Skip to contentSkip to conversation

perf tests: Test converting perf time to TSC

BUG Jiri Olsa 4 participants 13 messages Oct 2, 2013 – Oct 3, 2013

Subsystems (2)
lore
Share & tools
lore

Jiri Olsa

author
#0 lore raw

hi,
got a segfault in the tsc test on latest acme's tree.

I'm dealing with some other issues right now, so just reporting ;-)

thanks,
jirka

(gdb) r test 20
Starting program: /home/jolsa/linux-perf/tools/perf/perf test 20
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib64/libthread_db.so.1".
20: Test converting perf time to TSC :
Program received signal SIGFPE, Arithmetic exception.
0x00000000004cee73 in perf_time_to_tsc (ns=1460625577603, tc=0x7fffffffdf50) at arch/x86/util/tsc.c:16
16 quot = t / tc->time_mult;
Missing separate debuginfos, use: debuginfo-install audit-libs-2.2.1-1.fc16.x86_64 elfutils-libelf-0.154-2.fc16.x86_64 elfutils-libs-0.154-2.fc16.x86_64 glibc-2.14.90-24.fc16.9.x86_64 nss-softokn-freebl-3.14.1-3.fc16.x86_64 numactl-2.0.7-2.fc16.x86_64 perl-libs-5.14.3-205.fc16.x86_64 python-libs-2.7.3-4.fc16.x86_64 slang-2.2.4-1.fc16.x86_64 xz-libs-5.1.1-1alpha.fc16.x86_64 zlib-1.2.5-7.fc16.x86_64 (gdb) bt

#0  0x00000000004cee73 in perf_time_to_tsc (ns=1460625577603, tc=0x7fffffffdf50) at arch/x86/util/tsc.c:16
#1  0x00000000004cbdb8 in test__perf_time_to_tsc () at tests/perf-time-to-tsc.c:144
#2  0x00000000004491b7 in __cmd_test (argc=1, argv=0x7fffffffe400, skiplist=0x0) at tests/builtin-test.c:175
#3  0x0000000000449512 in cmd_test (argc=1, argv=0x7fffffffe400, prefix=0x0) at tests/builtin-test.c:240
#4  0x0000000000418dfd in run_builtin (p=0x802500, argc=2, argv=0x7fffffffe400) at perf.c:319
#5  0x0000000000419035 in handle_internal_command (argc=2, argv=0x7fffffffe400) at perf.c:376
#6  0x0000000000419181 in run_argv (argcp=0x7fffffffe2dc, argv=0x7fffffffe2d0) at perf.c:420
#7  0x0000000000419468 in main (argc=2, argv=0x7fffffffe400) at perf.c:529

(gdb) p tc->time_mult
$1 = 0

reviewer
#1 lore raw
Quoted 4 lines · Jiri Olsa

hi,
got a segfault in the tsc test on latest acme's tree.

I'm dealing with some other issues right now, so just reporting ;-)

The capability bits have changed positions. You need to have:

commit fa7315871046b9a4c48627905691dbde57e51033

Author: Peter Zijlstra <[email protected]>
Date:   Thu Sep 19 10:16:42 2013 +0200

perf: Fix capabilities bitfield compatibility in 'struct
perf_event_mmap_page'

in both (or neither) of your kernel and perf tools.

Quoted 26 lines

thanks,
jirka

(gdb) r test 20
Starting program: /home/jolsa/linux-perf/tools/perf/perf test 20
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib64/libthread_db.so.1".
20: Test converting perf time to TSC :
Program received signal SIGFPE, Arithmetic exception.
0x00000000004cee73 in perf_time_to_tsc (ns=1460625577603, tc=0x7fffffffdf50) at arch/x86/util/tsc.c:16
16 quot = t / tc->time_mult;
Missing separate debuginfos, use: debuginfo-install audit-libs-2.2.1-1.fc16.x86_64 elfutils-libelf-0.154-2.fc16.x86_64 elfutils-libs-0.154-2.fc16.x86_64 glibc-2.14.90-24.fc16.9.x86_64 nss-softokn-freebl-3.14.1-3.fc16.x86_64 numactl-2.0.7-2.fc16.x86_64 perl-libs-5.14.3-205.fc16.x86_64 python-libs-2.7.3-4.fc16.x86_64 slang-2.2.4-1.fc16.x86_64 xz-libs-5.1.1-1alpha.fc16.x86_64 zlib-1.2.5-7.fc16.x86_64 (gdb) bt

