[PATCH 35/37] perf evlist: Use unshare(CLONE_FS) in sb threads to let setns(CLONE_NEWNS) work
From: Arnaldo Carvalho de Melo
Date: Thu Aug 29 2019 - 10:41:35 EST
From: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>
When we started using a thread to catch the PERF_RECORD_BPF_EVENT meta
data events to then ask the kernel for further info (BTF, etc) for BPF
programs shortly after they get loaded, we forgot to use
unshare(CLONE_FS) as was done in:
868a832918f6 ("perf top: Support lookup of symbols in other mount namespaces.")
Do it so that we can enter the namespaces to read the build-ids at the
end of a 'perf record' session for the DSOs that had hits.
Before:
Starting a 'stress-ng --cpus 8' inside a container and then, outside the
container running:
# perf record -a --namespaces sleep 5
# perf buildid-list | grep stress-ng
#
We would end up with a 'perf.data' file that had no entry in its
build-id table for the /usr/bin/stress-ng binary inside the container
that got tons of PERF_RECORD_SAMPLEs.
After:
# perf buildid-list | grep stress-ng
f2ed02c68341183a124b9b0f6e2e6c493c465b29 /usr/bin/stress-ng
#
Then its just a matter of making sure that that binary debuginfo package
gets available in a place that 'perf report' will look at build-id keyed
ELF files, which, in my case, on a f30 notebook, was a matter of
installing the debuginfo file for the distro used in the container,
fedora 31:
# rpm -ivh http://fedora.c3sl.ufpr.br/linux/development/31/Everything/x86_64/debug/tree/Packages/s/stress-ng-debuginfo-0.07.29-10.fc31.x86_64.rpm
Then, because perf currently looks for those debuginfo files (richer ELF
symtab) inside that namespace (look at the setns calls):
openat(AT_FDCWD, "/proc/self/ns/mnt", O_RDONLY) = 137
openat(AT_FDCWD, "/proc/13169/ns/mnt", O_RDONLY) = 139
setns(139, CLONE_NEWNS) = 0
stat("/usr/bin/stress-ng", {st_mode=S_IFREG|0755, st_size=3065416, ...}) = 0
openat(AT_FDCWD, "/usr/bin/stress-ng", O_RDONLY) = 140
fcntl(140, F_GETFD) = 0
fstat(140, {st_mode=S_IFREG|0755, st_size=3065416, ...}) = 0
mmap(NULL, 3065416, PROT_READ, MAP_PRIVATE, 140, 0) = 0x7ff2fdc5b000
munmap(0x7ff2fdc5b000, 3065416) = 0
close(140) = 0
stat("stress-ng-0.07.29-10.fc31.x86_64.debug", 0x7fff45d71260) = -1 ENOENT (No such file or directory)
stat("/usr/bin/stress-ng-0.07.29-10.fc31.x86_64.debug", 0x7fff45d71260) = -1 ENOENT (No such file or directory)
stat("/usr/bin/.debug/stress-ng-0.07.29-10.fc31.x86_64.debug", 0x7fff45d71260) = -1 ENOENT (No such file or directory)
stat("/usr/lib/debug/usr/bin/stress-ng-0.07.29-10.fc31.x86_64.debug", 0x7fff45d71260) = -1 ENOENT (No such file or directory)
stat("/root/.debug/.build-id/f2/ed02c68341183a124b9b0f6e2e6c493c465b29", 0x7fff45d711e0) = -1 ENOENT (No such file or directory)
To only then go back to the "host" namespace to look just in the users's
~/.debug cache:
setns(137, CLONE_NEWNS) = 0
chdir("/root") = 0
close(137) = 0
close(139) = 0
stat("/root/.debug/.build-id/f2/ed02c68341183a124b9b0f6e2e6c493c465b29/elf", 0x7fff45d732e0) = -1 ENOENT (No such file or directory)
It continues to fail to resolve symbols:
# perf report | grep stress-ng | head -5
9.50% stress-ng-cpu stress-ng [.] 0x0000000000021ac1
8.58% stress-ng-cpu stress-ng [.] 0x0000000000021ab4
