v2->v3:
- Fix the commit log in the second patch which causes "git am" to fail.
- Remove the 1st patch in v2 as the problem will be fixed by another
patch from Arnaldo.
v1->v2:
- Include a compilation fix patch and a code streamlining patch
into the patch set.
- Use the __stringify() macro in stringify.h instead of adding a
duplicate macro.
- Add the --max-stack option to perf-top as well.
This perf patch set contains the following changes:
Patch 1 - Streamline the append_chain() function to make it run a bit
faster.
Patch 2 - Add a --max-stack option to perf-report to speed up its
processing at the expense of less backtrace information
available.
Patch 3 - Add a similar --max-stack option to perf-top.
Waiman Long (3):
perf: streamline append_chain() function
perf-report: add --max-stack option to limit callchain stack scan
perf-top: add --max-stack option to limit callchain stack scan
tools/perf/Documentation/perf-report.txt | 8 ++++++++
tools/perf/Documentation/perf-top.txt | 8 ++++++++
tools/perf/builtin-report.c | 22 +++++++++++++++++-----
tools/perf/builtin-top.c | 9 +++++++--
tools/perf/util/callchain.c | 9 ++++-----
tools/perf/util/machine.c | 14 +++++++++-----
tools/perf/util/machine.h | 3 ++-
tools/perf/util/session.c | 3 ++-
tools/perf/util/top.h | 1 +
9 files changed, 58 insertions(+), 19 deletions(-)
When the callgraph function is enabled (-G), it may take a long time to
scan all the stack data and merge them accordingly.
This patch adds a new --max-stack option to perf-top to limit the depth
of callchain stack data to look at to reduce the time it takes for
perf-top to finish its processing. It reduces the amount of information
provided to the user in exchange for faster speed.
Signed-off-by: Waiman Long <[email protected]>
Acked-by: David Ahern <[email protected]>
Tested-by: Davidlohr Bueso <[email protected]>
---
tools/perf/Documentation/perf-top.txt | 8 ++++++++
tools/perf/builtin-top.c | 8 ++++++--
tools/perf/util/top.h | 1 +
3 files changed, 15 insertions(+), 2 deletions(-)
diff --git a/tools/perf/Documentation/perf-top.txt b/tools/perf/Documentation/perf-top.txt
index 58d6598..3fd911c 100644
--- a/tools/perf/Documentation/perf-top.txt
+++ b/tools/perf/Documentation/perf-top.txt
@@ -155,6 +155,14 @@ Default is to monitor all CPUS.
Default: fractal,0.5,callee.
+--max-stack::
+ Set the stack depth limit when parsing the callchain, anything
+ beyond the specified depth will be ignored. This is a trade-off
+ between information loss and faster processing especially for
+ workloads that can have a very long callchain stack.
+
+ Default: 127
+
--ignore-callees=<regex>::
Ignore callees of the function(s) matching the given regex.
This has the effect of collecting the callers of each such
diff --git a/tools/perf/builtin-top.c b/tools/perf/builtin-top.c
index 2725aca..14902b0 100644
--- a/tools/perf/builtin-top.c
+++ b/tools/perf/builtin-top.c
@@ -772,7 +772,7 @@ static void perf_event__process_sample(struct perf_tool *tool,
err = machine__resolve_callchain(machine, evsel,
al.thread, sample,
&parent, &al,
- PERF_MAX_STACK_DEPTH);
+ top->max_stack);
if (err)
return;
}
@@ -1052,10 +1052,11 @@ int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
.user_freq = UINT_MAX,
.user_interval = ULLONG_MAX,
.freq = 4000, /* 4 KHz */
- .target = {
+ .target = {
.uses_mmap = true,
},
},
+ .max_stack = PERF_MAX_STACK_DEPTH,
.sym_pcnt_filter = 5,
};
struct perf_record_opts *opts = &top.record_opts;
@@ -1110,6 +1111,9 @@ int cmd_top(int argc, const char **argv, const char *prefix __maybe_unused)
OPT_CALLBACK_DEFAULT('G', "call-graph", &top.record_opts,
"mode[,dump_size]", record_callchain_help,
&parse_callchain_opt, "fp"),
+ OPT_INTEGER(0, "max-stack", &top.max_stack,
+ "Set the maximum stack depth when parsing the callchain. "
+ "Default: " __stringify(PERF_MAX_STACK_DEPTH)),
OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
"ignore callees of these functions in call graphs",
report_parse_ignore_callees_opt),
diff --git a/tools/perf/util/top.h b/tools/perf/util/top.h
index b554ffc..88cfeaf 100644
--- a/tools/perf/util/top.h
+++ b/tools/perf/util/top.h
@@ -24,6 +24,7 @@ struct perf_top {
u64 exact_samples;
u64 guest_us_samples, guest_kernel_samples;
int print_entries, count_filter, delay_secs;
+ int max_stack;
bool hide_kernel_symbols, hide_user_symbols, zero;
bool use_tui, use_stdio;
bool kptr_restrict_warned;
--
1.7.1
When callgraph data was included in the perf data file, it may take a
long time to scan all those data and merge them together especially
if the stored callchains are long and the perf data file itself is
large, like a Gbyte or so.