#0  0x00000000004cee73 in perf_time_to_tsc (ns=1460625577603, tc=0x7fffffffdf50) at arch/x86/util/tsc.c:16
#1  0x00000000004cbdb8 in test__perf_time_to_tsc () at tests/perf-time-to-tsc.c:144
#2  0x00000000004491b7 in __cmd_test (argc=1, argv=0x7fffffffe400, skiplist=0x0) at tests/builtin-test.c:175
#3  0x0000000000449512 in cmd_test (argc=1, argv=0x7fffffffe400, prefix=0x0) at tests/builtin-test.c:240
#4  0x0000000000418dfd in run_builtin (p=0x802500, argc=2, argv=0x7fffffffe400) at perf.c:319
#5  0x0000000000419035 in handle_internal_command (argc=2, argv=0x7fffffffe400) at perf.c:376
#6  0x0000000000419181 in run_argv (argcp=0x7fffffffe2dc, argv=0x7fffffffe2d0) at perf.c:420
#7  0x0000000000419468 in main (argc=2, argv=0x7fffffffe400) at perf.c:529

(gdb) p tc->time_mult
$1 = 0

#2 lore raw

Em Wed, Oct 02, 2013 at 04:46:59PM +0300, Adrian Hunter escreveu:

Quoted 8 lines · Jiri Olsa
Quoted 4 lines

hi,
got a segfault in the tsc test on latest acme's tree.

I'm dealing with some other issues right now, so just reporting ;-)

The capability bits have changed positions. You need to have:

Quoted 8 lines

commit fa7315871046b9a4c48627905691dbde57e51033

Author: Peter Zijlstra <[email protected]>
Date:   Thu Sep 19 10:16:42 2013 +0200

perf: Fix capabilities bitfield compatibility in 'struct
perf_event_mmap_page'

in both (or neither) of your kernel and perf tools.

Yeah, I'll merge perf/urgent into acme/perf/core soon.

- Arnaldo

author
#3 lore raw
In reply to Adrian Hunter
Quoted 6 lines · Adrian Hunter
Quoted 4 lines · Jiri Olsa

hi,
got a segfault in the tsc test on latest acme's tree.

I'm dealing with some other issues right now, so just reporting ;-)

Quoted 8 lines

The capability bits have changed positions. You need to have:

commit fa7315871046b9a4c48627905691dbde57e51033

Author: Peter Zijlstra <[email protected]>
Date:   Thu Sep 19 10:16:42 2013 +0200

perf: Fix capabilities bitfield compatibility in 'struct
perf_event_mmap_page'

ok, I'll try that.. but anyway, the test should
not crash in account of missing kernel change

jirka

reviewer
#4 lore raw
Quoted 13 lines · Jiri Olsa
Quoted 12 lines · Adrian Hunter
Quoted 4 lines · Jiri Olsa

hi,
got a segfault in the tsc test on latest acme's tree.

I'm dealing with some other issues right now, so just reporting ;-)

The capability bits have changed positions. You need to have:

commit fa7315871046b9a4c48627905691dbde57e51033

Author: Peter Zijlstra <[email protected]>
Date:   Thu Sep 19 10:16:42 2013 +0200
Quoted 5 lines
Quoted 2 lines

perf: Fix capabilities bitfield compatibility in 'struct
perf_event_mmap_page'

ok, I'll try that.. but anyway, the test should
not crash in account of missing kernel change

No the ABI is broken in that case - better to crash.

maintainer
#5 lore raw
Quoted 14 lines · Adrian Hunter
Quoted 13 lines · Jiri Olsa
Quoted 12 lines · Adrian Hunter
Quoted 4 lines · Jiri Olsa

hi,
got a segfault in the tsc test on latest acme's tree.

