perf/util/evsel.c +0 -0
index a6f572a40deb..426dfefeecda 100644
@@ -669,6 +669,9 @@ int perf_evsel__group_desc(struct perf_evsel *evsel, char *buf, size_t size)
 	return ret;
 }
 
Quoted 15 lines

tools/perf/util/evsel.c static void __perf_evsel__config_callchain(struct perf_evsel *evsel, · line 708

Quoted 12 lines
tools/perf/util/evsel.c +4 -1
+#define DWARF_REGS_MASK ((1ULL << PERF_REG_IP) | \
+			 (1ULL << PERF_REG_SP))
+
 static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
 					   struct record_opts *opts,
 					   struct callchain_param *param)
@@ -702,7 +705,13 @@ static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
 		if (!function) {
 			perf_evsel__set_sample_bit(evsel, REGS_USER);
 			perf_evsel__set_sample_bit(evsel, STACK_USER);
-			attr->sample_regs_user |= PERF_REGS_MASK;
+			if (opts->sample_user_regs) {

Where are you checking that opts->sample_user_regs doesn't have either IP or SP?

Sure. The the intention was to avoid such a complication, merge two masks and provide explicit warning that the resulting mask is extended.

If you still see the checking and auto detection of the exact mask extension as essential it can be implemented.

Quoted 23 lines

So, __perf_evsel__config_callchain its the routine that sets up the attr->sample_regs_user when callchains are asked for, and what was it doing? Asking for _all_ user regs, right?

I.e. what you're saying is that when --callgraph-dwarf is asked for, then only IP and BP are needed, and we should stop doing that, so that would be a first patch, if that is the case. I.e. a patch that doesn't even mention opts->sample_user_regs.

Then, a second patch would fix the opt->sample_user_regs request clash with --callgraph dwarf, i.e. it would do something like:

	      if ((opts->sample_regs_user & DWARF_REGS_MASK) != DWARF_REGS_MASK) {
	      		char * ip = (opts->sample_regs_user & (1ULL << PERF_REG_IP)) ? NULL : "IP",
	      		     * sp = (opts->sample_regs_user & (1ULL << PERF_REG_SP)) ? NULL : "SP",
			     * all = (!ip && !sp) ?  "s" : "";

			pr_warning("WARNING: specified --user-regs register set doesn't include register%s "
				   "needed by also specified --call-graph=dwarf, auto adding %s%s%s register%s.\n",
				   all, ip, all : ", " : "", sp, all);
		}
Quoted 6 lines

This if and only if all the registers that are needed to do DWARF unwinding are just IP and BP, which doesn't look like its true, since when no --user_regs is set (i.e. opts->user_regs is not set) then we continue asking for PERF_REGS_MASK...

Can you check where I'm missing something?

1.  -g call-graph dwarf,K                         full_regs
2.  --user-regs=user_regs                         user_regs
3.  -g call-graph dwarf,K --user-regs=user_regs	  user_regs + dwarf_regs

The default behavior stays the same for cases 1, 2 above. For case 3 register set becomes the one asked using --user_regs option. If the option value misses IP or SP or the both then they are explicitly added to the option value and a warning message mentioning the exact added registers is provided.

Quoted 4 lines

Jiri DWARF unwind uses just IP and SP? Looking at tools/perf/util/unwind-libunwind-local.c's access_reg() I don't think so, right?

If you ask me, AFAIK, DWARF unwind rules sometimes can refer additional general purpose registers for frames boundaries calculation.

Thanks,
Alexey

Quoted 15 lines

- Arnaldo

Quoted 11 lines
tools/perf/util/evsel.c +6 -0
+				attr->sample_regs_user |= DWARF_REGS_MASK;
+				pr_warning("WARNING: specified --user-regs register set doesn't include registers "
+					   "needed by also specified --call-graph=dwarf, auto adding IP, SP registers.\n");
+			} else {
+				attr->sample_regs_user |= PERF_REGS_MASK;
+			}
 			attr->sample_stack_user = param->dump_size;
 			attr->exclude_callchain_user = 1;
 		} else {
Signature · 1 line
2.20.1
#3 lore raw

Em Thu, May 30, 2019 at 11:24:49AM +0300, Alexey Budankov escreveu:

Quoted 20 lines · Arnaldo Carvalho de Melo
Quoted 1 line

Em Wed, May 29, 2019 at 05:30:49PM +0300, Alexey Budankov escreveu:

<SNIP>

Quoted 8 lines
Quoted 8 lines

+++ b/tools/perf/util/evsel.c

tools/perf/util/evsel.c +3 -0
+#define DWARF_REGS_MASK ((1ULL << PERF_REG_IP) | \
+			 (1ULL << PERF_REG_SP))
+
 static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
 					   struct record_opts *opts,
 					   struct callchain_param *param)
@@ -702,7 +705,13 @@ static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
Quoted 8 lines

tools/perf/util/evsel.c static void __perf_evsel__config_callchain(struct perf_evsel *evsel, · line 708

Quoted 5 lines
tools/perf/util/evsel.c +1 -1
 		if (!function) {
 			perf_evsel__set_sample_bit(evsel, REGS_USER);
 			perf_evsel__set_sample_bit(evsel, STACK_USER);
-			attr->sample_regs_user |= PERF_REGS_MASK;
+			if (opts->sample_user_regs) {

Where are you checking that opts->sample_user_regs doesn't have either IP or SP?

Quoted 2 lines

Sure. The the intention was to avoid such a complication, merge two masks and provide explicit warning that the resulting mask is extended.

s/is/may be/g

Quoted 2 lines

If you still see the checking and auto detection of the exact mask extension as essential it can be implemented.

perf, tracing, systems internals, etc are super complicated, full of details, the more precise we can make the messages, the better.

Quoted 39 lines
Quoted 22 lines

So, __perf_evsel__config_callchain its the routine that sets up the attr->sample_regs_user when callchains are asked for, and what was it doing? Asking for _all_ user regs, right?

I.e. what you're saying is that when --callgraph-dwarf is asked for, then only IP and BP are needed, and we should stop doing that, so that would be a first patch, if that is the case. I.e. a patch that doesn't even mention opts->sample_user_regs.

Then, a second patch would fix the opt->sample_user_regs request clash with --callgraph dwarf, i.e. it would do something like:

	      if ((opts->sample_regs_user & DWARF_REGS_MASK) != DWARF_REGS_MASK) {
	      		char * ip = (opts->sample_regs_user & (1ULL << PERF_REG_IP)) ? NULL : "IP",
	      		     * sp = (opts->sample_regs_user & (1ULL << PERF_REG_SP)) ? NULL : "SP",
			     * all = (!ip && !sp) ?  "s" : "";

			pr_warning("WARNING: specified --user-regs register set doesn't include register%s "
				   "needed by also specified --call-graph=dwarf, auto adding %s%s%s register%s.\n",
				   all, ip, all : ", " : "", sp, all);
		}
Quoted 6 lines

This if and only if all the registers that are needed to do DWARF unwinding are just IP and BP, which doesn't look like its true, since when no --user_regs is set (i.e. opts->user_regs is not set) then we continue asking for PERF_REGS_MASK...

Can you check where I'm missing something?

1.  -g call-graph dwarf,K                         full_regs
2.  --user-regs=user_regs                         user_regs
3.  -g call-graph dwarf,K --user-regs=user_regs	  user_regs + dwarf_regs

The default behavior stays the same for cases 1, 2 above. For case 3 register set becomes the one asked using --user_regs option. If the option value misses IP or SP or the both then they are explicitly added to the option value and a warning message mentioning the exact added registers is provided.

Quoted 6 lines
Quoted 3 lines

Jiri DWARF unwind uses just IP and SP? Looking at tools/perf/util/unwind-libunwind-local.c's access_reg() I don't think so, right?

If you ask me, AFAIK, DWARF unwind rules sometimes can refer additional general purpose registers for frames boundaries calculation.

:-) So that DWARF_REGS is misleading, should be something like DWARF_MINIMAL_REGS, as we may need other registers, so the original code was correct, right?

After all if the user asks for both --call-graph dwarf and --user-regs, then probably we should require --force? I.e. the message then would be:

"
WARNING: The use of --call-graph=dwarf may require all the user registers, specifying a subset with --user-regs may render DWARF unwinding unreliable, please use --force if you're sure that the subset specified via --user-regs is enough for your specific use case. "

And then plain refuse, if the user _really_ wants it, then we have --force/-f for those cases.