The callchain stack is currently limited to PERF_MAX_STACK_DEPTH (127).
This is a large value. Usually the callgraph data that developers are
most interested in are the first few levels, the rests are usually
not looked at.
This patch adds a new --max-stack option to perf-report to limit the
depth of callchain stack data to look at to reduce the time it takes
for perf-report to finish its processing. It trades the presence of
trailing stack information with faster speed.
The following table shows the elapsed time of doing perf-report on a
perf.data file of size 985,531,828 bytes.
--max_stack Elapsed Time Output data size
----------- ------------ ----------------
not set 88.0s 124,422,651
64 87.5s 116,303,213
32 87.2s 112,023,804
16 86.6s 94,326,380
8 59.9s 33,697,248
4 40.7s 10,116,637
-g none 27.1s 2,555,810
Signed-off-by: Waiman Long <[email protected]>
Acked-by: David Ahern <[email protected]>
---
tools/perf/Documentation/perf-report.txt | 8 ++++++++
tools/perf/builtin-report.c | 22 +++++++++++++++++-----
tools/perf/builtin-top.c | 3 ++-
tools/perf/util/machine.c | 14 +++++++++-----
tools/perf/util/machine.h | 3 ++-
tools/perf/util/session.c | 3 ++-
6 files changed, 40 insertions(+), 13 deletions(-)
diff --git a/tools/perf/Documentation/perf-report.txt b/tools/perf/Documentation/perf-report.txt
index 2b8097e..be3f196 100644
--- a/tools/perf/Documentation/perf-report.txt
+++ b/tools/perf/Documentation/perf-report.txt
@@ -135,6 +135,14 @@ OPTIONS
Default: fractal,0.5,callee,function.
+--max-stack::
+ Set the stack depth limit when parsing the callchain, anything
+ beyond the specified depth will be ignored. This is a trade-off
+ between information loss and faster processing especially for
+ workloads that can have a very long callchain stack.
+
+ Default: 127
+
-G::
--inverted::
alias for inverted caller based call graph.
diff --git a/tools/perf/builtin-report.c b/tools/perf/builtin-report.c
index 72eae74..d0c9504 100644
--- a/tools/perf/builtin-report.c
+++ b/tools/perf/builtin-report.c
@@ -47,6 +47,7 @@ struct perf_report {
bool show_threads;
bool inverted_callchain;
bool mem_mode;
+ int max_stack;
struct perf_read_values show_threads_values;
const char *pretty_printing_style;
const char *cpu_list;
@@ -88,7 +89,8 @@ static int perf_report__add_mem_hist_entry(struct perf_tool *tool,
if ((sort__has_parent || symbol_conf.use_callchain) &&
sample->callchain) {
err = machine__resolve_callchain(machine, evsel, al->thread,
- sample, &parent, al);
+ sample, &parent, al,
+ rep->max_stack);
if (err)
return err;
}
@@ -179,7 +181,8 @@ static int perf_report__add_branch_hist_entry(struct perf_tool *tool,
if ((sort__has_parent || symbol_conf.use_callchain)
&& sample->callchain) {
err = machine__resolve_callchain(machine, evsel, al->thread,
- sample, &parent, al);
+ sample, &parent, al,
+ rep->max_stack);
if (err)
return err;
}
@@ -242,18 +245,21 @@ out:
return err;
}
-static int perf_evsel__add_hist_entry(struct perf_evsel *evsel,
+static int perf_evsel__add_hist_entry(struct perf_tool *tool,
+ struct perf_evsel *evsel,
struct addr_location *al,
struct perf_sample *sample,
struct machine *machine)
{
+ struct perf_report *rep = container_of(tool, struct perf_report, tool);
struct symbol *parent = NULL;
int err = 0;
struct hist_entry *he;
if ((sort__has_parent || symbol_conf.use_callchain) && sample->callchain) {
err = machine__resolve_callchain(machine, evsel, al->thread,
- sample, &parent, al);
+ sample, &parent, al,
+ rep->max_stack);
if (err)
return err;
}
@@ -330,7 +336,8 @@ static int process_sample_event(struct perf_tool *tool,
if (al.map != NULL)
al.map->dso->hit = 1;
- ret = perf_evsel__add_hist_entry(evsel, &al, sample, machine);
+ ret = perf_evsel__add_hist_entry(tool, evsel, &al, sample,
+ machine);
if (ret < 0)
pr_debug("problem incrementing symbol period, skipping event\n");
}
@@ -757,6 +764,7 @@ int cmd_report(int argc, const char **argv, const char *prefix __maybe_unused)