I'm dealing with some other issues right now, so just reporting ;-)

The capability bits have changed positions. You need to have:

commit fa7315871046b9a4c48627905691dbde57e51033

Author: Peter Zijlstra <[email protected]>
Date:   Thu Sep 19 10:16:42 2013 +0200
Quoted 7 lines
Quoted 5 lines
Quoted 2 lines

perf: Fix capabilities bitfield compatibility in 'struct
perf_event_mmap_page'

ok, I'll try that.. but anyway, the test should
not crash in account of missing kernel change

No the ABI is broken in that case - better to crash.

No; neither case should crash.

Anyway; looking at this, why does time_zero have these different checks from the other time bits?

@@ -1897,6 +1898,11 @@ void arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)

        userpg->time_mult = this_cpu_read(cyc2ns);
        userpg->time_shift = CYC2NS_SCALE_FACTOR;
        userpg->time_offset = this_cpu_read(cyc2ns_offset) - now;

+
+ if (sched_clock_stable && !check_tsc_disabled()) {
+ userpg->cap_usr_time_zero = 1;
+ userpg->time_zero = this_cpu_read(cyc2ns_offset);
+ }
}

That doesn't make any kind of sense.. why is cyc2ns_offset differently tested from cyc2ns itself?

maintainer
#6 lore raw
Quoted 3 lines · Peter Zijlstra
Quoted 1 line

No the ABI is broken in that case - better to crash.

No; neither case should crash.

OK; previously we would report cap_usr_time = 1 even though we might not have written stuff into the fields; which would indeed explain the crash.

However; Jiri actually has the patch; as the code he posted had 'user' not 'usr'

reviewer
#7 lore raw
Quoted 11 lines · Peter Zijlstra
Quoted 3 lines · Peter Zijlstra
Quoted 1 line

No the ABI is broken in that case - better to crash.

No; neither case should crash.

OK; previously we would report cap_usr_time = 1 even though we might not have written stuff into the fields; which would indeed explain the crash.

However; Jiri actually has the patch; as the code he posted had 'user' not 'usr'

Where do you see that?

maintainer
#8 lore raw
Quoted 1 line · Adrian Hunter

Where do you see that?

my minds eye.. I should go back to sleep or so :-( Its going to be one of those days.

reviewer
#9 lore raw
In reply to Peter Zijlstra
Quoted 39 lines · Peter Zijlstra
Quoted 14 lines · Adrian Hunter
Quoted 13 lines · Jiri Olsa
Quoted 12 lines · Adrian Hunter
Quoted 4 lines · Jiri Olsa

hi,
got a segfault in the tsc test on latest acme's tree.

I'm dealing with some other issues right now, so just reporting ;-)

The capability bits have changed positions. You need to have:

commit fa7315871046b9a4c48627905691dbde57e51033

Author: Peter Zijlstra <[email protected]>
Date:   Thu Sep 19 10:16:42 2013 +0200
Quoted 7 lines
Quoted 5 lines
Quoted 2 lines

perf: Fix capabilities bitfield compatibility in 'struct
perf_event_mmap_page'

ok, I'll try that.. but anyway, the test should
not crash in account of missing kernel change

No the ABI is broken in that case - better to crash.

No; neither case should crash.

Anyway; looking at this, why does time_zero have these different checks from the other time bits?

(quoted diff) +5 -0
@@ -1897,6 +1898,11 @@ void arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
        userpg->time_mult = this_cpu_read(cyc2ns);
        userpg->time_shift = CYC2NS_SCALE_FACTOR;
        userpg->time_offset = this_cpu_read(cyc2ns_offset) - now;
+
+       if (sched_clock_stable && !check_tsc_disabled()) {
+               userpg->cap_usr_time_zero = 1;
+               userpg->time_zero = this_cpu_read(cyc2ns_offset);
+       }
 }
 
Quoted 2 lines

That doesn't make any kind of sense.. why is cyc2ns_offset differently tested from cyc2ns itself?