8.51% stress-ng-cpu stress-ng [.] 0x0000000000021489
7.17% stress-ng-cpu stress-ng [.] 0x00000000000219b6
3.93% stress-ng-cpu stress-ng [.] 0x0000000000021478
#
To overcome that we use:
# perf buildid-cache -v --add /usr/lib/debug/usr/bin/stress-ng-0.07.29-10.fc31.x86_64.debug
Adding f2ed02c68341183a124b9b0f6e2e6c493c465b29 /usr/lib/debug/usr/bin/stress-ng-0.07.29-10.fc31.x86_64.debug: Ok
#
# ls -la /root/.debug/.build-id/f2/ed02c68341183a124b9b0f6e2e6c493c465b29/elf
-rw-r--r--. 3 root root 2401184 Jul 27 07:03 /root/.debug/.build-id/f2/ed02c68341183a124b9b0f6e2e6c493c465b29/elf
# file /root/.debug/.build-id/f2/ed02c68341183a124b9b0f6e2e6c493c465b29/elf
/root/.debug/.build-id/f2/ed02c68341183a124b9b0f6e2e6c493c465b29/elf: ELF 64-bit LSB shared object, x86-64, version 1 (SYSV), dynamically linked, interpreter \004, BuildID[sha1]=f2ed02c68341183a124b9b0f6e2e6c493c465b29, for GNU/Linux 3.2.0, with debug_info, not stripped, too many notes (256)
#
Now it finally works:
# perf report | grep stress-ng | head -5
23.59% stress-ng-cpu stress-ng [.] ackermann
23.33% stress-ng-cpu stress-ng [.] is_prime
17.36% stress-ng-cpu stress-ng [.] stress_cpu_sieve
6.08% stress-ng-cpu stress-ng [.] stress_cpu_correlate
3.55% stress-ng-cpu stress-ng [.] queens_try
#
I'll make sure that it looks for the build-id keyed files in both the
"host" namespace (the namespace the user running 'perf record' was a the
time of the recording) and in the container namespace, as it shouldn't
matter where a content based key lookup finds the ELF file to use in
resolving symbols, etc.
Reported-by: Karl Rister <krister@xxxxxxxxxx>
Cc: Alexander Shishkin <alexander.shishkin@xxxxxxxxxxxxxxx>
Cc: Alexei Starovoitov <ast@xxxxxxxxxx>
Cc: Brendan Gregg <brendan.d.gregg@xxxxxxxxx>
Cc: Daniel Borkmann <daniel@xxxxxxxxxxxxx>
Cc: Krister Johansen <kjlx@xxxxxxxxxxxxxxxxxx>
Cc: Namhyung Kim <namhyung@xxxxxxxxxx>
Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
Cc: Song Liu <songliubraving@xxxxxx>
Cc: Stanislav Fomichev <sdf@xxxxxxxxxx>
Cc: Thomas-Mich Richter <tmricht@xxxxxxxxxxxxxxxxxx>
Fixes: 657ee5531903 ("perf evlist: Introduce side band thread")
Link: https://lkml.kernel.org/n/tip-g79k0jz41adiaeuqud742t2l@xxxxxxxxxxxxxx
Signed-off-by: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>
---
tools/perf/util/evlist.c | 9 +++++++++
1 file changed, 9 insertions(+)
diff --git a/tools/perf/util/evlist.c b/tools/perf/util/evlist.c
index 5ad92fa72e78..253dd8dd0e12 100644
--- a/tools/perf/util/evlist.c
+++ b/tools/perf/util/evlist.c
@@ -21,6 +21,7 @@
#include "bpf-event.h"
#include <signal.h>
#include <unistd.h>
+#include <sched.h>
#include "parse-events.h"
#include <subcmd/parse-options.h>
@@ -1824,6 +1825,14 @@ static void *perf_evlist__poll_thread(void *arg)
struct evlist *evlist = arg;
bool draining = false;
int i, done = 0;
+ /*
+ * In order to read symbols from other namespaces perf to needs to call
+ * setns(2). This isn't permitted if the struct_fs has multiple users.
+ * unshare(2) the fs so that we may continue to setns into namespaces
+ * that we're observing when, for instance, reading the build-ids at
+ * the end of a 'perf record' session.
+ */
+ unshare(CLONE_FS);
while (!done) {
bool got_data = false;
--
2.21.0