If using parallel threads to collect data, perf record needs at least 6 fds
per CPU. (one for sys_perf_event_open, four for pipe msg and ack of the
pipe, see record__thread_data_open_pipes(), and one for open perf.data.XXX)
For an environment with more than 100 cores, if perf record uses both
`-a` and `--threads` options, it is easy to exceed the upper limit of the
file descriptor number, when we run out of them try to increase the limits.
Before:
$ ulimit -n
1024
$ lscpu | grep 'On-line CPU(s)'
On-line CPU(s) list: 0-159
$ perf record --threads -a sleep 1
Failed to create data directory: Too many open files
After:
$ ulimit -n
1024
$ lscpu | grep 'On-line CPU(s)'
On-line CPU(s) list: 0-159
$ perf record --threads -a sleep 1
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.394 MB perf.data (1576 samples) ]
Signed-off-by: Yang Jihong <[email protected]>
---
tools/perf/util/data.c | 12 ++++++++++++
1 file changed, 12 insertions(+)
diff --git a/tools/perf/util/data.c b/tools/perf/util/data.c
index fc16299c915f..098f9e3bb2e7 100644
--- a/tools/perf/util/data.c
+++ b/tools/perf/util/data.c
@@ -17,6 +17,7 @@
#include "util.h" // rm_rf_perf_data()
#include "debug.h"
#include "header.h"
+#include "evsel.h"
#include <internal/lib.h>
static void close_dir(struct perf_data_file *files, int nr)
@@ -35,6 +36,7 @@ void perf_data__close_dir(struct perf_data *data)
int perf_data__create_dir(struct perf_data *data, int nr)
{
+ enum rlimit_action set_rlimit = NO_CHANGE;
struct perf_data_file *files = NULL;
int i, ret;
@@ -54,11 +56,21 @@ int perf_data__create_dir(struct perf_data *data, int nr)
goto out_err;
}
+retry_open:
ret = open(file->path, O_RDWR|O_CREAT|O_TRUNC, S_IRUSR|S_IWUSR);
if (ret < 0) {
+ /*
+ * If using parallel threads to collect data,
+ * perf record needs at least 6 fds per CPU.
+ * When we run out of them try to increase the limits.
+ */
+ if (errno == EMFILE && evsel__increase_rlimit(&set_rlimit))
+ goto retry_open;
+
ret = -errno;
goto out_err;
}
+ set_rlimit = NO_CHANGE;
file->fd = ret;
}
--
2.34.1
Hello,
On Fri, Oct 13, 2023 at 1:01 AM Yang Jihong <[email protected]> wrote:
>
> If using parallel threads to collect data, perf record needs at least 6 fds
> per CPU. (one for sys_perf_event_open, four for pipe msg and ack of the
> pipe, see record__thread_data_open_pipes(), and one for open perf.data.XXX)
Yep, probably one more for the dummy event.
> For an environment with more than 100 cores, if perf record uses both
> `-a` and `--threads` options, it is easy to exceed the upper limit of the
> file descriptor number, when we run out of them try to increase the limits.
>
> Before:
> $ ulimit -n
> 1024
> $ lscpu | grep 'On-line CPU(s)'
> On-line CPU(s) list: 0-159
> $ perf record --threads -a sleep 1
> Failed to create data directory: Too many open files
>
> After:
> $ ulimit -n
> 1024
> $ lscpu | grep 'On-line CPU(s)'
> On-line CPU(s) list: 0-159
> $ perf record --threads -a sleep 1
> [ perf record: Woken up 1 times to write data ]
> [ perf record: Captured and wrote 0.394 MB perf.data (1576 samples) ]
> Signed-off-by: Yang Jihong <[email protected]>
> ---
> tools/perf/util/data.c | 12 ++++++++++++
> 1 file changed, 12 insertions(+)
>
> diff --git a/tools/perf/util/data.c b/tools/perf/util/data.c
> index fc16299c915f..098f9e3bb2e7 100644
> --- a/tools/perf/util/data.c
> +++ b/tools/perf/util/data.c
> @@ -17,6 +17,7 @@
> #include "util.h" // rm_rf_perf_data()
> #include "debug.h"
> #include "header.h"
> +#include "evsel.h"
> #include <internal/lib.h>
>
> static void close_dir(struct perf_data_file *files, int nr)
> @@ -35,6 +36,7 @@ void perf_data__close_dir(struct perf_data *data)
>
> int perf_data__create_dir(struct perf_data *data, int nr)
> {
> + enum rlimit_action set_rlimit = NO_CHANGE;
> struct perf_data_file *files = NULL;
> int i, ret;
>
> @@ -54,11 +56,21 @@ int perf_data__create_dir(struct perf_data *data, int nr)
> goto out_err;
> }
>
> +retry_open:
> ret = open(file->path, O_RDWR|O_CREAT|O_TRUNC, S_IRUSR|S_IWUSR);
> if (ret < 0) {
> + /*
> + * If using parallel threads to collect data,
> + * perf record needs at least 6 fds per CPU.