I am afraid I don't understand the scaling calculations so I don't know if they make any sense.

cap_usr_time_zero (now cap_user_time_zero) means you can convert perf time to / from TSC. That only works if TSC is not disabled and sched_clock is stable (and you have constant, non-stop TSC)

As far as I can tell, assuming the hardware is not broken, sched_clock will be stable unless something (BIOS) or someone (meddling user) has changed TSC manually.

maintainer
#10 lore raw
Quoted 20 lines · Adrian Hunter
Quoted 14 lines

Anyway; looking at this, why does time_zero have these different checks from the other time bits?

(quoted diff) +5 -0
@@ -1897,6 +1898,11 @@ void arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
        userpg->time_mult = this_cpu_read(cyc2ns);
        userpg->time_shift = CYC2NS_SCALE_FACTOR;
        userpg->time_offset = this_cpu_read(cyc2ns_offset) - now;
+
+       if (sched_clock_stable && !check_tsc_disabled()) {
+               userpg->cap_usr_time_zero = 1;
+               userpg->time_zero = this_cpu_read(cyc2ns_offset);
+       }
 }
 
Quoted 2 lines

That doesn't make any kind of sense.. why is cyc2ns_offset differently tested from cyc2ns itself?

I am afraid I don't understand the scaling calculations so I don't know if they make any sense.

Quoted 7 lines

cap_usr_time_zero (now cap_user_time_zero) means you can convert perf time to / from TSC. That only works if TSC is not disabled and sched_clock is stable (and you have constant, non-stop TSC)

As far as I can tell, assuming the hardware is not broken, sched_clock will be stable unless something (BIOS) or someone (meddling user) has changed TSC manually.

Well all the scaling mess only applies to TSC.. furthermore note how time_offset even uses cyc2ns_offset.

So I don't see any point in cap_user_time and cap_user_time_zero having different conditions.

Also, I don't think that a cpu without cap_tsc will ever run perf code so I don't see the point of testing tsc_disabled; if someone is mad enough to boot with that he can bloody well keep the pieces.

Arguably the CONSTANT && NONSTOP test was to test the same thing as sched_clock_stable, but I suppose we can use that.

Which reminds me; I should go and fix the transition for
sched_clock_stable: 1->0.

---

arch/x86/kernel/cpu/perf_event.c +4 -9
index 897783b3302a..9db2b361a63d 100644
@@ -1888,21 +1888,16 @@ void arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
 	userpg->cap_user_rdpmc = x86_pmu.attr_rdpmc;
 	userpg->pmc_width = x86_pmu.cntval_bits;
 
-	if (!boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
-		return;
-
-	if (!boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
-		return;
+	if (!sched_clock_stable)
+		return
 
 	userpg->cap_user_time = 1;
 	userpg->time_mult = this_cpu_read(cyc2ns);
 	userpg->time_shift = CYC2NS_SCALE_FACTOR;
 	userpg->time_offset = this_cpu_read(cyc2ns_offset) - now;
 
-	if (sched_clock_stable && !check_tsc_disabled()) {
-		userpg->cap_user_time_zero = 1;
-		userpg->time_zero = this_cpu_read(cyc2ns_offset);
-	}
+	userpg->cap_user_time_zero = 1;
+	userpg->time_zero = this_cpu_read(cyc2ns_offset);
 }
 
 /*
reviewer
#11 lore raw
Quoted 32 lines · Peter Zijlstra
Quoted 20 lines · Adrian Hunter
Quoted 14 lines

Anyway; looking at this, why does time_zero have these different checks from the other time bits?

arch/x86/kernel/cpu/perf_event.c +5 -0
@@ -1897,6 +1898,11 @@ void arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
        userpg->time_mult = this_cpu_read(cyc2ns);
        userpg->time_shift = CYC2NS_SCALE_FACTOR;
        userpg->time_offset = this_cpu_read(cyc2ns_offset) - now;
+
+       if (sched_clock_stable && !check_tsc_disabled()) {
+               userpg->cap_usr_time_zero = 1;
+               userpg->time_zero = this_cpu_read(cyc2ns_offset);
+       }
 }
 
Quoted 2 lines

That doesn't make any kind of sense.. why is cyc2ns_offset differently tested from cyc2ns itself?