.ordered_samples = true,
.ordering_requires_timestamps = true,
},
+ .max_stack = PERF_MAX_STACK_DEPTH,
.pretty_printing_style = "normal",
};
const struct option options[] = {
@@ -797,6 +805,10 @@ int cmd_report(int argc, const char **argv, const char *prefix __maybe_unused)
OPT_CALLBACK_DEFAULT('g', "call-graph", &report, "output_type,min_percent[,print_limit],call_order",
"Display callchains using output_type (graph, flat, fractal, or none) , min percent threshold, optional print limit, callchain order, key (function or address). "
"Default: fractal,0.5,callee,function", &parse_callchain_opt, callchain_default_opt),
+ OPT_INTEGER(0, "max-stack", &report.max_stack,
+ "Set the maximum stack depth when parsing the callchain, "
+ "anything beyond the specified depth will be ignored. "
+ "Default: " __stringify(PERF_MAX_STACK_DEPTH)),
OPT_BOOLEAN('G', "inverted", &report.inverted_callchain,
"alias for inverted call graph"),
OPT_CALLBACK(0, "ignore-callees", NULL, "regex",
diff --git a/tools/perf/builtin-top.c b/tools/perf/builtin-top.c
index 2122141..2725aca 100644
--- a/tools/perf/builtin-top.c
+++ b/tools/perf/builtin-top.c
@@ -771,7 +771,8 @@ static void perf_event__process_sample(struct perf_tool *tool,
sample->callchain) {
err = machine__resolve_callchain(machine, evsel,
al.thread, sample,
- &parent, &al);
+ &parent, &al,
+ PERF_MAX_STACK_DEPTH);
if (err)
return;
}
diff --git a/tools/perf/util/machine.c b/tools/perf/util/machine.c
index 6188d28..9617c4a 100644
--- a/tools/perf/util/machine.c
+++ b/tools/perf/util/machine.c
@@ -1267,10 +1267,12 @@ static int machine__resolve_callchain_sample(struct machine *machine,
struct thread *thread,
struct ip_callchain *chain,
struct symbol **parent,
- struct addr_location *root_al)
+ struct addr_location *root_al,
+ int max_stack)
{
u8 cpumode = PERF_RECORD_MISC_USER;
- unsigned int i;
+ int chain_nr = min(max_stack, (int)chain->nr);
+ int i;
int err;
callchain_cursor_reset(&callchain_cursor);
@@ -1280,7 +1282,7 @@ static int machine__resolve_callchain_sample(struct machine *machine,
return 0;
}
- for (i = 0; i < chain->nr; i++) {
+ for (i = 0; i < chain_nr; i++) {
u64 ip;
struct addr_location al;
@@ -1352,12 +1354,14 @@ int machine__resolve_callchain(struct machine *machine,
struct thread *thread,
struct perf_sample *sample,
struct symbol **parent,
- struct addr_location *root_al)
+ struct addr_location *root_al,
+ int max_stack)
{
int ret;
ret = machine__resolve_callchain_sample(machine, thread,
- sample->callchain, parent, root_al);
+ sample->callchain, parent,
+ root_al, max_stack);
if (ret)
return ret;
diff --git a/tools/perf/util/machine.h b/tools/perf/util/machine.h
index 58a6be1..d09cce0 100644
--- a/tools/perf/util/machine.h
+++ b/tools/perf/util/machine.h
@@ -91,7 +91,8 @@ int machine__resolve_callchain(struct machine *machine,
struct thread *thread,
struct perf_sample *sample,
struct symbol **parent,
- struct addr_location *root_al);
+ struct addr_location *root_al,
+ int max_stack);
/*
* Default guest kernel is defined by parameter --guestkallsyms
diff --git a/tools/perf/util/session.c b/tools/perf/util/session.c
index 568b750..96e5449 100644
--- a/tools/perf/util/session.c
+++ b/tools/perf/util/session.c
@@ -1525,7 +1525,8 @@ void perf_evsel__print_ip(struct perf_evsel *evsel, union perf_event *event,
if (symbol_conf.use_callchain && sample->callchain) {
if (machine__resolve_callchain(machine, evsel, al.thread,
- sample, NULL, NULL) != 0) {
+ sample, NULL, NULL,
+ PERF_MAX_STACK_DEPTH) != 0) {
if (verbose)
error("Failed to resolve callchain. Skipping\n");
return;
--
1.7.1
When callgraph is enabled, the append_chain() function consumes a major
portion of the total CPU time. This patch tries to streamline the
append_chain() function by removing unneeded conditional test as well as
using ?: statement which can be more efficient than the regular if
statement in some architectures.