> + * When we run out of them try to increase the limits.
> + */
> + if (errno == EMFILE && evsel__increase_rlimit(&set_rlimit))
It seems weird that we have this helper with evsel prefix and
it does nothing with evsel. But it's a separate concern, so
Acked-by: Namhyung Kim <[email protected]>
Thanks,
Namhyung
> + goto retry_open;
> +
> ret = -errno;
> goto out_err;
> }
> + set_rlimit = NO_CHANGE;
>
> file->fd = ret;
> }
> --
> 2.34.1
>
On Fri, 13 Oct 2023 07:59:45 +0000, Yang Jihong wrote:
> If using parallel threads to collect data, perf record needs at least 6 fds
> per CPU. (one for sys_perf_event_open, four for pipe msg and ack of the
> pipe, see record__thread_data_open_pipes(), and one for open perf.data.XXX)
> For an environment with more than 100 cores, if perf record uses both
> `-a` and `--threads` options, it is easy to exceed the upper limit of the
> file descriptor number, when we run out of them try to increase the limits.
>
> [...]
Applied to perf-tools-next, thanks!
Hello,
On 2023/10/20 13:59, Namhyung Kim wrote:
> Hello,
>
> On Fri, Oct 13, 2023 at 1:01 AM Yang Jihong <[email protected]> wrote:
>>
>> If using parallel threads to collect data, perf record needs at least 6 fds
>> per CPU. (one for sys_perf_event_open, four for pipe msg and ack of the
>> pipe, see record__thread_data_open_pipes(), and one for open perf.data.XXX)
>
> Yep, probably one more for the dummy event.
>
>> For an environment with more than 100 cores, if perf record uses both
>> `-a` and `--threads` options, it is easy to exceed the upper limit of the
>> file descriptor number, when we run out of them try to increase the limits.
>>
>> Before:
>> $ ulimit -n
>> 1024
>> $ lscpu | grep 'On-line CPU(s)'
>> On-line CPU(s) list: 0-159
>> $ perf record --threads -a sleep 1
>> Failed to create data directory: Too many open files
>>
>> After:
>> $ ulimit -n
>> 1024
>> $ lscpu | grep 'On-line CPU(s)'
>> On-line CPU(s) list: 0-159
>> $ perf record --threads -a sleep 1
>> [ perf record: Woken up 1 times to write data ]
>> [ perf record: Captured and wrote 0.394 MB perf.data (1576 samples) ]
>> Signed-off-by: Yang Jihong <[email protected]>
>> ---
>> tools/perf/util/data.c | 12 ++++++++++++
>> 1 file changed, 12 insertions(+)
>>
>> diff --git a/tools/perf/util/data.c b/tools/perf/util/data.c
>> index fc16299c915f..098f9e3bb2e7 100644
>> --- a/tools/perf/util/data.c
>> +++ b/tools/perf/util/data.c
>> @@ -17,6 +17,7 @@
>> #include "util.h" // rm_rf_perf_data()
>> #include "debug.h"
>> #include "header.h"
>> +#include "evsel.h"
>> #include <internal/lib.h>
>>
>> static void close_dir(struct perf_data_file *files, int nr)
>> @@ -35,6 +36,7 @@ void perf_data__close_dir(struct perf_data *data)
>>
>> int perf_data__create_dir(struct perf_data *data, int nr)
>> {
>> + enum rlimit_action set_rlimit = NO_CHANGE;
>> struct perf_data_file *files = NULL;
>> int i, ret;
>>
>> @@ -54,11 +56,21 @@ int perf_data__create_dir(struct perf_data *data, int nr)
>> goto out_err;
>> }
>>
>> +retry_open:
>> ret = open(file->path, O_RDWR|O_CREAT|O_TRUNC, S_IRUSR|S_IWUSR);
>> if (ret < 0) {
>> + /*
>> + * If using parallel threads to collect data,
>> + * perf record needs at least 6 fds per CPU.
>> + * When we run out of them try to increase the limits.
>> + */
>> + if (errno == EMFILE && evsel__increase_rlimit(&set_rlimit))
>
> It seems weird that we have this helper with evsel prefix and
> it does nothing with evsel. But it's a separate concern, so
>
> Acked-by: Namhyung Kim <[email protected]>
Thanks for the Acked-by tag.
Uh... The name of this helper doesn't seem to be appropriate.
Okay, I'll submit a patch to fix this helper name.
Thanks,
Yang