I am afraid I don't understand the scaling calculations so I don't know if they make any sense.

Quoted 7 lines

cap_usr_time_zero (now cap_user_time_zero) means you can convert perf time to / from TSC. That only works if TSC is not disabled and sched_clock is stable (and you have constant, non-stop TSC)

As far as I can tell, assuming the hardware is not broken, sched_clock will be stable unless something (BIOS) or someone (meddling user) has changed TSC manually.

Well all the scaling mess only applies to TSC.. furthermore note how time_offset even uses cyc2ns_offset.

Quoted 6 lines

So I don't see any point in cap_user_time and cap_user_time_zero having different conditions.

Also, I don't think that a cpu without cap_tsc will ever run perf code so I don't see the point of testing tsc_disabled; if someone is mad enough to boot with that he can bloody well keep the pieces.

You can boot with notsc and perf still seems to work e.g.

$ dmesg | grep -i tsc
[    0.000000] Command line: BOOT_IMAGE=/boot/vmlinuz-3.11.0+ root=UUID=f1b4c71a-15aa-41a6-8898-cdde49966bce ro ignore_loglevel earlyprintk=ttyS0 console=ttyS0,115200n8 no_console_suspend notsc=1
[    0.000000] Kernel command line: BOOT_IMAGE=/boot/vmlinuz-3.11.0+ root=UUID=f1b4c71a-15aa-41a6-8898-cdde49966bce ro ignore_loglevel earlyprintk=ttyS0 console=ttyS0,115200n8 no_console_suspend notsc=1
[    0.000000] tsc: Kernel compiled with CONFIG_X86_TSC, cannot disable TSC completely
[    0.000000] tsc: Fast TSC calibration using PIT
[    0.004000] tsc: Detected 2593.932 MHz processor
[    0.224000] TSC deadline timer enabled
$ tools/perf/perf record -e instructions:u uname
Linux
[ perf record: Woken up 1 times to write data ]
[ perf record: Captured and wrote 0.032 MB perf.data (~1414 samples) ]
$ tools/perf/perf script
# ========
# captured on: Thu Oct  3 17:31:38 2013
# hostname : ivb01
# os release : 3.11.0+
# perf version : 3.11.rc4.gde8db8
# arch : x86_64
# nrcpus online : 8
# nrcpus avail : 8
# cpudesc : Intel(R) Core(TM) i7-3720QM CPU @ 2.60GHz
# cpuid : GenuineIntel,6,58,9
# total memory : 8074624 kB
# cmdline : /home/ahunter/git/linux-perf/tools/perf/perf record -e instructions:u uname 
# event : name = instructions:u, type = 0, config = 0x1, config1 = 0x0, config2 = 0x0, excl_usr = 0, excl_kern = 1, excl_host = 0, e
# HEADER_CPU_TOPOLOGY info available, use -I to display
# HEADER_NUMA_TOPOLOGY info available, use -I to display
# pmu mappings: cpu = 4, software = 1, tracepoint = 2, uncore_cbox_0 = 6, uncore_cbox_1 = 7, uncore_cbox_2 = 8, uncore_cbox_3 = 9, b
# ========
#
           uname  3914   451.484000: instructions:u:      7fe3f3315af3 [unknown] (/lib/x86_64-linux-gnu/ld-2.13.so)
           uname  3914   451.484000: instructions:u:  ffffffff813a1ad0 page_fault ([kernel.kallsyms])
           uname  3914   451.484000: instructions:u:      7fe3f3316136 _dl_start (/lib/x86_64-linux-gnu/ld-2.13.so)
           uname  3914   451.484000: instructions:u:  ffffffff813a1ad0 page_fault ([kernel.kallsyms])
           uname  3914   451.484000: instructions:u:  ffffffff813a1ad0 page_fault ([kernel.kallsyms])
           uname  3914   451.484000: instructions:u:  ffffffff813a1ad0 page_fault ([kernel.kallsyms])
           uname  3914   451.484000: instructions:u:      7fe3f33161b4 _dl_start (/lib/x86_64-linux-gnu/ld-2.13.so)
           uname  3914   451.484000: instructions:u:      7fe3f33161af _dl_start (/lib/x86_64-linux-gnu/ld-2.13.so)
           uname  3914   451.484000: instructions:u:      7fe3f33161af _dl_start (/lib/x86_64-linux-gnu/ld-2.13.so)
           uname  3914   451.484000: instructions:u:      7fe3f33161af _dl_start (/lib/x86_64-linux-gnu/ld-2.13.so)
           uname  3914   451.484000: instructions:u:      7fe3f33161bd _dl_start (/lib/x86_64-linux-gnu/ld-2.13.so)
           uname  3914   451.484000: instructions:u:      7fe3f3316388 _dl_start (/lib/x86_64-linux-gnu/ld-2.13.so)
           uname  3914   451.484000: instructions:u:      7fe3f3316590 _dl_start (/lib/x86_64-linux-gnu/ld-2.13.so)
           uname  3914   451.484000: instructions:u:      7fe3f33165a0 _dl_start (/lib/x86_64-linux-gnu/ld-2.13.so)
Quoted 40 lines