Signed-off-by: Waiman Long <[email protected]>
---
tools/perf/util/callchain.c | 9 ++++-----
1 files changed, 4 insertions(+), 5 deletions(-)
diff --git a/tools/perf/util/callchain.c b/tools/perf/util/callchain.c
index 482f680..1e79001 100644
--- a/tools/perf/util/callchain.c
+++ b/tools/perf/util/callchain.c
@@ -315,6 +315,7 @@ append_chain(struct callchain_node *root,
struct callchain_list *cnode;
u64 start = cursor->pos;
bool found = false;
+ bool func_mode = (callchain_param.key == CCKEY_FUNCTION);
u64 matches;
/*
@@ -331,17 +332,15 @@ append_chain(struct callchain_node *root,
if (!node)
break;
- sym = node->sym;
+ sym = func_mode ? node->sym : NULL;
- if (cnode->ms.sym && sym &&
- callchain_param.key == CCKEY_FUNCTION) {
+ if (cnode->ms.sym && sym) {
if (cnode->ms.sym->start != sym->start)
break;
} else if (cnode->ip != node->ip)
break;
- if (!found)
- found = true;
+ found = true;
callchain_cursor_advance(cursor);
}
--
1.7.1
Hi Waiman,
On Mon, 21 Oct 2013 11:03:36 -0400, Waiman Long wrote:
> v2->v3:
> - Fix the commit log in the second patch which causes "git am" to fail.
> - Remove the 1st patch in v2 as the problem will be fixed by another
> patch from Arnaldo.
>
> v1->v2:
> - Include a compilation fix patch and a code streamlining patch
> into the patch set.
> - Use the __stringify() macro in stringify.h instead of adding a
> duplicate macro.
> - Add the --max-stack option to perf-top as well.
>
> This perf patch set contains the following changes:
>
> Patch 1 - Streamline the append_chain() function to make it run a bit
> faster.
> Patch 2 - Add a --max-stack option to perf-report to speed up its
> processing at the expense of less backtrace information
> available.
> Patch 3 - Add a similar --max-stack option to perf-top.
At first, thank you for tackling this. It was really a problem using
callchain on non trivial perf session. But there's another effort to
address this problem in other direction. You might want to see below
link [1]. Patch 1/8 does the thing and it seems Arnaldo already merged
it into his tree.
Arnaldo, is there any issue prevents others from being merged?
[1] https://lkml.org/lkml/2013/10/11/16
Thanks,
Namhyung
On 10/22/2013 02:07 PM, Namhyung Kim wrote:
> Hi Waiman,
>
> On Mon, 21 Oct 2013 11:03:36 -0400, Waiman Long wrote:
>> v2->v3:
>> - Fix the commit log in the second patch which causes "git am" to fail.
>> - Remove the 1st patch in v2 as the problem will be fixed by another
>> patch from Arnaldo.
>>
>> v1->v2:
>> - Include a compilation fix patch and a code streamlining patch
>> into the patch set.
>> - Use the __stringify() macro in stringify.h instead of adding a
>> duplicate macro.
>> - Add the --max-stack option to perf-top as well.
>>
>> This perf patch set contains the following changes:
>>
>> Patch 1 - Streamline the append_chain() function to make it run a bit
>> faster.
>> Patch 2 - Add a --max-stack option to perf-report to speed up its
>> processing at the expense of less backtrace information
>> available.
>> Patch 3 - Add a similar --max-stack option to perf-top.
> At first, thank you for tackling this. It was really a problem using
> callchain on non trivial perf session. But there's another effort to
> address this problem in other direction. You might want to see below
> link [1]. Patch 1/8 does the thing and it seems Arnaldo already merged
> it into his tree.
Thank for letting me aware of your patch to address the perf-report
problem. As I am working on resolving scalability problem in large NUMA
machines, I often encounters large perf.data files that took more than
1/2 hour for perf-report to produce anything. It is kind of frustrating
to wait so long. I like the performance improvement in your patch 1/8,
coupled with the max-stack option, it can really speed thing up.
-Longman