Does this sound better?

- Arnaldo

#4 lore raw
Quoted 75 lines · Arnaldo Carvalho de Melo

Em Thu, May 30, 2019 at 11:24:49AM +0300, Alexey Budankov escreveu:

Quoted 20 lines · Arnaldo Carvalho de Melo
Quoted 1 line

Em Wed, May 29, 2019 at 05:30:49PM +0300, Alexey Budankov escreveu:

<SNIP>

Quoted 8 lines
Quoted 8 lines

+++ b/tools/perf/util/evsel.c

tools/perf/util/evsel.c +3 -0
+#define DWARF_REGS_MASK ((1ULL << PERF_REG_IP) | \
+			 (1ULL << PERF_REG_SP))
+
 static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
 					   struct record_opts *opts,
 					   struct callchain_param *param)
@@ -702,7 +705,13 @@ static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
Quoted 8 lines

tools/perf/util/evsel.c static void __perf_evsel__config_callchain(struct perf_evsel *evsel, · line 708

Quoted 5 lines
tools/perf/util/evsel.c +1 -1
 		if (!function) {
 			perf_evsel__set_sample_bit(evsel, REGS_USER);
 			perf_evsel__set_sample_bit(evsel, STACK_USER);
-			attr->sample_regs_user |= PERF_REGS_MASK;
+			if (opts->sample_user_regs) {

Where are you checking that opts->sample_user_regs doesn't have either IP or SP?

Quoted 2 lines

Sure. The the intention was to avoid such a complication, merge two masks and provide explicit warning that the resulting mask is extended.

s/is/may be/g

Quoted 2 lines

If you still see the checking and auto detection of the exact mask extension as essential it can be implemented.

perf, tracing, systems internals, etc are super complicated, full of details, the more precise we can make the messages, the better.

Quoted 42 lines
Quoted 22 lines

So, __perf_evsel__config_callchain its the routine that sets up the attr->sample_regs_user when callchains are asked for, and what was it doing? Asking for _all_ user regs, right?

I.e. what you're saying is that when --callgraph-dwarf is asked for, then only IP and BP are needed, and we should stop doing that, so that would be a first patch, if that is the case. I.e. a patch that doesn't even mention opts->sample_user_regs.

Then, a second patch would fix the opt->sample_user_regs request clash with --callgraph dwarf, i.e. it would do something like:

	      if ((opts->sample_regs_user & DWARF_REGS_MASK) != DWARF_REGS_MASK) {
	      		char * ip = (opts->sample_regs_user & (1ULL << PERF_REG_IP)) ? NULL : "IP",
	      		     * sp = (opts->sample_regs_user & (1ULL << PERF_REG_SP)) ? NULL : "SP",
			     * all = (!ip && !sp) ?  "s" : "";

			pr_warning("WARNING: specified --user-regs register set doesn't include register%s "
				   "needed by also specified --call-graph=dwarf, auto adding %s%s%s register%s.\n",
				   all, ip, all : ", " : "", sp, all);
		}
Quoted 6 lines

This if and only if all the registers that are needed to do DWARF unwinding are just IP and BP, which doesn't look like its true, since when no --user_regs is set (i.e. opts->user_regs is not set) then we continue asking for PERF_REGS_MASK...

Can you check where I'm missing something?

1.  -g call-graph dwarf,K                         full_regs
2.  --user-regs=user_regs                         user_regs
3.  -g call-graph dwarf,K --user-regs=user_regs	  user_regs + dwarf_regs

The default behavior stays the same for cases 1, 2 above. For case 3 register set becomes the one asked using --user_regs option. If the option value misses IP or SP or the both then they are explicitly added to the option value and a warning message mentioning the exact added registers is provided.

Quoted 3 lines

Jiri DWARF unwind uses just IP and SP? Looking at tools/perf/util/unwind-libunwind-local.c's access_reg() I don't think so, right?

Quoted 6 lines
Quoted 2 lines

If you ask me, AFAIK, DWARF unwind rules sometimes can refer additional general purpose registers for frames boundaries calculation.