Arguably the CONSTANT && NONSTOP test was to test the same thing as sched_clock_stable, but I suppose we can use that.

Which reminds me; I should go and fix the transition for
sched_clock_stable: 1->0.

---

arch/x86/kernel/cpu/perf_event.c +4 -9
index 897783b3302a..9db2b361a63d 100644
@@ -1888,21 +1888,16 @@ void arch_perf_update_userpage(struct perf_event_mmap_page *userpg, u64 now)
 	userpg->cap_user_rdpmc = x86_pmu.attr_rdpmc;
 	userpg->pmc_width = x86_pmu.cntval_bits;
 
-	if (!boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
-		return;
-
-	if (!boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
-		return;
+	if (!sched_clock_stable)
+		return
 
 	userpg->cap_user_time = 1;
 	userpg->time_mult = this_cpu_read(cyc2ns);
 	userpg->time_shift = CYC2NS_SCALE_FACTOR;
 	userpg->time_offset = this_cpu_read(cyc2ns_offset) - now;
 
-	if (sched_clock_stable && !check_tsc_disabled()) {
-		userpg->cap_user_time_zero = 1;
-		userpg->time_zero = this_cpu_read(cyc2ns_offset);
-	}
+	userpg->cap_user_time_zero = 1;
+	userpg->time_zero = this_cpu_read(cyc2ns_offset);
 }
 
 /*
maintainer
#12 lore raw
Quoted 10 lines · Adrian Hunter

You can boot with notsc and perf still seems to work e.g.

$ dmesg | grep -i tsc
[    0.000000] Command line: BOOT_IMAGE=/boot/vmlinuz-3.11.0+ root=UUID=f1b4c71a-15aa-41a6-8898-cdde49966bce ro ignore_loglevel earlyprintk=ttyS0 console=ttyS0,115200n8 no_console_suspend notsc=1
[    0.000000] Kernel command line: BOOT_IMAGE=/boot/vmlinuz-3.11.0+ root=UUID=f1b4c71a-15aa-41a6-8898-cdde49966bce ro ignore_loglevel earlyprintk=ttyS0 console=ttyS0,115200n8 no_console_suspend notsc=1
[    0.000000] tsc: Kernel compiled with CONFIG_X86_TSC, cannot disable TSC completely
[    0.000000] tsc: Fast TSC calibration using PIT
[    0.004000] tsc: Detected 2593.932 MHz processor
[    0.224000] TSC deadline timer enabled

Yeah, and we still use the TSC deadline timer too.. I think we should seriously consider removing the notsc option. I really don't know what its good for.

Thomas, Ingo, hpa?