Re: [PATCH 04/13] perf env: Introduce read_cpu_topology_map() method
From: Arnaldo Carvalho de Melo
Date: Fri Sep 11 2015 - 11:33:40 EST
Em Fri, Sep 11, 2015 at 11:40:28AM -0300, Arnaldo Carvalho de Melo escreveu:
> So this is a problem before and after my patches, i.e. If I go on and
> do, with what we have in acme/perf/core, i.e. none of the changes I'm
> playing with in perf/env:
>
> $ git remote update acme
> Fetching acme
> $ git checkout -b tmp acme/perf/core
> Branch tmp set up to track remote branch perf/core from acme.
> Switched to a new branch 'tmp'
> $ git log --oneline | head -5
> 7e150fb33a91 perf tests: Introduce iterator function for tests
> 1765d9b26f84 perf test: Add entry for hists socket filter
> 207bb55e9193 perf hists browser: Zoom in/out for processor socket
> 9a2843a5f421 perf report: Introduce --socket-filter option
> 99851c76436a perf tools: Introduce new sort type "socket" for the processor socket
> $ rm -rf /tmp/build/perf ; mkdir -p /tmp/build/perf ; make DEBUG=1 -C tools/perf O=/tmp/build/perf install-bin
> # echo 0 > /sys/devices/system/cpu/cpu2/online
> $ cat /sys/devices/system/cpu/cpu2/topology/core_id
> cat: /sys/devices/system/cpu/cpu2/topology/core_id: No such file or directory
> $ ls -la /sys/devices/system/cpu/cpu2/topology/
> ls: cannot access /sys/devices/system/cpu/cpu2/topology/: No such file or directory
> $ perf record usleep 1
> [ perf record: Woken up 1 times to write data ]
> [ perf record: Captured and wrote 0.012 MB perf.data (7 samples) ]
> [acme@felicio linux]$ perf report --header-only -I
> # ========
> # captured on: Fri Sep 11 11:34:18 2015
> # hostname : felicio.ghostprotocols.net
> # os release : 4.2.0
> # perf version : 4.2.g7e150f
> # arch : x86_64
> # nrcpus online : 4
> # nrcpus avail : 3
> # cpudesc : Intel(R) Core(TM) i5-2400 CPU @ 3.10GHz
> <SNIP>
> # node0 meminfo : total = 8085412 kB, free = 5317596 kB
> # node0 cpu list : 0-1,3
> <SNIP>
> $
Stephane, Namhyung, the bug report below is about the online/avail
above, they are swapped, full explanation below.
> We can see multiple bugs here, right? online/avail is swapped, and when
> online != avail we simply do not record the cpu topology info at all!
> If we get that CPU back online:
>
> # echo 1 > /sys/devices/system/cpu/cpu2/online
>
> Then all works:
>
> $ perf record usleep 1
> [ perf record: Woken up 1 times to write data ]
> [ perf record: Captured and wrote 0.012 MB perf.data (7 samples) ]
> $ perf report --header-only -I
> # ========
> # captured on: Fri Sep 11 11:37:31 2015
> # hostname : felicio.ghostprotocols.net
> # os release : 4.2.0
> # perf version : 4.2.g7e150f
> # arch : x86_64
> # nrcpus online : 4
> # nrcpus avail : 4
> # cpudesc : Intel(R) Core(TM) i5-2400 CPU @ 3.10GHz
> <SNIP>
> # sibling cores : 0-3
> # sibling threads : 0
> # sibling threads : 1
> # sibling threads : 2
> # sibling threads : 3
> # CPU 0: Core ID 0, Socket ID 0
> # CPU 1: Core ID 1, Socket ID 0
> # CPU 2: Core ID 2, Socket ID 0
> # CPU 3: Core ID 3, Socket ID 0
> # node0 meminfo : total = 8085412 kB, free = 5316992 kB
> # node0 cpu list : 0-3
>
> So, again, this was not introduced by this patchkit, but it is good that you did
> these offline tests, so we can fix it!
Stephane, this was introduced in:
commit fbe96f29ce4b33e0a22219cc7f5996d9157717e3
Author: Stephane Eranian <eranian@xxxxxxxxxx>
Date: Fri Sep 30 15:40:40 2011 +0200
perf tools: Make perf.data more self-descriptive (v8)
------------------
When you write this part:
- HEADER_NRCPUS: number of online/avail cpus
You do:
static int write_nrcpus(int fd, struct perf_header *h __used,
struct perf_evlist *evlist __used)
{
long nr;
u32 nrc, nra;
int ret;
nr = sysconf(_SC_NPROCESSORS_CONF);
if (nr < 0)
return -1;
nrc = (u32)(nr & UINT_MAX);
nr = sysconf(_SC_NPROCESSORS_ONLN);
if (nr < 0)
return -1;
nra = (u32)(nr & UINT_MAX);
ret = do_write(fd, &nrc, sizeof(nrc));
if (ret < 0)
return ret;
return do_write(fd, &nra, sizeof(nra));
}
I.e. write what you called 'nrc' using what is in SC_NRPROCESSORS_CONF, that in
the documentation for glibc reads:
---------------------
http://www.gnu.org/software/libc/manual/html_node/Processor-Resources.html
sysconf (_SC_NPROCESSORS_CONF)
which returns the number of processors the operating system configured. But it
might be possible for the operating system to disable individual processors and
so the call
---------------------
Which menas "NR_AVAILABLE", right?
But then you call a variable 'nra' which sounds like you think that what is in
_SC_NPROCESSORS_ONLN is the "available" number of CPUs, which is confused a bit
more by the glibc docs when refering to _SC_NPROCESSORS_ONLN:
---------------------
sysconf (_SC_NPROCESSORS_ONLN)
returns the number of processors which are currently online (i.e., available).
---------------------
Then, when printing the number of CPUs encoded in the perf.data file by the
above write_nrcpus() routine, you did:
static void print_nrcpus(struct perf_header *ph, int fd, FILE *fp)
{
ssize_t ret;
u32 nr;
ret = read(fd, &nr, sizeof(nr));
if (ret != (ssize_t)sizeof(nr))
nr = -1; /* interpreted as error */
if (ph->needs_swap)
nr = bswap_32(nr);
fprintf(fp, "# nrcpus online : %u\n", nr);
ret = read(fd, &nr, sizeof(nr));
if (ret != (ssize_t)sizeof(nr))
nr = -1; /* interpreted as error */
if (ph->needs_swap)
nr = bswap_32(nr);
fprintf(fp, "# nrcpus avail : %u\n", nr);
}
You inverted it, no?
So, could you please check if the below patch can have your Acked-by?
Namhyung?
Thanks,
- Arnaldo
diff --git a/tools/perf/util/header.c b/tools/perf/util/header.c
index 41814547da15..fce6634aebe2 100644
--- a/tools/perf/util/header.c
+++ b/tools/perf/util/header.c
@@ -1438,7 +1438,7 @@ static int process_nrcpus(struct perf_file_section *section __maybe_unused,
if (ph->needs_swap)
nr = bswap_32(nr);
- ph->env.nr_cpus_online = nr;
+ ph->env.nr_cpus_avail = nr;
ret = readn(fd, &nr, sizeof(nr));
if (ret != sizeof(nr))
@@ -1447,7 +1447,7 @@ static int process_nrcpus(struct perf_file_section *section __maybe_unused,
if (ph->needs_swap)
nr = bswap_32(nr);
- ph->env.nr_cpus_avail = nr;
+ ph->env.nr_cpus_online = nr;
return 0;
}
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/