:-) So that DWARF_REGS is misleading, should be something like DWARF_MINIMAL_REGS, as we may need other registers, so the original code was correct, right?

Right. Actually came to the same conclusion with the same naming for IP,SP mask :)

Quoted 11 lines

After all if the user asks for both --call-graph dwarf and --user-regs, then probably we should require --force? I.e. the message then would be:

"
WARNING: The use of --call-graph=dwarf may require all the user registers, specifying a subset with --user-regs may render DWARF unwinding unreliable, please use --force if you're sure that the subset specified via --user-regs is enough for your specific use case. "

Quoted 4 lines

And then plain refuse, if the user _really_ wants it, then we have --force/-f for those cases.

Does this sound better?

If --user-regs is specified jointly with --call-graph dwarf option then --user-regs already serves as the --force and, IMHO, a warning does the best.

The ideal solution, I could imagine, is to also dynamically calculate regs set extension and provide it in the warning, but it is only for two registers.

So, if --call-graph dwarf --user-regs=A,B,C are specified jointly then "
WARNING: The use of --call-graph=dwarf may require all the user registers, specifying a subset with --user-regs may render DWARF unwinding unreliable, so the minimal registers set (IP, SP) is explicitly forced.
"
The message is precise and it would fit the majority of use cases. Final decision is up to you.

~Alexey

Quoted 3 lines

- Arnaldo

#5 lore raw

Em Thu, May 30, 2019 at 07:24:57PM +0300, Alexey Budankov escreveu:

Quoted 108 lines · Arnaldo Carvalho de Melo
Quoted 75 lines

Em Thu, May 30, 2019 at 11:24:49AM +0300, Alexey Budankov escreveu:

Quoted 20 lines · Arnaldo Carvalho de Melo
Quoted 1 line

Em Wed, May 29, 2019 at 05:30:49PM +0300, Alexey Budankov escreveu:

<SNIP>

Quoted 8 lines
Quoted 8 lines

+++ b/tools/perf/util/evsel.c

tools/perf/util/evsel.c +3 -0
+#define DWARF_REGS_MASK ((1ULL << PERF_REG_IP) | \
+			 (1ULL << PERF_REG_SP))
+
 static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
 					   struct record_opts *opts,
 					   struct callchain_param *param)
@@ -702,7 +705,13 @@ static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
Quoted 8 lines

tools/perf/util/evsel.c static void __perf_evsel__config_callchain(struct perf_evsel *evsel, · line 708

Quoted 5 lines
tools/perf/util/evsel.c +1 -1
 		if (!function) {
 			perf_evsel__set_sample_bit(evsel, REGS_USER);
 			perf_evsel__set_sample_bit(evsel, STACK_USER);
-			attr->sample_regs_user |= PERF_REGS_MASK;
+			if (opts->sample_user_regs) {

Where are you checking that opts->sample_user_regs doesn't have either IP or SP?

Quoted 2 lines

Sure. The the intention was to avoid such a complication, merge two masks and provide explicit warning that the resulting mask is extended.

s/is/may be/g

Quoted 2 lines

If you still see the checking and auto detection of the exact mask extension as essential it can be implemented.

perf, tracing, systems internals, etc are super complicated, full of details, the more precise we can make the messages, the better.

Quoted 42 lines
Quoted 22 lines

So, __perf_evsel__config_callchain its the routine that sets up the attr->sample_regs_user when callchains are asked for, and what was it doing? Asking for _all_ user regs, right?

I.e. what you're saying is that when --callgraph-dwarf is asked for, then only IP and BP are needed, and we should stop doing that, so that would be a first patch, if that is the case. I.e. a patch that doesn't even mention opts->sample_user_regs.

Then, a second patch would fix the opt->sample_user_regs request clash with --callgraph dwarf, i.e. it would do something like:

	      if ((opts->sample_regs_user & DWARF_REGS_MASK) != DWARF_REGS_MASK) {
	      		char * ip = (opts->sample_regs_user & (1ULL << PERF_REG_IP)) ? NULL : "IP",
	      		     * sp = (opts->sample_regs_user & (1ULL << PERF_REG_SP)) ? NULL : "SP",
			     * all = (!ip && !sp) ?  "s" : "";

			pr_warning("WARNING: specified --user-regs register set doesn't include register%s "
				   "needed by also specified --call-graph=dwarf, auto adding %s%s%s register%s.\n",
				   all, ip, all : ", " : "", sp, all);
		}
Quoted 6 lines

This if and only if all the registers that are needed to do DWARF unwinding are just IP and BP, which doesn't look like its true, since when no --user_regs is set (i.e. opts->user_regs is not set) then we continue asking for PERF_REGS_MASK...

Can you check where I'm missing something?

1.  -g call-graph dwarf,K                         full_regs
2.  --user-regs=user_regs                         user_regs
3.  -g call-graph dwarf,K --user-regs=user_regs	  user_regs + dwarf_regs

The default behavior stays the same for cases 1, 2 above. For case 3 register set becomes the one asked using --user_regs option. If the option value misses IP or SP or the both then they are explicitly added to the option value and a warning message mentioning the exact added registers is provided.

Quoted 3 lines

Jiri DWARF unwind uses just IP and SP? Looking at tools/perf/util/unwind-libunwind-local.c's access_reg() I don't think so, right?

Quoted 6 lines
Quoted 2 lines

If you ask me, AFAIK, DWARF unwind rules sometimes can refer additional general purpose registers for frames boundaries calculation.

:-) So that DWARF_REGS is misleading, should be something like DWARF_MINIMAL_REGS, as we may need other registers, so the original code was correct, right?

Right. Actually came to the same conclusion with the same naming for IP,SP mask :)

Quoted 11 lines

After all if the user asks for both --call-graph dwarf and --user-regs, then probably we should require --force? I.e. the message then would be:

"
WARNING: The use of --call-graph=dwarf may require all the user registers, specifying a subset with --user-regs may render DWARF unwinding unreliable, please use --force if you're sure that the subset specified via --user-regs is enough for your specific use case. "

Quoted 4 lines

And then plain refuse, if the user _really_ wants it, then we have --force/-f for those cases.

Does this sound better?

If --user-regs is specified jointly with --call-graph dwarf option then --user-regs already serves as the --force and, IMHO, a warning does the best.

The ideal solution, I could imagine, is to also dynamically calculate regs set extension and provide it in the warning, but it is only for two registers.

Quoted 6 lines

So, if --call-graph dwarf --user-regs=A,B,C are specified jointly then "
WARNING: The use of --call-graph=dwarf may require all the user registers, specifying a subset with --user-regs may render DWARF unwinding unreliable, so the minimal registers set (IP, SP) is explicitly forced.
"

I think with this wording and the renaming of DWARF_REGS to DWARF_MINIMAL_REGS it should be enough.

- Arnaldo

#6 lore raw
Quoted 116 lines · Arnaldo Carvalho de Melo

Em Thu, May 30, 2019 at 07:24:57PM +0300, Alexey Budankov escreveu:

Quoted 108 lines · Arnaldo Carvalho de Melo
Quoted 75 lines

Em Thu, May 30, 2019 at 11:24:49AM +0300, Alexey Budankov escreveu:

Quoted 20 lines · Arnaldo Carvalho de Melo
Quoted 1 line

Em Wed, May 29, 2019 at 05:30:49PM +0300, Alexey Budankov escreveu:

<SNIP>

Quoted 8 lines
Quoted 8 lines

+++ b/tools/perf/util/evsel.c

tools/perf/util/evsel.c +3 -0
+#define DWARF_REGS_MASK ((1ULL << PERF_REG_IP) | \
+			 (1ULL << PERF_REG_SP))
+
 static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
 					   struct record_opts *opts,
 					   struct callchain_param *param)
@@ -702,7 +705,13 @@ static void __perf_evsel__config_callchain(struct perf_evsel *evsel,
Quoted 8 lines

tools/perf/util/evsel.c static void __perf_evsel__config_callchain(struct perf_evsel *evsel, · line 708

Quoted 5 lines
tools/perf/util/evsel.c +1 -1
 		if (!function) {
 			perf_evsel__set_sample_bit(evsel, REGS_USER);
 			perf_evsel__set_sample_bit(evsel, STACK_USER);
-			attr->sample_regs_user |= PERF_REGS_MASK;
+			if (opts->sample_user_regs) {

Where are you checking that opts->sample_user_regs doesn't have either IP or SP?

Quoted 2 lines

Sure. The the intention was to avoid such a complication, merge two masks and provide explicit warning that the resulting mask is extended.

s/is/may be/g

Quoted 2 lines

If you still see the checking and auto detection of the exact mask extension as essential it can be implemented.

perf, tracing, systems internals, etc are super complicated, full of details, the more precise we can make the messages, the better.

Quoted 42 lines
Quoted 22 lines

So, __perf_evsel__config_callchain its the routine that sets up the attr->sample_regs_user when callchains are asked for, and what was it doing? Asking for _all_ user regs, right?

I.e. what you're saying is that when --callgraph-dwarf is asked for, then only IP and BP are needed, and we should stop doing that, so that would be a first patch, if that is the case. I.e. a patch that doesn't even mention opts->sample_user_regs.

Then, a second patch would fix the opt->sample_user_regs request clash with --callgraph dwarf, i.e. it would do something like:

	      if ((opts->sample_regs_user & DWARF_REGS_MASK) != DWARF_REGS_MASK) {
	      		char * ip = (opts->sample_regs_user & (1ULL << PERF_REG_IP)) ? NULL : "IP",
	      		     * sp = (opts->sample_regs_user & (1ULL << PERF_REG_SP)) ? NULL : "SP",
			     * all = (!ip && !sp) ?  "s" : "";

			pr_warning("WARNING: specified --user-regs register set doesn't include register%s "
				   "needed by also specified --call-graph=dwarf, auto adding %s%s%s register%s.\n",
				   all, ip, all : ", " : "", sp, all);
		}
Quoted 6 lines

This if and only if all the registers that are needed to do DWARF unwinding are just IP and BP, which doesn't look like its true, since when no --user_regs is set (i.e. opts->user_regs is not set) then we continue asking for PERF_REGS_MASK...

Can you check where I'm missing something?

1.  -g call-graph dwarf,K                         full_regs
2.  --user-regs=user_regs                         user_regs
3.  -g call-graph dwarf,K --user-regs=user_regs	  user_regs + dwarf_regs

The default behavior stays the same for cases 1, 2 above. For case 3 register set becomes the one asked using --user_regs option. If the option value misses IP or SP or the both then they are explicitly added to the option value and a warning message mentioning the exact added registers is provided.

Quoted 3 lines

Jiri DWARF unwind uses just IP and SP? Looking at tools/perf/util/unwind-libunwind-local.c's access_reg() I don't think so, right?

Quoted 6 lines
Quoted 2 lines

If you ask me, AFAIK, DWARF unwind rules sometimes can refer additional general purpose registers for frames boundaries calculation.

:-) So that DWARF_REGS is misleading, should be something like DWARF_MINIMAL_REGS, as we may need other registers, so the original code was correct, right?

Right. Actually came to the same conclusion with the same naming for IP,SP mask :)

Quoted 11 lines

After all if the user asks for both --call-graph dwarf and --user-regs, then probably we should require --force? I.e. the message then would be:

"
WARNING: The use of --call-graph=dwarf may require all the user registers, specifying a subset with --user-regs may render DWARF unwinding unreliable, please use --force if you're sure that the subset specified via --user-regs is enough for your specific use case. "

Quoted 4 lines

And then plain refuse, if the user _really_ wants it, then we have --force/-f for those cases.

Does this sound better?

If --user-regs is specified jointly with --call-graph dwarf option then --user-regs already serves as the --force and, IMHO, a warning does the best.

The ideal solution, I could imagine, is to also dynamically calculate regs set extension and provide it in the warning, but it is only for two registers.

Quoted 6 lines

So, if --call-graph dwarf --user-regs=A,B,C are specified jointly then "
WARNING: The use of --call-graph=dwarf may require all the user registers, specifying a subset with --user-regs may render DWARF unwinding unreliable, so the minimal registers set (IP, SP) is explicitly forced.
"

Quoted 2 lines

I think with this wording and the renaming of DWARF_REGS to DWARF_MINIMAL_REGS it should be enough.

Well, let's have it like this in v5.

Thanks,
Alexey

Quoted 3 lines

- Arnaldo